From 72e9d57b15cf19c2be54d93f2ea50b779717b69d Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 14:09:18 +0100
Subject: [PATCH 01/47] Raw import of libfmt.

---
 dep/fmt/.clang-format            |    8 +
 dep/fmt/CMakeLists.txt           |  453 +++
 dep/fmt/CONTRIBUTING.md          |   20 +
 dep/fmt/ChangeLog.md             | 5533 ++++++++++++++++++++++++++++++
 dep/fmt/LICENSE                  |   27 +
 dep/fmt/README.md                |  490 +++
 dep/fmt/UPSTREAM.md              |    2 +
 dep/fmt/include/fmt/args.h       |  235 ++
 dep/fmt/include/fmt/chrono.h     | 2240 ++++++++++++
 dep/fmt/include/fmt/color.h      |  643 ++++
 dep/fmt/include/fmt/compile.h    |  535 +++
 dep/fmt/include/fmt/core.h       | 2969 ++++++++++++++++
 dep/fmt/include/fmt/format-inl.h | 1678 +++++++++
 dep/fmt/include/fmt/format.h     | 4535 ++++++++++++++++++++++++
 dep/fmt/include/fmt/os.h         |  455 +++
 dep/fmt/include/fmt/ostream.h    |  245 ++
 dep/fmt/include/fmt/printf.h     |  675 ++++
 dep/fmt/include/fmt/ranges.h     |  738 ++++
 dep/fmt/include/fmt/std.h        |  537 +++
 dep/fmt/include/fmt/xchar.h      |  259 ++
 dep/fmt/src/fmt.cc               |  108 +
 dep/fmt/src/format.cc            |   43 +
 dep/fmt/src/os.cc                |  402 +++
 23 files changed, 22830 insertions(+)
 create mode 100644 dep/fmt/.clang-format
 create mode 100644 dep/fmt/CMakeLists.txt
 create mode 100644 dep/fmt/CONTRIBUTING.md
 create mode 100644 dep/fmt/ChangeLog.md
 create mode 100644 dep/fmt/LICENSE
 create mode 100644 dep/fmt/README.md
 create mode 100644 dep/fmt/UPSTREAM.md
 create mode 100644 dep/fmt/include/fmt/args.h
 create mode 100644 dep/fmt/include/fmt/chrono.h
 create mode 100644 dep/fmt/include/fmt/color.h
 create mode 100644 dep/fmt/include/fmt/compile.h
 create mode 100644 dep/fmt/include/fmt/core.h
 create mode 100644 dep/fmt/include/fmt/format-inl.h
 create mode 100644 dep/fmt/include/fmt/format.h
 create mode 100644 dep/fmt/include/fmt/os.h
 create mode 100644 dep/fmt/include/fmt/ostream.h
 create mode 100644 dep/fmt/include/fmt/printf.h
 create mode 100644 dep/fmt/include/fmt/ranges.h
 create mode 100644 dep/fmt/include/fmt/std.h
 create mode 100644 dep/fmt/include/fmt/xchar.h
 create mode 100644 dep/fmt/src/fmt.cc
 create mode 100644 dep/fmt/src/format.cc
 create mode 100644 dep/fmt/src/os.cc

