diff --git a/src/core/SkBBoxHierarchyRecord.cpp b/src/core/SkBBoxHierarchyRecord.cpp
index 35949b6760..c27940af6f 100644
--- a/src/core/SkBBoxHierarchyRecord.cpp
+++ b/src/core/SkBBoxHierarchyRecord.cpp
@@ -10,8 +10,10 @@
 #include "SkPictureStateTree.h"
 #include "SkBBoxHierarchy.h"
 
-SkBBoxHierarchyRecord::SkBBoxHierarchyRecord(uint32_t recordFlags, SkBBoxHierarchy* h)
-    : INHERITED(recordFlags) {
+SkBBoxHierarchyRecord::SkBBoxHierarchyRecord(uint32_t recordFlags,
+                                             SkBBoxHierarchy* h,
+                                             SkDevice* device)
+    : INHERITED(recordFlags, device) {
     fStateTree = SkNEW(SkPictureStateTree);
     fBoundingHierarchy = h;
     fBoundingHierarchy->ref();
diff --git a/src/core/SkBBoxHierarchyRecord.h b/src/core/SkBBoxHierarchyRecord.h
index c0cf365e76..4741338f2a 100644
--- a/src/core/SkBBoxHierarchyRecord.h
+++ b/src/core/SkBBoxHierarchyRecord.h
@@ -18,7 +18,8 @@
 class SkBBoxHierarchyRecord : public SkBBoxRecord {
 public:
     /** This will take a ref of h */
-    SkBBoxHierarchyRecord(uint32_t recordFlags, SkBBoxHierarchy* h);
+    SkBBoxHierarchyRecord(uint32_t recordFlags, SkBBoxHierarchy* h,
+                          SkDevice*);
 
     virtual void handleBBox(const SkRect& bounds) SK_OVERRIDE;
 
diff --git a/src/core/SkBBoxRecord.h b/src/core/SkBBoxRecord.h
index c13a7f01c4..e6b85db5f7 100644
--- a/src/core/SkBBoxRecord.h
+++ b/src/core/SkBBoxRecord.h
@@ -19,7 +19,8 @@
 class SkBBoxRecord : public SkPictureRecord {
 public:
 
-    SkBBoxRecord(uint32_t recordFlags) :INHERITED(recordFlags) { }
+    SkBBoxRecord(uint32_t recordFlags, SkDevice* device)
+            : INHERITED(recordFlags, device) { }
     virtual ~SkBBoxRecord() { }
 
     /**
diff --git a/src/core/SkPicture.cpp b/src/core/SkPicture.cpp
index 7e8b371b58..b3d3751e36 100644
--- a/src/core/SkPicture.cpp
+++ b/src/core/SkPicture.cpp
@@ -13,6 +13,7 @@
 
 #include "SkCanvas.h"
 #include "SkChunkAlloc.h"
+#include "SkDevice.h"
 #include "SkPicture.h"
 #include "SkRegion.h"
 #include "SkStream.h"
@@ -185,25 +186,26 @@ SkCanvas* SkPicture::beginRecording(int width, int height,
         fRecord = NULL;
     }
 
+    SkBitmap bm;
+    bm.setConfig(SkBitmap::kNo_Config, width, height);
+    SkAutoTUnref<SkDevice> dev(SkNEW_ARGS(SkDevice, (bm)));
+
     if (recordingFlags & kOptimizeForClippedPlayback_RecordingFlag) {
         SkScalar aspectRatio = SkScalarDiv(SkIntToScalar(width),
                                            SkIntToScalar(height));
         SkRTree* tree = SkRTree::Create(kRTreeMinChildren, kRTreeMaxChildren,
                                         aspectRatio);
         SkASSERT(NULL != tree);
-        fRecord = SkNEW_ARGS(SkBBoxHierarchyRecord, (recordingFlags, tree));
+        fRecord = SkNEW_ARGS(SkBBoxHierarchyRecord, (recordingFlags, tree, dev));
         tree->unref();
     } else {
-        fRecord = SkNEW_ARGS(SkPictureRecord, (recordingFlags));
+        fRecord = SkNEW_ARGS(SkPictureRecord, (recordingFlags, dev));
     }
+    fRecord->beginRecording();
 
     fWidth = width;
     fHeight = height;
 
