Index: src/radeon.h
===================================================================
RCS file: /cvs/xorg/driver/xf86-video-ati/src/radeon.h,v
retrieving revision 1.42
diff -p -u -r1.42 radeon.h
--- src/radeon.h	26 Apr 2006 08:38:47 -0000	1.42
+++ src/radeon.h	30 Apr 2006 11:00:45 -0000
@@ -799,9 +807,12 @@ extern void        RADEONCPFlushIndirect
 extern void        RADEONCPReleaseIndirect(ScrnInfoPtr pScrn);
 extern int         RADEONCPStop(ScrnInfoPtr pScrn,  RADEONInfoPtr info);
 
-extern CARD8*      RADEONHostDataBlit(ScrnInfoPtr pScrn, unsigned int bpp,
-				      unsigned int w, CARD32 dstPitch,
-				      CARD32 *bufPitch, CARD8 **dst,
+extern void        RADEONHostDataParams(ScrnInfoPtr pScrn, CARD8 *dst,
+					CARD32 pitch, int cpp,
+					CARD32 *dstPitchOffset, int *x, int *y);
+extern CARD8*      RADEONHostDataBlit(ScrnInfoPtr pScrn, unsigned int cpp,
+				      unsigned int w, CARD32 dstPitchOff,
+				      CARD32 *bufPitch, int x, int *y,
 				      unsigned int *h, unsigned int *hpass);
 extern void        RADEONHostDataBlitCopyPass(ScrnInfoPtr pScrn,
 					      unsigned int bpp,
Index: src/radeon_accel.c
===================================================================
RCS file: /cvs/xorg/driver/xf86-video-ati/src/radeon_accel.c,v
retrieving revision 1.25
diff -p -u -r1.25 radeon_accel.c
--- src/radeon_accel.c	28 Apr 2006 02:55:08 -0000	1.25
+++ src/radeon_accel.c	30 Apr 2006 11:00:45 -0000
@@ -604,6 +604,22 @@ void RADEONCPReleaseIndirect(ScrnInfoPtr
 			&indirect, sizeof(drmRadeonIndirect));
 }
 
+/** \brief Calculate HostDataBlit parameters from pointer and pitch
+ *
+ * This is a helper for the trivial HostDataBlit users that don't need to worry
+ * about tiling etc.
+ */
+void
+RADEONHostDataParams(ScrnInfoPtr pScrn, CARD8 *dst, CARD32 pitch, int cpp,
+		     CARD32 *dstPitchOff, int *x, int *y)
+{
+    RADEONInfoPtr info = RADEONPTR( pScrn );
+    CARD32 dstOffs = dst - info->FB + info->fbLocation;
+
+    *dstPitchOff = pitch << 16 | dstOffs >> 10;
+    *y = ( dstOffs & RADEON_BUFFER_ALIGN ) / pitch;
+    *x = ( ( dstOffs & RADEON_BUFFER_ALIGN ) - ( *y * pitch ) ) / cpp;
+}
 
 /* Set up a hostdata blit to transfer data from system memory to the
  * framebuffer. Returns the address where the data can be written to and sets
@@ -612,16 +628,17 @@ void RADEONCPReleaseIndirect(ScrnInfoPtr
 CARD8*
 RADEONHostDataBlit(
     ScrnInfoPtr pScrn,
-    unsigned int bpp,
+    unsigned int cpp,
     unsigned int w,
-    CARD32 dstPitch,
+    CARD32 dstPitchOff,
     CARD32 *bufPitch,
-    CARD8* *dst,
+    int x,
+    int *y,
     unsigned int *h,
     unsigned int *hpass
 ){
     RADEONInfoPtr info = RADEONPTR( pScrn );
-    CARD32 format, dst_offs, dwords, x, y;
+    CARD32 format, dwords;
     CARD8 *ret;
     RING_LOCALS;
 
@@ -630,7 +647,7 @@ RADEONHostDataBlit(
 	return NULL;
     }
 
-    switch ( bpp )
+    switch ( cpp )
     {
     case 4:
 	format = RADEON_GMC_DST_32BPP;
@@ -638,17 +655,15 @@ RADEONHostDataBlit(
 	break;
     case 2:
 	format = RADEON_GMC_DST_16BPP;
-	w = (w + 1) & ~1;
-	*bufPitch = 2 * w;
+	*bufPitch = 2 * ((w + 1) & ~1);
 	break;
     case 1:
 	format = RADEON_GMC_DST_8BPP_CI;
-	w = (w + 3) & ~3;
-	*bufPitch = w;
+	*bufPitch = (w + 3) & ~3;
 	break;
     default:
 	xf86DrvMsg( pScrn->scrnIndex, X_ERROR,
-		    "%s: Unsupported bpp %d!\n", __func__, bpp );
+		    "%s: Unsupported cpp %d!\n", __func__, cpp );
 	return NULL;
     }
 
