refresh from skia, with blitrow_d32 bottlenecks for neon optimizations

12 files changed, 266 insertions, 147 deletions

diff --git a/Android.mk b/Android.mk
index de09fd4901..5d27c91fa1 100644
--- a/Android.mk
+++ b/Android.mk
@@ -90,6 +90,7 @@ LOCAL_SRC_FILES:= \
 	src/core/SkBitmapProcState_matrixProcs.cpp \
 	src/core/SkBitmapSampler.cpp \
 	src/core/SkBlitRow_D16.cpp \
+	src/core/SkBlitRow_D32.cpp \
 	src/core/SkBlitRow_D4444.cpp \
 	src/core/SkBlitter.cpp \
 	src/core/SkBlitter_4444.cpp \
diff --git a/include/views/SkOSWindow_Mac.h b/include/views/SkOSWindow_Mac.h
index 47411516bf..98e76b032c 100644
--- a/include/views/SkOSWindow_Mac.h
+++ b/include/views/SkOSWindow_Mac.h
@@ -30,8 +30,8 @@ public:
 
     static bool PostEvent(SkEvent* evt, SkEventSinkID, SkMSec delay);
 
-    static OSStatus SkOSWindow::EventHandler(EventHandlerCallRef inHandler,
-                                             EventRef inEvent, void* userData);
+    static OSStatus EventHandler(EventHandlerCallRef inHandler,
+                                 EventRef inEvent, void* userData);
 
     void    doPaint(void* ctx);
 
diff --git a/src/core/SkBlitRow.h b/src/core/SkBlitRow.h
index ec31ff5309..dbbd84d182 100644
--- a/src/core/SkBlitRow.h
+++ b/src/core/SkBlitRow.h
@@ -6,7 +6,7 @@
 
 class SkBlitRow {
 public:
-    enum {
+    enum Flags16 {
         //! If set, the alpha parameter will be != 255
         kGlobalAlpha_Flag   = 0x01,
         //! If set, the src colors may have alpha != 255
@@ -32,13 +32,51 @@ public:
     //! Public entry-point to return a blit function ptr
     static Proc Factory(unsigned flags, SkBitmap::Config);
 
+    ///////////// D32 version
+
+    enum Flags32 {
+        kGlobalAlpha_Flag32     = 1 << 0,
+        kSrcPixelAlpha_Flag32   = 1 << 1,
+    };
+
+    /** Function pointer that blends 32bit colors onto a 32bit destination.
+        @param dst  array of dst 32bit colors
+        @param src  array of src 32bit colors (w/ or w/o alpha)
+        @param count number of colors to blend
+        @param alpha global alpha to be applied to all src colors
+     */
+    typedef void (*Proc32)(uint32_t* SK_RESTRICT dst,
+                         const SkPMColor* SK_RESTRICT src,
+                         int count, U8CPU alpha);
+
+    static Proc32 Factory32(unsigned flags32);
+    
+    /** Blend a single color onto a row of S32 pixels, writing the result
+        into a row of D32 pixels. src and dst may be the same memory, but
+        if they are not, they may not overlap.
+     */
+    static void Color32(SkPMColor dst[], const SkPMColor src[], int count,
+                         SkPMColor color);
+
+    /** Blend a single color onto a row of 32bit pixels, writing the result
+        into the same row.
+     */
+    static void Color32(SkPMColor row[], int count, SkPMColor color) {
+        Color32(row, row, count, color);
+    }
+
 private:
+    enum {
+        kFlags16_Mask = 7,
+        kFlags32_Mask = 3
+    };
     /** These global arrays are indexed using the flags parameter to Factory,
         and contain either NULL, or a platform-specific function-ptr to be used
         in place of the system default.
      */
     static const Proc gPlatform_565_Procs[];
     static const Proc gPlatform_4444_Procs[];
+    static const Proc32 gPlatform_Procs32[];
 };
 
 #endif
diff --git a/src/core/SkBlitRow_D16.cpp b/src/core/SkBlitRow_D16.cpp
index 1860d3bf9a..66ac90e29a 100644
--- a/src/core/SkBlitRow_D16.cpp
+++ b/src/core/SkBlitRow_D16.cpp
@@ -235,6 +235,8 @@ extern SkBlitRow::Proc SkBlitRow_Factory_4444(unsigned flags);
     