diff --git a/dep/fmt/.clang-format b/dep/fmt/.clang-format
new file mode 100644
index 00000000..df55d340
--- /dev/null
+++ b/dep/fmt/.clang-format
@@ -0,0 +1,8 @@
+# Run manually to reformat a file:
+# clang-format -i --style=file <file>
+Language: Cpp
+BasedOnStyle: Google
+IndentPPDirectives: AfterHash
+IndentCaseLabels: false
+AlwaysBreakTemplateDeclarations: false
+DerivePointerAlignment: false
diff --git a/dep/fmt/CMakeLists.txt b/dep/fmt/CMakeLists.txt
new file mode 100644
index 00000000..4b928ae3
--- /dev/null
+++ b/dep/fmt/CMakeLists.txt
@@ -0,0 +1,453 @@
+cmake_minimum_required(VERSION 3.8...3.26)
+
+# Fallback for using newer policies on CMake <3.12.
+if (${CMAKE_VERSION} VERSION_LESS 3.12)
+  cmake_policy(VERSION ${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION})
+endif ()
+
+# Determine if fmt is built as a subproject (using add_subdirectory)
+# or if it is the master project.
+if (NOT DEFINED FMT_MASTER_PROJECT)
+  set(FMT_MASTER_PROJECT OFF)
+  if (CMAKE_CURRENT_SOURCE_DIR STREQUAL CMAKE_SOURCE_DIR)
+    set(FMT_MASTER_PROJECT ON)
+    message(STATUS "CMake version: ${CMAKE_VERSION}")
+  endif ()
+endif ()
+
+# Joins arguments and places the results in ${result_var}.
+function(join result_var)
+  set(result "")
+  foreach (arg ${ARGN})
+    set(result "${result}${arg}")
+  endforeach ()
+  set(${result_var} "${result}" PARENT_SCOPE)
+endfunction()
+
+# DEPRECATED! Should be merged into add_module_library.
+function(enable_module target)
+  if (MSVC)
+    set(BMI ${CMAKE_CURRENT_BINARY_DIR}/${target}.ifc)
+    target_compile_options(${target}
+      PRIVATE /interface /ifcOutput ${BMI}
+      INTERFACE /reference fmt=${BMI})
+    set_target_properties(${target} PROPERTIES ADDITIONAL_CLEAN_FILES ${BMI})
+    set_source_files_properties(${BMI} PROPERTIES GENERATED ON)
+  endif ()
+endfunction()
+
+# Adds a library compiled with C++20 module support.
+# `enabled` is a CMake variables that specifies if modules are enabled.
+# If modules are disabled `add_module_library` falls back to creating a
+# non-modular library.
+#
+# Usage:
+#   add_module_library(<name> [sources...] FALLBACK [sources...] [IF enabled])
+function(add_module_library name)
+  cmake_parse_arguments(AML "" "IF" "FALLBACK" ${ARGN})
+  set(sources ${AML_UNPARSED_ARGUMENTS})
+
+  add_library(${name})
+  set_target_properties(${name} PROPERTIES LINKER_LANGUAGE CXX)
+
+  if (NOT ${${AML_IF}})
+    # Create a non-modular library.
+    target_sources(${name} PRIVATE ${AML_FALLBACK})
+    return()
+  endif ()
+
+  # Modules require C++20.
+  target_compile_features(${name} PUBLIC cxx_std_20)
+  if (CMAKE_COMPILER_IS_GNUCXX)
+    target_compile_options(${name} PUBLIC -fmodules-ts)
+  endif ()
+
+  # `std` is affected by CMake options and may be higher than C++20.
+  get_target_property(std ${name} CXX_STANDARD)
+
+  if (CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+    set(pcms)
+    foreach (src ${sources})
+      get_filename_component(pcm ${src} NAME_WE)
+      set(pcm ${pcm}.pcm)
+
+      # Propagate -fmodule-file=*.pcm to targets that link with this library.
+      target_compile_options(
+        ${name} PUBLIC -fmodule-file=${CMAKE_CURRENT_BINARY_DIR}/${pcm})
+
+      # Use an absolute path to prevent target_link_libraries prepending -l
+      # to it.
+      set(pcms ${pcms} ${CMAKE_CURRENT_BINARY_DIR}/${pcm})
+      add_custom_command(
+        OUTPUT ${pcm}
+        COMMAND ${CMAKE_CXX_COMPILER}
+                -std=c++${std} -x c++-module --precompile -c
+                -o ${pcm} ${CMAKE_CURRENT_SOURCE_DIR}/${src}
+                "-I$<JOIN:$<TARGET_PROPERTY:${name},INCLUDE_DIRECTORIES>,;-I>"
+        # Required by the -I generator expression above.
+        COMMAND_EXPAND_LISTS
+        DEPENDS ${src})
+    endforeach ()
+
+    # Add .pcm files as sources to make sure they are built before the library.
+    set(sources)
+    foreach (pcm ${pcms})
+      get_filename_component(pcm_we ${pcm} NAME_WE)
+      set(obj ${pcm_we}.o)
+      # Use an absolute path to prevent target_link_libraries prepending -l.
+      set(sources ${sources} ${pcm} ${CMAKE_CURRENT_BINARY_DIR}/${obj})
+      add_custom_command(
+        OUTPUT ${obj}
+        COMMAND ${CMAKE_CXX_COMPILER} $<TARGET_PROPERTY:${name},COMPILE_OPTIONS>
+                -c -o ${obj} ${pcm}
+        DEPENDS ${pcm})
+    endforeach ()
+  endif ()
+  target_sources(${name} PRIVATE ${sources})
+endfunction()
+
+include(CMakeParseArguments)
+
+# Sets a cache variable with a docstring joined from multiple arguments:
+#   set(<variable> <value>... CACHE <type> <docstring>...)
+# This allows splitting a long docstring for readability.
+function(set_verbose)
+  # cmake_parse_arguments is broken in CMake 3.4 (cannot parse CACHE) so use
+  # list instead.
+  list(GET ARGN 0 var)
+  list(REMOVE_AT ARGN 0)
+  list(GET ARGN 0 val)
+  list(REMOVE_AT ARGN 0)
+  list(REMOVE_AT ARGN 0)
+  list(GET ARGN 0 type)
+  list(REMOVE_AT ARGN 0)
+  join(doc ${ARGN})
+  set(${var} ${val} CACHE ${type} ${doc})
+endfunction()
+
+# Set the default CMAKE_BUILD_TYPE to Release.
+# This should be done before the project command since the latter can set
+# CMAKE_BUILD_TYPE itself (it does so for nmake).
+if (FMT_MASTER_PROJECT AND NOT CMAKE_BUILD_TYPE)
+  set_verbose(CMAKE_BUILD_TYPE Release CACHE STRING
+              "Choose the type of build, options are: None(CMAKE_CXX_FLAGS or "
+              "CMAKE_C_FLAGS used) Debug Release RelWithDebInfo MinSizeRel.")
+endif ()
+
+project(FMT CXX)
+include(GNUInstallDirs)
+set_verbose(FMT_INC_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE STRING
+            "Installation directory for include files, a relative path that "
+            "will be joined with ${CMAKE_INSTALL_PREFIX} or an absolute path.")
+
+option(FMT_PEDANTIC "Enable extra warnings and expensive tests." OFF)
+option(FMT_WERROR "Halt the compilation with an error on compiler warnings."
+       OFF)
+
+# Options that control generation of various targets.
+option(FMT_DOC "Generate the doc target." ${FMT_MASTER_PROJECT})
+option(FMT_INSTALL "Generate the install target." ON)
+option(FMT_TEST "Generate the test target." ${FMT_MASTER_PROJECT})
+option(FMT_FUZZ "Generate the fuzz target." OFF)
+option(FMT_CUDA_TEST "Generate the cuda-test target." OFF)
+option(FMT_OS "Include core requiring OS (Windows/Posix) " ON)
+option(FMT_MODULE "Build a module instead of a traditional library." OFF)
+option(FMT_SYSTEM_HEADERS "Expose headers with marking them as system." OFF)
+
+if (FMT_TEST AND FMT_MODULE)
+  # The tests require {fmt} to be compiled as traditional library
+  message(STATUS "Testing is incompatible with build mode 'module'.")
+endif ()
+set(FMT_SYSTEM_HEADERS_ATTRIBUTE "")
+if (FMT_SYSTEM_HEADERS)
+  set(FMT_SYSTEM_HEADERS_ATTRIBUTE SYSTEM)
+endif ()
+if (CMAKE_SYSTEM_NAME STREQUAL "MSDOS")
+  set(FMT_TEST OFF)
+  message(STATUS "MSDOS is incompatible with gtest")
+endif ()
+
+# Get version from core.h
+file(READ include/fmt/core.h core_h)
+if (NOT core_h MATCHES "FMT_VERSION ([0-9]+)([0-9][0-9])([0-9][0-9])")
+  message(FATAL_ERROR "Cannot get FMT_VERSION from core.h.")
+endif ()
+# Use math to skip leading zeros if any.
+math(EXPR CPACK_PACKAGE_VERSION_MAJOR ${CMAKE_MATCH_1})
+math(EXPR CPACK_PACKAGE_VERSION_MINOR ${CMAKE_MATCH_2})
+math(EXPR CPACK_PACKAGE_VERSION_PATCH ${CMAKE_MATCH_3})
+join(FMT_VERSION ${CPACK_PACKAGE_VERSION_MAJOR}.${CPACK_PACKAGE_VERSION_MINOR}.
+                 ${CPACK_PACKAGE_VERSION_PATCH})
+message(STATUS "Version: ${FMT_VERSION}")
+
+message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
+
+if (NOT CMAKE_RUNTIME_OUTPUT_DIRECTORY)
+  set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
+endif ()
+
+set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH}
+  "${CMAKE_CURRENT_SOURCE_DIR}/support/cmake")
+
+include(CheckCXXCompilerFlag)
+include(JoinPaths)
+
+if (FMT_MASTER_PROJECT AND NOT DEFINED CMAKE_CXX_VISIBILITY_PRESET)
+  set_verbose(CMAKE_CXX_VISIBILITY_PRESET hidden CACHE STRING
+              "Preset for the export of private symbols")
+  set_property(CACHE CMAKE_CXX_VISIBILITY_PRESET PROPERTY STRINGS
+               hidden default)
+endif ()
+
+if (FMT_MASTER_PROJECT AND NOT DEFINED CMAKE_VISIBILITY_INLINES_HIDDEN)
+  set_verbose(CMAKE_VISIBILITY_INLINES_HIDDEN ON CACHE BOOL
+              "Whether to add a compile flag to hide symbols of inline functions")
+endif ()
+
+if (CMAKE_CXX_COMPILER_ID MATCHES "GNU")
+  set(PEDANTIC_COMPILE_FLAGS -pedantic-errors -Wall -Wextra -pedantic
+      -Wold-style-cast -Wundef
+      -Wredundant-decls -Wwrite-strings -Wpointer-arith
+      -Wcast-qual -Wformat=2 -Wmissing-include-dirs
+      -Wcast-align
+      -Wctor-dtor-privacy -Wdisabled-optimization
+      -Winvalid-pch -Woverloaded-virtual
+      -Wconversion -Wundef
+      -Wno-ctor-dtor-privacy -Wno-format-nonliteral)
+  if (NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS 4.6)
+      set(PEDANTIC_COMPILE_FLAGS ${PEDANTIC_COMPILE_FLAGS}
+         -Wno-dangling-else -Wno-unused-local-typedefs)
+  endif ()
+  if (NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS 5.0)
+      set(PEDANTIC_COMPILE_FLAGS ${PEDANTIC_COMPILE_FLAGS} -Wdouble-promotion
+          -Wtrampolines -Wzero-as-null-pointer-constant -Wuseless-cast
+          -Wvector-operation-performance -Wsized-deallocation -Wshadow)
+  endif ()
+  if (NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS 6.0)
+      set(PEDANTIC_COMPILE_FLAGS ${PEDANTIC_COMPILE_FLAGS} -Wshift-overflow=2
+          -Wnull-dereference -Wduplicated-cond)
+  endif ()
+  set(WERROR_FLAG -Werror)
+endif ()
+
+if (CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+  set(PEDANTIC_COMPILE_FLAGS -Wall -Wextra -pedantic -Wconversion -Wundef
+      -Wdeprecated -Wweak-vtables -Wshadow
+      -Wno-gnu-zero-variadic-macro-arguments)
+  check_cxx_compiler_flag(-Wzero-as-null-pointer-constant HAS_NULLPTR_WARNING)
+  if (HAS_NULLPTR_WARNING)
+    set(PEDANTIC_COMPILE_FLAGS ${PEDANTIC_COMPILE_FLAGS}
+        -Wzero-as-null-pointer-constant)
+  endif ()
+  set(WERROR_FLAG -Werror)
+endif ()
+
+if (MSVC)
+  set(PEDANTIC_COMPILE_FLAGS /W3)
+  set(WERROR_FLAG /WX)
+endif ()
+
+if (FMT_MASTER_PROJECT AND CMAKE_GENERATOR MATCHES "Visual Studio")
+  # If Microsoft SDK is installed create script run-msbuild.bat that
+  # calls SetEnv.cmd to set up build environment and runs msbuild.
+  # It is useful when building Visual Studio projects with the SDK
+  # toolchain rather than Visual Studio.
+  include(FindSetEnv)
+  if (WINSDK_SETENV)
+    set(MSBUILD_SETUP "call \"${WINSDK_SETENV}\"")
+  endif ()
+  # Set FrameworkPathOverride to get rid of MSB3644 warnings.
+  join(netfxpath
+       "C:\\Program Files\\Reference Assemblies\\Microsoft\\Framework\\"
+       ".NETFramework\\v4.0")
+  file(WRITE run-msbuild.bat "
+    ${MSBUILD_SETUP}
+    ${CMAKE_MAKE_PROGRAM} -p:FrameworkPathOverride=\"${netfxpath}\" %*")
+endif ()
+
+function(add_headers VAR)
+  set(headers ${${VAR}})
+  foreach (header ${ARGN})
+    set(headers ${headers} include/fmt/${header})
+  endforeach()
+  set(${VAR} ${headers} PARENT_SCOPE)
+endfunction()
+
+# Define the fmt library, its includes and the needed defines.
+add_headers(FMT_HEADERS args.h chrono.h color.h compile.h core.h format.h
+                        format-inl.h os.h ostream.h printf.h ranges.h std.h
+                        xchar.h)
+set(FMT_SOURCES src/format.cc)
+if (FMT_OS)
+  set(FMT_SOURCES ${FMT_SOURCES} src/os.cc)
+endif ()
+
+add_module_library(fmt src/fmt.cc FALLBACK
+                   ${FMT_SOURCES} ${FMT_HEADERS} README.md ChangeLog.md
+                   IF FMT_MODULE)
+add_library(fmt::fmt ALIAS fmt)
+if (FMT_MODULE)
+  enable_module(fmt)
+endif ()
+
+if (FMT_WERROR)
+  target_compile_options(fmt PRIVATE ${WERROR_FLAG})
+endif ()
+if (FMT_PEDANTIC)
+  target_compile_options(fmt PRIVATE ${PEDANTIC_COMPILE_FLAGS})
+endif ()
+
+if (cxx_std_11 IN_LIST CMAKE_CXX_COMPILE_FEATURES)
+  target_compile_features(fmt PUBLIC cxx_std_11)
+else ()
+  message(WARNING "Feature cxx_std_11 is unknown for the CXX compiler")
+endif ()
+
+target_include_directories(fmt ${FMT_SYSTEM_HEADERS_ATTRIBUTE} PUBLIC
+  $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
+  $<INSTALL_INTERFACE:${FMT_INC_DIR}>)
+
+set(FMT_DEBUG_POSTFIX d CACHE STRING "Debug library postfix.")
+
+set_target_properties(fmt PROPERTIES
+  VERSION ${FMT_VERSION} SOVERSION ${CPACK_PACKAGE_VERSION_MAJOR}
+  PUBLIC_HEADER "${FMT_HEADERS}"
+  DEBUG_POSTFIX "${FMT_DEBUG_POSTFIX}"
+
+  # Workaround for Visual Studio 2017:
+  # Ensure the .pdb is created with the same name and in the same directory
+  # as the .lib. Newer VS versions already do this by default, but there is no
+  # harm in setting it for those too. Ignored by other generators.
+  COMPILE_PDB_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}"
+  COMPILE_PDB_NAME "fmt"
+  COMPILE_PDB_NAME_DEBUG "fmt${FMT_DEBUG_POSTFIX}")
+
+# Set FMT_LIB_NAME for pkg-config fmt.pc. We cannot use the OUTPUT_NAME target
+# property because it's not set by default.
+set(FMT_LIB_NAME fmt)
+if (CMAKE_BUILD_TYPE STREQUAL "Debug")
+  set(FMT_LIB_NAME ${FMT_LIB_NAME}${FMT_DEBUG_POSTFIX})
+endif ()
+
+if (BUILD_SHARED_LIBS)
+  target_compile_definitions(fmt PRIVATE FMT_LIB_EXPORT INTERFACE FMT_SHARED)
+endif ()
+if (FMT_SAFE_DURATION_CAST)
+  target_compile_definitions(fmt PUBLIC FMT_SAFE_DURATION_CAST)
+endif ()
+
+add_library(fmt-header-only INTERFACE)
+add_library(fmt::fmt-header-only ALIAS fmt-header-only)
+
+target_compile_definitions(fmt-header-only INTERFACE FMT_HEADER_ONLY=1)
+target_compile_features(fmt-header-only INTERFACE cxx_std_11)
+
+target_include_directories(fmt-header-only
+  ${FMT_SYSTEM_HEADERS_ATTRIBUTE} INTERFACE
+  $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
+  $<INSTALL_INTERFACE:${FMT_INC_DIR}>)
+
+# Install targets.
+if (FMT_INSTALL)
+  include(CMakePackageConfigHelpers)
+  set_verbose(FMT_CMAKE_DIR ${CMAKE_INSTALL_LIBDIR}/cmake/fmt CACHE STRING
+              "Installation directory for cmake files, a relative path that "
+              "will be joined with ${CMAKE_INSTALL_PREFIX} or an absolute "
+              "path.")
+  set(version_config ${PROJECT_BINARY_DIR}/fmt-config-version.cmake)
+  set(project_config ${PROJECT_BINARY_DIR}/fmt-config.cmake)
+  set(pkgconfig ${PROJECT_BINARY_DIR}/fmt.pc)
+  set(targets_export_name fmt-targets)
+
+  set_verbose(FMT_LIB_DIR ${CMAKE_INSTALL_LIBDIR} CACHE STRING
+              "Installation directory for libraries, a relative path that "
+              "will be joined to ${CMAKE_INSTALL_PREFIX} or an absolute path.")
+
+  set_verbose(FMT_PKGCONFIG_DIR ${CMAKE_INSTALL_LIBDIR}/pkgconfig CACHE STRING
+              "Installation directory for pkgconfig (.pc) files, a relative "
+              "path that will be joined with ${CMAKE_INSTALL_PREFIX} or an "
+              "absolute path.")
+
+  # Generate the version, config and target files into the build directory.
+  write_basic_package_version_file(
+    ${version_config}
+    VERSION ${FMT_VERSION}
+    COMPATIBILITY AnyNewerVersion)
+
+  join_paths(libdir_for_pc_file "\${exec_prefix}" "${FMT_LIB_DIR}")
+  join_paths(includedir_for_pc_file "\${prefix}" "${FMT_INC_DIR}")
+
+  configure_file(
+    "${PROJECT_SOURCE_DIR}/support/cmake/fmt.pc.in"
+    "${pkgconfig}"
+    @ONLY)
+  configure_package_config_file(
+    ${PROJECT_SOURCE_DIR}/support/cmake/fmt-config.cmake.in
+    ${project_config}
+    INSTALL_DESTINATION ${FMT_CMAKE_DIR})
+
+  set(INSTALL_TARGETS fmt fmt-header-only)
+
+  # Install the library and headers.
+  install(TARGETS ${INSTALL_TARGETS} EXPORT ${targets_export_name}
+          LIBRARY DESTINATION ${FMT_LIB_DIR}
+          ARCHIVE DESTINATION ${FMT_LIB_DIR}
+          PUBLIC_HEADER DESTINATION "${FMT_INC_DIR}/fmt"
+          RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
+
+  # Use a namespace because CMake provides better diagnostics for namespaced
+  # imported targets.
+  export(TARGETS ${INSTALL_TARGETS} NAMESPACE fmt::
+         FILE ${PROJECT_BINARY_DIR}/${targets_export_name}.cmake)
+
+  # Install version, config and target files.
+  install(
+    FILES ${project_config} ${version_config}
+    DESTINATION ${FMT_CMAKE_DIR})
+  install(EXPORT ${targets_export_name} DESTINATION ${FMT_CMAKE_DIR}
+          NAMESPACE fmt::)
+
+  install(FILES "${pkgconfig}" DESTINATION "${FMT_PKGCONFIG_DIR}")
+endif ()
+
+if (FMT_DOC)
+  add_subdirectory(doc)
+endif ()
+
+if (FMT_TEST)
+  enable_testing()
+  add_subdirectory(test)
+endif ()
+
+# Control fuzzing independent of the unit tests.
+if (FMT_FUZZ)
+  add_subdirectory(test/fuzzing)
+
+  # The FMT_FUZZ macro is used to prevent resource exhaustion in fuzzing
+  # mode and make fuzzing practically possible. It is similar to
+  # FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION but uses a different name to
+  # avoid interfering with fuzzing of projects that use {fmt}.
+  # See also https://llvm.org/docs/LibFuzzer.html#fuzzer-friendly-build-mode.
+  target_compile_definitions(fmt PUBLIC FMT_FUZZ)
+endif ()
+
+set(gitignore ${PROJECT_SOURCE_DIR}/.gitignore)
+if (FMT_MASTER_PROJECT AND EXISTS ${gitignore})
+  # Get the list of ignored files from .gitignore.
+  file (STRINGS ${gitignore} lines)
+  list(REMOVE_ITEM lines /doc/html)
+  foreach (line ${lines})
+    string(REPLACE "." "[.]" line "${line}")
+    string(REPLACE "*" ".*" line "${line}")
+    set(ignored_files ${ignored_files} "${line}$" "${line}/")
+  endforeach ()
+  set(ignored_files ${ignored_files}
+    /.git /breathe /format-benchmark sphinx/ .buildinfo .doctrees)
+
+  set(CPACK_SOURCE_GENERATOR ZIP)
+  set(CPACK_SOURCE_IGNORE_FILES ${ignored_files})
+  set(CPACK_SOURCE_PACKAGE_FILE_NAME fmt-${FMT_VERSION})
+  set(CPACK_PACKAGE_NAME fmt)
+  set(CPACK_RESOURCE_FILE_README ${PROJECT_SOURCE_DIR}/README.md)
+  include(CPack)
+endif ()
diff --git a/dep/fmt/CONTRIBUTING.md b/dep/fmt/CONTRIBUTING.md
new file mode 100644
index 00000000..b82f1450
--- /dev/null
+++ b/dep/fmt/CONTRIBUTING.md
@@ -0,0 +1,20 @@
+Contributing to {fmt}
+=====================
+
+By submitting a pull request or a patch, you represent that you have the right
+to license your contribution to the {fmt} project owners and the community,
+agree that your contributions are licensed under the {fmt} license, and agree
+to future changes to the licensing.
+
+All C++ code must adhere to [Google C++ Style Guide](
+https://google.github.io/styleguide/cppguide.html) with the following
+exceptions:
+
+* Exceptions are permitted
+* snake_case should be used instead of UpperCamelCase for function and type
+  names
+
+All documentation must adhere to the [Google Developer Documentation Style
+Guide](https://developers.google.com/style).
+
+Thanks for contributing!
diff --git a/dep/fmt/ChangeLog.md b/dep/fmt/ChangeLog.md
new file mode 100644
index 00000000..515f9fdc
--- /dev/null
+++ b/dep/fmt/ChangeLog.md
@@ -0,0 +1,5533 @@
+# 10.2.1 - 2024-01-03
+
+-   Fixed ABI compatibility with earlier 10.x versions
+    (https://github.com/fmtlib/fmt/pull/3786). Thanks @saraedum.
+
+# 10.2.0 - 2024-01-01
+
+-   Added support for the `%j` specifier (the number of days) for
+    `std::chrono::duration` (https://github.com/fmtlib/fmt/issues/3643,
+    https://github.com/fmtlib/fmt/pull/3732). Thanks @intelfx.
+
+-   Added support for the chrono suffix for days and changed
+    the suffix for minutes from "m" to the correct "min"
+    (https://github.com/fmtlib/fmt/issues/3662,
+    https://github.com/fmtlib/fmt/pull/3664).
+    For example ([godbolt](https://godbolt.org/z/9KhMnq9ba)):
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      fmt::print("{}\n", std::chrono::days(42)); // prints "42d"
+    }
+    ```
+
+    Thanks @Richardk2n.
+
+-   Fixed an overflow in `std::chrono::time_point` formatting with large dates
+    (https://github.com/fmtlib/fmt/issues/3725,
+    https://github.com/fmtlib/fmt/pull/3727). Thanks @cschreib.
+
+-   Added a formatter for `std::source_location`
+    (https://github.com/fmtlib/fmt/pull/3730).
+    For example ([godbolt](https://godbolt.org/z/YajfKjhhr)):
+
+    ```c++
+    #include <source_location>
+    #include <fmt/std.h>
+
+    int main() {
+      fmt::print("{}\n", std::source_location::current());
+    }
+    ```
+
+    prints
+
+    ```
+    /app/example.cpp:5:51: int main()
+    ```
+
+    Thanks @felix642.
+
+-   Added a formatter for `std::bitset`
+    (https://github.com/fmtlib/fmt/pull/3660).
+    For example ([godbolt](https://godbolt.org/z/bdEaGeYxe)):
+
+    ```c++
+    #include <bitset>
+    #include <fmt/std.h>
+
+    int main() {
+      fmt::print("{}\n", std::bitset<6>(42)); // prints "101010"
+    }
+    ```
+
+    Thanks @muggenhor.
+
+-   Added an experimental `nested_formatter` that provides an easy way of
+    applying a formatter to one or more subobjects while automatically handling
+    width, fill and alignment. For example:
+
+    ```c++
+    #include <fmt/format.h>
+
+    struct point {
+      double x, y;
+    };
+
+    template <>
+    struct fmt::formatter<point> : nested_formatter<double> {
+      auto format(point p, format_context& ctx) const {
+        return write_padded(ctx, [=](auto out) {
+          return format_to(out, "({}, {})", nested(p.x), nested(p.y));
+        });
+      }
+    };
+
+    int main() {
+      fmt::print("[{:>20.2f}]", point{1, 2});
+    }
+    ```
+
+    prints
+
+    ```
+    [          (1.00, 2.00)]
+    ```
+
+-   Added the generic representation (`g`) to `std::filesystem::path`
+    (https://github.com/fmtlib/fmt/issues/3715,
+    https://github.com/fmtlib/fmt/pull/3729). For example:
+
+    ```c++
+    #include <filesystem>
+    #include <fmt/std.h>
+
+    int main() {
+      fmt::print("{:g}\n", std::filesystem::path("C:\\foo"));
+    }
+    ```
+
+    prints `"C:/foo"` on Windows.
+
+    Thanks @js324.
+
+-   Made `format_as` work with references
+    (https://github.com/fmtlib/fmt/pull/3739). Thanks @tchaikov.
+
+-   Fixed formatting of invalid UTF-8 with precision
+    (https://github.com/fmtlib/fmt/issues/3284).
+
+-   Fixed an inconsistency between `fmt::to_string` and `fmt::format`
+    (https://github.com/fmtlib/fmt/issues/3684).
+
+-   Disallowed unsafe uses of `fmt::styled`
+    (https://github.com/fmtlib/fmt/issues/3625):
+
+    ```c++
+    auto s = fmt::styled(std::string("dangle"), fmt::emphasis::bold);
+    fmt::print("{}\n", s); // compile error
+    ```
+
+    Pass `fmt::styled(...)` as a parameter instead.
+
+-   Added a null check when formatting a C string with the `s` specifier
+    (https://github.com/fmtlib/fmt/issues/3706).
+
+-   Disallowed the `c` specifier for `bool`
+    (https://github.com/fmtlib/fmt/issues/3726,
+    https://github.com/fmtlib/fmt/pull/3734). Thanks @js324.
+
+-   Made the default formatting unlocalized in `fmt::ostream_formatter` for
+    consistency with the rest of the library
+    (https://github.com/fmtlib/fmt/issues/3460).
+
+-   Fixed localized formatting in bases other than decimal
+    (https://github.com/fmtlib/fmt/issues/3693,
+    https://github.com/fmtlib/fmt/pull/3750). Thanks @js324.
+
+-   Fixed a performance regression in experimental `fmt::ostream::print`
+    (https://github.com/fmtlib/fmt/issues/3674).
+
+-   Added synchronization with the underlying output stream when writing to
+    the Windows console
+    (https://github.com/fmtlib/fmt/pull/3668,
+    https://github.com/fmtlib/fmt/issues/3688,
+    https://github.com/fmtlib/fmt/pull/3689).
+    Thanks @Roman-Koshelev and @dimztimz.
+
+-   Changed to only export `format_error` when {fmt} is built as a shared
+    library (https://github.com/fmtlib/fmt/issues/3626,
+    https://github.com/fmtlib/fmt/pull/3627). Thanks @phprus.
+
+-   Made `fmt::streamed` `constexpr`.
+    (https://github.com/fmtlib/fmt/pull/3650). Thanks @muggenhor.
+
+-   Enabled `consteval` on older versions of MSVC
+    (https://github.com/fmtlib/fmt/pull/3757). Thanks @phprus.
+
+-   Added an option to build without `wchar_t` support on Windows
+    (https://github.com/fmtlib/fmt/issues/3631,
+    https://github.com/fmtlib/fmt/pull/3636). Thanks @glebm.
+
+-   Improved build and CI configuration
+    (https://github.com/fmtlib/fmt/pull/3679,
+    https://github.com/fmtlib/fmt/issues/3701,
+    https://github.com/fmtlib/fmt/pull/3702,
+    https://github.com/fmtlib/fmt/pull/3749).
+    Thanks @jcar87, @pklima and @tchaikov.
+
+-   Fixed various warnings, compilation and test issues
+    (https://github.com/fmtlib/fmt/issues/3607,
+    https://github.com/fmtlib/fmt/pull/3610,
+    https://github.com/fmtlib/fmt/pull/3624,
+    https://github.com/fmtlib/fmt/pull/3630,
+    https://github.com/fmtlib/fmt/pull/3634,
+    https://github.com/fmtlib/fmt/pull/3638,
+    https://github.com/fmtlib/fmt/issues/3645,
+    https://github.com/fmtlib/fmt/issues/3646,
+    https://github.com/fmtlib/fmt/pull/3647,
+    https://github.com/fmtlib/fmt/pull/3652,
+    https://github.com/fmtlib/fmt/issues/3654,
+    https://github.com/fmtlib/fmt/pull/3663,
+    https://github.com/fmtlib/fmt/issues/3670,
+    https://github.com/fmtlib/fmt/pull/3680,
+    https://github.com/fmtlib/fmt/issues/3694,
+    https://github.com/fmtlib/fmt/pull/3695,
+    https://github.com/fmtlib/fmt/pull/3699,
+    https://github.com/fmtlib/fmt/issues/3705,
+    https://github.com/fmtlib/fmt/issues/3710,
+    https://github.com/fmtlib/fmt/issues/3712,
+    https://github.com/fmtlib/fmt/pull/3713,
+    https://github.com/fmtlib/fmt/issues/3714,
+    https://github.com/fmtlib/fmt/pull/3716,
+    https://github.com/fmtlib/fmt/pull/3723,
+    https://github.com/fmtlib/fmt/issues/3738,
+    https://github.com/fmtlib/fmt/issues/3740,
+    https://github.com/fmtlib/fmt/pull/3741,
+    https://github.com/fmtlib/fmt/pull/3743,
+    https://github.com/fmtlib/fmt/issues/3745,
+    https://github.com/fmtlib/fmt/pull/3747,
+    https://github.com/fmtlib/fmt/pull/3748,
+    https://github.com/fmtlib/fmt/pull/3751,
+    https://github.com/fmtlib/fmt/pull/3754,
+    https://github.com/fmtlib/fmt/pull/3755,
+    https://github.com/fmtlib/fmt/issues/3760,
+    https://github.com/fmtlib/fmt/pull/3762,
+    https://github.com/fmtlib/fmt/issues/3763,
+    https://github.com/fmtlib/fmt/pull/3764,
+    https://github.com/fmtlib/fmt/issues/3774,
+    https://github.com/fmtlib/fmt/pull/3779).
+    Thanks @danakj, @vinayyadav3016, @cyyever, @phprus, @qimiko, @saschasc,
+    @gsjaardema, @lazka, @Zhaojun-Liu, @carlsmedstad, @hotwatermorning,
+    @cptFracassa, @kuguma, @PeterJohnson, @H1X4Dev, @asantoni, @eltociear,
+    @msimberg, @tchaikov, @waywardmonkeys.
+
+-   Improved documentation and README
+    (https://github.com/fmtlib/fmt/issues/2086,
+    https://github.com/fmtlib/fmt/issues/3637,
+    https://github.com/fmtlib/fmt/pull/3642,
+    https://github.com/fmtlib/fmt/pull/3653,
+    https://github.com/fmtlib/fmt/pull/3655,
+    https://github.com/fmtlib/fmt/pull/3661,
+    https://github.com/fmtlib/fmt/issues/3673,
+    https://github.com/fmtlib/fmt/pull/3677,
+    https://github.com/fmtlib/fmt/pull/3737,
+    https://github.com/fmtlib/fmt/issues/3742,
+    https://github.com/fmtlib/fmt/pull/3744).
+    Thanks @idzm, @perlun, @joycebrum, @fennewald, @reinhardt1053, @GeorgeLS.
+
+-   Updated CI dependencies
+    (https://github.com/fmtlib/fmt/pull/3615,
+    https://github.com/fmtlib/fmt/pull/3622,
+    https://github.com/fmtlib/fmt/pull/3623,
+    https://github.com/fmtlib/fmt/pull/3666,
+    https://github.com/fmtlib/fmt/pull/3696,
+    https://github.com/fmtlib/fmt/pull/3697,
+    https://github.com/fmtlib/fmt/pull/3759,
+    https://github.com/fmtlib/fmt/pull/3782).
+
+# 10.1.1 - 2023-08-28
+
+-   Added formatters for `std::atomic` and `atomic_flag`
+    (https://github.com/fmtlib/fmt/pull/3574,
+    https://github.com/fmtlib/fmt/pull/3594).
+    Thanks @wangzw and @AlexGuteniev.
+-   Fixed an error about partial specialization of `formatter<string>`
+    after instantiation when compiled with gcc and C++20
+    (https://github.com/fmtlib/fmt/issues/3584).
+-   Fixed compilation as a C++20 module with gcc and clang
+    (https://github.com/fmtlib/fmt/issues/3587,
+    https://github.com/fmtlib/fmt/pull/3597,
+    https://github.com/fmtlib/fmt/pull/3605).
+    Thanks @MathewBensonCode.
+-   Made `fmt::to_string` work with types that have `format_as`
+    overloads (https://github.com/fmtlib/fmt/pull/3575). Thanks @phprus.
+-   Made `formatted_size` work with integral format specifiers at
+    compile time (https://github.com/fmtlib/fmt/pull/3591).
+    Thanks @elbeno.
+-   Fixed a warning about the `no_unique_address` attribute on clang-cl
+    (https://github.com/fmtlib/fmt/pull/3599). Thanks @lukester1975.
+-   Improved compatibility with the legacy GBK encoding
+    (https://github.com/fmtlib/fmt/issues/3598,
+    https://github.com/fmtlib/fmt/pull/3599). Thanks @YuHuanTin.
+-   Added OpenSSF Scorecard analysis
+    (https://github.com/fmtlib/fmt/issues/3530,
+    https://github.com/fmtlib/fmt/pull/3571). Thanks @joycebrum.
+-   Updated CI dependencies
+    (https://github.com/fmtlib/fmt/pull/3591,
+    https://github.com/fmtlib/fmt/pull/3592,
+    https://github.com/fmtlib/fmt/pull/3593,
+    https://github.com/fmtlib/fmt/pull/3602).
+
+# 10.1.0 - 2023-08-12
+
+-   Optimized format string compilation resulting in up to 40% speed up
+    in compiled `format_to` and \~4x speed up in compiled `format_to_n`
+    on a concatenation benchmark
+    (https://github.com/fmtlib/fmt/issues/3133,
+    https://github.com/fmtlib/fmt/issues/3484).
+
+    {fmt} 10.0:
+
+        ---------------------------------------------------------
+        Benchmark               Time             CPU   Iterations
+        ---------------------------------------------------------
+        BM_format_to         78.9 ns         78.9 ns      8881746
+        BM_format_to_n        568 ns          568 ns      1232089
+
+    {fmt} 10.1:
+
+        ---------------------------------------------------------
+        Benchmark               Time             CPU   Iterations
+        ---------------------------------------------------------
+        BM_format_to         54.9 ns         54.9 ns     12727944
+        BM_format_to_n        133 ns          133 ns      5257795
+
+-   Optimized storage of an empty allocator in `basic_memory_buffer`
+    (https://github.com/fmtlib/fmt/pull/3485). Thanks @Minty-Meeo.
+
+-   Added formatters for proxy references to elements of
+    `std::vector<bool>` and `std::bitset<N>`
+    (https://github.com/fmtlib/fmt/issues/3567,
+    https://github.com/fmtlib/fmt/pull/3570). For example
+    ([godbolt](https://godbolt.org/z/zYb79Pvn8)):
+
+    ```c++
+    #include <vector>
+    #include <fmt/std.h>
+
+    int main() {
+      auto v = std::vector<bool>{true};
+      fmt::print("{}", v[0]);
+    }
+    ```
+
+    Thanks @phprus and @felix642.
+
+-   Fixed an ambiguous formatter specialization for containers that look
+    like container adaptors such as `boost::flat_set`
+    (https://github.com/fmtlib/fmt/issues/3556,
+    https://github.com/fmtlib/fmt/pull/3561). Thanks @5chmidti.
+
+-   Fixed compilation when formatting durations not convertible from
+    `std::chrono::seconds`
+    (https://github.com/fmtlib/fmt/pull/3430). Thanks @patlkli.
+
+-   Made the `formatter` specialization for `char*` const-correct
+    (https://github.com/fmtlib/fmt/pull/3432). Thanks @timsong-cpp.
+
+-   Made `{}` and `{:}` handled consistently during compile-time checks
+    (https://github.com/fmtlib/fmt/issues/3526).
+
+-   Disallowed passing temporaries to `make_format_args` to improve API
+    safety by preventing dangling references.
+
+-   Improved the compile-time error for unformattable types
+    (https://github.com/fmtlib/fmt/pull/3478). Thanks @BRevzin.
+
+-   Improved the floating-point formatter
+    (https://github.com/fmtlib/fmt/pull/3448,
+    https://github.com/fmtlib/fmt/pull/3450).
+    Thanks @florimond-collette.
+
+-   Fixed handling of precision for `long double` larger than 64 bits.
+    (https://github.com/fmtlib/fmt/issues/3539,
+    https://github.com/fmtlib/fmt/issues/3564).
+
+-   Made floating-point and chrono tests less platform-dependent
+    (https://github.com/fmtlib/fmt/issues/3337,
+    https://github.com/fmtlib/fmt/issues/3433,
+    https://github.com/fmtlib/fmt/pull/3434). Thanks @phprus.
+
+-   Removed the remnants of the Grisu floating-point formatter that has
+    been replaced by Dragonbox in earlier versions.
+
+-   Added `throw_format_error` to the public API
+    (https://github.com/fmtlib/fmt/pull/3551). Thanks @mjerabek.
+
+-   Made `FMT_THROW` assert even if assertions are disabled when
+    compiling with exceptions disabled
+    (https://github.com/fmtlib/fmt/issues/3418,
+    https://github.com/fmtlib/fmt/pull/3439). Thanks @BRevzin.
+
+-   Made `format_as` and `std::filesystem::path` formatter work with
+    exotic code unit types.
+    (https://github.com/fmtlib/fmt/pull/3457,
+    https://github.com/fmtlib/fmt/pull/3476). Thanks @gix and @hmbj.
+
+-   Added support for the `?` format specifier to
+    `std::filesystem::path` and made the default unescaped for
+    consistency with strings.
+
+-   Deprecated the wide stream overload of `printf`.
+
+-   Removed unused `basic_printf_parse_context`.
+
+-   Improved RTTI detection used when formatting exceptions
+    (https://github.com/fmtlib/fmt/pull/3468). Thanks @danakj.
+
+-   Improved compatibility with VxWorks7
+    (https://github.com/fmtlib/fmt/pull/3467). Thanks @wenshan1.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/3174,
+    https://github.com/fmtlib/fmt/issues/3423,
+    https://github.com/fmtlib/fmt/pull/3454,
+    https://github.com/fmtlib/fmt/issues/3458,
+    https://github.com/fmtlib/fmt/pull/3461,
+    https://github.com/fmtlib/fmt/issues/3487,
+    https://github.com/fmtlib/fmt/pull/3515).
+    Thanks @zencatalyst, @rlalik and @mikecrowe.
+
+-   Improved build and CI configurations
+    (https://github.com/fmtlib/fmt/issues/3449,
+    https://github.com/fmtlib/fmt/pull/3451,
+    https://github.com/fmtlib/fmt/pull/3452,
+    https://github.com/fmtlib/fmt/pull/3453,
+    https://github.com/fmtlib/fmt/pull/3459,
+    https://github.com/fmtlib/fmt/issues/3481,
+    https://github.com/fmtlib/fmt/pull/3486,
+    https://github.com/fmtlib/fmt/issues/3489,
+    https://github.com/fmtlib/fmt/pull/3496,
+    https://github.com/fmtlib/fmt/issues/3517,
+    https://github.com/fmtlib/fmt/pull/3523,
+    https://github.com/fmtlib/fmt/pull/3563).
+    Thanks @joycebrum, @glebm, @phprus, @petrmanek, @setoye and @abouvier.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/issues/3408,
+    https://github.com/fmtlib/fmt/issues/3424,
+    https://github.com/fmtlib/fmt/issues/3444,
+    https://github.com/fmtlib/fmt/pull/3446,
+    https://github.com/fmtlib/fmt/pull/3475,
+    https://github.com/fmtlib/fmt/pull/3482,
+    https://github.com/fmtlib/fmt/issues/3492,
+    https://github.com/fmtlib/fmt/pull/3493,
+    https://github.com/fmtlib/fmt/pull/3508,
+    https://github.com/fmtlib/fmt/issues/3509,
+    https://github.com/fmtlib/fmt/issues/3533,
+    https://github.com/fmtlib/fmt/pull/3542,
+    https://github.com/fmtlib/fmt/issues/3543,
+    https://github.com/fmtlib/fmt/issues/3540,
+    https://github.com/fmtlib/fmt/pull/3544,
+    https://github.com/fmtlib/fmt/issues/3548,
+    https://github.com/fmtlib/fmt/pull/3549,
+    https://github.com/fmtlib/fmt/pull/3550,
+    https://github.com/fmtlib/fmt/pull/3552).
+    Thanks @adesitter, @hmbj, @Minty-Meeo, @phprus, @TobiSchluter,
+    @kieranclancy, @alexeedm, @jurihock, @Ozomahtli and @razaqq.
+
+# 10.0.0 - 2023-05-09
+
+-   Replaced Grisu with a new floating-point formatting algorithm for
+    given precision (https://github.com/fmtlib/fmt/issues/3262,
+    https://github.com/fmtlib/fmt/issues/2750,
+    https://github.com/fmtlib/fmt/pull/3269,
+    https://github.com/fmtlib/fmt/pull/3276). The new algorithm
+    is based on Dragonbox already used for the shortest representation
+    and gives substantial performance improvement:
+
+    ![](https://user-images.githubusercontent.com/33922675/211956670-84891a09-6867-47d9-82fc-3230da7abe0f.png)
+
+    -   Red: new algorithm
+    -   Green: new algorithm with `FMT_USE_FULL_CACHE_DRAGONBOX` defined
+        to 1
+    -   Blue: old algorithm
+
+    Thanks @jk-jeon.
+
+-   Replaced `snprintf`-based hex float formatter with an internal
+    implementation (https://github.com/fmtlib/fmt/pull/3179,
+    https://github.com/fmtlib/fmt/pull/3203). This removes the
+    last usage of `s(n)printf` in {fmt}. Thanks @phprus.
+
+-   Fixed alignment of floating-point numbers with localization
+    (https://github.com/fmtlib/fmt/issues/3263,
+    https://github.com/fmtlib/fmt/pull/3272). Thanks @ShawnZhong.
+
+-   Made handling of `#` consistent with `std::format`.
+
+-   Improved C++20 module support
+    (https://github.com/fmtlib/fmt/pull/3134,
+    https://github.com/fmtlib/fmt/pull/3254,
+    https://github.com/fmtlib/fmt/pull/3386,
+    https://github.com/fmtlib/fmt/pull/3387,
+    https://github.com/fmtlib/fmt/pull/3388,
+    https://github.com/fmtlib/fmt/pull/3392,
+    https://github.com/fmtlib/fmt/pull/3397,
+    https://github.com/fmtlib/fmt/pull/3399,
+    https://github.com/fmtlib/fmt/pull/3400).
+    Thanks @laitingsheng, @Orvid and @DanielaE.
+    
+-   Switched to the [modules CMake library](https://github.com/vitaut/modules)
+    which allows building {fmt} as a C++20 module with clang:
+
+        CXX=clang++ cmake -DFMT_MODULE=ON .
+        make
+
+-   Made `format_as` work with any user-defined type and not just enums.
+    For example ([godbolt](https://godbolt.org/z/b7rqhq5Kh)):
+
+    ```c++
+    #include <fmt/format.h>
+
+    struct floaty_mc_floatface {
+      double value;
+    };
+
+    auto format_as(floaty_mc_floatface f) { return f.value; }
+
+    int main() {
+      fmt::print("{:8}\n", floaty_mc_floatface{0.42}); // prints "    0.42"
+    }
+    ```
+
+-   Removed deprecated implicit conversions for enums and conversions to
+    primitive types for compatibility with `std::format` and to prevent
+    potential ODR violations. Use `format_as` instead.
+
+-   Added support for fill, align and width to the time point formatter
+    (https://github.com/fmtlib/fmt/issues/3237,
+    https://github.com/fmtlib/fmt/pull/3260,
+    https://github.com/fmtlib/fmt/pull/3275). For example
+    ([godbolt](https://godbolt.org/z/rKP6MGz6c)):
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      // prints "    2023"
+      fmt::print("{:>8%Y}\n", std::chrono::system_clock::now());
+    }
+    ```
+
+    Thanks @ShawnZhong.
+
+-   Implemented formatting of subseconds
+    (https://github.com/fmtlib/fmt/issues/2207,
+    https://github.com/fmtlib/fmt/issues/3117,
+    https://github.com/fmtlib/fmt/pull/3115,
+    https://github.com/fmtlib/fmt/pull/3143,
+    https://github.com/fmtlib/fmt/pull/3144,
+    https://github.com/fmtlib/fmt/pull/3349). For example
+    ([godbolt](https://godbolt.org/z/45738oGEo)):
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      // prints 01.234567
+      fmt::print("{:%S}\n", std::chrono::microseconds(1234567));
+    }
+    ```
+
+    Thanks @patrickroocks @phprus and @BRevzin.
+
+-   Added precision support to `%S`
+    (https://github.com/fmtlib/fmt/pull/3148). Thanks @SappyJoy
+
+-   Added support for `std::utc_time`
+    (https://github.com/fmtlib/fmt/issues/3098,
+    https://github.com/fmtlib/fmt/pull/3110). Thanks @patrickroocks.
+
+-   Switched formatting of `std::chrono::system_clock` from local time
+    to UTC for compatibility with the standard
+    (https://github.com/fmtlib/fmt/issues/3199,
+    https://github.com/fmtlib/fmt/pull/3230). Thanks @ned14.
+
+-   Added support for `%Ez` and `%Oz` to chrono formatters.
+    (https://github.com/fmtlib/fmt/issues/3220,
+    https://github.com/fmtlib/fmt/pull/3222). Thanks @phprus.
+
+-   Improved validation of format specifiers for `std::chrono::duration`
+    (https://github.com/fmtlib/fmt/issues/3219,
+    https://github.com/fmtlib/fmt/pull/3232). Thanks @ShawnZhong.
+
+-   Fixed formatting of time points before the epoch
+    (https://github.com/fmtlib/fmt/issues/3117,
+    https://github.com/fmtlib/fmt/pull/3261). For example
+    ([godbolt](https://godbolt.org/z/f7bcznb3W)):
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      auto t = std::chrono::system_clock::from_time_t(0) -
+               std::chrono::milliseconds(250);
+      fmt::print("{:%S}\n", t); // prints 59.750000000
+    }
+    ```
+
+    Thanks @ShawnZhong.
+
+-   Experimental: implemented glibc extension for padding seconds,
+    minutes and hours
+    (https://github.com/fmtlib/fmt/issues/2959,
+    https://github.com/fmtlib/fmt/pull/3271). Thanks @ShawnZhong.
+
+-   Added a formatter for `std::exception`
+    (https://github.com/fmtlib/fmt/issues/2977,
+    https://github.com/fmtlib/fmt/issues/3012,
+    https://github.com/fmtlib/fmt/pull/3062,
+    https://github.com/fmtlib/fmt/pull/3076,
+    https://github.com/fmtlib/fmt/pull/3119). For example
+    ([godbolt](https://godbolt.org/z/8xoWGs9e4)):
+
+    ```c++
+    #include <fmt/std.h>
+    #include <vector>
+
+    int main() {
+      try {
+        std::vector<bool>().at(0);
+      } catch(const std::exception& e) {
+        fmt::print("{}", e);
+      }
+    }
+    ```
+
+    prints:
+
+        vector<bool>::_M_range_check: __n (which is 0) >= this->size() (which is 0)
+
+    on libstdc++. Thanks @zach2good and @phprus.
+
+-   Moved `std::error_code` formatter from `fmt/os.h` to `fmt/std.h`.
+    (https://github.com/fmtlib/fmt/pull/3125). Thanks @phprus.
+
+-   Added formatters for standard container adapters:
+    `std::priority_queue`, `std::queue` and `std::stack`
+    (https://github.com/fmtlib/fmt/issues/3215,
+    https://github.com/fmtlib/fmt/pull/3279). For example
+    ([godbolt](https://godbolt.org/z/74h1xY9qK)):
+
+    ```c++
+    #include <fmt/ranges.h>
+    #include <stack>
+    #include <vector>
+
+    int main() {
+      auto s = std::stack<bool, std::vector<bool>>();
+      for (auto b: {true, false, true}) s.push(b);
+      fmt::print("{}\n", s); // prints [true, false, true]
+    }
+    ```
+
+    Thanks @ShawnZhong.
+
+-   Added a formatter for `std::optional` to `fmt/std.h`
+    (https://github.com/fmtlib/fmt/issues/1367,
+    https://github.com/fmtlib/fmt/pull/3303).
+    Thanks @tom-huntington.
+
+-   Fixed formatting of valueless by exception variants
+    (https://github.com/fmtlib/fmt/pull/3347). Thanks @TheOmegaCarrot.
+
+-   Made `fmt::ptr` accept `unique_ptr` with a custom deleter
+    (https://github.com/fmtlib/fmt/pull/3177). Thanks @hmbj.
+
+-   Fixed formatting of noncopyable ranges and nested ranges of chars
+    (https://github.com/fmtlib/fmt/pull/3158
+    https://github.com/fmtlib/fmt/issues/3286,
+    https://github.com/fmtlib/fmt/pull/3290). Thanks @BRevzin.
+
+-   Fixed issues with formatting of paths and ranges of paths
+    (https://github.com/fmtlib/fmt/issues/3319,
+    https://github.com/fmtlib/fmt/pull/3321
+    https://github.com/fmtlib/fmt/issues/3322). Thanks @phprus.
+
+-   Improved handling of invalid Unicode in paths.
+
+-   Enabled compile-time checks on Apple clang 14 and later
+    (https://github.com/fmtlib/fmt/pull/3331). Thanks @cloyce.
+
+-   Improved compile-time checks of named arguments
+    (https://github.com/fmtlib/fmt/issues/3105,
+    https://github.com/fmtlib/fmt/pull/3214). Thanks @rbrich.
+
+-   Fixed formatting when both alignment and `0` are given
+    (https://github.com/fmtlib/fmt/issues/3236,
+    https://github.com/fmtlib/fmt/pull/3248). Thanks @ShawnZhong.
+
+-   Improved Unicode support in the experimental file API on Windows
+    (https://github.com/fmtlib/fmt/issues/3234,
+    https://github.com/fmtlib/fmt/pull/3293). Thanks @Fros1er.
+
+-   Unified UTF transcoding
+    (https://github.com/fmtlib/fmt/pull/3416). Thanks @phprus.
+
+-   Added support for UTF-8 digit separators via an experimental locale
+    facet (https://github.com/fmtlib/fmt/issues/1861). For
+    example ([godbolt](https://godbolt.org/z/f7bcznb3W)):
+
+    ```c++
+    auto loc = std::locale(
+      std::locale(), new fmt::format_facet<std::locale>("’"));
+    auto s = fmt::format(loc, "{:L}", 1000);
+    ```
+
+    where `’` is U+2019 used as a digit separator in the de_CH locale.
+
+-   Added an overload of `formatted_size` that takes a locale
+    (https://github.com/fmtlib/fmt/issues/3084,
+    https://github.com/fmtlib/fmt/pull/3087). Thanks @gerboengels.
+
+-   Removed the deprecated `FMT_DEPRECATED_OSTREAM`.
+
+-   Fixed a UB when using a null `std::string_view` with
+    `fmt::to_string` or format string compilation
+    (https://github.com/fmtlib/fmt/issues/3241,
+    https://github.com/fmtlib/fmt/pull/3244). Thanks @phprus.
+
+-   Added `starts_with` to the fallback `string_view` implementation
+    (https://github.com/fmtlib/fmt/pull/3080). Thanks @phprus.
+
+-   Added `fmt::basic_format_string::get()` for compatibility with
+    `basic_format_string`
+    (https://github.com/fmtlib/fmt/pull/3111). Thanks @huangqinjin.
+
+-   Added `println` for compatibility with C++23
+    (https://github.com/fmtlib/fmt/pull/3267). Thanks @ShawnZhong.
+
+-   Renamed the `FMT_EXPORT` macro for shared library usage to
+    `FMT_LIB_EXPORT`.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/3108,
+    https://github.com/fmtlib/fmt/issues/3169,
+    https://github.com/fmtlib/fmt/pull/3243).
+    https://github.com/fmtlib/fmt/pull/3404).
+    Thanks @Cleroth and @Vertexwahn.
+
+-   Improved build configuration and tests
+    (https://github.com/fmtlib/fmt/pull/3118,
+    https://github.com/fmtlib/fmt/pull/3120,
+    https://github.com/fmtlib/fmt/pull/3188,
+    https://github.com/fmtlib/fmt/issues/3189,
+    https://github.com/fmtlib/fmt/pull/3198,
+    https://github.com/fmtlib/fmt/pull/3205,
+    https://github.com/fmtlib/fmt/pull/3207,
+    https://github.com/fmtlib/fmt/pull/3210,
+    https://github.com/fmtlib/fmt/pull/3240,
+    https://github.com/fmtlib/fmt/pull/3256,
+    https://github.com/fmtlib/fmt/pull/3264,
+    https://github.com/fmtlib/fmt/issues/3299,
+    https://github.com/fmtlib/fmt/pull/3302,
+    https://github.com/fmtlib/fmt/pull/3312,
+    https://github.com/fmtlib/fmt/issues/3317,
+    https://github.com/fmtlib/fmt/pull/3328,
+    https://github.com/fmtlib/fmt/pull/3333,
+    https://github.com/fmtlib/fmt/pull/3369,
+    https://github.com/fmtlib/fmt/issues/3373,
+    https://github.com/fmtlib/fmt/pull/3395,
+    https://github.com/fmtlib/fmt/pull/3406,
+    https://github.com/fmtlib/fmt/pull/3411).
+    Thanks @dimztimz, @phprus, @DavidKorczynski, @ChrisThrasher,
+    @FrancoisCarouge, @kennyweiss, @luzpaz, @codeinred, @Mixaill, @joycebrum,
+    @kevinhwang and @Vertexwahn.
+
+-   Fixed a regression in handling empty format specifiers after a colon
+    (`{:}`) (https://github.com/fmtlib/fmt/pull/3086). Thanks @oxidase.
+
+-   Worked around a broken implementation of
+    `std::is_constant_evaluated` in some versions of libstdc++ on clang
+    (https://github.com/fmtlib/fmt/issues/3247,
+    https://github.com/fmtlib/fmt/pull/3281). Thanks @phprus.
+
+-   Fixed formatting of volatile variables
+    (https://github.com/fmtlib/fmt/pull/3068).
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/pull/3057,
+    https://github.com/fmtlib/fmt/pull/3066,
+    https://github.com/fmtlib/fmt/pull/3072,
+    https://github.com/fmtlib/fmt/pull/3082,
+    https://github.com/fmtlib/fmt/pull/3091,
+    https://github.com/fmtlib/fmt/issues/3092,
+    https://github.com/fmtlib/fmt/pull/3093,
+    https://github.com/fmtlib/fmt/pull/3095,
+    https://github.com/fmtlib/fmt/issues/3096,
+    https://github.com/fmtlib/fmt/pull/3097,
+    https://github.com/fmtlib/fmt/issues/3128,
+    https://github.com/fmtlib/fmt/pull/3129,
+    https://github.com/fmtlib/fmt/pull/3137,
+    https://github.com/fmtlib/fmt/pull/3139,
+    https://github.com/fmtlib/fmt/issues/3140,
+    https://github.com/fmtlib/fmt/pull/3142,
+    https://github.com/fmtlib/fmt/issues/3149,
+    https://github.com/fmtlib/fmt/pull/3150,
+    https://github.com/fmtlib/fmt/issues/3154,
+    https://github.com/fmtlib/fmt/issues/3163,
+    https://github.com/fmtlib/fmt/issues/3178,
+    https://github.com/fmtlib/fmt/pull/3184,
+    https://github.com/fmtlib/fmt/pull/3196,
+    https://github.com/fmtlib/fmt/issues/3204,
+    https://github.com/fmtlib/fmt/pull/3206,
+    https://github.com/fmtlib/fmt/pull/3208,
+    https://github.com/fmtlib/fmt/issues/3213,
+    https://github.com/fmtlib/fmt/pull/3216,
+    https://github.com/fmtlib/fmt/issues/3224,
+    https://github.com/fmtlib/fmt/issues/3226,
+    https://github.com/fmtlib/fmt/issues/3228,
+    https://github.com/fmtlib/fmt/pull/3229,
+    https://github.com/fmtlib/fmt/pull/3259,
+    https://github.com/fmtlib/fmt/issues/3274,
+    https://github.com/fmtlib/fmt/issues/3287,
+    https://github.com/fmtlib/fmt/pull/3288,
+    https://github.com/fmtlib/fmt/issues/3292,
+    https://github.com/fmtlib/fmt/pull/3295,
+    https://github.com/fmtlib/fmt/pull/3296,
+    https://github.com/fmtlib/fmt/issues/3298,
+    https://github.com/fmtlib/fmt/issues/3325,
+    https://github.com/fmtlib/fmt/pull/3326,
+    https://github.com/fmtlib/fmt/issues/3334,
+    https://github.com/fmtlib/fmt/issues/3342,
+    https://github.com/fmtlib/fmt/pull/3343,
+    https://github.com/fmtlib/fmt/issues/3351,
+    https://github.com/fmtlib/fmt/pull/3352,
+    https://github.com/fmtlib/fmt/pull/3362,
+    https://github.com/fmtlib/fmt/issues/3365,
+    https://github.com/fmtlib/fmt/pull/3366,
+    https://github.com/fmtlib/fmt/pull/3374,
+    https://github.com/fmtlib/fmt/issues/3377,
+    https://github.com/fmtlib/fmt/pull/3378,
+    https://github.com/fmtlib/fmt/issues/3381,
+    https://github.com/fmtlib/fmt/pull/3398,
+    https://github.com/fmtlib/fmt/pull/3413,
+    https://github.com/fmtlib/fmt/issues/3415).
+    Thanks @phprus, @gsjaardema, @NewbieOrange, @EngineLessCC, @asmaloney,
+    @HazardyKnusperkeks, @sergiud, @Youw, @thesmurph, @czudziakm,
+    @Roman-Koshelev, @chronoxor, @ShawnZhong, @russelltg, @glebm, @tmartin-gh,
+    @Zhaojun-Liu, @louiswins and @mogemimi.
+
+# 9.1.0 - 2022-08-27
+
+-   `fmt::formatted_size` now works at compile time
+    (https://github.com/fmtlib/fmt/pull/3026). For example
+    ([godbolt](https://godbolt.org/z/1MW5rMdf8)):
+
+    ```c++
+    #include <fmt/compile.h>
+
+    int main() {
+      using namespace fmt::literals;
+      constexpr size_t n = fmt::formatted_size("{}"_cf, 42);
+      fmt::print("{}\n", n); // prints 2
+    }
+    ```
+
+    Thanks @marksantaniello.
+
+-   Fixed handling of invalid UTF-8
+    (https://github.com/fmtlib/fmt/pull/3038,
+    https://github.com/fmtlib/fmt/pull/3044,
+    https://github.com/fmtlib/fmt/pull/3056).
+    Thanks @phprus and @skeeto.
+
+-   Improved Unicode support in `ostream` overloads of `print`
+    (https://github.com/fmtlib/fmt/pull/2994,
+    https://github.com/fmtlib/fmt/pull/3001,
+    https://github.com/fmtlib/fmt/pull/3025). Thanks @dimztimz.
+
+-   Fixed handling of the sign specifier in localized formatting on
+    systems with 32-bit `wchar_t`
+    (https://github.com/fmtlib/fmt/issues/3041).
+
+-   Added support for wide streams to `fmt::streamed`
+    (https://github.com/fmtlib/fmt/pull/2994). Thanks @phprus.
+
+-   Added the `n` specifier that disables the output of delimiters when
+    formatting ranges (https://github.com/fmtlib/fmt/pull/2981,
+    https://github.com/fmtlib/fmt/pull/2983). For example
+    ([godbolt](https://godbolt.org/z/roKqGdj8c)):
+
+    ```c++
+    #include <fmt/ranges.h>
+    #include <vector>
+
+    int main() {
+      auto v = std::vector{1, 2, 3};
+      fmt::print("{:n}\n", v); // prints 1, 2, 3
+    }
+    ```
+
+    Thanks @BRevzin.
+
+-   Worked around problematic `std::string_view` constructors introduced
+    in C++23 (https://github.com/fmtlib/fmt/issues/3030,
+    https://github.com/fmtlib/fmt/issues/3050). Thanks @strega-nil-ms.
+
+-   Improve handling (exclusion) of recursive ranges
+    (https://github.com/fmtlib/fmt/issues/2968,
+    https://github.com/fmtlib/fmt/pull/2974). Thanks @Dani-Hub.
+
+-   Improved error reporting in format string compilation
+    (https://github.com/fmtlib/fmt/issues/3055).
+
+-   Improved the implementation of
+    [Dragonbox](https://github.com/jk-jeon/dragonbox), the algorithm
+    used for the default floating-point formatting
+    (https://github.com/fmtlib/fmt/pull/2984). Thanks @jk-jeon.
+
+-   Fixed issues with floating-point formatting on exotic platforms.
+
+-   Improved the implementation of chrono formatting
+    (https://github.com/fmtlib/fmt/pull/3010). Thanks @phprus.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/pull/2966,
+    https://github.com/fmtlib/fmt/pull/3009,
+    https://github.com/fmtlib/fmt/issues/3020,
+    https://github.com/fmtlib/fmt/pull/3037).
+    Thanks @mwinterb, @jcelerier and @remiburtin.
+
+-   Improved build configuration
+    (https://github.com/fmtlib/fmt/pull/2991,
+    https://github.com/fmtlib/fmt/pull/2995,
+    https://github.com/fmtlib/fmt/issues/3004,
+    https://github.com/fmtlib/fmt/pull/3007,
+    https://github.com/fmtlib/fmt/pull/3040).
+    Thanks @dimztimz and @hwhsu1231.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/issues/2969,
+    https://github.com/fmtlib/fmt/pull/2971,
+    https://github.com/fmtlib/fmt/issues/2975,
+    https://github.com/fmtlib/fmt/pull/2982,
+    https://github.com/fmtlib/fmt/pull/2985,
+    https://github.com/fmtlib/fmt/issues/2988,
+    https://github.com/fmtlib/fmt/issues/2989,
+    https://github.com/fmtlib/fmt/issues/3000,
+    https://github.com/fmtlib/fmt/issues/3006,
+    https://github.com/fmtlib/fmt/issues/3014,
+    https://github.com/fmtlib/fmt/issues/3015,
+    https://github.com/fmtlib/fmt/pull/3021,
+    https://github.com/fmtlib/fmt/issues/3023,
+    https://github.com/fmtlib/fmt/pull/3024,
+    https://github.com/fmtlib/fmt/pull/3029,
+    https://github.com/fmtlib/fmt/pull/3043,
+    https://github.com/fmtlib/fmt/issues/3052,
+    https://github.com/fmtlib/fmt/pull/3053,
+    https://github.com/fmtlib/fmt/pull/3054).
+    Thanks @h-friederich, @dimztimz, @olupton, @bernhardmgruber and @phprus.
+
+# 9.0.0 - 2022-07-04
+
+-   Switched to the internal floating point formatter for all decimal
+    presentation formats. In particular this results in consistent
+    rounding on all platforms and removing the `s[n]printf` fallback for
+    decimal FP formatting.
+
+-   Compile-time floating point formatting no longer requires the
+    header-only mode. For example
+    ([godbolt](https://godbolt.org/z/G37PTeG3b)):
+
+    ```c++
+    #include <array>
+    #include <fmt/compile.h>
+
+    consteval auto compile_time_dtoa(double value) -> std::array<char, 10> {
+      auto result = std::array<char, 10>();
+      fmt::format_to(result.data(), FMT_COMPILE("{}"), value);
+      return result;
+    }
+
+    constexpr auto answer = compile_time_dtoa(0.42);
+    ```
+
+    works with the default settings.
+
+-   Improved the implementation of
+    [Dragonbox](https://github.com/jk-jeon/dragonbox), the algorithm
+    used for the default floating-point formatting
+    (https://github.com/fmtlib/fmt/pull/2713,
+    https://github.com/fmtlib/fmt/pull/2750). Thanks @jk-jeon.
+
+-   Made `fmt::to_string` work with `__float128`. This uses the internal
+    FP formatter and works even on system without `__float128` support
+    in `[s]printf`.
+
+-   Disabled automatic `std::ostream` insertion operator (`operator<<`)
+    discovery when `fmt/ostream.h` is included to prevent ODR
+    violations. You can get the old behavior by defining
+    `FMT_DEPRECATED_OSTREAM` but this will be removed in the next major
+    release. Use `fmt::streamed` or `fmt::ostream_formatter` to enable
+    formatting via `std::ostream` instead.
+
+-   Added `fmt::ostream_formatter` that can be used to write `formatter`
+    specializations that perform formatting via `std::ostream`. For
+    example ([godbolt](https://godbolt.org/z/5sEc5qMsf)):
+
+    ```c++
+    #include <fmt/ostream.h>
+
+    struct date {
+      int year, month, day;
+
+      friend std::ostream& operator<<(std::ostream& os, const date& d) {
+        return os << d.year << '-' << d.month << '-' << d.day;
+      }
+    };
+
+    template <> struct fmt::formatter<date> : ostream_formatter {};
+
+    std::string s = fmt::format("The date is {}", date{2012, 12, 9});
+    // s == "The date is 2012-12-9"
+    ```
+
+-   Added the `fmt::streamed` function that takes an object and formats
+    it via `std::ostream`. For example
+    ([godbolt](https://godbolt.org/z/5G3346G1f)):
+
+    ```c++
+    #include <thread>
+    #include <fmt/ostream.h>
+
+    int main() {
+      fmt::print("Current thread id: {}\n",
+                 fmt::streamed(std::this_thread::get_id()));
+    }
+    ```
+
+    Note that `fmt/std.h` provides a `formatter` specialization for
+    `std::thread::id` so you don\'t need to format it via
+    `std::ostream`.
+
+-   Deprecated implicit conversions of unscoped enums to integers for
+    consistency with scoped enums.
+
+-   Added an argument-dependent lookup based `format_as` extension API
+    to simplify formatting of enums.
+
+-   Added experimental `std::variant` formatting support
+    (https://github.com/fmtlib/fmt/pull/2941). For example
+    ([godbolt](https://godbolt.org/z/KG9z6cq68)):
+
+    ```c++
+    #include <variant>
+    #include <fmt/std.h>
+
+    int main() {
+      auto v = std::variant<int, std::string>(42);
+      fmt::print("{}\n", v);
+    }
+    ```
+
+    prints:
+
+        variant(42)
+
+    Thanks @jehelset.
+
+-   Added experimental `std::filesystem::path` formatting support
+    (https://github.com/fmtlib/fmt/issues/2865,
+    https://github.com/fmtlib/fmt/pull/2902,
+    https://github.com/fmtlib/fmt/issues/2917,
+    https://github.com/fmtlib/fmt/pull/2918). For example
+    ([godbolt](https://godbolt.org/z/o44dMexEb)):
+
+    ```c++
+    #include <filesystem>
+    #include <fmt/std.h>
+
+    int main() {
+      fmt::print("There is no place like {}.", std::filesystem::path("/home"));
+    }
+    ```
+
+    prints:
+
+        There is no place like "/home".
+
+    Thanks @phprus.
+
+-   Added a `std::thread::id` formatter to `fmt/std.h`. For example
+    ([godbolt](https://godbolt.org/z/j1azbYf3E)):
+
+    ```c++
+    #include <thread>
+    #include <fmt/std.h>
+
+    int main() {
+      fmt::print("Current thread id: {}\n", std::this_thread::get_id());
+    }
+    ```
+
+-   Added `fmt::styled` that applies a text style to an individual
+    argument (https://github.com/fmtlib/fmt/pull/2793). For
+    example ([godbolt](https://godbolt.org/z/vWGW7v5M6)):
+
+    ```c++
+    #include <fmt/chrono.h>
+    #include <fmt/color.h>
+
+    int main() {
+      auto now = std::chrono::system_clock::now();
+      fmt::print(
+        "[{}] {}: {}\n",
+        fmt::styled(now, fmt::emphasis::bold),
+        fmt::styled("error", fg(fmt::color::red)),
+        "something went wrong");
+    }
+    ```
+
+    prints
+
+    ![](https://user-images.githubusercontent.com/576385/175071215-12809244-dab0-4005-96d8-7cd911c964d5.png)
+
+    Thanks @rbrugo.
+
+-   Made `fmt::print` overload for text styles correctly handle UTF-8
+    (https://github.com/fmtlib/fmt/issues/2681,
+    https://github.com/fmtlib/fmt/pull/2701). Thanks @AlexGuteniev.
+
+-   Fixed Unicode handling when writing to an ostream.
+
+-   Added support for nested specifiers to range formatting
+    (https://github.com/fmtlib/fmt/pull/2673). For example
+    ([godbolt](https://godbolt.org/z/xd3Gj38cf)):
+
+    ```c++
+    #include <vector>
+    #include <fmt/ranges.h>
+
+    int main() {
+      fmt::print("{::#x}\n", std::vector{10, 20, 30});
+    }
+    ```
+
+    prints `[0xa, 0x14, 0x1e]`.
+
+    Thanks @BRevzin.
+
+-   Implemented escaping of wide strings in ranges
+    (https://github.com/fmtlib/fmt/pull/2904). Thanks @phprus.
+
+-   Added support for ranges with `begin` / `end` found via the
+    argument-dependent lookup
+    (https://github.com/fmtlib/fmt/pull/2807). Thanks @rbrugo.
+
+-   Fixed formatting of certain kinds of ranges of ranges
+    (https://github.com/fmtlib/fmt/pull/2787). Thanks @BRevzin.
+
+-   Fixed handling of maps with element types other than `std::pair`
+    (https://github.com/fmtlib/fmt/pull/2944). Thanks @BrukerJWD.
+
+-   Made tuple formatter enabled only if elements are formattable
+    (https://github.com/fmtlib/fmt/issues/2939,
+    https://github.com/fmtlib/fmt/pull/2940). Thanks @jehelset.
+
+-   Made `fmt::join` compatible with format string compilation
+    (https://github.com/fmtlib/fmt/issues/2719,
+    https://github.com/fmtlib/fmt/pull/2720). Thanks @phprus.
+
+-   Made compile-time checks work with named arguments of custom types
+    and `std::ostream` `print` overloads
+    (https://github.com/fmtlib/fmt/issues/2816,
+    https://github.com/fmtlib/fmt/issues/2817,
+    https://github.com/fmtlib/fmt/pull/2819). Thanks @timsong-cpp.
+
+-   Removed `make_args_checked` because it is no longer needed for
+    compile-time checks
+    (https://github.com/fmtlib/fmt/pull/2760). Thanks @phprus.
+
+-   Removed the following deprecated APIs: `_format`, `arg_join`, the
+    `format_to` overload that takes a memory buffer, `[v]fprintf` that
+    takes an `ostream`.
+
+-   Removed the deprecated implicit conversion of `[const] signed char*`
+    and `[const] unsigned char*` to C strings.
+
+-   Removed the deprecated `fmt/locale.h`.
+
+-   Replaced the deprecated `fileno()` with `descriptor()` in
+    `buffered_file`.
+
+-   Moved `to_string_view` to the `detail` namespace since it\'s an
+    implementation detail.
+
+-   Made access mode of a created file consistent with `fopen` by
+    setting `S_IWGRP` and `S_IWOTH`
+    (https://github.com/fmtlib/fmt/pull/2733). Thanks @arogge.
+
+-   Removed a redundant buffer resize when formatting to `std::ostream`
+    (https://github.com/fmtlib/fmt/issues/2842,
+    https://github.com/fmtlib/fmt/pull/2843). Thanks @jcelerier.
+
+-   Made precision computation for strings consistent with width
+    (https://github.com/fmtlib/fmt/issues/2888).
+
+-   Fixed handling of locale separators in floating point formatting
+    (https://github.com/fmtlib/fmt/issues/2830).
+
+-   Made sign specifiers work with `__int128_t`
+    (https://github.com/fmtlib/fmt/issues/2773).
+
+-   Improved support for systems such as CHERI with extra data stored in
+    pointers (https://github.com/fmtlib/fmt/pull/2932).
+    Thanks @davidchisnall.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/pull/2706,
+    https://github.com/fmtlib/fmt/pull/2712,
+    https://github.com/fmtlib/fmt/pull/2789,
+    https://github.com/fmtlib/fmt/pull/2803,
+    https://github.com/fmtlib/fmt/pull/2805,
+    https://github.com/fmtlib/fmt/pull/2815,
+    https://github.com/fmtlib/fmt/pull/2924).
+    Thanks @BRevzin, @Pokechu22, @setoye, @rtobar, @rbrugo, @anoonD and
+    @leha-bot.
+
+-   Improved build configuration
+    (https://github.com/fmtlib/fmt/pull/2766,
+    https://github.com/fmtlib/fmt/pull/2772,
+    https://github.com/fmtlib/fmt/pull/2836,
+    https://github.com/fmtlib/fmt/pull/2852,
+    https://github.com/fmtlib/fmt/pull/2907,
+    https://github.com/fmtlib/fmt/pull/2913,
+    https://github.com/fmtlib/fmt/pull/2914).
+    Thanks @kambala-decapitator, @mattiasljungstrom, @kieselnb, @nathannaveen
+    and @Vertexwahn.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/issues/2408,
+    https://github.com/fmtlib/fmt/issues/2507,
+    https://github.com/fmtlib/fmt/issues/2697,
+    https://github.com/fmtlib/fmt/issues/2715,
+    https://github.com/fmtlib/fmt/issues/2717,
+    https://github.com/fmtlib/fmt/pull/2722,
+    https://github.com/fmtlib/fmt/pull/2724,
+    https://github.com/fmtlib/fmt/pull/2725,
+    https://github.com/fmtlib/fmt/issues/2726,
+    https://github.com/fmtlib/fmt/pull/2728,
+    https://github.com/fmtlib/fmt/pull/2732,
+    https://github.com/fmtlib/fmt/issues/2738,
+    https://github.com/fmtlib/fmt/pull/2742,
+    https://github.com/fmtlib/fmt/issues/2744,
+    https://github.com/fmtlib/fmt/issues/2745,
+    https://github.com/fmtlib/fmt/issues/2746,
+    https://github.com/fmtlib/fmt/issues/2754,
+    https://github.com/fmtlib/fmt/pull/2755,
+    https://github.com/fmtlib/fmt/issues/2757,
+    https://github.com/fmtlib/fmt/pull/2758,
+    https://github.com/fmtlib/fmt/issues/2761,
+    https://github.com/fmtlib/fmt/pull/2762,
+    https://github.com/fmtlib/fmt/issues/2763,
+    https://github.com/fmtlib/fmt/pull/2765,
+    https://github.com/fmtlib/fmt/issues/2769,
+    https://github.com/fmtlib/fmt/pull/2770,
+    https://github.com/fmtlib/fmt/issues/2771,
+    https://github.com/fmtlib/fmt/issues/2777,
+    https://github.com/fmtlib/fmt/pull/2779,
+    https://github.com/fmtlib/fmt/pull/2782,
+    https://github.com/fmtlib/fmt/pull/2783,
+    https://github.com/fmtlib/fmt/issues/2794,
+    https://github.com/fmtlib/fmt/issues/2796,
+    https://github.com/fmtlib/fmt/pull/2797,
+    https://github.com/fmtlib/fmt/pull/2801,
+    https://github.com/fmtlib/fmt/pull/2802,
+    https://github.com/fmtlib/fmt/issues/2808,
+    https://github.com/fmtlib/fmt/issues/2818,
+    https://github.com/fmtlib/fmt/pull/2819,
+    https://github.com/fmtlib/fmt/issues/2829,
+    https://github.com/fmtlib/fmt/issues/2835,
+    https://github.com/fmtlib/fmt/issues/2848,
+    https://github.com/fmtlib/fmt/issues/2860,
+    https://github.com/fmtlib/fmt/pull/2861,
+    https://github.com/fmtlib/fmt/pull/2882,
+    https://github.com/fmtlib/fmt/issues/2886,
+    https://github.com/fmtlib/fmt/issues/2891,
+    https://github.com/fmtlib/fmt/pull/2892,
+    https://github.com/fmtlib/fmt/issues/2895,
+    https://github.com/fmtlib/fmt/issues/2896,
+    https://github.com/fmtlib/fmt/pull/2903,
+    https://github.com/fmtlib/fmt/issues/2906,
+    https://github.com/fmtlib/fmt/issues/2908,
+    https://github.com/fmtlib/fmt/pull/2909,
+    https://github.com/fmtlib/fmt/issues/2920,
+    https://github.com/fmtlib/fmt/pull/2922,
+    https://github.com/fmtlib/fmt/pull/2927,
+    https://github.com/fmtlib/fmt/pull/2929,
+    https://github.com/fmtlib/fmt/issues/2936,
+    https://github.com/fmtlib/fmt/pull/2937,
+    https://github.com/fmtlib/fmt/pull/2938,
+    https://github.com/fmtlib/fmt/pull/2951,
+    https://github.com/fmtlib/fmt/issues/2954,
+    https://github.com/fmtlib/fmt/pull/2957,
+    https://github.com/fmtlib/fmt/issues/2958,
+    https://github.com/fmtlib/fmt/pull/2960).
+    Thanks @matrackif @Tobi823, @ivan-volnov, @VasiliPupkin256,
+    @federico-busato, @barcharcraz, @jk-jeon, @HazardyKnusperkeks, @dalboris,
+    @seanm, @gsjaardema, @timsong-cpp, @seanm, @frithrah, @chronoxor, @Agga,
+    @madmaxoft, @JurajX, @phprus and @Dani-Hub.
+
+# 8.1.1 - 2022-01-06
+
+-   Restored ABI compatibility with version 8.0.x
+    (https://github.com/fmtlib/fmt/issues/2695,
+    https://github.com/fmtlib/fmt/pull/2696). Thanks @saraedum.
+-   Fixed chrono formatting on big endian systems
+    (https://github.com/fmtlib/fmt/issues/2698,
+    https://github.com/fmtlib/fmt/pull/2699).
+    Thanks @phprus and @xvitaly.
+-   Fixed a linkage error with mingw
+    (https://github.com/fmtlib/fmt/issues/2691,
+    https://github.com/fmtlib/fmt/pull/2692). Thanks @rbberger.
+
+# 8.1.0 - 2022-01-02
+
+-   Optimized chrono formatting
+    (https://github.com/fmtlib/fmt/pull/2500,
+    https://github.com/fmtlib/fmt/pull/2537,
+    https://github.com/fmtlib/fmt/issues/2541,
+    https://github.com/fmtlib/fmt/pull/2544,
+    https://github.com/fmtlib/fmt/pull/2550,
+    https://github.com/fmtlib/fmt/pull/2551,
+    https://github.com/fmtlib/fmt/pull/2576,
+    https://github.com/fmtlib/fmt/issues/2577,
+    https://github.com/fmtlib/fmt/pull/2586,
+    https://github.com/fmtlib/fmt/pull/2591,
+    https://github.com/fmtlib/fmt/pull/2594,
+    https://github.com/fmtlib/fmt/pull/2602,
+    https://github.com/fmtlib/fmt/pull/2617,
+    https://github.com/fmtlib/fmt/issues/2628,
+    https://github.com/fmtlib/fmt/pull/2633,
+    https://github.com/fmtlib/fmt/issues/2670,
+    https://github.com/fmtlib/fmt/pull/2671).
+
+    Processing of some specifiers such as `%z` and `%Y` is now up to
+    10-20 times faster, for example on GCC 11 with libstdc++:
+
+        ----------------------------------------------------------------------------
+        Benchmark                                  Before             After
+        ----------------------------------------------------------------------------
+        FMTFormatter_z                             261 ns             26.3 ns
+        FMTFormatterCompile_z                      246 ns             11.6 ns
+        FMTFormatter_Y                             263 ns             26.1 ns
+        FMTFormatterCompile_Y                      244 ns             10.5 ns
+        ----------------------------------------------------------------------------
+
+    Thanks @phprus and @toughengineer.
+
+-   Implemented subsecond formatting for chrono durations
+    (https://github.com/fmtlib/fmt/pull/2623). For example
+    ([godbolt](https://godbolt.org/z/es7vWTETe)):
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      fmt::print("{:%S}", std::chrono::milliseconds(1234));
+    }
+    ```
+
+    prints \"01.234\".
+
+    Thanks @matrackif.
+
+-   Fixed handling of precision 0 when formatting chrono durations
+    (https://github.com/fmtlib/fmt/issues/2587,
+    https://github.com/fmtlib/fmt/pull/2588). Thanks @lukester1975.
+
+-   Fixed an overflow on invalid inputs in the `tm` formatter
+    (https://github.com/fmtlib/fmt/pull/2564). Thanks @phprus.
+
+-   Added `fmt::group_digits` that formats integers with a non-localized
+    digit separator (comma) for groups of three digits. For example
+    ([godbolt](https://godbolt.org/z/TxGxG9Poq)):
+
+    ```c++
+    #include <fmt/format.h>
+
+    int main() {
+      fmt::print("{} dollars", fmt::group_digits(1000000));
+    }
+    ```
+
+    prints \"1,000,000 dollars\".
+
+-   Added support for faint, conceal, reverse and blink text styles
+    (https://github.com/fmtlib/fmt/pull/2394):
+
+    <https://user-images.githubusercontent.com/576385/147710227-c68f5317-f8fa-42c3-9123-7c4ba3c398cb.mp4>
+
+    Thanks @benit8 and @data-man.
+
+-   Added experimental support for compile-time floating point
+    formatting (https://github.com/fmtlib/fmt/pull/2426,
+    https://github.com/fmtlib/fmt/pull/2470). It is currently
+    limited to the header-only mode. Thanks @alexezeder.
+
+-   Added UDL-based named argument support to compile-time format string
+    checks (https://github.com/fmtlib/fmt/issues/2640,
+    https://github.com/fmtlib/fmt/pull/2649). For example
+    ([godbolt](https://godbolt.org/z/ohGbbvonv)):
+
+    ```c++
+    #include <fmt/format.h>
+
+    int main() {
+      using namespace fmt::literals;
+      fmt::print("{answer:s}", "answer"_a=42);
+    }
+    ```
+
+    gives a compile-time error on compilers with C++20 `consteval` and
+    non-type template parameter support (gcc 10+) because `s` is not a
+    valid format specifier for an integer.
+
+    Thanks @alexezeder.
+
+-   Implemented escaping of string range elements. For example
+    ([godbolt](https://godbolt.org/z/rKvM1vKf3)):
+
+    ```c++
+    #include <fmt/ranges.h>
+    #include <vector>
+
+    int main() {
+      fmt::print("{}", std::vector<std::string>{"\naan"});
+    }
+    ```
+
+    is now printed as:
+
+        ["\naan"]
+
+    instead of:
+
+        ["
+        aan"]
+
+-   Added an experimental `?` specifier for escaping strings.
+    (https://github.com/fmtlib/fmt/pull/2674). Thanks @BRevzin.
+
+-   Switched to JSON-like representation of maps and sets for
+    consistency with Python\'s `str.format`. For example
+    ([godbolt](https://godbolt.org/z/seKjoY9W5)):
+
+    ```c++
+    #include <fmt/ranges.h>
+    #include <map>
+
+    int main() {
+      fmt::print("{}", std::map<std::string, int>{{"answer", 42}});
+    }
+    ```
+
+    is now printed as:
+
+        {"answer": 42}
+
+-   Extended `fmt::join` to support C++20-only ranges
+    (https://github.com/fmtlib/fmt/pull/2549). Thanks @BRevzin.
+
+-   Optimized handling of non-const-iterable ranges and implemented
+    initial support for non-const-formattable types.
+
+-   Disabled implicit conversions of scoped enums to integers that was
+    accidentally introduced in earlier versions
+    (https://github.com/fmtlib/fmt/pull/1841).
+
+-   Deprecated implicit conversion of `[const] signed char*` and
+    `[const] unsigned char*` to C strings.
+
+-   Deprecated `_format`, a legacy UDL-based format API
+    (https://github.com/fmtlib/fmt/pull/2646). Thanks @alexezeder.
+
+-   Marked `format`, `formatted_size` and `to_string` as `[[nodiscard]]`
+    (https://github.com/fmtlib/fmt/pull/2612). @0x8000-0000.
+
+-   Added missing diagnostic when trying to format function and member
+    pointers as well as objects convertible to pointers which is
+    explicitly disallowed
+    (https://github.com/fmtlib/fmt/issues/2598,
+    https://github.com/fmtlib/fmt/pull/2609,
+    https://github.com/fmtlib/fmt/pull/2610). Thanks @AlexGuteniev.
+
+-   Optimized writing to a contiguous buffer with `format_to_n`
+    (https://github.com/fmtlib/fmt/pull/2489). Thanks @Roman-Koshelev.
+
+-   Optimized writing to non-`char` buffers
+    (https://github.com/fmtlib/fmt/pull/2477). Thanks @Roman-Koshelev.
+
+-   Decimal point is now localized when using the `L` specifier.
+
+-   Improved floating point formatter implementation
+    (https://github.com/fmtlib/fmt/pull/2498,
+    https://github.com/fmtlib/fmt/pull/2499). Thanks @Roman-Koshelev.
+
+-   Fixed handling of very large precision in fixed format
+    (https://github.com/fmtlib/fmt/pull/2616).
+
+-   Made a table of cached powers used in FP formatting static
+    (https://github.com/fmtlib/fmt/pull/2509). Thanks @jk-jeon.
+
+-   Resolved a lookup ambiguity with C++20 format-related functions due
+    to ADL (https://github.com/fmtlib/fmt/issues/2639,
+    https://github.com/fmtlib/fmt/pull/2641). Thanks @mkurdej.
+
+-   Removed unnecessary inline namespace qualification
+    (https://github.com/fmtlib/fmt/issues/2642,
+    https://github.com/fmtlib/fmt/pull/2643). Thanks @mkurdej.
+
+-   Implemented argument forwarding in `format_to_n`
+    (https://github.com/fmtlib/fmt/issues/2462,
+    https://github.com/fmtlib/fmt/pull/2463). Thanks @owent.
+
+-   Fixed handling of implicit conversions in `fmt::to_string` and
+    format string compilation
+    (https://github.com/fmtlib/fmt/issues/2565).
+
+-   Changed the default access mode of files created by
+    `fmt::output_file` to `-rw-r--r--` for consistency with `fopen`
+    (https://github.com/fmtlib/fmt/issues/2530).
+
+-   Make `fmt::ostream::flush` public
+    (https://github.com/fmtlib/fmt/issues/2435).
+
+-   Improved C++14/17 attribute detection
+    (https://github.com/fmtlib/fmt/pull/2615). Thanks @AlexGuteniev.
+
+-   Improved `consteval` detection for MSVC
+    (https://github.com/fmtlib/fmt/pull/2559). Thanks @DanielaE.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/2406,
+    https://github.com/fmtlib/fmt/pull/2446,
+    https://github.com/fmtlib/fmt/issues/2493,
+    https://github.com/fmtlib/fmt/issues/2513,
+    https://github.com/fmtlib/fmt/pull/2515,
+    https://github.com/fmtlib/fmt/issues/2522,
+    https://github.com/fmtlib/fmt/pull/2562,
+    https://github.com/fmtlib/fmt/pull/2575,
+    https://github.com/fmtlib/fmt/pull/2606,
+    https://github.com/fmtlib/fmt/pull/2620,
+    https://github.com/fmtlib/fmt/issues/2676).
+    Thanks @sobolevn, @UnePierre, @zhsj, @phprus, @ericcurtin and @Lounarok.
+
+-   Improved fuzzers and added a fuzzer for chrono timepoint formatting
+    (https://github.com/fmtlib/fmt/pull/2461,
+    https://github.com/fmtlib/fmt/pull/2469). @pauldreik,
+
+-   Added the `FMT_SYSTEM_HEADERS` CMake option setting which marks
+    {fmt}\'s headers as system. It can be used to suppress warnings
+    (https://github.com/fmtlib/fmt/issues/2644,
+    https://github.com/fmtlib/fmt/pull/2651). Thanks @alexezeder.
+
+-   Added the Bazel build system support
+    (https://github.com/fmtlib/fmt/pull/2505,
+    https://github.com/fmtlib/fmt/pull/2516). Thanks @Vertexwahn.
+
+-   Improved build configuration and tests
+    (https://github.com/fmtlib/fmt/issues/2437,
+    https://github.com/fmtlib/fmt/pull/2558,
+    https://github.com/fmtlib/fmt/pull/2648,
+    https://github.com/fmtlib/fmt/pull/2650,
+    https://github.com/fmtlib/fmt/pull/2663,
+    https://github.com/fmtlib/fmt/pull/2677).
+    Thanks @DanielaE, @alexezeder and @phprus.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/pull/2353,
+    https://github.com/fmtlib/fmt/pull/2356,
+    https://github.com/fmtlib/fmt/pull/2399,
+    https://github.com/fmtlib/fmt/issues/2408,
+    https://github.com/fmtlib/fmt/pull/2414,
+    https://github.com/fmtlib/fmt/pull/2427,
+    https://github.com/fmtlib/fmt/pull/2432,
+    https://github.com/fmtlib/fmt/pull/2442,
+    https://github.com/fmtlib/fmt/pull/2434,
+    https://github.com/fmtlib/fmt/issues/2439,
+    https://github.com/fmtlib/fmt/pull/2447,
+    https://github.com/fmtlib/fmt/pull/2450,
+    https://github.com/fmtlib/fmt/issues/2455,
+    https://github.com/fmtlib/fmt/issues/2465,
+    https://github.com/fmtlib/fmt/issues/2472,
+    https://github.com/fmtlib/fmt/issues/2474,
+    https://github.com/fmtlib/fmt/pull/2476,
+    https://github.com/fmtlib/fmt/issues/2478,
+    https://github.com/fmtlib/fmt/issues/2479,
+    https://github.com/fmtlib/fmt/issues/2481,
+    https://github.com/fmtlib/fmt/pull/2482,
+    https://github.com/fmtlib/fmt/pull/2483,
+    https://github.com/fmtlib/fmt/issues/2490,
+    https://github.com/fmtlib/fmt/pull/2491,
+    https://github.com/fmtlib/fmt/pull/2510,
+    https://github.com/fmtlib/fmt/pull/2518,
+    https://github.com/fmtlib/fmt/issues/2528,
+    https://github.com/fmtlib/fmt/pull/2529,
+    https://github.com/fmtlib/fmt/pull/2539,
+    https://github.com/fmtlib/fmt/issues/2540,
+    https://github.com/fmtlib/fmt/pull/2545,
+    https://github.com/fmtlib/fmt/pull/2555,
+    https://github.com/fmtlib/fmt/issues/2557,
+    https://github.com/fmtlib/fmt/issues/2570,
+    https://github.com/fmtlib/fmt/pull/2573,
+    https://github.com/fmtlib/fmt/pull/2582,
+    https://github.com/fmtlib/fmt/issues/2605,
+    https://github.com/fmtlib/fmt/pull/2611,
+    https://github.com/fmtlib/fmt/pull/2647,
+    https://github.com/fmtlib/fmt/issues/2627,
+    https://github.com/fmtlib/fmt/pull/2630,
+    https://github.com/fmtlib/fmt/issues/2635,
+    https://github.com/fmtlib/fmt/issues/2638,
+    https://github.com/fmtlib/fmt/issues/2653,
+    https://github.com/fmtlib/fmt/issues/2654,
+    https://github.com/fmtlib/fmt/issues/2661,
+    https://github.com/fmtlib/fmt/pull/2664,
+    https://github.com/fmtlib/fmt/pull/2684).
+    Thanks @DanielaE, @mwinterb, @cdacamar, @TrebledJ, @bodomartin, @cquammen,
+    @white238, @mmarkeloff, @palacaze, @jcelerier, @mborn-adi, @BrukerJWD,
+    @spyridon97, @phprus, @oliverlee, @joshessman-llnl, @akohlmey, @timkalu,
+    @olupton, @Acretock, @alexezeder, @andrewcorrigan, @lucpelletier and
+    @HazardyKnusperkeks.
+
+# 8.0.1 - 2021-07-02
+
+-   Fixed the version number in the inline namespace
+    (https://github.com/fmtlib/fmt/issues/2374).
+-   Added a missing presentation type check for `std::string`
+    (https://github.com/fmtlib/fmt/issues/2402).
+-   Fixed a linkage error when mixing code built with clang and gcc
+    (https://github.com/fmtlib/fmt/issues/2377).
+-   Fixed documentation issues
+    (https://github.com/fmtlib/fmt/pull/2396,
+    https://github.com/fmtlib/fmt/issues/2403,
+    https://github.com/fmtlib/fmt/issues/2406). Thanks @mkurdej.
+-   Removed dead code in FP formatter (
+    https://github.com/fmtlib/fmt/pull/2398). Thanks @javierhonduco.
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/issues/2351,
+    https://github.com/fmtlib/fmt/issues/2359,
+    https://github.com/fmtlib/fmt/pull/2365,
+    https://github.com/fmtlib/fmt/issues/2368,
+    https://github.com/fmtlib/fmt/pull/2370,
+    https://github.com/fmtlib/fmt/pull/2376,
+    https://github.com/fmtlib/fmt/pull/2381,
+    https://github.com/fmtlib/fmt/pull/2382,
+    https://github.com/fmtlib/fmt/issues/2386,
+    https://github.com/fmtlib/fmt/pull/2389,
+    https://github.com/fmtlib/fmt/pull/2395,
+    https://github.com/fmtlib/fmt/pull/2397,
+    https://github.com/fmtlib/fmt/issues/2400,
+    https://github.com/fmtlib/fmt/issues/2401,
+    https://github.com/fmtlib/fmt/pull/2407).
+    Thanks @zx2c4, @AidanSun05, @mattiasljungstrom, @joemmett, @erengy,
+    @patlkli, @gsjaardema and @phprus.
+
+# 8.0.0 - 2021-06-21
+
+-   Enabled compile-time format string checks by default. For example
+    ([godbolt](https://godbolt.org/z/sMxcohGjz)):
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{:d}", "I am not a number");
+    }
+    ```
+
+    gives a compile-time error on compilers with C++20 `consteval`
+    support (gcc 10+, clang 11+) because `d` is not a valid format
+    specifier for a string.
+
+    To pass a runtime string wrap it in `fmt::runtime`:
+
+    ```c++
+    fmt::print(fmt::runtime("{:d}"), "I am not a number");
+    ```
+
+-   Added compile-time formatting
+    (https://github.com/fmtlib/fmt/pull/2019,
+    https://github.com/fmtlib/fmt/pull/2044,
+    https://github.com/fmtlib/fmt/pull/2056,
+    https://github.com/fmtlib/fmt/pull/2072,
+    https://github.com/fmtlib/fmt/pull/2075,
+    https://github.com/fmtlib/fmt/issues/2078,
+    https://github.com/fmtlib/fmt/pull/2129,
+    https://github.com/fmtlib/fmt/pull/2326). For example
+    ([godbolt](https://godbolt.org/z/Mxx9d89jM)):
+
+    ```c++
+    #include <fmt/compile.h>
+
+    consteval auto compile_time_itoa(int value) -> std::array<char, 10> {
+      auto result = std::array<char, 10>();
+      fmt::format_to(result.data(), FMT_COMPILE("{}"), value);
+      return result;
+    }
+
+    constexpr auto answer = compile_time_itoa(42);
+    ```
+
+    Most of the formatting functionality is available at compile time
+    with a notable exception of floating-point numbers and pointers.
+    Thanks @alexezeder.
+
+-   Optimized handling of format specifiers during format string
+    compilation. For example, hexadecimal formatting (`"{:x}"`) is now
+    3-7x faster than before when using `format_to` with format string
+    compilation and a stack-allocated buffer
+    (https://github.com/fmtlib/fmt/issues/1944).
+
+    Before (7.1.3):
+
+        ----------------------------------------------------------------------------
+        Benchmark                                  Time             CPU   Iterations
+        ----------------------------------------------------------------------------
+        FMTCompileOld/0                         15.5 ns         15.5 ns     43302898
+        FMTCompileOld/42                        16.6 ns         16.6 ns     43278267
+        FMTCompileOld/273123                    18.7 ns         18.6 ns     37035861
+        FMTCompileOld/9223372036854775807       19.4 ns         19.4 ns     35243000
+        ----------------------------------------------------------------------------
+
+    After (8.x):
+
+        ----------------------------------------------------------------------------
+        Benchmark                                  Time             CPU   Iterations
+        ----------------------------------------------------------------------------
+        FMTCompileNew/0                         1.99 ns         1.99 ns    360523686
+        FMTCompileNew/42                        2.33 ns         2.33 ns    279865664
+        FMTCompileNew/273123                    3.72 ns         3.71 ns    190230315
+        FMTCompileNew/9223372036854775807       5.28 ns         5.26 ns    130711631
+        ----------------------------------------------------------------------------
+
+    It is even faster than `std::to_chars` from libc++ compiled with
+    clang on macOS:
+
+        ----------------------------------------------------------------------------
+        Benchmark                                  Time             CPU   Iterations
+        ----------------------------------------------------------------------------
+        ToChars/0                               4.42 ns         4.41 ns    160196630
+        ToChars/42                              5.00 ns         4.98 ns    140735201
+        ToChars/273123                          7.26 ns         7.24 ns     95784130
+        ToChars/9223372036854775807             8.77 ns         8.75 ns     75872534
+        ----------------------------------------------------------------------------
+
+    In other cases, especially involving `std::string` construction, the
+    speed up is usually lower because handling format specifiers takes a
+    smaller fraction of the total time.
+
+-   Added the `_cf` user-defined literal to represent a compiled format
+    string. It can be used instead of the `FMT_COMPILE` macro
+    (https://github.com/fmtlib/fmt/pull/2043,
+    https://github.com/fmtlib/fmt/pull/2242):
+
+    ```c++
+    #include <fmt/compile.h>
+
+    using namespace fmt::literals;
+    auto s = fmt::format(FMT_COMPILE("{}"), 42); // 🙁 not modern
+    auto s = fmt::format("{}"_cf, 42);           // 🙂 modern as hell
+    ```
+
+    It requires compiler support for class types in non-type template
+    parameters (a C++20 feature) which is available in GCC 9.3+.
+    Thanks @alexezeder.
+
+-   Format string compilation now requires `format` functions of
+    `formatter` specializations for user-defined types to be `const`:
+
+    ```c++
+    template <> struct fmt::formatter<my_type>: formatter<string_view> {
+      template <typename FormatContext>
+      auto format(my_type obj, FormatContext& ctx) const {  // Note const here.
+        // ...
+      }
+    };
+    ```
+
+-   Added UDL-based named argument support to format string compilation
+    (https://github.com/fmtlib/fmt/pull/2243,
+    https://github.com/fmtlib/fmt/pull/2281). For example:
+
+    ```c++
+    #include <fmt/compile.h>
+
+    using namespace fmt::literals;
+    auto s = fmt::format(FMT_COMPILE("{answer}"), "answer"_a = 42);
+    ```
+
+    Here the argument named \"answer\" is resolved at compile time with
+    no runtime overhead. Thanks @alexezeder.
+
+-   Added format string compilation support to `fmt::print`
+    (https://github.com/fmtlib/fmt/issues/2280,
+    https://github.com/fmtlib/fmt/pull/2304). Thanks @alexezeder.
+
+-   Added initial support for compiling {fmt} as a C++20 module
+    (https://github.com/fmtlib/fmt/pull/2235,
+    https://github.com/fmtlib/fmt/pull/2240,
+    https://github.com/fmtlib/fmt/pull/2260,
+    https://github.com/fmtlib/fmt/pull/2282,
+    https://github.com/fmtlib/fmt/pull/2283,
+    https://github.com/fmtlib/fmt/pull/2288,
+    https://github.com/fmtlib/fmt/pull/2298,
+    https://github.com/fmtlib/fmt/pull/2306,
+    https://github.com/fmtlib/fmt/pull/2307,
+    https://github.com/fmtlib/fmt/pull/2309,
+    https://github.com/fmtlib/fmt/pull/2318,
+    https://github.com/fmtlib/fmt/pull/2324,
+    https://github.com/fmtlib/fmt/pull/2332,
+    https://github.com/fmtlib/fmt/pull/2340). Thanks @DanielaE.
+
+-   Made symbols private by default reducing shared library size
+    (https://github.com/fmtlib/fmt/pull/2301). For example
+    there was a \~15% reported reduction on one platform. Thanks @sergiud.
+
+-   Optimized includes making the result of preprocessing `fmt/format.h`
+    \~20% smaller with libstdc++/C++20 and slightly improving build
+    times (https://github.com/fmtlib/fmt/issues/1998).
+
+-   Added support of ranges with non-const `begin` / `end`
+    (https://github.com/fmtlib/fmt/pull/1953). Thanks @kitegi.
+
+-   Added support of `std::byte` and other formattable types to
+    `fmt::join` (https://github.com/fmtlib/fmt/issues/1981,
+    https://github.com/fmtlib/fmt/issues/2040,
+    https://github.com/fmtlib/fmt/pull/2050,
+    https://github.com/fmtlib/fmt/issues/2262). For example:
+
+    ```c++
+    #include <fmt/format.h>
+    #include <cstddef>
+    #include <vector>
+
+    int main() {
+      auto bytes = std::vector{std::byte(4), std::byte(2)};
+      fmt::print("{}", fmt::join(bytes, ""));
+    }
+    ```
+
+    prints \"42\".
+
+    Thanks @kamibo.
+
+-   Implemented the default format for `std::chrono::system_clock`
+    (https://github.com/fmtlib/fmt/issues/2319,
+    https://github.com/fmtlib/fmt/pull/2345). For example:
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      fmt::print("{}", std::chrono::system_clock::now());
+    }
+    ```
+
+    prints \"2021-06-18 15:22:00\" (the output depends on the current
+    date and time). Thanks @sunmy2019.
+
+-   Made more chrono specifiers locale independent by default. Use the
+    `'L'` specifier to get localized formatting. For example:
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      std::locale::global(std::locale("ru_RU.UTF-8"));
+      auto monday = std::chrono::weekday(1);
+      fmt::print("{}\n", monday);   // prints "Mon"
+      fmt::print("{:L}\n", monday); // prints "пн"
+    }
+    ```
+
+-   Improved locale handling in chrono formatting
+    (https://github.com/fmtlib/fmt/issues/2337,
+    https://github.com/fmtlib/fmt/pull/2349,
+    https://github.com/fmtlib/fmt/pull/2350). Thanks @phprus.
+
+-   Deprecated `fmt/locale.h` moving the formatting functions that take
+    a locale to `fmt/format.h` (`char`) and `fmt/xchar` (other
+    overloads). This doesn\'t introduce a dependency on `<locale>` so
+    there is virtually no compile time effect.
+
+-   Deprecated an undocumented `format_to` overload that takes
+    `basic_memory_buffer`.
+
+-   Made parameter order in `vformat_to` consistent with `format_to`
+    (https://github.com/fmtlib/fmt/issues/2327).
+
+-   Added support for time points with arbitrary durations
+    (https://github.com/fmtlib/fmt/issues/2208). For example:
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      using tp = std::chrono::time_point<
+        std::chrono::system_clock, std::chrono::seconds>;
+      fmt::print("{:%S}", tp(std::chrono::seconds(42)));
+    }
+    ```
+
+    prints \"42\".
+
+-   Formatting floating-point numbers no longer produces trailing zeros
+    by default for consistency with `std::format`. For example:
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{0:.3}", 1.1);
+    }
+    ```
+
+    prints \"1.1\". Use the `'#'` specifier to keep trailing zeros.
+
+-   Dropped a limit on the number of elements in a range and replaced
+    `{}` with `[]` as range delimiters for consistency with Python\'s
+    `str.format`.
+
+-   The `'L'` specifier for locale-specific numeric formatting can now
+    be combined with presentation specifiers as in `std::format`. For
+    example:
+
+    ```c++
+    #include <fmt/core.h>
+    #include <locale>
+
+    int main() {
+      std::locale::global(std::locale("fr_FR.UTF-8"));
+      fmt::print("{0:.2Lf}", 0.42);
+    }
+    ```
+
+    prints \"0,42\". The deprecated `'n'` specifier has been removed.
+
+-   Made the `0` specifier ignored for infinity and NaN
+    (https://github.com/fmtlib/fmt/issues/2305,
+    https://github.com/fmtlib/fmt/pull/2310). Thanks @Liedtke.
+
+-   Made the hexfloat formatting use the right alignment by default
+    (https://github.com/fmtlib/fmt/issues/2308,
+    https://github.com/fmtlib/fmt/pull/2317). Thanks @Liedtke.
+
+-   Removed the deprecated numeric alignment (`'='`). Use the `'0'`
+    specifier instead.
+
+-   Removed the deprecated `fmt/posix.h` header that has been replaced
+    with `fmt/os.h`.
+
+-   Removed the deprecated `format_to_n_context`, `format_to_n_args` and
+    `make_format_to_n_args`. They have been replaced with
+    `format_context`, `` format_args` and ``make_format_args\`\`
+    respectively.
+
+-   Moved `wchar_t`-specific functions and types to `fmt/xchar.h`. You
+    can define `FMT_DEPRECATED_INCLUDE_XCHAR` to automatically include
+    `fmt/xchar.h` from `fmt/format.h` but this will be disabled in the
+    next major release.
+
+-   Fixed handling of the `'+'` specifier in localized formatting
+    (https://github.com/fmtlib/fmt/issues/2133).
+
+-   Added support for the `'s'` format specifier that gives textual
+    representation of `bool`
+    (https://github.com/fmtlib/fmt/issues/2094,
+    https://github.com/fmtlib/fmt/pull/2109). For example:
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{:s}", true);
+    }
+    ```
+
+    prints \"true\". Thanks @powercoderlol.
+
+-   Made `fmt::ptr` work with function pointers
+    (https://github.com/fmtlib/fmt/pull/2131). For example:
+
+    ```c++
+    #include <fmt/format.h>
+
+    int main() {
+      fmt::print("My main: {}\n", fmt::ptr(main));
+    }
+    ```
+
+    Thanks @mikecrowe.
+
+-   The undocumented support for specializing `formatter` for pointer
+    types has been removed.
+
+-   Fixed `fmt::formatted_size` with format string compilation
+    (https://github.com/fmtlib/fmt/pull/2141,
+    https://github.com/fmtlib/fmt/pull/2161). Thanks @alexezeder.
+
+-   Fixed handling of empty format strings during format string
+    compilation (https://github.com/fmtlib/fmt/issues/2042):
+
+    ```c++
+    auto s = fmt::format(FMT_COMPILE(""));
+    ```
+
+    Thanks @alexezeder.
+
+-   Fixed handling of enums in `fmt::to_string`
+    (https://github.com/fmtlib/fmt/issues/2036).
+
+-   Improved width computation
+    (https://github.com/fmtlib/fmt/issues/2033,
+    https://github.com/fmtlib/fmt/issues/2091). For example:
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{:-<10}{}\n", "你好", "世界");
+      fmt::print("{:-<10}{}\n", "hello", "world");
+    }
+    ```
+
+    prints
+
+    ![](https://user-images.githubusercontent.com/576385/119840373-cea3ca80-beb9-11eb-91e0-54266c48e181.png)
+
+    on a modern terminal.
+
+-   The experimental fast output stream (`fmt::ostream`) is now
+    truncated by default for consistency with `fopen`
+    (https://github.com/fmtlib/fmt/issues/2018). For example:
+
+    ```c++
+    #include <fmt/os.h>
+
+    int main() {
+      fmt::ostream out1 = fmt::output_file("guide");
+      out1.print("Zaphod");
+      out1.close();
+      fmt::ostream out2 = fmt::output_file("guide");
+      out2.print("Ford");
+    }
+    ```
+
+    writes \"Ford\" to the file \"guide\". To preserve the old file
+    content if any pass `fmt::file::WRONLY | fmt::file::CREATE` flags to
+    `fmt::output_file`.
+
+-   Fixed moving of `fmt::ostream` that holds buffered data
+    (https://github.com/fmtlib/fmt/issues/2197,
+    https://github.com/fmtlib/fmt/pull/2198). Thanks @vtta.
+
+-   Replaced the `fmt::system_error` exception with a function of the
+    same name that constructs `std::system_error`
+    (https://github.com/fmtlib/fmt/issues/2266).
+
+-   Replaced the `fmt::windows_error` exception with a function of the
+    same name that constructs `std::system_error` with the category
+    returned by `fmt::system_category()`
+    (https://github.com/fmtlib/fmt/issues/2274,
+    https://github.com/fmtlib/fmt/pull/2275). The latter is
+    similar to `std::sytem_category` but correctly handles UTF-8.
+    Thanks @phprus.
+
+-   Replaced `fmt::error_code` with `std::error_code` and made it
+    formattable (https://github.com/fmtlib/fmt/issues/2269,
+    https://github.com/fmtlib/fmt/pull/2270,
+    https://github.com/fmtlib/fmt/pull/2273). Thanks @phprus.
+
+-   Added speech synthesis support
+    (https://github.com/fmtlib/fmt/pull/2206).
+
+-   Made `format_to` work with a memory buffer that has a custom
+    allocator (https://github.com/fmtlib/fmt/pull/2300).
+    Thanks @voxmea.
+
+-   Added `Allocator::max_size` support to `basic_memory_buffer`.
+    (https://github.com/fmtlib/fmt/pull/1960). Thanks @phprus.
+
+-   Added wide string support to `fmt::join`
+    (https://github.com/fmtlib/fmt/pull/2236). Thanks @crbrz.
+
+-   Made iterators passed to `formatter` specializations via a format
+    context satisfy C++20 `std::output_iterator` requirements
+    (https://github.com/fmtlib/fmt/issues/2156,
+    https://github.com/fmtlib/fmt/pull/2158,
+    https://github.com/fmtlib/fmt/issues/2195,
+    https://github.com/fmtlib/fmt/pull/2204). Thanks @randomnetcat.
+
+-   Optimized the `printf` implementation
+    (https://github.com/fmtlib/fmt/pull/1982,
+    https://github.com/fmtlib/fmt/pull/1984,
+    https://github.com/fmtlib/fmt/pull/2016,
+    https://github.com/fmtlib/fmt/pull/2164).
+    Thanks @rimathia and @moiwi.
+
+-   Improved detection of `constexpr` `char_traits`
+    (https://github.com/fmtlib/fmt/pull/2246,
+    https://github.com/fmtlib/fmt/pull/2257). Thanks @phprus.
+
+-   Fixed writing to `stdout` when it is redirected to `NUL` on Windows
+    (https://github.com/fmtlib/fmt/issues/2080).
+
+-   Fixed exception propagation from iterators
+    (https://github.com/fmtlib/fmt/issues/2097).
+
+-   Improved `strftime` error handling
+    (https://github.com/fmtlib/fmt/issues/2238,
+    https://github.com/fmtlib/fmt/pull/2244). Thanks @yumeyao.
+
+-   Stopped using deprecated GCC UDL template extension.
+
+-   Added `fmt/args.h` to the install target
+    (https://github.com/fmtlib/fmt/issues/2096).
+
+-   Error messages are now passed to assert when exceptions are disabled
+    (https://github.com/fmtlib/fmt/pull/2145). Thanks @NobodyXu.
+
+-   Added the `FMT_MASTER_PROJECT` CMake option to control build and
+    install targets when {fmt} is included via `add_subdirectory`
+    (https://github.com/fmtlib/fmt/issues/2098,
+    https://github.com/fmtlib/fmt/pull/2100).
+    Thanks @randomizedthinking.
+
+-   Improved build configuration
+    (https://github.com/fmtlib/fmt/pull/2026,
+    https://github.com/fmtlib/fmt/pull/2122).
+    Thanks @luncliff and @ibaned.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/issues/1947,
+    https://github.com/fmtlib/fmt/pull/1959,
+    https://github.com/fmtlib/fmt/pull/1963,
+    https://github.com/fmtlib/fmt/pull/1965,
+    https://github.com/fmtlib/fmt/issues/1966,
+    https://github.com/fmtlib/fmt/pull/1974,
+    https://github.com/fmtlib/fmt/pull/1975,
+    https://github.com/fmtlib/fmt/pull/1990,
+    https://github.com/fmtlib/fmt/issues/2000,
+    https://github.com/fmtlib/fmt/pull/2001,
+    https://github.com/fmtlib/fmt/issues/2002,
+    https://github.com/fmtlib/fmt/issues/2004,
+    https://github.com/fmtlib/fmt/pull/2006,
+    https://github.com/fmtlib/fmt/pull/2009,
+    https://github.com/fmtlib/fmt/pull/2010,
+    https://github.com/fmtlib/fmt/issues/2038,
+    https://github.com/fmtlib/fmt/issues/2039,
+    https://github.com/fmtlib/fmt/issues/2047,
+    https://github.com/fmtlib/fmt/pull/2053,
+    https://github.com/fmtlib/fmt/issues/2059,
+    https://github.com/fmtlib/fmt/pull/2065,
+    https://github.com/fmtlib/fmt/pull/2067,
+    https://github.com/fmtlib/fmt/pull/2068,
+    https://github.com/fmtlib/fmt/pull/2073,
+    https://github.com/fmtlib/fmt/issues/2103,
+    https://github.com/fmtlib/fmt/issues/2105,
+    https://github.com/fmtlib/fmt/pull/2106,
+    https://github.com/fmtlib/fmt/pull/2107,
+    https://github.com/fmtlib/fmt/issues/2116,
+    https://github.com/fmtlib/fmt/pull/2117,
+    https://github.com/fmtlib/fmt/issues/2118,
+    https://github.com/fmtlib/fmt/pull/2119,
+    https://github.com/fmtlib/fmt/issues/2127,
+    https://github.com/fmtlib/fmt/pull/2128,
+    https://github.com/fmtlib/fmt/issues/2140,
+    https://github.com/fmtlib/fmt/issues/2142,
+    https://github.com/fmtlib/fmt/pull/2143,
+    https://github.com/fmtlib/fmt/pull/2144,
+    https://github.com/fmtlib/fmt/issues/2147,
+    https://github.com/fmtlib/fmt/issues/2148,
+    https://github.com/fmtlib/fmt/issues/2149,
+    https://github.com/fmtlib/fmt/pull/2152,
+    https://github.com/fmtlib/fmt/pull/2160,
+    https://github.com/fmtlib/fmt/issues/2170,
+    https://github.com/fmtlib/fmt/issues/2175,
+    https://github.com/fmtlib/fmt/issues/2176,
+    https://github.com/fmtlib/fmt/pull/2177,
+    https://github.com/fmtlib/fmt/issues/2178,
+    https://github.com/fmtlib/fmt/pull/2179,
+    https://github.com/fmtlib/fmt/issues/2180,
+    https://github.com/fmtlib/fmt/issues/2181,
+    https://github.com/fmtlib/fmt/pull/2183,
+    https://github.com/fmtlib/fmt/issues/2184,
+    https://github.com/fmtlib/fmt/issues/2185,
+    https://github.com/fmtlib/fmt/pull/2186,
+    https://github.com/fmtlib/fmt/pull/2187,
+    https://github.com/fmtlib/fmt/pull/2190,
+    https://github.com/fmtlib/fmt/pull/2192,
+    https://github.com/fmtlib/fmt/pull/2194,
+    https://github.com/fmtlib/fmt/pull/2205,
+    https://github.com/fmtlib/fmt/issues/2210,
+    https://github.com/fmtlib/fmt/pull/2211,
+    https://github.com/fmtlib/fmt/pull/2215,
+    https://github.com/fmtlib/fmt/pull/2216,
+    https://github.com/fmtlib/fmt/pull/2218,
+    https://github.com/fmtlib/fmt/pull/2220,
+    https://github.com/fmtlib/fmt/issues/2228,
+    https://github.com/fmtlib/fmt/pull/2229,
+    https://github.com/fmtlib/fmt/pull/2230,
+    https://github.com/fmtlib/fmt/issues/2233,
+    https://github.com/fmtlib/fmt/pull/2239,
+    https://github.com/fmtlib/fmt/issues/2248,
+    https://github.com/fmtlib/fmt/issues/2252,
+    https://github.com/fmtlib/fmt/pull/2253,
+    https://github.com/fmtlib/fmt/pull/2255,
+    https://github.com/fmtlib/fmt/issues/2261,
+    https://github.com/fmtlib/fmt/issues/2278,
+    https://github.com/fmtlib/fmt/issues/2284,
+    https://github.com/fmtlib/fmt/pull/2287,
+    https://github.com/fmtlib/fmt/pull/2289,
+    https://github.com/fmtlib/fmt/pull/2290,
+    https://github.com/fmtlib/fmt/pull/2293,
+    https://github.com/fmtlib/fmt/issues/2295,
+    https://github.com/fmtlib/fmt/pull/2296,
+    https://github.com/fmtlib/fmt/pull/2297,
+    https://github.com/fmtlib/fmt/issues/2311,
+    https://github.com/fmtlib/fmt/pull/2313,
+    https://github.com/fmtlib/fmt/pull/2315,
+    https://github.com/fmtlib/fmt/issues/2320,
+    https://github.com/fmtlib/fmt/pull/2321,
+    https://github.com/fmtlib/fmt/pull/2323,
+    https://github.com/fmtlib/fmt/issues/2328,
+    https://github.com/fmtlib/fmt/pull/2329,
+    https://github.com/fmtlib/fmt/pull/2333,
+    https://github.com/fmtlib/fmt/pull/2338,
+    https://github.com/fmtlib/fmt/pull/2341).
+    Thanks @darklukee, @fagg, @killerbot242, @jgopel, @yeswalrus, @Finkman,
+    @HazardyKnusperkeks, @dkavolis, @concatime, @chronoxor, @summivox, @yNeo,
+    @Apache-HB, @alexezeder, @toojays, @Brainy0207, @vadz, @imsherlock, @phprus,
+    @white238, @yafshar, @BillyDonahue, @jstaahl, @denchat, @DanielaE,
+    @ilyakurdyukov, @ilmai, @JessyDL, @sergiud, @mwinterb, @sven-herrmann,
+    @jmelas, @twoixter, @crbrz and @upsj.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/1986,
+    https://github.com/fmtlib/fmt/pull/2051,
+    https://github.com/fmtlib/fmt/issues/2057,
+    https://github.com/fmtlib/fmt/pull/2081,
+    https://github.com/fmtlib/fmt/issues/2084,
+    https://github.com/fmtlib/fmt/pull/2312).
+    Thanks @imba-tjd, @0x416c69 and @mordante.
+
+-   Continuous integration and test improvements
+    (https://github.com/fmtlib/fmt/issues/1969,
+    https://github.com/fmtlib/fmt/pull/1991,
+    https://github.com/fmtlib/fmt/pull/2020,
+    https://github.com/fmtlib/fmt/pull/2110,
+    https://github.com/fmtlib/fmt/pull/2114,
+    https://github.com/fmtlib/fmt/issues/2196,
+    https://github.com/fmtlib/fmt/pull/2217,
+    https://github.com/fmtlib/fmt/pull/2247,
+    https://github.com/fmtlib/fmt/pull/2256,
+    https://github.com/fmtlib/fmt/pull/2336,
+    https://github.com/fmtlib/fmt/pull/2346).
+    Thanks @jgopel, @alexezeder and @DanielaE.
+
+# 7.1.3 - 2020-11-24
+
+-   Fixed handling of buffer boundaries in `format_to_n`
+    (https://github.com/fmtlib/fmt/issues/1996,
+    https://github.com/fmtlib/fmt/issues/2029).
+-   Fixed linkage errors when linking with a shared library
+    (https://github.com/fmtlib/fmt/issues/2011).
+-   Reintroduced ostream support to range formatters
+    (https://github.com/fmtlib/fmt/issues/2014).
+-   Worked around an issue with mixing std versions in gcc
+    (https://github.com/fmtlib/fmt/issues/2017).
+
+# 7.1.2 - 2020-11-04
+
+-   Fixed floating point formatting with large precision
+    (https://github.com/fmtlib/fmt/issues/1976).
+
+# 7.1.1 - 2020-11-01
+
+-   Fixed ABI compatibility with 7.0.x
+    (https://github.com/fmtlib/fmt/issues/1961).
+-   Added the `FMT_ARM_ABI_COMPATIBILITY` macro to work around ABI
+    incompatibility between GCC and Clang on ARM
+    (https://github.com/fmtlib/fmt/issues/1919).
+-   Worked around a SFINAE bug in GCC 8
+    (https://github.com/fmtlib/fmt/issues/1957).
+-   Fixed linkage errors when building with GCC\'s LTO
+    (https://github.com/fmtlib/fmt/issues/1955).
+-   Fixed a compilation error when building without `__builtin_clz` or
+    equivalent (https://github.com/fmtlib/fmt/pull/1968).
+    Thanks @tohammer.
+-   Fixed a sign conversion warning
+    (https://github.com/fmtlib/fmt/pull/1964). Thanks @OptoCloud.
+
+# 7.1.0 - 2020-10-25
+
+-   Switched from
+    [Grisu3](https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf)
+    to [Dragonbox](https://github.com/jk-jeon/dragonbox) for the default
+    floating-point formatting which gives the shortest decimal
+    representation with round-trip guarantee and correct rounding
+    (https://github.com/fmtlib/fmt/pull/1882,
+    https://github.com/fmtlib/fmt/pull/1887,
+    https://github.com/fmtlib/fmt/pull/1894). This makes {fmt}
+    up to 20-30x faster than common implementations of
+    `std::ostringstream` and `sprintf` on
+    [dtoa-benchmark](https://github.com/fmtlib/dtoa-benchmark) and
+    faster than double-conversion and Ryū:
+
+    ![](https://user-images.githubusercontent.com/576385/95684665-11719600-0ba8-11eb-8e5b-972ff4e49428.png)
+
+    It is possible to get even better performance at the cost of larger
+    binary size by compiling with the `FMT_USE_FULL_CACHE_DRAGONBOX`
+    macro set to 1.
+
+    Thanks @jk-jeon.
+
+-   Added an experimental unsynchronized file output API which, together
+    with [format string
+    compilation](https://fmt.dev/latest/api.html#compile-api), can give
+    [5-9 times speed up compared to
+    fprintf](https://www.zverovich.net/2020/08/04/optimal-file-buffer-size.html)
+    on common platforms ([godbolt](https://godbolt.org/z/nsTcG8)):
+
+    ```c++
+    #include <fmt/os.h>
+
+    int main() {
+      auto f = fmt::output_file("guide");
+      f.print("The answer is {}.", 42);
+    }
+    ```
+
+-   Added a formatter for `std::chrono::time_point<system_clock>`
+    (https://github.com/fmtlib/fmt/issues/1819,
+    https://github.com/fmtlib/fmt/pull/1837). For example
+    ([godbolt](https://godbolt.org/z/c4M6fh)):
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      auto now = std::chrono::system_clock::now();
+      fmt::print("The time is {:%H:%M:%S}.\n", now);
+    }
+    ```
+
+    Thanks @adamburgess.
+
+-   Added support for ranges with non-const `begin`/`end` to `fmt::join`
+    (https://github.com/fmtlib/fmt/issues/1784,
+    https://github.com/fmtlib/fmt/pull/1786). For example
+    ([godbolt](https://godbolt.org/z/jP63Tv)):
+
+    ```c++
+    #include <fmt/ranges.h>
+    #include <range/v3/view/filter.hpp>
+
+    int main() {
+      using std::literals::string_literals::operator""s;
+      auto strs = std::array{"a"s, "bb"s, "ccc"s};
+      auto range = strs | ranges::views::filter(
+        [] (const std::string &x) { return x.size() != 2; }
+      );
+      fmt::print("{}\n", fmt::join(range, ""));
+    }
+    ```
+
+    prints \"accc\".
+
+    Thanks @tonyelewis.
+
+-   Added a `memory_buffer::append` overload that takes a range
+    (https://github.com/fmtlib/fmt/pull/1806). Thanks @BRevzin.
+
+-   Improved handling of single code units in `FMT_COMPILE`. For
+    example:
+
+    ```c++
+    #include <fmt/compile.h>
+
+    char* f(char* buf) {
+      return fmt::format_to(buf, FMT_COMPILE("x{}"), 42);
+    }
+    ```
+
+    compiles to just ([godbolt](https://godbolt.org/z/5vncz3)):
+
+    ```asm
+    _Z1fPc:
+      movb $120, (%rdi)
+      xorl %edx, %edx
+      cmpl $42, _ZN3fmt2v76detail10basic_dataIvE23zero_or_powers_of_10_32E+8(%rip)
+      movl $3, %eax
+      seta %dl
+      subl %edx, %eax
+      movzwl _ZN3fmt2v76detail10basic_dataIvE6digitsE+84(%rip), %edx
+      cltq
+      addq %rdi, %rax
+      movw %dx, -2(%rax)
+      ret
+    ```
+
+    Here a single `mov` instruction writes `'x'` (`$120`) to the output
+    buffer.
+
+-   Added dynamic width support to format string compilation
+    (https://github.com/fmtlib/fmt/issues/1809).
+
+-   Improved error reporting for unformattable types: now you\'ll get
+    the type name directly in the error message instead of the note:
+
+    ```c++
+    #include <fmt/core.h>
+
+    struct how_about_no {};
+
+    int main() {
+      fmt::print("{}", how_about_no());
+    }
+    ```
+
+    Error ([godbolt](https://godbolt.org/z/GoxM4e)):
+
+    `fmt/core.h:1438:3: error: static_assert failed due to requirement 'fmt::v7::formattable<how_about_no>()' "Cannot format an argument. To make type T formattable provide a formatter<T> specialization: https://fmt.dev/latest/api.html#udt" ...`
+
+-   Added the
+    [make_args_checked](https://fmt.dev/7.1.0/api.html#argument-lists)
+    function template that allows you to write formatting functions with
+    compile-time format string checks and avoid binary code bloat
+    ([godbolt](https://godbolt.org/z/PEf9qr)):
+
+    ```c++
+    void vlog(const char* file, int line, fmt::string_view format,
+              fmt::format_args args) {
+      fmt::print("{}: {}: ", file, line);
+      fmt::vprint(format, args);
+    }
+
+    template <typename S, typename... Args>
+    void log(const char* file, int line, const S& format, Args&&... args) {
+      vlog(file, line, format,
+          fmt::make_args_checked<Args...>(format, args...));
+    }
+
+    #define MY_LOG(format, ...) \
+      log(__FILE__, __LINE__, FMT_STRING(format), __VA_ARGS__)
+
+    MY_LOG("invalid squishiness: {}", 42);
+    ```
+
+-   Replaced `snprintf` fallback with a faster internal IEEE 754 `float`
+    and `double` formatter for arbitrary precision. For example
+    ([godbolt](https://godbolt.org/z/dPhWvj)):
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{:.500}\n", 4.9406564584124654E-324);
+    }
+    ```
+
+    prints
+
+    `4.9406564584124654417656879286822137236505980261432476442558568250067550727020875186529983636163599237979656469544571773092665671035593979639877479601078187812630071319031140452784581716784898210368871863605699873072305000638740915356498438731247339727316961514003171538539807412623856559117102665855668676818703956031062493194527159149245532930545654440112748012970999954193198940908041656332452475714786901472678015935523861155013480352649347201937902681071074917033322268447533357208324319360923829e-324`.
+
+-   Made `format_to_n` and `formatted_size` part of the [core
+    API](https://fmt.dev/latest/api.html#core-api)
+    ([godbolt](https://godbolt.org/z/sPjY1K)):
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      char buffer[10];
+      auto result = fmt::format_to_n(buffer, sizeof(buffer), "{}", 42);
+    }
+    ```
+
+-   Added `fmt::format_to_n` overload with format string compilation
+    (https://github.com/fmtlib/fmt/issues/1764,
+    https://github.com/fmtlib/fmt/pull/1767,
+    https://github.com/fmtlib/fmt/pull/1869). For example
+    ([godbolt](https://godbolt.org/z/93h86q)):
+
+    ```c++
+    #include <fmt/compile.h>
+
+    int main() {
+      char buffer[8];
+      fmt::format_to_n(buffer, sizeof(buffer), FMT_COMPILE("{}"), 42);
+    }
+    ```
+
+    Thanks @Kurkin and @alexezeder.
+
+-   Added `fmt::format_to` overload that take `text_style`
+    (https://github.com/fmtlib/fmt/issues/1593,
+    https://github.com/fmtlib/fmt/issues/1842,
+    https://github.com/fmtlib/fmt/pull/1843). For example
+    ([godbolt](https://godbolt.org/z/91153r)):
+
+    ```c++
+    #include <fmt/color.h>
+
+    int main() {
+      std::string out;
+      fmt::format_to(std::back_inserter(out),
+                     fmt::emphasis::bold | fg(fmt::color::red),
+                     "The answer is {}.", 42);
+    }
+    ```
+
+    Thanks @Naios.
+
+-   Made the `'#'` specifier emit trailing zeros in addition to the
+    decimal point (https://github.com/fmtlib/fmt/issues/1797).
+    For example ([godbolt](https://godbolt.org/z/bhdcW9)):
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{:#.2g}", 0.5);
+    }
+    ```
+
+    prints `0.50`.
+
+-   Changed the default floating point format to not include `.0` for
+    consistency with `std::format` and `std::to_chars`
+    (https://github.com/fmtlib/fmt/issues/1893,
+    https://github.com/fmtlib/fmt/issues/1943). It is possible
+    to get the decimal point and trailing zero with the `#` specifier.
+
+-   Fixed an issue with floating-point formatting that could result in
+    addition of a non-significant trailing zero in rare cases e.g.
+    `1.00e-34` instead of `1.0e-34`
+    (https://github.com/fmtlib/fmt/issues/1873,
+    https://github.com/fmtlib/fmt/issues/1917).
+
+-   Made `fmt::to_string` fallback on `ostream` insertion operator if
+    the `formatter` specialization is not provided
+    (https://github.com/fmtlib/fmt/issues/1815,
+    https://github.com/fmtlib/fmt/pull/1829). Thanks @alexezeder.
+
+-   Added support for the append mode to the experimental file API and
+    improved `fcntl.h` detection.
+    (https://github.com/fmtlib/fmt/pull/1847,
+    https://github.com/fmtlib/fmt/pull/1848). Thanks @t-wiser.
+
+-   Fixed handling of types that have both an implicit conversion
+    operator and an overloaded `ostream` insertion operator
+    (https://github.com/fmtlib/fmt/issues/1766).
+
+-   Fixed a slicing issue in an internal iterator type
+    (https://github.com/fmtlib/fmt/pull/1822). Thanks @BRevzin.
+
+-   Fixed an issue in locale-specific integer formatting
+    (https://github.com/fmtlib/fmt/issues/1927).
+
+-   Fixed handling of exotic code unit types
+    (https://github.com/fmtlib/fmt/issues/1870,
+    https://github.com/fmtlib/fmt/issues/1932).
+
+-   Improved `FMT_ALWAYS_INLINE`
+    (https://github.com/fmtlib/fmt/pull/1878). Thanks @jk-jeon.
+
+-   Removed dependency on `windows.h`
+    (https://github.com/fmtlib/fmt/pull/1900). Thanks @bernd5.
+
+-   Optimized counting of decimal digits on MSVC
+    (https://github.com/fmtlib/fmt/pull/1890). Thanks @mwinterb.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/1772,
+    https://github.com/fmtlib/fmt/pull/1775,
+    https://github.com/fmtlib/fmt/pull/1792,
+    https://github.com/fmtlib/fmt/pull/1838,
+    https://github.com/fmtlib/fmt/pull/1888,
+    https://github.com/fmtlib/fmt/pull/1918,
+    https://github.com/fmtlib/fmt/pull/1939).
+    Thanks @leolchat, @pepsiman, @Klaim, @ravijanjam, @francesco-st and @udnaan.
+
+-   Added the `FMT_REDUCE_INT_INSTANTIATIONS` CMake option that reduces
+    the binary code size at the cost of some integer formatting
+    performance. This can be useful for extremely memory-constrained
+    embedded systems
+    (https://github.com/fmtlib/fmt/issues/1778,
+    https://github.com/fmtlib/fmt/pull/1781). Thanks @kammce.
+
+-   Added the `FMT_USE_INLINE_NAMESPACES` macro to control usage of
+    inline namespaces
+    (https://github.com/fmtlib/fmt/pull/1945). Thanks @darklukee.
+
+-   Improved build configuration
+    (https://github.com/fmtlib/fmt/pull/1760,
+    https://github.com/fmtlib/fmt/pull/1770,
+    https://github.com/fmtlib/fmt/issues/1779,
+    https://github.com/fmtlib/fmt/pull/1783,
+    https://github.com/fmtlib/fmt/pull/1823).
+    Thanks @dvetutnev, @xvitaly, @tambry, @medithe and @martinwuehrer.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/pull/1790,
+    https://github.com/fmtlib/fmt/pull/1802,
+    https://github.com/fmtlib/fmt/pull/1808,
+    https://github.com/fmtlib/fmt/issues/1810,
+    https://github.com/fmtlib/fmt/issues/1811,
+    https://github.com/fmtlib/fmt/pull/1812,
+    https://github.com/fmtlib/fmt/pull/1814,
+    https://github.com/fmtlib/fmt/pull/1816,
+    https://github.com/fmtlib/fmt/pull/1817,
+    https://github.com/fmtlib/fmt/pull/1818,
+    https://github.com/fmtlib/fmt/issues/1825,
+    https://github.com/fmtlib/fmt/pull/1836,
+    https://github.com/fmtlib/fmt/pull/1855,
+    https://github.com/fmtlib/fmt/pull/1856,
+    https://github.com/fmtlib/fmt/pull/1860,
+    https://github.com/fmtlib/fmt/pull/1877,
+    https://github.com/fmtlib/fmt/pull/1879,
+    https://github.com/fmtlib/fmt/pull/1880,
+    https://github.com/fmtlib/fmt/issues/1896,
+    https://github.com/fmtlib/fmt/pull/1897,
+    https://github.com/fmtlib/fmt/pull/1898,
+    https://github.com/fmtlib/fmt/issues/1904,
+    https://github.com/fmtlib/fmt/pull/1908,
+    https://github.com/fmtlib/fmt/issues/1911,
+    https://github.com/fmtlib/fmt/issues/1912,
+    https://github.com/fmtlib/fmt/issues/1928,
+    https://github.com/fmtlib/fmt/pull/1929,
+    https://github.com/fmtlib/fmt/issues/1935,
+    https://github.com/fmtlib/fmt/pull/1937,
+    https://github.com/fmtlib/fmt/pull/1942,
+    https://github.com/fmtlib/fmt/issues/1949).
+    Thanks @TheQwertiest, @medithe, @martinwuehrer, @n16h7hunt3r, @Othereum,
+    @gsjaardema, @AlexanderLanin, @gcerretani, @chronoxor, @noizefloor,
+    @akohlmey, @jk-jeon, @rimathia, @rglarix, @moiwi, @heckad, @MarcDirven.
+    @BartSiwek and @darklukee.
+
+# 7.0.3 - 2020-08-06
+
+-   Worked around broken `numeric_limits` for 128-bit integers
+    (https://github.com/fmtlib/fmt/issues/1787).
+-   Added error reporting on missing named arguments
+    (https://github.com/fmtlib/fmt/issues/1796).
+-   Stopped using 128-bit integers with clang-cl
+    (https://github.com/fmtlib/fmt/pull/1800). Thanks @Kingcom.
+-   Fixed issues in locale-specific integer formatting
+    (https://github.com/fmtlib/fmt/issues/1782,
+    https://github.com/fmtlib/fmt/issues/1801).
+
+# 7.0.2 - 2020-07-29
+
+-   Worked around broken `numeric_limits` for 128-bit integers
+    (https://github.com/fmtlib/fmt/issues/1725).
+-   Fixed compatibility with CMake 3.4
+    (https://github.com/fmtlib/fmt/issues/1779).
+-   Fixed handling of digit separators in locale-specific formatting
+    (https://github.com/fmtlib/fmt/issues/1782).
+
+# 7.0.1 - 2020-07-07
+
+-   Updated the inline version namespace name.
+-   Worked around a gcc bug in mangling of alias templates
+    (https://github.com/fmtlib/fmt/issues/1753).
+-   Fixed a linkage error on Windows
+    (https://github.com/fmtlib/fmt/issues/1757). Thanks @Kurkin.
+-   Fixed minor issues with the documentation.
+
+# 7.0.0 - 2020-07-05
+
+-   Reduced the library size. For example, on macOS a stripped test
+    binary statically linked with {fmt} [shrank from \~368k to less than
+    100k](http://www.zverovich.net/2020/05/21/reducing-library-size.html).
+
+-   Added a simpler and more efficient [format string compilation
+    API](https://fmt.dev/7.0.0/api.html#compile-api):
+
+    ```c++
+    #include <fmt/compile.h>
+
+    // Converts 42 into std::string using the most efficient method and no
+    // runtime format string processing.
+    std::string s = fmt::format(FMT_COMPILE("{}"), 42);
+    ```
+
+    The old `fmt::compile` API is now deprecated.
+
+-   Optimized integer formatting: `format_to` with format string
+    compilation and a stack-allocated buffer is now [faster than
+    to_chars on both libc++ and
+    libstdc++](http://www.zverovich.net/2020/06/13/fast-int-to-string-revisited.html).
+
+-   Optimized handling of small format strings. For example,
+
+    ```c++
+    fmt::format("Result: {}: ({},{},{},{})", str1, str2, str3, str4, str5)
+    ```
+
+    is now \~40% faster
+    (https://github.com/fmtlib/fmt/issues/1685).
+
+-   Applied extern templates to improve compile times when using the
+    core API and `fmt/format.h`
+    (https://github.com/fmtlib/fmt/issues/1452). For example,
+    on macOS with clang the compile time of a test translation unit
+    dropped from 2.3s to 0.3s with `-O2` and from 0.6s to 0.3s with the
+    default settings (`-O0`).
+
+    Before (`-O2`):
+
+        % time c++ -c test.cc -I include -std=c++17 -O2
+        c++ -c test.cc -I include -std=c++17 -O2  2.22s user 0.08s system 99% cpu 2.311 total
+
+    After (`-O2`):
+
+        % time c++ -c test.cc -I include -std=c++17 -O2
+        c++ -c test.cc -I include -std=c++17 -O2  0.26s user 0.04s system 98% cpu 0.303 total
+
+    Before (default):
+
+        % time c++ -c test.cc -I include -std=c++17
+        c++ -c test.cc -I include -std=c++17  0.53s user 0.06s system 98% cpu 0.601 total
+
+    After (default):
+
+        % time c++ -c test.cc -I include -std=c++17
+        c++ -c test.cc -I include -std=c++17  0.24s user 0.06s system 98% cpu 0.301 total
+
+    It is still recommended to use `fmt/core.h` instead of
+    `fmt/format.h` but the compile time difference is now smaller.
+    Thanks @alex3d for the suggestion.
+
+-   Named arguments are now stored on stack (no dynamic memory
+    allocations) and the compiled code is more compact and efficient.
+    For example
+
+    ```c++
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("The answer is {answer}\n", fmt::arg("answer", 42));
+    }
+    ```
+
+    compiles to just ([godbolt](https://godbolt.org/z/NcfEp_))
+
+    ```asm
+    .LC0:
+            .string "answer"
+    .LC1:
+            .string "The answer is {answer}\n"
+    main:
+            sub     rsp, 56
+            mov     edi, OFFSET FLAT:.LC1
+            mov     esi, 23
+            movabs  rdx, 4611686018427387905
+            lea     rax, [rsp+32]
+            lea     rcx, [rsp+16]
+            mov     QWORD PTR [rsp+8], 1
+            mov     QWORD PTR [rsp], rax
+            mov     DWORD PTR [rsp+16], 42
+            mov     QWORD PTR [rsp+32], OFFSET FLAT:.LC0
+            mov     DWORD PTR [rsp+40], 0
+            call    fmt::v6::vprint(fmt::v6::basic_string_view<char>,
+                                    fmt::v6::format_args)
+            xor     eax, eax
+            add     rsp, 56
+            ret
+
+        .L.str.1:
+                .asciz  "answer"
+    ```
+
+-   Implemented compile-time checks for dynamic width and precision
+    (https://github.com/fmtlib/fmt/issues/1614):
+
+    ```c++
+    #include <fmt/format.h>
+
+    int main() {
+      fmt::print(FMT_STRING("{0:{1}}"), 42);
+    }
+    ```
+
+    now gives a compilation error because argument 1 doesn\'t exist:
+
+        In file included from test.cc:1:
+        include/fmt/format.h:2726:27: error: constexpr variable 'invalid_format' must be
+        initialized by a constant expression
+          FMT_CONSTEXPR_DECL bool invalid_format =
+                                  ^
+        ...
+        include/fmt/core.h:569:26: note: in call to
+        '&checker(s, {}).context_->on_error(&"argument not found"[0])'
+            if (id >= num_args_) on_error("argument not found");
+                                ^
+
+-   Added sentinel support to `fmt::join`
+    (https://github.com/fmtlib/fmt/pull/1689)
+
+    ```c++
+    struct zstring_sentinel {};
+    bool operator==(const char* p, zstring_sentinel) { return *p == '\0'; }
+    bool operator!=(const char* p, zstring_sentinel) { return *p != '\0'; }
+
+    struct zstring {
+      const char* p;
+      const char* begin() const { return p; }
+      zstring_sentinel end() const { return {}; }
+    };
+
+    auto s = fmt::format("{}", fmt::join(zstring{"hello"}, "_"));
+    // s == "h_e_l_l_o"
+    ```
+
+    Thanks @BRevzin.
+
+-   Added support for named arguments, `clear` and `reserve` to
+    `dynamic_format_arg_store`
+    (https://github.com/fmtlib/fmt/issues/1655,
+    https://github.com/fmtlib/fmt/pull/1663,
+    https://github.com/fmtlib/fmt/pull/1674,
+    https://github.com/fmtlib/fmt/pull/1677). Thanks @vsolontsov-ll.
+
+-   Added support for the `'c'` format specifier to integral types for
+    compatibility with `std::format`
+    (https://github.com/fmtlib/fmt/issues/1652).
+
+-   Replaced the `'n'` format specifier with `'L'` for compatibility
+    with `std::format`
+    (https://github.com/fmtlib/fmt/issues/1624). The `'n'`
+    specifier can be enabled via the `FMT_DEPRECATED_N_SPECIFIER` macro.
+
+-   The `'='` format specifier is now disabled by default for
+    compatibility with `std::format`. It can be enabled via the
+    `FMT_DEPRECATED_NUMERIC_ALIGN` macro.
+
+-   Removed the following deprecated APIs:
+
+    -   `FMT_STRING_ALIAS` and `fmt` macros - replaced by `FMT_STRING`
+    -   `fmt::basic_string_view::char_type` - replaced by
+        `fmt::basic_string_view::value_type`
+    -   `convert_to_int`
+    -   `format_arg_store::types`
+    -   `*parse_context` - replaced by `*format_parse_context`
+    -   `FMT_DEPRECATED_INCLUDE_OS`
+    -   `FMT_DEPRECATED_PERCENT` - incompatible with `std::format`
+    -   `*writer` - replaced by compiled format API
+
+-   Renamed the `internal` namespace to `detail`
+    (https://github.com/fmtlib/fmt/issues/1538). The former is
+    still provided as an alias if the `FMT_USE_INTERNAL` macro is
+    defined.
+
+-   Improved compatibility between `fmt::printf` with the standard specs
+    (https://github.com/fmtlib/fmt/issues/1595,
+    https://github.com/fmtlib/fmt/pull/1682,
+    https://github.com/fmtlib/fmt/pull/1683,
+    https://github.com/fmtlib/fmt/pull/1687,
+    https://github.com/fmtlib/fmt/pull/1699). Thanks @rimathia.
+
+-   Fixed handling of `operator<<` overloads that use `copyfmt`
+    (https://github.com/fmtlib/fmt/issues/1666).
+
+-   Added the `FMT_OS` CMake option to control inclusion of OS-specific
+    APIs in the fmt target. This can be useful for embedded platforms
+    (https://github.com/fmtlib/fmt/issues/1654,
+    https://github.com/fmtlib/fmt/pull/1656). Thanks @kwesolowski.
+
+-   Replaced `FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION` with the
+    `FMT_FUZZ` macro to prevent interfering with fuzzing of projects
+    using {fmt} (https://github.com/fmtlib/fmt/pull/1650).
+    Thanks @asraa.
+
+-   Fixed compatibility with emscripten
+    (https://github.com/fmtlib/fmt/issues/1636,
+    https://github.com/fmtlib/fmt/pull/1637). Thanks @ArthurSonzogni.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/704,
+    https://github.com/fmtlib/fmt/pull/1643,
+    https://github.com/fmtlib/fmt/pull/1660,
+    https://github.com/fmtlib/fmt/pull/1681,
+    https://github.com/fmtlib/fmt/pull/1691,
+    https://github.com/fmtlib/fmt/pull/1706,
+    https://github.com/fmtlib/fmt/pull/1714,
+    https://github.com/fmtlib/fmt/pull/1721,
+    https://github.com/fmtlib/fmt/pull/1739,
+    https://github.com/fmtlib/fmt/pull/1740,
+    https://github.com/fmtlib/fmt/pull/1741,
+    https://github.com/fmtlib/fmt/pull/1751).
+    Thanks @senior7515, @lsr0, @puetzk, @fpelliccioni, Alexey Kuzmenko, @jelly,
+    @claremacrae, @jiapengwen, @gsjaardema and @alexey-milovidov.
+
+-   Implemented various build configuration fixes and improvements
+    (https://github.com/fmtlib/fmt/pull/1603,
+    https://github.com/fmtlib/fmt/pull/1657,
+    https://github.com/fmtlib/fmt/pull/1702,
+    https://github.com/fmtlib/fmt/pull/1728).
+    Thanks @scramsby, @jtojnar, @orivej and @flagarde.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/pull/1616,
+    https://github.com/fmtlib/fmt/issues/1620,
+    https://github.com/fmtlib/fmt/issues/1622,
+    https://github.com/fmtlib/fmt/issues/1625,
+    https://github.com/fmtlib/fmt/pull/1627,
+    https://github.com/fmtlib/fmt/issues/1628,
+    https://github.com/fmtlib/fmt/pull/1629,
+    https://github.com/fmtlib/fmt/issues/1631,
+    https://github.com/fmtlib/fmt/pull/1633,
+    https://github.com/fmtlib/fmt/pull/1649,
+    https://github.com/fmtlib/fmt/issues/1658,
+    https://github.com/fmtlib/fmt/pull/1661,
+    https://github.com/fmtlib/fmt/pull/1667,
+    https://github.com/fmtlib/fmt/issues/1668,
+    https://github.com/fmtlib/fmt/pull/1669,
+    https://github.com/fmtlib/fmt/issues/1692,
+    https://github.com/fmtlib/fmt/pull/1696,
+    https://github.com/fmtlib/fmt/pull/1697,
+    https://github.com/fmtlib/fmt/issues/1707,
+    https://github.com/fmtlib/fmt/pull/1712,
+    https://github.com/fmtlib/fmt/pull/1716,
+    https://github.com/fmtlib/fmt/pull/1722,
+    https://github.com/fmtlib/fmt/issues/1724,
+    https://github.com/fmtlib/fmt/pull/1729,
+    https://github.com/fmtlib/fmt/pull/1738,
+    https://github.com/fmtlib/fmt/issues/1742,
+    https://github.com/fmtlib/fmt/issues/1743,
+    https://github.com/fmtlib/fmt/pull/1744,
+    https://github.com/fmtlib/fmt/issues/1747,
+    https://github.com/fmtlib/fmt/pull/1750).
+    Thanks @gsjaardema, @gabime, @johnor, @Kurkin, @invexed, @peterbell10,
+    @daixtrose, @petrutlucian94, @Neargye, @ambitslix, @gabime, @erthink,
+    @tohammer and @0x8000-0000.
+
+# 6.2.1 - 2020-05-09
+
+-   Fixed ostream support in `sprintf`
+    (https://github.com/fmtlib/fmt/issues/1631).
+-   Fixed type detection when using implicit conversion to `string_view`
+    and ostream `operator<<` inconsistently
+    (https://github.com/fmtlib/fmt/issues/1662).
+
+# 6.2.0 - 2020-04-05
+
+-   Improved error reporting when trying to format an object of a
+    non-formattable type:
+
+    ```c++
+    fmt::format("{}", S());
+    ```
+
+    now gives:
+
+        include/fmt/core.h:1015:5: error: static_assert failed due to requirement
+        'formattable' "Cannot format argument. To make type T formattable provide a
+        formatter<T> specialization:
+        https://fmt.dev/latest/api.html#formatting-user-defined-types"
+            static_assert(
+            ^
+        ...
+        note: in instantiation of function template specialization
+        'fmt::v6::format<char [3], S, char>' requested here
+          fmt::format("{}", S());
+               ^
+
+    if `S` is not formattable.
+
+-   Reduced the library size by \~10%.
+
+-   Always print decimal point if `#` is specified
+    (https://github.com/fmtlib/fmt/issues/1476,
+    https://github.com/fmtlib/fmt/issues/1498):
+
+    ```c++
+    fmt::print("{:#.0f}", 42.0);
+    ```
+
+    now prints `42.`
+
+-   Implemented the `'L'` specifier for locale-specific numeric
+    formatting to improve compatibility with `std::format`. The `'n'`
+    specifier is now deprecated and will be removed in the next major
+    release.
+
+-   Moved OS-specific APIs such as `windows_error` from `fmt/format.h`
+    to `fmt/os.h`. You can define `FMT_DEPRECATED_INCLUDE_OS` to
+    automatically include `fmt/os.h` from `fmt/format.h` for
+    compatibility but this will be disabled in the next major release.
+
+-   Added precision overflow detection in floating-point formatting.
+
+-   Implemented detection of invalid use of `fmt::arg`.
+
+-   Used `type_identity` to block unnecessary template argument
+    deduction. Thanks Tim Song.
+
+-   Improved UTF-8 handling
+    (https://github.com/fmtlib/fmt/issues/1109):
+
+    ```c++
+    fmt::print("┌{0:─^{2}}┐\n"
+               "│{1: ^{2}}│\n"
+               "└{0:─^{2}}┘\n", "", "Привет, мир!", 20);
+    ```
+
+    now prints:
+
+        ┌────────────────────┐
+        │    Привет, мир!    │
+        └────────────────────┘
+
+    on systems that support Unicode.
+
+-   Added experimental dynamic argument storage
+    (https://github.com/fmtlib/fmt/issues/1170,
+    https://github.com/fmtlib/fmt/pull/1584):
+
+    ```c++
+    fmt::dynamic_format_arg_store<fmt::format_context> store;
+    store.push_back("answer");
+    store.push_back(42);
+    fmt::vprint("The {} is {}.\n", store);
+    ```
+
+    prints:
+
+        The answer is 42.
+
+    Thanks @vsolontsov-ll.
+
+-   Made `fmt::join` accept `initializer_list`
+    (https://github.com/fmtlib/fmt/pull/1591). Thanks @Rapotkinnik.
+
+-   Fixed handling of empty tuples
+    (https://github.com/fmtlib/fmt/issues/1588).
+
+-   Fixed handling of output iterators in `format_to_n`
+    (https://github.com/fmtlib/fmt/issues/1506).
+
+-   Fixed formatting of `std::chrono::duration` types to wide output
+    (https://github.com/fmtlib/fmt/pull/1533). Thanks @zeffy.
+
+-   Added const `begin` and `end` overload to buffers
+    (https://github.com/fmtlib/fmt/pull/1553). Thanks @dominicpoeschko.
+
+-   Added the ability to disable floating-point formatting via
+    `FMT_USE_FLOAT`, `FMT_USE_DOUBLE` and `FMT_USE_LONG_DOUBLE` macros
+    for extremely memory-constrained embedded system
+    (https://github.com/fmtlib/fmt/pull/1590). Thanks @albaguirre.
+
+-   Made `FMT_STRING` work with `constexpr` `string_view`
+    (https://github.com/fmtlib/fmt/pull/1589). Thanks @scramsby.
+
+-   Implemented a minor optimization in the format string parser
+    (https://github.com/fmtlib/fmt/pull/1560). Thanks @IkarusDeveloper.
+
+-   Improved attribute detection
+    (https://github.com/fmtlib/fmt/pull/1469,
+    https://github.com/fmtlib/fmt/pull/1475,
+    https://github.com/fmtlib/fmt/pull/1576).
+    Thanks @federico-busato, @chronoxor and @refnum.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/pull/1481,
+    https://github.com/fmtlib/fmt/pull/1523).
+    Thanks @JackBoosY and @imba-tjd.
+
+-   Fixed symbol visibility on Linux when compiling with
+    `-fvisibility=hidden`
+    (https://github.com/fmtlib/fmt/pull/1535). Thanks @milianw.
+
+-   Implemented various build configuration fixes and improvements
+    (https://github.com/fmtlib/fmt/issues/1264,
+    https://github.com/fmtlib/fmt/issues/1460,
+    https://github.com/fmtlib/fmt/pull/1534,
+    https://github.com/fmtlib/fmt/issues/1536,
+    https://github.com/fmtlib/fmt/issues/1545,
+    https://github.com/fmtlib/fmt/pull/1546,
+    https://github.com/fmtlib/fmt/issues/1566,
+    https://github.com/fmtlib/fmt/pull/1582,
+    https://github.com/fmtlib/fmt/issues/1597,
+    https://github.com/fmtlib/fmt/pull/1598).
+    Thanks @ambitslix, @jwillikers and @stac47.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/pull/1433,
+    https://github.com/fmtlib/fmt/issues/1461,
+    https://github.com/fmtlib/fmt/pull/1470,
+    https://github.com/fmtlib/fmt/pull/1480,
+    https://github.com/fmtlib/fmt/pull/1485,
+    https://github.com/fmtlib/fmt/pull/1492,
+    https://github.com/fmtlib/fmt/issues/1493,
+    https://github.com/fmtlib/fmt/issues/1504,
+    https://github.com/fmtlib/fmt/pull/1505,
+    https://github.com/fmtlib/fmt/pull/1512,
+    https://github.com/fmtlib/fmt/issues/1515,
+    https://github.com/fmtlib/fmt/pull/1516,
+    https://github.com/fmtlib/fmt/pull/1518,
+    https://github.com/fmtlib/fmt/pull/1519,
+    https://github.com/fmtlib/fmt/pull/1520,
+    https://github.com/fmtlib/fmt/pull/1521,
+    https://github.com/fmtlib/fmt/pull/1522,
+    https://github.com/fmtlib/fmt/issues/1524,
+    https://github.com/fmtlib/fmt/pull/1530,
+    https://github.com/fmtlib/fmt/issues/1531,
+    https://github.com/fmtlib/fmt/pull/1532,
+    https://github.com/fmtlib/fmt/issues/1539,
+    https://github.com/fmtlib/fmt/issues/1547,
+    https://github.com/fmtlib/fmt/issues/1548,
+    https://github.com/fmtlib/fmt/pull/1554,
+    https://github.com/fmtlib/fmt/issues/1567,
+    https://github.com/fmtlib/fmt/pull/1568,
+    https://github.com/fmtlib/fmt/pull/1569,
+    https://github.com/fmtlib/fmt/pull/1571,
+    https://github.com/fmtlib/fmt/pull/1573,
+    https://github.com/fmtlib/fmt/pull/1575,
+    https://github.com/fmtlib/fmt/pull/1581,
+    https://github.com/fmtlib/fmt/issues/1583,
+    https://github.com/fmtlib/fmt/issues/1586,
+    https://github.com/fmtlib/fmt/issues/1587,
+    https://github.com/fmtlib/fmt/issues/1594,
+    https://github.com/fmtlib/fmt/pull/1596,
+    https://github.com/fmtlib/fmt/issues/1604,
+    https://github.com/fmtlib/fmt/pull/1606,
+    https://github.com/fmtlib/fmt/issues/1607,
+    https://github.com/fmtlib/fmt/issues/1609).
+    Thanks @marti4d, @iPherian, @parkertomatoes, @gsjaardema, @chronoxor,
+    @DanielaE, @torsten48, @tohammer, @lefticus, @ryusakki, @adnsv, @fghzxm,
+    @refnum, @pramodk, @Spirrwell and @scramsby.
+
+# 6.1.2 - 2019-12-11
+
+-   Fixed ABI compatibility with `libfmt.so.6.0.0`
+    (https://github.com/fmtlib/fmt/issues/1471).
+-   Fixed handling types convertible to `std::string_view`
+    (https://github.com/fmtlib/fmt/pull/1451). Thanks @denizevrenci.
+-   Made CUDA test an opt-in enabled via the `FMT_CUDA_TEST` CMake
+    option.
+-   Fixed sign conversion warnings
+    (https://github.com/fmtlib/fmt/pull/1440). Thanks @0x8000-0000.
+
+# 6.1.1 - 2019-12-04
+
+-   Fixed shared library build on Windows
+    (https://github.com/fmtlib/fmt/pull/1443,
+    https://github.com/fmtlib/fmt/issues/1445,
+    https://github.com/fmtlib/fmt/pull/1446,
+    https://github.com/fmtlib/fmt/issues/1450).
+    Thanks @egorpugin and @bbolli.
+-   Added a missing decimal point in exponent notation with trailing
+    zeros.
+-   Removed deprecated `format_arg_store::TYPES`.
+
+# 6.1.0 - 2019-12-01
+
+-   {fmt} now formats IEEE 754 `float` and `double` using the shortest
+    decimal representation with correct rounding by default:
+
+    ```c++
+    #include <cmath>
+    #include <fmt/core.h>
+
+    int main() {
+      fmt::print("{}", M_PI);
+    }
+    ```
+
+    prints `3.141592653589793`.
+
+-   Made the fast binary to decimal floating-point formatter the
+    default, simplified it and improved performance. {fmt} is now 15
+    times faster than libc++\'s `std::ostringstream`, 11 times faster
+    than `printf` and 10% faster than double-conversion on
+    [dtoa-benchmark](https://github.com/fmtlib/dtoa-benchmark):
+
+    | Function      | Time (ns) | Speedup |
+    | ------------- | --------: | ------: |
+    | ostringstream | 1,346.30  | 1.00x   |
+    | ostrstream    | 1,195.74  | 1.13x   |
+    | sprintf       | 995.08    | 1.35x   |
+    | doubleconv    | 99.10     | 13.59x  |
+    | fmt           | 88.34     | 15.24x  |
+
+    ![](https://user-images.githubusercontent.com/576385/69767160-cdaca400-112f-11ea-9fc5-347c9f83caad.png)
+
+-   {fmt} no longer converts `float` arguments to `double`. In
+    particular this improves the default (shortest) representation of
+    floats and makes `fmt::format` consistent with `std::format` specs
+    (https://github.com/fmtlib/fmt/issues/1336,
+    https://github.com/fmtlib/fmt/issues/1353,
+    https://github.com/fmtlib/fmt/pull/1360,
+    https://github.com/fmtlib/fmt/pull/1361):
+
+    ```c++
+    fmt::print("{}", 0.1f);
+    ```
+
+    prints `0.1` instead of `0.10000000149011612`.
+
+    Thanks @orivej.
+
+-   Made floating-point formatting output consistent with
+    `printf`/iostreams
+    (https://github.com/fmtlib/fmt/issues/1376,
+    https://github.com/fmtlib/fmt/issues/1417).
+
+-   Added support for 128-bit integers
+    (https://github.com/fmtlib/fmt/pull/1287):
+
+    ```c++
+    fmt::print("{}", std::numeric_limits<__int128_t>::max());
+    ```
+
+    prints `170141183460469231731687303715884105727`.
+
+    Thanks @denizevrenci.
+
+-   The overload of `print` that takes `text_style` is now atomic, i.e.
+    the output from different threads doesn\'t interleave
+    (https://github.com/fmtlib/fmt/pull/1351). Thanks @tankiJong.
+
+-   Made compile time in the header-only mode \~20% faster by reducing
+    the number of template instantiations. `wchar_t` overload of
+    `vprint` was moved from `fmt/core.h` to `fmt/format.h`.
+
+-   Added an overload of `fmt::join` that works with tuples
+    (https://github.com/fmtlib/fmt/issues/1322,
+    https://github.com/fmtlib/fmt/pull/1330):
+
+    ```c++
+    #include <tuple>
+    #include <fmt/ranges.h>
+
+    int main() {
+      std::tuple<char, int, float> t{'a', 1, 2.0f};
+      fmt::print("{}", t);
+    }
+    ```
+
+    prints `('a', 1, 2.0)`.
+
+    Thanks @jeremyong.
+
+-   Changed formatting of octal zero with prefix from \"00\" to \"0\":
+
+    ```c++
+    fmt::print("{:#o}", 0);
+    ```
+
+    prints `0`.
+
+-   The locale is now passed to ostream insertion (`<<`) operators
+    (https://github.com/fmtlib/fmt/pull/1406):
+
+    ```c++
+    #include <fmt/locale.h>
+    #include <fmt/ostream.h>
+
+    struct S {
+      double value;
+    };
+
+    std::ostream& operator<<(std::ostream& os, S s) {
+      return os << s.value;
+    }
+
+    int main() {
+      auto s = fmt::format(std::locale("fr_FR.UTF-8"), "{}", S{0.42});
+      // s == "0,42"
+    }
+    ```
+
+    Thanks @dlaugt.
+
+-   Locale-specific number formatting now uses grouping
+    (https://github.com/fmtlib/fmt/issues/1393,
+    https://github.com/fmtlib/fmt/pull/1394). Thanks @skrdaniel.
+
+-   Fixed handling of types with deleted implicit rvalue conversion to
+    `const char**` (https://github.com/fmtlib/fmt/issues/1421):
+
+    ```c++
+    struct mystring {
+      operator const char*() const&;
+      operator const char*() &;
+      operator const char*() const&& = delete;
+      operator const char*() && = delete;
+    };
+    mystring str;
+    fmt::print("{}", str); // now compiles
+    ```
+
+-   Enums are now mapped to correct underlying types instead of `int`
+    (https://github.com/fmtlib/fmt/pull/1286). Thanks @agmt.
+
+-   Enum classes are no longer implicitly converted to `int`
+    (https://github.com/fmtlib/fmt/issues/1424).
+
+-   Added `basic_format_parse_context` for consistency with C++20
+    `std::format` and deprecated `basic_parse_context`.
+
+-   Fixed handling of UTF-8 in precision
+    (https://github.com/fmtlib/fmt/issues/1389,
+    https://github.com/fmtlib/fmt/pull/1390). Thanks @tajtiattila.
+
+-   {fmt} can now be installed on Linux, macOS and Windows with
+    [Conda](https://docs.conda.io/en/latest/) using its
+    [conda-forge](https://conda-forge.org)
+    [package](https://github.com/conda-forge/fmt-feedstock)
+    (https://github.com/fmtlib/fmt/pull/1410):
+
+        conda install -c conda-forge fmt
+
+    Thanks @tdegeus.
+
+-   Added a CUDA test (https://github.com/fmtlib/fmt/pull/1285,
+    https://github.com/fmtlib/fmt/pull/1317).
+    Thanks @luncliff and @risa2000.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/pull/1276,
+    https://github.com/fmtlib/fmt/issues/1291,
+    https://github.com/fmtlib/fmt/issues/1296,
+    https://github.com/fmtlib/fmt/pull/1315,
+    https://github.com/fmtlib/fmt/pull/1332,
+    https://github.com/fmtlib/fmt/pull/1337,
+    https://github.com/fmtlib/fmt/issues/1395
+    https://github.com/fmtlib/fmt/pull/1418).
+    Thanks @waywardmonkeys, @pauldreik and @jackoalan.
+
+-   Various code improvements
+    (https://github.com/fmtlib/fmt/pull/1358,
+    https://github.com/fmtlib/fmt/pull/1407).
+    Thanks @orivej and @dpacbach.
+
+-   Fixed compile-time format string checks for user-defined types
+    (https://github.com/fmtlib/fmt/issues/1292).
+
+-   Worked around a false positive in `unsigned-integer-overflow` sanitizer
+    (https://github.com/fmtlib/fmt/issues/1377).
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/issues/1273,
+    https://github.com/fmtlib/fmt/pull/1278,
+    https://github.com/fmtlib/fmt/pull/1280,
+    https://github.com/fmtlib/fmt/issues/1281,
+    https://github.com/fmtlib/fmt/issues/1288,
+    https://github.com/fmtlib/fmt/pull/1290,
+    https://github.com/fmtlib/fmt/pull/1301,
+    https://github.com/fmtlib/fmt/issues/1305,
+    https://github.com/fmtlib/fmt/issues/1306,
+    https://github.com/fmtlib/fmt/issues/1309,
+    https://github.com/fmtlib/fmt/pull/1312,
+    https://github.com/fmtlib/fmt/issues/1313,
+    https://github.com/fmtlib/fmt/issues/1316,
+    https://github.com/fmtlib/fmt/issues/1319,
+    https://github.com/fmtlib/fmt/pull/1320,
+    https://github.com/fmtlib/fmt/pull/1326,
+    https://github.com/fmtlib/fmt/pull/1328,
+    https://github.com/fmtlib/fmt/issues/1344,
+    https://github.com/fmtlib/fmt/pull/1345,
+    https://github.com/fmtlib/fmt/pull/1347,
+    https://github.com/fmtlib/fmt/pull/1349,
+    https://github.com/fmtlib/fmt/issues/1354,
+    https://github.com/fmtlib/fmt/issues/1362,
+    https://github.com/fmtlib/fmt/issues/1366,
+    https://github.com/fmtlib/fmt/pull/1364,
+    https://github.com/fmtlib/fmt/pull/1370,
+    https://github.com/fmtlib/fmt/pull/1371,
+    https://github.com/fmtlib/fmt/issues/1385,
+    https://github.com/fmtlib/fmt/issues/1388,
+    https://github.com/fmtlib/fmt/pull/1397,
+    https://github.com/fmtlib/fmt/pull/1414,
+    https://github.com/fmtlib/fmt/pull/1416,
+    https://github.com/fmtlib/fmt/issues/1422
+    https://github.com/fmtlib/fmt/pull/1427,
+    https://github.com/fmtlib/fmt/issues/1431,
+    https://github.com/fmtlib/fmt/pull/1433).
+    Thanks @hhb, @gsjaardema, @gabime, @neheb, @vedranmiletic, @dkavolis,
+    @mwinterb, @orivej, @denizevrenci, @leonklingele, @chronoxor, @kent-tri,
+    @0x8000-0000 and @marti4d.
+
+# 6.0.0 - 2019-08-26
+
+-   Switched to the [MIT license](
+    https://github.com/fmtlib/fmt/blob/5a4b24613ba16cc689977c3b5bd8274a3ba1dd1f/LICENSE.rst)
+    with an optional exception that allows distributing binary code
+    without attribution.
+
+-   Floating-point formatting is now locale-independent by default:
+
+    ```c++
+    #include <locale>
+    #include <fmt/core.h>
+
+    int main() {
+      std::locale::global(std::locale("ru_RU.UTF-8"));
+      fmt::print("value = {}", 4.2);
+    }
+    ```
+
+    prints \"value = 4.2\" regardless of the locale.
+
+    For locale-specific formatting use the `n` specifier:
+
+    ```c++
+    std::locale::global(std::locale("ru_RU.UTF-8"));
+    fmt::print("value = {:n}", 4.2);
+    ```
+
+    prints \"value = 4,2\".
+
+-   Added an experimental Grisu floating-point formatting algorithm
+    implementation (disabled by default). To enable it compile with the
+    `FMT_USE_GRISU` macro defined to 1:
+
+    ```c++
+    #define FMT_USE_GRISU 1
+    #include <fmt/format.h>
+
+    auto s = fmt::format("{}", 4.2); // formats 4.2 using Grisu
+    ```
+
+    With Grisu enabled, {fmt} is 13x faster than `std::ostringstream`
+    (libc++) and 10x faster than `sprintf` on
+    [dtoa-benchmark](https://github.com/fmtlib/dtoa-benchmark) ([full
+    results](https://fmt.dev/unknown_mac64_clang10.0.html)):
+
+    ![](https://user-images.githubusercontent.com/576385/54883977-9fe8c000-4e28-11e9-8bde-272d122e7c52.jpg)
+
+-   Separated formatting and parsing contexts for consistency with
+    [C++20 std::format](http://eel.is/c++draft/format), removing the
+    undocumented `basic_format_context::parse_context()` function.
+
+-   Added [oss-fuzz](https://github.com/google/oss-fuzz) support
+    (https://github.com/fmtlib/fmt/pull/1199). Thanks @pauldreik.
+
+-   `formatter` specializations now always take precedence over
+    `operator<<` (https://github.com/fmtlib/fmt/issues/952):
+
+    ```c++
+    #include <iostream>
+    #include <fmt/ostream.h>
+
+    struct S {};
+
+    std::ostream& operator<<(std::ostream& os, S) {
+      return os << 1;
+    }
+
+    template <>
+    struct fmt::formatter<S> : fmt::formatter<int> {
+      auto format(S, format_context& ctx) {
+        return formatter<int>::format(2, ctx);
+      }
+    };
+
+    int main() {
+      std::cout << S() << "\n"; // prints 1 using operator<<
+      fmt::print("{}\n", S());  // prints 2 using formatter
+    }
+    ```
+
+-   Introduced the experimental `fmt::compile` function that does format
+    string compilation
+    (https://github.com/fmtlib/fmt/issues/618,
+    https://github.com/fmtlib/fmt/issues/1169,
+    https://github.com/fmtlib/fmt/pull/1171):
+
+    ```c++
+    #include <fmt/compile.h>
+
+    auto f = fmt::compile<int>("{}");
+    std::string s = fmt::format(f, 42); // can be called multiple times to
+                                        // format different values
+    // s == "42"
+    ```
+
+    It moves the cost of parsing a format string outside of the format
+    function which can be beneficial when identically formatting many
+    objects of the same types. Thanks @stryku.
+
+-   Added experimental `%` format specifier that formats floating-point
+    values as percentages
+    (https://github.com/fmtlib/fmt/pull/1060,
+    https://github.com/fmtlib/fmt/pull/1069,
+    https://github.com/fmtlib/fmt/pull/1071):
+
+    ```c++
+    auto s = fmt::format("{:.1%}", 0.42); // s == "42.0%"
+    ```
+
+    Thanks @gawain-bolton.
+
+-   Implemented precision for floating-point durations
+    (https://github.com/fmtlib/fmt/issues/1004,
+    https://github.com/fmtlib/fmt/pull/1012):
+
+    ```c++
+    auto s = fmt::format("{:.1}", std::chrono::duration<double>(1.234));
+    // s == 1.2s
+    ```
+
+    Thanks @DanielaE.
+
+-   Implemented `chrono` format specifiers `%Q` and `%q` that give the
+    value and the unit respectively
+    (https://github.com/fmtlib/fmt/pull/1019):
+
+    ```c++
+    auto value = fmt::format("{:%Q}", 42s); // value == "42"
+    auto unit  = fmt::format("{:%q}", 42s); // unit == "s"
+    ```
+
+    Thanks @DanielaE.
+
+-   Fixed handling of dynamic width in chrono formatter:
+
+    ```c++
+    auto s = fmt::format("{0:{1}%H:%M:%S}", std::chrono::seconds(12345), 12);
+    //                        ^ width argument index                     ^ width
+    // s == "03:25:45    "
+    ```
+
+    Thanks Howard Hinnant.
+
+-   Removed deprecated `fmt/time.h`. Use `fmt/chrono.h` instead.
+
+-   Added `fmt::format` and `fmt::vformat` overloads that take
+    `text_style` (https://github.com/fmtlib/fmt/issues/993,
+    https://github.com/fmtlib/fmt/pull/994):
+
+    ```c++
+    #include <fmt/color.h>
+
+    std::string message = fmt::format(fmt::emphasis::bold | fg(fmt::color::red),
+                                      "The answer is {}.", 42);
+    ```
+
+    Thanks @Naios.
+
+-   Removed the deprecated color API (`print_colored`). Use the new API,
+    namely `print` overloads that take `text_style` instead.
+
+-   Made `std::unique_ptr` and `std::shared_ptr` formattable as pointers
+    via `fmt::ptr` (https://github.com/fmtlib/fmt/pull/1121):
+
+    ```c++
+    std::unique_ptr<int> p = ...;
+    fmt::print("{}", fmt::ptr(p)); // prints p as a pointer
+    ```
+
+    Thanks @sighingnow.
+
+-   Made `print` and `vprint` report I/O errors
+    (https://github.com/fmtlib/fmt/issues/1098,
+    https://github.com/fmtlib/fmt/pull/1099). Thanks @BillyDonahue.
+
+-   Marked deprecated APIs with the `[[deprecated]]` attribute and
+    removed internal uses of deprecated APIs
+    (https://github.com/fmtlib/fmt/pull/1022). Thanks @eliaskosunen.
+
+-   Modernized the codebase using more C++11 features and removing
+    workarounds. Most importantly, `buffer_context` is now an alias
+    template, so use `buffer_context<T>` instead of
+    `buffer_context<T>::type`. These features require GCC 4.8 or later.
+
+-   `formatter` specializations now always take precedence over implicit
+    conversions to `int` and the undocumented `convert_to_int` trait is
+    now deprecated.
+
+-   Moved the undocumented `basic_writer`, `writer`, and `wwriter` types
+    to the `internal` namespace.
+
+-   Removed deprecated `basic_format_context::begin()`. Use `out()`
+    instead.
+
+-   Disallowed passing the result of `join` as an lvalue to prevent
+    misuse.
+
+-   Refactored the undocumented structs that represent parsed format
+    specifiers to simplify the API and allow multibyte fill.
+
+-   Moved SFINAE to template parameters to reduce symbol sizes.
+
+-   Switched to `fputws` for writing wide strings so that it\'s no
+    longer required to call `_setmode` on Windows
+    (https://github.com/fmtlib/fmt/issues/1229,
+    https://github.com/fmtlib/fmt/pull/1243). Thanks @jackoalan.
+
+-   Improved literal-based API
+    (https://github.com/fmtlib/fmt/pull/1254). Thanks @sylveon.
+
+-   Added support for exotic platforms without `uintptr_t` such as IBM i
+    (AS/400) which has 128-bit pointers and only 64-bit integers
+    (https://github.com/fmtlib/fmt/issues/1059).
+
+-   Added [Sublime Text syntax highlighting config](
+    https://github.com/fmtlib/fmt/blob/master/support/C%2B%2B.sublime-syntax)
+    (https://github.com/fmtlib/fmt/issues/1037). Thanks @Kronuz.
+
+-   Added the `FMT_ENFORCE_COMPILE_STRING` macro to enforce the use of
+    compile-time format strings
+    (https://github.com/fmtlib/fmt/pull/1231). Thanks @jackoalan.
+
+-   Stopped setting `CMAKE_BUILD_TYPE` if {fmt} is a subproject
+    (https://github.com/fmtlib/fmt/issues/1081).
+
+-   Various build improvements
+    (https://github.com/fmtlib/fmt/pull/1039,
+    https://github.com/fmtlib/fmt/pull/1078,
+    https://github.com/fmtlib/fmt/pull/1091,
+    https://github.com/fmtlib/fmt/pull/1103,
+    https://github.com/fmtlib/fmt/pull/1177).
+    Thanks @luncliff, @jasonszang, @olafhering, @Lecetem and @pauldreik.
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/1049,
+    https://github.com/fmtlib/fmt/pull/1051,
+    https://github.com/fmtlib/fmt/pull/1083,
+    https://github.com/fmtlib/fmt/pull/1113,
+    https://github.com/fmtlib/fmt/pull/1114,
+    https://github.com/fmtlib/fmt/issues/1146,
+    https://github.com/fmtlib/fmt/issues/1180,
+    https://github.com/fmtlib/fmt/pull/1250,
+    https://github.com/fmtlib/fmt/pull/1252,
+    https://github.com/fmtlib/fmt/pull/1265).
+    Thanks @mikelui, @foonathan, @BillyDonahue, @jwakely, @kaisbe and
+    @sdebionne.
+
+-   Fixed ambiguous formatter specialization in `fmt/ranges.h`
+    (https://github.com/fmtlib/fmt/issues/1123).
+
+-   Fixed formatting of a non-empty `std::filesystem::path` which is an
+    infinitely deep range of its components
+    (https://github.com/fmtlib/fmt/issues/1268).
+
+-   Fixed handling of general output iterators when formatting
+    characters (https://github.com/fmtlib/fmt/issues/1056,
+    https://github.com/fmtlib/fmt/pull/1058). Thanks @abolz.
+
+-   Fixed handling of output iterators in `formatter` specialization for
+    ranges (https://github.com/fmtlib/fmt/issues/1064).
+
+-   Fixed handling of exotic character types
+    (https://github.com/fmtlib/fmt/issues/1188).
+
+-   Made chrono formatting work with exceptions disabled
+    (https://github.com/fmtlib/fmt/issues/1062).
+
+-   Fixed DLL visibility issues
+    (https://github.com/fmtlib/fmt/pull/1134,
+    https://github.com/fmtlib/fmt/pull/1147). Thanks @denchat.
+
+-   Disabled the use of UDL template extension on GCC 9
+    (https://github.com/fmtlib/fmt/issues/1148).
+
+-   Removed misplaced `format` compile-time checks from `printf`
+    (https://github.com/fmtlib/fmt/issues/1173).
+
+-   Fixed issues in the experimental floating-point formatter
+    (https://github.com/fmtlib/fmt/issues/1072,
+    https://github.com/fmtlib/fmt/issues/1129,
+    https://github.com/fmtlib/fmt/issues/1153,
+    https://github.com/fmtlib/fmt/pull/1155,
+    https://github.com/fmtlib/fmt/issues/1210,
+    https://github.com/fmtlib/fmt/issues/1222). Thanks @alabuzhev.
+
+-   Fixed bugs discovered by fuzzing or during fuzzing integration
+    (https://github.com/fmtlib/fmt/issues/1124,
+    https://github.com/fmtlib/fmt/issues/1127,
+    https://github.com/fmtlib/fmt/issues/1132,
+    https://github.com/fmtlib/fmt/pull/1135,
+    https://github.com/fmtlib/fmt/issues/1136,
+    https://github.com/fmtlib/fmt/issues/1141,
+    https://github.com/fmtlib/fmt/issues/1142,
+    https://github.com/fmtlib/fmt/issues/1178,
+    https://github.com/fmtlib/fmt/issues/1179,
+    https://github.com/fmtlib/fmt/issues/1194). Thanks @pauldreik.
+
+-   Fixed building tests on FreeBSD and Hurd
+    (https://github.com/fmtlib/fmt/issues/1043). Thanks @jackyf.
+
+-   Fixed various warnings and compilation issues
+    (https://github.com/fmtlib/fmt/pull/998,
+    https://github.com/fmtlib/fmt/pull/1006,
+    https://github.com/fmtlib/fmt/issues/1008,
+    https://github.com/fmtlib/fmt/issues/1011,
+    https://github.com/fmtlib/fmt/issues/1025,
+    https://github.com/fmtlib/fmt/pull/1027,
+    https://github.com/fmtlib/fmt/pull/1028,
+    https://github.com/fmtlib/fmt/pull/1029,
+    https://github.com/fmtlib/fmt/pull/1030,
+    https://github.com/fmtlib/fmt/pull/1031,
+    https://github.com/fmtlib/fmt/pull/1054,
+    https://github.com/fmtlib/fmt/issues/1063,
+    https://github.com/fmtlib/fmt/pull/1068,
+    https://github.com/fmtlib/fmt/pull/1074,
+    https://github.com/fmtlib/fmt/pull/1075,
+    https://github.com/fmtlib/fmt/pull/1079,
+    https://github.com/fmtlib/fmt/pull/1086,
+    https://github.com/fmtlib/fmt/issues/1088,
+    https://github.com/fmtlib/fmt/pull/1089,
+    https://github.com/fmtlib/fmt/pull/1094,
+    https://github.com/fmtlib/fmt/issues/1101,
+    https://github.com/fmtlib/fmt/pull/1102,
+    https://github.com/fmtlib/fmt/issues/1105,
+    https://github.com/fmtlib/fmt/pull/1107,
+    https://github.com/fmtlib/fmt/issues/1115,
+    https://github.com/fmtlib/fmt/issues/1117,
+    https://github.com/fmtlib/fmt/issues/1118,
+    https://github.com/fmtlib/fmt/issues/1120,
+    https://github.com/fmtlib/fmt/issues/1123,
+    https://github.com/fmtlib/fmt/pull/1139,
+    https://github.com/fmtlib/fmt/issues/1140,
+    https://github.com/fmtlib/fmt/issues/1143,
+    https://github.com/fmtlib/fmt/pull/1144,
+    https://github.com/fmtlib/fmt/pull/1150,
+    https://github.com/fmtlib/fmt/pull/1151,
+    https://github.com/fmtlib/fmt/issues/1152,
+    https://github.com/fmtlib/fmt/issues/1154,
+    https://github.com/fmtlib/fmt/issues/1156,
+    https://github.com/fmtlib/fmt/pull/1159,
+    https://github.com/fmtlib/fmt/issues/1175,
+    https://github.com/fmtlib/fmt/issues/1181,
+    https://github.com/fmtlib/fmt/issues/1186,
+    https://github.com/fmtlib/fmt/pull/1187,
+    https://github.com/fmtlib/fmt/pull/1191,
+    https://github.com/fmtlib/fmt/issues/1197,
+    https://github.com/fmtlib/fmt/issues/1200,
+    https://github.com/fmtlib/fmt/issues/1203,
+    https://github.com/fmtlib/fmt/issues/1205,
+    https://github.com/fmtlib/fmt/pull/1206,
+    https://github.com/fmtlib/fmt/issues/1213,
+    https://github.com/fmtlib/fmt/issues/1214,
+    https://github.com/fmtlib/fmt/pull/1217,
+    https://github.com/fmtlib/fmt/issues/1228,
+    https://github.com/fmtlib/fmt/pull/1230,
+    https://github.com/fmtlib/fmt/issues/1232,
+    https://github.com/fmtlib/fmt/pull/1235,
+    https://github.com/fmtlib/fmt/pull/1236,
+    https://github.com/fmtlib/fmt/issues/1240).
+    Thanks @DanielaE, @mwinterb, @eliaskosunen, @morinmorin, @ricco19,
+    @waywardmonkeys, @chronoxor, @remyabel, @pauldreik, @gsjaardema, @rcane,
+    @mocabe, @denchat, @cjdb, @HazardyKnusperkeks, @vedranmiletic, @jackoalan,
+    @DaanDeMeyer and @starkmapper.
+
+# 5.3.0 - 2018-12-28
+
+-   Introduced experimental chrono formatting support:
+
+    ```c++
+    #include <fmt/chrono.h>
+
+    int main() {
+      using namespace std::literals::chrono_literals;
+      fmt::print("Default format: {} {}\n", 42s, 100ms);
+      fmt::print("strftime-like format: {:%H:%M:%S}\n", 3h + 15min + 30s);
+    }
+    ```
+
+    prints:
+
+        Default format: 42s 100ms
+        strftime-like format: 03:15:30
+
+-   Added experimental support for emphasis (bold, italic, underline,
+    strikethrough), colored output to a file stream, and improved
+    colored formatting API
+    (https://github.com/fmtlib/fmt/pull/961,
+    https://github.com/fmtlib/fmt/pull/967,
+    https://github.com/fmtlib/fmt/pull/973):
+
+    ```c++
+    #include <fmt/color.h>
+
+    int main() {
+      print(fg(fmt::color::crimson) | fmt::emphasis::bold,
+            "Hello, {}!\n", "world");
+      print(fg(fmt::color::floral_white) | bg(fmt::color::slate_gray) |
+            fmt::emphasis::underline, "Hello, {}!\n", "мир");
+      print(fg(fmt::color::steel_blue) | fmt::emphasis::italic,
+            "Hello, {}!\n", "世界");
+    }
+    ```
+
+    prints the following on modern terminals with RGB color support:
+
+    ![](https://user-images.githubusercontent.com/576385/50405788-b66e7500-076e-11e9-9592-7324d1f951d8.png)
+
+    Thanks @Rakete1111.
+
+-   Added support for 4-bit terminal colors
+    (https://github.com/fmtlib/fmt/issues/968,
+    https://github.com/fmtlib/fmt/pull/974)
+
+    ```c++
+    #include <fmt/color.h>
+
+    int main() {
+      print(fg(fmt::terminal_color::red), "stop\n");
+    }
+    ```
+
+    Note that these colors vary by terminal:
+
+    ![](https://user-images.githubusercontent.com/576385/50405925-dbfc7e00-0770-11e9-9b85-333fab0af9ac.png)
+
+    Thanks @Rakete1111.
+
+-   Parameterized formatting functions on the type of the format string
+    (https://github.com/fmtlib/fmt/issues/880,
+    https://github.com/fmtlib/fmt/pull/881,
+    https://github.com/fmtlib/fmt/pull/883,
+    https://github.com/fmtlib/fmt/pull/885,
+    https://github.com/fmtlib/fmt/pull/897,
+    https://github.com/fmtlib/fmt/issues/920). Any object of
+    type `S` that has an overloaded `to_string_view(const S&)` returning
+    `fmt::string_view` can be used as a format string:
+
+    ```c++
+    namespace my_ns {
+    inline string_view to_string_view(const my_string& s) {
+      return {s.data(), s.length()};
+    }
+    }
+
+    std::string message = fmt::format(my_string("The answer is {}."), 42);
+    ```
+
+    Thanks @DanielaE.
+
+-   Made `std::string_view` work as a format string
+    (https://github.com/fmtlib/fmt/pull/898):
+
+    ```c++
+    auto message = fmt::format(std::string_view("The answer is {}."), 42);
+    ```
+
+    Thanks @DanielaE.
+
+-   Added wide string support to compile-time format string checks
+    (https://github.com/fmtlib/fmt/pull/924):
+
+    ```c++
+    print(fmt(L"{:f}"), 42); // compile-time error: invalid type specifier
+    ```
+
+    Thanks @XZiar.
+
+-   Made colored print functions work with wide strings
+    (https://github.com/fmtlib/fmt/pull/867):
+
+    ```c++
+    #include <fmt/color.h>
+
+    int main() {
+      print(fg(fmt::color::red), L"{}\n", 42);
+    }
+    ```
+
+    Thanks @DanielaE.
+
+-   Introduced experimental Unicode support
+    (https://github.com/fmtlib/fmt/issues/628,
+    https://github.com/fmtlib/fmt/pull/891):
+
+    ```c++
+    using namespace fmt::literals;
+    auto s = fmt::format("{:*^5}"_u, "🤡"_u); // s == "**🤡**"_u
+    ```
+
+-   Improved locale support:
+
+    ```c++
+    #include <fmt/locale.h>
+
+    struct numpunct : std::numpunct<char> {
+     protected:
+      char do_thousands_sep() const override { return '~'; }
+    };
+
+    std::locale loc;
+    auto s = fmt::format(std::locale(loc, new numpunct()), "{:n}", 1234567);
+    // s == "1~234~567"
+    ```
+
+-   Constrained formatting functions on proper iterator types
+    (https://github.com/fmtlib/fmt/pull/921). Thanks @DanielaE.
+
+-   Added `make_printf_args` and `make_wprintf_args` functions
+    (https://github.com/fmtlib/fmt/pull/934). Thanks @tnovotny.
+
+-   Deprecated `fmt::visit`, `parse_context`, and `wparse_context`. Use
+    `fmt::visit_format_arg`, `format_parse_context`, and
+    `wformat_parse_context` instead.
+
+-   Removed undocumented `basic_fixed_buffer` which has been superseded
+    by the iterator-based API
+    (https://github.com/fmtlib/fmt/issues/873,
+    https://github.com/fmtlib/fmt/pull/902). Thanks @superfunc.
+
+-   Disallowed repeated leading zeros in an argument ID:
+
+    ```c++
+    fmt::print("{000}", 42); // error
+    ```
+
+-   Reintroduced support for gcc 4.4.
+
+-   Fixed compilation on platforms with exotic `double`
+    (https://github.com/fmtlib/fmt/issues/878).
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/issues/164,
+    https://github.com/fmtlib/fmt/issues/877,
+    https://github.com/fmtlib/fmt/pull/901,
+    https://github.com/fmtlib/fmt/pull/906,
+    https://github.com/fmtlib/fmt/pull/979).
+    Thanks @kookjr, @DarkDimius and @HecticSerenity.
+
+-   Added pkgconfig support which makes it easier to consume the library
+    from meson and other build systems
+    (https://github.com/fmtlib/fmt/pull/916). Thanks @colemickens.
+
+-   Various build improvements
+    (https://github.com/fmtlib/fmt/pull/909,
+    https://github.com/fmtlib/fmt/pull/926,
+    https://github.com/fmtlib/fmt/pull/937,
+    https://github.com/fmtlib/fmt/pull/953,
+    https://github.com/fmtlib/fmt/pull/959).
