Don't batch large instanced rects with other shapes

src/gpu/instanced/InstancedRendering.cpp

Index: src/gpu/instanced/InstancedRendering.cpp
diff --git a/src/gpu/instanced/InstancedRendering.cpp b/src/gpu/instanced/InstancedRendering.cpp
index 165bff494b3fa0ab1f241c844ec59afb8b582da9..12edc986c3534a34906421dff313af830d340af8 100644
--- a/src/gpu/instanced/InstancedRendering.cpp
+++ b/src/gpu/instanced/InstancedRendering.cpp
@@ -184,6 +184,7 @@ InstancedRendering::Batch* InstancedRendering::recordShape(ShapeType type, const
     const float* rectAsFloats = localRect.asScalars(); // Ensure SkScalar == float.
     memcpy(&instance.fLocalRect, rectAsFloats, 4 * sizeof(float));
 
+    batch->fPixelLoad = batch->fBounds.height() * batch->fBounds.width();
     return batch;
 }
 
@@ -366,14 +367,29 @@ bool InstancedRendering::Batch::onCombineIfPossible(GrBatch* other, const GrCaps
     SkASSERT(fTailDraw);
     SkASSERT(that->fTailDraw);
 
-    if (!fInfo.canJoin(that->fInfo) ||
+    if (!BatchInfo::CanCombine(fInfo, that->fInfo) ||
         !GrPipeline::CanCombine(*this->pipeline(), this->bounds(),
                                 *that->pipeline(), that->bounds(), caps)) {
         return false;
     }
 
+    BatchInfo combinedInfo = fInfo | that->fInfo;
+    if (!combinedInfo.isSimpleRects()) {
+        // This threshold was chosen with the "shapes_mixed" bench on a MacBook with Intel graphics.
+        // There seems to be a wide range where it doesn't matter if we combine or not. What matters
+        // is that the itty bitty rects combine with other shapes and the giant ones don't.
+        constexpr SkScalar kMaxPixelsToGeneralizeRects = 256 * 256;
+        if (fInfo.isSimpleRects() && fPixelLoad > kMaxPixelsToGeneralizeRects) {
+            return false;
+        }
+        if (that->fInfo.isSimpleRects() && that->fPixelLoad > kMaxPixelsToGeneralizeRects) {
+            return false;
+        }
+    }
+
     fBounds.join(that->fBounds);
-    fInfo.join(that->fInfo);
+    fInfo = combinedInfo;
+    fPixelLoad += that->fPixelLoad;
 
     // Adopt the other batch's draws.
     fNumDraws += that->fNumDraws;