diff --git a/src/r600_exa.c b/src/r600_exa.c
index f722f2d..64b7dcb 100644
--- a/src/r600_exa.c
+++ b/src/r600_exa.c
@@ -716,16 +716,13 @@ R600PrepareCopy(PixmapPtr pSrc,   PixmapPtr pDst,
 	ErrorF("same surface!\n");
 #endif
     } else {
-
 	accel_state->same_surface = FALSE;
-
 	R600DoPrepareCopy(pScrn,
 			  accel_state->src_pitch[0], pSrc->drawable.width, pSrc->drawable.height,
 			  accel_state->src_mc_addr[0], pSrc->drawable.bitsPerPixel,
 			  accel_state->dst_pitch, pDst->drawable.height,
 			  accel_state->dst_mc_addr, pDst->drawable.bitsPerPixel,
 			  rop, planemask);
-
     }
 
     return TRUE;
@@ -744,136 +741,6 @@ is_overlap(int sx1, int sx2, int sy1, int sy2, int dx1, int dx2, int dy1, int dy
 }
 
 static void
-R600OverlapCopy(PixmapPtr pDst,
-		int srcX, int srcY,
-		int dstX, int dstY,
-		int w, int h)
-{
-    ScrnInfoPtr pScrn = xf86Screens[pDst->drawable.pScreen->myNum];
-    RHDPtr rhdPtr = RHDPTR(pScrn);
-    struct r6xx_accel_state *accel_state = rhdPtr->TwoDPrivate;
-    uint32_t dst_pitch = exaGetPixmapPitch(pDst) / (pDst->drawable.bitsPerPixel / 8);
-    uint32_t dst_offset = exaGetPixmapOffset(pDst) + rhdPtr->FbIntAddress + rhdPtr->FbScanoutStart;
-    int i, hchunk, vchunk;
-
-    if (is_overlap(srcX, srcX + w, srcY, srcY + h,
-		   dstX, dstX + w, dstY, dstY + h)) {
-        /* Calculate height/width of non-overlapping area */
-        hchunk = (srcX < dstX) ? (dstX - srcX) : (srcX - dstX);
-        vchunk = (srcY < dstY) ? (dstY - srcY) : (srcY - dstY);
-
-        /* Diagonally offset overlap is reduced to either horizontal or vertical offset-only
-         * by copying a part of the  non-overlapping portion, then adjusting coordinates
-         * Choose horizontal vs vertical to minimize the total number of copy operations
-         */
-        if (vchunk != 0 && hchunk != 0) { //diagonal
-            if ((w / hchunk) <= (h / vchunk)) { // reduce to horizontal
-                if (srcY > dstY ) { // diagonal up
-                    R600DoPrepareCopy(pScrn,
-                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      accel_state->rop, accel_state->planemask);
-                    R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, w, vchunk);
-                    R600DoCopy(pScrn);
-
-                    srcY = srcY + vchunk;
-                    dstY = dstY + vchunk;
-                } else { // diagonal down
-                    R600DoPrepareCopy(pScrn,
-                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      accel_state->rop, accel_state->planemask);
-                    R600AppendCopyVertex(pScrn, srcX, srcY + h - vchunk, dstX, dstY + h - vchunk, w, vchunk);
-                    R600DoCopy(pScrn);
-                }
-                h = h - vchunk;
-                vchunk = 0;
-            } else { //reduce to vertical
-                if (srcX > dstX ) { // diagonal left
-                    R600DoPrepareCopy(pScrn,
-                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      accel_state->rop, accel_state->planemask);
-                    R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, hchunk, h);
-                    R600DoCopy(pScrn);
-
-                    srcX = srcX + hchunk;
-                    dstX = dstX + hchunk;
-                } else { // diagonal right
-                    R600DoPrepareCopy(pScrn,
-                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      accel_state->rop, accel_state->planemask);
-                    R600AppendCopyVertex(pScrn, srcX + w - hchunk, srcY, dstX + w - hchunk, dstY, hchunk, h);
-                    R600DoCopy(pScrn);
-                }
-                w = w - hchunk;
-                hchunk = 0;
-            }
-        }
-
-	if (vchunk == 0) { // left/right
-	    if (srcX < dstX) { // right
-		// copy right to left
-		for (i = w; i > 0; i -= hchunk) {
-		    R600DoPrepareCopy(pScrn,
-				      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      accel_state->rop, accel_state->planemask);
-		    R600AppendCopyVertex(pScrn, srcX + i - hchunk, srcY, dstX + i - hchunk, dstY, hchunk, h);
-		    R600DoCopy(pScrn);
-		}
-	    } else { //left
-		// copy left to right
-		for (i = 0; i < w; i += hchunk) {
-		    R600DoPrepareCopy(pScrn,
-				      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-				      accel_state->rop, accel_state->planemask);
-
-		    R600AppendCopyVertex(pScrn, srcX + i, srcY, dstX + i, dstY, hchunk, h);
-		    R600DoCopy(pScrn);
-		}
-	    }
-	} else { //up/down
-	    if (srcY > dstY) { // up
-		// copy top to bottom
-                for (i = 0; i < h; i += vchunk) {
-                    R600DoPrepareCopy(pScrn,
-                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      accel_state->rop, accel_state->planemask);
-
-                    if (vchunk > h - i) vchunk = h - i;
-                    R600AppendCopyVertex(pScrn, srcX, srcY + i, dstX, dstY + i, w, vchunk);
-                    R600DoCopy(pScrn);
-                }
-	    } else { // down
-		// copy bottom to top
-                for (i = h; i > 0; i -= vchunk) {
-                    R600DoPrepareCopy(pScrn,
-                                      dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-                                      accel_state->rop, accel_state->planemask);
-
-                    if (vchunk > i) vchunk = i;
-                    R600AppendCopyVertex(pScrn, srcX, srcY + i - vchunk, dstX, dstY + i - vchunk, w, vchunk);
-                    R600DoCopy(pScrn);
-                }
-            }
-	}
-    } else {
-	R600DoPrepareCopy(pScrn,
-			  dst_pitch, pDst->drawable.width, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-			  dst_pitch, pDst->drawable.height, dst_offset, pDst->drawable.bitsPerPixel,
-			  accel_state->rop, accel_state->planemask);
-
-	R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, w, h);
-	R600DoCopy(pScrn);
-    }
-}
-
-static void
 R600Copy(PixmapPtr pDst,
 	 int srcX, int srcY,
 	 int dstX, int dstY,
@@ -883,11 +750,43 @@ R600Copy(PixmapPtr pDst,
     RHDPtr rhdPtr = RHDPTR(pScrn);
     struct r6xx_accel_state *accel_state = rhdPtr->TwoDPrivate;
 
