Add the state caching mechanism. It seems to work, from what I can see.

10 files changed, 605 insertions, 496 deletions

diff --git a/src/mesa/drivers/dri/nouveau/Makefile b/src/mesa/drivers/dri/nouveau/Makefile
index 2db6f8989d1..4d1e3e6c70c 100644
--- a/src/mesa/drivers/dri/nouveau/Makefile
+++ b/src/mesa/drivers/dri/nouveau/Makefile
@@ -22,7 +22,8 @@ DRIVER_SOURCES = \
 	nv10_swtcl.c             \
 	nv10_state.c             \
 	nv20_state.c             \
-	nv30_state.c
+	nv30_state.c             \
+	nouveau_state_cache.c
 
 C_SOURCES = \
 	$(COMMON_SOURCES) \
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_context.h b/src/mesa/drivers/dri/nouveau/nouveau_context.h
index 09972bebac4..8ae7be015dc 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_context.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_context.h
@@ -37,6 +37,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "tnl/t_vertex.h"
 
 #include "nouveau_screen.h"
+#include "nouveau_state_cache.h"
 
 #include "xmlconfig.h"
 
@@ -73,7 +74,6 @@ typedef void (*nouveau_line_func)( struct nouveau_context*,
 typedef void (*nouveau_point_func)( struct nouveau_context*,
 		nouveauVertex * );
 
-
 typedef struct nouveau_context {
 	/* Mesa context */
 	GLcontext *glCtx;
@@ -102,6 +102,9 @@ typedef struct nouveau_context {
 	GLboolean lighting_enabled;
 	uint32_t enabled_lights;
 
+	/* Cached state */
+	nouveau_state_cache state_cache;
+
 	/* The drawing fallbacks */
 	GLuint Fallback;
 	nouveau_tri_func draw_tri;
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_fifo.h b/src/mesa/drivers/dri/nouveau/nouveau_fifo.h
index da2dadada74..259e5a1dc76 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_fifo.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_fifo.h
@@ -45,7 +45,6 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
  * Ring/fifo interface
  *
  * - Begin a ring section with BEGIN_RING_SIZE (if you know the full size in advance)
- * - Begin a ring section with BEGIN_RING_PRIM otherwise (and then finish with FINISH_RING_PRIM)
  * - Output stuff to the ring with either OUT_RINGp (outputs a raw mem chunk), OUT_RING (1 uint32_t) or OUT_RINGf (1 float)
  * - RING_AVAILABLE returns the available fifo (in uint32_ts)
  * - RING_AHEAD returns how much ahead of the last submission point we are
@@ -59,7 +58,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #ifdef NOUVEAU_RING_DEBUG
 
 #define OUT_RINGp(ptr,sz) do {                                                  \
-uint32_t* p=(uint32_t*)ptr;							\
+uint32_t* p=(uint32_t*)(ptr);							\
 int i; printf("OUT_RINGp: (size 0x%x dwords)\n",sz); for(i=0;i<sz;i++) printf(" 0x%08x\n", *(p+i)); 	\
 }while(0)
 
@@ -74,33 +73,47 @@ int i; printf("OUT_RINGp: (size 0x%x dwords)\n",sz); for(i=0;i<sz;i++) printf("
 #else
 
 #define OUT_RINGp(ptr,sz) do{							\
-	memcpy(nmesa->fifo.buffer+nmesa->fifo.current,ptr,sz*4);		\
-	nmesa->fifo.current+=sz;						\
+	memcpy(nmesa->fifo.buffer+nmesa->fifo.current,ptr,(sz)*4);		\
+	nmesa->fifo.current+=(sz);						\
 }while(0)
 
 #define OUT_RING(n) do {							\
-nmesa->fifo.buffer[nmesa->fifo.current++]=n;					\
+nmesa->fifo.buffer[nmesa->fifo.current++]=(n);					\
 }while(0)
 
 #define OUT_RINGf(n) do {							\
-*((float*)(nmesa->fifo.buffer+nmesa->fifo.current++))=n;			\
+*((float*)(nmesa->fifo.buffer+nmesa->fifo.current++))=(n);			\
 }while(0)
 
 #endif
 
 extern void WAIT_RING(nouveauContextPtr nmesa,u_int32_t size);
+extern void nouveau_state_cache_flush(nouveauContextPtr nmesa);
+extern void nouveau_state_cache_init(nouveauContextPtr nmesa);
 
-#define BEGIN_RING_PRIM(subchannel,tag,size) do {					\
-	if (nmesa->fifo.free<size)							\
-		WAIT_RING(nmesa,(size));						\
-	OUT_RING( ((subchannel) << 13) | (tag));					\
+#define BEGIN_RING_CACHE(subchannel,tag,size) do {					\
+	nmesa->state_cache.dirty=1;	 						\
+	nmesa->state_cache.current_pos=((tag)/4);					\
+}while(0)
+
+#define OUT_RING_CACHE(n) do {									\
+	if (nmesa->state_cache.atoms[nmesa->state_cache.current_pos].value!=(n))	{	\
+		nmesa->state_cache.atoms[nmesa->state_cache.current_pos].dirty=1; 		\
+		nmesa->state_cache.atoms[nmesa->state_cache.current_pos].value=(n);		\
+	}											\
+	nmesa->state_cache.current_pos++;							\
 }while(0)
 
-#define FINISH_RING_PRIM() do{								\
-	nmesa->fifo.buffer[nmesa->fifo.put]|=((nmesa->fifo.current-nmesa->fifo.put) << 18);		\
+#define OUT_RING_CACHEf(n) do {									\
+	if ((*(float*)(&nmesa->state_cache.atoms[nmesa->state_cache.current_pos].value))!=(n)){	\
+		nmesa->state_cache.atoms[nmesa->state_cache.current_pos].dirty=1;	 	\
+		(*(float*)(&nmesa->state_cache.atoms[nmesa->state_cache.current_pos].value))=(n);\
+	}											\
+	nmesa->state_cache.current_pos++;							\
 }while(0)
 
 #define BEGIN_RING_SIZE(subchannel,tag,size) do {					\
+	nouveau_state_cache_flush(nmesa);						\
 	if (nmesa->fifo.free <= (size))							\
 		WAIT_RING(nmesa,(size));						\
 	OUT_RING( ((size)<<18) | ((subchannel) << 13) | (tag));				\
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_state.c b/src/mesa/drivers/dri/nouveau/nouveau_state.c
index 9811606311f..2f8f3248ce6 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_state.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_state.c
@@ -108,8 +108,6 @@ static void nouveauDDUpdateHWState(GLcontext *ctx)
 
     if ( new_state || nmesa->new_render_state & _NEW_TEXTURE )
     {
-        FINISH_RING_PRIM();
-
         nmesa->new_state = 0;
 
         /* Update the various parts of the context's state.
@@ -174,6 +172,7 @@ void nouveauDDInitState(nouveauContextPtr nmesa)
         default:
             break;
     }
+    nouveau_state_cache_init(nmesa);
 }
 
 /* Initialize the driver's state functions */
@@ -231,95 +230,98 @@ void nouveauDDInitStateFuncs(GLcontext *ctx)
    ctx->Driver.CopyConvolutionFilter2D = _swrast_CopyConvolutionFilter2D;
 }
 
+#define STATE_INIT(a) if (ctx->Driver.a) ctx->Driver.a
+
 void nouveauInitState(GLcontext *ctx)
 {
     /*
      * Mesa should do this for us:
      */
-    ctx->Driver.AlphaFunc( ctx, 
+
+    STATE_INIT(AlphaFunc)( ctx, 
             ctx->Color.AlphaFunc,
             ctx->Color.AlphaRef);
 
-    ctx->Driver.BlendColor( ctx,
+    STATE_INIT(BlendColor)( ctx,
             ctx->Color.BlendColor );
 
-    ctx->Driver.BlendEquationSeparate( ctx, 
+    STATE_INIT(BlendEquationSeparate)( ctx, 
             ctx->Color.BlendEquationRGB,
             ctx->Color.BlendEquationA);
 
-    ctx->Driver.BlendFuncSeparate( ctx,
+    STATE_INIT(BlendFuncSeparate)( ctx,
             ctx->Color.BlendSrcRGB,
             ctx->Color.BlendDstRGB,
             ctx->Color.BlendSrcA,
             ctx->Color.BlendDstA);
 
-    ctx->Driver.ColorMask( ctx, 
+    STATE_INIT(ColorMask)( ctx, 
             ctx->Color.ColorMask[RCOMP],
             ctx->Color.ColorMask[GCOMP],
             ctx->Color.ColorMask[BCOMP],
             ctx->Color.ColorMask[ACOMP]);
 
