Grow the json_view dump's output in steps and compare raw-number dumps

- The output string reserves the size estimate (the source extent) and grows
  in steps of 64 KiB within it, instead of being resized to the estimate at
  once: resize() zero-fills, and for a pretty-printed source the estimate is
  far larger than the compact output. citm_catalog dump: 342 -> 283 us on
  x86-64, 144 -> 134 us on Apple M1.
- bench_corpus compares the dump with source numbers with yyjson writing
  numbers read as raw text (YYJSON_READ_NUMBER_AS_RAW); both write the same
  bytes.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

include/nlohmann/detail/view/builder.hpp

@@ -828,14 +828,19 @@ indent_done:
             const auto idx = static_cast<std::uint32_t>(emit(k, 0, 0, static_cast<std::size_t>(p - b), 0) - base);
             if (depth != 0)
             {
-                const frame f = {cur_idx, cur_count, cur_is_object};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    cold.shallow[depth - 1] = f;
+                    // field by field: a frame put together on the stack and
+                    // copied would be read back wider than it was written,
+                    // and that load waits until the stores are done
+                    frame& f = cold.shallow[depth - 1];
+                    f.idx = cur_idx;
+                    f.count = cur_count;
+                    f.is_object = cur_is_object;
                 }
                 else
                 {
-                    cold.deep.push_back(f);
+                    cold.deep.push_back(frame{cur_idx, cur_count, cur_is_object});
                 }
             }
             ++depth;
@@ -851,19 +856,21 @@ indent_done:
             n.next = static_cast<std::uint32_t>(out - base) - cur_idx;
             if (--depth != 0)
             {
-                frame f{};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    f = cold.shallow[depth - 1];
+                    const frame& f = cold.shallow[depth - 1];
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
                 else
                 {
-                    f = cold.deep.back();
+                    const frame f = cold.deep.back();
                     cold.deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
-                cur_idx = f.idx;
-                cur_count = f.count;
-                cur_is_object = f.is_object;
             }
         }
 

include/nlohmann/detail/view/scan.hpp

@@ -67,13 +67,19 @@ NLOHMANN_VIEW_ALWAYS_INLINE std::uint16_t load16(const unsigned char* p) noexcep
 /// in predicted branches: 16 for keys, whose lengths repeat from record to
 /// record, and 8 for string values (Value) where a vector loop follows, as
 /// their lengths vary more. Longer runs continue 16 bytes at a time with NEON
-/// or SSE2, else eight bytes at a time.
+/// or SSE2, else eight bytes at a time. With SSE2, the run is checked 16 bytes
+/// at a time from its first byte instead: on x86-64, one compare that finds
+/// the end of most keys and short values is faster than a branch per byte (on
+/// AArch64, where a NEON mask costs more and branches predict well, slower).
 template<bool Value = false>
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned char* p, const unsigned char* e) noexcept
 {
     const std::uint8_t* plain = string_plain();
     for (;;)
     {
+#if NLOHMANN_VIEW_SSE2
+        p = vector_plain_run(p, e);
+#else
         if (e - p >= 16)
         {
 #define NLOHMANN_VIEW_STEP(i) if (NLOHMANN_VIEW_LIKELY(plain[p[i]] != 0)) {} else { p += (i); goto stop; }
@@ -107,6 +113,7 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
 #endif
             continue;
         }
+#endif
         while (p != e && plain[*p] != 0)
         {
             ++p;
@@ -115,7 +122,9 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
         {
             return p;
         }
+#if !NLOHMANN_VIEW_SSE2
 stop:
+#endif
         if (*p < 0x80)
         {
             return p; // quote, backslash, or control character

include/nlohmann/detail/view/serializer.hpp

@@ -31,7 +31,11 @@ namespace view
 {
 
 /// append-only output buffer: writes through a raw pointer into a string that
-/// is resized ahead, and trimmed by finish()
+/// is resized ahead, and trimmed by finish(). The estimate is reserved, and the
+/// string grows in steps of 64 KiB within it: resize() fills the new bytes
+/// with zeros (before C++23, a string cannot grow without), and a small step
+/// is filled while the writer is about to use it, in the cache, instead of
+/// filling the whole estimate in memory first.
 template<typename StringType>
 class output_buffer
 {
@@ -93,16 +97,31 @@ class output_buffer
     }
 
   private:
+    /// the size of a growth step (a function: std::min() takes a reference,
+    /// which a static constexpr member does not have before C++17)
+    static constexpr std::size_t step() noexcept
+    {
+        return 65536;
+    }
+
     static StringType& sized(StringType& out, std::size_t estimate)
     {
-        out.resize((std::max)(estimate, static_cast<std::size_t>(64)));
+        out.reserve(estimate);
+        out.resize((std::min)((std::max)(estimate, static_cast<std::size_t>(64)), step()));
         return out;
     }
 
