Added support for the 3rd texture coordinate for cubemaps and 3D

textures.  progs/demos/cubemap and progs/demos/stex3d seem to work
correctly at all tcl_mode settings.  x86 / SSE codegen is currently
disabled.

8 files changed, 294 insertions, 184 deletions

diff --git a/src/mesa/drivers/dri/r200/r200_context.h b/src/mesa/drivers/dri/r200/r200_context.h
index 69ac92b1f10..63d688718c8 100644
--- a/src/mesa/drivers/dri/r200/r200_context.h
+++ b/src/mesa/drivers/dri/r200/r200_context.h
@@ -49,7 +49,7 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "colormac.h"
 #include "r200_reg.h"
 
-#define ENABLE_HW_3D_TEXTURE 0  /* XXX this is temporary! */
+#define ENABLE_HW_3D_TEXTURE 1  /* XXX this is temporary! */
 
 struct r200_context;
 typedef struct r200_context r200ContextRec;
@@ -365,6 +365,12 @@ struct r200_state_atom {
 #define VTX_COLOR(v,n)   (((v)>>(R200_VTX_COLOR_0_SHIFT+(n)*2))&\
                          R200_VTX_COLOR_MASK)
 
+/**
+ * Given the \c R200_SE_VTX_FMT_1 for the current vertex state, determine
+ * how many components are in texture coordinate \c n.
+ */
+#define VTX_TEXn_COUNT(v,n)   (((v) >> (3 * n)) & 0x07)
+
 #define MAT_CMD_0              0
 #define MAT_ELT_0              1
 #define MAT_STATE_SIZE         17
@@ -703,10 +709,14 @@ struct dfn_lists {
    struct dynfn SecondaryColor3fvEXT;
    struct dynfn Normal3f;
    struct dynfn Normal3fv;
+   struct dynfn TexCoord3f;
+   struct dynfn TexCoord3fv;
    struct dynfn TexCoord2f;
    struct dynfn TexCoord2fv;
    struct dynfn TexCoord1f;
    struct dynfn TexCoord1fv;
+   struct dynfn MultiTexCoord3fARB;
+   struct dynfn MultiTexCoord3fvARB;
    struct dynfn MultiTexCoord2fARB;
    struct dynfn MultiTexCoord2fvARB;
    struct dynfn MultiTexCoord1fARB;
@@ -732,10 +742,14 @@ struct dfn_generators {
    struct dynfn *(*SecondaryColor3fvEXT)( GLcontext *, const int * );
    struct dynfn *(*Normal3f)( GLcontext *, const int * );
    struct dynfn *(*Normal3fv)( GLcontext *, const int * );
+   struct dynfn *(*TexCoord3f)( GLcontext *, const int * );
+   struct dynfn *(*TexCoord3fv)( GLcontext *, const int * );
    struct dynfn *(*TexCoord2f)( GLcontext *, const int * );
    struct dynfn *(*TexCoord2fv)( GLcontext *, const int * );
    struct dynfn *(*TexCoord1f)( GLcontext *, const int * );
    struct dynfn *(*TexCoord1fv)( GLcontext *, const int * );
+   struct dynfn *(*MultiTexCoord3fARB)( GLcontext *, const int * );
+   struct dynfn *(*MultiTexCoord3fvARB)( GLcontext *, const int * );
    struct dynfn *(*MultiTexCoord2fARB)( GLcontext *, const int * );
    struct dynfn *(*MultiTexCoord2fvARB)( GLcontext *, const int * );
    struct dynfn *(*MultiTexCoord1fARB)( GLcontext *, const int * );
@@ -756,17 +770,16 @@ struct r200_vbinfo {
    void (*notify)( void );
    GLint vertex_size;
 
-   /* A maximum total of 15 elements per vertex:  3 floats for position, 3
+   /* A maximum total of 17 elements per vertex:  3 floats for position, 3
     * floats for normal, 4 floats for color, 4 bytes for secondary color,
-    * 2 floats for each texture unit (4 floats total).
+    * 3 floats for each texture unit (6 floats total).
     * 
-    * As soon as the 3rd TMU is supported or cube maps (or 3D textures) are
-    * supported, this value will grow.
+    * As soon as the 3rd through 6th TMUs are supported, this value will grow.
     * 
     * The position data is never actually stored here, so 3 elements could be
     * trimmed out of the buffer.
     */
-   union { float f; int i; r200_color_t color; } vertex[15];
+   union { float f; int i; r200_color_t color; } vertex[17];
 
    GLfloat *normalptr;
    GLfloat *floatcolorptr;
diff --git a/src/mesa/drivers/dri/r200/r200_maos_arrays.c b/src/mesa/drivers/dri/r200/r200_maos_arrays.c
index 9d1724db7e7..301b265396a 100644
--- a/src/mesa/drivers/dri/r200/r200_maos_arrays.c
+++ b/src/mesa/drivers/dri/r200/r200_maos_arrays.c
@@ -343,6 +343,8 @@ void r200EmitArrays( GLcontext *ctx, GLuint inputs )
    GLuint nr = 0;
    GLuint vfmt0 = 0, vfmt1 = 0;
    GLuint count = VB->Count;
+   GLuint i;
+   GLuint re_cntl;
    
    if (1) {
       if (!rmesa->tcl.obj.buf) 
@@ -422,30 +424,34 @@ void r200EmitArrays( GLcontext *ctx, GLuint inputs )
 /*    vtx = (rmesa->hw.tcl.cmd[TCL_OUTPUT_VTXFMT] & */
 /* 	  ~(R200_TCL_VTX_Q0|R200_TCL_VTX_Q1)); */
       
