--- a/src/mesa/drivers/dri/r128/r128_context.h	2009-01-22 18:38:33.000000000 +0100
+++ b/src/mesa/drivers/dri/r128/r128_context.h	2009-07-10 08:18:26.000000000 +0200
@@ -122,13 +122,15 @@
    drm_r128_context_regs_t setup;
 
    /* Vertex state */
-   GLuint vertex_size;
+   GLuint sw_vertex_size;
+   GLuint hw_vertex_size;
    GLuint vertex_format;
    struct tnl_attr_map vertex_attrs[VERT_ATTRIB_MAX];
    GLuint vertex_attr_count;
    char *verts;			/* points to tnl->clipspace.vertex_buf */
    GLuint num_verts;
    int coloroffset, specoffset;
+   int tex0_offset, tex1_offset;
    DECLARE_RENDERINPUTS(tnl_state_bitset);	/* tnl->render_inputs for this _tnl_install_attrs */
 
    GLuint NewGLState;

--- a/src/mesa/drivers/dri/r128/r128_state.c	2009-01-22 18:38:33.000000000 +0100
+++ b/src/mesa/drivers/dri/r128/r128_state.c	2009-07-10 08:18:26.000000000 +0200
@@ -1176,7 +1176,7 @@
       tex->tex_border_color	= t1->setup.tex_border_color;
    }
 
-   sarea->vertsize = rmesa->vertex_size;
+   sarea->vertsize = rmesa->hw_vertex_size;
    sarea->vc_format = rmesa->vertex_format;
 
    /* Turn off the texture cache flushing */

--- a/src/mesa/drivers/dri/r128/r128_tris.c	2009-07-10 17:18:13.000000000 +0200
+++ b/src/mesa/drivers/dri/r128/r128_tris.c	2009-07-10 08:18:26.000000000 +0200
@@ -71,25 +71,157 @@
  *                    Emit primitives as inline vertices               *
  ***********************************************************************/
 	
