diff --git a/src/compiler/crystal/codegen/class_var.cr b/src/compiler/crystal/codegen/class_var.cr
index 07d8ed0f96b1..a31a9b00bac5 100644
--- a/src/compiler/crystal/codegen/class_var.cr
+++ b/src/compiler/crystal/codegen/class_var.cr
@@ -25,8 +25,8 @@ class Crystal::CodeGenVisitor
     initialized_flag_name = class_var_global_initialized_name(class_var)
     initialized_flag = @main_mod.globals[initialized_flag_name]?
     unless initialized_flag
-      initialized_flag = @main_mod.globals.add(@main_llvm_context.int1, initialized_flag_name)
-      initialized_flag.initializer = @main_llvm_context.int1.const_int(0)
+      initialized_flag = @main_mod.globals.add(@main_llvm_context.int8, initialized_flag_name)
+      initialized_flag.initializer = @main_llvm_context.int8.const_int(0)
       initialized_flag.linkage = LLVM::Linkage::Internal if @single_module
       initialized_flag.thread_local = true if class_var.thread_local?
     end
@@ -61,7 +61,7 @@ class Crystal::CodeGenVisitor
       initialized_flag_name = class_var_global_initialized_name(class_var)
       initialized_flag = @llvm_mod.globals[initialized_flag_name]?
       unless initialized_flag
-        initialized_flag = @llvm_mod.globals.add(llvm_context.int1, initialized_flag_name)
+        initialized_flag = @llvm_mod.globals.add(llvm_context.int8, initialized_flag_name)
         initialized_flag.thread_local = true if class_var.thread_local?
       end
     end
diff --git a/src/compiler/crystal/codegen/const.cr b/src/compiler/crystal/codegen/const.cr
index ec306e97296d..decfb3945e20 100644
--- a/src/compiler/crystal/codegen/const.cr
+++ b/src/compiler/crystal/codegen/const.cr
@@ -64,8 +64,8 @@ class Crystal::CodeGenVisitor
     initialized_flag_name = const.initialized_llvm_name
     initialized_flag = @main_mod.globals[initialized_flag_name]?
     unless initialized_flag
-      initialized_flag = @main_mod.globals.add(@main_llvm_context.int1, initialized_flag_name)
-      initialized_flag.initializer = @main_llvm_context.int1.const_int(0)
+      initialized_flag = @main_mod.globals.add(@main_llvm_context.int8, initialized_flag_name)
+      initialized_flag.initializer = @main_llvm_context.int8.const_int(0)
       initialized_flag.linkage = LLVM::Linkage::Internal if @single_module
     end
     initialized_flag
diff --git a/src/compiler/crystal/codegen/once.cr b/src/compiler/crystal/codegen/once.cr
index 2e91267c1f52..a15e84724787 100644
--- a/src/compiler/crystal/codegen/once.cr
+++ b/src/compiler/crystal/codegen/once.cr
@@ -1,37 +1,17 @@
 require "./codegen"
 
 class Crystal::CodeGenVisitor
-  ONCE_STATE = "~ONCE_STATE"
-
   def once_init
     if once_init_fun = typed_fun?(@main_mod, ONCE_INIT)
       once_init_fun = check_main_fun ONCE_INIT, once_init_fun
-
-      once_state_global = @main_mod.globals.add(once_init_fun.type.return_type, ONCE_STATE)
-      once_state_global.linkage = LLVM::Linkage::Internal if @single_module
-      once_state_global.initializer = once_init_fun.type.return_type.null
-
-      state = call once_init_fun
-      store state, once_state_global
+      call once_init_fun
     end
   end
 
   def run_once(flag, func : LLVMTypedFunction)
     once_fun = main_fun(ONCE)
-    once_init_fun = main_fun(ONCE_INIT)
-
-    # both of these should be Void*
-    once_state_type = once_init_fun.type.return_type
-    once_initializer_type = once_fun.func.params.last.type
-
-    once_state_global = @llvm_mod.globals[ONCE_STATE]? || begin
-      global = @llvm_mod.globals.add(once_state_type, ONCE_STATE)
-      global.linkage = LLVM::Linkage::External
-      global
-    end
-
-    state = load(once_state_type, once_state_global)
+    once_initializer_type = once_fun.func.params.last.type # must be Void*
     initializer = pointer_cast(func.func.to_value, once_initializer_type)
-    call once_fun, [state, flag, initializer]
+    call once_fun, [flag, initializer]
   end
 end
diff --git a/src/crystal/once.cr b/src/crystal/once.cr
index 56eea2be693a..5cab01980177 100644
--- a/src/crystal/once.cr
+++ b/src/crystal/once.cr
@@ -1,54 +1,125 @@
-# This file defines the functions `__crystal_once_init` and `__crystal_once` expected
-# by the compiler. `__crystal_once` is called each time a constant or class variable
-# has to be initialized and is its responsibility to verify the initializer is executed
-# only once. `__crystal_once_init` is executed only once at the beginning of the program
-# and the result is passed on each call to `__crystal_once`.
-
-# This implementation uses an array to store the initialization flag pointers for each value
-# to find infinite loops and raise an error. In multithread mode a mutex is used to
-# avoid race conditions between threads.
-
-# :nodoc:
-class Crystal::OnceState
-  @rec = [] of Bool*
-
-  def once(flag : Bool*, initializer : Void*)
-    unless flag.value
-      if @rec.includes?(flag)
-        raise "Recursion while initializing class variables and/or constants"
+# This file defines two functions expected by the compiler:
+#
+# - `__crystal_once_init`: executed only once at the beginning of the program
+#   and, for the legacy implementation, the result is passed on each call to
+#   `__crystal_once`.
+#
+# - `__crystal_once`: called each time a constant or class variable has to be
+#   initialized and is its responsibility to verify the initializer is executed
+#   only once and to fail on recursion.
+
+# In multithread mode a mutex is used to avoid race conditions between threads.
+#
+# On Win32, `Crystal::System::FileDescriptor#@@reader_thread` spawns a new
+# thread even without the `preview_mt` flag, and the thread can also reference
+# Crystal constants, leading to race conditions, so we always enable the mutex.
+
+{% if compare_versions(Crystal::VERSION, "1.15.0-dev") >= 0 %}
+  # This implementation uses an enum over the initialization flag pointer for
+  # each value to find infinite loops and raise an error.
+
+  module Crystal
+    enum OnceState : Int8
+      Processing    = -1
+      Uninitialized = 0
+      Initialized   = 1
+    end
+
+    {% if flag?(:preview_mt) || flag?(:win32) %}
+      @@once_mutex = uninitialized Mutex
+
+      def self.once_mutex : Mutex
+        @@once_mutex
       end
-      @rec << flag
 
