Xfermode: SSE2 implementation of multiply_modeproc

src/opts/SkXfermode_opts_SSE2.h

+#ifndef SkXfermode_opts_SSE2_DEFINED
+#define SkXfermode_opts_SSE2_DEFINED
+
+#include "SkXfermode_proccoeff.h"
+
+class SkSSE2ProcCoeffXfermode : public SkProcCoeffXfermode {
+public:
+    SkSSE2ProcCoeffXfermode(const ProcCoeff& rec, SkXfermode::Mode mode,
+                            void* procSIMD)
+            : INHERITED(rec, mode), fProcSIMD(procSIMD) {}
+
+    virtual void xfer32(SkPMColor dst[], const SkPMColor src[], int count,
+                        const SkAlpha aa[]) const SK_OVERRIDE;
+    virtual void xfer16(uint16_t* SK_RESTRICT dst, const SkPMColor* SK_RESTRICT src,
+                        int count, const SkAlpha* SK_RESTRICT aa) const SK_OVERRIDE;
+
+    SK_DEVELOPER_TO_STRING()
+    SK_DECLARE_PUBLIC_FLATTENABLE_DESERIALIZATION_PROCS(SkSSE2ProcCoeffXfermode)
+
+private:
+    SkSSE2ProcCoeffXfermode(SkReadBuffer& buffer);
+
+    void* fProcSIMD;

[mtklein, 2014/03/27 14:55:17]

I'd be much much happier if we could store this as SkXfermodeProcSIMD fProcSIMD,
or at least as __m128i (*fProcSIMD)(__m128i src, __m128i dst).  Is there
something technical preventing us from doing this?

Skia does sometimes resort to working with void*, but it's almost always as
const void*, implicitly meaning a byte array.  If we're forced to do anything
else with void*, I'd want to document it heavily.

[qiankun, 2014/03/28 07:52:57]

I referred the arm NEON code previously. Now, I avoid using void* and use
SkXfermodeProcSIMD instead. It works well.

+    typedef SkProcCoeffXfermode INHERITED;
+};
+
+#endif // SkXfermode_opts_SSE2_DEFINED