-   if (inputs & VERT_BIT_TEX0) {
-      if (!rmesa->tcl.tex[0].buf)
-	 emit_vector( ctx, 
-		      &(rmesa->tcl.tex[0]), 
-		      (char *)VB->TexCoordPtr[0]->data,
-		      VB->TexCoordPtr[0]->size,
-		      VB->TexCoordPtr[0]->stride,
-		      count );
-
-      vfmt1 |= VB->TexCoordPtr[0]->size << R200_VTX_TEX0_COMP_CNT_SHIFT;
-      component[nr++] = &rmesa->tcl.tex[0];
+   re_cntl = rmesa->hw.set.cmd[SET_RE_CNTL] & ~(R200_VTX_STQ0_D3D |
+						R200_VTX_STQ1_D3D |
+						R200_VTX_STQ2_D3D |
+						R200_VTX_STQ3_D3D |
+						R200_VTX_STQ4_D3D |
+						R200_VTX_STQ5_D3D );
+   for ( i = 0 ; i < ctx->Const.MaxTextureUnits ; i++ ) {
+      if (inputs & (VERT_BIT_TEX0 << i)) {
+	 if (!rmesa->tcl.tex[i].buf)
+	     emit_vector( ctx, 
+			  &(rmesa->tcl.tex[i]),
+			  (char *)VB->TexCoordPtr[i]->data,
+			  VB->TexCoordPtr[i]->size,
+			  VB->TexCoordPtr[i]->stride,
+			  count );
+
+	 if ( ctx->Texture.Unit[i]._ReallyEnabled == TEXTURE_CUBE_BIT ) {
+	    re_cntl |= R200_VTX_STQ0_D3D << (2 * i);
+	 }
+
+	 vfmt1 |= VB->TexCoordPtr[i]->size << (i * 3);
+	 component[nr++] = &rmesa->tcl.tex[i];
+      }
    }
 
-   if (inputs & VERT_BIT_TEX1) {
-      if (!rmesa->tcl.tex[1].buf)
-	 emit_vector( ctx, 
-		      &(rmesa->tcl.tex[1]), 
-		      (char *)VB->TexCoordPtr[1]->data,
-		      VB->TexCoordPtr[1]->size,
-		      VB->TexCoordPtr[1]->stride,
-		      count );
-	 
-      vfmt1 |= VB->TexCoordPtr[1]->size << R200_VTX_TEX1_COMP_CNT_SHIFT;
-      component[nr++] = &rmesa->tcl.tex[1];
+   if ( re_cntl != rmesa->hw.set.cmd[SET_RE_CNTL] ) {
+      R200_STATECHANGE( rmesa, set );
+      rmesa->hw.set.cmd[SET_RE_CNTL] = re_cntl;
    }
 
    if (vfmt0 != rmesa->hw.vtx.cmd[VTX_VTXFMT_0] ||
diff --git a/src/mesa/drivers/dri/r200/r200_swtcl.c b/src/mesa/drivers/dri/r200/r200_swtcl.c
index 2a0cd5b9d15..80f6f5f91cb 100644
--- a/src/mesa/drivers/dri/r200/r200_swtcl.c
+++ b/src/mesa/drivers/dri/r200/r200_swtcl.c
@@ -147,22 +147,15 @@ static void r200SetVertexFormat( GLcontext *ctx )
 	    GLuint sz = VB->TexCoordPtr[i]->size;
 	    GLuint emit;
 
-	    /* r200 doesn't like 1D or 4D texcoords (is that true?):
+	    /* r200 doesn't like 4D texcoords (is that true?):
 	     */
-	    switch (sz) {
-	    case 1: 
-	    case 2: 
-	    case 3: 		/* no attempt at cube texturing so far */
-	       emit = EMIT_2F; 
-	       sz = 2; 
-	       break;
-	    case 4: 
+	    if (sz != 4) {
+	       emit = EMIT_1F + (sz - 1);
+	    }
+	    else {
+	       sz = 3;
 	       emit = EMIT_3F_XYW; 
-	       sz = 3;     
-	       break;
-	    default: 
-	       continue;
-	    };
+	    }
 
 	    fmt_1 |= sz << (3 * i);
 	    EMIT_ATTR( _TNL_ATTRIB_TEX0+i, EMIT_SZ(sz), 0 );
diff --git a/src/mesa/drivers/dri/r200/r200_texstate.c b/src/mesa/drivers/dri/r200/r200_texstate.c
index dbd15ac7eac..164c17e0644 100644
--- a/src/mesa/drivers/dri/r200/r200_texstate.c
+++ b/src/mesa/drivers/dri/r200/r200_texstate.c
@@ -747,6 +747,7 @@ static GLboolean r200UpdateTextureEnv( GLcontext *ctx, int unit )
 
 #define TEXOBJ_TXFORMAT_X_MASK (R200_DEPTH_LOG2_MASK |		\
                                 R200_TEXCOORD_MASK |		\
+                                R200_CLAMP_Q_MASK | 		\
                                 R200_VOLUME_FILTER_MASK)
 