-    ctx->Driver.CullFace( ctx, ctx->Polygon.CullFaceMode );
-    ctx->Driver.DepthFunc( ctx, ctx->Depth.Func );
-    ctx->Driver.DepthMask( ctx, ctx->Depth.Mask );
-
-    ctx->Driver.Enable( ctx, GL_ALPHA_TEST, ctx->Color.AlphaEnabled );
-    ctx->Driver.Enable( ctx, GL_BLEND, ctx->Color.BlendEnabled );
-    ctx->Driver.Enable( ctx, GL_COLOR_LOGIC_OP, ctx->Color.ColorLogicOpEnabled );
-    ctx->Driver.Enable( ctx, GL_COLOR_SUM, ctx->Fog.ColorSumEnabled );
-    ctx->Driver.Enable( ctx, GL_CULL_FACE, ctx->Polygon.CullFlag );
-    ctx->Driver.Enable( ctx, GL_DEPTH_TEST, ctx->Depth.Test );
-    ctx->Driver.Enable( ctx, GL_DITHER, ctx->Color.DitherFlag );
-    ctx->Driver.Enable( ctx, GL_FOG, ctx->Fog.Enabled );
-    ctx->Driver.Enable( ctx, GL_LIGHTING, ctx->Light.Enabled );
-    ctx->Driver.Enable( ctx, GL_LINE_SMOOTH, ctx->Line.SmoothFlag );
-    ctx->Driver.Enable( ctx, GL_POLYGON_STIPPLE, ctx->Polygon.StippleFlag );
-    ctx->Driver.Enable( ctx, GL_SCISSOR_TEST, ctx->Scissor.Enabled );
-    ctx->Driver.Enable( ctx, GL_STENCIL_TEST, ctx->Stencil.Enabled );
-    ctx->Driver.Enable( ctx, GL_TEXTURE_1D, GL_FALSE );
-    ctx->Driver.Enable( ctx, GL_TEXTURE_2D, GL_FALSE );
-    ctx->Driver.Enable( ctx, GL_TEXTURE_RECTANGLE_NV, GL_FALSE );
-    ctx->Driver.Enable( ctx, GL_TEXTURE_3D, GL_FALSE );
-    ctx->Driver.Enable( ctx, GL_TEXTURE_CUBE_MAP, GL_FALSE );
-
-    ctx->Driver.Fogfv( ctx, GL_FOG_COLOR, ctx->Fog.Color );
-    ctx->Driver.Fogfv( ctx, GL_FOG_MODE, 0 );
-    ctx->Driver.Fogfv( ctx, GL_FOG_DENSITY, &ctx->Fog.Density );
-    ctx->Driver.Fogfv( ctx, GL_FOG_START, &ctx->Fog.Start );
-    ctx->Driver.Fogfv( ctx, GL_FOG_END, &ctx->Fog.End );
-
-    ctx->Driver.FrontFace( ctx, ctx->Polygon.FrontFace );
+    STATE_INIT(CullFace)( ctx, ctx->Polygon.CullFaceMode );
+    STATE_INIT(DepthFunc)( ctx, ctx->Depth.Func );
+    STATE_INIT(DepthMask)( ctx, ctx->Depth.Mask );
+
+    STATE_INIT(Enable)( ctx, GL_ALPHA_TEST, ctx->Color.AlphaEnabled );
+    STATE_INIT(Enable)( ctx, GL_BLEND, ctx->Color.BlendEnabled );
+    STATE_INIT(Enable)( ctx, GL_COLOR_LOGIC_OP, ctx->Color.ColorLogicOpEnabled );
+    STATE_INIT(Enable)( ctx, GL_COLOR_SUM, ctx->Fog.ColorSumEnabled );
+    STATE_INIT(Enable)( ctx, GL_CULL_FACE, ctx->Polygon.CullFlag );
+    STATE_INIT(Enable)( ctx, GL_DEPTH_TEST, ctx->Depth.Test );
+    STATE_INIT(Enable)( ctx, GL_DITHER, ctx->Color.DitherFlag );
+    STATE_INIT(Enable)( ctx, GL_FOG, ctx->Fog.Enabled );
+    STATE_INIT(Enable)( ctx, GL_LIGHTING, ctx->Light.Enabled );
+    STATE_INIT(Enable)( ctx, GL_LINE_SMOOTH, ctx->Line.SmoothFlag );
+    STATE_INIT(Enable)( ctx, GL_POLYGON_STIPPLE, ctx->Polygon.StippleFlag );
+    STATE_INIT(Enable)( ctx, GL_SCISSOR_TEST, ctx->Scissor.Enabled );
+    STATE_INIT(Enable)( ctx, GL_STENCIL_TEST, ctx->Stencil.Enabled );
+    STATE_INIT(Enable)( ctx, GL_TEXTURE_1D, GL_FALSE );
+    STATE_INIT(Enable)( ctx, GL_TEXTURE_2D, GL_FALSE );
+    STATE_INIT(Enable)( ctx, GL_TEXTURE_RECTANGLE_NV, GL_FALSE );
+    STATE_INIT(Enable)( ctx, GL_TEXTURE_3D, GL_FALSE );
+    STATE_INIT(Enable)( ctx, GL_TEXTURE_CUBE_MAP, GL_FALSE );
+
+    STATE_INIT(Fogfv)( ctx, GL_FOG_COLOR, ctx->Fog.Color );
+    STATE_INIT(Fogfv)( ctx, GL_FOG_MODE, 0 );
+    STATE_INIT(Fogfv)( ctx, GL_FOG_DENSITY, &ctx->Fog.Density );
+    STATE_INIT(Fogfv)( ctx, GL_FOG_START, &ctx->Fog.Start );
+    STATE_INIT(Fogfv)( ctx, GL_FOG_END, &ctx->Fog.End );
+
+    STATE_INIT(FrontFace)( ctx, ctx->Polygon.FrontFace );
 
     {
         GLfloat f = (GLfloat)ctx->Light.Model.ColorControl;
-        ctx->Driver.LightModelfv( ctx, GL_LIGHT_MODEL_COLOR_CONTROL, &f );
+        STATE_INIT(LightModelfv)( ctx, GL_LIGHT_MODEL_COLOR_CONTROL, &f );
     }
 
-    ctx->Driver.LineWidth( ctx, ctx->Line.Width );
-    ctx->Driver.LogicOpcode( ctx, ctx->Color.LogicOp );
-    ctx->Driver.PointSize( ctx, ctx->Point.Size );
-    ctx->Driver.PolygonStipple( ctx, (const GLubyte *)ctx->PolygonStipple );
-    ctx->Driver.Scissor( ctx, ctx->Scissor.X, ctx->Scissor.Y,
+    STATE_INIT(LineWidth)( ctx, ctx->Line.Width );
+    STATE_INIT(LogicOpcode)( ctx, ctx->Color.LogicOp );
+    STATE_INIT(PointSize)( ctx, ctx->Point.Size );
+    STATE_INIT(PolygonStipple)( ctx, (const GLubyte *)ctx->PolygonStipple );
+    STATE_INIT(Scissor)( ctx, ctx->Scissor.X, ctx->Scissor.Y,
             ctx->Scissor.Width, ctx->Scissor.Height );
-    ctx->Driver.ShadeModel( ctx, ctx->Light.ShadeModel );
-    ctx->Driver.StencilFuncSeparate( ctx, GL_FRONT,
+    STATE_INIT(ShadeModel)( ctx, ctx->Light.ShadeModel );
+    STATE_INIT(StencilFuncSeparate)( ctx, GL_FRONT,
             ctx->Stencil.Function[0],
             ctx->Stencil.Ref[0],
             ctx->Stencil.ValueMask[0] );
-    ctx->Driver.StencilFuncSeparate( ctx, GL_BACK,
+    STATE_INIT(StencilFuncSeparate)( ctx, GL_BACK,
             ctx->Stencil.Function[1],
             ctx->Stencil.Ref[1],
             ctx->Stencil.ValueMask[1] );
-    ctx->Driver.StencilMaskSeparate( ctx, GL_FRONT, ctx->Stencil.WriteMask[0] );
-    ctx->Driver.StencilMaskSeparate( ctx, GL_BACK, ctx->Stencil.WriteMask[1] );
-    ctx->Driver.StencilOpSeparate( ctx, GL_FRONT,
+    STATE_INIT(StencilMaskSeparate)( ctx, GL_FRONT, ctx->Stencil.WriteMask[0] );
+    STATE_INIT(StencilMaskSeparate)( ctx, GL_BACK, ctx->Stencil.WriteMask[1] );
+    STATE_INIT(StencilOpSeparate)( ctx, GL_FRONT,
             ctx->Stencil.FailFunc[0],
             ctx->Stencil.ZFailFunc[0],
             ctx->Stencil.ZPassFunc[0]);
-    ctx->Driver.StencilOpSeparate( ctx, GL_BACK,
+    STATE_INIT(StencilOpSeparate)( ctx, GL_BACK,
             ctx->Stencil.FailFunc[1],
             ctx->Stencil.ZFailFunc[1],
             ctx->Stencil.ZPassFunc[1]);
 
