diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 8aa30fcf..30da328c 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -64,6 +64,10 @@
 #define USE_ZERO_SPANS 1
 #define USE_BO_FOR_SCRATCH_PIXMAP 1
 
+#define MIGRATE_ALL 0
+
+#define ACCEL_PUT_IMAGE 1
+
 static int sna_font_key;
 
 static const uint8_t copy_ROP[] = {
@@ -1142,8 +1146,33 @@ sna_drawable_move_region_to_cpu(DrawablePtr drawable,
 		goto done;
 	}
 
+	if (MIGRATE_ALL && priv->gpu_damage) {
+		BoxPtr box;
+		int n = sna_damage_get_boxes(priv->gpu_damage, &box);
+		if (n) {
+			Bool ok;
+
+			DBG(("%s: forced migration\n", __FUNCTION__));
+
+			assert(pixmap_contains_damage(pixmap, priv->gpu_damage));
+
+			ok = FALSE;
+			if (priv->cpu_bo && sna->kgem.gen >= 60)
+				ok = sna->render.copy_boxes(sna, GXcopy,
+							    pixmap, priv->gpu_bo, 0, 0,
+							    pixmap, priv->cpu_bo, 0, 0,
+							    box, n);
+			if (!ok)
+				sna_read_boxes(sna,
+					       priv->gpu_bo, 0, 0,
+					       pixmap, 0, 0,
+					       box, n);
+		}
+		sna_damage_destroy(&priv->gpu_damage);
+	}
+
 	if (sna_damage_contains_box(priv->gpu_damage,
-				    REGION_EXTENTS(NULL, region)) != PIXMAN_REGION_OUT) {
+				    &region->extents) != PIXMAN_REGION_OUT) {
 		DBG(("%s: region (%dx%d) intersects gpu damage\n",
 		     __FUNCTION__,
 		     region->extents.x2 - region->extents.x1,
@@ -1194,11 +1223,13 @@ sna_drawable_move_region_to_cpu(DrawablePtr drawable,
 			}
 
 			if (region_subsumes_damage(r, priv->gpu_damage)) {
-				BoxPtr box = REGION_RECTS(&DAMAGE_PTR(priv->gpu_damage)->region);
-				int n = REGION_NUM_RECTS(&DAMAGE_PTR(priv->gpu_damage)->region);
+				BoxPtr box;
+				int n;
 				Bool ok;
 
 				ok = FALSE;
+				n = sna_damage_get_boxes(priv->gpu_damage,
+							 &box);
 				if (priv->cpu_bo && sna->kgem.gen >= 30)
 					ok = sna->render.copy_boxes(sna, GXcopy,
 								    pixmap, priv->gpu_bo, 0, 0,
@@ -1238,7 +1269,7 @@ sna_drawable_move_region_to_cpu(DrawablePtr drawable,
 				if (sna_damage_intersect(priv->gpu_damage, r, &need)) {
 					BoxPtr box = REGION_RECTS(&need);
 					int n = REGION_NUM_RECTS(&need);
-					Bool ok = FALSE;
+					Bool ok;
 
 					ok = FALSE;
 					if (priv->cpu_bo && sna->kgem.gen >= 30)
@@ -1369,11 +1400,11 @@ sna_pixmap_move_area_to_gpu(PixmapPtr pixmap, BoxPtr box)
 		goto done;
 
 	region_set(&r, box);
-	if (region_subsumes_damage(&r, priv->cpu_damage)) {
-		int n = REGION_NUM_RECTS(&DAMAGE_PTR(priv->cpu_damage)->region);
+	if (MIGRATE_ALL || region_subsumes_damage(&r, priv->cpu_damage)) {
 		Bool ok;
+		int n;
 
