This change is in preparation for updating how processor keys and meta keys are generated for frag procs.

BUG=skia:4182

Review URL: https://codereview.chromium.org/1298233002

4 files changed, 42 insertions, 19 deletions

diff --git a/include/gpu/GrFragmentProcessor.h b/include/gpu/GrFragmentProcessor.h
index 8d2a4d42b1..f08d3656b7 100644
--- a/include/gpu/GrFragmentProcessor.h
+++ b/include/gpu/GrFragmentProcessor.h
@@ -25,7 +25,9 @@ class GrFragmentProcessor : public GrProcessor {
 public:
     GrFragmentProcessor()
         : INHERITED()
-        , fUsesLocalCoords(false) {}
+        , fUsesLocalCoords(false)
+        , fNumTexturesExclChildren(0)
+        , fNumTransformsExclChildren(0) {}
 
     GrGLFragmentProcessor* createGLInstance() const;
 
@@ -40,6 +42,10 @@ public:
         }
     }
 
+    int numTexturesExclChildren() const { return fNumTexturesExclChildren; }
+
+    int numTransformsExclChildren() const { return fNumTransformsExclChildren; }
+
     int numTransforms() const { return fCoordTransforms.count(); }
 
     /** Returns the coordinate transformation at index. index must be valid according to
@@ -84,6 +90,8 @@ public:
     void computeInvariantOutput(GrInvariantOutput* inout) const;
 
 protected:
+    void addTextureAccess(const GrTextureAccess* textureAccess) override;
+
     /**
      * Fragment Processor subclasses call this from their constructor to register coordinate
      * transformations. Coord transforms provide a mechanism for a processor to receive coordinates
@@ -105,7 +113,8 @@ protected:
 
     /**
      * FragmentProcessor subclasses call this from their constructor to register any child
-     * FragmentProcessors they have.
+     * FragmentProcessors they have. This must be called AFTER all texture accesses and coord
+     * transforms have been added.
      * This is for processors whose shader code will be composed of nested processors whose output
      * colors will be combined somehow to produce its output color.  Registering these child
      * processors will allow the ProgramBuilder to automatically handle their transformed coords and
@@ -168,6 +177,9 @@ private:
      */
     SkSTArray<4, const GrCoordTransform*, true>  fCoordTransforms;
 
+    int                                          fNumTexturesExclChildren;
+    int                                          fNumTransformsExclChildren;
+
     SkTArray<GrFragmentStage, false>             fChildProcessors;
 
     typedef GrProcessor INHERITED;
diff --git a/include/gpu/GrProcessor.h b/include/gpu/GrProcessor.h
index 29d648cec8..c1f4aa0256 100644
--- a/include/gpu/GrProcessor.h
+++ b/include/gpu/GrProcessor.h
@@ -101,7 +101,7 @@ protected:
      * GrTextureAccess is typically a member field of the GrProcessor subclass. This must only be
      * called from the constructor because GrProcessors are immutable.
      */
-    void addTextureAccess(const GrTextureAccess* textureAccess);
+    virtual void addTextureAccess(const GrTextureAccess* textureAccess);
 
     bool hasSameTextureAccesses(const GrProcessor&) const;
 
diff --git a/src/gpu/GrProcessor.cpp b/src/gpu/GrProcessor.cpp
index 0eaab75b61..6716ab1d12 100644
--- a/src/gpu/GrProcessor.cpp
+++ b/src/gpu/GrProcessor.cpp
@@ -166,10 +166,24 @@ GrGLFragmentProcessor* GrFragmentProcessor::createGLInstance() const {
     return glFragProc;
 }
 
