diff --git a/gm/gradtext.cpp b/gm/gradtext.cpp
index 6c75db1b1a..a3769d2728 100644
--- a/gm/gradtext.cpp
+++ b/gm/gradtext.cpp
@@ -9,15 +9,16 @@
 #include "SkCanvas.h"
 #include "SkGradientShader.h"
 
+// test shader w/ transparency
 static SkShader* make_grad(SkScalar width) {
-    SkColor colors[] = { SK_ColorBLACK, SK_ColorBLACK, 0 };
-    SkScalar pos[] = { 0, SK_Scalar1 * 5 / 10, SK_Scalar1 };
+    SkColor colors[] = { SK_ColorRED, 0x0000FF00, SK_ColorBLUE };
     SkPoint pts[] = { { 0, 0 }, { width, 0 } };
-    return SkGradientShader::CreateLinear(pts, colors, pos,
+    return SkGradientShader::CreateLinear(pts, colors, NULL,
                                           SK_ARRAY_COUNT(colors),
                                           SkShader::kMirror_TileMode);
 }
 
+// test opaque shader
 static SkShader* make_grad2(SkScalar width) {
     SkColor colors[] = { SK_ColorRED, SK_ColorGREEN, SK_ColorBLUE };
     SkPoint pts[] = { { 0, 0 }, { width, 0 } };
@@ -38,7 +39,7 @@ protected:
         return SkString("gradtext");
     }
 
-    virtual SkISize onISize() { return make_isize(640, 480); }
+    virtual SkISize onISize() { return make_isize(500, 480); }
 
     static void draw_text(SkCanvas* canvas, const SkPaint& paint) {
         const char* text = "When in the course of human events";
@@ -52,26 +53,35 @@ protected:
         p.setAntiAlias(false);
         draw_text(canvas, p);
         p.setAntiAlias(true);
-        canvas->translate(0, SkIntToScalar(20));
+        canvas->translate(0, paint.getTextSize() * 4/3);
         draw_text(canvas, p);
         p.setLCDRenderText(true);
-        canvas->translate(0, SkIntToScalar(20));
+        canvas->translate(0, paint.getTextSize() * 4/3);
         draw_text(canvas, p);
     }
 
     virtual void onDraw(SkCanvas* canvas) {
-        canvas->translate(SkIntToScalar(20), SkIntToScalar(20));
-
-        //  The blits special-case opaque and non-paque shaders, so test both
-
         SkPaint paint;
-        paint.setTextSize(SkIntToScalar(16));
-        paint.setShader(make_grad(SkIntToScalar(80)))->unref();
+        paint.setTextSize(SkIntToScalar(26));
 
-        draw_text3(canvas, paint);
-        canvas->translate(0, SkIntToScalar(40));
-        paint.setShader(make_grad2(SkIntToScalar(80)))->unref();
-        draw_text3(canvas, paint);
+        const SkISize& size = this->getISize();
+        SkRect r = SkRect::MakeWH(SkIntToScalar(size.width()),
+                                  SkIntToScalar(size.height()) / 2);
+        canvas->drawRect(r, paint);
+
+        canvas->translate(SkIntToScalar(20), paint.getTextSize());
+
+        for (int i = 0; i < 2; ++i) {
+            paint.setShader(make_grad(SkIntToScalar(80)))->unref();
+            draw_text3(canvas, paint);
+
+            canvas->translate(0, paint.getTextSize() * 2);
+
+            paint.setShader(make_grad2(SkIntToScalar(80)))->unref();
+            draw_text3(canvas, paint);
+            
+            canvas->translate(0, paint.getTextSize() * 2);
+        }
     }
 
 private:
diff --git a/src/core/SkBlitMask_D32.cpp b/src/core/SkBlitMask_D32.cpp
index 1a006e5e1f..72a83ee2eb 100644
--- a/src/core/SkBlitMask_D32.cpp
+++ b/src/core/SkBlitMask_D32.cpp
@@ -552,6 +552,14 @@ static void LCD16_RowProc_Opaque(SkPMColor* SK_RESTRICT dst,
     }
 }
 