+    Thanks @tchaikov, @luncliff, @AndreasSchoenle, @hotwatermorning and @Zefz.
+
+-   Improved `string_view` construction performance
+    (https://github.com/fmtlib/fmt/pull/914). Thanks @gabime.
+
+-   Fixed non-matching char types
+    (https://github.com/fmtlib/fmt/pull/895). Thanks @DanielaE.
+
+-   Fixed `format_to_n` with `std::back_insert_iterator`
+    (https://github.com/fmtlib/fmt/pull/913). Thanks @DanielaE.
+
+-   Fixed locale-dependent formatting
+    (https://github.com/fmtlib/fmt/issues/905).
+
+-   Fixed various compiler warnings and errors
+    (https://github.com/fmtlib/fmt/pull/882,
+    https://github.com/fmtlib/fmt/pull/886,
+    https://github.com/fmtlib/fmt/pull/933,
+    https://github.com/fmtlib/fmt/pull/941,
+    https://github.com/fmtlib/fmt/issues/931,
+    https://github.com/fmtlib/fmt/pull/943,
+    https://github.com/fmtlib/fmt/pull/954,
+    https://github.com/fmtlib/fmt/pull/956,
+    https://github.com/fmtlib/fmt/pull/962,
+    https://github.com/fmtlib/fmt/issues/965,
+    https://github.com/fmtlib/fmt/issues/977,
+    https://github.com/fmtlib/fmt/pull/983,
+    https://github.com/fmtlib/fmt/pull/989).
+    Thanks @Luthaf, @stevenhoving, @christinaa, @lgritz, @DanielaE,
+    @0x8000-0000 and @liuping1997.
+
+# 5.2.1 - 2018-09-21
+
+-   Fixed `visit` lookup issues on gcc 7 & 8
+    (https://github.com/fmtlib/fmt/pull/870). Thanks @medithe.
+-   Fixed linkage errors on older gcc.
+-   Prevented `fmt/range.h` from specializing `fmt::basic_string_view`
+    (https://github.com/fmtlib/fmt/issues/865,
+    https://github.com/fmtlib/fmt/pull/868). Thanks @hhggit.
+-   Improved error message when formatting unknown types
+    (https://github.com/fmtlib/fmt/pull/872). Thanks @foonathan.
+-   Disabled templated user-defined literals when compiled under nvcc
+    (https://github.com/fmtlib/fmt/pull/875). Thanks @CandyGumdrop.
+-   Fixed `format_to` formatting to `wmemory_buffer`
+    (https://github.com/fmtlib/fmt/issues/874).
+
+# 5.2.0 - 2018-09-13
+
+-   Optimized format string parsing and argument processing which
+    resulted in up to 5x speed up on long format strings and significant
+    performance boost on various benchmarks. For example, version 5.2 is
+    2.22x faster than 5.1 on decimal integer formatting with `format_to`
+    (macOS, clang-902.0.39.2):
+
+    | Method                     | Time, s         | Speedup |
+    | -------------------------- | --------------: | ------: |
+    | fmt::format 5.1            | 0.58            |         |
+    | fmt::format 5.2            | 0.35            |   1.66x |
+    | fmt::format_to 5.1         | 0.51            |         |
+    | fmt::format_to 5.2         | 0.23            |   2.22x |
+    | sprintf                    | 0.71            |         |
+    | std::to_string             | 1.01            |         |
+    | std::stringstream          | 1.73            |         |
+
+-   Changed the `fmt` macro from opt-out to opt-in to prevent name
+    collisions. To enable it define the `FMT_STRING_ALIAS` macro to 1
+    before including `fmt/format.h`:
+
+    ```c++
+    #define FMT_STRING_ALIAS 1
+    #include <fmt/format.h>
+    std::string answer = format(fmt("{}"), 42);
+    ```
+
+-   Added compile-time format string checks to `format_to` overload that
+    takes `fmt::memory_buffer`
+    (https://github.com/fmtlib/fmt/issues/783):
+
+    ```c++
+    fmt::memory_buffer buf;
+    // Compile-time error: invalid type specifier.
+    fmt::format_to(buf, fmt("{:d}"), "foo");
+    ```
+
+-   Moved experimental color support to `fmt/color.h` and enabled the
+    new API by default. The old API can be enabled by defining the
+    `FMT_DEPRECATED_COLORS` macro.
+
+-   Added formatting support for types explicitly convertible to
+    `fmt::string_view`:
+
+    ```c++
+    struct foo {
+      explicit operator fmt::string_view() const { return "foo"; }
+    };
+    auto s = format("{}", foo());
+    ```
+
+    In particular, this makes formatting function work with
+    `folly::StringPiece`.
+
+-   Implemented preliminary support for `char*_t` by replacing the
+    `format` function overloads with a single function template
+    parameterized on the string type.
+
+-   Added support for dynamic argument lists
+    (https://github.com/fmtlib/fmt/issues/814,
+    https://github.com/fmtlib/fmt/pull/819). Thanks @MikePopoloski.
+
+-   Reduced executable size overhead for embedded targets using newlib
+    nano by making locale dependency optional
+    (https://github.com/fmtlib/fmt/pull/839). Thanks @teajay-fr.
+
+-   Keep `noexcept` specifier when exceptions are disabled
+    (https://github.com/fmtlib/fmt/issues/801,
+    https://github.com/fmtlib/fmt/pull/810). Thanks @qis.
+
+-   Fixed formatting of user-defined types providing `operator<<` with
+    `format_to_n` (https://github.com/fmtlib/fmt/pull/806).
+    Thanks @mkurdej.
+
+-   Fixed dynamic linkage of new symbols
+    (https://github.com/fmtlib/fmt/issues/808).
+
+-   Fixed global initialization issue
+    (https://github.com/fmtlib/fmt/issues/807):
+
+    ```c++
+    // This works on compilers with constexpr support.
+    static const std::string answer = fmt::format("{}", 42);
+    ```
+
+-   Fixed various compiler warnings and errors
+    (https://github.com/fmtlib/fmt/pull/804,
+    https://github.com/fmtlib/fmt/issues/809,
+    https://github.com/fmtlib/fmt/pull/811,
+    https://github.com/fmtlib/fmt/issues/822,
+    https://github.com/fmtlib/fmt/pull/827,
+    https://github.com/fmtlib/fmt/issues/830,
+    https://github.com/fmtlib/fmt/pull/838,
+    https://github.com/fmtlib/fmt/issues/843,
+    https://github.com/fmtlib/fmt/pull/844,
+    https://github.com/fmtlib/fmt/issues/851,
+    https://github.com/fmtlib/fmt/pull/852,
+    https://github.com/fmtlib/fmt/pull/854).
+    Thanks @henryiii, @medithe, and @eliasdaler.
+
+# 5.1.0 - 2018-07-05
+
+-   Added experimental support for RGB color output enabled with the
+    `FMT_EXTENDED_COLORS` macro:
+
+    ```c++
+    #define FMT_EXTENDED_COLORS
+    #define FMT_HEADER_ONLY // or compile fmt with FMT_EXTENDED_COLORS defined
+    #include <fmt/format.h>
+
+    fmt::print(fmt::color::steel_blue, "Some beautiful text");
+    ```
+
+    The old API (the `print_colored` and `vprint_colored` functions and
+    the `color` enum) is now deprecated.
+    (https://github.com/fmtlib/fmt/issues/762
+    https://github.com/fmtlib/fmt/pull/767). thanks @Remotion.
+
+-   Added quotes to strings in ranges and tuples
+    (https://github.com/fmtlib/fmt/pull/766). Thanks @Remotion.
+
+-   Made `format_to` work with `basic_memory_buffer`
+    (https://github.com/fmtlib/fmt/issues/776).
+
+-   Added `vformat_to_n` and `wchar_t` overload of `format_to_n`
+    (https://github.com/fmtlib/fmt/issues/764,
+    https://github.com/fmtlib/fmt/issues/769).
+
+-   Made `is_range` and `is_tuple_like` part of public (experimental)
+    API to allow specialization for user-defined types
+    (https://github.com/fmtlib/fmt/issues/751,
+    https://github.com/fmtlib/fmt/pull/759). Thanks @drrlvn.
+
+-   Added more compilers to continuous integration and increased
+    `FMT_PEDANTIC` warning levels
+    (https://github.com/fmtlib/fmt/pull/736). Thanks @eliaskosunen.
+
+-   Fixed compilation with MSVC 2013.
+
+-   Fixed handling of user-defined types in `format_to`
+    (https://github.com/fmtlib/fmt/issues/793).
+
+-   Forced linking of inline `vformat` functions into the library
+    (https://github.com/fmtlib/fmt/issues/795).
+
+-   Fixed incorrect call to on_align in `'{:}='`
+    (https://github.com/fmtlib/fmt/issues/750).
+
+-   Fixed floating-point formatting to a non-back_insert_iterator with
+    sign & numeric alignment specified
+    (https://github.com/fmtlib/fmt/issues/756).
+
+-   Fixed formatting to an array with `format_to_n`
+    (https://github.com/fmtlib/fmt/issues/778).
+
+-   Fixed formatting of more than 15 named arguments
+    (https://github.com/fmtlib/fmt/issues/754).
+
+-   Fixed handling of compile-time strings when including
+    `fmt/ostream.h`. (https://github.com/fmtlib/fmt/issues/768).
+
+-   Fixed various compiler warnings and errors
+    (https://github.com/fmtlib/fmt/issues/742,
+    https://github.com/fmtlib/fmt/issues/748,
+    https://github.com/fmtlib/fmt/issues/752,
+    https://github.com/fmtlib/fmt/issues/770,
+    https://github.com/fmtlib/fmt/pull/775,
+    https://github.com/fmtlib/fmt/issues/779,
+    https://github.com/fmtlib/fmt/pull/780,
+    https://github.com/fmtlib/fmt/pull/790,
+    https://github.com/fmtlib/fmt/pull/792,
+    https://github.com/fmtlib/fmt/pull/800).
+    Thanks @Remotion, @gabime, @foonathan, @Dark-Passenger and @0x8000-0000.
+
+# 5.0.0 - 2018-05-21
+
+-   Added a requirement for partial C++11 support, most importantly
+    variadic templates and type traits, and dropped `FMT_VARIADIC_*`
+    emulation macros. Variadic templates are available since GCC 4.4,
+    Clang 2.9 and MSVC 18.0 (2013). For older compilers use {fmt}
+    [version 4.x](https://github.com/fmtlib/fmt/releases/tag/4.1.0)
+    which continues to be maintained and works with C++98 compilers.
+
+-   Renamed symbols to follow standard C++ naming conventions and
+    proposed a subset of the library for standardization in [P0645R2
+    Text Formatting](https://wg21.link/P0645).
+
+-   Implemented `constexpr` parsing of format strings and [compile-time
+    format string
+    checks](https://fmt.dev/latest/api.html#compile-time-format-string-checks).
+    For example
+
+    ```c++
+    #include <fmt/format.h>
+
+    std::string s = format(fmt("{:d}"), "foo");
+    ```
+
+    gives a compile-time error because `d` is an invalid specifier for
+    strings ([godbolt](https://godbolt.org/g/rnCy9Q)):
+
+        ...
+        <source>:4:19: note: in instantiation of function template specialization 'fmt::v5::format<S, char [4]>' requested here
+          std::string s = format(fmt("{:d}"), "foo");
+                          ^
+        format.h:1337:13: note: non-constexpr function 'on_error' cannot be used in a constant expression
+            handler.on_error("invalid type specifier");
+
+    Compile-time checks require relaxed `constexpr` (C++14 feature)
+    support. If the latter is not available, checks will be performed at
+    runtime.
+
+-   Separated format string parsing and formatting in the extension API
+    to enable compile-time format string processing. For example
+
+    ```c++
+    struct Answer {};
+
+    namespace fmt {
+    template <>
+    struct formatter<Answer> {
+      constexpr auto parse(parse_context& ctx) {
+        auto it = ctx.begin();
+        spec = *it;
+        if (spec != 'd' && spec != 's')
+          throw format_error("invalid specifier");
+        return ++it;
+      }
+
+      template <typename FormatContext>
+      auto format(Answer, FormatContext& ctx) {
+        return spec == 's' ?
+          format_to(ctx.begin(), "{}", "fourty-two") :
+          format_to(ctx.begin(), "{}", 42);
+      }
+
+      char spec = 0;
+    };
+    }
+
+    std::string s = format(fmt("{:x}"), Answer());
+    ```
+
+    gives a compile-time error due to invalid format specifier
+    ([godbolt](https://godbolt.org/g/2jQ1Dv)):
+
+        ...
+        <source>:12:45: error: expression '<throw-expression>' is not a constant expression
+               throw format_error("invalid specifier");
+
+-   Added [iterator
+    support](https://fmt.dev/latest/api.html#output-iterator-support):
+
+    ```c++
+    #include <vector>
+    #include <fmt/format.h>
+
+    std::vector<char> out;
+    fmt::format_to(std::back_inserter(out), "{}", 42);
+    ```
+
+-   Added the
+    [format_to_n](https://fmt.dev/latest/api.html#_CPPv2N3fmt11format_to_nE8OutputItNSt6size_tE11string_viewDpRK4Args)
+    function that restricts the output to the specified number of
+    characters (https://github.com/fmtlib/fmt/issues/298):
+
+    ```c++
+    char out[4];
+    fmt::format_to_n(out, sizeof(out), "{}", 12345);
+    // out == "1234" (without terminating '\0')
+    ```
+
+-   Added the [formatted_size](
+    https://fmt.dev/latest/api.html#_CPPv2N3fmt14formatted_sizeE11string_viewDpRK4Args)
+    function for computing the output size:
+
+    ```c++
+    #include <fmt/format.h>
+
+    auto size = fmt::formatted_size("{}", 12345); // size == 5
+    ```
+
+-   Improved compile times by reducing dependencies on standard headers
+    and providing a lightweight [core
+    API](https://fmt.dev/latest/api.html#core-api):
+
+    ```c++
+    #include <fmt/core.h>
+
+    fmt::print("The answer is {}.", 42);
+    ```
+
+    See [Compile time and code
+    bloat](https://github.com/fmtlib/fmt#compile-time-and-code-bloat).
+
+-   Added the [make_format_args](
+    https://fmt.dev/latest/api.html#_CPPv2N3fmt16make_format_argsEDpRK4Args)
+    function for capturing formatting arguments:
+
+    ```c++
+    // Prints formatted error message.
+    void vreport_error(const char *format, fmt::format_args args) {
+      fmt::print("Error: ");
+      fmt::vprint(format, args);
+    }
+    template <typename... Args>
+    void report_error(const char *format, const Args & ... args) {
+      vreport_error(format, fmt::make_format_args(args...));
+    }
+    ```
+
+-   Added the `make_printf_args` function for capturing `printf`
+    arguments (https://github.com/fmtlib/fmt/issues/687,
+    https://github.com/fmtlib/fmt/pull/694). Thanks @Kronuz.
+
+-   Added prefix `v` to non-variadic functions taking `format_args` to
+    distinguish them from variadic ones:
+
+    ```c++
+    std::string vformat(string_view format_str, format_args args);
+
+    template <typename... Args>
+    std::string format(string_view format_str, const Args & ... args);
+    ```
+
+-   Added experimental support for formatting ranges, containers and
+    tuple-like types in `fmt/ranges.h`
+    (https://github.com/fmtlib/fmt/pull/735):
+
+    ```c++
+    #include <fmt/ranges.h>
+
+    std::vector<int> v = {1, 2, 3};
+    fmt::print("{}", v); // prints {1, 2, 3}
+    ```
+
+    Thanks @Remotion.
+
+-   Implemented `wchar_t` date and time formatting
+    (https://github.com/fmtlib/fmt/pull/712):
+
+    ```c++
+    #include <fmt/time.h>
+
+    std::time_t t = std::time(nullptr);
+    auto s = fmt::format(L"The date is {:%Y-%m-%d}.", *std::localtime(&t));
+    ```
+
+    Thanks @DanielaE.
+
+-   Provided more wide string overloads
+    (https://github.com/fmtlib/fmt/pull/724). Thanks @DanielaE.
+
+-   Switched from a custom null-terminated string view class to
+    `string_view` in the format API and provided `fmt::string_view`
+    which implements a subset of `std::string_view` API for pre-C++17
+    systems.
+
+-   Added support for `std::experimental::string_view`
+    (https://github.com/fmtlib/fmt/pull/607):
+
+    ```c++
+    #include <fmt/core.h>
+    #include <experimental/string_view>
+
+    fmt::print("{}", std::experimental::string_view("foo"));
+    ```
+
+    Thanks @virgiliofornazin.
+
+-   Allowed mixing named and automatic arguments:
+
+    ```c++
+    fmt::format("{} {two}", 1, fmt::arg("two", 2));
+    ```
+
+-   Removed the write API in favor of the [format
+    API](https://fmt.dev/latest/api.html#format-api) with compile-time
+    handling of format strings.
+
+-   Disallowed formatting of multibyte strings into a wide character
+    target (https://github.com/fmtlib/fmt/pull/606).
+
+-   Improved documentation
+    (https://github.com/fmtlib/fmt/pull/515,
+    https://github.com/fmtlib/fmt/issues/614,
+    https://github.com/fmtlib/fmt/pull/617,
+    https://github.com/fmtlib/fmt/pull/661,
+    https://github.com/fmtlib/fmt/pull/680).
+    Thanks @ibell, @mihaitodor and @johnthagen.
+
+-   Implemented more efficient handling of large number of format
+    arguments.
+
+-   Introduced an inline namespace for symbol versioning.
+
+-   Added debug postfix `d` to the `fmt` library name
+    (https://github.com/fmtlib/fmt/issues/636).
+
+-   Removed unnecessary `fmt/` prefix in includes
+    (https://github.com/fmtlib/fmt/pull/397). Thanks @chronoxor.
+
+-   Moved `fmt/*.h` to `include/fmt/*.h` to prevent irrelevant files and
+    directories appearing on the include search paths when fmt is used
+    as a subproject and moved source files to the `src` directory.
+
+-   Added qmake project file `support/fmt.pro`
+    (https://github.com/fmtlib/fmt/pull/641). Thanks @cowo78.
+
+-   Added Gradle build file `support/build.gradle`
+    (https://github.com/fmtlib/fmt/pull/649). Thanks @luncliff.
+
+-   Removed `FMT_CPPFORMAT` CMake option.
+
+-   Fixed a name conflict with the macro `CHAR_WIDTH` in glibc
+    (https://github.com/fmtlib/fmt/pull/616). Thanks @aroig.
+
+-   Fixed handling of nested braces in `fmt::join`
+    (https://github.com/fmtlib/fmt/issues/638).
+
+-   Added `SOURCELINK_SUFFIX` for compatibility with Sphinx 1.5
+    (https://github.com/fmtlib/fmt/pull/497). Thanks @ginggs.
+
+-   Added a missing `inline` in the header-only mode
+    (https://github.com/fmtlib/fmt/pull/626). Thanks @aroig.
+
+-   Fixed various compiler warnings
+    (https://github.com/fmtlib/fmt/pull/640,
+    https://github.com/fmtlib/fmt/pull/656,
+    https://github.com/fmtlib/fmt/pull/679,
+    https://github.com/fmtlib/fmt/pull/681,
+    https://github.com/fmtlib/fmt/pull/705,
+    https://github.com/fmtlib/fmt/issues/715,
+    https://github.com/fmtlib/fmt/pull/717,
+    https://github.com/fmtlib/fmt/pull/720,
+    https://github.com/fmtlib/fmt/pull/723,
+    https://github.com/fmtlib/fmt/pull/726,
+    https://github.com/fmtlib/fmt/pull/730,
+    https://github.com/fmtlib/fmt/pull/739).
+    Thanks @peterbell10, @LarsGullik, @foonathan, @eliaskosunen,
+    @christianparpart, @DanielaE and @mwinterb.
+
+-   Worked around an MSVC bug and fixed several warnings
+    (https://github.com/fmtlib/fmt/pull/653). Thanks @alabuzhev.
+
+-   Worked around GCC bug 67371
+    (https://github.com/fmtlib/fmt/issues/682).
+
+-   Fixed compilation with `-fno-exceptions`
+    (https://github.com/fmtlib/fmt/pull/655). Thanks @chenxiaolong.
+
+-   Made `constexpr remove_prefix` gcc version check tighter
+    (https://github.com/fmtlib/fmt/issues/648).
+
+-   Renamed internal type enum constants to prevent collision with
+    poorly written C libraries
+    (https://github.com/fmtlib/fmt/issues/644).
+
+-   Added detection of `wostream operator<<`
+    (https://github.com/fmtlib/fmt/issues/650).
+
+-   Fixed compilation on OpenBSD
+    (https://github.com/fmtlib/fmt/pull/660). Thanks @hubslave.
+
+-   Fixed compilation on FreeBSD 12
+    (https://github.com/fmtlib/fmt/pull/732). Thanks @dankm.
+
+-   Fixed compilation when there is a mismatch between `-std` options
+    between the library and user code
+    (https://github.com/fmtlib/fmt/issues/664).
+
+-   Fixed compilation with GCC 7 and `-std=c++11`
+    (https://github.com/fmtlib/fmt/issues/734).
+
+-   Improved generated binary code on GCC 7 and older
+    (https://github.com/fmtlib/fmt/issues/668).
+
+-   Fixed handling of numeric alignment with no width
+    (https://github.com/fmtlib/fmt/issues/675).
+
+-   Fixed handling of empty strings in UTF8/16 converters
+    (https://github.com/fmtlib/fmt/pull/676). Thanks @vgalka-sl.
+
+-   Fixed formatting of an empty `string_view`
+    (https://github.com/fmtlib/fmt/issues/689).
+
+-   Fixed detection of `string_view` on libc++
+    (https://github.com/fmtlib/fmt/issues/686).
+
+-   Fixed DLL issues (https://github.com/fmtlib/fmt/pull/696).
+    Thanks @sebkoenig.
+
+-   Fixed compile checks for mixing narrow and wide strings
+    (https://github.com/fmtlib/fmt/issues/690).
+
+-   Disabled unsafe implicit conversion to `std::string`
+    (https://github.com/fmtlib/fmt/issues/729).
+
+-   Fixed handling of reused format specs (as in `fmt::join`) for
+    pointers (https://github.com/fmtlib/fmt/pull/725). Thanks @mwinterb.
+
+-   Fixed installation of `fmt/ranges.h`
+    (https://github.com/fmtlib/fmt/pull/738). Thanks @sv1990.
+
+# 4.1.0 - 2017-12-20
+
+-   Added `fmt::to_wstring()` in addition to `fmt::to_string()`
+    (https://github.com/fmtlib/fmt/pull/559). Thanks @alabuzhev.
+-   Added support for C++17 `std::string_view`
+    (https://github.com/fmtlib/fmt/pull/571 and
+    https://github.com/fmtlib/fmt/pull/578).
+    Thanks @thelostt and @mwinterb.
+-   Enabled stream exceptions to catch errors
+    (https://github.com/fmtlib/fmt/issues/581). Thanks @crusader-mike.
+-   Allowed formatting of class hierarchies with `fmt::format_arg()`
+    (https://github.com/fmtlib/fmt/pull/547). Thanks @rollbear.
+-   Removed limitations on character types
+    (https://github.com/fmtlib/fmt/pull/563). Thanks @Yelnats321.
+-   Conditionally enabled use of `std::allocator_traits`
+    (https://github.com/fmtlib/fmt/pull/583). Thanks @mwinterb.
+-   Added support for `const` variadic member function emulation with
+    `FMT_VARIADIC_CONST`
+    (https://github.com/fmtlib/fmt/pull/591). Thanks @ludekvodicka.
+-   Various bugfixes: bad overflow check, unsupported implicit type
+    conversion when determining formatting function, test segfaults
+    (https://github.com/fmtlib/fmt/issues/551), ill-formed
+    macros (https://github.com/fmtlib/fmt/pull/542) and
+    ambiguous overloads
+    (https://github.com/fmtlib/fmt/issues/580). Thanks @xylosper.
+-   Prevented warnings on MSVC
+    (https://github.com/fmtlib/fmt/pull/605,
+    https://github.com/fmtlib/fmt/pull/602, and
+    https://github.com/fmtlib/fmt/pull/545), clang
+    (https://github.com/fmtlib/fmt/pull/582), GCC
+    (https://github.com/fmtlib/fmt/issues/573), various
+    conversion warnings (https://github.com/fmtlib/fmt/pull/609,
+    https://github.com/fmtlib/fmt/pull/567,
+    https://github.com/fmtlib/fmt/pull/553 and
+    https://github.com/fmtlib/fmt/pull/553), and added
+    `override` and `[[noreturn]]`
+    (https://github.com/fmtlib/fmt/pull/549 and
+    https://github.com/fmtlib/fmt/issues/555).
+    Thanks @alabuzhev, @virgiliofornazin, @alexanderbock, @yumetodo, @VaderY,
+    @jpcima, @thelostt and @Manu343726.
+-   Improved CMake: Used `GNUInstallDirs` to set installation location
+    (https://github.com/fmtlib/fmt/pull/610) and fixed warnings
+    (https://github.com/fmtlib/fmt/pull/536 and
+    https://github.com/fmtlib/fmt/pull/556).
+    Thanks @mikecrowe, @evgen231 and @henryiii.
+
+# 4.0.0 - 2017-06-27
+
+-   Removed old compatibility headers `cppformat/*.h` and CMake options
+    (https://github.com/fmtlib/fmt/pull/527). Thanks @maddinat0r.
+
+-   Added `string.h` containing `fmt::to_string()` as alternative to
+    `std::to_string()` as well as other string writer functionality
+    (https://github.com/fmtlib/fmt/issues/326 and
+    https://github.com/fmtlib/fmt/pull/441):
+
+    ```c++
+    #include "fmt/string.h"
+
+    std::string answer = fmt::to_string(42);
+    ```
+
+    Thanks @glebov-andrey.
+
+-   Moved `fmt::printf()` to new `printf.h` header and allowed `%s` as
+    generic specifier (https://github.com/fmtlib/fmt/pull/453),
+    made `%.f` more conformant to regular `printf()`
+    (https://github.com/fmtlib/fmt/pull/490), added custom
+    writer support (https://github.com/fmtlib/fmt/issues/476)
+    and implemented missing custom argument formatting
+    (https://github.com/fmtlib/fmt/pull/339 and
+    https://github.com/fmtlib/fmt/pull/340):
+
+    ```c++
+    #include "fmt/printf.h"
+
+    // %s format specifier can be used with any argument type.
+    fmt::printf("%s", 42);
+    ```
+
+    Thanks @mojoBrendan, @manylegged and @spacemoose.
+    See also https://github.com/fmtlib/fmt/issues/360,
+    https://github.com/fmtlib/fmt/issues/335 and
+    https://github.com/fmtlib/fmt/issues/331.
+
+-   Added `container.h` containing a `BasicContainerWriter` to write to
+    containers like `std::vector`
+    (https://github.com/fmtlib/fmt/pull/450). Thanks @polyvertex.
+
+-   Added `fmt::join()` function that takes a range and formats its
+    elements separated by a given string
+    (https://github.com/fmtlib/fmt/pull/466):
+
+    ```c++
+    #include "fmt/format.h"
+
+    std::vector<double> v = {1.2, 3.4, 5.6};
+    // Prints "(+01.20, +03.40, +05.60)".
+    fmt::print("({:+06.2f})", fmt::join(v.begin(), v.end(), ", "));
+    ```
+
+    Thanks @olivier80.
+
+-   Added support for custom formatting specifications to simplify
+    customization of built-in formatting
+    (https://github.com/fmtlib/fmt/pull/444). Thanks @polyvertex.
+    See also https://github.com/fmtlib/fmt/issues/439.
+
+-   Added `fmt::format_system_error()` for error code formatting
+    (https://github.com/fmtlib/fmt/issues/323 and
+    https://github.com/fmtlib/fmt/pull/526). Thanks @maddinat0r.
+
+-   Added thread-safe `fmt::localtime()` and `fmt::gmtime()` as
+    replacement for the standard version to `time.h`
+    (https://github.com/fmtlib/fmt/pull/396). Thanks @codicodi.
+
+-   Internal improvements to `NamedArg` and `ArgLists`
+    (https://github.com/fmtlib/fmt/pull/389 and
+    https://github.com/fmtlib/fmt/pull/390). Thanks @chronoxor.
+
+-   Fixed crash due to bug in `FormatBuf`
+    (https://github.com/fmtlib/fmt/pull/493). Thanks @effzeh. See also
+    https://github.com/fmtlib/fmt/issues/480 and
+    https://github.com/fmtlib/fmt/issues/491.
+
+-   Fixed handling of wide strings in `fmt::StringWriter`.
+
+-   Improved compiler error messages
+    (https://github.com/fmtlib/fmt/issues/357).
+
+-   Fixed various warnings and issues with various compilers
+    (https://github.com/fmtlib/fmt/pull/494,
+    https://github.com/fmtlib/fmt/pull/499,
+    https://github.com/fmtlib/fmt/pull/483,
+    https://github.com/fmtlib/fmt/pull/485,
+    https://github.com/fmtlib/fmt/pull/482,
+    https://github.com/fmtlib/fmt/pull/475,
+    https://github.com/fmtlib/fmt/pull/473 and
+    https://github.com/fmtlib/fmt/pull/414).
+    Thanks @chronoxor, @zhaohuaxishi, @pkestene, @dschmidt and @0x414c.
+
+-   Improved CMake: targets are now namespaced
+    (https://github.com/fmtlib/fmt/pull/511 and
+    https://github.com/fmtlib/fmt/pull/513), supported
+    header-only `printf.h`
+    (https://github.com/fmtlib/fmt/pull/354), fixed issue with
+    minimal supported library subset
+    (https://github.com/fmtlib/fmt/issues/418,
+    https://github.com/fmtlib/fmt/pull/419 and
+    https://github.com/fmtlib/fmt/pull/420).
+    Thanks @bjoernthiel, @niosHD, @LogicalKnight and @alabuzhev.
+
+-   Improved documentation (https://github.com/fmtlib/fmt/pull/393).
+    Thanks @pwm1234.
+
+# 3.0.2 - 2017-06-14
+
+-   Added `FMT_VERSION` macro
+    (https://github.com/fmtlib/fmt/issues/411).
+-   Used `FMT_NULL` instead of literal `0`
+    (https://github.com/fmtlib/fmt/pull/409). Thanks @alabuzhev.
+-   Added extern templates for `format_float`
+    (https://github.com/fmtlib/fmt/issues/413).
+-   Fixed implicit conversion issue
+    (https://github.com/fmtlib/fmt/issues/507).
+-   Fixed signbit detection
+    (https://github.com/fmtlib/fmt/issues/423).
+-   Fixed naming collision
+    (https://github.com/fmtlib/fmt/issues/425).
+-   Fixed missing intrinsic for C++/CLI
+    (https://github.com/fmtlib/fmt/pull/457). Thanks @calumr.
+-   Fixed Android detection
+    (https://github.com/fmtlib/fmt/pull/458). Thanks @Gachapen.
+-   Use lean `windows.h` if not in header-only mode
+    (https://github.com/fmtlib/fmt/pull/503). Thanks @Quentin01.
+-   Fixed issue with CMake exporting C++11 flag
+    (https://github.com/fmtlib/fmt/pull/455). Thanks @EricWF.
+-   Fixed issue with nvcc and MSVC compiler bug and MinGW
+    (https://github.com/fmtlib/fmt/issues/505).
+-   Fixed DLL issues (https://github.com/fmtlib/fmt/pull/469 and
+    https://github.com/fmtlib/fmt/pull/502).
+    Thanks @richardeakin and @AndreasSchoenle.
+-   Fixed test compilation under FreeBSD
+    (https://github.com/fmtlib/fmt/issues/433).
+-   Fixed various warnings
+    (https://github.com/fmtlib/fmt/pull/403,
+    https://github.com/fmtlib/fmt/pull/410 and
+    https://github.com/fmtlib/fmt/pull/510).
+    Thanks @Lecetem, @chenhayat and @trozen.
+-   Worked around a broken `__builtin_clz` in clang with MS codegen
+    (https://github.com/fmtlib/fmt/issues/519).
+-   Removed redundant include
+    (https://github.com/fmtlib/fmt/issues/479).
+-   Fixed documentation issues.
+
+# 3.0.1 - 2016-11-01
+
+-   Fixed handling of thousands separator
+    (https://github.com/fmtlib/fmt/issues/353).
+-   Fixed handling of `unsigned char` strings
+    (https://github.com/fmtlib/fmt/issues/373).
+-   Corrected buffer growth when formatting time
+    (https://github.com/fmtlib/fmt/issues/367).
+-   Removed warnings under MSVC and clang
+    (https://github.com/fmtlib/fmt/issues/318,
+    https://github.com/fmtlib/fmt/issues/250, also merged
+    https://github.com/fmtlib/fmt/pull/385 and
+    https://github.com/fmtlib/fmt/pull/361).
+    Thanks @jcelerier and @nmoehrle.
+-   Fixed compilation issues under Android
+    (https://github.com/fmtlib/fmt/pull/327,
+    https://github.com/fmtlib/fmt/issues/345 and
+    https://github.com/fmtlib/fmt/pull/381), FreeBSD
+    (https://github.com/fmtlib/fmt/pull/358), Cygwin
+    (https://github.com/fmtlib/fmt/issues/388), MinGW
+    (https://github.com/fmtlib/fmt/issues/355) as well as other
+    issues (https://github.com/fmtlib/fmt/issues/350,
+    https://github.com/fmtlib/fmt/issues/355,
+    https://github.com/fmtlib/fmt/pull/348,
+    https://github.com/fmtlib/fmt/pull/402,
+    https://github.com/fmtlib/fmt/pull/405).
+    Thanks @dpantele, @hghwng, @arvedarved, @LogicalKnight and @JanHellwig.
+-   Fixed some documentation issues and extended specification
+    (https://github.com/fmtlib/fmt/issues/320,
+    https://github.com/fmtlib/fmt/pull/333,
+    https://github.com/fmtlib/fmt/issues/347,
+    https://github.com/fmtlib/fmt/pull/362). Thanks @smellman.
+
+# 3.0.0 - 2016-05-07
+
+-   The project has been renamed from C++ Format (cppformat) to fmt for
+    consistency with the used namespace and macro prefix
+    (https://github.com/fmtlib/fmt/issues/307). Library headers
+    are now located in the `fmt` directory:
+
+    ```c++
+    #include "fmt/format.h"
+    ```
+
+    Including `format.h` from the `cppformat` directory is deprecated
+    but works via a proxy header which will be removed in the next major
+    version.
+
+    The documentation is now available at <https://fmt.dev>.
+
+-   Added support for
+    [strftime](http://en.cppreference.com/w/cpp/chrono/c/strftime)-like
+    [date and time
+    formatting](https://fmt.dev/3.0.0/api.html#date-and-time-formatting)
+    (https://github.com/fmtlib/fmt/issues/283):
+
+    ```c++
+    #include "fmt/time.h"
+
+    std::time_t t = std::time(nullptr);
+    // Prints "The date is 2016-04-29." (with the current date)
+    fmt::print("The date is {:%Y-%m-%d}.", *std::localtime(&t));
+    ```
+
+-   `std::ostream` support including formatting of user-defined types
+    that provide overloaded `operator<<` has been moved to
+    `fmt/ostream.h`:
+
+    ```c++
+    #include "fmt/ostream.h"
+
+    class Date {
+      int year_, month_, day_;
+    public:
+      Date(int year, int month, int day) : year_(year), month_(month), day_(day) {}
+
+      friend std::ostream &operator<<(std::ostream &os, const Date &d) {
+        return os << d.year_ << '-' << d.month_ << '-' << d.day_;
+      }
+    };
+
+    std::string s = fmt::format("The date is {}", Date(2012, 12, 9));
+    // s == "The date is 2012-12-9"
+    ```
+
+-   Added support for [custom argument
+    formatters](https://fmt.dev/3.0.0/api.html#argument-formatters)
+    (https://github.com/fmtlib/fmt/issues/235).
+
+-   Added support for locale-specific integer formatting with the `n`
+    specifier (https://github.com/fmtlib/fmt/issues/305):
+
+    ```c++
+    std::setlocale(LC_ALL, "en_US.utf8");
+    fmt::print("cppformat: {:n}\n", 1234567); // prints 1,234,567
+    ```
+
+-   Sign is now preserved when formatting an integer with an incorrect
+    `printf` format specifier
+    (https://github.com/fmtlib/fmt/issues/265):
+
+    ```c++
+    fmt::printf("%lld", -42); // prints -42
+    ```
+
+    Note that it would be an undefined behavior in `std::printf`.
+
+-   Length modifiers such as `ll` are now optional in printf formatting
+    functions and the correct type is determined automatically
+    (https://github.com/fmtlib/fmt/issues/255):
+
+    ```c++
+    fmt::printf("%d", std::numeric_limits<long long>::max());
+    ```
+
+    Note that it would be an undefined behavior in `std::printf`.
+
+-   Added initial support for custom formatters
+    (https://github.com/fmtlib/fmt/issues/231).
+
+-   Fixed detection of user-defined literal support on Intel C++
+    compiler (https://github.com/fmtlib/fmt/issues/311,
+    https://github.com/fmtlib/fmt/pull/312).
+    Thanks @dean0x7d and @speth.
+
+-   Reduced compile time
+    (https://github.com/fmtlib/fmt/pull/243,
+    https://github.com/fmtlib/fmt/pull/249,
+    https://github.com/fmtlib/fmt/issues/317):
+
+    ![](https://cloud.githubusercontent.com/assets/4831417/11614060/b9e826d2-9c36-11e5-8666-d4131bf503ef.png)
+
+    ![](https://cloud.githubusercontent.com/assets/4831417/11614080/6ac903cc-9c37-11e5-8165-26df6efae364.png)
+
+    Thanks @dean0x7d.
+
+-   Compile test fixes (https://github.com/fmtlib/fmt/pull/313).
+    Thanks @dean0x7d.
+
+-   Documentation fixes (https://github.com/fmtlib/fmt/pull/239,
+    https://github.com/fmtlib/fmt/issues/248,
+    https://github.com/fmtlib/fmt/issues/252,
+    https://github.com/fmtlib/fmt/pull/258,
+    https://github.com/fmtlib/fmt/issues/260,
+    https://github.com/fmtlib/fmt/issues/301,
+    https://github.com/fmtlib/fmt/pull/309).
+    Thanks @ReadmeCritic @Gachapen and @jwilk.
+
+-   Fixed compiler and sanitizer warnings
+    (https://github.com/fmtlib/fmt/issues/244,
+    https://github.com/fmtlib/fmt/pull/256,
+    https://github.com/fmtlib/fmt/pull/259,
+    https://github.com/fmtlib/fmt/issues/263,
+    https://github.com/fmtlib/fmt/issues/274,
+    https://github.com/fmtlib/fmt/pull/277,
+    https://github.com/fmtlib/fmt/pull/286,
+    https://github.com/fmtlib/fmt/issues/291,
+    https://github.com/fmtlib/fmt/issues/296,
+    https://github.com/fmtlib/fmt/issues/308).
+    Thanks @mwinterb, @pweiskircher and @Naios.
+
+-   Improved compatibility with Windows Store apps
+    (https://github.com/fmtlib/fmt/issues/280,
+    https://github.com/fmtlib/fmt/pull/285) Thanks @mwinterb.
+
+-   Added tests of compatibility with older C++ standards
+    (https://github.com/fmtlib/fmt/pull/273). Thanks @niosHD.
+
+-   Fixed Android build
+    (https://github.com/fmtlib/fmt/pull/271). Thanks @newnon.
+
+-   Changed `ArgMap` to be backed by a vector instead of a map.
+    (https://github.com/fmtlib/fmt/issues/261,
+    https://github.com/fmtlib/fmt/pull/262). Thanks @mwinterb.
+
+-   Added `fprintf` overload that writes to a `std::ostream`
+    (https://github.com/fmtlib/fmt/pull/251).
+    Thanks @nickhutchinson.
+
+-   Export symbols when building a Windows DLL
+    (https://github.com/fmtlib/fmt/pull/245).
+    Thanks @macdems.
+
+-   Fixed compilation on Cygwin
+    (https://github.com/fmtlib/fmt/issues/304).
+
+-   Implemented a workaround for a bug in Apple LLVM version 4.2 of
+    clang (https://github.com/fmtlib/fmt/issues/276).
+
+-   Implemented a workaround for Google Test bug
+    https://github.com/google/googletest/issues/705 on gcc 6
+    (https://github.com/fmtlib/fmt/issues/268). Thanks @octoploid.
+
+-   Removed Biicode support because the latter has been discontinued.
+
+# 2.1.1 - 2016-04-11
+
+-   The install location for generated CMake files is now configurable
+    via the `FMT_CMAKE_DIR` CMake variable
+    (https://github.com/fmtlib/fmt/pull/299). Thanks @niosHD.
+-   Documentation fixes
+    (https://github.com/fmtlib/fmt/issues/252).
+
+# 2.1.0 - 2016-03-21
+
+-   Project layout and build system improvements
+    (https://github.com/fmtlib/fmt/pull/267):
+
+    -   The code have been moved to the `cppformat` directory. Including
+        `format.h` from the top-level directory is deprecated but works
+        via a proxy header which will be removed in the next major
+        version.
+    -   C++ Format CMake targets now have proper interface definitions.
+    -   Installed version of the library now supports the header-only
+        configuration.
+    -   Targets `doc`, `install`, and `test` are now disabled if C++
+        Format is included as a CMake subproject. They can be enabled by
+        setting `FMT_DOC`, `FMT_INSTALL`, and `FMT_TEST` in the parent
+        project.
+
+    Thanks @niosHD.
+
+# 2.0.1 - 2016-03-13
+
+-   Improved CMake find and package support
+    (https://github.com/fmtlib/fmt/issues/264). Thanks @niosHD.
+-   Fix compile error with Android NDK and mingw32
+    (https://github.com/fmtlib/fmt/issues/241). Thanks @Gachapen.
+-   Documentation fixes
+    (https://github.com/fmtlib/fmt/issues/248,
+    https://github.com/fmtlib/fmt/issues/260).
+
+# 2.0.0 - 2015-12-01
+
+## General
+
+-   \[Breaking\] Named arguments
+    (https://github.com/fmtlib/fmt/pull/169,
+    https://github.com/fmtlib/fmt/pull/173,
+    https://github.com/fmtlib/fmt/pull/174):
+
+    ```c++
+    fmt::print("The answer is {answer}.", fmt::arg("answer", 42));
+    ```
+
+    Thanks @jamboree.
+
+-   \[Experimental\] User-defined literals for format and named
+    arguments (https://github.com/fmtlib/fmt/pull/204,
+    https://github.com/fmtlib/fmt/pull/206,
+    https://github.com/fmtlib/fmt/pull/207):
+
+    ```c++
+    using namespace fmt::literals;
+    fmt::print("The answer is {answer}.", "answer"_a=42);
+    ```
+
+    Thanks @dean0x7d.
+
+-   \[Breaking\] Formatting of more than 16 arguments is now supported
+    when using variadic templates
+    (https://github.com/fmtlib/fmt/issues/141). Thanks @Shauren.
+
+-   Runtime width specification
+    (https://github.com/fmtlib/fmt/pull/168):
+
+    ```c++
+    fmt::format("{0:{1}}", 42, 5); // gives "   42"
+    ```
+
+    Thanks @jamboree.
+
+-   \[Breaking\] Enums are now formatted with an overloaded
+    `std::ostream` insertion operator (`operator<<`) if available
+    (https://github.com/fmtlib/fmt/issues/232).
+
+-   \[Breaking\] Changed default `bool` format to textual, \"true\" or
+    \"false\" (https://github.com/fmtlib/fmt/issues/170):
+
+    ```c++
+    fmt::print("{}", true); // prints "true"
+    ```
+
+    To print `bool` as a number use numeric format specifier such as
+    `d`:
+
+    ```c++
+    fmt::print("{:d}", true); // prints "1"
+    ```
+
+-   `fmt::printf` and `fmt::sprintf` now support formatting of `bool`
+    with the `%s` specifier giving textual output, \"true\" or \"false\"
+    (https://github.com/fmtlib/fmt/pull/223):
+
+    ```c++
+    fmt::printf("%s", true); // prints "true"
+    ```
+
+    Thanks @LarsGullik.
+
+-   \[Breaking\] `signed char` and `unsigned char` are now formatted as
+    integers by default
+    (https://github.com/fmtlib/fmt/pull/217).
+
+-   \[Breaking\] Pointers to C strings can now be formatted with the `p`
+    specifier (https://github.com/fmtlib/fmt/pull/223):
+
+    ```c++
+    fmt::print("{:p}", "test"); // prints pointer value
+    ```
+
+    Thanks @LarsGullik.
+
+-   \[Breaking\] `fmt::printf` and `fmt::sprintf` now print null
+    pointers as `(nil)` and null strings as `(null)` for consistency
+    with glibc (https://github.com/fmtlib/fmt/pull/226).
+    Thanks @LarsGullik.
+
+-   \[Breaking\] `fmt::(s)printf` now supports formatting of objects of
+    user-defined types that provide an overloaded `std::ostream`
+    insertion operator (`operator<<`)
+    (https://github.com/fmtlib/fmt/issues/201):
+
+    ```c++
+    fmt::printf("The date is %s", Date(2012, 12, 9));
+    ```
+
+-   \[Breaking\] The `Buffer` template is now part of the public API and
+    can be used to implement custom memory buffers
+    (https://github.com/fmtlib/fmt/issues/140). Thanks @polyvertex.
+
+-   \[Breaking\] Improved compatibility between `BasicStringRef` and
+    [std::experimental::basic_string_view](
+    http://en.cppreference.com/w/cpp/experimental/basic_string_view)
+    (https://github.com/fmtlib/fmt/issues/100,
+    https://github.com/fmtlib/fmt/issues/159,
+    https://github.com/fmtlib/fmt/issues/183):
+
+    -   Comparison operators now compare string content, not pointers
+    -   `BasicStringRef::c_str` replaced by `BasicStringRef::data`
+    -   `BasicStringRef` is no longer assumed to be null-terminated
+
+    References to null-terminated strings are now represented by a new
+    class, `BasicCStringRef`.
+
+-   Dependency on pthreads introduced by Google Test is now optional
+    (https://github.com/fmtlib/fmt/issues/185).
+
+-   New CMake options `FMT_DOC`, `FMT_INSTALL` and `FMT_TEST` to control
+    generation of `doc`, `install` and `test` targets respectively, on
+    by default (https://github.com/fmtlib/fmt/issues/197,
+    https://github.com/fmtlib/fmt/issues/198,
+    https://github.com/fmtlib/fmt/issues/200). Thanks @maddinat0r.
+
+-   `noexcept` is now used when compiling with MSVC2015
+    (https://github.com/fmtlib/fmt/pull/215). Thanks @dmkrepo.
+
+-   Added an option to disable use of `windows.h` when
+    `FMT_USE_WINDOWS_H` is defined as 0 before including `format.h`
+    (https://github.com/fmtlib/fmt/issues/171). Thanks @alfps.
+
+-   \[Breaking\] `windows.h` is now included with `NOMINMAX` unless
+    `FMT_WIN_MINMAX` is defined. This is done to prevent breaking code
+    using `std::min` and `std::max` and only affects the header-only
+    configuration (https://github.com/fmtlib/fmt/issues/152,
+    https://github.com/fmtlib/fmt/pull/153,
+    https://github.com/fmtlib/fmt/pull/154). Thanks @DevO2012.
+
+-   Improved support for custom character types
+    (https://github.com/fmtlib/fmt/issues/171). Thanks @alfps.
+
+-   Added an option to disable use of IOStreams when `FMT_USE_IOSTREAMS`
+    is defined as 0 before including `format.h`
+    (https://github.com/fmtlib/fmt/issues/205,
+    https://github.com/fmtlib/fmt/pull/208). Thanks @JodiTheTigger.
+
+-   Improved detection of `isnan`, `isinf` and `signbit`.
+
+## Optimization
+
+-   Made formatting of user-defined types more efficient with a custom
+    stream buffer (https://github.com/fmtlib/fmt/issues/92,
+    https://github.com/fmtlib/fmt/pull/230). Thanks @NotImplemented.
+-   Further improved performance of `fmt::Writer` on integer formatting
+    and fixed a minor regression. Now it is \~7% faster than
+    `karma::generate` on Karma\'s benchmark
+    (https://github.com/fmtlib/fmt/issues/186).
+-   \[Breaking\] Reduced [compiled code
+    size](https://github.com/fmtlib/fmt#compile-time-and-code-bloat)
+    (https://github.com/fmtlib/fmt/issues/143,
+    https://github.com/fmtlib/fmt/pull/149).
+
+## Distribution
+
+-   \[Breaking\] Headers are now installed in
+    `${CMAKE_INSTALL_PREFIX}/include/cppformat`
+    (https://github.com/fmtlib/fmt/issues/178). Thanks @jackyf.
+
+-   \[Breaking\] Changed the library name from `format` to `cppformat`
+    for consistency with the project name and to avoid potential
+    conflicts (https://github.com/fmtlib/fmt/issues/178).
+    Thanks @jackyf.
+
+-   C++ Format is now available in [Debian](https://www.debian.org/)
+    GNU/Linux
+    ([stretch](https://packages.debian.org/source/stretch/cppformat),
+    [sid](https://packages.debian.org/source/sid/cppformat)) and derived
+    distributions such as
+    [Ubuntu](https://launchpad.net/ubuntu/+source/cppformat) 15.10 and
+    later (https://github.com/fmtlib/fmt/issues/155):
+
+        $ sudo apt-get install libcppformat1-dev
+
+    Thanks @jackyf.
+
+-   [Packages for Fedora and
+    RHEL](https://admin.fedoraproject.org/pkgdb/package/cppformat/) are
+    now available. Thanks Dave Johansen.
+
+-   C++ Format can now be installed via [Homebrew](http://brew.sh/) on
+    OS X (https://github.com/fmtlib/fmt/issues/157):
+
+        $ brew install cppformat
+
+    Thanks @ortho and Anatoliy Bulukin.
+
+## Documentation
+
+-   Migrated from ReadTheDocs to GitHub Pages for better responsiveness
+    and reliability (https://github.com/fmtlib/fmt/issues/128).
+    New documentation address is <http://cppformat.github.io/>.
+-   Added [Building thedocumentation](
+    https://fmt.dev/2.0.0/usage.html#building-the-documentation)
+    section to the documentation.
+-   Documentation build script is now compatible with Python 3 and newer
+    pip versions. (https://github.com/fmtlib/fmt/pull/189,
+    https://github.com/fmtlib/fmt/issues/209).
+    Thanks @JodiTheTigger and @xentec.
+-   Documentation fixes and improvements
+    (https://github.com/fmtlib/fmt/issues/36,
+    https://github.com/fmtlib/fmt/issues/75,
+    https://github.com/fmtlib/fmt/issues/125,
+    https://github.com/fmtlib/fmt/pull/160,
+    https://github.com/fmtlib/fmt/pull/161,
+    https://github.com/fmtlib/fmt/issues/162,
+    https://github.com/fmtlib/fmt/issues/165,
+    https://github.com/fmtlib/fmt/issues/210). 
+    Thanks @syohex.
+-   Fixed out-of-tree documentation build
+    (https://github.com/fmtlib/fmt/issues/177). Thanks @jackyf.
+
+## Fixes
+
+-   Fixed `initializer_list` detection
+    (https://github.com/fmtlib/fmt/issues/136). Thanks @Gachapen.
+
+-   \[Breaking\] Fixed formatting of enums with numeric format
+    specifiers in `fmt::(s)printf`
+    (https://github.com/fmtlib/fmt/issues/131,
+    https://github.com/fmtlib/fmt/issues/139):
+
+    ```c++
+    enum { ANSWER = 42 };
+    fmt::printf("%d", ANSWER);
+    ```
+
+    Thanks @Naios.
+
+-   Improved compatibility with old versions of MinGW
+    (https://github.com/fmtlib/fmt/issues/129,
+    https://github.com/fmtlib/fmt/pull/130,
+    https://github.com/fmtlib/fmt/issues/132). Thanks @cstamford.
+
+-   Fixed a compile error on MSVC with disabled exceptions
+    (https://github.com/fmtlib/fmt/issues/144).
+
+-   Added a workaround for broken implementation of variadic templates
+    in MSVC2012 (https://github.com/fmtlib/fmt/issues/148).
+
+-   Placed the anonymous namespace within `fmt` namespace for the
+    header-only configuration (https://github.com/fmtlib/fmt/issues/171).
+    Thanks @alfps.
+
+-   Fixed issues reported by Coverity Scan
+    (https://github.com/fmtlib/fmt/issues/187,
+    https://github.com/fmtlib/fmt/issues/192).
+
+-   Implemented a workaround for a name lookup bug in MSVC2010
+    (https://github.com/fmtlib/fmt/issues/188).
+
+-   Fixed compiler warnings
+    (https://github.com/fmtlib/fmt/issues/95,
+    https://github.com/fmtlib/fmt/issues/96,
+    https://github.com/fmtlib/fmt/pull/114,
+    https://github.com/fmtlib/fmt/issues/135,
+    https://github.com/fmtlib/fmt/issues/142,
+    https://github.com/fmtlib/fmt/issues/145,
+    https://github.com/fmtlib/fmt/issues/146,
+    https://github.com/fmtlib/fmt/issues/158,
+    https://github.com/fmtlib/fmt/issues/163,
+    https://github.com/fmtlib/fmt/issues/175,
+    https://github.com/fmtlib/fmt/issues/190,
+    https://github.com/fmtlib/fmt/pull/191,
+    https://github.com/fmtlib/fmt/issues/194,
+    https://github.com/fmtlib/fmt/pull/196,
+    https://github.com/fmtlib/fmt/issues/216,
+    https://github.com/fmtlib/fmt/pull/218,
+    https://github.com/fmtlib/fmt/pull/220,
+    https://github.com/fmtlib/fmt/pull/229,
+    https://github.com/fmtlib/fmt/issues/233,
+    https://github.com/fmtlib/fmt/issues/234,
+    https://github.com/fmtlib/fmt/pull/236,
+    https://github.com/fmtlib/fmt/issues/281,
+    https://github.com/fmtlib/fmt/issues/289).
+    Thanks @seanmiddleditch, @dixlorenz, @CarterLi, @Naios, @fmatthew5876,
+    @LevskiWeng, @rpopescu, @gabime, @cubicool, @jkflying, @LogicalKnight,
+    @inguin and @Jopie64.
+
+-   Fixed portability issues (mostly causing test failures) on ARM,
+    ppc64, ppc64le, s390x and SunOS 5.11 i386
+    (https://github.com/fmtlib/fmt/issues/138,
+    https://github.com/fmtlib/fmt/issues/179,
+    https://github.com/fmtlib/fmt/issues/180,
+    https://github.com/fmtlib/fmt/issues/202,
+    https://github.com/fmtlib/fmt/issues/225, [Red Hat Bugzilla
+    Bug 1260297](https://bugzilla.redhat.com/show_bug.cgi?id=1260297)).
+    Thanks @Naios, @jackyf and Dave Johansen.
+
+-   Fixed a name conflict with macro `free` defined in `crtdbg.h` when
+    `_CRTDBG_MAP_ALLOC` is set (https://github.com/fmtlib/fmt/issues/211).
+
+-   Fixed shared library build on OS X
+    (https://github.com/fmtlib/fmt/pull/212). Thanks @dean0x7d.
+
+-   Fixed an overload conflict on MSVC when `/Zc:wchar_t-` option is
+    specified (https://github.com/fmtlib/fmt/pull/214).
+    Thanks @slavanap.
+
+-   Improved compatibility with MSVC 2008
+    (https://github.com/fmtlib/fmt/pull/236). Thanks @Jopie64.
+
+-   Improved compatibility with bcc32
+    (https://github.com/fmtlib/fmt/issues/227).
+
+-   Fixed `static_assert` detection on Clang
+    (https://github.com/fmtlib/fmt/pull/228). Thanks @dean0x7d.
+
+# 1.1.0 - 2015-03-06
+
+-   Added `BasicArrayWriter`, a class template that provides operations
+    for formatting and writing data into a fixed-size array
+    (https://github.com/fmtlib/fmt/issues/105 and
+    https://github.com/fmtlib/fmt/issues/122):
+
+    ```c++
+    char buffer[100];
+    fmt::ArrayWriter w(buffer);
+    w.write("The answer is {}", 42);
+    ```
+
+-   Added [0 A.D.](http://play0ad.com/) and [PenUltima Online
+    (POL)](http://www.polserver.com/) to the list of notable projects
+    using C++ Format.
+
+-   C++ Format now uses MSVC intrinsics for better formatting performance
+    (https://github.com/fmtlib/fmt/pull/115,
+    https://github.com/fmtlib/fmt/pull/116,
+    https://github.com/fmtlib/fmt/pull/118 and
+    https://github.com/fmtlib/fmt/pull/121). Previously these
+    optimizations where only used on GCC and Clang.
+    Thanks @CarterLi and @objectx.
+
+-   CMake install target
+    (https://github.com/fmtlib/fmt/pull/119). Thanks @TrentHouliston.
+
+    You can now install C++ Format with `make install` command.
+
+-   Improved [Biicode](http://www.biicode.com/) support
+    (https://github.com/fmtlib/fmt/pull/98 and
+    https://github.com/fmtlib/fmt/pull/104).
+    Thanks @MariadeAnton and @franramirez688.
+
+-   Improved support for building with [Android NDK](
+    https://developer.android.com/tools/sdk/ndk/index.html)
+    (https://github.com/fmtlib/fmt/pull/107). Thanks @newnon.
+
+    The [android-ndk-example](https://github.com/fmtlib/android-ndk-example)
+    repository provides and example of using C++ Format with Android NDK:
+
+    ![](https://raw.githubusercontent.com/fmtlib/android-ndk-example/master/screenshot.png)
+
+-   Improved documentation of `SystemError` and `WindowsError`
+    (https://github.com/fmtlib/fmt/issues/54).
+
+-   Various code improvements
+    (https://github.com/fmtlib/fmt/pull/110,
+    https://github.com/fmtlib/fmt/pull/111
+    https://github.com/fmtlib/fmt/pull/112). Thanks @CarterLi.
+
+-   Improved compile-time errors when formatting wide into narrow
+    strings (https://github.com/fmtlib/fmt/issues/117).
+
+-   Fixed `BasicWriter::write` without formatting arguments when C++11
+    support is disabled
+    (https://github.com/fmtlib/fmt/issues/109).
+
+-   Fixed header-only build on OS X with GCC 4.9
+    (https://github.com/fmtlib/fmt/issues/124).
+
+-   Fixed packaging issues (https://github.com/fmtlib/fmt/issues/94).
+
+-   Added [changelog](https://github.com/fmtlib/fmt/blob/master/ChangeLog.md)
+    (https://github.com/fmtlib/fmt/issues/103).
+
+# 1.0.0 - 2015-02-05
+
+-   Add support for a header-only configuration when `FMT_HEADER_ONLY`
+    is defined before including `format.h`:
+
+    ```c++
+    #define FMT_HEADER_ONLY
+    #include "format.h"
+    ```
+
+-   Compute string length in the constructor of `BasicStringRef` instead
+    of the `size` method
+    (https://github.com/fmtlib/fmt/issues/79). This eliminates
+    size computation for string literals on reasonable optimizing
+    compilers.
+
+-   Fix formatting of types with overloaded `operator <<` for
+    `std::wostream` (https://github.com/fmtlib/fmt/issues/86):
+
+    ```c++
+    fmt::format(L"The date is {0}", Date(2012, 12, 9));
+    ```
+
+-   Fix linkage of tests on Arch Linux
+    (https://github.com/fmtlib/fmt/issues/89).
+
+-   Allow precision specifier for non-float arguments
+    (https://github.com/fmtlib/fmt/issues/90):
+
+    ```c++
+    fmt::print("{:.3}\n", "Carpet"); // prints "Car"
+    ```
+
+-   Fix build on Android NDK (https://github.com/fmtlib/fmt/issues/93).
+
+-   Improvements to documentation build procedure.
+
+-   Remove `FMT_SHARED` CMake variable in favor of standard [BUILD_SHARED_LIBS](
+    http://www.cmake.org/cmake/help/v3.0/variable/BUILD_SHARED_LIBS.html).
+
+-   Fix error handling in `fmt::fprintf`.
+
+-   Fix a number of warnings.
+
+# 0.12.0 - 2014-10-25
+
+-   \[Breaking\] Improved separation between formatting and buffer
+    management. `Writer` is now a base class that cannot be instantiated
+    directly. The new `MemoryWriter` class implements the default buffer
+    management with small allocations done on stack. So `fmt::Writer`
+    should be replaced with `fmt::MemoryWriter` in variable
+    declarations.
+
+    Old code:
+
+    ```c++
+    fmt::Writer w;
+    ```
+
+    New code:
+
+    ```c++
+    fmt::MemoryWriter w;
+    ```
+
+    If you pass `fmt::Writer` by reference, you can continue to do so:
+
+    ```c++
+    void f(fmt::Writer &w);
+    ```
+
+    This doesn\'t affect the formatting API.
+
+-   Support for custom memory allocators
+    (https://github.com/fmtlib/fmt/issues/69)
+
+-   Formatting functions now accept [signed char]{.title-ref} and
+    [unsigned char]{.title-ref} strings as arguments
+    (https://github.com/fmtlib/fmt/issues/73):
+
+    ```c++
+    auto s = format("GLSL version: {}", glGetString(GL_VERSION));
+    ```
+
+-   Reduced code bloat. According to the new [benchmark
+    results](https://github.com/fmtlib/fmt#compile-time-and-code-bloat),
+    cppformat is close to `printf` and by the order of magnitude better
+    than Boost Format in terms of compiled code size.
+
+-   Improved appearance of the documentation on mobile by using the
+    [Sphinx Bootstrap
+    theme](http://ryan-roemer.github.io/sphinx-bootstrap-theme/):
+
+    | Old | New |
+    | --- | --- |
+    | ![](https://cloud.githubusercontent.com/assets/576385/4792130/cd256436-5de3-11e4-9a62-c077d0c2b003.png) | ![](https://cloud.githubusercontent.com/assets/576385/4792131/cd29896c-5de3-11e4-8f59-cac952942bf0.png) |
+
+# 0.11.0 - 2014-08-21
+
+-   Safe printf implementation with a POSIX extension for positional
+    arguments:
+
+    ```c++
+    fmt::printf("Elapsed time: %.2f seconds", 1.23);
+    fmt::printf("%1$s, %3$d %2$s", weekday, month, day);
+    ```
+
+-   Arguments of `char` type can now be formatted as integers (Issue
+    https://github.com/fmtlib/fmt/issues/55):
+
+    ```c++
+    fmt::format("0x{0:02X}", 'a');
+    ```
+
+-   Deprecated parts of the API removed.
+
+-   The library is now built and tested on MinGW with Appveyor in
+    addition to existing test platforms Linux/GCC, OS X/Clang,
+    Windows/MSVC.
+
+# 0.10.0 - 2014-07-01
+
+**Improved API**
+
+-   All formatting methods are now implemented as variadic functions
+    instead of using `operator<<` for feeding arbitrary arguments into a
+    temporary formatter object. This works both with C++11 where
+    variadic templates are used and with older standards where variadic
+    functions are emulated by providing lightweight wrapper functions
+    defined with the `FMT_VARIADIC` macro. You can use this macro for
+    defining your own portable variadic functions:
+
+    ```c++
+    void report_error(const char *format, const fmt::ArgList &args) {
+      fmt::print("Error: {}");
+      fmt::print(format, args);
+    }
+    FMT_VARIADIC(void, report_error, const char *)
+
+    report_error("file not found: {}", path);
+    ```
+
+    Apart from a more natural syntax, this also improves performance as
+    there is no need to construct temporary formatter objects and
+    control arguments\' lifetimes. Because the wrapper functions are
+    very lightweight, this doesn\'t cause code bloat even in pre-C++11
+    mode.
+
+-   Simplified common case of formatting an `std::string`. Now it
+    requires a single function call:
+
+    ```c++
+    std::string s = format("The answer is {}.", 42);
+    ```
+
+    Previously it required 2 function calls:
+
+    ```c++
+    std::string s = str(Format("The answer is {}.") << 42);
+    ```
+
+    Instead of unsafe `c_str` function, `fmt::Writer` should be used
+    directly to bypass creation of `std::string`:
+
+    ```c++
+    fmt::Writer w;
+    w.write("The answer is {}.", 42);
+    w.c_str();  // returns a C string
+    ```
+
+    This doesn\'t do dynamic memory allocation for small strings and is
+    less error prone as the lifetime of the string is the same as for
+    `std::string::c_str` which is well understood (hopefully).
+
+-   Improved consistency in naming functions that are a part of the
+    public API. Now all public functions are lowercase following the
+    standard library conventions. Previously it was a combination of
+    lowercase and CapitalizedWords. Issue
+    https://github.com/fmtlib/fmt/issues/50.
+
+-   Old functions are marked as deprecated and will be removed in the
+    next release.
+
+**Other Changes**
+
+-   Experimental support for printf format specifications (work in
+    progress):
+
+    ```c++
+    fmt::printf("The answer is %d.", 42);
+    std::string s = fmt::sprintf("Look, a %s!", "string");
+    ```
+
+-   Support for hexadecimal floating point format specifiers `a` and
+    `A`:
+
+    ```c++
+    print("{:a}", -42.0); // Prints -0x1.5p+5
+    print("{:A}", -42.0); // Prints -0X1.5P+5
+    ```
+
+-   CMake option `FMT_SHARED` that specifies whether to build format as
+    a shared library (off by default).
+
+# 0.9.0 - 2014-05-13
+
+-   More efficient implementation of variadic formatting functions.
+
+-   `Writer::Format` now has a variadic overload:
+
+    ```c++
+    Writer out;
+    out.Format("Look, I'm {}!", "variadic");
+    ```
+
+-   For efficiency and consistency with other overloads, variadic
+    overload of the `Format` function now returns `Writer` instead of
+    `std::string`. Use the `str` function to convert it to
+    `std::string`:
+
+    ```c++
+    std::string s = str(Format("Look, I'm {}!", "variadic"));
+    ```
+
+-   Replaced formatter actions with output sinks: `NoAction` -\>
+    `NullSink`, `Write` -\> `FileSink`, `ColorWriter` -\>
+    `ANSITerminalSink`. This improves naming consistency and shouldn\'t
+    affect client code unless these classes are used directly which
+    should be rarely needed.
+
+-   Added `ThrowSystemError` function that formats a message and throws
+    `SystemError` containing the formatted message and system-specific
+    error description. For example, the following code
+
+    ```c++
+    FILE *f = fopen(filename, "r");
+    if (!f)
+      ThrowSystemError(errno, "Failed to open file '{}'") << filename;
+    ```
+
+    will throw `SystemError` exception with description \"Failed to open
+    file \'\<filename\>\': No such file or directory\" if file doesn\'t
+    exist.
+
+-   Support for AppVeyor continuous integration platform.
+
+-   `Format` now throws `SystemError` in case of I/O errors.
+
+-   Improve test infrastructure. Print functions are now tested by
+    redirecting the output to a pipe.
+
+# 0.8.0 - 2014-04-14
+
+-   Initial release
diff --git a/dep/fmt/LICENSE b/dep/fmt/LICENSE
new file mode 100644
index 00000000..1cd1ef92
--- /dev/null
+++ b/dep/fmt/LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2012 - present, Victor Zverovich and {fmt} contributors
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+--- Optional exception to the license ---
+
+As an exception, if, as a result of your compiling your source code, portions
+of this Software are embedded into a machine-executable object form of such
+source code, you may redistribute such embedded portions in such object form
+without including the above copyright and permission notices.
diff --git a/dep/fmt/README.md b/dep/fmt/README.md
new file mode 100644
index 00000000..dcfb16ec
--- /dev/null
+++ b/dep/fmt/README.md
@@ -0,0 +1,490 @@
+<img src="https://user-images.githubusercontent.com/576385/156254208-f5b743a9-88cf-439d-b0c0-923d53e8d551.png" alt="{fmt}" width="25%"/>
+
+[![image](https://github.com/fmtlib/fmt/workflows/linux/badge.svg)](https://github.com/fmtlib/fmt/actions?query=workflow%3Alinux)
+[![image](https://github.com/fmtlib/fmt/workflows/macos/badge.svg)](https://github.com/fmtlib/fmt/actions?query=workflow%3Amacos)
+[![image](https://github.com/fmtlib/fmt/workflows/windows/badge.svg)](https://github.com/fmtlib/fmt/actions?query=workflow%3Awindows)
+[![fmt is continuously fuzzed at oss-fuzz](https://oss-fuzz-build-logs.storage.googleapis.com/badges/fmt.svg)](https://bugs.chromium.org/p/oss-fuzz/issues/list?\%0Acolspec=ID%20Type%20Component%20Status%20Proj%20Reported%20Owner%20\%0ASummary&q=proj%3Dfmt&can=1)
+[![Ask questions at StackOverflow with the tag fmt](https://img.shields.io/badge/stackoverflow-fmt-blue.svg)](https://stackoverflow.com/questions/tagged/fmt)
+[![image](https://api.securityscorecards.dev/projects/github.com/fmtlib/fmt/badge)](https://securityscorecards.dev/viewer/?uri=github.com/fmtlib/fmt)
+
+**{fmt}** is an open-source formatting library providing a fast and safe
+alternative to C stdio and C++ iostreams.
+
+If you like this project, please consider donating to one of the funds
+that help victims of the war in Ukraine: <https://www.stopputin.net/>.
+
+[Documentation](https://fmt.dev)
+
+[Cheat Sheets](https://hackingcpp.com/cpp/libs/fmt.html)
+
+Q&A: ask questions on [StackOverflow with the tag
+fmt](https://stackoverflow.com/questions/tagged/fmt).
+
+Try {fmt} in [Compiler Explorer](https://godbolt.org/z/Eq5763).
+
+# Features
+
+- Simple [format API](https://fmt.dev/latest/api.html) with positional
+  arguments for localization
+- Implementation of [C++20
+  std::format](https://en.cppreference.com/w/cpp/utility/format) and
+  [C++23 std::print](https://en.cppreference.com/w/cpp/io/print)
+- [Format string syntax](https://fmt.dev/latest/syntax.html) similar
+  to Python\'s
+  [format](https://docs.python.org/3/library/stdtypes.html#str.format)
+- Fast IEEE 754 floating-point formatter with correct rounding,
+  shortness and round-trip guarantees using the
+  [Dragonbox](https://github.com/jk-jeon/dragonbox) algorithm
+- Portable Unicode support
+- Safe [printf
+  implementation](https://fmt.dev/latest/api.html#printf-formatting)
+  including the POSIX extension for positional arguments
+- Extensibility: [support for user-defined
+  types](https://fmt.dev/latest/api.html#formatting-user-defined-types)
+- High performance: faster than common standard library
+  implementations of `(s)printf`, iostreams, `to_string` and
+  `to_chars`, see [Speed tests](#speed-tests) and [Converting a
+  hundred million integers to strings per
+  second](http://www.zverovich.net/2020/06/13/fast-int-to-string-revisited.html)
+- Small code size both in terms of source code with the minimum
+  configuration consisting of just three files, `core.h`, `format.h`
+  and `format-inl.h`, and compiled code; see [Compile time and code
+  bloat](#compile-time-and-code-bloat)
+- Reliability: the library has an extensive set of
+  [tests](https://github.com/fmtlib/fmt/tree/master/test) and is
+  [continuously fuzzed](https://bugs.chromium.org/p/oss-fuzz/issues/list?colspec=ID%20Type%20Component%20Status%20Proj%20Reported%20Owner%20Summary&q=proj%3Dfmt&can=1)
+- Safety: the library is fully type-safe, errors in format strings can
+  be reported at compile time, automatic memory management prevents
+  buffer overflow errors
+- Ease of use: small self-contained code base, no external
+  dependencies, permissive MIT
+  [license](https://github.com/fmtlib/fmt/blob/master/LICENSE.rst)
+- [Portability](https://fmt.dev/latest/index.html#portability) with
+  consistent output across platforms and support for older compilers
+- Clean warning-free codebase even on high warning levels such as
+  `-Wall -Wextra -pedantic`
+- Locale independence by default
+- Optional header-only configuration enabled with the
+  `FMT_HEADER_ONLY` macro
+
+See the [documentation](https://fmt.dev) for more details.
+
+# Examples
+
+**Print to stdout** ([run](https://godbolt.org/z/Tevcjh))
+
+``` c++
+#include <fmt/core.h>
+
+int main() {
+  fmt::print("Hello, world!\n");
+}
+```
+
+**Format a string** ([run](https://godbolt.org/z/oK8h33))
+
+``` c++
+std::string s = fmt::format("The answer is {}.", 42);
+// s == "The answer is 42."
+```
+
+**Format a string using positional arguments**
+([run](https://godbolt.org/z/Yn7Txe))
+
+``` c++
+std::string s = fmt::format("I'd rather be {1} than {0}.", "right", "happy");
+// s == "I'd rather be happy than right."
+```
+
+**Print dates and times** ([run](https://godbolt.org/z/c31ExdY3W))
+
+``` c++
+#include <fmt/chrono.h>
+
+int main() {
+  auto now = std::chrono::system_clock::now();
+  fmt::print("Date and time: {}\n", now);
+  fmt::print("Time: {:%H:%M}\n", now);
+}
+```
+
+Output:
+
+    Date and time: 2023-12-26 19:10:31.557195597
+    Time: 19:10
+
+**Print a container** ([run](https://godbolt.org/z/MxM1YqjE7))
+
+``` c++
+#include <vector>
+#include <fmt/ranges.h>
+
+int main() {
+  std::vector<int> v = {1, 2, 3};
+  fmt::print("{}\n", v);
+}
+```
+
+Output:
+
+    [1, 2, 3]
+
+**Check a format string at compile time**
+
+``` c++
+std::string s = fmt::format("{:d}", "I am not a number");
+```
+
+This gives a compile-time error in C++20 because `d` is an invalid
+format specifier for a string.
+
+**Write a file from a single thread**
+
+``` c++
+#include <fmt/os.h>
+
+int main() {
+  auto out = fmt::output_file("guide.txt");
+  out.print("Don't {}", "Panic");
+}
+```
+
+This can be [5 to 9 times faster than
+fprintf](http://www.zverovich.net/2020/08/04/optimal-file-buffer-size.html).
+
+**Print with colors and text styles**
+
+``` c++
+#include <fmt/color.h>
+
+int main() {
+  fmt::print(fg(fmt::color::crimson) | fmt::emphasis::bold,
+             "Hello, {}!\n", "world");
+  fmt::print(fg(fmt::color::floral_white) | bg(fmt::color::slate_gray) |
+             fmt::emphasis::underline, "Olá, {}!\n", "Mundo");
+  fmt::print(fg(fmt::color::steel_blue) | fmt::emphasis::italic,
+             "你好{}！\n", "世界");
+}
+```
+
+Output on a modern terminal with Unicode support:
+
+![image](https://github.com/fmtlib/fmt/assets/%0A576385/2a93c904-d6fa-4aa6-b453-2618e1c327d7)
+
+# Benchmarks
+
+## Speed tests
+
+| Library           | Method        | Run Time, s |
+|-------------------|---------------|-------------|
+| libc              | printf        |   0.91      |
+| libc++            | std::ostream  |   2.49      |
+| {fmt} 9.1         | fmt::print    |   0.74      |
+| Boost Format 1.80 | boost::format |   6.26      |
+| Folly Format      | folly::format |   1.87      |
+
+{fmt} is the fastest of the benchmarked methods, \~20% faster than
+`printf`.
+
+The above results were generated by building `tinyformat_test.cpp` on
+macOS 12.6.1 with `clang++ -O3 -DNDEBUG -DSPEED_TEST -DHAVE_FORMAT`, and
+taking the best of three runs. In the test, the format string
+`"%0.10f:%04d:%+g:%s:%p:%c:%%\n"` or equivalent is filled 2,000,000
+times with output sent to `/dev/null`; for further details refer to the
+[source](https://github.com/fmtlib/format-benchmark/blob/master/src/tinyformat-test.cc).
+
+{fmt} is up to 20-30x faster than `std::ostringstream` and `sprintf` on
+IEEE754 `float` and `double` formatting
+([dtoa-benchmark](https://github.com/fmtlib/dtoa-benchmark)) and faster
+than [double-conversion](https://github.com/google/double-conversion)
+and [ryu](https://github.com/ulfjack/ryu):
+
+[![image](https://user-images.githubusercontent.com/576385/95684665-11719600-0ba8-11eb-8e5b-972ff4e49428.png)](https://fmt.dev/unknown_mac64_clang12.0.html)
+
+## Compile time and code bloat
+
+The script
+[bloat-test.py](https://github.com/fmtlib/format-benchmark/blob/master/bloat-test.py)
+from [format-benchmark](https://github.com/fmtlib/format-benchmark)
+tests compile time and code bloat for nontrivial projects. It generates
+100 translation units and uses `printf()` or its alternative five times
+in each to simulate a medium-sized project. The resulting executable
+size and compile time (Apple LLVM version 8.1.0 (clang-802.0.42), macOS
+Sierra, best of three) is shown in the following tables.
+
+**Optimized build (-O3)**
+
+| Method        | Compile Time, s | Executable size, KiB | Stripped size, KiB |
+|---------------|-----------------|----------------------|--------------------|
+| printf        |   2.6           |   29                 |   26               |
+| printf+string |   16.4          |   29                 |   26               |
+| iostreams     |   31.1          |   59                 |   55               |
+| {fmt}         |   19.0          |   37                 |   34               |
+| Boost Format  |   91.9          |   226                |   203              |
+| Folly Format  |   115.7         |   101                |   88               |
+
+As you can see, {fmt} has 60% less overhead in terms of resulting binary
+code size compared to iostreams and comes pretty close to `printf`.
+Boost Format and Folly Format have the largest overheads.
+
+`printf+string` is the same as `printf` but with an extra `<string>`
+include to measure the overhead of the latter.
+
+**Non-optimized build**
+
+| Method        | Compile Time, s | Executable size, KiB | Stripped size, KiB |
+|---------------|-----------------|----------------------|--------------------|
+| printf        |   2.2           |   33                 |   30               |
+| printf+string |   16.0          |   33                 |   30               |
+| iostreams     |   28.3          |   56                 |   52               |
+| {fmt}         |   18.2          |   59                 |   50               |
+| Boost Format  |   54.1          |   365                |   303              |
+| Folly Format  |   79.9          |   445                |   430              |
+
+`libc`, `lib(std)c++`, and `libfmt` are all linked as shared libraries
+to compare formatting function overhead only. Boost Format is a
+header-only library so it doesn\'t provide any linkage options.
+
+## Running the tests
+
+Please refer to [Building the
+library](https://fmt.dev/latest/usage.html#building-the-library) for
+instructions on how to build the library and run the unit tests.
+
+Benchmarks reside in a separate repository,
+[format-benchmarks](https://github.com/fmtlib/format-benchmark), so to
+run the benchmarks you first need to clone this repository and generate
+Makefiles with CMake:
+
+    $ git clone --recursive https://github.com/fmtlib/format-benchmark.git
+    $ cd format-benchmark
+    $ cmake .
+
+Then you can run the speed test:
+
+    $ make speed-test
+
+or the bloat test:
+
+    $ make bloat-test
+
+# Migrating code
+
+[clang-tidy](https://clang.llvm.org/extra/clang-tidy/) v17 (not yet
+released) provides the
+[modernize-use-std-print](https://clang.llvm.org/extra/clang-tidy/checks/modernize/use-std-print.html)
+check that is capable of converting occurrences of `printf` and
+`fprintf` to `fmt::print` if configured to do so. (By default it
+converts to `std::print`.)
+
+# Notable projects using this library
+
+- [0 A.D.](https://play0ad.com/): a free, open-source, cross-platform
+  real-time strategy game
+- [AMPL/MP](https://github.com/ampl/mp): an open-source library for
+  mathematical programming
+- [Apple's FoundationDB](https://github.com/apple/foundationdb): an open-source,
+  distributed, transactional key-value store
+- [Aseprite](https://github.com/aseprite/aseprite): animated sprite
+  editor & pixel art tool
+- [AvioBook](https://www.aviobook.aero/en): a comprehensive aircraft
+  operations suite
+- [Blizzard Battle.net](https://battle.net/): an online gaming
+  platform
+- [Celestia](https://celestia.space/): real-time 3D visualization of
+  space
+- [Ceph](https://ceph.com/): a scalable distributed storage system
+- [ccache](https://ccache.dev/): a compiler cache
+- [ClickHouse](https://github.com/ClickHouse/ClickHouse): an
+  analytical database management system
+- [Contour](https://github.com/contour-terminal/contour/): a modern
+  terminal emulator
+- [CUAUV](https://cuauv.org/): Cornell University\'s autonomous
+  underwater vehicle
+- [Drake](https://drake.mit.edu/): a planning, control, and analysis
+  toolbox for nonlinear dynamical systems (MIT)
+- [Envoy](https://lyft.github.io/envoy/): C++ L7 proxy and
+  communication bus (Lyft)
+- [FiveM](https://fivem.net/): a modification framework for GTA V
+- [fmtlog](https://github.com/MengRao/fmtlog): a performant
+  fmtlib-style logging library with latency in nanoseconds
+- [Folly](https://github.com/facebook/folly): Facebook open-source
+  library
+- [GemRB](https://gemrb.org/): a portable open-source implementation
+  of Bioware's Infinity Engine
+- [Grand Mountain
+  Adventure](https://store.steampowered.com/app/1247360/Grand_Mountain_Adventure/):
+  a beautiful open-world ski & snowboarding game
+- [HarpyWar/pvpgn](https://github.com/pvpgn/pvpgn-server): Player vs
+  Player Gaming Network with tweaks
+- [KBEngine](https://github.com/kbengine/kbengine): an open-source
+  MMOG server engine
+- [Keypirinha](https://keypirinha.com/): a semantic launcher for
+  Windows
+- [Kodi](https://kodi.tv/) (formerly xbmc): home theater software
+- [Knuth](https://kth.cash/): high-performance Bitcoin full-node
+- [libunicode](https://github.com/contour-terminal/libunicode/): a
+  modern C++17 Unicode library
+- [MariaDB](https://mariadb.org/): relational database management
+  system
+- [Microsoft Verona](https://github.com/microsoft/verona): research
+  programming language for concurrent ownership
+- [MongoDB](https://mongodb.com/): distributed document database
+- [MongoDB Smasher](https://github.com/duckie/mongo_smasher): a small
+  tool to generate randomized datasets
+- [OpenSpace](https://openspaceproject.com/): an open-source
+  astrovisualization framework
+- [PenUltima Online (POL)](https://www.polserver.com/): an MMO server,
+  compatible with most Ultima Online clients
+- [PyTorch](https://github.com/pytorch/pytorch): an open-source
+  machine learning library
+- [quasardb](https://www.quasardb.net/): a distributed,
+  high-performance, associative database
+- [Quill](https://github.com/odygrd/quill): asynchronous low-latency
+  logging library
+- [QKW](https://github.com/ravijanjam/qkw): generalizing aliasing to
+  simplify navigation, and executing complex multi-line terminal
+  command sequences
+- [redis-cerberus](https://github.com/HunanTV/redis-cerberus): a Redis
+  cluster proxy
+- [redpanda](https://vectorized.io/redpanda): a 10x faster Kafka®
+  replacement for mission-critical systems written in C++
+- [rpclib](http://rpclib.net/): a modern C++ msgpack-RPC server and
+  client library
+- [Salesforce Analytics
+  Cloud](https://www.salesforce.com/analytics-cloud/overview/):
+  business intelligence software
+- [Scylla](https://www.scylladb.com/): a Cassandra-compatible NoSQL
+  data store that can handle 1 million transactions per second on a
+  single server
+- [Seastar](http://www.seastar-project.org/): an advanced, open-source
+  C++ framework for high-performance server applications on modern
+  hardware
+- [spdlog](https://github.com/gabime/spdlog): super fast C++ logging
+  library
+- [Stellar](https://www.stellar.org/): financial platform
+- [Touch Surgery](https://www.touchsurgery.com/): surgery simulator
+- [TrinityCore](https://github.com/TrinityCore/TrinityCore):
+  open-source MMORPG framework
+- [🐙 userver framework](https://userver.tech/): open-source
+  asynchronous framework with a rich set of abstractions and database
+  drivers
+- [Windows Terminal](https://github.com/microsoft/terminal): the new
+  Windows terminal
+
+[More\...](https://github.com/search?q=fmtlib&type=Code)
+
+If you are aware of other projects using this library, please let me
+know by [email](mailto:victor.zverovich@gmail.com) or by submitting an
+[issue](https://github.com/fmtlib/fmt/issues).
+
+# Motivation
+
+So why yet another formatting library?
+
+There are plenty of methods for doing this task, from standard ones like
+the printf family of function and iostreams to Boost Format and
+FastFormat libraries. The reason for creating a new library is that
+every existing solution that I found either had serious issues or
+didn\'t provide all the features I needed.
+
+## printf
+
+The good thing about `printf` is that it is pretty fast and readily
+available being a part of the C standard library. The main drawback is
+that it doesn\'t support user-defined types. `printf` also has safety
+issues although they are somewhat mitigated with [\_\_attribute\_\_
+((format (printf,
+\...))](https://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html) in
+GCC. There is a POSIX extension that adds positional arguments required
+for
+[i18n](https://en.wikipedia.org/wiki/Internationalization_and_localization)
+to `printf` but it is not a part of C99 and may not be available on some
+platforms.
+
+## iostreams
+
+The main issue with iostreams is best illustrated with an example:
+
+``` c++
+std::cout << std::setprecision(2) << std::fixed << 1.23456 << "\n";
+```
+
+which is a lot of typing compared to printf:
+
+``` c++
+printf("%.2f\n", 1.23456);
+```
+
+Matthew Wilson, the author of FastFormat, called this \"chevron hell\".
+iostreams don\'t support positional arguments by design.
+
+The good part is that iostreams support user-defined types and are safe
+although error handling is awkward.
+
+## Boost Format
+
+This is a very powerful library that supports both `printf`-like format
+strings and positional arguments. Its main drawback is performance.
+According to various benchmarks, it is much slower than other methods
+considered here. Boost Format also has excessive build times and severe
+code bloat issues (see [Benchmarks](#benchmarks)).
+
+## FastFormat
+
+This is an interesting library that is fast, safe, and has positional
+arguments. However, it has significant limitations, citing its author:
+
+> Three features that have no hope of being accommodated within the
+> current design are:
+>
+> - Leading zeros (or any other non-space padding)
+> - Octal/hexadecimal encoding
+> - Runtime width/alignment specification
+
+It is also quite big and has a heavy dependency, STLSoft, which might be
+too restrictive for using it in some projects.
+
+## Boost Spirit.Karma
+
+This is not a formatting library but I decided to include it here for
+completeness. As iostreams, it suffers from the problem of mixing
+verbatim text with arguments. The library is pretty fast, but slower on
+integer formatting than `fmt::format_to` with format string compilation
+on Karma\'s own benchmark, see [Converting a hundred million integers to
+strings per
+second](http://www.zverovich.net/2020/06/13/fast-int-to-string-revisited.html).
+
+# License
+
+{fmt} is distributed under the MIT
+[license](https://github.com/fmtlib/fmt/blob/master/LICENSE).
+
+# Documentation License
+
+The [Format String Syntax](https://fmt.dev/latest/syntax.html) section
+in the documentation is based on the one from Python [string module
+documentation](https://docs.python.org/3/library/string.html#module-string).
+For this reason, the documentation is distributed under the Python
+Software Foundation license available in
+[doc/python-license.txt](https://raw.github.com/fmtlib/fmt/master/doc/python-license.txt).
+It only applies if you distribute the documentation of {fmt}.
+
+# Maintainers
+
+The {fmt} library is maintained by Victor Zverovich
+([vitaut](https://github.com/vitaut)) with contributions from many other
+people. See
+[Contributors](https://github.com/fmtlib/fmt/graphs/contributors) and
+[Releases](https://github.com/fmtlib/fmt/releases) for some of the
+names. Let us know if your contribution is not listed or mentioned
+incorrectly and we\'ll make it right.
+
+# Security Policy
+
+To report a security issue, please disclose it at [security
+advisory](https://github.com/fmtlib/fmt/security/advisories/new).
+
+This project is maintained by a team of volunteers on a
+reasonable-effort basis. As such, please give us at least 90 days to
+work on a fix before public exposure.
diff --git a/dep/fmt/UPSTREAM.md b/dep/fmt/UPSTREAM.md
new file mode 100644
index 00000000..4b3e02ed
--- /dev/null
+++ b/dep/fmt/UPSTREAM.md
@@ -0,0 +1,2 @@
+This is a pruned version of fmt 10.2.1, obtained from
+https://github.com/fmtlib/fmt/releases/tag/10.2.1.
\ No newline at end of file
diff --git a/dep/fmt/include/fmt/args.h b/dep/fmt/include/fmt/args.h
new file mode 100644
index 00000000..ad1654bb
--- /dev/null
+++ b/dep/fmt/include/fmt/args.h
@@ -0,0 +1,235 @@
+// Formatting library for C++ - dynamic argument lists
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_ARGS_H_
+#define FMT_ARGS_H_
+
+#include <functional>  // std::reference_wrapper
+#include <memory>      // std::unique_ptr
+#include <vector>
+
+#include "core.h"
+
+FMT_BEGIN_NAMESPACE
+
+namespace detail {
+
+template <typename T> struct is_reference_wrapper : std::false_type {};
+template <typename T>
+struct is_reference_wrapper<std::reference_wrapper<T>> : std::true_type {};
+
+template <typename T> auto unwrap(const T& v) -> const T& { return v; }
+template <typename T>
+auto unwrap(const std::reference_wrapper<T>& v) -> const T& {
+  return static_cast<const T&>(v);
+}
+
+class dynamic_arg_list {
+  // Workaround for clang's -Wweak-vtables. Unlike for regular classes, for
+  // templates it doesn't complain about inability to deduce single translation
+  // unit for placing vtable. So storage_node_base is made a fake template.
+  template <typename = void> struct node {
+    virtual ~node() = default;
+    std::unique_ptr<node<>> next;
+  };
+
+  template <typename T> struct typed_node : node<> {
+    T value;
+
+    template <typename Arg>
+    FMT_CONSTEXPR typed_node(const Arg& arg) : value(arg) {}
+
+    template <typename Char>
+    FMT_CONSTEXPR typed_node(const basic_string_view<Char>& arg)
+        : value(arg.data(), arg.size()) {}
+  };
+
+  std::unique_ptr<node<>> head_;
+
+ public:
+  template <typename T, typename Arg> auto push(const Arg& arg) -> const T& {
+    auto new_node = std::unique_ptr<typed_node<T>>(new typed_node<T>(arg));
+    auto& value = new_node->value;
+    new_node->next = std::move(head_);
+    head_ = std::move(new_node);
+    return value;
+  }
+};
+}  // namespace detail
+
+/**
+  \rst
+  A dynamic version of `fmt::format_arg_store`.
+  It's equipped with a storage to potentially temporary objects which lifetimes
+  could be shorter than the format arguments object.
+
+  It can be implicitly converted into `~fmt::basic_format_args` for passing
+  into type-erased formatting functions such as `~fmt::vformat`.
+  \endrst
+ */
+template <typename Context>
+class dynamic_format_arg_store
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+    // Workaround a GCC template argument substitution bug.
+    : public basic_format_args<Context>
+#endif
+{
+ private:
+  using char_type = typename Context::char_type;
+
+  template <typename T> struct need_copy {
+    static constexpr detail::type mapped_type =
+        detail::mapped_type_constant<T, Context>::value;
+
+    enum {
+      value = !(detail::is_reference_wrapper<T>::value ||
+                std::is_same<T, basic_string_view<char_type>>::value ||
+                std::is_same<T, detail::std_string_view<char_type>>::value ||
+                (mapped_type != detail::type::cstring_type &&
+                 mapped_type != detail::type::string_type &&
+                 mapped_type != detail::type::custom_type))
+    };
+  };
+
+  template <typename T>
+  using stored_type = conditional_t<
+      std::is_convertible<T, std::basic_string<char_type>>::value &&
+          !detail::is_reference_wrapper<T>::value,
+      std::basic_string<char_type>, T>;
+
+  // Storage of basic_format_arg must be contiguous.
+  std::vector<basic_format_arg<Context>> data_;
+  std::vector<detail::named_arg_info<char_type>> named_info_;
+
+  // Storage of arguments not fitting into basic_format_arg must grow
+  // without relocation because items in data_ refer to it.
+  detail::dynamic_arg_list dynamic_args_;
+
+  friend class basic_format_args<Context>;
+
+  auto get_types() const -> unsigned long long {
+    return detail::is_unpacked_bit | data_.size() |
+           (named_info_.empty()
+                ? 0ULL
+                : static_cast<unsigned long long>(detail::has_named_args_bit));
+  }
+
+  auto data() const -> const basic_format_arg<Context>* {
+    return named_info_.empty() ? data_.data() : data_.data() + 1;
+  }
+
+  template <typename T> void emplace_arg(const T& arg) {
+    data_.emplace_back(detail::make_arg<Context>(arg));
+  }
+
+  template <typename T>
+  void emplace_arg(const detail::named_arg<char_type, T>& arg) {
+    if (named_info_.empty()) {
+      constexpr const detail::named_arg_info<char_type>* zero_ptr{nullptr};
+      data_.insert(data_.begin(), {zero_ptr, 0});
+    }
+    data_.emplace_back(detail::make_arg<Context>(detail::unwrap(arg.value)));
+    auto pop_one = [](std::vector<basic_format_arg<Context>>* data) {
+      data->pop_back();
+    };
+    std::unique_ptr<std::vector<basic_format_arg<Context>>, decltype(pop_one)>
+        guard{&data_, pop_one};
+    named_info_.push_back({arg.name, static_cast<int>(data_.size() - 2u)});
+    data_[0].value_.named_args = {named_info_.data(), named_info_.size()};
+    guard.release();
+  }
+
+ public:
+  constexpr dynamic_format_arg_store() = default;
+
+  /**
+    \rst
+    Adds an argument into the dynamic store for later passing to a formatting
+    function.
+
+    Note that custom types and string types (but not string views) are copied
+    into the store dynamically allocating memory if necessary.
+
+    **Example**::
+
+      fmt::dynamic_format_arg_store<fmt::format_context> store;
+      store.push_back(42);
+      store.push_back("abc");
+      store.push_back(1.5f);
+      std::string result = fmt::vformat("{} and {} and {}", store);
+    \endrst
+  */
+  template <typename T> void push_back(const T& arg) {
+    if (detail::const_check(need_copy<T>::value))
+      emplace_arg(dynamic_args_.push<stored_type<T>>(arg));
+    else
+      emplace_arg(detail::unwrap(arg));
+  }
+
+  /**
+    \rst
+    Adds a reference to the argument into the dynamic store for later passing to
+    a formatting function.
+
+    **Example**::
+
+      fmt::dynamic_format_arg_store<fmt::format_context> store;
+      char band[] = "Rolling Stones";
+      store.push_back(std::cref(band));
+      band[9] = 'c'; // Changing str affects the output.
+      std::string result = fmt::vformat("{}", store);
+      // result == "Rolling Scones"
+    \endrst
+  */
+  template <typename T> void push_back(std::reference_wrapper<T> arg) {
+    static_assert(
+        need_copy<T>::value,
+        "objects of built-in types and string views are always copied");
+    emplace_arg(arg.get());
+  }
+
+  /**
+    Adds named argument into the dynamic store for later passing to a formatting
+    function. ``std::reference_wrapper`` is supported to avoid copying of the
+    argument. The name is always copied into the store.
+  */
+  template <typename T>
+  void push_back(const detail::named_arg<char_type, T>& arg) {
+    const char_type* arg_name =
+        dynamic_args_.push<std::basic_string<char_type>>(arg.name).c_str();
+    if (detail::const_check(need_copy<T>::value)) {
+      emplace_arg(
+          fmt::arg(arg_name, dynamic_args_.push<stored_type<T>>(arg.value)));
+    } else {
+      emplace_arg(fmt::arg(arg_name, arg.value));
+    }
+  }
+
+  /** Erase all elements from the store */
+  void clear() {
+    data_.clear();
+    named_info_.clear();
+    dynamic_args_ = detail::dynamic_arg_list();
+  }
+
+  /**
+    \rst
+    Reserves space to store at least *new_cap* arguments including
+    *new_cap_named* named arguments.
+    \endrst
+  */
+  void reserve(size_t new_cap, size_t new_cap_named) {
+    FMT_ASSERT(new_cap >= new_cap_named,
+               "Set of arguments includes set of named arguments");
+    data_.reserve(new_cap);
+    named_info_.reserve(new_cap_named);
+  }
+};
+
+FMT_END_NAMESPACE
+
+#endif  // FMT_ARGS_H_
diff --git a/dep/fmt/include/fmt/chrono.h b/dep/fmt/include/fmt/chrono.h
new file mode 100644
index 00000000..9d54574e
--- /dev/null
+++ b/dep/fmt/include/fmt/chrono.h
@@ -0,0 +1,2240 @@
+// Formatting library for C++ - chrono support
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_CHRONO_H_
+#define FMT_CHRONO_H_
+
+#include <algorithm>
+#include <chrono>
+#include <cmath>    // std::isfinite
+#include <cstring>  // std::memcpy
+#include <ctime>
+#include <iterator>
+#include <locale>
+#include <ostream>
+#include <type_traits>
+
+#include "ostream.h"  // formatbuf
+
+FMT_BEGIN_NAMESPACE
+
+// Check if std::chrono::local_t is available.
+#ifndef FMT_USE_LOCAL_TIME
+#  ifdef __cpp_lib_chrono
+#    define FMT_USE_LOCAL_TIME (__cpp_lib_chrono >= 201907L)
+#  else
+#    define FMT_USE_LOCAL_TIME 0
+#  endif
+#endif
+
+// Check if std::chrono::utc_timestamp is available.
+#ifndef FMT_USE_UTC_TIME
+#  ifdef __cpp_lib_chrono
+#    define FMT_USE_UTC_TIME (__cpp_lib_chrono >= 201907L)
+#  else
+#    define FMT_USE_UTC_TIME 0
+#  endif
+#endif
+
+// Enable tzset.
+#ifndef FMT_USE_TZSET
+// UWP doesn't provide _tzset.
+#  if FMT_HAS_INCLUDE("winapifamily.h")
+#    include <winapifamily.h>
+#  endif
+#  if defined(_WIN32) && (!defined(WINAPI_FAMILY) || \
+                          (WINAPI_FAMILY == WINAPI_FAMILY_DESKTOP_APP))
+#    define FMT_USE_TZSET 1
+#  else
+#    define FMT_USE_TZSET 0
+#  endif
+#endif
+
+// Enable safe chrono durations, unless explicitly disabled.
+#ifndef FMT_SAFE_DURATION_CAST
+#  define FMT_SAFE_DURATION_CAST 1
+#endif
+#if FMT_SAFE_DURATION_CAST
+
+// For conversion between std::chrono::durations without undefined
+// behaviour or erroneous results.
+// This is a stripped down version of duration_cast, for inclusion in fmt.
+// See https://github.com/pauldreik/safe_duration_cast
+//
+// Copyright Paul Dreik 2019
+namespace safe_duration_cast {
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value &&
+                        std::numeric_limits<From>::is_signed ==
+                            std::numeric_limits<To>::is_signed)>
+FMT_CONSTEXPR auto lossless_integral_conversion(const From from, int& ec)
+    -> To {
+  ec = 0;
+  using F = std::numeric_limits<From>;
+  using T = std::numeric_limits<To>;
+  static_assert(F::is_integer, "From must be integral");
+  static_assert(T::is_integer, "To must be integral");
+
+  // A and B are both signed, or both unsigned.
+  if (detail::const_check(F::digits <= T::digits)) {
+    // From fits in To without any problem.
+  } else {
+    // From does not always fit in To, resort to a dynamic check.
+    if (from < (T::min)() || from > (T::max)()) {
+      // outside range.
+      ec = 1;
+      return {};
+    }
+  }
+  return static_cast<To>(from);
+}
+
+/**
+ * converts From to To, without loss. If the dynamic value of from
+ * can't be converted to To without loss, ec is set.
+ */
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value &&
+                        std::numeric_limits<From>::is_signed !=
+                            std::numeric_limits<To>::is_signed)>
+FMT_CONSTEXPR auto lossless_integral_conversion(const From from, int& ec)
+    -> To {
+  ec = 0;
+  using F = std::numeric_limits<From>;
+  using T = std::numeric_limits<To>;
+  static_assert(F::is_integer, "From must be integral");
+  static_assert(T::is_integer, "To must be integral");
+
+  if (detail::const_check(F::is_signed && !T::is_signed)) {
+    // From may be negative, not allowed!
+    if (fmt::detail::is_negative(from)) {
+      ec = 1;
+      return {};
+    }
+    // From is positive. Can it always fit in To?
+    if (detail::const_check(F::digits > T::digits) &&
+        from > static_cast<From>(detail::max_value<To>())) {
+      ec = 1;
+      return {};
+    }
+  }
+
+  if (detail::const_check(!F::is_signed && T::is_signed &&
+                          F::digits >= T::digits) &&
+      from > static_cast<From>(detail::max_value<To>())) {
+    ec = 1;
+    return {};
+  }
+  return static_cast<To>(from);  // Lossless conversion.
+}
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(std::is_same<From, To>::value)>
+FMT_CONSTEXPR auto lossless_integral_conversion(const From from, int& ec)
+    -> To {
+  ec = 0;
+  return from;
+}  // function
+
+// clang-format off
+/**
+ * converts From to To if possible, otherwise ec is set.
+ *
+ * input                            |    output
+ * ---------------------------------|---------------
+ * NaN                              | NaN
+ * Inf                              | Inf
+ * normal, fits in output           | converted (possibly lossy)
+ * normal, does not fit in output   | ec is set
+ * subnormal                        | best effort
+ * -Inf                             | -Inf
+ */
+// clang-format on
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value)>
+FMT_CONSTEXPR auto safe_float_conversion(const From from, int& ec) -> To {
+  ec = 0;
+  using T = std::numeric_limits<To>;
+  static_assert(std::is_floating_point<From>::value, "From must be floating");
+  static_assert(std::is_floating_point<To>::value, "To must be floating");
+
+  // catch the only happy case
+  if (std::isfinite(from)) {
+    if (from >= T::lowest() && from <= (T::max)()) {
+      return static_cast<To>(from);
+    }
+    // not within range.
+    ec = 1;
+    return {};
+  }
+
+  // nan and inf will be preserved
+  return static_cast<To>(from);
+}  // function
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(std::is_same<From, To>::value)>
+FMT_CONSTEXPR auto safe_float_conversion(const From from, int& ec) -> To {
+  ec = 0;
+  static_assert(std::is_floating_point<From>::value, "From must be floating");
+  return from;
+}
+
+/**
+ * safe duration cast between integral durations
+ */
+template <typename To, typename FromRep, typename FromPeriod,
+          FMT_ENABLE_IF(std::is_integral<FromRep>::value),
+          FMT_ENABLE_IF(std::is_integral<typename To::rep>::value)>
+auto safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
+                        int& ec) -> To {
+  using From = std::chrono::duration<FromRep, FromPeriod>;
+  ec = 0;
+  // the basic idea is that we need to convert from count() in the from type
+  // to count() in the To type, by multiplying it with this:
+  struct Factor
+      : std::ratio_divide<typename From::period, typename To::period> {};
+
+  static_assert(Factor::num > 0, "num must be positive");
+  static_assert(Factor::den > 0, "den must be positive");
+
+  // the conversion is like this: multiply from.count() with Factor::num
+  // /Factor::den and convert it to To::rep, all this without
+  // overflow/underflow. let's start by finding a suitable type that can hold
+  // both To, From and Factor::num
+  using IntermediateRep =
+      typename std::common_type<typename From::rep, typename To::rep,
+                                decltype(Factor::num)>::type;
+
+  // safe conversion to IntermediateRep
+  IntermediateRep count =
+      lossless_integral_conversion<IntermediateRep>(from.count(), ec);
+  if (ec) return {};
+  // multiply with Factor::num without overflow or underflow
+  if (detail::const_check(Factor::num != 1)) {
+    const auto max1 = detail::max_value<IntermediateRep>() / Factor::num;
+    if (count > max1) {
+      ec = 1;
+      return {};
+    }
+    const auto min1 =
+        (std::numeric_limits<IntermediateRep>::min)() / Factor::num;
+    if (detail::const_check(!std::is_unsigned<IntermediateRep>::value) &&
+        count < min1) {
+      ec = 1;
+      return {};
+    }
+    count *= Factor::num;
+  }
+
+  if (detail::const_check(Factor::den != 1)) count /= Factor::den;
+  auto tocount = lossless_integral_conversion<typename To::rep>(count, ec);
+  return ec ? To() : To(tocount);
+}
+
+/**
+ * safe duration_cast between floating point durations
+ */
+template <typename To, typename FromRep, typename FromPeriod,
+          FMT_ENABLE_IF(std::is_floating_point<FromRep>::value),
+          FMT_ENABLE_IF(std::is_floating_point<typename To::rep>::value)>
+auto safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
+                        int& ec) -> To {
+  using From = std::chrono::duration<FromRep, FromPeriod>;
+  ec = 0;
+  if (std::isnan(from.count())) {
+    // nan in, gives nan out. easy.
+    return To{std::numeric_limits<typename To::rep>::quiet_NaN()};
+  }
+  // maybe we should also check if from is denormal, and decide what to do about
+  // it.
+
+  // +-inf should be preserved.
+  if (std::isinf(from.count())) {
+    return To{from.count()};
+  }
+
+  // the basic idea is that we need to convert from count() in the from type
+  // to count() in the To type, by multiplying it with this:
+  struct Factor
+      : std::ratio_divide<typename From::period, typename To::period> {};
+
+  static_assert(Factor::num > 0, "num must be positive");
+  static_assert(Factor::den > 0, "den must be positive");
+
+  // the conversion is like this: multiply from.count() with Factor::num
+  // /Factor::den and convert it to To::rep, all this without
+  // overflow/underflow. let's start by finding a suitable type that can hold
+  // both To, From and Factor::num
+  using IntermediateRep =
+      typename std::common_type<typename From::rep, typename To::rep,
+                                decltype(Factor::num)>::type;
+
+  // force conversion of From::rep -> IntermediateRep to be safe,
+  // even if it will never happen be narrowing in this context.
+  IntermediateRep count =
+      safe_float_conversion<IntermediateRep>(from.count(), ec);
+  if (ec) {
+    return {};
+  }
+
+  // multiply with Factor::num without overflow or underflow
+  if (detail::const_check(Factor::num != 1)) {
+    constexpr auto max1 = detail::max_value<IntermediateRep>() /
+                          static_cast<IntermediateRep>(Factor::num);
+    if (count > max1) {
+      ec = 1;
+      return {};
+    }
+    constexpr auto min1 = std::numeric_limits<IntermediateRep>::lowest() /
+                          static_cast<IntermediateRep>(Factor::num);
+    if (count < min1) {
+      ec = 1;
+      return {};
+    }
+    count *= static_cast<IntermediateRep>(Factor::num);
+  }
+
+  // this can't go wrong, right? den>0 is checked earlier.
+  if (detail::const_check(Factor::den != 1)) {
+    using common_t = typename std::common_type<IntermediateRep, intmax_t>::type;
+    count /= static_cast<common_t>(Factor::den);
+  }
+
+  // convert to the to type, safely
+  using ToRep = typename To::rep;
+
+  const ToRep tocount = safe_float_conversion<ToRep>(count, ec);
+  if (ec) {
+    return {};
+  }
+  return To{tocount};
+}
+}  // namespace safe_duration_cast
+#endif
+
+// Prevents expansion of a preceding token as a function-style macro.
+// Usage: f FMT_NOMACRO()
+#define FMT_NOMACRO
+
+namespace detail {
+template <typename T = void> struct null {};
+inline auto localtime_r FMT_NOMACRO(...) -> null<> { return null<>(); }
+inline auto localtime_s(...) -> null<> { return null<>(); }
+inline auto gmtime_r(...) -> null<> { return null<>(); }
+inline auto gmtime_s(...) -> null<> { return null<>(); }
+
+inline auto get_classic_locale() -> const std::locale& {
+  static const auto& locale = std::locale::classic();
+  return locale;
+}
+
+template <typename CodeUnit> struct codecvt_result {
+  static constexpr const size_t max_size = 32;
+  CodeUnit buf[max_size];
+  CodeUnit* end;
+};
+
+template <typename CodeUnit>
+void write_codecvt(codecvt_result<CodeUnit>& out, string_view in_buf,
+                   const std::locale& loc) {
+#if FMT_CLANG_VERSION
+#  pragma clang diagnostic push
+#  pragma clang diagnostic ignored "-Wdeprecated"
+  auto& f = std::use_facet<std::codecvt<CodeUnit, char, std::mbstate_t>>(loc);
+#  pragma clang diagnostic pop
+#else
+  auto& f = std::use_facet<std::codecvt<CodeUnit, char, std::mbstate_t>>(loc);
+#endif
+  auto mb = std::mbstate_t();
+  const char* from_next = nullptr;
+  auto result = f.in(mb, in_buf.begin(), in_buf.end(), from_next,
+                     std::begin(out.buf), std::end(out.buf), out.end);
+  if (result != std::codecvt_base::ok)
+    FMT_THROW(format_error("failed to format time"));
+}
+
+template <typename OutputIt>
+auto write_encoded_tm_str(OutputIt out, string_view in, const std::locale& loc)
+    -> OutputIt {
+  if (detail::is_utf8() && loc != get_classic_locale()) {
+    // char16_t and char32_t codecvts are broken in MSVC (linkage errors) and
+    // gcc-4.
+#if FMT_MSC_VERSION != 0 || \
+    (defined(__GLIBCXX__) && !defined(_GLIBCXX_USE_DUAL_ABI))
+    // The _GLIBCXX_USE_DUAL_ABI macro is always defined in libstdc++ from gcc-5
+    // and newer.
+    using code_unit = wchar_t;
+#else
+    using code_unit = char32_t;
+#endif
+
+    using unit_t = codecvt_result<code_unit>;
+    unit_t unit;
+    write_codecvt(unit, in, loc);
+    // In UTF-8 is used one to four one-byte code units.
+    auto u =
+        to_utf8<code_unit, basic_memory_buffer<char, unit_t::max_size * 4>>();
+    if (!u.convert({unit.buf, to_unsigned(unit.end - unit.buf)}))
+      FMT_THROW(format_error("failed to format time"));
+    return copy_str<char>(u.c_str(), u.c_str() + u.size(), out);
+  }
+  return copy_str<char>(in.data(), in.data() + in.size(), out);
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+auto write_tm_str(OutputIt out, string_view sv, const std::locale& loc)
+    -> OutputIt {
+  codecvt_result<Char> unit;
+  write_codecvt(unit, sv, loc);
+  return copy_str<Char>(unit.buf, unit.end, out);
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(std::is_same<Char, char>::value)>
+auto write_tm_str(OutputIt out, string_view sv, const std::locale& loc)
+    -> OutputIt {
+  return write_encoded_tm_str(out, sv, loc);
+}
+
+template <typename Char>
+inline void do_write(buffer<Char>& buf, const std::tm& time,
+                     const std::locale& loc, char format, char modifier) {
+  auto&& format_buf = formatbuf<std::basic_streambuf<Char>>(buf);
+  auto&& os = std::basic_ostream<Char>(&format_buf);
+  os.imbue(loc);
+  const auto& facet = std::use_facet<std::time_put<Char>>(loc);
+  auto end = facet.put(os, os, Char(' '), &time, format, modifier);
+  if (end.failed()) FMT_THROW(format_error("failed to format time"));
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+auto write(OutputIt out, const std::tm& time, const std::locale& loc,
+           char format, char modifier = 0) -> OutputIt {
+  auto&& buf = get_buffer<Char>(out);
+  do_write<Char>(buf, time, loc, format, modifier);
+  return get_iterator(buf, out);
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(std::is_same<Char, char>::value)>
+auto write(OutputIt out, const std::tm& time, const std::locale& loc,
+           char format, char modifier = 0) -> OutputIt {
+  auto&& buf = basic_memory_buffer<Char>();
+  do_write<char>(buf, time, loc, format, modifier);
+  return write_encoded_tm_str(out, string_view(buf.data(), buf.size()), loc);
+}
+
+template <typename Rep1, typename Rep2>
+struct is_same_arithmetic_type
+    : public std::integral_constant<bool,
+                                    (std::is_integral<Rep1>::value &&
+                                     std::is_integral<Rep2>::value) ||
+                                        (std::is_floating_point<Rep1>::value &&
+                                         std::is_floating_point<Rep2>::value)> {
+};
+
+template <
+    typename To, typename FromRep, typename FromPeriod,
+    FMT_ENABLE_IF(is_same_arithmetic_type<FromRep, typename To::rep>::value)>
+auto fmt_duration_cast(std::chrono::duration<FromRep, FromPeriod> from) -> To {
+#if FMT_SAFE_DURATION_CAST
+  // Throwing version of safe_duration_cast is only available for
+  // integer to integer or float to float casts.
+  int ec;
+  To to = safe_duration_cast::safe_duration_cast<To>(from, ec);
+  if (ec) FMT_THROW(format_error("cannot format duration"));
+  return to;
+#else
+  // Standard duration cast, may overflow.
+  return std::chrono::duration_cast<To>(from);
+#endif
+}
+
+template <
+    typename To, typename FromRep, typename FromPeriod,
+    FMT_ENABLE_IF(!is_same_arithmetic_type<FromRep, typename To::rep>::value)>
+auto fmt_duration_cast(std::chrono::duration<FromRep, FromPeriod> from) -> To {
+  // Mixed integer <-> float cast is not supported by safe_duration_cast.
+  return std::chrono::duration_cast<To>(from);
+}
+
+template <typename Duration>
+auto to_time_t(
+    std::chrono::time_point<std::chrono::system_clock, Duration> time_point)
+    -> std::time_t {
+  // Cannot use std::chrono::system_clock::to_time_t since this would first
+  // require a cast to std::chrono::system_clock::time_point, which could
+  // overflow.
+  return fmt_duration_cast<std::chrono::duration<std::time_t>>(
+             time_point.time_since_epoch())
+      .count();
+}
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+/**
+  Converts given time since epoch as ``std::time_t`` value into calendar time,
+  expressed in local time. Unlike ``std::localtime``, this function is
+  thread-safe on most platforms.
+ */
+inline auto localtime(std::time_t time) -> std::tm {
+  struct dispatcher {
+    std::time_t time_;
+    std::tm tm_;
+
+    dispatcher(std::time_t t) : time_(t) {}
+
+    auto run() -> bool {
+      using namespace fmt::detail;
+      return handle(localtime_r(&time_, &tm_));
+    }
+
+    auto handle(std::tm* tm) -> bool { return tm != nullptr; }
+
+    auto handle(detail::null<>) -> bool {
+      using namespace fmt::detail;
+      return fallback(localtime_s(&tm_, &time_));
+    }
+
+    auto fallback(int res) -> bool { return res == 0; }
+
+#if !FMT_MSC_VERSION
+    auto fallback(detail::null<>) -> bool {
+      using namespace fmt::detail;
+      std::tm* tm = std::localtime(&time_);
+      if (tm) tm_ = *tm;
+      return tm != nullptr;
+    }
+#endif
+  };
+  dispatcher lt(time);
+  // Too big time values may be unsupported.
+  if (!lt.run()) FMT_THROW(format_error("time_t value out of range"));
+  return lt.tm_;
+}
+
+#if FMT_USE_LOCAL_TIME
+template <typename Duration>
+inline auto localtime(std::chrono::local_time<Duration> time) -> std::tm {
+  return localtime(
+      detail::to_time_t(std::chrono::current_zone()->to_sys(time)));
+}
+#endif
+
+/**
+  Converts given time since epoch as ``std::time_t`` value into calendar time,
+  expressed in Coordinated Universal Time (UTC). Unlike ``std::gmtime``, this
+  function is thread-safe on most platforms.
+ */
+inline auto gmtime(std::time_t time) -> std::tm {
+  struct dispatcher {
+    std::time_t time_;
+    std::tm tm_;
+
+    dispatcher(std::time_t t) : time_(t) {}
+
+    auto run() -> bool {
+      using namespace fmt::detail;
+      return handle(gmtime_r(&time_, &tm_));
+    }
+
+    auto handle(std::tm* tm) -> bool { return tm != nullptr; }
+
+    auto handle(detail::null<>) -> bool {
+      using namespace fmt::detail;
+      return fallback(gmtime_s(&tm_, &time_));
+    }
+
+    auto fallback(int res) -> bool { return res == 0; }
+
+#if !FMT_MSC_VERSION
+    auto fallback(detail::null<>) -> bool {
+      std::tm* tm = std::gmtime(&time_);
+      if (tm) tm_ = *tm;
+      return tm != nullptr;
+    }
+#endif
+  };
+  auto gt = dispatcher(time);
+  // Too big time values may be unsupported.
+  if (!gt.run()) FMT_THROW(format_error("time_t value out of range"));
+  return gt.tm_;
+}
+
+template <typename Duration>
+inline auto gmtime(
+    std::chrono::time_point<std::chrono::system_clock, Duration> time_point)
+    -> std::tm {
+  return gmtime(detail::to_time_t(time_point));
+}
+
+namespace detail {
+
+// Writes two-digit numbers a, b and c separated by sep to buf.
+// The method by Pavel Novikov based on
+// https://johnnylee-sde.github.io/Fast-unsigned-integer-to-time-string/.
+inline void write_digit2_separated(char* buf, unsigned a, unsigned b,
+                                   unsigned c, char sep) {
+  unsigned long long digits =
+      a | (b << 24) | (static_cast<unsigned long long>(c) << 48);
+  // Convert each value to BCD.
+  // We have x = a * 10 + b and we want to convert it to BCD y = a * 16 + b.
+  // The difference is
+  //   y - x = a * 6
+  // a can be found from x:
+  //   a = floor(x / 10)
+  // then
+  //   y = x + a * 6 = x + floor(x / 10) * 6
+  // floor(x / 10) is (x * 205) >> 11 (needs 16 bits).
+  digits += (((digits * 205) >> 11) & 0x000f00000f00000f) * 6;
+  // Put low nibbles to high bytes and high nibbles to low bytes.
+  digits = ((digits & 0x00f00000f00000f0) >> 4) |
+           ((digits & 0x000f00000f00000f) << 8);
+  auto usep = static_cast<unsigned long long>(sep);
+  // Add ASCII '0' to each digit byte and insert separators.
+  digits |= 0x3030003030003030 | (usep << 16) | (usep << 40);
+
+  constexpr const size_t len = 8;
+  if (const_check(is_big_endian())) {
+    char tmp[len];
+    std::memcpy(tmp, &digits, len);
+    std::reverse_copy(tmp, tmp + len, buf);
+  } else {
+    std::memcpy(buf, &digits, len);
+  }
+}
+
+template <typename Period>
+FMT_CONSTEXPR inline auto get_units() -> const char* {
+  if (std::is_same<Period, std::atto>::value) return "as";
+  if (std::is_same<Period, std::femto>::value) return "fs";
+  if (std::is_same<Period, std::pico>::value) return "ps";
+  if (std::is_same<Period, std::nano>::value) return "ns";
+  if (std::is_same<Period, std::micro>::value) return "µs";
+  if (std::is_same<Period, std::milli>::value) return "ms";
+  if (std::is_same<Period, std::centi>::value) return "cs";
+  if (std::is_same<Period, std::deci>::value) return "ds";
+  if (std::is_same<Period, std::ratio<1>>::value) return "s";
+  if (std::is_same<Period, std::deca>::value) return "das";
+  if (std::is_same<Period, std::hecto>::value) return "hs";
+  if (std::is_same<Period, std::kilo>::value) return "ks";
+  if (std::is_same<Period, std::mega>::value) return "Ms";
+  if (std::is_same<Period, std::giga>::value) return "Gs";
+  if (std::is_same<Period, std::tera>::value) return "Ts";
+  if (std::is_same<Period, std::peta>::value) return "Ps";
+  if (std::is_same<Period, std::exa>::value) return "Es";
+  if (std::is_same<Period, std::ratio<60>>::value) return "min";
+  if (std::is_same<Period, std::ratio<3600>>::value) return "h";
+  if (std::is_same<Period, std::ratio<86400>>::value) return "d";
+  return nullptr;
+}
+
+enum class numeric_system {
+  standard,
+  // Alternative numeric system, e.g. 十二 instead of 12 in ja_JP locale.
+  alternative
+};
+
+// Glibc extensions for formatting numeric values.
+enum class pad_type {
+  unspecified,
+  // Do not pad a numeric result string.
+  none,
+  // Pad a numeric result string with zeros even if the conversion specifier
+  // character uses space-padding by default.
+  zero,
+  // Pad a numeric result string with spaces.
+  space,
+};
+
+template <typename OutputIt>
+auto write_padding(OutputIt out, pad_type pad, int width) -> OutputIt {
+  if (pad == pad_type::none) return out;
+  return std::fill_n(out, width, pad == pad_type::space ? ' ' : '0');
+}
+
+template <typename OutputIt>
+auto write_padding(OutputIt out, pad_type pad) -> OutputIt {
+  if (pad != pad_type::none) *out++ = pad == pad_type::space ? ' ' : '0';
+  return out;
+}
+
+// Parses a put_time-like format string and invokes handler actions.
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_chrono_format(const Char* begin, const Char* end,
+                                       Handler&& handler) -> const Char* {
+  if (begin == end || *begin == '}') return begin;
+  if (*begin != '%') FMT_THROW(format_error("invalid format"));
+  auto ptr = begin;
+  pad_type pad = pad_type::unspecified;
+  while (ptr != end) {
+    auto c = *ptr;
+    if (c == '}') break;
+    if (c != '%') {
+      ++ptr;
+      continue;
+    }
+    if (begin != ptr) handler.on_text(begin, ptr);
+    ++ptr;  // consume '%'
+    if (ptr == end) FMT_THROW(format_error("invalid format"));
+    c = *ptr;
+    switch (c) {
+    case '_':
+      pad = pad_type::space;
+      ++ptr;
+      break;
+    case '-':
+      pad = pad_type::none;
+      ++ptr;
+      break;
+    case '0':
+      pad = pad_type::zero;
+      ++ptr;
+      break;
+    }
+    if (ptr == end) FMT_THROW(format_error("invalid format"));
+    c = *ptr++;
+    switch (c) {
+    case '%':
+      handler.on_text(ptr - 1, ptr);
+      break;
+    case 'n': {
+      const Char newline[] = {'\n'};
+      handler.on_text(newline, newline + 1);
+      break;
+    }
+    case 't': {
+      const Char tab[] = {'\t'};
+      handler.on_text(tab, tab + 1);
+      break;
+    }
+    // Year:
+    case 'Y':
+      handler.on_year(numeric_system::standard);
+      break;
+    case 'y':
+      handler.on_short_year(numeric_system::standard);
+      break;
+    case 'C':
+      handler.on_century(numeric_system::standard);
+      break;
+    case 'G':
+      handler.on_iso_week_based_year();
+      break;
+    case 'g':
+      handler.on_iso_week_based_short_year();
+      break;
+    // Day of the week:
+    case 'a':
+      handler.on_abbr_weekday();
+      break;
+    case 'A':
+      handler.on_full_weekday();
+      break;
+    case 'w':
+      handler.on_dec0_weekday(numeric_system::standard);
+      break;
+    case 'u':
+      handler.on_dec1_weekday(numeric_system::standard);
+      break;
+    // Month:
+    case 'b':
+    case 'h':
+      handler.on_abbr_month();
+      break;
+    case 'B':
+      handler.on_full_month();
+      break;
+    case 'm':
+      handler.on_dec_month(numeric_system::standard);
+      break;
+    // Day of the year/month:
+    case 'U':
+      handler.on_dec0_week_of_year(numeric_system::standard);
+      break;
+    case 'W':
+      handler.on_dec1_week_of_year(numeric_system::standard);
+      break;
+    case 'V':
+      handler.on_iso_week_of_year(numeric_system::standard);
+      break;
+    case 'j':
+      handler.on_day_of_year();
+      break;
+    case 'd':
+      handler.on_day_of_month(numeric_system::standard);
+      break;
+    case 'e':
+      handler.on_day_of_month_space(numeric_system::standard);
+      break;
+    // Hour, minute, second:
+    case 'H':
+      handler.on_24_hour(numeric_system::standard, pad);
+      break;
+    case 'I':
+      handler.on_12_hour(numeric_system::standard, pad);
+      break;
+    case 'M':
+      handler.on_minute(numeric_system::standard, pad);
+      break;
+    case 'S':
+      handler.on_second(numeric_system::standard, pad);
+      break;
+    // Other:
+    case 'c':
+      handler.on_datetime(numeric_system::standard);
+      break;
+    case 'x':
+      handler.on_loc_date(numeric_system::standard);
+      break;
+    case 'X':
+      handler.on_loc_time(numeric_system::standard);
+      break;
+    case 'D':
+      handler.on_us_date();
+      break;
+    case 'F':
+      handler.on_iso_date();
+      break;
+    case 'r':
+      handler.on_12_hour_time();
+      break;
+    case 'R':
+      handler.on_24_hour_time();
+      break;
+    case 'T':
+      handler.on_iso_time();
+      break;
+    case 'p':
+      handler.on_am_pm();
+      break;
+    case 'Q':
+      handler.on_duration_value();
+      break;
+    case 'q':
+      handler.on_duration_unit();
+      break;
+    case 'z':
+      handler.on_utc_offset(numeric_system::standard);
+      break;
+    case 'Z':
+      handler.on_tz_name();
+      break;
+    // Alternative representation:
+    case 'E': {
+      if (ptr == end) FMT_THROW(format_error("invalid format"));
+      c = *ptr++;
+      switch (c) {
+      case 'Y':
+        handler.on_year(numeric_system::alternative);
+        break;
+      case 'y':
+        handler.on_offset_year();
+        break;
+      case 'C':
+        handler.on_century(numeric_system::alternative);
+        break;
+      case 'c':
+        handler.on_datetime(numeric_system::alternative);
+        break;
+      case 'x':
+        handler.on_loc_date(numeric_system::alternative);
+        break;
+      case 'X':
+        handler.on_loc_time(numeric_system::alternative);
+        break;
+      case 'z':
+        handler.on_utc_offset(numeric_system::alternative);
+        break;
+      default:
+        FMT_THROW(format_error("invalid format"));
+      }
+      break;
+    }
+    case 'O':
+      if (ptr == end) FMT_THROW(format_error("invalid format"));
+      c = *ptr++;
+      switch (c) {
+      case 'y':
+        handler.on_short_year(numeric_system::alternative);
+        break;
+      case 'm':
+        handler.on_dec_month(numeric_system::alternative);
+        break;
+      case 'U':
+        handler.on_dec0_week_of_year(numeric_system::alternative);
+        break;
+      case 'W':
+        handler.on_dec1_week_of_year(numeric_system::alternative);
+        break;
+      case 'V':
+        handler.on_iso_week_of_year(numeric_system::alternative);
+        break;
+      case 'd':
+        handler.on_day_of_month(numeric_system::alternative);
+        break;
+      case 'e':
+        handler.on_day_of_month_space(numeric_system::alternative);
+        break;
+      case 'w':
+        handler.on_dec0_weekday(numeric_system::alternative);
+        break;
+      case 'u':
+        handler.on_dec1_weekday(numeric_system::alternative);
+        break;
+      case 'H':
+        handler.on_24_hour(numeric_system::alternative, pad);
+        break;
+      case 'I':
+        handler.on_12_hour(numeric_system::alternative, pad);
+        break;
+      case 'M':
+        handler.on_minute(numeric_system::alternative, pad);
+        break;
+      case 'S':
+        handler.on_second(numeric_system::alternative, pad);
+        break;
+      case 'z':
+        handler.on_utc_offset(numeric_system::alternative);
+        break;
+      default:
+        FMT_THROW(format_error("invalid format"));
+      }
+      break;
+    default:
+      FMT_THROW(format_error("invalid format"));
+    }
+    begin = ptr;
+  }
+  if (begin != ptr) handler.on_text(begin, ptr);
+  return ptr;
+}
+
+template <typename Derived> struct null_chrono_spec_handler {
+  FMT_CONSTEXPR void unsupported() {
+    static_cast<Derived*>(this)->unsupported();
+  }
+  FMT_CONSTEXPR void on_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_short_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_offset_year() { unsupported(); }
+  FMT_CONSTEXPR void on_century(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_iso_week_based_year() { unsupported(); }
+  FMT_CONSTEXPR void on_iso_week_based_short_year() { unsupported(); }
+  FMT_CONSTEXPR void on_abbr_weekday() { unsupported(); }
+  FMT_CONSTEXPR void on_full_weekday() { unsupported(); }
+  FMT_CONSTEXPR void on_dec0_weekday(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_dec1_weekday(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_abbr_month() { unsupported(); }
+  FMT_CONSTEXPR void on_full_month() { unsupported(); }
+  FMT_CONSTEXPR void on_dec_month(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_dec0_week_of_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_dec1_week_of_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_iso_week_of_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_day_of_year() { unsupported(); }
+  FMT_CONSTEXPR void on_day_of_month(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_day_of_month_space(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_24_hour(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_12_hour(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_minute(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_second(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_datetime(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_loc_date(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_loc_time(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_us_date() { unsupported(); }
+  FMT_CONSTEXPR void on_iso_date() { unsupported(); }
+  FMT_CONSTEXPR void on_12_hour_time() { unsupported(); }
+  FMT_CONSTEXPR void on_24_hour_time() { unsupported(); }
+  FMT_CONSTEXPR void on_iso_time() { unsupported(); }
+  FMT_CONSTEXPR void on_am_pm() { unsupported(); }
+  FMT_CONSTEXPR void on_duration_value() { unsupported(); }
+  FMT_CONSTEXPR void on_duration_unit() { unsupported(); }
+  FMT_CONSTEXPR void on_utc_offset(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_tz_name() { unsupported(); }
+};
+
+struct tm_format_checker : null_chrono_spec_handler<tm_format_checker> {
+  FMT_NORETURN void unsupported() { FMT_THROW(format_error("no format")); }
+
+  template <typename Char>
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+  FMT_CONSTEXPR void on_year(numeric_system) {}
+  FMT_CONSTEXPR void on_short_year(numeric_system) {}
+  FMT_CONSTEXPR void on_offset_year() {}
+  FMT_CONSTEXPR void on_century(numeric_system) {}
+  FMT_CONSTEXPR void on_iso_week_based_year() {}
+  FMT_CONSTEXPR void on_iso_week_based_short_year() {}
+  FMT_CONSTEXPR void on_abbr_weekday() {}
+  FMT_CONSTEXPR void on_full_weekday() {}
+  FMT_CONSTEXPR void on_dec0_weekday(numeric_system) {}
+  FMT_CONSTEXPR void on_dec1_weekday(numeric_system) {}
+  FMT_CONSTEXPR void on_abbr_month() {}
+  FMT_CONSTEXPR void on_full_month() {}
+  FMT_CONSTEXPR void on_dec_month(numeric_system) {}
+  FMT_CONSTEXPR void on_dec0_week_of_year(numeric_system) {}
+  FMT_CONSTEXPR void on_dec1_week_of_year(numeric_system) {}
+  FMT_CONSTEXPR void on_iso_week_of_year(numeric_system) {}
+  FMT_CONSTEXPR void on_day_of_year() {}
+  FMT_CONSTEXPR void on_day_of_month(numeric_system) {}
+  FMT_CONSTEXPR void on_day_of_month_space(numeric_system) {}
+  FMT_CONSTEXPR void on_24_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_12_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_minute(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_second(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_datetime(numeric_system) {}
+  FMT_CONSTEXPR void on_loc_date(numeric_system) {}
+  FMT_CONSTEXPR void on_loc_time(numeric_system) {}
+  FMT_CONSTEXPR void on_us_date() {}
+  FMT_CONSTEXPR void on_iso_date() {}
+  FMT_CONSTEXPR void on_12_hour_time() {}
+  FMT_CONSTEXPR void on_24_hour_time() {}
+  FMT_CONSTEXPR void on_iso_time() {}
+  FMT_CONSTEXPR void on_am_pm() {}
+  FMT_CONSTEXPR void on_utc_offset(numeric_system) {}
+  FMT_CONSTEXPR void on_tz_name() {}
+};
+
+inline auto tm_wday_full_name(int wday) -> const char* {
+  static constexpr const char* full_name_list[] = {
+      "Sunday",   "Monday", "Tuesday", "Wednesday",
+      "Thursday", "Friday", "Saturday"};
+  return wday >= 0 && wday <= 6 ? full_name_list[wday] : "?";
+}
+inline auto tm_wday_short_name(int wday) -> const char* {
+  static constexpr const char* short_name_list[] = {"Sun", "Mon", "Tue", "Wed",
+                                                    "Thu", "Fri", "Sat"};
+  return wday >= 0 && wday <= 6 ? short_name_list[wday] : "???";
+}
+
+inline auto tm_mon_full_name(int mon) -> const char* {
+  static constexpr const char* full_name_list[] = {
+      "January", "February", "March",     "April",   "May",      "June",
+      "July",    "August",   "September", "October", "November", "December"};
+  return mon >= 0 && mon <= 11 ? full_name_list[mon] : "?";
+}
+inline auto tm_mon_short_name(int mon) -> const char* {
+  static constexpr const char* short_name_list[] = {
+      "Jan", "Feb", "Mar", "Apr", "May", "Jun",
+      "Jul", "Aug", "Sep", "Oct", "Nov", "Dec",
+  };
+  return mon >= 0 && mon <= 11 ? short_name_list[mon] : "???";
+}
+
+template <typename T, typename = void>
+struct has_member_data_tm_gmtoff : std::false_type {};
+template <typename T>
+struct has_member_data_tm_gmtoff<T, void_t<decltype(T::tm_gmtoff)>>
+    : std::true_type {};
+
+template <typename T, typename = void>
+struct has_member_data_tm_zone : std::false_type {};
+template <typename T>
+struct has_member_data_tm_zone<T, void_t<decltype(T::tm_zone)>>
+    : std::true_type {};
+
+#if FMT_USE_TZSET
+inline void tzset_once() {
+  static bool init = []() -> bool {
+    _tzset();
+    return true;
+  }();
+  ignore_unused(init);
+}
+#endif
+
+// Converts value to Int and checks that it's in the range [0, upper).
+template <typename T, typename Int, FMT_ENABLE_IF(std::is_integral<T>::value)>
+inline auto to_nonnegative_int(T value, Int upper) -> Int {
+  if (!std::is_unsigned<Int>::value &&
+      (value < 0 || to_unsigned(value) > to_unsigned(upper))) {
+    FMT_THROW(fmt::format_error("chrono value is out of range"));
+  }
+  return static_cast<Int>(value);
+}
+template <typename T, typename Int, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+inline auto to_nonnegative_int(T value, Int upper) -> Int {
+  if (value < 0 || value > static_cast<T>(upper))
+    FMT_THROW(format_error("invalid value"));
+  return static_cast<Int>(value);
+}
+
+constexpr auto pow10(std::uint32_t n) -> long long {
+  return n == 0 ? 1 : 10 * pow10(n - 1);
+}
+
+// Counts the number of fractional digits in the range [0, 18] according to the
+// C++20 spec. If more than 18 fractional digits are required then returns 6 for
+// microseconds precision.
+template <long long Num, long long Den, int N = 0,
+          bool Enabled = (N < 19) && (Num <= max_value<long long>() / 10)>
+struct count_fractional_digits {
+  static constexpr int value =
+      Num % Den == 0 ? N : count_fractional_digits<Num * 10, Den, N + 1>::value;
+};
+
+// Base case that doesn't instantiate any more templates
+// in order to avoid overflow.
+template <long long Num, long long Den, int N>
+struct count_fractional_digits<Num, Den, N, false> {
+  static constexpr int value = (Num % Den == 0) ? N : 6;
+};
+
+// Format subseconds which are given as an integer type with an appropriate
+// number of digits.
+template <typename Char, typename OutputIt, typename Duration>
+void write_fractional_seconds(OutputIt& out, Duration d, int precision = -1) {
+  constexpr auto num_fractional_digits =
+      count_fractional_digits<Duration::period::num,
+                              Duration::period::den>::value;
+
+  using subsecond_precision = std::chrono::duration<
+      typename std::common_type<typename Duration::rep,
+                                std::chrono::seconds::rep>::type,
+      std::ratio<1, detail::pow10(num_fractional_digits)>>;
+
+  const auto fractional = d - fmt_duration_cast<std::chrono::seconds>(d);
+  const auto subseconds =
+      std::chrono::treat_as_floating_point<
+          typename subsecond_precision::rep>::value
+          ? fractional.count()
+          : fmt_duration_cast<subsecond_precision>(fractional).count();
+  auto n = static_cast<uint32_or_64_or_128_t<long long>>(subseconds);
+  const int num_digits = detail::count_digits(n);
+
+  int leading_zeroes = (std::max)(0, num_fractional_digits - num_digits);
+  if (precision < 0) {
+    FMT_ASSERT(!std::is_floating_point<typename Duration::rep>::value, "");
+    if (std::ratio_less<typename subsecond_precision::period,
+                        std::chrono::seconds::period>::value) {
+      *out++ = '.';
+      out = std::fill_n(out, leading_zeroes, '0');
+      out = format_decimal<Char>(out, n, num_digits).end;
+    }
+  } else {
+    *out++ = '.';
+    leading_zeroes = (std::min)(leading_zeroes, precision);
+    out = std::fill_n(out, leading_zeroes, '0');
+    int remaining = precision - leading_zeroes;
+    if (remaining != 0 && remaining < num_digits) {
+      n /= to_unsigned(detail::pow10(to_unsigned(num_digits - remaining)));
+      out = format_decimal<Char>(out, n, remaining).end;
+      return;
+    }
+    out = format_decimal<Char>(out, n, num_digits).end;
+    remaining -= num_digits;
+    out = std::fill_n(out, remaining, '0');
+  }
+}
+
+// Format subseconds which are given as a floating point type with an
+// appropriate number of digits. We cannot pass the Duration here, as we
+// explicitly need to pass the Rep value in the chrono_formatter.
+template <typename Duration>
+void write_floating_seconds(memory_buffer& buf, Duration duration,
+                            int num_fractional_digits = -1) {
+  using rep = typename Duration::rep;
+  FMT_ASSERT(std::is_floating_point<rep>::value, "");
+
+  auto val = duration.count();
+
+  if (num_fractional_digits < 0) {
+    // For `std::round` with fallback to `round`:
+    // On some toolchains `std::round` is not available (e.g. GCC 6).
+    using namespace std;
+    num_fractional_digits =
+        count_fractional_digits<Duration::period::num,
+                                Duration::period::den>::value;
+    if (num_fractional_digits < 6 && static_cast<rep>(round(val)) != val)
+      num_fractional_digits = 6;
+  }
+
+  fmt::format_to(std::back_inserter(buf), FMT_STRING("{:.{}f}"),
+                 std::fmod(val * static_cast<rep>(Duration::period::num) /
+                               static_cast<rep>(Duration::period::den),
+                           static_cast<rep>(60)),
+                 num_fractional_digits);
+}
+
+template <typename OutputIt, typename Char,
+          typename Duration = std::chrono::seconds>
+class tm_writer {
+ private:
+  static constexpr int days_per_week = 7;
+
+  const std::locale& loc_;
+  const bool is_classic_;
+  OutputIt out_;
+  const Duration* subsecs_;
+  const std::tm& tm_;
+
+  auto tm_sec() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_sec >= 0 && tm_.tm_sec <= 61, "");
+    return tm_.tm_sec;
+  }
+  auto tm_min() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_min >= 0 && tm_.tm_min <= 59, "");
+    return tm_.tm_min;
+  }
+  auto tm_hour() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_hour >= 0 && tm_.tm_hour <= 23, "");
+    return tm_.tm_hour;
+  }
+  auto tm_mday() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_mday >= 1 && tm_.tm_mday <= 31, "");
+    return tm_.tm_mday;
+  }
+  auto tm_mon() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_mon >= 0 && tm_.tm_mon <= 11, "");
+    return tm_.tm_mon;
+  }
+  auto tm_year() const noexcept -> long long { return 1900ll + tm_.tm_year; }
+  auto tm_wday() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_wday >= 0 && tm_.tm_wday <= 6, "");
+    return tm_.tm_wday;
+  }
+  auto tm_yday() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_yday >= 0 && tm_.tm_yday <= 365, "");
+    return tm_.tm_yday;
+  }
+
+  auto tm_hour12() const noexcept -> int {
+    const auto h = tm_hour();
+    const auto z = h < 12 ? h : h - 12;
+    return z == 0 ? 12 : z;
+  }
+
+  // POSIX and the C Standard are unclear or inconsistent about what %C and %y
+  // do if the year is negative or exceeds 9999. Use the convention that %C
+  // concatenated with %y yields the same output as %Y, and that %Y contains at
+  // least 4 characters, with more only if necessary.
+  auto split_year_lower(long long year) const noexcept -> int {
+    auto l = year % 100;
+    if (l < 0) l = -l;  // l in [0, 99]
+    return static_cast<int>(l);
+  }
+
+  // Algorithm: https://en.wikipedia.org/wiki/ISO_week_date.
+  auto iso_year_weeks(long long curr_year) const noexcept -> int {
+    const auto prev_year = curr_year - 1;
+    const auto curr_p =
+        (curr_year + curr_year / 4 - curr_year / 100 + curr_year / 400) %
+        days_per_week;
+    const auto prev_p =
+        (prev_year + prev_year / 4 - prev_year / 100 + prev_year / 400) %
+        days_per_week;
+    return 52 + ((curr_p == 4 || prev_p == 3) ? 1 : 0);
+  }
+  auto iso_week_num(int tm_yday, int tm_wday) const noexcept -> int {
+    return (tm_yday + 11 - (tm_wday == 0 ? days_per_week : tm_wday)) /
+           days_per_week;
+  }
+  auto tm_iso_week_year() const noexcept -> long long {
+    const auto year = tm_year();
+    const auto w = iso_week_num(tm_yday(), tm_wday());
+    if (w < 1) return year - 1;
+    if (w > iso_year_weeks(year)) return year + 1;
+    return year;
+  }
+  auto tm_iso_week_of_year() const noexcept -> int {
+    const auto year = tm_year();
+    const auto w = iso_week_num(tm_yday(), tm_wday());
+    if (w < 1) return iso_year_weeks(year - 1);
+    if (w > iso_year_weeks(year)) return 1;
+    return w;
+  }
+
+  void write1(int value) {
+    *out_++ = static_cast<char>('0' + to_unsigned(value) % 10);
+  }
+  void write2(int value) {
+    const char* d = digits2(to_unsigned(value) % 100);
+    *out_++ = *d++;
+    *out_++ = *d;
+  }
+  void write2(int value, pad_type pad) {
+    unsigned int v = to_unsigned(value) % 100;
+    if (v >= 10) {
+      const char* d = digits2(v);
+      *out_++ = *d++;
+      *out_++ = *d;
+    } else {
+      out_ = detail::write_padding(out_, pad);
+      *out_++ = static_cast<char>('0' + v);
+    }
+  }
+
+  void write_year_extended(long long year) {
+    // At least 4 characters.
+    int width = 4;
+    if (year < 0) {
+      *out_++ = '-';
+      year = 0 - year;
+      --width;
+    }
+    uint32_or_64_or_128_t<long long> n = to_unsigned(year);
+    const int num_digits = count_digits(n);
+    if (width > num_digits) out_ = std::fill_n(out_, width - num_digits, '0');
+    out_ = format_decimal<Char>(out_, n, num_digits).end;
+  }
+  void write_year(long long year) {
+    if (year >= 0 && year < 10000) {
+      write2(static_cast<int>(year / 100));
+      write2(static_cast<int>(year % 100));
+    } else {
+      write_year_extended(year);
+    }
+  }
+
+  void write_utc_offset(long offset, numeric_system ns) {
+    if (offset < 0) {
+      *out_++ = '-';
+      offset = -offset;
+    } else {
+      *out_++ = '+';
+    }
+    offset /= 60;
+    write2(static_cast<int>(offset / 60));
+    if (ns != numeric_system::standard) *out_++ = ':';
+    write2(static_cast<int>(offset % 60));
+  }
+  template <typename T, FMT_ENABLE_IF(has_member_data_tm_gmtoff<T>::value)>
+  void format_utc_offset_impl(const T& tm, numeric_system ns) {
+    write_utc_offset(tm.tm_gmtoff, ns);
+  }
+  template <typename T, FMT_ENABLE_IF(!has_member_data_tm_gmtoff<T>::value)>
+  void format_utc_offset_impl(const T& tm, numeric_system ns) {
+#if defined(_WIN32) && defined(_UCRT)
+#  if FMT_USE_TZSET
+    tzset_once();
+#  endif
+    long offset = 0;
+    _get_timezone(&offset);
+    if (tm.tm_isdst) {
+      long dstbias = 0;
+      _get_dstbias(&dstbias);
+      offset += dstbias;
+    }
+    write_utc_offset(-offset, ns);
+#else
+    if (ns == numeric_system::standard) return format_localized('z');
+
+    // Extract timezone offset from timezone conversion functions.
+    std::tm gtm = tm;
+    std::time_t gt = std::mktime(&gtm);
+    std::tm ltm = gmtime(gt);
+    std::time_t lt = std::mktime(&ltm);
+    long offset = gt - lt;
+    write_utc_offset(offset, ns);
+#endif
+  }
+
+  template <typename T, FMT_ENABLE_IF(has_member_data_tm_zone<T>::value)>
+  void format_tz_name_impl(const T& tm) {
+    if (is_classic_)
+      out_ = write_tm_str<Char>(out_, tm.tm_zone, loc_);
+    else
+      format_localized('Z');
+  }
+  template <typename T, FMT_ENABLE_IF(!has_member_data_tm_zone<T>::value)>
+  void format_tz_name_impl(const T&) {
+    format_localized('Z');
+  }
+
+  void format_localized(char format, char modifier = 0) {
+    out_ = write<Char>(out_, tm_, loc_, format, modifier);
+  }
+
+ public:
+  tm_writer(const std::locale& loc, OutputIt out, const std::tm& tm,
+            const Duration* subsecs = nullptr)
+      : loc_(loc),
+        is_classic_(loc_ == get_classic_locale()),
+        out_(out),
+        subsecs_(subsecs),
+        tm_(tm) {}
+
+  auto out() const -> OutputIt { return out_; }
+
+  FMT_CONSTEXPR void on_text(const Char* begin, const Char* end) {
+    out_ = copy_str<Char>(begin, end, out_);
+  }
+
+  void on_abbr_weekday() {
+    if (is_classic_)
+      out_ = write(out_, tm_wday_short_name(tm_wday()));
+    else
+      format_localized('a');
+  }
+  void on_full_weekday() {
+    if (is_classic_)
+      out_ = write(out_, tm_wday_full_name(tm_wday()));
+    else
+      format_localized('A');
+  }
+  void on_dec0_weekday(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) return write1(tm_wday());
+    format_localized('w', 'O');
+  }
+  void on_dec1_weekday(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto wday = tm_wday();
+      write1(wday == 0 ? days_per_week : wday);
+    } else {
+      format_localized('u', 'O');
+    }
+  }
+
+  void on_abbr_month() {
+    if (is_classic_)
+      out_ = write(out_, tm_mon_short_name(tm_mon()));
+    else
+      format_localized('b');
+  }
+  void on_full_month() {
+    if (is_classic_)
+      out_ = write(out_, tm_mon_full_name(tm_mon()));
+    else
+      format_localized('B');
+  }
+
+  void on_datetime(numeric_system ns) {
+    if (is_classic_) {
+      on_abbr_weekday();
+      *out_++ = ' ';
+      on_abbr_month();
+      *out_++ = ' ';
+      on_day_of_month_space(numeric_system::standard);
+      *out_++ = ' ';
+      on_iso_time();
+      *out_++ = ' ';
+      on_year(numeric_system::standard);
+    } else {
+      format_localized('c', ns == numeric_system::standard ? '\0' : 'E');
+    }
+  }
+  void on_loc_date(numeric_system ns) {
+    if (is_classic_)
+      on_us_date();
+    else
+      format_localized('x', ns == numeric_system::standard ? '\0' : 'E');
+  }
+  void on_loc_time(numeric_system ns) {
+    if (is_classic_)
+      on_iso_time();
+    else
+      format_localized('X', ns == numeric_system::standard ? '\0' : 'E');
+  }
+  void on_us_date() {
+    char buf[8];
+    write_digit2_separated(buf, to_unsigned(tm_mon() + 1),
+                           to_unsigned(tm_mday()),
+                           to_unsigned(split_year_lower(tm_year())), '/');
+    out_ = copy_str<Char>(std::begin(buf), std::end(buf), out_);
+  }
+  void on_iso_date() {
+    auto year = tm_year();
+    char buf[10];
+    size_t offset = 0;
+    if (year >= 0 && year < 10000) {
+      copy2(buf, digits2(static_cast<size_t>(year / 100)));
+    } else {
+      offset = 4;
+      write_year_extended(year);
+      year = 0;
+    }
+    write_digit2_separated(buf + 2, static_cast<unsigned>(year % 100),
+                           to_unsigned(tm_mon() + 1), to_unsigned(tm_mday()),
+                           '-');
+    out_ = copy_str<Char>(std::begin(buf) + offset, std::end(buf), out_);
+  }
+
+  void on_utc_offset(numeric_system ns) { format_utc_offset_impl(tm_, ns); }
+  void on_tz_name() { format_tz_name_impl(tm_); }
+
+  void on_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write_year(tm_year());
+    format_localized('Y', 'E');
+  }
+  void on_short_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(split_year_lower(tm_year()));
+    format_localized('y', 'O');
+  }
+  void on_offset_year() {
+    if (is_classic_) return write2(split_year_lower(tm_year()));
+    format_localized('y', 'E');
+  }
+
+  void on_century(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto year = tm_year();
+      auto upper = year / 100;
+      if (year >= -99 && year < 0) {
+        // Zero upper on negative year.
+        *out_++ = '-';
+        *out_++ = '0';
+      } else if (upper >= 0 && upper < 100) {
+        write2(static_cast<int>(upper));
+      } else {
+        out_ = write<Char>(out_, upper);
+      }
+    } else {
+      format_localized('C', 'E');
+    }
+  }
+
+  void on_dec_month(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_mon() + 1);
+    format_localized('m', 'O');
+  }
+
+  void on_dec0_week_of_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2((tm_yday() + days_per_week - tm_wday()) / days_per_week);
+    format_localized('U', 'O');
+  }
+  void on_dec1_week_of_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto wday = tm_wday();
+      write2((tm_yday() + days_per_week -
+              (wday == 0 ? (days_per_week - 1) : (wday - 1))) /
+             days_per_week);
+    } else {
+      format_localized('W', 'O');
+    }
+  }
+  void on_iso_week_of_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_iso_week_of_year());
+    format_localized('V', 'O');
+  }
+
+  void on_iso_week_based_year() { write_year(tm_iso_week_year()); }
+  void on_iso_week_based_short_year() {
+    write2(split_year_lower(tm_iso_week_year()));
+  }
+
+  void on_day_of_year() {
+    auto yday = tm_yday() + 1;
+    write1(yday / 100);
+    write2(yday % 100);
+  }
+  void on_day_of_month(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) return write2(tm_mday());
+    format_localized('d', 'O');
+  }
+  void on_day_of_month_space(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto mday = to_unsigned(tm_mday()) % 100;
+      const char* d2 = digits2(mday);
+      *out_++ = mday < 10 ? ' ' : d2[0];
+      *out_++ = d2[1];
+    } else {
+      format_localized('e', 'O');
+    }
+  }
+
+  void on_24_hour(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_hour(), pad);
+    format_localized('H', 'O');
+  }
+  void on_12_hour(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_hour12(), pad);
+    format_localized('I', 'O');
+  }
+  void on_minute(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_min(), pad);
+    format_localized('M', 'O');
+  }
+
+  void on_second(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      write2(tm_sec(), pad);
+      if (subsecs_) {
+        if (std::is_floating_point<typename Duration::rep>::value) {
+          auto buf = memory_buffer();
+          write_floating_seconds(buf, *subsecs_);
+          if (buf.size() > 1) {
+            // Remove the leading "0", write something like ".123".
+            out_ = std::copy(buf.begin() + 1, buf.end(), out_);
+          }
+        } else {
+          write_fractional_seconds<Char>(out_, *subsecs_);
+        }
+      }
+    } else {
+      // Currently no formatting of subseconds when a locale is set.
+      format_localized('S', 'O');
+    }
+  }
+
+  void on_12_hour_time() {
+    if (is_classic_) {
+      char buf[8];
+      write_digit2_separated(buf, to_unsigned(tm_hour12()),
+                             to_unsigned(tm_min()), to_unsigned(tm_sec()), ':');
+      out_ = copy_str<Char>(std::begin(buf), std::end(buf), out_);
+      *out_++ = ' ';
+      on_am_pm();
+    } else {
+      format_localized('r');
+    }
+  }
+  void on_24_hour_time() {
+    write2(tm_hour());
+    *out_++ = ':';
+    write2(tm_min());
+  }
+  void on_iso_time() {
+    on_24_hour_time();
+    *out_++ = ':';
+    on_second(numeric_system::standard, pad_type::unspecified);
+  }
+
+  void on_am_pm() {
+    if (is_classic_) {
+      *out_++ = tm_hour() < 12 ? 'A' : 'P';
+      *out_++ = 'M';
+    } else {
+      format_localized('p');
+    }
+  }
+
+  // These apply to chrono durations but not tm.
+  void on_duration_value() {}
+  void on_duration_unit() {}
+};
+
+struct chrono_format_checker : null_chrono_spec_handler<chrono_format_checker> {
+  bool has_precision_integral = false;
+
+  FMT_NORETURN void unsupported() { FMT_THROW(format_error("no date")); }
+
+  template <typename Char>
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+  FMT_CONSTEXPR void on_day_of_year() {}
+  FMT_CONSTEXPR void on_24_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_12_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_minute(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_second(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_12_hour_time() {}
+  FMT_CONSTEXPR void on_24_hour_time() {}
+  FMT_CONSTEXPR void on_iso_time() {}
+  FMT_CONSTEXPR void on_am_pm() {}
+  FMT_CONSTEXPR void on_duration_value() const {
+    if (has_precision_integral) {
+      FMT_THROW(format_error("precision not allowed for this argument type"));
+    }
+  }
+  FMT_CONSTEXPR void on_duration_unit() {}
+};
+
+template <typename T,
+          FMT_ENABLE_IF(std::is_integral<T>::value&& has_isfinite<T>::value)>
+inline auto isfinite(T) -> bool {
+  return true;
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+inline auto mod(T x, int y) -> T {
+  return x % static_cast<T>(y);
+}
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+inline auto mod(T x, int y) -> T {
+  return std::fmod(x, static_cast<T>(y));
+}
+
+// If T is an integral type, maps T to its unsigned counterpart, otherwise
+// leaves it unchanged (unlike std::make_unsigned).
+template <typename T, bool INTEGRAL = std::is_integral<T>::value>
+struct make_unsigned_or_unchanged {
+  using type = T;
+};
+
+template <typename T> struct make_unsigned_or_unchanged<T, true> {
+  using type = typename std::make_unsigned<T>::type;
+};
+
+template <typename Rep, typename Period,
+          FMT_ENABLE_IF(std::is_integral<Rep>::value)>
+inline auto get_milliseconds(std::chrono::duration<Rep, Period> d)
+    -> std::chrono::duration<Rep, std::milli> {
+  // this may overflow and/or the result may not fit in the
+  // target type.
+#if FMT_SAFE_DURATION_CAST
+  using CommonSecondsType =
+      typename std::common_type<decltype(d), std::chrono::seconds>::type;
+  const auto d_as_common = fmt_duration_cast<CommonSecondsType>(d);
+  const auto d_as_whole_seconds =
+      fmt_duration_cast<std::chrono::seconds>(d_as_common);
+  // this conversion should be nonproblematic
+  const auto diff = d_as_common - d_as_whole_seconds;
+  const auto ms =
+      fmt_duration_cast<std::chrono::duration<Rep, std::milli>>(diff);
+  return ms;
+#else
+  auto s = fmt_duration_cast<std::chrono::seconds>(d);
+  return fmt_duration_cast<std::chrono::milliseconds>(d - s);
+#endif
+}
+
+template <typename Char, typename Rep, typename OutputIt,
+          FMT_ENABLE_IF(std::is_integral<Rep>::value)>
+auto format_duration_value(OutputIt out, Rep val, int) -> OutputIt {
+  return write<Char>(out, val);
+}
+
+template <typename Char, typename Rep, typename OutputIt,
+          FMT_ENABLE_IF(std::is_floating_point<Rep>::value)>
+auto format_duration_value(OutputIt out, Rep val, int precision) -> OutputIt {
+  auto specs = format_specs<Char>();
+  specs.precision = precision;
+  specs.type = precision >= 0 ? presentation_type::fixed_lower
+                              : presentation_type::general_lower;
+  return write<Char>(out, val, specs);
+}
+
+template <typename Char, typename OutputIt>
+auto copy_unit(string_view unit, OutputIt out, Char) -> OutputIt {
+  return std::copy(unit.begin(), unit.end(), out);
+}
+
+template <typename OutputIt>
+auto copy_unit(string_view unit, OutputIt out, wchar_t) -> OutputIt {
+  // This works when wchar_t is UTF-32 because units only contain characters
+  // that have the same representation in UTF-16 and UTF-32.
+  utf8_to_utf16 u(unit);
+  return std::copy(u.c_str(), u.c_str() + u.size(), out);
+}
+
+template <typename Char, typename Period, typename OutputIt>
+auto format_duration_unit(OutputIt out) -> OutputIt {
+  if (const char* unit = get_units<Period>())
+    return copy_unit(string_view(unit), out, Char());
+  *out++ = '[';
+  out = write<Char>(out, Period::num);
+  if (const_check(Period::den != 1)) {
+    *out++ = '/';
+    out = write<Char>(out, Period::den);
+  }
+  *out++ = ']';
+  *out++ = 's';
+  return out;
+}
+
+class get_locale {
+ private:
+  union {
+    std::locale locale_;
+  };
+  bool has_locale_ = false;
+
+ public:
+  get_locale(bool localized, locale_ref loc) : has_locale_(localized) {
+    if (localized)
+      ::new (&locale_) std::locale(loc.template get<std::locale>());
+  }
+  ~get_locale() {
+    if (has_locale_) locale_.~locale();
+  }
+  operator const std::locale&() const {
+    return has_locale_ ? locale_ : get_classic_locale();
+  }
+};
+
+template <typename FormatContext, typename OutputIt, typename Rep,
+          typename Period>
+struct chrono_formatter {
+  FormatContext& context;
+  OutputIt out;
+  int precision;
+  bool localized = false;
+  // rep is unsigned to avoid overflow.
+  using rep =
+      conditional_t<std::is_integral<Rep>::value && sizeof(Rep) < sizeof(int),
+                    unsigned, typename make_unsigned_or_unchanged<Rep>::type>;
+  rep val;
+  using seconds = std::chrono::duration<rep>;
+  seconds s;
+  using milliseconds = std::chrono::duration<rep, std::milli>;
+  bool negative;
+
+  using char_type = typename FormatContext::char_type;
+  using tm_writer_type = tm_writer<OutputIt, char_type>;
+
+  chrono_formatter(FormatContext& ctx, OutputIt o,
+                   std::chrono::duration<Rep, Period> d)
+      : context(ctx),
+        out(o),
+        val(static_cast<rep>(d.count())),
+        negative(false) {
+    if (d.count() < 0) {
+      val = 0 - val;
+      negative = true;
+    }
+
+    // this may overflow and/or the result may not fit in the
+    // target type.
+    // might need checked conversion (rep!=Rep)
+    s = fmt_duration_cast<seconds>(std::chrono::duration<rep, Period>(val));
+  }
+
+  // returns true if nan or inf, writes to out.
+  auto handle_nan_inf() -> bool {
+    if (isfinite(val)) {
+      return false;
+    }
+    if (isnan(val)) {
+      write_nan();
+      return true;
+    }
+    // must be +-inf
+    if (val > 0) {
+      write_pinf();
+    } else {
+      write_ninf();
+    }
+    return true;
+  }
+
+  auto days() const -> Rep { return static_cast<Rep>(s.count() / 86400); }
+  auto hour() const -> Rep {
+    return static_cast<Rep>(mod((s.count() / 3600), 24));
+  }
+
+  auto hour12() const -> Rep {
+    Rep hour = static_cast<Rep>(mod((s.count() / 3600), 12));
+    return hour <= 0 ? 12 : hour;
+  }
+
+  auto minute() const -> Rep {
+    return static_cast<Rep>(mod((s.count() / 60), 60));
+  }
+  auto second() const -> Rep { return static_cast<Rep>(mod(s.count(), 60)); }
+
+  auto time() const -> std::tm {
+    auto time = std::tm();
+    time.tm_hour = to_nonnegative_int(hour(), 24);
+    time.tm_min = to_nonnegative_int(minute(), 60);
+    time.tm_sec = to_nonnegative_int(second(), 60);
+    return time;
+  }
+
+  void write_sign() {
+    if (negative) {
+      *out++ = '-';
+      negative = false;
+    }
+  }
+
+  void write(Rep value, int width, pad_type pad = pad_type::unspecified) {
+    write_sign();
+    if (isnan(value)) return write_nan();
+    uint32_or_64_or_128_t<int> n =
+        to_unsigned(to_nonnegative_int(value, max_value<int>()));
+    int num_digits = detail::count_digits(n);
+    if (width > num_digits) {
+      out = detail::write_padding(out, pad, width - num_digits);
+    }
+    out = format_decimal<char_type>(out, n, num_digits).end;
+  }
+
+  void write_nan() { std::copy_n("nan", 3, out); }
+  void write_pinf() { std::copy_n("inf", 3, out); }
+  void write_ninf() { std::copy_n("-inf", 4, out); }
+
+  template <typename Callback, typename... Args>
+  void format_tm(const tm& time, Callback cb, Args... args) {
+    if (isnan(val)) return write_nan();
+    get_locale loc(localized, context.locale());
+    auto w = tm_writer_type(loc, out, time);
+    (w.*cb)(args...);
+    out = w.out();
+  }
+
+  void on_text(const char_type* begin, const char_type* end) {
+    std::copy(begin, end, out);
+  }
+
+  // These are not implemented because durations don't have date information.
+  void on_abbr_weekday() {}
+  void on_full_weekday() {}
+  void on_dec0_weekday(numeric_system) {}
+  void on_dec1_weekday(numeric_system) {}
+  void on_abbr_month() {}
+  void on_full_month() {}
+  void on_datetime(numeric_system) {}
+  void on_loc_date(numeric_system) {}
+  void on_loc_time(numeric_system) {}
+  void on_us_date() {}
+  void on_iso_date() {}
+  void on_utc_offset(numeric_system) {}
+  void on_tz_name() {}
+  void on_year(numeric_system) {}
+  void on_short_year(numeric_system) {}
+  void on_offset_year() {}
+  void on_century(numeric_system) {}
+  void on_iso_week_based_year() {}
+  void on_iso_week_based_short_year() {}
+  void on_dec_month(numeric_system) {}
+  void on_dec0_week_of_year(numeric_system) {}
+  void on_dec1_week_of_year(numeric_system) {}
+  void on_iso_week_of_year(numeric_system) {}
+  void on_day_of_month(numeric_system) {}
+  void on_day_of_month_space(numeric_system) {}
+
+  void on_day_of_year() {
+    if (handle_nan_inf()) return;
+    write(days(), 0);
+  }
+
+  void on_24_hour(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) return write(hour(), 2, pad);
+    auto time = tm();
+    time.tm_hour = to_nonnegative_int(hour(), 24);
+    format_tm(time, &tm_writer_type::on_24_hour, ns, pad);
+  }
+
+  void on_12_hour(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) return write(hour12(), 2, pad);
+    auto time = tm();
+    time.tm_hour = to_nonnegative_int(hour12(), 12);
+    format_tm(time, &tm_writer_type::on_12_hour, ns, pad);
+  }
+
+  void on_minute(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) return write(minute(), 2, pad);
+    auto time = tm();
+    time.tm_min = to_nonnegative_int(minute(), 60);
+    format_tm(time, &tm_writer_type::on_minute, ns, pad);
+  }
+
+  void on_second(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) {
+      if (std::is_floating_point<rep>::value) {
+        auto buf = memory_buffer();
+        write_floating_seconds(buf, std::chrono::duration<rep, Period>(val),
+                               precision);
+        if (negative) *out++ = '-';
+        if (buf.size() < 2 || buf[1] == '.') {
+          out = detail::write_padding(out, pad);
+        }
+        out = std::copy(buf.begin(), buf.end(), out);
+      } else {
+        write(second(), 2, pad);
+        write_fractional_seconds<char_type>(
+            out, std::chrono::duration<rep, Period>(val), precision);
+      }
+      return;
+    }
+    auto time = tm();
+    time.tm_sec = to_nonnegative_int(second(), 60);
+    format_tm(time, &tm_writer_type::on_second, ns, pad);
+  }
+
+  void on_12_hour_time() {
+    if (handle_nan_inf()) return;
+    format_tm(time(), &tm_writer_type::on_12_hour_time);
+  }
+
+  void on_24_hour_time() {
+    if (handle_nan_inf()) {
+      *out++ = ':';
+      handle_nan_inf();
+      return;
+    }
+
+    write(hour(), 2);
+    *out++ = ':';
+    write(minute(), 2);
+  }
+
+  void on_iso_time() {
+    on_24_hour_time();
+    *out++ = ':';
+    if (handle_nan_inf()) return;
+    on_second(numeric_system::standard, pad_type::unspecified);
+  }
+
+  void on_am_pm() {
+    if (handle_nan_inf()) return;
+    format_tm(time(), &tm_writer_type::on_am_pm);
+  }
+
+  void on_duration_value() {
+    if (handle_nan_inf()) return;
+    write_sign();
+    out = format_duration_value<char_type>(out, val, precision);
+  }
+
+  void on_duration_unit() {
+    out = format_duration_unit<char_type, Period>(out);
+  }
+};
+
+}  // namespace detail
+
+#if defined(__cpp_lib_chrono) && __cpp_lib_chrono >= 201907
+using weekday = std::chrono::weekday;
+#else
+// A fallback version of weekday.
+class weekday {
+ private:
+  unsigned char value;
+
+ public:
+  weekday() = default;
+  explicit constexpr weekday(unsigned wd) noexcept
+      : value(static_cast<unsigned char>(wd != 7 ? wd : 0)) {}
+  constexpr auto c_encoding() const noexcept -> unsigned { return value; }
+};
+
+class year_month_day {};
+#endif
+
+// A rudimentary weekday formatter.
+template <typename Char> struct formatter<weekday, Char> {
+ private:
+  bool localized = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto begin = ctx.begin(), end = ctx.end();
+    if (begin != end && *begin == 'L') {
+      ++begin;
+      localized = true;
+    }
+    return begin;
+  }
+
+  template <typename FormatContext>
+  auto format(weekday wd, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto time = std::tm();
+    time.tm_wday = static_cast<int>(wd.c_encoding());
+    detail::get_locale loc(localized, ctx.locale());
+    auto w = detail::tm_writer<decltype(ctx.out()), Char>(loc, ctx.out(), time);
+    w.on_abbr_weekday();
+    return w.out();
+  }
+};
+
+template <typename Rep, typename Period, typename Char>
+struct formatter<std::chrono::duration<Rep, Period>, Char> {
+ private:
+  format_specs<Char> specs_;
+  detail::arg_ref<Char> width_ref_;
+  detail::arg_ref<Char> precision_ref_;
+  bool localized_ = false;
+  basic_string_view<Char> format_str_;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it == end || *it == '}') return it;
+
+    it = detail::parse_align(it, end, specs_);
+    if (it == end) return it;
+
+    it = detail::parse_dynamic_spec(it, end, specs_.width, width_ref_, ctx);
+    if (it == end) return it;
+
+    auto checker = detail::chrono_format_checker();
+    if (*it == '.') {
+      checker.has_precision_integral = !std::is_floating_point<Rep>::value;
+      it = detail::parse_precision(it, end, specs_.precision, precision_ref_,
+                                   ctx);
+    }
+    if (it != end && *it == 'L') {
+      localized_ = true;
+      ++it;
+    }
+    end = detail::parse_chrono_format(it, end, checker);
+    format_str_ = {it, detail::to_unsigned(end - it)};
+    return end;
+  }
+
+  template <typename FormatContext>
+  auto format(std::chrono::duration<Rep, Period> d, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto specs = specs_;
+    auto precision = specs.precision;
+    specs.precision = -1;
+    auto begin = format_str_.begin(), end = format_str_.end();
+    // As a possible future optimization, we could avoid extra copying if width
+    // is not specified.
+    auto buf = basic_memory_buffer<Char>();
+    auto out = std::back_inserter(buf);
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref_,
+                                                       ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(precision,
+                                                           precision_ref_, ctx);
+    if (begin == end || *begin == '}') {
+      out = detail::format_duration_value<Char>(out, d.count(), precision);
+      detail::format_duration_unit<Char, Period>(out);
+    } else {
+      using chrono_formatter =
+          detail::chrono_formatter<FormatContext, decltype(out), Rep, Period>;
+      auto f = chrono_formatter(ctx, out, d);
+      f.precision = precision;
+      f.localized = localized_;
+      detail::parse_chrono_format(begin, end, f);
+    }
+    return detail::write(
+        ctx.out(), basic_string_view<Char>(buf.data(), buf.size()), specs);
+  }
+};
+
+template <typename Char, typename Duration>
+struct formatter<std::chrono::time_point<std::chrono::system_clock, Duration>,
+                 Char> : formatter<std::tm, Char> {
+  FMT_CONSTEXPR formatter() {
+    this->format_str_ = detail::string_literal<Char, '%', 'F', ' ', '%', 'T'>{};
+  }
+
+  template <typename FormatContext>
+  auto format(std::chrono::time_point<std::chrono::system_clock, Duration> val,
+              FormatContext& ctx) const -> decltype(ctx.out()) {
+    using period = typename Duration::period;
+    if (detail::const_check(
+            period::num != 1 || period::den != 1 ||
+            std::is_floating_point<typename Duration::rep>::value)) {
+      const auto epoch = val.time_since_epoch();
+      auto subsecs = detail::fmt_duration_cast<Duration>(
+          epoch - detail::fmt_duration_cast<std::chrono::seconds>(epoch));
+
+      if (subsecs.count() < 0) {
+        auto second =
+            detail::fmt_duration_cast<Duration>(std::chrono::seconds(1));
+        if (epoch.count() < ((Duration::min)() + second).count())
+          FMT_THROW(format_error("duration is too small"));
+        subsecs += second;
+        val -= second;
+      }
+
+      return formatter<std::tm, Char>::do_format(gmtime(val), ctx, &subsecs);
+    }
+
+    return formatter<std::tm, Char>::format(gmtime(val), ctx);
+  }
+};
+
+#if FMT_USE_LOCAL_TIME
+template <typename Char, typename Duration>
+struct formatter<std::chrono::local_time<Duration>, Char>
+    : formatter<std::tm, Char> {
+  FMT_CONSTEXPR formatter() {
+    this->format_str_ = detail::string_literal<Char, '%', 'F', ' ', '%', 'T'>{};
+  }
+
+  template <typename FormatContext>
+  auto format(std::chrono::local_time<Duration> val, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    using period = typename Duration::period;
+    if (period::num != 1 || period::den != 1 ||
+        std::is_floating_point<typename Duration::rep>::value) {
+      const auto epoch = val.time_since_epoch();
+      const auto subsecs = detail::fmt_duration_cast<Duration>(
+          epoch - detail::fmt_duration_cast<std::chrono::seconds>(epoch));
+
+      return formatter<std::tm, Char>::do_format(localtime(val), ctx, &subsecs);
+    }
+
+    return formatter<std::tm, Char>::format(localtime(val), ctx);
+  }
+};
+#endif
+
+#if FMT_USE_UTC_TIME
+template <typename Char, typename Duration>
+struct formatter<std::chrono::time_point<std::chrono::utc_clock, Duration>,
+                 Char>
+    : formatter<std::chrono::time_point<std::chrono::system_clock, Duration>,
+                Char> {
+  template <typename FormatContext>
+  auto format(std::chrono::time_point<std::chrono::utc_clock, Duration> val,
+              FormatContext& ctx) const -> decltype(ctx.out()) {
+    return formatter<
+        std::chrono::time_point<std::chrono::system_clock, Duration>,
+        Char>::format(std::chrono::utc_clock::to_sys(val), ctx);
+  }
+};
+#endif
+
+template <typename Char> struct formatter<std::tm, Char> {
+ private:
+  format_specs<Char> specs_;
+  detail::arg_ref<Char> width_ref_;
+
+ protected:
+  basic_string_view<Char> format_str_;
+
+  template <typename FormatContext, typename Duration>
+  auto do_format(const std::tm& tm, FormatContext& ctx,
+                 const Duration* subsecs) const -> decltype(ctx.out()) {
+    auto specs = specs_;
+    auto buf = basic_memory_buffer<Char>();
+    auto out = std::back_inserter(buf);
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref_,
+                                                       ctx);
+
+    auto loc_ref = ctx.locale();
+    detail::get_locale loc(static_cast<bool>(loc_ref), loc_ref);
+    auto w =
+        detail::tm_writer<decltype(out), Char, Duration>(loc, out, tm, subsecs);
+    detail::parse_chrono_format(format_str_.begin(), format_str_.end(), w);
+    return detail::write(
+        ctx.out(), basic_string_view<Char>(buf.data(), buf.size()), specs);
+  }
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it == end || *it == '}') return it;
+
+    it = detail::parse_align(it, end, specs_);
+    if (it == end) return it;
+
+    it = detail::parse_dynamic_spec(it, end, specs_.width, width_ref_, ctx);
+    if (it == end) return it;
+
+    end = detail::parse_chrono_format(it, end, detail::tm_format_checker());
+    // Replace the default format_str only if the new spec is not empty.
+    if (end != it) format_str_ = {it, detail::to_unsigned(end - it)};
+    return end;
+  }
+
+  template <typename FormatContext>
+  auto format(const std::tm& tm, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return do_format<FormatContext, std::chrono::seconds>(tm, ctx, nullptr);
+  }
+};
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_CHRONO_H_
diff --git a/dep/fmt/include/fmt/color.h b/dep/fmt/include/fmt/color.h
new file mode 100644
index 00000000..367849a8
--- /dev/null
+++ b/dep/fmt/include/fmt/color.h
@@ -0,0 +1,643 @@
+// Formatting library for C++ - color support
+//
+// Copyright (c) 2018 - present, Victor Zverovich and fmt contributors
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_COLOR_H_
+#define FMT_COLOR_H_
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+FMT_BEGIN_EXPORT
+
+enum class color : uint32_t {
+  alice_blue = 0xF0F8FF,               // rgb(240,248,255)
+  antique_white = 0xFAEBD7,            // rgb(250,235,215)
+  aqua = 0x00FFFF,                     // rgb(0,255,255)
+  aquamarine = 0x7FFFD4,               // rgb(127,255,212)
+  azure = 0xF0FFFF,                    // rgb(240,255,255)
+  beige = 0xF5F5DC,                    // rgb(245,245,220)
+  bisque = 0xFFE4C4,                   // rgb(255,228,196)
+  black = 0x000000,                    // rgb(0,0,0)
+  blanched_almond = 0xFFEBCD,          // rgb(255,235,205)
+  blue = 0x0000FF,                     // rgb(0,0,255)
+  blue_violet = 0x8A2BE2,              // rgb(138,43,226)
+  brown = 0xA52A2A,                    // rgb(165,42,42)
+  burly_wood = 0xDEB887,               // rgb(222,184,135)
+  cadet_blue = 0x5F9EA0,               // rgb(95,158,160)
+  chartreuse = 0x7FFF00,               // rgb(127,255,0)
+  chocolate = 0xD2691E,                // rgb(210,105,30)