 
diff --git a/src/mesa/drivers/dri/r200/r200_vtxfmt.c b/src/mesa/drivers/dri/r200/r200_vtxfmt.c
index f90a3a5cc11..c63518fefb6 100644
--- a/src/mesa/drivers/dri/r200/r200_vtxfmt.c
+++ b/src/mesa/drivers/dri/r200/r200_vtxfmt.c
@@ -88,10 +88,14 @@ static void count_funcs( r200ContextPtr rmesa )
    count_func( "SecondaryColor3ubv", &rmesa->vb.dfn_cache.SecondaryColor3ubvEXT );
    count_func( "Normal3f", &rmesa->vb.dfn_cache.Normal3f );
    count_func( "Normal3fv", &rmesa->vb.dfn_cache.Normal3fv );
+   count_func( "TexCoord3f", &rmesa->vb.dfn_cache.TexCoord3f );
+   count_func( "TexCoord3fv", &rmesa->vb.dfn_cache.TexCoord3fv );
    count_func( "TexCoord2f", &rmesa->vb.dfn_cache.TexCoord2f );
    count_func( "TexCoord2fv", &rmesa->vb.dfn_cache.TexCoord2fv );
    count_func( "TexCoord1f", &rmesa->vb.dfn_cache.TexCoord1f );
    count_func( "TexCoord1fv", &rmesa->vb.dfn_cache.TexCoord1fv );
+   count_func( "MultiTexCoord3fARB", &rmesa->vb.dfn_cache.MultiTexCoord3fARB );
+   count_func( "MultiTexCoord3fvARB", &rmesa->vb.dfn_cache.MultiTexCoord3fvARB );
    count_func( "MultiTexCoord2fARB", &rmesa->vb.dfn_cache.MultiTexCoord2fARB );
    count_func( "MultiTexCoord2fvARB", &rmesa->vb.dfn_cache.MultiTexCoord2fvARB );
    count_func( "MultiTexCoord1fARB", &rmesa->vb.dfn_cache.MultiTexCoord1fARB );
@@ -102,6 +106,7 @@ static void count_funcs( r200ContextPtr rmesa )
 void r200_copy_to_current( GLcontext *ctx ) 
 {
    r200ContextPtr rmesa = R200_CONTEXT(ctx);
+   unsigned i;
 
    if (R200_DEBUG & DEBUG_VFMT)
       fprintf(stderr, "%s\n", __FUNCTION__);
@@ -145,18 +150,29 @@ void r200_copy_to_current( GLcontext *ctx )
       ctx->Current.Attrib[VERT_ATTRIB_COLOR1][2] = UBYTE_TO_FLOAT( rmesa->vb.specptr->blue );
    } 
 
-   if (rmesa->vb.vtxfmt_1 & (7 << R200_VTX_TEX0_COMP_CNT_SHIFT)) {
-      ctx->Current.Attrib[VERT_ATTRIB_TEX0][0] = rmesa->vb.texcoordptr[0][0];
-      ctx->Current.Attrib[VERT_ATTRIB_TEX0][1] = rmesa->vb.texcoordptr[0][1];
-      ctx->Current.Attrib[VERT_ATTRIB_TEX0][2] = 0.0F;
-      ctx->Current.Attrib[VERT_ATTRIB_TEX0][3] = 1.0F;
-   }
+   for ( i = 0 ; i < ctx->Const.MaxTextureUnits ; i++ ) {
+      const unsigned count = VTX_TEXn_COUNT( rmesa->vb.vtxfmt_1, i );
+      GLfloat * const src = rmesa->vb.texcoordptr[i];
+
+      if ( count != 0 ) {
+	 switch( count ) {
+	 case 3:
+	    ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][1] = src[1];
+	    ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][2] = src[2];
+	    break;
+	 case 2:
+	    ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][1] = src[1];
+	    ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][2] = 0.0F;
+	    break;
+	 case 1:
+	    ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][1] = 0.0F;
+	    ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][2] = 0.0F;
+	    break;
+	 }
 
-   if (rmesa->vb.vtxfmt_1 & (7 << R200_VTX_TEX1_COMP_CNT_SHIFT)) {
-      ctx->Current.Attrib[VERT_ATTRIB_TEX1][0] = rmesa->vb.texcoordptr[1][0];
-      ctx->Current.Attrib[VERT_ATTRIB_TEX1][1] = rmesa->vb.texcoordptr[1][1];
-      ctx->Current.Attrib[VERT_ATTRIB_TEX1][2] = 0.0F;
-      ctx->Current.Attrib[VERT_ATTRIB_TEX1][3] = 1.0F;
+	 ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][0] = src[0];
+	 ctx->Current.Attrib[VERT_ATTRIB_TEX0+i][3] = 1.0F;
+      }
    }
 
    ctx->Driver.NeedFlush &= ~FLUSH_UPDATE_CURRENT;
@@ -378,6 +394,42 @@ static void VFMT_FALLBACK_OUTSIDE_BEGIN_END( const char *caller )
 }
 
 
+static void dispatch_texcoord( GLuint count, GLfloat * f )
+{
+   switch( count ) {
+   case 3:
+      _glapi_Dispatch->TexCoord3fv( f );
+      break;
+   case 2:
+      _glapi_Dispatch->TexCoord2fv( f );
+      break;
+   case 1:
+      _glapi_Dispatch->TexCoord1fv( f );
+      break;
+   default:
+      assert( count == 0 );
+      break;
+   }
+}
+
+static void dispatch_multitexcoord( GLuint count, GLuint unit, GLfloat * f )
+{
+   switch( count ) {
+   case 3:
+      _glapi_Dispatch->MultiTexCoord3fvARB( GL_TEXTURE0+unit, f );
+      break;
+   case 2:
+      _glapi_Dispatch->MultiTexCoord2fvARB( GL_TEXTURE0+unit, f );
+      break;
+   case 1:
+      _glapi_Dispatch->MultiTexCoord1fvARB( GL_TEXTURE0+unit, f );
+      break;
+   default:
+      assert( count == 0 );
+      break;
+   }
+}
+
 static void VFMT_FALLBACK( const char *caller )
 {
    GET_CURRENT_CONTEXT(ctx);
@@ -388,6 +440,8 @@ static void VFMT_FALLBACK( const char *caller )
    GLuint ind1 = rmesa->vb.vtxfmt_1;
    GLuint nrverts;
    GLfloat alpha = 1.0;
+   GLuint count;
+   GLuint unit;
 
    if (R200_DEBUG & (DEBUG_FALLBACKS|DEBUG_VFMT))
       fprintf(stderr, "%s from %s\n", __FUNCTION__, caller);
@@ -425,6 +479,7 @@ static void VFMT_FALLBACK( const char *caller )
     */
    for (i = 0 ; i < nrverts; i++) {
       GLuint offset = 3;
+
       if (ind0 & R200_VTX_N0) {
 	 _glapi_Dispatch->Normal3fv( &tmp[i][offset] ); 
 	 offset += 3;
@@ -448,14 +503,14 @@ static void VFMT_FALLBACK( const char *caller )
 	 offset++;
       }
 
