[struct_pack] add document, bench, example for fast_varint (#503)

* improve varint performance

* add bench for fast_var_int

* add document & example for fast_varint

README.md

@@ -165,7 +165,7 @@ More examples [here](https://github.com/alibaba/yalantinglibs/tree/main/src/coro
 
 Based on compile-time reflection, very easy to use, high performance serialization library, struct_pack is a header only library, it is used by coro_rpc now.
 
-Only one line code to finish serialization and deserialization, 2-50x faster than protobuf.
+Only one line code to finish serialization and deserialization, 2-20x faster than protobuf.
 
 ### quick example
 ```cpp

include/ylt/struct_pack/calculate_size.hpp

@@ -263,10 +263,21 @@ constexpr bool STRUCT_PACK_INLINE has_64bits_varint() {
   }
 }
 
-template <uint64_t parent_tag, typename Arg>
+template <uint64_t parent_tag, typename Arg, typename... Args>
+constexpr std::size_t STRUCT_PACK_INLINE get_int_len() {
+  if (has_64bits_varint<parent_tag, Arg, Args...>()) {
+    return 8;
+  }
+  else {
+    return 4;
+  }
+}
+
+template <uint64_t parent_tag, typename Arg, typename unsigned_t,
+          typename signed_t>
 constexpr void STRUCT_PACK_INLINE get_fast_varint_width_impl(
     const Arg &item, int &non_zero_cnt32, int &non_zero_cnt64,
-    uint64_t &unsigned_max, int64_t &signed_max) {
+    unsigned_t &unsigned_max, signed_t &signed_max) {
   if constexpr (varint_t<Arg, parent_tag>) {
     if (get_varint_value(item)) {
       if constexpr (sizeof(Arg) == 4) {
@@ -280,27 +291,29 @@ constexpr void STRUCT_PACK_INLINE get_fast_varint_width_impl(
       }
       if constexpr (std::is_unsigned_v<std::remove_reference_t<
                         decltype(get_varint_value(item))>>) {
-        unsigned_max = std::max<uint64_t>(unsigned_max, get_varint_value(item));
+        unsigned_max =
+            std::max<unsigned_t>(unsigned_max, get_varint_value(item));
       }
       else {
         signed_max =
-            std::max<int64_t>(signed_max, get_varint_value(item) > 0
-                                              ? get_varint_value(item)
-                                              : -(get_varint_value(item) + 1));
+            std::max<signed_t>(signed_max, get_varint_value(item) > 0
+                                               ? get_varint_value(item)
+                                               : -(get_varint_value(item) + 1));
       }
     }
   }
 }
 
-template <uint64_t parent_tag, typename... Args>
+template <uint64_t parent_tag, typename... Args, typename unsigned_t,
+          typename signed_t>
 constexpr int STRUCT_PACK_INLINE
-get_fast_varint_width_from_max(uint64_t unsigned_max, int64_t signed_max) {
+get_fast_varint_width_from_max(unsigned_t unsigned_max, signed_t signed_max) {
   int width_unsigned = 0, width_signed = 0;
   if constexpr (has_unsigned_varint<parent_tag, Args...>()) {
     if SP_LIKELY (unsigned_max <= UINT8_MAX) {
       width_unsigned = 0;
     }
-    else if (unsigned_max <= UINT16_MAX) {
+    else if SP_LIKELY (unsigned_max <= UINT16_MAX) {
       width_unsigned = 1;
     }
     else if (unsigned_max <= UINT32_MAX) {
@@ -314,7 +327,7 @@ get_fast_varint_width_from_max(uint64_t unsigned_max, int64_t signed_max) {
     if SP_LIKELY (signed_max <= INT8_MAX) {
       width_signed = 0;
     }
-    else if (signed_max <= INT16_MAX) {
+    else if SP_LIKELY (signed_max <= INT16_MAX) {
       width_signed = 1;
     }
     else if (signed_max <= INT32_MAX) {
@@ -342,8 +355,8 @@ get_fast_varint_width_from_max(uint64_t unsigned_max, int64_t signed_max) {
 