-		box = REGION_RECTS(&DAMAGE_PTR(priv->cpu_damage)->region);
+		n = sna_damage_get_boxes(priv->cpu_damage, &box);
 		ok = FALSE;
 		if (priv->cpu_bo)
 			ok = sna->render.copy_boxes(sna, GXcopy,
@@ -1386,8 +1417,7 @@ sna_pixmap_move_area_to_gpu(PixmapPtr pixmap, BoxPtr box)
 			    box->x2 >= pixmap->drawable.width &&
 			    box->y2 >= pixmap->drawable.height) {
 				priv->gpu_bo =
-					sna_replace(sna,
-						    pixmap,
+					sna_replace(sna, pixmap,
 						    priv->gpu_bo,
 						    pixmap->devPrivate.ptr,
 						    pixmap->devKind);
@@ -1403,7 +1433,7 @@ sna_pixmap_move_area_to_gpu(PixmapPtr pixmap, BoxPtr box)
 
 		sna_damage_destroy(&priv->cpu_damage);
 		list_del(&priv->list);
-	} else if (DAMAGE_IS_ALL(priv->gpu_damage) ||
+	} else if (DAMAGE_IS_ALL(priv->cpu_damage) ||
 		   sna_damage_contains_box__no_reduce(priv->cpu_damage, box)) {
 		Bool ok = FALSE;
 		if (priv->cpu_bo)
@@ -1817,8 +1847,7 @@ sna_pixmap_move_to_gpu(PixmapPtr pixmap, unsigned flags)
 			    (box->x2 - box->x1) >= pixmap->drawable.width &&
 			    (box->y2 - box->y1) >= pixmap->drawable.height) {
 				priv->gpu_bo =
-					sna_replace(sna,
-						    pixmap,
+					sna_replace(sna, pixmap,
 						    priv->gpu_bo,
 						    pixmap->devPrivate.ptr,
 						    pixmap->devKind);
@@ -2609,6 +2638,9 @@ sna_put_image(DrawablePtr drawable, GCPtr gc, int depth,
 	if (wedged(sna))
 		goto fallback;
 
+	if (!ACCEL_PUT_IMAGE)
+		goto fallback;
+
 	switch (format) {
 	case ZPixmap:
 		if (!PM_IS_SOLID(drawable, gc->planemask))
@@ -2714,7 +2746,10 @@ sna_self_copy_boxes(DrawablePtr src, DrawablePtr dst, GCPtr gc,
 	if (dst != src)
 		get_drawable_deltas(dst, pixmap, &tx, &ty);
 
-	if (priv && priv->gpu_bo) {
+	if (priv == NULL || DAMAGE_IS_ALL(priv->cpu_damage))
+		goto fallback;
+
+	if (priv->gpu_bo) {
 		if (!sna_pixmap_move_to_gpu(pixmap, MOVE_WRITE | MOVE_READ)) {
 			DBG(("%s: fallback - not a pure copy and failed to move dst to GPU\n",
 			     __FUNCTION__));
@@ -3040,8 +3075,7 @@ sna_copy_boxes(DrawablePtr src, DrawablePtr dst, GCPtr gc,
 				assert(src_dx + box->x1 + dst_pixmap->drawable.width <= src_pixmap->drawable.width);
 
 				dst_priv->gpu_bo =
-					sna_replace(sna,
-						    dst_pixmap,
+					sna_replace(sna, dst_pixmap,
 						    dst_priv->gpu_bo,
 						    bits, stride);
 
@@ -3052,7 +3086,7 @@ sna_copy_boxes(DrawablePtr src, DrawablePtr dst, GCPtr gc,
 						       dst_pixmap->drawable.height);
 				}
 			} else {
-				DBG(("%s: dst is on the GPU, src is on the CPU, uploading\n",
+				DBG(("%s: dst is on the GPU, src is on the CPU, uploading into dst\n",
 				     __FUNCTION__));
 				sna_write_boxes(sna, dst_pixmap,
 						dst_priv->gpu_bo, dst_dx, dst_dy,
diff --git a/src/sna/sna_damage.c b/src/sna/sna_damage.c
index ea981573..6e456cf5 100644
--- a/src/sna/sna_damage.c
+++ b/src/sna/sna_damage.c
@@ -248,7 +248,7 @@ _sna_damage_create_elt(struct sna_damage *damage,
 	 return damage;
 }
 