+  coral = 0xFF7F50,                    // rgb(255,127,80)
+  cornflower_blue = 0x6495ED,          // rgb(100,149,237)
+  cornsilk = 0xFFF8DC,                 // rgb(255,248,220)
+  crimson = 0xDC143C,                  // rgb(220,20,60)
+  cyan = 0x00FFFF,                     // rgb(0,255,255)
+  dark_blue = 0x00008B,                // rgb(0,0,139)
+  dark_cyan = 0x008B8B,                // rgb(0,139,139)
+  dark_golden_rod = 0xB8860B,          // rgb(184,134,11)
+  dark_gray = 0xA9A9A9,                // rgb(169,169,169)
+  dark_green = 0x006400,               // rgb(0,100,0)
+  dark_khaki = 0xBDB76B,               // rgb(189,183,107)
+  dark_magenta = 0x8B008B,             // rgb(139,0,139)
+  dark_olive_green = 0x556B2F,         // rgb(85,107,47)
+  dark_orange = 0xFF8C00,              // rgb(255,140,0)
+  dark_orchid = 0x9932CC,              // rgb(153,50,204)
+  dark_red = 0x8B0000,                 // rgb(139,0,0)
+  dark_salmon = 0xE9967A,              // rgb(233,150,122)
+  dark_sea_green = 0x8FBC8F,           // rgb(143,188,143)
+  dark_slate_blue = 0x483D8B,          // rgb(72,61,139)
+  dark_slate_gray = 0x2F4F4F,          // rgb(47,79,79)
+  dark_turquoise = 0x00CED1,           // rgb(0,206,209)
+  dark_violet = 0x9400D3,              // rgb(148,0,211)
+  deep_pink = 0xFF1493,                // rgb(255,20,147)
+  deep_sky_blue = 0x00BFFF,            // rgb(0,191,255)
+  dim_gray = 0x696969,                 // rgb(105,105,105)
+  dodger_blue = 0x1E90FF,              // rgb(30,144,255)
+  fire_brick = 0xB22222,               // rgb(178,34,34)
+  floral_white = 0xFFFAF0,             // rgb(255,250,240)
+  forest_green = 0x228B22,             // rgb(34,139,34)
+  fuchsia = 0xFF00FF,                  // rgb(255,0,255)
+  gainsboro = 0xDCDCDC,                // rgb(220,220,220)
+  ghost_white = 0xF8F8FF,              // rgb(248,248,255)
+  gold = 0xFFD700,                     // rgb(255,215,0)
+  golden_rod = 0xDAA520,               // rgb(218,165,32)
+  gray = 0x808080,                     // rgb(128,128,128)
+  green = 0x008000,                    // rgb(0,128,0)
+  green_yellow = 0xADFF2F,             // rgb(173,255,47)
+  honey_dew = 0xF0FFF0,                // rgb(240,255,240)
+  hot_pink = 0xFF69B4,                 // rgb(255,105,180)
+  indian_red = 0xCD5C5C,               // rgb(205,92,92)
+  indigo = 0x4B0082,                   // rgb(75,0,130)
+  ivory = 0xFFFFF0,                    // rgb(255,255,240)
+  khaki = 0xF0E68C,                    // rgb(240,230,140)
+  lavender = 0xE6E6FA,                 // rgb(230,230,250)
+  lavender_blush = 0xFFF0F5,           // rgb(255,240,245)
+  lawn_green = 0x7CFC00,               // rgb(124,252,0)
+  lemon_chiffon = 0xFFFACD,            // rgb(255,250,205)
+  light_blue = 0xADD8E6,               // rgb(173,216,230)
+  light_coral = 0xF08080,              // rgb(240,128,128)
+  light_cyan = 0xE0FFFF,               // rgb(224,255,255)
+  light_golden_rod_yellow = 0xFAFAD2,  // rgb(250,250,210)
+  light_gray = 0xD3D3D3,               // rgb(211,211,211)
+  light_green = 0x90EE90,              // rgb(144,238,144)
+  light_pink = 0xFFB6C1,               // rgb(255,182,193)
+  light_salmon = 0xFFA07A,             // rgb(255,160,122)
+  light_sea_green = 0x20B2AA,          // rgb(32,178,170)
+  light_sky_blue = 0x87CEFA,           // rgb(135,206,250)
+  light_slate_gray = 0x778899,         // rgb(119,136,153)
+  light_steel_blue = 0xB0C4DE,         // rgb(176,196,222)
+  light_yellow = 0xFFFFE0,             // rgb(255,255,224)
+  lime = 0x00FF00,                     // rgb(0,255,0)
+  lime_green = 0x32CD32,               // rgb(50,205,50)
+  linen = 0xFAF0E6,                    // rgb(250,240,230)
+  magenta = 0xFF00FF,                  // rgb(255,0,255)
+  maroon = 0x800000,                   // rgb(128,0,0)
+  medium_aquamarine = 0x66CDAA,        // rgb(102,205,170)
+  medium_blue = 0x0000CD,              // rgb(0,0,205)
+  medium_orchid = 0xBA55D3,            // rgb(186,85,211)
+  medium_purple = 0x9370DB,            // rgb(147,112,219)
+  medium_sea_green = 0x3CB371,         // rgb(60,179,113)
+  medium_slate_blue = 0x7B68EE,        // rgb(123,104,238)
+  medium_spring_green = 0x00FA9A,      // rgb(0,250,154)
+  medium_turquoise = 0x48D1CC,         // rgb(72,209,204)
+  medium_violet_red = 0xC71585,        // rgb(199,21,133)
+  midnight_blue = 0x191970,            // rgb(25,25,112)
+  mint_cream = 0xF5FFFA,               // rgb(245,255,250)
+  misty_rose = 0xFFE4E1,               // rgb(255,228,225)
+  moccasin = 0xFFE4B5,                 // rgb(255,228,181)
+  navajo_white = 0xFFDEAD,             // rgb(255,222,173)
+  navy = 0x000080,                     // rgb(0,0,128)
+  old_lace = 0xFDF5E6,                 // rgb(253,245,230)
+  olive = 0x808000,                    // rgb(128,128,0)
+  olive_drab = 0x6B8E23,               // rgb(107,142,35)
+  orange = 0xFFA500,                   // rgb(255,165,0)
+  orange_red = 0xFF4500,               // rgb(255,69,0)
+  orchid = 0xDA70D6,                   // rgb(218,112,214)
+  pale_golden_rod = 0xEEE8AA,          // rgb(238,232,170)
+  pale_green = 0x98FB98,               // rgb(152,251,152)
+  pale_turquoise = 0xAFEEEE,           // rgb(175,238,238)
+  pale_violet_red = 0xDB7093,          // rgb(219,112,147)
+  papaya_whip = 0xFFEFD5,              // rgb(255,239,213)
+  peach_puff = 0xFFDAB9,               // rgb(255,218,185)
+  peru = 0xCD853F,                     // rgb(205,133,63)
+  pink = 0xFFC0CB,                     // rgb(255,192,203)
+  plum = 0xDDA0DD,                     // rgb(221,160,221)
+  powder_blue = 0xB0E0E6,              // rgb(176,224,230)
+  purple = 0x800080,                   // rgb(128,0,128)
+  rebecca_purple = 0x663399,           // rgb(102,51,153)
+  red = 0xFF0000,                      // rgb(255,0,0)
+  rosy_brown = 0xBC8F8F,               // rgb(188,143,143)
+  royal_blue = 0x4169E1,               // rgb(65,105,225)
+  saddle_brown = 0x8B4513,             // rgb(139,69,19)
+  salmon = 0xFA8072,                   // rgb(250,128,114)
+  sandy_brown = 0xF4A460,              // rgb(244,164,96)
+  sea_green = 0x2E8B57,                // rgb(46,139,87)
+  sea_shell = 0xFFF5EE,                // rgb(255,245,238)
+  sienna = 0xA0522D,                   // rgb(160,82,45)
+  silver = 0xC0C0C0,                   // rgb(192,192,192)
+  sky_blue = 0x87CEEB,                 // rgb(135,206,235)
+  slate_blue = 0x6A5ACD,               // rgb(106,90,205)
+  slate_gray = 0x708090,               // rgb(112,128,144)
+  snow = 0xFFFAFA,                     // rgb(255,250,250)
+  spring_green = 0x00FF7F,             // rgb(0,255,127)
+  steel_blue = 0x4682B4,               // rgb(70,130,180)
+  tan = 0xD2B48C,                      // rgb(210,180,140)
+  teal = 0x008080,                     // rgb(0,128,128)
+  thistle = 0xD8BFD8,                  // rgb(216,191,216)
+  tomato = 0xFF6347,                   // rgb(255,99,71)
+  turquoise = 0x40E0D0,                // rgb(64,224,208)
+  violet = 0xEE82EE,                   // rgb(238,130,238)
+  wheat = 0xF5DEB3,                    // rgb(245,222,179)
+  white = 0xFFFFFF,                    // rgb(255,255,255)
+  white_smoke = 0xF5F5F5,              // rgb(245,245,245)
+  yellow = 0xFFFF00,                   // rgb(255,255,0)
+  yellow_green = 0x9ACD32              // rgb(154,205,50)
+};                                     // enum class color
+
+enum class terminal_color : uint8_t {
+  black = 30,
+  red,
+  green,
+  yellow,
+  blue,
+  magenta,
+  cyan,
+  white,
+  bright_black = 90,
+  bright_red,
+  bright_green,
+  bright_yellow,
+  bright_blue,
+  bright_magenta,
+  bright_cyan,
+  bright_white
+};
+
+enum class emphasis : uint8_t {
+  bold = 1,
+  faint = 1 << 1,
+  italic = 1 << 2,
+  underline = 1 << 3,
+  blink = 1 << 4,
+  reverse = 1 << 5,
+  conceal = 1 << 6,
+  strikethrough = 1 << 7,
+};
+
+// rgb is a struct for red, green and blue colors.
+// Using the name "rgb" makes some editors show the color in a tooltip.
+struct rgb {
+  FMT_CONSTEXPR rgb() : r(0), g(0), b(0) {}
+  FMT_CONSTEXPR rgb(uint8_t r_, uint8_t g_, uint8_t b_) : r(r_), g(g_), b(b_) {}
+  FMT_CONSTEXPR rgb(uint32_t hex)
+      : r((hex >> 16) & 0xFF), g((hex >> 8) & 0xFF), b(hex & 0xFF) {}
+  FMT_CONSTEXPR rgb(color hex)
+      : r((uint32_t(hex) >> 16) & 0xFF),
+        g((uint32_t(hex) >> 8) & 0xFF),
+        b(uint32_t(hex) & 0xFF) {}
+  uint8_t r;
+  uint8_t g;
+  uint8_t b;
+};
+
+namespace detail {
+
+// color is a struct of either a rgb color or a terminal color.
+struct color_type {
+  FMT_CONSTEXPR color_type() noexcept : is_rgb(), value{} {}
+  FMT_CONSTEXPR color_type(color rgb_color) noexcept : is_rgb(true), value{} {
+    value.rgb_color = static_cast<uint32_t>(rgb_color);
+  }
+  FMT_CONSTEXPR color_type(rgb rgb_color) noexcept : is_rgb(true), value{} {
+    value.rgb_color = (static_cast<uint32_t>(rgb_color.r) << 16) |
+                      (static_cast<uint32_t>(rgb_color.g) << 8) | rgb_color.b;
+  }
+  FMT_CONSTEXPR color_type(terminal_color term_color) noexcept
+      : is_rgb(), value{} {
+    value.term_color = static_cast<uint8_t>(term_color);
+  }
+  bool is_rgb;
+  union color_union {
+    uint8_t term_color;
+    uint32_t rgb_color;
+  } value;
+};
+}  // namespace detail
+
+/** A text style consisting of foreground and background colors and emphasis. */
+class text_style {
+ public:
+  FMT_CONSTEXPR text_style(emphasis em = emphasis()) noexcept
+      : set_foreground_color(), set_background_color(), ems(em) {}
+
+  FMT_CONSTEXPR auto operator|=(const text_style& rhs) -> text_style& {
+    if (!set_foreground_color) {
+      set_foreground_color = rhs.set_foreground_color;
+      foreground_color = rhs.foreground_color;
+    } else if (rhs.set_foreground_color) {
+      if (!foreground_color.is_rgb || !rhs.foreground_color.is_rgb)
+        FMT_THROW(format_error("can't OR a terminal color"));
+      foreground_color.value.rgb_color |= rhs.foreground_color.value.rgb_color;
+    }
+
+    if (!set_background_color) {
+      set_background_color = rhs.set_background_color;
+      background_color = rhs.background_color;
+    } else if (rhs.set_background_color) {
+      if (!background_color.is_rgb || !rhs.background_color.is_rgb)
+        FMT_THROW(format_error("can't OR a terminal color"));
+      background_color.value.rgb_color |= rhs.background_color.value.rgb_color;
+    }
+
+    ems = static_cast<emphasis>(static_cast<uint8_t>(ems) |
+                                static_cast<uint8_t>(rhs.ems));
+    return *this;
+  }
+
+  friend FMT_CONSTEXPR auto operator|(text_style lhs, const text_style& rhs)
+      -> text_style {
+    return lhs |= rhs;
+  }
+
+  FMT_CONSTEXPR auto has_foreground() const noexcept -> bool {
+    return set_foreground_color;
+  }
+  FMT_CONSTEXPR auto has_background() const noexcept -> bool {
+    return set_background_color;
+  }
+  FMT_CONSTEXPR auto has_emphasis() const noexcept -> bool {
+    return static_cast<uint8_t>(ems) != 0;
+  }
+  FMT_CONSTEXPR auto get_foreground() const noexcept -> detail::color_type {
+    FMT_ASSERT(has_foreground(), "no foreground specified for this style");
+    return foreground_color;
+  }
+  FMT_CONSTEXPR auto get_background() const noexcept -> detail::color_type {
+    FMT_ASSERT(has_background(), "no background specified for this style");
+    return background_color;
+  }
+  FMT_CONSTEXPR auto get_emphasis() const noexcept -> emphasis {
+    FMT_ASSERT(has_emphasis(), "no emphasis specified for this style");
+    return ems;
+  }
+
+ private:
+  FMT_CONSTEXPR text_style(bool is_foreground,
+                           detail::color_type text_color) noexcept
+      : set_foreground_color(), set_background_color(), ems() {
+    if (is_foreground) {
+      foreground_color = text_color;
+      set_foreground_color = true;
+    } else {
+      background_color = text_color;
+      set_background_color = true;
+    }
+  }
+
+  friend FMT_CONSTEXPR auto fg(detail::color_type foreground) noexcept
+      -> text_style;
+
+  friend FMT_CONSTEXPR auto bg(detail::color_type background) noexcept
+      -> text_style;
+
+  detail::color_type foreground_color;
+  detail::color_type background_color;
+  bool set_foreground_color;
+  bool set_background_color;
+  emphasis ems;
+};
+
+/** Creates a text style from the foreground (text) color. */
+FMT_CONSTEXPR inline auto fg(detail::color_type foreground) noexcept
+    -> text_style {
+  return text_style(true, foreground);
+}
+
+/** Creates a text style from the background color. */
+FMT_CONSTEXPR inline auto bg(detail::color_type background) noexcept
+    -> text_style {
+  return text_style(false, background);
+}
+
+FMT_CONSTEXPR inline auto operator|(emphasis lhs, emphasis rhs) noexcept
+    -> text_style {
+  return text_style(lhs) | rhs;
+}
+
+namespace detail {
+
+template <typename Char> struct ansi_color_escape {
+  FMT_CONSTEXPR ansi_color_escape(detail::color_type text_color,
+                                  const char* esc) noexcept {
+    // If we have a terminal color, we need to output another escape code
+    // sequence.
+    if (!text_color.is_rgb) {
+      bool is_background = esc == string_view("\x1b[48;2;");
+      uint32_t value = text_color.value.term_color;
+      // Background ASCII codes are the same as the foreground ones but with
+      // 10 more.
+      if (is_background) value += 10u;
+
+      size_t index = 0;
+      buffer[index++] = static_cast<Char>('\x1b');
+      buffer[index++] = static_cast<Char>('[');
+
+      if (value >= 100u) {
+        buffer[index++] = static_cast<Char>('1');
+        value %= 100u;
+      }
+      buffer[index++] = static_cast<Char>('0' + value / 10u);
+      buffer[index++] = static_cast<Char>('0' + value % 10u);
+
+      buffer[index++] = static_cast<Char>('m');
+      buffer[index++] = static_cast<Char>('\0');
+      return;
+    }
+
+    for (int i = 0; i < 7; i++) {
+      buffer[i] = static_cast<Char>(esc[i]);
+    }
+    rgb color(text_color.value.rgb_color);
+    to_esc(color.r, buffer + 7, ';');
+    to_esc(color.g, buffer + 11, ';');
+    to_esc(color.b, buffer + 15, 'm');
+    buffer[19] = static_cast<Char>(0);
+  }
+  FMT_CONSTEXPR ansi_color_escape(emphasis em) noexcept {
+    uint8_t em_codes[num_emphases] = {};
+    if (has_emphasis(em, emphasis::bold)) em_codes[0] = 1;
+    if (has_emphasis(em, emphasis::faint)) em_codes[1] = 2;
+    if (has_emphasis(em, emphasis::italic)) em_codes[2] = 3;
+    if (has_emphasis(em, emphasis::underline)) em_codes[3] = 4;
+    if (has_emphasis(em, emphasis::blink)) em_codes[4] = 5;
+    if (has_emphasis(em, emphasis::reverse)) em_codes[5] = 7;
+    if (has_emphasis(em, emphasis::conceal)) em_codes[6] = 8;
+    if (has_emphasis(em, emphasis::strikethrough)) em_codes[7] = 9;
+
+    size_t index = 0;
+    for (size_t i = 0; i < num_emphases; ++i) {
+      if (!em_codes[i]) continue;
+      buffer[index++] = static_cast<Char>('\x1b');
+      buffer[index++] = static_cast<Char>('[');
+      buffer[index++] = static_cast<Char>('0' + em_codes[i]);
+      buffer[index++] = static_cast<Char>('m');
+    }
+    buffer[index++] = static_cast<Char>(0);
+  }
+  FMT_CONSTEXPR operator const Char*() const noexcept { return buffer; }
+
+  FMT_CONSTEXPR auto begin() const noexcept -> const Char* { return buffer; }
+  FMT_CONSTEXPR_CHAR_TRAITS auto end() const noexcept -> const Char* {
+    return buffer + std::char_traits<Char>::length(buffer);
+  }
+
+ private:
+  static constexpr size_t num_emphases = 8;
+  Char buffer[7u + 3u * num_emphases + 1u];
+
+  static FMT_CONSTEXPR void to_esc(uint8_t c, Char* out,
+                                   char delimiter) noexcept {
+    out[0] = static_cast<Char>('0' + c / 100);
+    out[1] = static_cast<Char>('0' + c / 10 % 10);
+    out[2] = static_cast<Char>('0' + c % 10);
+    out[3] = static_cast<Char>(delimiter);
+  }
+  static FMT_CONSTEXPR auto has_emphasis(emphasis em, emphasis mask) noexcept
+      -> bool {
+    return static_cast<uint8_t>(em) & static_cast<uint8_t>(mask);
+  }
+};
+
+template <typename Char>
+FMT_CONSTEXPR auto make_foreground_color(detail::color_type foreground) noexcept
+    -> ansi_color_escape<Char> {
+  return ansi_color_escape<Char>(foreground, "\x1b[38;2;");
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto make_background_color(detail::color_type background) noexcept
+    -> ansi_color_escape<Char> {
+  return ansi_color_escape<Char>(background, "\x1b[48;2;");
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto make_emphasis(emphasis em) noexcept
+    -> ansi_color_escape<Char> {
+  return ansi_color_escape<Char>(em);
+}
+
+template <typename Char> inline void reset_color(buffer<Char>& buffer) {
+  auto reset_color = string_view("\x1b[0m");
+  buffer.append(reset_color.begin(), reset_color.end());
+}
+
+template <typename T> struct styled_arg : detail::view {
+  const T& value;
+  text_style style;
+  styled_arg(const T& v, text_style s) : value(v), style(s) {}
+};
+
+template <typename Char>
+void vformat_to(buffer<Char>& buf, const text_style& ts,
+                basic_string_view<Char> format_str,
+                basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  bool has_style = false;
+  if (ts.has_emphasis()) {
+    has_style = true;
+    auto emphasis = detail::make_emphasis<Char>(ts.get_emphasis());
+    buf.append(emphasis.begin(), emphasis.end());
+  }
+  if (ts.has_foreground()) {
+    has_style = true;
+    auto foreground = detail::make_foreground_color<Char>(ts.get_foreground());
+    buf.append(foreground.begin(), foreground.end());
+  }
+  if (ts.has_background()) {
+    has_style = true;
+    auto background = detail::make_background_color<Char>(ts.get_background());
+    buf.append(background.begin(), background.end());
+  }
+  detail::vformat_to(buf, format_str, args, {});
+  if (has_style) detail::reset_color<Char>(buf);
+}
+
+}  // namespace detail
+
+inline void vprint(std::FILE* f, const text_style& ts, string_view fmt,
+                   format_args args) {
+  // Legacy wide streams are not supported.
+  auto buf = memory_buffer();
+  detail::vformat_to(buf, ts, fmt, args);
+  if (detail::is_utf8()) {
+    detail::print(f, string_view(buf.begin(), buf.size()));
+    return;
+  }
+  buf.push_back('\0');
+  int result = std::fputs(buf.data(), f);
+  if (result < 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot write to file")));
+}
+
+/**
+  \rst
+  Formats a string and prints it to the specified file stream using ANSI
+  escape sequences to specify text formatting.
+
+  **Example**::
+
+    fmt::print(fmt::emphasis::bold | fg(fmt::color::red),
+               "Elapsed time: {0:.2f} seconds", 1.23);
+  \endrst
+ */
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_string<S>::value)>
+void print(std::FILE* f, const text_style& ts, const S& format_str,
+           const Args&... args) {
+  vprint(f, ts, format_str,
+         fmt::make_format_args<buffer_context<char_t<S>>>(args...));
+}
+
+/**
+  \rst
+  Formats a string and prints it to stdout using ANSI escape sequences to
+  specify text formatting.
+
+  **Example**::
+
+    fmt::print(fmt::emphasis::bold | fg(fmt::color::red),
+               "Elapsed time: {0:.2f} seconds", 1.23);
+  \endrst
+ */
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_string<S>::value)>
+void print(const text_style& ts, const S& format_str, const Args&... args) {
+  return print(stdout, ts, format_str, args...);
+}
+
+template <typename S, typename Char = char_t<S>>
+inline auto vformat(
+    const text_style& ts, const S& format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> std::basic_string<Char> {
+  basic_memory_buffer<Char> buf;
+  detail::vformat_to(buf, ts, detail::to_string_view(format_str), args);
+  return fmt::to_string(buf);
+}
+
+/**
+  \rst
+  Formats arguments and returns the result as a string using ANSI
+  escape sequences to specify text formatting.
+
+  **Example**::
+
+    #include <fmt/color.h>
+    std::string message = fmt::format(fmt::emphasis::bold | fg(fmt::color::red),
+                                      "The answer is {}", 42);
+  \endrst
+*/
+template <typename S, typename... Args, typename Char = char_t<S>>
+inline auto format(const text_style& ts, const S& format_str,
+                   const Args&... args) -> std::basic_string<Char> {
+  return fmt::vformat(ts, detail::to_string_view(format_str),
+                      fmt::make_format_args<buffer_context<Char>>(args...));
+}
+
+/**
+  Formats a string with the given text_style and writes the output to ``out``.
+ */
+template <typename OutputIt, typename Char,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
+auto vformat_to(OutputIt out, const text_style& ts,
+                basic_string_view<Char> format_str,
+                basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> OutputIt {
+  auto&& buf = detail::get_buffer<Char>(out);
+  detail::vformat_to(buf, ts, format_str, args);
+  return detail::get_iterator(buf, out);
+}
+
+/**
+  \rst
+  Formats arguments with the given text_style, writes the result to the output
+  iterator ``out`` and returns the iterator past the end of the output range.
+
+  **Example**::
+
+    std::vector<char> out;
+    fmt::format_to(std::back_inserter(out),
+                   fmt::emphasis::bold | fg(fmt::color::red), "{}", 42);
+  \endrst
+*/
+template <
+    typename OutputIt, typename S, typename... Args,
+    bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value &&
+                  detail::is_string<S>::value>
+inline auto format_to(OutputIt out, const text_style& ts, const S& format_str,
+                      Args&&... args) ->
+    typename std::enable_if<enable, OutputIt>::type {
+  return vformat_to(out, ts, detail::to_string_view(format_str),
+                    fmt::make_format_args<buffer_context<char_t<S>>>(args...));
+}
+
+template <typename T, typename Char>
+struct formatter<detail::styled_arg<T>, Char> : formatter<T, Char> {
+  template <typename FormatContext>
+  auto format(const detail::styled_arg<T>& arg, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    const auto& ts = arg.style;
+    const auto& value = arg.value;
+    auto out = ctx.out();
+
+    bool has_style = false;
+    if (ts.has_emphasis()) {
+      has_style = true;
+      auto emphasis = detail::make_emphasis<Char>(ts.get_emphasis());
+      out = std::copy(emphasis.begin(), emphasis.end(), out);
+    }
+    if (ts.has_foreground()) {
+      has_style = true;
+      auto foreground =
+          detail::make_foreground_color<Char>(ts.get_foreground());
+      out = std::copy(foreground.begin(), foreground.end(), out);
+    }
+    if (ts.has_background()) {
+      has_style = true;
+      auto background =
+          detail::make_background_color<Char>(ts.get_background());
+      out = std::copy(background.begin(), background.end(), out);
+    }
+    out = formatter<T, Char>::format(value, ctx);
+    if (has_style) {
+      auto reset_color = string_view("\x1b[0m");
+      out = std::copy(reset_color.begin(), reset_color.end(), out);
+    }
+    return out;
+  }
+};
+
+/**
+  \rst
+  Returns an argument that will be formatted using ANSI escape sequences,
+  to be used in a formatting function.
+
+  **Example**::
+
+    fmt::print("Elapsed time: {0:.2f} seconds",
+               fmt::styled(1.23, fmt::fg(fmt::color::green) |
+                                 fmt::bg(fmt::color::blue)));
+  \endrst
+ */
+template <typename T>
+FMT_CONSTEXPR auto styled(const T& value, text_style ts)
+    -> detail::styled_arg<remove_cvref_t<T>> {
+  return detail::styled_arg<remove_cvref_t<T>>{value, ts};
+}
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_COLOR_H_
diff --git a/dep/fmt/include/fmt/compile.h b/dep/fmt/include/fmt/compile.h
new file mode 100644
index 00000000..3b3f166e
--- /dev/null
+++ b/dep/fmt/include/fmt/compile.h
@@ -0,0 +1,535 @@
+// Formatting library for C++ - experimental format string compilation
+//
+// Copyright (c) 2012 - present, Victor Zverovich and fmt contributors
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_COMPILE_H_
+#define FMT_COMPILE_H_
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename Char, typename InputIt>
+FMT_CONSTEXPR inline auto copy_str(InputIt begin, InputIt end,
+                                   counting_iterator it) -> counting_iterator {
+  return it + (end - begin);
+}
+
+// A compile-time string which is compiled into fast formatting code.
+class compiled_string {};
+
+template <typename S>
+struct is_compiled_string : std::is_base_of<compiled_string, S> {};
+
+/**
+  \rst
+  Converts a string literal *s* into a format string that will be parsed at
+  compile time and converted into efficient formatting code. Requires C++17
+  ``constexpr if`` compiler support.
+
+  **Example**::
+
+    // Converts 42 into std::string using the most efficient method and no
+    // runtime format string processing.
+    std::string s = fmt::format(FMT_COMPILE("{}"), 42);
+  \endrst
+ */
+#if defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+#  define FMT_COMPILE(s) \
+    FMT_STRING_IMPL(s, fmt::detail::compiled_string, explicit)
+#else
+#  define FMT_COMPILE(s) FMT_STRING(s)
+#endif
+
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct udl_compiled_string : compiled_string {
+  using char_type = Char;
+  explicit constexpr operator basic_string_view<char_type>() const {
+    return {Str.data, N - 1};
+  }
+};
+#endif
+
+template <typename T, typename... Tail>
+auto first(const T& value, const Tail&...) -> const T& {
+  return value;
+}
+
+#if defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+template <typename... Args> struct type_list {};
+
+// Returns a reference to the argument at index N from [first, rest...].
+template <int N, typename T, typename... Args>
+constexpr const auto& get([[maybe_unused]] const T& first,
+                          [[maybe_unused]] const Args&... rest) {
+  static_assert(N < 1 + sizeof...(Args), "index is out of bounds");
+  if constexpr (N == 0)
+    return first;
+  else
+    return detail::get<N - 1>(rest...);
+}
+
+template <typename Char, typename... Args>
+constexpr int get_arg_index_by_name(basic_string_view<Char> name,
+                                    type_list<Args...>) {
+  return get_arg_index_by_name<Args...>(name);
+}
+
+template <int N, typename> struct get_type_impl;
+
+template <int N, typename... Args> struct get_type_impl<N, type_list<Args...>> {
+  using type =
+      remove_cvref_t<decltype(detail::get<N>(std::declval<Args>()...))>;
+};
+
+template <int N, typename T>
+using get_type = typename get_type_impl<N, T>::type;
+
+template <typename T> struct is_compiled_format : std::false_type {};
+
+template <typename Char> struct text {
+  basic_string_view<Char> data;
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&...) const {
+    return write<Char>(out, data);
+  }
+};
+
+template <typename Char>
+struct is_compiled_format<text<Char>> : std::true_type {};
+
+template <typename Char>
+constexpr text<Char> make_text(basic_string_view<Char> s, size_t pos,
+                               size_t size) {
+  return {{&s[pos], size}};
+}
+
+template <typename Char> struct code_unit {
+  Char value;
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&...) const {
+    *out++ = value;
+    return out;
+  }
+};
+
+// This ensures that the argument type is convertible to `const T&`.
+template <typename T, int N, typename... Args>
+constexpr const T& get_arg_checked(const Args&... args) {
+  const auto& arg = detail::get<N>(args...);
+  if constexpr (detail::is_named_arg<remove_cvref_t<decltype(arg)>>()) {
+    return arg.value;
+  } else {
+    return arg;
+  }
+}
+
+template <typename Char>
+struct is_compiled_format<code_unit<Char>> : std::true_type {};
+
+// A replacement field that refers to argument N.
+template <typename Char, typename T, int N> struct field {
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&... args) const {
+    const T& arg = get_arg_checked<T, N>(args...);
+    if constexpr (std::is_convertible_v<T, basic_string_view<Char>>) {
+      auto s = basic_string_view<Char>(arg);
+      return copy_str<Char>(s.begin(), s.end(), out);
+    }
+    return write<Char>(out, arg);
+  }
+};
+
+template <typename Char, typename T, int N>
+struct is_compiled_format<field<Char, T, N>> : std::true_type {};
+
+// A replacement field that refers to argument with name.
+template <typename Char> struct runtime_named_field {
+  using char_type = Char;
+  basic_string_view<Char> name;
+
+  template <typename OutputIt, typename T>
+  constexpr static bool try_format_argument(
+      OutputIt& out,
+      // [[maybe_unused]] due to unused-but-set-parameter warning in GCC 7,8,9
+      [[maybe_unused]] basic_string_view<Char> arg_name, const T& arg) {
+    if constexpr (is_named_arg<typename std::remove_cv<T>::type>::value) {
+      if (arg_name == arg.name) {
+        out = write<Char>(out, arg.value);
+        return true;
+      }
+    }
+    return false;
+  }
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&... args) const {
+    bool found = (try_format_argument(out, name, args) || ...);
+    if (!found) {
+      FMT_THROW(format_error("argument with specified name is not found"));
+    }
+    return out;
+  }
+};
+
+template <typename Char>
+struct is_compiled_format<runtime_named_field<Char>> : std::true_type {};
+
+// A replacement field that refers to argument N and has format specifiers.
+template <typename Char, typename T, int N> struct spec_field {
+  using char_type = Char;
+  formatter<T, Char> fmt;
+
+  template <typename OutputIt, typename... Args>
+  constexpr FMT_INLINE OutputIt format(OutputIt out,
+                                       const Args&... args) const {
+    const auto& vargs =
+        fmt::make_format_args<basic_format_context<OutputIt, Char>>(args...);
+    basic_format_context<OutputIt, Char> ctx(out, vargs);
+    return fmt.format(get_arg_checked<T, N>(args...), ctx);
+  }
+};
+
+template <typename Char, typename T, int N>
+struct is_compiled_format<spec_field<Char, T, N>> : std::true_type {};
+
+template <typename L, typename R> struct concat {
+  L lhs;
+  R rhs;
+  using char_type = typename L::char_type;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&... args) const {
+    out = lhs.format(out, args...);
+    return rhs.format(out, args...);
+  }
+};
+
+template <typename L, typename R>
+struct is_compiled_format<concat<L, R>> : std::true_type {};
+
+template <typename L, typename R>
+constexpr concat<L, R> make_concat(L lhs, R rhs) {
+  return {lhs, rhs};
+}
+
+struct unknown_format {};
+
+template <typename Char>
+constexpr size_t parse_text(basic_string_view<Char> str, size_t pos) {
+  for (size_t size = str.size(); pos != size; ++pos) {
+    if (str[pos] == '{' || str[pos] == '}') break;
+  }
+  return pos;
+}
+
+template <typename Args, size_t POS, int ID, typename S>
+constexpr auto compile_format_string(S format_str);
+
+template <typename Args, size_t POS, int ID, typename T, typename S>
+constexpr auto parse_tail(T head, S format_str) {
+  if constexpr (POS !=
+                basic_string_view<typename S::char_type>(format_str).size()) {
+    constexpr auto tail = compile_format_string<Args, POS, ID>(format_str);
+    if constexpr (std::is_same<remove_cvref_t<decltype(tail)>,
+                               unknown_format>())
+      return tail;
+    else
+      return make_concat(head, tail);
+  } else {
+    return head;
+  }
+}
+
+template <typename T, typename Char> struct parse_specs_result {
+  formatter<T, Char> fmt;
+  size_t end;
+  int next_arg_id;
+};
+
+enum { manual_indexing_id = -1 };
+
+template <typename T, typename Char>
+constexpr parse_specs_result<T, Char> parse_specs(basic_string_view<Char> str,
+                                                  size_t pos, int next_arg_id) {
+  str.remove_prefix(pos);
+  auto ctx =
+      compile_parse_context<Char>(str, max_value<int>(), nullptr, next_arg_id);
+  auto f = formatter<T, Char>();
+  auto end = f.parse(ctx);
+  return {f, pos + fmt::detail::to_unsigned(end - str.data()),
+          next_arg_id == 0 ? manual_indexing_id : ctx.next_arg_id()};
+}
+
+template <typename Char> struct arg_id_handler {
+  arg_ref<Char> arg_id;
+
+  constexpr int on_auto() {
+    FMT_ASSERT(false, "handler cannot be used with automatic indexing");
+    return 0;
+  }
+  constexpr int on_index(int id) {
+    arg_id = arg_ref<Char>(id);
+    return 0;
+  }
+  constexpr int on_name(basic_string_view<Char> id) {
+    arg_id = arg_ref<Char>(id);
+    return 0;
+  }
+};
+
+template <typename Char> struct parse_arg_id_result {
+  arg_ref<Char> arg_id;
+  const Char* arg_id_end;
+};
+
+template <int ID, typename Char>
+constexpr auto parse_arg_id(const Char* begin, const Char* end) {
+  auto handler = arg_id_handler<Char>{arg_ref<Char>{}};
+  auto arg_id_end = parse_arg_id(begin, end, handler);
+  return parse_arg_id_result<Char>{handler.arg_id, arg_id_end};
+}
+
+template <typename T, typename Enable = void> struct field_type {
+  using type = remove_cvref_t<T>;
+};
+
+template <typename T>
+struct field_type<T, enable_if_t<detail::is_named_arg<T>::value>> {
+  using type = remove_cvref_t<decltype(T::value)>;
+};
+
+template <typename T, typename Args, size_t END_POS, int ARG_INDEX, int NEXT_ID,
+          typename S>
+constexpr auto parse_replacement_field_then_tail(S format_str) {
+  using char_type = typename S::char_type;
+  constexpr auto str = basic_string_view<char_type>(format_str);
+  constexpr char_type c = END_POS != str.size() ? str[END_POS] : char_type();
+  if constexpr (c == '}') {
+    return parse_tail<Args, END_POS + 1, NEXT_ID>(
+        field<char_type, typename field_type<T>::type, ARG_INDEX>(),
+        format_str);
+  } else if constexpr (c != ':') {
+    FMT_THROW(format_error("expected ':'"));
+  } else {
+    constexpr auto result = parse_specs<typename field_type<T>::type>(
+        str, END_POS + 1, NEXT_ID == manual_indexing_id ? 0 : NEXT_ID);
+    if constexpr (result.end >= str.size() || str[result.end] != '}') {
+      FMT_THROW(format_error("expected '}'"));
+      return 0;
+    } else {
+      return parse_tail<Args, result.end + 1, result.next_arg_id>(
+          spec_field<char_type, typename field_type<T>::type, ARG_INDEX>{
+              result.fmt},
+          format_str);
+    }
+  }
+}
+
+// Compiles a non-empty format string and returns the compiled representation
+// or unknown_format() on unrecognized input.
+template <typename Args, size_t POS, int ID, typename S>
+constexpr auto compile_format_string(S format_str) {
+  using char_type = typename S::char_type;
+  constexpr auto str = basic_string_view<char_type>(format_str);
+  if constexpr (str[POS] == '{') {
+    if constexpr (POS + 1 == str.size())
+      FMT_THROW(format_error("unmatched '{' in format string"));
+    if constexpr (str[POS + 1] == '{') {
+      return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), format_str);
+    } else if constexpr (str[POS + 1] == '}' || str[POS + 1] == ':') {
+      static_assert(ID != manual_indexing_id,
+                    "cannot switch from manual to automatic argument indexing");
+      constexpr auto next_id =
+          ID != manual_indexing_id ? ID + 1 : manual_indexing_id;
+      return parse_replacement_field_then_tail<get_type<ID, Args>, Args,
+                                               POS + 1, ID, next_id>(
+          format_str);
+    } else {
+      constexpr auto arg_id_result =
+          parse_arg_id<ID>(str.data() + POS + 1, str.data() + str.size());
+      constexpr auto arg_id_end_pos = arg_id_result.arg_id_end - str.data();
+      constexpr char_type c =
+          arg_id_end_pos != str.size() ? str[arg_id_end_pos] : char_type();
+      static_assert(c == '}' || c == ':', "missing '}' in format string");
+      if constexpr (arg_id_result.arg_id.kind == arg_id_kind::index) {
+        static_assert(
+            ID == manual_indexing_id || ID == 0,
+            "cannot switch from automatic to manual argument indexing");
+        constexpr auto arg_index = arg_id_result.arg_id.val.index;
+        return parse_replacement_field_then_tail<get_type<arg_index, Args>,
+                                                 Args, arg_id_end_pos,
+                                                 arg_index, manual_indexing_id>(
+            format_str);
+      } else if constexpr (arg_id_result.arg_id.kind == arg_id_kind::name) {
+        constexpr auto arg_index =
+            get_arg_index_by_name(arg_id_result.arg_id.val.name, Args{});
+        if constexpr (arg_index >= 0) {
+          constexpr auto next_id =
+              ID != manual_indexing_id ? ID + 1 : manual_indexing_id;
+          return parse_replacement_field_then_tail<
+              decltype(get_type<arg_index, Args>::value), Args, arg_id_end_pos,
+              arg_index, next_id>(format_str);
+        } else if constexpr (c == '}') {
+          return parse_tail<Args, arg_id_end_pos + 1, ID>(
+              runtime_named_field<char_type>{arg_id_result.arg_id.val.name},
+              format_str);
+        } else if constexpr (c == ':') {
+          return unknown_format();  // no type info for specs parsing
+        }
+      }
+    }
+  } else if constexpr (str[POS] == '}') {
+    if constexpr (POS + 1 == str.size())
+      FMT_THROW(format_error("unmatched '}' in format string"));
+    return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), format_str);
+  } else {
+    constexpr auto end = parse_text(str, POS + 1);
+    if constexpr (end - POS > 1) {
+      return parse_tail<Args, end, ID>(make_text(str, POS, end - POS),
+                                       format_str);
+    } else {
+      return parse_tail<Args, end, ID>(code_unit<char_type>{str[POS]},
+                                       format_str);
+    }
+  }
+}
+
+template <typename... Args, typename S,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+constexpr auto compile(S format_str) {
+  constexpr auto str = basic_string_view<typename S::char_type>(format_str);
+  if constexpr (str.size() == 0) {
+    return detail::make_text(str, 0, 0);
+  } else {
+    constexpr auto result =
+        detail::compile_format_string<detail::type_list<Args...>, 0, 0>(
+            format_str);
+    return result;
+  }
+}
+#endif  // defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+#if defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+
+template <typename CompiledFormat, typename... Args,
+          typename Char = typename CompiledFormat::char_type,
+          FMT_ENABLE_IF(detail::is_compiled_format<CompiledFormat>::value)>
+FMT_INLINE std::basic_string<Char> format(const CompiledFormat& cf,
+                                          const Args&... args) {
+  auto s = std::basic_string<Char>();
+  cf.format(std::back_inserter(s), args...);
+  return s;
+}
+
+template <typename OutputIt, typename CompiledFormat, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_format<CompiledFormat>::value)>
+constexpr FMT_INLINE OutputIt format_to(OutputIt out, const CompiledFormat& cf,
+                                        const Args&... args) {
+  return cf.format(out, args...);
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_INLINE std::basic_string<typename S::char_type> format(const S&,
+                                                           Args&&... args) {
+  if constexpr (std::is_same<typename S::char_type, char>::value) {
+    constexpr auto str = basic_string_view<typename S::char_type>(S());
+    if constexpr (str.size() == 2 && str[0] == '{' && str[1] == '}') {
+      const auto& first = detail::first(args...);
+      if constexpr (detail::is_named_arg<
+                        remove_cvref_t<decltype(first)>>::value) {
+        return fmt::to_string(first.value);
+      } else {
+        return fmt::to_string(first);
+      }
+    }
+  }
+  constexpr auto compiled = detail::compile<Args...>(S());
+  if constexpr (std::is_same<remove_cvref_t<decltype(compiled)>,
+                             detail::unknown_format>()) {
+    return fmt::format(
+        static_cast<basic_string_view<typename S::char_type>>(S()),
+        std::forward<Args>(args)...);
+  } else {
+    return fmt::format(compiled, std::forward<Args>(args)...);
+  }
+}
+
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_CONSTEXPR OutputIt format_to(OutputIt out, const S&, Args&&... args) {
+  constexpr auto compiled = detail::compile<Args...>(S());
+  if constexpr (std::is_same<remove_cvref_t<decltype(compiled)>,
+                             detail::unknown_format>()) {
+    return fmt::format_to(
+        out, static_cast<basic_string_view<typename S::char_type>>(S()),
+        std::forward<Args>(args)...);
+  } else {
+    return fmt::format_to(out, compiled, std::forward<Args>(args)...);
+  }
+}
+#endif
+
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+auto format_to_n(OutputIt out, size_t n, const S& format_str, Args&&... args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, char, traits>(out, n);
+  fmt::format_to(std::back_inserter(buf), format_str,
+                 std::forward<Args>(args)...);
+  return {buf.out(), buf.count()};
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_CONSTEXPR20 auto formatted_size(const S& format_str, const Args&... args)
+    -> size_t {
+  return fmt::format_to(detail::counting_iterator(), format_str, args...)
+      .count();
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+void print(std::FILE* f, const S& format_str, const Args&... args) {
+  memory_buffer buffer;
+  fmt::format_to(std::back_inserter(buffer), format_str, args...);
+  detail::print(f, {buffer.data(), buffer.size()});
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+void print(const S& format_str, const Args&... args) {
+  print(stdout, format_str, args...);
+}
+
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+inline namespace literals {
+template <detail_exported::fixed_string Str> constexpr auto operator""_cf() {
+  using char_t = remove_cvref_t<decltype(Str.data[0])>;
+  return detail::udl_compiled_string<char_t, sizeof(Str.data) / sizeof(char_t),
+                                     Str>();
+}
+}  // namespace literals
+#endif
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_COMPILE_H_
diff --git a/dep/fmt/include/fmt/core.h b/dep/fmt/include/fmt/core.h
new file mode 100644
index 00000000..b51c1406
--- /dev/null
+++ b/dep/fmt/include/fmt/core.h
@@ -0,0 +1,2969 @@
+// Formatting library for C++ - the core API for char/UTF-8
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_CORE_H_
+#define FMT_CORE_H_
+
+#include <cstddef>  // std::byte
+#include <cstdio>   // std::FILE
+#include <cstring>  // std::strlen
+#include <iterator>
+#include <limits>
+#include <memory>  // std::addressof
+#include <string>
+#include <type_traits>
+
+// The fmt library version in the form major * 10000 + minor * 100 + patch.
+#define FMT_VERSION 100201
+
+#if defined(__clang__) && !defined(__ibmxl__)
+#  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
+#else
+#  define FMT_CLANG_VERSION 0
+#endif
+
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && \
+    !defined(__NVCOMPILER)
+#  define FMT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#else
+#  define FMT_GCC_VERSION 0
+#endif
+
+#ifndef FMT_GCC_PRAGMA
+// Workaround _Pragma bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59884.
+#  if FMT_GCC_VERSION >= 504
+#    define FMT_GCC_PRAGMA(arg) _Pragma(arg)
+#  else
+#    define FMT_GCC_PRAGMA(arg)
+#  endif
+#endif
+
+#ifdef __ICL
+#  define FMT_ICC_VERSION __ICL
+#elif defined(__INTEL_COMPILER)
+#  define FMT_ICC_VERSION __INTEL_COMPILER
+#else
+#  define FMT_ICC_VERSION 0
+#endif
+
+#ifdef _MSC_VER
+#  define FMT_MSC_VERSION _MSC_VER
+#  define FMT_MSC_WARNING(...) __pragma(warning(__VA_ARGS__))
+#else
+#  define FMT_MSC_VERSION 0
+#  define FMT_MSC_WARNING(...)
+#endif
+
+#ifdef _MSVC_LANG
+#  define FMT_CPLUSPLUS _MSVC_LANG
+#else
+#  define FMT_CPLUSPLUS __cplusplus
+#endif
+
+#ifdef __has_feature
+#  define FMT_HAS_FEATURE(x) __has_feature(x)
+#else
+#  define FMT_HAS_FEATURE(x) 0
+#endif
+
+#if defined(__has_include) || FMT_ICC_VERSION >= 1600 || FMT_MSC_VERSION > 1900
+#  define FMT_HAS_INCLUDE(x) __has_include(x)
+#else
+#  define FMT_HAS_INCLUDE(x) 0
+#endif
+
+#ifdef __has_cpp_attribute
+#  define FMT_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+#  define FMT_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+#define FMT_HAS_CPP14_ATTRIBUTE(attribute) \
+  (FMT_CPLUSPLUS >= 201402L && FMT_HAS_CPP_ATTRIBUTE(attribute))
+
+#define FMT_HAS_CPP17_ATTRIBUTE(attribute) \
+  (FMT_CPLUSPLUS >= 201703L && FMT_HAS_CPP_ATTRIBUTE(attribute))
+
+// Check if relaxed C++14 constexpr is supported.
+// GCC doesn't allow throw in constexpr until version 6 (bug 67371).
+#ifndef FMT_USE_CONSTEXPR
+#  if (FMT_HAS_FEATURE(cxx_relaxed_constexpr) || FMT_MSC_VERSION >= 1912 || \
+       (FMT_GCC_VERSION >= 600 && FMT_CPLUSPLUS >= 201402L)) &&             \
+      !FMT_ICC_VERSION && (!defined(__NVCC__) || FMT_CPLUSPLUS >= 202002L)
+#    define FMT_USE_CONSTEXPR 1
+#  else
+#    define FMT_USE_CONSTEXPR 0
+#  endif
+#endif
+#if FMT_USE_CONSTEXPR
+#  define FMT_CONSTEXPR constexpr
+#else
+#  define FMT_CONSTEXPR
+#endif
+
+#if (FMT_CPLUSPLUS >= 202002L ||                                \
+     (FMT_CPLUSPLUS >= 201709L && FMT_GCC_VERSION >= 1002)) &&  \
+    ((!defined(_GLIBCXX_RELEASE) || _GLIBCXX_RELEASE >= 10) &&  \
+     (!defined(_LIBCPP_VERSION) || _LIBCPP_VERSION >= 10000) && \
+     (!FMT_MSC_VERSION || FMT_MSC_VERSION >= 1928)) &&          \
+    defined(__cpp_lib_is_constant_evaluated)
+#  define FMT_CONSTEXPR20 constexpr
+#else
+#  define FMT_CONSTEXPR20
+#endif
+
+// Check if constexpr std::char_traits<>::{compare,length} are supported.
+#if defined(__GLIBCXX__)
+#  if FMT_CPLUSPLUS >= 201703L && defined(_GLIBCXX_RELEASE) && \
+      _GLIBCXX_RELEASE >= 7  // GCC 7+ libstdc++ has _GLIBCXX_RELEASE.
+#    define FMT_CONSTEXPR_CHAR_TRAITS constexpr
+#  endif
+#elif defined(_LIBCPP_VERSION) && FMT_CPLUSPLUS >= 201703L && \
+    _LIBCPP_VERSION >= 4000
+#  define FMT_CONSTEXPR_CHAR_TRAITS constexpr
+#elif FMT_MSC_VERSION >= 1914 && FMT_CPLUSPLUS >= 201703L
+#  define FMT_CONSTEXPR_CHAR_TRAITS constexpr
+#endif
+#ifndef FMT_CONSTEXPR_CHAR_TRAITS
+#  define FMT_CONSTEXPR_CHAR_TRAITS
+#endif
+
+// Check if exceptions are disabled.
+#ifndef FMT_EXCEPTIONS
+#  if (defined(__GNUC__) && !defined(__EXCEPTIONS)) || \
+      (FMT_MSC_VERSION && !_HAS_EXCEPTIONS)
+#    define FMT_EXCEPTIONS 0
+#  else
+#    define FMT_EXCEPTIONS 1
+#  endif
+#endif
+
+// Disable [[noreturn]] on MSVC/NVCC because of bogus unreachable code warnings.
+#if FMT_EXCEPTIONS && FMT_HAS_CPP_ATTRIBUTE(noreturn) && !FMT_MSC_VERSION && \
+    !defined(__NVCC__)
+#  define FMT_NORETURN [[noreturn]]
+#else
+#  define FMT_NORETURN
+#endif
+
+#ifndef FMT_NODISCARD
+#  if FMT_HAS_CPP17_ATTRIBUTE(nodiscard)
+#    define FMT_NODISCARD [[nodiscard]]
+#  else
+#    define FMT_NODISCARD
+#  endif
+#endif
+
+#ifndef FMT_INLINE
+#  if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#    define FMT_INLINE inline __attribute__((always_inline))
+#  else
+#    define FMT_INLINE inline
+#  endif
+#endif
+
+#ifdef _MSC_VER
+#  define FMT_UNCHECKED_ITERATOR(It) \
+    using _Unchecked_type = It  // Mark iterator as checked.
+#else
+#  define FMT_UNCHECKED_ITERATOR(It) using unchecked_type = It
+#endif
+
+#ifndef FMT_BEGIN_NAMESPACE
+#  define FMT_BEGIN_NAMESPACE \
+    namespace fmt {           \
+    inline namespace v10 {
+#  define FMT_END_NAMESPACE \
+    }                       \
+    }
+#endif
+
+#ifndef FMT_EXPORT
+#  define FMT_EXPORT
+#  define FMT_BEGIN_EXPORT
+#  define FMT_END_EXPORT
+#endif
+
+#if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#  define FMT_VISIBILITY(value) __attribute__((visibility(value)))
+#else
+#  define FMT_VISIBILITY(value)
+#endif
+
+#if !defined(FMT_HEADER_ONLY) && defined(_WIN32)
+#  if defined(FMT_LIB_EXPORT)
+#    define FMT_API __declspec(dllexport)
+#  elif defined(FMT_SHARED)
+#    define FMT_API __declspec(dllimport)
+#  endif
+#elif defined(FMT_LIB_EXPORT) || defined(FMT_SHARED)
+#  define FMT_API FMT_VISIBILITY("default")
+#endif
+#ifndef FMT_API
+#  define FMT_API
+#endif
+
+// libc++ supports string_view in pre-c++17.
+#if FMT_HAS_INCLUDE(<string_view>) && \
+    (FMT_CPLUSPLUS >= 201703L || defined(_LIBCPP_VERSION))
+#  include <string_view>
+#  define FMT_USE_STRING_VIEW
+#elif FMT_HAS_INCLUDE("experimental/string_view") && FMT_CPLUSPLUS >= 201402L
+#  include <experimental/string_view>
+#  define FMT_USE_EXPERIMENTAL_STRING_VIEW
+#endif
+
+#ifndef FMT_UNICODE
+#  define FMT_UNICODE !FMT_MSC_VERSION
+#endif
+
+#ifndef FMT_CONSTEVAL
+#  if ((FMT_GCC_VERSION >= 1000 || FMT_CLANG_VERSION >= 1101) && \
+       (!defined(__apple_build_version__) ||                     \
+        __apple_build_version__ >= 14000029L) &&                 \
+       FMT_CPLUSPLUS >= 202002L) ||                              \
+      (defined(__cpp_consteval) &&                               \
+       (!FMT_MSC_VERSION || FMT_MSC_VERSION >= 1929))
+// consteval is broken in MSVC before VS2019 version 16.10 and Apple clang
+// before 14.
+#    define FMT_CONSTEVAL consteval
+#    define FMT_HAS_CONSTEVAL
+#  else
+#    define FMT_CONSTEVAL
+#  endif
+#endif
+
+#ifndef FMT_USE_NONTYPE_TEMPLATE_ARGS
+#  if defined(__cpp_nontype_template_args) &&                  \
+      ((FMT_GCC_VERSION >= 903 && FMT_CPLUSPLUS >= 201709L) || \
+       __cpp_nontype_template_args >= 201911L) &&              \
+      !defined(__NVCOMPILER) && !defined(__LCC__)
+#    define FMT_USE_NONTYPE_TEMPLATE_ARGS 1
+#  else
+#    define FMT_USE_NONTYPE_TEMPLATE_ARGS 0
+#  endif
+#endif
+
+// GCC < 5 requires this-> in decltype
+#ifndef FMT_DECLTYPE_THIS
+#  if FMT_GCC_VERSION && FMT_GCC_VERSION < 500
+#    define FMT_DECLTYPE_THIS this->
+#  else
+#    define FMT_DECLTYPE_THIS
+#  endif
+#endif
+
+// Enable minimal optimizations for more compact code in debug mode.
+FMT_GCC_PRAGMA("GCC push_options")
+#if !defined(__OPTIMIZE__) && !defined(__NVCOMPILER) && !defined(__LCC__) && \
+    !defined(__CUDACC__)
+FMT_GCC_PRAGMA("GCC optimize(\"Og\")")
+#endif
+
+FMT_BEGIN_NAMESPACE
+
+// Implementations of enable_if_t and other metafunctions for older systems.
+template <bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+template <bool B, typename T, typename F>
+using conditional_t = typename std::conditional<B, T, F>::type;
+template <bool B> using bool_constant = std::integral_constant<bool, B>;
+template <typename T>
+using remove_reference_t = typename std::remove_reference<T>::type;
+template <typename T>
+using remove_const_t = typename std::remove_const<T>::type;
+template <typename T>
+using remove_cvref_t = typename std::remove_cv<remove_reference_t<T>>::type;
+template <typename T> struct type_identity {
+  using type = T;
+};
+template <typename T> using type_identity_t = typename type_identity<T>::type;
+template <typename T>
+using underlying_t = typename std::underlying_type<T>::type;
+
+// Checks whether T is a container with contiguous storage.
+template <typename T> struct is_contiguous : std::false_type {};
+template <typename Char>
+struct is_contiguous<std::basic_string<Char>> : std::true_type {};
+
+struct monostate {
+  constexpr monostate() {}
+};
+
+// An enable_if helper to be used in template parameters which results in much
+// shorter symbols: https://godbolt.org/z/sWw4vP. Extra parentheses are needed
+// to workaround a bug in MSVC 2019 (see #1140 and #1186).
+#ifdef FMT_DOC
+#  define FMT_ENABLE_IF(...)
+#else
+#  define FMT_ENABLE_IF(...) fmt::enable_if_t<(__VA_ARGS__), int> = 0
+#endif
+
+// This is defined in core.h instead of format.h to avoid injecting in std.
+// It is a template to avoid undesirable implicit conversions to std::byte.
+#ifdef __cpp_lib_byte
+template <typename T, FMT_ENABLE_IF(std::is_same<T, std::byte>::value)>
+inline auto format_as(T b) -> unsigned char {
+  return static_cast<unsigned char>(b);
+}
+#endif
+
+namespace detail {
+// Suppresses "unused variable" warnings with the method described in
+// https://herbsutter.com/2009/10/18/mailbag-shutting-up-compiler-warnings/.
+// (void)var does not work on many Intel compilers.
+template <typename... T> FMT_CONSTEXPR void ignore_unused(const T&...) {}
+
+constexpr FMT_INLINE auto is_constant_evaluated(
+    bool default_value = false) noexcept -> bool {
+// Workaround for incompatibility between libstdc++ consteval-based
+// std::is_constant_evaluated() implementation and clang-14.
+// https://github.com/fmtlib/fmt/issues/3247
+#if FMT_CPLUSPLUS >= 202002L && defined(_GLIBCXX_RELEASE) && \
+    _GLIBCXX_RELEASE >= 12 &&                                \
+    (FMT_CLANG_VERSION >= 1400 && FMT_CLANG_VERSION < 1500)
+  ignore_unused(default_value);
+  return __builtin_is_constant_evaluated();
+#elif defined(__cpp_lib_is_constant_evaluated)
+  ignore_unused(default_value);
+  return std::is_constant_evaluated();
+#else
+  return default_value;
+#endif
+}
+
+// Suppresses "conditional expression is constant" warnings.
+template <typename T> constexpr FMT_INLINE auto const_check(T value) -> T {
+  return value;
+}
+
+FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
+                                      const char* message);
+
+#ifndef FMT_ASSERT
+#  ifdef NDEBUG
+// FMT_ASSERT is not empty to avoid -Wempty-body.
+#    define FMT_ASSERT(condition, message) \
+      fmt::detail::ignore_unused((condition), (message))
+#  else
+#    define FMT_ASSERT(condition, message)                                    \
+      ((condition) /* void() fails with -Winvalid-constexpr on clang 4.0.1 */ \
+           ? (void)0                                                          \
+           : fmt::detail::assert_fail(__FILE__, __LINE__, (message)))
+#  endif
+#endif
+
+#if defined(FMT_USE_STRING_VIEW)
+template <typename Char> using std_string_view = std::basic_string_view<Char>;
+#elif defined(FMT_USE_EXPERIMENTAL_STRING_VIEW)
+template <typename Char>
+using std_string_view = std::experimental::basic_string_view<Char>;
+#else
+template <typename T> struct std_string_view {};
+#endif
+
+#ifdef FMT_USE_INT128
+// Do nothing.
+#elif defined(__SIZEOF_INT128__) && !defined(__NVCC__) && \
+    !(FMT_CLANG_VERSION && FMT_MSC_VERSION)
+#  define FMT_USE_INT128 1
+using int128_opt = __int128_t;  // An optional native 128-bit integer.
+using uint128_opt = __uint128_t;
+template <typename T> inline auto convert_for_visit(T value) -> T {
+  return value;
+}
+#else
+#  define FMT_USE_INT128 0
+#endif
+#if !FMT_USE_INT128
+enum class int128_opt {};
+enum class uint128_opt {};
+// Reduce template instantiations.
+template <typename T> auto convert_for_visit(T) -> monostate { return {}; }
+#endif
+
+// Casts a nonnegative integer to unsigned.
+template <typename Int>
+FMT_CONSTEXPR auto to_unsigned(Int value) ->
+    typename std::make_unsigned<Int>::type {
+  FMT_ASSERT(std::is_unsigned<Int>::value || value >= 0, "negative value");
+  return static_cast<typename std::make_unsigned<Int>::type>(value);
+}
+
+FMT_CONSTEXPR inline auto is_utf8() -> bool {
+  FMT_MSC_WARNING(suppress : 4566) constexpr unsigned char section[] = "\u00A7";
+
+  // Avoid buggy sign extensions in MSVC's constant evaluation mode (#2297).
+  using uchar = unsigned char;
+  return FMT_UNICODE || (sizeof(section) == 3 && uchar(section[0]) == 0xC2 &&
+                         uchar(section[1]) == 0xA7);
+}
+}  // namespace detail
+
+/**
+  An implementation of ``std::basic_string_view`` for pre-C++17. It provides a
+  subset of the API. ``fmt::basic_string_view`` is used for format strings even
+  if ``std::string_view`` is available to prevent issues when a library is
+  compiled with a different ``-std`` option than the client code (which is not
+  recommended).
+ */
+FMT_EXPORT
+template <typename Char> class basic_string_view {
+ private:
+  const Char* data_;
+  size_t size_;
+
+ public:
+  using value_type = Char;
+  using iterator = const Char*;
+
+  constexpr basic_string_view() noexcept : data_(nullptr), size_(0) {}
+
+  /** Constructs a string reference object from a C string and a size. */
+  constexpr basic_string_view(const Char* s, size_t count) noexcept
+      : data_(s), size_(count) {}
+
+  /**
+    \rst
+    Constructs a string reference object from a C string computing
+    the size with ``std::char_traits<Char>::length``.
+    \endrst
+   */
+  FMT_CONSTEXPR_CHAR_TRAITS
+  FMT_INLINE
+  basic_string_view(const Char* s)
+      : data_(s),
+        size_(detail::const_check(std::is_same<Char, char>::value &&
+                                  !detail::is_constant_evaluated(true))
+                  ? std::strlen(reinterpret_cast<const char*>(s))
+                  : std::char_traits<Char>::length(s)) {}
+
+  /** Constructs a string reference from a ``std::basic_string`` object. */
+  template <typename Traits, typename Alloc>
+  FMT_CONSTEXPR basic_string_view(
+      const std::basic_string<Char, Traits, Alloc>& s) noexcept
+      : data_(s.data()), size_(s.size()) {}
+
+  template <typename S, FMT_ENABLE_IF(std::is_same<
+                                      S, detail::std_string_view<Char>>::value)>
+  FMT_CONSTEXPR basic_string_view(S s) noexcept
+      : data_(s.data()), size_(s.size()) {}
+
+  /** Returns a pointer to the string data. */
+  constexpr auto data() const noexcept -> const Char* { return data_; }
+
+  /** Returns the string size. */
+  constexpr auto size() const noexcept -> size_t { return size_; }
+
+  constexpr auto begin() const noexcept -> iterator { return data_; }
+  constexpr auto end() const noexcept -> iterator { return data_ + size_; }
+
+  constexpr auto operator[](size_t pos) const noexcept -> const Char& {
+    return data_[pos];
+  }
+
+  FMT_CONSTEXPR void remove_prefix(size_t n) noexcept {
+    data_ += n;
+    size_ -= n;
+  }
+
+  FMT_CONSTEXPR_CHAR_TRAITS auto starts_with(
+      basic_string_view<Char> sv) const noexcept -> bool {
+    return size_ >= sv.size_ &&
+           std::char_traits<Char>::compare(data_, sv.data_, sv.size_) == 0;
+  }
+  FMT_CONSTEXPR_CHAR_TRAITS auto starts_with(Char c) const noexcept -> bool {
+    return size_ >= 1 && std::char_traits<Char>::eq(*data_, c);
+  }
+  FMT_CONSTEXPR_CHAR_TRAITS auto starts_with(const Char* s) const -> bool {
+    return starts_with(basic_string_view<Char>(s));
+  }
+
+  // Lexicographically compare this string reference to other.
+  FMT_CONSTEXPR_CHAR_TRAITS auto compare(basic_string_view other) const -> int {
+    size_t str_size = size_ < other.size_ ? size_ : other.size_;
+    int result = std::char_traits<Char>::compare(data_, other.data_, str_size);
+    if (result == 0)
+      result = size_ == other.size_ ? 0 : (size_ < other.size_ ? -1 : 1);
+    return result;
+  }
+
+  FMT_CONSTEXPR_CHAR_TRAITS friend auto operator==(basic_string_view lhs,
+                                                   basic_string_view rhs)
+      -> bool {
+    return lhs.compare(rhs) == 0;
+  }
+  friend auto operator!=(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) != 0;
+  }
+  friend auto operator<(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) < 0;
+  }
+  friend auto operator<=(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) <= 0;
+  }
+  friend auto operator>(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) > 0;
+  }
+  friend auto operator>=(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) >= 0;
+  }
+};
+
+FMT_EXPORT
+using string_view = basic_string_view<char>;
+
+/** Specifies if ``T`` is a character type. Can be specialized by users. */
+FMT_EXPORT
+template <typename T> struct is_char : std::false_type {};
+template <> struct is_char<char> : std::true_type {};
+
+namespace detail {
+
+// A base class for compile-time strings.
+struct compile_string {};
+
+template <typename S>
+struct is_compile_string : std::is_base_of<compile_string, S> {};
+
+template <typename Char, FMT_ENABLE_IF(is_char<Char>::value)>
+FMT_INLINE auto to_string_view(const Char* s) -> basic_string_view<Char> {
+  return s;
+}
+template <typename Char, typename Traits, typename Alloc>
+inline auto to_string_view(const std::basic_string<Char, Traits, Alloc>& s)
+    -> basic_string_view<Char> {
+  return s;
+}
+template <typename Char>
+constexpr auto to_string_view(basic_string_view<Char> s)
+    -> basic_string_view<Char> {
+  return s;
+}
+template <typename Char,
+          FMT_ENABLE_IF(!std::is_empty<std_string_view<Char>>::value)>
+inline auto to_string_view(std_string_view<Char> s) -> basic_string_view<Char> {
+  return s;
+}
+template <typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
+constexpr auto to_string_view(const S& s)
+    -> basic_string_view<typename S::char_type> {
+  return basic_string_view<typename S::char_type>(s);
+}
+void to_string_view(...);
+
+// Specifies whether S is a string type convertible to fmt::basic_string_view.
+// It should be a constexpr function but MSVC 2017 fails to compile it in
+// enable_if and MSVC 2015 fails to compile it as an alias template.
+// ADL is intentionally disabled as to_string_view is not an extension point.
+template <typename S>
+struct is_string
+    : std::is_class<decltype(detail::to_string_view(std::declval<S>()))> {};
+
+template <typename S, typename = void> struct char_t_impl {};
+template <typename S> struct char_t_impl<S, enable_if_t<is_string<S>::value>> {
+  using result = decltype(to_string_view(std::declval<S>()));
+  using type = typename result::value_type;
+};
+
+enum class type {
+  none_type,
+  // Integer types should go first,
+  int_type,
+  uint_type,
+  long_long_type,
+  ulong_long_type,
+  int128_type,
+  uint128_type,
+  bool_type,
+  char_type,
+  last_integer_type = char_type,
+  // followed by floating-point types.
+  float_type,
+  double_type,
+  long_double_type,
+  last_numeric_type = long_double_type,
+  cstring_type,
+  string_type,
+  pointer_type,
+  custom_type
+};
+
+// Maps core type T to the corresponding type enum constant.
+template <typename T, typename Char>
+struct type_constant : std::integral_constant<type, type::custom_type> {};
+
+#define FMT_TYPE_CONSTANT(Type, constant) \
+  template <typename Char>                \
+  struct type_constant<Type, Char>        \
+      : std::integral_constant<type, type::constant> {}
+
+FMT_TYPE_CONSTANT(int, int_type);
+FMT_TYPE_CONSTANT(unsigned, uint_type);
+FMT_TYPE_CONSTANT(long long, long_long_type);
+FMT_TYPE_CONSTANT(unsigned long long, ulong_long_type);
+FMT_TYPE_CONSTANT(int128_opt, int128_type);
+FMT_TYPE_CONSTANT(uint128_opt, uint128_type);
+FMT_TYPE_CONSTANT(bool, bool_type);
+FMT_TYPE_CONSTANT(Char, char_type);
+FMT_TYPE_CONSTANT(float, float_type);
+FMT_TYPE_CONSTANT(double, double_type);
+FMT_TYPE_CONSTANT(long double, long_double_type);
+FMT_TYPE_CONSTANT(const Char*, cstring_type);
+FMT_TYPE_CONSTANT(basic_string_view<Char>, string_type);
+FMT_TYPE_CONSTANT(const void*, pointer_type);
+
+constexpr auto is_integral_type(type t) -> bool {
+  return t > type::none_type && t <= type::last_integer_type;
+}
+constexpr auto is_arithmetic_type(type t) -> bool {
+  return t > type::none_type && t <= type::last_numeric_type;
+}
+
+constexpr auto set(type rhs) -> int { return 1 << static_cast<int>(rhs); }
+constexpr auto in(type t, int set) -> bool {
+  return ((set >> static_cast<int>(t)) & 1) != 0;
+}
+
+// Bitsets of types.
+enum {
+  sint_set =
+      set(type::int_type) | set(type::long_long_type) | set(type::int128_type),
+  uint_set = set(type::uint_type) | set(type::ulong_long_type) |
+             set(type::uint128_type),
+  bool_set = set(type::bool_type),
+  char_set = set(type::char_type),
+  float_set = set(type::float_type) | set(type::double_type) |
+              set(type::long_double_type),
+  string_set = set(type::string_type),
+  cstring_set = set(type::cstring_type),
+  pointer_set = set(type::pointer_type)
+};
+
+// DEPRECATED!
+FMT_NORETURN FMT_API void throw_format_error(const char* message);
+
+struct error_handler {
+  constexpr error_handler() = default;
+
+  // This function is intentionally not constexpr to give a compile-time error.
+  FMT_NORETURN void on_error(const char* message) {
+    throw_format_error(message);
+  }
+};
+}  // namespace detail
+
+/** Throws ``format_error`` with a given message. */
+using detail::throw_format_error;
+
+/** String's character type. */
+template <typename S> using char_t = typename detail::char_t_impl<S>::type;
+
+/**
+  \rst
+  Parsing context consisting of a format string range being parsed and an
+  argument counter for automatic indexing.
+  You can use the ``format_parse_context`` type alias for ``char`` instead.
+  \endrst
+ */
+FMT_EXPORT
+template <typename Char> class basic_format_parse_context {
+ private:
+  basic_string_view<Char> format_str_;
+  int next_arg_id_;
+
+  FMT_CONSTEXPR void do_check_arg_id(int id);
+
+ public:
+  using char_type = Char;
+  using iterator = const Char*;
+
+  explicit constexpr basic_format_parse_context(
+      basic_string_view<Char> format_str, int next_arg_id = 0)
+      : format_str_(format_str), next_arg_id_(next_arg_id) {}
+
+  /**
+    Returns an iterator to the beginning of the format string range being
+    parsed.
+   */
+  constexpr auto begin() const noexcept -> iterator {
+    return format_str_.begin();
+  }
+
+  /**
+    Returns an iterator past the end of the format string range being parsed.
+   */
+  constexpr auto end() const noexcept -> iterator { return format_str_.end(); }
+
+  /** Advances the begin iterator to ``it``. */
+  FMT_CONSTEXPR void advance_to(iterator it) {
+    format_str_.remove_prefix(detail::to_unsigned(it - begin()));
+  }
+
+  /**
+    Reports an error if using the manual argument indexing; otherwise returns
+    the next argument index and switches to the automatic indexing.
+   */
+  FMT_CONSTEXPR auto next_arg_id() -> int {
+    if (next_arg_id_ < 0) {
+      detail::throw_format_error(
+          "cannot switch from manual to automatic argument indexing");
+      return 0;
+    }
+    int id = next_arg_id_++;
+    do_check_arg_id(id);
+    return id;
+  }
+
+  /**
+    Reports an error if using the automatic argument indexing; otherwise
+    switches to the manual indexing.
+   */
+  FMT_CONSTEXPR void check_arg_id(int id) {
+    if (next_arg_id_ > 0) {
+      detail::throw_format_error(
+          "cannot switch from automatic to manual argument indexing");
+      return;
+    }
+    next_arg_id_ = -1;
+    do_check_arg_id(id);
+  }
+  FMT_CONSTEXPR void check_arg_id(basic_string_view<Char>) {}
+  FMT_CONSTEXPR void check_dynamic_spec(int arg_id);
+};
+
+FMT_EXPORT
+using format_parse_context = basic_format_parse_context<char>;
+
+namespace detail {
+// A parse context with extra data used only in compile-time checks.
+template <typename Char>
+class compile_parse_context : public basic_format_parse_context<Char> {
+ private:
+  int num_args_;
+  const type* types_;
+  using base = basic_format_parse_context<Char>;
+
+ public:
+  explicit FMT_CONSTEXPR compile_parse_context(
+      basic_string_view<Char> format_str, int num_args, const type* types,
+      int next_arg_id = 0)
+      : base(format_str, next_arg_id), num_args_(num_args), types_(types) {}
+
+  constexpr auto num_args() const -> int { return num_args_; }
+  constexpr auto arg_type(int id) const -> type { return types_[id]; }
+
+  FMT_CONSTEXPR auto next_arg_id() -> int {
+    int id = base::next_arg_id();
+    if (id >= num_args_) throw_format_error("argument not found");
+    return id;
+  }
+
+  FMT_CONSTEXPR void check_arg_id(int id) {
+    base::check_arg_id(id);
+    if (id >= num_args_) throw_format_error("argument not found");
+  }
+  using base::check_arg_id;
+
+  FMT_CONSTEXPR void check_dynamic_spec(int arg_id) {
+    detail::ignore_unused(arg_id);
+#if !defined(__LCC__)
+    if (arg_id < num_args_ && types_ && !is_integral_type(types_[arg_id]))
+      throw_format_error("width/precision is not integer");
+#endif
+  }
+};
+
+// Extracts a reference to the container from back_insert_iterator.
+template <typename Container>
+inline auto get_container(std::back_insert_iterator<Container> it)
+    -> Container& {
+  using base = std::back_insert_iterator<Container>;
+  struct accessor : base {
+    accessor(base b) : base(b) {}
+    using base::container;
+  };
+  return *accessor(it).container;
+}
+
+template <typename Char, typename InputIt, typename OutputIt>
+FMT_CONSTEXPR auto copy_str(InputIt begin, InputIt end, OutputIt out)
+    -> OutputIt {
+  while (begin != end) *out++ = static_cast<Char>(*begin++);
+  return out;
+}
+
+template <typename Char, typename T, typename U,
+          FMT_ENABLE_IF(
+              std::is_same<remove_const_t<T>, U>::value&& is_char<U>::value)>
+FMT_CONSTEXPR auto copy_str(T* begin, T* end, U* out) -> U* {
+  if (is_constant_evaluated()) return copy_str<Char, T*, U*>(begin, end, out);
+  auto size = to_unsigned(end - begin);
+  if (size > 0) memcpy(out, begin, size * sizeof(U));
+  return out + size;
+}
+
+/**
+  \rst
+  A contiguous memory buffer with an optional growing ability. It is an internal
+  class and shouldn't be used directly, only via `~fmt::basic_memory_buffer`.
+  \endrst
+ */
+template <typename T> class buffer {
+ private:
+  T* ptr_;
+  size_t size_;
+  size_t capacity_;
+
+ protected:
+  // Don't initialize ptr_ since it is not accessed to save a few cycles.
+  FMT_MSC_WARNING(suppress : 26495)
+  FMT_CONSTEXPR buffer(size_t sz) noexcept : size_(sz), capacity_(sz) {}
+
+  FMT_CONSTEXPR20 buffer(T* p = nullptr, size_t sz = 0, size_t cap = 0) noexcept
+      : ptr_(p), size_(sz), capacity_(cap) {}
+
+  FMT_CONSTEXPR20 ~buffer() = default;
+  buffer(buffer&&) = default;
+
+  /** Sets the buffer data and capacity. */
+  FMT_CONSTEXPR void set(T* buf_data, size_t buf_capacity) noexcept {
+    ptr_ = buf_data;
+    capacity_ = buf_capacity;
+  }
+
+  /** Increases the buffer capacity to hold at least *capacity* elements. */
+  // DEPRECATED!
+  virtual FMT_CONSTEXPR20 void grow(size_t capacity) = 0;
+
+ public:
+  using value_type = T;
+  using const_reference = const T&;
+
+  buffer(const buffer&) = delete;
+  void operator=(const buffer&) = delete;
+
+  FMT_INLINE auto begin() noexcept -> T* { return ptr_; }
+  FMT_INLINE auto end() noexcept -> T* { return ptr_ + size_; }
+
+  FMT_INLINE auto begin() const noexcept -> const T* { return ptr_; }
+  FMT_INLINE auto end() const noexcept -> const T* { return ptr_ + size_; }
+
+  /** Returns the size of this buffer. */
+  constexpr auto size() const noexcept -> size_t { return size_; }
+
+  /** Returns the capacity of this buffer. */
+  constexpr auto capacity() const noexcept -> size_t { return capacity_; }
+
+  /** Returns a pointer to the buffer data (not null-terminated). */
+  FMT_CONSTEXPR auto data() noexcept -> T* { return ptr_; }
+  FMT_CONSTEXPR auto data() const noexcept -> const T* { return ptr_; }
+
+  /** Clears this buffer. */
+  void clear() { size_ = 0; }
+
+  // Tries resizing the buffer to contain *count* elements. If T is a POD type
+  // the new elements may not be initialized.
+  FMT_CONSTEXPR20 void try_resize(size_t count) {
+    try_reserve(count);
+    size_ = count <= capacity_ ? count : capacity_;
+  }
+
+  // Tries increasing the buffer capacity to *new_capacity*. It can increase the
+  // capacity by a smaller amount than requested but guarantees there is space
+  // for at least one additional element either by increasing the capacity or by
+  // flushing the buffer if it is full.
+  FMT_CONSTEXPR20 void try_reserve(size_t new_capacity) {
+    if (new_capacity > capacity_) grow(new_capacity);
+  }
+
+  FMT_CONSTEXPR20 void push_back(const T& value) {
+    try_reserve(size_ + 1);
+    ptr_[size_++] = value;
+  }
+
+  /** Appends data to the end of the buffer. */
+  template <typename U> void append(const U* begin, const U* end);
+
+  template <typename Idx> FMT_CONSTEXPR auto operator[](Idx index) -> T& {
+    return ptr_[index];
+  }
+  template <typename Idx>
+  FMT_CONSTEXPR auto operator[](Idx index) const -> const T& {
+    return ptr_[index];
+  }
+};
+
+struct buffer_traits {
+  explicit buffer_traits(size_t) {}
+  auto count() const -> size_t { return 0; }
+  auto limit(size_t size) -> size_t { return size; }
+};
+
+class fixed_buffer_traits {
+ private:
+  size_t count_ = 0;
+  size_t limit_;
+
+ public:
+  explicit fixed_buffer_traits(size_t limit) : limit_(limit) {}
+  auto count() const -> size_t { return count_; }
+  auto limit(size_t size) -> size_t {
+    size_t n = limit_ > count_ ? limit_ - count_ : 0;
+    count_ += size;
+    return size < n ? size : n;
+  }
+};
+
+// A buffer that writes to an output iterator when flushed.
+template <typename OutputIt, typename T, typename Traits = buffer_traits>
+class iterator_buffer final : public Traits, public buffer<T> {
+ private:
+  OutputIt out_;
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t) override {
+    if (this->size() == buffer_size) flush();
+  }
+
+  void flush() {
+    auto size = this->size();
+    this->clear();
+    out_ = copy_str<T>(data_, data_ + this->limit(size), out_);
+  }
+
+ public:
+  explicit iterator_buffer(OutputIt out, size_t n = buffer_size)
+      : Traits(n), buffer<T>(data_, 0, buffer_size), out_(out) {}
+  iterator_buffer(iterator_buffer&& other)
+      : Traits(other), buffer<T>(data_, 0, buffer_size), out_(other.out_) {}
+  ~iterator_buffer() { flush(); }
+
+  auto out() -> OutputIt {
+    flush();
+    return out_;
+  }
+  auto count() const -> size_t { return Traits::count() + this->size(); }
+};
+
+template <typename T>
+class iterator_buffer<T*, T, fixed_buffer_traits> final
+    : public fixed_buffer_traits,
+      public buffer<T> {
+ private:
+  T* out_;
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t) override {
+    if (this->size() == this->capacity()) flush();
+  }
+
+  void flush() {
+    size_t n = this->limit(this->size());
+    if (this->data() == out_) {
+      out_ += n;
+      this->set(data_, buffer_size);
+    }
+    this->clear();
+  }
+
+ public:
+  explicit iterator_buffer(T* out, size_t n = buffer_size)
+      : fixed_buffer_traits(n), buffer<T>(out, 0, n), out_(out) {}
+  iterator_buffer(iterator_buffer&& other)
+      : fixed_buffer_traits(other),
+        buffer<T>(std::move(other)),
+        out_(other.out_) {
+    if (this->data() != out_) {
+      this->set(data_, buffer_size);
+      this->clear();
+    }
+  }
+  ~iterator_buffer() { flush(); }
+
+  auto out() -> T* {
+    flush();
+    return out_;
+  }
+  auto count() const -> size_t {
+    return fixed_buffer_traits::count() + this->size();
+  }
+};
+
+template <typename T> class iterator_buffer<T*, T> final : public buffer<T> {
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t) override {}
+
+ public:
+  explicit iterator_buffer(T* out, size_t = 0) : buffer<T>(out, 0, ~size_t()) {}
+
+  auto out() -> T* { return &*this->end(); }
+};
+
+// A buffer that writes to a container with the contiguous storage.
+template <typename Container>
+class iterator_buffer<std::back_insert_iterator<Container>,
+                      enable_if_t<is_contiguous<Container>::value,
+                                  typename Container::value_type>>
+    final : public buffer<typename Container::value_type> {
+ private:
+  Container& container_;
+
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t capacity) override {
+    container_.resize(capacity);
+    this->set(&container_[0], capacity);
+  }
+
+ public:
+  explicit iterator_buffer(Container& c)
+      : buffer<typename Container::value_type>(c.size()), container_(c) {}
+  explicit iterator_buffer(std::back_insert_iterator<Container> out, size_t = 0)
+      : iterator_buffer(get_container(out)) {}
+
+  auto out() -> std::back_insert_iterator<Container> {
+    return std::back_inserter(container_);
+  }
+};
+
+// A buffer that counts the number of code units written discarding the output.
+template <typename T = char> class counting_buffer final : public buffer<T> {
+ private:
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+  size_t count_ = 0;
+
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t) override {
+    if (this->size() != buffer_size) return;
+    count_ += this->size();
+    this->clear();
+  }
+
+ public:
+  counting_buffer() : buffer<T>(data_, 0, buffer_size) {}
+
+  auto count() -> size_t { return count_ + this->size(); }
+};
+}  // namespace detail
+
+template <typename Char>
+FMT_CONSTEXPR void basic_format_parse_context<Char>::do_check_arg_id(int id) {
+  // Argument id is only checked at compile-time during parsing because
+  // formatting has its own validation.
+  if (detail::is_constant_evaluated() &&
+      (!FMT_GCC_VERSION || FMT_GCC_VERSION >= 1200)) {
+    using context = detail::compile_parse_context<Char>;
+    if (id >= static_cast<context*>(this)->num_args())
+      detail::throw_format_error("argument not found");
+  }
+}
+
+template <typename Char>
+FMT_CONSTEXPR void basic_format_parse_context<Char>::check_dynamic_spec(
+    int arg_id) {
+  if (detail::is_constant_evaluated() &&
+      (!FMT_GCC_VERSION || FMT_GCC_VERSION >= 1200)) {
+    using context = detail::compile_parse_context<Char>;
+    static_cast<context*>(this)->check_dynamic_spec(arg_id);
+  }
+}
+
+FMT_EXPORT template <typename Context> class basic_format_arg;
+FMT_EXPORT template <typename Context> class basic_format_args;
+FMT_EXPORT template <typename Context> class dynamic_format_arg_store;
+
+// A formatter for objects of type T.
+FMT_EXPORT
+template <typename T, typename Char = char, typename Enable = void>
+struct formatter {
+  // A deleted default constructor indicates a disabled formatter.
+  formatter() = delete;
+};
+
+// Specifies if T has an enabled formatter specialization. A type can be
+// formattable even if it doesn't have a formatter e.g. via a conversion.
+template <typename T, typename Context>
+using has_formatter =
+    std::is_constructible<typename Context::template formatter_type<T>>;
+
+// An output iterator that appends to a buffer.
+// It is used to reduce symbol sizes for the common case.
+class appender : public std::back_insert_iterator<detail::buffer<char>> {
+  using base = std::back_insert_iterator<detail::buffer<char>>;
+
+ public:
+  using std::back_insert_iterator<detail::buffer<char>>::back_insert_iterator;
+  appender(base it) noexcept : base(it) {}
+  FMT_UNCHECKED_ITERATOR(appender);
+
+  auto operator++() noexcept -> appender& { return *this; }
+  auto operator++(int) noexcept -> appender { return *this; }
+};
+
+namespace detail {
+
+template <typename Context, typename T>
+constexpr auto has_const_formatter_impl(T*)
+    -> decltype(typename Context::template formatter_type<T>().format(
+                    std::declval<const T&>(), std::declval<Context&>()),
+                true) {
+  return true;
+}
+template <typename Context>
+constexpr auto has_const_formatter_impl(...) -> bool {
+  return false;
+}
+template <typename T, typename Context>
+constexpr auto has_const_formatter() -> bool {
+  return has_const_formatter_impl<Context>(static_cast<T*>(nullptr));
+}
+
+template <typename T>
+using buffer_appender = conditional_t<std::is_same<T, char>::value, appender,
+                                      std::back_insert_iterator<buffer<T>>>;
+
+// Maps an output iterator to a buffer.
+template <typename T, typename OutputIt>
+auto get_buffer(OutputIt out) -> iterator_buffer<OutputIt, T> {
+  return iterator_buffer<OutputIt, T>(out);
+}
+template <typename T, typename Buf,
+          FMT_ENABLE_IF(std::is_base_of<buffer<char>, Buf>::value)>
+auto get_buffer(std::back_insert_iterator<Buf> out) -> buffer<char>& {
+  return get_container(out);
+}
+
+template <typename Buf, typename OutputIt>
+FMT_INLINE auto get_iterator(Buf& buf, OutputIt) -> decltype(buf.out()) {
+  return buf.out();
+}
+template <typename T, typename OutputIt>
+auto get_iterator(buffer<T>&, OutputIt out) -> OutputIt {
+  return out;
+}
+
+struct view {};
+
+template <typename Char, typename T> struct named_arg : view {
+  const Char* name;
+  const T& value;
+  named_arg(const Char* n, const T& v) : name(n), value(v) {}
+};
+
+template <typename Char> struct named_arg_info {
+  const Char* name;
+  int id;
+};
+
+template <typename T, typename Char, size_t NUM_ARGS, size_t NUM_NAMED_ARGS>
+struct arg_data {
+  // args_[0].named_args points to named_args_ to avoid bloating format_args.
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  T args_[1 + (NUM_ARGS != 0 ? NUM_ARGS : +1)];
+  named_arg_info<Char> named_args_[NUM_NAMED_ARGS];
+
+  template <typename... U>
+  arg_data(const U&... init) : args_{T(named_args_, NUM_NAMED_ARGS), init...} {}
+  arg_data(const arg_data& other) = delete;
+  auto args() const -> const T* { return args_ + 1; }
+  auto named_args() -> named_arg_info<Char>* { return named_args_; }
+};
+
+template <typename T, typename Char, size_t NUM_ARGS>
+struct arg_data<T, Char, NUM_ARGS, 0> {
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  T args_[NUM_ARGS != 0 ? NUM_ARGS : +1];
+
+  template <typename... U>
+  FMT_CONSTEXPR FMT_INLINE arg_data(const U&... init) : args_{init...} {}
+  FMT_CONSTEXPR FMT_INLINE auto args() const -> const T* { return args_; }
+  FMT_CONSTEXPR FMT_INLINE auto named_args() -> std::nullptr_t {
+    return nullptr;
+  }
+};
+
+template <typename Char>
+inline void init_named_args(named_arg_info<Char>*, int, int) {}
+
+template <typename T> struct is_named_arg : std::false_type {};
+template <typename T> struct is_statically_named_arg : std::false_type {};
+
+template <typename T, typename Char>
+struct is_named_arg<named_arg<Char, T>> : std::true_type {};
+
+template <typename Char, typename T, typename... Tail,
+          FMT_ENABLE_IF(!is_named_arg<T>::value)>
+void init_named_args(named_arg_info<Char>* named_args, int arg_count,
+                     int named_arg_count, const T&, const Tail&... args) {
+  init_named_args(named_args, arg_count + 1, named_arg_count, args...);
+}
+
+template <typename Char, typename T, typename... Tail,
+          FMT_ENABLE_IF(is_named_arg<T>::value)>
+void init_named_args(named_arg_info<Char>* named_args, int arg_count,
+                     int named_arg_count, const T& arg, const Tail&... args) {
+  named_args[named_arg_count++] = {arg.name, arg_count};
+  init_named_args(named_args, arg_count + 1, named_arg_count, args...);
+}
+
+template <typename... Args>
+FMT_CONSTEXPR FMT_INLINE void init_named_args(std::nullptr_t, int, int,
+                                              const Args&...) {}
+
+template <bool B = false> constexpr auto count() -> size_t { return B ? 1 : 0; }
+template <bool B1, bool B2, bool... Tail> constexpr auto count() -> size_t {
+  return (B1 ? 1 : 0) + count<B2, Tail...>();
+}
+
+template <typename... Args> constexpr auto count_named_args() -> size_t {
+  return count<is_named_arg<Args>::value...>();
+}
+
+template <typename... Args>
+constexpr auto count_statically_named_args() -> size_t {
+  return count<is_statically_named_arg<Args>::value...>();
+}
+
+struct unformattable {};
+struct unformattable_char : unformattable {};
+struct unformattable_pointer : unformattable {};
+
+template <typename Char> struct string_value {
+  const Char* data;
+  size_t size;
+};
+
+template <typename Char> struct named_arg_value {
+  const named_arg_info<Char>* data;
+  size_t size;
+};
+
+template <typename Context> struct custom_value {
+  using parse_context = typename Context::parse_context_type;
+  void* value;
+  void (*format)(void* arg, parse_context& parse_ctx, Context& ctx);
+};
+
+// A formatting argument value.
+template <typename Context> class value {
+ public:
+  using char_type = typename Context::char_type;
+
+  union {
+    monostate no_value;
+    int int_value;
+    unsigned uint_value;
+    long long long_long_value;
+    unsigned long long ulong_long_value;
+    int128_opt int128_value;
+    uint128_opt uint128_value;
+    bool bool_value;
+    char_type char_value;
+    float float_value;
+    double double_value;
+    long double long_double_value;
+    const void* pointer;
+    string_value<char_type> string;
+    custom_value<Context> custom;
+    named_arg_value<char_type> named_args;
+  };
+
+  constexpr FMT_INLINE value() : no_value() {}
+  constexpr FMT_INLINE value(int val) : int_value(val) {}
+  constexpr FMT_INLINE value(unsigned val) : uint_value(val) {}
+  constexpr FMT_INLINE value(long long val) : long_long_value(val) {}
+  constexpr FMT_INLINE value(unsigned long long val) : ulong_long_value(val) {}
+  FMT_INLINE value(int128_opt val) : int128_value(val) {}
+  FMT_INLINE value(uint128_opt val) : uint128_value(val) {}
+  constexpr FMT_INLINE value(float val) : float_value(val) {}
+  constexpr FMT_INLINE value(double val) : double_value(val) {}
+  FMT_INLINE value(long double val) : long_double_value(val) {}
+  constexpr FMT_INLINE value(bool val) : bool_value(val) {}
+  constexpr FMT_INLINE value(char_type val) : char_value(val) {}
+  FMT_CONSTEXPR FMT_INLINE value(const char_type* val) {
+    string.data = val;
+    if (is_constant_evaluated()) string.size = {};
+  }
+  FMT_CONSTEXPR FMT_INLINE value(basic_string_view<char_type> val) {
+    string.data = val.data();
+    string.size = val.size();
+  }
+  FMT_INLINE value(const void* val) : pointer(val) {}
+  FMT_INLINE value(const named_arg_info<char_type>* args, size_t size)
+      : named_args{args, size} {}
+
+  template <typename T> FMT_CONSTEXPR20 FMT_INLINE value(T& val) {
+    using value_type = remove_const_t<T>;
+    custom.value = const_cast<value_type*>(std::addressof(val));
+    // Get the formatter type through the context to allow different contexts
+    // have different extension points, e.g. `formatter<T>` for `format` and
+    // `printf_formatter<T>` for `printf`.
+    custom.format = format_custom_arg<
+        value_type, typename Context::template formatter_type<value_type>>;
+  }
+  value(unformattable);
+  value(unformattable_char);
+  value(unformattable_pointer);
+
+ private:
+  // Formats an argument of a custom type, such as a user-defined class.
+  template <typename T, typename Formatter>
+  static void format_custom_arg(void* arg,
+                                typename Context::parse_context_type& parse_ctx,
+                                Context& ctx) {
+    auto f = Formatter();
+    parse_ctx.advance_to(f.parse(parse_ctx));
+    using qualified_type =
+        conditional_t<has_const_formatter<T, Context>(), const T, T>;
+    // Calling format through a mutable reference is deprecated.
+    ctx.advance_to(f.format(*static_cast<qualified_type*>(arg), ctx));
+  }
+};
+
+// To minimize the number of types we need to deal with, long is translated
+// either to int or to long long depending on its size.
+enum { long_short = sizeof(long) == sizeof(int) };
+using long_type = conditional_t<long_short, int, long long>;
+using ulong_type = conditional_t<long_short, unsigned, unsigned long long>;
+
+template <typename T> struct format_as_result {
+  template <typename U,
+            FMT_ENABLE_IF(std::is_enum<U>::value || std::is_class<U>::value)>
+  static auto map(U*) -> remove_cvref_t<decltype(format_as(std::declval<U>()))>;
+  static auto map(...) -> void;
+
+  using type = decltype(map(static_cast<T*>(nullptr)));
+};
+template <typename T> using format_as_t = typename format_as_result<T>::type;
+
+template <typename T>
+struct has_format_as
+    : bool_constant<!std::is_same<format_as_t<T>, void>::value> {};
+
+// Maps formatting arguments to core types.
+// arg_mapper reports errors by returning unformattable instead of using
+// static_assert because it's used in the is_formattable trait.
+template <typename Context> struct arg_mapper {
+  using char_type = typename Context::char_type;
+
+  FMT_CONSTEXPR FMT_INLINE auto map(signed char val) -> int { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned char val) -> unsigned {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(short val) -> int { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned short val) -> unsigned {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(int val) -> int { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned val) -> unsigned { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(long val) -> long_type { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned long val) -> ulong_type {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(long long val) -> long long { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned long long val)
+      -> unsigned long long {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(int128_opt val) -> int128_opt {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(uint128_opt val) -> uint128_opt {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(bool val) -> bool { return val; }
+
+  template <typename T, FMT_ENABLE_IF(std::is_same<T, char>::value ||
+                                      std::is_same<T, char_type>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(T val) -> char_type {
+    return val;
+  }
+  template <typename T, enable_if_t<(std::is_same<T, wchar_t>::value ||
+#ifdef __cpp_char8_t
+                                     std::is_same<T, char8_t>::value ||
+#endif
+                                     std::is_same<T, char16_t>::value ||
+                                     std::is_same<T, char32_t>::value) &&
+                                        !std::is_same<T, char_type>::value,
+                                    int> = 0>
+  FMT_CONSTEXPR FMT_INLINE auto map(T) -> unformattable_char {
+    return {};
+  }
+
+  FMT_CONSTEXPR FMT_INLINE auto map(float val) -> float { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(double val) -> double { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(long double val) -> long double {
+    return val;
+  }
+
+  FMT_CONSTEXPR FMT_INLINE auto map(char_type* val) -> const char_type* {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(const char_type* val) -> const char_type* {
+    return val;
+  }
+  template <typename T,
+            FMT_ENABLE_IF(is_string<T>::value && !std::is_pointer<T>::value &&
+                          std::is_same<char_type, char_t<T>>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& val)
+      -> basic_string_view<char_type> {
+    return to_string_view(val);
+  }
+  template <typename T,
+            FMT_ENABLE_IF(is_string<T>::value && !std::is_pointer<T>::value &&
+                          !std::is_same<char_type, char_t<T>>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T&) -> unformattable_char {
+    return {};
+  }
+
+  FMT_CONSTEXPR FMT_INLINE auto map(void* val) -> const void* { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(const void* val) -> const void* {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(std::nullptr_t val) -> const void* {
+    return val;
+  }
+
+  // Use SFINAE instead of a const T* parameter to avoid a conflict with the
+  // array overload.
+  template <
+      typename T,
+      FMT_ENABLE_IF(
+          std::is_pointer<T>::value || std::is_member_pointer<T>::value ||
+          std::is_function<typename std::remove_pointer<T>::type>::value ||
+          (std::is_array<T>::value &&
+           !std::is_convertible<T, const char_type*>::value))>
+  FMT_CONSTEXPR auto map(const T&) -> unformattable_pointer {
+    return {};
+  }
+
+  template <typename T, std::size_t N,
+            FMT_ENABLE_IF(!std::is_same<T, wchar_t>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T (&values)[N]) -> const T (&)[N] {
+    return values;
+  }
+
+  // Only map owning types because mapping views can be unsafe.
+  template <typename T, typename U = format_as_t<T>,
+            FMT_ENABLE_IF(std::is_arithmetic<U>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& val)
+      -> decltype(FMT_DECLTYPE_THIS map(U())) {
+    return map(format_as(val));
+  }
+
+  template <typename T, typename U = remove_const_t<T>>
+  struct formattable : bool_constant<has_const_formatter<U, Context>() ||
+                                     (has_formatter<U, Context>::value &&
+                                      !std::is_const<T>::value)> {};
+
+  template <typename T, FMT_ENABLE_IF(formattable<T>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto do_map(T& val) -> T& {
+    return val;
+  }
+  template <typename T, FMT_ENABLE_IF(!formattable<T>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto do_map(T&) -> unformattable {
+    return {};
+  }
+
+  template <typename T, typename U = remove_const_t<T>,
+            FMT_ENABLE_IF((std::is_class<U>::value || std::is_enum<U>::value ||
+                           std::is_union<U>::value) &&
+                          !is_string<U>::value && !is_char<U>::value &&
+                          !is_named_arg<U>::value &&
+                          !std::is_arithmetic<format_as_t<U>>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(T& val)
+      -> decltype(FMT_DECLTYPE_THIS do_map(val)) {
+    return do_map(val);
+  }
+
+  template <typename T, FMT_ENABLE_IF(is_named_arg<T>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& named_arg)
+      -> decltype(FMT_DECLTYPE_THIS map(named_arg.value)) {
+    return map(named_arg.value);
+  }
+
+  auto map(...) -> unformattable { return {}; }
+};
+
+// A type constant after applying arg_mapper<Context>.
+template <typename T, typename Context>
+using mapped_type_constant =
+    type_constant<decltype(arg_mapper<Context>().map(std::declval<const T&>())),
+                  typename Context::char_type>;
+
+enum { packed_arg_bits = 4 };
+// Maximum number of arguments with packed types.
+enum { max_packed_args = 62 / packed_arg_bits };
+enum : unsigned long long { is_unpacked_bit = 1ULL << 63 };
+enum : unsigned long long { has_named_args_bit = 1ULL << 62 };
+
+template <typename Char, typename InputIt>
+auto copy_str(InputIt begin, InputIt end, appender out) -> appender {
+  get_container(out).append(begin, end);
+  return out;
+}
+template <typename Char, typename InputIt>
+auto copy_str(InputIt begin, InputIt end,
+              std::back_insert_iterator<std::string> out)
+    -> std::back_insert_iterator<std::string> {
+  get_container(out).append(begin, end);
+  return out;
+}
+
+template <typename Char, typename R, typename OutputIt>
+FMT_CONSTEXPR auto copy_str(R&& rng, OutputIt out) -> OutputIt {
+  return detail::copy_str<Char>(rng.begin(), rng.end(), out);
+}
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 500
+// A workaround for gcc 4.8 to make void_t work in a SFINAE context.
+template <typename...> struct void_t_impl {
+  using type = void;
+};
+template <typename... T> using void_t = typename void_t_impl<T...>::type;
+#else
+template <typename...> using void_t = void;
+#endif
+
+template <typename It, typename T, typename Enable = void>
+struct is_output_iterator : std::false_type {};
+
+template <typename It, typename T>
+struct is_output_iterator<
+    It, T,
+    void_t<typename std::iterator_traits<It>::iterator_category,
+           decltype(*std::declval<It>() = std::declval<T>())>>
+    : std::true_type {};
+
+template <typename It> struct is_back_insert_iterator : std::false_type {};
+template <typename Container>
+struct is_back_insert_iterator<std::back_insert_iterator<Container>>
+    : std::true_type {};
+
+// A type-erased reference to an std::locale to avoid a heavy <locale> include.
+class locale_ref {
+ private:
+  const void* locale_;  // A type-erased pointer to std::locale.
+
+ public:
+  constexpr FMT_INLINE locale_ref() : locale_(nullptr) {}
+  template <typename Locale> explicit locale_ref(const Locale& loc);
+
+  explicit operator bool() const noexcept { return locale_ != nullptr; }
+
+  template <typename Locale> auto get() const -> Locale;
+};
+
+template <typename> constexpr auto encode_types() -> unsigned long long {
+  return 0;
+}
+
+template <typename Context, typename Arg, typename... Args>
+constexpr auto encode_types() -> unsigned long long {
+  return static_cast<unsigned>(mapped_type_constant<Arg, Context>::value) |
+         (encode_types<Context, Args...>() << packed_arg_bits);
+}
+
+#if defined(__cpp_if_constexpr)
+// This type is intentionally undefined, only used for errors
+template <typename T, typename Char> struct type_is_unformattable_for;
+#endif
+
+template <bool PACKED, typename Context, typename T, FMT_ENABLE_IF(PACKED)>
+FMT_CONSTEXPR FMT_INLINE auto make_arg(T& val) -> value<Context> {
+  using arg_type = remove_cvref_t<decltype(arg_mapper<Context>().map(val))>;
+
+  constexpr bool formattable_char =
+      !std::is_same<arg_type, unformattable_char>::value;
+  static_assert(formattable_char, "Mixing character types is disallowed.");
+
+  // Formatting of arbitrary pointers is disallowed. If you want to format a
+  // pointer cast it to `void*` or `const void*`. In particular, this forbids
+  // formatting of `[const] volatile char*` printed as bool by iostreams.
+  constexpr bool formattable_pointer =
+      !std::is_same<arg_type, unformattable_pointer>::value;
+  static_assert(formattable_pointer,
+                "Formatting of non-void pointers is disallowed.");
+
+  constexpr bool formattable = !std::is_same<arg_type, unformattable>::value;
+#if defined(__cpp_if_constexpr)
+  if constexpr (!formattable) {
+    type_is_unformattable_for<T, typename Context::char_type> _;
+  }
+#endif
+  static_assert(
+      formattable,
+      "Cannot format an argument. To make type T formattable provide a "
+      "formatter<T> specialization: https://fmt.dev/latest/api.html#udt");
+  return {arg_mapper<Context>().map(val)};
+}
+
+template <typename Context, typename T>
+FMT_CONSTEXPR auto make_arg(T& val) -> basic_format_arg<Context> {
+  auto arg = basic_format_arg<Context>();
+  arg.type_ = mapped_type_constant<T, Context>::value;
+  arg.value_ = make_arg<true, Context>(val);
+  return arg;
+}
+
+template <bool PACKED, typename Context, typename T, FMT_ENABLE_IF(!PACKED)>
+FMT_CONSTEXPR inline auto make_arg(T& val) -> basic_format_arg<Context> {
+  return make_arg<Context>(val);
+}
+}  // namespace detail
+FMT_BEGIN_EXPORT
+
+// A formatting argument. Context is a template parameter for the compiled API
+// where output can be unbuffered.
+template <typename Context> class basic_format_arg {
+ private:
+  detail::value<Context> value_;
+  detail::type type_;
+
+  template <typename ContextType, typename T>
+  friend FMT_CONSTEXPR auto detail::make_arg(T& value)
+      -> basic_format_arg<ContextType>;
+
+  template <typename Visitor, typename Ctx>
+  friend FMT_CONSTEXPR auto visit_format_arg(Visitor&& vis,
+                                             const basic_format_arg<Ctx>& arg)
+      -> decltype(vis(0));
+
+  friend class basic_format_args<Context>;
+  friend class dynamic_format_arg_store<Context>;
+
+  using char_type = typename Context::char_type;
+
+  template <typename T, typename Char, size_t NUM_ARGS, size_t NUM_NAMED_ARGS>
+  friend struct detail::arg_data;
+
+  basic_format_arg(const detail::named_arg_info<char_type>* args, size_t size)
+      : value_(args, size) {}
+
+ public:
+  class handle {
+   public:
+    explicit handle(detail::custom_value<Context> custom) : custom_(custom) {}
+
+    void format(typename Context::parse_context_type& parse_ctx,
+                Context& ctx) const {
+      custom_.format(custom_.value, parse_ctx, ctx);
+    }
+
+   private:
+    detail::custom_value<Context> custom_;
+  };
+
+  constexpr basic_format_arg() : type_(detail::type::none_type) {}
+
+  constexpr explicit operator bool() const noexcept {
+    return type_ != detail::type::none_type;
+  }
+
+  auto type() const -> detail::type { return type_; }
+
+  auto is_integral() const -> bool { return detail::is_integral_type(type_); }
+  auto is_arithmetic() const -> bool {
+    return detail::is_arithmetic_type(type_);
+  }
+
+  FMT_INLINE auto format_custom(const char_type* parse_begin,
+                                typename Context::parse_context_type& parse_ctx,
+                                Context& ctx) -> bool {
+    if (type_ != detail::type::custom_type) return false;
+    parse_ctx.advance_to(parse_begin);
+    value_.custom.format(value_.custom.value, parse_ctx, ctx);
+    return true;
+  }
+};
+
+/**
+  \rst
+  Visits an argument dispatching to the appropriate visit method based on
+  the argument type. For example, if the argument type is ``double`` then
+  ``vis(value)`` will be called with the value of type ``double``.
+  \endrst
+ */
+// DEPRECATED!
+template <typename Visitor, typename Context>
+FMT_CONSTEXPR FMT_INLINE auto visit_format_arg(
+    Visitor&& vis, const basic_format_arg<Context>& arg) -> decltype(vis(0)) {
+  switch (arg.type_) {
+  case detail::type::none_type:
+    break;
+  case detail::type::int_type:
+    return vis(arg.value_.int_value);
+  case detail::type::uint_type:
+    return vis(arg.value_.uint_value);
+  case detail::type::long_long_type:
+    return vis(arg.value_.long_long_value);
+  case detail::type::ulong_long_type:
+    return vis(arg.value_.ulong_long_value);
+  case detail::type::int128_type:
+    return vis(detail::convert_for_visit(arg.value_.int128_value));
+  case detail::type::uint128_type:
+    return vis(detail::convert_for_visit(arg.value_.uint128_value));
+  case detail::type::bool_type:
+    return vis(arg.value_.bool_value);
+  case detail::type::char_type:
+    return vis(arg.value_.char_value);
+  case detail::type::float_type:
+    return vis(arg.value_.float_value);
+  case detail::type::double_type:
+    return vis(arg.value_.double_value);
+  case detail::type::long_double_type:
+    return vis(arg.value_.long_double_value);
+  case detail::type::cstring_type:
+    return vis(arg.value_.string.data);
+  case detail::type::string_type:
+    using sv = basic_string_view<typename Context::char_type>;
+    return vis(sv(arg.value_.string.data, arg.value_.string.size));
+  case detail::type::pointer_type:
+    return vis(arg.value_.pointer);
+  case detail::type::custom_type:
+    return vis(typename basic_format_arg<Context>::handle(arg.value_.custom));
+  }
+  return vis(monostate());
+}
+
+// Formatting context.
+template <typename OutputIt, typename Char> class basic_format_context {
+ private:
+  OutputIt out_;
+  basic_format_args<basic_format_context> args_;
+  detail::locale_ref loc_;
+
+ public:
+  using iterator = OutputIt;
+  using format_arg = basic_format_arg<basic_format_context>;
+  using format_args = basic_format_args<basic_format_context>;
+  using parse_context_type = basic_format_parse_context<Char>;
+  template <typename T> using formatter_type = formatter<T, Char>;
+
+  /** The character type for the output. */
+  using char_type = Char;
+
+  basic_format_context(basic_format_context&&) = default;
+  basic_format_context(const basic_format_context&) = delete;
+  void operator=(const basic_format_context&) = delete;
+  /**
+    Constructs a ``basic_format_context`` object. References to the arguments
+    are stored in the object so make sure they have appropriate lifetimes.
+   */
+  constexpr basic_format_context(OutputIt out, format_args ctx_args,
+                                 detail::locale_ref loc = {})
+      : out_(out), args_(ctx_args), loc_(loc) {}
+
+  constexpr auto arg(int id) const -> format_arg { return args_.get(id); }
+  FMT_CONSTEXPR auto arg(basic_string_view<Char> name) -> format_arg {
+    return args_.get(name);
+  }
+  FMT_CONSTEXPR auto arg_id(basic_string_view<Char> name) -> int {
+    return args_.get_id(name);
+  }
+  auto args() const -> const format_args& { return args_; }
+
+  // DEPRECATED!
+  FMT_CONSTEXPR auto error_handler() -> detail::error_handler { return {}; }
+  void on_error(const char* message) { error_handler().on_error(message); }
+
+  // Returns an iterator to the beginning of the output range.
+  FMT_CONSTEXPR auto out() -> iterator { return out_; }
+
+  // Advances the begin iterator to ``it``.
+  void advance_to(iterator it) {
+    if (!detail::is_back_insert_iterator<iterator>()) out_ = it;
+  }
+
+  FMT_CONSTEXPR auto locale() -> detail::locale_ref { return loc_; }
+};
+
+template <typename Char>
+using buffer_context =
+    basic_format_context<detail::buffer_appender<Char>, Char>;
+using format_context = buffer_context<char>;
+
+template <typename T, typename Char = char>
+using is_formattable = bool_constant<!std::is_base_of<
+    detail::unformattable, decltype(detail::arg_mapper<buffer_context<Char>>()
+                                        .map(std::declval<T&>()))>::value>;
+
+/**
+  \rst
+  An array of references to arguments. It can be implicitly converted into
+  `~fmt::basic_format_args` for passing into type-erased formatting functions
+  such as `~fmt::vformat`.
+  \endrst
+ */
+template <typename Context, typename... Args>
+class format_arg_store
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+    // Workaround a GCC template argument substitution bug.
+    : public basic_format_args<Context>
+#endif
+{
+ private:
+  static const size_t num_args = sizeof...(Args);
+  static constexpr size_t num_named_args = detail::count_named_args<Args...>();
+  static const bool is_packed = num_args <= detail::max_packed_args;
+
+  using value_type = conditional_t<is_packed, detail::value<Context>,
+                                   basic_format_arg<Context>>;
+
+  detail::arg_data<value_type, typename Context::char_type, num_args,
+                   num_named_args>
+      data_;
+
+  friend class basic_format_args<Context>;
+
+  static constexpr unsigned long long desc =
+      (is_packed ? detail::encode_types<Context, Args...>()
+                 : detail::is_unpacked_bit | num_args) |
+      (num_named_args != 0
+           ? static_cast<unsigned long long>(detail::has_named_args_bit)
+           : 0);
+
+ public:
+  template <typename... T>
+  FMT_CONSTEXPR FMT_INLINE format_arg_store(T&... args)
+      :
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+        basic_format_args<Context>(*this),
+#endif
+        data_{detail::make_arg<is_packed, Context>(args)...} {
+    if (detail::const_check(num_named_args != 0))
+      detail::init_named_args(data_.named_args(), 0, 0, args...);
+  }
+};
+
+/**
+  \rst
+  Constructs a `~fmt::format_arg_store` object that contains references to
+  arguments and can be implicitly converted to `~fmt::format_args`. `Context`
+  can be omitted in which case it defaults to `~fmt::format_context`.
+  See `~fmt::arg` for lifetime considerations.
+  \endrst
+ */
+// Arguments are taken by lvalue references to avoid some lifetime issues.
+template <typename Context = format_context, typename... T>
+constexpr auto make_format_args(T&... args)
+    -> format_arg_store<Context, remove_cvref_t<T>...> {
+  return {args...};
+}
+
+/**
+  \rst
+  Returns a named argument to be used in a formatting function.
+  It should only be used in a call to a formatting function or
+  `dynamic_format_arg_store::push_back`.
+
+  **Example**::
+
+    fmt::print("Elapsed time: {s:.2f} seconds", fmt::arg("s", 1.23));
+  \endrst
+ */
+template <typename Char, typename T>
+inline auto arg(const Char* name, const T& arg) -> detail::named_arg<Char, T> {
+  static_assert(!detail::is_named_arg<T>(), "nested named arguments");
+  return {name, arg};
+}
+FMT_END_EXPORT
+
+/**
+  \rst
+  A view of a collection of formatting arguments. To avoid lifetime issues it
+  should only be used as a parameter type in type-erased functions such as
+  ``vformat``::
+
+    void vlog(string_view format_str, format_args args);  // OK
+    format_args args = make_format_args();  // Error: dangling reference
+  \endrst
+ */
+template <typename Context> class basic_format_args {
+ public:
+  using size_type = int;
+  using format_arg = basic_format_arg<Context>;
+
+ private:
+  // A descriptor that contains information about formatting arguments.
+  // If the number of arguments is less or equal to max_packed_args then
+  // argument types are passed in the descriptor. This reduces binary code size
+  // per formatting function call.
+  unsigned long long desc_;
+  union {
+    // If is_packed() returns true then argument values are stored in values_;
+    // otherwise they are stored in args_. This is done to improve cache
+    // locality and reduce compiled code size since storing larger objects
+    // may require more code (at least on x86-64) even if the same amount of
+    // data is actually copied to stack. It saves ~10% on the bloat test.
+    const detail::value<Context>* values_;
+    const format_arg* args_;
+  };
+
+  constexpr auto is_packed() const -> bool {
+    return (desc_ & detail::is_unpacked_bit) == 0;
+  }
+  auto has_named_args() const -> bool {
+    return (desc_ & detail::has_named_args_bit) != 0;
+  }
+
+  FMT_CONSTEXPR auto type(int index) const -> detail::type {
+    int shift = index * detail::packed_arg_bits;
+    unsigned int mask = (1 << detail::packed_arg_bits) - 1;
+    return static_cast<detail::type>((desc_ >> shift) & mask);
+  }
+
+  constexpr FMT_INLINE basic_format_args(unsigned long long desc,
+                                         const detail::value<Context>* values)
+      : desc_(desc), values_(values) {}
+  constexpr basic_format_args(unsigned long long desc, const format_arg* args)
+      : desc_(desc), args_(args) {}
+
+ public:
+  constexpr basic_format_args() : desc_(0), args_(nullptr) {}
+
+  /**
+   \rst
+   Constructs a `basic_format_args` object from `~fmt::format_arg_store`.
+   \endrst
+   */
+  template <typename... Args>
+  constexpr FMT_INLINE basic_format_args(
+      const format_arg_store<Context, Args...>& store)
+      : basic_format_args(format_arg_store<Context, Args...>::desc,
+                          store.data_.args()) {}
+
+  /**
+   \rst
+   Constructs a `basic_format_args` object from
+   `~fmt::dynamic_format_arg_store`.
+   \endrst
+   */
+  constexpr FMT_INLINE basic_format_args(
+      const dynamic_format_arg_store<Context>& store)
+      : basic_format_args(store.get_types(), store.data()) {}
+
+  /**
+   \rst
+   Constructs a `basic_format_args` object from a dynamic set of arguments.
+   \endrst
+   */
+  constexpr basic_format_args(const format_arg* args, int count)
+      : basic_format_args(detail::is_unpacked_bit | detail::to_unsigned(count),
+                          args) {}
+
+  /** Returns the argument with the specified id. */
+  FMT_CONSTEXPR auto get(int id) const -> format_arg {
+    format_arg arg;
+    if (!is_packed()) {
+      if (id < max_size()) arg = args_[id];
+      return arg;
+    }
+    if (id >= detail::max_packed_args) return arg;
+    arg.type_ = type(id);
+    if (arg.type_ == detail::type::none_type) return arg;
+    arg.value_ = values_[id];
+    return arg;
+  }
+
+  template <typename Char>
+  auto get(basic_string_view<Char> name) const -> format_arg {
+    int id = get_id(name);
+    return id >= 0 ? get(id) : format_arg();
+  }
+
+  template <typename Char>
+  auto get_id(basic_string_view<Char> name) const -> int {
+    if (!has_named_args()) return -1;
+    const auto& named_args =
+        (is_packed() ? values_[-1] : args_[-1].value_).named_args;
+    for (size_t i = 0; i < named_args.size; ++i) {
+      if (named_args.data[i].name == name) return named_args.data[i].id;
+    }
+    return -1;
+  }
+
+  auto max_size() const -> int {
+    unsigned long long max_packed = detail::max_packed_args;
+    return static_cast<int>(is_packed() ? max_packed
+                                        : desc_ & ~detail::is_unpacked_bit);
+  }
+};
+
+/** An alias to ``basic_format_args<format_context>``. */
+// A separate type would result in shorter symbols but break ABI compatibility
+// between clang and gcc on ARM (#1919).
+FMT_EXPORT using format_args = basic_format_args<format_context>;
+
+// We cannot use enum classes as bit fields because of a gcc bug, so we put them
+// in namespaces instead (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61414).
+// Additionally, if an underlying type is specified, older gcc incorrectly warns
+// that the type is too small. Both bugs are fixed in gcc 9.3.
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 903
+#  define FMT_ENUM_UNDERLYING_TYPE(type)
+#else
+#  define FMT_ENUM_UNDERLYING_TYPE(type) : type
+#endif
+namespace align {
+enum type FMT_ENUM_UNDERLYING_TYPE(unsigned char){none, left, right, center,
+                                                  numeric};
+}
+using align_t = align::type;
+namespace sign {
+enum type FMT_ENUM_UNDERLYING_TYPE(unsigned char){none, minus, plus, space};
+}
+using sign_t = sign::type;
+
+namespace detail {
+
+// Workaround an array initialization issue in gcc 4.8.
+template <typename Char> struct fill_t {
+ private:
+  enum { max_size = 4 };
+  Char data_[max_size] = {Char(' '), Char(0), Char(0), Char(0)};
+  unsigned char size_ = 1;
+
+ public:
+  FMT_CONSTEXPR void operator=(basic_string_view<Char> s) {
+    auto size = s.size();
+    FMT_ASSERT(size <= max_size, "invalid fill");
+    for (size_t i = 0; i < size; ++i) data_[i] = s[i];
+    size_ = static_cast<unsigned char>(size);
+  }
+
+  constexpr auto size() const -> size_t { return size_; }
+  constexpr auto data() const -> const Char* { return data_; }
+
+  FMT_CONSTEXPR auto operator[](size_t index) -> Char& { return data_[index]; }
+  FMT_CONSTEXPR auto operator[](size_t index) const -> const Char& {
+    return data_[index];
+  }
+};
+}  // namespace detail
+
+enum class presentation_type : unsigned char {
+  none,
+  dec,             // 'd'
+  oct,             // 'o'
+  hex_lower,       // 'x'
+  hex_upper,       // 'X'
+  bin_lower,       // 'b'
+  bin_upper,       // 'B'
+  hexfloat_lower,  // 'a'
+  hexfloat_upper,  // 'A'
+  exp_lower,       // 'e'
+  exp_upper,       // 'E'
+  fixed_lower,     // 'f'
+  fixed_upper,     // 'F'
+  general_lower,   // 'g'
+  general_upper,   // 'G'
+  chr,             // 'c'
+  string,          // 's'
+  pointer,         // 'p'
+  debug            // '?'
+};
+
+// Format specifiers for built-in and string types.
+template <typename Char = char> struct format_specs {
+  int width;
+  int precision;
+  presentation_type type;
+  align_t align : 4;
+  sign_t sign : 3;
+  bool alt : 1;  // Alternate form ('#').
+  bool localized : 1;
+  detail::fill_t<Char> fill;
+
+  constexpr format_specs()
+      : width(0),
+        precision(-1),
+        type(presentation_type::none),
+        align(align::none),
+        sign(sign::none),
+        alt(false),
+        localized(false) {}
+};
+
+namespace detail {
+
+enum class arg_id_kind { none, index, name };
+
+// An argument reference.
+template <typename Char> struct arg_ref {
+  FMT_CONSTEXPR arg_ref() : kind(arg_id_kind::none), val() {}
+
+  FMT_CONSTEXPR explicit arg_ref(int index)
+      : kind(arg_id_kind::index), val(index) {}
+  FMT_CONSTEXPR explicit arg_ref(basic_string_view<Char> name)
+      : kind(arg_id_kind::name), val(name) {}
+
+  FMT_CONSTEXPR auto operator=(int idx) -> arg_ref& {
+    kind = arg_id_kind::index;
+    val.index = idx;
+    return *this;
+  }
+
+  arg_id_kind kind;
+  union value {
+    FMT_CONSTEXPR value(int idx = 0) : index(idx) {}
+    FMT_CONSTEXPR value(basic_string_view<Char> n) : name(n) {}
+
+    int index;
+    basic_string_view<Char> name;
+  } val;
+};
+
+// Format specifiers with width and precision resolved at formatting rather
+// than parsing time to allow reusing the same parsed specifiers with
+// different sets of arguments (precompilation of format strings).
+template <typename Char = char>
+struct dynamic_format_specs : format_specs<Char> {
+  arg_ref<Char> width_ref;
+  arg_ref<Char> precision_ref;
+};
+
+// Converts a character to ASCII. Returns '\0' on conversion failure.
+template <typename Char, FMT_ENABLE_IF(std::is_integral<Char>::value)>
+constexpr auto to_ascii(Char c) -> char {
+  return c <= 0xff ? static_cast<char>(c) : '\0';
+}
+template <typename Char, FMT_ENABLE_IF(std::is_enum<Char>::value)>
+constexpr auto to_ascii(Char c) -> char {
+  return c <= 0xff ? static_cast<char>(c) : '\0';
+}
+
+// Returns the number of code units in a code point or 1 on error.
+template <typename Char>
+FMT_CONSTEXPR auto code_point_length(const Char* begin) -> int {
+  if (const_check(sizeof(Char) != 1)) return 1;
+  auto c = static_cast<unsigned char>(*begin);
+  return static_cast<int>((0x3a55000000000000ull >> (2 * (c >> 3))) & 0x3) + 1;
+}
+
+// Return the result via the out param to workaround gcc bug 77539.
+template <bool IS_CONSTEXPR, typename T, typename Ptr = const T*>
+FMT_CONSTEXPR auto find(Ptr first, Ptr last, T value, Ptr& out) -> bool {
+  for (out = first; out != last; ++out) {
+    if (*out == value) return true;
+  }
+  return false;
+}
+
+template <>
+inline auto find<false, char>(const char* first, const char* last, char value,
+                              const char*& out) -> bool {
+  out = static_cast<const char*>(
+      std::memchr(first, value, to_unsigned(last - first)));
+  return out != nullptr;
+}
+
+// Parses the range [begin, end) as an unsigned integer. This function assumes
+// that the range is non-empty and the first character is a digit.
+template <typename Char>
+FMT_CONSTEXPR auto parse_nonnegative_int(const Char*& begin, const Char* end,
+                                         int error_value) noexcept -> int {
+  FMT_ASSERT(begin != end && '0' <= *begin && *begin <= '9', "");
+  unsigned value = 0, prev = 0;
+  auto p = begin;
+  do {
+    prev = value;
+    value = value * 10 + unsigned(*p - '0');
+    ++p;
+  } while (p != end && '0' <= *p && *p <= '9');
+  auto num_digits = p - begin;
+  begin = p;
+  if (num_digits <= std::numeric_limits<int>::digits10)
+    return static_cast<int>(value);
+  // Check for overflow.
+  const unsigned max = to_unsigned((std::numeric_limits<int>::max)());
+  return num_digits == std::numeric_limits<int>::digits10 + 1 &&
+                 prev * 10ull + unsigned(p[-1] - '0') <= max
+             ? static_cast<int>(value)
+             : error_value;
+}
+
+FMT_CONSTEXPR inline auto parse_align(char c) -> align_t {
+  switch (c) {
+  case '<':
+    return align::left;
+  case '>':
+    return align::right;
+  case '^':
+    return align::center;
+  }
+  return align::none;
+}
+
+template <typename Char> constexpr auto is_name_start(Char c) -> bool {
+  return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '_';
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto do_parse_arg_id(const Char* begin, const Char* end,
+                                   Handler&& handler) -> const Char* {
+  Char c = *begin;
+  if (c >= '0' && c <= '9') {
+    int index = 0;
+    constexpr int max = (std::numeric_limits<int>::max)();
+    if (c != '0')
+      index = parse_nonnegative_int(begin, end, max);
+    else
+      ++begin;
+    if (begin == end || (*begin != '}' && *begin != ':'))
+      throw_format_error("invalid format string");
+    else
+      handler.on_index(index);
+    return begin;
+  }
+  if (!is_name_start(c)) {
+    throw_format_error("invalid format string");
+    return begin;
+  }
+  auto it = begin;
+  do {
+    ++it;
+  } while (it != end && (is_name_start(*it) || ('0' <= *it && *it <= '9')));
+  handler.on_name({begin, to_unsigned(it - begin)});
+  return it;
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR FMT_INLINE auto parse_arg_id(const Char* begin, const Char* end,
+                                           Handler&& handler) -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  Char c = *begin;
+  if (c != '}' && c != ':') return do_parse_arg_id(begin, end, handler);
+  handler.on_auto();
+  return begin;
+}
+
+template <typename Char> struct dynamic_spec_id_handler {
+  basic_format_parse_context<Char>& ctx;
+  arg_ref<Char>& ref;
+
+  FMT_CONSTEXPR void on_auto() {
+    int id = ctx.next_arg_id();
+    ref = arg_ref<Char>(id);
+    ctx.check_dynamic_spec(id);
+  }
+  FMT_CONSTEXPR void on_index(int id) {
+    ref = arg_ref<Char>(id);
+    ctx.check_arg_id(id);
+    ctx.check_dynamic_spec(id);
+  }
+  FMT_CONSTEXPR void on_name(basic_string_view<Char> id) {
+    ref = arg_ref<Char>(id);
+    ctx.check_arg_id(id);
+  }
+};
+
+// Parses [integer | "{" [arg_id] "}"].
+template <typename Char>
+FMT_CONSTEXPR auto parse_dynamic_spec(const Char* begin, const Char* end,
+                                      int& value, arg_ref<Char>& ref,
+                                      basic_format_parse_context<Char>& ctx)
+    -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  if ('0' <= *begin && *begin <= '9') {
+    int val = parse_nonnegative_int(begin, end, -1);
+    if (val != -1)
+      value = val;
+    else
+      throw_format_error("number is too big");
+  } else if (*begin == '{') {
+    ++begin;
+    auto handler = dynamic_spec_id_handler<Char>{ctx, ref};
+    if (begin != end) begin = parse_arg_id(begin, end, handler);
+    if (begin != end && *begin == '}') return ++begin;
+    throw_format_error("invalid format string");
+  }
+  return begin;
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto parse_precision(const Char* begin, const Char* end,
+                                   int& value, arg_ref<Char>& ref,
+                                   basic_format_parse_context<Char>& ctx)
+    -> const Char* {
+  ++begin;
+  if (begin == end || *begin == '}') {
+    throw_format_error("invalid precision");
+    return begin;
+  }
+  return parse_dynamic_spec(begin, end, value, ref, ctx);
+}
+
+enum class state { start, align, sign, hash, zero, width, precision, locale };
+
+// Parses standard format specifiers.
+template <typename Char>
+FMT_CONSTEXPR FMT_INLINE auto parse_format_specs(
+    const Char* begin, const Char* end, dynamic_format_specs<Char>& specs,
+    basic_format_parse_context<Char>& ctx, type arg_type) -> const Char* {
+  auto c = '\0';
+  if (end - begin > 1) {
+    auto next = to_ascii(begin[1]);
+    c = parse_align(next) == align::none ? to_ascii(*begin) : '\0';
+  } else {
+    if (begin == end) return begin;
+    c = to_ascii(*begin);
+  }
+
+  struct {
+    state current_state = state::start;
+    FMT_CONSTEXPR void operator()(state s, bool valid = true) {
+      if (current_state >= s || !valid)
+        throw_format_error("invalid format specifier");
+      current_state = s;
+    }
+  } enter_state;
+
+  using pres = presentation_type;
+  constexpr auto integral_set = sint_set | uint_set | bool_set | char_set;
+  struct {
+    const Char*& begin;
+    dynamic_format_specs<Char>& specs;
+    type arg_type;
+
+    FMT_CONSTEXPR auto operator()(pres pres_type, int set) -> const Char* {
+      if (!in(arg_type, set)) {
+        if (arg_type == type::none_type) return begin;
+        throw_format_error("invalid format specifier");
+      }
+      specs.type = pres_type;
+      return begin + 1;
+    }
+  } parse_presentation_type{begin, specs, arg_type};
+
+  for (;;) {
+    switch (c) {
+    case '<':
+    case '>':
+    case '^':
+      enter_state(state::align);
+      specs.align = parse_align(c);
+      ++begin;
+      break;
+    case '+':
+    case '-':
+    case ' ':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::sign, in(arg_type, sint_set | float_set));
+      switch (c) {
+      case '+':
+        specs.sign = sign::plus;
+        break;
+      case '-':
+        specs.sign = sign::minus;
+        break;
+      case ' ':
+        specs.sign = sign::space;
+        break;
+      }
+      ++begin;
+      break;
+    case '#':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::hash, is_arithmetic_type(arg_type));
+      specs.alt = true;
+      ++begin;
+      break;
+    case '0':
+      enter_state(state::zero);
+      if (!is_arithmetic_type(arg_type)) {
+        if (arg_type == type::none_type) return begin;
+        throw_format_error("format specifier requires numeric argument");
+      }
+      if (specs.align == align::none) {
+        // Ignore 0 if align is specified for compatibility with std::format.
+        specs.align = align::numeric;
+        specs.fill[0] = Char('0');
+      }
+      ++begin;
+      break;
+    case '1':
+    case '2':
+    case '3':
+    case '4':
+    case '5':
+    case '6':
+    case '7':
+    case '8':
+    case '9':
+    case '{':
+      enter_state(state::width);
+      begin = parse_dynamic_spec(begin, end, specs.width, specs.width_ref, ctx);
+      break;
+    case '.':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::precision,
+                  in(arg_type, float_set | string_set | cstring_set));
+      begin = parse_precision(begin, end, specs.precision, specs.precision_ref,
+                              ctx);
+      break;
+    case 'L':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::locale, is_arithmetic_type(arg_type));
+      specs.localized = true;
+      ++begin;
+      break;
+    case 'd':
+      return parse_presentation_type(pres::dec, integral_set);
+    case 'o':
+      return parse_presentation_type(pres::oct, integral_set);
+    case 'x':
+      return parse_presentation_type(pres::hex_lower, integral_set);
+    case 'X':
+      return parse_presentation_type(pres::hex_upper, integral_set);
+    case 'b':
+      return parse_presentation_type(pres::bin_lower, integral_set);
+    case 'B':
+      return parse_presentation_type(pres::bin_upper, integral_set);
+    case 'a':
+      return parse_presentation_type(pres::hexfloat_lower, float_set);
+    case 'A':
+      return parse_presentation_type(pres::hexfloat_upper, float_set);
+    case 'e':
+      return parse_presentation_type(pres::exp_lower, float_set);
+    case 'E':
+      return parse_presentation_type(pres::exp_upper, float_set);
+    case 'f':
+      return parse_presentation_type(pres::fixed_lower, float_set);
+    case 'F':
+      return parse_presentation_type(pres::fixed_upper, float_set);
+    case 'g':
+      return parse_presentation_type(pres::general_lower, float_set);
+    case 'G':
+      return parse_presentation_type(pres::general_upper, float_set);
+    case 'c':
+      if (arg_type == type::bool_type)
+        throw_format_error("invalid format specifier");
+      return parse_presentation_type(pres::chr, integral_set);
+    case 's':
+      return parse_presentation_type(pres::string,
+                                     bool_set | string_set | cstring_set);
+    case 'p':
+      return parse_presentation_type(pres::pointer, pointer_set | cstring_set);
+    case '?':
+      return parse_presentation_type(pres::debug,
+                                     char_set | string_set | cstring_set);
+    case '}':
+      return begin;
+    default: {
+      if (*begin == '}') return begin;
+      // Parse fill and alignment.
+      auto fill_end = begin + code_point_length(begin);
+      if (end - fill_end <= 0) {
+        throw_format_error("invalid format specifier");
+        return begin;
+      }
+      if (*begin == '{') {
+        throw_format_error("invalid fill character '{'");
+        return begin;
+      }
+      auto align = parse_align(to_ascii(*fill_end));
+      enter_state(state::align, align != align::none);
+      specs.fill = {begin, to_unsigned(fill_end - begin)};
+      specs.align = align;
+      begin = fill_end + 1;
+    }
+    }
+    if (begin == end) return begin;
+    c = to_ascii(*begin);
+  }
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_replacement_field(const Char* begin, const Char* end,
+                                           Handler&& handler) -> const Char* {
+  struct id_adapter {
+    Handler& handler;
+    int arg_id;
+
+    FMT_CONSTEXPR void on_auto() { arg_id = handler.on_arg_id(); }
+    FMT_CONSTEXPR void on_index(int id) { arg_id = handler.on_arg_id(id); }
+    FMT_CONSTEXPR void on_name(basic_string_view<Char> id) {
+      arg_id = handler.on_arg_id(id);
+    }
+  };
+
+  ++begin;
+  if (begin == end) return handler.on_error("invalid format string"), end;
+  if (*begin == '}') {
+    handler.on_replacement_field(handler.on_arg_id(), begin);
+  } else if (*begin == '{') {
+    handler.on_text(begin, begin + 1);
+  } else {
+    auto adapter = id_adapter{handler, 0};
+    begin = parse_arg_id(begin, end, adapter);
+    Char c = begin != end ? *begin : Char();
+    if (c == '}') {
+      handler.on_replacement_field(adapter.arg_id, begin);
+    } else if (c == ':') {
+      begin = handler.on_format_specs(adapter.arg_id, begin + 1, end);
+      if (begin == end || *begin != '}')
+        return handler.on_error("unknown format specifier"), end;
+    } else {
+      return handler.on_error("missing '}' in format string"), end;
+    }
+  }
+  return begin + 1;
+}
+
+template <bool IS_CONSTEXPR, typename Char, typename Handler>
+FMT_CONSTEXPR FMT_INLINE void parse_format_string(
+    basic_string_view<Char> format_str, Handler&& handler) {
+  auto begin = format_str.data();
+  auto end = begin + format_str.size();
+  if (end - begin < 32) {
+    // Use a simple loop instead of memchr for small strings.
+    const Char* p = begin;
+    while (p != end) {
+      auto c = *p++;
+      if (c == '{') {
+        handler.on_text(begin, p - 1);
+        begin = p = parse_replacement_field(p - 1, end, handler);
+      } else if (c == '}') {
+        if (p == end || *p != '}')
+          return handler.on_error("unmatched '}' in format string");
+        handler.on_text(begin, p);
+        begin = ++p;
+      }
+    }
+    handler.on_text(begin, end);
+    return;
+  }
+  struct writer {
+    FMT_CONSTEXPR void operator()(const Char* from, const Char* to) {
+      if (from == to) return;
+      for (;;) {
+        const Char* p = nullptr;
+        if (!find<IS_CONSTEXPR>(from, to, Char('}'), p))
+          return handler_.on_text(from, to);
+        ++p;
+        if (p == to || *p != '}')
+          return handler_.on_error("unmatched '}' in format string");
+        handler_.on_text(from, p);
+        from = p + 1;
+      }
+    }
+    Handler& handler_;
+  } write = {handler};
+  while (begin != end) {
+    // Doing two passes with memchr (one for '{' and another for '}') is up to
+    // 2.5x faster than the naive one-pass implementation on big format strings.
+    const Char* p = begin;
+    if (*begin != '{' && !find<IS_CONSTEXPR>(begin + 1, end, Char('{'), p))
+      return write(begin, end);
+    write(begin, p);
+    begin = parse_replacement_field(p, end, handler);
+  }
+}
+
+template <typename T, bool = is_named_arg<T>::value> struct strip_named_arg {
+  using type = T;
+};
+template <typename T> struct strip_named_arg<T, true> {
+  using type = remove_cvref_t<decltype(T::value)>;
+};
+
+template <typename T, typename ParseContext>
+FMT_CONSTEXPR auto parse_format_specs(ParseContext& ctx)
+    -> decltype(ctx.begin()) {
+  using char_type = typename ParseContext::char_type;
+  using context = buffer_context<char_type>;
+  using mapped_type = conditional_t<
+      mapped_type_constant<T, context>::value != type::custom_type,
+      decltype(arg_mapper<context>().map(std::declval<const T&>())),
+      typename strip_named_arg<T>::type>;
+#if defined(__cpp_if_constexpr)
+  if constexpr (std::is_default_constructible<
+                    formatter<mapped_type, char_type>>::value) {
+    return formatter<mapped_type, char_type>().parse(ctx);
+  } else {
+    type_is_unformattable_for<T, char_type> _;
+    return ctx.begin();
+  }
+#else
+  return formatter<mapped_type, char_type>().parse(ctx);
+#endif
+}
+
+// Checks char specs and returns true iff the presentation type is char-like.
+template <typename Char>
+FMT_CONSTEXPR auto check_char_specs(const format_specs<Char>& specs) -> bool {
+  if (specs.type != presentation_type::none &&
+      specs.type != presentation_type::chr &&
+      specs.type != presentation_type::debug) {
+    return false;
+  }
+  if (specs.align == align::numeric || specs.sign != sign::none || specs.alt)
+    throw_format_error("invalid format specifier for char");
+  return true;
+}
+
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <int N, typename T, typename... Args, typename Char>
+constexpr auto get_arg_index_by_name(basic_string_view<Char> name) -> int {
+  if constexpr (is_statically_named_arg<T>()) {
+    if (name == T::name) return N;
+  }
+  if constexpr (sizeof...(Args) > 0)
+    return get_arg_index_by_name<N + 1, Args...>(name);
+  (void)name;  // Workaround an MSVC bug about "unused" parameter.
+  return -1;
+}
+#endif
+
+template <typename... Args, typename Char>
+FMT_CONSTEXPR auto get_arg_index_by_name(basic_string_view<Char> name) -> int {
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+  if constexpr (sizeof...(Args) > 0)
+    return get_arg_index_by_name<0, Args...>(name);
+#endif
+  (void)name;
+  return -1;
+}
+
+template <typename Char, typename... Args> class format_string_checker {
+ private:
+  using parse_context_type = compile_parse_context<Char>;
+  static constexpr int num_args = sizeof...(Args);
+
+  // Format specifier parsing function.
+  // In the future basic_format_parse_context will replace compile_parse_context
+  // here and will use is_constant_evaluated and downcasting to access the data
+  // needed for compile-time checks: https://godbolt.org/z/GvWzcTjh1.
+  using parse_func = const Char* (*)(parse_context_type&);
+
+  type types_[num_args > 0 ? static_cast<size_t>(num_args) : 1];
+  parse_context_type context_;
+  parse_func parse_funcs_[num_args > 0 ? static_cast<size_t>(num_args) : 1];
+
+ public:
+  explicit FMT_CONSTEXPR format_string_checker(basic_string_view<Char> fmt)
+      : types_{mapped_type_constant<Args, buffer_context<Char>>::value...},
+        context_(fmt, num_args, types_),
+        parse_funcs_{&parse_format_specs<Args, parse_context_type>...} {}
+
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+
+  FMT_CONSTEXPR auto on_arg_id() -> int { return context_.next_arg_id(); }
+  FMT_CONSTEXPR auto on_arg_id(int id) -> int {
+    return context_.check_arg_id(id), id;
+  }
+  FMT_CONSTEXPR auto on_arg_id(basic_string_view<Char> id) -> int {
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+    auto index = get_arg_index_by_name<Args...>(id);
+    if (index < 0) on_error("named argument is not found");
+    return index;
+#else
+    (void)id;
+    on_error("compile-time checks for named arguments require C++20 support");
+    return 0;
+#endif
+  }
+
+  FMT_CONSTEXPR void on_replacement_field(int id, const Char* begin) {
+    on_format_specs(id, begin, begin);  // Call parse() on empty specs.
+  }
+
+  FMT_CONSTEXPR auto on_format_specs(int id, const Char* begin, const Char*)
+      -> const Char* {
+    context_.advance_to(begin);
+    // id >= 0 check is a workaround for gcc 10 bug (#2065).
+    return id >= 0 && id < num_args ? parse_funcs_[id](context_) : begin;
+  }
+
+  FMT_CONSTEXPR void on_error(const char* message) {
+    throw_format_error(message);
+  }
+};
+
+// Reports a compile-time error if S is not a valid format string.
+template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
+FMT_INLINE void check_format_string(const S&) {
+#ifdef FMT_ENFORCE_COMPILE_STRING
+  static_assert(is_compile_string<S>::value,
+                "FMT_ENFORCE_COMPILE_STRING requires all format strings to use "
+                "FMT_STRING.");
+#endif
+}
+template <typename... Args, typename S,
+          FMT_ENABLE_IF(is_compile_string<S>::value)>
+void check_format_string(S format_str) {
+  using char_t = typename S::char_type;
+  FMT_CONSTEXPR auto s = basic_string_view<char_t>(format_str);
+  using checker = format_string_checker<char_t, remove_cvref_t<Args>...>;
+  FMT_CONSTEXPR bool error = (parse_format_string<true>(s, checker(s)), true);
+  ignore_unused(error);
+}
+
+template <typename Char = char> struct vformat_args {
+  using type = basic_format_args<
+      basic_format_context<std::back_insert_iterator<buffer<Char>>, Char>>;
+};
+template <> struct vformat_args<char> {
+  using type = format_args;
+};
+
+// Use vformat_args and avoid type_identity to keep symbols short.
+template <typename Char>
+void vformat_to(buffer<Char>& buf, basic_string_view<Char> fmt,
+                typename vformat_args<Char>::type args, locale_ref loc = {});
+
+FMT_API void vprint_mojibake(std::FILE*, string_view, format_args);
+#ifndef _WIN32
+inline void vprint_mojibake(std::FILE*, string_view, format_args) {}
+#endif
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+// A formatter specialization for natively supported types.
+template <typename T, typename Char>
+struct formatter<T, Char,
+                 enable_if_t<detail::type_constant<T, Char>::value !=
+                             detail::type::custom_type>> {
+ private:
+  detail::dynamic_format_specs<Char> specs_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const Char* {
+    auto type = detail::type_constant<T, Char>::value;
+    auto end =
+        detail::parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx, type);
+    if (type == detail::type::char_type) detail::check_char_specs(specs_);
+    return end;
+  }
+
+  template <detail::type U = detail::type_constant<T, Char>::value,
+            FMT_ENABLE_IF(U == detail::type::string_type ||
+                          U == detail::type::cstring_type ||
+                          U == detail::type::char_type)>
+  FMT_CONSTEXPR void set_debug_format(bool set = true) {
+    specs_.type = set ? presentation_type::debug : presentation_type::none;
+  }
+
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const T& val, FormatContext& ctx) const
+      -> decltype(ctx.out());
+};
+
+template <typename Char = char> struct runtime_format_string {
+  basic_string_view<Char> str;
+};
+
+/** A compile-time format string. */
+template <typename Char, typename... Args> class basic_format_string {
+ private:
+  basic_string_view<Char> str_;
+
+ public:
+  template <typename S,
+            FMT_ENABLE_IF(
+                std::is_convertible<const S&, basic_string_view<Char>>::value)>
+  FMT_CONSTEVAL FMT_INLINE basic_format_string(const S& s) : str_(s) {
+    static_assert(
+        detail::count<
+            (std::is_base_of<detail::view, remove_reference_t<Args>>::value &&
+             std::is_reference<Args>::value)...>() == 0,
+        "passing views as lvalues is disallowed");
+#ifdef FMT_HAS_CONSTEVAL
+    if constexpr (detail::count_named_args<Args...>() ==
+                  detail::count_statically_named_args<Args...>()) {
+      using checker =
+          detail::format_string_checker<Char, remove_cvref_t<Args>...>;
+      detail::parse_format_string<true>(str_, checker(s));
+    }
+#else
+    detail::check_format_string<Args...>(s);
+#endif
+  }
+  basic_format_string(runtime_format_string<Char> fmt) : str_(fmt.str) {}
+
+  FMT_INLINE operator basic_string_view<Char>() const { return str_; }
+  FMT_INLINE auto get() const -> basic_string_view<Char> { return str_; }
+};
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+// Workaround broken conversion on older gcc.
+template <typename...> using format_string = string_view;
+inline auto runtime(string_view s) -> string_view { return s; }
+#else
+template <typename... Args>
+using format_string = basic_format_string<char, type_identity_t<Args>...>;
+/**
+  \rst
+  Creates a runtime format string.
+
+  **Example**::
+
+    // Check format string at runtime instead of compile-time.
+    fmt::print(fmt::runtime("{:d}"), "I am not a number");
+  \endrst
+ */
+inline auto runtime(string_view s) -> runtime_format_string<> { return {{s}}; }
+#endif
+
+FMT_API auto vformat(string_view fmt, format_args args) -> std::string;
+
+/**
+  \rst
+  Formats ``args`` according to specifications in ``fmt`` and returns the result
+  as a string.
+
+  **Example**::
+
+    #include <fmt/core.h>
+    std::string message = fmt::format("The answer is {}.", 42);
+  \endrst
+*/
+template <typename... T>
+FMT_NODISCARD FMT_INLINE auto format(format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return vformat(fmt, fmt::make_format_args(args...));
+}
+
+/** Formats a string and writes the output to ``out``. */
+template <typename OutputIt,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+auto vformat_to(OutputIt out, string_view fmt, format_args args) -> OutputIt {
+  auto&& buf = detail::get_buffer<char>(out);
+  detail::vformat_to(buf, fmt, args, {});
+  return detail::get_iterator(buf, out);
+}
+
+/**
+ \rst
+ Formats ``args`` according to specifications in ``fmt``, writes the result to
+ the output iterator ``out`` and returns the iterator past the end of the output
+ range. `format_to` does not append a terminating null character.
+
+ **Example**::
+
+   auto out = std::vector<char>();
+   fmt::format_to(std::back_inserter(out), "{}", 42);
+ \endrst
+ */
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+FMT_INLINE auto format_to(OutputIt out, format_string<T...> fmt, T&&... args)
+    -> OutputIt {
+  return vformat_to(out, fmt, fmt::make_format_args(args...));
+}
+
+template <typename OutputIt> struct format_to_n_result {
+  /** Iterator past the end of the output range. */
+  OutputIt out;
+  /** Total (not truncated) output size. */
+  size_t size;
+};
+
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+auto vformat_to_n(OutputIt out, size_t n, string_view fmt, format_args args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, char, traits>(out, n);
+  detail::vformat_to(buf, fmt, args, {});
+  return {buf.out(), buf.count()};
+}
+
+/**
+  \rst
+  Formats ``args`` according to specifications in ``fmt``, writes up to ``n``
+  characters of the result to the output iterator ``out`` and returns the total
+  (not truncated) output size and the iterator past the end of the output range.
+  `format_to_n` does not append a terminating null character.
+  \endrst
+ */
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+FMT_INLINE auto format_to_n(OutputIt out, size_t n, format_string<T...> fmt,
+                            T&&... args) -> format_to_n_result<OutputIt> {
+  return vformat_to_n(out, n, fmt, fmt::make_format_args(args...));
+}
+
+/** Returns the number of chars in the output of ``format(fmt, args...)``. */
+template <typename... T>
+FMT_NODISCARD FMT_INLINE auto formatted_size(format_string<T...> fmt,
+                                             T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<>();
+  detail::vformat_to<char>(buf, fmt, fmt::make_format_args(args...), {});
+  return buf.count();
+}
+
+FMT_API void vprint(string_view fmt, format_args args);
+FMT_API void vprint(std::FILE* f, string_view fmt, format_args args);
+
+/**
+  \rst
+  Formats ``args`` according to specifications in ``fmt`` and writes the output
+  to ``stdout``.
+
+  **Example**::
+
+    fmt::print("Elapsed time: {0:.2f} seconds", 1.23);
+  \endrst
+ */
+template <typename... T>
+FMT_INLINE void print(format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  return detail::is_utf8() ? vprint(fmt, vargs)
+                           : detail::vprint_mojibake(stdout, fmt, vargs);
+}
+
+/**
+  \rst
+  Formats ``args`` according to specifications in ``fmt`` and writes the
+  output to the file ``f``.
+
+  **Example**::
+
+    fmt::print(stderr, "Don't {}!", "panic");
+  \endrst
+ */
+template <typename... T>
+FMT_INLINE void print(std::FILE* f, format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  return detail::is_utf8() ? vprint(f, fmt, vargs)
+                           : detail::vprint_mojibake(f, fmt, vargs);
+}
+
+/**
+  Formats ``args`` according to specifications in ``fmt`` and writes the
+  output to the file ``f`` followed by a newline.
+ */
+template <typename... T>
+FMT_INLINE void println(std::FILE* f, format_string<T...> fmt, T&&... args) {
+  return fmt::print(f, "{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+/**
+  Formats ``args`` according to specifications in ``fmt`` and writes the output
+  to ``stdout`` followed by a newline.
+ */
+template <typename... T>
+FMT_INLINE void println(format_string<T...> fmt, T&&... args) {
+  return fmt::println(stdout, fmt, std::forward<T>(args)...);
+}
+
+FMT_END_EXPORT
+FMT_GCC_PRAGMA("GCC pop_options")
+FMT_END_NAMESPACE
+
+#ifdef FMT_HEADER_ONLY
+#  include "format.h"
+#endif
+#endif  // FMT_CORE_H_
diff --git a/dep/fmt/include/fmt/format-inl.h b/dep/fmt/include/fmt/format-inl.h
new file mode 100644
index 00000000..efac5d1f
--- /dev/null
+++ b/dep/fmt/include/fmt/format-inl.h
@@ -0,0 +1,1678 @@
+// Formatting library for C++ - implementation
+//
+// Copyright (c) 2012 - 2016, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_FORMAT_INL_H_
+#define FMT_FORMAT_INL_H_
+
+#include <algorithm>
+#include <cerrno>  // errno
+#include <climits>
+#include <cmath>
+#include <exception>
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+#  include <locale>
+#endif
+
+#if defined(_WIN32) && !defined(FMT_WINDOWS_NO_WCHAR)
+#  include <io.h>  // _isatty
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+FMT_FUNC void assert_fail(const char* file, int line, const char* message) {
+  // Use unchecked std::fprintf to avoid triggering another assertion when
+  // writing to stderr fails
+  std::fprintf(stderr, "%s:%d: assertion failed: %s", file, line, message);
+  // Chosen instead of std::abort to satisfy Clang in CUDA mode during device
+  // code pass.
+  std::terminate();
+}
+
+FMT_FUNC void throw_format_error(const char* message) {
+  FMT_THROW(format_error(message));
+}
+
+FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
+                                string_view message) noexcept {
+  // Report error code making sure that the output fits into
+  // inline_buffer_size to avoid dynamic memory allocation and potential
+  // bad_alloc.
+  out.try_resize(0);
+  static const char SEP[] = ": ";
+  static const char ERROR_STR[] = "error ";
+  // Subtract 2 to account for terminating null characters in SEP and ERROR_STR.
+  size_t error_code_size = sizeof(SEP) + sizeof(ERROR_STR) - 2;
+  auto abs_value = static_cast<uint32_or_64_or_128_t<int>>(error_code);
+  if (detail::is_negative(error_code)) {
+    abs_value = 0 - abs_value;
+    ++error_code_size;
+  }
+  error_code_size += detail::to_unsigned(detail::count_digits(abs_value));
+  auto it = buffer_appender<char>(out);
+  if (message.size() <= inline_buffer_size - error_code_size)
+    fmt::format_to(it, FMT_STRING("{}{}"), message, SEP);
+  fmt::format_to(it, FMT_STRING("{}{}"), ERROR_STR, error_code);
+  FMT_ASSERT(out.size() <= inline_buffer_size, "");
+}
+
+FMT_FUNC void report_error(format_func func, int error_code,
+                           const char* message) noexcept {
+  memory_buffer full_message;
+  func(full_message, error_code, message);
+  // Don't use fwrite_fully because the latter may throw.
+  if (std::fwrite(full_message.data(), full_message.size(), 1, stderr) > 0)
+    std::fputc('\n', stderr);
+}
+
+// A wrapper around fwrite that throws on error.
+inline void fwrite_fully(const void* ptr, size_t count, FILE* stream) {
+  size_t written = std::fwrite(ptr, 1, count, stream);
+  if (written < count)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot write to file")));
+}
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+template <typename Locale>
+locale_ref::locale_ref(const Locale& loc) : locale_(&loc) {
+  static_assert(std::is_same<Locale, std::locale>::value, "");
+}
+
+template <typename Locale> auto locale_ref::get() const -> Locale {
+  static_assert(std::is_same<Locale, std::locale>::value, "");
+  return locale_ ? *static_cast<const std::locale*>(locale_) : std::locale();
+}
+
+template <typename Char>
+FMT_FUNC auto thousands_sep_impl(locale_ref loc) -> thousands_sep_result<Char> {
+  auto& facet = std::use_facet<std::numpunct<Char>>(loc.get<std::locale>());
+  auto grouping = facet.grouping();
+  auto thousands_sep = grouping.empty() ? Char() : facet.thousands_sep();
+  return {std::move(grouping), thousands_sep};
+}
+template <typename Char>
+FMT_FUNC auto decimal_point_impl(locale_ref loc) -> Char {
+  return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>())
+      .decimal_point();
+}
+#else
+template <typename Char>
+FMT_FUNC auto thousands_sep_impl(locale_ref) -> thousands_sep_result<Char> {
+  return {"\03", FMT_STATIC_THOUSANDS_SEPARATOR};
+}
+template <typename Char> FMT_FUNC Char decimal_point_impl(locale_ref) {
+  return '.';
+}
+#endif
+
+FMT_FUNC auto write_loc(appender out, loc_value value,
+                        const format_specs<>& specs, locale_ref loc) -> bool {
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+  auto locale = loc.get<std::locale>();
+  // We cannot use the num_put<char> facet because it may produce output in
+  // a wrong encoding.
+  using facet = format_facet<std::locale>;
+  if (std::has_facet<facet>(locale))
+    return std::use_facet<facet>(locale).put(out, value, specs);
+  return facet(locale).put(out, value, specs);
+#endif
+  return false;
+}
+}  // namespace detail
+
+template <typename Locale> typename Locale::id format_facet<Locale>::id;
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+template <typename Locale> format_facet<Locale>::format_facet(Locale& loc) {
+  auto& numpunct = std::use_facet<std::numpunct<char>>(loc);
+  grouping_ = numpunct.grouping();
+  if (!grouping_.empty()) separator_ = std::string(1, numpunct.thousands_sep());
+}
+
+template <>
+FMT_API FMT_FUNC auto format_facet<std::locale>::do_put(
+    appender out, loc_value val, const format_specs<>& specs) const -> bool {
+  return val.visit(
+      detail::loc_writer<>{out, specs, separator_, grouping_, decimal_point_});
+}
+#endif
+
+FMT_FUNC auto vsystem_error(int error_code, string_view fmt, format_args args)
+    -> std::system_error {
+  auto ec = std::error_code(error_code, std::generic_category());
+  return std::system_error(ec, vformat(fmt, args));
+}
+
+namespace detail {
+
+template <typename F>
+inline auto operator==(basic_fp<F> x, basic_fp<F> y) -> bool {
+  return x.f == y.f && x.e == y.e;
+}
+
+// Compilers should be able to optimize this into the ror instruction.
+FMT_CONSTEXPR inline auto rotr(uint32_t n, uint32_t r) noexcept -> uint32_t {
+  r &= 31;
+  return (n >> r) | (n << (32 - r));
+}
+FMT_CONSTEXPR inline auto rotr(uint64_t n, uint32_t r) noexcept -> uint64_t {
+  r &= 63;
+  return (n >> r) | (n << (64 - r));
+}
+
+// Implementation of Dragonbox algorithm: https://github.com/jk-jeon/dragonbox.
+namespace dragonbox {
+// Computes upper 64 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline auto umul96_upper64(uint32_t x, uint64_t y) noexcept -> uint64_t {
+  return umul128_upper64(static_cast<uint64_t>(x) << 32, y);
+}
+
+// Computes lower 128 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+inline auto umul192_lower128(uint64_t x, uint128_fallback y) noexcept
+    -> uint128_fallback {
+  uint64_t high = x * y.high();
+  uint128_fallback high_low = umul128(x, y.low());
+  return {high + high_low.high(), high_low.low()};
+}
+
+// Computes lower 64 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline auto umul96_lower64(uint32_t x, uint64_t y) noexcept -> uint64_t {
+  return x * y;
+}
+
+// Various fast log computations.
+inline auto floor_log10_pow2_minus_log10_4_over_3(int e) noexcept -> int {
+  FMT_ASSERT(e <= 2936 && e >= -2985, "too large exponent");
+  return (e * 631305 - 261663) >> 21;
+}
+
+FMT_INLINE_VARIABLE constexpr struct {
+  uint32_t divisor;
+  int shift_amount;
+} div_small_pow10_infos[] = {{10, 16}, {100, 16}};
+
+// Replaces n by floor(n / pow(10, N)) returning true if and only if n is
+// divisible by pow(10, N).
+// Precondition: n <= pow(10, N + 1).
+template <int N>
+auto check_divisibility_and_divide_by_pow10(uint32_t& n) noexcept -> bool {
+  // The numbers below are chosen such that:
+  //   1. floor(n/d) = floor(nm / 2^k) where d=10 or d=100,
+  //   2. nm mod 2^k < m if and only if n is divisible by d,
+  // where m is magic_number, k is shift_amount
+  // and d is divisor.
+  //
+  // Item 1 is a common technique of replacing division by a constant with
+  // multiplication, see e.g. "Division by Invariant Integers Using
+  // Multiplication" by Granlund and Montgomery (1994). magic_number (m) is set
+  // to ceil(2^k/d) for large enough k.
+  // The idea for item 2 originates from Schubfach.
+  constexpr auto info = div_small_pow10_infos[N - 1];
+  FMT_ASSERT(n <= info.divisor * 10, "n is too large");
+  constexpr uint32_t magic_number =
+      (1u << info.shift_amount) / info.divisor + 1;
+  n *= magic_number;
+  const uint32_t comparison_mask = (1u << info.shift_amount) - 1;
+  bool result = (n & comparison_mask) < magic_number;
+  n >>= info.shift_amount;
+  return result;
+}
+
+// Computes floor(n / pow(10, N)) for small n and N.
+// Precondition: n <= pow(10, N + 1).
+template <int N> auto small_division_by_pow10(uint32_t n) noexcept -> uint32_t {
+  constexpr auto info = div_small_pow10_infos[N - 1];
+  FMT_ASSERT(n <= info.divisor * 10, "n is too large");
+  constexpr uint32_t magic_number =
+      (1u << info.shift_amount) / info.divisor + 1;
+  return (n * magic_number) >> info.shift_amount;
+}
+
+// Computes floor(n / 10^(kappa + 1)) (float)
+inline auto divide_by_10_to_kappa_plus_1(uint32_t n) noexcept -> uint32_t {
+  // 1374389535 = ceil(2^37/100)
+  return static_cast<uint32_t>((static_cast<uint64_t>(n) * 1374389535) >> 37);
+}
+// Computes floor(n / 10^(kappa + 1)) (double)
+inline auto divide_by_10_to_kappa_plus_1(uint64_t n) noexcept -> uint64_t {
+  // 2361183241434822607 = ceil(2^(64+7)/1000)
+  return umul128_upper64(n, 2361183241434822607ull) >> 7;
+}
+
+// Various subroutines using pow10 cache
+template <typename T> struct cache_accessor;
+
+template <> struct cache_accessor<float> {
+  using carrier_uint = float_info<float>::carrier_uint;
+  using cache_entry_type = uint64_t;
+
+  static auto get_cached_power(int k) noexcept -> uint64_t {
+    FMT_ASSERT(k >= float_info<float>::min_k && k <= float_info<float>::max_k,
+               "k is out of range");
+    static constexpr const uint64_t pow10_significands[] = {
+        0x81ceb32c4b43fcf5, 0xa2425ff75e14fc32, 0xcad2f7f5359a3b3f,
+        0xfd87b5f28300ca0e, 0x9e74d1b791e07e49, 0xc612062576589ddb,
+        0xf79687aed3eec552, 0x9abe14cd44753b53, 0xc16d9a0095928a28,
+        0xf1c90080baf72cb2, 0x971da05074da7bef, 0xbce5086492111aeb,
+        0xec1e4a7db69561a6, 0x9392ee8e921d5d08, 0xb877aa3236a4b44a,
+        0xe69594bec44de15c, 0x901d7cf73ab0acda, 0xb424dc35095cd810,
+        0xe12e13424bb40e14, 0x8cbccc096f5088cc, 0xafebff0bcb24aaff,
+        0xdbe6fecebdedd5bf, 0x89705f4136b4a598, 0xabcc77118461cefd,
+        0xd6bf94d5e57a42bd, 0x8637bd05af6c69b6, 0xa7c5ac471b478424,
+        0xd1b71758e219652c, 0x83126e978d4fdf3c, 0xa3d70a3d70a3d70b,
+        0xcccccccccccccccd, 0x8000000000000000, 0xa000000000000000,
+        0xc800000000000000, 0xfa00000000000000, 0x9c40000000000000,
+        0xc350000000000000, 0xf424000000000000, 0x9896800000000000,
+        0xbebc200000000000, 0xee6b280000000000, 0x9502f90000000000,
+        0xba43b74000000000, 0xe8d4a51000000000, 0x9184e72a00000000,
+        0xb5e620f480000000, 0xe35fa931a0000000, 0x8e1bc9bf04000000,
+        0xb1a2bc2ec5000000, 0xde0b6b3a76400000, 0x8ac7230489e80000,
+        0xad78ebc5ac620000, 0xd8d726b7177a8000, 0x878678326eac9000,
+        0xa968163f0a57b400, 0xd3c21bcecceda100, 0x84595161401484a0,
+        0xa56fa5b99019a5c8, 0xcecb8f27f4200f3a, 0x813f3978f8940985,
+        0xa18f07d736b90be6, 0xc9f2c9cd04674edf, 0xfc6f7c4045812297,
+        0x9dc5ada82b70b59e, 0xc5371912364ce306, 0xf684df56c3e01bc7,
+        0x9a130b963a6c115d, 0xc097ce7bc90715b4, 0xf0bdc21abb48db21,
+        0x96769950b50d88f5, 0xbc143fa4e250eb32, 0xeb194f8e1ae525fe,
+        0x92efd1b8d0cf37bf, 0xb7abc627050305ae, 0xe596b7b0c643c71a,
+        0x8f7e32ce7bea5c70, 0xb35dbf821ae4f38c, 0xe0352f62a19e306f};
+    return pow10_significands[k - float_info<float>::min_k];
+  }
+
+  struct compute_mul_result {
+    carrier_uint result;
+    bool is_integer;
+  };
+  struct compute_mul_parity_result {
+    bool parity;
+    bool is_integer;
+  };
+
+  static auto compute_mul(carrier_uint u,
+                          const cache_entry_type& cache) noexcept
+      -> compute_mul_result {
+    auto r = umul96_upper64(u, cache);
+    return {static_cast<carrier_uint>(r >> 32),
+            static_cast<carrier_uint>(r) == 0};
+  }
+
+  static auto compute_delta(const cache_entry_type& cache, int beta) noexcept
+      -> uint32_t {
+    return static_cast<uint32_t>(cache >> (64 - 1 - beta));
+  }
+
+  static auto compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta) noexcept
+      -> compute_mul_parity_result {
+    FMT_ASSERT(beta >= 1, "");
+    FMT_ASSERT(beta < 64, "");
+
+    auto r = umul96_lower64(two_f, cache);
+    return {((r >> (64 - beta)) & 1) != 0,
+            static_cast<uint32_t>(r >> (32 - beta)) == 0};
+  }
+
+  static auto compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return static_cast<carrier_uint>(
+        (cache - (cache >> (num_significand_bits<float>() + 2))) >>
+        (64 - num_significand_bits<float>() - 1 - beta));
+  }
+
+  static auto compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return static_cast<carrier_uint>(
+        (cache + (cache >> (num_significand_bits<float>() + 1))) >>
+        (64 - num_significand_bits<float>() - 1 - beta));
+  }
+
+  static auto compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return (static_cast<carrier_uint>(
+                cache >> (64 - num_significand_bits<float>() - 2 - beta)) +
+            1) /
+           2;
+  }
+};
+
+template <> struct cache_accessor<double> {
+  using carrier_uint = float_info<double>::carrier_uint;
+  using cache_entry_type = uint128_fallback;
+
+  static auto get_cached_power(int k) noexcept -> uint128_fallback {
+    FMT_ASSERT(k >= float_info<double>::min_k && k <= float_info<double>::max_k,
+               "k is out of range");
+
+    static constexpr const uint128_fallback pow10_significands[] = {
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+      {0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7b},
+      {0x9faacf3df73609b1, 0x77b191618c54e9ad},
+      {0xc795830d75038c1d, 0xd59df5b9ef6a2418},
+      {0xf97ae3d0d2446f25, 0x4b0573286b44ad1e},
+      {0x9becce62836ac577, 0x4ee367f9430aec33},
+      {0xc2e801fb244576d5, 0x229c41f793cda740},
+      {0xf3a20279ed56d48a, 0x6b43527578c11110},
+      {0x9845418c345644d6, 0x830a13896b78aaaa},
+      {0xbe5691ef416bd60c, 0x23cc986bc656d554},
+      {0xedec366b11c6cb8f, 0x2cbfbe86b7ec8aa9},
+      {0x94b3a202eb1c3f39, 0x7bf7d71432f3d6aa},
+      {0xb9e08a83a5e34f07, 0xdaf5ccd93fb0cc54},
+      {0xe858ad248f5c22c9, 0xd1b3400f8f9cff69},
+      {0x91376c36d99995be, 0x23100809b9c21fa2},
+      {0xb58547448ffffb2d, 0xabd40a0c2832a78b},
+      {0xe2e69915b3fff9f9, 0x16c90c8f323f516d},
+      {0x8dd01fad907ffc3b, 0xae3da7d97f6792e4},
+      {0xb1442798f49ffb4a, 0x99cd11cfdf41779d},
+      {0xdd95317f31c7fa1d, 0x40405643d711d584},
+      {0x8a7d3eef7f1cfc52, 0x482835ea666b2573},
+      {0xad1c8eab5ee43b66, 0xda3243650005eed0},
+      {0xd863b256369d4a40, 0x90bed43e40076a83},
+      {0x873e4f75e2224e68, 0x5a7744a6e804a292},
+      {0xa90de3535aaae202, 0x711515d0a205cb37},
+      {0xd3515c2831559a83, 0x0d5a5b44ca873e04},
+      {0x8412d9991ed58091, 0xe858790afe9486c3},
+      {0xa5178fff668ae0b6, 0x626e974dbe39a873},
+      {0xce5d73ff402d98e3, 0xfb0a3d212dc81290},
+      {0x80fa687f881c7f8e, 0x7ce66634bc9d0b9a},
+      {0xa139029f6a239f72, 0x1c1fffc1ebc44e81},
+      {0xc987434744ac874e, 0xa327ffb266b56221},
+      {0xfbe9141915d7a922, 0x4bf1ff9f0062baa9},
+      {0x9d71ac8fada6c9b5, 0x6f773fc3603db4aa},
+      {0xc4ce17b399107c22, 0xcb550fb4384d21d4},
+      {0xf6019da07f549b2b, 0x7e2a53a146606a49},
+      {0x99c102844f94e0fb, 0x2eda7444cbfc426e},
+      {0xc0314325637a1939, 0xfa911155fefb5309},
+      {0xf03d93eebc589f88, 0x793555ab7eba27cb},
+      {0x96267c7535b763b5, 0x4bc1558b2f3458df},
+      {0xbbb01b9283253ca2, 0x9eb1aaedfb016f17},
+      {0xea9c227723ee8bcb, 0x465e15a979c1cadd},
+      {0x92a1958a7675175f, 0x0bfacd89ec191eca},
+      {0xb749faed14125d36, 0xcef980ec671f667c},
+      {0xe51c79a85916f484, 0x82b7e12780e7401b},
+      {0x8f31cc0937ae58d2, 0xd1b2ecb8b0908811},
+      {0xb2fe3f0b8599ef07, 0x861fa7e6dcb4aa16},
+      {0xdfbdcece67006ac9, 0x67a791e093e1d49b},
+      {0x8bd6a141006042bd, 0xe0c8bb2c5c6d24e1},
+      {0xaecc49914078536d, 0x58fae9f773886e19},
+      {0xda7f5bf590966848, 0xaf39a475506a899f},
+      {0x888f99797a5e012d, 0x6d8406c952429604},
+      {0xaab37fd7d8f58178, 0xc8e5087ba6d33b84},
+      {0xd5605fcdcf32e1d6, 0xfb1e4a9a90880a65},
+      {0x855c3be0a17fcd26, 0x5cf2eea09a550680},
+      {0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481f},
+      {0xd0601d8efc57b08b, 0xf13b94daf124da27},
+      {0x823c12795db6ce57, 0x76c53d08d6b70859},
+      {0xa2cb1717b52481ed, 0x54768c4b0c64ca6f},
+      {0xcb7ddcdda26da268, 0xa9942f5dcf7dfd0a},
+      {0xfe5d54150b090b02, 0xd3f93b35435d7c4d},
+      {0x9efa548d26e5a6e1, 0xc47bc5014a1a6db0},
+      {0xc6b8e9b0709f109a, 0x359ab6419ca1091c},
+      {0xf867241c8cc6d4c0, 0xc30163d203c94b63},
+      {0x9b407691d7fc44f8, 0x79e0de63425dcf1e},
+      {0xc21094364dfb5636, 0x985915fc12f542e5},
+      {0xf294b943e17a2bc4, 0x3e6f5b7b17b2939e},
+      {0x979cf3ca6cec5b5a, 0xa705992ceecf9c43},
+      {0xbd8430bd08277231, 0x50c6ff782a838354},
+      {0xece53cec4a314ebd, 0xa4f8bf5635246429},
+      {0x940f4613ae5ed136, 0x871b7795e136be9a},
+      {0xb913179899f68584, 0x28e2557b59846e40},
+      {0xe757dd7ec07426e5, 0x331aeada2fe589d0},
+      {0x9096ea6f3848984f, 0x3ff0d2c85def7622},
+      {0xb4bca50b065abe63, 0x0fed077a756b53aa},
+      {0xe1ebce4dc7f16dfb, 0xd3e8495912c62895},
+      {0x8d3360f09cf6e4bd, 0x64712dd7abbbd95d},
+      {0xb080392cc4349dec, 0xbd8d794d96aacfb4},
+      {0xdca04777f541c567, 0xecf0d7a0fc5583a1},
+      {0x89e42caaf9491b60, 0xf41686c49db57245},
+      {0xac5d37d5b79b6239, 0x311c2875c522ced6},
+      {0xd77485cb25823ac7, 0x7d633293366b828c},
+      {0x86a8d39ef77164bc, 0xae5dff9c02033198},
+      {0xa8530886b54dbdeb, 0xd9f57f830283fdfd},
+      {0xd267caa862a12d66, 0xd072df63c324fd7c},
+      {0x8380dea93da4bc60, 0x4247cb9e59f71e6e},
+      {0xa46116538d0deb78, 0x52d9be85f074e609},
+      {0xcd795be870516656, 0x67902e276c921f8c},
+      {0x806bd9714632dff6, 0x00ba1cd8a3db53b7},
+      {0xa086cfcd97bf97f3, 0x80e8a40eccd228a5},
+      {0xc8a883c0fdaf7df0, 0x6122cd128006b2ce},
+      {0xfad2a4b13d1b5d6c, 0x796b805720085f82},
+      {0x9cc3a6eec6311a63, 0xcbe3303674053bb1},
+      {0xc3f490aa77bd60fc, 0xbedbfc4411068a9d},
+      {0xf4f1b4d515acb93b, 0xee92fb5515482d45},
+      {0x991711052d8bf3c5, 0x751bdd152d4d1c4b},
+      {0xbf5cd54678eef0b6, 0xd262d45a78a0635e},
+      {0xef340a98172aace4, 0x86fb897116c87c35},
+      {0x9580869f0e7aac0e, 0xd45d35e6ae3d4da1},
+      {0xbae0a846d2195712, 0x8974836059cca10a},
+      {0xe998d258869facd7, 0x2bd1a438703fc94c},
+      {0x91ff83775423cc06, 0x7b6306a34627ddd0},
+      {0xb67f6455292cbf08, 0x1a3bc84c17b1d543},
+      {0xe41f3d6a7377eeca, 0x20caba5f1d9e4a94},
+      {0x8e938662882af53e, 0x547eb47b7282ee9d},
+      {0xb23867fb2a35b28d, 0xe99e619a4f23aa44},
+      {0xdec681f9f4c31f31, 0x6405fa00e2ec94d5},
+      {0x8b3c113c38f9f37e, 0xde83bc408dd3dd05},
+      {0xae0b158b4738705e, 0x9624ab50b148d446},
+      {0xd98ddaee19068c76, 0x3badd624dd9b0958},
+      {0x87f8a8d4cfa417c9, 0xe54ca5d70a80e5d7},
+      {0xa9f6d30a038d1dbc, 0x5e9fcf4ccd211f4d},
+      {0xd47487cc8470652b, 0x7647c32000696720},
+      {0x84c8d4dfd2c63f3b, 0x29ecd9f40041e074},
+      {0xa5fb0a17c777cf09, 0xf468107100525891},
+      {0xcf79cc9db955c2cc, 0x7182148d4066eeb5},
+      {0x81ac1fe293d599bf, 0xc6f14cd848405531},
+      {0xa21727db38cb002f, 0xb8ada00e5a506a7d},
+      {0xca9cf1d206fdc03b, 0xa6d90811f0e4851d},
+      {0xfd442e4688bd304a, 0x908f4a166d1da664},
+      {0x9e4a9cec15763e2e, 0x9a598e4e043287ff},
+      {0xc5dd44271ad3cdba, 0x40eff1e1853f29fe},
+      {0xf7549530e188c128, 0xd12bee59e68ef47d},
+      {0x9a94dd3e8cf578b9, 0x82bb74f8301958cf},
+      {0xc13a148e3032d6e7, 0xe36a52363c1faf02},
+      {0xf18899b1bc3f8ca1, 0xdc44e6c3cb279ac2},
+      {0x96f5600f15a7b7e5, 0x29ab103a5ef8c0ba},
+      {0xbcb2b812db11a5de, 0x7415d448f6b6f0e8},
+      {0xebdf661791d60f56, 0x111b495b3464ad22},
+      {0x936b9fcebb25c995, 0xcab10dd900beec35},
+      {0xb84687c269ef3bfb, 0x3d5d514f40eea743},
+      {0xe65829b3046b0afa, 0x0cb4a5a3112a5113},
+      {0x8ff71a0fe2c2e6dc, 0x47f0e785eaba72ac},