-      if (ind1 & (7 << R200_VTX_TEX0_COMP_CNT_SHIFT)) {
-	 _glapi_Dispatch->TexCoord2fv( &tmp[i][offset] ); 
-	 offset += 2;
-      }
+      count = VTX_TEXn_COUNT( ind1, 0 );
+      dispatch_texcoord( count, &tmp[i][offset] );
+      offset += count;
 
-      if (ind1 & (7 << R200_VTX_TEX1_COMP_CNT_SHIFT)) {
-	 _glapi_Dispatch->MultiTexCoord2fvARB( GL_TEXTURE1, &tmp[i][offset] );
-	 offset += 2;
+      for ( unit = 1 ; unit < ctx->Const.MaxTextureUnits ; unit++ ) {
+	 count = VTX_TEXn_COUNT( ind1, unit );
+	 dispatch_multitexcoord( count, unit, &tmp[i][offset] );
+	 offset += count;
       }
 
       _glapi_Dispatch->Vertex3fv( &tmp[i][0] );
@@ -492,12 +547,12 @@ static void VFMT_FALLBACK( const char *caller )
 					     rmesa->vb.specptr->green,
 					     rmesa->vb.specptr->blue ); 
 
-   if (ind1 & (7 << R200_VTX_TEX0_COMP_CNT_SHIFT)) {
-      _glapi_Dispatch->TexCoord2fv( rmesa->vb.texcoordptr[0] );
-   }
+   count = VTX_TEXn_COUNT( ind1, 0 );
+   dispatch_texcoord( count, rmesa->vb.texcoordptr[0] );
 
-   if (ind1 & (7 << R200_VTX_TEX1_COMP_CNT_SHIFT)) {
-      _glapi_Dispatch->MultiTexCoord2fvARB( GL_TEXTURE1, rmesa->vb.texcoordptr[1] );
+   for ( unit = 1 ; unit < ctx->Const.MaxTextureUnits ; unit++ ) {
+      count = VTX_TEXn_COUNT( ind1, unit );
+      dispatch_multitexcoord( count, unit, rmesa->vb.texcoordptr[unit] );
    }
 }
 
@@ -582,12 +637,34 @@ static void wrap_buffer( void )
 }
 
 
-
+/**
+ * Determines the hardware vertex format based on the current state vector.
+ * 
+ * \returns
+ * If the hardware TCL unit is capable of handling the current state vector,
+ * \c GL_TRUE is returned.  Otherwise, \c GL_FALSE is returned.
+ *
+ * \todo
+ * Make this color format selection data driven.  If we receive only ubytes,
+ * send color as ubytes.  Also check if converting (with free checking for
+ * overflow) is cheaper than sending floats directly.
+ *
+ * \todo
+ * When intializing texture coordinates, it might be faster to just copy the
+ * entire \c VERT_ATTRIB_TEX0 vector into the vertex buffer.  It may mean that
+ * some of the data (i.e., the last texture coordinate components) get copied
+ * over, but that still may be faster than the conditional branching.  If
+ * nothing else, the code will be smaller and easier to follow.
+ */
 static GLboolean check_vtx_fmt( GLcontext *ctx )
 {
    r200ContextPtr rmesa = R200_CONTEXT(ctx);
    GLuint ind0 = R200_VTX_Z0;
    GLuint ind1 = 0;
+   GLuint i;
+   GLuint count[R200_MAX_TEXTURE_UNITS];
+   GLuint re_cntl;
+
 
    if (rmesa->TclFallback || rmesa->vb.fell_back || ctx->CompileFlag)
       return GL_FALSE;
@@ -600,11 +677,6 @@ static GLboolean check_vtx_fmt( GLcontext *ctx )
    if (ctx->Light.Enabled) {
       ind0 |= R200_VTX_N0;
 
-      /* TODO: make this data driven: If we receive only ubytes, send
-       * color as ubytes.  Also check if converting (with free
-       * checking for overflow) is cheaper than sending floats
-       * directly.
-       */
       if (ctx->Light.ColorMaterialEnabled) 
 	 ind0 |= R200_VTX_FP_RGBA << R200_VTX_COLOR_0_SHIFT;
       else
@@ -620,36 +692,46 @@ static GLboolean check_vtx_fmt( GLcontext *ctx )
       }
    }
 
-   if (ctx->Texture.Unit[0]._ReallyEnabled) {
-      if (ctx->Texture.Unit[0].TexGenEnabled) {
-	 if (rmesa->TexGenNeedNormals[0]) {
-	    ind0 |= R200_VTX_N0;
+   re_cntl = rmesa->hw.set.cmd[SET_RE_CNTL] & ~(R200_VTX_STQ0_D3D |
+						R200_VTX_STQ1_D3D |
+						R200_VTX_STQ2_D3D |
+						R200_VTX_STQ3_D3D |
+						R200_VTX_STQ4_D3D |
+						R200_VTX_STQ5_D3D );
+   for ( i = 0 ; i < ctx->Const.MaxTextureUnits ; i++ ) {
+      count[i] = 0;
+
+      if (ctx->Texture.Unit[i]._ReallyEnabled) {
+	 if (ctx->Texture.Unit[i].TexGenEnabled) {
+	    if (rmesa->TexGenNeedNormals[i]) {
+	       ind0 |= R200_VTX_N0;
+	    }
 	 }
-      } else {
-	 if (ctx->Current.Attrib[VERT_ATTRIB_TEX0][2] != 0.0F ||
-	     ctx->Current.Attrib[VERT_ATTRIB_TEX0][3] != 1.0) {
-	    if (R200_DEBUG & (DEBUG_VFMT|DEBUG_FALLBACKS))
-	       fprintf(stderr, "%s: rq0\n", __FUNCTION__);
-	    return GL_FALSE;
+	 else {
+	    switch( ctx->Texture.Unit[i]._ReallyEnabled ) {
+	    case TEXTURE_CUBE_BIT:
+	       re_cntl |= R200_VTX_STQ0_D3D << (2 * i);
+	       /* FALLTHROUGH */
+	    case TEXTURE_3D_BIT:
+	       count[i] = 3;
+	       break;
+	    case TEXTURE_2D_BIT:
+	    case TEXTURE_RECT_BIT:
+	       count[i] = 2;
+	       break;
+	    case TEXTURE_1D_BIT:
+	       count[i] = 1;
+	       break;
+	    }
+
+	    ind1 |= count[i] << (3 * i);
 	 }
-	 ind1 |= 2 << R200_VTX_TEX0_COMP_CNT_SHIFT;
       }
    }
 