 SkBlitRow::Proc SkBlitRow::Factory(unsigned flags, SkBitmap::Config config) {
     SkASSERT(flags < SK_ARRAY_COUNT(gDefault_565_Procs));
+    // just so we don't crash
+    flags &= kFlags16_Mask;
 
     SkBlitRow::Proc proc = NULL;
 
diff --git a/src/core/SkBlitRow_D32.cpp b/src/core/SkBlitRow_D32.cpp
new file mode 100644
index 0000000000..f67bb9abae
--- /dev/null
+++ b/src/core/SkBlitRow_D32.cpp
@@ -0,0 +1,112 @@
+#include "SkBlitRow.h"
+#include "SkColorPriv.h"
+#include "SkUtils.h"
+
+static void S32_Opaque_BlitRow32(SkPMColor* SK_RESTRICT dst,
+                                 const SkPMColor* SK_RESTRICT src,
+                                 int count, U8CPU alpha) {
+    SkASSERT(255 == alpha);
+    memcpy(dst, src, count * sizeof(SkPMColor));
+}
+
+static void S32_Blend_BlitRow32(SkPMColor* SK_RESTRICT dst,
+                                const SkPMColor* SK_RESTRICT src,
+                                int count, U8CPU alpha) {
+    SkASSERT(alpha <= 255);
+    if (count > 0) {
+        unsigned src_scale = SkAlpha255To256(alpha);
+        unsigned dst_scale = 256 - src_scale;
+        do {
+            *dst = SkAlphaMulQ(*src, src_scale) + SkAlphaMulQ(*dst, dst_scale);
+            src += 1;
+            dst += 1;
+        } while (--count > 0);
+    }
+}
+
+//#define TEST_SRC_ALPHA
+
+static void S32A_Opaque_BlitRow32(SkPMColor* SK_RESTRICT dst,
+                                  const SkPMColor* SK_RESTRICT src,
+                                  int count, U8CPU alpha) {
+    SkASSERT(255 == alpha);
+    if (count > 0) {
+        do {
+#ifdef TEST_SRC_ALPHA
+            SkPMColor sc = *src;
+            if (sc) {
+                unsigned srcA = SkGetPackedA32(sc);
+                SkPMColor result = sc;
+                if (srcA != 255) {
+                    result = SkPMSrcOver(sc, *dst);
+                }
+                *dst = result;
+            }
+#else
+            *dst = SkPMSrcOver(*src, *dst);
+#endif
+            src += 1;
+            dst += 1;
+        } while (--count > 0);
+    }
+}
+
+static void S32A_Blend_BlitRow32(SkPMColor* SK_RESTRICT dst,
+                                 const SkPMColor* SK_RESTRICT src,
+                                 int count, U8CPU alpha) {
+    SkASSERT(alpha <= 255);
+    if (count > 0) {
+        do {
+            *dst = SkBlendARGB32(*src, *dst, alpha);
+            src += 1;
+            dst += 1;
+        } while (--count > 0);
+    }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+
+static const SkBlitRow::Proc32 gDefault_Procs32[] = {
+    S32_Opaque_BlitRow32,
+    S32_Blend_BlitRow32,
+    S32A_Opaque_BlitRow32,
+    S32A_Blend_BlitRow32
+};
+
+SkBlitRow::Proc32 SkBlitRow::Factory32(unsigned flags) {
+    SkASSERT(flags < SK_ARRAY_COUNT(gDefault_Procs32));
+    // just so we don't crash
+    flags &= kFlags32_Mask;
+    
+    SkBlitRow::Proc32 proc = gPlatform_Procs32[flags];
+    if (NULL == proc) {
+        proc = gDefault_Procs32[flags];
+    }
+    SkASSERT(proc);
+    return proc;
+}
+
+void SkBlitRow::Color32(SkPMColor dst[], const SkPMColor src[], int count,
+                        SkPMColor color) {
+    if (count > 0) {
+        if (0 == color) {
+            if (src != dst) {
+                memcpy(dst, src, count * sizeof(SkPMColor));
+            }
+        }
+        unsigned colorA = SkGetPackedA32(color);
+        if (255 == colorA) {
+            sk_memset32(dst, color, count);
+        } else {
+            unsigned scale = 256 - SkAlpha255To256(colorA);
+            do {
+                *dst = color + SkAlphaMulQ(*src, scale);
+                src += 1;
+                dst += 1;
+            } while (--count);
+        }
+    }
+}
+
+
+
diff --git a/src/core/SkBlitter_ARGB32.cpp b/src/core/SkBlitter_ARGB32.cpp
index d3e5714e8f..37bd0da684 100644
--- a/src/core/SkBlitter_ARGB32.cpp
+++ b/src/core/SkBlitter_ARGB32.cpp
@@ -69,29 +69,7 @@ const SkBitmap* SkARGB32_Blitter::justAnOpaqueColor(uint32_t* value) {
 void SkARGB32_Blitter::blitH(int x, int y, int width) {
     SkASSERT(x >= 0 && y >= 0 && x + width <= fDevice.width());
 