-#define HAVE_QUADS 0
-#define HAVE_LINES 1
-#define HAVE_POINTS 1
-#define HAVE_LE32_VERTS 1
-#define CTX_ARG r128ContextPtr rmesa
-#define GET_VERTEX_DWORDS() rmesa->vertex_size
-#define ALLOC_VERTS( n, size ) r128AllocDmaLow( rmesa, (n), (size) * 4 )
-#undef LOCAL_VARS
-#define LOCAL_VARS						\
-   r128ContextPtr rmesa = R128_CONTEXT(ctx);			\
+#ifndef DO_DEBUG_VERTS
+#define DO_DEBUG_VERTS 0
+#endif
+
+#ifndef PRINT_VERTEX
+#define PRINT_VERTEX(x)
+#endif
+
+static __inline void r128EmitHwVertex( r128ContextPtr rmesa,
+				       GLuint **vb,
+				       const r128VertexPtr v )
+{
+   GLfloat *hv = (GLfloat *) (*vb);
+
+   for (int i = 0; i < rmesa->hw_vertex_size; i++) {
+      (*vb)[i] = v->ui[i];
+   }
+
+   if (rmesa->tex0_offset) {
+      GLfloat rq0 = 1.0f / v->f[rmesa->tex0_offset + 2];
+
+      hv[3] *= v->f[rmesa->tex0_offset + 2]; // rhw0
+      hv[rmesa->tex0_offset] *= rq0; // u0
+      hv[rmesa->tex0_offset + 1] *= rq0; // v0
+   }
+
+   if (rmesa->tex1_offset) {
+      GLfloat rq1 = 1.0f / v->f[rmesa->tex1_offset + 2];
+
+      hv[rmesa->hw_vertex_size - 1] = v->f[3] * v->f[rmesa->tex1_offset + 2]; // rhw1
+      hv[rmesa->hw_vertex_size - 2] = v->f[rmesa->tex1_offset + 1] * rq1; // v1
+      hv[rmesa->hw_vertex_size - 3] = v->f[rmesa->tex1_offset] * rq1; // u1
+   }
+
+   for (int i = 0; i < rmesa->hw_vertex_size; i++) {
+      (*vb)[i] = CPU_TO_LE32(((*vb)[i]));
+   }
+
+   (*vb) += rmesa->hw_vertex_size;
+}
+
+static __inline void r128_quad( r128ContextPtr rmesa,
+				r128VertexPtr v0,
+				r128VertexPtr v1,
+				r128VertexPtr v2,
+				r128VertexPtr v3 )
+{
+   GLuint *vb = (GLuint *)r128AllocDmaLow( rmesa,
+					   6,
+					   rmesa->hw_vertex_size * 4 );
+
+   if (DO_DEBUG_VERTS) {
+      fprintf(stderr, "%s\n", __FUNCTION__);
+      PRINT_VERTEX(v0);
+      PRINT_VERTEX(v1);
+      PRINT_VERTEX(v2);
+      PRINT_VERTEX(v3);
+   }
+
+   r128EmitHwVertex( rmesa, &vb, v0 );
+   r128EmitHwVertex( rmesa, &vb, v1 );
+   r128EmitHwVertex( rmesa, &vb, v3 );
+   r128EmitHwVertex( rmesa, &vb, v1 );
+   r128EmitHwVertex( rmesa, &vb, v2 );
+   r128EmitHwVertex( rmesa, &vb, v3 );
+}
+
+static __inline void r128_triangle( r128ContextPtr rmesa,
+				    r128VertexPtr v0,
+				    r128VertexPtr v1,
+				    r128VertexPtr v2 )
+{
+   GLuint *vb = (GLuint *)r128AllocDmaLow( rmesa,
+					   3,
+					   rmesa->hw_vertex_size * 4 );
+
+   if (DO_DEBUG_VERTS) {
+      fprintf(stderr, "%s\n", __FUNCTION__);
+      PRINT_VERTEX(v0);
+      PRINT_VERTEX(v1);
+      PRINT_VERTEX(v2);
+   }
+
+   r128EmitHwVertex( rmesa, &vb, v0 );
+   r128EmitHwVertex( rmesa, &vb, v1 );
+   r128EmitHwVertex( rmesa, &vb, v2 );
+}
+
+static __inline void r128_line( r128ContextPtr rmesa,
+				r128VertexPtr v0,
+				r128VertexPtr v1 )
+{
+   GLuint *vb = (GLuint *)r128AllocDmaLow( rmesa,
+					   2,
+					   rmesa->hw_vertex_size * 4 );
+
+   if (DO_DEBUG_VERTS) {
+      fprintf(stderr, "%s\n", __FUNCTION__);
+      PRINT_VERTEX(v0);
+      PRINT_VERTEX(v1);
+   }
+
+   r128EmitHwVertex( rmesa, &vb, v0 );
+   r128EmitHwVertex( rmesa, &vb, v1 );
+}
+
+static __inline void r128_point( r128ContextPtr rmesa,
+				 r128VertexPtr v0 )
+{
+   GLuint *vb = (GLuint *)r128AllocDmaLow( rmesa,
+					   1,
+					   rmesa->hw_vertex_size * 4 );
+
+   if (DO_DEBUG_VERTS) {
+      fprintf(stderr, "%s\n", __FUNCTION__);
+      PRINT_VERTEX(v0);
+   }
+
+   r128EmitHwVertex( rmesa, &vb, v0 );
+}
+
+#define VERT(x) (const r128VertexPtr)(vertptr + ((x) * vertsize * 4))
+
+static void r128_fast_clipped_poly( GLcontext *ctx,
+				    const GLuint *elts,
+				    GLuint n )
+{
+   r128ContextPtr rmesa = R128_CONTEXT(ctx);
+   GLuint vertsize = rmesa->sw_vertex_size;
    const char *vertptr = rmesa->verts;
-#define VERT(x) (r128Vertex *)(vertptr + ((x) * vertsize * 4))
-#define VERTEX r128Vertex
-#undef TAG
-#define TAG(x) r128_##x
-#include "tnl_dd/t_dd_triemit.h"
-#undef TAG
-#undef LOCAL_VARS
 