-static void
+static struct sna_damage *
 _sna_damage_create_elt_from_boxes(struct sna_damage *damage,
 				  const BoxRec *boxes, int count,
 				  int16_t dx, int16_t dy)
@@ -273,13 +273,13 @@ _sna_damage_create_elt_from_boxes(struct sna_damage *damage,
 		count -=n;
 		boxes += n;
 		if (count == 0)
-			return;
+			return damage;
 	}
 
 	DBG(("    %s(): new elt\n", __FUNCTION__));
 
 	if (!_sna_damage_create_boxes(damage, count))
-		return;
+		return damage;
 
 	for (i = 0; i < count; i++) {
 		damage->box[i].x1 = boxes[i].x1 + dx;
@@ -289,9 +289,11 @@ _sna_damage_create_elt_from_boxes(struct sna_damage *damage,
 	}
 	damage->box += i;
 	damage->remain -= i;
+
+	return damage;
 }
 
-static void
+static struct sna_damage *
 _sna_damage_create_elt_from_rectangles(struct sna_damage *damage,
 				       const xRectangle *r, int count,
 				       int16_t dx, int16_t dy)
@@ -317,13 +319,13 @@ _sna_damage_create_elt_from_rectangles(struct sna_damage *damage,
 		count -=n;
 		r += n;
 		if (count == 0)
-			return;
+			return damage;
 	}
 
 	DBG(("    %s(): new elt\n", __FUNCTION__));
 
 	if (!_sna_damage_create_boxes(damage, count))
-		return;
+		return damage;
 
 	for (i = 0; i < count; i++) {
 		damage->box[i].x1 = r[i].x + dx;
@@ -333,9 +335,11 @@ _sna_damage_create_elt_from_rectangles(struct sna_damage *damage,
 	}
 	damage->box += n;
 	damage->remain -= n;
+
+	return damage;
 }
 
-static void
+static struct sna_damage *
 _sna_damage_create_elt_from_points(struct sna_damage *damage,
 				   const DDXPointRec *p, int count,
 				   int16_t dx, int16_t dy)
@@ -361,13 +365,13 @@ _sna_damage_create_elt_from_points(struct sna_damage *damage,
 		count -=n;
 		p += n;
 		if (count == 0)
-			return;
+			return damage;
 	}
 
 	DBG(("    %s(): new elt\n", __FUNCTION__));
 
 	if (! _sna_damage_create_boxes(damage, count))
-		return;
+		return damage;
 
 	for (i = 0; i < count; i++) {
 		damage->box[i].x1 = p[i].x + dx;
@@ -377,6 +381,8 @@ _sna_damage_create_elt_from_points(struct sna_damage *damage,
 	}
 	damage->box += count;
 	damage->remain -= count;
+
+	return damage;
 }
 
 static void free_list(struct list *head)
@@ -407,6 +413,21 @@ static void __sna_damage_reduce(struct sna_damage *damage)
 	nboxes -= damage->remain;
 	if (nboxes == 0)
 		goto done;
+	if (nboxes == 1) {
+		pixman_region16_t tmp;
+
+		tmp.extents = damage->embedded_box.box[0];
+		tmp.data = NULL;
+
+		if (damage->mode == DAMAGE_ADD)
+			pixman_region_union(region, region, &tmp);
+		else
+			pixman_region_subtract(region, region, &tmp);
+		damage->extents = region->extents;
+
+		goto done;
+	}
+
 	if (damage->mode == DAMAGE_ADD)
 		nboxes += REGION_NUM_RECTS(region);
 
@@ -496,6 +517,23 @@ done:
 	DBG(("    reduce: after region.n=%d\n", REGION_NUM_RECTS(region)));
 }
 