-    if (fSrcA == 0) {
-        return;
-    }
-
-    uint32_t* device = fDevice.getAddr32(x, y);
-
-    if (fSrcA == 255) {
-        sk_memset32(device, fPMColor, width);
-    } else {
-        uint32_t color = fPMColor;
-        unsigned dst_scale = SkAlpha255To256(255 - fSrcA);
-        uint32_t prevDst = ~device[0];  // so we always fail the test the first time
-        uint32_t result SK_INIT_TO_AVOID_WARNING;
-
-        for (int i = 0; i < width; i++) {
-            uint32_t currDst = device[i];
-            if (currDst != prevDst) {
-                result = color + SkAlphaMulQ(currDst, dst_scale);
-                prevDst = currDst;
-            }
-            device[i] = result;
-        }
-    }
+    SkBlitRow::Color32(fDevice.getAddr32(x, y), width, fPMColor);
 }
 
 void SkARGB32_Blitter::blitAntiH(int x, int y, const SkAlpha antialias[],
@@ -115,13 +93,8 @@ void SkARGB32_Blitter::blitAntiH(int x, int y, const SkAlpha antialias[],
             if ((opaqueMask & aa) == 255) {
                 sk_memset32(device, color, count);
             } else {
-                uint32_t sc = SkAlphaMulQ(color, aa);
-                unsigned dst_scale = 255 - SkGetPackedA32(sc);
-                int n = count;
-                do {
-                    --n;
-                    device[n] = sc + SkAlphaMulQ(device[n], dst_scale);
-                } while (n > 0);
+                uint32_t sc = SkAlphaMulQ(color, SkAlpha255To256(aa));
+                SkBlitRow::Color32(device, count, sc);
             }
         }
         runs += count;
@@ -310,29 +283,11 @@ void SkARGB32_Blitter::blitRect(int x, int y, int width, int height) {
 
     uint32_t*   device = fDevice.getAddr32(x, y);
     uint32_t    color = fPMColor;
+    size_t      rowBytes = fDevice.rowBytes();
 
-    if (fSrcA == 255) {
-        while (--height >= 0) {
-            sk_memset32(device, color, width);
-            device = (uint32_t*)((char*)device + fDevice.rowBytes());
-        }
-    } else {
-        unsigned dst_scale = SkAlpha255To256(255 - fSrcA);
-
-        while (--height >= 0) {
-            uint32_t prevDst = ~device[0];
-            uint32_t result SK_INIT_TO_AVOID_WARNING;
-
-            for (int i = 0; i < width; i++) {
-                uint32_t dst = device[i];
-                if (dst != prevDst) {
-                    result = color + SkAlphaMulQ(dst, dst_scale);
-                    prevDst = dst;
-                }
-                device[i] = result;
-            }
-            device = (uint32_t*)((char*)device + fDevice.rowBytes());
-        }
+    while (--height >= 0) {
+        SkBlitRow::Color32(device, width, color);
+        device = (uint32_t*)((char*)device + rowBytes);
     }
 }
 