+   GLuint *vb = (GLuint *)r128AllocDmaLow( rmesa,
+					   (n-2) * 3,
+					   rmesa->hw_vertex_size * 4 );
+
+   const r128VertexPtr start = VERT(elts[0]);
+
+   if (DO_DEBUG_VERTS) {
+      fprintf(stderr, "%s\n", __FUNCTION__);
+
+      for (GLuint i = 0 ; i < n ; i++) {
+	 PRINT_VERTEX((VERT(elts[i])));
+      }
+   }
+
+   for (GLuint i = 2; i < n; i++) {
+      r128EmitHwVertex( rmesa, &vb, VERT(elts[i-1]) );
+      r128EmitHwVertex( rmesa, &vb, VERT(elts[i]) );
+      r128EmitHwVertex( rmesa, &vb, start );
+   }
+}
 
 /***********************************************************************
  *          Macros for t_dd_tritmp.h to draw basic primitives          *
@@ -173,7 +305,7 @@
 #define VERT_Y(_v) _v->v.y
 #define VERT_Z(_v) _v->v.z
 #define AREA_IS_CCW( a ) (a > 0)
-#define GET_VERTEX(e) (rmesa->verts + (e * rmesa->vertex_size * sizeof(int)))
+#define GET_VERTEX(e) (rmesa->verts + (e * rmesa->sw_vertex_size * sizeof(int)))
 
 #define VERT_SET_RGBA( v, c )  					\
 do {								\
@@ -405,7 +537,7 @@
 #undef LOCAL_VARS
 #define LOCAL_VARS						\
     r128ContextPtr rmesa = R128_CONTEXT(ctx);		\
-    const GLuint vertsize = rmesa->vertex_size;		\
+    const GLuint vertsize = rmesa->sw_vertex_size;		\
     const char *vertptr = (char *)rmesa->verts;		\
     const GLuint * const elt = TNL_CONTEXT(ctx)->vb.Elts;	\
     (void) elt;
@@ -598,7 +730,6 @@
    struct vertex_buffer *VB = &tnl->vb;
    DECLARE_RENDERINPUTS(index_bitset);
    GLuint vc_frmt = 0;
-   GLboolean fallback_projtex = GL_FALSE;
    GLuint offset = 0;
 
    RENDERINPUTS_COPY( index_bitset, tnl->render_inputs_bitset );
@@ -607,16 +738,19 @@
    VB->AttribPtr[VERT_ATTRIB_POS] = VB->NdcPtr;
    rmesa->vertex_attr_count = 0;
    rmesa->specoffset = 0;
+   rmesa->tex0_offset = 0;
+   rmesa->tex1_offset = 0;
 
    /* EMIT_ATTR's must be in order as they tell t_vertex.c how to
     * build up a hardware vertex.
     */
    if (RENDERINPUTS_TEST_RANGE( index_bitset, _TNL_FIRST_TEX, _TNL_LAST_TEX ))
-      EMIT_ATTR( _TNL_ATTRIB_POS, EMIT_4F_VIEWPORT, R128_CCE_VC_FRMT_RHW, 4 );
+      EMIT_ATTR( _TNL_ATTRIB_POS, EMIT_4F_VIEWPORT, R128_CCE_VC_FRMT_RHW, 16 );
    else
-      EMIT_ATTR( _TNL_ATTRIB_POS, EMIT_3F_VIEWPORT, 0, 3 );
+      EMIT_ATTR( _TNL_ATTRIB_POS, EMIT_3F_VIEWPORT, 0, 12 );
+
+   rmesa->coloroffset = (offset >> 2);
 