+static void damage_union(struct sna_damage *damage, const BoxRec *box)
+{
+	if (damage->extents.x2 < damage->extents.x1) {
+		damage->extents = *box;
+	} else {
+		if (damage->extents.x1 > box->x1)
+			damage->extents.x1 = box->x1;
+		if (damage->extents.x2 < box->x2)
+			damage->extents.x2 = box->x2;
+
+		if (damage->extents.y1 > box->y1)
+			damage->extents.y1 = box->y1;
+		if (damage->extents.y2 < box->y2)
+			damage->extents.y2 = box->y2;
+	}
+}
+
 inline static struct sna_damage *__sna_damage_add(struct sna_damage *damage,
 						  RegionPtr region)
 {
@@ -516,7 +554,7 @@ inline static struct sna_damage *__sna_damage_add(struct sna_damage *damage,
 
 	if (REGION_NUM_RECTS(&damage->region) <= 1) {
 		pixman_region_union(&damage->region, &damage->region, region);
-		damage->extents = damage->region.extents;
+		damage_union(damage, &region->extents);
 		return damage;
 	}
 
@@ -524,17 +562,7 @@ inline static struct sna_damage *__sna_damage_add(struct sna_damage *damage,
 					     &region->extents) == PIXMAN_REGION_IN)
 		return damage;
 
-
-	if (damage->extents.x1 > region->extents.x1)
-		damage->extents.x1 = region->extents.x1;
-	if (damage->extents.x2 < region->extents.x2)
-		damage->extents.x2 = region->extents.x2;
-
-	if (damage->extents.y1 > region->extents.y1)
-		damage->extents.y1 = region->extents.y1;
-	if (damage->extents.y2 < region->extents.y2)
-		damage->extents.y2 = region->extents.y2;
-
+	damage_union(damage, &region->extents);
 	return _sna_damage_create_elt(damage,
 				      REGION_RECTS(region),
 				      REGION_NUM_RECTS(region));
@@ -566,6 +594,50 @@ fastcall struct sna_damage *_sna_damage_add(struct sna_damage *damage,
 }
 #endif
 
+static void _pixman_region_union_box(RegionRec *region, const BoxRec *box)
+{
+	RegionRec u = { *box, NULL };
+	pixman_region_union(region, region, &u);
+}
+
+static struct sna_damage *__sna_damage_add_box(struct sna_damage *damage,
+					       const BoxRec *box)
+{
+	if (box->y2 <= box->y1 || box->x2 <= box->x1)
+		return damage;
+
+	if (!damage) {
+		damage = _sna_damage_create();
+		if (damage == NULL)
+			return NULL;
+	} else switch (damage->mode) {
+	case DAMAGE_ALL:
+		return damage;
+	case DAMAGE_SUBTRACT:
+		__sna_damage_reduce(damage);
+	case DAMAGE_ADD:
+		break;
+	}
+
+	switch (REGION_NUM_RECTS(&damage->region)) {
+	case 0:
+		pixman_region_init_rects(&damage->region, box, 1);
+		damage_union(damage, box);
+		return damage;
+	case 1:
+		_pixman_region_union_box(&damage->region, box);
+		damage_union(damage, box);
+		return damage;
+	}
+
+	if (pixman_region_contains_rectangle(&damage->region,
+					     (BoxPtr)box) == PIXMAN_REGION_IN)
+		return damage;
+
+	damage_union(damage, box);
+	return _sna_damage_create_elt(damage, box, 1);
+}
+
 inline static struct sna_damage *
 __sna_damage_add_boxes(struct sna_damage *damage,
 		       const BoxRec *box, int n,
@@ -608,29 +680,15 @@ __sna_damage_add_boxes(struct sna_damage *damage,
 	extents.y1 += dy;
 	extents.y2 += dy;
 
+	if (n == 1)
+		return __sna_damage_add_box(damage, &extents);
+
 	if (pixman_region_contains_rectangle(&damage->region,
 					     &extents) == PIXMAN_REGION_IN)
 		return damage;
 