@@ -440,11 +395,20 @@ void SkARGB32_Black_Blitter::blitAntiH(int x, int y, const SkAlpha antialias[],
 //////////////////////////////////////////////////////////////////////////////////////////
 
 SkARGB32_Shader_Blitter::SkARGB32_Shader_Blitter(const SkBitmap& device,
-                                                 const SkPaint& paint)
-        : INHERITED(device, paint) {
+                            const SkPaint& paint) : INHERITED(device, paint) {
     fBuffer = (SkPMColor*)sk_malloc_throw(device.width() * (sizeof(SkPMColor)));
 
-    (fXfermode = paint.getXfermode())->safeRef();
+    fXfermode = paint.getXfermode();
+    SkSafeRef(fXfermode);
+
+    int flags = 0;
+    if (!(fShader->getFlags() & SkShader::kOpaqueAlpha_Flag)) {
+        flags |= SkBlitRow::kSrcPixelAlpha_Flag32;
+    }
+    // we call this on the output from the shader
+    fProc32 = SkBlitRow::Factory32(flags);
+    // we call this on the output from the shader + alpha from the aa buffer
+    fProc32Blend = SkBlitRow::Factory32(flags | SkBlitRow::kGlobalAlpha_Flag32);
 }
 
 SkARGB32_Shader_Blitter::~SkARGB32_Shader_Blitter() {
@@ -465,16 +429,7 @@ void SkARGB32_Shader_Blitter::blitH(int x, int y, int width) {
         if (fXfermode) {
             fXfermode->xfer32(device, span, width, NULL);
         } else {
-            for (int i = 0; i < width; i++) {
-                uint32_t src = span[i];
-                if (src) {
-                    unsigned srcA = SkGetPackedA32(src);
-                    if (srcA != 0xFF) {
-                        src += SkAlphaMulQ(device[i], SkAlpha255To256(255 - srcA));
-                    }
-                    device[i] = src;
-                }
-            }
+            fProc32(device, span, width, 255);
         }
     }
 }
@@ -519,15 +474,12 @@ void SkARGB32_Shader_Blitter::blitAntiH(int x, int y, const SkAlpha antialias[],
             }
             int aa = *antialias;
             if (aa) {
-                if (aa == 255) {  // cool, have the shader draw right into the device
+                if (aa == 255) {
+                    // cool, have the shader draw right into the device
                     shader->shadeSpan(x, y, device, count);
                 } else {
                     shader->shadeSpan(x, y, span, count);
-                    for (int i = count - 1; i >= 0; --i) {
-                        if (span[i]) {
-                            device[i] = SkBlendARGB32(span[i], device[i], aa);
-                        }
-                    }
+                    fProc32Blend(device, span, count, aa);
                 }
             }
             device += count;
@@ -535,7 +487,7 @@ void SkARGB32_Shader_Blitter::blitAntiH(int x, int y, const SkAlpha antialias[],
             antialias += count;
             x += count;
         } 
-    } else {    // no xfermode but we are not opaque
+    } else {    // no xfermode but the shader not opaque
         for (;;) {
             int count = *runs;
             if (count <= 0) {
@@ -545,17 +497,9 @@ void SkARGB32_Shader_Blitter::blitAntiH(int x, int y, const SkAlpha antialias[],
             if (aa) {
                 fShader->shadeSpan(x, y, span, count);
                 if (aa == 255) {
-                    for (int i = count - 1; i >= 0; --i) {
-                        if (span[i]) {
-                            device[i] = SkPMSrcOver(span[i], device[i]);
-                        }
-                    }
+                    fProc32(device, span, count, 255);
                 } else {
-                    for (int i = count - 1; i >= 0; --i) {
-                        if (span[i]) {
-                            device[i] = SkBlendARGB32(span[i], device[i], aa);
-                        }
-                    }
+                    fProc32Blend(device, span, count, aa);
                 }
             }
             device += count;
diff --git a/src/core/SkComposeShader.cpp b/src/core/SkComposeShader.cpp
index 1e5e202b8e..7e566a3d2e 100644
--- a/src/core/SkComposeShader.cpp
+++ b/src/core/SkComposeShader.cpp
@@ -124,6 +124,8 @@ void SkComposeShader::shadeSpan(int x, int y, SkPMColor result[], int count)
 
