Merge commit 'origin/master' into gallium-texture-transfergallium-texture-transfer

Conflicts:

	src/gallium/drivers/softpipe/sp_tile_cache.c

105 files changed, 1066 insertions, 514 deletions

diff --git a/configs/default b/configs/default
index a90d46f18e..2566a860a2 100644
--- a/configs/default
+++ b/configs/default
@@ -92,7 +92,7 @@ EGL_DRIVERS_DIRS = demo
 GALLIUM_DIRS = auxiliary drivers state_trackers
 GALLIUM_AUXILIARY_DIRS = draw translate cso_cache pipebuffer tgsi sct rtasm util indices
 GALLIUM_AUXILIARIES = $(foreach DIR,$(GALLIUM_AUXILIARY_DIRS),$(TOP)/src/gallium/auxiliary/$(DIR)/lib$(DIR).a)
-GALLIUM_DRIVER_DIRS = softpipe i915simple i965simple failover trace
+GALLIUM_DRIVER_DIRS = softpipe #i915simple i965simple failover trace
 GALLIUM_DRIVERS = $(foreach DIR,$(GALLIUM_DRIVER_DIRS),$(TOP)/src/gallium/drivers/$(DIR)/lib$(DIR).a)
 GALLIUM_WINSYS_DIRS = xlib egl_xlib
 GALLIUM_WINSYS_DRM_DIRS =
diff --git a/progs/demos/isosurf.c b/progs/demos/isosurf.c
index 393741cc9d..e280d8f507 100644
--- a/progs/demos/isosurf.c
+++ b/progs/demos/isosurf.c
@@ -69,6 +69,7 @@
 #define NO_STIPPLE	0x08000000
 #define POLYGON_FILL	0x10000000
 #define POLYGON_LINE	0x20000000
+#define POLYGON_POINT	0x40000000
 
 #define LIGHT_MASK		(LIT|UNLIT|REFLECT)
 #define FILTER_MASK		(POINT_FILTER|LINEAR_FILTER)
@@ -81,7 +82,7 @@
 #define SHADE_MASK		(SHADE_SMOOTH|SHADE_FLAT)
 #define FOG_MASK		(FOG|NO_FOG)
 #define STIPPLE_MASK		(STIPPLE|NO_STIPPLE)
-#define POLYGON_MASK		(POLYGON_FILL|POLYGON_LINE)
+#define POLYGON_MASK		(POLYGON_FILL|POLYGON_LINE|POLYGON_POINT)
 
 #define MAXVERTS 10000
 static GLint maxverts = MAXVERTS;
@@ -147,7 +148,7 @@ static void read_surface( char *filename )
 static void print_flags( const char *msg, GLuint flags ) 
 {
    fprintf(stderr, 
-	   "%s (0x%x): %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
+	   "%s (0x%x): %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
 	   msg, flags,
 	   (flags & GLVERTEX) ? "glVertex, " : "",
 	   (flags & DRAW_ARRAYS) ? "glDrawArrays, " : "",
@@ -166,7 +167,8 @@ static void print_flags( const char *msg, GLuint flags )
 	   (flags & MATERIALS) ? "materials, " : "",
 	   (flags & FOG) ? "fog, " : "",
 	   (flags & STIPPLE) ? "stipple, " : "",
-	   (flags & POLYGON_LINE) ? "polygon mode line, " : "");
+	   (flags & POLYGON_LINE) ? "polygon mode line, " : "",
+	   (flags & POLYGON_POINT) ? "polygon mode point, " : "");
 }
 
 
@@ -711,9 +713,12 @@ static void ModeMenu(int m)
       if (m & POLYGON_FILL) {
 	 glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
       }
-      else {
+      else if (m & POLYGON_LINE) {
 	 glPolygonMode(GL_FRONT_AND_BACK, GL_LINE);
       }
+      else {
+	 glPolygonMode(GL_FRONT_AND_BACK, GL_POINT);
+      }
    }
 
 #ifdef GL_EXT_vertex_array
@@ -1089,6 +1094,7 @@ int main(int argc, char **argv)
    glutAddMenuEntry("", 0);
    glutAddMenuEntry("Polygon Mode Fill",     POLYGON_FILL);
    glutAddMenuEntry("Polygon Mode Line",     POLYGON_LINE);
+   glutAddMenuEntry("Polygon Mode Points",   POLYGON_POINT);
    glutAddMenuEntry("", 0);
    glutAddMenuEntry("Point Filtered",        POINT_FILTER);
    glutAddMenuEntry("Linear Filtered",       LINEAR_FILTER);
diff --git a/progs/trivial/.gitignore b/progs/trivial/.gitignore
index d77f120b96..f58345ce3b 100644
--- a/progs/trivial/.gitignore
+++ b/progs/trivial/.gitignore
@@ -102,6 +102,7 @@ tri-tri
 tri-unfilled
 tri-unfilled-clip
 tri-unfilled-edgeflag
+tri-unfilled-fog
 tri-unfilled-smooth
 tri-unfilled-tri
 tri-unfilled-tri-lit
diff --git a/progs/trivial/Makefile b/progs/trivial/Makefile
index dce96f6bc8..2800f3ef82 100644
--- a/progs/trivial/Makefile
+++ b/progs/trivial/Makefile
@@ -108,6 +108,7 @@ SOURCES = \
 	tri-tex.c \
 	tri-tex-3d.c \
 	tri-tri.c \
+	tri-unfilled-fog.c \
 	tri-unfilled-edgeflag.c \
 	tri-unfilled-clip.c \
 	tri-unfilled-smooth.c \
diff --git a/progs/trivial/SConscript b/progs/trivial/SConscript
index 76826d9be0..d699e2b46b 100644
--- a/progs/trivial/SConscript
+++ b/progs/trivial/SConscript
@@ -103,6 +103,7 @@ progs = [
 	'tri-tex',
 	'tri-tex-3d',
 	'tri-tri',
+	'tri-unfilled-fog',
 	'tri-unfilled-edgeflag',
 	'tri-unfilled-clip',
 	'tri-unfilled-smooth',
diff --git a/progs/trivial/tri-unfilled-fog.c b/progs/trivial/tri-unfilled-fog.c
new file mode 100644
index 0000000000..90444aecdf
--- /dev/null
+++ b/progs/trivial/tri-unfilled-fog.c
@@ -0,0 +1,152 @@
+/*
+ * Copyright (c) 1991, 1992, 1993 Silicon Graphics, Inc.
+ *
+ * Permission to use, copy, modify, distribute, and sell this software and
+ * its documentation for any purpose is hereby granted without fee, provided
+ * that (i) the above copyright notices and this permission notice appear in
+ * all copies of the software and related documentation, and (ii) the name of
+ * Silicon Graphics may not be used in any advertising or
+ * publicity relating to the software without the specific, prior written
+ * permission of Silicon Graphics.
+ *
+ * THE SOFTWARE IS PROVIDED "AS-IS" AND WITHOUT WARRANTY OF
+ * ANY KIND,
+ * EXPRESS, IMPLIED OR OTHERWISE, INCLUDING WITHOUT LIMITATION, ANY
+ * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * IN NO EVENT SHALL SILICON GRAPHICS BE LIABLE FOR
+ * ANY SPECIAL, INCIDENTAL, INDIRECT OR CONSEQUENTIAL DAMAGES OF ANY KIND,
+ * OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
+ * WHETHER OR NOT ADVISED OF THE POSSIBILITY OF DAMAGE, AND ON ANY THEORY OF
+ * LIABILITY, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
+ * OF THIS SOFTWARE.
+ */
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <GL/glut.h>
+
+
+#define CI_OFFSET_1 16
+#define CI_OFFSET_2 32
+
+GLint Width = 250, Height = 250;
+
+GLenum doubleBuffer;
+
+static void Init(void)
+{
+   fprintf(stderr, "GL_RENDERER   = %s\n", (char *) glGetString(GL_RENDERER));
+   fprintf(stderr, "GL_VERSION    = %s\n", (char *) glGetString(GL_VERSION));
+   fprintf(stderr, "GL_VENDOR     = %s\n", (char *) glGetString(GL_VENDOR));
+
+   glClearColor(0.0, 0.0, 1.0, 0.0);
+}
+
+static void Reshape(int width, int height)
+{
+   glViewport(0, 0, (GLint)width, (GLint)height);
+
+   glMatrixMode(GL_PROJECTION);
+   glLoadIdentity();
+   glOrtho(-1.0, 1.0, -1.0, 1.0, -0.5, 1000.0);
+   glMatrixMode(GL_MODELVIEW);
+}
+
+static void Key(unsigned char key, int x, int y)
+{
+   switch (key) {
+      case 27:
+         exit(1);
+   }
+
+   glutPostRedisplay();
+}
+
+static void Draw(void)
+{
+   glClear(GL_COLOR_BUFFER_BIT); 
+
+   glEnable(GL_FOG);
+
+   glPolygonMode(GL_FRONT, GL_FILL);
+   glPolygonMode(GL_BACK, GL_FILL);
+
+   glBegin(GL_TRIANGLES);
+   glColor3f(0,0,.7); 
+   glVertex3f( 0.7, -0.7, -30.0);
+   glColor3f(.8,0,0); 
+   glVertex3f( 0.7,  0.7, -30.0);
+   glColor3f(0,.9,0); 
+   glVertex3f(-0.7,  0.0, -30.0);
+   glEnd();
+
+   glPolygonMode(GL_FRONT, GL_LINE);
+   glPolygonMode(GL_BACK, GL_LINE);
+
+   glBegin(GL_TRIANGLES);
+   glColor3f(0,0,.7); 
+   glVertex3f( 0.9, -0.9, -30.0);
+   glColor3f(.8,0,0); 
+   glVertex3f( 0.9,  0.9, -30.0);
+   glColor3f(0,.9,0); 
+   glVertex3f(-0.9,  0.0, -30.0);
+   glEnd();
+
+   glDisable(GL_FOG);
+
+   glFlush();
+
+   if (doubleBuffer) {
+      glutSwapBuffers();
+   }
+}
+
+static GLenum Args(int argc, char **argv)
+{
+   GLint i;
+
+   doubleBuffer = GL_FALSE;
+
+   for (i = 1; i < argc; i++) {
+      if (strcmp(argv[i], "-sb") == 0) {
+         doubleBuffer = GL_FALSE;
+      } else if (strcmp(argv[i], "-db") == 0) {
+         doubleBuffer = GL_TRUE;
+      } else {
+         fprintf(stderr, "%s (Bad option).\n", argv[i]);
+         return GL_FALSE;
+      }
+   }
+   return GL_TRUE;
+}
+
+int main(int argc, char **argv)
+{
+   GLenum type;
+
+   glutInit(&argc, argv);
+
+   if (Args(argc, argv) == GL_FALSE) {
+      exit(1);
+   }
+
+   glutInitWindowPosition(100, 0); glutInitWindowSize(Width, Height);
+
+   type = GLUT_RGB;
+   type |= (doubleBuffer) ? GLUT_DOUBLE : GLUT_SINGLE;
+   glutInitDisplayMode(type);
+
+   if (glutCreateWindow("Filled and unfilled fog tri") == GL_FALSE) {
+      exit(1);
+   }
+
+   Init();
+
+   glutReshapeFunc(Reshape);
+   glutKeyboardFunc(Key);
+   glutDisplayFunc(Draw);
+   glutMainLoop();
+   return 0;
+}
diff --git a/src/gallium/auxiliary/cso_cache/cso_cache.c b/src/gallium/auxiliary/cso_cache/cso_cache.c
index 6b1754ea00..0bc77a5728 100644
--- a/src/gallium/auxiliary/cso_cache/cso_cache.c
+++ b/src/gallium/auxiliary/cso_cache/cso_cache.c
@@ -28,7 +28,7 @@
 /* Authors:  Zack Rusin <zack@tungstengraphics.com>
  */
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "util/u_memory.h"
 
diff --git a/src/gallium/auxiliary/cso_cache/cso_hash.c b/src/gallium/auxiliary/cso_cache/cso_hash.c
index 4e7664f9bf..288cef7b6f 100644
--- a/src/gallium/auxiliary/cso_cache/cso_hash.c
+++ b/src/gallium/auxiliary/cso_cache/cso_hash.c
@@ -30,7 +30,7 @@
   *   Zack Rusin <zack@tungstengraphics.com>
   */
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 
 #include "cso_hash.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
index 9153bc2f86..12325d30d6 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
@@ -29,12 +29,12 @@
  * \file
  * Vertex buffer drawing stage.
  * 
- * \author José Fonseca <jrfonsec@tungstengraphics.com>
+ * \author Jose Fonseca <jrfonsec@tungstengraphics.com>
  * \author Keith Whitwell <keith@tungstengraphics.com>
  */
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
 
