Prefetch destination cache lines in ExtendWithPclmul temporal copy mode.
We can only prefetch on x86 and in temporal mode because both in non-temporal mode and on aarch64, writes to full cachelines are streamed to lower level caches or DRAM, and prefetching them actually causes more installations and hurts performance.

Also reduce the number of copy streams to 2, which helps performance slightly overall.

PiperOrigin-RevId: 1010616216

absl/crc/internal/crc_x86_arm_combined.cc

diff --git a/absl/crc/internal/crc_x86_arm_combined.cc b/absl/crc/internal/crc_x86_arm_combined.cc
index 2d68ceb..0d1dfdc 100644
--- a/absl/crc/internal/crc_x86_arm_combined.cc
+++ b/absl/crc/internal/crc_x86_arm_combined.cc
@@ -516,7 +516,15 @@
   // are combined in the end.
   static constexpr size_t kMaxStreams = 3;
   // When copying, at most kMaxCopyStreams streams are used in total.
-  static constexpr size_t kMaxCopyStreams = 3;
+  static constexpr size_t kMaxCopyStreams = 2;
+  // On AArch64, full-cache-line stores engage hardware write-streaming, and
+  // software destination prefetches force unnecessary DRAM reads.
+#if defined(__aarch64__)
+  static constexpr bool kShouldPrefetchDst = false;
+#else
+  static constexpr bool kShouldPrefetchDst = true;
+#endif
+  static constexpr size_t kDstPrefetchHorizon = 2 * ABSL_CACHELINE_SIZE;
 };
 
 template <size_t num_crc_streams, size_t num_pclmul_streams,
@@ -789,9 +797,15 @@
       // Prefetch data for next iterations.
       for (size_t j = 0; j < kNumActiveCrcStreams; ++j) {
         PrefetchToLocalCache(crc_streams[j] + kPrefetchHorizon);
+        if constexpr (copy_type == CopyType::kTemporal && kShouldPrefetchDst) {
+          PrefetchToLocalCache(dst_crc_streams[j] + kDstPrefetchHorizon);
+        }
       }
       for (size_t j = 0; j < kNumActivePclmulStreams; ++j) {
         PrefetchToLocalCache(pclmul_streams[j] + kPrefetchHorizon);
+        if constexpr (copy_type == CopyType::kTemporal && kShouldPrefetchDst) {
+          PrefetchToLocalCache(dst_pclmul_streams[j] + kDstPrefetchHorizon);
+        }
       }
 
       const uint8_t* pclmul_copy[kMaxStreams];