 template <uint64_t parent_tag, typename... Args>
 constexpr int STRUCT_PACK_INLINE get_fast_varint_width(const Args &...items) {
-  uint64_t unsigned_max = 0;
-  int64_t signed_max = 0;
+  typename uint_t<get_int_len<parent_tag, Args...>()>::type unsigned_max = 0;
+  typename int_t<get_int_len<parent_tag, Args...>()>::type signed_max = 0;
   int non_zero_cnt32 = 0, non_zero_cnt64 = 0;
   (get_fast_varint_width_impl<parent_tag>(items, non_zero_cnt32, non_zero_cnt64,
                                           unsigned_max, signed_max),

include/ylt/struct_pack/packer.hpp

@@ -100,9 +100,9 @@ class packer {
     }
     if constexpr (hash_head % 2) {  // has more metainfo
       auto metainfo = info.metainfo();
-      std::size_t sz = info.size();
       write_wrapper<sizeof(char)>(writer_, (char *)&metainfo);
       if constexpr (serialize_static_config<serialize_type>::has_compatible) {
+        std::size_t sz = info.size();
         switch (metainfo & 0b11) {
           case 1:
             low_bytes_write_wrapper<2>(writer_, sz);
@@ -148,23 +148,24 @@ class packer {
     }
   }
 
-  template <uint64_t parent_tag, std::size_t sz, typename Arg>
+  template <uint64_t parent_tag, std::size_t sz, typename Arg,
+            typename unsigned_t, typename signed_t>
   static constexpr void STRUCT_PACK_INLINE
   get_fast_varint_width_impl(char (&vec)[sz], unsigned int &i, const Arg &item,
-                             uint64_t &unsigned_max, int64_t &signed_max) {
+                             unsigned_t &unsigned_max, signed_t &signed_max) {
     if constexpr (varint_t<Arg, parent_tag>) {
       if (get_varint_value(item) != 0) {
         vec[i / 8] |= (0b1 << (i % 8));
         if constexpr (std::is_unsigned_v<std::remove_reference_t<
                           decltype(get_varint_value(item))>>) {
           unsigned_max =
-              std::max<uint64_t>(unsigned_max, get_varint_value(item));
+              std::max<unsigned_t>(unsigned_max, get_varint_value(item));
         }
         else {
-          signed_max = std::max<int64_t>(signed_max,
-                                         get_varint_value(item) > 0
-                                             ? get_varint_value(item)
-                                             : -(get_varint_value(item) + 1));
+          signed_max = std::max<signed_t>(signed_max,
+                                          get_varint_value(item) > 0
+                                              ? get_varint_value(item)
+                                              : -(get_varint_value(item) + 1));
         }
       }
       ++i;
@@ -174,8 +175,8 @@ class packer {
   template <uint64_t parent_tag, std::size_t sz, typename... Args>
   static constexpr int STRUCT_PACK_INLINE
   get_fast_varint_width(char (&vec)[sz], const Args &...items) {
-    uint64_t unsigned_max = 0;
-    int64_t signed_max = 0;
+    typename uint_t<get_int_len<parent_tag, Args...>()>::type unsigned_max = 0;
+    typename int_t<get_int_len<parent_tag, Args...>()>::type signed_max = 0;
     unsigned int i = 0;
     (get_fast_varint_width_impl<parent_tag>(vec, i, items, unsigned_max,
                                             signed_max),
@@ -535,35 +536,40 @@ STRUCT_PACK_MAY_INLINE void serialize_to(Writer &writer,
 #endif
   static_assert(sizeof...(args) > 0);
   detail::packer<Writer, detail::get_args_type<Args...>> o(writer, info);
-  switch ((info.metainfo() & 0b11000) >> 3) {
-    case 0:
-      o.template serialize<conf, 1>(args...);
-      break;
+  if constexpr (!check_if_has_container<detail::get_args_type<Args...>>()) {
