diff --git a/core/SkPaintPriv.cpp b/core/SkPaintPriv.cpp
index ce05389..65fd0e7 100644
--- a/core/SkPaintPriv.cpp
+++ b/core/SkPaintPriv.cpp
@@ -76,3 +76,24 @@
     }
     return false;
 }
+
+bool NeedsDeepCopy(const SkPaint& paint) {
+    /*
+     *  These fields are known to be immutable, and so can be shallow-copied
+     *
+     *  getTypeface()
+     *  getAnnotation()
+     *  paint.getColorFilter()
+     *  getXfermode()
+     *  getPathEffect()
+     *  getMaskFilter()
+     */
+
+    return paint.getShader() ||
+#ifdef SK_SUPPORT_LEGACY_LAYERRASTERIZER_API
+           paint.getRasterizer() ||
+#endif
+           paint.getLooper() || // needs to hide its addLayer...
+           paint.getImageFilter();
+}
+
diff --git a/core/SkPaintPriv.h b/core/SkPaintPriv.h
index 38c9063..9668fef 100644
--- a/core/SkPaintPriv.h
+++ b/core/SkPaintPriv.h
@@ -22,4 +22,11 @@
 */
 bool isPaintOpaque(const SkPaint* paint,
                    const SkBitmap* bmpReplacesShader = NULL);
+
+/** Returns true if the provided paint has fields which are not
+    immutable (and will thus require deep copying).
+    @param paint the paint to be analyzed
+    @return true if the paint requires a deep copy
+*/
+bool NeedsDeepCopy(const SkPaint& paint);
 #endif
diff --git a/core/SkPicture.cpp b/core/SkPicture.cpp
index 3b04906..6843430 100644
--- a/core/SkPicture.cpp
+++ b/core/SkPicture.cpp
@@ -15,6 +15,7 @@
 #include "SkBitmapDevice.h"
 #include "SkCanvas.h"
 #include "SkChunkAlloc.h"
+#include "SkPaintPriv.h"
 #include "SkPicture.h"
 #include "SkRegion.h"
 #include "SkStream.h"
@@ -217,26 +218,6 @@
     return clonedPicture;
 }
 
-static bool needs_deep_copy(const SkPaint& paint) {
-    /*
-     *  These fields are known to be immutable, and so can be shallow-copied
-     *
-     *  getTypeface()
-     *  getAnnotation()
-     *  paint.getColorFilter()
-     *  getXfermode()
-     *  getPathEffect()
-     *  getMaskFilter()
-     */
-
-    return paint.getShader() ||
-#ifdef SK_SUPPORT_LEGACY_LAYERRASTERIZER_API
-           paint.getRasterizer() ||
-#endif
-           paint.getLooper() || // needs to hide its addLayer...
-           paint.getImageFilter();
-}
-
 void SkPicture::clone(SkPicture* pictures, int count) const {
     SkPictCopyInfo copyInfo;
     SkPictInfo info;
@@ -282,7 +263,7 @@
 
                 SkDEBUGCODE(int heapSize = SafeCount(fPlayback->fBitmapHeap.get());)
                 for (int i = 0; i < paintCount; i++) {
-                    if (needs_deep_copy(fPlayback->fPaints->at(i))) {
+                    if (NeedsDeepCopy(fPlayback->fPaints->at(i))) {
                         copyInfo.paintData[i] =
                             SkFlatData::Create<SkPaint::FlatteningTraits>(&copyInfo.controller,
                                                               fPlayback->fPaints->at(i), 0);
diff --git a/gpu/GrPictureUtils.cpp b/gpu/GrPictureUtils.cpp
index e8c3b50..d8e4161 100644
--- a/gpu/GrPictureUtils.cpp
+++ b/gpu/GrPictureUtils.cpp
@@ -7,6 +7,8 @@
 
 #include "GrPictureUtils.h"
 #include "SkDevice.h"
+#include "SkDraw.h"
+#include "SkPaintPriv.h"
 
 // The GrGather device performs GPU-backend-specific preprocessing on
 // a picture. The results are stored in a GPUAccelData.
@@ -20,12 +22,17 @@
 public:
     SK_DECLARE_INST_COUNT(GrGatherDevice)
 