+void GrFragmentProcessor::addTextureAccess(const GrTextureAccess* textureAccess) {
+    // Can't add texture accesses after registering any children since their texture accesses have
+    // already been bubbled up into our fTextureAccesses array
+    SkASSERT(fChildProcessors.empty());
+
+    INHERITED::addTextureAccess(textureAccess);
+    fNumTexturesExclChildren++;
+}
+
 void GrFragmentProcessor::addCoordTransform(const GrCoordTransform* transform) {
+    // Can't add transforms after registering any children since their transforms have already been
+    // bubbled up into our fCoordTransforms array
+    SkASSERT(fChildProcessors.empty());
+
     fCoordTransforms.push_back(transform);
     fUsesLocalCoords = fUsesLocalCoords || transform->sourceCoords() == kLocal_GrCoordSet;
     SkDEBUGCODE(transform->setInProcessor();)
+    fNumTransformsExclChildren++;
 }
 
 int GrFragmentProcessor::registerChildProcessor(const GrFragmentProcessor* child) {
diff --git a/src/gpu/gl/GrGLFragmentProcessor.cpp b/src/gpu/gl/GrGLFragmentProcessor.cpp
index d2902c26ac..078491a5bf 100644
--- a/src/gpu/gl/GrGLFragmentProcessor.cpp
+++ b/src/gpu/gl/GrGLFragmentProcessor.cpp
@@ -32,8 +32,8 @@ void GrGLFragmentProcessor::emitChild(int childIndex, const char* inputColor,
 
     /*
      * We now want to find the subset of coords and samplers that belong to the child and its
-     * descendants and put that into childCoords and childSamplers. To do so, we must do a
-     * backwards linear search on coords and samplers.
+     * descendants and put that into childCoords and childSamplers. To do so, we'll do a forwards
+     * linear search.
      *
      * Explanation:
      * Each GrFragmentProcessor has a copy of all the transforms and textures of itself and
@@ -53,23 +53,20 @@ void GrGLFragmentProcessor::emitChild(int childIndex, const char* inputColor,
      *                      (C)          (E)          (F)
      *                     [c1]      [e1,e2,e3]      [f1,f2]
      *
-     * So if we're inside proc A's emitCode, and A is about to call emitCode on proc B, we want the
-     * EmitArgs that's passed onto B to only contain its and its descendants' coords. The
+     * So if we're inside proc A's emitCode, and A is about to call emitCode on proc D, we want the
+     * EmitArgs that's passed onto D to only contain its and its descendants' coords. The
      * EmitArgs given to A would contain the transforms [a1,b1,b2,c1,d1,e1,e2,e3,f1,f2], and we want
-     * to extract the subset [b1,b2,c1] to pass on to B. We can do this with a backwards linear
-     * search since we know that D's subtree has 6 transforms and B's subtree has 3 transforms (by
-     * calling D.numTextures() and B.numTextures()), so we know the start of B's transforms is 9
-     * from the end of A's transforms.  We cannot do this with a forwards linear search since we
-     * don't know how many transforms belong to A (A.numTextures() will return 10, not 1), so
-     * we wouldn't know how many transforms to initially skip in A's array if using a forward linear
-     * search.
+     * to extract the subset [d1,e1,e2,e3,f1,f2] to pass on to D. We can do this with a linear
+     * search since we know that A has 1 transform (using A.numTransformsExclChildren()), and B's
+     * subtree has 3 transforms (using B.numTransforms()), so we know the start of D's transforms is
+     * 4 after the start of A's transforms.
      * Textures work the same way as transforms.
      */
-    int firstCoordAt = args.fFp.numTransforms();
-    int firstSamplerAt = args.fFp.numTextures();
-    for (int i = args.fFp.numChildProcessors() - 1; i >= childIndex; --i) {
-        firstCoordAt -= args.fFp.childProcessor(i).numTransforms();
-        firstSamplerAt -= args.fFp.childProcessor(i).numTextures();
+    int firstCoordAt = args.fFp.numTransformsExclChildren();
+    int firstSamplerAt = args.fFp.numTexturesExclChildren();
+    for (int i = 0; i < childIndex; ++i) {
+        firstCoordAt += args.fFp.childProcessor(i).numTransforms();
+        firstSamplerAt += args.fFp.childProcessor(i).numTextures();
     }
     TransformedCoordsArray childCoords;
     TextureSamplerArray childSamplers;