r600: split state emit into block specific functions

We probably want to go finer grained eventually, but
this is a good start.

6 files changed, 379 insertions, 320 deletions

diff --git a/src/mesa/drivers/dri/r600/r600_context.c b/src/mesa/drivers/dri/r600/r600_context.c
index 7009374b0ca..4489064c0d0 100644
--- a/src/mesa/drivers/dri/r600/r600_context.c
+++ b/src/mesa/drivers/dri/r600/r600_context.c
@@ -247,8 +247,6 @@ GLboolean r600CreateContext(const __GLcontextModes * glVisual,
 	 */
 	_mesa_init_driver_functions(&functions);
 
-	r700InitChipObject(r600);  /* let the eag... */
-
 	r700InitStateFuncs(&functions);
 	r600InitTextureFuncs(&functions);
 	r700InitShaderFuncs(&functions);
@@ -386,18 +384,4 @@ GLboolean r600CreateContext(const __GLcontextModes * glVisual,
 	return GL_TRUE;
 }
 
-/* Clean our own things only, radeonDestroyContext will do every thing else. */
-void
-r600DestroyContext (__DRIcontextPrivate * driContextPriv)
-{
-    GET_CURRENT_CONTEXT (ctx);
-    context_t *context = ctx ? R700_CONTEXT(ctx) : NULL;
-
-    if (context)
-	    FREE(context->hw.pStateList);
-
-    radeonDestroyContext(driContextPriv);
-}
-
-
 
diff --git a/src/mesa/drivers/dri/r600/r600_context.h b/src/mesa/drivers/dri/r600/r600_context.h
index 4373254dd68..17ac0889011 100644
--- a/src/mesa/drivers/dri/r600/r600_context.h
+++ b/src/mesa/drivers/dri/r600/r600_context.h
@@ -131,15 +131,20 @@ struct r600_context {
 #define R700_CONTEXT(ctx)		((context_t *)(ctx->DriverCtx))
 #define GL_CONTEXT(context)     ((GLcontext *)(context->radeon.glCtx))
 
-extern void r600DestroyContext(__DRIcontextPrivate * driContextPriv);
 extern GLboolean r600CreateContext(const __GLcontextModes * glVisual,
 				   __DRIcontextPrivate * driContextPriv,
 				   void *sharedContextPrivate);
 
 #define R700_CONTEXT_STATES(context) ((R700_CHIP_CONTEXT *)(&context->hw))
 
-extern GLboolean r700InitChipObject(context_t *context);
-extern GLboolean r700SendContextStates(context_t *context);
+extern GLboolean r700SendSPIState(context_t *context);
+extern GLboolean r700SendVGTState(context_t *context);
+extern GLboolean r700SendSXState(context_t *context);
+extern GLboolean r700SendDBState(context_t *context);
+extern GLboolean r700SendCBState(context_t *context);
+extern GLboolean r700SendSUState(context_t *context);
+extern GLboolean r700SendCLState(context_t *context);
+extern GLboolean r700SendSCState(context_t *context);
 extern GLboolean r700SendViewportState(context_t *context, int id);
 extern GLboolean r700SendRenderTargetState(context_t *context, int id);
 extern GLboolean r700SendTextureState(context_t *context);
diff --git a/src/mesa/drivers/dri/r600/r700_chip.c b/src/mesa/drivers/dri/r600/r700_chip.c
index 43661ec714c..e67e544d538 100644
--- a/src/mesa/drivers/dri/r600/r700_chip.c
+++ b/src/mesa/drivers/dri/r600/r700_chip.c
@@ -40,215 +40,6 @@
 
 #include "radeon_mipmap_tree.h"
 