-    GrGatherDevice(int width, int height, SkPicture* picture, GPUAccelData* accelData) {
+    GrGatherDevice(int width, int height, SkPicture* picture, GPUAccelData* accelData,
+                   int saveLayerDepth) {
         fPicture = picture;
+        fSaveLayerDepth = saveLayerDepth;
+        fInfo.fValid = true;
         fInfo.fSize.set(width, height);
+        fInfo.fPaint = NULL;
         fInfo.fSaveLayerOpID = fPicture->EXPERIMENTAL_curOpID();
         fInfo.fRestoreOpID = 0;
         fInfo.fHasNestedLayers = false;
+        fInfo.fIsNested = (2 == fSaveLayerDepth);
 
         fEmptyBitmap.setConfig(SkImageInfo::Make(fInfo.fSize.fWidth,
                                                  fInfo.fSize.fHeight,
@@ -110,7 +117,8 @@
                               const SkPaint& paint) SK_OVERRIDE {
     }
     virtual void drawDevice(const SkDraw& draw, SkBaseDevice* deviceIn, int x, int y,
-                            const SkPaint&) SK_OVERRIDE {
+                            const SkPaint& paint) SK_OVERRIDE {
+        // deviceIn is the one that is being "restored" back to its parent
         GrGatherDevice* device = static_cast<GrGatherDevice*>(deviceIn);
 
         if (device->fAlreadyDrawn) {
@@ -118,6 +126,29 @@
         }
 
         device->fInfo.fRestoreOpID = fPicture->EXPERIMENTAL_curOpID();
+        device->fInfo.fCTM = *draw.fMatrix;
+        device->fInfo.fCTM.postTranslate(SkIntToScalar(-device->getOrigin().fX),
+                                         SkIntToScalar(-device->getOrigin().fY));
+
+        // We need the x & y values that will yield 'getOrigin' when transformed
+        // by 'draw.fMatrix'.
+        device->fInfo.fOffset.iset(device->getOrigin());
+
+        SkMatrix invMatrix;
+        if (draw.fMatrix->invert(&invMatrix)) {
+            invMatrix.mapPoints(&device->fInfo.fOffset, 1);
+        } else {
+            device->fInfo.fValid = false;
+        }
+
+        if (NeedsDeepCopy(paint)) {
+            // This NULL acts as a signal that the paint was uncopyable (for now)
+            device->fInfo.fPaint = NULL;
+            device->fInfo.fValid = false;
+        } else {
+            device->fInfo.fPaint = SkNEW_ARGS(SkPaint, (paint));
+        }
+
         fAccelData->addSaveLayerInfo(device->fInfo);
         device->fAlreadyDrawn = true;
     }
@@ -158,6 +189,9 @@
     // The information regarding the saveLayer call this device represents.
     GPUAccelData::SaveLayerInfo fInfo;
 
+    // The depth of this device in the saveLayer stack
+    int fSaveLayerDepth;
+
     virtual void replaceBitmapBackendForRasterSurface(const SkBitmap&) SK_OVERRIDE {
         NotSupported();
     }
@@ -167,7 +201,8 @@
         SkASSERT(kSaveLayer_Usage == usage);
 
         fInfo.fHasNestedLayers = true;
-        return SkNEW_ARGS(GrGatherDevice, (info.width(), info.height(), fPicture, fAccelData));
+        return SkNEW_ARGS(GrGatherDevice, (info.width(), info.height(), fPicture, 
+                                           fAccelData, fSaveLayerDepth+1));
     }
 
     virtual void flush() SK_OVERRIDE {}
@@ -239,7 +274,7 @@
         return ;
     }
 
-    GrGatherDevice device(pict->width(), pict->height(), pict, accelData);
+    GrGatherDevice device(pict->width(), pict->height(), pict, accelData, 0);
     GrGatherCanvas canvas(&device, pict);
 
     canvas.gather();
diff --git a/gpu/GrPictureUtils.h b/gpu/GrPictureUtils.h
index 6b4d901..5ca4132 100644
--- a/gpu/GrPictureUtils.h
+++ b/gpu/GrPictureUtils.h
@@ -17,8 +17,21 @@
 public:
     // Information about a given saveLayer in an SkPicture
     struct SaveLayerInfo {
+        // True if the SaveLayerInfo is valid. False if either 'fOffset' is
+        // invalid (due to a non-invertible CTM) or 'fPaint' is NULL (due
+        // to a non-copyable paint).
+        bool fValid;
         // The size of the saveLayer
         SkISize fSize;
+        // The CTM in which this layer's draws must occur. It already incorporates
+        // the translation needed to map the layer's top-left point to the origin.
+        SkMatrix fCTM;
+        // The offset that needs to be passed to drawBitmap to correctly
+        // position the pre-rendered layer.
+        SkPoint fOffset;
+        // The paint to use on restore. NULL if the paint was not copyable (and
+        // thus that this layer should not be pulled forward).
+        const SkPaint* fPaint;
         // The ID of this saveLayer in the picture. 0 is an invalid ID.
         size_t  fSaveLayerOpID;
         // The ID of the matching restore in the picture. 0 is an invalid ID.
@@ -26,6 +39,8 @@
         // True if this saveLayer has at least one other saveLayer nested within it.
         // False otherwise.
         bool    fHasNestedLayers;
+        // True if this saveLayer is nested within another. False otherwise.
+        bool    fIsNested;
     };
 