@@ -93,7 +93,6 @@ vbuf_stage( struct draw_stage *stage )
 }
 
 
-static void vbuf_flush_indices( struct vbuf_stage *vbuf );
 static void vbuf_flush_vertices( struct vbuf_stage *vbuf );
 static void vbuf_alloc_vertices( struct vbuf_stage *vbuf );
 
@@ -109,13 +108,12 @@ overflow( void *map, void *ptr, unsigned bytes, unsigned bufsz )
 static INLINE void 
 check_space( struct vbuf_stage *vbuf, unsigned nr )
 {
-   if (vbuf->nr_vertices + nr > vbuf->max_vertices ) {
-      vbuf_flush_vertices(vbuf);
-      vbuf_alloc_vertices(vbuf);
+   if (vbuf->nr_vertices + nr > vbuf->max_vertices ||
+       vbuf->nr_indices + nr > vbuf->max_indices)
+   {
+      vbuf_flush_vertices( vbuf );
+      vbuf_alloc_vertices( vbuf );
    }
-
-   if (vbuf->nr_indices + nr > vbuf->max_indices )
-      vbuf_flush_indices(vbuf);
 }
 
 
@@ -202,7 +200,7 @@ vbuf_point( struct draw_stage *stage,
  * will be flushed if needed and a new one allocated.
  */
 static void
-vbuf_set_prim( struct vbuf_stage *vbuf, uint prim )
+vbuf_start_prim( struct vbuf_stage *vbuf, uint prim )
 {
    struct translate_key hw_key;
    unsigned dst_offset;
@@ -217,11 +215,7 @@ vbuf_set_prim( struct vbuf_stage *vbuf, uint prim )
     * state change.
     */
    vbuf->vinfo = vbuf->render->get_vertex_info(vbuf->render);
-
-   if (vbuf->vertex_size != vbuf->vinfo->size * sizeof(float)) {
-      vbuf_flush_vertices(vbuf);
-      vbuf->vertex_size = vbuf->vinfo->size * sizeof(float);
-   }
+   vbuf->vertex_size = vbuf->vinfo->size * sizeof(float);
 
    /* Translate from pipeline vertices to hw vertices.
     */
@@ -294,8 +288,8 @@ vbuf_set_prim( struct vbuf_stage *vbuf, uint prim )
 
    /* Allocate new buffer?
     */
-   if (!vbuf->vertices)
-      vbuf_alloc_vertices(vbuf);
+   assert(vbuf->vertices == NULL);
+   vbuf_alloc_vertices(vbuf);
 }
 
 
@@ -305,9 +299,9 @@ vbuf_first_tri( struct draw_stage *stage,
 {
    struct vbuf_stage *vbuf = vbuf_stage( stage );
 
-   vbuf_flush_indices( vbuf );   
+   vbuf_flush_vertices( vbuf );
+   vbuf_start_prim(vbuf, PIPE_PRIM_TRIANGLES);
    stage->tri = vbuf_tri;
-   vbuf_set_prim(vbuf, PIPE_PRIM_TRIANGLES);
    stage->tri( stage, prim );
 }
 
@@ -318,9 +312,9 @@ vbuf_first_line( struct draw_stage *stage,
 {
    struct vbuf_stage *vbuf = vbuf_stage( stage );
 
-   vbuf_flush_indices( vbuf );
+   vbuf_flush_vertices( vbuf );
+   vbuf_start_prim(vbuf, PIPE_PRIM_LINES);
    stage->line = vbuf_line;
-   vbuf_set_prim(vbuf, PIPE_PRIM_LINES);
    stage->line( stage, prim );
 }
 
@@ -331,53 +325,42 @@ vbuf_first_point( struct draw_stage *stage,
 {
    struct vbuf_stage *vbuf = vbuf_stage( stage );
 
-   vbuf_flush_indices( vbuf );
+   vbuf_flush_vertices(vbuf);
+   vbuf_start_prim(vbuf, PIPE_PRIM_POINTS);
    stage->point = vbuf_point;
-   vbuf_set_prim(vbuf, PIPE_PRIM_POINTS);
    stage->point( stage, prim );
 }
 
 
-static void 
-vbuf_flush_indices( struct vbuf_stage *vbuf ) 
-{
-   if(!vbuf->nr_indices)
-      return;
-   
-   assert((uint) (vbuf->vertex_ptr - vbuf->vertices) == 
-          vbuf->nr_vertices * vbuf->vertex_size / sizeof(unsigned));
-
-   vbuf->render->draw(vbuf->render, vbuf->indices, vbuf->nr_indices);
-   
-   vbuf->nr_indices = 0;
-}
-
 
 /**
  * Flush existing vertex buffer and allocate a new one.
- * 
- * XXX: We separate flush-on-index-full and flush-on-vb-full, but may 
- * raise issues uploading vertices if the hardware wants to flush when
- * we flush.
  */
 static void 
 vbuf_flush_vertices( struct vbuf_stage *vbuf )
 {
-   if(vbuf->vertices) {      
-      vbuf_flush_indices(vbuf);
-      
+   if(vbuf->vertices) {
+
+      vbuf->render->unmap_vertices( vbuf->render, 0, vbuf->nr_vertices - 1 );
+
+      if (vbuf->nr_indices) 
+      {
+         vbuf->render->draw(vbuf->render, 
+                            vbuf->indices, 
+                            vbuf->nr_indices );
+   
+         vbuf->nr_indices = 0;
+      }
+     
       /* Reset temporary vertices ids */
       if(vbuf->nr_vertices)
 	 draw_reset_vertex_ids( vbuf->stage.draw );
       
       /* Free the vertex buffer */
-      vbuf->render->release_vertices(vbuf->render,
-                                     vbuf->vertices,
-                                     vbuf->vertex_size,
-                                     vbuf->nr_vertices);
+      vbuf->render->release_vertices( vbuf->render );
+
       vbuf->max_vertices = vbuf->nr_vertices = 0;
       vbuf->vertex_ptr = vbuf->vertices = NULL;
-      
    }
 }
    
@@ -402,9 +385,12 @@ vbuf_alloc_vertices( struct vbuf_stage *vbuf )
     * and it will flush itself if necessary to do so.  If this does
     * fail, we are basically without usable hardware.
     */
-   vbuf->vertices = (uint *) vbuf->render->allocate_vertices(vbuf->render,
-							     (ushort) vbuf->vertex_size,
-							     (ushort) vbuf->max_vertices);
+   vbuf->render->allocate_vertices(vbuf->render,
+                                   (ushort) vbuf->vertex_size,
+                                   (ushort) vbuf->max_vertices);
+
+   vbuf->vertices = (uint *) vbuf->render->map_vertices( vbuf->render );
+   
    vbuf->vertex_ptr = vbuf->vertices;
 }
 
@@ -415,14 +401,11 @@ vbuf_flush( struct draw_stage *stage, unsigned flags )
 {
    struct vbuf_stage *vbuf = vbuf_stage( stage );
 
-   vbuf_flush_indices( vbuf );
+   vbuf_flush_vertices( vbuf );
 
    stage->point = vbuf_first_point;
    stage->line = vbuf_first_line;
    stage->tri = vbuf_first_tri;
-
-   if (flags & DRAW_FLUSH_BACKEND)
-      vbuf_flush_vertices( vbuf );
 }
 
 
diff --git a/src/gallium/auxiliary/draw/draw_pt.h b/src/gallium/auxiliary/draw/draw_pt.h
index c02f229110..aecaeee5b9 100644
--- a/src/gallium/auxiliary/draw/draw_pt.h
+++ b/src/gallium/auxiliary/draw/draw_pt.h
@@ -173,9 +173,7 @@ void draw_pt_emit( struct pt_emit *emit,
 
 void draw_pt_emit_linear( struct pt_emit *emit,
                           const float (*vertex_data)[4],
-                          unsigned vertex_count,
                           unsigned stride,
-                          unsigned start,
                           unsigned count );
 
 void draw_pt_emit_destroy( struct pt_emit *emit );
diff --git a/src/gallium/auxiliary/draw/draw_pt_emit.c b/src/gallium/auxiliary/draw/draw_pt_emit.c
index 232dfdaed2..064e16c295 100644
--- a/src/gallium/auxiliary/draw/draw_pt_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_emit.c
@@ -165,6 +165,9 @@ void draw_pt_emit( struct pt_emit *emit,
     */
    draw_do_flush( draw, DRAW_FLUSH_BACKEND );
 
+   if (vertex_count == 0)
+      return;
+
    if (vertex_count >= UNDEFINED_VERTEX_ID) {
       assert(0);
       return;
@@ -178,9 +181,11 @@ void draw_pt_emit( struct pt_emit *emit,
       return;
    }
 
-   hw_verts = render->allocate_vertices(render,
-					(ushort)translate->key.output_stride,
-					(ushort)vertex_count);
+   render->allocate_vertices(render,
+                             (ushort)translate->key.output_stride,
+                             (ushort)vertex_count);
+
+   hw_verts = render->map_vertices( render );
    if (!hw_verts) {
       assert(0);
       return;
@@ -201,22 +206,21 @@ void draw_pt_emit( struct pt_emit *emit,
 		   vertex_count,
 		   hw_verts );
 
+   render->unmap_vertices( render, 
+                           0, 
+                           vertex_count - 1 );
+
    render->draw(render,
 		elts,
 		count);
 
-   render->release_vertices(render,
-			    hw_verts,
-			    translate->key.output_stride,
-			    vertex_count);
+   render->release_vertices(render);
 }
 
 
 void draw_pt_emit_linear(struct pt_emit *emit,
                          const float (*vertex_data)[4],
-                         unsigned vertex_count,
                          unsigned stride,
-                         unsigned start,
                          unsigned count)
 {
    struct draw_context *draw = emit->draw;
@@ -231,26 +235,23 @@ void draw_pt_emit_linear(struct pt_emit *emit,
     */
    draw_do_flush( draw, DRAW_FLUSH_BACKEND );
 
-   if (count >= UNDEFINED_VERTEX_ID) {
-      assert(0);
-      return;
-   }
+   if (count >= UNDEFINED_VERTEX_ID)
+      goto fail;
 
    /* XXX: and work out some way to coordinate the render primitive
     * between vbuf.c and here...
     */
-   if (!draw->render->set_primitive(draw->render, emit->prim)) {
-      assert(0);
-      return;
-   }
+   if (!draw->render->set_primitive(draw->render, emit->prim)) 
+      goto fail;
 