     if (NULL == mode)   // implied SRC_OVER
     {
+        // TODO: when we have a good test-case, should use SkBlitRow::Proc32
+        // for these loops
         do {
             int n = count;
             if (n > TMP_COLOR_COUNT)
diff --git a/src/core/SkCoreBlitters.h b/src/core/SkCoreBlitters.h
index 5e34685a08..84f8fa6524 100644
--- a/src/core/SkCoreBlitters.h
+++ b/src/core/SkCoreBlitters.h
@@ -141,8 +141,10 @@ public:
     virtual void blitAntiH(int x, int y, const SkAlpha antialias[], const int16_t runs[]);
 
 private:
-    SkXfermode* fXfermode;
-    SkPMColor*  fBuffer;
+    SkXfermode*         fXfermode;
+    SkPMColor*          fBuffer;
+    SkBlitRow::Proc32   fProc32;
+    SkBlitRow::Proc32   fProc32Blend;
 
     // illegal
     SkARGB32_Shader_Blitter& operator=(const SkARGB32_Shader_Blitter&);
diff --git a/src/core/SkSpriteBlitter_ARGB32.cpp b/src/core/SkSpriteBlitter_ARGB32.cpp
index 6addde7b0f..01ee9f1f69 100644
--- a/src/core/SkSpriteBlitter_ARGB32.cpp
+++ b/src/core/SkSpriteBlitter_ARGB32.cpp
@@ -16,64 +16,58 @@
 */
 
 #include "SkSpriteBlitter.h"
+#include "SkBlitRow.h"
+#include "SkColorFilter.h"
+#include "SkColorPriv.h"
 #include "SkTemplates.h"
 #include "SkUtils.h"
-#include "SkColorPriv.h"
-
-#define D32_S32A_Opaque_Pixel(dst, sc) \
-do {                                                                  \
-    if (sc) {                                                         \
-        unsigned srcA = SkGetPackedA32(sc);                           \
-        uint32_t result = sc;                                         \
-        if (srcA != 0xFF) {                                           \
-            result += SkAlphaMulQ(*dst, SkAlpha255To256(255 - srcA)); \
-        }                                                             \
-        *dst = result;                                                \
-    }                                                                 \
-} while (0)
-
-#define SkSPRITE_CLASSNAME                  Sprite_D32_S32A_Opaque
-#define SkSPRITE_ARGS
-#define SkSPRITE_FIELDS
-#define SkSPRITE_INIT
-#define SkSPRITE_DST_TYPE                   uint32_t
-#define SkSPRITE_SRC_TYPE                   uint32_t
-#define SkSPRITE_DST_GETADDR                getAddr32
-#define SkSPRITE_SRC_GETADDR                getAddr32
-#define SkSPRITE_PREAMBLE(srcBM, x, y)
-#define SkSPRITE_BLIT_PIXEL(dst, src)       D32_S32A_Opaque_Pixel(dst, src)
-#define SkSPRITE_NEXT_ROW
-#define SkSPRITE_POSTAMBLE(srcBM)
-#include "SkSpriteBlitterTemplate.h"
+#include "SkXfermode.h"
 
 ///////////////////////////////////////////////////////////////////////////////
 
-class Sprite_D32_S32_Opaque : public SkSpriteBlitter {
+class Sprite_D32_S32 : public SkSpriteBlitter {
 public:
-    Sprite_D32_S32_Opaque(const SkBitmap& source) : SkSpriteBlitter(source) {}
+    Sprite_D32_S32(const SkBitmap& src, U8CPU alpha)  : INHERITED(src) {
+        SkASSERT(src.config() == SkBitmap::kARGB_8888_Config);
+
+        unsigned flags32 = 0;
+        if (255 != alpha) {
+            flags32 |= SkBlitRow::kGlobalAlpha_Flag32;
+        }
+        if (!src.isOpaque()) {
+            flags32 |= SkBlitRow::kSrcPixelAlpha_Flag32;
+        }
 
