Properly enable S32_D16_filter_DX_SSE2 optimization.

src/opts/opts_check_SSE2.cpp

Index: src/opts/opts_check_SSE2.cpp
diff --git a/src/opts/opts_check_SSE2.cpp b/src/opts/opts_check_SSE2.cpp
index e7677220cb9326480e4fea631e2249398547e1a9..2bd19851b7d5a69a08f32615adb303f790b31f9f 100644
--- a/src/opts/opts_check_SSE2.cpp
+++ b/src/opts/opts_check_SSE2.cpp
@@ -123,46 +123,54 @@ void SkBitmapProcState::platformConvolutionProcs(SkConvolutionProcs* procs) {
 }
 
 void SkBitmapProcState::platformProcs() {
-    if (cachedHasSSSE3()) {
-        if (fSampleProc32 == S32_opaque_D32_filter_DX) {
+    /* Every optimization in the function requires at least SSE2 */
+    if (!cachedHasSSE2()) {
+        return;
+    }
+
+    /* Check fSampleProc32 */
+    if (fSampleProc32 == S32_opaque_D32_filter_DX) {            // Check S32_opaque_D32_filter_DX

[tomhudson, 2014/04/22 09:47:28]

Nit: '/* Check fSampleProc32 */' and similar block-heading comments might be
useful, but these line-ending comments strike me as useless noise? The point of
the comment shouldn't be *what* the code does if that's obvious by inspecition
(which I believe it is here), but whether *why* the code does something.

I'd like to increase the comment density in Skia, but not at the expense of
comment quality. Other reviewers?

[henrik.smiding, 2014/04/22 12:18:58]

I'll remove the four end-of-line comments. I only added them to make it slightly
easier to find the correct location to put future optimizations.

[henrik.smiding, 2014/04/23 13:12:50]

Done.

+        if (cachedHasSSSE3()) {
             fSampleProc32 = S32_opaque_D32_filter_DX_SSSE3;
-        } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
-            fSampleProc32 = S32_alpha_D32_filter_DX_SSSE3;
+        } else {
+            fSampleProc32 = S32_opaque_D32_filter_DX_SSE2;
         }
-
-        if (fSampleProc32 == S32_opaque_D32_filter_DXDY) {
+    } else if (fSampleProc32 == S32_opaque_D32_filter_DXDY) {   // Check S32_opaque_D32_filter_DXDY
+        if (cachedHasSSSE3()) {
             fSampleProc32 = S32_opaque_D32_filter_DXDY_SSSE3;
-        } else if (fSampleProc32 == S32_alpha_D32_filter_DXDY) {
-            fSampleProc32 = S32_alpha_D32_filter_DXDY_SSSE3;
         }
-    } else if (cachedHasSSE2()) {
-        if (fSampleProc32 == S32_opaque_D32_filter_DX) {
-            fSampleProc32 = S32_opaque_D32_filter_DX_SSE2;
-        } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {
+    } else if (fSampleProc32 == S32_alpha_D32_filter_DX) {      // Check S32_alpha_D32_filter_DX
+        if (cachedHasSSSE3()) {
+            fSampleProc32 = S32_alpha_D32_filter_DX_SSSE3;
+        } else {
             fSampleProc32 = S32_alpha_D32_filter_DX_SSE2;
         }
-
-        if (fSampleProc16 == S32_D16_filter_DX) {
-            fSampleProc16 = S32_D16_filter_DX_SSE2;
+    } else if (fSampleProc32 == S32_alpha_D32_filter_DXDY) {    // Check S32_alpha_D32_filter_DXDY
+        if (cachedHasSSSE3()) {
+            fSampleProc32 = S32_alpha_D32_filter_DXDY_SSSE3;
         }
     }
 
-    if (cachedHasSSSE3() || cachedHasSSE2()) {
-        if (fMatrixProc == ClampX_ClampY_filter_scale) {
-            fMatrixProc = ClampX_ClampY_filter_scale_SSE2;
-        } else if (fMatrixProc == ClampX_ClampY_nofilter_scale) {
-            fMatrixProc = ClampX_ClampY_nofilter_scale_SSE2;
-        }
+    /* Check fSampleProc16 */
+    if (fSampleProc16 == S32_D16_filter_DX) {
+        fSampleProc16 = S32_D16_filter_DX_SSE2;
+    }
 
-        if (fMatrixProc == ClampX_ClampY_filter_affine) {
-            fMatrixProc = ClampX_ClampY_filter_affine_SSE2;
-        } else if (fMatrixProc == ClampX_ClampY_nofilter_affine) {
-            fMatrixProc = ClampX_ClampY_nofilter_affine_SSE2;
-        }
-        if (c_hqfilter_sse) {
-            if (fShaderProc32 == highQualityFilter32) {
-                fShaderProc32 = highQualityFilter_SSE2;
-            }
+    /* Check fMatrixProc */
+    if (fMatrixProc == ClampX_ClampY_filter_scale) {
+        fMatrixProc = ClampX_ClampY_filter_scale_SSE2;
+    } else if (fMatrixProc == ClampX_ClampY_nofilter_scale) {
+        fMatrixProc = ClampX_ClampY_nofilter_scale_SSE2;
+    } else if (fMatrixProc == ClampX_ClampY_filter_affine) {
+        fMatrixProc = ClampX_ClampY_filter_affine_SSE2;
+    } else if (fMatrixProc == ClampX_ClampY_nofilter_affine) {
+        fMatrixProc = ClampX_ClampY_nofilter_affine_SSE2;
+    }
+
+    /* Check fShaderProc32 */
+    if (c_hqfilter_sse) {
+        if (fShaderProc32 == highQualityFilter32) {
+            fShaderProc32 = highQualityFilter_SSE2;
         }
     }
 }