-   hw_verts = render->allocate_vertices(render,
-					(ushort)translate->key.output_stride,
-					(ushort)count);
-   if (!hw_verts) {
-      assert(0);
-      return;
-   }
+   if (!render->allocate_vertices(render,
+                                  (ushort)translate->key.output_stride,
+                                  (ushort)count))
+      goto fail;
+
+   hw_verts = render->map_vertices( render );
+   if (!hw_verts)
+      goto fail;
 
    translate->set_buffer(translate, 0,
 			 vertex_data, stride);
@@ -261,12 +262,12 @@ void draw_pt_emit_linear(struct pt_emit *emit,
 
    translate->run(translate,
                   0,
-                  vertex_count,
+                  count,
                   hw_verts);
 
    if (0) {
       unsigned i;
-      for (i = 0; i < vertex_count; i++) {
+      for (i = 0; i < count; i++) {
          debug_printf("\n\n%s vertex %d:\n", __FUNCTION__, i);
          draw_dump_emitted_vertex( emit->vinfo, 
                                    (const uint8_t *)hw_verts + 
@@ -274,13 +275,17 @@ void draw_pt_emit_linear(struct pt_emit *emit,
       }
    }
 
+   render->unmap_vertices( render, 0, count - 1 );
+
+   render->draw_arrays(render, 0, count);
+
+   render->release_vertices(render);
 
-   render->draw_arrays(render, start, count);
+   return;
 
-   render->release_vertices(render,
-			    hw_verts,
-			    translate->key.output_stride,
-			    vertex_count);
+fail:
+   assert(0);
+   return;
 }
 
 struct pt_emit *draw_pt_emit_create( struct draw_context *draw )
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
index dcb7744b17..6b7d02a19b 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
@@ -234,9 +234,11 @@ static void fetch_emit_run( struct draw_pt_middle_end *middle,
       return;
    }
 
-   hw_verts = draw->render->allocate_vertices( draw->render,
-                                               (ushort)feme->translate->key.output_stride,
-                                               (ushort)fetch_count );
+   draw->render->allocate_vertices( draw->render,
+                                    (ushort)feme->translate->key.output_stride,
+                                    (ushort)fetch_count );
+
+   hw_verts = draw->render->map_vertices( draw->render );
    if (!hw_verts) {
       assert(0);
       return;
@@ -259,6 +261,10 @@ static void fetch_emit_run( struct draw_pt_middle_end *middle,
       }
    }
 
+   draw->render->unmap_vertices( draw->render, 
+                                 0, 
+                                 (ushort)(fetch_count - 1) );
+
    /* XXX: Draw arrays path to avoid re-emitting index list again and
     * again.
     */
@@ -268,10 +274,7 @@ static void fetch_emit_run( struct draw_pt_middle_end *middle,
 
    /* Done -- that was easy, wasn't it: 
     */
-   draw->render->release_vertices( draw->render, 
-                                   hw_verts, 
-                                   feme->translate->key.output_stride, 
-                                   fetch_count );
+   draw->render->release_vertices( draw->render );
 
 }
 
@@ -288,18 +291,17 @@ static void fetch_emit_run_linear( struct draw_pt_middle_end *middle,
     */
    draw_do_flush( draw, DRAW_FLUSH_BACKEND );
 
-   if (count >= UNDEFINED_VERTEX_ID) {
-      assert(0);
-      return;
-   }
+   if (count >= UNDEFINED_VERTEX_ID) 
+      goto fail;
 
-   hw_verts = draw->render->allocate_vertices( draw->render,
-                                               (ushort)feme->translate->key.output_stride,
-                                               (ushort)count );
-   if (!hw_verts) {
-      assert(0);
-      return;
-   }
+   if (!draw->render->allocate_vertices( draw->render,
+                                         (ushort)feme->translate->key.output_stride,
+                                         (ushort)count )) 
+      goto fail;
+
+   hw_verts = draw->render->map_vertices( draw->render );
+   if (!hw_verts) 
+      goto fail;
 
    /* Single routine to fetch vertices and emit HW verts.
     */
@@ -317,20 +319,21 @@ static void fetch_emit_run_linear( struct draw_pt_middle_end *middle,
       }
    }
 
+   draw->render->unmap_vertices( draw->render, 0, count - 1 );
+
    /* XXX: Draw arrays path to avoid re-emitting index list again and
     * again.
     */
-   draw->render->draw_arrays( draw->render,
-                              0, /*start*/
-                              count );
+   draw->render->draw_arrays( draw->render, 0, count );
 
    /* Done -- that was easy, wasn't it:
     */
-   draw->render->release_vertices( draw->render,
-                                   hw_verts,
-                                   feme->translate->key.output_stride,
-                                   count );
+   draw->render->release_vertices( draw->render );
+   return;
 
+fail:
+   assert(0);
+   return;
 }
 
 
@@ -351,9 +354,12 @@ static boolean fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
    if (count >= UNDEFINED_VERTEX_ID)
       return FALSE;
 
-   hw_verts = draw->render->allocate_vertices( draw->render,
-                                               (ushort)feme->translate->key.output_stride,
-                                               (ushort)count );
+   if (!draw->render->allocate_vertices( draw->render,
+                                         (ushort)feme->translate->key.output_stride,
+                                         (ushort)count ))
+      return FALSE;
+
+   hw_verts = draw->render->map_vertices( draw->render );
    if (!hw_verts) 
       return FALSE;
 
@@ -364,6 +370,8 @@ static boolean fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
                          count,
                          hw_verts );
 
+   draw->render->unmap_vertices( draw->render, 0, (ushort)(count - 1) );
+
    /* XXX: Draw arrays path to avoid re-emitting index list again and
     * again.
     */
@@ -373,10 +381,7 @@ static boolean fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
 
    /* Done -- that was easy, wasn't it:
     */
-   draw->render->release_vertices( draw->render,
-                                   hw_verts,
-                                   feme->translate->key.output_stride,
-                                   count );
+   draw->render->release_vertices( draw->render );
 
    return TRUE;
 }
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c
index 84ffe3296a..cd9cd4b53f 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c
@@ -234,19 +234,17 @@ static void fse_run_linear( struct draw_pt_middle_end *middle,
     */
    draw_do_flush( draw, DRAW_FLUSH_BACKEND );
 
-   if (count >= UNDEFINED_VERTEX_ID) {
-      assert(0);
-      return;
-   }
+   if (count >= UNDEFINED_VERTEX_ID) 
+      goto fail;
 
-   hw_verts = draw->render->allocate_vertices( draw->render,
-                                               (ushort)fse->key.output_stride,
-                                               (ushort)count );
+   if (!draw->render->allocate_vertices( draw->render,
+                                         (ushort)fse->key.output_stride,
+                                         (ushort)count ))
+      goto fail;
 
-   if (!hw_verts) {
-      assert(0);
-      return;
-   }
+   hw_verts = draw->render->map_vertices( draw->render );
+   if (!hw_verts)
+      goto fail;
 
    /* Single routine to fetch vertices, run shader and emit HW verts.
     * Clipping is done elsewhere -- either by the API or on hardware,
@@ -256,13 +254,7 @@ static void fse_run_linear( struct draw_pt_middle_end *middle,
                             start, count,
                             hw_verts );
 
-   /* Draw arrays path to avoid re-emitting index list again and
-    * again.
-    */
-   draw->render->draw_arrays( draw->render,
-                              0,
-                              count );
-   
+
    if (0) {
       unsigned i;
       for (i = 0; i < count; i++) {
@@ -274,12 +266,24 @@ static void fse_run_linear( struct draw_pt_middle_end *middle,
                                    (const uint8_t *)hw_verts + fse->key.output_stride * i );
       }
    }
+   
+   draw->render->unmap_vertices( draw->render, 0, (ushort)(count - 1) );
 
+   /* Draw arrays path to avoid re-emitting index list again and
+    * again.
+    */
+   draw->render->draw_arrays( draw->render,
+                              0,
+                              count );
+   
+
+   draw->render->release_vertices( draw->render );
+
+   return;
 
-   draw->render->release_vertices( draw->render, 
-				   hw_verts, 
-				   fse->key.output_stride, 
-				   count );
+fail:
+   assert(0);
+   return;
 }
 
 
@@ -298,18 +302,17 @@ fse_run(struct draw_pt_middle_end *middle,
     */
    draw_do_flush( draw, DRAW_FLUSH_BACKEND );
 
-   if (fetch_count >= UNDEFINED_VERTEX_ID) {
-      assert(0);
-      return;
-   }
+   if (fetch_count >= UNDEFINED_VERTEX_ID) 
+      goto fail;
 
-   hw_verts = draw->render->allocate_vertices( draw->render,
-                                               (ushort)fse->key.output_stride,
-                                               (ushort)fetch_count );
-   if (!hw_verts) {
-      assert(0);
-      return;
-   }
+   if (!draw->render->allocate_vertices( draw->render,
+                                         (ushort)fse->key.output_stride,
+                                         (ushort)fetch_count ))
+      goto fail;
+
+   hw_verts = draw->render->map_vertices( draw->render ); 
+   if (!hw_verts) 
+      goto fail;
          
 					
    /* Single routine to fetch vertices, run shader and emit HW verts.
@@ -319,9 +322,6 @@ fse_run(struct draw_pt_middle_end *middle,
                           fetch_count,
                           hw_verts );
 
-   draw->render->draw( draw->render, 
-                       draw_elts, 
-                       draw_count );
 
    if (0) {
       unsigned i;
@@ -333,12 +333,19 @@ fse_run(struct draw_pt_middle_end *middle,
       }
    }
 
+   draw->render->unmap_vertices( draw->render, 0, (ushort)(fetch_count - 1) );
+   
+   draw->render->draw( draw->render, 
+                       draw_elts, 
+                       draw_count );
+
 
-   draw->render->release_vertices( draw->render, 
-                                   hw_verts, 
-                                   fse->key.output_stride, 
-                                   fetch_count );
+   draw->render->release_vertices( draw->render );
+   return;
 
+fail:
+   assert(0);
+   return;
 }
 
 
@@ -360,13 +367,14 @@ static boolean fse_run_linear_elts( struct draw_pt_middle_end *middle,
    if (count >= UNDEFINED_VERTEX_ID)
       return FALSE;
 
-   hw_verts = draw->render->allocate_vertices( draw->render,
-                                               (ushort)fse->key.output_stride,
-                                               (ushort)count );
+   if (!draw->render->allocate_vertices( draw->render,
+                                         (ushort)fse->key.output_stride,
+                                         (ushort)count ))
+      return FALSE;
 
-   if (!hw_verts) {
+   hw_verts = draw->render->map_vertices( draw->render );
+   if (!hw_verts) 
       return FALSE;
-   }
 
    /* Single routine to fetch vertices, run shader and emit HW verts.
     * Clipping is done elsewhere -- either by the API or on hardware,
@@ -382,11 +390,9 @@ static boolean fse_run_linear_elts( struct draw_pt_middle_end *middle,
                        draw_count );
    
 
+   draw->render->unmap_vertices( draw->render, 0, (ushort)(count - 1) );
 
-   draw->render->release_vertices( draw->render, 
-				   hw_verts, 
-				   fse->key.output_stride, 
-				   count );
+   draw->render->release_vertices( draw->render );
 
    return TRUE;
 }
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline.c b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline.c
index ec3b41c320..38f9b604d3 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline.c
@@ -251,9 +251,7 @@ static void fetch_pipeline_linear_run( struct draw_pt_middle_end *middle,
    else {
       draw_pt_emit_linear( fpme->emit,
                            (const float (*)[4])pipeline_verts->data,
-                           count,
                            fpme->vertex_size,
-                           0, /*start*/
                            count );
    }
 
diff --git a/src/gallium/auxiliary/draw/draw_vbuf.h b/src/gallium/auxiliary/draw/draw_vbuf.h
index a1c4c14445..cccd3bf435 100644
--- a/src/gallium/auxiliary/draw/draw_vbuf.h
+++ b/src/gallium/auxiliary/draw/draw_vbuf.h
@@ -80,9 +80,14 @@ struct vbuf_render {
     * Hardware renderers will use ttm memory, others will just malloc
     * something.
     */
-   void *(*allocate_vertices)( struct vbuf_render *,
-			       ushort vertex_size,
-			       ushort nr_vertices );
+   boolean (*allocate_vertices)( struct vbuf_render *,
+                                 ushort vertex_size,
+                                 ushort nr_vertices );
+   
+   void *(*map_vertices)( struct vbuf_render * );
+   void (*unmap_vertices)( struct vbuf_render *, 
+                           ushort min_index,
+                           ushort max_index );
 