+    o.template serialize<conf, 1>(args...);
+  }
+  else {
+    switch ((info.metainfo() & 0b11000) >> 3) {
+      case 0:
+        o.template serialize<conf, 1>(args...);
+        break;
 #ifdef STRUCT_PACK_OPTIMIZE
-    case 1:
-      o.template serialize<conf, 2>(args...);
-      break;
-    case 2:
-      o.template serialize<conf, 4>(args...);
-      break;
-    case 3:
-      if constexpr (sizeof(std::size_t) >= 8) {
-        o.template serialize<conf, 8>(args...);
-      }
-      else {
-        unreachable();
-      }
-      break;
+      case 1:
+        o.template serialize<conf, 2>(args...);
+        break;
+      case 2:
+        o.template serialize<conf, 4>(args...);
+        break;
+      case 3:
+        if constexpr (sizeof(std::size_t) >= 8) {
+          o.template serialize<conf, 8>(args...);
+        }
+        else {
+          unreachable();
+        }
+        break;
 #else
-    case 1:
-    case 2:
-    case 3:
-      o.template serialize<conf, 2>(args...);
-      break;
+      case 1:
+      case 2:
+      case 3:
+        o.template serialize<conf, 2>(args...);
+        break;
 #endif
-    default:
-      detail::unreachable();
-      break;
-  };
+      default:
+        detail::unreachable();
+        break;
+    };
+  }
 }
 }  // namespace struct_pack::detail

include/ylt/struct_pack/util.h

@@ -137,7 +137,9 @@ void string_set_length_hacker(std::string &, std::size_t);
 template <typename ch>
 inline void resize(std::basic_string<ch> &raw_str, std::size_t sz) {
   std::string &str = *reinterpret_cast<std::string *>(&raw_str);
-#if defined(__GLIBCXX__) || defined(_LIBCPP_VERSION) || \
+#if defined(__SANITIZE_ADDRESS__)
+  raw_str.resize(sz);
+#elif defined(__GLIBCXX__) || defined(_LIBCPP_VERSION) || \
     defined(_MSVC_STL_VERSION)
   if (sz > str.capacity()) {
     str.reserve(sz);
@@ -193,7 +195,9 @@ template class vector_thief<decltype(&std::vector<char>::_Mypair),
 
 template <typename ch>
 inline void resize(std::vector<ch> &raw_vec, std::size_t sz) {
-#if defined(__GLIBCXX__) ||                                       \
+#if defined(__SANITIZE_ADDRESS__)
+  raw_vec.resize(sz);
+#elif defined(__GLIBCXX__) ||                                     \
     (defined(_LIBCPP_VERSION) && defined(_LIBCPP_HAS_NO_ASAN)) || \
     defined(_MSVC_STL_VERSION)
   std::vector<char> &vec = *reinterpret_cast<std::vector<char> *>(&raw_vec);

include/ylt/struct_pack/varint.hpp

@@ -52,6 +52,29 @@ struct int_t<8> {
   using type = int64_t;
 };
 
+template <std::size_t bytes_width>
+struct uint_t;
+
+template <>
+struct uint_t<1> {
+  using type = uint8_t;
+};
+
+template <>
+struct uint_t<2> {
+  using type = uint16_t;
+};
+
+template <>
+struct uint_t<4> {
+  using type = uint32_t;
+};
+
+template <>
+struct uint_t<8> {
+  using type = uint64_t;
+};
+
 template <typename T>
 class varint {
  public:

src/struct_pack/benchmark/config.hpp

@@ -15,11 +15,23 @@ enum class LibType {
   PROTOBUF,
   FLATBUFFER,
 };
-enum class SampleType { RECT, RECTS, PERSON, PERSONS, MONSTER, MONSTERS };
+enum class SampleType {
+  RECT,
+  RECTS,
+  VAR_RECT,
+  VAR_RECTS,
+  PERSON,
+  PERSONS,
+  MONSTER,
+  MONSTERS
+};
 
 inline const std::unordered_map<SampleType, std::string> g_sample_name_map = {
     {SampleType::RECT, "1 rect"},
     {SampleType::RECTS, std::to_string(OBJECT_COUNT) + " rects"},
+    {SampleType::VAR_RECT, "1 rect(with fast varint edcode)"},
+    {SampleType::VAR_RECTS,
+     std::to_string(OBJECT_COUNT) + " rects(with fast varint edcode)"},
     {SampleType::PERSON, "1 person"},
