Improve PNG decode performance: avoid branching in the row write loop

Source/core/platform/image-decoders/ImageDecoder.h

Index: Source/core/platform/image-decoders/ImageDecoder.h
diff --git a/Source/core/platform/image-decoders/ImageDecoder.h b/Source/core/platform/image-decoders/ImageDecoder.h
index 19587f5c2ea485f879fcd25f23fc8c861c752af5..8952e42da5c6d050787f3a763d822677217ee0bf 100644
--- a/Source/core/platform/image-decoders/ImageDecoder.h
+++ b/Source/core/platform/image-decoders/ImageDecoder.h
@@ -174,6 +174,23 @@ namespace WebCore {
             *dest = SkPackARGB32NoCheck(a, r, g, b);
         }
 
+        inline void setRGBAPremultiply(PixelData* dest, unsigned r, unsigned g, unsigned b, unsigned a)

[Peter Kasting, 2013/07/10 18:24:47]

I think we ought to be able to define setRGBA() in terms of these next two
functions.

I'm aware that we're trying to avoid the GIF decoder regression that came from
doing this sort of thing previously.  I think the main problem with that patch
was that its version of setRGBAPremultiply() moved the "if (!a)" clause outside
the "if (a < 255)" clause, resulting in the common case doing two comparisons
instead of one.  Since you haven't done that here, I think this patch ought to
be more performant than that one.  Can we try it?

[Noel Gordon, 2013/07/11 08:23:40]

Seems perfectly reasonable ...
In https://chromiumcodereview.appspot.com/17448010#msg12 I mentioned the exact
same thing to b.keleman post-facto in review after his patch landed, viz., why
was the if (!a) test moved?  He prepared a change to put it back to where it
was, and we landed that:

https://src.chromium.org/viewvc/blink?revision=153264&view=revision

but no joy: it did not fix the perf issue as noted in
https://code.google.com/p/chromium/issues/detail?id=254462#c9 
so I believe we'd see the perf regression if we went this way again.

+        {
+            if (a < 255) {
+                if (!a) {
+                    *dest = 0;
+                    return;
+                }
+
+                unsigned alpha = a * div255;
+                r = (r * alpha) >> 24;
+                g = (g * alpha) >> 24;
+                b = (b * alpha) >> 24;
+            }
+
+            *dest = SkPackARGB32NoCheck(a, r, g, b);
+        }
+
         inline void setRGBARaw(PixelData* dest, unsigned r, unsigned g, unsigned b, unsigned a)
         {
             *dest = SkPackARGB32NoCheck(a, r, g, b);