+        fProc32 = SkBlitRow::Factory32(flags32);
+        fAlpha = alpha;
+    }
+    
     virtual void blitRect(int x, int y, int width, int height) {
         SkASSERT(width > 0 && height > 0);
         SK_RESTRICT uint32_t* dst = fDevice->getAddr32(x, y);
         const SK_RESTRICT uint32_t* src = fSource->getAddr32(x - fLeft,
                                                              y - fTop);
-        unsigned dstRB = fDevice->rowBytes();
-        unsigned srcRB = fSource->rowBytes();
-        size_t size = width * sizeof(uint32_t);
+        size_t dstRB = fDevice->rowBytes();
+        size_t srcRB = fSource->rowBytes();
+        SkBlitRow::Proc32 proc = fProc32;
+        U8CPU             alpha = fAlpha;
 
         do {
-            memcpy(dst, src, size);
+            proc(dst, src, width, alpha);
             dst = (SK_RESTRICT uint32_t*)((char*)dst + dstRB);
             src = (const SK_RESTRICT uint32_t*)((const char*)src + srcRB);
         } while (--height != 0);
     }
+
+private:
+    SkBlitRow::Proc32   fProc32;
+    U8CPU               fAlpha;
+    
+    typedef SkSpriteBlitter INHERITED;
 };
 
 ///////////////////////////////////////////////////////////////////////////////
 
-#include "SkColorFilter.h"
-#include "SkXfermode.h"
-
 class Sprite_D32_XferFilter : public SkSpriteBlitter {
 public:
     Sprite_D32_XferFilter(const SkBitmap& source, const SkPaint& paint)
@@ -86,6 +80,17 @@ public:
         
         fBufferSize = 0;
         fBuffer = NULL;
+
+        unsigned flags32 = 0;
+        if (255 != paint.getAlpha()) {
+            flags32 |= SkBlitRow::kGlobalAlpha_Flag32;
+        }
+        if (!source.isOpaque()) {
+            flags32 |= SkBlitRow::kSrcPixelAlpha_Flag32;
+        }
+        
+        fProc32 = SkBlitRow::Factory32(flags32);
+        fAlpha = paint.getAlpha();
     }
     
     virtual ~Sprite_D32_XferFilter() {
@@ -107,11 +112,13 @@ public:
     }
 
 protected:
-    SkColorFilter*  fColorFilter;
-    SkXfermode*     fXfermode;
-    int             fBufferSize;
-    SkPMColor*      fBuffer;
-    
+    SkColorFilter*      fColorFilter;
+    SkXfermode*         fXfermode;
+    int                 fBufferSize;
+    SkPMColor*          fBuffer;
+    SkBlitRow::Proc32   fProc32;
+    U8CPU               fAlpha;
+
 private:
     typedef SkSpriteBlitter INHERITED;
 };
@@ -144,9 +151,7 @@ public:
             if (NULL != xfermode) {
                 xfermode->xfer32(dst, tmp, width, NULL);
             } else {
-                for (int i = 0; i < width; i++) {
-                    dst[i] = SkPMSrcOver(tmp[i], dst[i]);
-                }
+                fProc32(dst, tmp, width, fAlpha);
             }
 
             dst = (SK_RESTRICT uint32_t*)((char*)dst + dstRB);
@@ -192,9 +197,7 @@ public:
             if (NULL != xfermode) {
                 xfermode->xfer32(dst, buffer, width, NULL);
             } else {
-                for (int i = 0; i < width; i++) {
-                    dst[i] = SkPMSrcOver(buffer[i], dst[i]);
-                }
+                fProc32(dst, buffer, width, fAlpha);
             }
             
             dst = (SK_RESTRICT SkPMColor*)((char*)dst + dstRB);
@@ -273,16 +276,20 @@ public:
 SkSpriteBlitter* SkSpriteBlitter::ChooseD32(const SkBitmap& source,
                                             const SkPaint& paint,
                                             void* storage, size_t storageSize) {
-    if (paint.getMaskFilter() != NULL || paint.getAlpha() != 0xFF) {
+    if (paint.getMaskFilter() != NULL) {
         return NULL;
     }
 