-   if (ctx->Texture.Unit[1]._ReallyEnabled) {
-      if (ctx->Texture.Unit[1].TexGenEnabled) {
-	 if (rmesa->TexGenNeedNormals[1]) {
-	    ind0 |= R200_VTX_N0;
-	 }
-      } else {
-	 if (ctx->Current.Attrib[VERT_ATTRIB_TEX1][2] != 0.0F ||
-	     ctx->Current.Attrib[VERT_ATTRIB_TEX1][3] != 1.0) {
-	    if (R200_DEBUG & (DEBUG_VFMT|DEBUG_FALLBACKS))
-	       fprintf(stderr, "%s: rq1\n", __FUNCTION__);
-	    return GL_FALSE;
-	 }
-	 ind1 |= 2 << R200_VTX_TEX1_COMP_CNT_SHIFT;
-      }
+   if ( re_cntl != rmesa->hw.set.cmd[SET_RE_CNTL] ) {
+      R200_STATECHANGE( rmesa, set );
+      rmesa->hw.set.cmd[SET_RE_CNTL] = re_cntl;
    }
 
    if (R200_DEBUG & (DEBUG_VFMT|DEBUG_STATE))
@@ -713,19 +795,20 @@ static GLboolean check_vtx_fmt( GLcontext *ctx )
    }
 
 
-   if (ind1 & (7 << R200_VTX_TEX0_COMP_CNT_SHIFT)) {
-      rmesa->vb.texcoordptr[0] = &rmesa->vb.vertex[rmesa->vb.vertex_size].f;
-      rmesa->vb.vertex_size += 2;
-      rmesa->vb.texcoordptr[0][0] = ctx->Current.Attrib[VERT_ATTRIB_TEX0][0];
-      rmesa->vb.texcoordptr[0][1] = ctx->Current.Attrib[VERT_ATTRIB_TEX0][1];   
-   } 
+   for ( i = 0 ; i < ctx->Const.MaxTextureUnits ; i++ ) {
+      if ( count[i] != 0 ) {
+	 float * const attr = ctx->Current.Attrib[VERT_ATTRIB_TEX0+i];
+	 unsigned  j;
 
-   if (ind1 & (7 << R200_VTX_TEX1_COMP_CNT_SHIFT)) {
-      rmesa->vb.texcoordptr[1] = &rmesa->vb.vertex[rmesa->vb.vertex_size].f;
-      rmesa->vb.vertex_size += 2;
-      rmesa->vb.texcoordptr[1][0] = ctx->Current.Attrib[VERT_ATTRIB_TEX1][0];
-      rmesa->vb.texcoordptr[1][1] = ctx->Current.Attrib[VERT_ATTRIB_TEX1][1];
-   } 
+	 rmesa->vb.texcoordptr[i] = &rmesa->vb.vertex[rmesa->vb.vertex_size].f;
+
+	 for ( j = 0 ; j < count[i] ; j++ ) {
+	    rmesa->vb.texcoordptr[i][j] = attr[j];
+	 }
+
+	 rmesa->vb.vertex_size += count[i];
+      }
+   }
 
    if (rmesa->vb.installed_vertex_format != rmesa->vb.vtxfmt_0) {
       if (R200_DEBUG & DEBUG_VFMT)
@@ -736,7 +819,10 @@ static GLboolean check_vtx_fmt( GLcontext *ctx )
 
    if (R200_DEBUG & DEBUG_VFMT)
       fprintf(stderr, "%s -- success\n", __FUNCTION__);
-   
+
+   fprintf( stderr, "[%s:%u] vertex size = %u\n", __func__, __LINE__,
+	    rmesa->vb.vertex_size );
+
    return GL_TRUE;
 }
 