-    ctx->Driver.DrawBuffer( ctx, ctx->Color.DrawBuffer[0] );
+    STATE_INIT(DrawBuffer)( ctx, ctx->Color.DrawBuffer[0] );
 }
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_state_cache.c b/src/mesa/drivers/dri/nouveau/nouveau_state_cache.c
new file mode 100644
index 00000000000..36f0c1024b6
--- /dev/null
+++ b/src/mesa/drivers/dri/nouveau/nouveau_state_cache.c
@@ -0,0 +1,64 @@
+
+#include "nouveau_state_cache.h"
+#include "nouveau_context.h"
+#include "nouveau_object.h"
+#include "nouveau_fifo.h"
+
+#define BEGIN_RING_NOFLUSH(subchannel,tag,size) do {					\
+	if (nmesa->fifo.free <= (size))							\
+		WAIT_RING(nmesa,(size));						\
+	OUT_RING( ((size)<<18) | ((subchannel) << 13) | (tag));				\
+	nmesa->fifo.free -= ((size) + 1);                                               \
+}while(0)
+
+// flush all the dirty state
+void nouveau_state_cache_flush(nouveauContextPtr nmesa)
+{
+	int i=0;
+	int run=0;
+
+	// fast-path no state changes
+	if (!nmesa->state_cache.dirty)
+		return;
+	nmesa->state_cache.dirty=0;
+
+	do
+	{
+		// jump to a dirty state
+		while((nmesa->state_cache.atoms[i].dirty==0)&&(i<NOUVEAU_STATE_CACHE_ENTRIES))
+			i++;
+
+		// figure out a run of dirty values
+		run=0;
+		while((nmesa->state_cache.atoms[i+run].dirty)&&(i+run<NOUVEAU_STATE_CACHE_ENTRIES))
+			run++;
+
+		// output everything as a single run
+		if (run>0) {
+			int j;
+
+			BEGIN_RING_NOFLUSH(NvSub3D, i*4, run);
+			for(j=0;j<run;j++)
+			{
+				OUT_RING(nmesa->state_cache.atoms[i+j].value);
+				nmesa->state_cache.atoms[i+j].dirty=0;
+			}
+			i+=run;
+		}
+	}
+	while(i<NOUVEAU_STATE_CACHE_ENTRIES);
+}
+
+
+// inits the state cache
+void nouveau_state_cache_init(nouveauContextPtr nmesa)
+{
+	int i;
+	for(i=0;i<NOUVEAU_STATE_CACHE_ENTRIES;i++)
+	{
+		nmesa->state_cache.atoms[i].dirty=0;
+		nmesa->state_cache.atoms[i].value=0xDEADBEEF; // nvidia cards like beef
+	}
+	nmesa->state_cache.dirty=0;
+}
+
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_state_cache.h b/src/mesa/drivers/dri/nouveau/nouveau_state_cache.h
new file mode 100644
index 00000000000..24882748468
--- /dev/null
+++ b/src/mesa/drivers/dri/nouveau/nouveau_state_cache.h
@@ -0,0 +1,23 @@
+
+#ifndef __NOUVEAU_STATE_CACHE_H__
+#define __NOUVEAU_STATE_CACHE_H__
+
+#include "mtypes.h"
+
+#define NOUVEAU_STATE_CACHE_ENTRIES 2048
+
+typedef struct nouveau_state_atom_t{
+	uint32_t value;
+	uint32_t dirty;
+}nouveau_state_atom;
+
+typedef struct nouveau_state_cache_t{
+	nouveau_state_atom atoms[NOUVEAU_STATE_CACHE_ENTRIES];
+	uint32_t current_pos;
+	// master dirty flag
+	uint32_t dirty;
+}nouveau_state_cache;
+
+
+#endif
+
diff --git a/src/mesa/drivers/dri/nouveau/nv10_state.c b/src/mesa/drivers/dri/nouveau/nv10_state.c
index 368235ac585..0d19ce94af4 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_state.c
+++ b/src/mesa/drivers/dri/nouveau/nv10_state.c
@@ -40,9 +40,9 @@ void nv10AlphaFunc(GLcontext *ctx, GLenum func, GLfloat ref)
 	GLubyte ubRef;
 	CLAMPED_FLOAT_TO_UBYTE(ubRef, ref);
 
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC, 2);
-	OUT_RING(func);     /* NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC */
-	OUT_RING(ubRef);    /* NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_REF  */
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC, 2);
+	OUT_RING_CACHE(func);     /* NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC */
+	OUT_RING_CACHE(ubRef);    /* NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_REF  */
 }
 
 void nv10BlendColor(GLcontext *ctx, const GLfloat color[4])
@@ -55,15 +55,15 @@ void nv10BlendColor(GLcontext *ctx, const GLfloat color[4])
 	CLAMPED_FLOAT_TO_UBYTE(cf[2], color[2]);
 	CLAMPED_FLOAT_TO_UBYTE(cf[3], color[3]);
 
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_COLOR, 1);
-	OUT_RING(PACK_COLOR_8888(cf[3], cf[1], cf[2], cf[0]));
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_COLOR, 1);
+	OUT_RING_CACHE(PACK_COLOR_8888(cf[3], cf[1], cf[2], cf[0]));
 }
 
 void nv10BlendEquationSeparate(GLcontext *ctx, GLenum modeRGB, GLenum modeA)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_EQUATION, 1);
-	OUT_RING((modeA<<16) | modeRGB);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_EQUATION, 1);
+	OUT_RING_CACHE((modeA<<16) | modeRGB);
 }
 
 
@@ -71,9 +71,9 @@ void nv10BlendFuncSeparate(GLcontext *ctx, GLenum sfactorRGB, GLenum dfactorRGB,
 		GLenum sfactorA, GLenum dfactorA)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_FUNC_SRC, 2);
-	OUT_RING((sfactorA<<16) | sfactorRGB);
-	OUT_RING((dfactorA<<16) | dfactorRGB);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_FUNC_SRC, 2);
+	OUT_RING_CACHE((sfactorA<<16) | sfactorRGB);
+	OUT_RING_CACHE((dfactorA<<16) | dfactorRGB);
 }
 
 /*
@@ -82,16 +82,16 @@ void nv30ClearColor(GLcontext *ctx, const GLfloat color[4])
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	GLubyte c[4];
 	UNCLAMPED_FLOAT_TO_RGBA_CHAN(c,color);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_ARGB, 1);
-	OUT_RING(PACK_COLOR_8888(c[3],c[0],c[1],c[2]));
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_ARGB, 1);
+	OUT_RING_CACHE(PACK_COLOR_8888(c[3],c[0],c[1],c[2]));
 }
 
 void nv30ClearDepth(GLcontext *ctx, GLclampd d)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	nmesa->clear_value=((nmesa->clear_value&0x000000FF)|(((uint32_t)(d*0xFFFFFF))<<8));
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
-	OUT_RING(nmesa->clear_value);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
+	OUT_RING_CACHE(nmesa->clear_value);
 }
 */
 
@@ -104,19 +104,19 @@ void nv30ClearStencil(GLcontext *ctx, GLint s)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	nmesa->clear_value=((nmesa->clear_value&0xFFFFFF00)|(s&0x000000FF));
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
-	OUT_RING(nmesa->clear_value);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
+	OUT_RING_CACHE(nmesa->clear_value);
 }
 */
 
 void nv10ClipPlane(GLcontext *ctx, GLenum plane, const GLfloat *equation)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CLIP_PLANE_A(plane), 4);
-	OUT_RINGf(equation[0]);
-	OUT_RINGf(equation[1]);
-	OUT_RINGf(equation[2]);
-	OUT_RINGf(equation[3]);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CLIP_PLANE_A(plane), 4);
+	OUT_RING_CACHEf(equation[0]);
+	OUT_RING_CACHEf(equation[1]);
+	OUT_RING_CACHEf(equation[2]);
+	OUT_RING_CACHEf(equation[3]);
 }
 
 /* Seems does not support alpha in color mask */
@@ -124,8 +124,8 @@ void nv10ColorMask(GLcontext *ctx, GLboolean rmask, GLboolean gmask,
 		GLboolean bmask, GLboolean amask )
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_COLOR_MASK, 1);
-	OUT_RING(/*((amask && 0x01) << 24) |*/ ((rmask && 0x01) << 16) | ((gmask && 0x01)<< 8) | ((bmask && 0x01) << 0));
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_COLOR_MASK, 1);
+	OUT_RING_CACHE(/*((amask && 0x01) << 24) |*/ ((rmask && 0x01) << 16) | ((gmask && 0x01)<< 8) | ((bmask && 0x01) << 0));
 }
 
 void nv10ColorMaterial(GLcontext *ctx, GLenum face, GLenum mode)
@@ -136,37 +136,37 @@ void nv10ColorMaterial(GLcontext *ctx, GLenum face, GLenum mode)
 void nv10CullFace(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CULL_FACE, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CULL_FACE, 1);
+	OUT_RING_CACHE(mode);
 }
 
 void nv10FrontFace(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_FRONT_FACE, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_FRONT_FACE, 1);
+	OUT_RING_CACHE(mode);
 }
 
 void nv10DepthFunc(GLcontext *ctx, GLenum func)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_FUNC, 1);
-	OUT_RING(func);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_FUNC, 1);
+	OUT_RING_CACHE(func);
 }
 
 void nv10DepthMask(GLcontext *ctx, GLboolean flag)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_WRITE_ENABLE, 1);
-	OUT_RING(flag);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_WRITE_ENABLE, 1);
+	OUT_RING_CACHE(flag);
 }
 
 void nv10DepthRange(GLcontext *ctx, GLclampd nearval, GLclampd farval)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_RANGE_NEAR, 2);
-	OUT_RINGf(nearval);
-	OUT_RINGf(farval);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_RANGE_NEAR, 2);
+	OUT_RING_CACHEf(nearval);
+	OUT_RING_CACHEf(farval);
 }
 