+    U8CPU       alpha = paint.getAlpha();
     SkXfermode* xfermode = paint.getXfermode();
     SkColorFilter* filter = paint.getColorFilter();
     SkSpriteBlitter* blitter = NULL;
 
     switch (source.getConfig()) {
         case SkBitmap::kARGB_4444_Config:
+            if (alpha != 0xFF) {
+                return NULL;    // we only have opaque sprites
+            }
             if (xfermode || filter) {
                 SK_PLACEMENT_NEW_ARGS(blitter, Sprite_D32_S4444_XferFilter,
                                       storage, storageSize, (source, paint));
@@ -296,14 +303,15 @@ SkSpriteBlitter* SkSpriteBlitter::ChooseD32(const SkBitmap& source,
             break;
         case SkBitmap::kARGB_8888_Config:
             if (xfermode || filter) {
-                SK_PLACEMENT_NEW_ARGS(blitter, Sprite_D32_S32A_XferFilter,
+                if (255 == alpha) {
+                    // this can handle xfermode or filter, but not alpha
+                    SK_PLACEMENT_NEW_ARGS(blitter, Sprite_D32_S32A_XferFilter,
                                       storage, storageSize, (source, paint));
-            } else if (source.isOpaque()) {
-                SK_PLACEMENT_NEW_ARGS(blitter, Sprite_D32_S32_Opaque,
-                                      storage, storageSize, (source));
+                }
             } else {
-                SK_PLACEMENT_NEW_ARGS(blitter, Sprite_D32_S32A_Opaque,
-                                      storage, storageSize, (source));
+                // this can handle alpha, but not xfermode or filter
+                SK_PLACEMENT_NEW_ARGS(blitter, Sprite_D32_S32,
+                              storage, storageSize, (source, alpha));
             }
             break;
         default:
diff --git a/src/effects/SkColorFilters.cpp b/src/effects/SkColorFilters.cpp
index 0bb3ae595a..733e1aea8e 100644
--- a/src/effects/SkColorFilters.cpp
+++ b/src/effects/SkColorFilters.cpp
@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+#include "SkBlitRow.h"
 #include "SkColorFilter.h"
 #include "SkColorPriv.h"
 #include "SkUtils.h"
@@ -86,12 +87,7 @@ public:
     
     virtual void filterSpan(const SkPMColor shader[], int count,
                             SkPMColor result[]) {
-        SkPMColor   src = fPMColor;
-        unsigned    scale = SkAlpha255To256(255 - SkGetPackedA32(src));
-
-        for (int i = 0; i < count; i++) {
-            result[i] = src + SkAlphaMulQ(shader[i], scale);
-        }
+        SkBlitRow::Color32(result, shader, count, fPMColor);
     }
 
     virtual void filterSpan16(const uint16_t shader[], int count,
diff --git a/src/opts/SkBlitRow_opts_arm.cpp b/src/opts/SkBlitRow_opts_arm.cpp
index 9ed6ef5885..0c38113adf 100644
--- a/src/opts/SkBlitRow_opts_arm.cpp
+++ b/src/opts/SkBlitRow_opts_arm.cpp
@@ -443,3 +443,10 @@ const SkBlitRow::Proc SkBlitRow::gPlatform_4444_Procs[] = {
     NULL,   // S32A_D4444_Blend_Dither
 };
 
+const SkBlitRow::Proc32 SkBlitRow::gPlatform_Procs32[] = {
+    NULL,   // S32_Opaque,
+    NULL,   // S32_Blend,
+    NULL,   // S32A_Opaque,
+    NULL,   // S32A_Blend,
+};
+
diff --git a/src/opts/SkBlitRow_opts_none.cpp b/src/opts/SkBlitRow_opts_none.cpp
index 88ab7c4203..7a777597c8 100644
--- a/src/opts/SkBlitRow_opts_none.cpp
+++ b/src/opts/SkBlitRow_opts_none.cpp
@@ -30,3 +30,10 @@ const SkBlitRow::Proc SkBlitRow::gPlatform_4444_Procs[] = {
     NULL,   // S32A_D4444_Blend_Dither
 };
 
+const SkBlitRow::Proc32 SkBlitRow::gPlatform_Procs32[] = {
+    NULL,   // S32_Opaque,
+    NULL,   // S32_Blend,
+    NULL,   // S32A_Opaque,
+    NULL,   // S32A_Blend,
+};
+