-	_sna_damage_create_elt_from_boxes(damage, box, n, dx, dy);
-
-	if (REGION_NUM_RECTS(&damage->region) == 0) {
-		damage->region.extents = damage->embedded_box.box[0];
-		damage->region.data = NULL;
-		damage->extents = extents;
-	} else {
-		if (damage->extents.x1 > extents.x1)
-			damage->extents.x1 = extents.x1;
-		if (damage->extents.x2 < extents.x2)
-			damage->extents.x2 = extents.x2;
-
-		if (damage->extents.y1 > extents.y1)
-			damage->extents.y1 = extents.y1;
-		if (damage->extents.y2 < extents.y2)
-			damage->extents.y2 = extents.y2;
-	}
-
-	return damage;
+	damage_union(damage, &extents);
+	return _sna_damage_create_elt_from_boxes(damage, box, n, dx, dy);
 }
 
 #if DEBUG_DAMAGE
@@ -660,12 +718,6 @@ struct sna_damage *_sna_damage_add_boxes(struct sna_damage *damage,
 }
 #endif
 
-static void _pixman_region_union_box(RegionRec *region, const BoxRec *box)
-{
-	RegionRec u = { *box, NULL };
-	pixman_region_union(region, region, &u);
-}
-
 inline static struct sna_damage *
 __sna_damage_add_rectangles(struct sna_damage *damage,
 			    const xRectangle *r, int n,
@@ -698,6 +750,9 @@ __sna_damage_add_rectangles(struct sna_damage *damage,
 	extents.y1 += dy;
 	extents.y2 += dy;
 
+	if (n == 1)
+		return __sna_damage_add_box(damage, &extents);
+
 	if (!damage) {
 		damage = _sna_damage_create();
 		if (damage == NULL)
@@ -715,25 +770,8 @@ __sna_damage_add_rectangles(struct sna_damage *damage,
 					     &extents) == PIXMAN_REGION_IN)
 		return damage;
 
-	_sna_damage_create_elt_from_rectangles(damage, r, n, dx, dy);
-
-	if (REGION_NUM_RECTS(&damage->region) == 0) {
-		damage->region.extents = damage->embedded_box.box[0];
-		damage->region.data = NULL;
-		damage->extents = extents;
-	} else {
-		if (damage->extents.x1 > extents.x1)
-			damage->extents.x1 = extents.x1;
-		if (damage->extents.x2 < extents.x2)
-			damage->extents.x2 = extents.x2;
-
-		if (damage->extents.y1 > extents.y1)
-			damage->extents.y1 = extents.y1;
-		if (damage->extents.y2 < extents.y2)
-			damage->extents.y2 = extents.y2;
-	}
-
-	return damage;
+	damage_union(damage, &extents);
+	return _sna_damage_create_elt_from_rectangles(damage, r, n, dx, dy);
 }
 
 #if DEBUG_DAMAGE
@@ -792,6 +830,9 @@ __sna_damage_add_points(struct sna_damage *damage,
 	extents.y1 += dy;
 	extents.y2 += dy + 1;
 
+	if (n == 1)
+		return __sna_damage_add_box(damage, &extents);
+
 	if (!damage) {
 		damage = _sna_damage_create();
 		if (damage == NULL)
@@ -809,24 +850,9 @@ __sna_damage_add_points(struct sna_damage *damage,
 					     &extents) == PIXMAN_REGION_IN)
 		return damage;
 
+	damage_union(damage, &extents);
 	_sna_damage_create_elt_from_points(damage, p, n, dx, dy);
 
-	if (REGION_NUM_RECTS(&damage->region) == 0) {
-		damage->region.extents = damage->embedded_box.box[0];
-		damage->region.data = NULL;
-		damage->extents = extents;
-	} else {
-		if (damage->extents.x1 > extents.x1)
-			damage->extents.x1 = extents.x1;
-		if (damage->extents.x2 < extents.x2)
-			damage->extents.x2 = extents.x2;
-
-		if (damage->extents.y1 > extents.y1)
-			damage->extents.y1 = extents.y1;
-		if (damage->extents.y2 < extents.y2)
-			damage->extents.y2 = extents.y2;
-	}
-
 	return damage;
 }
 
@@ -857,53 +883,6 @@ struct sna_damage *_sna_damage_add_points(struct sna_damage *damage,
 }
 #endif
 