 /** Specify the current buffer for writing */
@@ -180,13 +180,13 @@ void nv10Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 	switch(cap)
 	{
 		case GL_ALPHA_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ALPHA_FUNC_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_AUTO_NORMAL:
 		case GL_BLEND:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_BLEND_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_CLIP_PLANE0:
 		case GL_CLIP_PLANE1:
@@ -194,12 +194,12 @@ void nv10Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 		case GL_CLIP_PLANE3:
 		case GL_CLIP_PLANE4:
 		case GL_CLIP_PLANE5:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CLIP_PLANE_ENABLE(cap-GL_CLIP_PLANE0), 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CLIP_PLANE_ENABLE(cap-GL_CLIP_PLANE0), 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_COLOR_LOGIC_OP:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_COLOR_MATERIAL:
 //		case GL_COLOR_SUM_EXT:
@@ -207,20 +207,20 @@ void nv10Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_CONVOLUTION_1D:
 //		case GL_CONVOLUTION_2D:
 		case GL_CULL_FACE:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CULL_FACE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_CULL_FACE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_DEPTH_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_TEST_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DEPTH_TEST_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_DITHER:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DITHER_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_DITHER_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_FOG:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_FOG_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_FOG_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_HISTOGRAM:
 //		case GL_INDEX_LOGIC_OP:
@@ -237,22 +237,22 @@ void nv10Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 			nmesa->enabled_lights=((nmesa->enabled_lights&mask)|(mask*state));
 			if (nmesa->lighting_enabled)
 			{
-				BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
-				OUT_RING(nmesa->enabled_lights);
+				BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
+				OUT_RING_CACHE(nmesa->enabled_lights);
 			}
 			break;
 			}
 		case GL_LIGHTING:
 			nmesa->lighting_enabled=state;
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
 			if (nmesa->lighting_enabled)
-				OUT_RING(nmesa->enabled_lights);
+				OUT_RING_CACHE(nmesa->enabled_lights);
 			else
-				OUT_RING(0x0);
+				OUT_RING_CACHE(0x0);
 			break;
 		case GL_LINE_SMOOTH:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LINE_SMOOTH_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LINE_SMOOTH_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_LINE_STIPPLE:
 //		case GL_MAP1_COLOR_4:
@@ -275,29 +275,29 @@ void nv10Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_MAP2_VERTEX_4:
 //		case GL_MINMAX:
 		case GL_NORMALIZE:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_NORMALIZE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_NORMALIZE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_POINT_SMOOTH:
 		case GL_POLYGON_OFFSET_POINT:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_OFFSET_POINT_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_OFFSET_POINT_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_OFFSET_LINE:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_OFFSET_LINE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_OFFSET_LINE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_OFFSET_FILL:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_OFFSET_FILL_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_OFFSET_FILL_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_SMOOTH:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_SMOOTH_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POLYGON_SMOOTH_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POINT_SMOOTH:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POINT_SMOOTH_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POINT_SMOOTH_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_POLYGON_STIPPLE:
 //		case GL_POST_COLOR_MATRIX_COLOR_TABLE:
@@ -306,8 +306,8 @@ void nv10Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_SCISSOR_TEST:
 //		case GL_SEPARABLE_2D:
 		case GL_STENCIL_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_TEXTURE_GEN_Q:
 //		case GL_TEXTURE_GEN_R:
@@ -325,8 +325,8 @@ void nv10Fogfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
     switch(pname)
     {
         case GL_FOG_MODE:
-            BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_FOG_MODE, 1);
-            //OUT_RING (params);
+            BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_FOG_MODE, 1);
+            //OUT_RING_CACHE (params);
             break;
             /* TODO: unsure about the rest.*/
         default:
@@ -349,60 +349,60 @@ void nv10Lightfv(GLcontext *ctx, GLenum light, GLenum pname, const GLfloat *para
 	switch(pname)
 	{
 		case GL_AMBIENT:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LIGHT_AMBIENT(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LIGHT_AMBIENT(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_DIFFUSE:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LIGHT_DIFFUSE(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LIGHT_DIFFUSE(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPECULAR:
-			BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LIGHT_SPECULAR(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LIGHT_SPECULAR(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 #if 0
 		case GL_SPOT_DIRECTION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_DIR_X(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_DIR_X(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_POSITION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_POSITION_X(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_POSITION_X(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPOT_EXPONENT:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_EXPONENT(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_EXPONENT(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_SPOT_CUTOFF:
 			/* you can't factor these */
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_A(light), 1);
-			OUT_RINGf(params[0]);
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_B(light), 1);
-			OUT_RINGf(params[1]);
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_C(light), 1);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_A(light), 1);
+			OUT_RING_CACHEf(params[0]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_B(light), 1);
+			OUT_RING_CACHEf(params[1]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_C(light), 1);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_CONSTANT_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_CONSTANT_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_CONSTANT_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_LINEAR_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_LINEAR_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_LINEAR_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_QUADRATIC_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_QUADRATIC_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_QUADRATIC_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 #endif
 		default:
@@ -417,23 +417,23 @@ void (*LightModelfv)(GLcontext *ctx, GLenum pname, const GLfloat *params);
 void nv30LineStipple(GLcontext *ctx, GLint factor, GLushort pattern )
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_STIPPLE_PATTERN, 1);
-	OUT_RING((pattern << 16) | factor);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_STIPPLE_PATTERN, 1);
+	OUT_RING_CACHE((pattern << 16) | factor);
 }
 
 void nv30LineWidth(GLcontext *ctx, GLfloat width)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_WIDTH_SMOOTH, 1);
-	OUT_RINGf(width);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_WIDTH_SMOOTH, 1);
+	OUT_RING_CACHEf(width);
 }
 */
 
 void nv10LogicOpcode(GLcontext *ctx, GLenum opcode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LOGIC_OP, 1);
-	OUT_RING(opcode);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_LOGIC_OP, 1);
+	OUT_RING_CACHE(opcode);
 }
 
 void nv10PointParameterfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
@@ -447,8 +447,8 @@ void nv10PointParameterfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
 void nv10PointSize(GLcontext *ctx, GLfloat size)
 {
     nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-    BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POINT_SIZE, 1);
-    OUT_RINGf(size);
+    BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_POINT_SIZE, 1);
+    OUT_RING_CACHEf(size);
 }
 
 /** Select a polygon rasterization mode */
@@ -469,8 +469,8 @@ void nv10ShadeModel(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_SHADE_MODEL, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_SHADE_MODEL, 1);
+	OUT_RING_CACHE(mode);
 }
 
 /** OpenGL 2.0 two-sided StencilFunc */
@@ -479,10 +479,10 @@ static void nv10StencilFuncSeparate(GLcontext *ctx, GLenum face, GLenum func,
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_FUNC_FUNC, 3);
-	OUT_RING(func);
-	OUT_RING(ref);
-	OUT_RING(mask);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_FUNC_FUNC, 3);
+	OUT_RING_CACHE(func);
+	OUT_RING_CACHE(ref);
+	OUT_RING_CACHE(mask);
 }
 
 /** OpenGL 2.0 two-sided StencilMask */
@@ -490,8 +490,8 @@ static void nv10StencilMaskSeparate(GLcontext *ctx, GLenum face, GLuint mask)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_MASK, 1);
-	OUT_RING(mask);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_MASK, 1);
+	OUT_RING_CACHE(mask);
 }
 
 /** OpenGL 2.0 two-sided StencilOp */
@@ -500,10 +500,10 @@ static void nv10StencilOpSeparate(GLcontext *ctx, GLenum face, GLenum fail,
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_OP_FAIL, 1);
-	OUT_RING(fail);
-	OUT_RING(zfail);
-	OUT_RING(zpass);
+	BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_STENCIL_OP_FAIL, 1);
+	OUT_RING_CACHE(fail);
+	OUT_RING_CACHE(zfail);
+	OUT_RING_CACHE(zpass);
 }
 
 /** Control the generation of texture coordinates */
@@ -523,8 +523,8 @@ void nv10Viewport(GLcontext *ctx, GLint x, GLint y, GLsizei w, GLsizei h)
 {
     /* TODO: Where do the VIEWPORT_XFRM_* regs come in? */
     nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-    BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_VIEWPORT_HORIZ, 2);
-    OUT_RING((w << 16) | x);
-    OUT_RING((h << 16) | y);
+    BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_VIEWPORT_HORIZ, 2);
+    OUT_RING_CACHE((w << 16) | x);
+    OUT_RING_CACHE((h << 16) | y);
 }
 
diff --git a/src/mesa/drivers/dri/nouveau/nv10_swtcl.c b/src/mesa/drivers/dri/nouveau/nv10_swtcl.c
index 832ce4ad21d..198e3a2668a 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_swtcl.c
+++ b/src/mesa/drivers/dri/nouveau/nv10_swtcl.c
@@ -61,6 +61,7 @@ static void nv10ResetLineStipple( GLcontext *ctx );
 
 static inline void nv10StartPrimitive(struct nouveau_context* nmesa,uint32_t primitive,uint32_t size)
 {
+	// FIXME the primitive type can probably go trough the caching system as well
 	if (nmesa->screen->card->type==NV_10)
 		BEGIN_RING_SIZE(NvSub3D,NV10_TCL_PRIMITIVE_3D_BEGIN_END,1);
 	else if (nmesa->screen->card->type==NV_20)
@@ -79,6 +80,7 @@ static inline void nv10StartPrimitive(struct nouveau_context* nmesa,uint32_t pri
 
 inline void nv10FinishPrimitive(struct nouveau_context *nmesa)
 {
+	// FIXME this is probably not needed
 	if (nmesa->screen->card->type==NV_10)
 		BEGIN_RING_SIZE(NvSub3D,NV10_TCL_PRIMITIVE_3D_BEGIN_END,1);
 	else if (nmesa->screen->card->type==NV_20)
@@ -216,23 +218,6 @@ static void nv10_render_noop_verts(GLcontext *ctx,GLuint start,GLuint count,GLui
 {
 }
 
-static inline void nv10_render_generic_primitive_elts(GLcontext *ctx,GLuint start,GLuint count,GLuint flags,GLuint prim)
-{
-	struct nouveau_context *nmesa = NOUVEAU_CONTEXT(ctx);
-	GLubyte *vertptr = (GLubyte *)nmesa->verts;
-	GLuint vertsize = nmesa->vertex_size;
-	GLuint size_dword = vertsize*(count-start);
-	const GLuint * const elt = TNL_CONTEXT(ctx)->vb.Elts;
-	GLuint j;
-
-	nv10ExtendPrimitive(nmesa, size_dword);
-	nv10StartPrimitive(nmesa,prim+1,size_dword);
-	for (j=start; j<count; j++ ) {
-		OUT_RINGp((nouveauVertex*)(vertptr+(elt[j]*vertsize*4)),vertsize);
-	}
-	nv10FinishPrimitive(nmesa);
-}
-
 static void (*nv10_render_tab_verts[GL_POLYGON+2])(GLcontext *,
 							   GLuint,
 							   GLuint,
@@ -252,6 +237,23 @@ static void (*nv10_render_tab_verts[GL_POLYGON+2])(GLcontext *,
 };
 