+      {0xb3f4e093db73a093, 0x59ed216765690f57},
+      {0xe0f218b8d25088b8, 0x306869c13ec3532d},
+      {0x8c974f7383725573, 0x1e414218c73a13fc},
+      {0xafbd2350644eeacf, 0xe5d1929ef90898fb},
+      {0xdbac6c247d62a583, 0xdf45f746b74abf3a},
+      {0x894bc396ce5da772, 0x6b8bba8c328eb784},
+      {0xab9eb47c81f5114f, 0x066ea92f3f326565},
+      {0xd686619ba27255a2, 0xc80a537b0efefebe},
+      {0x8613fd0145877585, 0xbd06742ce95f5f37},
+      {0xa798fc4196e952e7, 0x2c48113823b73705},
+      {0xd17f3b51fca3a7a0, 0xf75a15862ca504c6},
+      {0x82ef85133de648c4, 0x9a984d73dbe722fc},
+      {0xa3ab66580d5fdaf5, 0xc13e60d0d2e0ebbb},
+      {0xcc963fee10b7d1b3, 0x318df905079926a9},
+      {0xffbbcfe994e5c61f, 0xfdf17746497f7053},
+      {0x9fd561f1fd0f9bd3, 0xfeb6ea8bedefa634},
+      {0xc7caba6e7c5382c8, 0xfe64a52ee96b8fc1},
+      {0xf9bd690a1b68637b, 0x3dfdce7aa3c673b1},
+      {0x9c1661a651213e2d, 0x06bea10ca65c084f},
+      {0xc31bfa0fe5698db8, 0x486e494fcff30a63},
+      {0xf3e2f893dec3f126, 0x5a89dba3c3efccfb},
+      {0x986ddb5c6b3a76b7, 0xf89629465a75e01d},
+      {0xbe89523386091465, 0xf6bbb397f1135824},
+      {0xee2ba6c0678b597f, 0x746aa07ded582e2d},
+      {0x94db483840b717ef, 0xa8c2a44eb4571cdd},
+      {0xba121a4650e4ddeb, 0x92f34d62616ce414},
+      {0xe896a0d7e51e1566, 0x77b020baf9c81d18},
+      {0x915e2486ef32cd60, 0x0ace1474dc1d122f},
+      {0xb5b5ada8aaff80b8, 0x0d819992132456bb},
+      {0xe3231912d5bf60e6, 0x10e1fff697ed6c6a},
+      {0x8df5efabc5979c8f, 0xca8d3ffa1ef463c2},
+      {0xb1736b96b6fd83b3, 0xbd308ff8a6b17cb3},
+      {0xddd0467c64bce4a0, 0xac7cb3f6d05ddbdf},
+      {0x8aa22c0dbef60ee4, 0x6bcdf07a423aa96c},
+      {0xad4ab7112eb3929d, 0x86c16c98d2c953c7},
+      {0xd89d64d57a607744, 0xe871c7bf077ba8b8},
+      {0x87625f056c7c4a8b, 0x11471cd764ad4973},
+      {0xa93af6c6c79b5d2d, 0xd598e40d3dd89bd0},
+      {0xd389b47879823479, 0x4aff1d108d4ec2c4},
+      {0x843610cb4bf160cb, 0xcedf722a585139bb},
+      {0xa54394fe1eedb8fe, 0xc2974eb4ee658829},
+      {0xce947a3da6a9273e, 0x733d226229feea33},
+      {0x811ccc668829b887, 0x0806357d5a3f5260},
+      {0xa163ff802a3426a8, 0xca07c2dcb0cf26f8},
+      {0xc9bcff6034c13052, 0xfc89b393dd02f0b6},
+      {0xfc2c3f3841f17c67, 0xbbac2078d443ace3},
+      {0x9d9ba7832936edc0, 0xd54b944b84aa4c0e},
+      {0xc5029163f384a931, 0x0a9e795e65d4df12},
+      {0xf64335bcf065d37d, 0x4d4617b5ff4a16d6},
+      {0x99ea0196163fa42e, 0x504bced1bf8e4e46},
+      {0xc06481fb9bcf8d39, 0xe45ec2862f71e1d7},
+      {0xf07da27a82c37088, 0x5d767327bb4e5a4d},
+      {0x964e858c91ba2655, 0x3a6a07f8d510f870},
+      {0xbbe226efb628afea, 0x890489f70a55368c},
+      {0xeadab0aba3b2dbe5, 0x2b45ac74ccea842f},
+      {0x92c8ae6b464fc96f, 0x3b0b8bc90012929e},
+      {0xb77ada0617e3bbcb, 0x09ce6ebb40173745},
+      {0xe55990879ddcaabd, 0xcc420a6a101d0516},
+      {0x8f57fa54c2a9eab6, 0x9fa946824a12232e},
+      {0xb32df8e9f3546564, 0x47939822dc96abfa},
+      {0xdff9772470297ebd, 0x59787e2b93bc56f8},
+      {0x8bfbea76c619ef36, 0x57eb4edb3c55b65b},
+      {0xaefae51477a06b03, 0xede622920b6b23f2},
+      {0xdab99e59958885c4, 0xe95fab368e45ecee},
+      {0x88b402f7fd75539b, 0x11dbcb0218ebb415},
+      {0xaae103b5fcd2a881, 0xd652bdc29f26a11a},
+      {0xd59944a37c0752a2, 0x4be76d3346f04960},
+      {0x857fcae62d8493a5, 0x6f70a4400c562ddc},
+      {0xa6dfbd9fb8e5b88e, 0xcb4ccd500f6bb953},
+      {0xd097ad07a71f26b2, 0x7e2000a41346a7a8},
+      {0x825ecc24c873782f, 0x8ed400668c0c28c9},
+      {0xa2f67f2dfa90563b, 0x728900802f0f32fb},
+      {0xcbb41ef979346bca, 0x4f2b40a03ad2ffba},
+      {0xfea126b7d78186bc, 0xe2f610c84987bfa9},
+      {0x9f24b832e6b0f436, 0x0dd9ca7d2df4d7ca},
+      {0xc6ede63fa05d3143, 0x91503d1c79720dbc},
+      {0xf8a95fcf88747d94, 0x75a44c6397ce912b},
+      {0x9b69dbe1b548ce7c, 0xc986afbe3ee11abb},
+      {0xc24452da229b021b, 0xfbe85badce996169},
+      {0xf2d56790ab41c2a2, 0xfae27299423fb9c4},
+      {0x97c560ba6b0919a5, 0xdccd879fc967d41b},
+      {0xbdb6b8e905cb600f, 0x5400e987bbc1c921},
+      {0xed246723473e3813, 0x290123e9aab23b69},
+      {0x9436c0760c86e30b, 0xf9a0b6720aaf6522},
+      {0xb94470938fa89bce, 0xf808e40e8d5b3e6a},
+      {0xe7958cb87392c2c2, 0xb60b1d1230b20e05},
+      {0x90bd77f3483bb9b9, 0xb1c6f22b5e6f48c3},
+      {0xb4ecd5f01a4aa828, 0x1e38aeb6360b1af4},
+      {0xe2280b6c20dd5232, 0x25c6da63c38de1b1},
+      {0x8d590723948a535f, 0x579c487e5a38ad0f},
+      {0xb0af48ec79ace837, 0x2d835a9df0c6d852},
+      {0xdcdb1b2798182244, 0xf8e431456cf88e66},
+      {0x8a08f0f8bf0f156b, 0x1b8e9ecb641b5900},
+      {0xac8b2d36eed2dac5, 0xe272467e3d222f40},
+      {0xd7adf884aa879177, 0x5b0ed81dcc6abb10},
+      {0x86ccbb52ea94baea, 0x98e947129fc2b4ea},
+      {0xa87fea27a539e9a5, 0x3f2398d747b36225},
+      {0xd29fe4b18e88640e, 0x8eec7f0d19a03aae},
+      {0x83a3eeeef9153e89, 0x1953cf68300424ad},
+      {0xa48ceaaab75a8e2b, 0x5fa8c3423c052dd8},
+      {0xcdb02555653131b6, 0x3792f412cb06794e},
+      {0x808e17555f3ebf11, 0xe2bbd88bbee40bd1},
+      {0xa0b19d2ab70e6ed6, 0x5b6aceaeae9d0ec5},
+      {0xc8de047564d20a8b, 0xf245825a5a445276},
+      {0xfb158592be068d2e, 0xeed6e2f0f0d56713},
+      {0x9ced737bb6c4183d, 0x55464dd69685606c},
+      {0xc428d05aa4751e4c, 0xaa97e14c3c26b887},
+      {0xf53304714d9265df, 0xd53dd99f4b3066a9},
+      {0x993fe2c6d07b7fab, 0xe546a8038efe402a},
+      {0xbf8fdb78849a5f96, 0xde98520472bdd034},
+      {0xef73d256a5c0f77c, 0x963e66858f6d4441},
+      {0x95a8637627989aad, 0xdde7001379a44aa9},
+      {0xbb127c53b17ec159, 0x5560c018580d5d53},
+      {0xe9d71b689dde71af, 0xaab8f01e6e10b4a7},
+      {0x9226712162ab070d, 0xcab3961304ca70e9},
+      {0xb6b00d69bb55c8d1, 0x3d607b97c5fd0d23},
+      {0xe45c10c42a2b3b05, 0x8cb89a7db77c506b},
+      {0x8eb98a7a9a5b04e3, 0x77f3608e92adb243},
+      {0xb267ed1940f1c61c, 0x55f038b237591ed4},
+      {0xdf01e85f912e37a3, 0x6b6c46dec52f6689},
+      {0x8b61313bbabce2c6, 0x2323ac4b3b3da016},
+      {0xae397d8aa96c1b77, 0xabec975e0a0d081b},
+      {0xd9c7dced53c72255, 0x96e7bd358c904a22},
+      {0x881cea14545c7575, 0x7e50d64177da2e55},
+      {0xaa242499697392d2, 0xdde50bd1d5d0b9ea},
+      {0xd4ad2dbfc3d07787, 0x955e4ec64b44e865},
+      {0x84ec3c97da624ab4, 0xbd5af13bef0b113f},
+      {0xa6274bbdd0fadd61, 0xecb1ad8aeacdd58f},
+      {0xcfb11ead453994ba, 0x67de18eda5814af3},
+      {0x81ceb32c4b43fcf4, 0x80eacf948770ced8},
+      {0xa2425ff75e14fc31, 0xa1258379a94d028e},
+      {0xcad2f7f5359a3b3e, 0x096ee45813a04331},
+      {0xfd87b5f28300ca0d, 0x8bca9d6e188853fd},
+      {0x9e74d1b791e07e48, 0x775ea264cf55347e},
+      {0xc612062576589dda, 0x95364afe032a819e},
+      {0xf79687aed3eec551, 0x3a83ddbd83f52205},
+      {0x9abe14cd44753b52, 0xc4926a9672793543},
+      {0xc16d9a0095928a27, 0x75b7053c0f178294},
+      {0xf1c90080baf72cb1, 0x5324c68b12dd6339},
+      {0x971da05074da7bee, 0xd3f6fc16ebca5e04},
+      {0xbce5086492111aea, 0x88f4bb1ca6bcf585},
+      {0xec1e4a7db69561a5, 0x2b31e9e3d06c32e6},
+      {0x9392ee8e921d5d07, 0x3aff322e62439fd0},
+      {0xb877aa3236a4b449, 0x09befeb9fad487c3},
+      {0xe69594bec44de15b, 0x4c2ebe687989a9b4},
+      {0x901d7cf73ab0acd9, 0x0f9d37014bf60a11},
+      {0xb424dc35095cd80f, 0x538484c19ef38c95},
+      {0xe12e13424bb40e13, 0x2865a5f206b06fba},
+      {0x8cbccc096f5088cb, 0xf93f87b7442e45d4},
+      {0xafebff0bcb24aafe, 0xf78f69a51539d749},
+      {0xdbe6fecebdedd5be, 0xb573440e5a884d1c},
+      {0x89705f4136b4a597, 0x31680a88f8953031},
+      {0xabcc77118461cefc, 0xfdc20d2b36ba7c3e},
+      {0xd6bf94d5e57a42bc, 0x3d32907604691b4d},
+      {0x8637bd05af6c69b5, 0xa63f9a49c2c1b110},
+      {0xa7c5ac471b478423, 0x0fcf80dc33721d54},
+      {0xd1b71758e219652b, 0xd3c36113404ea4a9},
+      {0x83126e978d4fdf3b, 0x645a1cac083126ea},
+      {0xa3d70a3d70a3d70a, 0x3d70a3d70a3d70a4},
+      {0xcccccccccccccccc, 0xcccccccccccccccd},
+      {0x8000000000000000, 0x0000000000000000},
+      {0xa000000000000000, 0x0000000000000000},
+      {0xc800000000000000, 0x0000000000000000},
+      {0xfa00000000000000, 0x0000000000000000},
+      {0x9c40000000000000, 0x0000000000000000},
+      {0xc350000000000000, 0x0000000000000000},
+      {0xf424000000000000, 0x0000000000000000},
+      {0x9896800000000000, 0x0000000000000000},
+      {0xbebc200000000000, 0x0000000000000000},
+      {0xee6b280000000000, 0x0000000000000000},
+      {0x9502f90000000000, 0x0000000000000000},
+      {0xba43b74000000000, 0x0000000000000000},
+      {0xe8d4a51000000000, 0x0000000000000000},
+      {0x9184e72a00000000, 0x0000000000000000},
+      {0xb5e620f480000000, 0x0000000000000000},
+      {0xe35fa931a0000000, 0x0000000000000000},
+      {0x8e1bc9bf04000000, 0x0000000000000000},
+      {0xb1a2bc2ec5000000, 0x0000000000000000},
+      {0xde0b6b3a76400000, 0x0000000000000000},
+      {0x8ac7230489e80000, 0x0000000000000000},
+      {0xad78ebc5ac620000, 0x0000000000000000},
+      {0xd8d726b7177a8000, 0x0000000000000000},
+      {0x878678326eac9000, 0x0000000000000000},
+      {0xa968163f0a57b400, 0x0000000000000000},
+      {0xd3c21bcecceda100, 0x0000000000000000},
+      {0x84595161401484a0, 0x0000000000000000},
+      {0xa56fa5b99019a5c8, 0x0000000000000000},
+      {0xcecb8f27f4200f3a, 0x0000000000000000},
+      {0x813f3978f8940984, 0x4000000000000000},
+      {0xa18f07d736b90be5, 0x5000000000000000},
+      {0xc9f2c9cd04674ede, 0xa400000000000000},
+      {0xfc6f7c4045812296, 0x4d00000000000000},
+      {0x9dc5ada82b70b59d, 0xf020000000000000},
+      {0xc5371912364ce305, 0x6c28000000000000},
+      {0xf684df56c3e01bc6, 0xc732000000000000},
+      {0x9a130b963a6c115c, 0x3c7f400000000000},
+      {0xc097ce7bc90715b3, 0x4b9f100000000000},
+      {0xf0bdc21abb48db20, 0x1e86d40000000000},
+      {0x96769950b50d88f4, 0x1314448000000000},
+      {0xbc143fa4e250eb31, 0x17d955a000000000},
+      {0xeb194f8e1ae525fd, 0x5dcfab0800000000},
+      {0x92efd1b8d0cf37be, 0x5aa1cae500000000},
+      {0xb7abc627050305ad, 0xf14a3d9e40000000},
+      {0xe596b7b0c643c719, 0x6d9ccd05d0000000},
+      {0x8f7e32ce7bea5c6f, 0xe4820023a2000000},
+      {0xb35dbf821ae4f38b, 0xdda2802c8a800000},
+      {0xe0352f62a19e306e, 0xd50b2037ad200000},
+      {0x8c213d9da502de45, 0x4526f422cc340000},
+      {0xaf298d050e4395d6, 0x9670b12b7f410000},
+      {0xdaf3f04651d47b4c, 0x3c0cdd765f114000},
+      {0x88d8762bf324cd0f, 0xa5880a69fb6ac800},
+      {0xab0e93b6efee0053, 0x8eea0d047a457a00},
+      {0xd5d238a4abe98068, 0x72a4904598d6d880},
+      {0x85a36366eb71f041, 0x47a6da2b7f864750},
+      {0xa70c3c40a64e6c51, 0x999090b65f67d924},
+      {0xd0cf4b50cfe20765, 0xfff4b4e3f741cf6d},
+      {0x82818f1281ed449f, 0xbff8f10e7a8921a5},
+      {0xa321f2d7226895c7, 0xaff72d52192b6a0e},
+      {0xcbea6f8ceb02bb39, 0x9bf4f8a69f764491},
+      {0xfee50b7025c36a08, 0x02f236d04753d5b5},
+      {0x9f4f2726179a2245, 0x01d762422c946591},
+      {0xc722f0ef9d80aad6, 0x424d3ad2b7b97ef6},
+      {0xf8ebad2b84e0d58b, 0xd2e0898765a7deb3},
+      {0x9b934c3b330c8577, 0x63cc55f49f88eb30},
+      {0xc2781f49ffcfa6d5, 0x3cbf6b71c76b25fc},
+      {0xf316271c7fc3908a, 0x8bef464e3945ef7b},
+      {0x97edd871cfda3a56, 0x97758bf0e3cbb5ad},
+      {0xbde94e8e43d0c8ec, 0x3d52eeed1cbea318},
+      {0xed63a231d4c4fb27, 0x4ca7aaa863ee4bde},
+      {0x945e455f24fb1cf8, 0x8fe8caa93e74ef6b},
+      {0xb975d6b6ee39e436, 0xb3e2fd538e122b45},
+      {0xe7d34c64a9c85d44, 0x60dbbca87196b617},
+      {0x90e40fbeea1d3a4a, 0xbc8955e946fe31ce},
+      {0xb51d13aea4a488dd, 0x6babab6398bdbe42},
+      {0xe264589a4dcdab14, 0xc696963c7eed2dd2},
+      {0x8d7eb76070a08aec, 0xfc1e1de5cf543ca3},
+      {0xb0de65388cc8ada8, 0x3b25a55f43294bcc},
+      {0xdd15fe86affad912, 0x49ef0eb713f39ebf},
+      {0x8a2dbf142dfcc7ab, 0x6e3569326c784338},
+      {0xacb92ed9397bf996, 0x49c2c37f07965405},
+      {0xd7e77a8f87daf7fb, 0xdc33745ec97be907},
+      {0x86f0ac99b4e8dafd, 0x69a028bb3ded71a4},
+      {0xa8acd7c0222311bc, 0xc40832ea0d68ce0d},
+      {0xd2d80db02aabd62b, 0xf50a3fa490c30191},
+      {0x83c7088e1aab65db, 0x792667c6da79e0fb},
+      {0xa4b8cab1a1563f52, 0x577001b891185939},
+      {0xcde6fd5e09abcf26, 0xed4c0226b55e6f87},
+      {0x80b05e5ac60b6178, 0x544f8158315b05b5},
+      {0xa0dc75f1778e39d6, 0x696361ae3db1c722},
+      {0xc913936dd571c84c, 0x03bc3a19cd1e38ea},
+      {0xfb5878494ace3a5f, 0x04ab48a04065c724},
+      {0x9d174b2dcec0e47b, 0x62eb0d64283f9c77},
+      {0xc45d1df942711d9a, 0x3ba5d0bd324f8395},
+      {0xf5746577930d6500, 0xca8f44ec7ee3647a},
+      {0x9968bf6abbe85f20, 0x7e998b13cf4e1ecc},
+      {0xbfc2ef456ae276e8, 0x9e3fedd8c321a67f},
+      {0xefb3ab16c59b14a2, 0xc5cfe94ef3ea101f},
+      {0x95d04aee3b80ece5, 0xbba1f1d158724a13},
+      {0xbb445da9ca61281f, 0x2a8a6e45ae8edc98},
+      {0xea1575143cf97226, 0xf52d09d71a3293be},
+      {0x924d692ca61be758, 0x593c2626705f9c57},
+      {0xb6e0c377cfa2e12e, 0x6f8b2fb00c77836d},
+      {0xe498f455c38b997a, 0x0b6dfb9c0f956448},
+      {0x8edf98b59a373fec, 0x4724bd4189bd5ead},
+      {0xb2977ee300c50fe7, 0x58edec91ec2cb658},
+      {0xdf3d5e9bc0f653e1, 0x2f2967b66737e3ee},
+      {0x8b865b215899f46c, 0xbd79e0d20082ee75},
+      {0xae67f1e9aec07187, 0xecd8590680a3aa12},
+      {0xda01ee641a708de9, 0xe80e6f4820cc9496},
+      {0x884134fe908658b2, 0x3109058d147fdcde},
+      {0xaa51823e34a7eede, 0xbd4b46f0599fd416},
+      {0xd4e5e2cdc1d1ea96, 0x6c9e18ac7007c91b},
+      {0x850fadc09923329e, 0x03e2cf6bc604ddb1},
+      {0xa6539930bf6bff45, 0x84db8346b786151d},
+      {0xcfe87f7cef46ff16, 0xe612641865679a64},
+      {0x81f14fae158c5f6e, 0x4fcb7e8f3f60c07f},
+      {0xa26da3999aef7749, 0xe3be5e330f38f09e},
+      {0xcb090c8001ab551c, 0x5cadf5bfd3072cc6},
+      {0xfdcb4fa002162a63, 0x73d9732fc7c8f7f7},
+      {0x9e9f11c4014dda7e, 0x2867e7fddcdd9afb},
+      {0xc646d63501a1511d, 0xb281e1fd541501b9},
+      {0xf7d88bc24209a565, 0x1f225a7ca91a4227},
+      {0x9ae757596946075f, 0x3375788de9b06959},
+      {0xc1a12d2fc3978937, 0x0052d6b1641c83af},
+      {0xf209787bb47d6b84, 0xc0678c5dbd23a49b},
+      {0x9745eb4d50ce6332, 0xf840b7ba963646e1},
+      {0xbd176620a501fbff, 0xb650e5a93bc3d899},
+      {0xec5d3fa8ce427aff, 0xa3e51f138ab4cebf},
+      {0x93ba47c980e98cdf, 0xc66f336c36b10138},
+      {0xb8a8d9bbe123f017, 0xb80b0047445d4185},
+      {0xe6d3102ad96cec1d, 0xa60dc059157491e6},
+      {0x9043ea1ac7e41392, 0x87c89837ad68db30},
+      {0xb454e4a179dd1877, 0x29babe4598c311fc},
+      {0xe16a1dc9d8545e94, 0xf4296dd6fef3d67b},
+      {0x8ce2529e2734bb1d, 0x1899e4a65f58660d},
+      {0xb01ae745b101e9e4, 0x5ec05dcff72e7f90},
+      {0xdc21a1171d42645d, 0x76707543f4fa1f74},
+      {0x899504ae72497eba, 0x6a06494a791c53a9},
+      {0xabfa45da0edbde69, 0x0487db9d17636893},
+      {0xd6f8d7509292d603, 0x45a9d2845d3c42b7},
+      {0x865b86925b9bc5c2, 0x0b8a2392ba45a9b3},
+      {0xa7f26836f282b732, 0x8e6cac7768d7141f},
+      {0xd1ef0244af2364ff, 0x3207d795430cd927},
+      {0x8335616aed761f1f, 0x7f44e6bd49e807b9},
+      {0xa402b9c5a8d3a6e7, 0x5f16206c9c6209a7},
+      {0xcd036837130890a1, 0x36dba887c37a8c10},
+      {0x802221226be55a64, 0xc2494954da2c978a},
+      {0xa02aa96b06deb0fd, 0xf2db9baa10b7bd6d},
+      {0xc83553c5c8965d3d, 0x6f92829494e5acc8},
+      {0xfa42a8b73abbf48c, 0xcb772339ba1f17fa},
+      {0x9c69a97284b578d7, 0xff2a760414536efc},
+      {0xc38413cf25e2d70d, 0xfef5138519684abb},
+      {0xf46518c2ef5b8cd1, 0x7eb258665fc25d6a},
+      {0x98bf2f79d5993802, 0xef2f773ffbd97a62},
+      {0xbeeefb584aff8603, 0xaafb550ffacfd8fb},
+      {0xeeaaba2e5dbf6784, 0x95ba2a53f983cf39},
+      {0x952ab45cfa97a0b2, 0xdd945a747bf26184},
+      {0xba756174393d88df, 0x94f971119aeef9e5},
+      {0xe912b9d1478ceb17, 0x7a37cd5601aab85e},
+      {0x91abb422ccb812ee, 0xac62e055c10ab33b},
+      {0xb616a12b7fe617aa, 0x577b986b314d600a},
+      {0xe39c49765fdf9d94, 0xed5a7e85fda0b80c},
+      {0x8e41ade9fbebc27d, 0x14588f13be847308},
+      {0xb1d219647ae6b31c, 0x596eb2d8ae258fc9},
+      {0xde469fbd99a05fe3, 0x6fca5f8ed9aef3bc},
+      {0x8aec23d680043bee, 0x25de7bb9480d5855},
+      {0xada72ccc20054ae9, 0xaf561aa79a10ae6b},
+      {0xd910f7ff28069da4, 0x1b2ba1518094da05},
+      {0x87aa9aff79042286, 0x90fb44d2f05d0843},
+      {0xa99541bf57452b28, 0x353a1607ac744a54},
+      {0xd3fa922f2d1675f2, 0x42889b8997915ce9},
+      {0x847c9b5d7c2e09b7, 0x69956135febada12},
+      {0xa59bc234db398c25, 0x43fab9837e699096},
+      {0xcf02b2c21207ef2e, 0x94f967e45e03f4bc},
+      {0x8161afb94b44f57d, 0x1d1be0eebac278f6},
+      {0xa1ba1ba79e1632dc, 0x6462d92a69731733},
+      {0xca28a291859bbf93, 0x7d7b8f7503cfdcff},
+      {0xfcb2cb35e702af78, 0x5cda735244c3d43f},
+      {0x9defbf01b061adab, 0x3a0888136afa64a8},
+      {0xc56baec21c7a1916, 0x088aaa1845b8fdd1},
+      {0xf6c69a72a3989f5b, 0x8aad549e57273d46},
+      {0x9a3c2087a63f6399, 0x36ac54e2f678864c},
+      {0xc0cb28a98fcf3c7f, 0x84576a1bb416a7de},
+      {0xf0fdf2d3f3c30b9f, 0x656d44a2a11c51d6},
+      {0x969eb7c47859e743, 0x9f644ae5a4b1b326},
+      {0xbc4665b596706114, 0x873d5d9f0dde1fef},
+      {0xeb57ff22fc0c7959, 0xa90cb506d155a7eb},
+      {0x9316ff75dd87cbd8, 0x09a7f12442d588f3},
+      {0xb7dcbf5354e9bece, 0x0c11ed6d538aeb30},
+      {0xe5d3ef282a242e81, 0x8f1668c8a86da5fb},
+      {0x8fa475791a569d10, 0xf96e017d694487bd},
+      {0xb38d92d760ec4455, 0x37c981dcc395a9ad},
+      {0xe070f78d3927556a, 0x85bbe253f47b1418},
+      {0x8c469ab843b89562, 0x93956d7478ccec8f},
+      {0xaf58416654a6babb, 0x387ac8d1970027b3},
+      {0xdb2e51bfe9d0696a, 0x06997b05fcc0319f},
+      {0x88fcf317f22241e2, 0x441fece3bdf81f04},
+      {0xab3c2fddeeaad25a, 0xd527e81cad7626c4},
+      {0xd60b3bd56a5586f1, 0x8a71e223d8d3b075},
+      {0x85c7056562757456, 0xf6872d5667844e4a},
+      {0xa738c6bebb12d16c, 0xb428f8ac016561dc},
+      {0xd106f86e69d785c7, 0xe13336d701beba53},
+      {0x82a45b450226b39c, 0xecc0024661173474},
+      {0xa34d721642b06084, 0x27f002d7f95d0191},
+      {0xcc20ce9bd35c78a5, 0x31ec038df7b441f5},
+      {0xff290242c83396ce, 0x7e67047175a15272},
+      {0x9f79a169bd203e41, 0x0f0062c6e984d387},
+      {0xc75809c42c684dd1, 0x52c07b78a3e60869},
+      {0xf92e0c3537826145, 0xa7709a56ccdf8a83},
+      {0x9bbcc7a142b17ccb, 0x88a66076400bb692},
+      {0xc2abf989935ddbfe, 0x6acff893d00ea436},
+      {0xf356f7ebf83552fe, 0x0583f6b8c4124d44},
+      {0x98165af37b2153de, 0xc3727a337a8b704b},
+      {0xbe1bf1b059e9a8d6, 0x744f18c0592e4c5d},
+      {0xeda2ee1c7064130c, 0x1162def06f79df74},
+      {0x9485d4d1c63e8be7, 0x8addcb5645ac2ba9},
+      {0xb9a74a0637ce2ee1, 0x6d953e2bd7173693},
+      {0xe8111c87c5c1ba99, 0xc8fa8db6ccdd0438},
+      {0x910ab1d4db9914a0, 0x1d9c9892400a22a3},
+      {0xb54d5e4a127f59c8, 0x2503beb6d00cab4c},
+      {0xe2a0b5dc971f303a, 0x2e44ae64840fd61e},
+      {0x8da471a9de737e24, 0x5ceaecfed289e5d3},
+      {0xb10d8e1456105dad, 0x7425a83e872c5f48},
+      {0xdd50f1996b947518, 0xd12f124e28f7771a},
+      {0x8a5296ffe33cc92f, 0x82bd6b70d99aaa70},
+      {0xace73cbfdc0bfb7b, 0x636cc64d1001550c},
+      {0xd8210befd30efa5a, 0x3c47f7e05401aa4f},
+      {0x8714a775e3e95c78, 0x65acfaec34810a72},
+      {0xa8d9d1535ce3b396, 0x7f1839a741a14d0e},
+      {0xd31045a8341ca07c, 0x1ede48111209a051},
+      {0x83ea2b892091e44d, 0x934aed0aab460433},
+      {0xa4e4b66b68b65d60, 0xf81da84d56178540},
+      {0xce1de40642e3f4b9, 0x36251260ab9d668f},
+      {0x80d2ae83e9ce78f3, 0xc1d72b7c6b42601a},
+      {0xa1075a24e4421730, 0xb24cf65b8612f820},
+      {0xc94930ae1d529cfc, 0xdee033f26797b628},
+      {0xfb9b7cd9a4a7443c, 0x169840ef017da3b2},
+      {0x9d412e0806e88aa5, 0x8e1f289560ee864f},
+      {0xc491798a08a2ad4e, 0xf1a6f2bab92a27e3},
+      {0xf5b5d7ec8acb58a2, 0xae10af696774b1dc},
+      {0x9991a6f3d6bf1765, 0xacca6da1e0a8ef2a},
+      {0xbff610b0cc6edd3f, 0x17fd090a58d32af4},
+      {0xeff394dcff8a948e, 0xddfc4b4cef07f5b1},
+      {0x95f83d0a1fb69cd9, 0x4abdaf101564f98f},
+      {0xbb764c4ca7a4440f, 0x9d6d1ad41abe37f2},
+      {0xea53df5fd18d5513, 0x84c86189216dc5ee},
+      {0x92746b9be2f8552c, 0x32fd3cf5b4e49bb5},
+      {0xb7118682dbb66a77, 0x3fbc8c33221dc2a2},
+      {0xe4d5e82392a40515, 0x0fabaf3feaa5334b},
+      {0x8f05b1163ba6832d, 0x29cb4d87f2a7400f},
+      {0xb2c71d5bca9023f8, 0x743e20e9ef511013},
+      {0xdf78e4b2bd342cf6, 0x914da9246b255417},
+      {0x8bab8eefb6409c1a, 0x1ad089b6c2f7548f},
+      {0xae9672aba3d0c320, 0xa184ac2473b529b2},
+      {0xda3c0f568cc4f3e8, 0xc9e5d72d90a2741f},
+      {0x8865899617fb1871, 0x7e2fa67c7a658893},
+      {0xaa7eebfb9df9de8d, 0xddbb901b98feeab8},
+      {0xd51ea6fa85785631, 0x552a74227f3ea566},
+      {0x8533285c936b35de, 0xd53a88958f872760},
+      {0xa67ff273b8460356, 0x8a892abaf368f138},
+      {0xd01fef10a657842c, 0x2d2b7569b0432d86},
+      {0x8213f56a67f6b29b, 0x9c3b29620e29fc74},
+      {0xa298f2c501f45f42, 0x8349f3ba91b47b90},
+      {0xcb3f2f7642717713, 0x241c70a936219a74},
+      {0xfe0efb53d30dd4d7, 0xed238cd383aa0111},
+      {0x9ec95d1463e8a506, 0xf4363804324a40ab},
+      {0xc67bb4597ce2ce48, 0xb143c6053edcd0d6},
+      {0xf81aa16fdc1b81da, 0xdd94b7868e94050b},
+      {0x9b10a4e5e9913128, 0xca7cf2b4191c8327},
+      {0xc1d4ce1f63f57d72, 0xfd1c2f611f63a3f1},
+      {0xf24a01a73cf2dccf, 0xbc633b39673c8ced},
+      {0x976e41088617ca01, 0xd5be0503e085d814},
+      {0xbd49d14aa79dbc82, 0x4b2d8644d8a74e19},
+      {0xec9c459d51852ba2, 0xddf8e7d60ed1219f},
+      {0x93e1ab8252f33b45, 0xcabb90e5c942b504},
+      {0xb8da1662e7b00a17, 0x3d6a751f3b936244},
+      {0xe7109bfba19c0c9d, 0x0cc512670a783ad5},
+      {0x906a617d450187e2, 0x27fb2b80668b24c6},
+      {0xb484f9dc9641e9da, 0xb1f9f660802dedf7},
+      {0xe1a63853bbd26451, 0x5e7873f8a0396974},
+      {0x8d07e33455637eb2, 0xdb0b487b6423e1e9},
+      {0xb049dc016abc5e5f, 0x91ce1a9a3d2cda63},
+      {0xdc5c5301c56b75f7, 0x7641a140cc7810fc},
+      {0x89b9b3e11b6329ba, 0xa9e904c87fcb0a9e},
+      {0xac2820d9623bf429, 0x546345fa9fbdcd45},
+      {0xd732290fbacaf133, 0xa97c177947ad4096},
+      {0x867f59a9d4bed6c0, 0x49ed8eabcccc485e},
+      {0xa81f301449ee8c70, 0x5c68f256bfff5a75},
+      {0xd226fc195c6a2f8c, 0x73832eec6fff3112},
+      {0x83585d8fd9c25db7, 0xc831fd53c5ff7eac},
+      {0xa42e74f3d032f525, 0xba3e7ca8b77f5e56},
+      {0xcd3a1230c43fb26f, 0x28ce1bd2e55f35ec},
+      {0x80444b5e7aa7cf85, 0x7980d163cf5b81b4},
+      {0xa0555e361951c366, 0xd7e105bcc3326220},
+      {0xc86ab5c39fa63440, 0x8dd9472bf3fefaa8},
+      {0xfa856334878fc150, 0xb14f98f6f0feb952},
+      {0x9c935e00d4b9d8d2, 0x6ed1bf9a569f33d4},
+      {0xc3b8358109e84f07, 0x0a862f80ec4700c9},
+      {0xf4a642e14c6262c8, 0xcd27bb612758c0fb},
+      {0x98e7e9cccfbd7dbd, 0x8038d51cb897789d},
+      {0xbf21e44003acdd2c, 0xe0470a63e6bd56c4},
+      {0xeeea5d5004981478, 0x1858ccfce06cac75},
+      {0x95527a5202df0ccb, 0x0f37801e0c43ebc9},
+      {0xbaa718e68396cffd, 0xd30560258f54e6bb},
+      {0xe950df20247c83fd, 0x47c6b82ef32a206a},
+      {0x91d28b7416cdd27e, 0x4cdc331d57fa5442},
+      {0xb6472e511c81471d, 0xe0133fe4adf8e953},
+      {0xe3d8f9e563a198e5, 0x58180fddd97723a7},
+      {0x8e679c2f5e44ff8f, 0x570f09eaa7ea7649},
+      {0xb201833b35d63f73, 0x2cd2cc6551e513db},
+      {0xde81e40a034bcf4f, 0xf8077f7ea65e58d2},
+      {0x8b112e86420f6191, 0xfb04afaf27faf783},
+      {0xadd57a27d29339f6, 0x79c5db9af1f9b564},
+      {0xd94ad8b1c7380874, 0x18375281ae7822bd},
+      {0x87cec76f1c830548, 0x8f2293910d0b15b6},
+      {0xa9c2794ae3a3c69a, 0xb2eb3875504ddb23},
+      {0xd433179d9c8cb841, 0x5fa60692a46151ec},
+      {0x849feec281d7f328, 0xdbc7c41ba6bcd334},
+      {0xa5c7ea73224deff3, 0x12b9b522906c0801},
+      {0xcf39e50feae16bef, 0xd768226b34870a01},
+      {0x81842f29f2cce375, 0xe6a1158300d46641},
+      {0xa1e53af46f801c53, 0x60495ae3c1097fd1},
+      {0xca5e89b18b602368, 0x385bb19cb14bdfc5},
+      {0xfcf62c1dee382c42, 0x46729e03dd9ed7b6},
+      {0x9e19db92b4e31ba9, 0x6c07a2c26a8346d2},
+      {0xc5a05277621be293, 0xc7098b7305241886},
+      {0xf70867153aa2db38, 0xb8cbee4fc66d1ea8},
+      {0x9a65406d44a5c903, 0x737f74f1dc043329},
+      {0xc0fe908895cf3b44, 0x505f522e53053ff3},
+      {0xf13e34aabb430a15, 0x647726b9e7c68ff0},
+      {0x96c6e0eab509e64d, 0x5eca783430dc19f6},
+      {0xbc789925624c5fe0, 0xb67d16413d132073},
+      {0xeb96bf6ebadf77d8, 0xe41c5bd18c57e890},
+      {0x933e37a534cbaae7, 0x8e91b962f7b6f15a},
+      {0xb80dc58e81fe95a1, 0x723627bbb5a4adb1},
+      {0xe61136f2227e3b09, 0xcec3b1aaa30dd91d},
+      {0x8fcac257558ee4e6, 0x213a4f0aa5e8a7b2},
+      {0xb3bd72ed2af29e1f, 0xa988e2cd4f62d19e},
+      {0xe0accfa875af45a7, 0x93eb1b80a33b8606},
+      {0x8c6c01c9498d8b88, 0xbc72f130660533c4},
+      {0xaf87023b9bf0ee6a, 0xeb8fad7c7f8680b5},
+      {0xdb68c2ca82ed2a05, 0xa67398db9f6820e2},
+#else
+      {0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7b},
+      {0xce5d73ff402d98e3, 0xfb0a3d212dc81290},
+      {0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481f},
+      {0x86a8d39ef77164bc, 0xae5dff9c02033198},
+      {0xd98ddaee19068c76, 0x3badd624dd9b0958},
+      {0xafbd2350644eeacf, 0xe5d1929ef90898fb},
+      {0x8df5efabc5979c8f, 0xca8d3ffa1ef463c2},
+      {0xe55990879ddcaabd, 0xcc420a6a101d0516},
+      {0xb94470938fa89bce, 0xf808e40e8d5b3e6a},
+      {0x95a8637627989aad, 0xdde7001379a44aa9},
+      {0xf1c90080baf72cb1, 0x5324c68b12dd6339},
+      {0xc350000000000000, 0x0000000000000000},
+      {0x9dc5ada82b70b59d, 0xf020000000000000},
+      {0xfee50b7025c36a08, 0x02f236d04753d5b5},
+      {0xcde6fd5e09abcf26, 0xed4c0226b55e6f87},
+      {0xa6539930bf6bff45, 0x84db8346b786151d},
+      {0x865b86925b9bc5c2, 0x0b8a2392ba45a9b3},
+      {0xd910f7ff28069da4, 0x1b2ba1518094da05},
+      {0xaf58416654a6babb, 0x387ac8d1970027b3},
+      {0x8da471a9de737e24, 0x5ceaecfed289e5d3},
+      {0xe4d5e82392a40515, 0x0fabaf3feaa5334b},
+      {0xb8da1662e7b00a17, 0x3d6a751f3b936244},
+      {0x95527a5202df0ccb, 0x0f37801e0c43ebc9},
+      {0xf13e34aabb430a15, 0x647726b9e7c68ff0}
+#endif
+    };
+
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+    return pow10_significands[k - float_info<double>::min_k];
+#else
+    static constexpr const uint64_t powers_of_5_64[] = {
+        0x0000000000000001, 0x0000000000000005, 0x0000000000000019,
+        0x000000000000007d, 0x0000000000000271, 0x0000000000000c35,
+        0x0000000000003d09, 0x000000000001312d, 0x000000000005f5e1,
+        0x00000000001dcd65, 0x00000000009502f9, 0x0000000002e90edd,
+        0x000000000e8d4a51, 0x0000000048c27395, 0x000000016bcc41e9,
+        0x000000071afd498d, 0x0000002386f26fc1, 0x000000b1a2bc2ec5,
+        0x000003782dace9d9, 0x00001158e460913d, 0x000056bc75e2d631,
+        0x0001b1ae4d6e2ef5, 0x000878678326eac9, 0x002a5a058fc295ed,
+        0x00d3c21bcecceda1, 0x0422ca8b0a00a425, 0x14adf4b7320334b9};
+
+    static const int compression_ratio = 27;
+
+    // Compute base index.
+    int cache_index = (k - float_info<double>::min_k) / compression_ratio;
+    int kb = cache_index * compression_ratio + float_info<double>::min_k;
+    int offset = k - kb;
+
+    // Get base cache.
+    uint128_fallback base_cache = pow10_significands[cache_index];
+    if (offset == 0) return base_cache;
+
+    // Compute the required amount of bit-shift.
+    int alpha = floor_log2_pow10(kb + offset) - floor_log2_pow10(kb) - offset;
+    FMT_ASSERT(alpha > 0 && alpha < 64, "shifting error detected");
+
+    // Try to recover the real cache.
+    uint64_t pow5 = powers_of_5_64[offset];
+    uint128_fallback recovered_cache = umul128(base_cache.high(), pow5);
+    uint128_fallback middle_low = umul128(base_cache.low(), pow5);
+
+    recovered_cache += middle_low.high();
+
+    uint64_t high_to_middle = recovered_cache.high() << (64 - alpha);
+    uint64_t middle_to_low = recovered_cache.low() << (64 - alpha);
+
+    recovered_cache =
+        uint128_fallback{(recovered_cache.low() >> alpha) | high_to_middle,
+                         ((middle_low.low() >> alpha) | middle_to_low)};
+    FMT_ASSERT(recovered_cache.low() + 1 != 0, "");
+    return {recovered_cache.high(), recovered_cache.low() + 1};
+#endif
+  }
+
+  struct compute_mul_result {
+    carrier_uint result;
+    bool is_integer;
+  };
+  struct compute_mul_parity_result {
+    bool parity;
+    bool is_integer;
+  };
+
+  static auto compute_mul(carrier_uint u,
+                          const cache_entry_type& cache) noexcept
+      -> compute_mul_result {
+    auto r = umul192_upper128(u, cache);
+    return {r.high(), r.low() == 0};
+  }
+
+  static auto compute_delta(cache_entry_type const& cache, int beta) noexcept
+      -> uint32_t {
+    return static_cast<uint32_t>(cache.high() >> (64 - 1 - beta));
+  }
+
+  static auto compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta) noexcept
+      -> compute_mul_parity_result {
+    FMT_ASSERT(beta >= 1, "");
+    FMT_ASSERT(beta < 64, "");
+
+    auto r = umul192_lower128(two_f, cache);
+    return {((r.high() >> (64 - beta)) & 1) != 0,
+            ((r.high() << beta) | (r.low() >> (64 - beta))) == 0};
+  }
+
+  static auto compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return (cache.high() -
+            (cache.high() >> (num_significand_bits<double>() + 2))) >>
+           (64 - num_significand_bits<double>() - 1 - beta);
+  }
+
+  static auto compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return (cache.high() +
+            (cache.high() >> (num_significand_bits<double>() + 1))) >>
+           (64 - num_significand_bits<double>() - 1 - beta);
+  }
+
+  static auto compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return ((cache.high() >> (64 - num_significand_bits<double>() - 2 - beta)) +
+            1) /
+           2;
+  }
+};
+
+FMT_FUNC auto get_cached_power(int k) noexcept -> uint128_fallback {
+  return cache_accessor<double>::get_cached_power(k);
+}
+
+// Various integer checks
+template <typename T>
+auto is_left_endpoint_integer_shorter_interval(int exponent) noexcept -> bool {
+  const int case_shorter_interval_left_endpoint_lower_threshold = 2;
+  const int case_shorter_interval_left_endpoint_upper_threshold = 3;
+  return exponent >= case_shorter_interval_left_endpoint_lower_threshold &&
+         exponent <= case_shorter_interval_left_endpoint_upper_threshold;
+}
+
+// Remove trailing zeros from n and return the number of zeros removed (float)
+FMT_INLINE int remove_trailing_zeros(uint32_t& n, int s = 0) noexcept {
+  FMT_ASSERT(n != 0, "");
+  // Modular inverse of 5 (mod 2^32): (mod_inv_5 * 5) mod 2^32 = 1.
+  constexpr uint32_t mod_inv_5 = 0xcccccccd;
+  constexpr uint32_t mod_inv_25 = 0xc28f5c29;  // = mod_inv_5 * mod_inv_5
+
+  while (true) {
+    auto q = rotr(n * mod_inv_25, 2);
+    if (q > max_value<uint32_t>() / 100) break;
+    n = q;
+    s += 2;
+  }
+  auto q = rotr(n * mod_inv_5, 1);
+  if (q <= max_value<uint32_t>() / 10) {
+    n = q;
+    s |= 1;
+  }
+  return s;
+}
+
+// Removes trailing zeros and returns the number of zeros removed (double)
+FMT_INLINE int remove_trailing_zeros(uint64_t& n) noexcept {
+  FMT_ASSERT(n != 0, "");
+
+  // This magic number is ceil(2^90 / 10^8).
+  constexpr uint64_t magic_number = 12379400392853802749ull;
+  auto nm = umul128(n, magic_number);
+
+  // Is n is divisible by 10^8?
+  if ((nm.high() & ((1ull << (90 - 64)) - 1)) == 0 && nm.low() < magic_number) {
+    // If yes, work with the quotient...
+    auto n32 = static_cast<uint32_t>(nm.high() >> (90 - 64));
+    // ... and use the 32 bit variant of the function
+    int s = remove_trailing_zeros(n32, 8);
+    n = n32;
+    return s;
+  }
+
+  // If n is not divisible by 10^8, work with n itself.
+  constexpr uint64_t mod_inv_5 = 0xcccccccccccccccd;
+  constexpr uint64_t mod_inv_25 = 0x8f5c28f5c28f5c29;  // mod_inv_5 * mod_inv_5
+
+  int s = 0;
+  while (true) {
+    auto q = rotr(n * mod_inv_25, 2);
+    if (q > max_value<uint64_t>() / 100) break;
+    n = q;
+    s += 2;
+  }
+  auto q = rotr(n * mod_inv_5, 1);
+  if (q <= max_value<uint64_t>() / 10) {
+    n = q;
+    s |= 1;
+  }
+
+  return s;
+}
+
+// The main algorithm for shorter interval case
+template <typename T>
+FMT_INLINE decimal_fp<T> shorter_interval_case(int exponent) noexcept {
+  decimal_fp<T> ret_value;
+  // Compute k and beta
+  const int minus_k = floor_log10_pow2_minus_log10_4_over_3(exponent);
+  const int beta = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute xi and zi
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+
+  auto xi = cache_accessor<T>::compute_left_endpoint_for_shorter_interval_case(
+      cache, beta);
+  auto zi = cache_accessor<T>::compute_right_endpoint_for_shorter_interval_case(
+      cache, beta);
+
+  // If the left endpoint is not an integer, increase it
+  if (!is_left_endpoint_integer_shorter_interval<T>(exponent)) ++xi;
+
+  // Try bigger divisor
+  ret_value.significand = zi / 10;
+
+  // If succeed, remove trailing zeros if necessary and return
+  if (ret_value.significand * 10 >= xi) {
+    ret_value.exponent = minus_k + 1;
+    ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+    return ret_value;
+  }
+
+  // Otherwise, compute the round-up of y
+  ret_value.significand =
+      cache_accessor<T>::compute_round_up_for_shorter_interval_case(cache,
+                                                                    beta);
+  ret_value.exponent = minus_k;
+
+  // When tie occurs, choose one of them according to the rule
+  if (exponent >= float_info<T>::shorter_interval_tie_lower_threshold &&
+      exponent <= float_info<T>::shorter_interval_tie_upper_threshold) {
+    ret_value.significand = ret_value.significand % 2 == 0
+                                ? ret_value.significand
+                                : ret_value.significand - 1;
+  } else if (ret_value.significand < xi) {
+    ++ret_value.significand;
+  }
+  return ret_value;
+}
+
+template <typename T> auto to_decimal(T x) noexcept -> decimal_fp<T> {
+  // Step 1: integer promotion & Schubfach multiplier calculation.
+
+  using carrier_uint = typename float_info<T>::carrier_uint;
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  auto br = bit_cast<carrier_uint>(x);
+
+  // Extract significand bits and exponent bits.
+  const carrier_uint significand_mask =
+      (static_cast<carrier_uint>(1) << num_significand_bits<T>()) - 1;
+  carrier_uint significand = (br & significand_mask);
+  int exponent =
+      static_cast<int>((br & exponent_mask<T>()) >> num_significand_bits<T>());
+
+  if (exponent != 0) {  // Check if normal.
+    exponent -= exponent_bias<T>() + num_significand_bits<T>();
+
+    // Shorter interval case; proceed like Schubfach.
+    // In fact, when exponent == 1 and significand == 0, the interval is
+    // regular. However, it can be shown that the end-results are anyway same.
+    if (significand == 0) return shorter_interval_case<T>(exponent);
+
+    significand |= (static_cast<carrier_uint>(1) << num_significand_bits<T>());
+  } else {
+    // Subnormal case; the interval is always regular.
+    if (significand == 0) return {0, 0};
+    exponent =
+        std::numeric_limits<T>::min_exponent - num_significand_bits<T>() - 1;
+  }
+
+  const bool include_left_endpoint = (significand % 2 == 0);
+  const bool include_right_endpoint = include_left_endpoint;
+
+  // Compute k and beta.
+  const int minus_k = floor_log10_pow2(exponent) - float_info<T>::kappa;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+  const int beta = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute zi and deltai.
+  // 10^kappa <= deltai < 10^(kappa + 1)
+  const uint32_t deltai = cache_accessor<T>::compute_delta(cache, beta);
+  const carrier_uint two_fc = significand << 1;
+
+  // For the case of binary32, the result of integer check is not correct for
+  // 29711844 * 2^-82
+  // = 6.1442653300000000008655037797566933477355632930994033813476... * 10^-18
+  // and 29711844 * 2^-81
+  // = 1.2288530660000000001731007559513386695471126586198806762695... * 10^-17,
+  // and they are the unique counterexamples. However, since 29711844 is even,
+  // this does not cause any problem for the endpoints calculations; it can only
+  // cause a problem when we need to perform integer check for the center.
+  // Fortunately, with these inputs, that branch is never executed, so we are
+  // fine.
+  const typename cache_accessor<T>::compute_mul_result z_mul =
+      cache_accessor<T>::compute_mul((two_fc | 1) << beta, cache);
+
+  // Step 2: Try larger divisor; remove trailing zeros if necessary.
+
+  // Using an upper bound on zi, we might be able to optimize the division
+  // better than the compiler; we are computing zi / big_divisor here.
+  decimal_fp<T> ret_value;
+  ret_value.significand = divide_by_10_to_kappa_plus_1(z_mul.result);
+  uint32_t r = static_cast<uint32_t>(z_mul.result - float_info<T>::big_divisor *
+                                                        ret_value.significand);
+
+  if (r < deltai) {
+    // Exclude the right endpoint if necessary.
+    if (r == 0 && (z_mul.is_integer & !include_right_endpoint)) {
+      --ret_value.significand;
+      r = float_info<T>::big_divisor;
+      goto small_divisor_case_label;
+    }
+  } else if (r > deltai) {
+    goto small_divisor_case_label;
+  } else {
+    // r == deltai; compare fractional parts.
+    const typename cache_accessor<T>::compute_mul_parity_result x_mul =
+        cache_accessor<T>::compute_mul_parity(two_fc - 1, cache, beta);
+
+    if (!(x_mul.parity | (x_mul.is_integer & include_left_endpoint)))
+      goto small_divisor_case_label;
+  }
+  ret_value.exponent = minus_k + float_info<T>::kappa + 1;
+
+  // We may need to remove trailing zeros.
+  ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+  return ret_value;
+
+  // Step 3: Find the significand with the smaller divisor.
+
+small_divisor_case_label:
+  ret_value.significand *= 10;
+  ret_value.exponent = minus_k + float_info<T>::kappa;
+
+  uint32_t dist = r - (deltai / 2) + (float_info<T>::small_divisor / 2);
+  const bool approx_y_parity =
+      ((dist ^ (float_info<T>::small_divisor / 2)) & 1) != 0;
+
+  // Is dist divisible by 10^kappa?
+  const bool divisible_by_small_divisor =
+      check_divisibility_and_divide_by_pow10<float_info<T>::kappa>(dist);
+
+  // Add dist / 10^kappa to the significand.
+  ret_value.significand += dist;
+
+  if (!divisible_by_small_divisor) return ret_value;
+
+  // Check z^(f) >= epsilon^(f).
+  // We have either yi == zi - epsiloni or yi == (zi - epsiloni) - 1,
+  // where yi == zi - epsiloni if and only if z^(f) >= epsilon^(f).
+  // Since there are only 2 possibilities, we only need to care about the
+  // parity. Also, zi and r should have the same parity since the divisor
+  // is an even number.
+  const auto y_mul = cache_accessor<T>::compute_mul_parity(two_fc, cache, beta);
+
+  // If z^(f) >= epsilon^(f), we might have a tie when z^(f) == epsilon^(f),
+  // or equivalently, when y is an integer.
+  if (y_mul.parity != approx_y_parity)
+    --ret_value.significand;
+  else if (y_mul.is_integer & (ret_value.significand % 2 != 0))
+    --ret_value.significand;
+  return ret_value;
+}
+}  // namespace dragonbox
+}  // namespace detail
+
+template <> struct formatter<detail::bigint> {
+  FMT_CONSTEXPR auto parse(format_parse_context& ctx)
+      -> format_parse_context::iterator {
+    return ctx.begin();
+  }
+
+  auto format(const detail::bigint& n, format_context& ctx) const
+      -> format_context::iterator {
+    auto out = ctx.out();
+    bool first = true;
+    for (auto i = n.bigits_.size(); i > 0; --i) {
+      auto value = n.bigits_[i - 1u];
+      if (first) {
+        out = fmt::format_to(out, FMT_STRING("{:x}"), value);
+        first = false;
+        continue;
+      }
+      out = fmt::format_to(out, FMT_STRING("{:08x}"), value);
+    }
+    if (n.exp_ > 0)
+      out = fmt::format_to(out, FMT_STRING("p{}"),
+                           n.exp_ * detail::bigint::bigit_bits);
+    return out;
+  }
+};
+
+FMT_FUNC detail::utf8_to_utf16::utf8_to_utf16(string_view s) {
+  for_each_codepoint(s, [this](uint32_t cp, string_view) {
+    if (cp == invalid_code_point) FMT_THROW(std::runtime_error("invalid utf8"));
+    if (cp <= 0xFFFF) {
+      buffer_.push_back(static_cast<wchar_t>(cp));
+    } else {
+      cp -= 0x10000;
+      buffer_.push_back(static_cast<wchar_t>(0xD800 + (cp >> 10)));
+      buffer_.push_back(static_cast<wchar_t>(0xDC00 + (cp & 0x3FF)));
+    }
+    return true;
+  });
+  buffer_.push_back(0);
+}
+
+FMT_FUNC void format_system_error(detail::buffer<char>& out, int error_code,
+                                  const char* message) noexcept {
+  FMT_TRY {
+    auto ec = std::error_code(error_code, std::generic_category());
+    write(std::back_inserter(out), std::system_error(ec, message).what());
+    return;
+  }
+  FMT_CATCH(...) {}
+  format_error_code(out, error_code, message);
+}
+
+FMT_FUNC void report_system_error(int error_code,
+                                  const char* message) noexcept {
+  report_error(format_system_error, error_code, message);
+}
+
+FMT_FUNC auto vformat(string_view fmt, format_args args) -> std::string {
+  // Don't optimize the "{}" case to keep the binary size small and because it
+  // can be better optimized in fmt::format anyway.
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  return to_string(buffer);
+}
+
+namespace detail {
+#if !defined(_WIN32) || defined(FMT_WINDOWS_NO_WCHAR)
+FMT_FUNC auto write_console(int, string_view) -> bool { return false; }
+FMT_FUNC auto write_console(std::FILE*, string_view) -> bool { return false; }
+#else
+using dword = conditional_t<sizeof(long) == 4, unsigned long, unsigned>;
+extern "C" __declspec(dllimport) int __stdcall WriteConsoleW(  //
+    void*, const void*, dword, dword*, void*);
+
+FMT_FUNC bool write_console(int fd, string_view text) {
+  auto u16 = utf8_to_utf16(text);
+  return WriteConsoleW(reinterpret_cast<void*>(_get_osfhandle(fd)), u16.c_str(),
+                       static_cast<dword>(u16.size()), nullptr, nullptr) != 0;
+}
+
+FMT_FUNC auto write_console(std::FILE* f, string_view text) -> bool {
+  return write_console(_fileno(f), text);
+}
+#endif
+
+#ifdef _WIN32
+// Print assuming legacy (non-Unicode) encoding.
+FMT_FUNC void vprint_mojibake(std::FILE* f, string_view fmt, format_args args) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  fwrite_fully(buffer.data(), buffer.size(), f);
+}
+#endif
+
+FMT_FUNC void print(std::FILE* f, string_view text) {
+#ifdef _WIN32
+  int fd = _fileno(f);
+  if (_isatty(fd)) {
+    std::fflush(f);
+    if (write_console(fd, text)) return;
+  }
+#endif
+  fwrite_fully(text.data(), text.size(), f);
+}
+}  // namespace detail
+
+FMT_FUNC void vprint(std::FILE* f, string_view fmt, format_args args) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  detail::print(f, {buffer.data(), buffer.size()});
+}
+
+FMT_FUNC void vprint(string_view fmt, format_args args) {
+  vprint(stdout, fmt, args);
+}
+
+namespace detail {
+
+struct singleton {
+  unsigned char upper;
+  unsigned char lower_count;
+};
+
+inline auto is_printable(uint16_t x, const singleton* singletons,
+                         size_t singletons_size,
+                         const unsigned char* singleton_lowers,
+                         const unsigned char* normal, size_t normal_size)
+    -> bool {
+  auto upper = x >> 8;
+  auto lower_start = 0;
+  for (size_t i = 0; i < singletons_size; ++i) {
+    auto s = singletons[i];
+    auto lower_end = lower_start + s.lower_count;
+    if (upper < s.upper) break;
+    if (upper == s.upper) {
+      for (auto j = lower_start; j < lower_end; ++j) {
+        if (singleton_lowers[j] == (x & 0xff)) return false;
+      }
+    }
+    lower_start = lower_end;
+  }
+
+  auto xsigned = static_cast<int>(x);
+  auto current = true;
+  for (size_t i = 0; i < normal_size; ++i) {
+    auto v = static_cast<int>(normal[i]);
+    auto len = (v & 0x80) != 0 ? (v & 0x7f) << 8 | normal[++i] : v;
+    xsigned -= len;
+    if (xsigned < 0) break;
+    current = !current;
+  }
+  return current;
+}
+
+// This code is generated by support/printable.py.
+FMT_FUNC auto is_printable(uint32_t cp) -> bool {
+  static constexpr singleton singletons0[] = {
+      {0x00, 1},  {0x03, 5},  {0x05, 6},  {0x06, 3},  {0x07, 6},  {0x08, 8},
+      {0x09, 17}, {0x0a, 28}, {0x0b, 25}, {0x0c, 20}, {0x0d, 16}, {0x0e, 13},
+      {0x0f, 4},  {0x10, 3},  {0x12, 18}, {0x13, 9},  {0x16, 1},  {0x17, 5},
+      {0x18, 2},  {0x19, 3},  {0x1a, 7},  {0x1c, 2},  {0x1d, 1},  {0x1f, 22},
+      {0x20, 3},  {0x2b, 3},  {0x2c, 2},  {0x2d, 11}, {0x2e, 1},  {0x30, 3},
+      {0x31, 2},  {0x32, 1},  {0xa7, 2},  {0xa9, 2},  {0xaa, 4},  {0xab, 8},
+      {0xfa, 2},  {0xfb, 5},  {0xfd, 4},  {0xfe, 3},  {0xff, 9},
+  };
+  static constexpr unsigned char singletons0_lower[] = {
+      0xad, 0x78, 0x79, 0x8b, 0x8d, 0xa2, 0x30, 0x57, 0x58, 0x8b, 0x8c, 0x90,
+      0x1c, 0x1d, 0xdd, 0x0e, 0x0f, 0x4b, 0x4c, 0xfb, 0xfc, 0x2e, 0x2f, 0x3f,
+      0x5c, 0x5d, 0x5f, 0xb5, 0xe2, 0x84, 0x8d, 0x8e, 0x91, 0x92, 0xa9, 0xb1,
+      0xba, 0xbb, 0xc5, 0xc6, 0xc9, 0xca, 0xde, 0xe4, 0xe5, 0xff, 0x00, 0x04,
+      0x11, 0x12, 0x29, 0x31, 0x34, 0x37, 0x3a, 0x3b, 0x3d, 0x49, 0x4a, 0x5d,
+      0x84, 0x8e, 0x92, 0xa9, 0xb1, 0xb4, 0xba, 0xbb, 0xc6, 0xca, 0xce, 0xcf,
+      0xe4, 0xe5, 0x00, 0x04, 0x0d, 0x0e, 0x11, 0x12, 0x29, 0x31, 0x34, 0x3a,
+      0x3b, 0x45, 0x46, 0x49, 0x4a, 0x5e, 0x64, 0x65, 0x84, 0x91, 0x9b, 0x9d,
+      0xc9, 0xce, 0xcf, 0x0d, 0x11, 0x29, 0x45, 0x49, 0x57, 0x64, 0x65, 0x8d,
+      0x91, 0xa9, 0xb4, 0xba, 0xbb, 0xc5, 0xc9, 0xdf, 0xe4, 0xe5, 0xf0, 0x0d,
+      0x11, 0x45, 0x49, 0x64, 0x65, 0x80, 0x84, 0xb2, 0xbc, 0xbe, 0xbf, 0xd5,
+      0xd7, 0xf0, 0xf1, 0x83, 0x85, 0x8b, 0xa4, 0xa6, 0xbe, 0xbf, 0xc5, 0xc7,
+      0xce, 0xcf, 0xda, 0xdb, 0x48, 0x98, 0xbd, 0xcd, 0xc6, 0xce, 0xcf, 0x49,
+      0x4e, 0x4f, 0x57, 0x59, 0x5e, 0x5f, 0x89, 0x8e, 0x8f, 0xb1, 0xb6, 0xb7,
+      0xbf, 0xc1, 0xc6, 0xc7, 0xd7, 0x11, 0x16, 0x17, 0x5b, 0x5c, 0xf6, 0xf7,
+      0xfe, 0xff, 0x80, 0x0d, 0x6d, 0x71, 0xde, 0xdf, 0x0e, 0x0f, 0x1f, 0x6e,
+      0x6f, 0x1c, 0x1d, 0x5f, 0x7d, 0x7e, 0xae, 0xaf, 0xbb, 0xbc, 0xfa, 0x16,
+      0x17, 0x1e, 0x1f, 0x46, 0x47, 0x4e, 0x4f, 0x58, 0x5a, 0x5c, 0x5e, 0x7e,
+      0x7f, 0xb5, 0xc5, 0xd4, 0xd5, 0xdc, 0xf0, 0xf1, 0xf5, 0x72, 0x73, 0x8f,
+      0x74, 0x75, 0x96, 0x2f, 0x5f, 0x26, 0x2e, 0x2f, 0xa7, 0xaf, 0xb7, 0xbf,
+      0xc7, 0xcf, 0xd7, 0xdf, 0x9a, 0x40, 0x97, 0x98, 0x30, 0x8f, 0x1f, 0xc0,
+      0xc1, 0xce, 0xff, 0x4e, 0x4f, 0x5a, 0x5b, 0x07, 0x08, 0x0f, 0x10, 0x27,
+      0x2f, 0xee, 0xef, 0x6e, 0x6f, 0x37, 0x3d, 0x3f, 0x42, 0x45, 0x90, 0x91,
+      0xfe, 0xff, 0x53, 0x67, 0x75, 0xc8, 0xc9, 0xd0, 0xd1, 0xd8, 0xd9, 0xe7,
+      0xfe, 0xff,
+  };
+  static constexpr singleton singletons1[] = {
+      {0x00, 6},  {0x01, 1}, {0x03, 1},  {0x04, 2}, {0x08, 8},  {0x09, 2},
+      {0x0a, 5},  {0x0b, 2}, {0x0e, 4},  {0x10, 1}, {0x11, 2},  {0x12, 5},
+      {0x13, 17}, {0x14, 1}, {0x15, 2},  {0x17, 2}, {0x19, 13}, {0x1c, 5},
+      {0x1d, 8},  {0x24, 1}, {0x6a, 3},  {0x6b, 2}, {0xbc, 2},  {0xd1, 2},
+      {0xd4, 12}, {0xd5, 9}, {0xd6, 2},  {0xd7, 2}, {0xda, 1},  {0xe0, 5},
+      {0xe1, 2},  {0xe8, 2}, {0xee, 32}, {0xf0, 4}, {0xf8, 2},  {0xf9, 2},
+      {0xfa, 2},  {0xfb, 1},
+  };
+  static constexpr unsigned char singletons1_lower[] = {
+      0x0c, 0x27, 0x3b, 0x3e, 0x4e, 0x4f, 0x8f, 0x9e, 0x9e, 0x9f, 0x06, 0x07,
+      0x09, 0x36, 0x3d, 0x3e, 0x56, 0xf3, 0xd0, 0xd1, 0x04, 0x14, 0x18, 0x36,
+      0x37, 0x56, 0x57, 0x7f, 0xaa, 0xae, 0xaf, 0xbd, 0x35, 0xe0, 0x12, 0x87,
+      0x89, 0x8e, 0x9e, 0x04, 0x0d, 0x0e, 0x11, 0x12, 0x29, 0x31, 0x34, 0x3a,
+      0x45, 0x46, 0x49, 0x4a, 0x4e, 0x4f, 0x64, 0x65, 0x5c, 0xb6, 0xb7, 0x1b,
+      0x1c, 0x07, 0x08, 0x0a, 0x0b, 0x14, 0x17, 0x36, 0x39, 0x3a, 0xa8, 0xa9,
+      0xd8, 0xd9, 0x09, 0x37, 0x90, 0x91, 0xa8, 0x07, 0x0a, 0x3b, 0x3e, 0x66,
+      0x69, 0x8f, 0x92, 0x6f, 0x5f, 0xee, 0xef, 0x5a, 0x62, 0x9a, 0x9b, 0x27,
+      0x28, 0x55, 0x9d, 0xa0, 0xa1, 0xa3, 0xa4, 0xa7, 0xa8, 0xad, 0xba, 0xbc,
+      0xc4, 0x06, 0x0b, 0x0c, 0x15, 0x1d, 0x3a, 0x3f, 0x45, 0x51, 0xa6, 0xa7,
+      0xcc, 0xcd, 0xa0, 0x07, 0x19, 0x1a, 0x22, 0x25, 0x3e, 0x3f, 0xc5, 0xc6,
+      0x04, 0x20, 0x23, 0x25, 0x26, 0x28, 0x33, 0x38, 0x3a, 0x48, 0x4a, 0x4c,
+      0x50, 0x53, 0x55, 0x56, 0x58, 0x5a, 0x5c, 0x5e, 0x60, 0x63, 0x65, 0x66,
+      0x6b, 0x73, 0x78, 0x7d, 0x7f, 0x8a, 0xa4, 0xaa, 0xaf, 0xb0, 0xc0, 0xd0,
+      0xae, 0xaf, 0x79, 0xcc, 0x6e, 0x6f, 0x93,
+  };
+  static constexpr unsigned char normal0[] = {
+      0x00, 0x20, 0x5f, 0x22, 0x82, 0xdf, 0x04, 0x82, 0x44, 0x08, 0x1b, 0x04,
+      0x06, 0x11, 0x81, 0xac, 0x0e, 0x80, 0xab, 0x35, 0x28, 0x0b, 0x80, 0xe0,
+      0x03, 0x19, 0x08, 0x01, 0x04, 0x2f, 0x04, 0x34, 0x04, 0x07, 0x03, 0x01,
+      0x07, 0x06, 0x07, 0x11, 0x0a, 0x50, 0x0f, 0x12, 0x07, 0x55, 0x07, 0x03,
+      0x04, 0x1c, 0x0a, 0x09, 0x03, 0x08, 0x03, 0x07, 0x03, 0x02, 0x03, 0x03,
+      0x03, 0x0c, 0x04, 0x05, 0x03, 0x0b, 0x06, 0x01, 0x0e, 0x15, 0x05, 0x3a,
+      0x03, 0x11, 0x07, 0x06, 0x05, 0x10, 0x07, 0x57, 0x07, 0x02, 0x07, 0x15,
+      0x0d, 0x50, 0x04, 0x43, 0x03, 0x2d, 0x03, 0x01, 0x04, 0x11, 0x06, 0x0f,
+      0x0c, 0x3a, 0x04, 0x1d, 0x25, 0x5f, 0x20, 0x6d, 0x04, 0x6a, 0x25, 0x80,
+      0xc8, 0x05, 0x82, 0xb0, 0x03, 0x1a, 0x06, 0x82, 0xfd, 0x03, 0x59, 0x07,
+      0x15, 0x0b, 0x17, 0x09, 0x14, 0x0c, 0x14, 0x0c, 0x6a, 0x06, 0x0a, 0x06,
+      0x1a, 0x06, 0x59, 0x07, 0x2b, 0x05, 0x46, 0x0a, 0x2c, 0x04, 0x0c, 0x04,
+      0x01, 0x03, 0x31, 0x0b, 0x2c, 0x04, 0x1a, 0x06, 0x0b, 0x03, 0x80, 0xac,
+      0x06, 0x0a, 0x06, 0x21, 0x3f, 0x4c, 0x04, 0x2d, 0x03, 0x74, 0x08, 0x3c,
+      0x03, 0x0f, 0x03, 0x3c, 0x07, 0x38, 0x08, 0x2b, 0x05, 0x82, 0xff, 0x11,
+      0x18, 0x08, 0x2f, 0x11, 0x2d, 0x03, 0x20, 0x10, 0x21, 0x0f, 0x80, 0x8c,
+      0x04, 0x82, 0x97, 0x19, 0x0b, 0x15, 0x88, 0x94, 0x05, 0x2f, 0x05, 0x3b,
+      0x07, 0x02, 0x0e, 0x18, 0x09, 0x80, 0xb3, 0x2d, 0x74, 0x0c, 0x80, 0xd6,
+      0x1a, 0x0c, 0x05, 0x80, 0xff, 0x05, 0x80, 0xdf, 0x0c, 0xee, 0x0d, 0x03,
+      0x84, 0x8d, 0x03, 0x37, 0x09, 0x81, 0x5c, 0x14, 0x80, 0xb8, 0x08, 0x80,
+      0xcb, 0x2a, 0x38, 0x03, 0x0a, 0x06, 0x38, 0x08, 0x46, 0x08, 0x0c, 0x06,
+      0x74, 0x0b, 0x1e, 0x03, 0x5a, 0x04, 0x59, 0x09, 0x80, 0x83, 0x18, 0x1c,
+      0x0a, 0x16, 0x09, 0x4c, 0x04, 0x80, 0x8a, 0x06, 0xab, 0xa4, 0x0c, 0x17,
+      0x04, 0x31, 0xa1, 0x04, 0x81, 0xda, 0x26, 0x07, 0x0c, 0x05, 0x05, 0x80,
+      0xa5, 0x11, 0x81, 0x6d, 0x10, 0x78, 0x28, 0x2a, 0x06, 0x4c, 0x04, 0x80,
+      0x8d, 0x04, 0x80, 0xbe, 0x03, 0x1b, 0x03, 0x0f, 0x0d,
+  };
+  static constexpr unsigned char normal1[] = {
+      0x5e, 0x22, 0x7b, 0x05, 0x03, 0x04, 0x2d, 0x03, 0x66, 0x03, 0x01, 0x2f,
+      0x2e, 0x80, 0x82, 0x1d, 0x03, 0x31, 0x0f, 0x1c, 0x04, 0x24, 0x09, 0x1e,
+      0x05, 0x2b, 0x05, 0x44, 0x04, 0x0e, 0x2a, 0x80, 0xaa, 0x06, 0x24, 0x04,
+      0x24, 0x04, 0x28, 0x08, 0x34, 0x0b, 0x01, 0x80, 0x90, 0x81, 0x37, 0x09,
+      0x16, 0x0a, 0x08, 0x80, 0x98, 0x39, 0x03, 0x63, 0x08, 0x09, 0x30, 0x16,
+      0x05, 0x21, 0x03, 0x1b, 0x05, 0x01, 0x40, 0x38, 0x04, 0x4b, 0x05, 0x2f,
+      0x04, 0x0a, 0x07, 0x09, 0x07, 0x40, 0x20, 0x27, 0x04, 0x0c, 0x09, 0x36,
+      0x03, 0x3a, 0x05, 0x1a, 0x07, 0x04, 0x0c, 0x07, 0x50, 0x49, 0x37, 0x33,
+      0x0d, 0x33, 0x07, 0x2e, 0x08, 0x0a, 0x81, 0x26, 0x52, 0x4e, 0x28, 0x08,
+      0x2a, 0x56, 0x1c, 0x14, 0x17, 0x09, 0x4e, 0x04, 0x1e, 0x0f, 0x43, 0x0e,
+      0x19, 0x07, 0x0a, 0x06, 0x48, 0x08, 0x27, 0x09, 0x75, 0x0b, 0x3f, 0x41,
+      0x2a, 0x06, 0x3b, 0x05, 0x0a, 0x06, 0x51, 0x06, 0x01, 0x05, 0x10, 0x03,
+      0x05, 0x80, 0x8b, 0x62, 0x1e, 0x48, 0x08, 0x0a, 0x80, 0xa6, 0x5e, 0x22,
+      0x45, 0x0b, 0x0a, 0x06, 0x0d, 0x13, 0x39, 0x07, 0x0a, 0x36, 0x2c, 0x04,
+      0x10, 0x80, 0xc0, 0x3c, 0x64, 0x53, 0x0c, 0x48, 0x09, 0x0a, 0x46, 0x45,
+      0x1b, 0x48, 0x08, 0x53, 0x1d, 0x39, 0x81, 0x07, 0x46, 0x0a, 0x1d, 0x03,
+      0x47, 0x49, 0x37, 0x03, 0x0e, 0x08, 0x0a, 0x06, 0x39, 0x07, 0x0a, 0x81,
+      0x36, 0x19, 0x80, 0xb7, 0x01, 0x0f, 0x32, 0x0d, 0x83, 0x9b, 0x66, 0x75,
+      0x0b, 0x80, 0xc4, 0x8a, 0xbc, 0x84, 0x2f, 0x8f, 0xd1, 0x82, 0x47, 0xa1,
+      0xb9, 0x82, 0x39, 0x07, 0x2a, 0x04, 0x02, 0x60, 0x26, 0x0a, 0x46, 0x0a,
+      0x28, 0x05, 0x13, 0x82, 0xb0, 0x5b, 0x65, 0x4b, 0x04, 0x39, 0x07, 0x11,
+      0x40, 0x05, 0x0b, 0x02, 0x0e, 0x97, 0xf8, 0x08, 0x84, 0xd6, 0x2a, 0x09,
+      0xa2, 0xf7, 0x81, 0x1f, 0x31, 0x03, 0x11, 0x04, 0x08, 0x81, 0x8c, 0x89,
+      0x04, 0x6b, 0x05, 0x0d, 0x03, 0x09, 0x07, 0x10, 0x93, 0x60, 0x80, 0xf6,
+      0x0a, 0x73, 0x08, 0x6e, 0x17, 0x46, 0x80, 0x9a, 0x14, 0x0c, 0x57, 0x09,
+      0x19, 0x80, 0x87, 0x81, 0x47, 0x03, 0x85, 0x42, 0x0f, 0x15, 0x85, 0x50,
+      0x2b, 0x80, 0xd5, 0x2d, 0x03, 0x1a, 0x04, 0x02, 0x81, 0x70, 0x3a, 0x05,
+      0x01, 0x85, 0x00, 0x80, 0xd7, 0x29, 0x4c, 0x04, 0x0a, 0x04, 0x02, 0x83,
+      0x11, 0x44, 0x4c, 0x3d, 0x80, 0xc2, 0x3c, 0x06, 0x01, 0x04, 0x55, 0x05,
+      0x1b, 0x34, 0x02, 0x81, 0x0e, 0x2c, 0x04, 0x64, 0x0c, 0x56, 0x0a, 0x80,
+      0xae, 0x38, 0x1d, 0x0d, 0x2c, 0x04, 0x09, 0x07, 0x02, 0x0e, 0x06, 0x80,
+      0x9a, 0x83, 0xd8, 0x08, 0x0d, 0x03, 0x0d, 0x03, 0x74, 0x0c, 0x59, 0x07,
+      0x0c, 0x14, 0x0c, 0x04, 0x38, 0x08, 0x0a, 0x06, 0x28, 0x08, 0x22, 0x4e,
+      0x81, 0x54, 0x0c, 0x15, 0x03, 0x03, 0x05, 0x07, 0x09, 0x19, 0x07, 0x07,
+      0x09, 0x03, 0x0d, 0x07, 0x29, 0x80, 0xcb, 0x25, 0x0a, 0x84, 0x06,
+  };
+  auto lower = static_cast<uint16_t>(cp);
+  if (cp < 0x10000) {
+    return is_printable(lower, singletons0,
+                        sizeof(singletons0) / sizeof(*singletons0),
+                        singletons0_lower, normal0, sizeof(normal0));
+  }
+  if (cp < 0x20000) {
+    return is_printable(lower, singletons1,
+                        sizeof(singletons1) / sizeof(*singletons1),
+                        singletons1_lower, normal1, sizeof(normal1));
+  }
+  if (0x2a6de <= cp && cp < 0x2a700) return false;
+  if (0x2b735 <= cp && cp < 0x2b740) return false;
+  if (0x2b81e <= cp && cp < 0x2b820) return false;
+  if (0x2cea2 <= cp && cp < 0x2ceb0) return false;
+  if (0x2ebe1 <= cp && cp < 0x2f800) return false;
+  if (0x2fa1e <= cp && cp < 0x30000) return false;
+  if (0x3134b <= cp && cp < 0xe0100) return false;
+  if (0xe01f0 <= cp && cp < 0x110000) return false;
+  return cp < 0x110000;
+}
+
+}  // namespace detail
+
+FMT_END_NAMESPACE
+
+#endif  // FMT_FORMAT_INL_H_
diff --git a/dep/fmt/include/fmt/format.h b/dep/fmt/include/fmt/format.h
new file mode 100644
index 00000000..7637c8a0
--- /dev/null
+++ b/dep/fmt/include/fmt/format.h
@@ -0,0 +1,4535 @@
+/*
+  Formatting library for C++
+
+  Copyright (c) 2012 - present, Victor Zverovich
+
+  Permission is hereby granted, free of charge, to any person obtaining
+  a copy of this software and associated documentation files (the
+  "Software"), to deal in the Software without restriction, including
+  without limitation the rights to use, copy, modify, merge, publish,
+  distribute, sublicense, and/or sell copies of the Software, and to
+  permit persons to whom the Software is furnished to do so, subject to
+  the following conditions:
+
+  The above copyright notice and this permission notice shall be
+  included in all copies or substantial portions of the Software.
+
+  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+  NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+  --- Optional exception to the license ---
+
+  As an exception, if, as a result of your compiling your source code, portions
+  of this Software are embedded into a machine-executable object form of such
+  source code, you may redistribute such embedded portions in such object form
+  without including the above copyright and permission notices.
+ */
+
+#ifndef FMT_FORMAT_H_
+#define FMT_FORMAT_H_
+
+#include <cmath>             // std::signbit
+#include <cstdint>           // uint32_t
+#include <cstring>           // std::memcpy
+#include <initializer_list>  // std::initializer_list
+#include <limits>            // std::numeric_limits
+#include <memory>            // std::uninitialized_copy
+#include <stdexcept>         // std::runtime_error
+#include <system_error>      // std::system_error
+
+#ifdef __cpp_lib_bit_cast
+#  include <bit>  // std::bit_cast
+#endif
+
+#include "core.h"
+
+#if defined __cpp_inline_variables && __cpp_inline_variables >= 201606L
+#  define FMT_INLINE_VARIABLE inline
+#else
+#  define FMT_INLINE_VARIABLE
+#endif
+
+#if FMT_HAS_CPP17_ATTRIBUTE(fallthrough)
+#  define FMT_FALLTHROUGH [[fallthrough]]
+#elif defined(__clang__)
+#  define FMT_FALLTHROUGH [[clang::fallthrough]]
+#elif FMT_GCC_VERSION >= 700 && \
+    (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
+#  define FMT_FALLTHROUGH [[gnu::fallthrough]]
+#else
+#  define FMT_FALLTHROUGH
+#endif
+
+#ifndef FMT_DEPRECATED
+#  if FMT_HAS_CPP14_ATTRIBUTE(deprecated) || FMT_MSC_VERSION >= 1900
+#    define FMT_DEPRECATED [[deprecated]]
+#  else
+#    if (defined(__GNUC__) && !defined(__LCC__)) || defined(__clang__)
+#      define FMT_DEPRECATED __attribute__((deprecated))
+#    elif FMT_MSC_VERSION
+#      define FMT_DEPRECATED __declspec(deprecated)
+#    else
+#      define FMT_DEPRECATED /* deprecated */
+#    endif
+#  endif
+#endif
+
+#ifndef FMT_NO_UNIQUE_ADDRESS
+#  if FMT_CPLUSPLUS >= 202002L
+#    if FMT_HAS_CPP_ATTRIBUTE(no_unique_address)
+#      define FMT_NO_UNIQUE_ADDRESS [[no_unique_address]]
+// VS2019 v16.10 and later except clang-cl (https://reviews.llvm.org/D110485)
+#    elif (FMT_MSC_VERSION >= 1929) && !FMT_CLANG_VERSION
+#      define FMT_NO_UNIQUE_ADDRESS [[msvc::no_unique_address]]
+#    endif
+#  endif
+#endif
+#ifndef FMT_NO_UNIQUE_ADDRESS
+#  define FMT_NO_UNIQUE_ADDRESS
+#endif
+
+// Visibility when compiled as a shared library/object.
+#if defined(FMT_LIB_EXPORT) || defined(FMT_SHARED)
+#  define FMT_SO_VISIBILITY(value) FMT_VISIBILITY(value)
+#else
+#  define FMT_SO_VISIBILITY(value)
+#endif
+
+#ifdef __has_builtin
+#  define FMT_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define FMT_HAS_BUILTIN(x) 0
+#endif
+
+#if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#  define FMT_NOINLINE __attribute__((noinline))
+#else
+#  define FMT_NOINLINE
+#endif
+
+#ifndef FMT_THROW
+#  if FMT_EXCEPTIONS
+#    if FMT_MSC_VERSION || defined(__NVCC__)
+FMT_BEGIN_NAMESPACE
+namespace detail {
+template <typename Exception> inline void do_throw(const Exception& x) {
+  // Silence unreachable code warnings in MSVC and NVCC because these
+  // are nearly impossible to fix in a generic code.
+  volatile bool b = true;
+  if (b) throw x;
+}
+}  // namespace detail
+FMT_END_NAMESPACE
+#      define FMT_THROW(x) detail::do_throw(x)
+#    else
+#      define FMT_THROW(x) throw x
+#    endif
+#  else
+#    define FMT_THROW(x) \
+      ::fmt::detail::assert_fail(__FILE__, __LINE__, (x).what())
+#  endif
+#endif
+
+#if FMT_EXCEPTIONS
+#  define FMT_TRY try
+#  define FMT_CATCH(x) catch (x)
+#else
+#  define FMT_TRY if (true)
+#  define FMT_CATCH(x) if (false)
+#endif
+
+#ifndef FMT_MAYBE_UNUSED
+#  if FMT_HAS_CPP17_ATTRIBUTE(maybe_unused)
+#    define FMT_MAYBE_UNUSED [[maybe_unused]]
+#  else
+#    define FMT_MAYBE_UNUSED
+#  endif
+#endif
+
+#ifndef FMT_USE_USER_DEFINED_LITERALS
+// EDG based compilers (Intel, NVIDIA, Elbrus, etc), GCC and MSVC support UDLs.
+//
+// GCC before 4.9 requires a space in `operator"" _a` which is invalid in later
+// compiler versions.
+#  if (FMT_HAS_FEATURE(cxx_user_literals) || FMT_GCC_VERSION >= 409 || \
+       FMT_MSC_VERSION >= 1900) &&                                     \
+      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= /* UDL feature */ 480)
+#    define FMT_USE_USER_DEFINED_LITERALS 1
+#  else
+#    define FMT_USE_USER_DEFINED_LITERALS 0
+#  endif
+#endif
+
+// Defining FMT_REDUCE_INT_INSTANTIATIONS to 1, will reduce the number of
+// integer formatter template instantiations to just one by only using the
+// largest integer type. This results in a reduction in binary size but will
+// cause a decrease in integer formatting performance.
+#if !defined(FMT_REDUCE_INT_INSTANTIATIONS)
+#  define FMT_REDUCE_INT_INSTANTIATIONS 0
+#endif
+
+// __builtin_clz is broken in clang with Microsoft CodeGen:
+// https://github.com/fmtlib/fmt/issues/519.
+#if !FMT_MSC_VERSION
+#  if FMT_HAS_BUILTIN(__builtin_clz) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CLZ(n) __builtin_clz(n)
+#  endif
+#  if FMT_HAS_BUILTIN(__builtin_clzll) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CLZLL(n) __builtin_clzll(n)
+#  endif
+#endif
+
+// __builtin_ctz is broken in Intel Compiler Classic on Windows:
+// https://github.com/fmtlib/fmt/issues/2510.
+#ifndef __ICL
+#  if FMT_HAS_BUILTIN(__builtin_ctz) || FMT_GCC_VERSION || FMT_ICC_VERSION || \
+      defined(__NVCOMPILER)
+#    define FMT_BUILTIN_CTZ(n) __builtin_ctz(n)
+#  endif
+#  if FMT_HAS_BUILTIN(__builtin_ctzll) || FMT_GCC_VERSION || \
+      FMT_ICC_VERSION || defined(__NVCOMPILER)
+#    define FMT_BUILTIN_CTZLL(n) __builtin_ctzll(n)
+#  endif
+#endif
+
+#if FMT_MSC_VERSION
+#  include <intrin.h>  // _BitScanReverse[64], _BitScanForward[64], _umul128
+#endif
+
+// Some compilers masquerade as both MSVC and GCC-likes or otherwise support
+// __builtin_clz and __builtin_clzll, so only define FMT_BUILTIN_CLZ using the
+// MSVC intrinsics if the clz and clzll builtins are not available.
+#if FMT_MSC_VERSION && !defined(FMT_BUILTIN_CLZLL) && \
+    !defined(FMT_BUILTIN_CTZLL)
+FMT_BEGIN_NAMESPACE
+namespace detail {
+// Avoid Clang with Microsoft CodeGen's -Wunknown-pragmas warning.
+#  if !defined(__clang__)
+#    pragma intrinsic(_BitScanForward)
+#    pragma intrinsic(_BitScanReverse)
+#    if defined(_WIN64)
+#      pragma intrinsic(_BitScanForward64)
+#      pragma intrinsic(_BitScanReverse64)
+#    endif
+#  endif
+
+inline auto clz(uint32_t x) -> int {
+  unsigned long r = 0;
+  _BitScanReverse(&r, x);
+  FMT_ASSERT(x != 0, "");
+  // Static analysis complains about using uninitialized data
+  // "r", but the only way that can happen is if "x" is 0,
+  // which the callers guarantee to not happen.
+  FMT_MSC_WARNING(suppress : 6102)
+  return 31 ^ static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CLZ(n) detail::clz(n)
+
+inline auto clzll(uint64_t x) -> int {
+  unsigned long r = 0;
+#  ifdef _WIN64
+  _BitScanReverse64(&r, x);
+#  else
+  // Scan the high 32 bits.
+  if (_BitScanReverse(&r, static_cast<uint32_t>(x >> 32)))
+    return 63 ^ static_cast<int>(r + 32);
+  // Scan the low 32 bits.
+  _BitScanReverse(&r, static_cast<uint32_t>(x));
+#  endif
+  FMT_ASSERT(x != 0, "");
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
+  return 63 ^ static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CLZLL(n) detail::clzll(n)
+
+inline auto ctz(uint32_t x) -> int {
+  unsigned long r = 0;
+  _BitScanForward(&r, x);
+  FMT_ASSERT(x != 0, "");
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZ(n) detail::ctz(n)
+
+inline auto ctzll(uint64_t x) -> int {
+  unsigned long r = 0;
+  FMT_ASSERT(x != 0, "");
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
+#  ifdef _WIN64
+  _BitScanForward64(&r, x);
+#  else
+  // Scan the low 32 bits.
+  if (_BitScanForward(&r, static_cast<uint32_t>(x))) return static_cast<int>(r);
+  // Scan the high 32 bits.
+  _BitScanForward(&r, static_cast<uint32_t>(x >> 32));
+  r += 32;
+#  endif
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZLL(n) detail::ctzll(n)
+}  // namespace detail
+FMT_END_NAMESPACE
+#endif
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+FMT_CONSTEXPR inline void abort_fuzzing_if(bool condition) {
+  ignore_unused(condition);
+#ifdef FMT_FUZZ
+  if (condition) throw std::runtime_error("fuzzing limit reached");
+#endif
+}
+
+template <typename CharT, CharT... C> struct string_literal {
+  static constexpr CharT value[sizeof...(C)] = {C...};
+  constexpr operator basic_string_view<CharT>() const {
+    return {value, sizeof...(C)};
+  }
+};
+
+#if FMT_CPLUSPLUS < 201703L
+template <typename CharT, CharT... C>
+constexpr CharT string_literal<CharT, C...>::value[sizeof...(C)];
+#endif
+
+// Implementation of std::bit_cast for pre-C++20.
+template <typename To, typename From, FMT_ENABLE_IF(sizeof(To) == sizeof(From))>
+FMT_CONSTEXPR20 auto bit_cast(const From& from) -> To {
+#ifdef __cpp_lib_bit_cast
+  if (is_constant_evaluated()) return std::bit_cast<To>(from);
+#endif
+  auto to = To();
+  // The cast suppresses a bogus -Wclass-memaccess on GCC.
+  std::memcpy(static_cast<void*>(&to), &from, sizeof(to));
+  return to;
+}
+
+inline auto is_big_endian() -> bool {
+#ifdef _WIN32
+  return false;
+#elif defined(__BIG_ENDIAN__)
+  return true;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__)
+  return __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__;
+#else
+  struct bytes {
+    char data[sizeof(int)];
+  };
+  return bit_cast<bytes>(1).data[0] == 0;
+#endif
+}
+
+class uint128_fallback {
+ private:
+  uint64_t lo_, hi_;
+
+ public:
+  constexpr uint128_fallback(uint64_t hi, uint64_t lo) : lo_(lo), hi_(hi) {}
+  constexpr uint128_fallback(uint64_t value = 0) : lo_(value), hi_(0) {}
+
+  constexpr auto high() const noexcept -> uint64_t { return hi_; }
+  constexpr auto low() const noexcept -> uint64_t { return lo_; }
+
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  constexpr explicit operator T() const {
+    return static_cast<T>(lo_);
+  }
+
+  friend constexpr auto operator==(const uint128_fallback& lhs,
+                                   const uint128_fallback& rhs) -> bool {
+    return lhs.hi_ == rhs.hi_ && lhs.lo_ == rhs.lo_;
+  }
+  friend constexpr auto operator!=(const uint128_fallback& lhs,
+                                   const uint128_fallback& rhs) -> bool {
+    return !(lhs == rhs);
+  }
+  friend constexpr auto operator>(const uint128_fallback& lhs,
+                                  const uint128_fallback& rhs) -> bool {
+    return lhs.hi_ != rhs.hi_ ? lhs.hi_ > rhs.hi_ : lhs.lo_ > rhs.lo_;
+  }
+  friend constexpr auto operator|(const uint128_fallback& lhs,
+                                  const uint128_fallback& rhs)
+      -> uint128_fallback {
+    return {lhs.hi_ | rhs.hi_, lhs.lo_ | rhs.lo_};
+  }
+  friend constexpr auto operator&(const uint128_fallback& lhs,
+                                  const uint128_fallback& rhs)
+      -> uint128_fallback {
+    return {lhs.hi_ & rhs.hi_, lhs.lo_ & rhs.lo_};
+  }
+  friend constexpr auto operator~(const uint128_fallback& n)
+      -> uint128_fallback {
+    return {~n.hi_, ~n.lo_};
+  }
+  friend auto operator+(const uint128_fallback& lhs,
+                        const uint128_fallback& rhs) -> uint128_fallback {
+    auto result = uint128_fallback(lhs);
+    result += rhs;
+    return result;
+  }
+  friend auto operator*(const uint128_fallback& lhs, uint32_t rhs)
+      -> uint128_fallback {
+    FMT_ASSERT(lhs.hi_ == 0, "");
+    uint64_t hi = (lhs.lo_ >> 32) * rhs;
+    uint64_t lo = (lhs.lo_ & ~uint32_t()) * rhs;
+    uint64_t new_lo = (hi << 32) + lo;
+    return {(hi >> 32) + (new_lo < lo ? 1 : 0), new_lo};
+  }
+  friend auto operator-(const uint128_fallback& lhs, uint64_t rhs)
+      -> uint128_fallback {
+    return {lhs.hi_ - (lhs.lo_ < rhs ? 1 : 0), lhs.lo_ - rhs};
+  }
+  FMT_CONSTEXPR auto operator>>(int shift) const -> uint128_fallback {
+    if (shift == 64) return {0, hi_};
+    if (shift > 64) return uint128_fallback(0, hi_) >> (shift - 64);
+    return {hi_ >> shift, (hi_ << (64 - shift)) | (lo_ >> shift)};
+  }
+  FMT_CONSTEXPR auto operator<<(int shift) const -> uint128_fallback {
+    if (shift == 64) return {lo_, 0};
+    if (shift > 64) return uint128_fallback(lo_, 0) << (shift - 64);
+    return {hi_ << shift | (lo_ >> (64 - shift)), (lo_ << shift)};
+  }
+  FMT_CONSTEXPR auto operator>>=(int shift) -> uint128_fallback& {
+    return *this = *this >> shift;
+  }
+  FMT_CONSTEXPR void operator+=(uint128_fallback n) {
+    uint64_t new_lo = lo_ + n.lo_;
+    uint64_t new_hi = hi_ + n.hi_ + (new_lo < lo_ ? 1 : 0);
+    FMT_ASSERT(new_hi >= hi_, "");
+    lo_ = new_lo;
+    hi_ = new_hi;
+  }
+  FMT_CONSTEXPR void operator&=(uint128_fallback n) {
+    lo_ &= n.lo_;
+    hi_ &= n.hi_;
+  }
+
+  FMT_CONSTEXPR20 auto operator+=(uint64_t n) noexcept -> uint128_fallback& {
+    if (is_constant_evaluated()) {
+      lo_ += n;
+      hi_ += (lo_ < n ? 1 : 0);
+      return *this;
+    }
+#if FMT_HAS_BUILTIN(__builtin_addcll) && !defined(__ibmxl__)
+    unsigned long long carry;
+    lo_ = __builtin_addcll(lo_, n, 0, &carry);
+    hi_ += carry;
+#elif FMT_HAS_BUILTIN(__builtin_ia32_addcarryx_u64) && !defined(__ibmxl__)
+    unsigned long long result;
+    auto carry = __builtin_ia32_addcarryx_u64(0, lo_, n, &result);
+    lo_ = result;
+    hi_ += carry;
+#elif defined(_MSC_VER) && defined(_M_X64)
+    auto carry = _addcarry_u64(0, lo_, n, &lo_);
+    _addcarry_u64(carry, hi_, 0, &hi_);
+#else
+    lo_ += n;
+    hi_ += (lo_ < n ? 1 : 0);
+#endif
+    return *this;
+  }
+};
+
+using uint128_t = conditional_t<FMT_USE_INT128, uint128_opt, uint128_fallback>;
+
+#ifdef UINTPTR_MAX
+using uintptr_t = ::uintptr_t;
+#else
+using uintptr_t = uint128_t;
+#endif
+
+// Returns the largest possible value for type T. Same as
+// std::numeric_limits<T>::max() but shorter and not affected by the max macro.
+template <typename T> constexpr auto max_value() -> T {
+  return (std::numeric_limits<T>::max)();
+}
+template <typename T> constexpr auto num_bits() -> int {
+  return std::numeric_limits<T>::digits;
+}
+// std::numeric_limits<T>::digits may return 0 for 128-bit ints.
+template <> constexpr auto num_bits<int128_opt>() -> int { return 128; }
+template <> constexpr auto num_bits<uint128_t>() -> int { return 128; }
+
+// A heterogeneous bit_cast used for converting 96-bit long double to uint128_t
+// and 128-bit pointers to uint128_fallback.
+template <typename To, typename From, FMT_ENABLE_IF(sizeof(To) > sizeof(From))>
+inline auto bit_cast(const From& from) -> To {
+  constexpr auto size = static_cast<int>(sizeof(From) / sizeof(unsigned));
+  struct data_t {
+    unsigned value[static_cast<unsigned>(size)];
+  } data = bit_cast<data_t>(from);
+  auto result = To();
+  if (const_check(is_big_endian())) {
+    for (int i = 0; i < size; ++i)
+      result = (result << num_bits<unsigned>()) | data.value[i];
+  } else {
+    for (int i = size - 1; i >= 0; --i)
+      result = (result << num_bits<unsigned>()) | data.value[i];
+  }
+  return result;
+}
+
+template <typename UInt>
+FMT_CONSTEXPR20 inline auto countl_zero_fallback(UInt n) -> int {
+  int lz = 0;
+  constexpr UInt msb_mask = static_cast<UInt>(1) << (num_bits<UInt>() - 1);
+  for (; (n & msb_mask) == 0; n <<= 1) lz++;
+  return lz;
+}
+
+FMT_CONSTEXPR20 inline auto countl_zero(uint32_t n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated()) return FMT_BUILTIN_CLZ(n);
+#endif
+  return countl_zero_fallback(n);
+}
+
+FMT_CONSTEXPR20 inline auto countl_zero(uint64_t n) -> int {
+#ifdef FMT_BUILTIN_CLZLL
+  if (!is_constant_evaluated()) return FMT_BUILTIN_CLZLL(n);
+#endif
+  return countl_zero_fallback(n);
+}
+
+FMT_INLINE void assume(bool condition) {
+  (void)condition;
+#if FMT_HAS_BUILTIN(__builtin_assume) && !FMT_ICC_VERSION
+  __builtin_assume(condition);
+#elif FMT_GCC_VERSION
+  if (!condition) __builtin_unreachable();
+#endif
+}
+
+// An approximation of iterator_t for pre-C++20 systems.
+template <typename T>
+using iterator_t = decltype(std::begin(std::declval<T&>()));
+template <typename T> using sentinel_t = decltype(std::end(std::declval<T&>()));
+
+// A workaround for std::string not having mutable data() until C++17.
+template <typename Char>
+inline auto get_data(std::basic_string<Char>& s) -> Char* {
+  return &s[0];
+}
+template <typename Container>
+inline auto get_data(Container& c) -> typename Container::value_type* {
+  return c.data();
+}
+
+// Attempts to reserve space for n extra characters in the output range.
+// Returns a pointer to the reserved range or a reference to it.
+template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
+#if FMT_CLANG_VERSION >= 307 && !FMT_ICC_VERSION
+__attribute__((no_sanitize("undefined")))
+#endif
+inline auto
+reserve(std::back_insert_iterator<Container> it, size_t n) ->
+    typename Container::value_type* {
+  Container& c = get_container(it);
+  size_t size = c.size();
+  c.resize(size + n);
+  return get_data(c) + size;
+}
+
+template <typename T>
+inline auto reserve(buffer_appender<T> it, size_t n) -> buffer_appender<T> {
+  buffer<T>& buf = get_container(it);
+  buf.try_reserve(buf.size() + n);
+  return it;
+}
+
+template <typename Iterator>
+constexpr auto reserve(Iterator& it, size_t) -> Iterator& {
+  return it;
+}
+
+template <typename OutputIt>
+using reserve_iterator =
+    remove_reference_t<decltype(reserve(std::declval<OutputIt&>(), 0))>;
+
+template <typename T, typename OutputIt>
+constexpr auto to_pointer(OutputIt, size_t) -> T* {
+  return nullptr;
+}
+template <typename T> auto to_pointer(buffer_appender<T> it, size_t n) -> T* {
+  buffer<T>& buf = get_container(it);
+  auto size = buf.size();
+  if (buf.capacity() < size + n) return nullptr;
+  buf.try_resize(size + n);
+  return buf.data() + size;
+}
+
+template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
+inline auto base_iterator(std::back_insert_iterator<Container> it,
+                          typename Container::value_type*)
+    -> std::back_insert_iterator<Container> {
+  return it;
+}
+
+template <typename Iterator>
+constexpr auto base_iterator(Iterator, Iterator it) -> Iterator {
+  return it;
+}
+
+// <algorithm> is spectacularly slow to compile in C++20 so use a simple fill_n
+// instead (#1998).
+template <typename OutputIt, typename Size, typename T>
+FMT_CONSTEXPR auto fill_n(OutputIt out, Size count, const T& value)
+    -> OutputIt {
+  for (Size i = 0; i < count; ++i) *out++ = value;
+  return out;
+}
+template <typename T, typename Size>
+FMT_CONSTEXPR20 auto fill_n(T* out, Size count, char value) -> T* {
+  if (is_constant_evaluated()) {
+    return fill_n<T*, Size, T>(out, count, value);
+  }
+  std::memset(out, value, to_unsigned(count));
+  return out + count;
+}
+
+#ifdef __cpp_char8_t
+using char8_type = char8_t;
+#else
+enum char8_type : unsigned char {};
+#endif
+
+template <typename OutChar, typename InputIt, typename OutputIt>
+FMT_CONSTEXPR FMT_NOINLINE auto copy_str_noinline(InputIt begin, InputIt end,
+                                                  OutputIt out) -> OutputIt {
+  return copy_str<OutChar>(begin, end, out);
+}
+
+// A public domain branchless UTF-8 decoder by Christopher Wellons:
+// https://github.com/skeeto/branchless-utf8
+/* Decode the next character, c, from s, reporting errors in e.
+ *
+ * Since this is a branchless decoder, four bytes will be read from the
+ * buffer regardless of the actual length of the next character. This
+ * means the buffer _must_ have at least three bytes of zero padding
+ * following the end of the data stream.
+ *
+ * Errors are reported in e, which will be non-zero if the parsed
+ * character was somehow invalid: invalid byte sequence, non-canonical
+ * encoding, or a surrogate half.
+ *
+ * The function returns a pointer to the next character. When an error
+ * occurs, this pointer will be a guess that depends on the particular
+ * error, but it will always advance at least one byte.
+ */
+FMT_CONSTEXPR inline auto utf8_decode(const char* s, uint32_t* c, int* e)
+    -> const char* {
+  constexpr const int masks[] = {0x00, 0x7f, 0x1f, 0x0f, 0x07};
+  constexpr const uint32_t mins[] = {4194304, 0, 128, 2048, 65536};
+  constexpr const int shiftc[] = {0, 18, 12, 6, 0};
+  constexpr const int shifte[] = {0, 6, 4, 2, 0};
+
+  int len = "\1\1\1\1\1\1\1\1\1\1\1\1\1\1\1\1\0\0\0\0\0\0\0\0\2\2\2\2\3\3\4"
+      [static_cast<unsigned char>(*s) >> 3];
+  // Compute the pointer to the next character early so that the next
+  // iteration can start working on the next character. Neither Clang
+  // nor GCC figure out this reordering on their own.
+  const char* next = s + len + !len;
+
+  using uchar = unsigned char;
+
+  // Assume a four-byte character and load four bytes. Unused bits are
+  // shifted out.
+  *c = uint32_t(uchar(s[0]) & masks[len]) << 18;
+  *c |= uint32_t(uchar(s[1]) & 0x3f) << 12;
+  *c |= uint32_t(uchar(s[2]) & 0x3f) << 6;
+  *c |= uint32_t(uchar(s[3]) & 0x3f) << 0;
+  *c >>= shiftc[len];
+
+  // Accumulate the various error conditions.
+  *e = (*c < mins[len]) << 6;       // non-canonical encoding
+  *e |= ((*c >> 11) == 0x1b) << 7;  // surrogate half?
+  *e |= (*c > 0x10FFFF) << 8;       // out of range?
+  *e |= (uchar(s[1]) & 0xc0) >> 2;
+  *e |= (uchar(s[2]) & 0xc0) >> 4;
+  *e |= uchar(s[3]) >> 6;
+  *e ^= 0x2a;  // top two bits of each tail byte correct?
+  *e >>= shifte[len];
+
+  return next;
+}
+
+constexpr FMT_INLINE_VARIABLE uint32_t invalid_code_point = ~uint32_t();
+
+// Invokes f(cp, sv) for every code point cp in s with sv being the string view
+// corresponding to the code point. cp is invalid_code_point on error.
+template <typename F>
+FMT_CONSTEXPR void for_each_codepoint(string_view s, F f) {
+  auto decode = [f](const char* buf_ptr, const char* ptr) {
+    auto cp = uint32_t();
+    auto error = 0;
+    auto end = utf8_decode(buf_ptr, &cp, &error);
+    bool result = f(error ? invalid_code_point : cp,
+                    string_view(ptr, error ? 1 : to_unsigned(end - buf_ptr)));
+    return result ? (error ? buf_ptr + 1 : end) : nullptr;
+  };
+  auto p = s.data();
+  const size_t block_size = 4;  // utf8_decode always reads blocks of 4 chars.
+  if (s.size() >= block_size) {
+    for (auto end = p + s.size() - block_size + 1; p < end;) {
+      p = decode(p, p);
+      if (!p) return;
+    }
+  }
+  if (auto num_chars_left = s.data() + s.size() - p) {
+    char buf[2 * block_size - 1] = {};
+    copy_str<char>(p, p + num_chars_left, buf);
+    const char* buf_ptr = buf;
+    do {
+      auto end = decode(buf_ptr, p);
+      if (!end) return;
+      p += end - buf_ptr;
+      buf_ptr = end;
+    } while (buf_ptr - buf < num_chars_left);
+  }
+}
+
+template <typename Char>
+inline auto compute_width(basic_string_view<Char> s) -> size_t {
+  return s.size();
+}
+
+// Computes approximate display width of a UTF-8 string.
+FMT_CONSTEXPR inline auto compute_width(string_view s) -> size_t {
+  size_t num_code_points = 0;
+  // It is not a lambda for compatibility with C++14.
+  struct count_code_points {
+    size_t* count;
+    FMT_CONSTEXPR auto operator()(uint32_t cp, string_view) const -> bool {
+      *count += detail::to_unsigned(
+          1 +
+          (cp >= 0x1100 &&
+           (cp <= 0x115f ||  // Hangul Jamo init. consonants
+            cp == 0x2329 ||  // LEFT-POINTING ANGLE BRACKET
+            cp == 0x232a ||  // RIGHT-POINTING ANGLE BRACKET
+            // CJK ... Yi except IDEOGRAPHIC HALF FILL SPACE:
+            (cp >= 0x2e80 && cp <= 0xa4cf && cp != 0x303f) ||
+            (cp >= 0xac00 && cp <= 0xd7a3) ||    // Hangul Syllables
+            (cp >= 0xf900 && cp <= 0xfaff) ||    // CJK Compatibility Ideographs
+            (cp >= 0xfe10 && cp <= 0xfe19) ||    // Vertical Forms
+            (cp >= 0xfe30 && cp <= 0xfe6f) ||    // CJK Compatibility Forms
+            (cp >= 0xff00 && cp <= 0xff60) ||    // Fullwidth Forms
+            (cp >= 0xffe0 && cp <= 0xffe6) ||    // Fullwidth Forms
+            (cp >= 0x20000 && cp <= 0x2fffd) ||  // CJK
+            (cp >= 0x30000 && cp <= 0x3fffd) ||
+            // Miscellaneous Symbols and Pictographs + Emoticons:
+            (cp >= 0x1f300 && cp <= 0x1f64f) ||
+            // Supplemental Symbols and Pictographs:
+            (cp >= 0x1f900 && cp <= 0x1f9ff))));
+      return true;
+    }
+  };
+  // We could avoid branches by using utf8_decode directly.
+  for_each_codepoint(s, count_code_points{&num_code_points});
+  return num_code_points;
+}
+
+inline auto compute_width(basic_string_view<char8_type> s) -> size_t {
+  return compute_width(
+      string_view(reinterpret_cast<const char*>(s.data()), s.size()));
+}
+
+template <typename Char>
+inline auto code_point_index(basic_string_view<Char> s, size_t n) -> size_t {
+  size_t size = s.size();
+  return n < size ? n : size;
+}
+
+// Calculates the index of the nth code point in a UTF-8 string.
+inline auto code_point_index(string_view s, size_t n) -> size_t {
+  size_t result = s.size();
+  const char* begin = s.begin();
+  for_each_codepoint(s, [begin, &n, &result](uint32_t, string_view sv) {
+    if (n != 0) {
+      --n;
+      return true;
+    }
+    result = to_unsigned(sv.begin() - begin);
+    return false;
+  });
+  return result;
+}
+
+inline auto code_point_index(basic_string_view<char8_type> s, size_t n)
+    -> size_t {
+  return code_point_index(
+      string_view(reinterpret_cast<const char*>(s.data()), s.size()), n);
+}
+
+template <typename T> struct is_integral : std::is_integral<T> {};
+template <> struct is_integral<int128_opt> : std::true_type {};
+template <> struct is_integral<uint128_t> : std::true_type {};
+
+template <typename T>
+using is_signed =
+    std::integral_constant<bool, std::numeric_limits<T>::is_signed ||
+                                     std::is_same<T, int128_opt>::value>;
+
+template <typename T>
+using is_integer =
+    bool_constant<is_integral<T>::value && !std::is_same<T, bool>::value &&
+                  !std::is_same<T, char>::value &&
+                  !std::is_same<T, wchar_t>::value>;
+
+#ifndef FMT_USE_FLOAT
+#  define FMT_USE_FLOAT 1
+#endif
+#ifndef FMT_USE_DOUBLE
+#  define FMT_USE_DOUBLE 1
+#endif
+#ifndef FMT_USE_LONG_DOUBLE
+#  define FMT_USE_LONG_DOUBLE 1
+#endif
+
+#ifndef FMT_USE_FLOAT128
+#  ifdef __clang__
+// Clang emulates GCC, so it has to appear early.
+#    if FMT_HAS_INCLUDE(<quadmath.h>)
+#      define FMT_USE_FLOAT128 1
+#    endif
+#  elif defined(__GNUC__)
+// GNU C++:
+#    if defined(_GLIBCXX_USE_FLOAT128) && !defined(__STRICT_ANSI__)
+#      define FMT_USE_FLOAT128 1
+#    endif
+#  endif
+#  ifndef FMT_USE_FLOAT128
+#    define FMT_USE_FLOAT128 0
+#  endif
+#endif
+
+#if FMT_USE_FLOAT128
+using float128 = __float128;
+#else
+using float128 = void;
+#endif
+template <typename T> using is_float128 = std::is_same<T, float128>;
+
+template <typename T>
+using is_floating_point =
+    bool_constant<std::is_floating_point<T>::value || is_float128<T>::value>;
+
+template <typename T, bool = std::is_floating_point<T>::value>
+struct is_fast_float : bool_constant<std::numeric_limits<T>::is_iec559 &&
+                                     sizeof(T) <= sizeof(double)> {};
+template <typename T> struct is_fast_float<T, false> : std::false_type {};
+
+template <typename T>
+using is_double_double = bool_constant<std::numeric_limits<T>::digits == 106>;
+
+#ifndef FMT_USE_FULL_CACHE_DRAGONBOX
+#  define FMT_USE_FULL_CACHE_DRAGONBOX 0
+#endif
+
+template <typename T>
+template <typename U>
+void buffer<T>::append(const U* begin, const U* end) {
+  while (begin != end) {
+    auto count = to_unsigned(end - begin);
+    try_reserve(size_ + count);
+    auto free_cap = capacity_ - size_;
+    if (free_cap < count) count = free_cap;
+    std::uninitialized_copy_n(begin, count, ptr_ + size_);
+    size_ += count;
+    begin += count;
+  }
+}
+
+template <typename T, typename Enable = void>
+struct is_locale : std::false_type {};
+template <typename T>
+struct is_locale<T, void_t<decltype(T::classic())>> : std::true_type {};
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+// The number of characters to store in the basic_memory_buffer object itself
+// to avoid dynamic memory allocation.
+enum { inline_buffer_size = 500 };
+
+/**
+  \rst
+  A dynamically growing memory buffer for trivially copyable/constructible types
+  with the first ``SIZE`` elements stored in the object itself.
+
+  You can use the ``memory_buffer`` type alias for ``char`` instead.
+
+  **Example**::
+
+     auto out = fmt::memory_buffer();
+     fmt::format_to(std::back_inserter(out), "The answer is {}.", 42);
+
+  This will append the following output to the ``out`` object:
+
+  .. code-block:: none
+
+     The answer is 42.
+
+  The output can be converted to an ``std::string`` with ``to_string(out)``.
+  \endrst
+ */
+template <typename T, size_t SIZE = inline_buffer_size,
+          typename Allocator = std::allocator<T>>
+class basic_memory_buffer final : public detail::buffer<T> {
+ private:
+  T store_[SIZE];
+
+  // Don't inherit from Allocator to avoid generating type_info for it.
+  FMT_NO_UNIQUE_ADDRESS Allocator alloc_;
+
+  // Deallocate memory allocated by the buffer.
+  FMT_CONSTEXPR20 void deallocate() {
+    T* data = this->data();
+    if (data != store_) alloc_.deallocate(data, this->capacity());
+  }
+
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t size) override {
+    detail::abort_fuzzing_if(size > 5000);
+    const size_t max_size = std::allocator_traits<Allocator>::max_size(alloc_);
+    size_t old_capacity = this->capacity();
+    size_t new_capacity = old_capacity + old_capacity / 2;
+    if (size > new_capacity)
+      new_capacity = size;
+    else if (new_capacity > max_size)
+      new_capacity = size > max_size ? size : max_size;
+    T* old_data = this->data();
+    T* new_data =
+        std::allocator_traits<Allocator>::allocate(alloc_, new_capacity);
+    // Suppress a bogus -Wstringop-overflow in gcc 13.1 (#3481).
+    detail::assume(this->size() <= new_capacity);
+    // The following code doesn't throw, so the raw pointer above doesn't leak.
+    std::uninitialized_copy_n(old_data, this->size(), new_data);
+    this->set(new_data, new_capacity);
+    // deallocate must not throw according to the standard, but even if it does,
+    // the buffer already uses the new storage and will deallocate it in
+    // destructor.
+    if (old_data != store_) alloc_.deallocate(old_data, old_capacity);
+  }
+
+ public:
+  using value_type = T;
+  using const_reference = const T&;
+
+  FMT_CONSTEXPR20 explicit basic_memory_buffer(
+      const Allocator& alloc = Allocator())
+      : alloc_(alloc) {
+    this->set(store_, SIZE);
+    if (detail::is_constant_evaluated()) detail::fill_n(store_, SIZE, T());
+  }
+  FMT_CONSTEXPR20 ~basic_memory_buffer() { deallocate(); }
+
+ private:
+  // Move data from other to this buffer.
+  FMT_CONSTEXPR20 void move(basic_memory_buffer& other) {
+    alloc_ = std::move(other.alloc_);
+    T* data = other.data();
+    size_t size = other.size(), capacity = other.capacity();
+    if (data == other.store_) {
+      this->set(store_, capacity);
+      detail::copy_str<T>(other.store_, other.store_ + size, store_);
+    } else {
+      this->set(data, capacity);
+      // Set pointer to the inline array so that delete is not called
+      // when deallocating.
+      other.set(other.store_, 0);
+      other.clear();
+    }
+    this->resize(size);
+  }
+
+ public:
+  /**
+    \rst
+    Constructs a :class:`fmt::basic_memory_buffer` object moving the content
+    of the other object to it.
+    \endrst
+   */
+  FMT_CONSTEXPR20 basic_memory_buffer(basic_memory_buffer&& other) noexcept {
+    move(other);
+  }
+
+  /**
+    \rst
+    Moves the content of the other ``basic_memory_buffer`` object to this one.
+    \endrst
+   */
+  auto operator=(basic_memory_buffer&& other) noexcept -> basic_memory_buffer& {
+    FMT_ASSERT(this != &other, "");
+    deallocate();
+    move(other);
+    return *this;
+  }
+
+  // Returns a copy of the allocator associated with this buffer.
+  auto get_allocator() const -> Allocator { return alloc_; }
+
+  /**
+    Resizes the buffer to contain *count* elements. If T is a POD type new
+    elements may not be initialized.
+   */
+  FMT_CONSTEXPR20 void resize(size_t count) { this->try_resize(count); }
+
+  /** Increases the buffer capacity to *new_capacity*. */
+  void reserve(size_t new_capacity) { this->try_reserve(new_capacity); }
+
+  using detail::buffer<T>::append;
+  template <typename ContiguousRange>
+  void append(const ContiguousRange& range) {
+    append(range.data(), range.data() + range.size());
+  }
+};
+
+using memory_buffer = basic_memory_buffer<char>;
+
+template <typename T, size_t SIZE, typename Allocator>
+struct is_contiguous<basic_memory_buffer<T, SIZE, Allocator>> : std::true_type {
+};
+
+FMT_END_EXPORT
+namespace detail {
+FMT_API auto write_console(int fd, string_view text) -> bool;
+FMT_API auto write_console(std::FILE* f, string_view text) -> bool;
+FMT_API void print(std::FILE*, string_view);
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+// Suppress a misleading warning in older versions of clang.
+#if FMT_CLANG_VERSION
+#  pragma clang diagnostic ignored "-Wweak-vtables"
+#endif
+
+/** An error reported from a formatting function. */
+class FMT_SO_VISIBILITY("default") format_error : public std::runtime_error {
+ public:
+  using std::runtime_error::runtime_error;
+};
+
+namespace detail_exported {
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <typename Char, size_t N> struct fixed_string {
+  constexpr fixed_string(const Char (&str)[N]) {
+    detail::copy_str<Char, const Char*, Char*>(static_cast<const Char*>(str),
+                                               str + N, data);
+  }
+  Char data[N] = {};
+};
+#endif
+
+// Converts a compile-time string to basic_string_view.
+template <typename Char, size_t N>
+constexpr auto compile_string_to_view(const Char (&s)[N])
+    -> basic_string_view<Char> {
+  // Remove trailing NUL character if needed. Won't be present if this is used
+  // with a raw character array (i.e. not defined as a string).
+  return {s, N - (std::char_traits<Char>::to_int_type(s[N - 1]) == 0 ? 1 : 0)};
+}
+template <typename Char>
+constexpr auto compile_string_to_view(detail::std_string_view<Char> s)
+    -> basic_string_view<Char> {
+  return {s.data(), s.size()};
+}
+}  // namespace detail_exported
+
+class loc_value {
+ private:
+  basic_format_arg<format_context> value_;
+
+ public:
+  template <typename T, FMT_ENABLE_IF(!detail::is_float128<T>::value)>
+  loc_value(T value) : value_(detail::make_arg<format_context>(value)) {}
+
+  template <typename T, FMT_ENABLE_IF(detail::is_float128<T>::value)>
+  loc_value(T) {}
+
+  template <typename Visitor> auto visit(Visitor&& vis) -> decltype(vis(0)) {
+    return visit_format_arg(vis, value_);
+  }
+};
+
+// A locale facet that formats values in UTF-8.
+// It is parameterized on the locale to avoid the heavy <locale> include.
+template <typename Locale> class format_facet : public Locale::facet {
+ private:
+  std::string separator_;
+  std::string grouping_;
+  std::string decimal_point_;
+
+ protected:
+  virtual auto do_put(appender out, loc_value val,
+                      const format_specs<>& specs) const -> bool;
+
+ public:
+  static FMT_API typename Locale::id id;
+
+  explicit format_facet(Locale& loc);
+  explicit format_facet(string_view sep = "",
+                        std::initializer_list<unsigned char> g = {3},
+                        std::string decimal_point = ".")
+      : separator_(sep.data(), sep.size()),
+        grouping_(g.begin(), g.end()),
+        decimal_point_(decimal_point) {}
+
+  auto put(appender out, loc_value val, const format_specs<>& specs) const
+      -> bool {
+    return do_put(out, val, specs);
+  }
+};
+
+namespace detail {
+
+// Returns true if value is negative, false otherwise.
+// Same as `value < 0` but doesn't produce warnings if T is an unsigned type.
+template <typename T, FMT_ENABLE_IF(is_signed<T>::value)>
+constexpr auto is_negative(T value) -> bool {
+  return value < 0;
+}
+template <typename T, FMT_ENABLE_IF(!is_signed<T>::value)>
+constexpr auto is_negative(T) -> bool {
+  return false;
+}
+
+template <typename T>
+FMT_CONSTEXPR auto is_supported_floating_point(T) -> bool {
+  if (std::is_same<T, float>()) return FMT_USE_FLOAT;
+  if (std::is_same<T, double>()) return FMT_USE_DOUBLE;
+  if (std::is_same<T, long double>()) return FMT_USE_LONG_DOUBLE;
+  return true;
+}
+
+// Smallest of uint32_t, uint64_t, uint128_t that is large enough to
+// represent all values of an integral type T.
+template <typename T>
+using uint32_or_64_or_128_t =
+    conditional_t<num_bits<T>() <= 32 && !FMT_REDUCE_INT_INSTANTIATIONS,
+                  uint32_t,
+                  conditional_t<num_bits<T>() <= 64, uint64_t, uint128_t>>;
+template <typename T>
+using uint64_or_128_t = conditional_t<num_bits<T>() <= 64, uint64_t, uint128_t>;
+
+#define FMT_POWERS_OF_10(factor)                                  \
+  factor * 10, (factor) * 100, (factor) * 1000, (factor) * 10000, \
+      (factor) * 100000, (factor) * 1000000, (factor) * 10000000, \
+      (factor) * 100000000, (factor) * 1000000000
+
+// Converts value in the range [0, 100) to a string.
+constexpr auto digits2(size_t value) -> const char* {
+  // GCC generates slightly better code when value is pointer-size.
+  return &"0001020304050607080910111213141516171819"
+         "2021222324252627282930313233343536373839"
+         "4041424344454647484950515253545556575859"
+         "6061626364656667686970717273747576777879"
+         "8081828384858687888990919293949596979899"[value * 2];
+}
+
+// Sign is a template parameter to workaround a bug in gcc 4.8.
+template <typename Char, typename Sign> constexpr auto sign(Sign s) -> Char {
+#if !FMT_GCC_VERSION || FMT_GCC_VERSION >= 604
+  static_assert(std::is_same<Sign, sign_t>::value, "");
+#endif
+  return static_cast<Char>("\0-+ "[s]);
+}
+
+template <typename T> FMT_CONSTEXPR auto count_digits_fallback(T n) -> int {
+  int count = 1;
+  for (;;) {
+    // Integer division is slow so do it for a group of four digits instead
+    // of for every digit. The idea comes from the talk by Alexandrescu
+    // "Three Optimization Tips for C++". See speed-test for a comparison.
+    if (n < 10) return count;
+    if (n < 100) return count + 1;
+    if (n < 1000) return count + 2;
+    if (n < 10000) return count + 3;
+    n /= 10000u;
+    count += 4;
+  }
+}
+#if FMT_USE_INT128
+FMT_CONSTEXPR inline auto count_digits(uint128_opt n) -> int {
+  return count_digits_fallback(n);
+}
+#endif
+
+#ifdef FMT_BUILTIN_CLZLL
+// It is a separate function rather than a part of count_digits to workaround
+// the lack of static constexpr in constexpr functions.
+inline auto do_count_digits(uint64_t n) -> int {
+  // This has comparable performance to the version by Kendall Willets
+  // (https://github.com/fmtlib/format-benchmark/blob/master/digits10)
+  // but uses smaller tables.
+  // Maps bsr(n) to ceil(log10(pow(2, bsr(n) + 1) - 1)).
+  static constexpr uint8_t bsr2log10[] = {
+      1,  1,  1,  2,  2,  2,  3,  3,  3,  4,  4,  4,  4,  5,  5,  5,
+      6,  6,  6,  7,  7,  7,  7,  8,  8,  8,  9,  9,  9,  10, 10, 10,
+      10, 11, 11, 11, 12, 12, 12, 13, 13, 13, 13, 14, 14, 14, 15, 15,
+      15, 16, 16, 16, 16, 17, 17, 17, 18, 18, 18, 19, 19, 19, 19, 20};
+  auto t = bsr2log10[FMT_BUILTIN_CLZLL(n | 1) ^ 63];
+  static constexpr const uint64_t zero_or_powers_of_10[] = {
+      0, 0, FMT_POWERS_OF_10(1U), FMT_POWERS_OF_10(1000000000ULL),
+      10000000000000000000ULL};
+  return t - (n < zero_or_powers_of_10[t]);
+}
+#endif
+
+// Returns the number of decimal digits in n. Leading zeros are not counted
+// except for n == 0 in which case count_digits returns 1.
+FMT_CONSTEXPR20 inline auto count_digits(uint64_t n) -> int {
+#ifdef FMT_BUILTIN_CLZLL
+  if (!is_constant_evaluated()) {
+    return do_count_digits(n);
+  }
+#endif
+  return count_digits_fallback(n);
+}
+
+// Counts the number of digits in n. BITS = log2(radix).
+template <int BITS, typename UInt>
+FMT_CONSTEXPR auto count_digits(UInt n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated() && num_bits<UInt>() == 32)
+    return (FMT_BUILTIN_CLZ(static_cast<uint32_t>(n) | 1) ^ 31) / BITS + 1;
+#endif
+  // Lambda avoids unreachable code warnings from NVHPC.
+  return [](UInt m) {
+    int num_digits = 0;
+    do {
+      ++num_digits;
+    } while ((m >>= BITS) != 0);
+    return num_digits;
+  }(n);
+}
+
+#ifdef FMT_BUILTIN_CLZ
+// It is a separate function rather than a part of count_digits to workaround
+// the lack of static constexpr in constexpr functions.
+FMT_INLINE auto do_count_digits(uint32_t n) -> int {
+// An optimization by Kendall Willets from https://bit.ly/3uOIQrB.
+// This increments the upper 32 bits (log10(T) - 1) when >= T is added.
+#  define FMT_INC(T) (((sizeof(#T) - 1ull) << 32) - T)
+  static constexpr uint64_t table[] = {
+      FMT_INC(0),          FMT_INC(0),          FMT_INC(0),           // 8
+      FMT_INC(10),         FMT_INC(10),         FMT_INC(10),          // 64
+      FMT_INC(100),        FMT_INC(100),        FMT_INC(100),         // 512
+      FMT_INC(1000),       FMT_INC(1000),       FMT_INC(1000),        // 4096
+      FMT_INC(10000),      FMT_INC(10000),      FMT_INC(10000),       // 32k
+      FMT_INC(100000),     FMT_INC(100000),     FMT_INC(100000),      // 256k
+      FMT_INC(1000000),    FMT_INC(1000000),    FMT_INC(1000000),     // 2048k
+      FMT_INC(10000000),   FMT_INC(10000000),   FMT_INC(10000000),    // 16M
+      FMT_INC(100000000),  FMT_INC(100000000),  FMT_INC(100000000),   // 128M
+      FMT_INC(1000000000), FMT_INC(1000000000), FMT_INC(1000000000),  // 1024M
+      FMT_INC(1000000000), FMT_INC(1000000000)                        // 4B
+  };
+  auto inc = table[FMT_BUILTIN_CLZ(n | 1) ^ 31];
+  return static_cast<int>((n + inc) >> 32);
+}
+#endif
+
+// Optional version of count_digits for better performance on 32-bit platforms.
+FMT_CONSTEXPR20 inline auto count_digits(uint32_t n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated()) {
+    return do_count_digits(n);
+  }
+#endif
+  return count_digits_fallback(n);
+}
+
+template <typename Int> constexpr auto digits10() noexcept -> int {
+  return std::numeric_limits<Int>::digits10;
+}
+template <> constexpr auto digits10<int128_opt>() noexcept -> int { return 38; }
+template <> constexpr auto digits10<uint128_t>() noexcept -> int { return 38; }
+
+template <typename Char> struct thousands_sep_result {
+  std::string grouping;
+  Char thousands_sep;
+};
+
+template <typename Char>
+FMT_API auto thousands_sep_impl(locale_ref loc) -> thousands_sep_result<Char>;
+template <typename Char>
+inline auto thousands_sep(locale_ref loc) -> thousands_sep_result<Char> {
+  auto result = thousands_sep_impl<char>(loc);
+  return {result.grouping, Char(result.thousands_sep)};
+}
+template <>
+inline auto thousands_sep(locale_ref loc) -> thousands_sep_result<wchar_t> {
+  return thousands_sep_impl<wchar_t>(loc);
+}
+
+template <typename Char>
+FMT_API auto decimal_point_impl(locale_ref loc) -> Char;
+template <typename Char> inline auto decimal_point(locale_ref loc) -> Char {
+  return Char(decimal_point_impl<char>(loc));
+}
+template <> inline auto decimal_point(locale_ref loc) -> wchar_t {
+  return decimal_point_impl<wchar_t>(loc);
+}
+
+// Compares two characters for equality.
+template <typename Char> auto equal2(const Char* lhs, const char* rhs) -> bool {
+  return lhs[0] == Char(rhs[0]) && lhs[1] == Char(rhs[1]);
+}
+inline auto equal2(const char* lhs, const char* rhs) -> bool {
+  return memcmp(lhs, rhs, 2) == 0;
+}
+
+// Copies two characters from src to dst.
+template <typename Char>
+FMT_CONSTEXPR20 FMT_INLINE void copy2(Char* dst, const char* src) {
+  if (!is_constant_evaluated() && sizeof(Char) == sizeof(char)) {
+    memcpy(dst, src, 2);
+    return;
+  }
+  *dst++ = static_cast<Char>(*src++);
+  *dst = static_cast<Char>(*src);
+}
+
+template <typename Iterator> struct format_decimal_result {
+  Iterator begin;
+  Iterator end;
+};
+
+// Formats a decimal unsigned integer value writing into out pointing to a
+// buffer of specified size. The caller must ensure that the buffer is large
+// enough.
+template <typename Char, typename UInt>
+FMT_CONSTEXPR20 auto format_decimal(Char* out, UInt value, int size)
+    -> format_decimal_result<Char*> {
+  FMT_ASSERT(size >= count_digits(value), "invalid digit count");
+  out += size;
+  Char* end = out;
+  while (value >= 100) {
+    // Integer division is slow so do it for a group of two digits instead
+    // of for every digit. The idea comes from the talk by Alexandrescu
+    // "Three Optimization Tips for C++". See speed-test for a comparison.
+    out -= 2;
+    copy2(out, digits2(static_cast<size_t>(value % 100)));
+    value /= 100;
+  }
+  if (value < 10) {
+    *--out = static_cast<Char>('0' + value);
+    return {out, end};
+  }
+  out -= 2;
+  copy2(out, digits2(static_cast<size_t>(value)));
+  return {out, end};
+}
+
+template <typename Char, typename UInt, typename Iterator,
+          FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<Iterator>>::value)>
+FMT_CONSTEXPR inline auto format_decimal(Iterator out, UInt value, int size)
+    -> format_decimal_result<Iterator> {
+  // Buffer is large enough to hold all digits (digits10 + 1).
+  Char buffer[digits10<UInt>() + 1] = {};
+  auto end = format_decimal(buffer, value, size).end;
+  return {out, detail::copy_str_noinline<Char>(buffer, end, out)};
+}
+
+template <unsigned BASE_BITS, typename Char, typename UInt>
+FMT_CONSTEXPR auto format_uint(Char* buffer, UInt value, int num_digits,
+                               bool upper = false) -> Char* {
+  buffer += num_digits;
+  Char* end = buffer;
+  do {
+    const char* digits = upper ? "0123456789ABCDEF" : "0123456789abcdef";
+    unsigned digit = static_cast<unsigned>(value & ((1 << BASE_BITS) - 1));
+    *--buffer = static_cast<Char>(BASE_BITS < 4 ? static_cast<char>('0' + digit)
+                                                : digits[digit]);
+  } while ((value >>= BASE_BITS) != 0);
+  return end;
+}
+
+template <unsigned BASE_BITS, typename Char, typename It, typename UInt>
+FMT_CONSTEXPR inline auto format_uint(It out, UInt value, int num_digits,
+                                      bool upper = false) -> It {
+  if (auto ptr = to_pointer<Char>(out, to_unsigned(num_digits))) {
+    format_uint<BASE_BITS>(ptr, value, num_digits, upper);
+    return out;
+  }
+  // Buffer should be large enough to hold all digits (digits / BASE_BITS + 1).
+  char buffer[num_bits<UInt>() / BASE_BITS + 1] = {};
+  format_uint<BASE_BITS>(buffer, value, num_digits, upper);
+  return detail::copy_str_noinline<Char>(buffer, buffer + num_digits, out);
+}
+
+// A converter from UTF-8 to UTF-16.
+class utf8_to_utf16 {
+ private:
+  basic_memory_buffer<wchar_t> buffer_;
+
+ public:
+  FMT_API explicit utf8_to_utf16(string_view s);
+  operator basic_string_view<wchar_t>() const { return {&buffer_[0], size()}; }
+  auto size() const -> size_t { return buffer_.size() - 1; }
+  auto c_str() const -> const wchar_t* { return &buffer_[0]; }
+  auto str() const -> std::wstring { return {&buffer_[0], size()}; }
+};
+
+enum class to_utf8_error_policy { abort, replace };
+
+// A converter from UTF-16/UTF-32 (host endian) to UTF-8.
+template <typename WChar, typename Buffer = memory_buffer> class to_utf8 {
+ private:
+  Buffer buffer_;
+
+ public:
+  to_utf8() {}
+  explicit to_utf8(basic_string_view<WChar> s,
+                   to_utf8_error_policy policy = to_utf8_error_policy::abort) {
+    static_assert(sizeof(WChar) == 2 || sizeof(WChar) == 4,
+                  "Expect utf16 or utf32");
+    if (!convert(s, policy))
+      FMT_THROW(std::runtime_error(sizeof(WChar) == 2 ? "invalid utf16"
+                                                      : "invalid utf32"));
+  }
+  operator string_view() const { return string_view(&buffer_[0], size()); }
+  auto size() const -> size_t { return buffer_.size() - 1; }
+  auto c_str() const -> const char* { return &buffer_[0]; }
+  auto str() const -> std::string { return std::string(&buffer_[0], size()); }
+
+  // Performs conversion returning a bool instead of throwing exception on
+  // conversion error. This method may still throw in case of memory allocation
+  // error.
+  auto convert(basic_string_view<WChar> s,
+               to_utf8_error_policy policy = to_utf8_error_policy::abort)
+      -> bool {
+    if (!convert(buffer_, s, policy)) return false;
+    buffer_.push_back(0);
+    return true;
+  }
+  static auto convert(Buffer& buf, basic_string_view<WChar> s,
+                      to_utf8_error_policy policy = to_utf8_error_policy::abort)
+      -> bool {
+    for (auto p = s.begin(); p != s.end(); ++p) {
+      uint32_t c = static_cast<uint32_t>(*p);
+      if (sizeof(WChar) == 2 && c >= 0xd800 && c <= 0xdfff) {
+        // Handle a surrogate pair.
+        ++p;
+        if (p == s.end() || (c & 0xfc00) != 0xd800 || (*p & 0xfc00) != 0xdc00) {
+          if (policy == to_utf8_error_policy::abort) return false;
+          buf.append(string_view("\xEF\xBF\xBD"));
+          --p;
+        } else {
+          c = (c << 10) + static_cast<uint32_t>(*p) - 0x35fdc00;
+        }
+      } else if (c < 0x80) {
+        buf.push_back(static_cast<char>(c));
+      } else if (c < 0x800) {
+        buf.push_back(static_cast<char>(0xc0 | (c >> 6)));
+        buf.push_back(static_cast<char>(0x80 | (c & 0x3f)));
+      } else if ((c >= 0x800 && c <= 0xd7ff) || (c >= 0xe000 && c <= 0xffff)) {
+        buf.push_back(static_cast<char>(0xe0 | (c >> 12)));
+        buf.push_back(static_cast<char>(0x80 | ((c & 0xfff) >> 6)));
+        buf.push_back(static_cast<char>(0x80 | (c & 0x3f)));
+      } else if (c >= 0x10000 && c <= 0x10ffff) {
+        buf.push_back(static_cast<char>(0xf0 | (c >> 18)));
+        buf.push_back(static_cast<char>(0x80 | ((c & 0x3ffff) >> 12)));
+        buf.push_back(static_cast<char>(0x80 | ((c & 0xfff) >> 6)));
+        buf.push_back(static_cast<char>(0x80 | (c & 0x3f)));
+      } else {
+        return false;
+      }
+    }
+    return true;
+  }
+};
+
+// Computes 128-bit result of multiplication of two 64-bit unsigned integers.
+inline auto umul128(uint64_t x, uint64_t y) noexcept -> uint128_fallback {
+#if FMT_USE_INT128
+  auto p = static_cast<uint128_opt>(x) * static_cast<uint128_opt>(y);
+  return {static_cast<uint64_t>(p >> 64), static_cast<uint64_t>(p)};
+#elif defined(_MSC_VER) && defined(_M_X64)
+  auto hi = uint64_t();
+  auto lo = _umul128(x, y, &hi);
+  return {hi, lo};
+#else
+  const uint64_t mask = static_cast<uint64_t>(max_value<uint32_t>());
+
+  uint64_t a = x >> 32;
+  uint64_t b = x & mask;
+  uint64_t c = y >> 32;
+  uint64_t d = y & mask;
+
+  uint64_t ac = a * c;
+  uint64_t bc = b * c;
+  uint64_t ad = a * d;
+  uint64_t bd = b * d;
+
+  uint64_t intermediate = (bd >> 32) + (ad & mask) + (bc & mask);
+
+  return {ac + (intermediate >> 32) + (ad >> 32) + (bc >> 32),
+          (intermediate << 32) + (bd & mask)};
+#endif
+}
+
+namespace dragonbox {
+// Computes floor(log10(pow(2, e))) for e in [-2620, 2620] using the method from
+// https://fmt.dev/papers/Dragonbox.pdf#page=28, section 6.1.
+inline auto floor_log10_pow2(int e) noexcept -> int {
+  FMT_ASSERT(e <= 2620 && e >= -2620, "too large exponent");
+  static_assert((-1 >> 1) == -1, "right shift is not arithmetic");
+  return (e * 315653) >> 20;
+}
+
+inline auto floor_log2_pow10(int e) noexcept -> int {
+  FMT_ASSERT(e <= 1233 && e >= -1233, "too large exponent");
+  return (e * 1741647) >> 19;
+}
+
+// Computes upper 64 bits of multiplication of two 64-bit unsigned integers.
+inline auto umul128_upper64(uint64_t x, uint64_t y) noexcept -> uint64_t {
+#if FMT_USE_INT128
+  auto p = static_cast<uint128_opt>(x) * static_cast<uint128_opt>(y);
+  return static_cast<uint64_t>(p >> 64);
+#elif defined(_MSC_VER) && defined(_M_X64)
+  return __umulh(x, y);
+#else
+  return umul128(x, y).high();
+#endif
+}
+
+// Computes upper 128 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+inline auto umul192_upper128(uint64_t x, uint128_fallback y) noexcept
+    -> uint128_fallback {
+  uint128_fallback r = umul128(x, y.high());
+  r += umul128_upper64(x, y.low());
+  return r;
+}
+
+FMT_API auto get_cached_power(int k) noexcept -> uint128_fallback;
+
+// Type-specific information that Dragonbox uses.
+template <typename T, typename Enable = void> struct float_info;
+
+template <> struct float_info<float> {
+  using carrier_uint = uint32_t;
+  static const int exponent_bits = 8;
+  static const int kappa = 1;
+  static const int big_divisor = 100;
+  static const int small_divisor = 10;
+  static const int min_k = -31;
+  static const int max_k = 46;
+  static const int shorter_interval_tie_lower_threshold = -35;
+  static const int shorter_interval_tie_upper_threshold = -35;
+};
+
+template <> struct float_info<double> {
+  using carrier_uint = uint64_t;
+  static const int exponent_bits = 11;
+  static const int kappa = 2;
+  static const int big_divisor = 1000;
+  static const int small_divisor = 100;
+  static const int min_k = -292;
+  static const int max_k = 341;
+  static const int shorter_interval_tie_lower_threshold = -77;
+  static const int shorter_interval_tie_upper_threshold = -77;
+};
+
+// An 80- or 128-bit floating point number.
+template <typename T>
+struct float_info<T, enable_if_t<std::numeric_limits<T>::digits == 64 ||
+                                 std::numeric_limits<T>::digits == 113 ||
+                                 is_float128<T>::value>> {
+  using carrier_uint = detail::uint128_t;
+  static const int exponent_bits = 15;
+};
+
+// A double-double floating point number.
+template <typename T>
+struct float_info<T, enable_if_t<is_double_double<T>::value>> {
+  using carrier_uint = detail::uint128_t;
+};
+
+template <typename T> struct decimal_fp {
+  using significand_type = typename float_info<T>::carrier_uint;
+  significand_type significand;
+  int exponent;
+};
+
+template <typename T> FMT_API auto to_decimal(T x) noexcept -> decimal_fp<T>;
+}  // namespace dragonbox
+
+// Returns true iff Float has the implicit bit which is not stored.
+template <typename Float> constexpr auto has_implicit_bit() -> bool {
+  // An 80-bit FP number has a 64-bit significand an no implicit bit.
+  return std::numeric_limits<Float>::digits != 64;
+}
+
+// Returns the number of significand bits stored in Float. The implicit bit is
+// not counted since it is not stored.
+template <typename Float> constexpr auto num_significand_bits() -> int {
+  // std::numeric_limits may not support __float128.
+  return is_float128<Float>() ? 112
+                              : (std::numeric_limits<Float>::digits -
+                                 (has_implicit_bit<Float>() ? 1 : 0));
+}
+
+template <typename Float>
+constexpr auto exponent_mask() ->
+    typename dragonbox::float_info<Float>::carrier_uint {
+  using float_uint = typename dragonbox::float_info<Float>::carrier_uint;
+  return ((float_uint(1) << dragonbox::float_info<Float>::exponent_bits) - 1)
+         << num_significand_bits<Float>();
+}
+template <typename Float> constexpr auto exponent_bias() -> int {
+  // std::numeric_limits may not support __float128.
+  return is_float128<Float>() ? 16383
+                              : std::numeric_limits<Float>::max_exponent - 1;
+}
+
+// Writes the exponent exp in the form "[+-]d{2,3}" to buffer.
+template <typename Char, typename It>
+FMT_CONSTEXPR auto write_exponent(int exp, It it) -> It {
+  FMT_ASSERT(-10000 < exp && exp < 10000, "exponent out of range");
+  if (exp < 0) {
+    *it++ = static_cast<Char>('-');
+    exp = -exp;
+  } else {
+    *it++ = static_cast<Char>('+');
+  }
+  if (exp >= 100) {
+    const char* top = digits2(to_unsigned(exp / 100));
+    if (exp >= 1000) *it++ = static_cast<Char>(top[0]);
+    *it++ = static_cast<Char>(top[1]);
+    exp %= 100;
+  }
+  const char* d = digits2(to_unsigned(exp));
+  *it++ = static_cast<Char>(d[0]);
+  *it++ = static_cast<Char>(d[1]);
+  return it;
+}
+
+// A floating-point number f * pow(2, e) where F is an unsigned type.
+template <typename F> struct basic_fp {
+  F f;
+  int e;
+
+  static constexpr const int num_significand_bits =
+      static_cast<int>(sizeof(F) * num_bits<unsigned char>());
+
+  constexpr basic_fp() : f(0), e(0) {}
+  constexpr basic_fp(uint64_t f_val, int e_val) : f(f_val), e(e_val) {}
+
+  // Constructs fp from an IEEE754 floating-point number.
+  template <typename Float> FMT_CONSTEXPR basic_fp(Float n) { assign(n); }
+
+  // Assigns n to this and return true iff predecessor is closer than successor.
+  template <typename Float, FMT_ENABLE_IF(!is_double_double<Float>::value)>
+  FMT_CONSTEXPR auto assign(Float n) -> bool {
+    static_assert(std::numeric_limits<Float>::digits <= 113, "unsupported FP");
+    // Assume Float is in the format [sign][exponent][significand].
+    using carrier_uint = typename dragonbox::float_info<Float>::carrier_uint;
+    const auto num_float_significand_bits =
+        detail::num_significand_bits<Float>();
+    const auto implicit_bit = carrier_uint(1) << num_float_significand_bits;
+    const auto significand_mask = implicit_bit - 1;
+    auto u = bit_cast<carrier_uint>(n);
+    f = static_cast<F>(u & significand_mask);
+    auto biased_e = static_cast<int>((u & exponent_mask<Float>()) >>
+                                     num_float_significand_bits);
+    // The predecessor is closer if n is a normalized power of 2 (f == 0)
+    // other than the smallest normalized number (biased_e > 1).
+    auto is_predecessor_closer = f == 0 && biased_e > 1;
+    if (biased_e == 0)
+      biased_e = 1;  // Subnormals use biased exponent 1 (min exponent).
+    else if (has_implicit_bit<Float>())
+      f += static_cast<F>(implicit_bit);
+    e = biased_e - exponent_bias<Float>() - num_float_significand_bits;
+    if (!has_implicit_bit<Float>()) ++e;
+    return is_predecessor_closer;
+  }
+
+  template <typename Float, FMT_ENABLE_IF(is_double_double<Float>::value)>
+  FMT_CONSTEXPR auto assign(Float n) -> bool {
+    static_assert(std::numeric_limits<double>::is_iec559, "unsupported FP");
+    return assign(static_cast<double>(n));
+  }
+};
+
+using fp = basic_fp<unsigned long long>;
+
+// Normalizes the value converted from double and multiplied by (1 << SHIFT).
+template <int SHIFT = 0, typename F>
+FMT_CONSTEXPR auto normalize(basic_fp<F> value) -> basic_fp<F> {
+  // Handle subnormals.
+  const auto implicit_bit = F(1) << num_significand_bits<double>();
+  const auto shifted_implicit_bit = implicit_bit << SHIFT;
+  while ((value.f & shifted_implicit_bit) == 0) {
+    value.f <<= 1;
+    --value.e;
+  }
+  // Subtract 1 to account for hidden bit.
+  const auto offset = basic_fp<F>::num_significand_bits -
+                      num_significand_bits<double>() - SHIFT - 1;
+  value.f <<= offset;
+  value.e -= offset;
+  return value;
+}
+
+// Computes lhs * rhs / pow(2, 64) rounded to nearest with half-up tie breaking.
+FMT_CONSTEXPR inline auto multiply(uint64_t lhs, uint64_t rhs) -> uint64_t {
+#if FMT_USE_INT128
+  auto product = static_cast<__uint128_t>(lhs) * rhs;
+  auto f = static_cast<uint64_t>(product >> 64);
+  return (static_cast<uint64_t>(product) & (1ULL << 63)) != 0 ? f + 1 : f;
+#else
+  // Multiply 32-bit parts of significands.
+  uint64_t mask = (1ULL << 32) - 1;
+  uint64_t a = lhs >> 32, b = lhs & mask;
+  uint64_t c = rhs >> 32, d = rhs & mask;
+  uint64_t ac = a * c, bc = b * c, ad = a * d, bd = b * d;
+  // Compute mid 64-bit of result and round.
+  uint64_t mid = (bd >> 32) + (ad & mask) + (bc & mask) + (1U << 31);
+  return ac + (ad >> 32) + (bc >> 32) + (mid >> 32);
+#endif
+}
+
+FMT_CONSTEXPR inline auto operator*(fp x, fp y) -> fp {
+  return {multiply(x.f, y.f), x.e + y.e + 64};
+}
+
+template <typename T, bool doublish = num_bits<T>() == num_bits<double>()>
+using convert_float_result =
+    conditional_t<std::is_same<T, float>::value || doublish, double, T>;
+
+template <typename T>
+constexpr auto convert_float(T value) -> convert_float_result<T> {
+  return static_cast<convert_float_result<T>>(value);
+}
+
+template <typename OutputIt, typename Char>
+FMT_NOINLINE FMT_CONSTEXPR auto fill(OutputIt it, size_t n,
+                                     const fill_t<Char>& fill) -> OutputIt {
+  auto fill_size = fill.size();
+  if (fill_size == 1) return detail::fill_n(it, n, fill[0]);
+  auto data = fill.data();
+  for (size_t i = 0; i < n; ++i)
+    it = copy_str<Char>(data, data + fill_size, it);
+  return it;
+}
+
+// Writes the output of f, padded according to format specifications in specs.
+// size: output size in code units.
+// width: output display width in (terminal) column positions.
+template <align::type align = align::left, typename OutputIt, typename Char,
+          typename F>
+FMT_CONSTEXPR auto write_padded(OutputIt out, const format_specs<Char>& specs,
+                                size_t size, size_t width, F&& f) -> OutputIt {
+  static_assert(align == align::left || align == align::right, "");
+  unsigned spec_width = to_unsigned(specs.width);
+  size_t padding = spec_width > width ? spec_width - width : 0;
+  // Shifts are encoded as string literals because static constexpr is not
+  // supported in constexpr functions.
+  auto* shifts = align == align::left ? "\x1f\x1f\x00\x01" : "\x00\x1f\x00\x01";
+  size_t left_padding = padding >> shifts[specs.align];
+  size_t right_padding = padding - left_padding;
+  auto it = reserve(out, size + padding * specs.fill.size());
+  if (left_padding != 0) it = fill(it, left_padding, specs.fill);
+  it = f(it);
+  if (right_padding != 0) it = fill(it, right_padding, specs.fill);
+  return base_iterator(out, it);
+}
+
+template <align::type align = align::left, typename OutputIt, typename Char,
+          typename F>
+constexpr auto write_padded(OutputIt out, const format_specs<Char>& specs,
+                            size_t size, F&& f) -> OutputIt {
+  return write_padded<align>(out, specs, size, size, f);
+}
+
+template <align::type align = align::left, typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write_bytes(OutputIt out, string_view bytes,
+                               const format_specs<Char>& specs) -> OutputIt {
+  return write_padded<align>(
+      out, specs, bytes.size(), [bytes](reserve_iterator<OutputIt> it) {
+        const char* data = bytes.data();
+        return copy_str<Char>(data, data + bytes.size(), it);
+      });
+}
+
+template <typename Char, typename OutputIt, typename UIntPtr>
+auto write_ptr(OutputIt out, UIntPtr value, const format_specs<Char>* specs)
+    -> OutputIt {
+  int num_digits = count_digits<4>(value);
+  auto size = to_unsigned(num_digits) + size_t(2);
+  auto write = [=](reserve_iterator<OutputIt> it) {
+    *it++ = static_cast<Char>('0');
+    *it++ = static_cast<Char>('x');
+    return format_uint<4, Char>(it, value, num_digits);
+  };
+  return specs ? write_padded<align::right>(out, *specs, size, write)
+               : base_iterator(out, write(reserve(out, size)));
+}
+
+// Returns true iff the code point cp is printable.
+FMT_API auto is_printable(uint32_t cp) -> bool;
+
+inline auto needs_escape(uint32_t cp) -> bool {
+  return cp < 0x20 || cp == 0x7f || cp == '"' || cp == '\\' ||
+         !is_printable(cp);
+}
+
+template <typename Char> struct find_escape_result {
+  const Char* begin;
+  const Char* end;
+  uint32_t cp;
+};
+
+template <typename Char>
+using make_unsigned_char =
+    typename conditional_t<std::is_integral<Char>::value,
+                           std::make_unsigned<Char>,
+                           type_identity<uint32_t>>::type;
+
+template <typename Char>
+auto find_escape(const Char* begin, const Char* end)
+    -> find_escape_result<Char> {
+  for (; begin != end; ++begin) {
+    uint32_t cp = static_cast<make_unsigned_char<Char>>(*begin);
+    if (const_check(sizeof(Char) == 1) && cp >= 0x80) continue;
+    if (needs_escape(cp)) return {begin, begin + 1, cp};
+  }
+  return {begin, nullptr, 0};
+}
+
+inline auto find_escape(const char* begin, const char* end)
+    -> find_escape_result<char> {
+  if (!is_utf8()) return find_escape<char>(begin, end);
+  auto result = find_escape_result<char>{end, nullptr, 0};
+  for_each_codepoint(string_view(begin, to_unsigned(end - begin)),
+                     [&](uint32_t cp, string_view sv) {
+                       if (needs_escape(cp)) {
+                         result = {sv.begin(), sv.end(), cp};
+                         return false;
+                       }
+                       return true;
+                     });
+  return result;
+}
+
+#define FMT_STRING_IMPL(s, base, explicit)                                    \
+  [] {                                                                        \
+    /* Use the hidden visibility as a workaround for a GCC bug (#1973). */    \
+    /* Use a macro-like name to avoid shadowing warnings. */                  \
+    struct FMT_VISIBILITY("hidden") FMT_COMPILE_STRING : base {               \
+      using char_type FMT_MAYBE_UNUSED = fmt::remove_cvref_t<decltype(s[0])>; \
+      FMT_MAYBE_UNUSED FMT_CONSTEXPR explicit                                 \
+      operator fmt::basic_string_view<char_type>() const {                    \
+        return fmt::detail_exported::compile_string_to_view<char_type>(s);    \
+      }                                                                       \
+    };                                                                        \
+    return FMT_COMPILE_STRING();                                              \
+  }()
+
+/**
+  \rst
+  Constructs a compile-time format string from a string literal *s*.
+
+  **Example**::
+
+    // A compile-time error because 'd' is an invalid specifier for strings.
+    std::string s = fmt::format(FMT_STRING("{:d}"), "foo");
+  \endrst
+ */
+#define FMT_STRING(s) FMT_STRING_IMPL(s, fmt::detail::compile_string, )
+
+template <size_t width, typename Char, typename OutputIt>
+auto write_codepoint(OutputIt out, char prefix, uint32_t cp) -> OutputIt {
+  *out++ = static_cast<Char>('\\');
+  *out++ = static_cast<Char>(prefix);
+  Char buf[width];
+  fill_n(buf, width, static_cast<Char>('0'));
+  format_uint<4>(buf, cp, width);
+  return copy_str<Char>(buf, buf + width, out);
+}
+
+template <typename OutputIt, typename Char>
+auto write_escaped_cp(OutputIt out, const find_escape_result<Char>& escape)
+    -> OutputIt {
+  auto c = static_cast<Char>(escape.cp);
+  switch (escape.cp) {
+  case '\n':
+    *out++ = static_cast<Char>('\\');
+    c = static_cast<Char>('n');
+    break;
+  case '\r':
+    *out++ = static_cast<Char>('\\');
+    c = static_cast<Char>('r');
+    break;
+  case '\t':
+    *out++ = static_cast<Char>('\\');
+    c = static_cast<Char>('t');
+    break;
+  case '"':
+    FMT_FALLTHROUGH;
+  case '\'':
+    FMT_FALLTHROUGH;
+  case '\\':
+    *out++ = static_cast<Char>('\\');
+    break;
+  default:
+    if (escape.cp < 0x100) {
+      return write_codepoint<2, Char>(out, 'x', escape.cp);
+    }
+    if (escape.cp < 0x10000) {
+      return write_codepoint<4, Char>(out, 'u', escape.cp);
+    }
+    if (escape.cp < 0x110000) {
+      return write_codepoint<8, Char>(out, 'U', escape.cp);
+    }
+    for (Char escape_char : basic_string_view<Char>(
+             escape.begin, to_unsigned(escape.end - escape.begin))) {
+      out = write_codepoint<2, Char>(out, 'x',
+                                     static_cast<uint32_t>(escape_char) & 0xFF);
+    }
+    return out;
+  }
+  *out++ = c;
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+auto write_escaped_string(OutputIt out, basic_string_view<Char> str)
+    -> OutputIt {
+  *out++ = static_cast<Char>('"');
+  auto begin = str.begin(), end = str.end();
+  do {
+    auto escape = find_escape(begin, end);
+    out = copy_str<Char>(begin, escape.begin, out);
+    begin = escape.end;
+    if (!begin) break;
+    out = write_escaped_cp<OutputIt, Char>(out, escape);
+  } while (begin != end);
+  *out++ = static_cast<Char>('"');
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+auto write_escaped_char(OutputIt out, Char v) -> OutputIt {
+  Char v_array[1] = {v};
+  *out++ = static_cast<Char>('\'');
+  if ((needs_escape(static_cast<uint32_t>(v)) && v != static_cast<Char>('"')) ||
+      v == static_cast<Char>('\'')) {
+    out = write_escaped_cp(out,
+                           find_escape_result<Char>{v_array, v_array + 1,
+                                                    static_cast<uint32_t>(v)});
+  } else {
+    *out++ = v;
+  }
+  *out++ = static_cast<Char>('\'');
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write_char(OutputIt out, Char value,
+                              const format_specs<Char>& specs) -> OutputIt {
+  bool is_debug = specs.type == presentation_type::debug;
+  return write_padded(out, specs, 1, [=](reserve_iterator<OutputIt> it) {
+    if (is_debug) return write_escaped_char(it, value);
+    *it++ = value;
+    return it;
+  });
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, Char value,
+                         const format_specs<Char>& specs, locale_ref loc = {})
+    -> OutputIt {
+  // char is formatted as unsigned char for consistency across platforms.
+  using unsigned_type =
+      conditional_t<std::is_same<Char, char>::value, unsigned char, unsigned>;
+  return check_char_specs(specs)
+             ? write_char(out, value, specs)
+             : write(out, static_cast<unsigned_type>(value), specs, loc);
+}
+
+// Data for write_int that doesn't depend on output iterator type. It is used to
+// avoid template code bloat.
+template <typename Char> struct write_int_data {
+  size_t size;
+  size_t padding;
+
+  FMT_CONSTEXPR write_int_data(int num_digits, unsigned prefix,
+                               const format_specs<Char>& specs)
+      : size((prefix >> 24) + to_unsigned(num_digits)), padding(0) {
+    if (specs.align == align::numeric) {
+      auto width = to_unsigned(specs.width);
+      if (width > size) {
+        padding = width - size;
+        size = width;
+      }
+    } else if (specs.precision > num_digits) {
+      size = (prefix >> 24) + to_unsigned(specs.precision);
+      padding = to_unsigned(specs.precision - num_digits);
+    }
+  }
+};
+
+// Writes an integer in the format
+//   <left-padding><prefix><numeric-padding><digits><right-padding>
+// where <digits> are written by write_digits(it).
+// prefix contains chars in three lower bytes and the size in the fourth byte.
+template <typename OutputIt, typename Char, typename W>
+FMT_CONSTEXPR FMT_INLINE auto write_int(OutputIt out, int num_digits,
+                                        unsigned prefix,
+                                        const format_specs<Char>& specs,
+                                        W write_digits) -> OutputIt {
+  // Slightly faster check for specs.width == 0 && specs.precision == -1.
+  if ((specs.width | (specs.precision + 1)) == 0) {
+    auto it = reserve(out, to_unsigned(num_digits) + (prefix >> 24));
+    if (prefix != 0) {
+      for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+        *it++ = static_cast<Char>(p & 0xff);
+    }
+    return base_iterator(out, write_digits(it));
+  }
+  auto data = write_int_data<Char>(num_digits, prefix, specs);
+  return write_padded<align::right>(
+      out, specs, data.size, [=](reserve_iterator<OutputIt> it) {
+        for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+          *it++ = static_cast<Char>(p & 0xff);
+        it = detail::fill_n(it, data.padding, static_cast<Char>('0'));
+        return write_digits(it);
+      });
+}
+
+template <typename Char> class digit_grouping {
+ private:
+  std::string grouping_;
+  std::basic_string<Char> thousands_sep_;
+
+  struct next_state {
+    std::string::const_iterator group;
+    int pos;
+  };
+  auto initial_state() const -> next_state { return {grouping_.begin(), 0}; }
+
+  // Returns the next digit group separator position.
+  auto next(next_state& state) const -> int {
+    if (thousands_sep_.empty()) return max_value<int>();
+    if (state.group == grouping_.end()) return state.pos += grouping_.back();
+    if (*state.group <= 0 || *state.group == max_value<char>())
+      return max_value<int>();
+    state.pos += *state.group++;
+    return state.pos;
+  }
+
+ public:
+  explicit digit_grouping(locale_ref loc, bool localized = true) {
+    if (!localized) return;
+    auto sep = thousands_sep<Char>(loc);
+    grouping_ = sep.grouping;
+    if (sep.thousands_sep) thousands_sep_.assign(1, sep.thousands_sep);
+  }
+  digit_grouping(std::string grouping, std::basic_string<Char> sep)
+      : grouping_(std::move(grouping)), thousands_sep_(std::move(sep)) {}
+
+  auto has_separator() const -> bool { return !thousands_sep_.empty(); }
+
+  auto count_separators(int num_digits) const -> int {
+    int count = 0;
+    auto state = initial_state();
+    while (num_digits > next(state)) ++count;
+    return count;
+  }
+
+  // Applies grouping to digits and write the output to out.
+  template <typename Out, typename C>
+  auto apply(Out out, basic_string_view<C> digits) const -> Out {
+    auto num_digits = static_cast<int>(digits.size());
+    auto separators = basic_memory_buffer<int>();
+    separators.push_back(0);
+    auto state = initial_state();
+    while (int i = next(state)) {
+      if (i >= num_digits) break;
+      separators.push_back(i);
+    }
+    for (int i = 0, sep_index = static_cast<int>(separators.size() - 1);
+         i < num_digits; ++i) {
+      if (num_digits - i == separators[sep_index]) {
+        out =
+            copy_str<Char>(thousands_sep_.data(),
+                           thousands_sep_.data() + thousands_sep_.size(), out);
+        --sep_index;
+      }
+      *out++ = static_cast<Char>(digits[to_unsigned(i)]);
+    }
+    return out;
+  }
+};
+
+FMT_CONSTEXPR inline void prefix_append(unsigned& prefix, unsigned value) {
+  prefix |= prefix != 0 ? value << 8 : value;
+  prefix += (1u + (value > 0xff ? 1 : 0)) << 24;
+}
+
+// Writes a decimal integer with digit grouping.
+template <typename OutputIt, typename UInt, typename Char>
+auto write_int(OutputIt out, UInt value, unsigned prefix,
+               const format_specs<Char>& specs,
+               const digit_grouping<Char>& grouping) -> OutputIt {
+  static_assert(std::is_same<uint64_or_128_t<UInt>, UInt>::value, "");
+  int num_digits = 0;
+  auto buffer = memory_buffer();
+  switch (specs.type) {
+  case presentation_type::none:
+  case presentation_type::dec: {
+    num_digits = count_digits(value);
+    format_decimal<char>(appender(buffer), value, num_digits);
+    break;
+  }
+  case presentation_type::hex_lower:
+  case presentation_type::hex_upper: {
+    bool upper = specs.type == presentation_type::hex_upper;
+    if (specs.alt)
+      prefix_append(prefix, unsigned(upper ? 'X' : 'x') << 8 | '0');
+    num_digits = count_digits<4>(value);
+    format_uint<4, char>(appender(buffer), value, num_digits, upper);
+    break;
+  }
+  case presentation_type::bin_lower:
+  case presentation_type::bin_upper: {
+    bool upper = specs.type == presentation_type::bin_upper;
+    if (specs.alt)
+      prefix_append(prefix, unsigned(upper ? 'B' : 'b') << 8 | '0');
+    num_digits = count_digits<1>(value);
+    format_uint<1, char>(appender(buffer), value, num_digits);
+    break;
+  }
+  case presentation_type::oct: {
+    num_digits = count_digits<3>(value);
+    // Octal prefix '0' is counted as a digit, so only add it if precision
+    // is not greater than the number of digits.
+    if (specs.alt && specs.precision <= num_digits && value != 0)
+      prefix_append(prefix, '0');
+    format_uint<3, char>(appender(buffer), value, num_digits);
+    break;
+  }
+  case presentation_type::chr:
+    return write_char(out, static_cast<Char>(value), specs);
+  default:
+    throw_format_error("invalid format specifier");
+  }
+
+  unsigned size = (prefix != 0 ? prefix >> 24 : 0) + to_unsigned(num_digits) +
+                  to_unsigned(grouping.count_separators(num_digits));
+  return write_padded<align::right>(
+      out, specs, size, size, [&](reserve_iterator<OutputIt> it) {
+        for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+          *it++ = static_cast<Char>(p & 0xff);
+        return grouping.apply(it, string_view(buffer.data(), buffer.size()));
+      });
+}
+
+// Writes a localized value.
+FMT_API auto write_loc(appender out, loc_value value,
+                       const format_specs<>& specs, locale_ref loc) -> bool;
+template <typename OutputIt, typename Char>
+inline auto write_loc(OutputIt, loc_value, const format_specs<Char>&,
+                      locale_ref) -> bool {
+  return false;
+}
+
+template <typename UInt> struct write_int_arg {
+  UInt abs_value;
+  unsigned prefix;
+};
+
+template <typename T>
+FMT_CONSTEXPR auto make_write_int_arg(T value, sign_t sign)
+    -> write_int_arg<uint32_or_64_or_128_t<T>> {
+  auto prefix = 0u;
+  auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
+  if (is_negative(value)) {
+    prefix = 0x01000000 | '-';
+    abs_value = 0 - abs_value;
+  } else {
+    constexpr const unsigned prefixes[4] = {0, 0, 0x1000000u | '+',
+                                            0x1000000u | ' '};
+    prefix = prefixes[sign];
+  }
+  return {abs_value, prefix};
+}
+
+template <typename Char = char> struct loc_writer {
+  buffer_appender<Char> out;
+  const format_specs<Char>& specs;
+  std::basic_string<Char> sep;
+  std::string grouping;
+  std::basic_string<Char> decimal_point;
+
+  template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
+  auto operator()(T value) -> bool {
+    auto arg = make_write_int_arg(value, specs.sign);
+    write_int(out, static_cast<uint64_or_128_t<T>>(arg.abs_value), arg.prefix,
+              specs, digit_grouping<Char>(grouping, sep));
+    return true;
+  }
+
+  template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
+  auto operator()(T) -> bool {
+    return false;
+  }
+};
+
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR FMT_INLINE auto write_int(OutputIt out, write_int_arg<T> arg,
+                                        const format_specs<Char>& specs,
+                                        locale_ref) -> OutputIt {
+  static_assert(std::is_same<T, uint32_or_64_or_128_t<T>>::value, "");
+  auto abs_value = arg.abs_value;
+  auto prefix = arg.prefix;
+  switch (specs.type) {
+  case presentation_type::none:
+  case presentation_type::dec: {
+    auto num_digits = count_digits(abs_value);
+    return write_int(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_decimal<Char>(it, abs_value, num_digits).end;
+        });
+  }
+  case presentation_type::hex_lower:
+  case presentation_type::hex_upper: {
+    bool upper = specs.type == presentation_type::hex_upper;
+    if (specs.alt)
+      prefix_append(prefix, unsigned(upper ? 'X' : 'x') << 8 | '0');
+    int num_digits = count_digits<4>(abs_value);
+    return write_int(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_uint<4, Char>(it, abs_value, num_digits, upper);
+        });
+  }
+  case presentation_type::bin_lower:
+  case presentation_type::bin_upper: {
+    bool upper = specs.type == presentation_type::bin_upper;
+    if (specs.alt)
+      prefix_append(prefix, unsigned(upper ? 'B' : 'b') << 8 | '0');
+    int num_digits = count_digits<1>(abs_value);
+    return write_int(out, num_digits, prefix, specs,
+                     [=](reserve_iterator<OutputIt> it) {
+                       return format_uint<1, Char>(it, abs_value, num_digits);
+                     });
+  }
+  case presentation_type::oct: {
+    int num_digits = count_digits<3>(abs_value);
+    // Octal prefix '0' is counted as a digit, so only add it if precision
+    // is not greater than the number of digits.
+    if (specs.alt && specs.precision <= num_digits && abs_value != 0)
+      prefix_append(prefix, '0');
+    return write_int(out, num_digits, prefix, specs,
+                     [=](reserve_iterator<OutputIt> it) {
+                       return format_uint<3, Char>(it, abs_value, num_digits);
+                     });
+  }
+  case presentation_type::chr:
+    return write_char(out, static_cast<Char>(abs_value), specs);
+  default:
+    throw_format_error("invalid format specifier");
+  }
+  return out;
+}
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR FMT_NOINLINE auto write_int_noinline(
+    OutputIt out, write_int_arg<T> arg, const format_specs<Char>& specs,
+    locale_ref loc) -> OutputIt {
+  return write_int(out, arg, specs, loc);
+}
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        std::is_same<OutputIt, buffer_appender<Char>>::value)>
+FMT_CONSTEXPR FMT_INLINE auto write(OutputIt out, T value,
+                                    const format_specs<Char>& specs,
+                                    locale_ref loc) -> OutputIt {
+  if (specs.localized && write_loc(out, value, specs, loc)) return out;
+  return write_int_noinline(out, make_write_int_arg(value, specs.sign), specs,
+                            loc);
+}
+// An inlined version of write used in format string compilation.
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        !std::is_same<OutputIt, buffer_appender<Char>>::value)>
+FMT_CONSTEXPR FMT_INLINE auto write(OutputIt out, T value,
+                                    const format_specs<Char>& specs,
+                                    locale_ref loc) -> OutputIt {
+  if (specs.localized && write_loc(out, value, specs, loc)) return out;
+  return write_int(out, make_write_int_arg(value, specs.sign), specs, loc);
+}
+
+// An output iterator that counts the number of objects written to it and
+// discards them.
+class counting_iterator {
+ private:
+  size_t count_;
+
+ public:
+  using iterator_category = std::output_iterator_tag;
+  using difference_type = std::ptrdiff_t;
+  using pointer = void;
+  using reference = void;
+  FMT_UNCHECKED_ITERATOR(counting_iterator);
+
+  struct value_type {
+    template <typename T> FMT_CONSTEXPR void operator=(const T&) {}
+  };
+
+  FMT_CONSTEXPR counting_iterator() : count_(0) {}
+
+  FMT_CONSTEXPR auto count() const -> size_t { return count_; }
+
+  FMT_CONSTEXPR auto operator++() -> counting_iterator& {
+    ++count_;
+    return *this;
+  }
+  FMT_CONSTEXPR auto operator++(int) -> counting_iterator {
+    auto it = *this;
+    ++*this;
+    return it;
+  }
+
+  FMT_CONSTEXPR friend auto operator+(counting_iterator it, difference_type n)
+      -> counting_iterator {
+    it.count_ += static_cast<size_t>(n);
+    return it;
+  }
+
+  FMT_CONSTEXPR auto operator*() const -> value_type { return {}; }
+};
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, basic_string_view<Char> s,
+                         const format_specs<Char>& specs) -> OutputIt {
+  auto data = s.data();
+  auto size = s.size();
+  if (specs.precision >= 0 && to_unsigned(specs.precision) < size)
+    size = code_point_index(s, to_unsigned(specs.precision));
+  bool is_debug = specs.type == presentation_type::debug;
+  size_t width = 0;
+  if (specs.width != 0) {
+    if (is_debug)
+      width = write_escaped_string(counting_iterator{}, s).count();
+    else
+      width = compute_width(basic_string_view<Char>(data, size));
+  }
+  return write_padded(out, specs, size, width,
+                      [=](reserve_iterator<OutputIt> it) {
+                        if (is_debug) return write_escaped_string(it, s);
+                        return copy_str<Char>(data, data + size, it);
+                      });
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out,
+                         basic_string_view<type_identity_t<Char>> s,
+                         const format_specs<Char>& specs, locale_ref)
+    -> OutputIt {
+  return write(out, s, specs);
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, const Char* s,
+                         const format_specs<Char>& specs, locale_ref)
+    -> OutputIt {
+  if (specs.type == presentation_type::pointer)
+    return write_ptr<Char>(out, bit_cast<uintptr_t>(s), &specs);
+  if (!s) throw_format_error("string pointer is null");
+  return write(out, basic_string_view<Char>(s), specs, {});
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        !std::is_same<T, Char>::value)>
+FMT_CONSTEXPR auto write(OutputIt out, T value) -> OutputIt {
+  auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
+  bool negative = is_negative(value);
+  // Don't do -abs_value since it trips unsigned-integer-overflow sanitizer.
+  if (negative) abs_value = ~abs_value + 1;
+  int num_digits = count_digits(abs_value);
+  auto size = (negative ? 1 : 0) + static_cast<size_t>(num_digits);
+  auto it = reserve(out, size);
+  if (auto ptr = to_pointer<Char>(it, size)) {
+    if (negative) *ptr++ = static_cast<Char>('-');
+    format_decimal<Char>(ptr, abs_value, num_digits);
+    return out;
+  }
+  if (negative) *it++ = static_cast<Char>('-');
+  it = format_decimal<Char>(it, abs_value, num_digits).end;
+  return base_iterator(out, it);
+}
+
+// DEPRECATED!
+template <typename Char>
+FMT_CONSTEXPR auto parse_align(const Char* begin, const Char* end,
+                               format_specs<Char>& specs) -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  auto align = align::none;
+  auto p = begin + code_point_length(begin);
+  if (end - p <= 0) p = begin;
+  for (;;) {
+    switch (to_ascii(*p)) {
+    case '<':
+      align = align::left;
+      break;
+    case '>':
+      align = align::right;
+      break;
+    case '^':
+      align = align::center;
+      break;
+    }
+    if (align != align::none) {
+      if (p != begin) {
+        auto c = *begin;
+        if (c == '}') return begin;
+        if (c == '{') {
+          throw_format_error("invalid fill character '{'");
+          return begin;
+        }
+        specs.fill = {begin, to_unsigned(p - begin)};
+        begin = p + 1;
+      } else {
+        ++begin;
+      }
+      break;
+    } else if (p == begin) {
+      break;
+    }
+    p = begin;
+  }
+  specs.align = align;
+  return begin;
+}
+
+// A floating-point presentation format.
+enum class float_format : unsigned char {
+  general,  // General: exponent notation or fixed point based on magnitude.
+  exp,      // Exponent notation with the default precision of 6, e.g. 1.2e-3.
+  fixed,    // Fixed point with the default precision of 6, e.g. 0.0012.
+  hex
+};
+
+struct float_specs {
+  int precision;
+  float_format format : 8;
+  sign_t sign : 8;
+  bool upper : 1;
+  bool locale : 1;
+  bool binary32 : 1;
+  bool showpoint : 1;
+};
+
+template <typename Char>
+FMT_CONSTEXPR auto parse_float_type_spec(const format_specs<Char>& specs)
+    -> float_specs {
+  auto result = float_specs();
+  result.showpoint = specs.alt;
+  result.locale = specs.localized;
+  switch (specs.type) {
+  case presentation_type::none:
+    result.format = float_format::general;
+    break;
+  case presentation_type::general_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::general_lower:
+    result.format = float_format::general;
+    break;
+  case presentation_type::exp_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::exp_lower:
+    result.format = float_format::exp;
+    result.showpoint |= specs.precision != 0;
+    break;
+  case presentation_type::fixed_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::fixed_lower:
+    result.format = float_format::fixed;
+    result.showpoint |= specs.precision != 0;
+    break;
+  case presentation_type::hexfloat_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::hexfloat_lower:
+    result.format = float_format::hex;
+    break;
+  default:
+    throw_format_error("invalid format specifier");
+    break;
+  }
+  return result;
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR20 auto write_nonfinite(OutputIt out, bool isnan,
+                                     format_specs<Char> specs,
+                                     const float_specs& fspecs) -> OutputIt {
+  auto str =
+      isnan ? (fspecs.upper ? "NAN" : "nan") : (fspecs.upper ? "INF" : "inf");
+  constexpr size_t str_size = 3;
+  auto sign = fspecs.sign;
+  auto size = str_size + (sign ? 1 : 0);
+  // Replace '0'-padding with space for non-finite values.
+  const bool is_zero_fill =
+      specs.fill.size() == 1 && *specs.fill.data() == static_cast<Char>('0');
+  if (is_zero_fill) specs.fill[0] = static_cast<Char>(' ');
+  return write_padded(out, specs, size, [=](reserve_iterator<OutputIt> it) {
+    if (sign) *it++ = detail::sign<Char>(sign);
+    return copy_str<Char>(str, str + str_size, it);
+  });
+}
+
+// A decimal floating-point number significand * pow(10, exp).
+struct big_decimal_fp {
+  const char* significand;
+  int significand_size;
+  int exponent;
+};
+
+constexpr auto get_significand_size(const big_decimal_fp& f) -> int {
+  return f.significand_size;
+}
+template <typename T>
+inline auto get_significand_size(const dragonbox::decimal_fp<T>& f) -> int {
+  return count_digits(f.significand);
+}
+
+template <typename Char, typename OutputIt>
+constexpr auto write_significand(OutputIt out, const char* significand,
+                                 int significand_size) -> OutputIt {
+  return copy_str<Char>(significand, significand + significand_size, out);
+}
+template <typename Char, typename OutputIt, typename UInt>
+inline auto write_significand(OutputIt out, UInt significand,
+                              int significand_size) -> OutputIt {
+  return format_decimal<Char>(out, significand, significand_size).end;
+}
+template <typename Char, typename OutputIt, typename T, typename Grouping>
+FMT_CONSTEXPR20 auto write_significand(OutputIt out, T significand,
+                                       int significand_size, int exponent,
+                                       const Grouping& grouping) -> OutputIt {
+  if (!grouping.has_separator()) {
+    out = write_significand<Char>(out, significand, significand_size);
+    return detail::fill_n(out, exponent, static_cast<Char>('0'));
+  }
+  auto buffer = memory_buffer();
+  write_significand<char>(appender(buffer), significand, significand_size);
+  detail::fill_n(appender(buffer), exponent, '0');
+  return grouping.apply(out, string_view(buffer.data(), buffer.size()));
+}
+
+template <typename Char, typename UInt,
+          FMT_ENABLE_IF(std::is_integral<UInt>::value)>
+inline auto write_significand(Char* out, UInt significand, int significand_size,
+                              int integral_size, Char decimal_point) -> Char* {
+  if (!decimal_point)
+    return format_decimal(out, significand, significand_size).end;
+  out += significand_size + 1;
+  Char* end = out;
+  int floating_size = significand_size - integral_size;
+  for (int i = floating_size / 2; i > 0; --i) {
+    out -= 2;
+    copy2(out, digits2(static_cast<std::size_t>(significand % 100)));
+    significand /= 100;
+  }
+  if (floating_size % 2 != 0) {
+    *--out = static_cast<Char>('0' + significand % 10);
+    significand /= 10;
+  }
+  *--out = decimal_point;
+  format_decimal(out - integral_size, significand, integral_size);
+  return end;
+}
+
+template <typename OutputIt, typename UInt, typename Char,
+          FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<OutputIt>>::value)>
+inline auto write_significand(OutputIt out, UInt significand,
+                              int significand_size, int integral_size,
+                              Char decimal_point) -> OutputIt {
+  // Buffer is large enough to hold digits (digits10 + 1) and a decimal point.
+  Char buffer[digits10<UInt>() + 2];
+  auto end = write_significand(buffer, significand, significand_size,
+                               integral_size, decimal_point);
+  return detail::copy_str_noinline<Char>(buffer, end, out);
+}
+
+template <typename OutputIt, typename Char>
+FMT_CONSTEXPR auto write_significand(OutputIt out, const char* significand,
+                                     int significand_size, int integral_size,
+                                     Char decimal_point) -> OutputIt {
+  out = detail::copy_str_noinline<Char>(significand,
+                                        significand + integral_size, out);
+  if (!decimal_point) return out;
+  *out++ = decimal_point;
+  return detail::copy_str_noinline<Char>(significand + integral_size,
+                                         significand + significand_size, out);
+}
+
+template <typename OutputIt, typename Char, typename T, typename Grouping>
+FMT_CONSTEXPR20 auto write_significand(OutputIt out, T significand,
+                                       int significand_size, int integral_size,
+                                       Char decimal_point,
+                                       const Grouping& grouping) -> OutputIt {
+  if (!grouping.has_separator()) {
+    return write_significand(out, significand, significand_size, integral_size,
+                             decimal_point);
+  }
+  auto buffer = basic_memory_buffer<Char>();
+  write_significand(buffer_appender<Char>(buffer), significand,
+                    significand_size, integral_size, decimal_point);
+  grouping.apply(
+      out, basic_string_view<Char>(buffer.data(), to_unsigned(integral_size)));
+  return detail::copy_str_noinline<Char>(buffer.data() + integral_size,
+                                         buffer.end(), out);
+}
+
+template <typename OutputIt, typename DecimalFP, typename Char,
+          typename Grouping = digit_grouping<Char>>
+FMT_CONSTEXPR20 auto do_write_float(OutputIt out, const DecimalFP& f,
+                                    const format_specs<Char>& specs,
+                                    float_specs fspecs, locale_ref loc)
+    -> OutputIt {
+  auto significand = f.significand;
+  int significand_size = get_significand_size(f);
+  const Char zero = static_cast<Char>('0');
+  auto sign = fspecs.sign;
+  size_t size = to_unsigned(significand_size) + (sign ? 1 : 0);
+  using iterator = reserve_iterator<OutputIt>;
+
+  Char decimal_point =
+      fspecs.locale ? detail::decimal_point<Char>(loc) : static_cast<Char>('.');
+
+  int output_exp = f.exponent + significand_size - 1;
+  auto use_exp_format = [=]() {
+    if (fspecs.format == float_format::exp) return true;
+    if (fspecs.format != float_format::general) return false;
+    // Use the fixed notation if the exponent is in [exp_lower, exp_upper),
+    // e.g. 0.0001 instead of 1e-04. Otherwise use the exponent notation.
+    const int exp_lower = -4, exp_upper = 16;
+    return output_exp < exp_lower ||
+           output_exp >= (fspecs.precision > 0 ? fspecs.precision : exp_upper);
+  };
+  if (use_exp_format()) {
+    int num_zeros = 0;
+    if (fspecs.showpoint) {
+      num_zeros = fspecs.precision - significand_size;
+      if (num_zeros < 0) num_zeros = 0;
+      size += to_unsigned(num_zeros);
+    } else if (significand_size == 1) {
+      decimal_point = Char();
+    }
+    auto abs_output_exp = output_exp >= 0 ? output_exp : -output_exp;
+    int exp_digits = 2;
+    if (abs_output_exp >= 100) exp_digits = abs_output_exp >= 1000 ? 4 : 3;
+
+    size += to_unsigned((decimal_point ? 1 : 0) + 2 + exp_digits);
+    char exp_char = fspecs.upper ? 'E' : 'e';
+    auto write = [=](iterator it) {
+      if (sign) *it++ = detail::sign<Char>(sign);
+      // Insert a decimal point after the first digit and add an exponent.
+      it = write_significand(it, significand, significand_size, 1,
+                             decimal_point);
+      if (num_zeros > 0) it = detail::fill_n(it, num_zeros, zero);
+      *it++ = static_cast<Char>(exp_char);
+      return write_exponent<Char>(output_exp, it);
+    };
+    return specs.width > 0 ? write_padded<align::right>(out, specs, size, write)
+                           : base_iterator(out, write(reserve(out, size)));
+  }
+
+  int exp = f.exponent + significand_size;
+  if (f.exponent >= 0) {
+    // 1234e5 -> 123400000[.0+]
+    size += to_unsigned(f.exponent);
+    int num_zeros = fspecs.precision - exp;
+    abort_fuzzing_if(num_zeros > 5000);
+    if (fspecs.showpoint) {
+      ++size;
+      if (num_zeros <= 0 && fspecs.format != float_format::fixed) num_zeros = 0;
+      if (num_zeros > 0) size += to_unsigned(num_zeros);
+    }
+    auto grouping = Grouping(loc, fspecs.locale);
+    size += to_unsigned(grouping.count_separators(exp));
+    return write_padded<align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = detail::sign<Char>(sign);
+      it = write_significand<Char>(it, significand, significand_size,
+                                   f.exponent, grouping);
+      if (!fspecs.showpoint) return it;
+      *it++ = decimal_point;
+      return num_zeros > 0 ? detail::fill_n(it, num_zeros, zero) : it;
+    });
+  } else if (exp > 0) {
+    // 1234e-2 -> 12.34[0+]
+    int num_zeros = fspecs.showpoint ? fspecs.precision - significand_size : 0;
+    size += 1 + to_unsigned(num_zeros > 0 ? num_zeros : 0);
+    auto grouping = Grouping(loc, fspecs.locale);
+    size += to_unsigned(grouping.count_separators(exp));
+    return write_padded<align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = detail::sign<Char>(sign);
+      it = write_significand(it, significand, significand_size, exp,
+                             decimal_point, grouping);
+      return num_zeros > 0 ? detail::fill_n(it, num_zeros, zero) : it;
+    });
+  }
+  // 1234e-6 -> 0.001234
+  int num_zeros = -exp;
+  if (significand_size == 0 && fspecs.precision >= 0 &&
+      fspecs.precision < num_zeros) {
+    num_zeros = fspecs.precision;
+  }
+  bool pointy = num_zeros != 0 || significand_size != 0 || fspecs.showpoint;
+  size += 1 + (pointy ? 1 : 0) + to_unsigned(num_zeros);
+  return write_padded<align::right>(out, specs, size, [&](iterator it) {
+    if (sign) *it++ = detail::sign<Char>(sign);
+    *it++ = zero;
+    if (!pointy) return it;
+    *it++ = decimal_point;
+    it = detail::fill_n(it, num_zeros, zero);
+    return write_significand<Char>(it, significand, significand_size);
+  });
+}
+
+template <typename Char> class fallback_digit_grouping {
+ public:
+  constexpr fallback_digit_grouping(locale_ref, bool) {}
+
+  constexpr auto has_separator() const -> bool { return false; }
+
+  constexpr auto count_separators(int) const -> int { return 0; }
+
+  template <typename Out, typename C>
+  constexpr auto apply(Out out, basic_string_view<C>) const -> Out {
+    return out;
+  }
+};
+
+template <typename OutputIt, typename DecimalFP, typename Char>
+FMT_CONSTEXPR20 auto write_float(OutputIt out, const DecimalFP& f,
+                                 const format_specs<Char>& specs,
+                                 float_specs fspecs, locale_ref loc)
+    -> OutputIt {
+  if (is_constant_evaluated()) {
+    return do_write_float<OutputIt, DecimalFP, Char,
+                          fallback_digit_grouping<Char>>(out, f, specs, fspecs,
+                                                         loc);
+  } else {
+    return do_write_float(out, f, specs, fspecs, loc);
+  }
+}
+
+template <typename T> constexpr auto isnan(T value) -> bool {
+  return !(value >= value);  // std::isnan doesn't support __float128.
+}
+
+template <typename T, typename Enable = void>
+struct has_isfinite : std::false_type {};
+
+template <typename T>
+struct has_isfinite<T, enable_if_t<sizeof(std::isfinite(T())) != 0>>
+    : std::true_type {};
+
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value&&
+                                        has_isfinite<T>::value)>
+FMT_CONSTEXPR20 auto isfinite(T value) -> bool {
+  constexpr T inf = T(std::numeric_limits<double>::infinity());
+  if (is_constant_evaluated())
+    return !detail::isnan(value) && value < inf && value > -inf;
+  return std::isfinite(value);
+}
+template <typename T, FMT_ENABLE_IF(!has_isfinite<T>::value)>
+FMT_CONSTEXPR auto isfinite(T value) -> bool {
+  T inf = T(std::numeric_limits<double>::infinity());
+  // std::isfinite doesn't support __float128.
+  return !detail::isnan(value) && value < inf && value > -inf;
+}
+
+template <typename T, FMT_ENABLE_IF(is_floating_point<T>::value)>
+FMT_INLINE FMT_CONSTEXPR bool signbit(T value) {
+  if (is_constant_evaluated()) {
+#ifdef __cpp_if_constexpr
+    if constexpr (std::numeric_limits<double>::is_iec559) {
+      auto bits = detail::bit_cast<uint64_t>(static_cast<double>(value));
+      return (bits >> (num_bits<uint64_t>() - 1)) != 0;
+    }
+#endif
+  }
+  return std::signbit(static_cast<double>(value));
+}
+
+inline FMT_CONSTEXPR20 void adjust_precision(int& precision, int exp10) {
+  // Adjust fixed precision by exponent because it is relative to decimal
+  // point.
+  if (exp10 > 0 && precision > max_value<int>() - exp10)
+    FMT_THROW(format_error("number is too big"));
+  precision += exp10;
+}
+
+class bigint {
+ private:
+  // A bigint is stored as an array of bigits (big digits), with bigit at index
+  // 0 being the least significant one.
+  using bigit = uint32_t;
+  using double_bigit = uint64_t;
+  enum { bigits_capacity = 32 };
+  basic_memory_buffer<bigit, bigits_capacity> bigits_;
+  int exp_;
+
+  FMT_CONSTEXPR20 auto operator[](int index) const -> bigit {
+    return bigits_[to_unsigned(index)];
+  }
+  FMT_CONSTEXPR20 auto operator[](int index) -> bigit& {
+    return bigits_[to_unsigned(index)];
+  }
+
+  static constexpr const int bigit_bits = num_bits<bigit>();
+
+  friend struct formatter<bigint>;
+
+  FMT_CONSTEXPR20 void subtract_bigits(int index, bigit other, bigit& borrow) {
+    auto result = static_cast<double_bigit>((*this)[index]) - other - borrow;
+    (*this)[index] = static_cast<bigit>(result);
+    borrow = static_cast<bigit>(result >> (bigit_bits * 2 - 1));
+  }
+
+  FMT_CONSTEXPR20 void remove_leading_zeros() {
+    int num_bigits = static_cast<int>(bigits_.size()) - 1;
+    while (num_bigits > 0 && (*this)[num_bigits] == 0) --num_bigits;
+    bigits_.resize(to_unsigned(num_bigits + 1));
+  }
+
+  // Computes *this -= other assuming aligned bigints and *this >= other.
+  FMT_CONSTEXPR20 void subtract_aligned(const bigint& other) {
+    FMT_ASSERT(other.exp_ >= exp_, "unaligned bigints");
+    FMT_ASSERT(compare(*this, other) >= 0, "");
+    bigit borrow = 0;
+    int i = other.exp_ - exp_;
+    for (size_t j = 0, n = other.bigits_.size(); j != n; ++i, ++j)
+      subtract_bigits(i, other.bigits_[j], borrow);
+    while (borrow > 0) subtract_bigits(i, 0, borrow);
+    remove_leading_zeros();
+  }
+
+  FMT_CONSTEXPR20 void multiply(uint32_t value) {
+    const double_bigit wide_value = value;
+    bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      double_bigit result = bigits_[i] * wide_value + carry;
+      bigits_[i] = static_cast<bigit>(result);
+      carry = static_cast<bigit>(result >> bigit_bits);
+    }
+    if (carry != 0) bigits_.push_back(carry);
+  }
+
+  template <typename UInt, FMT_ENABLE_IF(std::is_same<UInt, uint64_t>::value ||
+                                         std::is_same<UInt, uint128_t>::value)>
+  FMT_CONSTEXPR20 void multiply(UInt value) {
+    using half_uint =
+        conditional_t<std::is_same<UInt, uint128_t>::value, uint64_t, uint32_t>;
+    const int shift = num_bits<half_uint>() - bigit_bits;
+    const UInt lower = static_cast<half_uint>(value);
+    const UInt upper = value >> num_bits<half_uint>();
+    UInt carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      UInt result = lower * bigits_[i] + static_cast<bigit>(carry);
+      carry = (upper * bigits_[i] << shift) + (result >> bigit_bits) +
+              (carry >> bigit_bits);
+      bigits_[i] = static_cast<bigit>(result);
+    }
+    while (carry != 0) {
+      bigits_.push_back(static_cast<bigit>(carry));
+      carry >>= bigit_bits;
+    }
+  }
+
+  template <typename UInt, FMT_ENABLE_IF(std::is_same<UInt, uint64_t>::value ||
+                                         std::is_same<UInt, uint128_t>::value)>
+  FMT_CONSTEXPR20 void assign(UInt n) {
+    size_t num_bigits = 0;
+    do {
+      bigits_[num_bigits++] = static_cast<bigit>(n);
+      n >>= bigit_bits;
+    } while (n != 0);
+    bigits_.resize(num_bigits);
+    exp_ = 0;
+  }
+
+ public:
+  FMT_CONSTEXPR20 bigint() : exp_(0) {}
+  explicit bigint(uint64_t n) { assign(n); }
+
+  bigint(const bigint&) = delete;
+  void operator=(const bigint&) = delete;
+
+  FMT_CONSTEXPR20 void assign(const bigint& other) {
+    auto size = other.bigits_.size();
+    bigits_.resize(size);
+    auto data = other.bigits_.data();
+    copy_str<bigit>(data, data + size, bigits_.data());
+    exp_ = other.exp_;
+  }
+
+  template <typename Int> FMT_CONSTEXPR20 void operator=(Int n) {
+    FMT_ASSERT(n > 0, "");
+    assign(uint64_or_128_t<Int>(n));
+  }
+
+  FMT_CONSTEXPR20 auto num_bigits() const -> int {
+    return static_cast<int>(bigits_.size()) + exp_;
+  }
+
+  FMT_NOINLINE FMT_CONSTEXPR20 auto operator<<=(int shift) -> bigint& {
+    FMT_ASSERT(shift >= 0, "");
+    exp_ += shift / bigit_bits;
+    shift %= bigit_bits;
+    if (shift == 0) return *this;
+    bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      bigit c = bigits_[i] >> (bigit_bits - shift);
+      bigits_[i] = (bigits_[i] << shift) + carry;
+      carry = c;
+    }
+    if (carry != 0) bigits_.push_back(carry);
+    return *this;
+  }
+
+  template <typename Int>
+  FMT_CONSTEXPR20 auto operator*=(Int value) -> bigint& {
+    FMT_ASSERT(value > 0, "");
+    multiply(uint32_or_64_or_128_t<Int>(value));
+    return *this;
+  }
+
+  friend FMT_CONSTEXPR20 auto compare(const bigint& lhs, const bigint& rhs)
+      -> int {
+    int num_lhs_bigits = lhs.num_bigits(), num_rhs_bigits = rhs.num_bigits();
+    if (num_lhs_bigits != num_rhs_bigits)
+      return num_lhs_bigits > num_rhs_bigits ? 1 : -1;
+    int i = static_cast<int>(lhs.bigits_.size()) - 1;
+    int j = static_cast<int>(rhs.bigits_.size()) - 1;
+    int end = i - j;
+    if (end < 0) end = 0;
+    for (; i >= end; --i, --j) {
+      bigit lhs_bigit = lhs[i], rhs_bigit = rhs[j];
+      if (lhs_bigit != rhs_bigit) return lhs_bigit > rhs_bigit ? 1 : -1;
+    }
+    if (i != j) return i > j ? 1 : -1;
+    return 0;
+  }
+
+  // Returns compare(lhs1 + lhs2, rhs).
+  friend FMT_CONSTEXPR20 auto add_compare(const bigint& lhs1,
+                                          const bigint& lhs2, const bigint& rhs)
+      -> int {
+    auto minimum = [](int a, int b) { return a < b ? a : b; };
+    auto maximum = [](int a, int b) { return a > b ? a : b; };
+    int max_lhs_bigits = maximum(lhs1.num_bigits(), lhs2.num_bigits());
+    int num_rhs_bigits = rhs.num_bigits();
+    if (max_lhs_bigits + 1 < num_rhs_bigits) return -1;
+    if (max_lhs_bigits > num_rhs_bigits) return 1;
+    auto get_bigit = [](const bigint& n, int i) -> bigit {
+      return i >= n.exp_ && i < n.num_bigits() ? n[i - n.exp_] : 0;
+    };
+    double_bigit borrow = 0;
+    int min_exp = minimum(minimum(lhs1.exp_, lhs2.exp_), rhs.exp_);
+    for (int i = num_rhs_bigits - 1; i >= min_exp; --i) {
+      double_bigit sum =
+          static_cast<double_bigit>(get_bigit(lhs1, i)) + get_bigit(lhs2, i);
+      bigit rhs_bigit = get_bigit(rhs, i);
+      if (sum > rhs_bigit + borrow) return 1;
+      borrow = rhs_bigit + borrow - sum;
+      if (borrow > 1) return -1;
+      borrow <<= bigit_bits;
+    }
+    return borrow != 0 ? -1 : 0;
+  }
+
+  // Assigns pow(10, exp) to this bigint.
+  FMT_CONSTEXPR20 void assign_pow10(int exp) {
+    FMT_ASSERT(exp >= 0, "");
+    if (exp == 0) return *this = 1;
+    // Find the top bit.
+    int bitmask = 1;
+    while (exp >= bitmask) bitmask <<= 1;
+    bitmask >>= 1;
+    // pow(10, exp) = pow(5, exp) * pow(2, exp). First compute pow(5, exp) by
+    // repeated squaring and multiplication.
+    *this = 5;
+    bitmask >>= 1;
+    while (bitmask != 0) {
+      square();
+      if ((exp & bitmask) != 0) *this *= 5;
+      bitmask >>= 1;
+    }
+    *this <<= exp;  // Multiply by pow(2, exp) by shifting.
+  }
+
+  FMT_CONSTEXPR20 void square() {
+    int num_bigits = static_cast<int>(bigits_.size());
+    int num_result_bigits = 2 * num_bigits;
+    basic_memory_buffer<bigit, bigits_capacity> n(std::move(bigits_));
+    bigits_.resize(to_unsigned(num_result_bigits));
+    auto sum = uint128_t();
+    for (int bigit_index = 0; bigit_index < num_bigits; ++bigit_index) {
+      // Compute bigit at position bigit_index of the result by adding
+      // cross-product terms n[i] * n[j] such that i + j == bigit_index.
+      for (int i = 0, j = bigit_index; j >= 0; ++i, --j) {
+        // Most terms are multiplied twice which can be optimized in the future.
+        sum += static_cast<double_bigit>(n[i]) * n[j];
+      }
+      (*this)[bigit_index] = static_cast<bigit>(sum);
+      sum >>= num_bits<bigit>();  // Compute the carry.
+    }
+    // Do the same for the top half.
+    for (int bigit_index = num_bigits; bigit_index < num_result_bigits;
+         ++bigit_index) {
+      for (int j = num_bigits - 1, i = bigit_index - j; i < num_bigits;)
+        sum += static_cast<double_bigit>(n[i++]) * n[j--];
+      (*this)[bigit_index] = static_cast<bigit>(sum);