    /**
     * Notify the renderer of the current primitive when it changes.
@@ -109,10 +114,7 @@ struct vbuf_render {
    /**
     * Called when vbuf is done with this set of vertices:
     */
-   void (*release_vertices)( struct vbuf_render *,
-			     void *vertices, 
-			     unsigned vertex_size,
-			     unsigned vertices_used );
+   void (*release_vertices)( struct vbuf_render * );
 
    void (*destroy)( struct vbuf_render * );
 };
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index 0c693a4a65..46a36af418 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -32,7 +32,7 @@
 #include "util/u_memory.h"
 #include "util/u_math.h"
 #include "pipe/p_shader_tokens.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_util.h"
 #include "tgsi/tgsi_exec.h"
@@ -284,6 +284,14 @@ void aos_release_xmm_reg( struct aos_compilation *cp,
 }
 
 
+static void aos_soft_release_xmm( struct aos_compilation *cp,
+                                  struct x86_reg reg )
+{
+   if (reg.file == file_XMM) {
+      cp->xmm[reg.idx].last_used = cp->insn_counter - 1;
+   }
+}
+
 
      
 /* Mark an xmm reg as holding the current copy of a shader reg.
@@ -584,10 +592,12 @@ static struct x86_reg fetch_src( struct aos_compilation *cp,
          sse_mulps(cp->func, dst, tmp);
 
          aos_release_xmm_reg(cp, tmp.idx);
+         aos_soft_release_xmm(cp, imm_swz);
       }
       else if (negs) {
          struct x86_reg imm_negs = aos_get_internal_xmm(cp, IMM_NEGS);
          sse_mulps(cp->func, dst, imm_negs);
+         aos_soft_release_xmm(cp, imm_negs);
       }
 
 
@@ -603,8 +613,10 @@ static struct x86_reg fetch_src( struct aos_compilation *cp,
          sse_maxps(cp->func, dst, tmp);
 
          aos_release_xmm_reg(cp, tmp.idx);
+         aos_soft_release_xmm(cp, neg);
       }
 
+      aos_soft_release_xmm(cp, arg0);
       return dst;
    }
       
@@ -1559,7 +1571,6 @@ static boolean emit_RCP( struct aos_compilation *cp, const struct tgsi_full_inst
  */
 static boolean emit_RSQ( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
 {
-
    if (0) {
       struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
       struct x86_reg r = aos_get_xmm_reg(cp);
@@ -1568,21 +1579,30 @@ static boolean emit_RSQ( struct aos_compilation *cp, const struct tgsi_full_inst
       return TRUE;
    }
    else {
-      struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
-      struct x86_reg r = aos_get_xmm_reg(cp);
+      struct x86_reg arg0           = fetch_src(cp, &op->FullSrcRegisters[0]);
+      struct x86_reg r              = aos_get_xmm_reg(cp);
 
       struct x86_reg neg_half       = get_reg_ptr( cp, AOS_FILE_INTERNAL, IMM_RSQ );
       struct x86_reg one_point_five = x86_make_disp( neg_half, 4 );
       struct x86_reg src            = get_xmm_writable( cp, arg0 );
-      
-      sse_rsqrtss( cp->func, r, src  );             /* rsqrtss(a) */
-      sse_mulss(   cp->func, src, neg_half  );      /* -.5 * a */
-      sse_mulss(   cp->func, src,  r );             /* -.5 * a * r */
-      sse_mulss(   cp->func, src,  r );             /* -.5 * a * r * r */
-      sse_addss(   cp->func, src, one_point_five ); /* 1.5 - .5 * a * r * r */
-      sse_mulss(   cp->func, r,  src );             /* r * (1.5 - .5 * a * r * r) */
+      struct x86_reg neg            = aos_get_internal(cp, IMM_NEGS);
+      struct x86_reg tmp            = aos_get_xmm_reg(cp);
+
+      sse_movaps(cp->func, tmp, src);
+      sse_mulps(cp->func, tmp, neg);
+      sse_maxps(cp->func, tmp, src);
+   
+      sse_rsqrtss( cp->func, r, tmp  );             /* rsqrtss(a) */
+      sse_mulss(   cp->func, tmp, neg_half  );      /* -.5 * a */
+      sse_mulss(   cp->func, tmp,  r );             /* -.5 * a * r */
+      sse_mulss(   cp->func, tmp,  r );             /* -.5 * a * r * r */
+      sse_addss(   cp->func, tmp, one_point_five ); /* 1.5 - .5 * a * r * r */
+      sse_mulss(   cp->func, r,  tmp );             /* r * (1.5 - .5 * a * r * r) */
 
       store_scalar_dest(cp, &op->FullDstRegisters[0], r);
+
+      aos_release_xmm_reg(cp, tmp.idx);
+
       return TRUE;
    }
 }
diff --git a/src/gallium/auxiliary/gallivm/storagesoa.cpp b/src/gallium/auxiliary/gallivm/storagesoa.cpp
index e1e5cabcf5..4984ce985c 100644
--- a/src/gallium/auxiliary/gallivm/storagesoa.cpp
+++ b/src/gallium/auxiliary/gallivm/storagesoa.cpp
@@ -30,7 +30,7 @@
 #include "gallivm_p.h"
 
 #include "pipe/p_shader_tokens.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include <llvm/BasicBlock.h>
 #include <llvm/Module.h>
diff --git a/src/gallium/auxiliary/indices/u_indices_gen.c b/src/gallium/auxiliary/indices/u_indices_gen.c
index 4c05b3eedb..3c981e5d7f 100644
--- a/src/gallium/auxiliary/indices/u_indices_gen.c
+++ b/src/gallium/auxiliary/indices/u_indices_gen.c
@@ -34,7 +34,7 @@
 #include "indices/u_indices.h"
 #include "indices/u_indices_priv.h"
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_defines.h"
 #include "util/u_memory.h"
 
diff --git a/src/gallium/auxiliary/indices/u_indices_gen.py b/src/gallium/auxiliary/indices/u_indices_gen.py
index 0dc58d0cd0..af63d09930 100644
--- a/src/gallium/auxiliary/indices/u_indices_gen.py
+++ b/src/gallium/auxiliary/indices/u_indices_gen.py
@@ -72,7 +72,7 @@ def prolog():
 #include "indices/u_indices.h"
 #include "indices/u_indices_priv.h"
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_defines.h"
 #include "util/u_memory.h"
 
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer.h b/src/gallium/auxiliary/pipebuffer/pb_buffer.h
index d8f1f02d68..e6b0b30ff4 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer.h
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer.h
@@ -45,7 +45,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_error.h"
 #include "pipe/p_state.h"
 #include "pipe/p_inlines.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
index f9e6226436..0cddc95aa6 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
@@ -43,7 +43,7 @@
 
 #include "pipe/p_compiler.h"
 #include "pipe/p_error.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "util/u_memory.h"
 #include "util/u_double_list.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h
index d1c9d4c17d..c2b29e974b 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h
@@ -51,7 +51,7 @@
 #define PB_BUFFER_FENCED_H_
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 
 #ifdef __cplusplus
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c b/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c
index 53f497cfb0..282802b171 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c
@@ -34,7 +34,7 @@
  */
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 #include "pb_buffer.h"
 #include "pb_bufmgr.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c
index c956924cc7..db67d46c56 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c
@@ -34,7 +34,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 
 #include "pb_buffer.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c
index a168853713..29117efe9b 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c
@@ -35,7 +35,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "util/u_memory.h"
 #include "util/u_double_list.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c
index 26d9c24aec..070bf3f517 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c
@@ -34,7 +34,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
index 47e9fee533..e352f5357b 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
@@ -34,7 +34,7 @@
  */
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 
 #include "pb_buffer.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c
index 2f5a5d8ea0..f3b1ca73b0 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c
@@ -34,7 +34,7 @@
 
 
 #include "pipe/p_defines.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "util/u_memory.h"
 #include "util/u_double_list.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c
index ba02a84e62..3d9c7bba0b 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c
@@ -34,7 +34,7 @@
  */
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 #include "pb_buffer.h"
 #include "pb_bufmgr.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c
index a6ff37653e..12447acfd9 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c
@@ -36,7 +36,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "pipe/p_defines.h"
 #include "util/u_memory.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c
index 9b9fedccb4..a3259351b9 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c
@@ -38,7 +38,7 @@
 
 #include "pipe/p_compiler.h"
 #include "pipe/p_error.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "pipe/p_defines.h"
 #include "util/u_memory.h"
diff --git a/src/gallium/auxiliary/pipebuffer/pb_validate.c b/src/gallium/auxiliary/pipebuffer/pb_validate.c
index 94532bb4ce..150fd50618 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_validate.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_validate.c
@@ -36,7 +36,7 @@
 #include "pipe/p_compiler.h"
 #include "pipe/p_error.h"
 #include "util/u_memory.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "pb_buffer.h"
 #include "pb_buffer_fenced.h"
diff --git a/src/gallium/auxiliary/rtasm/rtasm_cpu.c b/src/gallium/auxiliary/rtasm/rtasm_cpu.c
index 5499018b21..03bdd47238 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_cpu.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_cpu.c
@@ -26,7 +26,7 @@
  **************************************************************************/
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "rtasm_cpu.h"
 
 
diff --git a/src/gallium/auxiliary/rtasm/rtasm_execmem.c b/src/gallium/auxiliary/rtasm/rtasm_execmem.c
index be7433baf8..5acc5bcb7b 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_execmem.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_execmem.c
@@ -31,7 +31,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 #include "util/u_memory.h"
 
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.c b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
index 1bb9026205..e3586482db 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
@@ -38,7 +38,7 @@
 
 #include <stdio.h>
 #include "util/u_memory.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "rtasm_execmem.h"
 #include "rtasm_ppc.h"
 
diff --git a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c
index 99ee74cf14..57fcf6de2a 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_x86sse.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_x86sse.c
@@ -26,7 +26,7 @@
 #if defined(PIPE_ARCH_X86)
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_pointer.h"
 
 #include "rtasm_execmem.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_build.c b/src/gallium/auxiliary/tgsi/tgsi_build.c
index 17886540cf..a1891a140a 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_build.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_build.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_shader_tokens.h"
 #include "tgsi_build.h"
 #include "tgsi_parse.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump.c b/src/gallium/auxiliary/tgsi/tgsi_dump.c
index ab2b1f2c58..d57cb9139f 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_string.h"
 #include "tgsi_dump.h"
 #include "tgsi_info.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump_c.c b/src/gallium/auxiliary/tgsi/tgsi_dump_c.c
index 2ecf1e2f14..3dc61c48ca 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump_c.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump_c.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_string.h"
 #include "tgsi_dump_c.h"
 #include "tgsi_build.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c
index ab641efb60..5c5d8d2550 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
@@ -1889,6 +1889,7 @@ exec_instruction(
    case TGSI_OPCODE_RSQ:
    /* TGSI_OPCODE_RECIPSQRT */
       FETCH( &r[0], 0, CHAN_X );
+      micro_abs( &r[0], &r[0] );
       micro_sqrt( &r[0], &r[0] );
       micro_div( &r[0], &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &r[0] );
       FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
diff --git a/src/gallium/auxiliary/tgsi/tgsi_info.c b/src/gallium/auxiliary/tgsi/tgsi_info.c
index 68c7a6b7f5..2b8a6f0fb1 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_info.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_info.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "tgsi_info.h"
 
 static const struct tgsi_opcode_info opcode_info[TGSI_OPCODE_LAST] =
diff --git a/src/gallium/auxiliary/tgsi/tgsi_iterate.c b/src/gallium/auxiliary/tgsi/tgsi_iterate.c
index 5371a88b96..d88c2558d8 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_iterate.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_iterate.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "tgsi_iterate.h"
 
 boolean
diff --git a/src/gallium/auxiliary/tgsi/tgsi_parse.c b/src/gallium/auxiliary/tgsi/tgsi_parse.c
index d374b16f9a..22006edf3d 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_parse.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_parse.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_shader_tokens.h"
 #include "tgsi_parse.h"
 #include "tgsi_build.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_ppc.c b/src/gallium/auxiliary/tgsi/tgsi_ppc.c
index f365030e52..0c64ae5713 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_ppc.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_ppc.c
@@ -33,7 +33,7 @@
 