-inline static struct sna_damage *__sna_damage_add_box(struct sna_damage *damage,
-						      const BoxRec *box)
-{
-	if (box->y2 <= box->y1 || box->x2 <= box->x1)
-		return damage;
-
-	if (!damage) {
-		damage = _sna_damage_create();
-		if (damage == NULL)
-			return NULL;
-	} else switch (damage->mode) {
-	case DAMAGE_ALL:
-		return damage;
-	case DAMAGE_SUBTRACT:
-		__sna_damage_reduce(damage);
-	case DAMAGE_ADD:
-		break;
-	}
-
-	switch (REGION_NUM_RECTS(&damage->region)) {
-	case 0:
-		pixman_region_init_rects(&damage->region, box, 1);
-		damage->extents = *box;
-		return damage;
-	case 1:
-		_pixman_region_union_box(&damage->region, box);
-		damage->extents = damage->region.extents;
-		return damage;
-	}
-
-	if (pixman_region_contains_rectangle(&damage->region,
-					     (BoxPtr)box) == PIXMAN_REGION_IN)
-		return damage;
-
-	if (damage->extents.x1 > box->x1)
-		damage->extents.x1 = box->x1;
-	if (damage->extents.x2 < box->x2)
-		damage->extents.x2 = box->x2;
-
-	if (damage->extents.y1 > box->y1)
-		damage->extents.y1 = box->y1;
-	if (damage->extents.y2 < box->y2)
-		damage->extents.y2 = box->y2;
-
-	return _sna_damage_create_elt(damage, box, 1);
-}
-
 #if DEBUG_DAMAGE
 fastcall struct sna_damage *_sna_damage_add_box(struct sna_damage *damage,
 						const BoxRec *box)
@@ -971,12 +950,16 @@ struct sna_damage *_sna_damage_is_all(struct sna_damage *damage,
 	if (damage->region.data)
 		return damage;
 
+	DBG(("%s: (%d, %d), (%d, %d)\n", __FUNCTION__,
+	     damage->extents.x1, damage->extents.y1,
+	     damage->extents.x2, damage->extents.y2));
+
 	assert(damage->extents.x1 == 0 &&
 	       damage->extents.y1 == 0 &&
 	       damage->extents.x2 == width &&
 	       damage->extents.y2 == height);
 
-	return _sna_damage_all(damage, width, height);
+	return __sna_damage_all(damage, width, height);
 }
 
 static bool box_contains(const BoxRec *a, const BoxRec *b)
@@ -1213,16 +1196,15 @@ int _sna_damage_contains_box(struct sna_damage *damage,
 bool _sna_damage_contains_box__no_reduce(const struct sna_damage *damage,
 					 const BoxRec *box)
 {
-	int ret;
-
 	assert(damage && damage->mode != DAMAGE_ALL);
+	if (damage->mode == DAMAGE_SUBTRACT)
+		return false;
+
 	if (!sna_damage_maybe_contains_box(damage, box))
 		return false;
 
-	ret = pixman_region_contains_rectangle((RegionPtr)&damage->region,
-					       (BoxPtr)box);
-	return (!damage->dirty || damage->mode == DAMAGE_ADD) &&
-		ret == PIXMAN_REGION_IN;
+	return pixman_region_contains_rectangle((RegionPtr)&damage->region,
+						(BoxPtr)box) == PIXMAN_REGION_IN;
 }
 
 static Bool __sna_damage_intersect(struct sna_damage *damage,
diff --git a/src/sna/sna_render.c b/src/sna/sna_render.c
index 142c1aa9..304ff0f6 100644
--- a/src/sna/sna_render.c
+++ b/src/sna/sna_render.c
@@ -69,7 +69,7 @@ no_render_composite(struct sna *sna,
 		    int16_t width, int16_t height,
 		    struct sna_composite_op *tmp)
 {
-	DBG(("%s ()\n", __FUNCTION__));
+	DBG(("%s (op=%d, mask? %d)\n", __FUNCTION__, op, mask != NULL));
 
 	if (mask == NULL &&
 	    sna_blt_composite(sna,