diff --git a/src/radeon_exa_render.c b/src/radeon_exa_render.c
index 4736e4f..d4cd77c 100644
--- a/src/radeon_exa_render.c
+++ b/src/radeon_exa_render.c
@@ -824,7 +824,8 @@ static Bool FUNC_NAME(R200PrepareComposite)(int op, PicturePtr pSrcPicture,
 {
     RINFO_FROM_SCREEN(pDst->drawable.pScreen);
     uint32_t dst_format, dst_offset, dst_pitch;
-    uint32_t pp_cntl, blendcntl, cblend, ablend, colorpitch;
+    uint32_t pp_cntl, blendcntl, colorpitch;
+    uint32_t src_color, src_alpha, mask_color, mask_alpha;
     int pixel_shift;
     ACCEL_PREAMBLE();
 
@@ -896,37 +897,82 @@ static Bool FUNC_NAME(R200PrepareComposite)(int op, PicturePtr pSrcPicture,
      * be the source alpha (and there's no source value used), we have to zero
      * the source's color values.
      */
-    cblend = R200_TXC_OP_MADD | R200_TXC_ARG_C_ZERO;
-    ablend = R200_TXA_OP_MADD | R200_TXA_ARG_C_ZERO;
 
-    if (pDstPicture->format == PICT_a8 ||
-	(pMask && pMaskPicture->componentAlpha && RadeonBlendOp[op].src_alpha))
-    {
-	cblend |= R200_TXC_ARG_A_R0_ALPHA;
-    } else if (pSrcPicture->format == PICT_a8)
-	cblend |= R200_TXC_ARG_A_ZERO;
+    if (PICT_FORMAT_RGB(pSrcPicture->format) == 0)
+	src_color = R200_TXC_ARG_A_ZERO;
     else
-	cblend |= R200_TXC_ARG_A_R0_COLOR;
-    ablend |= R200_TXA_ARG_A_R0_ALPHA;
+	src_color = R200_TXC_ARG_A_R0_COLOR;
 
-    if (pMask) {
-	if (pMaskPicture->componentAlpha &&
-	    pDstPicture->format != PICT_a8)
-	    cblend |= R200_TXC_ARG_B_R1_COLOR;
+    if (PICT_FORMAT_A(pSrcPicture->format) == 0)
+	src_alpha = R200_TXA_ARG_A_ZERO | R200_TXA_COMP_ARG_A;
+    else
+	src_alpha = R200_TXA_ARG_A_R0_ALPHA;
+
+    if (pMask && pMaskPicture->componentAlpha) {
+	if (RadeonBlendOp[op].src_alpha) {
+	    if (PICT_FORMAT_A(pSrcPicture->format) == 0) {
+		src_color = R200_TXC_ARG_A_ZERO | R200_TXC_COMP_ARG_A;
+		src_alpha = R200_TXA_ARG_A_ZERO | R200_TXA_COMP_ARG_A;
+	    } else {
+		src_color = R200_TXC_ARG_A_R0_ALPHA;
+		src_alpha = R200_TXA_ARG_A_R0_ALPHA;
+	    }
+
+	    mask_color = R200_TXC_ARG_B_R1_COLOR;
+
+	    if (PICT_FORMAT_A(pMaskPicture->format) == 0)
+		mask_alpha = R200_TXA_ARG_B_ZERO | R200_TXA_COMP_ARG_B;
+	    else
+		mask_alpha = R200_TXA_ARG_B_R1_ALPHA;
+
+	} else {
+	    src_color = R200_TXC_ARG_A_R0_COLOR;
+
+	    if (PICT_FORMAT_A(pSrcPicture->format) == 0)
+		src_alpha = R200_TXA_ARG_A_ZERO | R200_TXA_COMP_ARG_A;
+	    else
+		src_alpha = R200_TXA_ARG_A_R0_ALPHA;
+
+	    mask_color = R200_TXC_ARG_B_R1_COLOR;
+
+	    if (PICT_FORMAT_A(pMaskPicture->format) == 0)
+		mask_alpha = R200_TXA_ARG_B_ZERO | R200_TXA_COMP_ARG_B;
+	    else
+		mask_alpha = R200_TXA_ARG_B_R1_ALPHA;
+
+	}
+    } else if (pMask) {
+	if (PICT_FORMAT_A(pMaskPicture->format) == 0)
+	    mask_color = R200_TXC_ARG_B_ZERO | R200_TXC_COMP_ARG_B;
 	else
-	    cblend |= R200_TXC_ARG_B_R1_ALPHA;
-	ablend |= R200_TXA_ARG_B_R1_ALPHA;
+	    mask_color = R200_TXC_ARG_B_R1_ALPHA;
+
+	if (PICT_FORMAT_A(pMaskPicture->format) == 0)
+	    mask_alpha = R200_TXA_ARG_B_ZERO | R200_TXA_COMP_ARG_B;
+	else
+	    mask_alpha = R200_TXA_ARG_B_R1_ALPHA;
     } else {
-	cblend |= R200_TXC_ARG_B_ZERO | R200_TXC_COMP_ARG_B;
-	ablend |= R200_TXA_ARG_B_ZERO | R200_TXA_COMP_ARG_B;
+	mask_color = R200_TXC_ARG_B_ZERO | R200_TXC_COMP_ARG_B;
+	mask_alpha = R200_TXA_ARG_B_ZERO | R200_TXA_COMP_ARG_B;
     }
 