 #if defined(PIPE_ARCH_PPC)
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_shader_tokens.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_sanity.c b/src/gallium/auxiliary/tgsi/tgsi_sanity.c
index bc7b941b78..76e773da91 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_sanity.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sanity.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "tgsi_sanity.h"
 #include "tgsi_info.h"
 #include "tgsi_iterate.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
index 481ba89c5e..d70bcd03c5 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
@@ -29,7 +29,7 @@
 
 #if defined(PIPE_ARCH_X86)
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_shader_tokens.h"
 #include "util/u_math.h"
 #if defined(PIPE_ARCH_SSE)
@@ -1575,6 +1575,7 @@ emit_instruction(
    case TGSI_OPCODE_RSQ:
    /* TGSI_OPCODE_RECIPSQRT */
       FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_abs( func, 0 );
       emit_rsqrt( func, 1, 0 );
       FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
          STORE( func, *inst, 1, 0, chan_index );
diff --git a/src/gallium/auxiliary/tgsi/tgsi_text.c b/src/gallium/auxiliary/tgsi/tgsi_text.c
index 1e822fbbea..58fe07c11d 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_text.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_text.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "tgsi_text.h"
 #include "tgsi_build.h"
 #include "tgsi_info.h"
diff --git a/src/gallium/auxiliary/tgsi/tgsi_transform.c b/src/gallium/auxiliary/tgsi/tgsi_transform.c
index ea87da31e5..062c1be938 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_transform.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_transform.c
@@ -31,7 +31,7 @@
  * Authors:  Brian Paul
  */
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "tgsi_transform.h"
 
diff --git a/src/gallium/auxiliary/tgsi/tgsi_util.c b/src/gallium/auxiliary/tgsi/tgsi_util.c
index 50101a9bb0..71f8a6ca40 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_util.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_util.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_shader_tokens.h"
 #include "tgsi_parse.h"
 #include "tgsi_build.h"
diff --git a/src/gallium/auxiliary/util/Makefile b/src/gallium/auxiliary/util/Makefile
index 44c2377721..671e671df2 100644
--- a/src/gallium/auxiliary/util/Makefile
+++ b/src/gallium/auxiliary/util/Makefile
@@ -4,7 +4,7 @@ include $(TOP)/configs/current
 LIBNAME = util
 
 C_SOURCES = \
-	p_debug.c \
+	u_debug.c \
 	u_blit.c \
 	u_cache.c \
 	u_draw_quad.c \
diff --git a/src/gallium/auxiliary/util/SConscript b/src/gallium/auxiliary/util/SConscript
index 35f683fb8e..84e4c48476 100644
--- a/src/gallium/auxiliary/util/SConscript
+++ b/src/gallium/auxiliary/util/SConscript
@@ -3,11 +3,11 @@ Import('*')
 util = env.ConvenienceLibrary(
 	target = 'util',
 	source = [
-		'p_debug.c',
-		'p_debug_mem.c',
-		'p_debug_prof.c',
 		'u_blit.c',
 		'u_cache.c',
+		'u_debug.c',
+		'u_debug_memory.c',
+		'u_debug_profile.c',
 		'u_draw_quad.c',
 		'u_gen_mipmap.c',
 		'u_handle_table.c',
diff --git a/src/gallium/auxiliary/util/u_blit.c b/src/gallium/auxiliary/util/u_blit.c
index 841e9c01e7..efc3a874cc 100644
--- a/src/gallium/auxiliary/util/u_blit.c
+++ b/src/gallium/auxiliary/util/u_blit.c
@@ -34,7 +34,7 @@
 
 
 #include "pipe/p_context.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_inlines.h"
 #include "pipe/p_shader_tokens.h"
diff --git a/src/gallium/auxiliary/util/u_cache.c b/src/gallium/auxiliary/util/u_cache.c
index 0a1a64259f..41cd38171f 100644
--- a/src/gallium/auxiliary/util/u_cache.c
+++ b/src/gallium/auxiliary/util/u_cache.c
@@ -36,7 +36,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "util/u_math.h"
 #include "util/u_memory.h"
diff --git a/src/gallium/auxiliary/util/p_debug.c b/src/gallium/auxiliary/util/u_debug.c
index 2a716c4e30..7f2a113dea 100644
--- a/src/gallium/auxiliary/util/p_debug.c
+++ b/src/gallium/auxiliary/util/u_debug.c
@@ -59,7 +59,7 @@
 #endif
 
 #include "pipe/p_compiler.h" 
-#include "pipe/p_debug.h" 
+#include "util/u_debug.h" 
 #include "pipe/p_format.h" 
 #include "pipe/p_state.h" 
 #include "pipe/p_inlines.h" 
diff --git a/src/gallium/include/pipe/p_debug.h b/src/gallium/auxiliary/util/u_debug.h
index e9c95982dd..b298b9b66d 100644
--- a/src/gallium/include/pipe/p_debug.h
+++ b/src/gallium/auxiliary/util/u_debug.h
@@ -35,13 +35,13 @@
  * @author Jose Fonseca <jrfonseca@tungstengraphics.com>
  */
 
-#ifndef P_DEBUG_H_
-#define P_DEBUG_H_
+#ifndef U_DEBUG_H_
+#define U_DEBUG_H_
 
 
 #include <stdarg.h>
 
-#include "p_compiler.h"
+#include "pipe/p_compiler.h"
 
 
 #ifdef	__cplusplus
@@ -358,4 +358,4 @@ void debug_dump_surface_bmp(const char *filename,
 }
 #endif
 
-#endif /* P_DEBUG_H_ */
+#endif /* U_DEBUG_H_ */
diff --git a/src/gallium/auxiliary/util/p_debug_mem.c b/src/gallium/auxiliary/util/u_debug_memory.c
index 250fd60f63..f6c136f6e5 100644
--- a/src/gallium/auxiliary/util/p_debug_mem.c
+++ b/src/gallium/auxiliary/util/u_debug_memory.c
@@ -44,7 +44,7 @@
 #include <stdlib.h>
 #endif
 
-#include "pipe/p_debug.h" 
+#include "util/u_debug.h" 
 #include "util/u_double_list.h" 
 
 
diff --git a/src/gallium/auxiliary/util/p_debug_prof.c b/src/gallium/auxiliary/util/u_debug_profile.c
index 5f9772ef91..5f9772ef91 100644
--- a/src/gallium/auxiliary/util/p_debug_prof.c
+++ b/src/gallium/auxiliary/util/u_debug_profile.c
diff --git a/src/gallium/auxiliary/util/u_gen_mipmap.c b/src/gallium/auxiliary/util/u_gen_mipmap.c
index 47dd3fc61f..90483fcb21 100644
--- a/src/gallium/auxiliary/util/u_gen_mipmap.c
+++ b/src/gallium/auxiliary/util/u_gen_mipmap.c
@@ -35,7 +35,7 @@
 
 
 #include "pipe/p_context.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_inlines.h"
 #include "pipe/p_shader_tokens.h"
diff --git a/src/gallium/auxiliary/util/u_handle_table.c b/src/gallium/auxiliary/util/u_handle_table.c
index 2d15932ce3..6da7353e25 100644
--- a/src/gallium/auxiliary/util/u_handle_table.c
+++ b/src/gallium/auxiliary/util/u_handle_table.c
@@ -34,7 +34,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "util/u_memory.h"
 #include "util/u_handle_table.h"
diff --git a/src/gallium/auxiliary/util/u_hash_table.c b/src/gallium/auxiliary/util/u_hash_table.c
index 0bc8de9632..2f83e318e4 100644
--- a/src/gallium/auxiliary/util/u_hash_table.c
+++ b/src/gallium/auxiliary/util/u_hash_table.c
@@ -39,7 +39,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "cso_cache/cso_hash.h"
 
diff --git a/src/gallium/auxiliary/util/u_keymap.c b/src/gallium/auxiliary/util/u_keymap.c
index 01b17ddb1b..3f70809efd 100644
--- a/src/gallium/auxiliary/util/u_keymap.c
+++ b/src/gallium/auxiliary/util/u_keymap.c
@@ -35,7 +35,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_error.h"
 
 #include "cso_cache/cso_hash.h"
diff --git a/src/gallium/auxiliary/util/u_linear.c b/src/gallium/auxiliary/util/u_linear.c
index e999cefe74..6be365e53b 100644
--- a/src/gallium/auxiliary/util/u_linear.c
+++ b/src/gallium/auxiliary/util/u_linear.c
@@ -1,5 +1,5 @@
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "u_linear.h"
 
 void
diff --git a/src/gallium/auxiliary/util/u_math.h b/src/gallium/auxiliary/util/u_math.h
index ab6f39ac31..1ecde7a912 100644
--- a/src/gallium/auxiliary/util/u_math.h
+++ b/src/gallium/auxiliary/util/u_math.h
@@ -40,7 +40,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 
 #ifdef __cplusplus
diff --git a/src/gallium/auxiliary/util/u_memory.h b/src/gallium/auxiliary/util/u_memory.h
index 1a6b596421..ceb3a1cb61 100644
--- a/src/gallium/auxiliary/util/u_memory.h
+++ b/src/gallium/auxiliary/util/u_memory.h
@@ -36,7 +36,7 @@
 
 
 #include "util/u_pointer.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 
 #ifdef __cplusplus
@@ -52,11 +52,11 @@ extern "C" {
 #endif
 
 
-#if defined(PIPE_SUBSYSTEM_WINDOWS_DISPLAY) && defined(DEBUG) 
+#if defined(PIPE_OS_WINDOWS) && defined(DEBUG) 
 
 /* memory debugging */
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #define MALLOC( _size ) \
    debug_malloc( __FILE__, __LINE__, __FUNCTION__, _size )
diff --git a/src/gallium/auxiliary/util/u_mm.c b/src/gallium/auxiliary/util/u_mm.c
index 45ce257b5e..151a480d34 100644
--- a/src/gallium/auxiliary/util/u_mm.c
+++ b/src/gallium/auxiliary/util/u_mm.c
@@ -24,7 +24,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "util/u_memory.h"
 #include "util/u_mm.h"
diff --git a/src/gallium/auxiliary/util/u_simple_shaders.c b/src/gallium/auxiliary/util/u_simple_shaders.c
index 706155e99a..3cd2d52c64 100644
--- a/src/gallium/auxiliary/util/u_simple_shaders.c
+++ b/src/gallium/auxiliary/util/u_simple_shaders.c
@@ -34,7 +34,7 @@
 
 
 #include "pipe/p_context.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_inlines.h"
 #include "pipe/p_screen.h"
diff --git a/src/gallium/drivers/cell/spu/spu_util.c b/src/gallium/drivers/cell/spu/spu_util.c
index b8a0d4a265..af25dd3718 100644
--- a/src/gallium/drivers/cell/spu/spu_util.c
+++ b/src/gallium/drivers/cell/spu/spu_util.c
@@ -1,7 +1,7 @@
 