-#define LINK_STATES(reg)                                            \
-do                                                                  \
-{                                                                   \
-    pStateListWork->puiValue = (unsigned int*)&(r700->reg);         \
-    pStateListWork->unOffset = mm##reg - ASIC_CONTEXT_BASE_INDEX; \
-    pStateListWork->pNext    = pStateListWork + 1;                  \
-    pStateListWork++;                                               \
-}while(0)
-
-GLboolean r700InitChipObject(context_t *context)
-{
-    ContextState * pStateListWork;
-
-    R700_CHIP_CONTEXT *r700 = &context->hw;
-
-    /* init state list */
-    r700->pStateList = (ContextState*) MALLOC (sizeof(ContextState)*sizeof(R700_CHIP_CONTEXT)/sizeof(unsigned int));
-    pStateListWork = r700->pStateList;
-
-    // SC
-    LINK_STATES(PA_SC_SCREEN_SCISSOR_TL);
-    LINK_STATES(PA_SC_SCREEN_SCISSOR_BR);
-    LINK_STATES(PA_SC_WINDOW_OFFSET);
-    LINK_STATES(PA_SC_WINDOW_SCISSOR_TL);
-    LINK_STATES(PA_SC_WINDOW_SCISSOR_BR);
-    LINK_STATES(PA_SC_CLIPRECT_RULE);
-    LINK_STATES(PA_SC_CLIPRECT_0_TL);
-    LINK_STATES(PA_SC_CLIPRECT_0_BR);
-    LINK_STATES(PA_SC_CLIPRECT_1_TL);
-    LINK_STATES(PA_SC_CLIPRECT_1_BR);
-    LINK_STATES(PA_SC_CLIPRECT_2_TL);
-    LINK_STATES(PA_SC_CLIPRECT_2_BR);
-    LINK_STATES(PA_SC_CLIPRECT_3_TL);
-    LINK_STATES(PA_SC_CLIPRECT_3_BR);
-    LINK_STATES(PA_SC_EDGERULE);
-    LINK_STATES(PA_SC_GENERIC_SCISSOR_TL);
-    LINK_STATES(PA_SC_GENERIC_SCISSOR_BR);
-    LINK_STATES(PA_SC_LINE_STIPPLE);
-    LINK_STATES(PA_SC_MPASS_PS_CNTL);
-    LINK_STATES(PA_SC_MODE_CNTL);
-    LINK_STATES(PA_SC_LINE_CNTL);
-    LINK_STATES(PA_SC_AA_CONFIG);
-    LINK_STATES(PA_SC_AA_SAMPLE_LOCS_MCTX);
-    LINK_STATES(PA_SC_AA_SAMPLE_LOCS_8S_WD1_MCTX);
-    LINK_STATES(PA_SC_AA_MASK);
-
-    // SU
-    LINK_STATES(PA_SU_POINT_SIZE);
-    LINK_STATES(PA_SU_POINT_MINMAX);
-    LINK_STATES(PA_SU_LINE_CNTL);
-    LINK_STATES(PA_SU_SC_MODE_CNTL);
-    LINK_STATES(PA_SU_VTX_CNTL);
-    LINK_STATES(PA_SU_POLY_OFFSET_DB_FMT_CNTL);
-    LINK_STATES(PA_SU_POLY_OFFSET_CLAMP);
-    LINK_STATES(PA_SU_POLY_OFFSET_FRONT_SCALE);
-    LINK_STATES(PA_SU_POLY_OFFSET_FRONT_OFFSET);
-    LINK_STATES(PA_SU_POLY_OFFSET_BACK_SCALE);
-    LINK_STATES(PA_SU_POLY_OFFSET_BACK_OFFSET);
-
-    // CL
-    LINK_STATES(PA_CL_CLIP_CNTL);
-    LINK_STATES(PA_CL_VTE_CNTL);
-    LINK_STATES(PA_CL_VS_OUT_CNTL);
-    LINK_STATES(PA_CL_NANINF_CNTL);
-    LINK_STATES(PA_CL_GB_VERT_CLIP_ADJ);
-    LINK_STATES(PA_CL_GB_VERT_DISC_ADJ);
-    LINK_STATES(PA_CL_GB_HORZ_CLIP_ADJ);
-    LINK_STATES(PA_CL_GB_HORZ_DISC_ADJ);
-
-    // CB
-    LINK_STATES(CB_CLEAR_RED_R6XX);
-    LINK_STATES(CB_CLEAR_GREEN_R6XX);
-    LINK_STATES(CB_CLEAR_BLUE_R6XX);
-    LINK_STATES(CB_CLEAR_ALPHA_R6XX);
-    LINK_STATES(CB_TARGET_MASK);
-    LINK_STATES(CB_SHADER_MASK);
-    LINK_STATES(CB_BLEND_RED);
-    LINK_STATES(CB_BLEND_GREEN);
-    LINK_STATES(CB_BLEND_BLUE);
-    LINK_STATES(CB_BLEND_ALPHA);
-    LINK_STATES(CB_FOG_RED_R6XX);
-    LINK_STATES(CB_FOG_GREEN_R6XX);
-    LINK_STATES(CB_FOG_BLUE_R6XX);
-    LINK_STATES(CB_SHADER_CONTROL);
-    LINK_STATES(CB_COLOR_CONTROL);
-    LINK_STATES(CB_CLRCMP_CONTROL);
-    LINK_STATES(CB_CLRCMP_SRC);
-    LINK_STATES(CB_CLRCMP_DST);
-    LINK_STATES(CB_CLRCMP_MSK);
-    LINK_STATES(CB_BLEND_CONTROL);
-
-    //DB
-    LINK_STATES(DB_HTILE_DATA_BASE);
-    LINK_STATES(DB_STENCIL_CLEAR);
-    LINK_STATES(DB_DEPTH_CLEAR);
-    LINK_STATES(DB_STENCILREFMASK);
-    LINK_STATES(DB_STENCILREFMASK_BF);
-    LINK_STATES(DB_DEPTH_CONTROL);
-    LINK_STATES(DB_SHADER_CONTROL);
-    LINK_STATES(DB_RENDER_CONTROL);
-    LINK_STATES(DB_RENDER_OVERRIDE);
-    LINK_STATES(DB_HTILE_SURFACE);
-    LINK_STATES(DB_ALPHA_TO_MASK);
-
-    // SX
-    LINK_STATES(SX_MISC);
-    LINK_STATES(SX_ALPHA_TEST_CONTROL);
-    LINK_STATES(SX_ALPHA_REF);
-
-    // VGT
-    LINK_STATES(VGT_MAX_VTX_INDX);
-    LINK_STATES(VGT_MIN_VTX_INDX);
-    LINK_STATES(VGT_INDX_OFFSET);
-    LINK_STATES(VGT_MULTI_PRIM_IB_RESET_INDX);
-    LINK_STATES(VGT_OUTPUT_PATH_CNTL);
-    LINK_STATES(VGT_HOS_CNTL);
-    LINK_STATES(VGT_HOS_MAX_TESS_LEVEL);
-    LINK_STATES(VGT_HOS_MIN_TESS_LEVEL);
-    LINK_STATES(VGT_HOS_REUSE_DEPTH);
-    LINK_STATES(VGT_GROUP_PRIM_TYPE);
-    LINK_STATES(VGT_GROUP_FIRST_DECR);
-    LINK_STATES(VGT_GROUP_DECR);
-    LINK_STATES(VGT_GROUP_VECT_0_CNTL);
-    LINK_STATES(VGT_GROUP_VECT_1_CNTL);
-    LINK_STATES(VGT_GROUP_VECT_0_FMT_CNTL);
-    LINK_STATES(VGT_GROUP_VECT_1_FMT_CNTL);
-    LINK_STATES(VGT_GS_MODE);
-    LINK_STATES(VGT_PRIMITIVEID_EN);
-    LINK_STATES(VGT_MULTI_PRIM_IB_RESET_EN);
-    LINK_STATES(VGT_INSTANCE_STEP_RATE_0);
-    LINK_STATES(VGT_INSTANCE_STEP_RATE_1);
-    LINK_STATES(VGT_STRMOUT_EN);
-    LINK_STATES(VGT_REUSE_OFF);
-    LINK_STATES(VGT_VTX_CNT_EN);
-    LINK_STATES(VGT_STRMOUT_BUFFER_EN);
-
-    LINK_STATES(SQ_VTX_SEMANTIC_0);
-    LINK_STATES(SQ_VTX_SEMANTIC_1);
-    LINK_STATES(SQ_VTX_SEMANTIC_2);
-    LINK_STATES(SQ_VTX_SEMANTIC_3);
-    LINK_STATES(SQ_VTX_SEMANTIC_4);
-    LINK_STATES(SQ_VTX_SEMANTIC_5);
-    LINK_STATES(SQ_VTX_SEMANTIC_6);
-    LINK_STATES(SQ_VTX_SEMANTIC_7);
-    LINK_STATES(SQ_VTX_SEMANTIC_8);
-    LINK_STATES(SQ_VTX_SEMANTIC_9);
-    LINK_STATES(SQ_VTX_SEMANTIC_10);
-    LINK_STATES(SQ_VTX_SEMANTIC_11);
-    LINK_STATES(SQ_VTX_SEMANTIC_12);
-    LINK_STATES(SQ_VTX_SEMANTIC_13);
-    LINK_STATES(SQ_VTX_SEMANTIC_14);
-    LINK_STATES(SQ_VTX_SEMANTIC_15);
-    LINK_STATES(SQ_VTX_SEMANTIC_16);
-    LINK_STATES(SQ_VTX_SEMANTIC_17);
-    LINK_STATES(SQ_VTX_SEMANTIC_18);
-    LINK_STATES(SQ_VTX_SEMANTIC_19);
-    LINK_STATES(SQ_VTX_SEMANTIC_20);
-    LINK_STATES(SQ_VTX_SEMANTIC_21);
-    LINK_STATES(SQ_VTX_SEMANTIC_22);
-    LINK_STATES(SQ_VTX_SEMANTIC_23);
-    LINK_STATES(SQ_VTX_SEMANTIC_24);
-    LINK_STATES(SQ_VTX_SEMANTIC_25);
-    LINK_STATES(SQ_VTX_SEMANTIC_26);
-    LINK_STATES(SQ_VTX_SEMANTIC_27);
-    LINK_STATES(SQ_VTX_SEMANTIC_28);
-    LINK_STATES(SQ_VTX_SEMANTIC_29);
-    LINK_STATES(SQ_VTX_SEMANTIC_30);
-    LINK_STATES(SQ_VTX_SEMANTIC_31);
-
-    // SPI
-    LINK_STATES(SPI_VS_OUT_ID_0);
-    LINK_STATES(SPI_VS_OUT_ID_1);
-    LINK_STATES(SPI_VS_OUT_ID_2);
-    LINK_STATES(SPI_VS_OUT_ID_3);
-    LINK_STATES(SPI_VS_OUT_ID_4);
-    LINK_STATES(SPI_VS_OUT_ID_5);
-    LINK_STATES(SPI_VS_OUT_ID_6);
-    LINK_STATES(SPI_VS_OUT_ID_7);
-    LINK_STATES(SPI_VS_OUT_ID_8);
-    LINK_STATES(SPI_VS_OUT_ID_9);
-
-    LINK_STATES(SPI_VS_OUT_CONFIG);
-    LINK_STATES(SPI_THREAD_GROUPING);
-    LINK_STATES(SPI_PS_IN_CONTROL_0);
-    LINK_STATES(SPI_PS_IN_CONTROL_1);
-    LINK_STATES(SPI_INTERP_CONTROL_0);
-    LINK_STATES(SPI_INPUT_Z);
-    LINK_STATES(SPI_FOG_CNTL);
-    LINK_STATES(SPI_FOG_FUNC_SCALE);
-    LINK_STATES(SPI_FOG_FUNC_BIAS);
-
-    // SQ
-    LINK_STATES(SQ_ESGS_RING_ITEMSIZE);
-    LINK_STATES(SQ_GSVS_RING_ITEMSIZE);
-    LINK_STATES(SQ_ESTMP_RING_ITEMSIZE);
-    LINK_STATES(SQ_GSTMP_RING_ITEMSIZE);
-    LINK_STATES(SQ_VSTMP_RING_ITEMSIZE);
-    LINK_STATES(SQ_PSTMP_RING_ITEMSIZE);
-    LINK_STATES(SQ_FBUF_RING_ITEMSIZE);
-    LINK_STATES(SQ_REDUC_RING_ITEMSIZE);
-    //LINK_STATES(SQ_GS_VERT_ITEMSIZE);
-
-    pStateListWork->puiValue = (unsigned int*)&(r700->SQ_GS_VERT_ITEMSIZE);
-    pStateListWork->unOffset = mmSQ_GS_VERT_ITEMSIZE - ASIC_CONTEXT_BASE_INDEX;
-    pStateListWork->pNext    = NULL;  /* END OF STATE LIST */
-
-    return GL_TRUE;
-}
-
 GLboolean r700SendTextureState(context_t *context)
 {
     unsigned int i;
@@ -429,74 +220,6 @@ int r700SetupStreams(GLcontext * ctx)
     return R600_FALLBACK_NONE;
 }
 