+static int mul(int a, int b) {
+    return a * b >> 8;
+}
+
+static int src_alpha_blend(int src, int dst, int srcA, int mask) {
+    return dst + mul(src - mul(srcA, dst), mask);
+}
+
 static void LCD32_RowProc_Blend(SkPMColor* SK_RESTRICT dst,
                                 const SkPMColor* SK_RESTRICT mask,
                                 const SkPMColor* SK_RESTRICT src, int count) {
@@ -564,7 +572,7 @@ static void LCD32_RowProc_Blend(SkPMColor* SK_RESTRICT dst,
         SkPMColor s = src[i];
         int srcA = SkGetPackedA32(s);
         int srcR = SkGetPackedR32(s);
-        int srcG = SkGetPackedB32(s);
+        int srcG = SkGetPackedG32(s);
         int srcB = SkGetPackedB32(s);
         
         srcA = SkAlpha255To256(srcA);
@@ -575,15 +583,11 @@ static void LCD32_RowProc_Blend(SkPMColor* SK_RESTRICT dst,
         int maskG = SkGetPackedG32(m);
         int maskB = SkGetPackedB32(m);
 
-        // Now upscale them to 0..256, so we can use SkAlphaBlend
+        // Now upscale them to 0..256
         maskR = SkAlpha255To256(maskR);
         maskG = SkAlpha255To256(maskG);
         maskB = SkAlpha255To256(maskB);
-        
-        maskR = maskR * srcA >> 8;
-        maskG = maskG * srcA >> 8;
-        maskB = maskB * srcA >> 8;
-        
+
         int dstR = SkGetPackedR32(d);
         int dstG = SkGetPackedG32(d);
         int dstB = SkGetPackedB32(d);
@@ -591,9 +595,9 @@ static void LCD32_RowProc_Blend(SkPMColor* SK_RESTRICT dst,
         // LCD blitting is only supported if the dst is known/required
         // to be opaque
         dst[i] = SkPackARGB32(0xFF,
-                              SkAlphaBlend(srcR, dstR, maskR),
-                              SkAlphaBlend(srcG, dstG, maskG),
-                              SkAlphaBlend(srcB, dstB, maskB));
+                              src_alpha_blend(srcR, dstR, srcA, maskR),
+                              src_alpha_blend(srcG, dstG, srcA, maskG),
+                              src_alpha_blend(srcB, dstB, srcA, maskB));
     }
 }
 
@@ -612,9 +616,11 @@ static void LCD32_RowProc_Opaque(SkPMColor* SK_RESTRICT dst,
         int maskR = SkGetPackedR32(m);
         int maskG = SkGetPackedG32(m);
         int maskB = SkGetPackedB32(m);
+
         int srcR = SkGetPackedR32(s);
-        int srcG = SkGetPackedB32(s);
-        int srcB = SkGetPackedB32(s);        
+        int srcG = SkGetPackedG32(s);
+        int srcB = SkGetPackedB32(s);
+
         int dstR = SkGetPackedR32(d);
         int dstG = SkGetPackedG32(d);
         int dstB = SkGetPackedB32(d);
@@ -637,7 +643,6 @@ SkBlitMask::RowProc SkBlitMask::RowFactory(SkBitmap::Config config,
                                            SkMask::Format format,
                                            RowFlags flags) {
 // make this opt-in until chrome can rebaseline
-#ifdef SK_ENABLE_FAST_SHADERMASK
     RowProc proc = PlatformRowProcs(config, format, flags);
     if (proc) {
         return proc;
@@ -670,7 +675,6 @@ SkBlitMask::RowProc SkBlitMask::RowFactory(SkBitmap::Config config,
         default:
             break;
     }
-#endif
     return NULL;
 }
 
diff --git a/src/core/SkDevice.cpp b/src/core/SkDevice.cpp
index 860e45b0ed..1865212e6e 100644
--- a/src/core/SkDevice.cpp
+++ b/src/core/SkDevice.cpp
@@ -365,11 +365,6 @@ bool SkDevice::filterTextFlags(const SkPaint& paint, TextFlags* flags) {
     }
 
     if (SkBitmap::kARGB_8888_Config != fBitmap.config() ||
-#ifndef SK_ENABLE_FAST_SHADERMASK
-        paint.getShader() ||
-        paint.getMaskFilter() ||
-        paint.getColorFilter() ||
-#endif
         paint.getRasterizer() ||
         paint.getPathEffect() ||
         paint.isFakeBoldText() ||