     NLOHMANN_VIEW_NOINLINE void grow(std::size_t n)
     {
         const auto used = static_cast<std::size_t>(m_pos - m_out.data());
-        m_out.resize((std::max)(m_out.size() * 2, used + n + 256));
+        // (a step does not go beyond the reserved estimate, so that a good
+        // estimate is never copied to a larger allocation)
+        const std::size_t size = (std::max)((std::min)(m_out.size() + step(), m_out.capacity()), used + n + 256);
+        if (size > m_out.capacity())
+        {
+            m_out.reserve((std::max)(m_out.capacity() * 2, size));
+        }
+        m_out.resize(size);
         m_pos = &m_out[0] + used;
         m_end = &m_out[0] + m_out.size();
     }

single_include/nlohmann/json_view.hpp

@@ -917,13 +917,19 @@ NLOHMANN_VIEW_ALWAYS_INLINE std::uint16_t load16(const unsigned char* p) noexcep
 /// in predicted branches: 16 for keys, whose lengths repeat from record to
 /// record, and 8 for string values (Value) where a vector loop follows, as
 /// their lengths vary more. Longer runs continue 16 bytes at a time with NEON
-/// or SSE2, else eight bytes at a time.
+/// or SSE2, else eight bytes at a time. With SSE2, the run is checked 16 bytes
+/// at a time from its first byte instead: on x86-64, one compare that finds
+/// the end of most keys and short values is faster than a branch per byte (on
+/// AArch64, where a NEON mask costs more and branches predict well, slower).
 template<bool Value = false>
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned char* p, const unsigned char* e) noexcept
 {
     const std::uint8_t* plain = string_plain();
     for (;;)
     {
+#if NLOHMANN_VIEW_SSE2
+        p = vector_plain_run(p, e);
+#else
         if (e - p >= 16)
         {
 #define NLOHMANN_VIEW_STEP(i) if (NLOHMANN_VIEW_LIKELY(plain[p[i]] != 0)) {} else { p += (i); goto stop; }
@@ -957,6 +963,7 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
 #endif
             continue;
         }
+#endif
         while (p != e && plain[*p] != 0)
         {
             ++p;
@@ -965,7 +972,9 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
         {
             return p;
         }
+#if !NLOHMANN_VIEW_SSE2
 stop:
+#endif
         if (*p < 0x80)
         {
             return p; // quote, backslash, or control character
@@ -1873,14 +1882,19 @@ indent_done:
             const auto idx = static_cast<std::uint32_t>(emit(k, 0, 0, static_cast<std::size_t>(p - b), 0) - base);
             if (depth != 0)
             {
-                const frame f = {cur_idx, cur_count, cur_is_object};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    cold.shallow[depth - 1] = f;
+                    // field by field: a frame put together on the stack and
+                    // copied would be read back wider than it was written,
+                    // and that load waits until the stores are done
+                    frame& f = cold.shallow[depth - 1];
+                    f.idx = cur_idx;
+                    f.count = cur_count;
+                    f.is_object = cur_is_object;
                 }
                 else
                 {
-                    cold.deep.push_back(f);
+                    cold.deep.push_back(frame{cur_idx, cur_count, cur_is_object});
                 }
             }
             ++depth;
@@ -1896,19 +1910,21 @@ indent_done:
             n.next = static_cast<std::uint32_t>(out - base) - cur_idx;
             if (--depth != 0)
             {
-                frame f{};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    f = cold.shallow[depth - 1];
+                    const frame& f = cold.shallow[depth - 1];
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
                 else
                 {
-                    f = cold.deep.back();
+                    const frame f = cold.deep.back();
                     cold.deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
-                cur_idx = f.idx;
-                cur_count = f.count;
-                cur_is_object = f.is_object;
             }
         }
 
