QUIC - add QuicPrefetchT0 to platform.

Profiled on android arm64, seems to speed up
net::QuicUtils::CopyToBuffer() by about 10%, translating to about 0.1%
overall.

Merge internal change: 177612619

R=rch@chromium.org

Bug: 
Change-Id: I7b85a4c91538a73af58c3fdbe32637e51693bd4a
Reviewed-on: https://chromium-review.googlesource.com/804744Reviewed-by: Ryan Hamilton <rch@chromium.org>
Commit-Queue: Buck Krasic <ckrasic@chromium.org>
Cr-Commit-Position: refs/heads/master@{#521430}

net/BUILD.gn

@@ -1459,6 +1459,7 @@ component("net") {
       "quic/platform/api/quic_mutex.h",
       "quic/platform/api/quic_optional.h",
       "quic/platform/api/quic_pcc_sender.h",
+      "quic/platform/api/quic_prefetch.h",
       "quic/platform/api/quic_ptr_util.h",
       "quic/platform/api/quic_reconstruct_object.h",
       "quic/platform/api/quic_reference_counted.h",
@@ -1499,6 +1500,7 @@ component("net") {
       "quic/platform/impl/quic_mutex_impl.h",
       "quic/platform/impl/quic_optional_impl.h",
       "quic/platform/impl/quic_pcc_sender_impl.h",
+      "quic/platform/impl/quic_prefetch_impl.h",
       "quic/platform/impl/quic_ptr_util_impl.h",
       "quic/platform/impl/quic_reconstruct_object_impl.h",
       "quic/platform/impl/quic_reference_counted_impl.h",

net/quic/core/quic_utils.cc

@@ -11,8 +11,10 @@
 #include "base/containers/adapters.h"
 #include "base/logging.h"
 #include "net/quic/core/quic_constants.h"
+#include "net/quic/platform/api/quic_aligned.h"
 #include "net/quic/platform/api/quic_bug_tracker.h"
 #include "net/quic/platform/api/quic_flags.h"
+#include "net/quic/platform/api/quic_prefetch.h"
 
 using std::string;
 
@@ -237,16 +239,12 @@ void QuicUtils::CopyToBuffer(const struct iovec* iov,
   // generally, the iov_offset is not 0, input iov consists of 2K buffers and
   // the output buffer is ~1.4K.
   if (copy_len == iov_available && iovnum + 1 < iov_count) {
-    // TODO(ckrasic) - this is unused without prefetch()
-    // char* next_base = static_cast<char*>(iov.iov[iovnum + 1].iov_base);
-    // char* next_base = static_cast<char*>(iov.iov[iovnum + 1].iov_base);
     // Prefetch 2 cachelines worth of data to get the prefetcher started; leave
     // it to the hardware prefetcher after that.
-    // TODO(ckrasic) - investigate what to do about prefetch directives.
-    // ::base::PrefetchT0(next_base);
+    char* next_base = static_cast<char*>(iov[iovnum + 1].iov_base);
+    QuicPrefetchT0(next_base);
     if (iov[iovnum + 1].iov_len >= 64) {
-      // TODO(ckrasic) - investigate what to do about prefetch directives.
-      // ::base::PrefetchT0(next_base + ABSL_CACHELINE_SIZE);
+      QuicPrefetchT0(next_base + QUIC_CACHELINE_SIZE);
     }
   }
 

net/quic/platform/api/quic_aligned.h

@@ -10,5 +10,6 @@
 #define QUIC_ALIGN_OF QUIC_ALIGN_OF_IMPL
 #define QUIC_ALIGNED(X) QUIC_ALIGNED_IMPL(X)
 #define QUIC_CACHELINE_ALIGNED QUIC_CACHELINE_ALIGNED_IMPL
+#define QUIC_CACHELINE_SIZE QUIC_CACHELINE_SIZE_IMPL
 
 #endif  // NET_QUIC_PLATFORM_API_QUIC_ALIGNED_H_

net/quic/platform/api/quic_prefetch.h

+// Copyright 2017 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef NET_QUIC_PLATFORM_API_QUIC_PREFETCH_H_
+#define NET_QUIC_PLATFORM_API_QUIC_PREFETCH_H_
+
+#include "net/quic/platform/impl/quic_prefetch_impl.h"
+
+namespace net {
+
+// Move data into the cache before it is read, or "prefetch" it.
+//
+// The value of `addr` is the address of the memory to prefetch. If
+// the target and compiler support it, data prefetch instructions are
+// generated. If the prefetch is done some time before the memory is
+// read, it may be in the cache by the time the read occurs.
+//
+// The function names specify the temporal locality heuristic applied,
+// using the names of Intel prefetch instructions:
+//
+//   T0 - high degree of temporal locality; data should be left in as
+//        many levels of the cache possible
+//   T1 - moderate degree of temporal locality
+//   T2 - low degree of temporal locality
+//   Nta - no temporal locality, data need not be left in the cache
+//         after the read
+//
+// Incorrect or gratuitous use of these functions can degrade
+// performance, so use them only when representative benchmarks show
+// an improvement.
+
+inline void QuicPrefetchT0(const void* addr) {
+  return QuicPrefetchT0Impl(addr);
+}
+
+}  // namespace net
+
+#endif  // NET_QUIC_PLATFORM_API_QUIC_PREFETCH_H_

net/quic/platform/impl/quic_aligned_impl.h

@@ -17,6 +17,7 @@
 
 // TODO(rtenneti): Change the default 64 alignas value (used the default
 // value from ABSL_CACHELINE_SIZE).
-#define QUIC_CACHELINE_ALIGNED_IMPL ALIGNAS(64)
+#define QUIC_CACHELINE_SIZE_IMPL (64)
+#define QUIC_CACHELINE_ALIGNED_IMPL ALIGNAS(QUIC_CACHELINE_SIZE)
 
 #endif  // NET_QUIC_PLATFORM_IMPL_QUIC_ALIGNED_IMPL_H_

net/quic/platform/impl/quic_prefetch_impl.h

+// Copyright 2017 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef NET_QUIC_PLATFORM_IMPL_QUIC_PREFETCH_IMPL_H_
+#define NET_QUIC_PLATFORM_IMPL_QUIC_PREFETCH_IMPL_H_
+
+#if defined(_MSC_VER)
+#include <intrin.h>
+#endif
+
+namespace net {
+
+inline void QuicPrefetchT0Impl(const void* addr) {
+#if defined(__GNUC__)
+  __builtin_prefetch(addr, 0, 3);
+#elif defined(_MSC_VER)
+  _mm_prefetch(reinterpret_cast<const char*>(addr), _MM_HINT_T0);
+#endif
+}
+
+}  // namespace net
+
+#endif  // NET_QUIC_PLATFORM_IMPL_QUIC_PREFETCH_IMPL_H_