+      sum >>= num_bits<bigit>();
+    }
+    remove_leading_zeros();
+    exp_ *= 2;
+  }
+
+  // If this bigint has a bigger exponent than other, adds trailing zero to make
+  // exponents equal. This simplifies some operations such as subtraction.
+  FMT_CONSTEXPR20 void align(const bigint& other) {
+    int exp_difference = exp_ - other.exp_;
+    if (exp_difference <= 0) return;
+    int num_bigits = static_cast<int>(bigits_.size());
+    bigits_.resize(to_unsigned(num_bigits + exp_difference));
+    for (int i = num_bigits - 1, j = i + exp_difference; i >= 0; --i, --j)
+      bigits_[j] = bigits_[i];
+    std::uninitialized_fill_n(bigits_.data(), exp_difference, 0u);
+    exp_ -= exp_difference;
+  }
+
+  // Divides this bignum by divisor, assigning the remainder to this and
+  // returning the quotient.
+  FMT_CONSTEXPR20 auto divmod_assign(const bigint& divisor) -> int {
+    FMT_ASSERT(this != &divisor, "");
+    if (compare(*this, divisor) < 0) return 0;
+    FMT_ASSERT(divisor.bigits_[divisor.bigits_.size() - 1u] != 0, "");
+    align(divisor);
+    int quotient = 0;
+    do {
+      subtract_aligned(divisor);
+      ++quotient;
+    } while (compare(*this, divisor) >= 0);
+    return quotient;
+  }
+};
+
+// format_dragon flags.
+enum dragon {
+  predecessor_closer = 1,
+  fixup = 2,  // Run fixup to correct exp10 which can be off by one.
+  fixed = 4,
+};
+
+// Formats a floating-point number using a variation of the Fixed-Precision
+// Positive Floating-Point Printout ((FPP)^2) algorithm by Steele & White:
+// https://fmt.dev/papers/p372-steele.pdf.
+FMT_CONSTEXPR20 inline void format_dragon(basic_fp<uint128_t> value,
+                                          unsigned flags, int num_digits,
+                                          buffer<char>& buf, int& exp10) {
+  bigint numerator;    // 2 * R in (FPP)^2.
+  bigint denominator;  // 2 * S in (FPP)^2.
+  // lower and upper are differences between value and corresponding boundaries.
+  bigint lower;             // (M^- in (FPP)^2).
+  bigint upper_store;       // upper's value if different from lower.
+  bigint* upper = nullptr;  // (M^+ in (FPP)^2).
+  // Shift numerator and denominator by an extra bit or two (if lower boundary
+  // is closer) to make lower and upper integers. This eliminates multiplication
+  // by 2 during later computations.
+  bool is_predecessor_closer = (flags & dragon::predecessor_closer) != 0;
+  int shift = is_predecessor_closer ? 2 : 1;
+  if (value.e >= 0) {
+    numerator = value.f;
+    numerator <<= value.e + shift;
+    lower = 1;
+    lower <<= value.e;
+    if (is_predecessor_closer) {
+      upper_store = 1;
+      upper_store <<= value.e + 1;
+      upper = &upper_store;
+    }
+    denominator.assign_pow10(exp10);
+    denominator <<= shift;
+  } else if (exp10 < 0) {
+    numerator.assign_pow10(-exp10);
+    lower.assign(numerator);
+    if (is_predecessor_closer) {
+      upper_store.assign(numerator);
+      upper_store <<= 1;
+      upper = &upper_store;
+    }
+    numerator *= value.f;
+    numerator <<= shift;
+    denominator = 1;
+    denominator <<= shift - value.e;
+  } else {
+    numerator = value.f;
+    numerator <<= shift;
+    denominator.assign_pow10(exp10);
+    denominator <<= shift - value.e;
+    lower = 1;
+    if (is_predecessor_closer) {
+      upper_store = 1ULL << 1;
+      upper = &upper_store;
+    }
+  }
+  int even = static_cast<int>((value.f & 1) == 0);
+  if (!upper) upper = &lower;
+  bool shortest = num_digits < 0;
+  if ((flags & dragon::fixup) != 0) {
+    if (add_compare(numerator, *upper, denominator) + even <= 0) {
+      --exp10;
+      numerator *= 10;
+      if (num_digits < 0) {
+        lower *= 10;
+        if (upper != &lower) *upper *= 10;
+      }
+    }
+    if ((flags & dragon::fixed) != 0) adjust_precision(num_digits, exp10 + 1);
+  }
+  // Invariant: value == (numerator / denominator) * pow(10, exp10).
+  if (shortest) {
+    // Generate the shortest representation.
+    num_digits = 0;
+    char* data = buf.data();
+    for (;;) {
+      int digit = numerator.divmod_assign(denominator);
+      bool low = compare(numerator, lower) - even < 0;  // numerator <[=] lower.
+      // numerator + upper >[=] pow10:
+      bool high = add_compare(numerator, *upper, denominator) + even > 0;
+      data[num_digits++] = static_cast<char>('0' + digit);
+      if (low || high) {
+        if (!low) {
+          ++data[num_digits - 1];
+        } else if (high) {
+          int result = add_compare(numerator, numerator, denominator);
+          // Round half to even.
+          if (result > 0 || (result == 0 && (digit % 2) != 0))
+            ++data[num_digits - 1];
+        }
+        buf.try_resize(to_unsigned(num_digits));
+        exp10 -= num_digits - 1;
+        return;
+      }
+      numerator *= 10;
+      lower *= 10;
+      if (upper != &lower) *upper *= 10;
+    }
+  }
+  // Generate the given number of digits.
+  exp10 -= num_digits - 1;
+  if (num_digits <= 0) {
+    denominator *= 10;
+    auto digit = add_compare(numerator, numerator, denominator) > 0 ? '1' : '0';
+    buf.push_back(digit);
+    return;
+  }
+  buf.try_resize(to_unsigned(num_digits));
+  for (int i = 0; i < num_digits - 1; ++i) {
+    int digit = numerator.divmod_assign(denominator);
+    buf[i] = static_cast<char>('0' + digit);
+    numerator *= 10;
+  }
+  int digit = numerator.divmod_assign(denominator);
+  auto result = add_compare(numerator, numerator, denominator);
+  if (result > 0 || (result == 0 && (digit % 2) != 0)) {
+    if (digit == 9) {
+      const auto overflow = '0' + 10;
+      buf[num_digits - 1] = overflow;
+      // Propagate the carry.
+      for (int i = num_digits - 1; i > 0 && buf[i] == overflow; --i) {
+        buf[i] = '0';
+        ++buf[i - 1];
+      }
+      if (buf[0] == overflow) {
+        buf[0] = '1';
+        if ((flags & dragon::fixed) != 0)
+          buf.push_back('0');
+        else
+          ++exp10;
+      }
+      return;
+    }
+    ++digit;
+  }
+  buf[num_digits - 1] = static_cast<char>('0' + digit);
+}
+
+// Formats a floating-point number using the hexfloat format.
+template <typename Float, FMT_ENABLE_IF(!is_double_double<Float>::value)>
+FMT_CONSTEXPR20 void format_hexfloat(Float value, int precision,
+                                     float_specs specs, buffer<char>& buf) {
+  // float is passed as double to reduce the number of instantiations and to
+  // simplify implementation.
+  static_assert(!std::is_same<Float, float>::value, "");
+
+  using info = dragonbox::float_info<Float>;
+
+  // Assume Float is in the format [sign][exponent][significand].
+  using carrier_uint = typename info::carrier_uint;
+
+  constexpr auto num_float_significand_bits =
+      detail::num_significand_bits<Float>();
+
+  basic_fp<carrier_uint> f(value);
+  f.e += num_float_significand_bits;
+  if (!has_implicit_bit<Float>()) --f.e;
+
+  constexpr auto num_fraction_bits =
+      num_float_significand_bits + (has_implicit_bit<Float>() ? 1 : 0);
+  constexpr auto num_xdigits = (num_fraction_bits + 3) / 4;
+
+  constexpr auto leading_shift = ((num_xdigits - 1) * 4);
+  const auto leading_mask = carrier_uint(0xF) << leading_shift;
+  const auto leading_xdigit =
+      static_cast<uint32_t>((f.f & leading_mask) >> leading_shift);
+  if (leading_xdigit > 1) f.e -= (32 - countl_zero(leading_xdigit) - 1);
+
+  int print_xdigits = num_xdigits - 1;
+  if (precision >= 0 && print_xdigits > precision) {
+    const int shift = ((print_xdigits - precision - 1) * 4);
+    const auto mask = carrier_uint(0xF) << shift;
+    const auto v = static_cast<uint32_t>((f.f & mask) >> shift);
+
+    if (v >= 8) {
+      const auto inc = carrier_uint(1) << (shift + 4);
+      f.f += inc;
+      f.f &= ~(inc - 1);
+    }
+
+    // Check long double overflow
+    if (!has_implicit_bit<Float>()) {
+      const auto implicit_bit = carrier_uint(1) << num_float_significand_bits;
+      if ((f.f & implicit_bit) == implicit_bit) {
+        f.f >>= 4;
+        f.e += 4;
+      }
+    }
+
+    print_xdigits = precision;
+  }
+
+  char xdigits[num_bits<carrier_uint>() / 4];
+  detail::fill_n(xdigits, sizeof(xdigits), '0');
+  format_uint<4>(xdigits, f.f, num_xdigits, specs.upper);
+
+  // Remove zero tail
+  while (print_xdigits > 0 && xdigits[print_xdigits] == '0') --print_xdigits;
+
+  buf.push_back('0');
+  buf.push_back(specs.upper ? 'X' : 'x');
+  buf.push_back(xdigits[0]);
+  if (specs.showpoint || print_xdigits > 0 || print_xdigits < precision)
+    buf.push_back('.');
+  buf.append(xdigits + 1, xdigits + 1 + print_xdigits);
+  for (; print_xdigits < precision; ++print_xdigits) buf.push_back('0');
+
+  buf.push_back(specs.upper ? 'P' : 'p');
+
+  uint32_t abs_e;
+  if (f.e < 0) {
+    buf.push_back('-');
+    abs_e = static_cast<uint32_t>(-f.e);
+  } else {
+    buf.push_back('+');
+    abs_e = static_cast<uint32_t>(f.e);
+  }
+  format_decimal<char>(appender(buf), abs_e, detail::count_digits(abs_e));
+}
+
+template <typename Float, FMT_ENABLE_IF(is_double_double<Float>::value)>
+FMT_CONSTEXPR20 void format_hexfloat(Float value, int precision,
+                                     float_specs specs, buffer<char>& buf) {
+  format_hexfloat(static_cast<double>(value), precision, specs, buf);
+}
+
+constexpr auto fractional_part_rounding_thresholds(int index) -> uint32_t {
+  // For checking rounding thresholds.
+  // The kth entry is chosen to be the smallest integer such that the
+  // upper 32-bits of 10^(k+1) times it is strictly bigger than 5 * 10^k.
+  // It is equal to ceil(2^31 + 2^32/10^(k + 1)).
+  // These are stored in a string literal because we cannot have static arrays
+  // in constexpr functions and non-static ones are poorly optimized.
+  return U"\x9999999a\x828f5c29\x80418938\x80068db9\x8000a7c6\x800010c7"
+         U"\x800001ae\x8000002b"[index];
+}
+
+template <typename Float>
+FMT_CONSTEXPR20 auto format_float(Float value, int precision, float_specs specs,
+                                  buffer<char>& buf) -> int {
+  // float is passed as double to reduce the number of instantiations.
+  static_assert(!std::is_same<Float, float>::value, "");
+  FMT_ASSERT(value >= 0, "value is negative");
+  auto converted_value = convert_float(value);
+
+  const bool fixed = specs.format == float_format::fixed;
+  if (value <= 0) {  // <= instead of == to silence a warning.
+    if (precision <= 0 || !fixed) {
+      buf.push_back('0');
+      return 0;
+    }
+    buf.try_resize(to_unsigned(precision));
+    fill_n(buf.data(), precision, '0');
+    return -precision;
+  }
+
+  int exp = 0;
+  bool use_dragon = true;
+  unsigned dragon_flags = 0;
+  if (!is_fast_float<Float>() || is_constant_evaluated()) {
+    const auto inv_log2_10 = 0.3010299956639812;  // 1 / log2(10)
+    using info = dragonbox::float_info<decltype(converted_value)>;
+    const auto f = basic_fp<typename info::carrier_uint>(converted_value);
+    // Compute exp, an approximate power of 10, such that
+    //   10^(exp - 1) <= value < 10^exp or 10^exp <= value < 10^(exp + 1).
+    // This is based on log10(value) == log2(value) / log2(10) and approximation
+    // of log2(value) by e + num_fraction_bits idea from double-conversion.
+    auto e = (f.e + count_digits<1>(f.f) - 1) * inv_log2_10 - 1e-10;
+    exp = static_cast<int>(e);
+    if (e > exp) ++exp;  // Compute ceil.
+    dragon_flags = dragon::fixup;
+  } else if (precision < 0) {
+    // Use Dragonbox for the shortest format.
+    if (specs.binary32) {
+      auto dec = dragonbox::to_decimal(static_cast<float>(value));
+      write<char>(buffer_appender<char>(buf), dec.significand);
+      return dec.exponent;
+    }
+    auto dec = dragonbox::to_decimal(static_cast<double>(value));
+    write<char>(buffer_appender<char>(buf), dec.significand);
+    return dec.exponent;
+  } else {
+    // Extract significand bits and exponent bits.
+    using info = dragonbox::float_info<double>;
+    auto br = bit_cast<uint64_t>(static_cast<double>(value));
+
+    const uint64_t significand_mask =
+        (static_cast<uint64_t>(1) << num_significand_bits<double>()) - 1;
+    uint64_t significand = (br & significand_mask);
+    int exponent = static_cast<int>((br & exponent_mask<double>()) >>
+                                    num_significand_bits<double>());
+
+    if (exponent != 0) {  // Check if normal.
+      exponent -= exponent_bias<double>() + num_significand_bits<double>();
+      significand |=
+          (static_cast<uint64_t>(1) << num_significand_bits<double>());
+      significand <<= 1;
+    } else {
+      // Normalize subnormal inputs.
+      FMT_ASSERT(significand != 0, "zeros should not appear here");
+      int shift = countl_zero(significand);
+      FMT_ASSERT(shift >= num_bits<uint64_t>() - num_significand_bits<double>(),
+                 "");
+      shift -= (num_bits<uint64_t>() - num_significand_bits<double>() - 2);
+      exponent = (std::numeric_limits<double>::min_exponent -
+                  num_significand_bits<double>()) -
+                 shift;
+      significand <<= shift;
+    }
+
+    // Compute the first several nonzero decimal significand digits.
+    // We call the number we get the first segment.
+    const int k = info::kappa - dragonbox::floor_log10_pow2(exponent);
+    exp = -k;
+    const int beta = exponent + dragonbox::floor_log2_pow10(k);
+    uint64_t first_segment;
+    bool has_more_segments;
+    int digits_in_the_first_segment;
+    {
+      const auto r = dragonbox::umul192_upper128(
+          significand << beta, dragonbox::get_cached_power(k));
+      first_segment = r.high();
+      has_more_segments = r.low() != 0;
+
+      // The first segment can have 18 ~ 19 digits.
+      if (first_segment >= 1000000000000000000ULL) {
+        digits_in_the_first_segment = 19;
+      } else {
+        // When it is of 18-digits, we align it to 19-digits by adding a bogus
+        // zero at the end.
+        digits_in_the_first_segment = 18;
+        first_segment *= 10;
+      }
+    }
+
+    // Compute the actual number of decimal digits to print.
+    if (fixed) adjust_precision(precision, exp + digits_in_the_first_segment);
+
+    // Use Dragon4 only when there might be not enough digits in the first
+    // segment.
+    if (digits_in_the_first_segment > precision) {
+      use_dragon = false;
+
+      if (precision <= 0) {
+        exp += digits_in_the_first_segment;
+
+        if (precision < 0) {
+          // Nothing to do, since all we have are just leading zeros.
+          buf.try_resize(0);
+        } else {
+          // We may need to round-up.
+          buf.try_resize(1);
+          if ((first_segment | static_cast<uint64_t>(has_more_segments)) >
+              5000000000000000000ULL) {
+            buf[0] = '1';
+          } else {
+            buf[0] = '0';
+          }
+        }
+      }  // precision <= 0
+      else {
+        exp += digits_in_the_first_segment - precision;
+
+        // When precision > 0, we divide the first segment into three
+        // subsegments, each with 9, 9, and 0 ~ 1 digits so that each fits
+        // in 32-bits which usually allows faster calculation than in
+        // 64-bits. Since some compiler (e.g. MSVC) doesn't know how to optimize
+        // division-by-constant for large 64-bit divisors, we do it here
+        // manually. The magic number 7922816251426433760 below is equal to
+        // ceil(2^(64+32) / 10^10).
+        const uint32_t first_subsegment = static_cast<uint32_t>(
+            dragonbox::umul128_upper64(first_segment, 7922816251426433760ULL) >>
+            32);
+        const uint64_t second_third_subsegments =
+            first_segment - first_subsegment * 10000000000ULL;
+
+        uint64_t prod;
+        uint32_t digits;
+        bool should_round_up;
+        int number_of_digits_to_print = precision > 9 ? 9 : precision;
+
+        // Print a 9-digits subsegment, either the first or the second.
+        auto print_subsegment = [&](uint32_t subsegment, char* buffer) {
+          int number_of_digits_printed = 0;
+
+          // If we want to print an odd number of digits from the subsegment,
+          if ((number_of_digits_to_print & 1) != 0) {
+            // Convert to 64-bit fixed-point fractional form with 1-digit
+            // integer part. The magic number 720575941 is a good enough
+            // approximation of 2^(32 + 24) / 10^8; see
+            // https://jk-jeon.github.io/posts/2022/12/fixed-precision-formatting/#fixed-length-case
+            // for details.
+            prod = ((subsegment * static_cast<uint64_t>(720575941)) >> 24) + 1;
+            digits = static_cast<uint32_t>(prod >> 32);
+            *buffer = static_cast<char>('0' + digits);
+            number_of_digits_printed++;
+          }
+          // If we want to print an even number of digits from the
+          // first_subsegment,
+          else {
+            // Convert to 64-bit fixed-point fractional form with 2-digits
+            // integer part. The magic number 450359963 is a good enough
+            // approximation of 2^(32 + 20) / 10^7; see
+            // https://jk-jeon.github.io/posts/2022/12/fixed-precision-formatting/#fixed-length-case
+            // for details.
+            prod = ((subsegment * static_cast<uint64_t>(450359963)) >> 20) + 1;
+            digits = static_cast<uint32_t>(prod >> 32);
+            copy2(buffer, digits2(digits));
+            number_of_digits_printed += 2;
+          }
+
+          // Print all digit pairs.
+          while (number_of_digits_printed < number_of_digits_to_print) {
+            prod = static_cast<uint32_t>(prod) * static_cast<uint64_t>(100);
+            digits = static_cast<uint32_t>(prod >> 32);
+            copy2(buffer + number_of_digits_printed, digits2(digits));
+            number_of_digits_printed += 2;
+          }
+        };
+
+        // Print first subsegment.
+        print_subsegment(first_subsegment, buf.data());
+
+        // Perform rounding if the first subsegment is the last subsegment to
+        // print.
+        if (precision <= 9) {
+          // Rounding inside the subsegment.
+          // We round-up if:
+          //  - either the fractional part is strictly larger than 1/2, or
+          //  - the fractional part is exactly 1/2 and the last digit is odd.
+          // We rely on the following observations:
+          //  - If fractional_part >= threshold, then the fractional part is
+          //    strictly larger than 1/2.
+          //  - If the MSB of fractional_part is set, then the fractional part
+          //    must be at least 1/2.
+          //  - When the MSB of fractional_part is set, either
+          //    second_third_subsegments being nonzero or has_more_segments
+          //    being true means there are further digits not printed, so the
+          //    fractional part is strictly larger than 1/2.
+          if (precision < 9) {
+            uint32_t fractional_part = static_cast<uint32_t>(prod);
+            should_round_up =
+                fractional_part >= fractional_part_rounding_thresholds(
+                                       8 - number_of_digits_to_print) ||
+                ((fractional_part >> 31) &
+                 ((digits & 1) | (second_third_subsegments != 0) |
+                  has_more_segments)) != 0;
+          }
+          // Rounding at the subsegment boundary.
+          // In this case, the fractional part is at least 1/2 if and only if
+          // second_third_subsegments >= 5000000000ULL, and is strictly larger
+          // than 1/2 if we further have either second_third_subsegments >
+          // 5000000000ULL or has_more_segments == true.
+          else {
+            should_round_up = second_third_subsegments > 5000000000ULL ||
+                              (second_third_subsegments == 5000000000ULL &&
+                               ((digits & 1) != 0 || has_more_segments));
+          }
+        }
+        // Otherwise, print the second subsegment.
+        else {
+          // Compilers are not aware of how to leverage the maximum value of
+          // second_third_subsegments to find out a better magic number which
+          // allows us to eliminate an additional shift. 1844674407370955162 =
+          // ceil(2^64/10) < ceil(2^64*(10^9/(10^10 - 1))).
+          const uint32_t second_subsegment =
+              static_cast<uint32_t>(dragonbox::umul128_upper64(
+                  second_third_subsegments, 1844674407370955162ULL));
+          const uint32_t third_subsegment =
+              static_cast<uint32_t>(second_third_subsegments) -
+              second_subsegment * 10;
+
+          number_of_digits_to_print = precision - 9;
+          print_subsegment(second_subsegment, buf.data() + 9);
+
+          // Rounding inside the subsegment.
+          if (precision < 18) {
+            // The condition third_subsegment != 0 implies that the segment was
+            // of 19 digits, so in this case the third segment should be
+            // consisting of a genuine digit from the input.
+            uint32_t fractional_part = static_cast<uint32_t>(prod);
+            should_round_up =
+                fractional_part >= fractional_part_rounding_thresholds(
+                                       8 - number_of_digits_to_print) ||
+                ((fractional_part >> 31) &
+                 ((digits & 1) | (third_subsegment != 0) |
+                  has_more_segments)) != 0;
+          }
+          // Rounding at the subsegment boundary.
+          else {
+            // In this case, the segment must be of 19 digits, thus
+            // the third subsegment should be consisting of a genuine digit from
+            // the input.
+            should_round_up = third_subsegment > 5 ||
+                              (third_subsegment == 5 &&
+                               ((digits & 1) != 0 || has_more_segments));
+          }
+        }
+
+        // Round-up if necessary.
+        if (should_round_up) {
+          ++buf[precision - 1];
+          for (int i = precision - 1; i > 0 && buf[i] > '9'; --i) {
+            buf[i] = '0';
+            ++buf[i - 1];
+          }
+          if (buf[0] > '9') {
+            buf[0] = '1';
+            if (fixed)
+              buf[precision++] = '0';
+            else
+              ++exp;
+          }
+        }
+        buf.try_resize(to_unsigned(precision));
+      }
+    }  // if (digits_in_the_first_segment > precision)
+    else {
+      // Adjust the exponent for its use in Dragon4.
+      exp += digits_in_the_first_segment - 1;
+    }
+  }
+  if (use_dragon) {
+    auto f = basic_fp<uint128_t>();
+    bool is_predecessor_closer = specs.binary32
+                                     ? f.assign(static_cast<float>(value))
+                                     : f.assign(converted_value);
+    if (is_predecessor_closer) dragon_flags |= dragon::predecessor_closer;
+    if (fixed) dragon_flags |= dragon::fixed;
+    // Limit precision to the maximum possible number of significant digits in
+    // an IEEE754 double because we don't need to generate zeros.
+    const int max_double_digits = 767;
+    if (precision > max_double_digits) precision = max_double_digits;
+    format_dragon(f, dragon_flags, precision, buf, exp);
+  }
+  if (!fixed && !specs.showpoint) {
+    // Remove trailing zeros.
+    auto num_digits = buf.size();
+    while (num_digits > 0 && buf[num_digits - 1] == '0') {
+      --num_digits;
+      ++exp;
+    }
+    buf.try_resize(num_digits);
+  }
+  return exp;
+}
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR20 auto write_float(OutputIt out, T value,
+                                 format_specs<Char> specs, locale_ref loc)
+    -> OutputIt {
+  float_specs fspecs = parse_float_type_spec(specs);
+  fspecs.sign = specs.sign;
+  if (detail::signbit(value)) {  // value < 0 is false for NaN so use signbit.
+    fspecs.sign = sign::minus;
+    value = -value;
+  } else if (fspecs.sign == sign::minus) {
+    fspecs.sign = sign::none;
+  }
+
+  if (!detail::isfinite(value))
+    return write_nonfinite(out, detail::isnan(value), specs, fspecs);
+
+  if (specs.align == align::numeric && fspecs.sign) {
+    auto it = reserve(out, 1);
+    *it++ = detail::sign<Char>(fspecs.sign);
+    out = base_iterator(out, it);
+    fspecs.sign = sign::none;
+    if (specs.width != 0) --specs.width;
+  }
+
+  memory_buffer buffer;
+  if (fspecs.format == float_format::hex) {
+    if (fspecs.sign) buffer.push_back(detail::sign<char>(fspecs.sign));
+    format_hexfloat(convert_float(value), specs.precision, fspecs, buffer);
+    return write_bytes<align::right>(out, {buffer.data(), buffer.size()},
+                                     specs);
+  }
+  int precision = specs.precision >= 0 || specs.type == presentation_type::none
+                      ? specs.precision
+                      : 6;
+  if (fspecs.format == float_format::exp) {
+    if (precision == max_value<int>())
+      throw_format_error("number is too big");
+    else
+      ++precision;
+  } else if (fspecs.format != float_format::fixed && precision == 0) {
+    precision = 1;
+  }
+  if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
+  int exp = format_float(convert_float(value), precision, fspecs, buffer);
+  fspecs.precision = precision;
+  auto f = big_decimal_fp{buffer.data(), static_cast<int>(buffer.size()), exp};
+  return write_float(out, f, specs, fspecs, loc);
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_floating_point<T>::value)>
+FMT_CONSTEXPR20 auto write(OutputIt out, T value, format_specs<Char> specs,
+                           locale_ref loc = {}) -> OutputIt {
+  if (const_check(!is_supported_floating_point(value))) return out;
+  return specs.localized && write_loc(out, value, specs, loc)
+             ? out
+             : write_float(out, value, specs, loc);
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_fast_float<T>::value)>
+FMT_CONSTEXPR20 auto write(OutputIt out, T value) -> OutputIt {
+  if (is_constant_evaluated()) return write(out, value, format_specs<Char>());
+  if (const_check(!is_supported_floating_point(value))) return out;
+
+  auto fspecs = float_specs();
+  if (detail::signbit(value)) {
+    fspecs.sign = sign::minus;
+    value = -value;
+  }
+
+  constexpr auto specs = format_specs<Char>();
+  using floaty = conditional_t<std::is_same<T, long double>::value, double, T>;
+  using floaty_uint = typename dragonbox::float_info<floaty>::carrier_uint;
+  floaty_uint mask = exponent_mask<floaty>();
+  if ((bit_cast<floaty_uint>(value) & mask) == mask)
+    return write_nonfinite(out, std::isnan(value), specs, fspecs);
+
+  auto dec = dragonbox::to_decimal(static_cast<floaty>(value));
+  return write_float(out, dec, specs, fspecs, {});
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_floating_point<T>::value &&
+                        !is_fast_float<T>::value)>
+inline auto write(OutputIt out, T value) -> OutputIt {
+  return write(out, value, format_specs<Char>());
+}
+
+template <typename Char, typename OutputIt>
+auto write(OutputIt out, monostate, format_specs<Char> = {}, locale_ref = {})
+    -> OutputIt {
+  FMT_ASSERT(false, "");
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, basic_string_view<Char> value)
+    -> OutputIt {
+  auto it = reserve(out, value.size());
+  it = copy_str_noinline<Char>(value.begin(), value.end(), it);
+  return base_iterator(out, it);
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_string<T>::value)>
+constexpr auto write(OutputIt out, const T& value) -> OutputIt {
+  return write<Char>(out, to_string_view(value));
+}
+
+// FMT_ENABLE_IF() condition separated to workaround an MSVC bug.
+template <
+    typename Char, typename OutputIt, typename T,
+    bool check =
+        std::is_enum<T>::value && !std::is_same<T, Char>::value &&
+        mapped_type_constant<T, basic_format_context<OutputIt, Char>>::value !=
+            type::custom_type,
+    FMT_ENABLE_IF(check)>
+FMT_CONSTEXPR auto write(OutputIt out, T value) -> OutputIt {
+  return write<Char>(out, static_cast<underlying_t<T>>(value));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_same<T, bool>::value)>
+FMT_CONSTEXPR auto write(OutputIt out, T value,
+                         const format_specs<Char>& specs = {}, locale_ref = {})
+    -> OutputIt {
+  return specs.type != presentation_type::none &&
+                 specs.type != presentation_type::string
+             ? write(out, value ? 1 : 0, specs, {})
+             : write_bytes(out, value ? "true" : "false", specs);
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, Char value) -> OutputIt {
+  auto it = reserve(out, 1);
+  *it++ = value;
+  return base_iterator(out, it);
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR_CHAR_TRAITS auto write(OutputIt out, const Char* value)
+    -> OutputIt {
+  if (value) return write(out, basic_string_view<Char>(value));
+  throw_format_error("string pointer is null");
+  return out;
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_same<T, void>::value)>
+auto write(OutputIt out, const T* value, const format_specs<Char>& specs = {},
+           locale_ref = {}) -> OutputIt {
+  return write_ptr<Char>(out, bit_cast<uintptr_t>(value), &specs);
+}
+
+// A write overload that handles implicit conversions.
+template <typename Char, typename OutputIt, typename T,
+          typename Context = basic_format_context<OutputIt, Char>>
+FMT_CONSTEXPR auto write(OutputIt out, const T& value) -> enable_if_t<
+    std::is_class<T>::value && !is_string<T>::value &&
+        !is_floating_point<T>::value && !std::is_same<T, Char>::value &&
+        !std::is_same<T, remove_cvref_t<decltype(arg_mapper<Context>().map(
+                             value))>>::value,
+    OutputIt> {
+  return write<Char>(out, arg_mapper<Context>().map(value));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          typename Context = basic_format_context<OutputIt, Char>>
+FMT_CONSTEXPR auto write(OutputIt out, const T& value)
+    -> enable_if_t<mapped_type_constant<T, Context>::value == type::custom_type,
+                   OutputIt> {
+  auto formatter = typename Context::template formatter_type<T>();
+  auto parse_ctx = typename Context::parse_context_type({});
+  formatter.parse(parse_ctx);
+  auto ctx = Context(out, {}, {});
+  return formatter.format(value, ctx);
+}
+
+// An argument visitor that formats the argument and writes it via the output
+// iterator. It's a class and not a generic lambda for compatibility with C++11.
+template <typename Char> struct default_arg_formatter {
+  using iterator = buffer_appender<Char>;
+  using context = buffer_context<Char>;
+
+  iterator out;
+  basic_format_args<context> args;
+  locale_ref loc;
+
+  template <typename T> auto operator()(T value) -> iterator {
+    return write<Char>(out, value);
+  }
+  auto operator()(typename basic_format_arg<context>::handle h) -> iterator {
+    basic_format_parse_context<Char> parse_ctx({});
+    context format_ctx(out, args, loc);
+    h.format(parse_ctx, format_ctx);
+    return format_ctx.out();
+  }
+};
+
+template <typename Char> struct arg_formatter {
+  using iterator = buffer_appender<Char>;
+  using context = buffer_context<Char>;
+
+  iterator out;
+  const format_specs<Char>& specs;
+  locale_ref locale;
+
+  template <typename T>
+  FMT_CONSTEXPR FMT_INLINE auto operator()(T value) -> iterator {
+    return detail::write(out, value, specs, locale);
+  }
+  auto operator()(typename basic_format_arg<context>::handle) -> iterator {
+    // User-defined types are handled separately because they require access
+    // to the parse context.
+    return out;
+  }
+};
+
+struct width_checker {
+  template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T value) -> unsigned long long {
+    if (is_negative(value)) throw_format_error("negative width");
+    return static_cast<unsigned long long>(value);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T) -> unsigned long long {
+    throw_format_error("width is not integer");
+    return 0;
+  }
+};
+
+struct precision_checker {
+  template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T value) -> unsigned long long {
+    if (is_negative(value)) throw_format_error("negative precision");
+    return static_cast<unsigned long long>(value);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T) -> unsigned long long {
+    throw_format_error("precision is not integer");
+    return 0;
+  }
+};
+
+template <typename Handler, typename FormatArg>
+FMT_CONSTEXPR auto get_dynamic_spec(FormatArg arg) -> int {
+  unsigned long long value = visit_format_arg(Handler(), arg);
+  if (value > to_unsigned(max_value<int>()))
+    throw_format_error("number is too big");
+  return static_cast<int>(value);
+}
+
+template <typename Context, typename ID>
+FMT_CONSTEXPR auto get_arg(Context& ctx, ID id) -> decltype(ctx.arg(id)) {
+  auto arg = ctx.arg(id);
+  if (!arg) ctx.on_error("argument not found");
+  return arg;
+}
+
+template <typename Handler, typename Context>
+FMT_CONSTEXPR void handle_dynamic_spec(int& value,
+                                       arg_ref<typename Context::char_type> ref,
+                                       Context& ctx) {
+  switch (ref.kind) {
+  case arg_id_kind::none:
+    break;
+  case arg_id_kind::index:
+    value = detail::get_dynamic_spec<Handler>(get_arg(ctx, ref.val.index));
+    break;
+  case arg_id_kind::name:
+    value = detail::get_dynamic_spec<Handler>(get_arg(ctx, ref.val.name));
+    break;
+  }
+}
+
+#if FMT_USE_USER_DEFINED_LITERALS
+#  if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct statically_named_arg : view {
+  static constexpr auto name = Str.data;
+
+  const T& value;
+  statically_named_arg(const T& v) : value(v) {}
+};
+
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct is_named_arg<statically_named_arg<T, Char, N, Str>> : std::true_type {};
+
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct is_statically_named_arg<statically_named_arg<T, Char, N, Str>>
+    : std::true_type {};
+
+template <typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct udl_arg {
+  template <typename T> auto operator=(T&& value) const {
+    return statically_named_arg<T, Char, N, Str>(std::forward<T>(value));
+  }
+};
+#  else
+template <typename Char> struct udl_arg {
+  const Char* str;
+
+  template <typename T> auto operator=(T&& value) const -> named_arg<Char, T> {
+    return {str, std::forward<T>(value)};
+  }
+};
+#  endif
+#endif  // FMT_USE_USER_DEFINED_LITERALS
+
+template <typename Locale, typename Char>
+auto vformat(const Locale& loc, basic_string_view<Char> fmt,
+             basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> std::basic_string<Char> {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vformat_to(buf, fmt, args, detail::locale_ref(loc));
+  return {buf.data(), buf.size()};
+}
+
+using format_func = void (*)(detail::buffer<char>&, int, const char*);
+
+FMT_API void format_error_code(buffer<char>& out, int error_code,
+                               string_view message) noexcept;
+
+FMT_API void report_error(format_func func, int error_code,
+                          const char* message) noexcept;
+}  // namespace detail
+
+FMT_API auto vsystem_error(int error_code, string_view format_str,
+                           format_args args) -> std::system_error;
+
+/**
+  \rst
+  Constructs :class:`std::system_error` with a message formatted with
+  ``fmt::format(fmt, args...)``.
+  *error_code* is a system error code as given by ``errno``.
+
+  **Example**::
+
+    // This throws std::system_error with the description
+    //   cannot open file 'madeup': No such file or directory
+    // or similar (system message may vary).
+    const char* filename = "madeup";
+    std::FILE* file = std::fopen(filename, "r");
+    if (!file)
+      throw fmt::system_error(errno, "cannot open file '{}'", filename);
+  \endrst
+ */
+template <typename... T>
+auto system_error(int error_code, format_string<T...> fmt, T&&... args)
+    -> std::system_error {
+  return vsystem_error(error_code, fmt, fmt::make_format_args(args...));
+}
+
+/**
+  \rst
+  Formats an error message for an error returned by an operating system or a
+  language runtime, for example a file opening error, and writes it to *out*.
+  The format is the same as the one used by ``std::system_error(ec, message)``
+  where ``ec`` is ``std::error_code(error_code, std::generic_category()})``.
+  It is implementation-defined but normally looks like:
+
+  .. parsed-literal::
+     *<message>*: *<system-message>*
+
+  where *<message>* is the passed message and *<system-message>* is the system
+  message corresponding to the error code.
+  *error_code* is a system error code as given by ``errno``.
+  \endrst
+ */
+FMT_API void format_system_error(detail::buffer<char>& out, int error_code,
+                                 const char* message) noexcept;
+
+// Reports a system error without throwing an exception.
+// Can be used to report errors from destructors.
+FMT_API void report_system_error(int error_code, const char* message) noexcept;
+
+/** Fast integer formatter. */
+class format_int {
+ private:
+  // Buffer should be large enough to hold all digits (digits10 + 1),
+  // a sign and a null character.
+  enum { buffer_size = std::numeric_limits<unsigned long long>::digits10 + 3 };
+  mutable char buffer_[buffer_size];
+  char* str_;
+
+  template <typename UInt> auto format_unsigned(UInt value) -> char* {
+    auto n = static_cast<detail::uint32_or_64_or_128_t<UInt>>(value);
+    return detail::format_decimal(buffer_, n, buffer_size - 1).begin;
+  }
+
+  template <typename Int> auto format_signed(Int value) -> char* {
+    auto abs_value = static_cast<detail::uint32_or_64_or_128_t<Int>>(value);
+    bool negative = value < 0;
+    if (negative) abs_value = 0 - abs_value;
+    auto begin = format_unsigned(abs_value);
+    if (negative) *--begin = '-';
+    return begin;
+  }
+
+ public:
+  explicit format_int(int value) : str_(format_signed(value)) {}
+  explicit format_int(long value) : str_(format_signed(value)) {}
+  explicit format_int(long long value) : str_(format_signed(value)) {}
+  explicit format_int(unsigned value) : str_(format_unsigned(value)) {}
+  explicit format_int(unsigned long value) : str_(format_unsigned(value)) {}
+  explicit format_int(unsigned long long value)
+      : str_(format_unsigned(value)) {}
+
+  /** Returns the number of characters written to the output buffer. */
+  auto size() const -> size_t {
+    return detail::to_unsigned(buffer_ - str_ + buffer_size - 1);
+  }
+
+  /**
+    Returns a pointer to the output buffer content. No terminating null
+    character is appended.
+   */
+  auto data() const -> const char* { return str_; }
+
+  /**
+    Returns a pointer to the output buffer content with terminating null
+    character appended.
+   */
+  auto c_str() const -> const char* {
+    buffer_[buffer_size - 1] = '\0';
+    return str_;
+  }
+
+  /**
+    \rst
+    Returns the content of the output buffer as an ``std::string``.
+    \endrst
+   */
+  auto str() const -> std::string { return std::string(str_, size()); }
+};
+
+template <typename T, typename Char>
+struct formatter<T, Char, enable_if_t<detail::has_format_as<T>::value>>
+    : formatter<detail::format_as_t<T>, Char> {
+  template <typename FormatContext>
+  auto format(const T& value, FormatContext& ctx) const -> decltype(ctx.out()) {
+    using base = formatter<detail::format_as_t<T>, Char>;
+    return base::format(format_as(value), ctx);
+  }
+};
+
+#define FMT_FORMAT_AS(Type, Base) \
+  template <typename Char>        \
+  struct formatter<Type, Char> : formatter<Base, Char> {}
+
+FMT_FORMAT_AS(signed char, int);
+FMT_FORMAT_AS(unsigned char, unsigned);
+FMT_FORMAT_AS(short, int);
+FMT_FORMAT_AS(unsigned short, unsigned);
+FMT_FORMAT_AS(long, detail::long_type);
+FMT_FORMAT_AS(unsigned long, detail::ulong_type);
+FMT_FORMAT_AS(Char*, const Char*);
+FMT_FORMAT_AS(std::basic_string<Char>, basic_string_view<Char>);
+FMT_FORMAT_AS(std::nullptr_t, const void*);
+FMT_FORMAT_AS(detail::std_string_view<Char>, basic_string_view<Char>);
+FMT_FORMAT_AS(void*, const void*);
+
+template <typename Char, size_t N>
+struct formatter<Char[N], Char> : formatter<basic_string_view<Char>, Char> {};
+
+/**
+  \rst
+  Converts ``p`` to ``const void*`` for pointer formatting.
+
+  **Example**::
+
+    auto s = fmt::format("{}", fmt::ptr(p));
+  \endrst
+ */
+template <typename T> auto ptr(T p) -> const void* {
+  static_assert(std::is_pointer<T>::value, "");
+  return detail::bit_cast<const void*>(p);
+}
+template <typename T, typename Deleter>
+auto ptr(const std::unique_ptr<T, Deleter>& p) -> const void* {
+  return p.get();
+}
+template <typename T> auto ptr(const std::shared_ptr<T>& p) -> const void* {
+  return p.get();
+}
+
+/**
+  \rst
+  Converts ``e`` to the underlying type.
+
+  **Example**::
+
+    enum class color { red, green, blue };
+    auto s = fmt::format("{}", fmt::underlying(color::red));
+  \endrst
+ */
+template <typename Enum>
+constexpr auto underlying(Enum e) noexcept -> underlying_t<Enum> {
+  return static_cast<underlying_t<Enum>>(e);
+}
+
+namespace enums {
+template <typename Enum, FMT_ENABLE_IF(std::is_enum<Enum>::value)>
+constexpr auto format_as(Enum e) noexcept -> underlying_t<Enum> {
+  return static_cast<underlying_t<Enum>>(e);
+}
+}  // namespace enums
+
+class bytes {
+ private:
+  string_view data_;
+  friend struct formatter<bytes>;
+
+ public:
+  explicit bytes(string_view data) : data_(data) {}
+};
+
+template <> struct formatter<bytes> {
+ private:
+  detail::dynamic_format_specs<> specs_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const char* {
+    return parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx,
+                              detail::type::string_type);
+  }
+
+  template <typename FormatContext>
+  auto format(bytes b, FormatContext& ctx) -> decltype(ctx.out()) {
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+    return detail::write_bytes(ctx.out(), b.data_, specs_);
+  }
+};
+
+// group_digits_view is not derived from view because it copies the argument.
+template <typename T> struct group_digits_view {
+  T value;
+};
+
+/**
+  \rst
+  Returns a view that formats an integer value using ',' as a locale-independent
+  thousands separator.
+
+  **Example**::
+
+    fmt::print("{}", fmt::group_digits(12345));
+    // Output: "12,345"
+  \endrst
+ */
+template <typename T> auto group_digits(T value) -> group_digits_view<T> {
+  return {value};
+}
+
+template <typename T> struct formatter<group_digits_view<T>> : formatter<T> {
+ private:
+  detail::dynamic_format_specs<> specs_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const char* {
+    return parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx,
+                              detail::type::int_type);
+  }
+
+  template <typename FormatContext>
+  auto format(group_digits_view<T> t, FormatContext& ctx)
+      -> decltype(ctx.out()) {
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+    return detail::write_int(
+        ctx.out(), static_cast<detail::uint64_or_128_t<T>>(t.value), 0, specs_,
+        detail::digit_grouping<char>("\3", ","));
+  }
+};
+
+template <typename T> struct nested_view {
+  const formatter<T>* fmt;
+  const T* value;
+};
+
+template <typename T> struct formatter<nested_view<T>> {
+  FMT_CONSTEXPR auto parse(format_parse_context& ctx) -> const char* {
+    return ctx.begin();
+  }
+  auto format(nested_view<T> view, format_context& ctx) const
+      -> decltype(ctx.out()) {
+    return view.fmt->format(*view.value, ctx);
+  }
+};
+
+template <typename T> struct nested_formatter {
+ private:
+  int width_;
+  detail::fill_t<char> fill_;
+  align_t align_ : 4;
+  formatter<T> formatter_;
+
+ public:
+  constexpr nested_formatter() : width_(0), align_(align_t::none) {}
+
+  FMT_CONSTEXPR auto parse(format_parse_context& ctx) -> const char* {
+    auto specs = detail::dynamic_format_specs<char>();
+    auto it = parse_format_specs(ctx.begin(), ctx.end(), specs, ctx,
+                                 detail::type::none_type);
+    width_ = specs.width;
+    fill_ = specs.fill;
+    align_ = specs.align;
+    ctx.advance_to(it);
+    return formatter_.parse(ctx);
+  }
+
+  template <typename F>
+  auto write_padded(format_context& ctx, F write) const -> decltype(ctx.out()) {
+    if (width_ == 0) return write(ctx.out());
+    auto buf = memory_buffer();
+    write(std::back_inserter(buf));
+    auto specs = format_specs<>();
+    specs.width = width_;
+    specs.fill = fill_;
+    specs.align = align_;
+    return detail::write(ctx.out(), string_view(buf.data(), buf.size()), specs);
+  }
+
+  auto nested(const T& value) const -> nested_view<T> {
+    return nested_view<T>{&formatter_, &value};
+  }
+};
+
+// DEPRECATED! join_view will be moved to ranges.h.
+template <typename It, typename Sentinel, typename Char = char>
+struct join_view : detail::view {
+  It begin;
+  Sentinel end;
+  basic_string_view<Char> sep;
+
+  join_view(It b, Sentinel e, basic_string_view<Char> s)
+      : begin(b), end(e), sep(s) {}
+};
+
+template <typename It, typename Sentinel, typename Char>
+struct formatter<join_view<It, Sentinel, Char>, Char> {
+ private:
+  using value_type =
+#ifdef __cpp_lib_ranges
+      std::iter_value_t<It>;
+#else
+      typename std::iterator_traits<It>::value_type;
+#endif
+  formatter<remove_cvref_t<value_type>, Char> value_formatter_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const Char* {
+    return value_formatter_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(const join_view<It, Sentinel, Char>& value,
+              FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto it = value.begin;
+    auto out = ctx.out();
+    if (it != value.end) {
+      out = value_formatter_.format(*it, ctx);
+      ++it;
+      while (it != value.end) {
+        out = detail::copy_str<Char>(value.sep.begin(), value.sep.end(), out);
+        ctx.advance_to(out);
+        out = value_formatter_.format(*it, ctx);
+        ++it;
+      }
+    }
+    return out;
+  }
+};
+
+/**
+  Returns a view that formats the iterator range `[begin, end)` with elements
+  separated by `sep`.
+ */
+template <typename It, typename Sentinel>
+auto join(It begin, Sentinel end, string_view sep) -> join_view<It, Sentinel> {
+  return {begin, end, sep};
+}
+
+/**
+  \rst
+  Returns a view that formats `range` with elements separated by `sep`.
+
+  **Example**::
+
+    std::vector<int> v = {1, 2, 3};
+    fmt::print("{}", fmt::join(v, ", "));
+    // Output: "1, 2, 3"
+
+  ``fmt::join`` applies passed format specifiers to the range elements::
+
+    fmt::print("{:02}", fmt::join(v, ", "));
+    // Output: "01, 02, 03"
+  \endrst
+ */
+template <typename Range>
+auto join(Range&& range, string_view sep)
+    -> join_view<detail::iterator_t<Range>, detail::sentinel_t<Range>> {
+  return join(std::begin(range), std::end(range), sep);
+}
+
+/**
+  \rst
+  Converts *value* to ``std::string`` using the default format for type *T*.
+
+  **Example**::
+
+    #include <fmt/format.h>
+
+    std::string answer = fmt::to_string(42);
+  \endrst
+ */
+template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value &&
+                                    !detail::has_format_as<T>::value)>
+inline auto to_string(const T& value) -> std::string {
+  auto buffer = memory_buffer();
+  detail::write<char>(appender(buffer), value);
+  return {buffer.data(), buffer.size()};
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+FMT_NODISCARD inline auto to_string(T value) -> std::string {
+  // The buffer should be large enough to store the number including the sign
+  // or "false" for bool.
+  constexpr int max_size = detail::digits10<T>() + 2;
+  char buffer[max_size > 5 ? static_cast<unsigned>(max_size) : 5];
+  char* begin = buffer;
+  return std::string(begin, detail::write<char>(begin, value));
+}
+
+template <typename Char, size_t SIZE>
+FMT_NODISCARD auto to_string(const basic_memory_buffer<Char, SIZE>& buf)
+    -> std::basic_string<Char> {
+  auto size = buf.size();
+  detail::assume(size < std::basic_string<Char>().max_size());
+  return std::basic_string<Char>(buf.data(), size);
+}
+
+template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value &&
+                                    detail::has_format_as<T>::value)>
+inline auto to_string(const T& value) -> std::string {
+  return to_string(format_as(value));
+}
+
+FMT_END_EXPORT
+
+namespace detail {
+
+template <typename Char>
+void vformat_to(buffer<Char>& buf, basic_string_view<Char> fmt,
+                typename vformat_args<Char>::type args, locale_ref loc) {
+  auto out = buffer_appender<Char>(buf);
+  if (fmt.size() == 2 && equal2(fmt.data(), "{}")) {
+    auto arg = args.get(0);
+    if (!arg) throw_format_error("argument not found");
+    visit_format_arg(default_arg_formatter<Char>{out, args, loc}, arg);
+    return;
+  }
+
+  struct format_handler : error_handler {
+    basic_format_parse_context<Char> parse_context;
+    buffer_context<Char> context;
+
+    format_handler(buffer_appender<Char> p_out, basic_string_view<Char> str,
+                   basic_format_args<buffer_context<Char>> p_args,
+                   locale_ref p_loc)
+        : parse_context(str), context(p_out, p_args, p_loc) {}
+
+    void on_text(const Char* begin, const Char* end) {
+      auto text = basic_string_view<Char>(begin, to_unsigned(end - begin));
+      context.advance_to(write<Char>(context.out(), text));
+    }
+
+    FMT_CONSTEXPR auto on_arg_id() -> int {
+      return parse_context.next_arg_id();
+    }
+    FMT_CONSTEXPR auto on_arg_id(int id) -> int {
+      return parse_context.check_arg_id(id), id;
+    }
+    FMT_CONSTEXPR auto on_arg_id(basic_string_view<Char> id) -> int {
+      int arg_id = context.arg_id(id);
+      if (arg_id < 0) throw_format_error("argument not found");
+      return arg_id;
+    }
+
+    FMT_INLINE void on_replacement_field(int id, const Char*) {
+      auto arg = get_arg(context, id);
+      context.advance_to(visit_format_arg(
+          default_arg_formatter<Char>{context.out(), context.args(),
+                                      context.locale()},
+          arg));
+    }
+
+    auto on_format_specs(int id, const Char* begin, const Char* end)
+        -> const Char* {
+      auto arg = get_arg(context, id);
+      // Not using a visitor for custom types gives better codegen.
+      if (arg.format_custom(begin, parse_context, context))
+        return parse_context.begin();
+      auto specs = detail::dynamic_format_specs<Char>();
+      begin = parse_format_specs(begin, end, specs, parse_context, arg.type());
+      detail::handle_dynamic_spec<detail::width_checker>(
+          specs.width, specs.width_ref, context);
+      detail::handle_dynamic_spec<detail::precision_checker>(
+          specs.precision, specs.precision_ref, context);
+      if (begin == end || *begin != '}')
+        throw_format_error("missing '}' in format string");
+      auto f = arg_formatter<Char>{context.out(), specs, context.locale()};
+      context.advance_to(visit_format_arg(f, arg));
+      return begin;
+    }
+  };
+  detail::parse_format_string<false>(fmt, format_handler(out, fmt, args, loc));
+}
+
+FMT_BEGIN_EXPORT
+
+#ifndef FMT_HEADER_ONLY
+extern template FMT_API void vformat_to(buffer<char>&, string_view,
+                                        typename vformat_args<>::type,
+                                        locale_ref);
+extern template FMT_API auto thousands_sep_impl<char>(locale_ref)
+    -> thousands_sep_result<char>;
+extern template FMT_API auto thousands_sep_impl<wchar_t>(locale_ref)
+    -> thousands_sep_result<wchar_t>;
+extern template FMT_API auto decimal_point_impl(locale_ref) -> char;
+extern template FMT_API auto decimal_point_impl(locale_ref) -> wchar_t;
+#endif  // FMT_HEADER_ONLY
+
+}  // namespace detail
+
+#if FMT_USE_USER_DEFINED_LITERALS
+inline namespace literals {
+/**
+  \rst
+  User-defined literal equivalent of :func:`fmt::arg`.
+
+  **Example**::
+
+    using namespace fmt::literals;
+    fmt::print("Elapsed time: {s:.2f} seconds", "s"_a=1.23);
+  \endrst
+ */
+#  if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <detail_exported::fixed_string Str> constexpr auto operator""_a() {
+  using char_t = remove_cvref_t<decltype(Str.data[0])>;
+  return detail::udl_arg<char_t, sizeof(Str.data) / sizeof(char_t), Str>();
+}
+#  else
+constexpr auto operator""_a(const char* s, size_t) -> detail::udl_arg<char> {
+  return {s};
+}
+#  endif
+}  // namespace literals
+#endif  // FMT_USE_USER_DEFINED_LITERALS
+
+template <typename Locale, FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+inline auto vformat(const Locale& loc, string_view fmt, format_args args)
+    -> std::string {
+  return detail::vformat(loc, fmt, args);
+}
+
+template <typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+inline auto format(const Locale& loc, format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return fmt::vformat(loc, string_view(fmt), fmt::make_format_args(args...));
+}
+
+template <typename OutputIt, typename Locale,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value&&
+                            detail::is_locale<Locale>::value)>
+auto vformat_to(OutputIt out, const Locale& loc, string_view fmt,
+                format_args args) -> OutputIt {
+  using detail::get_buffer;
+  auto&& buf = get_buffer<char>(out);
+  detail::vformat_to(buf, fmt, args, detail::locale_ref(loc));
+  return detail::get_iterator(buf, out);
+}
+
+template <typename OutputIt, typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value&&
+                            detail::is_locale<Locale>::value)>
+FMT_INLINE auto format_to(OutputIt out, const Locale& loc,
+                          format_string<T...> fmt, T&&... args) -> OutputIt {
+  return vformat_to(out, loc, fmt, fmt::make_format_args(args...));
+}
+
+template <typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+FMT_NODISCARD FMT_INLINE auto formatted_size(const Locale& loc,
+                                             format_string<T...> fmt,
+                                             T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<>();
+  detail::vformat_to<char>(buf, fmt, fmt::make_format_args(args...),
+                           detail::locale_ref(loc));
+  return buf.count();
+}
+
+FMT_END_EXPORT
+
+template <typename T, typename Char>
+template <typename FormatContext>
+FMT_CONSTEXPR FMT_INLINE auto
+formatter<T, Char,
+          enable_if_t<detail::type_constant<T, Char>::value !=
+                      detail::type::custom_type>>::format(const T& val,
+                                                          FormatContext& ctx)
+    const -> decltype(ctx.out()) {
+  if (specs_.width_ref.kind == detail::arg_id_kind::none &&
+      specs_.precision_ref.kind == detail::arg_id_kind::none) {
+    return detail::write<Char>(ctx.out(), val, specs_, ctx.locale());
+  }
+  auto specs = specs_;
+  detail::handle_dynamic_spec<detail::width_checker>(specs.width,
+                                                     specs.width_ref, ctx);
+  detail::handle_dynamic_spec<detail::precision_checker>(
+      specs.precision, specs.precision_ref, ctx);
+  return detail::write<Char>(ctx.out(), val, specs, ctx.locale());
+}
+
+FMT_END_NAMESPACE
+
+#ifdef FMT_HEADER_ONLY
+#  define FMT_FUNC inline
+#  include "format-inl.h"
+#else
+#  define FMT_FUNC
+#endif
+
+#endif  // FMT_FORMAT_H_
diff --git a/dep/fmt/include/fmt/os.h b/dep/fmt/include/fmt/os.h
new file mode 100644
index 00000000..3c7b3ccb
--- /dev/null
+++ b/dep/fmt/include/fmt/os.h
@@ -0,0 +1,455 @@
+// Formatting library for C++ - optional OS-specific functionality
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_OS_H_
+#define FMT_OS_H_
+
+#include <cerrno>
+#include <cstddef>
+#include <cstdio>
+#include <system_error>  // std::system_error
+
+#include "format.h"
+
+#if defined __APPLE__ || defined(__FreeBSD__)
+#  if FMT_HAS_INCLUDE(<xlocale.h>)
+#    include <xlocale.h>  // for LC_NUMERIC_MASK on OS X
+#  endif
+#endif
+
+#ifndef FMT_USE_FCNTL
+// UWP doesn't provide _pipe.
+#  if FMT_HAS_INCLUDE("winapifamily.h")
+#    include <winapifamily.h>
+#  endif
+#  if (FMT_HAS_INCLUDE(<fcntl.h>) || defined(__APPLE__) || \
+       defined(__linux__)) &&                              \
+      (!defined(WINAPI_FAMILY) ||                          \
+       (WINAPI_FAMILY == WINAPI_FAMILY_DESKTOP_APP))
+#    include <fcntl.h>  // for O_RDONLY
+#    define FMT_USE_FCNTL 1
+#  else
+#    define FMT_USE_FCNTL 0
+#  endif
+#endif
+
+#ifndef FMT_POSIX
+#  if defined(_WIN32) && !defined(__MINGW32__)
+// Fix warnings about deprecated symbols.
+#    define FMT_POSIX(call) _##call
+#  else
+#    define FMT_POSIX(call) call
+#  endif
+#endif
+
+// Calls to system functions are wrapped in FMT_SYSTEM for testability.
+#ifdef FMT_SYSTEM
+#  define FMT_HAS_SYSTEM
+#  define FMT_POSIX_CALL(call) FMT_SYSTEM(call)
+#else
+#  define FMT_SYSTEM(call) ::call
+#  ifdef _WIN32
+// Fix warnings about deprecated symbols.
+#    define FMT_POSIX_CALL(call) ::_##call
+#  else
+#    define FMT_POSIX_CALL(call) ::call
+#  endif
+#endif
+
+// Retries the expression while it evaluates to error_result and errno
+// equals to EINTR.
+#ifndef _WIN32
+#  define FMT_RETRY_VAL(result, expression, error_result) \
+    do {                                                  \
+      (result) = (expression);                            \
+    } while ((result) == (error_result) && errno == EINTR)
+#else
+#  define FMT_RETRY_VAL(result, expression, error_result) result = (expression)
+#endif
+
+#define FMT_RETRY(result, expression) FMT_RETRY_VAL(result, expression, -1)
+
+FMT_BEGIN_NAMESPACE
+FMT_BEGIN_EXPORT
+
+/**
+  \rst
+  A reference to a null-terminated string. It can be constructed from a C
+  string or ``std::string``.
+
+  You can use one of the following type aliases for common character types:
+
+  +---------------+-----------------------------+
+  | Type          | Definition                  |
+  +===============+=============================+
+  | cstring_view  | basic_cstring_view<char>    |
+  +---------------+-----------------------------+
+  | wcstring_view | basic_cstring_view<wchar_t> |
+  +---------------+-----------------------------+
+
+  This class is most useful as a parameter type to allow passing
+  different types of strings to a function, for example::
+
+    template <typename... Args>
+    std::string format(cstring_view format_str, const Args & ... args);
+
+    format("{}", 42);
+    format(std::string("{}"), 42);
+  \endrst
+ */
+template <typename Char> class basic_cstring_view {
+ private:
+  const Char* data_;
+
+ public:
+  /** Constructs a string reference object from a C string. */
+  basic_cstring_view(const Char* s) : data_(s) {}
+
+  /**
+    \rst
+    Constructs a string reference from an ``std::string`` object.
+    \endrst
+   */
+  basic_cstring_view(const std::basic_string<Char>& s) : data_(s.c_str()) {}
+
+  /** Returns the pointer to a C string. */
+  auto c_str() const -> const Char* { return data_; }
+};
+
+using cstring_view = basic_cstring_view<char>;
+using wcstring_view = basic_cstring_view<wchar_t>;
+
+#ifdef _WIN32
+FMT_API const std::error_category& system_category() noexcept;
+
+namespace detail {
+FMT_API void format_windows_error(buffer<char>& out, int error_code,
+                                  const char* message) noexcept;
+}
+
+FMT_API std::system_error vwindows_error(int error_code, string_view format_str,
+                                         format_args args);
+
+/**
+ \rst
+ Constructs a :class:`std::system_error` object with the description
+ of the form
+
+ .. parsed-literal::
+   *<message>*: *<system-message>*
+
+ where *<message>* is the formatted message and *<system-message>* is the
+ system message corresponding to the error code.
+ *error_code* is a Windows error code as given by ``GetLastError``.
+ If *error_code* is not a valid error code such as -1, the system message
+ will look like "error -1".
+
+ **Example**::
+
+   // This throws a system_error with the description
+   //   cannot open file 'madeup': The system cannot find the file specified.
+   // or similar (system message may vary).
+   const char *filename = "madeup";
+   LPOFSTRUCT of = LPOFSTRUCT();
+   HFILE file = OpenFile(filename, &of, OF_READ);
+   if (file == HFILE_ERROR) {
+     throw fmt::windows_error(GetLastError(),
+                              "cannot open file '{}'", filename);
+   }
+ \endrst
+*/
+template <typename... Args>
+std::system_error windows_error(int error_code, string_view message,
+                                const Args&... args) {
+  return vwindows_error(error_code, message, fmt::make_format_args(args...));
+}
+
+// Reports a Windows error without throwing an exception.
+// Can be used to report errors from destructors.
+FMT_API void report_windows_error(int error_code, const char* message) noexcept;
+#else
+inline auto system_category() noexcept -> const std::error_category& {
+  return std::system_category();
+}
+#endif  // _WIN32
+
+// std::system is not available on some platforms such as iOS (#2248).
+#ifdef __OSX__
+template <typename S, typename... Args, typename Char = char_t<S>>
+void say(const S& format_str, Args&&... args) {
+  std::system(format("say \"{}\"", format(format_str, args...)).c_str());
+}
+#endif
+
+// A buffered file.
+class buffered_file {
+ private:
+  FILE* file_;
+
+  friend class file;
+
+  explicit buffered_file(FILE* f) : file_(f) {}
+
+ public:
+  buffered_file(const buffered_file&) = delete;
+  void operator=(const buffered_file&) = delete;
+
+  // Constructs a buffered_file object which doesn't represent any file.
+  buffered_file() noexcept : file_(nullptr) {}
+
+  // Destroys the object closing the file it represents if any.
+  FMT_API ~buffered_file() noexcept;
+
+ public:
+  buffered_file(buffered_file&& other) noexcept : file_(other.file_) {
+    other.file_ = nullptr;
+  }
+
+  auto operator=(buffered_file&& other) -> buffered_file& {
+    close();
+    file_ = other.file_;
+    other.file_ = nullptr;
+    return *this;
+  }
+
+  // Opens a file.
+  FMT_API buffered_file(cstring_view filename, cstring_view mode);
+
+  // Closes the file.
+  FMT_API void close();
+
+  // Returns the pointer to a FILE object representing this file.
+  auto get() const noexcept -> FILE* { return file_; }
+
+  FMT_API auto descriptor() const -> int;
+
+  void vprint(string_view format_str, format_args args) {
+    fmt::vprint(file_, format_str, args);
+  }
+
+  template <typename... Args>
+  inline void print(string_view format_str, const Args&... args) {
+    vprint(format_str, fmt::make_format_args(args...));
+  }
+};
+
+#if FMT_USE_FCNTL
+// A file. Closed file is represented by a file object with descriptor -1.
+// Methods that are not declared with noexcept may throw
+// fmt::system_error in case of failure. Note that some errors such as
+// closing the file multiple times will cause a crash on Windows rather
+// than an exception. You can get standard behavior by overriding the
+// invalid parameter handler with _set_invalid_parameter_handler.
+class FMT_API file {
+ private:
+  int fd_;  // File descriptor.
+
+  // Constructs a file object with a given descriptor.
+  explicit file(int fd) : fd_(fd) {}
+
+ public:
+  // Possible values for the oflag argument to the constructor.
+  enum {
+    RDONLY = FMT_POSIX(O_RDONLY),  // Open for reading only.
+    WRONLY = FMT_POSIX(O_WRONLY),  // Open for writing only.
+    RDWR = FMT_POSIX(O_RDWR),      // Open for reading and writing.
+    CREATE = FMT_POSIX(O_CREAT),   // Create if the file doesn't exist.
+    APPEND = FMT_POSIX(O_APPEND),  // Open in append mode.
+    TRUNC = FMT_POSIX(O_TRUNC)     // Truncate the content of the file.
+  };
+
+  // Constructs a file object which doesn't represent any file.
+  file() noexcept : fd_(-1) {}
+
+  // Opens a file and constructs a file object representing this file.
+  file(cstring_view path, int oflag);
+
+ public:
+  file(const file&) = delete;
+  void operator=(const file&) = delete;
+
+  file(file&& other) noexcept : fd_(other.fd_) { other.fd_ = -1; }
+
+  // Move assignment is not noexcept because close may throw.
+  auto operator=(file&& other) -> file& {
+    close();
+    fd_ = other.fd_;
+    other.fd_ = -1;
+    return *this;
+  }
+
+  // Destroys the object closing the file it represents if any.
+  ~file() noexcept;
+
+  // Returns the file descriptor.
+  auto descriptor() const noexcept -> int { return fd_; }
+
+  // Closes the file.
+  void close();
+
+  // Returns the file size. The size has signed type for consistency with
+  // stat::st_size.
+  auto size() const -> long long;
+
+  // Attempts to read count bytes from the file into the specified buffer.
+  auto read(void* buffer, size_t count) -> size_t;
+
+  // Attempts to write count bytes from the specified buffer to the file.
+  auto write(const void* buffer, size_t count) -> size_t;
+
+  // Duplicates a file descriptor with the dup function and returns
+  // the duplicate as a file object.
+  static auto dup(int fd) -> file;
+
+  // Makes fd be the copy of this file descriptor, closing fd first if
+  // necessary.
+  void dup2(int fd);
+
+  // Makes fd be the copy of this file descriptor, closing fd first if
+  // necessary.
+  void dup2(int fd, std::error_code& ec) noexcept;
+
+  // Creates a pipe setting up read_end and write_end file objects for reading
+  // and writing respectively.
+  // DEPRECATED! Taking files as out parameters is deprecated.
+  static void pipe(file& read_end, file& write_end);
+
+  // Creates a buffered_file object associated with this file and detaches
+  // this file object from the file.
+  auto fdopen(const char* mode) -> buffered_file;
+
+#  if defined(_WIN32) && !defined(__MINGW32__)
+  // Opens a file and constructs a file object representing this file by
+  // wcstring_view filename. Windows only.
+  static file open_windows_file(wcstring_view path, int oflag);
+#  endif
+};
+
+// Returns the memory page size.
+auto getpagesize() -> long;
+
+namespace detail {
+
+struct buffer_size {
+  buffer_size() = default;
+  size_t value = 0;
+  auto operator=(size_t val) const -> buffer_size {
+    auto bs = buffer_size();
+    bs.value = val;
+    return bs;
+  }
+};
+
+struct ostream_params {
+  int oflag = file::WRONLY | file::CREATE | file::TRUNC;
+  size_t buffer_size = BUFSIZ > 32768 ? BUFSIZ : 32768;
+
+  ostream_params() {}
+
+  template <typename... T>
+  ostream_params(T... params, int new_oflag) : ostream_params(params...) {
+    oflag = new_oflag;
+  }
+
+  template <typename... T>
+  ostream_params(T... params, detail::buffer_size bs)
+      : ostream_params(params...) {
+    this->buffer_size = bs.value;
+  }
+
+// Intel has a bug that results in failure to deduce a constructor
+// for empty parameter packs.
+#  if defined(__INTEL_COMPILER) && __INTEL_COMPILER < 2000
+  ostream_params(int new_oflag) : oflag(new_oflag) {}
+  ostream_params(detail::buffer_size bs) : buffer_size(bs.value) {}
+#  endif
+};
+
+class file_buffer final : public buffer<char> {
+  file file_;
+
+  FMT_API void grow(size_t) override;
+
+ public:
+  FMT_API file_buffer(cstring_view path, const ostream_params& params);
+  FMT_API file_buffer(file_buffer&& other);
+  FMT_API ~file_buffer();
+
+  void flush() {
+    if (size() == 0) return;
+    file_.write(data(), size() * sizeof(data()[0]));
+    clear();
+  }
+
+  void close() {
+    flush();
+    file_.close();
+  }
+};
+
+}  // namespace detail
+
+// Added {} below to work around default constructor error known to
+// occur in Xcode versions 7.2.1 and 8.2.1.
+constexpr detail::buffer_size buffer_size{};
+
+/** A fast output stream which is not thread-safe. */
+class FMT_API ostream {
+ private:
+  FMT_MSC_WARNING(suppress : 4251)
+  detail::file_buffer buffer_;
+
+  ostream(cstring_view path, const detail::ostream_params& params)
+      : buffer_(path, params) {}
+
+ public:
+  ostream(ostream&& other) : buffer_(std::move(other.buffer_)) {}
+
+  ~ostream();
+
+  void flush() { buffer_.flush(); }
+
+  template <typename... T>
+  friend auto output_file(cstring_view path, T... params) -> ostream;
+
+  void close() { buffer_.close(); }
+
+  /**
+    Formats ``args`` according to specifications in ``fmt`` and writes the
+    output to the file.
+   */
+  template <typename... T> void print(format_string<T...> fmt, T&&... args) {
+    vformat_to(std::back_inserter(buffer_), fmt,
+               fmt::make_format_args(args...));
+  }
+};
+
+/**
+  \rst
+  Opens a file for writing. Supported parameters passed in *params*:
+
+  * ``<integer>``: Flags passed to `open
+    <https://pubs.opengroup.org/onlinepubs/007904875/functions/open.html>`_
+    (``file::WRONLY | file::CREATE | file::TRUNC`` by default)
+  * ``buffer_size=<integer>``: Output buffer size
+
+  **Example**::
+
+    auto out = fmt::output_file("guide.txt");
+    out.print("Don't {}", "Panic");
+  \endrst
+ */
+template <typename... T>
+inline auto output_file(cstring_view path, T... params) -> ostream {
+  return {path, detail::ostream_params(params...)};
+}
+#endif  // FMT_USE_FCNTL
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_OS_H_
diff --git a/dep/fmt/include/fmt/ostream.h b/dep/fmt/include/fmt/ostream.h
new file mode 100644
index 00000000..26fb3b5a
--- /dev/null
+++ b/dep/fmt/include/fmt/ostream.h
@@ -0,0 +1,245 @@
+// Formatting library for C++ - std::ostream support
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_OSTREAM_H_
+#define FMT_OSTREAM_H_
+
+#include <fstream>  // std::filebuf
+
+#ifdef _WIN32
+#  ifdef __GLIBCXX__
+#    include <ext/stdio_filebuf.h>
+#    include <ext/stdio_sync_filebuf.h>
+#  endif
+#  include <io.h>
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename Streambuf> class formatbuf : public Streambuf {
+ private:
+  using char_type = typename Streambuf::char_type;
+  using streamsize = decltype(std::declval<Streambuf>().sputn(nullptr, 0));
+  using int_type = typename Streambuf::int_type;
+  using traits_type = typename Streambuf::traits_type;
+
+  buffer<char_type>& buffer_;
+
+ public:
+  explicit formatbuf(buffer<char_type>& buf) : buffer_(buf) {}
+
+ protected:
+  // The put area is always empty. This makes the implementation simpler and has
+  // the advantage that the streambuf and the buffer are always in sync and
+  // sputc never writes into uninitialized memory. A disadvantage is that each
+  // call to sputc always results in a (virtual) call to overflow. There is no
+  // disadvantage here for sputn since this always results in a call to xsputn.
+
+  auto overflow(int_type ch) -> int_type override {
+    if (!traits_type::eq_int_type(ch, traits_type::eof()))
+      buffer_.push_back(static_cast<char_type>(ch));
+    return ch;
+  }
+
+  auto xsputn(const char_type* s, streamsize count) -> streamsize override {
+    buffer_.append(s, s + count);
+    return count;
+  }
+};
+
+// Generate a unique explicit instantion in every translation unit using a tag
+// type in an anonymous namespace.
+namespace {
+struct file_access_tag {};
+}  // namespace
+template <typename Tag, typename BufType, FILE* BufType::*FileMemberPtr>
+class file_access {
+  friend auto get_file(BufType& obj) -> FILE* { return obj.*FileMemberPtr; }
+};
+
+#if FMT_MSC_VERSION
+template class file_access<file_access_tag, std::filebuf,
+                           &std::filebuf::_Myfile>;
+auto get_file(std::filebuf&) -> FILE*;
+#endif
+
+inline auto write_ostream_unicode(std::ostream& os, fmt::string_view data)
+    -> bool {
+  FILE* f = nullptr;
+#if FMT_MSC_VERSION
+  if (auto* buf = dynamic_cast<std::filebuf*>(os.rdbuf()))
+    f = get_file(*buf);
+  else
+    return false;
+#elif defined(_WIN32) && defined(__GLIBCXX__)
+  auto* rdbuf = os.rdbuf();
+  if (auto* sfbuf = dynamic_cast<__gnu_cxx::stdio_sync_filebuf<char>*>(rdbuf))
+    f = sfbuf->file();
+  else if (auto* fbuf = dynamic_cast<__gnu_cxx::stdio_filebuf<char>*>(rdbuf))
+    f = fbuf->file();
+  else
+    return false;
+#else
+  ignore_unused(os, data, f);
+#endif
+#ifdef _WIN32
+  if (f) {
+    int fd = _fileno(f);
+    if (_isatty(fd)) {
+      os.flush();
+      return write_console(fd, data);
+    }
+  }
+#endif
+  return false;
+}
+inline auto write_ostream_unicode(std::wostream&,
+                                  fmt::basic_string_view<wchar_t>) -> bool {
+  return false;
+}
+
+// Write the content of buf to os.
+// It is a separate function rather than a part of vprint to simplify testing.
+template <typename Char>
+void write_buffer(std::basic_ostream<Char>& os, buffer<Char>& buf) {
+  const Char* buf_data = buf.data();
+  using unsigned_streamsize = std::make_unsigned<std::streamsize>::type;
+  unsigned_streamsize size = buf.size();
+  unsigned_streamsize max_size = to_unsigned(max_value<std::streamsize>());
+  do {
+    unsigned_streamsize n = size <= max_size ? size : max_size;
+    os.write(buf_data, static_cast<std::streamsize>(n));
+    buf_data += n;
+    size -= n;
+  } while (size != 0);
+}
+
+template <typename Char, typename T>
+void format_value(buffer<Char>& buf, const T& value) {
+  auto&& format_buf = formatbuf<std::basic_streambuf<Char>>(buf);
+  auto&& output = std::basic_ostream<Char>(&format_buf);
+#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+  output.imbue(std::locale::classic());  // The default is always unlocalized.
+#endif
+  output << value;
+  output.exceptions(std::ios_base::failbit | std::ios_base::badbit);
+}
+
+template <typename T> struct streamed_view {
+  const T& value;
+};
+
+}  // namespace detail
+
+// Formats an object of type T that has an overloaded ostream operator<<.
+template <typename Char>
+struct basic_ostream_formatter : formatter<basic_string_view<Char>, Char> {
+  void set_debug_format() = delete;
+
+  template <typename T, typename OutputIt>
+  auto format(const T& value, basic_format_context<OutputIt, Char>& ctx) const
+      -> OutputIt {
+    auto buffer = basic_memory_buffer<Char>();
+    detail::format_value(buffer, value);
+    return formatter<basic_string_view<Char>, Char>::format(
+        {buffer.data(), buffer.size()}, ctx);
+  }
+};
+
+using ostream_formatter = basic_ostream_formatter<char>;
+
+template <typename T, typename Char>
+struct formatter<detail::streamed_view<T>, Char>
+    : basic_ostream_formatter<Char> {
+  template <typename OutputIt>
+  auto format(detail::streamed_view<T> view,
+              basic_format_context<OutputIt, Char>& ctx) const -> OutputIt {
+    return basic_ostream_formatter<Char>::format(view.value, ctx);
+  }
+};
+
+/**
+  \rst
+  Returns a view that formats `value` via an ostream ``operator<<``.
+
+  **Example**::
+
+    fmt::print("Current thread id: {}\n",
+               fmt::streamed(std::this_thread::get_id()));
+  \endrst
+ */
+template <typename T>
+constexpr auto streamed(const T& value) -> detail::streamed_view<T> {
+  return {value};
+}
+
+namespace detail {
+
+inline void vprint_directly(std::ostream& os, string_view format_str,
+                            format_args args) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, format_str, args);
+  detail::write_buffer(os, buffer);
+}
+
+}  // namespace detail
+
+FMT_EXPORT template <typename Char>
+void vprint(std::basic_ostream<Char>& os,
+            basic_string_view<type_identity_t<Char>> format_str,
+            basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  auto buffer = basic_memory_buffer<Char>();
+  detail::vformat_to(buffer, format_str, args);
+  if (detail::write_ostream_unicode(os, {buffer.data(), buffer.size()})) return;
+  detail::write_buffer(os, buffer);
+}
+
+/**
+  \rst
+  Prints formatted data to the stream *os*.
+
+  **Example**::
+
+    fmt::print(cerr, "Don't {}!", "panic");
+  \endrst
+ */
+FMT_EXPORT template <typename... T>
+void print(std::ostream& os, format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  if (detail::is_utf8())
+    vprint(os, fmt, vargs);
+  else
+    detail::vprint_directly(os, fmt, vargs);
+}
+
+FMT_EXPORT
+template <typename... Args>
+void print(std::wostream& os,
+           basic_format_string<wchar_t, type_identity_t<Args>...> fmt,
+           Args&&... args) {
+  vprint(os, fmt, fmt::make_format_args<buffer_context<wchar_t>>(args...));
+}
+
+FMT_EXPORT template <typename... T>
+void println(std::ostream& os, format_string<T...> fmt, T&&... args) {
+  fmt::print(os, "{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+FMT_EXPORT
+template <typename... Args>
+void println(std::wostream& os,
+             basic_format_string<wchar_t, type_identity_t<Args>...> fmt,
+             Args&&... args) {
+  print(os, L"{}\n", fmt::format(fmt, std::forward<Args>(args)...));
+}
+
+FMT_END_NAMESPACE
+
+#endif  // FMT_OSTREAM_H_
diff --git a/dep/fmt/include/fmt/printf.h b/dep/fmt/include/fmt/printf.h
new file mode 100644
index 00000000..07e81577
--- /dev/null
+++ b/dep/fmt/include/fmt/printf.h
@@ -0,0 +1,675 @@
+// Formatting library for C++ - legacy printf implementation
+//
+// Copyright (c) 2012 - 2016, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_PRINTF_H_
+#define FMT_PRINTF_H_
+
+#include <algorithm>  // std::max
+#include <limits>     // std::numeric_limits
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+FMT_BEGIN_EXPORT
+
+template <typename T> struct printf_formatter {
+  printf_formatter() = delete;
+};
+
+template <typename Char> class basic_printf_context {
+ private:
+  detail::buffer_appender<Char> out_;
+  basic_format_args<basic_printf_context> args_;
+
+  static_assert(std::is_same<Char, char>::value ||
+                    std::is_same<Char, wchar_t>::value,
+                "Unsupported code unit type.");
+
+ public:
+  using char_type = Char;
+  using parse_context_type = basic_format_parse_context<Char>;
+  template <typename T> using formatter_type = printf_formatter<T>;
+
+  /**
+    \rst
+    Constructs a ``printf_context`` object. References to the arguments are
+    stored in the context object so make sure they have appropriate lifetimes.
+    \endrst
+   */
+  basic_printf_context(detail::buffer_appender<Char> out,
+                       basic_format_args<basic_printf_context> args)
+      : out_(out), args_(args) {}
+
+  auto out() -> detail::buffer_appender<Char> { return out_; }
+  void advance_to(detail::buffer_appender<Char>) {}
+
+  auto locale() -> detail::locale_ref { return {}; }
+
+  auto arg(int id) const -> basic_format_arg<basic_printf_context> {
+    return args_.get(id);
+  }
+
+  FMT_CONSTEXPR void on_error(const char* message) {
+    detail::error_handler().on_error(message);
+  }
+};
+
+namespace detail {
+
+// Checks if a value fits in int - used to avoid warnings about comparing
+// signed and unsigned integers.
+template <bool IsSigned> struct int_checker {
+  template <typename T> static auto fits_in_int(T value) -> bool {
+    unsigned max = max_value<int>();
+    return value <= max;
+  }
+  static auto fits_in_int(bool) -> bool { return true; }
+};
+
+template <> struct int_checker<true> {
+  template <typename T> static auto fits_in_int(T value) -> bool {
+    return value >= (std::numeric_limits<int>::min)() &&
+           value <= max_value<int>();
+  }
+  static auto fits_in_int(int) -> bool { return true; }
+};
+
+struct printf_precision_handler {
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  auto operator()(T value) -> int {
+    if (!int_checker<std::numeric_limits<T>::is_signed>::fits_in_int(value))
+      throw_format_error("number is too big");
+    return (std::max)(static_cast<int>(value), 0);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  auto operator()(T) -> int {
+    throw_format_error("precision is not integer");
+    return 0;
+  }
+};
+
+// An argument visitor that returns true iff arg is a zero integer.
+struct is_zero_int {
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  auto operator()(T value) -> bool {
+    return value == 0;
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  auto operator()(T) -> bool {
+    return false;
+  }
+};
+
+template <typename T> struct make_unsigned_or_bool : std::make_unsigned<T> {};
+
+template <> struct make_unsigned_or_bool<bool> {
+  using type = bool;
+};
+
+template <typename T, typename Context> class arg_converter {
+ private:
+  using char_type = typename Context::char_type;
+
+  basic_format_arg<Context>& arg_;
+  char_type type_;
+
+ public:
+  arg_converter(basic_format_arg<Context>& arg, char_type type)
+      : arg_(arg), type_(type) {}
+
+  void operator()(bool value) {
+    if (type_ != 's') operator()<bool>(value);
+  }
+
+  template <typename U, FMT_ENABLE_IF(std::is_integral<U>::value)>
+  void operator()(U value) {
+    bool is_signed = type_ == 'd' || type_ == 'i';
+    using target_type = conditional_t<std::is_same<T, void>::value, U, T>;
+    if (const_check(sizeof(target_type) <= sizeof(int))) {
+      // Extra casts are used to silence warnings.
+      if (is_signed) {
+        auto n = static_cast<int>(static_cast<target_type>(value));
+        arg_ = detail::make_arg<Context>(n);
+      } else {
+        using unsigned_type = typename make_unsigned_or_bool<target_type>::type;
+        auto n = static_cast<unsigned>(static_cast<unsigned_type>(value));
+        arg_ = detail::make_arg<Context>(n);
+      }
+    } else {
+      if (is_signed) {
+        // glibc's printf doesn't sign extend arguments of smaller types:
+        //   std::printf("%lld", -42);  // prints "4294967254"
+        // but we don't have to do the same because it's a UB.
+        auto n = static_cast<long long>(value);
+        arg_ = detail::make_arg<Context>(n);
+      } else {
+        auto n = static_cast<typename make_unsigned_or_bool<U>::type>(value);
+        arg_ = detail::make_arg<Context>(n);
+      }
+    }
+  }
+
+  template <typename U, FMT_ENABLE_IF(!std::is_integral<U>::value)>
+  void operator()(U) {}  // No conversion needed for non-integral types.
+};
+
+// Converts an integer argument to T for printf, if T is an integral type.
+// If T is void, the argument is converted to corresponding signed or unsigned
+// type depending on the type specifier: 'd' and 'i' - signed, other -
+// unsigned).
+template <typename T, typename Context, typename Char>
+void convert_arg(basic_format_arg<Context>& arg, Char type) {
+  visit_format_arg(arg_converter<T, Context>(arg, type), arg);
+}
+
+// Converts an integer argument to char for printf.
+template <typename Context> class char_converter {
+ private:
+  basic_format_arg<Context>& arg_;
+
+ public:
+  explicit char_converter(basic_format_arg<Context>& arg) : arg_(arg) {}
+
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  void operator()(T value) {
+    auto c = static_cast<typename Context::char_type>(value);
+    arg_ = detail::make_arg<Context>(c);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  void operator()(T) {}  // No conversion needed for non-integral types.
+};
+
+// An argument visitor that return a pointer to a C string if argument is a
+// string or null otherwise.
+template <typename Char> struct get_cstring {
+  template <typename T> auto operator()(T) -> const Char* { return nullptr; }
+  auto operator()(const Char* s) -> const Char* { return s; }
+};
+
+// Checks if an argument is a valid printf width specifier and sets
+// left alignment if it is negative.
+template <typename Char> class printf_width_handler {
+ private:
+  format_specs<Char>& specs_;
+
+ public:
+  explicit printf_width_handler(format_specs<Char>& specs) : specs_(specs) {}
+
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  auto operator()(T value) -> unsigned {
+    auto width = static_cast<uint32_or_64_or_128_t<T>>(value);
+    if (detail::is_negative(value)) {
+      specs_.align = align::left;
+      width = 0 - width;
+    }
+    unsigned int_max = max_value<int>();
+    if (width > int_max) throw_format_error("number is too big");
+    return static_cast<unsigned>(width);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  auto operator()(T) -> unsigned {
+    throw_format_error("width is not integer");
+    return 0;
+  }
+};
+
+// Workaround for a bug with the XL compiler when initializing
+// printf_arg_formatter's base class.
+template <typename Char>
+auto make_arg_formatter(buffer_appender<Char> iter, format_specs<Char>& s)
+    -> arg_formatter<Char> {
+  return {iter, s, locale_ref()};
+}
+
+// The ``printf`` argument formatter.
+template <typename Char>
+class printf_arg_formatter : public arg_formatter<Char> {
+ private:
+  using base = arg_formatter<Char>;
+  using context_type = basic_printf_context<Char>;
+
+  context_type& context_;
+
+  void write_null_pointer(bool is_string = false) {
+    auto s = this->specs;
+    s.type = presentation_type::none;
+    write_bytes(this->out, is_string ? "(null)" : "(nil)", s);
+  }
+
+ public:
+  printf_arg_formatter(buffer_appender<Char> iter, format_specs<Char>& s,
+                       context_type& ctx)
+      : base(make_arg_formatter(iter, s)), context_(ctx) {}
+
+  void operator()(monostate value) { base::operator()(value); }
+
+  template <typename T, FMT_ENABLE_IF(detail::is_integral<T>::value)>
+  void operator()(T value) {
+    // MSVC2013 fails to compile separate overloads for bool and Char so use
+    // std::is_same instead.
+    if (!std::is_same<T, Char>::value) {
+      base::operator()(value);
+      return;
+    }
+    format_specs<Char> fmt_specs = this->specs;
+    if (fmt_specs.type != presentation_type::none &&
+        fmt_specs.type != presentation_type::chr) {
+      return (*this)(static_cast<int>(value));
+    }
+    fmt_specs.sign = sign::none;
+    fmt_specs.alt = false;
+    fmt_specs.fill[0] = ' ';  // Ignore '0' flag for char types.
+    // align::numeric needs to be overwritten here since the '0' flag is
+    // ignored for non-numeric types
+    if (fmt_specs.align == align::none || fmt_specs.align == align::numeric)
+      fmt_specs.align = align::right;
+    write<Char>(this->out, static_cast<Char>(value), fmt_specs);
+  }
+
+  template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+  void operator()(T value) {
+    base::operator()(value);
+  }
+
+  /** Formats a null-terminated C string. */
+  void operator()(const char* value) {
+    if (value)
+      base::operator()(value);
+    else
+      write_null_pointer(this->specs.type != presentation_type::pointer);
+  }
+
+  /** Formats a null-terminated wide C string. */
+  void operator()(const wchar_t* value) {
+    if (value)
+      base::operator()(value);
+    else
+      write_null_pointer(this->specs.type != presentation_type::pointer);
+  }
+
+  void operator()(basic_string_view<Char> value) { base::operator()(value); }
+
+  /** Formats a pointer. */
+  void operator()(const void* value) {
+    if (value)
+      base::operator()(value);
+    else
+      write_null_pointer();
+  }
+
+  /** Formats an argument of a custom (user-defined) type. */
+  void operator()(typename basic_format_arg<context_type>::handle handle) {
+    auto parse_ctx = basic_format_parse_context<Char>({});
+    handle.format(parse_ctx, context_);
+  }
+};
+
+template <typename Char>
+void parse_flags(format_specs<Char>& specs, const Char*& it, const Char* end) {
+  for (; it != end; ++it) {
+    switch (*it) {
+    case '-':
+      specs.align = align::left;
+      break;
+    case '+':
+      specs.sign = sign::plus;
+      break;
+    case '0':
+      specs.fill[0] = '0';
+      break;
+    case ' ':
+      if (specs.sign != sign::plus) specs.sign = sign::space;
+      break;
+    case '#':
+      specs.alt = true;
+      break;
+    default:
+      return;
+    }
+  }
+}
+
+template <typename Char, typename GetArg>
+auto parse_header(const Char*& it, const Char* end, format_specs<Char>& specs,
+                  GetArg get_arg) -> int {
+  int arg_index = -1;
+  Char c = *it;
+  if (c >= '0' && c <= '9') {
+    // Parse an argument index (if followed by '$') or a width possibly
+    // preceded with '0' flag(s).
+    int value = parse_nonnegative_int(it, end, -1);
+    if (it != end && *it == '$') {  // value is an argument index
+      ++it;
+      arg_index = value != -1 ? value : max_value<int>();
+    } else {
+      if (c == '0') specs.fill[0] = '0';
+      if (value != 0) {
+        // Nonzero value means that we parsed width and don't need to
+        // parse it or flags again, so return now.
+        if (value == -1) throw_format_error("number is too big");
+        specs.width = value;
+        return arg_index;
+      }
+    }
+  }
+  parse_flags(specs, it, end);
+  // Parse width.
+  if (it != end) {
+    if (*it >= '0' && *it <= '9') {
+      specs.width = parse_nonnegative_int(it, end, -1);
+      if (specs.width == -1) throw_format_error("number is too big");
+    } else if (*it == '*') {
+      ++it;
+      specs.width = static_cast<int>(visit_format_arg(
+          detail::printf_width_handler<Char>(specs), get_arg(-1)));
+    }
+  }
+  return arg_index;
+}
+
+inline auto parse_printf_presentation_type(char c, type t)
+    -> presentation_type {
+  using pt = presentation_type;
+  constexpr auto integral_set = sint_set | uint_set | bool_set | char_set;
+  switch (c) {
+  case 'd':
+    return in(t, integral_set) ? pt::dec : pt::none;
+  case 'o':
+    return in(t, integral_set) ? pt::oct : pt::none;
+  case 'x':
+    return in(t, integral_set) ? pt::hex_lower : pt::none;
+  case 'X':
+    return in(t, integral_set) ? pt::hex_upper : pt::none;
+  case 'a':
+    return in(t, float_set) ? pt::hexfloat_lower : pt::none;
+  case 'A':
+    return in(t, float_set) ? pt::hexfloat_upper : pt::none;
+  case 'e':
+    return in(t, float_set) ? pt::exp_lower : pt::none;
+  case 'E':
+    return in(t, float_set) ? pt::exp_upper : pt::none;
+  case 'f':
+    return in(t, float_set) ? pt::fixed_lower : pt::none;
+  case 'F':
+    return in(t, float_set) ? pt::fixed_upper : pt::none;
+  case 'g':
+    return in(t, float_set) ? pt::general_lower : pt::none;
+  case 'G':
+    return in(t, float_set) ? pt::general_upper : pt::none;
+  case 'c':
+    return in(t, integral_set) ? pt::chr : pt::none;
+  case 's':
+    return in(t, string_set | cstring_set) ? pt::string : pt::none;
+  case 'p':
+    return in(t, pointer_set | cstring_set) ? pt::pointer : pt::none;
+  default:
+    return pt::none;
+  }
+}
+
+template <typename Char, typename Context>
+void vprintf(buffer<Char>& buf, basic_string_view<Char> format,
+             basic_format_args<Context> args) {
+  using iterator = buffer_appender<Char>;
+  auto out = iterator(buf);
+  auto context = basic_printf_context<Char>(out, args);
+  auto parse_ctx = basic_format_parse_context<Char>(format);
+
+  // Returns the argument with specified index or, if arg_index is -1, the next
+  // argument.
+  auto get_arg = [&](int arg_index) {
+    if (arg_index < 0)
+      arg_index = parse_ctx.next_arg_id();
+    else
+      parse_ctx.check_arg_id(--arg_index);
+    return detail::get_arg(context, arg_index);
+  };
+
+  const Char* start = parse_ctx.begin();
+  const Char* end = parse_ctx.end();
+  auto it = start;
+  while (it != end) {
+    if (!find<false, Char>(it, end, '%', it)) {
+      it = end;  // find leaves it == nullptr if it doesn't find '%'.
+      break;
+    }
+    Char c = *it++;
+    if (it != end && *it == c) {
+      write(out, basic_string_view<Char>(start, to_unsigned(it - start)));
+      start = ++it;
+      continue;
+    }
+    write(out, basic_string_view<Char>(start, to_unsigned(it - 1 - start)));
+
+    auto specs = format_specs<Char>();
+    specs.align = align::right;
+
+    // Parse argument index, flags and width.
+    int arg_index = parse_header(it, end, specs, get_arg);
+    if (arg_index == 0) throw_format_error("argument not found");
+
+    // Parse precision.
+    if (it != end && *it == '.') {
+      ++it;
+      c = it != end ? *it : 0;
+      if ('0' <= c && c <= '9') {
+        specs.precision = parse_nonnegative_int(it, end, 0);
+      } else if (c == '*') {
+        ++it;
+        specs.precision = static_cast<int>(
+            visit_format_arg(printf_precision_handler(), get_arg(-1)));
+      } else {
+        specs.precision = 0;
+      }
+    }
+
+    auto arg = get_arg(arg_index);
+    // For d, i, o, u, x, and X conversion specifiers, if a precision is
+    // specified, the '0' flag is ignored
+    if (specs.precision >= 0 && arg.is_integral()) {
+      // Ignore '0' for non-numeric types or if '-' present.
+      specs.fill[0] = ' ';
+    }
+    if (specs.precision >= 0 && arg.type() == type::cstring_type) {
+      auto str = visit_format_arg(get_cstring<Char>(), arg);
+      auto str_end = str + specs.precision;
+      auto nul = std::find(str, str_end, Char());
+      auto sv = basic_string_view<Char>(
+          str, to_unsigned(nul != str_end ? nul - str : specs.precision));
+      arg = make_arg<basic_printf_context<Char>>(sv);
+    }
+    if (specs.alt && visit_format_arg(is_zero_int(), arg)) specs.alt = false;
+    if (specs.fill[0] == '0') {
+      if (arg.is_arithmetic() && specs.align != align::left)
+        specs.align = align::numeric;
+      else
+        specs.fill[0] = ' ';  // Ignore '0' flag for non-numeric types or if '-'
+                              // flag is also present.
+    }
+
+    // Parse length and convert the argument to the required type.
+    c = it != end ? *it++ : 0;
+    Char t = it != end ? *it : 0;
+    switch (c) {
+    case 'h':
+      if (t == 'h') {
+        ++it;
+        t = it != end ? *it : 0;
+        convert_arg<signed char>(arg, t);
+      } else {
+        convert_arg<short>(arg, t);
+      }
+      break;
+    case 'l':
+      if (t == 'l') {
+        ++it;
+        t = it != end ? *it : 0;
+        convert_arg<long long>(arg, t);
+      } else {
+        convert_arg<long>(arg, t);
+      }
+      break;
+    case 'j':
+      convert_arg<intmax_t>(arg, t);
+      break;
+    case 'z':
+      convert_arg<size_t>(arg, t);
+      break;
+    case 't':
+      convert_arg<std::ptrdiff_t>(arg, t);
+      break;
+    case 'L':
+      // printf produces garbage when 'L' is omitted for long double, no
+      // need to do the same.
+      break;
+    default:
+      --it;
+      convert_arg<void>(arg, c);
+    }
+
+    // Parse type.
+    if (it == end) throw_format_error("invalid format string");
+    char type = static_cast<char>(*it++);
+    if (arg.is_integral()) {
+      // Normalize type.
+      switch (type) {
+      case 'i':
+      case 'u':
+        type = 'd';
+        break;
+      case 'c':
+        visit_format_arg(char_converter<basic_printf_context<Char>>(arg), arg);
+        break;
+      }
+    }
+    specs.type = parse_printf_presentation_type(type, arg.type());
+    if (specs.type == presentation_type::none)
+      throw_format_error("invalid format specifier");
+
+    start = it;
+
+    // Format argument.
+    visit_format_arg(printf_arg_formatter<Char>(out, specs, context), arg);
+  }
+  write(out, basic_string_view<Char>(start, to_unsigned(it - start)));
+}
+}  // namespace detail
+
+using printf_context = basic_printf_context<char>;
+using wprintf_context = basic_printf_context<wchar_t>;
+
+using printf_args = basic_format_args<printf_context>;
+using wprintf_args = basic_format_args<wprintf_context>;
+
+/**
+  \rst
+  Constructs an `~fmt::format_arg_store` object that contains references to
+  arguments and can be implicitly converted to `~fmt::printf_args`.
+  \endrst
+ */
+template <typename... T>
+inline auto make_printf_args(const T&... args)
+    -> format_arg_store<printf_context, T...> {
+  return {args...};
+}
+
+// DEPRECATED!
+template <typename... T>
+inline auto make_wprintf_args(const T&... args)
+    -> format_arg_store<wprintf_context, T...> {
+  return {args...};
+}
+
+template <typename Char>
+inline auto vsprintf(
+    basic_string_view<Char> fmt,
+    basic_format_args<basic_printf_context<type_identity_t<Char>>> args)
+    -> std::basic_string<Char> {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vprintf(buf, fmt, args);
+  return to_string(buf);
+}
+
+/**
+  \rst
+  Formats arguments and returns the result as a string.
+
+  **Example**::
+
+    std::string message = fmt::sprintf("The answer is %d", 42);
+  \endrst
+*/
+template <typename S, typename... T,
+          typename Char = enable_if_t<detail::is_string<S>::value, char_t<S>>>
+inline auto sprintf(const S& fmt, const T&... args) -> std::basic_string<Char> {
+  return vsprintf(detail::to_string_view(fmt),
+                  fmt::make_format_args<basic_printf_context<Char>>(args...));
+}
+
+template <typename Char>
+inline auto vfprintf(
+    std::FILE* f, basic_string_view<Char> fmt,
+    basic_format_args<basic_printf_context<type_identity_t<Char>>> args)
+    -> int {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vprintf(buf, fmt, args);
+  size_t size = buf.size();
+  return std::fwrite(buf.data(), sizeof(Char), size, f) < size
+             ? -1
+             : static_cast<int>(size);
+}
+
+/**
+  \rst
+  Prints formatted data to the file *f*.
+
+  **Example**::
+
+    fmt::fprintf(stderr, "Don't %s!", "panic");
+  \endrst
+ */
+template <typename S, typename... T, typename Char = char_t<S>>
+inline auto fprintf(std::FILE* f, const S& fmt, const T&... args) -> int {
+  return vfprintf(f, detail::to_string_view(fmt),
+                  fmt::make_format_args<basic_printf_context<Char>>(args...));
+}
+
+template <typename Char>
+FMT_DEPRECATED inline auto vprintf(
+    basic_string_view<Char> fmt,
+    basic_format_args<basic_printf_context<type_identity_t<Char>>> args)
+    -> int {
+  return vfprintf(stdout, fmt, args);
+}
+
+/**
+  \rst
+  Prints formatted data to ``stdout``.
+
+  **Example**::
+
+    fmt::printf("Elapsed time: %.2f seconds", 1.23);
+  \endrst
+ */
+template <typename... T>
+inline auto printf(string_view fmt, const T&... args) -> int {
+  return vfprintf(stdout, fmt, make_printf_args(args...));
+}
+template <typename... T>
+FMT_DEPRECATED inline auto printf(basic_string_view<wchar_t> fmt,
+                                  const T&... args) -> int {
+  return vfprintf(stdout, fmt, make_wprintf_args(args...));
+}
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_PRINTF_H_
diff --git a/dep/fmt/include/fmt/ranges.h b/dep/fmt/include/fmt/ranges.h
new file mode 100644
index 00000000..3638fffb
--- /dev/null
+++ b/dep/fmt/include/fmt/ranges.h
@@ -0,0 +1,738 @@
+// Formatting library for C++ - range and tuple support
+//
+// Copyright (c) 2012 - present, Victor Zverovich and {fmt} contributors
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_RANGES_H_
+#define FMT_RANGES_H_
+
+#include <initializer_list>
+#include <tuple>
+#include <type_traits>
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+
+namespace detail {
+
+template <typename Range, typename OutputIt>
+auto copy(const Range& range, OutputIt out) -> OutputIt {
+  for (auto it = range.begin(), end = range.end(); it != end; ++it)
+    *out++ = *it;
+  return out;
+}
+
+template <typename OutputIt>
+auto copy(const char* str, OutputIt out) -> OutputIt {
+  while (*str) *out++ = *str++;
+  return out;
+}
+
+template <typename OutputIt> auto copy(char ch, OutputIt out) -> OutputIt {
+  *out++ = ch;
+  return out;
+}
+
+template <typename OutputIt> auto copy(wchar_t ch, OutputIt out) -> OutputIt {
+  *out++ = ch;
+  return out;
+}
+
+// Returns true if T has a std::string-like interface, like std::string_view.
+template <typename T> class is_std_string_like {
+  template <typename U>
+  static auto check(U* p)
+      -> decltype((void)p->find('a'), p->length(), (void)p->data(), int());
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      is_string<T>::value ||
+      std::is_convertible<T, std_string_view<char>>::value ||
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+};
+
+template <typename Char>
+struct is_std_string_like<fmt::basic_string_view<Char>> : std::true_type {};
+
+template <typename T> class is_map {
+  template <typename U> static auto check(U*) -> typename U::mapped_type;
+  template <typename> static void check(...);
+
+ public:
+#ifdef FMT_FORMAT_MAP_AS_LIST  // DEPRECATED!
+  static constexpr const bool value = false;
+#else
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+#endif
+};
+
+template <typename T> class is_set {
+  template <typename U> static auto check(U*) -> typename U::key_type;
+  template <typename> static void check(...);
+
+ public:
+#ifdef FMT_FORMAT_SET_AS_LIST  // DEPRECATED!
+  static constexpr const bool value = false;
+#else
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value && !is_map<T>::value;
+#endif
+};
+
+template <typename... Ts> struct conditional_helper {};
+
+template <typename T, typename _ = void> struct is_range_ : std::false_type {};
+
+#if !FMT_MSC_VERSION || FMT_MSC_VERSION > 1800
+
+#  define FMT_DECLTYPE_RETURN(val)  \
+    ->decltype(val) { return val; } \
+    static_assert(                  \
+        true, "")  // This makes it so that a semicolon is required after the
+                   // macro, which helps clang-format handle the formatting.
+
+// C array overload
+template <typename T, std::size_t N>
+auto range_begin(const T (&arr)[N]) -> const T* {
+  return arr;
+}
+template <typename T, std::size_t N>
+auto range_end(const T (&arr)[N]) -> const T* {
+  return arr + N;
+}
+
+template <typename T, typename Enable = void>
+struct has_member_fn_begin_end_t : std::false_type {};
+
+template <typename T>
+struct has_member_fn_begin_end_t<T, void_t<decltype(std::declval<T>().begin()),
+                                           decltype(std::declval<T>().end())>>
+    : std::true_type {};
+
+// Member function overload
+template <typename T>
+auto range_begin(T&& rng) FMT_DECLTYPE_RETURN(static_cast<T&&>(rng).begin());
+template <typename T>
+auto range_end(T&& rng) FMT_DECLTYPE_RETURN(static_cast<T&&>(rng).end());
+
+// ADL overload. Only participates in overload resolution if member functions
+// are not found.
+template <typename T>
+auto range_begin(T&& rng)
+    -> enable_if_t<!has_member_fn_begin_end_t<T&&>::value,
+                   decltype(begin(static_cast<T&&>(rng)))> {
+  return begin(static_cast<T&&>(rng));
+}
+template <typename T>
+auto range_end(T&& rng) -> enable_if_t<!has_member_fn_begin_end_t<T&&>::value,
+                                       decltype(end(static_cast<T&&>(rng)))> {
+  return end(static_cast<T&&>(rng));
+}
+
+template <typename T, typename Enable = void>
+struct has_const_begin_end : std::false_type {};
+template <typename T, typename Enable = void>
+struct has_mutable_begin_end : std::false_type {};
+
+template <typename T>
+struct has_const_begin_end<
+    T,
+    void_t<
+        decltype(detail::range_begin(std::declval<const remove_cvref_t<T>&>())),
+        decltype(detail::range_end(std::declval<const remove_cvref_t<T>&>()))>>
+    : std::true_type {};
+
+template <typename T>
+struct has_mutable_begin_end<
+    T, void_t<decltype(detail::range_begin(std::declval<T>())),
+              decltype(detail::range_end(std::declval<T>())),
+              // the extra int here is because older versions of MSVC don't
+              // SFINAE properly unless there are distinct types
+              int>> : std::true_type {};
+
+template <typename T>
+struct is_range_<T, void>
+    : std::integral_constant<bool, (has_const_begin_end<T>::value ||
+                                    has_mutable_begin_end<T>::value)> {};
+#  undef FMT_DECLTYPE_RETURN
+#endif
+
+// tuple_size and tuple_element check.
+template <typename T> class is_tuple_like_ {
+  template <typename U>
+  static auto check(U* p) -> decltype(std::tuple_size<U>::value, int());
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+};
+
+// Check for integer_sequence
+#if defined(__cpp_lib_integer_sequence) || FMT_MSC_VERSION >= 1900
+template <typename T, T... N>
+using integer_sequence = std::integer_sequence<T, N...>;
+template <size_t... N> using index_sequence = std::index_sequence<N...>;
+template <size_t N> using make_index_sequence = std::make_index_sequence<N>;
+#else
+template <typename T, T... N> struct integer_sequence {
+  using value_type = T;
+
+  static FMT_CONSTEXPR auto size() -> size_t { return sizeof...(N); }
+};
+
+template <size_t... N> using index_sequence = integer_sequence<size_t, N...>;
+
+template <typename T, size_t N, T... Ns>
+struct make_integer_sequence : make_integer_sequence<T, N - 1, N - 1, Ns...> {};
+template <typename T, T... Ns>
+struct make_integer_sequence<T, 0, Ns...> : integer_sequence<T, Ns...> {};
+
+template <size_t N>
+using make_index_sequence = make_integer_sequence<size_t, N>;
+#endif
+
+template <typename T>
+using tuple_index_sequence = make_index_sequence<std::tuple_size<T>::value>;
+
+template <typename T, typename C, bool = is_tuple_like_<T>::value>
+class is_tuple_formattable_ {
+ public:
+  static constexpr const bool value = false;
+};
+template <typename T, typename C> class is_tuple_formattable_<T, C, true> {
+  template <std::size_t... Is>
+  static auto check2(index_sequence<Is...>,
+                     integer_sequence<bool, (Is == Is)...>) -> std::true_type;
+  static auto check2(...) -> std::false_type;
+  template <std::size_t... Is>
+  static auto check(index_sequence<Is...>) -> decltype(check2(
+      index_sequence<Is...>{},
+      integer_sequence<bool,
+                       (is_formattable<typename std::tuple_element<Is, T>::type,
+                                       C>::value)...>{}));
+
+ public:
+  static constexpr const bool value =
+      decltype(check(tuple_index_sequence<T>{}))::value;
+};
+
+template <typename Tuple, typename F, size_t... Is>
+FMT_CONSTEXPR void for_each(index_sequence<Is...>, Tuple&& t, F&& f) {
+  using std::get;
+  // Using a free function get<Is>(Tuple) now.
+  const int unused[] = {0, ((void)f(get<Is>(t)), 0)...};
+  ignore_unused(unused);
+}
+
+template <typename Tuple, typename F>
+FMT_CONSTEXPR void for_each(Tuple&& t, F&& f) {
+  for_each(tuple_index_sequence<remove_cvref_t<Tuple>>(),
+           std::forward<Tuple>(t), std::forward<F>(f));
+}
+
+template <typename Tuple1, typename Tuple2, typename F, size_t... Is>
+void for_each2(index_sequence<Is...>, Tuple1&& t1, Tuple2&& t2, F&& f) {
+  using std::get;
+  const int unused[] = {0, ((void)f(get<Is>(t1), get<Is>(t2)), 0)...};
+  ignore_unused(unused);
+}
+
+template <typename Tuple1, typename Tuple2, typename F>
+void for_each2(Tuple1&& t1, Tuple2&& t2, F&& f) {
+  for_each2(tuple_index_sequence<remove_cvref_t<Tuple1>>(),
+            std::forward<Tuple1>(t1), std::forward<Tuple2>(t2),
+            std::forward<F>(f));
+}
+
+namespace tuple {
+// Workaround a bug in MSVC 2019 (v140).
+template <typename Char, typename... T>
+using result_t = std::tuple<formatter<remove_cvref_t<T>, Char>...>;
+
+using std::get;
+template <typename Tuple, typename Char, std::size_t... Is>
+auto get_formatters(index_sequence<Is...>)
+    -> result_t<Char, decltype(get<Is>(std::declval<Tuple>()))...>;
+}  // namespace tuple
+
+#if FMT_MSC_VERSION && FMT_MSC_VERSION < 1920
+// Older MSVC doesn't get the reference type correctly for arrays.
+template <typename R> struct range_reference_type_impl {
+  using type = decltype(*detail::range_begin(std::declval<R&>()));
+};
+
+template <typename T, std::size_t N> struct range_reference_type_impl<T[N]> {
+  using type = T&;
+};
+
+template <typename T>
+using range_reference_type = typename range_reference_type_impl<T>::type;
+#else
+template <typename Range>
+using range_reference_type =
+    decltype(*detail::range_begin(std::declval<Range&>()));
+#endif
+
+// We don't use the Range's value_type for anything, but we do need the Range's
+// reference type, with cv-ref stripped.
+template <typename Range>
+using uncvref_type = remove_cvref_t<range_reference_type<Range>>;
+
+template <typename Formatter>
+FMT_CONSTEXPR auto maybe_set_debug_format(Formatter& f, bool set)
+    -> decltype(f.set_debug_format(set)) {
+  f.set_debug_format(set);
+}
+template <typename Formatter>
+FMT_CONSTEXPR void maybe_set_debug_format(Formatter&, ...) {}
+
+// These are not generic lambdas for compatibility with C++11.
+template <typename ParseContext> struct parse_empty_specs {
+  template <typename Formatter> FMT_CONSTEXPR void operator()(Formatter& f) {
+    f.parse(ctx);
+    detail::maybe_set_debug_format(f, true);
+  }
+  ParseContext& ctx;
+};
+template <typename FormatContext> struct format_tuple_element {
+  using char_type = typename FormatContext::char_type;
+
+  template <typename T>
+  void operator()(const formatter<T, char_type>& f, const T& v) {
+    if (i > 0)
+      ctx.advance_to(detail::copy_str<char_type>(separator, ctx.out()));
+    ctx.advance_to(f.format(v, ctx));
+    ++i;
+  }
+
+  int i;
+  FormatContext& ctx;
+  basic_string_view<char_type> separator;
+};
+
+}  // namespace detail
+
+template <typename T> struct is_tuple_like {
+  static constexpr const bool value =
+      detail::is_tuple_like_<T>::value && !detail::is_range_<T>::value;
+};
+
+template <typename T, typename C> struct is_tuple_formattable {
+  static constexpr const bool value =
+      detail::is_tuple_formattable_<T, C>::value;
+};
+
+template <typename Tuple, typename Char>
+struct formatter<Tuple, Char,
+                 enable_if_t<fmt::is_tuple_like<Tuple>::value &&
+                             fmt::is_tuple_formattable<Tuple, Char>::value>> {
+ private:
+  decltype(detail::tuple::get_formatters<Tuple, Char>(
+      detail::tuple_index_sequence<Tuple>())) formatters_;
+
+  basic_string_view<Char> separator_ = detail::string_literal<Char, ',', ' '>{};
+  basic_string_view<Char> opening_bracket_ =
+      detail::string_literal<Char, '('>{};
+  basic_string_view<Char> closing_bracket_ =
+      detail::string_literal<Char, ')'>{};
+
+ public:
+  FMT_CONSTEXPR formatter() {}
+
+  FMT_CONSTEXPR void set_separator(basic_string_view<Char> sep) {
+    separator_ = sep;
+  }
+
+  FMT_CONSTEXPR void set_brackets(basic_string_view<Char> open,
+                                  basic_string_view<Char> close) {
+    opening_bracket_ = open;
+    closing_bracket_ = close;
+  }
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    if (it != ctx.end() && *it != '}')
+      FMT_THROW(format_error("invalid format specifier"));
+    detail::for_each(formatters_, detail::parse_empty_specs<ParseContext>{ctx});
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(const Tuple& value, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    ctx.advance_to(detail::copy_str<Char>(opening_bracket_, ctx.out()));
+    detail::for_each2(
+        formatters_, value,
+        detail::format_tuple_element<FormatContext>{0, ctx, separator_});
+    return detail::copy_str<Char>(closing_bracket_, ctx.out());
+  }
+};
+
+template <typename T, typename Char> struct is_range {
+  static constexpr const bool value =
+      detail::is_range_<T>::value && !detail::is_std_string_like<T>::value &&
+      !std::is_convertible<T, std::basic_string<Char>>::value &&
+      !std::is_convertible<T, detail::std_string_view<Char>>::value;
+};
+
+namespace detail {
+template <typename Context> struct range_mapper {
+  using mapper = arg_mapper<Context>;
+
+  template <typename T,
+            FMT_ENABLE_IF(has_formatter<remove_cvref_t<T>, Context>::value)>
+  static auto map(T&& value) -> T&& {
+    return static_cast<T&&>(value);
+  }
+  template <typename T,
+            FMT_ENABLE_IF(!has_formatter<remove_cvref_t<T>, Context>::value)>
+  static auto map(T&& value)
+      -> decltype(mapper().map(static_cast<T&&>(value))) {
+    return mapper().map(static_cast<T&&>(value));
+  }
+};
+
+template <typename Char, typename Element>
+using range_formatter_type =
+    formatter<remove_cvref_t<decltype(range_mapper<buffer_context<Char>>{}.map(
+                  std::declval<Element>()))>,
+              Char>;
+
+template <typename R>
+using maybe_const_range =
+    conditional_t<has_const_begin_end<R>::value, const R, R>;
+
+// Workaround a bug in MSVC 2015 and earlier.
+#if !FMT_MSC_VERSION || FMT_MSC_VERSION >= 1910
+template <typename R, typename Char>
+struct is_formattable_delayed
+    : is_formattable<uncvref_type<maybe_const_range<R>>, Char> {};
+#endif
+}  // namespace detail
+
+template <typename...> struct conjunction : std::true_type {};
+template <typename P> struct conjunction<P> : P {};
+template <typename P1, typename... Pn>
+struct conjunction<P1, Pn...>
+    : conditional_t<bool(P1::value), conjunction<Pn...>, P1> {};
+
+template <typename T, typename Char, typename Enable = void>
+struct range_formatter;
+
+template <typename T, typename Char>
+struct range_formatter<
+    T, Char,
+    enable_if_t<conjunction<std::is_same<T, remove_cvref_t<T>>,
+                            is_formattable<T, Char>>::value>> {
+ private:
+  detail::range_formatter_type<Char, T> underlying_;
+  basic_string_view<Char> separator_ = detail::string_literal<Char, ',', ' '>{};
+  basic_string_view<Char> opening_bracket_ =
+      detail::string_literal<Char, '['>{};
+  basic_string_view<Char> closing_bracket_ =
+      detail::string_literal<Char, ']'>{};
+
+ public:
+  FMT_CONSTEXPR range_formatter() {}
+
+  FMT_CONSTEXPR auto underlying() -> detail::range_formatter_type<Char, T>& {
+    return underlying_;
+  }
+
+  FMT_CONSTEXPR void set_separator(basic_string_view<Char> sep) {
+    separator_ = sep;
+  }
+
+  FMT_CONSTEXPR void set_brackets(basic_string_view<Char> open,
+                                  basic_string_view<Char> close) {
+    opening_bracket_ = open;
+    closing_bracket_ = close;
+  }
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    auto end = ctx.end();
+
+    if (it != end && *it == 'n') {
+      set_brackets({}, {});
+      ++it;
+    }
+
+    if (it != end && *it != '}') {
+      if (*it != ':') FMT_THROW(format_error("invalid format specifier"));
+      ++it;
+    } else {
+      detail::maybe_set_debug_format(underlying_, true);
+    }
+
+    ctx.advance_to(it);
+    return underlying_.parse(ctx);
+  }
+
+  template <typename R, typename FormatContext>
+  auto format(R&& range, FormatContext& ctx) const -> decltype(ctx.out()) {
+    detail::range_mapper<buffer_context<Char>> mapper;
+    auto out = ctx.out();
+    out = detail::copy_str<Char>(opening_bracket_, out);
+    int i = 0;
+    auto it = detail::range_begin(range);
+    auto end = detail::range_end(range);
+    for (; it != end; ++it) {
+      if (i > 0) out = detail::copy_str<Char>(separator_, out);
+      ctx.advance_to(out);
+      auto&& item = *it;
+      out = underlying_.format(mapper.map(item), ctx);
+      ++i;
+    }
+    out = detail::copy_str<Char>(closing_bracket_, out);
+    return out;
+  }
+};
+
+enum class range_format { disabled, map, set, sequence, string, debug_string };
+
+namespace detail {
+template <typename T>
+struct range_format_kind_
+    : std::integral_constant<range_format,
+                             std::is_same<uncvref_type<T>, T>::value
+                                 ? range_format::disabled
+                             : is_map<T>::value ? range_format::map
+                             : is_set<T>::value ? range_format::set
+                                                : range_format::sequence> {};
+
+template <range_format K, typename R, typename Char, typename Enable = void>
+struct range_default_formatter;
+
+template <range_format K>
+using range_format_constant = std::integral_constant<range_format, K>;
+
+template <range_format K, typename R, typename Char>
+struct range_default_formatter<
+    K, R, Char,
+    enable_if_t<(K == range_format::sequence || K == range_format::map ||
+                 K == range_format::set)>> {
+  using range_type = detail::maybe_const_range<R>;
+  range_formatter<detail::uncvref_type<range_type>, Char> underlying_;
+
+  FMT_CONSTEXPR range_default_formatter() { init(range_format_constant<K>()); }
+
+  FMT_CONSTEXPR void init(range_format_constant<range_format::set>) {
+    underlying_.set_brackets(detail::string_literal<Char, '{'>{},
+                             detail::string_literal<Char, '}'>{});
+  }
+
+  FMT_CONSTEXPR void init(range_format_constant<range_format::map>) {
+    underlying_.set_brackets(detail::string_literal<Char, '{'>{},
+                             detail::string_literal<Char, '}'>{});
+    underlying_.underlying().set_brackets({}, {});
+    underlying_.underlying().set_separator(
+        detail::string_literal<Char, ':', ' '>{});
+  }
+
+  FMT_CONSTEXPR void init(range_format_constant<range_format::sequence>) {}
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return underlying_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(range_type& range, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return underlying_.format(range, ctx);
+  }
+};
+}  // namespace detail
+
+template <typename T, typename Char, typename Enable = void>
+struct range_format_kind
+    : conditional_t<
+          is_range<T, Char>::value, detail::range_format_kind_<T>,
+          std::integral_constant<range_format, range_format::disabled>> {};
+
+template <typename R, typename Char>
+struct formatter<
+    R, Char,
+    enable_if_t<conjunction<bool_constant<range_format_kind<R, Char>::value !=
+                                          range_format::disabled>
+// Workaround a bug in MSVC 2015 and earlier.
+#if !FMT_MSC_VERSION || FMT_MSC_VERSION >= 1910
+                            ,
+                            detail::is_formattable_delayed<R, Char>
+#endif
+                            >::value>>
+    : detail::range_default_formatter<range_format_kind<R, Char>::value, R,
+                                      Char> {
+};
+
+template <typename Char, typename... T> struct tuple_join_view : detail::view {
+  const std::tuple<T...>& tuple;
+  basic_string_view<Char> sep;
+
+  tuple_join_view(const std::tuple<T...>& t, basic_string_view<Char> s)
+      : tuple(t), sep{s} {}
+};
+
+// Define FMT_TUPLE_JOIN_SPECIFIERS to enable experimental format specifiers
+// support in tuple_join. It is disabled by default because of issues with
+// the dynamic width and precision.
+#ifndef FMT_TUPLE_JOIN_SPECIFIERS
+#  define FMT_TUPLE_JOIN_SPECIFIERS 0
+#endif
+
+template <typename Char, typename... T>
+struct formatter<tuple_join_view<Char, T...>, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return do_parse(ctx, std::integral_constant<size_t, sizeof...(T)>());
+  }
+
+  template <typename FormatContext>
+  auto format(const tuple_join_view<Char, T...>& value,
+              FormatContext& ctx) const -> typename FormatContext::iterator {
+    return do_format(value, ctx,
+                     std::integral_constant<size_t, sizeof...(T)>());
+  }
+
+ private:
+  std::tuple<formatter<typename std::decay<T>::type, Char>...> formatters_;
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto do_parse(ParseContext& ctx,
+                              std::integral_constant<size_t, 0>)
+      -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename ParseContext, size_t N>
+  FMT_CONSTEXPR auto do_parse(ParseContext& ctx,
+                              std::integral_constant<size_t, N>)
+      -> decltype(ctx.begin()) {
+    auto end = ctx.begin();
+#if FMT_TUPLE_JOIN_SPECIFIERS
+    end = std::get<sizeof...(T) - N>(formatters_).parse(ctx);
+    if (N > 1) {
+      auto end1 = do_parse(ctx, std::integral_constant<size_t, N - 1>());
+      if (end != end1)
+        FMT_THROW(format_error("incompatible format specs for tuple elements"));
+    }
+#endif
+    return end;
+  }
+
+  template <typename FormatContext>
+  auto do_format(const tuple_join_view<Char, T...>&, FormatContext& ctx,
+                 std::integral_constant<size_t, 0>) const ->
+      typename FormatContext::iterator {
+    return ctx.out();
+  }
+
+  template <typename FormatContext, size_t N>
+  auto do_format(const tuple_join_view<Char, T...>& value, FormatContext& ctx,
+                 std::integral_constant<size_t, N>) const ->
+      typename FormatContext::iterator {
+    auto out = std::get<sizeof...(T) - N>(formatters_)
+                   .format(std::get<sizeof...(T) - N>(value.tuple), ctx);
+    if (N > 1) {
+      out = std::copy(value.sep.begin(), value.sep.end(), out);
+      ctx.advance_to(out);
+      return do_format(value, ctx, std::integral_constant<size_t, N - 1>());
+    }
+    return out;
+  }
+};
+
+namespace detail {
+// Check if T has an interface like a container adaptor (e.g. std::stack,
+// std::queue, std::priority_queue).
+template <typename T> class is_container_adaptor_like {
+  template <typename U> static auto check(U* p) -> typename U::container_type;
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+};
+
+template <typename Container> struct all {
+  const Container& c;
+  auto begin() const -> typename Container::const_iterator { return c.begin(); }
+  auto end() const -> typename Container::const_iterator { return c.end(); }
+};
+}  // namespace detail
+
+template <typename T, typename Char>
+struct formatter<
+    T, Char,
+    enable_if_t<conjunction<detail::is_container_adaptor_like<T>,
+                            bool_constant<range_format_kind<T, Char>::value ==
+                                          range_format::disabled>>::value>>
+    : formatter<detail::all<typename T::container_type>, Char> {
+  using all = detail::all<typename T::container_type>;
+  template <typename FormatContext>
+  auto format(const T& t, FormatContext& ctx) const -> decltype(ctx.out()) {
+    struct getter : T {
+      static auto get(const T& t) -> all {
+        return {t.*(&getter::c)};  // Access c through the derived class.
+      }
+    };
+    return formatter<all>::format(getter::get(t), ctx);
+  }
+};
+
+FMT_BEGIN_EXPORT
+
+/**
+  \rst
+  Returns an object that formats `tuple` with elements separated by `sep`.
+
+  **Example**::
+
+    std::tuple<int, char> t = {1, 'a'};
+    fmt::print("{}", fmt::join(t, ", "));
+    // Output: "1, a"
+  \endrst
+ */
+template <typename... T>
+FMT_CONSTEXPR auto join(const std::tuple<T...>& tuple, string_view sep)
+    -> tuple_join_view<char, T...> {
+  return {tuple, sep};
+}
+
+template <typename... T>
+FMT_CONSTEXPR auto join(const std::tuple<T...>& tuple,
+                        basic_string_view<wchar_t> sep)
+    -> tuple_join_view<wchar_t, T...> {
+  return {tuple, sep};
+}
+
+/**
+  \rst
+  Returns an object that formats `initializer_list` with elements separated by
+  `sep`.
+
+  **Example**::
+
+    fmt::print("{}", fmt::join({1, 2, 3}, ", "));
+    // Output: "1, 2, 3"
+  \endrst
+ */
+template <typename T>
+auto join(std::initializer_list<T> list, string_view sep)
+    -> join_view<const T*, const T*> {
+  return join(std::begin(list), std::end(list), sep);
+}
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_RANGES_H_
diff --git a/dep/fmt/include/fmt/std.h b/dep/fmt/include/fmt/std.h
new file mode 100644
index 00000000..7cff1159
--- /dev/null
+++ b/dep/fmt/include/fmt/std.h
@@ -0,0 +1,537 @@
+// Formatting library for C++ - formatters for standard library types
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_STD_H_
+#define FMT_STD_H_
+
+#include <atomic>
+#include <bitset>
+#include <cstdlib>
+#include <exception>
+#include <memory>
+#include <thread>
+#include <type_traits>
+#include <typeinfo>
+#include <utility>
+#include <vector>
+
+#include "format.h"
+#include "ostream.h"
+
+#if FMT_HAS_INCLUDE(<version>)
+#  include <version>
+#endif
+// Checking FMT_CPLUSPLUS for warning suppression in MSVC.
+#if FMT_CPLUSPLUS >= 201703L
+#  if FMT_HAS_INCLUDE(<filesystem>)
+#    include <filesystem>
+#  endif
+#  if FMT_HAS_INCLUDE(<variant>)
+#    include <variant>
+#  endif
+#  if FMT_HAS_INCLUDE(<optional>)
+#    include <optional>
+#  endif
+#endif
+
+#if FMT_CPLUSPLUS > 201703L && FMT_HAS_INCLUDE(<source_location>)
+#  include <source_location>
+#endif
+
+// GCC 4 does not support FMT_HAS_INCLUDE.
+#if FMT_HAS_INCLUDE(<cxxabi.h>) || defined(__GLIBCXX__)
+#  include <cxxabi.h>
+// Android NDK with gabi++ library on some architectures does not implement
+// abi::__cxa_demangle().
+#  ifndef __GABIXX_CXXABI_H__
+#    define FMT_HAS_ABI_CXA_DEMANGLE
+#  endif
+#endif
+
+// Check if typeid is available.
+#ifndef FMT_USE_TYPEID
+// __RTTI is for EDG compilers. In MSVC typeid is available without RTTI.
+#  if defined(__GXX_RTTI) || FMT_HAS_FEATURE(cxx_rtti) || FMT_MSC_VERSION || \
+      defined(__INTEL_RTTI__) || defined(__RTTI)
+#    define FMT_USE_TYPEID 1
+#  else
+#    define FMT_USE_TYPEID 0
+#  endif
+#endif
+
+// For older Xcode versions, __cpp_lib_xxx flags are inaccurately defined.
+#ifndef FMT_CPP_LIB_FILESYSTEM
+#  ifdef __cpp_lib_filesystem
+#    define FMT_CPP_LIB_FILESYSTEM __cpp_lib_filesystem
+#  else
+#    define FMT_CPP_LIB_FILESYSTEM 0
+#  endif
+#endif
+
+#ifndef FMT_CPP_LIB_VARIANT
+#  ifdef __cpp_lib_variant
+#    define FMT_CPP_LIB_VARIANT __cpp_lib_variant
+#  else
+#    define FMT_CPP_LIB_VARIANT 0
+#  endif
+#endif
+
+#if FMT_CPP_LIB_FILESYSTEM
+FMT_BEGIN_NAMESPACE
+
+namespace detail {
+
+template <typename Char, typename PathChar>
+auto get_path_string(const std::filesystem::path& p,
+                     const std::basic_string<PathChar>& native) {
+  if constexpr (std::is_same_v<Char, char> && std::is_same_v<PathChar, wchar_t>)
+    return to_utf8<wchar_t>(native, to_utf8_error_policy::replace);
+  else
+    return p.string<Char>();
+}
+
+template <typename Char, typename PathChar>
+void write_escaped_path(basic_memory_buffer<Char>& quoted,
+                        const std::filesystem::path& p,
+                        const std::basic_string<PathChar>& native) {
+  if constexpr (std::is_same_v<Char, char> &&
+                std::is_same_v<PathChar, wchar_t>) {
+    auto buf = basic_memory_buffer<wchar_t>();
+    write_escaped_string<wchar_t>(std::back_inserter(buf), native);
+    bool valid = to_utf8<wchar_t>::convert(quoted, {buf.data(), buf.size()});
+    FMT_ASSERT(valid, "invalid utf16");
+  } else if constexpr (std::is_same_v<Char, PathChar>) {
+    write_escaped_string<std::filesystem::path::value_type>(
+        std::back_inserter(quoted), native);
+  } else {
+    write_escaped_string<Char>(std::back_inserter(quoted), p.string<Char>());
+  }
+}
+
+}  // namespace detail
+
+FMT_EXPORT
+template <typename Char> struct formatter<std::filesystem::path, Char> {
+ private:
+  format_specs<Char> specs_;
+  detail::arg_ref<Char> width_ref_;
+  bool debug_ = false;
+  char path_type_ = 0;
+
+ public:
+  FMT_CONSTEXPR void set_debug_format(bool set = true) { debug_ = set; }
+
+  template <typename ParseContext> FMT_CONSTEXPR auto parse(ParseContext& ctx) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it == end) return it;
+
+    it = detail::parse_align(it, end, specs_);
+    if (it == end) return it;
+
+    it = detail::parse_dynamic_spec(it, end, specs_.width, width_ref_, ctx);
+    if (it != end && *it == '?') {
+      debug_ = true;
+      ++it;
+    }
+    if (it != end && (*it == 'g')) path_type_ = *it++;
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(const std::filesystem::path& p, FormatContext& ctx) const {
+    auto specs = specs_;
+#  ifdef _WIN32
+    auto path_string = !path_type_ ? p.native() : p.generic_wstring();
+#  else
+    auto path_string = !path_type_ ? p.native() : p.generic_string();
+#  endif
+
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref_,
+                                                       ctx);
+    if (!debug_) {
+      auto s = detail::get_path_string<Char>(p, path_string);
+      return detail::write(ctx.out(), basic_string_view<Char>(s), specs);
+    }
+    auto quoted = basic_memory_buffer<Char>();
+    detail::write_escaped_path(quoted, p, path_string);
+    return detail::write(ctx.out(),
+                         basic_string_view<Char>(quoted.data(), quoted.size()),
+                         specs);
+  }
+};
+FMT_END_NAMESPACE
+#endif  // FMT_CPP_LIB_FILESYSTEM
+
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <std::size_t N, typename Char>
+struct formatter<std::bitset<N>, Char> : nested_formatter<string_view> {
+ private:
+  // Functor because C++11 doesn't support generic lambdas.
+  struct writer {
+    const std::bitset<N>& bs;
+
+    template <typename OutputIt>
+    FMT_CONSTEXPR auto operator()(OutputIt out) -> OutputIt {
+      for (auto pos = N; pos > 0; --pos) {
+        out = detail::write<Char>(out, bs[pos - 1] ? Char('1') : Char('0'));
+      }
+
+      return out;
+    }
+  };
+
+ public:
+  template <typename FormatContext>
+  auto format(const std::bitset<N>& bs, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return write_padded(ctx, writer{bs});
+  }
+};
+
+FMT_EXPORT
+template <typename Char>
+struct formatter<std::thread::id, Char> : basic_ostream_formatter<Char> {};
+FMT_END_NAMESPACE
+
+#ifdef __cpp_lib_optional
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <typename T, typename Char>
+struct formatter<std::optional<T>, Char,
+                 std::enable_if_t<is_formattable<T, Char>::value>> {
+ private:
+  formatter<T, Char> underlying_;
+  static constexpr basic_string_view<Char> optional =
+      detail::string_literal<Char, 'o', 'p', 't', 'i', 'o', 'n', 'a', 'l',
+                             '('>{};
+  static constexpr basic_string_view<Char> none =
+      detail::string_literal<Char, 'n', 'o', 'n', 'e'>{};
+
+  template <class U>
+  FMT_CONSTEXPR static auto maybe_set_debug_format(U& u, bool set)
+      -> decltype(u.set_debug_format(set)) {
+    u.set_debug_format(set);
+  }
+
+  template <class U>
+  FMT_CONSTEXPR static void maybe_set_debug_format(U&, ...) {}
+
+ public:
+  template <typename ParseContext> FMT_CONSTEXPR auto parse(ParseContext& ctx) {
+    maybe_set_debug_format(underlying_, true);
+    return underlying_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(const std::optional<T>& opt, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    if (!opt) return detail::write<Char>(ctx.out(), none);
+
+    auto out = ctx.out();
+    out = detail::write<Char>(out, optional);
+    ctx.advance_to(out);
+    out = underlying_.format(*opt, ctx);
+    return detail::write(out, ')');
+  }
+};
+FMT_END_NAMESPACE
+#endif  // __cpp_lib_optional
+
+#ifdef __cpp_lib_source_location
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <> struct formatter<std::source_location> {
+  template <typename ParseContext> FMT_CONSTEXPR auto parse(ParseContext& ctx) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const std::source_location& loc, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    out = detail::write(out, loc.file_name());
+    out = detail::write(out, ':');
+    out = detail::write<char>(out, loc.line());
+    out = detail::write(out, ':');
+    out = detail::write<char>(out, loc.column());
+    out = detail::write(out, ": ");
+    out = detail::write(out, loc.function_name());
+    return out;
+  }
+};
+FMT_END_NAMESPACE
+#endif
+
+#if FMT_CPP_LIB_VARIANT
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename T>
+using variant_index_sequence =
+    std::make_index_sequence<std::variant_size<T>::value>;
+
+template <typename> struct is_variant_like_ : std::false_type {};
+template <typename... Types>
+struct is_variant_like_<std::variant<Types...>> : std::true_type {};
+
+// formattable element check.
+template <typename T, typename C> class is_variant_formattable_ {
+  template <std::size_t... Is>
+  static std::conjunction<
+      is_formattable<std::variant_alternative_t<Is, T>, C>...>
+      check(std::index_sequence<Is...>);
+
+ public:
+  static constexpr const bool value =
+      decltype(check(variant_index_sequence<T>{}))::value;
+};
+
+template <typename Char, typename OutputIt, typename T>
+auto write_variant_alternative(OutputIt out, const T& v) -> OutputIt {
+  if constexpr (is_string<T>::value)
+    return write_escaped_string<Char>(out, detail::to_string_view(v));
+  else if constexpr (std::is_same_v<T, Char>)
+    return write_escaped_char(out, v);
+  else
+    return write<Char>(out, v);
+}
+
+}  // namespace detail
+
+template <typename T> struct is_variant_like {
+  static constexpr const bool value = detail::is_variant_like_<T>::value;
+};
+
+template <typename T, typename C> struct is_variant_formattable {
+  static constexpr const bool value =
+      detail::is_variant_formattable_<T, C>::value;
+};
+
+FMT_EXPORT
+template <typename Char> struct formatter<std::monostate, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const std::monostate&, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return detail::write<Char>(ctx.out(), "monostate");
+  }
+};
+
+FMT_EXPORT
+template <typename Variant, typename Char>
+struct formatter<
+    Variant, Char,
+    std::enable_if_t<std::conjunction_v<
+        is_variant_like<Variant>, is_variant_formattable<Variant, Char>>>> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const Variant& value, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+
+    out = detail::write<Char>(out, "variant(");
+    FMT_TRY {
+      std::visit(
+          [&](const auto& v) {
+            out = detail::write_variant_alternative<Char>(out, v);
+          },
+          value);
+    }
+    FMT_CATCH(const std::bad_variant_access&) {
+      detail::write<Char>(out, "valueless by exception");
+    }
+    *out++ = ')';
+    return out;
+  }
+};
+FMT_END_NAMESPACE
+#endif  // FMT_CPP_LIB_VARIANT
+
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <typename Char> struct formatter<std::error_code, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const std::error_code& ec, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    out = detail::write_bytes(out, ec.category().name(), format_specs<Char>());
+    out = detail::write<Char>(out, Char(':'));
+    out = detail::write<Char>(out, ec.value());
+    return out;
+  }
+};
+
+FMT_EXPORT
+template <typename T, typename Char>
+struct formatter<
+    T, Char,  // DEPRECATED! Mixing code unit types.
+    typename std::enable_if<std::is_base_of<std::exception, T>::value>::type> {
+ private:
+  bool with_typename_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    auto end = ctx.end();
+    if (it == end || *it == '}') return it;
+    if (*it == 't') {
+      ++it;
+      with_typename_ = FMT_USE_TYPEID != 0;
+    }
+    return it;
+  }
+
+  template <typename OutputIt>
+  auto format(const std::exception& ex,
+              basic_format_context<OutputIt, Char>& ctx) const -> OutputIt {
+    format_specs<Char> spec;
+    auto out = ctx.out();
+    if (!with_typename_)
+      return detail::write_bytes(out, string_view(ex.what()), spec);
+
+#if FMT_USE_TYPEID
+    const std::type_info& ti = typeid(ex);
+#  ifdef FMT_HAS_ABI_CXA_DEMANGLE
+    int status = 0;
+    std::size_t size = 0;
+    std::unique_ptr<char, void (*)(void*)> demangled_name_ptr(
+        abi::__cxa_demangle(ti.name(), nullptr, &size, &status), &std::free);
+
+    string_view demangled_name_view;
+    if (demangled_name_ptr) {
+      demangled_name_view = demangled_name_ptr.get();
+
+      // Normalization of stdlib inline namespace names.
+      // libc++ inline namespaces.
+      //  std::__1::*       -> std::*
+      //  std::__1::__fs::* -> std::*
+      // libstdc++ inline namespaces.
+      //  std::__cxx11::*             -> std::*
+      //  std::filesystem::__cxx11::* -> std::filesystem::*
+      if (demangled_name_view.starts_with("std::")) {
+        char* begin = demangled_name_ptr.get();
+        char* to = begin + 5;  // std::
+        for (char *from = to, *end = begin + demangled_name_view.size();
+             from < end;) {
+          // This is safe, because demangled_name is NUL-terminated.
+          if (from[0] == '_' && from[1] == '_') {
+            char* next = from + 1;
+            while (next < end && *next != ':') next++;
+            if (next[0] == ':' && next[1] == ':') {
+              from = next + 2;
+              continue;
+            }
+          }
+          *to++ = *from++;
+        }
+        demangled_name_view = {begin, detail::to_unsigned(to - begin)};
+      }
+    } else {
+      demangled_name_view = string_view(ti.name());
+    }
+    out = detail::write_bytes(out, demangled_name_view, spec);
+#  elif FMT_MSC_VERSION
+    string_view demangled_name_view(ti.name());
+    if (demangled_name_view.starts_with("class "))
+      demangled_name_view.remove_prefix(6);
+    else if (demangled_name_view.starts_with("struct "))
+      demangled_name_view.remove_prefix(7);
+    out = detail::write_bytes(out, demangled_name_view, spec);
+#  else
+    out = detail::write_bytes(out, string_view(ti.name()), spec);
+#  endif
+    *out++ = ':';
+    *out++ = ' ';
+    return detail::write_bytes(out, string_view(ex.what()), spec);
+#endif
+  }
+};
+
+namespace detail {
+
+template <typename T, typename Enable = void>
+struct has_flip : std::false_type {};
+
+template <typename T>
+struct has_flip<T, void_t<decltype(std::declval<T>().flip())>>
+    : std::true_type {};
+
+template <typename T> struct is_bit_reference_like {
+  static constexpr const bool value =
+      std::is_convertible<T, bool>::value &&
+      std::is_nothrow_assignable<T, bool>::value && has_flip<T>::value;
+};
+
+#ifdef _LIBCPP_VERSION
+
+// Workaround for libc++ incompatibility with C++ standard.
+// According to the Standard, `bitset::operator[] const` returns bool.
+template <typename C>
+struct is_bit_reference_like<std::__bit_const_reference<C>> {
+  static constexpr const bool value = true;
+};
+
+#endif
+
+}  // namespace detail
+
+// We can't use std::vector<bool, Allocator>::reference and
+// std::bitset<N>::reference because the compiler can't deduce Allocator and N
+// in partial specialization.
+FMT_EXPORT
+template <typename BitRef, typename Char>
+struct formatter<BitRef, Char,
+                 enable_if_t<detail::is_bit_reference_like<BitRef>::value>>
+    : formatter<bool, Char> {
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const BitRef& v, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return formatter<bool, Char>::format(v, ctx);
+  }
+};
+
+FMT_EXPORT
+template <typename T, typename Char>
+struct formatter<std::atomic<T>, Char,
+                 enable_if_t<is_formattable<T, Char>::value>>
+    : formatter<T, Char> {
+  template <typename FormatContext>
+  auto format(const std::atomic<T>& v, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return formatter<T, Char>::format(v.load(), ctx);
+  }
+};
+
+#ifdef __cpp_lib_atomic_flag_test
+FMT_EXPORT
+template <typename Char>
+struct formatter<std::atomic_flag, Char> : formatter<bool, Char> {
+  template <typename FormatContext>
+  auto format(const std::atomic_flag& v, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return formatter<bool, Char>::format(v.test(), ctx);
+  }
+};
+#endif  // __cpp_lib_atomic_flag_test
+
+FMT_END_NAMESPACE
+#endif  // FMT_STD_H_
diff --git a/dep/fmt/include/fmt/xchar.h b/dep/fmt/include/fmt/xchar.h
new file mode 100644
index 00000000..f609c5c4
--- /dev/null
+++ b/dep/fmt/include/fmt/xchar.h
@@ -0,0 +1,259 @@
+// Formatting library for C++ - optional wchar_t and exotic character support
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_XCHAR_H_
+#define FMT_XCHAR_H_
+
+#include <cwchar>
+
+#include "format.h"
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+#  include <locale>
+#endif
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename T>
+using is_exotic_char = bool_constant<!std::is_same<T, char>::value>;
+
+inline auto write_loc(std::back_insert_iterator<detail::buffer<wchar_t>> out,
+                      loc_value value, const format_specs<wchar_t>& specs,
+                      locale_ref loc) -> bool {
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+  auto& numpunct =
+      std::use_facet<std::numpunct<wchar_t>>(loc.get<std::locale>());
+  auto separator = std::wstring();
+  auto grouping = numpunct.grouping();
+  if (!grouping.empty()) separator = std::wstring(1, numpunct.thousands_sep());
+  return value.visit(loc_writer<wchar_t>{out, specs, separator, grouping, {}});
+#endif
+  return false;
+}
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+using wstring_view = basic_string_view<wchar_t>;
+using wformat_parse_context = basic_format_parse_context<wchar_t>;
+using wformat_context = buffer_context<wchar_t>;
+using wformat_args = basic_format_args<wformat_context>;
+using wmemory_buffer = basic_memory_buffer<wchar_t>;
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+// Workaround broken conversion on older gcc.
+template <typename... Args> using wformat_string = wstring_view;
+inline auto runtime(wstring_view s) -> wstring_view { return s; }
+#else
+template <typename... Args>
+using wformat_string = basic_format_string<wchar_t, type_identity_t<Args>...>;
+inline auto runtime(wstring_view s) -> runtime_format_string<wchar_t> {
+  return {{s}};
+}
+#endif
+
+template <> struct is_char<wchar_t> : std::true_type {};
+template <> struct is_char<detail::char8_type> : std::true_type {};
+template <> struct is_char<char16_t> : std::true_type {};
+template <> struct is_char<char32_t> : std::true_type {};
+
+template <typename... T>
+constexpr auto make_wformat_args(const T&... args)
+    -> format_arg_store<wformat_context, T...> {
+  return {args...};
+}
+
+inline namespace literals {
+#if FMT_USE_USER_DEFINED_LITERALS && !FMT_USE_NONTYPE_TEMPLATE_ARGS
+constexpr auto operator""_a(const wchar_t* s, size_t)
+    -> detail::udl_arg<wchar_t> {
+  return {s};
+}
+#endif
+}  // namespace literals
+
+template <typename It, typename Sentinel>
+auto join(It begin, Sentinel end, wstring_view sep)
+    -> join_view<It, Sentinel, wchar_t> {
+  return {begin, end, sep};
+}
+
+template <typename Range>
+auto join(Range&& range, wstring_view sep)
+    -> join_view<detail::iterator_t<Range>, detail::sentinel_t<Range>,
+                 wchar_t> {
+  return join(std::begin(range), std::end(range), sep);
+}
+
+template <typename T>
+auto join(std::initializer_list<T> list, wstring_view sep)
+    -> join_view<const T*, const T*, wchar_t> {
+  return join(std::begin(list), std::end(list), sep);
+}
+
+template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+auto vformat(basic_string_view<Char> format_str,
+             basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> std::basic_string<Char> {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vformat_to(buf, format_str, args);
+  return to_string(buf);
+}
+
+template <typename... T>
+auto format(wformat_string<T...> fmt, T&&... args) -> std::wstring {
+  return vformat(fmt::wstring_view(fmt), fmt::make_wformat_args(args...));
+}
+
+// Pass char_t as a default template parameter instead of using
+// std::basic_string<char_t<S>> to reduce the symbol size.
+template <typename S, typename... T, typename Char = char_t<S>,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value &&
+                        !std::is_same<Char, wchar_t>::value)>
+auto format(const S& format_str, T&&... args) -> std::basic_string<Char> {
+  return vformat(detail::to_string_view(format_str),
+                 fmt::make_format_args<buffer_context<Char>>(args...));
+}
+
+template <typename Locale, typename S, typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto vformat(
+    const Locale& loc, const S& format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> std::basic_string<Char> {
+  return detail::vformat(loc, detail::to_string_view(format_str), args);
+}
+
+template <typename Locale, typename S, typename... T, typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto format(const Locale& loc, const S& format_str, T&&... args)
+    -> std::basic_string<Char> {
+  return detail::vformat(loc, detail::to_string_view(format_str),
+                         fmt::make_format_args<buffer_context<Char>>(args...));
+}
+
+template <typename OutputIt, typename S, typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+auto vformat_to(OutputIt out, const S& format_str,
+                basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> OutputIt {
+  auto&& buf = detail::get_buffer<Char>(out);
+  detail::vformat_to(buf, detail::to_string_view(format_str), args);
+  return detail::get_iterator(buf, out);
+}
+
+template <typename OutputIt, typename S, typename... T,
+          typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto format_to(OutputIt out, const S& fmt, T&&... args) -> OutputIt {
+  return vformat_to(out, detail::to_string_view(fmt),
+                    fmt::make_format_args<buffer_context<Char>>(args...));
+}
+
+template <typename Locale, typename S, typename OutputIt, typename... Args,
+          typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_locale<Locale>::value&&
+                                detail::is_exotic_char<Char>::value)>
+inline auto vformat_to(
+    OutputIt out, const Locale& loc, const S& format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) -> OutputIt {
+  auto&& buf = detail::get_buffer<Char>(out);
+  vformat_to(buf, detail::to_string_view(format_str), args,
+             detail::locale_ref(loc));
+  return detail::get_iterator(buf, out);
+}
+
+template <typename OutputIt, typename Locale, typename S, typename... T,
+          typename Char = char_t<S>,
+          bool enable = detail::is_output_iterator<OutputIt, Char>::value &&
+                        detail::is_locale<Locale>::value &&
+                        detail::is_exotic_char<Char>::value>
+inline auto format_to(OutputIt out, const Locale& loc, const S& format_str,
+                      T&&... args) ->
+    typename std::enable_if<enable, OutputIt>::type {
+  return vformat_to(out, loc, detail::to_string_view(format_str),
+                    fmt::make_format_args<buffer_context<Char>>(args...));
+}
+
+template <typename OutputIt, typename Char, typename... Args,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto vformat_to_n(
+    OutputIt out, size_t n, basic_string_view<Char> format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, Char, traits>(out, n);
+  detail::vformat_to(buf, format_str, args);
+  return {buf.out(), buf.count()};
+}
+
+template <typename OutputIt, typename S, typename... T,
+          typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto format_to_n(OutputIt out, size_t n, const S& fmt, T&&... args)
+    -> format_to_n_result<OutputIt> {
+  return vformat_to_n(out, n, detail::to_string_view(fmt),
+                      fmt::make_format_args<buffer_context<Char>>(args...));
+}
+
+template <typename S, typename... T, typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_exotic_char<Char>::value)>
+inline auto formatted_size(const S& fmt, T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<Char>();
+  detail::vformat_to(buf, detail::to_string_view(fmt),
+                     fmt::make_format_args<buffer_context<Char>>(args...));
+  return buf.count();
+}
+
+inline void vprint(std::FILE* f, wstring_view fmt, wformat_args args) {
+  auto buf = wmemory_buffer();
+  detail::vformat_to(buf, fmt, args);
+  buf.push_back(L'\0');
+  if (std::fputws(buf.data(), f) == -1)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot write to file")));
+}
+
+inline void vprint(wstring_view fmt, wformat_args args) {
+  vprint(stdout, fmt, args);
+}
+
+template <typename... T>
+void print(std::FILE* f, wformat_string<T...> fmt, T&&... args) {
+  return vprint(f, wstring_view(fmt), fmt::make_wformat_args(args...));
+}
+
+template <typename... T> void print(wformat_string<T...> fmt, T&&... args) {
+  return vprint(wstring_view(fmt), fmt::make_wformat_args(args...));
+}
+
+template <typename... T>
+void println(std::FILE* f, wformat_string<T...> fmt, T&&... args) {
+  return print(f, L"{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+template <typename... T> void println(wformat_string<T...> fmt, T&&... args) {
+  return print(L"{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+/**
+  Converts *value* to ``std::wstring`` using the default format for type *T*.
+ */
+template <typename T> inline auto to_wstring(const T& value) -> std::wstring {
+  return format(FMT_STRING(L"{}"), value);
+}
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_XCHAR_H_
diff --git a/dep/fmt/src/fmt.cc b/dep/fmt/src/fmt.cc
new file mode 100644
index 00000000..5330463a
--- /dev/null
+++ b/dep/fmt/src/fmt.cc
@@ -0,0 +1,108 @@
+module;
+
+// Put all implementation-provided headers into the global module fragment
+// to prevent attachment to this module.
+#include <algorithm>
+#include <cerrno>
+#include <chrono>
+#include <climits>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <exception>
+#include <filesystem>
+#include <fstream>
+#include <functional>
+#include <iterator>
+#include <limits>
+#include <locale>
+#include <memory>
+#include <optional>
+#include <ostream>
+#include <stdexcept>
+#include <string>
+#include <string_view>
+#include <system_error>
+#include <thread>
+#include <type_traits>
+#include <typeinfo>
+#include <utility>
+#include <variant>
+#include <vector>
+#include <version>
+
+#if __has_include(<cxxabi.h>)
+#  include <cxxabi.h>
+#endif
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#  include <intrin.h>
+#endif
+#if defined __APPLE__ || defined(__FreeBSD__)
+#  include <xlocale.h>
+#endif
+#if __has_include(<winapifamily.h>)
+#  include <winapifamily.h>
+#endif
+#if (__has_include(<fcntl.h>) || defined(__APPLE__) || \
+     defined(__linux__)) &&                            \
+    (!defined(WINAPI_FAMILY) || (WINAPI_FAMILY == WINAPI_FAMILY_DESKTOP_APP))
+#  include <fcntl.h>
+#  include <sys/stat.h>
+#  include <sys/types.h>
+#  ifndef _WIN32
+#    include <unistd.h>
+#  else
+#    include <io.h>
+#  endif
+#endif
+#ifdef _WIN32
+#  if defined(__GLIBCXX__)
+#    include <ext/stdio_filebuf.h>
+#    include <ext/stdio_sync_filebuf.h>
+#  endif
+#  define WIN32_LEAN_AND_MEAN
+#  include <windows.h>
+#endif
+
+export module fmt;
+
+#define FMT_EXPORT export
+#define FMT_BEGIN_EXPORT export {
+#define FMT_END_EXPORT }
+
+// If you define FMT_ATTACH_TO_GLOBAL_MODULE
+//  - all declarations are detached from module 'fmt'
+//  - the module behaves like a traditional static library, too
+//  - all library symbols are mangled traditionally
+//  - you can mix TUs with either importing or #including the {fmt} API
+#ifdef FMT_ATTACH_TO_GLOBAL_MODULE
+extern "C++" {
+#endif
+
+// All library-provided declarations and definitions must be in the module
+// purview to be exported.
+#include "fmt/args.h"
+#include "fmt/chrono.h"
+#include "fmt/color.h"
+#include "fmt/compile.h"
+#include "fmt/format.h"
+#include "fmt/os.h"
+#include "fmt/printf.h"
+#include "fmt/std.h"
+#include "fmt/xchar.h"
+
+#ifdef FMT_ATTACH_TO_GLOBAL_MODULE
+}
+#endif
+
+// gcc doesn't yet implement private module fragments
+#if !FMT_GCC_VERSION
+module :private;
+#endif
+
+#include "format.cc"
+#include "os.cc"
diff --git a/dep/fmt/src/format.cc b/dep/fmt/src/format.cc
new file mode 100644
index 00000000..391d3a24
--- /dev/null
+++ b/dep/fmt/src/format.cc
@@ -0,0 +1,43 @@
+// Formatting library for C++
+//
+// Copyright (c) 2012 - 2016, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#include "fmt/format-inl.h"
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template FMT_API auto dragonbox::to_decimal(float x) noexcept
+    -> dragonbox::decimal_fp<float>;
+template FMT_API auto dragonbox::to_decimal(double x) noexcept
+    -> dragonbox::decimal_fp<double>;
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+template FMT_API locale_ref::locale_ref(const std::locale& loc);
+template FMT_API auto locale_ref::get<std::locale>() const -> std::locale;
+#endif
+
+// Explicit instantiations for char.
+
+template FMT_API auto thousands_sep_impl(locale_ref)
+    -> thousands_sep_result<char>;
+template FMT_API auto decimal_point_impl(locale_ref) -> char;
+
+template FMT_API void buffer<char>::append(const char*, const char*);
+
+template FMT_API void vformat_to(buffer<char>&, string_view,
+                                 typename vformat_args<>::type, locale_ref);
+
+// Explicit instantiations for wchar_t.
+
+template FMT_API auto thousands_sep_impl(locale_ref)
+    -> thousands_sep_result<wchar_t>;
+template FMT_API auto decimal_point_impl(locale_ref) -> wchar_t;
+
+template FMT_API void buffer<wchar_t>::append(const wchar_t*, const wchar_t*);
+
+}  // namespace detail
+FMT_END_NAMESPACE
diff --git a/dep/fmt/src/os.cc b/dep/fmt/src/os.cc
new file mode 100644
index 00000000..a639e78c
--- /dev/null
+++ b/dep/fmt/src/os.cc
@@ -0,0 +1,402 @@
+// Formatting library for C++ - optional OS-specific functionality
+//
+// Copyright (c) 2012 - 2016, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+// Disable bogus MSVC warnings.
+#if !defined(_CRT_SECURE_NO_WARNINGS) && defined(_MSC_VER)
+#  define _CRT_SECURE_NO_WARNINGS
+#endif
+
+#include "fmt/os.h"
+
+#include <climits>
+
+#if FMT_USE_FCNTL
+#  include <sys/stat.h>
+#  include <sys/types.h>
+
+#  ifdef _WRS_KERNEL    // VxWorks7 kernel
+#    include <ioLib.h>  // getpagesize
+#  endif
+
+#  ifndef _WIN32
+#    include <unistd.h>
+#  else
+#    ifndef WIN32_LEAN_AND_MEAN
+#      define WIN32_LEAN_AND_MEAN
+#    endif
+#    include <io.h>
+
+#    ifndef S_IRUSR
+#      define S_IRUSR _S_IREAD
+#    endif
+#    ifndef S_IWUSR
+#      define S_IWUSR _S_IWRITE
+#    endif
+#    ifndef S_IRGRP
+#      define S_IRGRP 0
+#    endif
+#    ifndef S_IWGRP
+#      define S_IWGRP 0
+#    endif
+#    ifndef S_IROTH
+#      define S_IROTH 0
+#    endif
+#    ifndef S_IWOTH
+#      define S_IWOTH 0
+#    endif
+#  endif  // _WIN32
+#endif    // FMT_USE_FCNTL
+
+#ifdef _WIN32
+#  include <windows.h>
+#endif
+
+namespace {
+#ifdef _WIN32
+// Return type of read and write functions.
+using rwresult = int;
+
+// On Windows the count argument to read and write is unsigned, so convert
+// it from size_t preventing integer overflow.
+inline unsigned convert_rwcount(std::size_t count) {
+  return count <= UINT_MAX ? static_cast<unsigned>(count) : UINT_MAX;
+}
+#elif FMT_USE_FCNTL
+// Return type of read and write functions.
+using rwresult = ssize_t;
+
+inline std::size_t convert_rwcount(std::size_t count) { return count; }
+#endif
+}  // namespace
+
+FMT_BEGIN_NAMESPACE
+
+#ifdef _WIN32
+namespace detail {
+
+class system_message {
+  system_message(const system_message&) = delete;
+  void operator=(const system_message&) = delete;
+
+  unsigned long result_;
+  wchar_t* message_;
+
+  static bool is_whitespace(wchar_t c) noexcept {
+    return c == L' ' || c == L'\n' || c == L'\r' || c == L'\t' || c == L'\0';
+  }
+
+ public:
+  explicit system_message(unsigned long error_code)
+      : result_(0), message_(nullptr) {
+    result_ = FormatMessageW(
+        FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM |
+            FORMAT_MESSAGE_IGNORE_INSERTS,
+        nullptr, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
+        reinterpret_cast<wchar_t*>(&message_), 0, nullptr);
+    if (result_ != 0) {
+      while (result_ != 0 && is_whitespace(message_[result_ - 1])) {
+        --result_;
+      }
+    }
+  }
+  ~system_message() { LocalFree(message_); }
+  explicit operator bool() const noexcept { return result_ != 0; }
+  operator basic_string_view<wchar_t>() const noexcept {
+    return basic_string_view<wchar_t>(message_, result_);
+  }
+};
+
+class utf8_system_category final : public std::error_category {
+ public:
+  const char* name() const noexcept override { return "system"; }
+  std::string message(int error_code) const override {
+    auto&& msg = system_message(error_code);
+    if (msg) {
+      auto utf8_message = to_utf8<wchar_t>();
+      if (utf8_message.convert(msg)) {
+        return utf8_message.str();
+      }
+    }
+    return "unknown error";
+  }
+};
+
+}  // namespace detail
+
+FMT_API const std::error_category& system_category() noexcept {
+  static const detail::utf8_system_category category;
+  return category;
+}
+
+std::system_error vwindows_error(int err_code, string_view format_str,
+                                 format_args args) {
+  auto ec = std::error_code(err_code, system_category());
+  return std::system_error(ec, vformat(format_str, args));
+}
+
+void detail::format_windows_error(detail::buffer<char>& out, int error_code,
+                                  const char* message) noexcept {
+  FMT_TRY {
+    auto&& msg = system_message(error_code);
+    if (msg) {
+      auto utf8_message = to_utf8<wchar_t>();
+      if (utf8_message.convert(msg)) {
+        fmt::format_to(appender(out), FMT_STRING("{}: {}"), message,
+                       string_view(utf8_message));
+        return;
+      }
+    }
+  }
+  FMT_CATCH(...) {}
+  format_error_code(out, error_code, message);
+}
+
+void report_windows_error(int error_code, const char* message) noexcept {
+  report_error(detail::format_windows_error, error_code, message);
+}
+#endif  // _WIN32
+
+buffered_file::~buffered_file() noexcept {
+  if (file_ && FMT_SYSTEM(fclose(file_)) != 0)
+    report_system_error(errno, "cannot close file");
+}
+
+buffered_file::buffered_file(cstring_view filename, cstring_view mode) {
+  FMT_RETRY_VAL(file_, FMT_SYSTEM(fopen(filename.c_str(), mode.c_str())),
+                nullptr);
+  if (!file_)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot open file {}"),
+                           filename.c_str()));
+}
+
+void buffered_file::close() {
+  if (!file_) return;
+  int result = FMT_SYSTEM(fclose(file_));
+  file_ = nullptr;
+  if (result != 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot close file")));
+}
+
+int buffered_file::descriptor() const {
+#if !defined(fileno)
+  int fd = FMT_POSIX_CALL(fileno(file_));
+#elif defined(FMT_HAS_SYSTEM)
+  // fileno is a macro on OpenBSD so we cannot use FMT_POSIX_CALL.
+#  define FMT_DISABLE_MACRO
+  int fd = FMT_SYSTEM(fileno FMT_DISABLE_MACRO(file_));
+#else
+  int fd = fileno(file_);
+#endif
+  if (fd == -1)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot get file descriptor")));
+  return fd;
+}
+
+#if FMT_USE_FCNTL
+#  ifdef _WIN32
+using mode_t = int;
+#  endif
+constexpr mode_t default_open_mode =
+    S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH | S_IWOTH;
+
+file::file(cstring_view path, int oflag) {
+#  if defined(_WIN32) && !defined(__MINGW32__)
+  fd_ = -1;
+  auto converted = detail::utf8_to_utf16(string_view(path.c_str()));
+  *this = file::open_windows_file(converted.c_str(), oflag);
+#  else
+  FMT_RETRY(fd_, FMT_POSIX_CALL(open(path.c_str(), oflag, default_open_mode)));
+  if (fd_ == -1)
+    FMT_THROW(
+        system_error(errno, FMT_STRING("cannot open file {}"), path.c_str()));
+#  endif
+}
+
+file::~file() noexcept {
+  // Don't retry close in case of EINTR!
+  // See http://linux.derkeiler.com/Mailing-Lists/Kernel/2005-09/3000.html
+  if (fd_ != -1 && FMT_POSIX_CALL(close(fd_)) != 0)
+    report_system_error(errno, "cannot close file");
+}
+
+void file::close() {
+  if (fd_ == -1) return;
+  // Don't retry close in case of EINTR!
+  // See http://linux.derkeiler.com/Mailing-Lists/Kernel/2005-09/3000.html
+  int result = FMT_POSIX_CALL(close(fd_));
+  fd_ = -1;
+  if (result != 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot close file")));
+}
+
+long long file::size() const {
+#  ifdef _WIN32
+  // Use GetFileSize instead of GetFileSizeEx for the case when _WIN32_WINNT
+  // is less than 0x0500 as is the case with some default MinGW builds.
+  // Both functions support large file sizes.
+  DWORD size_upper = 0;
+  HANDLE handle = reinterpret_cast<HANDLE>(_get_osfhandle(fd_));
+  DWORD size_lower = FMT_SYSTEM(GetFileSize(handle, &size_upper));
+  if (size_lower == INVALID_FILE_SIZE) {
+    DWORD error = GetLastError();
+    if (error != NO_ERROR)
+      FMT_THROW(windows_error(GetLastError(), "cannot get file size"));
+  }
+  unsigned long long long_size = size_upper;
+  return (long_size << sizeof(DWORD) * CHAR_BIT) | size_lower;
+#  else
+  using Stat = struct stat;
+  Stat file_stat = Stat();
+  if (FMT_POSIX_CALL(fstat(fd_, &file_stat)) == -1)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot get file attributes")));
+  static_assert(sizeof(long long) >= sizeof(file_stat.st_size),
+                "return type of file::size is not large enough");
+  return file_stat.st_size;
+#  endif
+}
+
+std::size_t file::read(void* buffer, std::size_t count) {
+  rwresult result = 0;
+  FMT_RETRY(result, FMT_POSIX_CALL(read(fd_, buffer, convert_rwcount(count))));
+  if (result < 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot read from file")));
+  return detail::to_unsigned(result);
+}
+
+std::size_t file::write(const void* buffer, std::size_t count) {
+  rwresult result = 0;
+  FMT_RETRY(result, FMT_POSIX_CALL(write(fd_, buffer, convert_rwcount(count))));
+  if (result < 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot write to file")));
+  return detail::to_unsigned(result);
+}
+
+file file::dup(int fd) {
+  // Don't retry as dup doesn't return EINTR.
+  // http://pubs.opengroup.org/onlinepubs/009695399/functions/dup.html
+  int new_fd = FMT_POSIX_CALL(dup(fd));
+  if (new_fd == -1)
+    FMT_THROW(system_error(
+        errno, FMT_STRING("cannot duplicate file descriptor {}"), fd));
+  return file(new_fd);
+}
+
+void file::dup2(int fd) {
+  int result = 0;
+  FMT_RETRY(result, FMT_POSIX_CALL(dup2(fd_, fd)));
+  if (result == -1) {
+    FMT_THROW(system_error(
+        errno, FMT_STRING("cannot duplicate file descriptor {} to {}"), fd_,
+        fd));
+  }
+}
+
+void file::dup2(int fd, std::error_code& ec) noexcept {
+  int result = 0;
+  FMT_RETRY(result, FMT_POSIX_CALL(dup2(fd_, fd)));
+  if (result == -1) ec = std::error_code(errno, std::generic_category());
+}
+
+void file::pipe(file& read_end, file& write_end) {
+  // Close the descriptors first to make sure that assignments don't throw
+  // and there are no leaks.
+  read_end.close();
+  write_end.close();
+  int fds[2] = {};
+#  ifdef _WIN32
+  // Make the default pipe capacity same as on Linux 2.6.11+.
+  enum { DEFAULT_CAPACITY = 65536 };
+  int result = FMT_POSIX_CALL(pipe(fds, DEFAULT_CAPACITY, _O_BINARY));
+#  else
+  // Don't retry as the pipe function doesn't return EINTR.
+  // http://pubs.opengroup.org/onlinepubs/009696799/functions/pipe.html
+  int result = FMT_POSIX_CALL(pipe(fds));
+#  endif
+  if (result != 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot create pipe")));
+  // The following assignments don't throw because read_fd and write_fd
+  // are closed.
+  read_end = file(fds[0]);
+  write_end = file(fds[1]);
+}
+
+buffered_file file::fdopen(const char* mode) {
+// Don't retry as fdopen doesn't return EINTR.
+#  if defined(__MINGW32__) && defined(_POSIX_)
+  FILE* f = ::fdopen(fd_, mode);
+#  else
+  FILE* f = FMT_POSIX_CALL(fdopen(fd_, mode));
+#  endif
+  if (!f) {
+    FMT_THROW(system_error(
+        errno, FMT_STRING("cannot associate stream with file descriptor")));
+  }
+  buffered_file bf(f);
+  fd_ = -1;
+  return bf;
+}
+
+#  if defined(_WIN32) && !defined(__MINGW32__)
+file file::open_windows_file(wcstring_view path, int oflag) {
+  int fd = -1;
+  auto err = _wsopen_s(&fd, path.c_str(), oflag, _SH_DENYNO, default_open_mode);
+  if (fd == -1) {
+    FMT_THROW(system_error(err, FMT_STRING("cannot open file {}"),
+                           detail::to_utf8<wchar_t>(path.c_str()).c_str()));
+  }
+  return file(fd);
+}
+#  endif
+
+#  if !defined(__MSDOS__)
+long getpagesize() {
+#    ifdef _WIN32
+  SYSTEM_INFO si;
+  GetSystemInfo(&si);
+  return si.dwPageSize;
+#    else
+#      ifdef _WRS_KERNEL
+  long size = FMT_POSIX_CALL(getpagesize());
+#      else
+  long size = FMT_POSIX_CALL(sysconf(_SC_PAGESIZE));
+#      endif
+
+  if (size < 0)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot get memory page size")));
+  return size;
+#    endif
+}
+#  endif
+
+namespace detail {
+
+void file_buffer::grow(size_t) {
+  if (this->size() == this->capacity()) flush();
+}
+
+file_buffer::file_buffer(cstring_view path,
+                         const detail::ostream_params& params)
+    : file_(path, params.oflag) {
+  set(new char[params.buffer_size], params.buffer_size);
+}
+
+file_buffer::file_buffer(file_buffer&& other)
+    : detail::buffer<char>(other.data(), other.size(), other.capacity()),
+      file_(std::move(other.file_)) {
+  other.clear();
+  other.set(nullptr, 0);
+}
+
+file_buffer::~file_buffer() {
+  flush();
+  delete[] data();
+}
+}  // namespace detail
+
+ostream::~ostream() = default;
+#endif  // FMT_USE_FCNTL
+FMT_END_NAMESPACE