@@ -1026,12 +1112,8 @@ void r200VtxfmtInit( GLcontext *ctx, GLboolean useCodegen )
    vfmt->EvalMesh2 = r200_fallback_EvalMesh2;
    vfmt->EvalPoint1 = r200_fallback_EvalPoint1;
    vfmt->EvalPoint2 = r200_fallback_EvalPoint2;
-   vfmt->TexCoord3f = r200_fallback_TexCoord3f;
-   vfmt->TexCoord3fv = r200_fallback_TexCoord3fv;
    vfmt->TexCoord4f = r200_fallback_TexCoord4f;
    vfmt->TexCoord4fv = r200_fallback_TexCoord4fv;
-   vfmt->MultiTexCoord3fARB = r200_fallback_MultiTexCoord3fARB;
-   vfmt->MultiTexCoord3fvARB = r200_fallback_MultiTexCoord3fvARB;
    vfmt->MultiTexCoord4fARB = r200_fallback_MultiTexCoord4fARB;
    vfmt->MultiTexCoord4fvARB = r200_fallback_MultiTexCoord4fvARB;
    vfmt->Vertex4f = r200_fallback_Vertex4f;
@@ -1071,10 +1153,14 @@ void r200VtxfmtInit( GLcontext *ctx, GLboolean useCodegen )
    make_empty_list( &rmesa->vb.dfn_cache.SecondaryColor3ubvEXT );
    make_empty_list( &rmesa->vb.dfn_cache.Normal3f );
    make_empty_list( &rmesa->vb.dfn_cache.Normal3fv );
+   make_empty_list( &rmesa->vb.dfn_cache.TexCoord3f );
+   make_empty_list( &rmesa->vb.dfn_cache.TexCoord3fv );
    make_empty_list( &rmesa->vb.dfn_cache.TexCoord2f );
    make_empty_list( &rmesa->vb.dfn_cache.TexCoord2fv );
    make_empty_list( &rmesa->vb.dfn_cache.TexCoord1f );
    make_empty_list( &rmesa->vb.dfn_cache.TexCoord1fv );
+   make_empty_list( &rmesa->vb.dfn_cache.MultiTexCoord3fARB );
+   make_empty_list( &rmesa->vb.dfn_cache.MultiTexCoord3fvARB );
    make_empty_list( &rmesa->vb.dfn_cache.MultiTexCoord2fARB );
    make_empty_list( &rmesa->vb.dfn_cache.MultiTexCoord2fvARB );
    make_empty_list( &rmesa->vb.dfn_cache.MultiTexCoord1fARB );
@@ -1126,10 +1212,14 @@ void r200VtxfmtDestroy( GLcontext *ctx )
    free_funcs( &rmesa->vb.dfn_cache.SecondaryColor3fvEXT );
    free_funcs( &rmesa->vb.dfn_cache.Normal3f );
    free_funcs( &rmesa->vb.dfn_cache.Normal3fv );
+   free_funcs( &rmesa->vb.dfn_cache.TexCoord3f );
+   free_funcs( &rmesa->vb.dfn_cache.TexCoord3fv );
    free_funcs( &rmesa->vb.dfn_cache.TexCoord2f );
    free_funcs( &rmesa->vb.dfn_cache.TexCoord2fv );
    free_funcs( &rmesa->vb.dfn_cache.TexCoord1f );
    free_funcs( &rmesa->vb.dfn_cache.TexCoord1fv );
+   free_funcs( &rmesa->vb.dfn_cache.MultiTexCoord3fARB );
+   free_funcs( &rmesa->vb.dfn_cache.MultiTexCoord3fvARB );
    free_funcs( &rmesa->vb.dfn_cache.MultiTexCoord2fARB );
    free_funcs( &rmesa->vb.dfn_cache.MultiTexCoord2fvARB );
    free_funcs( &rmesa->vb.dfn_cache.MultiTexCoord1fARB );
diff --git a/src/mesa/drivers/dri/r200/r200_vtxfmt_c.c b/src/mesa/drivers/dri/r200/r200_vtxfmt_c.c
index 776cb9a5276..e9f66fb5d77 100644
--- a/src/mesa/drivers/dri/r200/r200_vtxfmt_c.c
+++ b/src/mesa/drivers/dri/r200/r200_vtxfmt_c.c
@@ -505,41 +505,30 @@ static void r200_Normal3fv( const GLfloat *v )
 
 /* TexCoord
  */
-static void r200_TexCoord1f( GLfloat s )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[0];
-   dest[0] = s;
-   dest[1] = 0;
-}
 
-static void r200_TexCoord1fv( const GLfloat *v )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[0];
-   dest[0] = v[0];
-   dest[1] = 0;
-}
-
-static void r200_TexCoord2f( GLfloat s, GLfloat t )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[0];
-   dest[0] = s;
-   dest[1] = t;
-}
+#define TEX_to_nF(N, P, S, T, R)					\
+   static void r200_TexCoord ## N P					\
+   { 									\
+      GET_CURRENT_CONTEXT(ctx); r200ContextPtr rmesa = R200_CONTEXT(ctx); \
+      GLfloat * const dest = rmesa->vb.texcoordptr[0];			\
+      switch( ctx->Texture.Unit[0]._ReallyEnabled ) {			\
+      case TEXTURE_CUBE_BIT:						\
+      case TEXTURE_3D_BIT:						\
+	 dest[2] = R; 							\
+      case TEXTURE_2D_BIT:						\
+      case TEXTURE_RECT_BIT:						\
+	 dest[1] = T; 							\
+      case TEXTURE_1D_BIT:						\
+	 dest[0] = S; 							\
+      }									\
+   }
 
-static void r200_TexCoord2fv( const GLfloat *v )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[0];
-   dest[0] = v[0];
-   dest[1] = v[1];
-}
+TEX_to_nF( 1f,  (GLfloat s),                       s,    0.0,  0.0 )
+TEX_to_nF( 2f,  (GLfloat s, GLfloat t),            s,    t,    0.0 )
+TEX_to_nF( 3f,  (GLfloat s, GLfloat t, GLfloat r), s,    t,    r )
+TEX_to_nF( 1fv, (const GLfloat * v),               v[0], 0.0,  0.0 )
+TEX_to_nF( 2fv, (const GLfloat * v),               v[0], v[1], 0.0 )
+TEX_to_nF( 3fv, (const GLfloat * v),               v[0], v[1], v[2] )
 
 
 /* MultiTexcoord
@@ -552,41 +541,30 @@ static void r200_TexCoord2fv( const GLfloat *v )
  * the subtraction has been omitted.
  */
 