 #include "cell/common.h"
 #include "pipe/p_shader_tokens.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "tgsi/tgsi_parse.h"
 //#include "tgsi_build.h"
 #include "tgsi/tgsi_util.h"
diff --git a/src/gallium/drivers/i915simple/i915_debug.c b/src/gallium/drivers/i915simple/i915_debug.c
index a300b61c3b..e08582efab 100644
--- a/src/gallium/drivers/i915simple/i915_debug.c
+++ b/src/gallium/drivers/i915simple/i915_debug.c
@@ -31,7 +31,7 @@
 #include "i915_debug.h"
 #include "i915_batch.h"
 #include "pipe/internal/p_winsys_screen.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 
 static void
diff --git a/src/gallium/drivers/i915simple/i915_prim_vbuf.c b/src/gallium/drivers/i915simple/i915_prim_vbuf.c
index f49f6d6ed1..58c41840e1 100644
--- a/src/gallium/drivers/i915simple/i915_prim_vbuf.c
+++ b/src/gallium/drivers/i915simple/i915_prim_vbuf.c
@@ -40,7 +40,7 @@
 
 #include "draw/draw_context.h"
 #include "draw/draw_vbuf.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_inlines.h"
 #include "pipe/internal/p_winsys_screen.h"
 #include "util/u_math.h"
@@ -62,7 +62,7 @@ struct i915_vbuf_render {
    struct i915_context *i915;   
 
    /** Vertex size in bytes */
-   unsigned vertex_size;
+   size_t vertex_size;
 
    /** Software primitive */
    unsigned prim;
@@ -79,6 +79,7 @@ struct i915_vbuf_render {
    size_t vbo_offset;
    void *vbo_ptr;
    size_t vbo_alloc_size;
+   size_t vbo_max_used;
 };
 
 
@@ -108,7 +109,7 @@ i915_vbuf_render_get_vertex_info( struct vbuf_render *render )
 }
 
 
-static void *
+static boolean
 i915_vbuf_render_allocate_vertices( struct vbuf_render *render,
                                     ushort vertex_size,
                                     ushort nr_vertices )
@@ -124,7 +125,8 @@ i915_vbuf_render_allocate_vertices( struct vbuf_render *render,
    if (i915_render->vbo_size > size + i915_render->vbo_offset && !i915->vbo_flushed) {
    } else {
       i915->vbo_flushed = 0;
-      pipe_buffer_reference(screen, &i915_render->vbo, NULL);
+      if (i915_render->vbo)
+         pipe_buffer_reference(screen, &i915_render->vbo, NULL);
    }
 
    if (!i915_render->vbo) {
@@ -134,19 +136,49 @@ i915_vbuf_render_allocate_vertices( struct vbuf_render *render,
                                             64,
                                             I915_BUFFER_USAGE_LIT_VERTEX,
                                             i915_render->vbo_size);
-      i915_render->vbo_ptr = pipe_buffer_map(screen,
-                                             i915_render->vbo,
-                                             PIPE_BUFFER_USAGE_CPU_WRITE);
-      pipe_buffer_unmap(screen, i915_render->vbo);
+
    }
 
+   i915_render->vertex_size = vertex_size;
    i915->vbo = i915_render->vbo;
    i915->vbo_offset = i915_render->vbo_offset;
    i915->dirty |= I915_NEW_VBO;
 
+   if (!i915_render->vbo)
+      return FALSE;
+   return TRUE;
+}
+
+
+static void *
+i915_vbuf_render_map_vertices( struct vbuf_render *render )
+{
+   struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
+   struct i915_context *i915 = i915_render->i915;
+   struct pipe_screen *screen = i915->pipe.screen;
+
+   if (i915->vbo_flushed)
+      debug_printf("%s bad vbo flush occured stalling on hw\n");
+
+   i915_render->vbo_ptr = pipe_buffer_map(screen,
+                                          i915_render->vbo,
+                                          PIPE_BUFFER_USAGE_CPU_WRITE);
+
    return (unsigned char *)i915_render->vbo_ptr + i915->vbo_offset;
 }
 
+static void
+i915_vbuf_render_unmap_vertices( struct vbuf_render *render,
+                                 ushort min_index,
+                                 ushort max_index )
+{
+   struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
+   struct i915_context *i915 = i915_render->i915;
+   struct pipe_screen *screen = i915->pipe.screen;
+
+   i915_render->vbo_max_used = MAX2(i915_render->vbo_max_used, i915_render->vertex_size * (max_index + 1));
+   pipe_buffer_unmap(screen, i915_render->vbo);
+}
 
 static boolean
 i915_vbuf_render_set_primitive( struct vbuf_render *render, 
@@ -454,18 +486,15 @@ out:
 
 
 static void
-i915_vbuf_render_release_vertices( struct vbuf_render *render,
-			           void *vertices, 
-			           unsigned vertex_size,
-			           unsigned vertices_used )
+i915_vbuf_render_release_vertices( struct vbuf_render *render )
 {
    struct i915_vbuf_render *i915_render = i915_vbuf_render(render);
    struct i915_context *i915 = i915_render->i915;
-   size_t size = (size_t)vertex_size * (size_t)vertices_used;
 
    assert(i915->vbo);
 
-   i915_render->vbo_offset += size;
+   i915_render->vbo_offset += i915_render->vbo_max_used;
+   i915_render->vbo_max_used = 0;
    i915->vbo = NULL;
    i915->dirty |= I915_NEW_VBO;
 }
@@ -499,6 +528,8 @@ i915_vbuf_render_create( struct i915_context *i915 )
 
    i915_render->base.get_vertex_info = i915_vbuf_render_get_vertex_info;
    i915_render->base.allocate_vertices = i915_vbuf_render_allocate_vertices;
+   i915_render->base.map_vertices = i915_vbuf_render_map_vertices;
+   i915_render->base.unmap_vertices = i915_vbuf_render_unmap_vertices;
    i915_render->base.set_primitive = i915_vbuf_render_set_primitive;
    i915_render->base.draw = i915_vbuf_render_draw;
    i915_render->base.draw_arrays = i915_vbuf_render_draw_arrays;
diff --git a/src/gallium/drivers/i965simple/brw_eu_debug.c b/src/gallium/drivers/i965simple/brw_eu_debug.c
index 4a94ddefa6..4adfb0c02f 100644
--- a/src/gallium/drivers/i965simple/brw_eu_debug.c
+++ b/src/gallium/drivers/i965simple/brw_eu_debug.c
@@ -30,7 +30,7 @@
   */
     
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 #include "brw_eu.h"
 
diff --git a/src/gallium/drivers/nouveau/nouveau_stateobj.h b/src/gallium/drivers/nouveau/nouveau_stateobj.h
index 4ae4ff4940..029b01e17d 100644
--- a/src/gallium/drivers/nouveau/nouveau_stateobj.h
+++ b/src/gallium/drivers/nouveau/nouveau_stateobj.h
@@ -1,7 +1,7 @@
 #ifndef __NOUVEAU_STATEOBJ_H__
 #define __NOUVEAU_STATEOBJ_H__
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 struct nouveau_stateobj_reloc {
 	struct pipe_buffer *bo;
diff --git a/src/gallium/drivers/nv04/nv04_prim_vbuf.c b/src/gallium/drivers/nv04/nv04_prim_vbuf.c
index 18a8872ae3..f6458232ae 100644
--- a/src/gallium/drivers/nv04/nv04_prim_vbuf.c
+++ b/src/gallium/drivers/nv04/nv04_prim_vbuf.c
@@ -1,5 +1,5 @@
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_inlines.h"
 #include "pipe/internal/p_winsys_screen.h"
 #include "pipe/p_compiler.h"
@@ -51,7 +51,7 @@ nv04_vbuf_render_get_vertex_info( struct vbuf_render *render )
 }
 
 
-static void *
+static boolean
 nv04_vbuf_render_allocate_vertices( struct vbuf_render *render,
 		ushort vertex_size,
 		ushort nr_vertices )
@@ -61,9 +61,22 @@ nv04_vbuf_render_allocate_vertices( struct vbuf_render *render,
 	nv04_render->buffer = (unsigned char*) MALLOC(VERTEX_BUFFER_SIZE);
 	assert(!nv04_render->buffer);
 
+	return nv04_render->buffer ? TRUE : FALSE;
+}
+
+static void *
+nv04_vbuf_render_map_vertices( struct vbuf_render *render )
+{
+	struct nv04_vbuf_render *nv04_render = nv04_vbuf_render(render);
 	return nv04_render->buffer;
 }
 
+static void
+nv04_vbuf_render_unmap_vertices( struct vbuf_render *render,
+		ushort min_index,
+		ushort max_index )
+{
+}
 
 static boolean 
 nv04_vbuf_render_set_primitive( struct vbuf_render *render, 
@@ -244,10 +257,7 @@ nv04_vbuf_render_draw( struct vbuf_render *render,
 
 
 static void
-nv04_vbuf_render_release_vertices( struct vbuf_render *render,
-		void *vertices, 
-		unsigned vertex_size,
-		unsigned vertices_used )
+nv04_vbuf_render_release_vertices( struct vbuf_render *render )
 {
 	struct nv04_vbuf_render *nv04_render = nv04_vbuf_render(render);
 
@@ -278,6 +288,8 @@ nv04_vbuf_render_create( struct nv04_context *nv04 )
 	nv04_render->base.max_indices = 65536; 
 	nv04_render->base.get_vertex_info = nv04_vbuf_render_get_vertex_info;
 	nv04_render->base.allocate_vertices = nv04_vbuf_render_allocate_vertices;
+	nv04_render->base.map_vertices = nv04_vbuf_render_map_vertices;
+	nv04_render->base.unmap_vertices = nv04_vbuf_render_unmap_vertices;
 	nv04_render->base.set_primitive = nv04_vbuf_render_set_primitive;
 	nv04_render->base.draw = nv04_vbuf_render_draw;
 	nv04_render->base.release_vertices = nv04_vbuf_render_release_vertices;
diff --git a/src/gallium/drivers/nv10/nv10_prim_vbuf.c b/src/gallium/drivers/nv10/nv10_prim_vbuf.c
index 7435d87315..491a881806 100644
--- a/src/gallium/drivers/nv10/nv10_prim_vbuf.c
+++ b/src/gallium/drivers/nv10/nv10_prim_vbuf.c
@@ -38,7 +38,7 @@
  */
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_inlines.h"
 #include "pipe/internal/p_winsys_screen.h"
 
@@ -99,8 +99,7 @@ nv10_vbuf_render_get_vertex_info( struct vbuf_render *render )
 	return &nv10->vertex_info;
 }
 
-
-static void *
+static boolean
 nv10_vbuf_render_allocate_vertices( struct vbuf_render *render,
 		ushort vertex_size,
 		ushort nr_vertices )
@@ -115,11 +114,35 @@ nv10_vbuf_render_allocate_vertices( struct vbuf_render *render,
 
 	nv10->dirty |= NV10_NEW_VTXARRAYS;
 
+	if (nv10_render->buffer)
+		return FALSE;
+	return TRUE;
+}
+
+static void *
+nv10_vbuf_render_map_vertices( struct vbuf_render *render )
+{
+	struct nv10_vbuf_render *nv10_render = nv10_vbuf_render(render);
+	struct nv10_context *nv10 = nv10_render->nv10;
+	struct pipe_winsys *winsys = nv10->pipe.winsys;
+
 	return winsys->buffer_map(winsys, 
 			nv10_render->buffer, 
 			PIPE_BUFFER_USAGE_CPU_WRITE);
 }
 