From d960b020ea86a6a8d2e202c3fc76e647d0482fc4 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 19:14:02 +0100
Subject: [PATCH 02/47] Get the executables building on WSL2 Fedora.

---
 Makefile             | 41 +++++++++++++++++---
 README.md            | 78 ++++++++++++++++++-------------------
 build.py             | 12 ++++--
 build/ab.mk          |  2 +-
 build/ab.py          |  1 +
 build/c.py           | 91 +++++++++++++++++++++++++++++++++++++-------
 build/pkg.py         | 44 +++++++++++++++++++--
 config.py            | 12 ++++--
 dep/fmt/build.py     | 17 +++++++++
 dep/libusbp/build.py |  2 +-
 scripts/build.py     |  2 +-
 src/gui/windres.rc   |  2 +-
 12 files changed, 232 insertions(+), 72 deletions(-)
 create mode 100644 dep/fmt/build.py

diff --git a/Makefile b/Makefile
index a83a38d6..f974edcc 100644
--- a/Makefile
+++ b/Makefile
@@ -1,9 +1,37 @@
-CC = gcc
-CXX = g++ -std=c++17
-CFLAGS = -g -O3
-LDFLAGS =
+export BUILDTYPE
+BUILDTYPE ?= host
 
-OBJ = .obj
+ifeq ($(BUILDTYPE),windows)
+	MINGW = i686-w64-mingw32-
+	CC = $(MINGW)gcc
+	CXX = $(MINGW)g++ -std=c++17
+	CFLAGS += -g -O3
+	CXXFLAGS += \
+		-fext-numeric-literals \
+		-Wno-deprecated-enum-float-conversion \
+		-Wno-deprecated-enum-enum-conversion
+	LDFLAGS += -static
+	AR = $(MINGW)ar
+	PKG_CONFIG = $(MINGW)pkg-config -static
+	WINDRES = $(MINGW)windres
+	WX_CONFIG = /usr/i686-w64-mingw32/sys-root/mingw/bin/wx-config-3.0 --static=yes
+	EXT = .exe
+else
+	CC = gcc
+	CXX = g++ -std=c++17
+	CFLAGS = -g -O3
+	LDFLAGS =
+	AR = ar
+	PKG_CONFIG = pkg-config
+endif
+
+HOSTCC = gcc
+HOSTCXX = g++ -std=c++17
+HOSTCFLAGS = -g -O3
+HOSTLDFLAGS =
+
+REALOBJ = .obj
+OBJ = $(REALOBJ)/$(BUILDTYPE)
 DESTDIR ?=
 PREFIX ?= /usr/local
 BINDIR ?= $(PREFIX)/bin
@@ -54,6 +82,9 @@ README.md: $(OBJ)/scripts/+mkdocindex/+mkdocindex$(EXT)
 .PHONY: install install-bin
 install:: all install-bin
 