-GLboolean r700SendContextStates(context_t *context)
-{
-    BATCH_LOCALS(&context->radeon);
-
-    R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
-
-    ContextState * pState = r700->pStateList;
-    ContextState * pInit;
-    unsigned int   toSend;
-    unsigned int   ui;
-
-    while(NULL != pState)
-    {
-        toSend = 1;
-
-        pInit = pState;
-
-	while(NULL != pState->pNext)
-	{
-                if ((pState->pNext->unOffset - pState->unOffset) > 1)
-                {
-			break;
-                }
-                else
-                {
-			pState = pState->pNext;
-			toSend++;
-                }
-	}
-
-        pState = pState->pNext;
-
-        BEGIN_BATCH_NO_AUTOSTATE(toSend + 2);
-        R600_OUT_BATCH_REGSEQ(((pInit->unOffset + ASIC_CONTEXT_BASE_INDEX)<<2), toSend);
-        for(ui=0; ui<toSend; ui++)
-        {
-                R600_OUT_BATCH(*(pInit->puiValue));
-		pInit = pInit->pNext;
-        };
-        END_BATCH();
-    };
-
-    /* todo:
-     * - split this into a separate function?
-     * - only emit the ones we use
-     */
-    BEGIN_BATCH_NO_AUTOSTATE(2 + R700_MAX_SHADER_EXPORTS);
-    R600_OUT_BATCH_REGSEQ(SPI_PS_INPUT_CNTL_0, R700_MAX_SHADER_EXPORTS);
-    for(ui = 0; ui < R700_MAX_SHADER_EXPORTS; ui++)
-	    R600_OUT_BATCH(r700->SPI_PS_INPUT_CNTL[ui].u32All);
-    END_BATCH();
-
-    if (context->radeon.radeonScreen->chip_family > CHIP_FAMILY_R600) {
-	    for (ui = 0; ui < R700_MAX_RENDER_TARGETS; ui++) {
-		    if (r700->render_target[ui].enabled) {
-			    BEGIN_BATCH_NO_AUTOSTATE(3);
-			    R600_OUT_BATCH_REGVAL(CB_BLEND0_CONTROL + (4 * ui),
-						  r700->render_target[ui].CB_BLEND0_CONTROL.u32All);
-			    END_BATCH();
-		    }
-	    }
-    }
-
-    COMMIT_BATCH();
-
-    return GL_TRUE;
-}
-
 GLboolean r700SendDepthTargetState(context_t *context)
 {
 	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
@@ -733,7 +456,7 @@ GLboolean r700SendSQConfig(context_t *context)
 	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
 	BATCH_LOCALS(&context->radeon);
 