-static void r200_MultiTexCoord1fARB( GLenum target, GLfloat s  )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[target & 1];
-   dest[0] = s;
-   dest[1] = 0;
-}
-
-static void r200_MultiTexCoord1fvARB( GLenum target, const GLfloat *v )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[target & 1];
-   dest[0] = v[0];
-   dest[1] = 0;
-}
+#define MTEX_to_nF(N, P, U, S, T, R)					\
+   static void r200_MultiTexCoord ## N ## ARB P				\
+   { 									\
+      GET_CURRENT_CONTEXT(ctx); r200ContextPtr rmesa = R200_CONTEXT(ctx); \
+      GLfloat * const dest = rmesa->vb.texcoordptr[U];			\
+      switch( ctx->Texture.Unit[U]._ReallyEnabled ) {			\
+      case TEXTURE_CUBE_BIT:						\
+      case TEXTURE_3D_BIT:						\
+	 dest[2] = R; 							\
+      case TEXTURE_2D_BIT:						\
+      case TEXTURE_RECT_BIT:						\
+	 dest[1] = T; 							\
+      case TEXTURE_1D_BIT:						\
+	 dest[0] = S; 							\
+      }									\
+   }
 
-static void r200_MultiTexCoord2fARB( GLenum target, GLfloat s, GLfloat t )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[target & 1];
-   dest[0] = s;
-   dest[1] = t;
-}
+MTEX_to_nF( 1f, (GLenum target, GLfloat s), (target & 1), s, 0.0, 0.0 )
+MTEX_to_nF( 2f, (GLenum target, GLfloat s, GLfloat t), (target & 1), s, t, 0.0 )
+MTEX_to_nF( 3f, (GLenum target, GLfloat s, GLfloat t, GLfloat r), (target & 1), s, t, r )
 
-static void r200_MultiTexCoord2fvARB( GLenum target, const GLfloat *v )
-{
-   GET_CURRENT_CONTEXT(ctx);
-   r200ContextPtr rmesa = R200_CONTEXT(ctx);
-   GLfloat *dest = rmesa->vb.texcoordptr[target & 1];
-   dest[0] = v[0];
-   dest[1] = v[1];
-}
+MTEX_to_nF( 1fv, (GLenum target, const GLfloat *v), (target & 1), v[0], 0.0, 0.0 )
+MTEX_to_nF( 2fv, (GLenum target, const GLfloat *v), (target & 1), v[0], v[1], 0.0 )
+MTEX_to_nF( 3fv, (GLenum target, const GLfloat *v), (target & 1), v[0], v[1], v[2] )
 
 static struct dynfn *lookup( struct dynfn *l, const int *key )
 {
@@ -732,6 +710,8 @@ static void choose_##FN ARGS1						\
 
 
 
+
+
 /* VTXFMT_0
  */
 #define MASK_XYZW  (R200_VTX_W0|R200_VTX_Z0)
@@ -749,6 +729,7 @@ typedef void (*p4f)( GLfloat, GLfloat, GLfloat, GLfloat );
 typedef void (*p3f)( GLfloat, GLfloat, GLfloat );
 typedef void (*p2f)( GLfloat, GLfloat );
 typedef void (*p1f)( GLfloat );
+typedef void (*pe3f)( GLenum, GLfloat, GLfloat, GLfloat );
 typedef void (*pe2f)( GLenum, GLfloat, GLfloat );
 typedef void (*pe1f)( GLenum, GLfloat );
 typedef void (*p4ub)( GLubyte, GLubyte, GLubyte, GLubyte );
@@ -793,6 +774,10 @@ CHOOSE_SECONDARY_COLOR(SecondaryColor3fEXT, p3f, MASK_SPEC, 0,
 CHOOSE_SECONDARY_COLOR(SecondaryColor3fvEXT, pfv, MASK_SPEC, 0,
 	(const GLfloat *v), (v))
 
+CHOOSE(TexCoord3f, p3f, ~0, MASK_ST0, 
+       (GLfloat a,GLfloat b,GLfloat c), (a,b,c))
+CHOOSE(TexCoord3fv, pfv, ~0, MASK_ST0, 
+       (const GLfloat *v), (v))
 CHOOSE(TexCoord2f, p2f, ~0, MASK_ST0, 
        (GLfloat a,GLfloat b), (a,b))
 CHOOSE(TexCoord2fv, pfv, ~0, MASK_ST0, 
@@ -802,6 +787,10 @@ CHOOSE(TexCoord1f, p1f, ~0, MASK_ST0,
 CHOOSE(TexCoord1fv, pfv, ~0, MASK_ST0, 
        (const GLfloat *v), (v))
 