 
+static inline void nv10_render_generic_primitive_elts(GLcontext *ctx,GLuint start,GLuint count,GLuint flags,GLuint prim)
+{
+	struct nouveau_context *nmesa = NOUVEAU_CONTEXT(ctx);
+	GLubyte *vertptr = (GLubyte *)nmesa->verts;
+	GLuint vertsize = nmesa->vertex_size;
+	GLuint size_dword = vertsize*(count-start);
+	const GLuint * const elt = TNL_CONTEXT(ctx)->vb.Elts;
+	GLuint j;
+
+	nv10ExtendPrimitive(nmesa, size_dword);
+	nv10StartPrimitive(nmesa,prim+1,size_dword);
+	for (j=start; j<count; j++ ) {
+		OUT_RINGp((nouveauVertex*)(vertptr+(elt[j]*vertsize*4)),vertsize);
+	}
+	nv10FinishPrimitive(nmesa);
+}
+
 static void nv10_render_points_elts(GLcontext *ctx,GLuint start,GLuint count,GLuint flags)
 {
 	nv10_render_generic_primitive_elts(ctx,start,count,flags,GL_POINTS);
@@ -440,8 +442,8 @@ static inline void nv10OutputVertexFormat(struct nouveau_context* nmesa)
 		size = attr_size[j] << 4;	\
 		size |= (attr_size[j]*4) << 8;	\
 		size |= NV_VERTEX_ATTRIBUTE_TYPE_FLOAT;	\
-		BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_VERTEX_ATTR(i),1);	\
-		OUT_RING(size);	\
+		BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_VERTEX_ATTR(i),1);	\
+		OUT_RING_CACHE(size);	\
 	} while (0)
 
 		NV10_SET_VERTEX_ATTRIB(0, _TNL_ATTRIB_POS);
@@ -453,22 +455,23 @@ static inline void nv10OutputVertexFormat(struct nouveau_context* nmesa)
 		NV10_SET_VERTEX_ATTRIB(6, _TNL_ATTRIB_WEIGHT);
 		NV10_SET_VERTEX_ATTRIB(7, _TNL_ATTRIB_FOG);
 
-		BEGIN_RING_SIZE(NvSub3D, NV10_TCL_PRIMITIVE_3D_VERTEX_ARRAY_VALIDATE,1);
-		OUT_RING(0);
+		BEGIN_RING_CACHE(NvSub3D, NV10_TCL_PRIMITIVE_3D_VERTEX_ARRAY_VALIDATE,1);
+		OUT_RING_CACHE(0);
 	} else if (nmesa->screen->card->type==NV_20) {
 		for(i=0;i<16;i++)
 		{
 			int size=attr_size[i];
-			BEGIN_RING_SIZE(NvSub3D,NV20_TCL_PRIMITIVE_3D_VERTEX_ATTR(i),1);
-			OUT_RING(NV_VERTEX_ATTRIBUTE_TYPE_FLOAT|(size*0x10));
+			BEGIN_RING_CACHE(NvSub3D,NV20_TCL_PRIMITIVE_3D_VERTEX_ATTR(i),1);
+			OUT_RING_CACHE(NV_VERTEX_ATTRIBUTE_TYPE_FLOAT|(size*0x10));
 		}
 	} else {
-		BEGIN_RING_SIZE(NvSub3D,NV30_TCL_PRIMITIVE_3D_VERTEX_ATTR0_POS,slots);
+		BEGIN_RING_CACHE(NvSub3D,NV30_TCL_PRIMITIVE_3D_VERTEX_ATTR0_POS,slots);
 		for(i=0;i<slots;i++)
 		{
 			int size=attr_size[i];
-			OUT_RING(NV_VERTEX_ATTRIBUTE_TYPE_FLOAT|(size*0x10));
+			OUT_RING_CACHE(NV_VERTEX_ATTRIBUTE_TYPE_FLOAT|(size*0x10));
 		}
+		// FIXME this is probably not needed
 		BEGIN_RING_SIZE(NvSub3D,NV30_TCL_PRIMITIVE_3D_VERTEX_UNK_0,1);
 		OUT_RING(0);
 		BEGIN_RING_SIZE(NvSub3D,NV30_TCL_PRIMITIVE_3D_VERTEX_UNK_0,1);
diff --git a/src/mesa/drivers/dri/nouveau/nv20_state.c b/src/mesa/drivers/dri/nouveau/nv20_state.c
index 3ad5f5bc44b..7d93368a536 100644
--- a/src/mesa/drivers/dri/nouveau/nv20_state.c
+++ b/src/mesa/drivers/dri/nouveau/nv20_state.c
@@ -40,9 +40,9 @@ static void nv20AlphaFunc(GLcontext *ctx, GLenum func, GLfloat ref)
 	GLubyte ubRef;
 	CLAMPED_FLOAT_TO_UBYTE(ubRef, ref);
 
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC, 2);
-	OUT_RING(func);     /* NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC */
-	OUT_RING(ubRef);    /* NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_REF  */
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC, 2);
+	OUT_RING_CACHE(func);     /* NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC */
+	OUT_RING_CACHE(ubRef);    /* NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_REF  */
 }
 
 static void nv20BlendColor(GLcontext *ctx, const GLfloat color[4])
@@ -55,15 +55,15 @@ static void nv20BlendColor(GLcontext *ctx, const GLfloat color[4])
 	CLAMPED_FLOAT_TO_UBYTE(cf[2], color[2]);
 	CLAMPED_FLOAT_TO_UBYTE(cf[3], color[3]);
 
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_COLOR, 1);
-	OUT_RING(PACK_COLOR_8888(cf[3], cf[1], cf[2], cf[0]));
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_COLOR, 1);
+	OUT_RING_CACHE(PACK_COLOR_8888(cf[3], cf[1], cf[2], cf[0]));
 }
 
 static void nv20BlendEquationSeparate(GLcontext *ctx, GLenum modeRGB, GLenum modeA)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_EQUATION, 1);
-	OUT_RING((modeA<<16) | modeRGB);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_EQUATION, 1);
+	OUT_RING_CACHE((modeA<<16) | modeRGB);
 }
 
 
@@ -71,9 +71,9 @@ static void nv20BlendFuncSeparate(GLcontext *ctx, GLenum sfactorRGB, GLenum dfac
 		GLenum sfactorA, GLenum dfactorA)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_FUNC_SRC, 2);
-	OUT_RING((sfactorA<<16) | sfactorRGB);
-	OUT_RING((dfactorA<<16) | dfactorRGB);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_FUNC_SRC, 2);
+	OUT_RING_CACHE((sfactorA<<16) | sfactorRGB);
+	OUT_RING_CACHE((dfactorA<<16) | dfactorRGB);
 }
 
 static void nv20ClearColor(GLcontext *ctx, const GLfloat color[4])
@@ -81,16 +81,16 @@ static void nv20ClearColor(GLcontext *ctx, const GLfloat color[4])
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	GLubyte c[4];
 	UNCLAMPED_FLOAT_TO_RGBA_CHAN(c,color);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLEAR_VALUE_ARGB, 1);
-	OUT_RING(PACK_COLOR_8888(c[3],c[0],c[1],c[2]));
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLEAR_VALUE_ARGB, 1);
+	OUT_RING_CACHE(PACK_COLOR_8888(c[3],c[0],c[1],c[2]));
 }
 
 static void nv20ClearDepth(GLcontext *ctx, GLclampd d)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	nmesa->clear_value=((nmesa->clear_value&0x000000FF)|(((uint32_t)(d*0xFFFFFF))<<8));
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
-	OUT_RING(nmesa->clear_value);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
+	OUT_RING_CACHE(nmesa->clear_value);
 }
 