+static void
+nv10_vbuf_render_unmap_vertices( struct vbuf_render *render,
+		ushort min_index,
+		ushort max_index )
+{
+	struct nv10_vbuf_render *nv10_render = nv10_vbuf_render(render);
+	struct nv10_context *nv10 = nv10_render->nv10;
+	struct pipe_winsys *winsys = nv10->pipe.winsys;
+
+	assert(!nv10_render->buffer);
+	winsys->buffer_unmap(winsys, nv10_render->buffer);
+}
 
 static boolean
 nv10_vbuf_render_set_primitive( struct vbuf_render *render, 
@@ -176,18 +199,13 @@ nv10_vbuf_render_draw( struct vbuf_render *render,
 
 
 static void
-nv10_vbuf_render_release_vertices( struct vbuf_render *render,
-		void *vertices, 
-		unsigned vertex_size,
-		unsigned vertices_used )
+nv10_vbuf_render_release_vertices( struct vbuf_render *render )
 {
 	struct nv10_vbuf_render *nv10_render = nv10_vbuf_render(render);
 	struct nv10_context *nv10 = nv10_render->nv10;
-	struct pipe_winsys *winsys = nv10->pipe.winsys;
 	struct pipe_screen *pscreen = &nv10->screen->pipe;
 
 	assert(nv10_render->buffer);
-	winsys->buffer_unmap(winsys, nv10_render->buffer);
 	pipe_buffer_reference(pscreen, &nv10_render->buffer, NULL);
 }
 
@@ -214,6 +232,8 @@ nv10_vbuf_render_create( struct nv10_context *nv10 )
 	nv10_render->base.max_indices = 1024;
 	nv10_render->base.get_vertex_info = nv10_vbuf_render_get_vertex_info;
 	nv10_render->base.allocate_vertices = nv10_vbuf_render_allocate_vertices;
+	nv10_render->base.map_vertices = nv10_vbuf_render_map_vertices;
+	nv10_render->base.unmap_vertices = nv10_vbuf_render_unmap_vertices;
 	nv10_render->base.set_primitive = nv10_vbuf_render_set_primitive;
 	nv10_render->base.draw = nv10_vbuf_render_draw;
 	nv10_render->base.release_vertices = nv10_vbuf_render_release_vertices;
diff --git a/src/gallium/drivers/nv20/nv20_prim_vbuf.c b/src/gallium/drivers/nv20/nv20_prim_vbuf.c
index 4dd7052814..319e1f6557 100644
--- a/src/gallium/drivers/nv20/nv20_prim_vbuf.c
+++ b/src/gallium/drivers/nv20/nv20_prim_vbuf.c
@@ -38,7 +38,7 @@
  */
 
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_inlines.h"
 #include "pipe/internal/p_winsys_screen.h"
 
@@ -109,18 +109,15 @@ nv20__allocate_mbuffer(struct nv20_vbuf_render *nv20_render, size_t size)
 	return nv20_render->mbuffer;
 }
 
-static void *
+static void
 nv20__allocate_pbuffer(struct nv20_vbuf_render *nv20_render, size_t size)
 {
 	struct pipe_winsys *winsys = nv20_render->nv20->pipe.winsys;
 	nv20_render->pbuffer = winsys->buffer_create(winsys, 64,
 					PIPE_BUFFER_USAGE_VERTEX, size);
-	return winsys->buffer_map(winsys,
-			nv20_render->pbuffer,
-			PIPE_BUFFER_USAGE_CPU_WRITE);
 }
 
-static void *
+static boolean
 nv20_vbuf_render_allocate_vertices( struct vbuf_render *render,
 		ushort vertex_size,
 		ushort nr_vertices )
@@ -137,15 +134,49 @@ nv20_vbuf_render_allocate_vertices( struct vbuf_render *render,
 	 * buffer, the data will be passed directly via the fifo.
 	 */
 	/* XXX: Pipe vertex buffers don't work. */
-	if (0 && size > 16 * 1024)
-		buf = nv20__allocate_pbuffer(nv20_render, size);
-	else
+	if (0 && size > 16 * 1024) {
+		nv20__allocate_pbuffer(nv20_render, size);
+		/* umm yeah so this is ugly */
+		buf = nv20_render->pbuffer;
+	} else {
 		buf = nv20__allocate_mbuffer(nv20_render, size);
+	}
 
 	if (buf)
 		nv20_render->nv20->dirty |= NV20_NEW_VTXARRAYS;
 
-	return buf;
+	return buf ? TRUE : FALSE;
+}
+
+static void *
+nv20_vbuf_render_map_vertices( struct vbuf_render *render )
+{
+	struct nv20_vbuf_render *nv20_render = nv20_vbuf_render(render);
+	struct pipe_winsys *winsys = nv20_render->nv20->pipe.winsys;
+
+	if (nv20_render->pbuffer) {
+		return winsys->buffer_map(winsys,
+				nv20_render->pbuffer,
+				PIPE_BUFFER_USAGE_CPU_WRITE);
+	} else if (nv20_render->mbuffer) {
+		return nv20_render->mbuffer;
+	} else
+		assert(0);
+
+	/* warnings be gone */
+	return NULL;
+}
+
+static void
+nv20_vbuf_render_unmap_vertices( struct vbuf_render *render,
+		ushort min_index,
+		ushort max_index )
+{
+	struct nv20_vbuf_render *nv20_render = nv20_vbuf_render(render);
+	struct pipe_winsys *winsys = nv20_render->nv20->pipe.winsys;
+
+	if (nv20_render->pbuffer)
+		winsys->buffer_unmap(winsys, nv20_render->pbuffer);
 }
 
 static boolean
@@ -323,18 +354,13 @@ nv20_vbuf_render_draw( struct vbuf_render *render,
 
 
 static void
-nv20_vbuf_render_release_vertices( struct vbuf_render *render,
-		void *vertices, 
-		unsigned vertex_size,
-		unsigned vertices_used )
+nv20_vbuf_render_release_vertices( struct vbuf_render *render )
 {
 	struct nv20_vbuf_render *nv20_render = nv20_vbuf_render(render);
 	struct nv20_context *nv20 = nv20_render->nv20;
-	struct pipe_winsys *winsys = nv20->pipe.winsys;
 	struct pipe_screen *pscreen = &nv20->screen->pipe;
 
 	if (nv20_render->pbuffer) {
-		winsys->buffer_unmap(winsys, nv20_render->pbuffer);
 		pipe_buffer_reference(pscreen, &nv20_render->pbuffer, NULL);
 	} else if (nv20_render->mbuffer) {
 		FREE(nv20_render->mbuffer);
@@ -371,6 +397,8 @@ nv20_vbuf_render_create( struct nv20_context *nv20 )
 	nv20_render->base.get_vertex_info = nv20_vbuf_render_get_vertex_info;
 	nv20_render->base.allocate_vertices =
 					nv20_vbuf_render_allocate_vertices;
+	nv20_render->base.map_vertices = nv20_vbuf_render_map_vertices;
+	nv20_render->base.unmap_vertices = nv20_vbuf_render_unmap_vertices;
 	nv20_render->base.set_primitive = nv20_vbuf_render_set_primitive;
 	nv20_render->base.draw = nv20_vbuf_render_draw;
 	nv20_render->base.release_vertices = nv20_vbuf_render_release_vertices;
diff --git a/src/gallium/drivers/r300/r300_chipset.c b/src/gallium/drivers/r300/r300_chipset.c
index 794fa2b9b8..196537a432 100644
--- a/src/gallium/drivers/r300/r300_chipset.c
+++ b/src/gallium/drivers/r300/r300_chipset.c
@@ -21,7 +21,7 @@
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
 #include "r300_chipset.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 /* r300_chipset: A file all to itself for deducing the various properties of
  * Radeons. */
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index a2819294a4..0c86a9c92d 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -230,18 +230,23 @@ void r300_emit_dirty_state(struct r300_context* r300)
         return;
     }
 
+    r300_update_derived_state(r300);
+
     /* XXX check size */
 
     if (r300->dirty_state & R300_NEW_BLEND) {
         r300_emit_blend_state(r300, r300->blend_state);
+        r300->dirty_state &= ~R300_NEW_BLEND;
     }
 
     if (r300->dirty_state & R300_NEW_BLEND_COLOR) {
         r300_emit_blend_color_state(r300, r300->blend_color_state);
+        r300->dirty_state &= ~R300_NEW_BLEND_COLOR;
     }
 
     if (r300->dirty_state & R300_NEW_DSA) {
         r300_emit_dsa_state(r300, r300->dsa_state);
+        r300->dirty_state &= ~R300_NEW_DSA;
     }
 
     if (r300->dirty_state & R300_NEW_FRAGMENT_SHADER) {
@@ -252,15 +257,16 @@ void r300_emit_dirty_state(struct r300_context* r300)
             r300_emit_fragment_shader(r300,
                 (struct r300_fragment_shader*)r300->fs);
         }
+        r300->dirty_state &= ~R300_NEW_FRAGMENT_SHADER;
     }
 
     if (r300->dirty_state & R300_NEW_RASTERIZER) {
         r300_emit_rs_state(r300, r300->rs_state);
+        r300->dirty_state &= ~R300_NEW_RASTERIZER;
     }
 
     if (r300->dirty_state & R300_NEW_SCISSOR) {
         r300_emit_scissor_state(r300, r300->scissor_state);
+        r300->dirty_state &= ~R300_NEW_SCISSOR;
     }
-
-    r300->dirty_state = 0;
 }
diff --git a/src/gallium/drivers/r300/r300_reg.h b/src/gallium/drivers/r300/r300_reg.h
index 468e0a2e44..be26b13b0b 100644
--- a/src/gallium/drivers/r300/r300_reg.h
+++ b/src/gallium/drivers/r300/r300_reg.h
@@ -64,7 +64,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define R300_SE_VPORT_ZSCALE                0x1DA8
 #define R300_SE_VPORT_ZOFFSET               0x1DAC
 
-
+#define R300_VAP_PORT_IDX0		    0x2040
 /*
  * Vertex Array Processing (VAP) Control
  */
@@ -732,8 +732,8 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define R500_RS_IP_TEX_PTR_Q_SHIFT 			18
 #define R500_RS_IP_COL_PTR_SHIFT 			24
 #define R500_RS_IP_COL_FMT_SHIFT 			27
-#	define R500_RS_COL_PTR(x)		        (x << 24)
-#       define R500_RS_COL_FMT(x)                       (x << 27)
+#	define R500_RS_COL_PTR(x)		        ((x) << 24)
+#       define R500_RS_COL_FMT(x)                       ((x) << 27)
 /* gap */
 #define R500_RS_IP_OFFSET_DIS 				(0 << 31)
 #define R500_RS_IP_OFFSET_EN 				(1 << 31)
@@ -1175,8 +1175,8 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #       define R300_RS_INTERP_SRC_SHIFT          2 /* TODO: check for removal */
 #       define R300_RS_INTERP_SRC_MASK           (7 << 2) /* TODO: check for removal */
 #	define R300_RS_TEX_PTR(x)		        (x << 0)
-#	define R300_RS_COL_PTR(x)		        (x << 6)
-#	define R300_RS_COL_FMT(x)		        (x << 9)
+#	define R300_RS_COL_PTR(x)		        ((x) << 6)
+#	define R300_RS_COL_FMT(x)		        ((x) << 9)
 #	define R300_RS_COL_FMT_RGBA		        0
 #	define R300_RS_COL_FMT_RGB0		        1
 #	define R300_RS_COL_FMT_RGB1		        2
@@ -1186,10 +1186,10 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #	define R300_RS_COL_FMT_111A		        8
 #	define R300_RS_COL_FMT_1110		        9
 #	define R300_RS_COL_FMT_1111		        10
-#	define R300_RS_SEL_S(x)		                (x << 13)
-#	define R300_RS_SEL_T(x)		                (x << 16)
-#	define R300_RS_SEL_R(x)		                (x << 19)
-#	define R300_RS_SEL_Q(x)		                (x << 22)
+#	define R300_RS_SEL_S(x)		                ((x) << 13)
+#	define R300_RS_SEL_T(x)		                ((x) << 16)
+#	define R300_RS_SEL_R(x)		                ((x) << 19)
+#	define R300_RS_SEL_Q(x)		                ((x) << 22)
 #	define R300_RS_SEL_C0		                0
 #	define R300_RS_SEL_C1		                1
 #	define R300_RS_SEL_C2		                2
@@ -1708,7 +1708,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #   define R300_C3_SEL_R				(1 << 14)
 #   define R300_C3_SEL_G				(2 << 14)
 #   define R300_C3_SEL_B				(3 << 14)
-#   define R300_OUT_SIGN(x)				(x << 16)
+#   define R300_OUT_SIGN(x)				((x) << 16)
 #   define R500_ROUND_ADJ				(1 << 20)
 
