Add software prefetch to ByteSizeLong

Mkaes ByteSizeLong faster: BM_ByteSizeLong 15.7µs ± 4% 14.9µs ± 6% -5.17% (p=0.000 n=20+20) PiperOrigin-RevId: 605067418
protocolbuffers · Feb 7, 2024 · 948ca2e · 948ca2e
1 parent 85a8048
commit 948ca2e
Show file tree

Hide file tree

Showing 10 changed files with 77 additions and 2 deletions.
diff --git a/src/google/protobuf/compiler/cpp/message.cc b/src/google/protobuf/compiler/cpp/message.cc
@@ -4513,6 +4513,15 @@ void MessageGenerator::GenerateByteSize(io::Printer* p) {
       "// Prevent compiler warnings about cached_has_bits being unused\n"
       "(void) cached_has_bits;\n\n");
 
+  // See comment in third_party/protobuf/port.h for details,
+  // on how much we are prefetching. Only insert prefetch once per
+  // function, since advancing is actually slower. We sometimes
+  // prefetch more than sizeof(message), because it helps with
+  // next message on arena.
+  p->Emit(R"cc(
+    ::_pbi::Prefetch5LinesFrom7Lines(reinterpret_cast<const void*>(this));
+  )cc");
+
   while (it != end) {
     auto next = FindNextUnequalChunk(it, end, MayGroupChunksForHaswordsCheck);
     bool has_haswords_check = MaybeEmitHaswordsCheck(

diff --git a/src/google/protobuf/compiler/java/java_features.pb.cc b/src/google/protobuf/compiler/java/java_features.pb.cc
diff --git a/src/google/protobuf/compiler/plugin.pb.cc b/src/google/protobuf/compiler/plugin.pb.cc
diff --git a/src/google/protobuf/cpp_features.pb.cc b/src/google/protobuf/cpp_features.pb.cc
diff --git a/src/google/protobuf/descriptor.pb.cc b/src/google/protobuf/descriptor.pb.cc
diff --git a/src/google/protobuf/editions/golden/compare_cpp_codegen_failure.txt b/src/google/protobuf/editions/golden/compare_cpp_codegen_failure.txt
@@ -25,9 +25,9 @@
      target = ::proto2::internal::WireFormatLite::
          WriteInt32ToArrayWithField<1>(
 @@ @@
-   // Prevent compiler warnings about cached_has_bits being unused
    (void) cached_has_bits;
 
+   ::_pbi::Prefetch5LinesFrom7Lines(reinterpret_cast<const void*>(this));
 -  // optional int32 int32_field = 1;
 +  // int32 int32_field = 1;
    cached_has_bits = _impl_._has_bits_[0];