-        BEGIN_BATCH_NO_AUTOSTATE(8);
+        BEGIN_BATCH_NO_AUTOSTATE(34);
 	R600_OUT_BATCH_REGSEQ(SQ_CONFIG, 6);
 	R600_OUT_BATCH(r700->sq_config.SQ_CONFIG.u32All);
 	R600_OUT_BATCH(r700->sq_config.SQ_GPR_RESOURCE_MGMT_1.u32All);
@@ -741,14 +464,23 @@ GLboolean r700SendSQConfig(context_t *context)
 	R600_OUT_BATCH(r700->sq_config.SQ_THREAD_RESOURCE_MGMT.u32All);
 	R600_OUT_BATCH(r700->sq_config.SQ_STACK_RESOURCE_MGMT_1.u32All);
 	R600_OUT_BATCH(r700->sq_config.SQ_STACK_RESOURCE_MGMT_2.u32All);
-        END_BATCH();
 
-	BEGIN_BATCH_NO_AUTOSTATE(15);
 	R600_OUT_BATCH_REGVAL(TA_CNTL_AUX, r700->TA_CNTL_AUX.u32All);
 	R600_OUT_BATCH_REGVAL(VC_ENHANCE, r700->VC_ENHANCE.u32All);
 	R600_OUT_BATCH_REGVAL(R7xx_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, r700->SQ_DYN_GPR_CNTL_PS_FLUSH_REQ.u32All);
 	R600_OUT_BATCH_REGVAL(DB_DEBUG, r700->DB_DEBUG.u32All);
 	R600_OUT_BATCH_REGVAL(DB_WATERMARKS, r700->DB_WATERMARKS.u32All);
+
+	R600_OUT_BATCH_REGSEQ(SQ_ESGS_RING_ITEMSIZE, 9);
+	R600_OUT_BATCH(r700->SQ_ESGS_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_GSVS_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_ESTMP_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_GSTMP_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_VSTMP_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_PSTMP_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_FBUF_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_REDUC_RING_ITEMSIZE.u32All);
+	R600_OUT_BATCH(r700->SQ_GS_VERT_ITEMSIZE.u32All);
         END_BATCH();
 
 	COMMIT_BATCH();
@@ -778,3 +510,327 @@ GLboolean r700SendUCPState(context_t *context)
 	return GL_TRUE;
 }
 