     {SampleType::PERSONS, std::to_string(OBJECT_COUNT) + " persons"},
     {SampleType::MONSTER, "1 monster"},

src/struct_pack/benchmark/data_def.hpp

@@ -15,6 +15,7 @@
  */
 #pragma once
 
+#include "ylt/struct_pack.hpp"
 #include "ylt/struct_pack/reflection.hpp"
 #if __has_include(<msgpack.hpp>)
 #define HAVE_MSGPACK 1
@@ -60,6 +61,28 @@ inline constexpr struct_pack::sp_config set_sp_config(
   return struct_pack::DISABLE_ALL_META_INFO;
 }
 
+template <typename T>
+struct rect2 {
+  T x;
+  T y;
+  T width;
+  T height;
+#ifdef HAVE_MSGPACK
+  MSGPACK_DEFINE(x, y, width, height);
+#endif
+};
+
+inline constexpr struct_pack::sp_config set_sp_config(rect2<int32_t> *) {
+  return struct_pack::sp_config{struct_pack::sp_config::DISABLE_ALL_META_INFO |
+                                struct_pack::sp_config::USE_FAST_VARINT |
+                                struct_pack::sp_config::ENCODING_WITH_VARINT};
+}
+
+inline constexpr struct_pack::sp_config set_sp_config(
+    std::vector<rect2<int32_t>> *) {
+  return struct_pack::sp_config{struct_pack::DISABLE_ALL_META_INFO};
+}
+
 enum Color : uint8_t { Red, Green, Blue };
 
 struct Vec3 {

src/struct_pack/benchmark/flatbuffer_sample.hpp

@@ -35,7 +35,7 @@ inline void create<fb::rect32s>(flatbuffers::FlatBufferBuilder &builder,
                                 size_t object_count) {
   static std::vector<fb::rect32> rects_vector;
   rects_vector.clear();
-  rects_vector.assign(object_count, fb::rect32{1, 11, 1111, 1111111});
+  rects_vector.assign(object_count, fb::rect32{1, 0, 11, 1});
   auto rects = builder.CreateVectorOfStructs(rects_vector);
   auto orc = fb::Createrect32s(builder, rects);
   builder.Finish(orc);
@@ -44,7 +44,7 @@ inline void create<fb::rect32s>(flatbuffers::FlatBufferBuilder &builder,
 template <>
 inline void create<fb::rect32>(flatbuffers::FlatBufferBuilder &builder,
                                size_t object_count) {
-  const auto orc = fb::rect32{1, 11, 1111, 1111111};
+  const auto orc = fb::rect32{1, 0, 11, 1};
   auto offset = builder.CreateStruct(orc);
   builder.Finish(offset);
 }

src/struct_pack/benchmark/protobuf_sample.hpp

@@ -27,9 +27,9 @@ inline auto create_rects(size_t object_count) {
   for (int i = 0; i < object_count; i++) {
     mygame::rect32 *rc = rcs.add_rect32_list();
     rc->set_x(1);
-    rc->set_y(11);
-    rc->set_width(1111);
-    rc->set_height(111111);
+    rc->set_y(0);
+    rc->set_width(11);
+    rc->set_height(1);
   }
   return rcs;
 }

src/struct_pack/benchmark/sample.hpp

@@ -19,25 +19,16 @@ struct base_sample {
 
   void print_buffer_size(LibType type) {
     auto lib_name = get_lib_name(type);
-    for (auto sample_type : g_sample_type_vec) {
+    for (auto iter : buf_size_map_) {
       std::string prefix = lib_name;
-      prefix.append(" serialize ").append(get_sample_name(sample_type));
+      prefix.append(" serialize ").append(get_sample_name(iter.first));
 
       auto space_str = get_space_str(prefix.size(), 36);
-      std::cout << prefix << space_str
-                << " buffer size = " << buffer_size(sample_type) << "\n";
+      std::cout << prefix << space_str << " buffer size = " << iter.second
+                << "\n";
     }
   }
 
-  virtual std::size_t buffer_size(SampleType type) const {
-    auto it = buf_size_map_.find(type);
-    if (it == buf_size_map_.end()) {
-      throw std::runtime_error("unknown sample type");
-    }
-
-    return it->second;
-  }
-
   auto &get_ser_time_elapsed_map() { return ser_time_elapsed_map_; }
   auto &get_deser_time_elapsed_map() { return deser_time_elapsed_map_; }