Fix filter primitive bounds computations.

src/core/SkRecordDraw.cpp

Index: src/core/SkRecordDraw.cpp
diff --git a/src/core/SkRecordDraw.cpp b/src/core/SkRecordDraw.cpp
index 4847273f52a826f4eb2510abb75b4e7860cc82d2..90b85d1e969e63277b0eef6679576049c0e15d11 100644
--- a/src/core/SkRecordDraw.cpp
+++ b/src/core/SkRecordDraw.cpp
@@ -221,6 +221,7 @@ private:
         int controlOps;        // Number of control ops in this Save block, including the Save.
         Bounds bounds;         // Bounds of everything in the block.
         const SkPaint* paint;  // Unowned.  If set, adjusts the bounds of all ops in this block.
+        SkMatrix ctm;
     };
 
     // Only Restore, SetMatrix, and Concat change the CTM.
@@ -301,6 +302,7 @@ private:
         sb.bounds =
             PaintMayAffectTransparentBlack(paint) ? fCurrentClipBounds : Bounds::MakeEmpty();
         sb.paint = paint;
+        sb.ctm = this->fCTM;
 
         fSaveStack.push(sb);
         this->pushControl();
@@ -563,9 +565,15 @@ private:
 
     bool adjustForSaveLayerPaints(SkRect* rect, int savesToIgnore = 0) const {
         for (int i = fSaveStack.count() - 1 - savesToIgnore; i >= 0; i--) {
+            SkMatrix inverse;
+            if (!fSaveStack[i].ctm.invert(&inverse)) {
+                return false;
+            }
+            inverse.mapRect(rect);

[reed1, 2015/12/07 20:24:11]

Is this change a fix in its own right, or is it connected to the change for
imagefilters?

- I assume AdjustForPaint wants its rectangle in local coordinates
- I assume the parameter passed to adjustForSaveLayerPaints is in global
coordinates

If both of these are true, then it seems like we have a bug outside of this
CL...?

[Stephen White, 2015/12/07 20:29:40]

IIRC the problem/bug was being papered over by the join() calls, which
were previously making adjustForSaveLayerPaints() work regardless. 
When I removed the join() calls, tiled picture playback started
showing artifacts.

I'm going to take another look to see if this change is still necessary.

[Stephen White, 2015/12/07 20:51:15]

To follow up: this breaks layout tests css3/filters/effect-reference-tile.html
and effect-reference-subregion-hw.html over in Chrome,
both with respect to SkTileImageFilter. There's no Skia test coverage
for this case ATM; I'll try to construct some.

The problem is that filters apply the CTM that was in effect
at saveLayer() time to their filter parameters, not the current CTM.
So we need to capture that CTM, and use it when transforming rects 
backwards through the saveLayer stack.

This affects SkTileImageFilter, since its computeFastBounds()
basically ignores the rect passed to it, and always returns its 
fDstRect, since it always draws its result to that location.

That said, I'm not 100% certain that this change is correct.

             if (!AdjustForPaint(fSaveStack[i].paint, rect)) {
                 return false;
             }
+            fSaveStack[i].ctm.mapRect(rect);
         }
         return true;
     }