+GLboolean r700SendSPIState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+	unsigned int ui;
+
+	BEGIN_BATCH_NO_AUTOSTATE(59 + R700_MAX_SHADER_EXPORTS);
+
+	R600_OUT_BATCH_REGSEQ(SQ_VTX_SEMANTIC_0, 32);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_0.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_1.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_2.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_3.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_4.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_5.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_6.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_7.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_8.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_9.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_10.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_11.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_12.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_13.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_14.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_15.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_16.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_17.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_18.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_19.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_20.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_21.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_22.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_23.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_24.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_25.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_26.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_27.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_28.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_29.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_30.u32All);
+	R600_OUT_BATCH(r700->SQ_VTX_SEMANTIC_31.u32All);
+
+	R600_OUT_BATCH_REGSEQ(SPI_VS_OUT_ID_0, 10);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_0.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_1.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_2.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_3.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_4.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_5.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_6.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_7.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_8.u32All);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_ID_9.u32All);
+
+	R600_OUT_BATCH_REGSEQ(SPI_VS_OUT_CONFIG, 9);
+	R600_OUT_BATCH(r700->SPI_VS_OUT_CONFIG.u32All);
+	R600_OUT_BATCH(r700->SPI_THREAD_GROUPING.u32All);
+	R600_OUT_BATCH(r700->SPI_PS_IN_CONTROL_0.u32All);
+	R600_OUT_BATCH(r700->SPI_PS_IN_CONTROL_1.u32All);
+	R600_OUT_BATCH(r700->SPI_INTERP_CONTROL_0.u32All);
+	R600_OUT_BATCH(r700->SPI_INPUT_Z.u32All);
+	R600_OUT_BATCH(r700->SPI_FOG_CNTL.u32All);
+	R600_OUT_BATCH(r700->SPI_FOG_FUNC_SCALE.u32All);
+	R600_OUT_BATCH(r700->SPI_FOG_FUNC_BIAS.u32All);
+
+	R600_OUT_BATCH_REGSEQ(SPI_PS_INPUT_CNTL_0, R700_MAX_SHADER_EXPORTS);
+	for(ui = 0; ui < R700_MAX_SHADER_EXPORTS; ui++)
+		R600_OUT_BATCH(r700->SPI_PS_INPUT_CNTL[ui].u32All);
+
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+GLboolean r700SendVGTState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+
+        BEGIN_BATCH_NO_AUTOSTATE(41);
+
+	R600_OUT_BATCH_REGSEQ(VGT_MAX_VTX_INDX, 4);
+	R600_OUT_BATCH(r700->VGT_MAX_VTX_INDX.u32All);
+	R600_OUT_BATCH(r700->VGT_MIN_VTX_INDX.u32All);
+	R600_OUT_BATCH(r700->VGT_INDX_OFFSET.u32All);
+	R600_OUT_BATCH(r700->VGT_MULTI_PRIM_IB_RESET_INDX.u32All);
+
+	R600_OUT_BATCH_REGSEQ(VGT_OUTPUT_PATH_CNTL, 13);
+	R600_OUT_BATCH(r700->VGT_OUTPUT_PATH_CNTL.u32All);
+	R600_OUT_BATCH(r700->VGT_HOS_CNTL.u32All);
+	R600_OUT_BATCH(r700->VGT_HOS_MAX_TESS_LEVEL.u32All);
+	R600_OUT_BATCH(r700->VGT_HOS_MIN_TESS_LEVEL.u32All);
+	R600_OUT_BATCH(r700->VGT_HOS_REUSE_DEPTH.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_PRIM_TYPE.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_FIRST_DECR.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_DECR.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_VECT_0_CNTL.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_VECT_1_CNTL.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_VECT_0_FMT_CNTL.u32All);
+	R600_OUT_BATCH(r700->VGT_GROUP_VECT_1_FMT_CNTL.u32All);
+	R600_OUT_BATCH(r700->VGT_GS_MODE.u32All);
+
+	R600_OUT_BATCH_REGVAL(VGT_PRIMITIVEID_EN, r700->VGT_PRIMITIVEID_EN.u32All);
+	R600_OUT_BATCH_REGVAL(VGT_MULTI_PRIM_IB_RESET_EN, r700->VGT_MULTI_PRIM_IB_RESET_EN.u32All);
+	R600_OUT_BATCH_REGVAL(VGT_INSTANCE_STEP_RATE_0, r700->VGT_INSTANCE_STEP_RATE_0.u32All);
+	R600_OUT_BATCH_REGVAL(VGT_INSTANCE_STEP_RATE_1, r700->VGT_INSTANCE_STEP_RATE_1.u32All);
+
+	R600_OUT_BATCH_REGSEQ(VGT_STRMOUT_EN, 3);
+	R600_OUT_BATCH(r700->VGT_STRMOUT_EN.u32All);
+	R600_OUT_BATCH(r700->VGT_REUSE_OFF.u32All);
+	R600_OUT_BATCH(r700->VGT_VTX_CNT_EN.u32All);
+
+	R600_OUT_BATCH_REGVAL(VGT_STRMOUT_BUFFER_EN, r700->VGT_STRMOUT_BUFFER_EN.u32All);
+
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+GLboolean r700SendSXState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+
+        BEGIN_BATCH_NO_AUTOSTATE(9);
+	R600_OUT_BATCH_REGVAL(SX_MISC, r700->SX_MISC.u32All);
+	R600_OUT_BATCH_REGVAL(SX_ALPHA_TEST_CONTROL, r700->SX_ALPHA_TEST_CONTROL.u32All);
+	R600_OUT_BATCH_REGVAL(SX_ALPHA_REF, r700->SX_ALPHA_REF.u32All);
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+GLboolean r700SendDBState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+
+        BEGIN_BATCH_NO_AUTOSTATE(27);
+	R600_OUT_BATCH_REGVAL(DB_HTILE_DATA_BASE, r700->DB_HTILE_DATA_BASE.u32All);
+
+	R600_OUT_BATCH_REGSEQ(DB_STENCIL_CLEAR, 2);
+	R600_OUT_BATCH(r700->DB_STENCIL_CLEAR.u32All);
+	R600_OUT_BATCH(r700->DB_DEPTH_CLEAR.u32All);
+
+	R600_OUT_BATCH_REGSEQ(DB_STENCILREFMASK, 2);
+	R600_OUT_BATCH(r700->DB_STENCILREFMASK.u32All);
+	R600_OUT_BATCH(r700->DB_STENCILREFMASK_BF.u32All);
+
+	R600_OUT_BATCH_REGVAL(DB_DEPTH_CONTROL, r700->DB_DEPTH_CONTROL.u32All);
+	R600_OUT_BATCH_REGVAL(DB_SHADER_CONTROL, r700->DB_SHADER_CONTROL.u32All);
+
+	R600_OUT_BATCH_REGSEQ(DB_RENDER_CONTROL, 2);
+	R600_OUT_BATCH(r700->DB_RENDER_CONTROL.u32All);
+	R600_OUT_BATCH(r700->DB_RENDER_OVERRIDE.u32All);
+
+	R600_OUT_BATCH_REGVAL(DB_HTILE_SURFACE, r700->DB_HTILE_SURFACE.u32All);