@@ -658,10 +673,10 @@ RADEONHostDataBlit(
      */
     if (info->ChipFamily < CHIP_FAMILY_R300) {
         BEGIN_RING(2);
-	if (bpp == 2)
+	if (cpp == 2)
 	    OUT_RING_REG(RADEON_RBBM_GUICNTL,
 			 RADEON_HOST_DATA_SWAP_HDW);
-	else if (bpp == 1)
+	else if (cpp == 1)
 	    OUT_RING_REG(RADEON_RBBM_GUICNTL,
 			 RADEON_HOST_DATA_SWAP_32BIT);
 	else
@@ -674,16 +689,13 @@ RADEONHostDataBlit(
     /*RADEON_PURGE_CACHE();
       RADEON_WAIT_UNTIL_IDLE();*/
 
-    dst_offs = *dst - info->FB + info->fbLocation;
-    *hpass = min( *h, ( ( RADEON_BUFFER_SIZE - 8 * 4 ) / *bufPitch ) );
+    *hpass = min( *h, ( ( RADEON_BUFFER_SIZE - 10 * 4 ) / *bufPitch ) );
     dwords = *hpass * *bufPitch / 4;
 
-    y = ( dst_offs & 1023 ) / dstPitch;
-    x = ( ( dst_offs & 1023 ) - ( y * dstPitch ) ) / bpp;
-
-    BEGIN_RING( dwords + 8 );
-    OUT_RING( CP_PACKET3( RADEON_CP_PACKET3_CNTL_HOSTDATA_BLT, dwords + 8 - 2 ) );
+    BEGIN_RING( dwords + 10 );
+    OUT_RING( CP_PACKET3( RADEON_CP_PACKET3_CNTL_HOSTDATA_BLT, dwords + 10 - 2 ) );
     OUT_RING( RADEON_GMC_DST_PITCH_OFFSET_CNTL
+	    | RADEON_GMC_DST_CLIPPING
 	    | RADEON_GMC_BRUSH_NONE
 	    | format
 	    | RADEON_GMC_SRC_DATATYPE_COLOR
@@ -691,11 +703,13 @@ RADEONHostDataBlit(
 	    | RADEON_DP_SRC_SOURCE_HOST_DATA
 	    | RADEON_GMC_CLR_CMP_CNTL_DIS
 	    | RADEON_GMC_WR_MSK_DIS );
-    OUT_RING( dstPitch << 16 | dst_offs >> 10 );
+    OUT_RING( dstPitchOff );
+    OUT_RING( (*y << 16) | x );
+    OUT_RING( ((*y + *hpass) << 16) | (x + w) );
     OUT_RING( 0xffffffff );
     OUT_RING( 0xffffffff );
-    OUT_RING( y << 16 | x );
-    OUT_RING( *hpass << 16 | w );
+    OUT_RING( *y << 16 | x );
+    OUT_RING( *hpass << 16 | (*bufPitch / cpp) );
     OUT_RING( dwords );
 
     ret = ( CARD8* )&__head[__count];
@@ -703,7 +717,7 @@ RADEONHostDataBlit(
     __count += dwords;
     ADVANCE_RING();
 
-    *dst += *hpass * dstPitch;
+    *y += *hpass;
     *h -= *hpass;
 
     return ret;
@@ -763,7 +777,7 @@ void RADEONCopySwap(CARD8 *dst, CARD8 *s
 void
 RADEONHostDataBlitCopyPass(
     ScrnInfoPtr pScrn,
-    unsigned int bpp,
+    unsigned int cpp,
     CARD8 *dst,
     CARD8 *src,
     unsigned int hpass,
@@ -780,7 +794,7 @@ RADEONHostDataBlitCopyPass(
     {
 #if X_BYTE_ORDER == X_BIG_ENDIAN
         if (info->ChipFamily >= CHIP_FAMILY_R300) {
-	    switch(bpp) {
+	    switch(cpp) {
 	    case 1:
 		RADEONCopySwap(dst, src, hpass * dstPitch,
 			       RADEON_HOST_DATA_SWAP_32BIT);
@@ -801,7 +815,7 @@ RADEONHostDataBlitCopyPass(
 	{
 #if X_BYTE_ORDER == X_BIG_ENDIAN
             if (info->ChipFamily >= CHIP_FAMILY_R300) {
-		switch(bpp) {
+		switch(cpp) {
 		case 1:
 		    RADEONCopySwap(dst, src, minPitch,
 				   RADEON_HOST_DATA_SWAP_32BIT);
Index: src/radeon_exa_funcs.c
===================================================================
RCS file: /cvs/xorg/driver/xf86-video-ati/src/radeon_exa_funcs.c,v
retrieving revision 1.5
diff -p -u -r1.5 radeon_exa_funcs.c
--- src/radeon_exa_funcs.c	22 Mar 2006 22:30:14 -0000	1.5
+++ src/radeon_exa_funcs.c	30 Apr 2006 11:00:46 -0000
@@ -219,7 +217,7 @@ FUNC_NAME(RADEONUploadToScreen)(PixmapPt
     unsigned int   bpp	     = pDst->drawable.bitsPerPixel;
 #ifdef ACCEL_CP
     unsigned int   hpass;
-    CARD32	   buf_pitch;
+    CARD32	   buf_pitch, dst_pitch_off;
 #endif
 #if X_BYTE_ORDER == X_BIG_ENDIAN 
     unsigned char *RADEONMMIO = info->MMIO;
@@ -234,17 +232,17 @@ FUNC_NAME(RADEONUploadToScreen)(PixmapPt
 	return FALSE;
 