     GPUAccelData(Key key) : INHERITED(key) { }
@@ -43,6 +58,14 @@
         return fSaveLayerInfo[index];
     }
 
+    // We may, in the future, need to pass in the GPUDevice in order to
+    // incorporate the clip and matrix state into the key
+    static SkPicture::AccelData::Key ComputeAccelDataKey() {
+        static const SkPicture::AccelData::Key gGPUID = SkPicture::AccelData::GenerateDomain();
+
+        return gGPUID;
+    }
+
 protected:
     SkTDArray<SaveLayerInfo> fSaveLayerInfo;
 
diff --git a/gpu/SkGpuDevice.cpp b/gpu/SkGpuDevice.cpp
index 714a6da..3119a9e 100644
--- a/gpu/SkGpuDevice.cpp
+++ b/gpu/SkGpuDevice.cpp
@@ -1910,16 +1910,8 @@
     return SkSurface::NewRenderTarget(fContext, info, fRenderTarget->numSamples());
 }
 
-// In the future this may not be a static method if we need to incorporate the
-// clip and matrix state into the key
-SkPicture::AccelData::Key SkGpuDevice::ComputeAccelDataKey() {
-    static const SkPicture::AccelData::Key gGPUID = SkPicture::AccelData::GenerateDomain();
-
-    return gGPUID;
-}
-
 void SkGpuDevice::EXPERIMENTAL_optimize(SkPicture* picture) {
-    SkPicture::AccelData::Key key = ComputeAccelDataKey();
+    SkPicture::AccelData::Key key = GPUAccelData::ComputeAccelDataKey();
 
     GPUAccelData* data = SkNEW_ARGS(GPUAccelData, (key));
 
@@ -1934,7 +1926,7 @@
 
 bool SkGpuDevice::EXPERIMENTAL_drawPicture(SkCanvas* canvas, SkPicture* picture) {
 
-    SkPicture::AccelData::Key key = ComputeAccelDataKey();
+    SkPicture::AccelData::Key key = GPUAccelData::ComputeAccelDataKey();
 
     const SkPicture::AccelData* data = picture->EXPERIMENTAL_getAccelData(key);
     if (NULL == data) {
@@ -1943,27 +1935,6 @@
 
     const GPUAccelData *gpuData = static_cast<const GPUAccelData*>(data);
 
-//#define SK_PRINT_PULL_FORWARD_INFO 1
-
-#ifdef SK_PRINT_PULL_FORWARD_INFO
-    static bool gPrintedAccelData = false;
-
-    if (!gPrintedAccelData) {
-        for (int i = 0; i < gpuData->numSaveLayers(); ++i) {
-            const GPUAccelData::SaveLayerInfo& info = gpuData->saveLayerInfo(i);
-
-            SkDebugf("%d: Width: %d Height: %d SL: %d R: %d hasNestedLayers: %s\n",
-                                            i,
-                                            info.fSize.fWidth,
-                                            info.fSize.fHeight,
-                                            info.fSaveLayerOpID,
-                                            info.fRestoreOpID,
-                                            info.fHasNestedLayers ? "T" : "F");
-        }
-        gPrintedAccelData = true;
-    }
-#endif
-
     SkAutoTArray<bool> pullForward(gpuData->numSaveLayers());
     for (int i = 0; i < gpuData->numSaveLayers(); ++i) {
         pullForward[i] = false;
@@ -1984,10 +1955,6 @@
 
     const SkPicture::OperationList& ops = picture->EXPERIMENTAL_getActiveOps(clip);
 
-#ifdef SK_PRINT_PULL_FORWARD_INFO
-    SkDebugf("rect: %d %d %d %d\n", clip.fLeft, clip.fTop, clip.fRight, clip.fBottom);
-#endif
-
     for (int i = 0; i < ops.numOps(); ++i) {
         for (int j = 0; j < gpuData->numSaveLayers(); ++j) {
             const GPUAccelData::SaveLayerInfo& info = gpuData->saveLayerInfo(j);
@@ -1998,17 +1965,5 @@
         }
     }
 
-#ifdef SK_PRINT_PULL_FORWARD_INFO
-    SkDebugf("Need SaveLayers: ");
-    for (int i = 0; i < gpuData->numSaveLayers(); ++i) {
-        if (pullForward[i]) {
-            const GrCachedLayer* layer = fContext->getLayerCache()->findLayerOrCreate(picture, i);
-
-            SkDebugf("%d (%d), ", i, layer->layerID());
-        }
-    }
-    SkDebugf("\n");
-#endif
-
     return false;
 }