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

Source/platform/image-decoders/ImageFrame.h

Index: Source/platform/image-decoders/ImageFrame.h
diff --git a/Source/platform/image-decoders/ImageFrame.h b/Source/platform/image-decoders/ImageFrame.h
index bb665a9140c1f5e3d180b55379e4106f36f9acbb..0246efe21c91994b4f602870a0a4cad4cb531e08 100644
--- a/Source/platform/image-decoders/ImageFrame.h
+++ b/Source/platform/image-decoders/ImageFrame.h
@@ -162,11 +162,20 @@ public:
         setRGBA(getAddr(x, y), r, g, b, a);
     }
 
+    inline void setRGBA(PixelData* dest, unsigned r, unsigned g, unsigned b, unsigned a)
+    {
+        if (m_premultiplyAlpha)

[Peter Kasting, 2014/05/12 19:36:11]

We could simplify this to just:

    setRGBA<m_premultiplyAlpha>(dest, r, g, b, a);

[Noel Gordon, 2014/05/13 01:39:19]

Doesn't compile on any C++ compiler I know. VS2010, VS2013 complain with
"Template parameter must be a compile-time constant expression".

[Peter Kasting, 2014/05/13 01:54:56]

Huh, that's odd.  I could swear I've done precisely this before.  Oh well.  I
guess this should just manually call the two forms of setRGBA(), then.  (See
below)

+            setRGBA<true>(dest, r, g, b, a);
+        else
+            *dest = SkPackARGB32NoCheck(a, r, g, b);
+    }
+
     static const unsigned div255 = static_cast<unsigned>(1.0 / 255 * (1 << 24)) + 1;
 
+    template <bool Premultiply>
     inline void setRGBA(PixelData* dest, unsigned r, unsigned g, unsigned b, unsigned a)

[Peter Kasting, 2014/05/12 19:36:11]

This can be static, can't it?

[Noel Gordon, 2014/05/13 01:39:19]

Could be, but you'd then add function calling overhead to every pixel.

[Peter Kasting, 2014/05/13 01:54:56]

Why?  Can't something be both static and inline?  All marking it in static
should do is make it clear to the reader that we don't touch anything in |this|.

     {
-        if (m_premultiplyAlpha && a < 255) {
+        if (Premultiply && a < 255) {
             if (!a) {
                 *dest = 0;
                 return;