Minor sk_memset{16|32}_SSE2 optimization.

Using explicitly indexed references allows some compilers to generate more efficient loops. For gcc 4.6.3:

613c18: 83 ea 10 sub $0x10,%edx
613c1b: 66 0f 7f 07 movdqa %xmm0,(%rdi)
613c1f: 66 0f 7f 47 10 movdqa %xmm0,0x10(%rdi)
613c24: 66 0f 7f 47 20 movdqa %xmm0,0x20(%rdi)
613c29: 66 0f 7f 47 30 movdqa %xmm0,0x30(%rdi)
613c2e: 48 83 c7 40 add $0x40,%rdi
613c32: 83 fa 0f cmp $0xf,%edx
613c35: 7f e1 jg 613c18 <_Z16sk_memset32_SSE2Pjji+0x38>

vs. previous:

613c18: 83 ea 10 sub $0x10,%edx
613c1b: 66 0f 7f 07 movdqa %xmm0,(%rdi)
613c1f: 66 0f 7f 47 10 movdqa %xmm0,0x10(%rdi)
613c24: 66 0f 7f 47 20 movdqa %xmm0,0x20(%rdi)
613c29: 48 83 c7 40 add $0x40,%rdi
613c2d: 83 fa 0f cmp $0xf,%edx
613c30: 66 0f 7f 47 f0 movdqa %xmm0,-0x10(%rdi)
613c35: 7f e1 jg 613c18 <_Z16sk_memset32_SSE2Pjji+0x38>

This yields a 0.2% - 1% improvement with the memset micro benchmarks, presumably due to avoiding a stall on the next store after the %rdi increment.

R=reed@google.com, senorblanco@chromium.org

Author: fmalita@chromium.org

Review URL: https://chromiumcodereview.appspot.com/21703003

git-svn-id: http://skia.googlecode.com/svn/trunk@10545 2bbb7eff-a529-9590-31e7-b0007b416f81

diff --git a/src/opts/SkUtils_opts_SSE2.cpp b/src/opts/SkUtils_opts_SSE2.cpp
index 08e4f66..e22044d 100644 (file)
--- a/src/opts/SkUtils_opts_SSE2.cpp
+++ b/src/opts/SkUtils_opts_SSE2.cpp
@@ -25,10 +25,11 @@ void sk_memset16_SSE2(uint16_t *dst, uint16_t value, int count)
         __m128i *d = reinterpret_cast<__m128i*>(dst);
         __m128i value_wide = _mm_set1_epi16(value);
         while (count >= 32) {
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
+            _mm_store_si128(d    , value_wide);
+            _mm_store_si128(d + 1, value_wide);
+            _mm_store_si128(d + 2, value_wide);
+            _mm_store_si128(d + 3, value_wide);
+            d += 4;
             count -= 32;
         }
         dst = reinterpret_cast<uint16_t*>(d);
@@ -54,10 +55,11 @@ void sk_memset32_SSE2(uint32_t *dst, uint32_t value, int count)
         __m128i *d = reinterpret_cast<__m128i*>(dst);
         __m128i value_wide = _mm_set1_epi32(value);
         while (count >= 16) {
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
-            _mm_store_si128(d++, value_wide);
+            _mm_store_si128(d    , value_wide);
+            _mm_store_si128(d + 1, value_wide);
+            _mm_store_si128(d + 2, value_wide);
+            _mm_store_si128(d + 3, value_wide);
+            d += 4;
             count -= 16;
         }
         dst = reinterpret_cast<uint32_t*>(d);