 /* we're don't support indexed buffers
@@ -101,26 +101,26 @@ static void nv20ClearStencil(GLcontext *ctx, GLint s)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	nmesa->clear_value=((nmesa->clear_value&0xFFFFFF00)|(s&0x000000FF));
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
-	OUT_RING(nmesa->clear_value);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
+	OUT_RING_CACHE(nmesa->clear_value);
 }
 
 static void nv20ClipPlane(GLcontext *ctx, GLenum plane, const GLfloat *equation)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLIP_PLANE_A(plane), 4);
-	OUT_RINGf(equation[0]);
-	OUT_RINGf(equation[1]);
-	OUT_RINGf(equation[2]);
-	OUT_RINGf(equation[3]);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLIP_PLANE_A(plane), 4);
+	OUT_RING_CACHEf(equation[0]);
+	OUT_RING_CACHEf(equation[1]);
+	OUT_RING_CACHEf(equation[2]);
+	OUT_RING_CACHEf(equation[3]);
 }
 
 static void nv20ColorMask(GLcontext *ctx, GLboolean rmask, GLboolean gmask,
 		GLboolean bmask, GLboolean amask )
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_COLOR_MASK, 1);
-	OUT_RING(((amask && 0x01) << 24) | ((rmask && 0x01) << 16) | ((gmask && 0x01)<< 8) | ((bmask && 0x01) << 0));
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_COLOR_MASK, 1);
+	OUT_RING_CACHE(((amask && 0x01) << 24) | ((rmask && 0x01) << 16) | ((gmask && 0x01)<< 8) | ((bmask && 0x01) << 0));
 }
 
 static void nv20ColorMaterial(GLcontext *ctx, GLenum face, GLenum mode)
@@ -131,37 +131,37 @@ static void nv20ColorMaterial(GLcontext *ctx, GLenum face, GLenum mode)
 static void nv20CullFace(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CULL_FACE, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CULL_FACE, 1);
+	OUT_RING_CACHE(mode);
 }
 
 static void nv20FrontFace(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_FRONT_FACE, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_FRONT_FACE, 1);
+	OUT_RING_CACHE(mode);
 }
 
 static void nv20DepthFunc(GLcontext *ctx, GLenum func)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_FUNC, 1);
-	OUT_RING(func);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_FUNC, 1);
+	OUT_RING_CACHE(func);
 }
 
 static void nv20DepthMask(GLcontext *ctx, GLboolean flag)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_WRITE_ENABLE, 1);
-	OUT_RING(flag);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_WRITE_ENABLE, 1);
+	OUT_RING_CACHE(flag);
 }
 
 static void nv20DepthRange(GLcontext *ctx, GLclampd nearval, GLclampd farval)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_RANGE_NEAR, 2);
-	OUT_RINGf(nearval);
-	OUT_RINGf(farval);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_RANGE_NEAR, 2);
+	OUT_RING_CACHEf(nearval);
+	OUT_RING_CACHEf(farval);
 }
 
 /** Specify the current buffer for writing */
@@ -175,13 +175,13 @@ static void nv20Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 	switch(cap)
 	{
 		case GL_ALPHA_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ALPHA_FUNC_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_AUTO_NORMAL:
 		case GL_BLEND:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_FUNC_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_BLEND_FUNC_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_CLIP_PLANE0:
 		case GL_CLIP_PLANE1:
@@ -189,12 +189,12 @@ static void nv20Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 		case GL_CLIP_PLANE3:
 		case GL_CLIP_PLANE4:
 		case GL_CLIP_PLANE5:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLIP_PLANE_ENABLE(cap-GL_CLIP_PLANE0), 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CLIP_PLANE_ENABLE(cap-GL_CLIP_PLANE0), 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_COLOR_LOGIC_OP:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_COLOR_MATERIAL:
 //		case GL_COLOR_SUM_EXT:
@@ -202,20 +202,20 @@ static void nv20Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_CONVOLUTION_1D:
 //		case GL_CONVOLUTION_2D:
 		case GL_CULL_FACE:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CULL_FACE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_CULL_FACE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_DEPTH_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_TEST_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DEPTH_TEST_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_DITHER:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DITHER_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_DITHER_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_FOG:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_FOG_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_FOG_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_HISTOGRAM:
 //		case GL_INDEX_LOGIC_OP:
@@ -232,22 +232,22 @@ static void nv20Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 			nmesa->enabled_lights=((nmesa->enabled_lights&mask)|(mask*state));
 			if (nmesa->lighting_enabled)
 			{
-				BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
-				OUT_RING(nmesa->enabled_lights);
+				BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
+				OUT_RING_CACHE(nmesa->enabled_lights);
 			}
 			break;
 			}
 		case GL_LIGHTING:
 			nmesa->lighting_enabled=state;
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
 			if (nmesa->lighting_enabled)
-				OUT_RING(nmesa->enabled_lights);
+				OUT_RING_CACHE(nmesa->enabled_lights);
 			else
-				OUT_RING(0x0);
+				OUT_RING_CACHE(0x0);
 			break;
 		case GL_LINE_SMOOTH:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LINE_SMOOTH_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LINE_SMOOTH_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_LINE_STIPPLE:
 //		case GL_MAP1_COLOR_4:
@@ -270,29 +270,29 @@ static void nv20Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_MAP2_VERTEX_4:
 //		case GL_MINMAX:
 		case GL_NORMALIZE:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_NORMALIZE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_NORMALIZE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_POINT_SMOOTH:
 		case GL_POLYGON_OFFSET_POINT:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_OFFSET_POINT_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_OFFSET_POINT_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_OFFSET_LINE:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_OFFSET_LINE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_OFFSET_LINE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_OFFSET_FILL:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_OFFSET_FILL_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_OFFSET_FILL_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_SMOOTH:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_SMOOTH_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_SMOOTH_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_STIPPLE:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_STIPPLE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_STIPPLE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_POST_COLOR_MATRIX_COLOR_TABLE:
 //		case GL_POST_CONVOLUTION_COLOR_TABLE:
@@ -301,8 +301,8 @@ static void nv20Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_SEPARABLE_2D:
 		case GL_STENCIL_TEST:
 			// TODO BACK and FRONT ?
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_TEXTURE_GEN_Q:
 //		case GL_TEXTURE_GEN_R:
@@ -320,8 +320,8 @@ static void nv20Fogfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
     switch(pname)
     {
         case GL_FOG_MODE:
-            BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_FOG_MODE, 1);
-            //OUT_RING (params);
+            BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_FOG_MODE, 1);
+            //OUT_RING_CACHE (params);
             break;
             /* TODO: unsure about the rest.*/
         default:
@@ -344,59 +344,59 @@ static void nv20Lightfv(GLcontext *ctx, GLenum light, GLenum pname, const GLfloa
 	switch(pname)
 	{
 		case GL_AMBIENT:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_AMBIENT_A(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_AMBIENT_A(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_DIFFUSE:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_DIFFUSE_A(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_DIFFUSE_A(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPECULAR:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_SPECULAR_A(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_SPECULAR_A(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPOT_DIRECTION:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_DIR_X(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_DIR_X(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_POSITION:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_POSITION_X(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_POSITION_X(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPOT_EXPONENT:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_EXPONENT(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_EXPONENT(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_SPOT_CUTOFF:
 			/* you can't factor these */
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_A(light), 1);
-			OUT_RINGf(params[0]);
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_B(light), 1);
-			OUT_RINGf(params[1]);
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_C(light), 1);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_A(light), 1);
+			OUT_RING_CACHEf(params[0]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_B(light), 1);
+			OUT_RING_CACHEf(params[1]);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_C(light), 1);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_CONSTANT_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_CONSTANT_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_CONSTANT_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_LINEAR_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_LINEAR_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_LINEAR_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_QUADRATIC_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_QUADRATIC_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LIGHT_QUADRATIC_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		default:
 			break;
@@ -410,22 +410,22 @@ static void (*LightModelfv)(GLcontext *ctx, GLenum pname, const GLfloat *params)
 static void nv20LineStipple(GLcontext *ctx, GLint factor, GLushort pattern )
 {
 /*	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LINE_STIPPLE_PATTERN, 1);
-	OUT_RING((pattern << 16) | factor);*/
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LINE_STIPPLE_PATTERN, 1);
+	OUT_RING_CACHE((pattern << 16) | factor);*/
 }
 
 static void nv20LineWidth(GLcontext *ctx, GLfloat width)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LINE_WIDTH, 1);
-	OUT_RINGf(width);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_LINE_WIDTH, 1);
+	OUT_RING_CACHEf(width);
 }
 
 static void nv20LogicOpcode(GLcontext *ctx, GLenum opcode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_OP, 1);
-	OUT_RING(opcode);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_OP, 1);
+	OUT_RING_CACHE(opcode);
 }
 
 static void nv20PointParameterfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
@@ -439,8 +439,8 @@ static void nv20PointParameterfv(GLcontext *ctx, GLenum pname, const GLfloat *pa
 static void nv20PointSize(GLcontext *ctx, GLfloat size)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POINT_SIZE, 1);
-	OUT_RINGf(size);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POINT_SIZE, 1);
+	OUT_RING_CACHEf(size);
 }
 
 /** Select a polygon rasterization mode */
@@ -449,12 +449,12 @@ static void nv20PolygonMode(GLcontext *ctx, GLenum face, GLenum mode)
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
 	if (face == GL_FRONT || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_MODE_FRONT, 1);
-		OUT_RING(mode);
+		BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_MODE_FRONT, 1);
+		OUT_RING_CACHE(mode);
 	}
 	if (face == GL_BACK || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_MODE_BACK, 1);
-		OUT_RING(mode);
+		BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_POLYGON_MODE_BACK, 1);
+		OUT_RING_CACHE(mode);
 	}
 }
 
@@ -474,8 +474,8 @@ void nv20ShadeModel(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_SHADE_MODEL, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_SHADE_MODEL, 1);
+	OUT_RING_CACHE(mode);
 }
 
 /** OpenGL 2.0 two-sided StencilFunc */
@@ -484,10 +484,10 @@ static void nv20StencilFuncSeparate(GLcontext *ctx, GLenum face, GLenum func,
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_FUNC_FUNC, 3);
-	OUT_RING(func);
-	OUT_RING(ref);
-	OUT_RING(mask);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_FUNC_FUNC, 3);
+	OUT_RING_CACHE(func);
+	OUT_RING_CACHE(ref);
+	OUT_RING_CACHE(mask);
 }
 
 /** OpenGL 2.0 two-sided StencilMask */
@@ -495,8 +495,8 @@ static void nv20StencilMaskSeparate(GLcontext *ctx, GLenum face, GLuint mask)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_MASK, 1);
-	OUT_RING(mask);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_MASK, 1);
+	OUT_RING_CACHE(mask);
 }
 