-    OUT_ACCEL_REG(R200_PP_TXCBLEND_0, cblend);
-    OUT_ACCEL_REG(R200_PP_TXCBLEND2_0,
-	R200_TXC_CLAMP_0_1 | R200_TXC_OUTPUT_REG_R0);
-    OUT_ACCEL_REG(R200_PP_TXABLEND_0, ablend);
+    OUT_ACCEL_REG(R200_PP_TXCBLEND_0, (src_color |
+				       mask_color |
+				       R200_TXC_OP_MADD |
+				       R200_TXC_ARG_C_ZERO));
+    if (pDstPicture->format == PICT_a8)
+	OUT_ACCEL_REG(R200_PP_TXCBLEND2_0, (R200_TXC_CLAMP_0_1 |
+					    R200_TXC_OUTPUT_REG_R0 |
+					    R200_TXC_OUTPUT_ROTATE_ARG));
+    else
+	OUT_ACCEL_REG(R200_PP_TXCBLEND2_0, (R200_TXC_CLAMP_0_1 |
+					    R200_TXC_OUTPUT_REG_R0));
+    OUT_ACCEL_REG(R200_PP_TXABLEND_0, (src_alpha |
+				       mask_alpha |
+				       R200_TXA_OP_MADD |
+				       R200_TXA_ARG_C_ZERO));
     OUT_ACCEL_REG(R200_PP_TXABLEND2_0,
-	R200_TXA_CLAMP_0_1 | R200_TXA_OUTPUT_REG_R0);
+		  R200_TXA_CLAMP_0_1 | R200_TXA_OUTPUT_REG_R0);
 
     /* Op operator. */
     blendcntl = RADEONGetBlendCntl(op, pMaskPicture, pDstPicture->format);
diff --git a/src/radeon_reg.h b/src/radeon_reg.h
index fbc724a..41c90b6 100644
--- a/src/radeon_reg.h
+++ b/src/radeon_reg.h
@@ -2862,6 +2862,10 @@
 #       define R200_TXC_OUTPUT_MASK_G		(5 << 20)
 #       define R200_TXC_OUTPUT_MASK_B		(6 << 20)
 #       define R200_TXC_OUTPUT_MASK_NONE	(7 << 20)
+#       define R200_TXC_OUTPUT_ROTATE_RGB	(0 << 24)
+#       define R200_TXC_OUTPUT_ROTATE_ARG	(1 << 24)
+#       define R200_TXC_OUTPUT_ROTATE_GBA	(2 << 24)
+#       define R200_TXC_OUTPUT_ROTATE_RGA	(3 << 24)
 #       define R200_TXC_REPL_NORMAL		0
 #       define R200_TXC_REPL_RED		1
 #       define R200_TXC_REPL_GREEN		2