Be less aggressive about prefetching in bytesizelong

PiperOrigin-RevId: 609454594

src/google/protobuf/compiler/cpp/message.cc

@@ -4526,9 +4526,26 @@ void MessageGenerator::GenerateByteSize(io::Printer* p) {
   // 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(
+  bool generate_prefetch = false;
-    ::_pbi::Prefetch5LinesFrom7Lines(reinterpret_cast<const void*>(this));
+  // Skip trivial messages with 0 or 1 fields, unless they are repeated,
-  )cc");
+  // to reduce codesize.
+  switch (optimized_order_.size()) {
+    case 1:
+      generate_prefetch = optimized_order_[0]->is_repeated();
+      break;
+    case 0:
+      break;
+    default:
+      generate_prefetch = true;
+  }
+  if (!IsPresentMessage(descriptor_, options_)) {
+    generate_prefetch = false;
+  }
+  if (generate_prefetch) {
+    p->Emit(R"cc(
+      ::_pbi::Prefetch5LinesFrom7Lines(reinterpret_cast<const void*>(this));
+    )cc");
+  }
 
   while (it != end) {
     auto next = FindNextUnequalChunk(it, end, MayGroupChunksForHaswordsCheck);

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];

src/google/protobuf/editions/golden/compare_cpp_codegen_failure.xml

@@ -2,7 +2,7 @@
 <testsuites tests="1" name="AllTests">
 <testsuite name="EditionsCodegenTests">
 <testcase name="third_party/protobuf/editions/golden/simple_proto3.pb.cc" status="run" result="completed" classname="DiffTest">
-<failure message="Value of: third_party/protobuf/editions/golden/simple_proto3.pb.cc&#x0A;Expected: &#x0A;// Generated by the protocol buffer compiler.  DO NOT EDIT!&#x0A;// source: third_party/protobuf/editions/golden/simple_proto3.proto&#x0A;// Protobuf C++ Version: 0.20240110.0&#x0A;&#x0A;#include &quot;third_party/protobuf/editions/golden/simple_proto3.pb.h&quot;&#x0A;&#x0A;#include &lt;algorithm&gt;&#x0A;#include &quot;third_party/protobuf/io/coded_stream.h&quot;&#x0A;#include &quot;third_party/protobuf/generated_message_tctable_impl.h&quot;&#x0A;#include &quot;third_party/protobuf/extension_set.h&quot;&#x0A;#include &quot;third_party/protobuf/wire_format_lite.h&quot;&#x0A;#include &quot;third_party/protobuf/io/zero_copy_stream_impl_lite.h&quot;&#x0A;// @@protoc_insertion_point(includes)&#x0A;&#x0A;// Must be included last.&#x0A;, with the difference:&#x0A;@@ @@&#x0A;     ::_pbi::TcParser::GetTable&lt;::protobuf_editions_test::golden::SimpleProto3&gt;(),  // to_prefetch&#x0A;     #endif  // PROTOBUF_PREFETCH_PARSE_TABLE&#x0A;   }, {{&#x0A;-    // optional int32 int32_field = 1;&#x0A;+    // int32 int32_field = 1;&#x0A;     {::_pbi::TcParser::FastV32S1,&#x0A;      {8, 0, 0, PROTOBUF_FIELD_OFFSET(SimpleProto3, _impl_.int32_field_)}},&#x0A;   }}, {{&#x0A;     65535, 65535&#x0A;   }}, {{&#x0A;-    // optional int32 int32_field = 1;&#x0A;+    // int32 int32_field = 1;&#x0A;     {PROTOBUF_FIELD_OFFSET(SimpleProto3, _impl_.int32_field_), _Internal::kHasBitsOffset + 0, 0,&#x0A;     (0 | ::_fl::kFcOptional | ::_fl::kInt32)},&#x0A;   }},&#x0A;@@ @@&#x0A;   (void)cached_has_bits;&#x0A; &#x0A;   cached_has_bits = _impl_._has_bits_[0];&#x0A;-  // optional int32 int32_field = 1;&#x0A;+  // int32 int32_field = 1;&#x0A;   if (cached_has_bits &amp; 0x00000001u) {&#x0A;     target = ::proto2::internal::WireFormatLite::&#x0A;         WriteInt32ToArrayWithField&lt;1&gt;(&#x0A;@@ @@&#x0A;   (void) cached_has_bits;&#x0A; &#x0A;   ::_pbi::Prefetch5LinesFrom7Lines(reinterpret_cast&lt;const void*&gt;(this));&#x0A;-  // optional int32 int32_field = 1;&#x0A;+  // int32 int32_field = 1;&#x0A;   cached_has_bits = _impl_._has_bits_[0];&#x0A;   if (cached_has_bits &amp; 0x00000001u) {&#x0A;     total_size += ::_pbi::WireFormatLite::Int32SizePlusOne(" type=""></failure>