 /** OpenGL 2.0 two-sided StencilOp */
@@ -505,10 +505,10 @@ static void nv20StencilOpSeparate(GLcontext *ctx, GLenum face, GLenum fail,
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_OP_FAIL, 1);
-	OUT_RING(fail);
-	OUT_RING(zfail);
-	OUT_RING(zpass);
+	BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_STENCIL_OP_FAIL, 1);
+	OUT_RING_CACHE(fail);
+	OUT_RING_CACHE(zfail);
+	OUT_RING_CACHE(zpass);
 }
 
 /** Control the generation of texture coordinates */
@@ -528,9 +528,9 @@ static void nv20Viewport(GLcontext *ctx, GLint x, GLint y, GLsizei w, GLsizei h)
 {
     /* TODO: Where do the VIEWPORT_XFRM_* regs come in? */
     nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-    BEGIN_RING_SIZE(NvSub3D, NV20_TCL_PRIMITIVE_3D_VIEWPORT_HORIZ, 2);
-    OUT_RING((w << 16) | x);
-    OUT_RING((h << 16) | y);
+    BEGIN_RING_CACHE(NvSub3D, NV20_TCL_PRIMITIVE_3D_VIEWPORT_HORIZ, 2);
+    OUT_RING_CACHE((w << 16) | x);
+    OUT_RING_CACHE((h << 16) | y);
 }
 
 void nv20InitStateFuncs(struct dd_function_table *func)
diff --git a/src/mesa/drivers/dri/nouveau/nv30_state.c b/src/mesa/drivers/dri/nouveau/nv30_state.c
index a6cbcb8c8d4..7b50cbc4611 100644
--- a/src/mesa/drivers/dri/nouveau/nv30_state.c
+++ b/src/mesa/drivers/dri/nouveau/nv30_state.c
@@ -43,9 +43,9 @@ static void nv30AlphaFunc(GLcontext *ctx, GLenum func, GLfloat ref)
 	GLubyte ubRef;
 	CLAMPED_FLOAT_TO_UBYTE(ubRef, ref);
 
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC, 2);
-	OUT_RING(func);     /* NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC */
-	OUT_RING(ubRef);    /* NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_REF  */
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC, 2);
+	OUT_RING_CACHE(func);     /* NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_FUNC */
+	OUT_RING_CACHE(ubRef);    /* NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_REF  */
 }
 
 static void nv30BlendColor(GLcontext *ctx, const GLfloat color[4])
@@ -58,15 +58,15 @@ static void nv30BlendColor(GLcontext *ctx, const GLfloat color[4])
 	CLAMPED_FLOAT_TO_UBYTE(cf[2], color[2]);
 	CLAMPED_FLOAT_TO_UBYTE(cf[3], color[3]);
 
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_COLOR, 1);
-	OUT_RING(PACK_COLOR_8888(cf[3], cf[1], cf[2], cf[0]));
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_COLOR, 1);
+	OUT_RING_CACHE(PACK_COLOR_8888(cf[3], cf[1], cf[2], cf[0]));
 }
 
 static void nv30BlendEquationSeparate(GLcontext *ctx, GLenum modeRGB, GLenum modeA)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_EQUATION, 1);
-	OUT_RING((modeA<<16) | modeRGB);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_EQUATION, 1);
+	OUT_RING_CACHE((modeA<<16) | modeRGB);
 }
 
 
@@ -74,9 +74,9 @@ static void nv30BlendFuncSeparate(GLcontext *ctx, GLenum sfactorRGB, GLenum dfac
 		GLenum sfactorA, GLenum dfactorA)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_FUNC_SRC, 2);
-	OUT_RING((sfactorA<<16) | sfactorRGB);
-	OUT_RING((dfactorA<<16) | dfactorRGB);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_FUNC_SRC, 2);
+	OUT_RING_CACHE((sfactorA<<16) | sfactorRGB);
+	OUT_RING_CACHE((dfactorA<<16) | dfactorRGB);
 }
 
 static void nv30ClearColor(GLcontext *ctx, const GLfloat color[4])
@@ -84,16 +84,16 @@ static void nv30ClearColor(GLcontext *ctx, const GLfloat color[4])
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	GLubyte c[4];
 	UNCLAMPED_FLOAT_TO_RGBA_CHAN(c,color);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_ARGB, 1);
-	OUT_RING(PACK_COLOR_8888(c[3],c[0],c[1],c[2]));
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_ARGB, 1);
+	OUT_RING_CACHE(PACK_COLOR_8888(c[3],c[0],c[1],c[2]));
 }
 
 static void nv30ClearDepth(GLcontext *ctx, GLclampd d)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	nmesa->clear_value=((nmesa->clear_value&0x000000FF)|(((uint32_t)(d*0xFFFFFF))<<8));
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
-	OUT_RING(nmesa->clear_value);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
+	OUT_RING_CACHE(nmesa->clear_value);
 }
 
 /* we're don't support indexed buffers
@@ -104,26 +104,26 @@ static void nv30ClearStencil(GLcontext *ctx, GLint s)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 	nmesa->clear_value=((nmesa->clear_value&0xFFFFFF00)|(s&0x000000FF));
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
-	OUT_RING(nmesa->clear_value);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLEAR_VALUE_DEPTH, 1);
+	OUT_RING_CACHE(nmesa->clear_value);
 }
 
 static void nv30ClipPlane(GLcontext *ctx, GLenum plane, const GLfloat *equation)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLIP_PLANE_A(plane), 4);
-	OUT_RINGf(equation[0]);
-	OUT_RINGf(equation[1]);
-	OUT_RINGf(equation[2]);
-	OUT_RINGf(equation[3]);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLIP_PLANE_A(plane), 4);
+	OUT_RING_CACHEf(equation[0]);
+	OUT_RING_CACHEf(equation[1]);
+	OUT_RING_CACHEf(equation[2]);
+	OUT_RING_CACHEf(equation[3]);
 }
 
 static void nv30ColorMask(GLcontext *ctx, GLboolean rmask, GLboolean gmask,
 		GLboolean bmask, GLboolean amask )
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_COLOR_MASK, 1);
-	OUT_RING(((amask && 0x01) << 24) | ((rmask && 0x01) << 16) | ((gmask && 0x01)<< 8) | ((bmask && 0x01) << 0));
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_COLOR_MASK, 1);
+	OUT_RING_CACHE(((amask && 0x01) << 24) | ((rmask && 0x01) << 16) | ((gmask && 0x01)<< 8) | ((bmask && 0x01) << 0));
 }
 
 static void nv30ColorMaterial(GLcontext *ctx, GLenum face, GLenum mode)
@@ -134,37 +134,37 @@ static void nv30ColorMaterial(GLcontext *ctx, GLenum face, GLenum mode)
 static void nv30CullFace(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CULL_FACE, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CULL_FACE, 1);
+	OUT_RING_CACHE(mode);
 }
 
 static void nv30FrontFace(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_FRONT_FACE, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_FRONT_FACE, 1);
+	OUT_RING_CACHE(mode);
 }
 
 static void nv30DepthFunc(GLcontext *ctx, GLenum func)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_FUNC, 1);
-	OUT_RING(func);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_FUNC, 1);
+	OUT_RING_CACHE(func);
 }
 
 static void nv30DepthMask(GLcontext *ctx, GLboolean flag)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_WRITE_ENABLE, 1);
-	OUT_RING(flag);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_WRITE_ENABLE, 1);
+	OUT_RING_CACHE(flag);
 }
 
 static void nv30DepthRange(GLcontext *ctx, GLclampd nearval, GLclampd farval)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_RANGE_NEAR, 2);
-	OUT_RINGf(nearval);
-	OUT_RINGf(farval);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_RANGE_NEAR, 2);
+	OUT_RING_CACHEf(nearval);
+	OUT_RING_CACHEf(farval);
 }
 
 /** Specify the current buffer for writing */
@@ -178,13 +178,13 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 	switch(cap)
 	{
 		case GL_ALPHA_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ALPHA_FUNC_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_AUTO_NORMAL:
 		case GL_BLEND:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_FUNC_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_BLEND_FUNC_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_CLIP_PLANE0:
 		case GL_CLIP_PLANE1:
@@ -192,12 +192,12 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 		case GL_CLIP_PLANE3:
 		case GL_CLIP_PLANE4:
 		case GL_CLIP_PLANE5:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLIP_PLANE_ENABLE(cap-GL_CLIP_PLANE0), 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CLIP_PLANE_ENABLE(cap-GL_CLIP_PLANE0), 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_COLOR_LOGIC_OP:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_COLOR_MATERIAL:
 //		case GL_COLOR_SUM_EXT:
@@ -205,20 +205,20 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_CONVOLUTION_1D:
 //		case GL_CONVOLUTION_2D:
 		case GL_CULL_FACE:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CULL_FACE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_CULL_FACE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_DEPTH_TEST:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_TEST_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DEPTH_TEST_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_DITHER:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DITHER_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_DITHER_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_FOG:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_FOG_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_FOG_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_HISTOGRAM:
 //		case GL_INDEX_LOGIC_OP:
@@ -239,8 +239,8 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 			nmesa->enabled_lights=((nmesa->enabled_lights&mask)|(mask*state));
 			if (nmesa->lighting_enabled)
 			{
-				BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
-				OUT_RING(nmesa->enabled_lights);
+				BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
+				OUT_RING_CACHE(nmesa->enabled_lights);
 			}
 			break;
 			}
@@ -249,11 +249,11 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 				break;
 
