update memset16/32 inlining heuristics

I spent some time looking at perf.skia.org and it looks like we can do better.

It is weird, weird, weird that on x86, we see three completely different behaviors:
  - x86 Android: inlining better for small N, custom better for large N;
  - Windows:     inlining better for large N, custom better for small N;
  - other x86:   inlining generally better

BUG=skia:4316,chromium:516426

Committed: https://skia.googlesource.com/skia/+/b68fa409fc00ce2f38e2a0fd6f9dc2379b372481

Summaries: https://perf.skia.org/#4179
All traces, log scale: https://perf.skia.org/#4180

TBR=reed@google.com
No public API changes.

Review URL: https://codereview.chromium.org/1357193002

1 files changed, 23 insertions, 15 deletions

diff --git a/include/core/SkUtils.h b/include/core/SkUtils.h
index 4e24bd0883..2b390f04e5 100644
--- a/include/core/SkUtils.h
+++ b/include/core/SkUtils.h
@@ -17,13 +17,12 @@ namespace SkOpts {
 
 ///////////////////////////////////////////////////////////////////////////////
 
-// The inlining heuristics below were determined using bench/MemsetBench.cpp
-// on a x86 desktop, a Nexus 7 with and without NEON, and a Nexus 9:
-//   - on x86, inlining was never faster,
-//   - on ARMv7, inlining was faster for N<=10.  Putting this check inside the NEON
-//     code was not helpful; it's got to be here outside.
-//   - NEON code generation for ARMv8 with GCC 4.9 is terrible,
-//     making the NEON code ~8x slower that just a serial loop.
+// Inlining heuristics were determined by using perf.skia.org and bench/MemsetBench.cpp.
+// When using MSVC, inline is better >= 1K and worse <= 100.  The Nexus Player was the opposite.
+// Otherwise, when NEON or SSE is available to GCC or Clang, they can handle it best.
+// See https://code.google.com/p/chromium/issues/detail?id=516426#c15 for more details.
+// See also skia:4316; it might be a good idea to use rep stosw/stosd here.
+#define INLINE_IF(cond) if (cond) { while (count --> 0) { *buffer++ = value; } return; }
 
 /** Similar to memset(), but it assigns a 16bit value into the buffer.
     @param buffer   The memory to have value copied into it
@@ -31,10 +30,14 @@ namespace SkOpts {
     @param count    The number of times value should be copied into the buffer.
 */
 static inline void sk_memset16(uint16_t buffer[], uint16_t value, int count) {
-#if defined(SK_CPU_ARM64)
-    while (count --> 0) { *buffer++ = value; } return;
-#elif defined(SK_CPU_ARM32)
-    if (count <= 10) { while (count --> 0) { *buffer++ = value; } return; }
+#if defined(_MSC_VER)
+    INLINE_IF(count > 300)
+#elif defined(SK_BUILD_FOR_ANDROID) && defined(SK_CPU_X86)
+    INLINE_IF(count < 300)
+#elif defined(SK_ARM_HAS_NEON) || SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    INLINE_IF(true)
+#else
+    INLINE_IF(count <= 10)
 #endif
     SkOpts::memset16(buffer, value, count);
 }
@@ -45,14 +48,19 @@ static inline void sk_memset16(uint16_t buffer[], uint16_t value, int count) {
     @param count    The number of times value should be copied into the buffer.
 */
 static inline void sk_memset32(uint32_t buffer[], uint32_t value, int count) {
-#if defined(SK_CPU_ARM64)
-    while (count --> 0) { *buffer++ = value; } return;
-#elif defined(SK_CPU_ARM32)
-    if (count <= 10) { while (count --> 0) { *buffer++ = value; } return; }
+#if defined(_MSC_VER)
+    INLINE_IF(count > 300)
+#elif defined(SK_BUILD_FOR_ANDROID) && defined(SK_CPU_X86)
+    INLINE_IF(count < 300)
+#elif defined(SK_ARM_HAS_NEON) || SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    INLINE_IF(true)
+#else
+    INLINE_IF(count <= 10)
 #endif
     SkOpts::memset32(buffer, value, count);
 }
 
+#undef INLINE_IF
 
 ///////////////////////////////////////////////////////////////////////////////