+	R600_OUT_BATCH_REGVAL(DB_ALPHA_TO_MASK, r700->DB_ALPHA_TO_MASK.u32All);
+
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+GLboolean r700SendCBState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+	unsigned int ui;
+
+	if (context->radeon.radeonScreen->chip_family < CHIP_FAMILY_RV770) {
+		BEGIN_BATCH_NO_AUTOSTATE(14);
+		R600_OUT_BATCH_REGSEQ(CB_CLEAR_RED, 4);
+		R600_OUT_BATCH(r700->CB_CLEAR_RED_R6XX.u32All);
+		R600_OUT_BATCH(r700->CB_CLEAR_GREEN_R6XX.u32All);
+		R600_OUT_BATCH(r700->CB_CLEAR_BLUE_R6XX.u32All);
+		R600_OUT_BATCH(r700->CB_CLEAR_ALPHA_R6XX.u32All);
+		R600_OUT_BATCH_REGSEQ(CB_FOG_RED, 3);
+		R600_OUT_BATCH(r700->CB_FOG_RED_R6XX.u32All);
+		R600_OUT_BATCH(r700->CB_FOG_GREEN_R6XX.u32All);
+		R600_OUT_BATCH(r700->CB_FOG_BLUE_R6XX.u32All);
+		/* R600 does not have per-MRT blend */
+		R600_OUT_BATCH_REGVAL(CB_BLEND_CONTROL, r700->CB_BLEND_CONTROL.u32All);
+		END_BATCH();
+	}
+
+	BEGIN_BATCH_NO_AUTOSTATE(22);
+	R600_OUT_BATCH_REGSEQ(CB_TARGET_MASK, 2);
+	R600_OUT_BATCH(r700->CB_TARGET_MASK.u32All);
+	R600_OUT_BATCH(r700->CB_SHADER_MASK.u32All);
+
+	R600_OUT_BATCH_REGSEQ(CB_BLEND_RED, 4);
+	R600_OUT_BATCH(r700->CB_BLEND_RED.u32All);
+	R600_OUT_BATCH(r700->CB_BLEND_GREEN.u32All);
+	R600_OUT_BATCH(r700->CB_BLEND_BLUE.u32All);
+	R600_OUT_BATCH(r700->CB_BLEND_ALPHA.u32All);
+
+	R600_OUT_BATCH_REGVAL(R7xx_CB_SHADER_CONTROL, r700->CB_SHADER_CONTROL.u32All);
+	R600_OUT_BATCH_REGVAL(CB_COLOR_CONTROL, r700->CB_COLOR_CONTROL.u32All);
+
+	R600_OUT_BATCH_REGSEQ(CB_CLRCMP_CONTROL, 4);
+	R600_OUT_BATCH(r700->CB_CLRCMP_CONTROL.u32All);
+	R600_OUT_BATCH(r700->CB_CLRCMP_SRC.u32All);
+	R600_OUT_BATCH(r700->CB_CLRCMP_DST.u32All);
+	R600_OUT_BATCH(r700->CB_CLRCMP_MSK.u32All);
+	END_BATCH();
+
+	if (context->radeon.radeonScreen->chip_family > CHIP_FAMILY_R600) {
+		for (ui = 0; ui < R700_MAX_RENDER_TARGETS; ui++) {
+			if (r700->render_target[ui].enabled) {
+				BEGIN_BATCH_NO_AUTOSTATE(3);
+				R600_OUT_BATCH_REGVAL(CB_BLEND0_CONTROL + (4 * ui),
+						      r700->render_target[ui].CB_BLEND0_CONTROL.u32All);
+				END_BATCH();
+			}
+		}
+	}
+
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+GLboolean r700SendSUState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+
+	BEGIN_BATCH_NO_AUTOSTATE(19);
+	R600_OUT_BATCH_REGVAL(PA_SU_SC_MODE_CNTL, r700->PA_SU_SC_MODE_CNTL.u32All);
+
+	R600_OUT_BATCH_REGSEQ(PA_SU_POINT_SIZE, 4);
+	R600_OUT_BATCH(r700->PA_SU_POINT_SIZE.u32All);
+	R600_OUT_BATCH(r700->PA_SU_POINT_MINMAX.u32All);
+	R600_OUT_BATCH(r700->PA_SU_LINE_CNTL.u32All);
+	R600_OUT_BATCH(r700->PA_SU_VTX_CNTL.u32All);
+
+	R600_OUT_BATCH_REGSEQ(PA_SU_POLY_OFFSET_DB_FMT_CNTL, 2);
+	R600_OUT_BATCH(r700->PA_SU_POLY_OFFSET_DB_FMT_CNTL.u32All);
+	R600_OUT_BATCH(r700->PA_SU_POLY_OFFSET_CLAMP.u32All);
+
+	R600_OUT_BATCH_REGSEQ(PA_SU_POLY_OFFSET_FRONT_SCALE, 4);
+	R600_OUT_BATCH(r700->PA_SU_POLY_OFFSET_FRONT_SCALE.u32All);
+	R600_OUT_BATCH(r700->PA_SU_POLY_OFFSET_FRONT_OFFSET.u32All);
+	R600_OUT_BATCH(r700->PA_SU_POLY_OFFSET_BACK_SCALE.u32All);
+	R600_OUT_BATCH(r700->PA_SU_POLY_OFFSET_BACK_OFFSET.u32All);
+
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+GLboolean r700SendCLState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+
+	BEGIN_BATCH_NO_AUTOSTATE(18);
+	R600_OUT_BATCH_REGVAL(PA_CL_CLIP_CNTL, r700->PA_CL_CLIP_CNTL.u32All);
+	R600_OUT_BATCH_REGVAL(PA_CL_VTE_CNTL, r700->PA_CL_VTE_CNTL.u32All);
+	R600_OUT_BATCH_REGVAL(PA_CL_VS_OUT_CNTL, r700->PA_CL_VS_OUT_CNTL.u32All);
+	R600_OUT_BATCH_REGVAL(PA_CL_NANINF_CNTL, r700->PA_CL_NANINF_CNTL.u32All);
+
+	R600_OUT_BATCH_REGSEQ(PA_CL_GB_VERT_CLIP_ADJ, 4);
+	R600_OUT_BATCH(r700->PA_CL_GB_VERT_CLIP_ADJ.u32All);
+	R600_OUT_BATCH(r700->PA_CL_GB_VERT_DISC_ADJ.u32All);
+	R600_OUT_BATCH(r700->PA_CL_GB_HORZ_CLIP_ADJ.u32All);
+	R600_OUT_BATCH(r700->PA_CL_GB_HORZ_DISC_ADJ.u32All);
+
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
+
+// XXX need to split this up
+GLboolean r700SendSCState(context_t *context)
+{
+	R700_CHIP_CONTEXT *r700 = R700_CONTEXT_STATES(context);
+	BATCH_LOCALS(&context->radeon);
+
+	BEGIN_BATCH_NO_AUTOSTATE(47);
+	R600_OUT_BATCH_REGSEQ(PA_SC_SCREEN_SCISSOR_TL, 2);
+	R600_OUT_BATCH(r700->PA_SC_SCREEN_SCISSOR_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_SCREEN_SCISSOR_BR.u32All);
+
+	R600_OUT_BATCH_REGSEQ(PA_SC_WINDOW_OFFSET, 13);
+	R600_OUT_BATCH(r700->PA_SC_WINDOW_OFFSET.u32All);
+	R600_OUT_BATCH(r700->PA_SC_WINDOW_SCISSOR_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_WINDOW_SCISSOR_BR.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_RULE.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_0_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_0_BR.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_1_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_1_BR.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_2_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_2_BR.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_3_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_CLIPRECT_3_BR.u32All);
+	R600_OUT_BATCH(r700->PA_SC_EDGERULE.u32All);
+
+	R600_OUT_BATCH_REGSEQ(PA_SC_GENERIC_SCISSOR_TL, 2);
+	R600_OUT_BATCH(r700->PA_SC_GENERIC_SCISSOR_TL.u32All);
+	R600_OUT_BATCH(r700->PA_SC_GENERIC_SCISSOR_BR.u32All);
+
+	R600_OUT_BATCH_REGVAL(PA_SC_LINE_STIPPLE, r700->PA_SC_LINE_STIPPLE.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_MPASS_PS_CNTL, r700->PA_SC_MPASS_PS_CNTL.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_MODE_CNTL, r700->PA_SC_MODE_CNTL.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_LINE_CNTL, r700->PA_SC_LINE_CNTL.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_AA_CONFIG, r700->PA_SC_AA_CONFIG.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_AA_SAMPLE_LOCS_MCTX, r700->PA_SC_AA_SAMPLE_LOCS_MCTX.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_AA_SAMPLE_LOCS_8S_WD1_MCTX, r700->PA_SC_AA_SAMPLE_LOCS_8S_WD1_MCTX.u32All);
+	R600_OUT_BATCH_REGVAL(PA_SC_AA_MASK, r700->PA_SC_AA_MASK.u32All);
+
+	END_BATCH();
+	COMMIT_BATCH();
+
+	return GL_TRUE;
+}
diff --git a/src/mesa/drivers/dri/r600/r700_chip.h b/src/mesa/drivers/dri/r600/r700_chip.h
index 4e89c75f2f9..545e4a188a2 100644
--- a/src/mesa/drivers/dri/r600/r700_chip.h
+++ b/src/mesa/drivers/dri/r600/r700_chip.h
@@ -188,6 +188,7 @@ typedef struct _RENDER_TARGET_STATE_STRUCT
 	union UINT_FLOAT            	CB_COLOR0_MASK;  /* 0xA040 */
 	union UINT_FLOAT         	CB_BLEND0_CONTROL;  /* 0xA1E0 */
 	GLboolean                         enabled;
+	GLboolean                         dirty;
 } RENDER_TARGET_STATE_STRUCT;
 