-      Proc(Nil).new(initializer, Pointer(Void).null).call
-      flag.value = true
+      def self.once_mutex=(@@once_mutex : Mutex)
+      end
+    {% end %}
 
-      @rec.pop
+    # Using @[NoInline] so LLVM optimizes for the hot path (var already
+    # initialized).
+    @[NoInline]
+    def self.once(flag : OnceState*, initializer : Void*) : Nil
+      {% if flag?(:preview_mt) || flag?(:win32) %}
+        Crystal.once_mutex.synchronize { once_exec(flag, initializer) }
+      {% else %}
+        once_exec(flag, initializer)
+      {% end %}
+    end
+
+    private def self.once_exec(flag : OnceState*, initializer : Void*) : Nil
+      case flag.value
+      in .initialized?
+        return
+      in .uninitialized?
+        flag.value = :processing
+        Proc(Nil).new(initializer, Pointer(Void).null).call
+        flag.value = :initialized
+      in .processing?
+        raise "Recursion while initializing class variables and/or constants"
+      end
     end
   end
 
-  # on Win32, `Crystal::System::FileDescriptor#@@reader_thread` spawns a new
-  # thread even without the `preview_mt` flag, and the thread can also reference
-  # Crystal constants, leading to race conditions, so we always enable the mutex
-  # TODO: can this be improved?
-  {% if flag?(:preview_mt) || flag?(:win32) %}
-    @mutex = Mutex.new(:reentrant)
+  # :nodoc:
+  fun __crystal_once_init : Nil
+    {% if flag?(:preview_mt) || flag?(:win32) %}
+      Crystal.once_mutex = Mutex.new(:reentrant)
+    {% end %}
+  end
+
+  # :nodoc:
+  #
+  # Using `@[AlwaysInline]` allows LLVM to optimize const accesses. Since this
+  # is a `fun` the function will still appear in the symbol table, though it
+  # will never be called.
+  @[AlwaysInline]
+  fun __crystal_once(flag : Int8*, initializer : Void*) : Nil
+    flag = flag.as(Crystal::OnceState*)
+    Crystal.once(flag, initializer) unless flag.value.initialized?
+  end
+{% else %}
+  # This implementation uses a global array to store the initialization flag
+  # pointers for each value to find infinite loops and raise an error.
+
+  # :nodoc:
+  class Crystal::OnceState
+    @rec = [] of Bool*
 
     def once(flag : Bool*, initializer : Void*)
       unless flag.value
-        @mutex.synchronize do
-          previous_def
+        if @rec.includes?(flag)
+          raise "Recursion while initializing class variables and/or constants"
         end
+        @rec << flag
+
+        Proc(Nil).new(initializer, Pointer(Void).null).call
+        flag.value = true
+
+        @rec.pop
       end
     end
-  {% end %}
-end
-
-# :nodoc:
-fun __crystal_once_init : Void*
-  Crystal::OnceState.new.as(Void*)
-end
-
-# :nodoc:
-fun __crystal_once(state : Void*, flag : Bool*, initializer : Void*)
-  state.as(Crystal::OnceState).once(flag, initializer)
-end
+
+    {% if flag?(:preview_mt) || flag?(:win32) %}
+      @mutex = Mutex.new(:reentrant)
+
+      def once(flag : Bool*, initializer : Void*)
+        unless flag.value
+          @mutex.synchronize do
+            previous_def
+          end
+        end
+      end
+    {% end %}
+  end
+
+  # :nodoc:
+  fun __crystal_once_init : Void*
+    Crystal::OnceState.new.as(Void*)
+  end
+
+  # :nodoc:
+  fun __crystal_once(state : Void*, flag : Bool*, initializer : Void*)
+    state.as(Crystal::OnceState).once(flag, initializer)
+  end
+{% end %}