@@ -5393,7 +5409,11 @@ namespace view
 {
 
 /// append-only output buffer: writes through a raw pointer into a string that
-/// is resized ahead, and trimmed by finish()
+/// is resized ahead, and trimmed by finish(). The estimate is reserved, and the
+/// string grows in steps of 64 KiB within it: resize() fills the new bytes
+/// with zeros (before C++23, a string cannot grow without), and a small step
+/// is filled while the writer is about to use it, in the cache, instead of
+/// filling the whole estimate in memory first.
 template<typename StringType>
 class output_buffer
 {
@@ -5455,16 +5475,31 @@ class output_buffer
     }
 
   private:
+    /// the size of a growth step (a function: std::min() takes a reference,
+    /// which a static constexpr member does not have before C++17)
+    static constexpr std::size_t step() noexcept
+    {
+        return 65536;
+    }
+
     static StringType& sized(StringType& out, std::size_t estimate)
     {
-        out.resize((std::max)(estimate, static_cast<std::size_t>(64)));
+        out.reserve(estimate);
+        out.resize((std::min)((std::max)(estimate, static_cast<std::size_t>(64)), step()));
         return out;
     }
 
     NLOHMANN_VIEW_NOINLINE void grow(std::size_t n)
     {
         const auto used = static_cast<std::size_t>(m_pos - m_out.data());
-        m_out.resize((std::max)(m_out.size() * 2, used + n + 256));
+        // (a step does not go beyond the reserved estimate, so that a good
+        // estimate is never copied to a larger allocation)
+        const std::size_t size = (std::max)((std::min)(m_out.size() + step(), m_out.capacity()), used + n + 256);
+        if (size > m_out.capacity())
+        {
+            m_out.reserve((std::max)(m_out.capacity() * 2, size));
+        }
+        m_out.resize(size);
         m_pos = &m_out[0] + used;
         m_end = &m_out[0] + m_out.size();
     }

tests/benchmarks/json_view/README.md

@@ -50,7 +50,8 @@ JSON-RPC request (`rpc`):
 | dump | serialize a parsed document (compact) |
 
 `bench_corpus.cpp` runs parse, traverse, and dump on any list of files, so that no library is tuned to a handful of
-documents.
+documents. Its dump also writes the numbers as they are in the input: `json_view` with `number_format::source`, and
+yyjson with numbers read as raw text (`YYJSON_READ_NUMBER_AS_RAW`), without converting them.
 
 `bench_edit.cpp` measures read-modify-write: parse, apply the same logical edits with each library's own API, and
 serialize (compact). Workloads: `patch` (a handful of edits at fixed places) and `update` (edits in every record).

tests/benchmarks/json_view/bench_corpus.cpp

@@ -15,7 +15,8 @@
 // count, string bytes, sum of numbers) before anything is timed. Workloads:
 // parse (build and free a document), traverse (visit every value, convert
 // every number), dump (compact), and for json_view also dump with the source
-// number text. Results go to bench_corpus.csv.
+// number text, compared with yyjson writing numbers read as raw text
+// (YYJSON_READ_NUMBER_AS_RAW). Results go to bench_corpus.csv.
 #include <nlohmann/json_view.hpp>
 
 #if JSON_VIEW_BENCH_BOOST
@@ -272,6 +273,7 @@ int main(int argc, char** argv)
         };
         json_document vd = json_document::parse(s);
         yyjson_doc* yd = yyjson_read(s.data(), s.size(), 0);
+        yyjson_doc* yd_raw = yyjson_read(s.data(), s.size(), YYJSON_READ_NUMBER_AS_RAW);
         simdjson::dom::parser sjd;
         const simdjson::dom::element se = sjd.parse(ps).value_unsafe();
         const std::vector<std::pair<std::string, std::vector<engine>>> workloads =
@@ -304,6 +306,7 @@ int main(int argc, char** argv)
                     {"yyjson", [&] { std::size_t n = 0; char* o = yyjson_write(yd, 0, &n); g_sink = static_cast<double>(n); std::free(o); }},
                     {"simdjson DOM", [&] { std::string o = simdjson::to_string(se); g_sink = static_cast<double>(o.size()); }},
                     {"json_view (source numbers)", [&] { std::string o = vd.root().dump(-1, ' ', false, json_view::number_format::source); g_sink = static_cast<double>(o.size()); }},
+                    {"yyjson (raw numbers)", [&] { std::size_t n = 0; char* o = yyjson_write(yd_raw, 0, &n); g_sink = static_cast<double>(n); std::free(o); }},
                 }
             },
         };
@@ -333,6 +336,7 @@ int main(int argc, char** argv)
             std::fflush(stdout);
         }
         yyjson_doc_free(yd);
+        yyjson_doc_free(yd_raw);
     }
     std::fclose(csv);
 }