+clean::
+	$(hide) rm -rf $(REALOBJ)
+
 install-bin:
 	@echo "INSTALL"
 	$(hide) install -D -v "$(OBJ)/src+fluxengine/src+fluxengine" "$(DESTDIR)$(BINDIR)/fluxengine"
diff --git a/README.md b/README.md
index dd2cc632..8f2f3320 100644
--- a/README.md
+++ b/README.md
@@ -103,45 +103,45 @@ particular filesystem and can read (and sometimes write, support varies) files
 directly from disks, flux files or disk images. Some formats have multiple
 choices because they can store multiple types of file system.
 
-<!-- FORMATSSTART -->
-<!-- This section is automatically generated. Do not edit. -->
-
-| Profile | Format | Read? | Write? | Filesystem? |
-|:--------|:-------|:-----:|:------:|:------------|
-| [`acornadfs`](doc/disk-acornadfs.md) | Acorn ADFS: BBC Micro, Archimedes | 🦖 |  |  |
-| [`acorndfs`](doc/disk-acorndfs.md) | Acorn DFS: Acorn Atom, BBC Micro series | 🦄 |  | ACORNDFS  |
-| [`aeslanier`](doc/disk-aeslanier.md) | AES Lanier "No Problem": 616kB 5.25" 77-track SSDD hard sectored | 🦖 |  |  |
-| [`agat`](doc/disk-agat.md) | Agat: 840kB 5.25" 80-track DS | 🦖 | 🦖 |  |
-| [`amiga`](doc/disk-amiga.md) | Amiga: 880kB 3.5" DSDD | 🦄 | 🦄 | AMIGAFFS  |
-| [`ampro`](doc/disk-ampro.md) | Ampro Little Board: CP/M | 🦖 |  | CPMFS  |
-| [`apple2`](doc/disk-apple2.md) | Apple II: Prodos, Appledos, and CP/M | 🦄 | 🦄 | APPLEDOS CPMFS PRODOS  |
-| [`atarist`](doc/disk-atarist.md) | Atari ST: Almost PC compatible | 🦄 | 🦄 |  |
-| [`bk`](doc/disk-bk.md) | BK: 800kB 5.25"/3.5" 80-track 10-sector DSDD | 🦖 | 🦖 |  |
-| [`brother`](doc/disk-brother.md) | Brother word processors: GCR family | 🦄 | 🦄 | BROTHER120 FATFS  |
-| [`commodore`](doc/disk-commodore.md) | Commodore: 1541, 1581, 8050 and variations | 🦄 | 🦄 | CBMFS  |
-| [`eco1`](doc/disk-eco1.md) | VDS Eco1: CP/M; 1210kB 77-track mixed format DSHD | 🦖 |  | CPMFS  |
-| [`epsonpf10`](doc/disk-epsonpf10.md) | Epson PF-10: CP/M; 3.5" 40-track DSDD | 🦖 |  | CPMFS  |
-| [`f85`](doc/disk-f85.md) | Durango F85: 461kB 5.25" 77-track SS | 🦖 |  |  |
-| [`fb100`](doc/disk-fb100.md) | Brother FB-100: 100kB 3.5" 40-track SSSD | 🦖 |  |  |
-| [`hplif`](doc/disk-hplif.md) | Hewlett-Packard LIF: a variety of disk formats used by HP | 🦄 | 🦄 | LIF  |
-| [`ibm`](doc/disk-ibm.md) | IBM PC: Generic PC 3.5"/5.25" disks | 🦄 | 🦄 | FATFS  |
-| [`icl30`](doc/disk-icl30.md) | ICL Model 30: CP/M; 263kB 35-track DSSD | 🦖 |  | CPMFS  |
-| [`mac`](doc/disk-mac.md) | Macintosh: 400kB/800kB 3.5" GCR | 🦄 | 🦄 | MACHFS  |
-| [`micropolis`](doc/disk-micropolis.md) | Micropolis: 100tpi MetaFloppy disks | 🦄 | 🦄 |  |
-| [`ms2000`](doc/disk-ms2000.md) | : MS2000 Microdisk Development System |  |  | MICRODOS  |
-| [`mx`](doc/disk-mx.md) | DVK MX: Soviet-era PDP-11 clone | 🦖 |  |  |
-| [`n88basic`](doc/disk-n88basic.md) | N88-BASIC: PC8800/PC98 5.25" 77-track 26-sector DSHD | 🦄 | 🦄 |  |
-| [`northstar`](doc/disk-northstar.md) | Northstar: 5.25" hard sectored | 🦄 | 🦄 |  |
-| [`psos`](doc/disk-psos.md) | pSOS: 800kB DSDD with PHILE | 🦄 | 🦄 | PHILE  |
-| [`rolandd20`](doc/disk-rolandd20.md) | Roland D20: 3.5" electronic synthesiser disks | 🦄 | 🦖 | ROLAND  |
-| [`rx50`](doc/disk-rx50.md) | Digital RX50: 400kB 5.25" 80-track 10-sector SSDD | 🦖 | 🦖 |  |
-| [`smaky6`](doc/disk-smaky6.md) | Smaky 6: 308kB 5.25" 77-track 16-sector SSDD, hard sectored | 🦖 |  | SMAKY6  |
-| [`tids990`](doc/disk-tids990.md) | Texas Instruments DS990: 1126kB 8" DSSD | 🦖 | 🦖 |  |
-| [`tiki`](doc/disk-tiki.md) | Tiki 100: CP/M |  |  | CPMFS  |
-| [`victor9k`](doc/disk-victor9k.md) | Victor 9000 / Sirius One: 1224kB 5.25" DSDD GCR | 🦖 | 🦖 |  |
-| [`zilogmcz`](doc/disk-zilogmcz.md) | Zilog MCZ: 320kB 8" 77-track SSSD hard-sectored | 🦖 |  | ZDOS  |
-{: .datatable }
-
+<!-- FORMATSSTART -->
+<!-- This section is automatically generated. Do not edit. -->
+
+| Profile | Format | Read? | Write? | Filesystem? |
+|:--------|:-------|:-----:|:------:|:------------|
+| [`acornadfs`](doc/disk-acornadfs.md) | Acorn ADFS: BBC Micro, Archimedes | 🦖 |  |  |
+| [`acorndfs`](doc/disk-acorndfs.md) | Acorn DFS: Acorn Atom, BBC Micro series | 🦄 |  | ACORNDFS  |
+| [`aeslanier`](doc/disk-aeslanier.md) | AES Lanier "No Problem": 616kB 5.25" 77-track SSDD hard sectored | 🦖 |  |  |
+| [`agat`](doc/disk-agat.md) | Agat: 840kB 5.25" 80-track DS | 🦖 | 🦖 |  |
+| [`amiga`](doc/disk-amiga.md) | Amiga: 880kB 3.5" DSDD | 🦄 | 🦄 | AMIGAFFS  |
+| [`ampro`](doc/disk-ampro.md) | Ampro Little Board: CP/M | 🦖 |  | CPMFS  |
+| [`apple2`](doc/disk-apple2.md) | Apple II: Prodos, Appledos, and CP/M | 🦄 | 🦄 | APPLEDOS CPMFS PRODOS  |
+| [`atarist`](doc/disk-atarist.md) | Atari ST: Almost PC compatible | 🦄 | 🦄 |  |
+| [`bk`](doc/disk-bk.md) | BK: 800kB 5.25"/3.5" 80-track 10-sector DSDD | 🦖 | 🦖 |  |
+| [`brother`](doc/disk-brother.md) | Brother word processors: GCR family | 🦄 | 🦄 | BROTHER120 FATFS  |
+| [`commodore`](doc/disk-commodore.md) | Commodore: 1541, 1581, 8050 and variations | 🦄 | 🦄 | CBMFS  |
+| [`eco1`](doc/disk-eco1.md) | VDS Eco1: CP/M; 1210kB 77-track mixed format DSHD | 🦖 |  | CPMFS  |
+| [`epsonpf10`](doc/disk-epsonpf10.md) | Epson PF-10: CP/M; 3.5" 40-track DSDD | 🦖 |  | CPMFS  |
+| [`f85`](doc/disk-f85.md) | Durango F85: 461kB 5.25" 77-track SS | 🦖 |  |  |
+| [`fb100`](doc/disk-fb100.md) | Brother FB-100: 100kB 3.5" 40-track SSSD | 🦖 |  |  |
+| [`hplif`](doc/disk-hplif.md) | Hewlett-Packard LIF: a variety of disk formats used by HP | 🦄 | 🦄 | LIF  |
+| [`ibm`](doc/disk-ibm.md) | IBM PC: Generic PC 3.5"/5.25" disks | 🦄 | 🦄 | FATFS  |
+| [`icl30`](doc/disk-icl30.md) | ICL Model 30: CP/M; 263kB 35-track DSSD | 🦖 |  | CPMFS  |
+| [`mac`](doc/disk-mac.md) | Macintosh: 400kB/800kB 3.5" GCR | 🦄 | 🦄 | MACHFS  |
+| [`micropolis`](doc/disk-micropolis.md) | Micropolis: 100tpi MetaFloppy disks | 🦄 | 🦄 |  |
+| [`ms2000`](doc/disk-ms2000.md) | : MS2000 Microdisk Development System |  |  | MICRODOS  |
+| [`mx`](doc/disk-mx.md) | DVK MX: Soviet-era PDP-11 clone | 🦖 |  |  |
+| [`n88basic`](doc/disk-n88basic.md) | N88-BASIC: PC8800/PC98 5.25" 77-track 26-sector DSHD | 🦄 | 🦄 |  |
+| [`northstar`](doc/disk-northstar.md) | Northstar: 5.25" hard sectored | 🦄 | 🦄 |  |
+| [`psos`](doc/disk-psos.md) | pSOS: 800kB DSDD with PHILE | 🦄 | 🦄 | PHILE  |
+| [`rolandd20`](doc/disk-rolandd20.md) | Roland D20: 3.5" electronic synthesiser disks | 🦄 | 🦖 | ROLAND  |
+| [`rx50`](doc/disk-rx50.md) | Digital RX50: 400kB 5.25" 80-track 10-sector SSDD | 🦖 | 🦖 |  |
+| [`smaky6`](doc/disk-smaky6.md) | Smaky 6: 308kB 5.25" 77-track 16-sector SSDD, hard sectored | 🦖 |  | SMAKY6  |
+| [`tids990`](doc/disk-tids990.md) | Texas Instruments DS990: 1126kB 8" DSSD | 🦖 | 🦖 |  |
+| [`tiki`](doc/disk-tiki.md) | Tiki 100: CP/M |  |  | CPMFS  |
+| [`victor9k`](doc/disk-victor9k.md) | Victor 9000 / Sirius One: 1224kB 5.25" DSDD GCR | 🦖 | 🦖 |  |
+| [`zilogmcz`](doc/disk-zilogmcz.md) | Zilog MCZ: 320kB 8" 77-track SSSD hard-sectored | 🦖 |  | ZDOS  |
+{: .datatable }
+
 <!-- FORMATSEND -->
 
 ### Notes
diff --git a/build.py b/build.py
index d92c3845..6b19314c 100644
--- a/build.py
+++ b/build.py
@@ -1,7 +1,7 @@
 from build.ab import export
 from build.c import clibrary, cxxlibrary
 from build.protobuf import proto, protocc
-from build.pkg import package
+from build.pkg import package, hostpackage
 from build.utils import test
 from glob import glob
 import config
@@ -9,9 +9,14 @@
 
 package(name="protobuf_lib", package="protobuf")
 package(name="z_lib", package="zlib")
-package(name="fmt_lib", package="fmt")
+package(name="fmt_lib", package="fmt", fallback="dep/fmt")
 package(name="sqlite3_lib", package="sqlite3")
 
+hostpackage(name="protobuf_host_lib", package="protobuf")
+hostpackage(name="z_host_lib", package="zlib")
+hostpackage(name="fmt_host_lib", package="fmt")
+hostpackage(name="sqlite3_host_lib", package="sqlite3")
+
 clibrary(name="protocol", hdrs={"protocol.h": "./protocol.h"})
 
 proto(name="fl2_proto", srcs=["lib/fl2.proto"])
@@ -211,14 +216,15 @@
     },
     deps=[
         "+fl2_proto_lib",
+        "+fmt_lib",
         "+protocol",
-        "lib+config_proto_lib",
         "dep/adflib",
         "dep/agg",
         "dep/fatfs",
         "dep/hfsutils",
         "dep/libusbp",
         "dep/stb",
+        "lib+config_proto_lib",
     ],
 )
 
diff --git a/build/ab.mk b/build/ab.mk
index 61da4fb0..5674b1a9 100644
--- a/build/ab.mk
+++ b/build/ab.mk
@@ -43,7 +43,7 @@ update-ab:
 .PHONY: clean
 clean::
 	@echo CLEAN
-	$(hide) rm -rf $(OBJ) bin
+	$(hide) rm -rf $(OBJ)
 
 export PYTHONHASHSEED = 1
 build-files = $(shell find . -name 'build.py') $(wildcard build/*.py) $(wildcard config.py)
diff --git a/build/ab.py b/build/ab.py
index 4b485745..025b33a0 100644
--- a/build/ab.py
+++ b/build/ab.py
@@ -401,6 +401,7 @@ def emitter_endrule(rule, outs):
     emit("\t$(hide) touch", rule.sentinel)
 
     for f in filenamesof(outs):
+        emit(".SECONDARY:", f)
         emit(f, ":", rule.sentinel, ";")
 
 
diff --git a/build/c.py b/build/c.py
index 3cb32be0..e55bae5d 100644
--- a/build/c.py
+++ b/build/c.py
@@ -18,6 +18,26 @@
 from types import SimpleNamespace
 
 
+class Toolchain:
+    label = ""
+    cfile = ["$(CC) -c -o {outs[0]} {ins[0]} $(CFLAGS) {cflags}"]
+    cxxfile = ["$(CXX) -c -o {outs[0]} {ins[0]} $(CFLAGS) {cflags}"]
+    clibrary = ["$(AR) cqs {outs[0]} {ins}"]
+    cxxlibrary = ["$(AR) cqs {outs[0]} {ins}"]
+    cprogram = ["$(CC) -o {outs[0]} {ins} {ldflags} $(LDFLAGS)"]
+    cxxprogram = ["$(CXX) -o {outs[0]} {ins} {ldflags} $(LDFLAGS)"]
+
+
+class HostToolchain:
+    label = "HOST "
+    cfile = ["$(HOSTCC) -c -o {outs[0]} {ins[0]} $(HOSTCFLAGS) {cflags}"]
+    cxxfile = ["$(HOSTCXX) -c -o {outs[0]} {ins[0]} $(HOSTCFLAGS) {cflags}"]
+    clibrary = ["$(HOSTAR) cqs {outs[0]} {ins}"]
+    cxxlibrary = ["$(HOSTAR) cqs {outs[0]} {ins}"]
+    cprogram = ["$(HOSTCC) -o {outs[0]} {ins} {ldflags} $(HOSTLDFLAGS)"]
+    cxxprogram = ["$(HOSTCXX) -o {outs[0]} {ins} {ldflags} $(HOSTLDFLAGS)"]
+
+
 def cfileimpl(self, name, srcs, deps, suffix, commands, label, kind, cflags):
     outleaf = stripext(basename(filenameof(srcs[0]))) + suffix
 
@@ -40,9 +60,14 @@ def cfile(
     deps: Targets = None,
     cflags: List = [],
     suffix=".o",
-    commands=["$(CC) -c -o {outs[0]} {ins[0]} $(CFLAGS) {cflags}"],
-    label="CC",
+    toolchain=Toolchain,
+    commands=None,
+    label=None,
 ):
+    if not label:
+        label = toolchain.label + "CC"
+    if not commands:
+        commands = toolchain.cfile
     cfileimpl(self, name, srcs, deps, suffix, commands, label, "cfile", cflags)
 
 
@@ -54,15 +79,20 @@ def cxxfile(
     deps: Targets = None,
     cflags: List = [],
     suffix=".o",
-    commands=["$(CXX) -c -o {outs[0]} {ins[0]} $(CFLAGS) {cflags}"],
-    label="CXX",
+    toolchain=Toolchain,
+    commands=None,
+    label=None,
 ):
+    if not label:
+        label = toolchain.label + "CC"
+    if not commands:
+        commands = toolchain.cfile
     cfileimpl(
         self, name, srcs, deps, suffix, commands, label, "cxxfile", cflags
     )
 
 
-def findsources(name, srcs, deps, cflags, filerule):
+def findsources(name, srcs, deps, cflags, toolchain, filerule):
     objs = []
     for s in flatten(srcs):
         objs += [
@@ -71,6 +101,7 @@ def findsources(name, srcs, deps, cflags, filerule):
                 srcs=[f],
                 deps=deps,
                 cflags=cflags,
+                toolchain=toolchain,
             )
             for f in filenamesof(s)
             if f.endswith(".c")
@@ -131,6 +162,7 @@ def libraryimpl(
     caller_ldflags,
     cflags,
     ldflags,
+    toolchain,
     commands,
     label,
     kind,
@@ -159,6 +191,7 @@ def libraryimpl(
         srcs,
         targetswithtraitsof(deps, "cheaders"),
         cflags + bubbledattrsof(deps, "caller_cflags"),
+        toolchain,
         kind,
     )
 
@@ -188,10 +221,15 @@ def clibrary(
     caller_ldflags: List = [],
     cflags: List = [],
     ldflags: List = [],
-    commands=["$(AR) cqs {outs[0]} {ins}"],
-    label="LIB",
+    toolchain=Toolchain,
+    commands=None,
+    label=None,
     cfilerule=cfile,
 ):
+    if not label:
+        label = toolchain.label + "LIB"
+    if not commands:
+        commands = toolchain.clibrary
     libraryimpl(
         self,
         name,
@@ -202,6 +240,7 @@ def clibrary(
         caller_ldflags,
         cflags,
         ldflags,
+        toolchain,
         commands,
         label,
         cfilerule,
@@ -219,9 +258,14 @@ def cxxlibrary(
     caller_ldflags: List = [],
     cflags: List = [],
     ldflags: List = [],
-    commands=["$(AR) cqs {outs[0]} {ins}"],
-    label="LIB",
+    toolchain=Toolchain,
+    commands=None,
+    label=None,
 ):
+    if not label:
+        label = toolchain.label + "LIB"
+    if not commands:
+        commands = toolchain.clibrary
     libraryimpl(
         self,
         name,
@@ -232,6 +276,7 @@ def cxxlibrary(
         caller_ldflags,
         cflags,
         ldflags,
+        toolchain,
         commands,
         label,
         cxxfile,
@@ -239,19 +284,29 @@ def cxxlibrary(
 
 
 def programimpl(
-    self, name, srcs, deps, cflags, ldflags, commands, label, filerule, kind
+    self,
+    name,
+    srcs,
+    deps,
+    cflags,
+    ldflags,
+    toolchain,
+    commands,
+    label,
+    filerule,
+    kind,
 ):
     ars = filenamesmatchingof(deps, "*.a")
     deps = deps + filenamesmatchingof(srcs, "*.h")
     ldflags = ldflags + bubbledattrsof(deps, "caller_ldflags")
 
-    cfiles = findsources(name, srcs, deps, cflags, filerule)
+    cfiles = findsources(name, srcs, deps, cflags, toolchain, filerule)
     normalrule(
         replaces=self,
         ins=cfiles + ars + ars,
         outs=[basename(name) + "$(EXT)"],
         deps=deps,
-        label=label,
+        label=toolchain.label + label,
         commands=commands,
         ldflags=ldflags,
     )
@@ -265,11 +320,14 @@ def cprogram(
     deps: Targets = None,
     cflags: List = [],
     ldflags: List = [],
-    commands=["$(CC) -o {outs[0]} {ins} {ldflags} $(LDFLAGS)"],
+    toolchain=Toolchain,
+    commands=None,
     label="CLINK",
     cfilerule=cfile,
     cfilekind="cprogram",
 ):
+    if not commands:
+        commands = toolchain.cprogram
     programimpl(
         self,
         name,
@@ -277,6 +335,7 @@ def cprogram(
         deps,
         cflags,
         ldflags,
+        toolchain,
         commands,
         label,
         cfilerule,
@@ -292,9 +351,12 @@ def cxxprogram(
     deps: Targets = None,
     cflags: List = [],
     ldflags: List = [],
-    commands=["$(CXX) -o {outs[0]} {ins} {ldflags} $(LDFLAGS)"],
+    toolchain=Toolchain,
+    commands=None,
     label="CXXLINK",
 ):
+    if not commands:
+        commands = toolchain.cxxprogram
     programimpl(
         self,
         name,
@@ -302,6 +364,7 @@ def cxxprogram(
         deps,
         cflags,
         ldflags,
+        toolchain,
         commands,
         label,
         cxxfile,
diff --git a/build/pkg.py b/build/pkg.py
index a4200ad3..f7677bca 100644
--- a/build/pkg.py
+++ b/build/pkg.py
@@ -1,4 +1,4 @@
-from build.ab import Rule, emit, Target, bubbledattrsof
+from build.ab import Rule, emit, Target, bubbledattrsof, filenamesof
 from types import SimpleNamespace
 import os
 import subprocess
@@ -7,12 +7,18 @@
     """
 PKG_CONFIG ?= pkg-config
 PACKAGES := $(shell $(PKG_CONFIG) --list-all | cut -d' ' -f1 | sort)
+
+HOST_PKG_CONFIG ?= pkg-config
+HOST_PACKAGES := $(shell $(HOST_PKG_CONFIG) --list-all | cut -d' ' -f1 | sort)
 """
 )
 
 
 @Rule
 def package(self, name, package=None, fallback: Target = None):
+    self.ins = []
+    self.outs = []
+
     emit("ifeq ($(filter %s, $(PACKAGES)),)" % package)
     if fallback:
         emit(
@@ -23,7 +29,7 @@ def package(self, name, package=None, fallback: Target = None):
             f"PACKAGE_LDFLAGS_{package} := ",
             bubbledattrsof(fallback, "caller_ldflags"),
         )
-        emit(f"PACKAGE_DEP_{package} := ", fallback.name)
+        self.outs = filenamesof(fallback)
     else:
         emit(f"$(error Required package '{package}' not installed.)")
     emit("else")
@@ -33,11 +39,41 @@ def package(self, name, package=None, fallback: Target = None):
     emit(
         f"PACKAGE_LDFLAGS_{package} := $(shell $(PKG_CONFIG) --libs {package})"
     )
-    emit(f"PACKAGE_DEP_{package} := ")
     emit("endif")
 
     self.attr.caller_cflags = [f"$(PACKAGE_CFLAGS_{package})"]
     self.attr.caller_ldflags = [f"$(PACKAGE_LDFLAGS_{package})"]
+    self.traits.add("clibrary")
+    self.traits.add("cheaders")
+
+
+@Rule
+def hostpackage(self, name, package=None, fallback: Target = None):
+    emit("ifeq ($(filter %s, $(HOST_PACKAGES)),)" % package)
+    if fallback:
+        emit(
+            f"HOST_PACKAGE_CFLAGS_{package} :=",
+            bubbledattrsof(fallback, "caller_cflags"),
+        )
+        emit(
+            f"HOST_PACKAGE_LDFLAGS_{package} := ",
+            bubbledattrsof(fallback, "caller_ldflags"),
+        )
+        emit(f"HOST_PACKAGE_DEP_{package} := ", fallback.name)
+    else:
+        emit(f"$(error Required host package '{package}' not installed.)")
+    emit("else")
+    emit(
+        f"HOST_PACKAGE_CFLAGS_{package} := $(shell $(HOST_PKG_CONFIG) --cflags {package})"
+    )
+    emit(
+        f"HOST_PACKAGE_LDFLAGS_{package} := $(shell $(HOST_PKG_CONFIG) --libs {package})"
+    )
+    emit(f"HOST_PACKAGE_DEP_{package} := ")
+    emit("endif")
+
+    self.attr.caller_cflags = [f"$(HOST_PACKAGE_CFLAGS_{package})"]
+    self.attr.caller_ldflags = [f"$(HOST_PACKAGE_LDFLAGS_{package})"]
 
     self.ins = []
-    self.outs = [f"$(PACKAGE_DEP_{package})"]
+    self.outs = [f"$(HOST_PACKAGE_DEP_{package})"]
diff --git a/config.py b/config.py
index 08073692..062d0f28 100644
--- a/config.py
+++ b/config.py
@@ -1,5 +1,11 @@
 import platform
+import os
 
-windows = platform.system() == "Windows"
-osx = platform.system() == "Darwin"
-unix = not windows
+if os.getenv("BUILDTYPE") == "windows":
+    windows = True
+    osx = False
+    unix = False
+else:
+    windows = False
+    osx = platform.system() == "Darwin"
+    unix = True
diff --git a/dep/fmt/build.py b/dep/fmt/build.py
new file mode 100644
index 00000000..9bd272f9
--- /dev/null
+++ b/dep/fmt/build.py
@@ -0,0 +1,17 @@
+from build.c import cxxlibrary, HostToolchain
+
+cxxlibrary(
+    name="fmt",
+    srcs=[
+        "./src/format.cc",
+        "./src/os.cc",
+    ],
+    cflags=["-Idep/fmt/include"],
+    hdrs={
+        "fmt/args.h": "./include/fmt/args.h",
+        "fmt/chrono.h": "./include/fmt/chrono.h",
+        "fmt/core.h": "./include/fmt/core.h",
+        "fmt/format.h": "./include/fmt/format.h",
+        "fmt/ostream.h": "./include/fmt/ostream.h",
+        "fmt/ranges.h": "./include/fmt/ranges.h",
+        })
\ No newline at end of file
diff --git a/dep/libusbp/build.py b/dep/libusbp/build.py
index 2fe67486..491f7b0e 100644
--- a/dep/libusbp/build.py
+++ b/dep/libusbp/build.py
@@ -61,7 +61,7 @@
     name="libusbp",
     srcs=srcs,
     cflags=["-Idep/libusbp/include", "-Idep/libusbp/src"],
-    ldflags=ldflags,
+    caller_ldflags=ldflags,
     deps=deps,
     hdrs={
         "libusbp_internal.h": "./src/libusbp_internal.h",
diff --git a/scripts/build.py b/scripts/build.py
index 021b829e..7d53348a 100644
--- a/scripts/build.py
+++ b/scripts/build.py
@@ -1,5 +1,5 @@
 from build.ab import Rule, normalrule, Targets
-from build.c import cxxprogram
+from build.c import cxxprogram, HostToolchain
 
 encoders = {}
 
diff --git a/src/gui/windres.rc b/src/gui/windres.rc
index 9d2a5059..410df629 100644
--- a/src/gui/windres.rc
+++ b/src/gui/windres.rc
@@ -6,4 +6,4 @@
 #include "winuser.h"
 
 CREATEPROCESS_MANIFEST_RESOURCE_ID RT_MANIFEST "src/gui/manifest.xml"
-101 ICON DISCARDABLE ".obj/extras+fluxengine_ico/fluxengine.ico"
+101 ICON DISCARDABLE ".obj/windows/extras/+fluxengine_ico/fluxengine.ico"

From f9c1816e6fb9b21f770436b99127aeefb2870a9f Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 19:29:48 +0100
Subject: [PATCH 03/47] Format.

---
 dep/fmt/build.py |  3 ++-
 extras/build.py  |  4 +---
 lib/build.py     |  5 ++++-
 scripts/build.py | 16 ++++++++++++++--
 4 files changed, 21 insertions(+), 7 deletions(-)

diff --git a/dep/fmt/build.py b/dep/fmt/build.py
index 9bd272f9..c968265f 100644
--- a/dep/fmt/build.py
+++ b/dep/fmt/build.py
@@ -14,4 +14,5 @@
         "fmt/format.h": "./include/fmt/format.h",
         "fmt/ostream.h": "./include/fmt/ostream.h",
         "fmt/ranges.h": "./include/fmt/ranges.h",
-        })
\ No newline at end of file
+    },
+)
diff --git a/extras/build.py b/extras/build.py
index 2c37467f..2543208f 100644
--- a/extras/build.py
+++ b/extras/build.py
@@ -37,8 +37,6 @@
     name="fluxengine_ico",
     ins=["./icon.png"],
     outs=["fluxengine.ico"],
-    commands=[
-        "png2ico {outs[0]} {ins[0]}"
-    ],
+    commands=["png2ico {outs[0]} {ins[0]}"],
     label="MAKEICON",
 )
diff --git a/lib/build.py b/lib/build.py
index 1fbed16e..bc82936e 100644
--- a/lib/build.py
+++ b/lib/build.py
@@ -22,4 +22,7 @@
     deps=[".+common_proto", "+fl2_proto"],
 )
 
-protocc(name="config_proto_lib", srcs=[".+common_proto", ".+config_proto", "arch+arch_proto"])
+protocc(
+    name="config_proto_lib",
+    srcs=[".+common_proto", ".+config_proto", "arch+arch_proto"],
+)
diff --git a/scripts/build.py b/scripts/build.py
index 7d53348a..652d52aa 100644
--- a/scripts/build.py
+++ b/scripts/build.py
@@ -37,11 +37,23 @@ def protoencode(self, name, srcs: Targets, proto, symbol):
 cxxprogram(
     name="mkdoc",
     srcs=["./mkdoc.cc"],
-    deps=["src/formats", "lib+config_proto_lib", "+lib", "+fmt_lib", "+protobuf_lib"],
+    deps=[
+        "src/formats",
+        "lib+config_proto_lib",
+        "+lib",
+        "+fmt_lib",
+        "+protobuf_lib",
+    ],
 )
 
 cxxprogram(
     name="mkdocindex",
     srcs=["./mkdocindex.cc"],
-    deps=["src/formats", "lib+config_proto_lib", "+lib", "+fmt_lib", "+protobuf_lib"],
+    deps=[
+        "src/formats",
+        "lib+config_proto_lib",
+        "+lib",
+        "+fmt_lib",
+        "+protobuf_lib",
+    ],
 )

From b15d6cb8e525fa02264e80582f0f505afaffcb44 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:25:42 +0100
Subject: [PATCH 04/47] Try out a WSL github workflow.

---
 .github/workflows/ccpp.yml | 84 +++++++++++++++++++++++---------------
 1 file changed, 50 insertions(+), 34 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 469b2857..95452cee 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -7,44 +7,60 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  build-linux:
-    runs-on: ubuntu-22.04
-    steps:
-    - uses: actions/checkout@v2
-      with:
-        repository: 'davidgiven/fluxengine'
-        path: 'fluxengine'
-    - uses: actions/checkout@v2
-      with:
-        repository: 'davidgiven/fluxengine-testdata'
-        path: 'fluxengine-testdata'
-    - name: apt
-      run: |
-        sudo apt install libudev-dev libsqlite3-dev protobuf-compiler libwxgtk3.0-gtk3-dev libfmt-dev libprotobuf-dev wx-common
-    - name: make
-      run: CXXFLAGS="-Wp,-D_GLIBCXX_ASSERTIONS" make -j`nproc` -C fluxengine
+  #build-linux:
+  #  runs-on: ubuntu-22.04
+  #  steps:
+  #  - uses: actions/checkout@v2
+  #    with:
+  #      repository: 'davidgiven/fluxengine'
+  #      path: 'fluxengine'
+  #  - uses: actions/checkout@v2
+  #    with:
+  #      repository: 'davidgiven/fluxengine-testdata'
+  #      path: 'fluxengine-testdata'
+  #  - name: apt
+  #    run: |
+  #      sudo apt install libudev-dev libsqlite3-dev protobuf-compiler libwxgtk3.0-gtk3-dev libfmt-dev libprotobuf-dev wx-common
+  #  - name: make
+  #    run: CXXFLAGS="-Wp,-D_GLIBCXX_ASSERTIONS" make -j`nproc` -C fluxengine
+
+  #build-macos-current:
+  #  runs-on: macos-latest
+  #  steps:
+  #  - uses: actions/checkout@v2
+  #    with:
+  #      repository: 'davidgiven/fluxengine'
+  #      path: 'fluxengine'
+  #  - uses: actions/checkout@v2
+  #    with:
+  #      repository: 'davidgiven/fluxengine-testdata'
+  #      path: 'fluxengine-testdata'
+  #  - name: brew
+  #    run: brew install sqlite pkg-config libusb protobuf wxwidgets fmt make coreutils dylibbundler libjpeg
+  #  - name: make
+  #    run: gmake -j`nproc` -C fluxengine
+
+  #  - name: Upload build artifacts
+  #    uses: actions/upload-artifact@v2
+  #    with:
+  #      name: ${{ github.event.repository.name }}.${{ github.sha }}
+  #      path: fluxengine/FluxEngine.pkg
+
+  build-windows:
+    runs-on: windows-latest
+    defaults:
+      run:
+        shell: wsl-bash {0}
 
-  build-macos-current:
-    runs-on: macos-latest
     steps:
-    - uses: actions/checkout@v2
-      with:
-        repository: 'davidgiven/fluxengine'
-        path: 'fluxengine'
-    - uses: actions/checkout@v2
+    - uses: Vampire/setup-wsl@v3
       with:
-        repository: 'davidgiven/fluxengine-testdata'
-        path: 'fluxengine-testdata'
-    - name: brew
-      run: brew install sqlite pkg-config libusb protobuf wxwidgets fmt make coreutils dylibbundler libjpeg
-    - name: make
-      run: gmake -j`nproc` -C fluxengine
+        distribution: openSUSE-Leap-15.2
+        use-cache: true
 
-    - name: Upload build artifacts
-      uses: actions/upload-artifact@v2
-      with:
-        name: ${{ github.event.repository.name }}.${{ github.sha }}
-        path: fluxengine/FluxEngine.pkg
+    - name: go
+      run: |
+        ls
 
   #build-windows:
   #  runs-on: windows-latest

From 9d03951da9dafcf9f203851fd469a3b64725456b Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:28:53 +0100
Subject: [PATCH 05/47] Try installing OpenSUSE packages.

---
 .github/workflows/ccpp.yml | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 95452cee..6ee85cf5 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -57,6 +57,9 @@ jobs:
       with:
         distribution: openSUSE-Leap-15.2
         use-cache: true
+        additional-packages:
+          gcc
+          gcc-c++
 
     - name: go
       run: |

From aafad0a628bcadc1d6e27c82a56fd0943830399f Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:35:08 +0100
Subject: [PATCH 06/47] Install the rest of the packages.

---
 .github/workflows/ccpp.yml | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 6ee85cf5..5913b372 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -60,6 +60,32 @@ jobs:
         additional-packages:
           gcc
           gcc-c++
+          protobuf-c-compiler
+          protobuf-devel
+          fmt-devel
+          systemd-devel
+          sqlite-devel
+          wxGTK-devel
+          mingw32-gcc
+          mingw32-gcc-c++
+          mingw32-zlib-static
+          mingw32-protobuf-static
+          mingw32-sqlite-static
+          mingw32-wxWidgets3-static
+          mingw32-libpng-static
+          mingw32-libjpeg-static
+          mingw32-libtiff-static
+          mingw32-nsis
+
+    - uses: actions/checkout@v3
+      with:
+        repository: 'davidgiven/fluxengine'
+        path: 'fluxengine'
+
+    - uses: actions/checkout@v3
+      with:
+        repository: 'davidgiven/fluxengine-testdata'
+        path: 'fluxengine-testdata'
 
     - name: go
       run: |

From 0f584ee343c3c6181163422c15baa50b17cd0d58 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:50:18 +0100
Subject: [PATCH 07/47] Can we install the Fedora WSL system?

---
 .github/workflows/ccpp.yml | 75 +++++++++++++++++++-------------------
 1 file changed, 38 insertions(+), 37 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 5913b372..1ba9ec6d 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -48,48 +48,49 @@ jobs:
 
   build-windows:
     runs-on: windows-latest
-    defaults:
-      run:
-        shell: wsl-bash {0}
 
     steps:
-    - uses: Vampire/setup-wsl@v3
-      with:
-        distribution: openSUSE-Leap-15.2
-        use-cache: true
-        additional-packages:
-          gcc
-          gcc-c++
-          protobuf-c-compiler
-          protobuf-devel
-          fmt-devel
-          systemd-devel
-          sqlite-devel
-          wxGTK-devel
-          mingw32-gcc
-          mingw32-gcc-c++
-          mingw32-zlib-static
-          mingw32-protobuf-static
-          mingw32-sqlite-static
-          mingw32-wxWidgets3-static
-          mingw32-libpng-static
-          mingw32-libjpeg-static
-          mingw32-libtiff-static
-          mingw32-nsis
+    - name: setup WSL
+      run: |
+        wsl --install --distribution whitewaterfoundry.fedora-remix-for-wsl
 
-    - uses: actions/checkout@v3
-      with:
-        repository: 'davidgiven/fluxengine'
-        path: 'fluxengine'
+    #- uses: Vampire/setup-wsl@v3
+    #  with:
+    #    distribution: openSUSE-Leap-15.2
+    #    use-cache: true
+    #    additional-packages:
+    #      gcc
+    #      gcc-c++
+    #      protobuf-c-compiler
+    #      protobuf-devel
+    #      fmt-devel
+    #      systemd-devel
+    #      sqlite-devel
+    #      wxGTK-devel
+    #      mingw32-gcc
+    #      mingw32-gcc-c++
+    #      mingw32-zlib-static
+    #      mingw32-protobuf-static
+    #      mingw32-sqlite-static
+    #      mingw32-wxWidgets3-static
+    #      mingw32-libpng-static
+    #      mingw32-libjpeg-static
+    #      mingw32-libtiff-static
+    #      mingw32-nsis
 
-    - uses: actions/checkout@v3
-      with:
-        repository: 'davidgiven/fluxengine-testdata'
-        path: 'fluxengine-testdata'
+    #- uses: actions/checkout@v3
+    #  with:
+    #    repository: 'davidgiven/fluxengine'
+    #    path: 'fluxengine'
 
-    - name: go
-      run: |
-        ls
+    #- uses: actions/checkout@v3
+    #  with:
+    #    repository: 'davidgiven/fluxengine-testdata'
+    #    path: 'fluxengine-testdata'
+
+    #- name: go
+    #  run: |
+    #    ls
 
   #build-windows:
   #  runs-on: windows-latest

From 8303a1fbcac04f5a6f2e84117d46ec3623b412cf Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:51:12 +0100
Subject: [PATCH 08/47] Again.

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 1ba9ec6d..3bd76edd 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -52,7 +52,7 @@ jobs:
     steps:
     - name: setup WSL
       run: |
-        wsl --install --distribution whitewaterfoundry.fedora-remix-for-wsl
+        winget install whitewaterfoundry.fedora-remix-for-wsl
 
     #- uses: Vampire/setup-wsl@v3
     #  with:

From 9ff3aedf180bed429eacaf884c63c259fe160fc1 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:53:20 +0100
Subject: [PATCH 09/47] Again.

---
 .github/workflows/ccpp.yml | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 3bd76edd..4ad8477a 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -50,6 +50,11 @@ jobs:
     runs-on: windows-latest
 
     steps:
+    - name: Install winget
+      uses: Cyberboss/install-winget@v1
+      with:
+        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
     - name: setup WSL
       run: |
         winget install whitewaterfoundry.fedora-remix-for-wsl

From dbc691726de8753fe9548af83a48678665f366b9 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 20:55:04 +0100
Subject: [PATCH 10/47] Again.

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 4ad8477a..1e6aafb8 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -57,7 +57,7 @@ jobs:
 
     - name: setup WSL
       run: |
-        winget install whitewaterfoundry.fedora-remix-for-wsl
+        winget install  --accept-source-agreements --accept-source-agreements whitewaterfoundry.fedora-remix-for-wsl
 
     #- uses: Vampire/setup-wsl@v3
     #  with:

From 222a88f097684f9e865dce1111fc72e4badcaa24 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 21:01:14 +0100
Subject: [PATCH 11/47] Again.

---
 .github/workflows/ccpp.yml | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 1e6aafb8..7422ad6b 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -59,6 +59,10 @@ jobs:
       run: |
         winget install  --accept-source-agreements --accept-source-agreements whitewaterfoundry.fedora-remix-for-wsl
 
+    - name: run
+      run: |
+        wsl -e sh 'pwd; ls'
+        
     #- uses: Vampire/setup-wsl@v3
     #  with:
     #    distribution: openSUSE-Leap-15.2

From 2636352a49b3e3da95730beba44c080d6bc2e1ab Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 23:03:57 +0100
Subject: [PATCH 12/47] Try doing real work.

---
 .github/workflows/ccpp.yml | 42 ++++++++++++--------------------------
 1 file changed, 13 insertions(+), 29 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 7422ad6b..44791141 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -55,47 +55,31 @@ jobs:
       with:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
+    - uses: actions/checkout@v3
+      with:
+        repository: 'davidgiven/fluxengine'
+        path: 'fluxengine'
+
+    - uses: actions/checkout@v3
+      with:
+        repository: 'davidgiven/fluxengine-testdata'
+        path: 'fluxengine-testdata'
+
     - name: setup WSL
       run: |
         winget install  --accept-source-agreements --accept-source-agreements whitewaterfoundry.fedora-remix-for-wsl
+        wsl --set-default fedoraremix
+        wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 
     - name: run
       run: |
-        wsl -e sh 'pwd; ls'
+        wsl sh -c 'make -C fluxengine BUILDTYPE=windows -j$(nproc)'
         
     #- uses: Vampire/setup-wsl@v3
     #  with:
     #    distribution: openSUSE-Leap-15.2
     #    use-cache: true
     #    additional-packages:
-    #      gcc
-    #      gcc-c++
-    #      protobuf-c-compiler
-    #      protobuf-devel
-    #      fmt-devel
-    #      systemd-devel
-    #      sqlite-devel
-    #      wxGTK-devel
-    #      mingw32-gcc
-    #      mingw32-gcc-c++
-    #      mingw32-zlib-static
-    #      mingw32-protobuf-static
-    #      mingw32-sqlite-static
-    #      mingw32-wxWidgets3-static
-    #      mingw32-libpng-static
-    #      mingw32-libjpeg-static
-    #      mingw32-libtiff-static
-    #      mingw32-nsis
-
-    #- uses: actions/checkout@v3
-    #  with:
-    #    repository: 'davidgiven/fluxengine'
-    #    path: 'fluxengine'
-
-    #- uses: actions/checkout@v3
-    #  with:
-    #    repository: 'davidgiven/fluxengine-testdata'
-    #    path: 'fluxengine-testdata'
 
     #- name: go
     #  run: |

From c2369bcc31938c1e44e215ce259449a94416a95c Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:23:26 +0100
Subject: [PATCH 13/47] Again.

---
 .github/workflows/ccpp.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 44791141..d45d5921 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -68,6 +68,7 @@ jobs:
     - name: setup WSL
       run: |
         winget install  --accept-source-agreements --accept-source-agreements whitewaterfoundry.fedora-remix-for-wsl
+        wsl --install -d fedoraremix
         wsl --set-default fedoraremix
         wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 

From e74b16d28b85a95df325034ecef55e478cb3b888 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:29:37 +0100
Subject: [PATCH 14/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index d45d5921..bb4e7ae3 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -68,7 +68,7 @@ jobs:
     - name: setup WSL
       run: |
         winget install  --accept-source-agreements --accept-source-agreements whitewaterfoundry.fedora-remix-for-wsl
-        wsl --install -d fedoraremix
+        wsl --list --all
         wsl --set-default fedoraremix
         wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 

From 52c0c409e99ca5f69565aae68bfadfad87e9f220 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:35:28 +0100
Subject: [PATCH 15/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 16 +++++-----------
 1 file changed, 5 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index bb4e7ae3..46d0f63e 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -50,10 +50,11 @@ jobs:
     runs-on: windows-latest
 
     steps:
-    - name: Install winget
-      uses: Cyberboss/install-winget@v1
-      with:
-        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+    - name: setup WSL
+      run: |
+        curl https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
+        wsl --set-default fedora
+        wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 
     - uses: actions/checkout@v3
       with:
@@ -65,13 +66,6 @@ jobs:
         repository: 'davidgiven/fluxengine-testdata'
         path: 'fluxengine-testdata'
 
-    - name: setup WSL
-      run: |
-        winget install  --accept-source-agreements --accept-source-agreements whitewaterfoundry.fedora-remix-for-wsl
-        wsl --list --all
-        wsl --set-default fedoraremix
-        wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
-
     - name: run
       run: |
         wsl sh -c 'make -C fluxengine BUILDTYPE=windows -j$(nproc)'

From 1a59e5065d5d9fc81ac091a6511478811ed96921 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:36:39 +0100
Subject: [PATCH 16/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 46d0f63e..5c42c1dc 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -52,7 +52,8 @@ jobs:
     steps:
     - name: setup WSL
       run: |
-        curl https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
+        curl https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz -o fedora.tar.gz
+        wsl --import fedora fedora fedora.tar.gz
         wsl --set-default fedora
         wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 

From 62ddca2bb430e1ff04ac8a65be3a4c7b569e6186 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:44:20 +0100
Subject: [PATCH 17/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 5c42c1dc..42c8e981 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -52,8 +52,7 @@ jobs:
     steps:
     - name: setup WSL
       run: |
-        curl https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz -o fedora.tar.gz
-        wsl --import fedora fedora fedora.tar.gz
+        curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
         wsl --set-default fedora
         wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 

From 7e40dfa5da87819baa5f4a70893c9ad65a9e4806 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:45:15 +0100
Subject: [PATCH 18/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 42c8e981..f3cd2045 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -54,7 +54,7 @@ jobs:
       run: |
         curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
         wsl --set-default fedora
-        wsl sh -c 'dnf -y gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
+        wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 
     - uses: actions/checkout@v3
       with:

From 3adbfd7ef57ab4895d1f98ac99f2a237278b71da Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:52:56 +0100
Subject: [PATCH 19/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index f3cd2045..8fcfd3cb 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -56,6 +56,11 @@ jobs:
         wsl --set-default fedora
         wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
 
+    - name: fix line endings
+      run: |
+        git config --global core.autocrlf false
+        git config --global core.eol lf
+        
     - uses: actions/checkout@v3
       with:
         repository: 'davidgiven/fluxengine'

From 3807d22a78fdb11ee98fa23d31017933b0e30d9b Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 22:57:18 +0100
Subject: [PATCH 20/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 8fcfd3cb..7009fac7 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -54,7 +54,8 @@ jobs:
       run: |
         curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
         wsl --set-default fedora
-        wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis'
+        wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
+        wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
     - name: fix line endings
       run: |

From 6366ac66392f48ba510222856b383bcdb3174609 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 23:08:14 +0100
Subject: [PATCH 21/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 7009fac7..0f96bd6f 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -74,7 +74,7 @@ jobs:
 
     - name: run
       run: |
-        wsl sh -c 'make -C fluxengine BUILDTYPE=windows -j$(nproc)'
+        wsl sh -c 'make -C fluxengine BUILDTYPE=windows V=1'
         
     #- uses: Vampire/setup-wsl@v3
     #  with:

From 08982aae7ab84b532edd509b44da640e1e30884d Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 23:34:38 +0100
Subject: [PATCH 22/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 0f96bd6f..19afbcbb 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -54,6 +54,7 @@ jobs:
       run: |
         curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
         wsl --set-default fedora
+        wsl --set-version fedora 2
         wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
         wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 

From c0ef3215df78b5b8b8e208ecb2e1217d3604999c Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sat, 30 Mar 2024 23:48:40 +0100
Subject: [PATCH 23/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 19afbcbb..f17ec7d5 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -53,8 +53,8 @@ jobs:
     - name: setup WSL
       run: |
         curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
+        wsl --set-version fedora 1
         wsl --set-default fedora
-        wsl --set-version fedora 2
         wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
         wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
@@ -75,7 +75,9 @@ jobs:
 
     - name: run
       run: |
-        wsl sh -c 'make -C fluxengine BUILDTYPE=windows V=1'
+        wsl sh -c 'cp -av fluxengine fluxengine-testdata /tmp/'
+        wsl sh -c 'make -C /tmp/fluxengine BUILDTYPE=windows V=1'
+        wsl sh -c 'cp /tmp/*.exe .'
         
     #- uses: Vampire/setup-wsl@v3
     #  with:

From 4cad34a8a4ecd6ec11e32da20cb040d6bb3dad58 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 00:30:14 +0100
Subject: [PATCH 24/47] Update ccpp.yml

---
 .github/workflows/ccpp.yml | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index f17ec7d5..54f0f608 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -75,9 +75,7 @@ jobs:
 
     - name: run
       run: |
-        wsl sh -c 'cp -av fluxengine fluxengine-testdata /tmp/'
-        wsl sh -c 'make -C /tmp/fluxengine BUILDTYPE=windows V=1'
-        wsl sh -c 'cp /tmp/*.exe .'
+        wsl sh -c 'make -C fluxengine BUILDTYPE=windows V=1 CFLAGS=-v'
         
     #- uses: Vampire/setup-wsl@v3
     #  with:

From 7a94123f0d94eee393fe0e85da5057591bff7c65 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 12:31:58 +0200
Subject: [PATCH 25/47] Again.

---
 .github/workflows/ccpp.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 54f0f608..c086ba75 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -53,6 +53,7 @@ jobs:
     - name: setup WSL
       run: |
         curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
+        wsl --update
         wsl --set-version fedora 1
         wsl --set-default fedora
         wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'

From 586c09f5c35648bbefa1a5dbecf6c012e38e7470 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 12:53:49 +0200
Subject: [PATCH 26/47] Again.

---
 .github/workflows/ccpp.yml | 85 ++++++++------------------------------
 1 file changed, 17 insertions(+), 68 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index c086ba75..b18d83a4 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -47,7 +47,7 @@ jobs:
   #      path: fluxengine/FluxEngine.pkg
 
   build-windows:
-    runs-on: windows-latest
+    runs-on: windows-2019
 
     steps:
     - name: setup WSL
@@ -76,73 +76,22 @@ jobs:
 
     - name: run
       run: |
-        wsl sh -c 'make -C fluxengine BUILDTYPE=windows V=1 CFLAGS=-v'
-        
-    #- uses: Vampire/setup-wsl@v3
-    #  with:
-    #    distribution: openSUSE-Leap-15.2
-    #    use-cache: true
-    #    additional-packages:
-
-    #- name: go
-    #  run: |
-    #    ls
+        wsl sh -c 'make -C fluxengine BUILDTYPE=windows -j$(nproc)'
 
-  #build-windows:
-  #  runs-on: windows-latest
-  #  defaults:
-  #    run:
-  #      shell: msys2 {0}
-  #  steps:
-  #  - uses: msys2/setup-msys2@v2
-  #    with:
-  #      update: true
-  #      msystem: MINGW32
-  #      install: >-
-  #        diffutils
-  #        make
-  #        mingw-w64-i686-binutils
-  #        mingw-w64-i686-fmt
-  #        mingw-w64-i686-gcc
-  #        mingw-w64-i686-libusb
-  #        mingw-w64-i686-nsis
-  #        mingw-w64-i686-pkg-config
-  #        mingw-w64-i686-protobuf
-  #        mingw-w64-i686-python
-  #        mingw-w64-i686-sqlite3
-  #        mingw-w64-i686-wxWidgets
-  #        mingw-w64-i686-zlib
-  #        mingw-w64-i686-png2ico
-  #        vim
-  #        zip
-  #  - name: update-protobuf
-  #    run: |
-  #       pacman -U --noconfirm https://repo.msys2.org/mingw/mingw32/mingw-w64-i686-protobuf-21.9-1-any.pkg.tar.zst
-  #  - uses: actions/checkout@v2
-  #    with:
-  #      repository: 'davidgiven/fluxengine'
-  #      path: 'fluxengine'
-  #  - uses: actions/checkout@v2
-  #    with:
-  #      repository: 'davidgiven/fluxengine-testdata'
-  #      path: 'fluxengine-testdata'
-  #  - name: build
-  #    run: MAGICK_TIME_LIMIT=100 make -j`nproc` -C fluxengine
-
-  #  - name: nsis
-  #    run: |
-  #      cd fluxengine
-  #      strip fluxengine.exe -o fluxengine-stripped.exe
-  #      strip fluxengine-gui.exe -o fluxengine-gui-stripped.exe
-  #      makensis -v2 -nocd -dOUTFILE=fluxengine-installer.exe extras/windows-installer.nsi
+    - name: nsis
+      run: |
+        cd fluxengine
+        strip fluxengine.exe -o fluxengine-stripped.exe
+        strip fluxengine-gui.exe -o fluxengine-gui-stripped.exe
+        makensis -v2 -nocd -dOUTFILE=fluxengine-installer.exe extras/windows-installer.nsi
 
-  #  - name: zip
-  #    run: |
-  #      cd fluxengine
-  #      zip -9 fluxengine-windows.zip fluxengine.exe fluxengine-gui.exe upgrade-flux-file.exe brother120tool.exe brother240tool.exe FluxEngine.cydsn/CortexM3/ARM_GCC_541/Release/FluxEngine.hex fluxengine-installer.exe
+    - name: zip
+      run: |
+        cd fluxengine
+        zip -9 fluxengine-windows.zip fluxengine.exe fluxengine-gui.exe upgrade-flux-file.exe brother120tool.exe brother240tool.exe FluxEngine.cydsn/CortexM3/ARM_GCC_541/Release/FluxEngine.hex fluxengine-installer.exe
 
-  #  - name: Upload build artifacts
-  #    uses: actions/upload-artifact@v2
-  #    with:
-  #      name: ${{ github.event.repository.name }}.${{ github.sha }}
-  #      path: fluxengine/fluxengine-windows.zip
+    - name: Upload build artifacts
+      uses: actions/upload-artifact@v2
+      with:
+        name: ${{ github.event.repository.name }}.${{ github.sha }}
+        path: fluxengine/fluxengine-windows.zip

From 5c2702c6abbbc86d993b1f34f3310e18bc1dedaf Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 13:07:07 +0200
Subject: [PATCH 27/47] Again.

---
 .github/workflows/ccpp.yml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index b18d83a4..af050f47 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -47,7 +47,7 @@ jobs:
   #      path: fluxengine/FluxEngine.pkg
 
   build-windows:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     steps:
     - name: setup WSL
@@ -64,12 +64,12 @@ jobs:
         git config --global core.autocrlf false
         git config --global core.eol lf
         
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         repository: 'davidgiven/fluxengine'
         path: 'fluxengine'
 
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
       with:
         repository: 'davidgiven/fluxengine-testdata'
         path: 'fluxengine-testdata'

From 94372a7f09d2d52d69f6a6d11e9c08a376cd645d Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 13:12:39 +0200
Subject: [PATCH 28/47] Again.

---
 .github/workflows/ccpp.yml | 2 +-
 build/c.py                 | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index af050f47..44e47096 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -57,7 +57,7 @@ jobs:
         wsl --set-version fedora 1
         wsl --set-default fedora
         wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
-        wsl sh -c 'dnf -y install gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
+        wsl sh -c 'dnf -y install --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
     - name: fix line endings
       run: |
diff --git a/build/c.py b/build/c.py
index e55bae5d..42af1501 100644
--- a/build/c.py
+++ b/build/c.py
@@ -84,7 +84,7 @@ def cxxfile(
     label=None,
 ):
     if not label:
-        label = toolchain.label + "CC"
+        label = toolchain.label + "CXX"
     if not commands:
         commands = toolchain.cfile
     cfileimpl(

From 989a11931b62273464d1c7609ff5cdfd6ba4f047 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 13:27:00 +0200
Subject: [PATCH 29/47] Add libssp.

---
 Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Makefile b/Makefile
index f974edcc..c1b2aa4e 100644
--- a/Makefile
+++ b/Makefile
@@ -10,7 +10,7 @@ ifeq ($(BUILDTYPE),windows)
 		-fext-numeric-literals \
 		-Wno-deprecated-enum-float-conversion \
 		-Wno-deprecated-enum-enum-conversion
-	LDFLAGS += -static
+	LDFLAGS += -static -lssp
 	AR = $(MINGW)ar
 	PKG_CONFIG = $(MINGW)pkg-config -static
 	WINDRES = $(MINGW)windres

From f0663030e153210737f931173ad55d83c7888b7d Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 13:39:19 +0200
Subject: [PATCH 30/47] Again.

---
 Makefile | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/Makefile b/Makefile
index c1b2aa4e..04218afd 100644
--- a/Makefile
+++ b/Makefile
@@ -9,8 +9,9 @@ ifeq ($(BUILDTYPE),windows)
 	CXXFLAGS += \
 		-fext-numeric-literals \
 		-Wno-deprecated-enum-float-conversion \
-		-Wno-deprecated-enum-enum-conversion
-	LDFLAGS += -static -lssp
+		-Wno-deprecated-enum-enum-conversion \
+		-fstack-protector-all
+	LDFLAGS += -static -fstack-protector-all
 	AR = $(MINGW)ar
 	PKG_CONFIG = $(MINGW)pkg-config -static
 	WINDRES = $(MINGW)windres

From 61476ea0ccda571d028f6ae0c1453350597352cc Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 13:59:19 +0200
Subject: [PATCH 31/47] Again.

---
 Makefile | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/Makefile b/Makefile
index 04218afd..800f2bad 100644
--- a/Makefile
+++ b/Makefile
@@ -5,13 +5,14 @@ ifeq ($(BUILDTYPE),windows)
 	MINGW = i686-w64-mingw32-
 	CC = $(MINGW)gcc
 	CXX = $(MINGW)g++ -std=c++17
-	CFLAGS += -g -O3
+	CFLAGS += -g -O3 \
+		-fstack-protector
 	CXXFLAGS += \
 		-fext-numeric-literals \
 		-Wno-deprecated-enum-float-conversion \
 		-Wno-deprecated-enum-enum-conversion \
-		-fstack-protector-all
-	LDFLAGS += -static -fstack-protector-all
+		-fstack-protector
+	LDFLAGS += -static -fstack-protector
 	AR = $(MINGW)ar
 	PKG_CONFIG = $(MINGW)pkg-config -static
 	WINDRES = $(MINGW)windres

From bc17c624d3463d20cbe49bc0d0f84edc0f56ad58 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 14:10:53 +0200
Subject: [PATCH 32/47] Again.

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 44e47096..7f03b6ec 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -76,7 +76,7 @@ jobs:
 
     - name: run
       run: |
-        wsl sh -c 'make -C fluxengine BUILDTYPE=windows -j$(nproc)'
+        wsl sh -c 'make -C fluxengine BUILDTYPE=windows V=1'
 
     - name: nsis
       run: |

From 43624796db4facb85430f0eebe26040aaa2daa51 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 14:30:52 +0200
Subject: [PATCH 33/47] Again.

---
 Makefile | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/Makefile b/Makefile
index 800f2bad..a555e766 100644
--- a/Makefile
+++ b/Makefile
@@ -3,16 +3,15 @@ BUILDTYPE ?= host
 
 ifeq ($(BUILDTYPE),windows)
 	MINGW = i686-w64-mingw32-
-	CC = $(MINGW)gcc
-	CXX = $(MINGW)g++ -std=c++17
+	CC = $(MINGW)gcc -fstack-protector
+	CXX = $(MINGW)g++ -std=c++17 -fstack-protector
 	CFLAGS += -g -O3 \
 		-fstack-protector
 	CXXFLAGS += \
 		-fext-numeric-literals \
 		-Wno-deprecated-enum-float-conversion \
-		-Wno-deprecated-enum-enum-conversion \
-		-fstack-protector
-	LDFLAGS += -static -fstack-protector
+		-Wno-deprecated-enum-enum-conversion
+	LDFLAGS += -static
 	AR = $(MINGW)ar
 	PKG_CONFIG = $(MINGW)pkg-config -static
 	WINDRES = $(MINGW)windres

From 7b35232cad8b7c00fa5ab95094daf12f310a1790 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 14:47:44 +0200
Subject: [PATCH 34/47] Again.

---
 Makefile | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Makefile b/Makefile
index a555e766..d0bb0c66 100644
--- a/Makefile
+++ b/Makefile
@@ -3,8 +3,8 @@ BUILDTYPE ?= host
 
 ifeq ($(BUILDTYPE),windows)
 	MINGW = i686-w64-mingw32-
-	CC = $(MINGW)gcc -fstack-protector
-	CXX = $(MINGW)g++ -std=c++17 -fstack-protector
+	CC = $(MINGW)gcc -fstack-protector-strong
+	CXX = $(MINGW)g++ -std=c++17 -fstack-protector-strong
 	CFLAGS += -g -O3 \
 		-fstack-protector
 	CXXFLAGS += \

From 0f8c7d696999b0f24e68e90f2392769492563dcd Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 15:13:43 +0200
Subject: [PATCH 35/47] Again.

---
 .github/workflows/ccpp.yml | 2 +-
 Makefile                   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 7f03b6ec..44e47096 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -76,7 +76,7 @@ jobs:
 
     - name: run
       run: |
-        wsl sh -c 'make -C fluxengine BUILDTYPE=windows V=1'
+        wsl sh -c 'make -C fluxengine BUILDTYPE=windows -j$(nproc)'
 
     - name: nsis
       run: |
diff --git a/Makefile b/Makefile
index d0bb0c66..5fa9f1cc 100644
--- a/Makefile
+++ b/Makefile
@@ -11,7 +11,7 @@ ifeq ($(BUILDTYPE),windows)
 		-fext-numeric-literals \
 		-Wno-deprecated-enum-float-conversion \
 		-Wno-deprecated-enum-enum-conversion
-	LDFLAGS += -static
+	LDFLAGS += -static -lssp
 	AR = $(MINGW)ar
 	PKG_CONFIG = $(MINGW)pkg-config -static
 	WINDRES = $(MINGW)windres

From cacb3971985dbf9bbedb225a5d1666f1919bcf3f Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 15:36:28 +0200
Subject: [PATCH 36/47] Again.

---
 .github/workflows/ccpp.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 44e47096..17a8c1bb 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -57,7 +57,7 @@ jobs:
         wsl --set-version fedora 1
         wsl --set-default fedora
         wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
-        wsl sh -c 'dnf -y install --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
+        wsl sh -c 'dnf -y install --releasever=39 --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
     - name: fix line endings
       run: |

From b30eff294d5da51c7743f844c74da68b73aea7ce Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 15:45:54 +0200
Subject: [PATCH 37/47] Again.

---
 .github/workflows/ccpp.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 17a8c1bb..c565a3c3 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -56,6 +56,7 @@ jobs:
         wsl --update
         wsl --set-version fedora 1
         wsl --set-default fedora
+        wsl sh -c 'dnf -y upgrade'
         wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
         wsl sh -c 'dnf -y install --releasever=39 --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 

From b23b21a7bff59590190eb8f732ee63e607282ca1 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 17:38:36 +0200
Subject: [PATCH 38/47] Again.

---
 .github/workflows/ccpp.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index c565a3c3..cc6811a5 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -56,8 +56,8 @@ jobs:
         wsl --update
         wsl --set-version fedora 1
         wsl --set-default fedora
-        wsl sh -c 'dnf -y upgrade'
-        wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
+        wsl sh -c 'dnf -y install --releasever=39 --setop=install_weak_deps=False dnf'
+        wsl sh -c 'dnf -y install --releasever=39 https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
         wsl sh -c 'dnf -y install --releasever=39 --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
     - name: fix line endings

From 8180c1f79e327969ea731f5a27e7cabb0ab17b83 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 17:47:16 +0200
Subject: [PATCH 39/47] Again.

---
 .github/workflows/ccpp.yml | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index cc6811a5..0b232922 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -56,9 +56,8 @@ jobs:
         wsl --update
         wsl --set-version fedora 1
         wsl --set-default fedora
-        wsl sh -c 'dnf -y install --releasever=39 --setop=install_weak_deps=False dnf'
         wsl sh -c 'dnf -y install --releasever=39 https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
-        wsl sh -c 'dnf -y install --releasever=39 --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
+        wsl sh -c 'dnf -y install --nogpgcheck --releasever=39 --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
     - name: fix line endings
       run: |

From d4cb8f414a2642c877519701a87890f15e81ecaa Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 18:06:19 +0200
Subject: [PATCH 40/47] Again.

---
 .github/workflows/ccpp.yml | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index 0b232922..a782da50 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -52,12 +52,14 @@ jobs:
     steps:
     - name: setup WSL
       run: |
-        curl -L https://github.com/VSWSL/Fedora-WSL/releases/download/v37.0.3.0/rootfs.tar.gz | wsl --import fedora fedora -
+        curl -L https://github.com/WhitewaterFoundry/Fedora-Remix-for-WSL/releases/download/39.0.1/Fedora-Remix-for-WSL-SL_39.0.1.0_x64_arm64.msixbundle -o fedora.msixbundle
+        unzip fedora.msixbundle Fedora-Remix-for-WSL-SL_39.0.1.0_x64.msix
+        unzip Fedora-Remix-for-WSL-SL_39.0.1.0_x64.msix install.tar.gz
         wsl --update
-        wsl --set-version fedora 1
+        wsl --import fedora fedora install.tar.gz
         wsl --set-default fedora
-        wsl sh -c 'dnf -y install --releasever=39 https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
-        wsl sh -c 'dnf -y install --nogpgcheck --releasever=39 --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
+        wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
+        wsl sh -c 'dnf -y install --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
 
     - name: fix line endings
       run: |

From b84ecd289ddaaa3d6705ff5661e3e348455897b2 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 18:25:36 +0200
Subject: [PATCH 41/47] Again.

---
 .github/workflows/ccpp.yml | 10 ++++------
 Makefile                   |  9 ++++-----
 2 files changed, 8 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index a782da50..eca0f09a 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -82,15 +82,13 @@ jobs:
 
     - name: nsis
       run: |
-        cd fluxengine
-        strip fluxengine.exe -o fluxengine-stripped.exe
-        strip fluxengine-gui.exe -o fluxengine-gui-stripped.exe
-        makensis -v2 -nocd -dOUTFILE=fluxengine-installer.exe extras/windows-installer.nsi
+        wsl sh -c 'cd fluxengine && strip fluxengine.exe -o fluxengine-stripped.exe'
+        wsl sh -c 'cd fluxengine && strip fluxengine-gui.exe -o fluxengine-gui-stripped.exe'
+        wsl sh -c 'cd fluxengine && makensis -v2 -nocd -dOUTFILE=fluxengine-installer.exe extras/windows-installer.nsi'
 
     - name: zip
       run: |
-        cd fluxengine
-        zip -9 fluxengine-windows.zip fluxengine.exe fluxengine-gui.exe upgrade-flux-file.exe brother120tool.exe brother240tool.exe FluxEngine.cydsn/CortexM3/ARM_GCC_541/Release/FluxEngine.hex fluxengine-installer.exe
+        wsl sh -c 'cd fluxengine && zip -9 fluxengine-windows.zip fluxengine.exe fluxengine-gui.exe upgrade-flux-file.exe brother120tool.exe brother240tool.exe FluxEngine.cydsn/CortexM3/ARM_GCC_541/Release/FluxEngine.hex fluxengine-installer.exe'
 
     - name: Upload build artifacts
       uses: actions/upload-artifact@v2
diff --git a/Makefile b/Makefile
index 5fa9f1cc..f974edcc 100644
--- a/Makefile
+++ b/Makefile
@@ -3,15 +3,14 @@ BUILDTYPE ?= host
 
 ifeq ($(BUILDTYPE),windows)
 	MINGW = i686-w64-mingw32-
-	CC = $(MINGW)gcc -fstack-protector-strong
-	CXX = $(MINGW)g++ -std=c++17 -fstack-protector-strong
-	CFLAGS += -g -O3 \
-		-fstack-protector
+	CC = $(MINGW)gcc
+	CXX = $(MINGW)g++ -std=c++17
+	CFLAGS += -g -O3
 	CXXFLAGS += \
 		-fext-numeric-literals \
 		-Wno-deprecated-enum-float-conversion \
 		-Wno-deprecated-enum-enum-conversion
-	LDFLAGS += -static -lssp
+	LDFLAGS += -static
 	AR = $(MINGW)ar
 	PKG_CONFIG = $(MINGW)pkg-config -static
 	WINDRES = $(MINGW)windres

From cdf9cc387c38ead3a76dfc7bd1f456176d728232 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 18:42:37 +0200
Subject: [PATCH 42/47] Finally, it works!

---
 .github/workflows/ccpp.yml | 72 +++++++++++++++++++-------------------
 1 file changed, 36 insertions(+), 36 deletions(-)

diff --git a/.github/workflows/ccpp.yml b/.github/workflows/ccpp.yml
index eca0f09a..acb6fbe8 100644
--- a/.github/workflows/ccpp.yml
+++ b/.github/workflows/ccpp.yml
@@ -7,44 +7,44 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  #build-linux:
-  #  runs-on: ubuntu-22.04
-  #  steps:
-  #  - uses: actions/checkout@v2
-  #    with:
-  #      repository: 'davidgiven/fluxengine'
-  #      path: 'fluxengine'
-  #  - uses: actions/checkout@v2
-  #    with:
-  #      repository: 'davidgiven/fluxengine-testdata'
-  #      path: 'fluxengine-testdata'
-  #  - name: apt
-  #    run: |
-  #      sudo apt install libudev-dev libsqlite3-dev protobuf-compiler libwxgtk3.0-gtk3-dev libfmt-dev libprotobuf-dev wx-common
-  #  - name: make
-  #    run: CXXFLAGS="-Wp,-D_GLIBCXX_ASSERTIONS" make -j`nproc` -C fluxengine
+  build-linux:
+    runs-on: ubuntu-22.04
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        repository: 'davidgiven/fluxengine'
+        path: 'fluxengine'
+    - uses: actions/checkout@v2
+      with:
+        repository: 'davidgiven/fluxengine-testdata'
+        path: 'fluxengine-testdata'
+    - name: apt
+      run: |
+        sudo apt install libudev-dev libsqlite3-dev protobuf-compiler libwxgtk3.0-gtk3-dev libfmt-dev libprotobuf-dev wx-common
+    - name: make
+      run: CXXFLAGS="-Wp,-D_GLIBCXX_ASSERTIONS" make -j`nproc` -C fluxengine
 
-  #build-macos-current:
-  #  runs-on: macos-latest
-  #  steps:
-  #  - uses: actions/checkout@v2
-  #    with:
-  #      repository: 'davidgiven/fluxengine'
-  #      path: 'fluxengine'
-  #  - uses: actions/checkout@v2
-  #    with:
-  #      repository: 'davidgiven/fluxengine-testdata'
-  #      path: 'fluxengine-testdata'
-  #  - name: brew
-  #    run: brew install sqlite pkg-config libusb protobuf wxwidgets fmt make coreutils dylibbundler libjpeg
-  #  - name: make
-  #    run: gmake -j`nproc` -C fluxengine
+  build-macos-current:
+    runs-on: macos-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        repository: 'davidgiven/fluxengine'
+        path: 'fluxengine'
+    - uses: actions/checkout@v2
+      with:
+        repository: 'davidgiven/fluxengine-testdata'
+        path: 'fluxengine-testdata'
+    - name: brew
+      run: brew install sqlite pkg-config libusb protobuf wxwidgets fmt make coreutils dylibbundler libjpeg
+    - name: make
+      run: gmake -j`nproc` -C fluxengine
 
-  #  - name: Upload build artifacts
-  #    uses: actions/upload-artifact@v2
-  #    with:
-  #      name: ${{ github.event.repository.name }}.${{ github.sha }}
-  #      path: fluxengine/FluxEngine.pkg
+    - name: Upload build artifacts
+      uses: actions/upload-artifact@v2
+      with:
+        name: ${{ github.event.repository.name }}.${{ github.sha }}
+        path: fluxengine/FluxEngine.pkg
 
   build-windows:
     runs-on: windows-latest

From 621523db6238fff5362aded06abc6c9e78170a22 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 21:30:26 +0200
Subject: [PATCH 43/47] Try not to build fallback packages unless necessary.

---
 build/pkg.py | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/build/pkg.py b/build/pkg.py
index f7677bca..8b41a9f4 100644
--- a/build/pkg.py
+++ b/build/pkg.py
@@ -16,9 +16,6 @@
 
 @Rule
 def package(self, name, package=None, fallback: Target = None):
-    self.ins = []
-    self.outs = []
-
     emit("ifeq ($(filter %s, $(PACKAGES)),)" % package)
     if fallback:
         emit(
@@ -29,7 +26,7 @@ def package(self, name, package=None, fallback: Target = None):
             f"PACKAGE_LDFLAGS_{package} := ",
             bubbledattrsof(fallback, "caller_ldflags"),
         )
-        self.outs = filenamesof(fallback)
+        emit(f"PACKAGE_DEPS_{package} := ", filenamesof(fallback))
     else:
         emit(f"$(error Required package '{package}' not installed.)")
     emit("else")
@@ -39,6 +36,7 @@ def package(self, name, package=None, fallback: Target = None):
     emit(
         f"PACKAGE_LDFLAGS_{package} := $(shell $(PKG_CONFIG) --libs {package})"
     )
+    emit(f"PACKAGE_DEPS_{package} :=")
     emit("endif")
 
     self.attr.caller_cflags = [f"$(PACKAGE_CFLAGS_{package})"]
@@ -46,6 +44,9 @@ def package(self, name, package=None, fallback: Target = None):
     self.traits.add("clibrary")
     self.traits.add("cheaders")
 
+    self.ins = []
+    self.outs = [f"$(PACKAGE_DEPS_{package})"]
+
 
 @Rule
 def hostpackage(self, name, package=None, fallback: Target = None):

From 72ab957134c798af5109287465839617f0015a8f Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 21:44:12 +0200
Subject: [PATCH 44/47] Remember to build C++ files with g++.

---
 build/c.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/build/c.py b/build/c.py
index 42af1501..3a2205f3 100644
--- a/build/c.py
+++ b/build/c.py
@@ -86,7 +86,7 @@ def cxxfile(
     if not label:
         label = toolchain.label + "CXX"
     if not commands:
-        commands = toolchain.cfile
+        commands = toolchain.cxxfile
     cfileimpl(
         self, name, srcs, deps, suffix, commands, label, "cxxfile", cflags
     )

From 1a612c3db515d07522dc812b11fb0b6558ea654d Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 21:53:40 +0200
Subject: [PATCH 45/47] Pull the .a file out of fallback targets at make time,
 not at config time.

---
 build/pkg.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/build/pkg.py b/build/pkg.py
index 8b41a9f4..03ca0073 100644
--- a/build/pkg.py
+++ b/build/pkg.py
@@ -18,6 +18,7 @@
 def package(self, name, package=None, fallback: Target = None):
     emit("ifeq ($(filter %s, $(PACKAGES)),)" % package)
     if fallback:
+        emit(f"PACKAGE_DEPS_{package} := ", filenamesof(fallback))
         emit(
             f"PACKAGE_CFLAGS_{package} :=",
             bubbledattrsof(fallback, "caller_cflags"),
@@ -25,8 +26,8 @@ def package(self, name, package=None, fallback: Target = None):
         emit(
             f"PACKAGE_LDFLAGS_{package} := ",
             bubbledattrsof(fallback, "caller_ldflags"),
+            f"$(filter %.a, $(PACKAGE_DEPS_{package})",
         )
-        emit(f"PACKAGE_DEPS_{package} := ", filenamesof(fallback))
     else:
         emit(f"$(error Required package '{package}' not installed.)")
     emit("else")

From cdb4e9ed21c624a9f3591ad87dec131cfeef9b38 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 22:00:27 +0200
Subject: [PATCH 46/47] Typo fix.

---
 build/pkg.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/build/pkg.py b/build/pkg.py
index 03ca0073..bff53cd7 100644
--- a/build/pkg.py
+++ b/build/pkg.py
@@ -26,7 +26,7 @@ def package(self, name, package=None, fallback: Target = None):
         emit(
             f"PACKAGE_LDFLAGS_{package} := ",
             bubbledattrsof(fallback, "caller_ldflags"),
-            f"$(filter %.a, $(PACKAGE_DEPS_{package})",
+            f"$(filter %.a, $(PACKAGE_DEPS_{package}))",
         )
     else:
         emit(f"$(error Required package '{package}' not installed.)")

From 8f9c12b26cd35b4a650b09feff79bfed592a4d94 Mon Sep 17 00:00:00 2001
From: David Given <dg@cowlark.com>
Date: Sun, 31 Mar 2024 22:18:56 +0200
Subject: [PATCH 47/47] Update the release build script.

---
 .github/workflows/release.yml | 149 ++++++++++++++++------------------
 1 file changed, 70 insertions(+), 79 deletions(-)

diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 500eb90b..f4d8434e 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -10,85 +10,76 @@ on:
       - "master"
 
 jobs:
-  #dev-release:
-  #  runs-on: windows-latest
-  #  defaults:
-  #    run:
-  #      shell: msys2 {0}
-  #  steps:
-  #  - uses: msys2/setup-msys2@v2
-  #    with:
-  #      update: true
-  #      msystem: MINGW32
-  #      install: >-
-  #        diffutils
-  #        make
-  #        mingw-w64-i686-binutils
-  #        mingw-w64-i686-fmt
-  #        mingw-w64-i686-gcc
-  #        mingw-w64-i686-libusb
-  #        mingw-w64-i686-nsis
-  #        mingw-w64-i686-pkg-config
-  #        mingw-w64-i686-protobuf
-  #        mingw-w64-i686-python
-  #        mingw-w64-i686-sqlite3
-  #        mingw-w64-i686-wxWidgets
-  #        mingw-w64-i686-zlib
-  #        mingw-w64-i686-png2ico
-  #        vim
-  #        zip
-  #  - uses: actions/checkout@v3
-
-  #  - name: update-protobuf
-  #    run: |
-  #       pacman -U --noconfirm https://repo.msys2.org/mingw/mingw32/mingw-w64-i686-protobuf-21.9-1-any.pkg.tar.zst
-
-  #  - name: build
-  #    run: |
-  #      MAGICK_TIME_LIMIT=100 make -j`nproc`
-
-  #  - name: nsis
-  #    run: |
-  #      strip fluxengine.exe -o fluxengine-stripped.exe
-  #      strip fluxengine-gui.exe -o fluxengine-gui-stripped.exe
-  #      makensis -v2 -nocd -dOUTFILE=fluxengine-installer.exe extras/windows-installer.nsi
-
-  #  - name: zip
-  #    run: |
-  #      zip -9 fluxengine.zip fluxengine.exe fluxengine-gui.exe upgrade-flux-file.exe brother120tool.exe brother240tool.exe FluxEngine.cydsn/CortexM3/ARM_GCC_541/Release/FluxEngine.hex
-
-  #  - name: date
-  #    run: |
-  #      echo "RELEASE_DATE=$(date --rfc-3339=date)" >> ${GITHUB_ENV}
-
-  #  - name: tag
-  #    uses: EndBug/latest-tag@latest
-  #    with:
-  #      tag-name: dev
-  #      force-branch: false
-  #    env:
-  #      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-
-  #  - name: delete-old-assets
-  #    uses: mknejp/delete-release-assets@v1
-  #    with:
-  #      token: ${{ github.token }}
-  #      tag: dev
-  #      assets: | 
-  #        fluxengine.zip
-  #        fluxengine-installer.exe
-  #      fail-if-no-assets: false
-
-  #  - name: release
-  #    uses: softprops/action-gh-release@v1
-  #    with:
-  #      name: Development build ${{ env.RELEASE_DATE }}
-  #      files: |
-  #        fluxengine.zip
-  #        fluxengine-installer.exe
-  #      tag_name: dev
-  #    env:
-  #      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+  dev-release:
+    runs-on: windows-latest
+
+    steps:
+    - name: setup WSL
+      run: |
+        curl -L https://github.com/WhitewaterFoundry/Fedora-Remix-for-WSL/releases/download/39.0.1/Fedora-Remix-for-WSL-SL_39.0.1.0_x64_arm64.msixbundle -o fedora.msixbundle
+        unzip fedora.msixbundle Fedora-Remix-for-WSL-SL_39.0.1.0_x64.msix
+        unzip Fedora-Remix-for-WSL-SL_39.0.1.0_x64.msix install.tar.gz
+        wsl --update
+        wsl --import fedora fedora install.tar.gz
+        wsl --set-default fedora
+        wsl sh -c 'dnf -y install https://github.com/rpmsphere/noarch/raw/master/r/rpmsphere-release-38-1.noarch.rpm'
+        wsl sh -c 'dnf -y install --setop=install_weak_deps=False gcc gcc-c++ protobuf-c-compiler protobuf-devel fmt-devel systemd-devel sqlite-devel wxGTK-devel mingw32-gcc mingw32-gcc-c++ mingw32-zlib-static mingw32-protobuf-static mingw32-sqlite-static mingw32-wxWidgets3-static mingw32-libpng-static mingw32-libjpeg-static mingw32-libtiff-static mingw32-nsis png2ico'
+
+    - name: fix line endings
+      run: |
+        git config --global core.autocrlf false
+        git config --global core.eol lf
+        
+    - uses: actions/checkout@v4
+      with:
+        repository: 'davidgiven/fluxengine'
+
+    - name: run
+      run: |
+        wsl sh -c 'make BUILDTYPE=windows -j$(nproc)'
+
+    - name: nsis
+      run: |
+        wsl sh -c 'strip fluxengine.exe -o fluxengine-stripped.exe'
+        wsl sh -c 'strip fluxengine-gui.exe -o fluxengine-gui-stripped.exe'
+        wsl sh -c 'makensis -v2 -nocd -dOUTFILE=fluxengine-installer.exe extras/windows-installer.nsi'
+
+    - name: zip
+      run: |
+        wsl sh -c 'zip -9 fluxengine-windows.zip fluxengine.exe fluxengine-gui.exe upgrade-flux-file.exe brother120tool.exe brother240tool.exe FluxEngine.cydsn/CortexM3/ARM_GCC_541/Release/FluxEngine.hex fluxengine-installer.exe'
+
+    - name: date
+      run: |
+        echo "RELEASE_DATE=$(date --rfc-3339=date)" >> ${GITHUB_ENV}
+
+    - name: tag
+      uses: EndBug/latest-tag@latest
+      with:
+        tag-name: dev
+        force-branch: false
+      env:
+        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+    - name: delete-old-assets
+      uses: mknejp/delete-release-assets@v1
+      with:
+        token: ${{ github.token }}
+        tag: dev
+        assets: | 
+          fluxengine.zip
+          fluxengine-installer.exe
+        fail-if-no-assets: false
+
+    - name: release
+      uses: softprops/action-gh-release@v1
+      with:
+        name: Development build ${{ env.RELEASE_DATE }}
+        files: |
+          fluxengine.zip
+          fluxengine-installer.exe
+        tag_name: dev
+      env:
+        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
   build-macos:
     runs-on: macos-latest