 			nmesa->lighting_enabled=state;
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_ENABLED_LIGHTS, 1);
 			if (nmesa->lighting_enabled)
-				OUT_RING(nmesa->enabled_lights);
+				OUT_RING_CACHE(nmesa->enabled_lights);
 			else
-				OUT_RING(0x0);
+				OUT_RING_CACHE(0x0);
 			break;
 //		case GL_LINE_SMOOTH:
 //		case GL_LINE_STIPPLE:
@@ -277,29 +277,29 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_MAP2_VERTEX_4:
 //		case GL_MINMAX:
 		case GL_NORMALIZE:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_NORMALIZE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_NORMALIZE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_POINT_SMOOTH:
 		case GL_POLYGON_OFFSET_POINT:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_OFFSET_POINT_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_OFFSET_POINT_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_OFFSET_LINE:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_OFFSET_LINE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_OFFSET_LINE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_OFFSET_FILL:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_OFFSET_FILL_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_OFFSET_FILL_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_SMOOTH:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_SMOOTH_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_SMOOTH_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 		case GL_POLYGON_STIPPLE:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_STIPPLE_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_STIPPLE_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_POST_COLOR_MATRIX_COLOR_TABLE:
 //		case GL_POST_CONVOLUTION_COLOR_TABLE:
@@ -308,8 +308,8 @@ static void nv30Enable(GLcontext *ctx, GLenum cap, GLboolean state)
 //		case GL_SEPARABLE_2D:
 		case GL_STENCIL_TEST:
 			// TODO BACK and FRONT ?
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_ENABLE, 1);
-			OUT_RING(state);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_ENABLE, 1);
+			OUT_RING_CACHE(state);
 			break;
 //		case GL_TEXTURE_GEN_Q:
 //		case GL_TEXTURE_GEN_R:
@@ -327,8 +327,8 @@ static void nv30Fogfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
     switch(pname)
     {
         case GL_FOG_MODE:
-            BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_FOG_MODE, 1);
-            //OUT_RING (params);
+            BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_FOG_MODE, 1);
+            //OUT_RING_CACHE (params);
             break;
             /* TODO: unsure about the rest.*/
         default:
@@ -355,59 +355,59 @@ static void nv30Lightfv(GLcontext *ctx, GLenum light, GLenum pname, const GLfloa
 	switch(pname)
 	{
 		case GL_AMBIENT:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_AMBIENT_A(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_AMBIENT_A(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_DIFFUSE:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_DIFFUSE_A(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_DIFFUSE_A(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPECULAR:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_SPECULAR_A(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_FRONT_SIDE_PRODUCT_SPECULAR_A(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPOT_DIRECTION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_DIR_X(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_DIR_X(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_POSITION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_POSITION_X(light), 3);
-			OUT_RINGf(params[0]);
-			OUT_RINGf(params[1]);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_POSITION_X(light), 3);
+			OUT_RING_CACHEf(params[0]);
+			OUT_RING_CACHEf(params[1]);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_SPOT_EXPONENT:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_EXPONENT(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_EXPONENT(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_SPOT_CUTOFF:
 			/* you can't factor these */
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_A(light), 1);
-			OUT_RINGf(params[0]);
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_B(light), 1);
-			OUT_RINGf(params[1]);
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_C(light), 1);
-			OUT_RINGf(params[2]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_A(light), 1);
+			OUT_RING_CACHEf(params[0]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_B(light), 1);
+			OUT_RING_CACHEf(params[1]);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_SPOT_CUTOFF_C(light), 1);
+			OUT_RING_CACHEf(params[2]);
 			break;
 		case GL_CONSTANT_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_CONSTANT_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_CONSTANT_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_LINEAR_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_LINEAR_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_LINEAR_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		case GL_QUADRATIC_ATTENUATION:
-			BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_QUADRATIC_ATTENUATION(light), 1);
-			OUT_RINGf(*params);
+			BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LIGHT_QUADRATIC_ATTENUATION(light), 1);
+			OUT_RING_CACHEf(*params);
 			break;
 		default:
 			break;
@@ -421,8 +421,8 @@ void (*LightModelfv)(GLcontext *ctx, GLenum pname, const GLfloat *params);
 static void nv30LineStipple(GLcontext *ctx, GLint factor, GLushort pattern )
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_STIPPLE_PATTERN, 1);
-	OUT_RING((pattern << 16) | factor);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_STIPPLE_PATTERN, 1);
+	OUT_RING_CACHE((pattern << 16) | factor);
 }
 
 static void nv30LineWidth(GLcontext *ctx, GLfloat width)
@@ -432,15 +432,15 @@ static void nv30LineWidth(GLcontext *ctx, GLfloat width)
 
 	CLAMPED_FLOAT_TO_UBYTE(ubWidth, width);
 
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_WIDTH_SMOOTH, 1);
-	OUT_RING(ubWidth);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_LINE_WIDTH_SMOOTH, 1);
+	OUT_RING_CACHE(ubWidth);
 }
 
 static void nv30LogicOpcode(GLcontext *ctx, GLenum opcode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_OP, 1);
-	OUT_RING(opcode);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_COLOR_LOGIC_OP_OP, 1);
+	OUT_RING_CACHE(opcode);
 }
 
 static void nv30PointParameterfv(GLcontext *ctx, GLenum pname, const GLfloat *params)
@@ -454,8 +454,8 @@ static void nv30PointParameterfv(GLcontext *ctx, GLenum pname, const GLfloat *pa
 static void nv30PointSize(GLcontext *ctx, GLfloat size)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POINT_SIZE, 1);
-	OUT_RINGf(size);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POINT_SIZE, 1);
+	OUT_RING_CACHEf(size);
 }
 
 /** Select a polygon rasterization mode */
@@ -464,12 +464,12 @@ static void nv30PolygonMode(GLcontext *ctx, GLenum face, GLenum mode)
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
 	if (face == GL_FRONT || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_MODE_FRONT, 1);
-		OUT_RING(mode);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_MODE_FRONT, 1);
+		OUT_RING_CACHE(mode);
 	}
 	if (face == GL_BACK || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_MODE_BACK, 1);
-		OUT_RING(mode);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_POLYGON_MODE_BACK, 1);
+		OUT_RING_CACHE(mode);
 	}
 }
 
@@ -511,8 +511,8 @@ static void nv30ShadeModel(GLcontext *ctx, GLenum mode)
 {
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
-	BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_SHADE_MODEL, 1);
-	OUT_RING(mode);
+	BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_SHADE_MODEL, 1);
+	OUT_RING_CACHE(mode);
 }
 
 /** OpenGL 2.0 two-sided StencilFunc */
@@ -522,16 +522,16 @@ static void nv30StencilFuncSeparate(GLcontext *ctx, GLenum face, GLenum func,
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
 	if (face == GL_FRONT || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_FRONT_FUNC_FUNC, 3);
-		OUT_RING(func);
-		OUT_RING(ref);
-		OUT_RING(mask);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_FRONT_FUNC_FUNC, 3);
+		OUT_RING_CACHE(func);
+		OUT_RING_CACHE(ref);
+		OUT_RING_CACHE(mask);
 	}
 	if (face == GL_BACK || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_FUNC_FUNC, 3);
-		OUT_RING(func);
-		OUT_RING(ref);
-		OUT_RING(mask);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_FUNC_FUNC, 3);
+		OUT_RING_CACHE(func);
+		OUT_RING_CACHE(ref);
+		OUT_RING_CACHE(mask);
 	}
 }
 
@@ -541,12 +541,12 @@ static void nv30StencilMaskSeparate(GLcontext *ctx, GLenum face, GLuint mask)
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
 	if (face == GL_FRONT || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_FRONT_MASK, 1);
-		OUT_RING(mask);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_FRONT_MASK, 1);
+		OUT_RING_CACHE(mask);
 	}
 	if (face == GL_BACK || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_MASK, 1);
-		OUT_RING(mask);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_MASK, 1);
+		OUT_RING_CACHE(mask);
 	}
 }
 
@@ -557,16 +557,16 @@ static void nv30StencilOpSeparate(GLcontext *ctx, GLenum face, GLenum fail,
 	nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
 
 	if (face == GL_FRONT || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_FRONT_OP_FAIL, 3);
-		OUT_RING(fail);
-		OUT_RING(zfail);
-		OUT_RING(zpass);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_FRONT_OP_FAIL, 3);
+		OUT_RING_CACHE(fail);
+		OUT_RING_CACHE(zfail);
+		OUT_RING_CACHE(zpass);
 	}
 	if (face == GL_BACK || face == GL_FRONT_AND_BACK) {
-		BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_OP_FAIL, 3);
-		OUT_RING(fail);
-		OUT_RING(zfail);
-		OUT_RING(zpass);
+		BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_STENCIL_BACK_OP_FAIL, 3);
+		OUT_RING_CACHE(fail);
+		OUT_RING_CACHE(zfail);
+		OUT_RING_CACHE(zpass);
 	}
 }
 
@@ -594,9 +594,9 @@ static void nv30Viewport(GLcontext *ctx, GLint x, GLint y, GLsizei w, GLsizei h)
 {
         /* TODO: Where do the VIEWPORT_XFRM_* regs come in? */
         nouveauContextPtr nmesa = NOUVEAU_CONTEXT(ctx);
-        BEGIN_RING_SIZE(NvSub3D, NV30_TCL_PRIMITIVE_3D_VIEWPORT_DIMS_0, 2);
-        OUT_RING((w << 16) | x);
-        OUT_RING((h << 16) | y);
+        BEGIN_RING_CACHE(NvSub3D, NV30_TCL_PRIMITIVE_3D_VIEWPORT_DIMS_0, 2);
+        OUT_RING_CACHE((w << 16) | x);
+        OUT_RING_CACHE((h << 16) | y);
 }
 
 void nv30InitStateFuncs(struct dd_function_table *func)