 typedef struct _VIEWPORT_STATE_STRUCT
@@ -203,6 +204,7 @@ typedef struct _VIEWPORT_STATE_STRUCT
 	union UINT_FLOAT        PA_CL_VPORT_ZSCALE;        /* 0xA113 */
 	union UINT_FLOAT       	PA_CL_VPORT_ZOFFSET;       /* 0xA114 */
 	GLboolean                         enabled;
+	GLboolean                         dirty;
 } VIEWPORT_STATE_STRUCT;
 
 typedef struct _UCP_STATE_STRUCT
@@ -212,6 +214,7 @@ typedef struct _UCP_STATE_STRUCT
 	union UINT_FLOAT        PA_CL_UCP_0_Z;
 	union UINT_FLOAT        PA_CL_UCP_0_W;
 	GLboolean                         enabled;
+	GLboolean                         dirty;
 } UCP_STATE_STRUCT;
 
 typedef struct _PS_STATE_STRUCT
@@ -220,6 +223,7 @@ typedef struct _PS_STATE_STRUCT
 	union UINT_FLOAT       	        SQ_PGM_RESOURCES_PS       ;  /* 0xA214 */
 	union UINT_FLOAT         	SQ_PGM_EXPORTS_PS         ;  /* 0xA215 */
 	union UINT_FLOAT       	        SQ_PGM_CF_OFFSET_PS       ;  /* 0xA233 */
+	GLboolean                         dirty;
 } PS_STATE_STRUCT;
 
 typedef struct _VS_STATE_STRUCT
@@ -227,6 +231,7 @@ typedef struct _VS_STATE_STRUCT
  	union UINT_FLOAT           	SQ_PGM_START_VS           ;  /* 0xA216 */
 	union UINT_FLOAT  		SQ_PGM_RESOURCES_VS       ;  /* 0xA21A */
 	union UINT_FLOAT       	        SQ_PGM_CF_OFFSET_VS       ;  /* 0xA234 */
+	GLboolean                         dirty;
 } VS_STATE_STRUCT;
 
 typedef struct _GS_STATE_STRUCT
@@ -234,6 +239,7 @@ typedef struct _GS_STATE_STRUCT
 	union UINT_FLOAT           	SQ_PGM_START_GS           ;  /* 0xA21B */
 	union UINT_FLOAT       	        SQ_PGM_RESOURCES_GS       ;  /* 0xA21F */
 	union UINT_FLOAT       	        SQ_PGM_CF_OFFSET_GS       ;  /* 0xA235 */
+	GLboolean                         dirty;
 } GS_STATE_STRUCT;
 
 typedef struct _ES_STATE_STRUCT
@@ -241,6 +247,7 @@ typedef struct _ES_STATE_STRUCT
 	union UINT_FLOAT           	SQ_PGM_START_ES           ;  /* 0xA220 */
 	union UINT_FLOAT       	        SQ_PGM_RESOURCES_ES       ;  /* 0xA224 */
 	union UINT_FLOAT       	        SQ_PGM_CF_OFFSET_ES       ;  /* 0xA236 */
+	GLboolean                         dirty;
 } ES_STATE_STRUCT;
 
 typedef struct _FS_STATE_STRUCT
@@ -248,6 +255,7 @@ typedef struct _FS_STATE_STRUCT
 	union UINT_FLOAT           	SQ_PGM_START_FS           ;  /* 0xA225 */
 	union UINT_FLOAT       	        SQ_PGM_RESOURCES_FS       ;  /* 0xA229 */
 	union UINT_FLOAT       	        SQ_PGM_CF_OFFSET_FS       ;  /* 0xA237 */
+	GLboolean                         dirty;
 } FS_STATE_STRUCT;
 
 typedef struct _SQ_CONFIG_STRUCT
@@ -260,22 +268,8 @@ typedef struct _SQ_CONFIG_STRUCT
 	union UINT_FLOAT     	        SQ_STACK_RESOURCE_MGMT_2  ;  /* 0x2305 */
 } SQ_CONFIG_STRUCT;
 
-typedef struct ContextState
-{
-    unsigned int * puiValue;
-    unsigned int   unOffset;
-    struct ContextState * pNext;
-} ContextState;
-
 typedef struct _R700_CHIP_CONTEXT
 {
-	// misc
-	union UINT_FLOAT             	TA_CNTL_AUX               ;  /* 0x2542 */
-	union UINT_FLOAT             	VC_ENHANCE                ;  /* 0x25C5 */
-	union UINT_FLOAT             	SQ_DYN_GPR_CNTL_PS_FLUSH_REQ;  /* 0x2363 */
-	union UINT_FLOAT             	DB_DEBUG                  ;  /* 0x260C */
-	union UINT_FLOAT             	DB_WATERMARKS             ;  /* 0x260E */
-
 	// DB
 	union UINT_FLOAT             	DB_DEPTH_SIZE             ;  /* 0xA000 */
 	union UINT_FLOAT             	DB_DEPTH_VIEW             ;  /* 0xA001 */
@@ -292,6 +286,7 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT          	DB_ALPHA_TO_MASK          ;  /* 0xA351 */
 	union UINT_FLOAT          	DB_DEPTH_CONTROL          ;  /* 0xA200 */
 	union UINT_FLOAT         	DB_SHADER_CONTROL         ;  /* 0xA203 */
+	GLboolean                       db_dirty;
 
 	// SC
 	union UINT_FLOAT   	        PA_SC_SCREEN_SCISSOR_TL   ;  /* 0xA00C */
@@ -311,6 +306,8 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT            	PA_SC_EDGERULE            ;  /* 0xA08C */
 	union UINT_FLOAT  	        PA_SC_GENERIC_SCISSOR_TL  ;  /* 0xA090 */
 	union UINT_FLOAT  	        PA_SC_GENERIC_SCISSOR_BR  ;  /* 0xA091 */
+	GLboolean                       scissor_dirty;
+
 	union UINT_FLOAT        	PA_SC_LINE_STIPPLE        ;  /* 0xA283 */
 	union UINT_FLOAT           	PA_SC_LINE_CNTL           ;  /* 0xA300 */
 	union UINT_FLOAT           	PA_SC_AA_CONFIG           ;  /* 0xA301 */
@@ -319,6 +316,7 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT 	        PA_SC_AA_SAMPLE_LOCS_MCTX ;  /* 0xA307 */
 	union UINT_FLOAT                PA_SC_AA_SAMPLE_LOCS_8S_WD1_MCTX; /* 0xA308 */
 	union UINT_FLOAT             	PA_SC_AA_MASK             ;  /* 0xA312 */
+	GLboolean                       sc_dirty;
 
 	// CL
 	union UINT_FLOAT           	PA_CL_CLIP_CNTL           ;  /* 0xA204 */
@@ -329,6 +327,7 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT    	        PA_CL_GB_VERT_DISC_ADJ    ;  /* 0xA304 */
 	union UINT_FLOAT    	        PA_CL_GB_HORZ_CLIP_ADJ    ;  /* 0xA305 */
 	union UINT_FLOAT    	        PA_CL_GB_HORZ_DISC_ADJ    ;  /* 0xA306 */
+	GLboolean                       cl_dirty;
 