 #ifdef ACCEL_CP
-    if (info->directRenderingEnabled) {
+    if (info->directRenderingEnabled &&
+	RADEONGetPixmapOffsetPitch(pDst, &dst_pitch_off)) {
 	CARD8 *buf;
 	int cpp = bpp / 8;
 	ACCEL_PREAMBLE();
 
-	dst += (x * cpp) + (y * dst_pitch);
 	RADEON_SWITCH_TO_2D();
 	while ((buf = RADEONHostDataBlit(pScrn,
-					cpp, w, dst_pitch, &buf_pitch,
-					&dst, &h, &hpass)) != 0) {
-	    RADEONHostDataBlitCopyPass(pScrn, cpp, buf, (unsigned char *)src,
+					 cpp, w, dst_pitch_off, &buf_pitch,
+					 x, &y, (unsigned int*)&h, &hpass)) != 0) {
+	    RADEONHostDataBlitCopyPass(pScrn, cpp, buf, (CARD8 *)src,
 				       hpass, buf_pitch, src_pitch);
 	    src += hpass * src_pitch;
 	}
Index: src/radeon_render.c
===================================================================
RCS file: /cvs/xorg/driver/xf86-video-ati/src/radeon_render.c,v
retrieving revision 1.19
diff -p -u -r1.19 radeon_render.c
--- src/radeon_render.c	22 Mar 2006 22:30:14 -0000	1.19
+++ src/radeon_render.c	30 Apr 2006 11:00:47 -0000
@@ -383,9 +383,10 @@ static Bool FUNC_NAME(R100SetupTexture)(
     RADEONInfoPtr info = RADEONPTR(pScrn);
     CARD8 *dst;
     CARD32 tex_size = 0, txformat;
-    int dst_pitch, offset, size, i, tex_bytepp;
+    int dst_pitch, offset, size, tex_bytepp;
 #ifdef ACCEL_CP
-    CARD32 buf_pitch;
+    CARD32 buf_pitch, dst_pitch_off;
+    int x, y;
     unsigned int hpass;
     CARD8 *tmp_dst;
 #endif
@@ -430,11 +431,13 @@ static Bool FUNC_NAME(R100SetupTexture)(
 
 #ifdef ACCEL_CP
 
+    RADEONHostDataParams( pScrn, dst, dst_pitch, tex_bytepp, &dst_pitch_off, &x, &y );
+
     while ( height )
     {
     	tmp_dst = RADEONHostDataBlit( pScrn, tex_bytepp, width,
-				      dst_pitch, &buf_pitch,
-				      &dst, &height, &hpass);
+				      dst_pitch_off, &buf_pitch,
+				      x, &y, &height, &hpass );
 	RADEONHostDataBlitCopyPass( pScrn, tex_bytepp, tmp_dst, src,
 				    hpass, buf_pitch, src_pitch );
 	src += hpass * src_pitch;
@@ -445,12 +448,10 @@ static Bool FUNC_NAME(R100SetupTexture)(
 
 #else
 
-    i = height;
-
     if (info->accel->NeedToSync)
 	info->accel->Sync(pScrn);
 
-    while(i--) {
+    while (height--) {
 	memcpy(dst, src, width * tex_bytepp);
 	src += src_pitch;
 	dst += dst_pitch;
@@ -715,9 +716,10 @@ static Bool FUNC_NAME(R200SetupTexture)(
     RADEONInfoPtr info = RADEONPTR(pScrn);
     CARD8 *dst;
     CARD32 tex_size = 0, txformat;
-    int dst_pitch, offset, size, i, tex_bytepp;
+    int dst_pitch, offset, size, tex_bytepp;
 #ifdef ACCEL_CP
-    CARD32 buf_pitch;
+    CARD32 buf_pitch, dst_pitch_off;
+    int x, y;
     unsigned int hpass;
     CARD8 *tmp_dst;
 #endif
@@ -762,11 +764,13 @@ static Bool FUNC_NAME(R200SetupTexture)(
 