-   rmesa->coloroffset = offset;
 #if MESA_LITTLE_ENDIAN 
    EMIT_ATTR( _TNL_ATTRIB_COLOR0, EMIT_4UB_4F_BGRA,
       R128_CCE_VC_FRMT_DIFFUSE_ARGB, 4 );
@@ -629,7 +763,7 @@
        RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_FOG )) {
 #if MESA_LITTLE_ENDIAN
       if (RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_COLOR1 )) {
-	 rmesa->specoffset = offset;
+	 rmesa->specoffset = (offset >> 2);
 	 EMIT_ATTR( _TNL_ATTRIB_COLOR1, EMIT_3UB_3F_BGR,
 	    R128_CCE_VC_FRMT_SPEC_FRGB, 3 );
       } else 
@@ -648,7 +782,7 @@
 	 EMIT_PAD( 1 );
 
       if (RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_COLOR1 )) {
-	 rmesa->specoffset = offset;
+	 rmesa->specoffset = (offset >> 2);
 	 EMIT_ATTR( _TNL_ATTRIB_COLOR1, EMIT_3UB_3F_RGB,
 	    R128_CCE_VC_FRMT_SPEC_FRGB, 3 );
       } else 
@@ -656,25 +790,27 @@
 #endif
    }
 
-   if (RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_TEX(rmesa->tmu_source[0]) )) {
-      if ( VB->TexCoordPtr[rmesa->tmu_source[0]]->size > 2 )
-	 fallback_projtex = GL_TRUE;
+   if (RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_TEX0 )) {
+      if ( VB->TexCoordPtr[0]->size > 2 )
+      {
+	 rmesa->tex0_offset = (offset >> 2);
+	 EMIT_ATTR( _TNL_ATTRIB_TEX0, EMIT_3F_XYW, R128_CCE_VC_FRMT_S_T, 12 );
+      }
+      else
       EMIT_ATTR( _TNL_ATTRIB_TEX0, EMIT_2F, R128_CCE_VC_FRMT_S_T, 8 );
    }
-   if (RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_TEX(rmesa->tmu_source[1]) )) {
-      if ( VB->TexCoordPtr[rmesa->tmu_source[1]]->size > 2 )
-	 fallback_projtex = GL_TRUE;
+
+   if (RENDERINPUTS_TEST( index_bitset, _TNL_ATTRIB_TEX1 )) {
+      if ( VB->TexCoordPtr[1]->size > 2 )
+      {
+	 rmesa->tex1_offset = (offset >> 2);
+	 EMIT_ATTR( _TNL_ATTRIB_TEX1, EMIT_3F_XYW,
+		    (R128_CCE_VC_FRMT_S2_T2 | R128_CCE_VC_FRMT_RHW2), 12 );
+      }
+      else
       EMIT_ATTR( _TNL_ATTRIB_TEX1, EMIT_2F, R128_CCE_VC_FRMT_S2_T2, 8 );
    }
 
-   /* projective textures are not supported by the hardware */
-   // FALLBACK( rmesa, R128_FALLBACK_PROJTEX, fallback_projtex );
-   //
-   // if (fallback_projtex) {
-   //     tnl->Driver.Render.Start(ctx);
-   //     return;
-   // }
-
    /* Only need to change the vertex emit code if there has been a
     * statechange to a TNL index.
     */
@@ -682,13 +818,13 @@
       FLUSH_BATCH( rmesa );
       rmesa->dirty |= R128_UPLOAD_CONTEXT;
 
-      rmesa->vertex_size = 
+      rmesa->sw_vertex_size =
 	 _tnl_install_attrs( ctx, 
 			     rmesa->vertex_attrs, 
 			     rmesa->vertex_attr_count,
 			     rmesa->hw_viewport, 0 );
-      rmesa->vertex_size >>= 2;
-
+      rmesa->sw_vertex_size >>= 2;
+      rmesa->hw_vertex_size = (rmesa->tex0_offset) ? (rmesa->sw_vertex_size - 1) : rmesa->sw_vertex_size;
       rmesa->vertex_format = vc_frmt;
    }
 }