Insert software prefetches into merge functions. This improves performance when hardware prefetchers are disabled on AMD machines.

PiperOrigin-RevId: 671468969

src/google/protobuf/extension_set.cc

@@ -971,6 +971,7 @@ size_t SizeOfUnion(ItX it_dest, ItX end_dest, ItY it_source, ItY end_source) {
 
 void ExtensionSet::MergeFrom(const MessageLite* extendee,
                              const ExtensionSet& other) {
+  Prefetch5LinesFrom1Line(&other);
   if (PROTOBUF_PREDICT_TRUE(!is_large())) {
     if (PROTOBUF_PREDICT_TRUE(!other.is_large())) {
       GrowCapacity(SizeOfUnion(flat_begin(), flat_end(), other.flat_begin(),

src/google/protobuf/port.h

@@ -306,6 +306,15 @@ inline PROTOBUF_ALWAYS_INLINE void Prefetch5LinesFrom7Lines(const void* ptr) {
   PROTOBUF_PREFETCH_WITH_OFFSET(ptr, 704);
 }
 
+// Prefetch 5 64-byte cache lines starting from 1 cache-line ahead.
+inline PROTOBUF_ALWAYS_INLINE void Prefetch5LinesFrom1Line(const void* ptr) {
+  PROTOBUF_PREFETCH_WITH_OFFSET(ptr, 64);
+  PROTOBUF_PREFETCH_WITH_OFFSET(ptr, 128);
+  PROTOBUF_PREFETCH_WITH_OFFSET(ptr, 192);
+  PROTOBUF_PREFETCH_WITH_OFFSET(ptr, 256);
+  PROTOBUF_PREFETCH_WITH_OFFSET(ptr, 320);
+}
+
 #if defined(NDEBUG) && ABSL_HAVE_BUILTIN(__builtin_unreachable)
 [[noreturn]] ABSL_ATTRIBUTE_COLD PROTOBUF_ALWAYS_INLINE inline void
 Unreachable() {

src/google/protobuf/repeated_ptr_field.cc

@@ -132,6 +132,7 @@ memswap<ArenaOffsetHelper<RepeatedPtrFieldBase>::value>(
 template <>
 void RepeatedPtrFieldBase::MergeFrom<std::string>(
     const RepeatedPtrFieldBase& from) {
+  Prefetch5LinesFrom1Line(&from);
   ABSL_DCHECK_NE(&from, this);
   int new_size = current_size_ + from.current_size_;
   auto dst = reinterpret_cast<std::string**>(InternalReserve(new_size));
@@ -159,6 +160,7 @@ void RepeatedPtrFieldBase::MergeFrom<std::string>(
 
 int RepeatedPtrFieldBase::MergeIntoClearedMessages(
     const RepeatedPtrFieldBase& from) {
+  Prefetch5LinesFrom1Line(&from);
   auto dst = reinterpret_cast<MessageLite**>(elements() + current_size_);
   auto src = reinterpret_cast<MessageLite* const*>(from.elements());
   int count = std::min(ClearedCount(), from.current_size_);
@@ -173,6 +175,7 @@ int RepeatedPtrFieldBase::MergeIntoClearedMessages(
 
 void RepeatedPtrFieldBase::MergeFromConcreteMessage(
     const RepeatedPtrFieldBase& from, CopyFn copy_fn) {
+  Prefetch5LinesFrom1Line(&from);
   ABSL_DCHECK_NE(&from, this);
   int new_size = current_size_ + from.current_size_;
   void** dst = InternalReserve(new_size);
@@ -196,6 +199,7 @@ void RepeatedPtrFieldBase::MergeFromConcreteMessage(
 template <>
 void RepeatedPtrFieldBase::MergeFrom<MessageLite>(
     const RepeatedPtrFieldBase& from) {
+  Prefetch5LinesFrom1Line(&from);
   ABSL_DCHECK_NE(&from, this);
   ABSL_DCHECK(from.current_size_ > 0);
   int new_size = current_size_ + from.current_size_;