+<failure message="Value of: third_party/protobuf/editions/golden/simple_proto3.pb.cc&#x0A;Expected: &#x0A;// Generated by the protocol buffer compiler.  DO NOT EDIT!&#x0A;// source: third_party/protobuf/editions/golden/simple_proto3.proto&#x0A;// Protobuf C++ Version: 0.20240110.0&#x0A;&#x0A;#include &quot;third_party/protobuf/editions/golden/simple_proto3.pb.h&quot;&#x0A;&#x0A;#include &lt;algorithm&gt;&#x0A;#include &quot;third_party/protobuf/io/coded_stream.h&quot;&#x0A;#include &quot;third_party/protobuf/generated_message_tctable_impl.h&quot;&#x0A;#include &quot;third_party/protobuf/extension_set.h&quot;&#x0A;#include &quot;third_party/protobuf/wire_format_lite.h&quot;&#x0A;#include &quot;third_party/protobuf/io/zero_copy_stream_impl_lite.h&quot;&#x0A;// @@protoc_insertion_point(includes)&#x0A;&#x0A;// Must be included last.&#x0A;, with the difference:&#x0A;@@ @@&#x0A;     ::_pbi::TcParser::GetTable&lt;::protobuf_editions_test::golden::SimpleProto3&gt;(),  // to_prefetch&#x0A;     #endif  // PROTOBUF_PREFETCH_PARSE_TABLE&#x0A;   }, {{&#x0A;-    // optional int32 int32_field = 1;&#x0A;+    // int32 int32_field = 1;&#x0A;     {::_pbi::TcParser::FastV32S1,&#x0A;      {8, 0, 0, PROTOBUF_FIELD_OFFSET(SimpleProto3, _impl_.int32_field_)}},&#x0A;   }}, {{&#x0A;     65535, 65535&#x0A;   }}, {{&#x0A;-    // optional int32 int32_field = 1;&#x0A;+    // int32 int32_field = 1;&#x0A;     {PROTOBUF_FIELD_OFFSET(SimpleProto3, _impl_.int32_field_), _Internal::kHasBitsOffset + 0, 0,&#x0A;     (0 | ::_fl::kFcOptional | ::_fl::kInt32)},&#x0A;   }},&#x0A;@@ @@&#x0A;   (void)cached_has_bits;&#x0A; &#x0A;   cached_has_bits = _impl_._has_bits_[0];&#x0A;-  // optional int32 int32_field = 1;&#x0A;+  // int32 int32_field = 1;&#x0A;   if (cached_has_bits &amp; 0x00000001u) {&#x0A;     target = ::proto2::internal::WireFormatLite::&#x0A;         WriteInt32ToArrayWithField&lt;1&gt;(&#x0A;@@ @@&#x0A;   // Prevent compiler warnings about cached_has_bits being unused&#x0A;   (void) cached_has_bits;&#x0A; &#x0A;-  // optional int32 int32_field = 1;&#x0A;+  // int32 int32_field = 1;&#x0A;   cached_has_bits = _impl_._has_bits_[0];&#x0A;   if (cached_has_bits &amp; 0x00000001u) {&#x0A;     total_size += ::_pbi::WireFormatLite::Int32SizePlusOne(" type=""></failure>
 </testcase>
 <testcase name="third_party/protobuf/editions/golden/simple_proto3.pb.h" status="run" result="completed" classname="DiffTest">
 <failure message="Value of: third_party/protobuf/editions/golden/simple_proto3.pb.h&#x0A;Expected: &#x0A;// Generated by the protocol buffer compiler.  DO NOT EDIT!&#x0A;// source: third_party/protobuf/editions/golden/simple_proto3.proto&#x0A;// Protobuf C++ Version: 0.20240110.0&#x0A;&#x0A;#ifndef GOOGLE_PROTOBUF_INCLUDED_third_5fparty_2fprotobuf_2feditions_2fgolden_2fsimple_5fproto3_2eproto_2epb_2eh&#x0A;#define GOOGLE_PROTOBUF_INCLUDED_third_5fparty_2fprotobuf_2feditions_2fgolden_2fsimple_5fproto3_2eproto_2epb_2eh&#x0A;&#x0A;#include &lt;limits&gt;&#x0A;#include &lt;string&gt;&#x0A;#include &lt;type_traits&gt;&#x0A;#include &lt;utility&gt;&#x0A;&#x0A;#include &quot;third_party/protobuf/io/coded_stream.h&quot;&#x0A;#include &quot;third_party/protobuf/arena.h&quot;&#x0A;#include &quot;third_party/protobuf/arenastring.h&quot;&#x0A;#include &quot;third_party/protobuf/generated_message_tctable_decl.h&quot;&#x0A;#include &quot;third_party/protobuf/generated_message_util.h&quot;&#x0A;#include &quot;third_party/protobuf/metadata_lite.h&quot;&#x0A;#include &quot;third_party/protobuf/message_lite.h&quot;&#x0A;// @@protoc_insertion_point(includes)&#x0A;&#x0A;// Must be included last.&#x0A;&#x0A;#endif  // GOOGLE_PROTOBUF_INCLUDED_third_5fparty_2fprotobuf_2feditions_2fgolden_2fsimple_5fproto3_2eproto_2epb_2eh&#x0A;, with the difference:&#x0A;@@ @@&#x0A;   enum : int {&#x0A;     kInt32FieldFieldNumber = 1,&#x0A;   };&#x0A;-  // optional int32 int32_field = 1;&#x0A;+  // int32 int32_field = 1;&#x0A;   bool has_int32_field() const;&#x0A;   void clear_int32_field() ;&#x0A;   ::int32_t int32_field() const;&#x0A;@@ @@&#x0A; &#x0A; // SimpleProto3&#x0A; &#x0A;-// optional int32 int32_field = 1;&#x0A;+// int32 int32_field = 1;&#x0A; inline bool SimpleProto3::has_int32_field() const {&#x0A;   bool value = (_impl_._has_bits_[0] &amp; 0x00000001u) != 0;&#x0A;   return value;" type=""></failure>