Use texture barriers to read directly from the RT

src/gpu/GrTargetCommands.h

Index: src/gpu/GrTargetCommands.h
diff --git a/src/gpu/GrTargetCommands.h b/src/gpu/GrTargetCommands.h
index 4b3dd574e3a537c05b0aaaee4ef33caf5100000e..3a21afe938d70e357b7968cbdc437af3d5aafd15 100644
--- a/src/gpu/GrTargetCommands.h
+++ b/src/gpu/GrTargetCommands.h
@@ -47,6 +47,7 @@ public:
             kDrawPath_CmdType          = 6,
             kDrawPaths_CmdType         = 7,
             kDrawBatch_CmdType         = 8,
+            kBlendBarrier_CmdType      = 9,
         };
 
         Cmd(CmdType type) : fMarkerID(-1), fType(type) {}
@@ -121,6 +122,7 @@ public:
                            GrSurface* src,
                            const SkIRect& srcRect,
                            const SkIPoint& dstPoint);
+    Cmd* recordBlendBarrier(GrInOrderDrawBuffer*);
 
 protected:
     void willReserveVertexAndIndexSpace(int vertexCount,
@@ -295,6 +297,11 @@ private:
         GrBatchTarget*         fBatchTarget;
     };
 
+    struct BlendBarrier : public Cmd {
+        BlendBarrier() : Cmd(kBlendBarrier_CmdType) {}
+        void execute(GrGpu*, const SetState*) override;
+    };
+
      static const int kCmdBufferInitialSizeInBytes = 8 * 1024;
 
      typedef void* TCmdAlign; // This wouldn't be enough align if a command used long double.