-    SkBitmap bm;
-    bm.setConfig(SkBitmap::kNo_Config, width, height);
-    fRecord->setBitmapDevice(bm);
-
     return fRecord;
 }
 
diff --git a/src/core/SkPictureRecord.cpp b/src/core/SkPictureRecord.cpp
index 944fd1898f..60e51d1829 100644
--- a/src/core/SkPictureRecord.cpp
+++ b/src/core/SkPictureRecord.cpp
@@ -15,10 +15,12 @@
 #define HEAP_BLOCK_SIZE 4096
 
 enum {
+    // just need a value that save or getSaveCount would never return
     kNoInitialSave = -1,
 };
 
-SkPictureRecord::SkPictureRecord(uint32_t flags) :
+SkPictureRecord::SkPictureRecord(uint32_t flags, SkDevice* device) :
+        INHERITED(device),
         fBoundingHierarchy(NULL),
         fStateTree(NULL),
         fFlattenableHeap(HEAP_BLOCK_SIZE),
@@ -33,12 +35,13 @@ SkPictureRecord::SkPictureRecord(uint32_t flags) :
 #endif
 
     fRestoreOffsetStack.setReserve(32);
-    fInitialSaveCount = kNoInitialSave;
 
     fBitmapHeap = SkNEW(SkBitmapHeap);
     fFlattenableHeap.setBitmapStorage(fBitmapHeap);
     fPathHeap = NULL;   // lazy allocate
     fFirstSavedLayerIndex = kNoSavedLayerIndex;
+
+    fInitialSaveCount = kNoInitialSave;
 }
 
 SkPictureRecord::~SkPictureRecord() {
@@ -53,13 +56,8 @@ SkPictureRecord::~SkPictureRecord() {
 ///////////////////////////////////////////////////////////////////////////////
 
 SkDevice* SkPictureRecord::setDevice(SkDevice* device) {
-    SkASSERT(kNoInitialSave == fInitialSaveCount);
-    this->INHERITED::setDevice(device);
-
-    // The bracketting save() call needs to be recorded after setting the
-    // device otherwise the clip stack will get messed-up
-    fInitialSaveCount = this->save(SkCanvas::kMatrixClip_SaveFlag);
-    return device;
+    SkASSERT(!"eeek, don't try to change the device on a recording canvas");
+    return this->INHERITED::setDevice(device);
 }
 
 int SkPictureRecord::save(SaveFlags flags) {
@@ -321,6 +319,13 @@ void SkPictureRecord::fillRestoreOffsetPlaceholdersForCurrentStackLevel(
 #endif
 }
 
+void SkPictureRecord::beginRecording() {
+    // we have to call this *after* our constructor, to ensure that it gets
+    // recorded. This is balanced by restoreToCount() call from endRecording,
+    // which in-turn calls our overridden restore(), so those get recorded too.
+    fInitialSaveCount = this->save(kMatrixClip_SaveFlag);
+}
+
 void SkPictureRecord::endRecording() {
     SkASSERT(kNoInitialSave != fInitialSaveCount);
     this->restoreToCount(fInitialSaveCount);
diff --git a/src/core/SkPictureRecord.h b/src/core/SkPictureRecord.h
index fec1a829f3..c36c24ab3d 100644
--- a/src/core/SkPictureRecord.h
+++ b/src/core/SkPictureRecord.h
@@ -21,7 +21,7 @@ class SkBBoxHierarchy;
 
 class SkPictureRecord : public SkCanvas {
 public:
-    SkPictureRecord(uint32_t recordFlags);
+    SkPictureRecord(uint32_t recordFlags, SkDevice*);
     virtual ~SkPictureRecord();
 
     virtual SkDevice* setDevice(SkDevice* device) SK_OVERRIDE;
@@ -86,7 +86,9 @@ public:
         return fWriter;
     }
 
+    void beginRecording();
     void endRecording();
+
 private:
     void recordRestoreOffsetPlaceholder(SkRegion::Op);
     void fillRestoreOffsetPlaceholdersForCurrentStackLevel(