-    //blit to/from same surfacce
-    if (accel_state->same_surface)
-	R600OverlapCopy(pDst, srcX, srcY, dstX, dstY, w, h);
-    else
+    if (accel_state->same_surface && is_overlap(srcX, srcX + w, srcY, srcY + h, dstX, dstX + w, dstY, dstY + h)) {
+	uint32_t pitch = exaGetPixmapPitch(pDst) / (pDst->drawable.bitsPerPixel / 8);
+	uint32_t orig_offset, tmp_offset;
+
+	if(!(accel_state->copy_area)) {
+	    unsigned long size=pDst->drawable.height*pitch*pDst->drawable.bitsPerPixel/8;
+	    accel_state->copy_area=exaOffscreenAlloc(pDst->drawable.pScreen, size, 256, TRUE, NULL, NULL);
+	}
+
+	tmp_offset = accel_state->copy_area->offset + rhdPtr->FbIntAddress + rhdPtr->FbScanoutStart;
+	orig_offset = exaGetPixmapOffset(pDst) + rhdPtr->FbIntAddress + rhdPtr->FbScanoutStart;
+
+	R600DoPrepareCopy(pScrn,
+			  pitch, pDst->drawable.width, pDst->drawable.height, orig_offset, pDst->drawable.bitsPerPixel,
+			  pitch,                       pDst->drawable.height, tmp_offset, pDst->drawable.bitsPerPixel,
+			  accel_state->rop, accel_state->planemask);
 	R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, w, h);
+	R600DoCopy(pScrn); 
+	R600DoPrepareCopy(pScrn,
+			  pitch, pDst->drawable.width, pDst->drawable.height, tmp_offset, pDst->drawable.bitsPerPixel,
+			  pitch,                       pDst->drawable.height, orig_offset, pDst->drawable.bitsPerPixel,
+			  accel_state->rop, accel_state->planemask);
+	R600AppendCopyVertex(pScrn, dstX, dstY, dstX, dstY, w, h);
+	R600DoCopy(pScrn); 
+    } else if(accel_state->same_surface) {
+	uint32_t pitch = exaGetPixmapPitch(pDst) / (pDst->drawable.bitsPerPixel / 8);
+	uint32_t offset = exaGetPixmapOffset(pDst) + rhdPtr->FbIntAddress + rhdPtr->FbScanoutStart;
+
+	R600DoPrepareCopy(pScrn,
+			  pitch, pDst->drawable.width, pDst->drawable.height, offset, pDst->drawable.bitsPerPixel,
+			  pitch,                       pDst->drawable.height, offset, pDst->drawable.bitsPerPixel,
+			  accel_state->rop, accel_state->planemask);
+	R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, w, h);
+	R600DoCopy(pScrn); 
+    } else {
+	R600AppendCopyVertex(pScrn, srcX, srcY, dstX, dstY, w, h);
+    }
 }
 
 static void
@@ -897,10 +796,13 @@ R600DoneCopy(PixmapPtr pDst)
     RHDPtr rhdPtr = RHDPTR(pScrn);
     struct r6xx_accel_state *accel_state = rhdPtr->TwoDPrivate;
 
-    if (accel_state->same_surface)
-	return;
-    else
+    if(!(accel_state->same_surface))
 	R600DoCopy(pScrn);
+
+    if(accel_state->copy_area) {
+	exaOffscreenFree(pDst->drawable.pScreen, accel_state->copy_area);
+	accel_state->copy_area=NULL;
+    }
 }
 
 #define RADEON_TRACE_FALL 0
@@ -2365,7 +2267,6 @@ R600LoadShaders(ScrnInfoPtr pScrn, ScreenPtr pScreen)
 
     accel_state->shaders = exaOffscreenAlloc(pScreen, size, 256,
 					     TRUE, NULL, NULL);
-
     if (accel_state->shaders == NULL)
 	return FALSE;
 
diff --git a/src/r6xx_accel.h b/src/r6xx_accel.h
index 51e3109..eb6a06d 100644
--- a/src/r6xx_accel.h
+++ b/src/r6xx_accel.h
@@ -91,6 +91,7 @@ struct r6xx_accel_state {
     drmBufPtr         scratch;
 
     // copy
+    ExaOffscreenArea  *copy_area;
     Bool              same_surface;
     int               rop;
     uint32_t          planemask;