+CHOOSE(MultiTexCoord3fARB, pe3f, ~0, ~0,
+	 (GLenum u,GLfloat a,GLfloat b,GLfloat c), (u,a,b,c))
+CHOOSE(MultiTexCoord3fvARB, pefv, ~0, ~0,
+	(GLenum u,const GLfloat *v), (u,v))
 CHOOSE(MultiTexCoord2fARB, pe2f, ~0, ~0,
 	 (GLenum u,GLfloat a,GLfloat b), (u,a,b))
 CHOOSE(MultiTexCoord2fvARB, pefv, ~0, ~0,
@@ -836,12 +825,16 @@ void r200VtxfmtInitChoosers( GLvertexformat *vfmt )
    vfmt->MultiTexCoord1fvARB = choose_MultiTexCoord1fvARB;
    vfmt->MultiTexCoord2fARB = choose_MultiTexCoord2fARB;
    vfmt->MultiTexCoord2fvARB = choose_MultiTexCoord2fvARB;
+   vfmt->MultiTexCoord3fARB = choose_MultiTexCoord3fARB;
+   vfmt->MultiTexCoord3fvARB = choose_MultiTexCoord3fvARB;
    vfmt->Normal3f = choose_Normal3f;
    vfmt->Normal3fv = choose_Normal3fv;
    vfmt->TexCoord1f = choose_TexCoord1f;
    vfmt->TexCoord1fv = choose_TexCoord1fv;
    vfmt->TexCoord2f = choose_TexCoord2f;
    vfmt->TexCoord2fv = choose_TexCoord2fv;
+   vfmt->TexCoord3f = choose_TexCoord3f;
+   vfmt->TexCoord3fv = choose_TexCoord3fv;
    vfmt->Vertex2f = choose_Vertex2f;
    vfmt->Vertex2fv = choose_Vertex2fv;
    vfmt->Vertex3f = choose_Vertex3f;
@@ -874,10 +867,21 @@ void r200InitCodegen( struct dfn_generators *gen, GLboolean useCodegen )
    gen->Color4ubv = codegen_noop;
    gen->Normal3f = codegen_noop;
    gen->Normal3fv = codegen_noop;
+
+   gen->TexCoord3f = codegen_noop;
+   gen->TexCoord3fv = codegen_noop;
    gen->TexCoord2f = codegen_noop;
    gen->TexCoord2fv = codegen_noop;
+   gen->TexCoord1f = codegen_noop;
+   gen->TexCoord1fv = codegen_noop;
+
+   gen->MultiTexCoord3fARB = codegen_noop;
+   gen->MultiTexCoord3fvARB = codegen_noop;
    gen->MultiTexCoord2fARB = codegen_noop;
    gen->MultiTexCoord2fvARB = codegen_noop;
+   gen->MultiTexCoord1fARB = codegen_noop;
+   gen->MultiTexCoord1fvARB = codegen_noop;
+
    gen->Vertex2f = codegen_noop;
    gen->Vertex2fv = codegen_noop;
    gen->Color3ub = codegen_noop;
@@ -890,10 +894,6 @@ void r200InitCodegen( struct dfn_generators *gen, GLboolean useCodegen )
    gen->SecondaryColor3fvEXT = codegen_noop;
    gen->SecondaryColor3ubEXT = codegen_noop;
    gen->SecondaryColor3ubvEXT = codegen_noop;
-   gen->TexCoord1f = codegen_noop;
-   gen->TexCoord1fv = codegen_noop;
-   gen->MultiTexCoord1fARB = codegen_noop;
-   gen->MultiTexCoord1fvARB = codegen_noop;
 
    if (useCodegen) {
 #if defined(USE_X86_ASM)
diff --git a/src/mesa/drivers/dri/r200/r200_vtxfmt_sse.c b/src/mesa/drivers/dri/r200/r200_vtxfmt_sse.c
index 70859b627af..45950159f2d 100644
--- a/src/mesa/drivers/dri/r200/r200_vtxfmt_sse.c
+++ b/src/mesa/drivers/dri/r200/r200_vtxfmt_sse.c
@@ -154,6 +154,7 @@ static struct dynfn *r200_makeSSEColor3f( GLcontext *ctx, const int * key )
    }
 }
 
+#if 0 /* Temporarily disabled as it is broken w/the new cubemap code. - idr */
 static struct dynfn *r200_makeSSETexCoord2fv( GLcontext *ctx, const int * key )
 {
    r200ContextPtr rmesa = R200_CONTEXT(ctx);
@@ -205,6 +206,7 @@ static struct dynfn *r200_makeSSEMultiTexCoord2f( GLcontext *ctx, const int * ke
    }
    return dfn;
 }
+#endif
 
 void r200InitSSECodegen( struct dfn_generators *gen )
 {
@@ -213,10 +215,12 @@ void r200InitSSECodegen( struct dfn_generators *gen )
       gen->Normal3f = (void *) r200_makeSSENormal3f;
       gen->Color3fv = (void *) r200_makeSSEColor3fv;
       gen->Color3f = (void *) r200_makeSSEColor3f;
+#if 0 /* Temporarily disabled as it is broken w/the new cubemap code. - idr */
       gen->TexCoord2fv = (void *) r200_makeSSETexCoord2fv;
       gen->TexCoord2f = (void *) r200_makeSSETexCoord2f;
       gen->MultiTexCoord2fvARB = (void *) r200_makeSSEMultiTexCoord2fv;
       gen->MultiTexCoord2fARB = (void *) r200_makeSSEMultiTexCoord2f;
+#endif
    }
 }
 
diff --git a/src/mesa/drivers/dri/r200/r200_vtxfmt_x86.c b/src/mesa/drivers/dri/r200/r200_vtxfmt_x86.c
index c4d8af85e16..33331781834 100644
--- a/src/mesa/drivers/dri/r200/r200_vtxfmt_x86.c
+++ b/src/mesa/drivers/dri/r200/r200_vtxfmt_x86.c
@@ -337,6 +337,7 @@ struct dynfn *r200_makeX86Color3f( GLcontext *ctx, const int *key )
 
 
 
+#if 0 /* Temporarily disabled as it is broken w/the new cubemap code. - idr */
 struct dynfn *r200_makeX86TexCoord2fv( GLcontext *ctx, const int *key )
 {
    r200ContextPtr rmesa = R200_CONTEXT(ctx);
@@ -395,7 +396,7 @@ struct dynfn *r200_makeX86MultiTexCoord2fARB( GLcontext *ctx,
    }      
    return dfn;
 }
-
+#endif
 
 void r200InitX86Codegen( struct dfn_generators *gen )
 {
@@ -405,10 +406,12 @@ void r200InitX86Codegen( struct dfn_generators *gen )
    gen->Color4ubv = r200_makeX86Color4ubv; /* PKCOLOR only */
    gen->Normal3f = r200_makeX86Normal3f;
    gen->Normal3fv = r200_makeX86Normal3fv;
+#if 0 /* Temporarily disabled as it is broken w/the new cubemap code. - idr */
    gen->TexCoord2f = r200_makeX86TexCoord2f;
    gen->TexCoord2fv = r200_makeX86TexCoord2fv;
    gen->MultiTexCoord2fARB = r200_makeX86MultiTexCoord2fARB;
    gen->MultiTexCoord2fvARB = r200_makeX86MultiTexCoord2fvARB;
+#endif
    gen->Color3f = r200_makeX86Color3f;
    gen->Color3fv = r200_makeX86Color3fv;