 #ifdef ACCEL_CP
 
+    RADEONHostDataParams( pScrn, dst, dst_pitch, tex_bytepp, &dst_pitch_off, &x, &y );
+
     while ( height )
     {
         tmp_dst = RADEONHostDataBlit( pScrn, tex_bytepp, width,
-				      dst_pitch, &buf_pitch,
-				      &dst, &height, &hpass );
+				      dst_pitch_off, &buf_pitch,
+				      x, &y, &height, &hpass );
 	RADEONHostDataBlitCopyPass( pScrn, tex_bytepp, tmp_dst, src,
 				    hpass, buf_pitch, src_pitch );
 	src += hpass * src_pitch;
@@ -777,11 +781,10 @@ static Bool FUNC_NAME(R200SetupTexture)(
 
 #else
 
-    i = height;
     if (info->accel->NeedToSync)
 	info->accel->Sync(pScrn);
 
-    while(i--) {
+    while (height--) {
 	memcpy(dst, src, width * tex_bytepp);
 	src += src_pitch;
 	dst += dst_pitch;
Index: src/radeon_video.c
===================================================================
RCS file: /cvs/xorg/driver/xf86-video-ati/src/radeon_video.c,v
retrieving revision 1.33
diff -p -u -r1.33 radeon_video.c
--- src/radeon_video.c	7 Apr 2006 23:06:02 -0000	1.33
+++ src/radeon_video.c	30 Apr 2006 11:00:48 -0000
@@ -2048,7 +2048,8 @@ RADEONCopyData(
     if ( info->directRenderingEnabled && info->DMAForXv )
     {
 	CARD8 *buf;
-	CARD32 bufPitch;
+	CARD32 bufPitch, dstPitchOff;
+	int x, y;
 	unsigned int hpass;
 
 	/* Get the byte-swapping right for big endian systems */
@@ -2058,8 +2059,10 @@ RADEONCopyData(
 	    bpp = 1;
 	}
 
-	while ( buf = RADEONHostDataBlit( pScrn, bpp, w, dstPitch,
-					  &bufPitch, &dst, &h, &hpass ) )
+	RADEONHostDataParams( pScrn, dst, dstPitch, bpp, &dstPitchOff, &x, &y );
+
+	while ( (buf = RADEONHostDataBlit( pScrn, bpp, w, dstPitchOff, &bufPitch,
+					   x, &y, &h, &hpass )) )
 	{
 	    RADEONHostDataBlitCopyPass( pScrn, bpp, buf, src, hpass, bufPitch,
 					srcPitch );
@@ -2140,14 +2143,17 @@ RADEONCopyRGB24Data(
 
     if ( info->directRenderingEnabled && info->DMAForXv )
     {
-	CARD32 bufPitch;
+	CARD32 bufPitch, dstPitchOff;
+	int x, y;
 	unsigned int hpass;
 
 	/* XXX Fix endian flip on R300 */
 
-	while ( dptr = ( CARD32* )RADEONHostDataBlit( pScrn, 4, w, dstPitch,
-						      &bufPitch, &dst, &h,
-						      &hpass ) )
+	RADEONHostDataParams( pScrn, dst, dstPitch, 4, &dstPitchOff, &x, &y );
+
+	while ( (dptr = ( CARD32* )RADEONHostDataBlit( pScrn, 4, w, dstPitch,
+						       &bufPitch, x, &y, &h,
+						       &hpass )) )
 	{
 	    for( j = 0; j < hpass; j++ )
 	    {
@@ -2215,13 +2221,16 @@ RADEONCopyMungedData(
     if ( info->directRenderingEnabled && info->DMAForXv )
     {
 	CARD8 *buf;
-	CARD32 y = 0, bufPitch;
+	CARD32 y = 0, bufPitch, dstPitchOff;
+	int blitX, blitY;
 	unsigned int hpass;
 
 	/* XXX Fix endian flip on R300 */
 
-	while ( buf = RADEONHostDataBlit( pScrn, 4, w/2, dstPitch,
-					  &bufPitch, &dst1, &h, &hpass ) )
+	RADEONHostDataParams( pScrn, dst1, dstPitch, 4, &dstPitchOff, &blitX, &blitY );
+
+	while ( (buf = RADEONHostDataBlit( pScrn, 4, w/2, dstPitchOff, &bufPitch,
+					   blitX, &blitY, &h, &hpass )) )
 	{
 	    while ( hpass-- )
 	    {