 	// SU
 	union UINT_FLOAT        	PA_SU_SC_MODE_CNTL        ;  /* 0xA205 */
@@ -342,6 +341,7 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT                PA_SU_POLY_OFFSET_FRONT_OFFSET; /* 0xA381 */
 	union UINT_FLOAT                PA_SU_POLY_OFFSET_BACK_SCALE;    /* 0xA382 */
 	union UINT_FLOAT                PA_SU_POLY_OFFSET_BACK_OFFSET;   /* 0xA383 */
+	GLboolean                       su_dirty;
 
 	VIEWPORT_STATE_STRUCT           viewport[R700_MAX_VIEWPORTS];
 	UCP_STATE_STRUCT                ucp[R700_MAX_UCP];
@@ -367,12 +367,14 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT             	CB_CLRCMP_DST             ;  /* 0xA30E */
 	union UINT_FLOAT             	CB_CLRCMP_MSK             ;  /* 0xA30F */
 	union UINT_FLOAT             	CB_BLEND_CONTROL          ;  /* 0xABD0 */
+	GLboolean                       cb_dirty;
 	RENDER_TARGET_STATE_STRUCT      render_target[R700_MAX_RENDER_TARGETS];
 
 	// SX
 	union UINT_FLOAT                SX_MISC                   ;  /* 0xA0D4 */
 	union UINT_FLOAT     	        SX_ALPHA_TEST_CONTROL     ;  /* 0xA104 */
 	union UINT_FLOAT     	        SX_ALPHA_REF              ;  /* 0xA10E */
+	GLboolean                       sx_dirty;
 
 	// VGT
 	union UINT_FLOAT          	VGT_MAX_VTX_INDX          ;  /* 0xA100 */
@@ -400,6 +402,7 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT             	VGT_REUSE_OFF             ;  /* 0xA2AD */
 	union UINT_FLOAT             	VGT_VTX_CNT_EN            ;  /* 0xA2AE */
 	union UINT_FLOAT            	VGT_STRMOUT_BUFFER_EN     ;  /* 0xA2C8 */
+	GLboolean                       vgt_dirty;
 
 	// SPI
 	union UINT_FLOAT           	SPI_VS_OUT_ID_0           ;  /* 0xA185 */
@@ -454,8 +457,8 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT        	SQ_VTX_SEMANTIC_29        ;  /* 0xA0FD */
 	union UINT_FLOAT        	SQ_VTX_SEMANTIC_30        ;  /* 0xA0FE */
 	union UINT_FLOAT        	SQ_VTX_SEMANTIC_31        ;  /* 0xA0FF */
-
-	union UINT_FLOAT       	SPI_PS_INPUT_CNTL[R700_MAX_SHADER_EXPORTS];
+	union UINT_FLOAT       	        SPI_PS_INPUT_CNTL[R700_MAX_SHADER_EXPORTS];
+	GLboolean                       spi_dirty;
 
 	// shaders
 	PS_STATE_STRUCT                 ps;
@@ -466,7 +469,12 @@ typedef struct _R700_CHIP_CONTEXT
 
 	// SQ CONFIG
 	SQ_CONFIG_STRUCT                sq_config;
-
+	// misc
+	union UINT_FLOAT             	TA_CNTL_AUX               ;  /* 0x2542 */
+	union UINT_FLOAT             	VC_ENHANCE                ;  /* 0x25C5 */
+	union UINT_FLOAT             	SQ_DYN_GPR_CNTL_PS_FLUSH_REQ;  /* 0x2363 */
+	union UINT_FLOAT             	DB_DEBUG                  ;  /* 0x260C */
+	union UINT_FLOAT             	DB_WATERMARKS             ;  /* 0x260E */
 	// SQ
 	union UINT_FLOAT     	        SQ_ESGS_RING_ITEMSIZE     ;  /* 0xA22A */
 	union UINT_FLOAT     	        SQ_GSVS_RING_ITEMSIZE     ;  /* 0xA22B */
@@ -477,8 +485,7 @@ typedef struct _R700_CHIP_CONTEXT
 	union UINT_FLOAT     	        SQ_FBUF_RING_ITEMSIZE     ;  /* 0xA230 */
 	union UINT_FLOAT    	        SQ_REDUC_RING_ITEMSIZE    ;  /* 0xA231 */
 	union UINT_FLOAT       	        SQ_GS_VERT_ITEMSIZE       ;  /* 0xA232 */
-
-	ContextState*                   pStateList;
+	GLboolean                       sq_dirty;
 
 	radeonTexObj*                   textures[R700_TEXTURE_NUMBERUNITS];
 
diff --git a/src/mesa/drivers/dri/r600/r700_render.c b/src/mesa/drivers/dri/r600/r700_render.c
index 80df78f1233..1cc886d5e9c 100644
--- a/src/mesa/drivers/dri/r600/r700_render.c
+++ b/src/mesa/drivers/dri/r600/r700_render.c
@@ -345,7 +345,14 @@ void r700EmitState(GLcontext * ctx)
 
 	r700SendSQConfig(context);
 	r700SendUCPState(context);
-	r700SendContextStates(context);
+	r700SendSCState(context);
+	r700SendSUState(context);
+	r700SendCLState(context);
+	r700SendCBState(context);
+	r700SendDBState(context);
+	r700SendSXState(context);
+	r700SendVGTState(context);
+	r700SendSPIState(context);
 	r700SendViewportState(context, 0);
 	r700SendRenderTargetState(context, 0);
 	r700SendDepthTargetState(context);
diff --git a/src/mesa/drivers/dri/radeon/radeon_screen.c b/src/mesa/drivers/dri/radeon/radeon_screen.c
index c8d491621a8..e28543d855a 100644
--- a/src/mesa/drivers/dri/radeon/radeon_screen.c
+++ b/src/mesa/drivers/dri/radeon/radeon_screen.c
@@ -1797,7 +1797,7 @@ const struct __DriverAPIRec driDriverAPI = {
    .DestroyContext  = r200DestroyContext,
 #elif RADEON_COMMON && defined(RADEON_COMMON_FOR_R600)
    .CreateContext   = r600CreateContext,
-   .DestroyContext  = r600DestroyContext,
+   .DestroyContext  = radeonDestroyContext,
 #else
    .CreateContext   = radeonCreateContext,
    .DestroyContext  = radeonDestroyContext,