 /* ALU
@@ -2732,7 +2732,7 @@ enum {
 #   define R500_ALPHA_OP_COS				13
 #   define R500_ALPHA_OP_MDH				14
 #   define R500_ALPHA_OP_MDV				15
-#   define R500_ALPHA_ADDRD(x)				(x << 4)
+#   define R500_ALPHA_ADDRD(x)				((x) << 4)
 #   define R500_ALPHA_ADDRD_REL				(1 << 11)
 #  define R500_ALPHA_SEL_A_SHIFT			12
 #   define R500_ALPHA_SEL_A_SRC0			(0 << 12)
@@ -2776,16 +2776,16 @@ enum {
 #   define R500_ALPHA_OMOD_DIV_4			(5 << 26)
 #   define R500_ALPHA_OMOD_DIV_8			(6 << 26)
 #   define R500_ALPHA_OMOD_DISABLE			(7 << 26)
-#   define R500_ALPHA_TARGET(x)				(x << 29)
+#   define R500_ALPHA_TARGET(x)				((x) << 29)
 #   define R500_ALPHA_W_OMASK				(1 << 31)
 #define R500_US_ALU_ALPHA_ADDR_0			0x9800
-#   define R500_ALPHA_ADDR0(x)				(x << 0)
+#   define R500_ALPHA_ADDR0(x)				((x) << 0)
 #   define R500_ALPHA_ADDR0_CONST			(1 << 8)
 #   define R500_ALPHA_ADDR0_REL				(1 << 9)
-#   define R500_ALPHA_ADDR1(x)				(x << 10)
+#   define R500_ALPHA_ADDR1(x)				((x) << 10)
 #   define R500_ALPHA_ADDR1_CONST			(1 << 18)
 #   define R500_ALPHA_ADDR1_REL				(1 << 19)
-#   define R500_ALPHA_ADDR2(x)				(x << 20)
+#   define R500_ALPHA_ADDR2(x)				((x) << 20)
 #   define R500_ALPHA_ADDR2_CONST			(1 << 28)
 #   define R500_ALPHA_ADDR2_REL				(1 << 29)
 #   define R500_ALPHA_SRCP_OP_1_MINUS_2A0		(0 << 30)
@@ -2806,7 +2806,7 @@ enum {
 #   define R500_ALU_RGBA_OP_SOP				(10 << 0)
 #   define R500_ALU_RGBA_OP_MDH				(11 << 0)
 #   define R500_ALU_RGBA_OP_MDV				(12 << 0)
-#   define R500_ALU_RGBA_ADDRD(x)			(x << 4)
+#   define R500_ALU_RGBA_ADDRD(x)			((x) << 4)
 #   define R500_ALU_RGBA_ADDRD_REL			(1 << 11)
 #  define R500_ALU_RGBA_SEL_C_SHIFT			12
 #   define R500_ALU_RGBA_SEL_C_SRC0			(0 << 12)
@@ -2933,16 +2933,16 @@ enum {
 #   define R500_ALU_RGB_OMOD_DIV_4			(5 << 26)
 #   define R500_ALU_RGB_OMOD_DIV_8			(6 << 26)
 #   define R500_ALU_RGB_OMOD_DISABLE			(7 << 26)
-#   define R500_ALU_RGB_TARGET(x)			(x << 29)
+#   define R500_ALU_RGB_TARGET(x)			((x) << 29)
 #   define R500_ALU_RGB_WMASK				(1 << 31)
 #define R500_US_ALU_RGB_ADDR_0				0x9000
-#   define R500_RGB_ADDR0(x)				(x << 0)
+#   define R500_RGB_ADDR0(x)				((x) << 0)
 #   define R500_RGB_ADDR0_CONST				(1 << 8)
 #   define R500_RGB_ADDR0_REL				(1 << 9)
-#   define R500_RGB_ADDR1(x)				(x << 10)
+#   define R500_RGB_ADDR1(x)				((x) << 10)
 #   define R500_RGB_ADDR1_CONST				(1 << 18)
 #   define R500_RGB_ADDR1_REL				(1 << 19)
-#   define R500_RGB_ADDR2(x)				(x << 20)
+#   define R500_RGB_ADDR2(x)				((x) << 20)
 #   define R500_RGB_ADDR2_CONST				(1 << 28)
 #   define R500_RGB_ADDR2_REL				(1 << 29)
 #   define R500_RGB_SRCP_OP_1_MINUS_2RGB0		(0 << 30)
@@ -2998,19 +2998,19 @@ enum {
 
 /* note that these are 8 bit lengths, despite the offsets, at least for R500 */
 #define R500_US_CODE_ADDR				0x4630
-#   define R500_US_CODE_START_ADDR(x)			(x << 0)
-#   define R500_US_CODE_END_ADDR(x)			(x << 16)
+#   define R500_US_CODE_START_ADDR(x)			((x) << 0)
+#   define R500_US_CODE_END_ADDR(x)			((x) << 16)
 #define R500_US_CODE_OFFSET				0x4638
-#   define R500_US_CODE_OFFSET_ADDR(x)			(x << 0)
+#   define R500_US_CODE_OFFSET_ADDR(x)			((x) << 0)
 #define R500_US_CODE_RANGE				0x4634
-#   define R500_US_CODE_RANGE_ADDR(x)			(x << 0)
-#   define R500_US_CODE_RANGE_SIZE(x)			(x << 16)
+#   define R500_US_CODE_RANGE_ADDR(x)			((x) << 0)
+#   define R500_US_CODE_RANGE_SIZE(x)			((x) << 16)
 #define R500_US_CONFIG					0x4600
 #   define R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO		(1 << 1)
 #define R500_US_FC_ADDR_0				0xa000
-#   define R500_FC_BOOL_ADDR(x)				(x << 0)
-#   define R500_FC_INT_ADDR(x)				(x << 8)
-#   define R500_FC_JUMP_ADDR(x)				(x << 16)
+#   define R500_FC_BOOL_ADDR(x)				((x) << 0)
+#   define R500_FC_INT_ADDR(x)				((x) << 8)
+#   define R500_FC_JUMP_ADDR(x)				((x) << 16)
 #   define R500_FC_JUMP_GLOBAL				(1 << 31)
 #define R500_US_FC_BOOL_CONST				0x4620
 #   define R500_FC_KBOOL(x)				(x)
@@ -3031,8 +3031,8 @@ enum {
 #   define R500_FC_A_OP_NONE				(0 << 6)
 #   define R500_FC_A_OP_POP				(1 << 6)
 #   define R500_FC_A_OP_PUSH				(2 << 6)
-#   define R500_FC_JUMP_FUNC(x)				(x << 8)
-#   define R500_FC_B_POP_CNT(x)				(x << 16)
+#   define R500_FC_JUMP_FUNC(x)				((x) << 8)
+#   define R500_FC_B_POP_CNT(x)				((x) << 16)
 #   define R500_FC_B_OP0_NONE				(0 << 24)
 #   define R500_FC_B_OP0_DECR				(1 << 24)
 #   define R500_FC_B_OP0_INCR				(2 << 24)
@@ -3041,18 +3041,18 @@ enum {
 #   define R500_FC_B_OP1_INCR				(2 << 26)
 #   define R500_FC_IGNORE_UNCOVERED			(1 << 28)
 #define R500_US_FC_INT_CONST_0				0x4c00
-#   define R500_FC_INT_CONST_KR(x)			(x << 0)
-#   define R500_FC_INT_CONST_KG(x)			(x << 8)
-#   define R500_FC_INT_CONST_KB(x)			(x << 16)
+#   define R500_FC_INT_CONST_KR(x)			((x) << 0)
+#   define R500_FC_INT_CONST_KG(x)			((x) << 8)
+#   define R500_FC_INT_CONST_KB(x)			((x) << 16)
 /* _0 through _15 */
 #define R500_US_FORMAT0_0				0x4640
-#   define R500_FORMAT_TXWIDTH(x)			(x << 0)
-#   define R500_FORMAT_TXHEIGHT(x)			(x << 11)
-#   define R500_FORMAT_TXDEPTH(x)			(x << 22)
+#   define R500_FORMAT_TXWIDTH(x)			((x) << 0)
+#   define R500_FORMAT_TXHEIGHT(x)			((x) << 11)
+#   define R500_FORMAT_TXDEPTH(x)			((x) << 22)
 #define R500_US_PIXSIZE					0x4604
 #   define R500_PIX_SIZE(x)				(x)
 #define R500_US_TEX_ADDR_0				0x9800
-#   define R500_TEX_SRC_ADDR(x)				(x << 0)
+#   define R500_TEX_SRC_ADDR(x)				((x) << 0)
 #   define R500_TEX_SRC_ADDR_REL			(1 << 7)
 #   define R500_TEX_SRC_S_SWIZ_R			(0 << 8)
 #   define R500_TEX_SRC_S_SWIZ_G			(1 << 8)
@@ -3070,7 +3070,7 @@ enum {
 #   define R500_TEX_SRC_Q_SWIZ_G			(1 << 14)
 #   define R500_TEX_SRC_Q_SWIZ_B			(2 << 14)
 #   define R500_TEX_SRC_Q_SWIZ_A			(3 << 14)
-#   define R500_TEX_DST_ADDR(x)				(x << 16)
+#   define R500_TEX_DST_ADDR(x)				((x) << 16)
 #   define R500_TEX_DST_ADDR_REL			(1 << 23)
 #   define R500_TEX_DST_R_SWIZ_R			(0 << 24)
 #   define R500_TEX_DST_R_SWIZ_G			(1 << 24)
@@ -3089,7 +3089,7 @@ enum {
 #   define R500_TEX_DST_A_SWIZ_B			(2 << 30)
 #   define R500_TEX_DST_A_SWIZ_A			(3 << 30)
 #define R500_US_TEX_ADDR_DXDY_0				0xa000
-#   define R500_DX_ADDR(x)				(x << 0)
+#   define R500_DX_ADDR(x)				((x) << 0)
 #   define R500_DX_ADDR_REL				(1 << 7)
 #   define R500_DX_S_SWIZ_R				(0 << 8)
 #   define R500_DX_S_SWIZ_G				(1 << 8)
@@ -3107,7 +3107,7 @@ enum {
 #   define R500_DX_Q_SWIZ_G				(1 << 14)
 #   define R500_DX_Q_SWIZ_B				(2 << 14)
 #   define R500_DX_Q_SWIZ_A				(3 << 14)
-#   define R500_DY_ADDR(x)				(x << 16)
+#   define R500_DY_ADDR(x)				((x) << 16)
 #   define R500_DY_ADDR_REL				(1 << 17)
 #   define R500_DY_S_SWIZ_R				(0 << 24)
 #   define R500_DY_S_SWIZ_G				(1 << 24)
@@ -3126,7 +3126,7 @@ enum {
 #   define R500_DY_Q_SWIZ_B				(2 << 30)
 #   define R500_DY_Q_SWIZ_A				(3 << 30)
 #define R500_US_TEX_INST_0				0x9000
-#   define R500_TEX_ID(x)				(x << 16)
+#   define R500_TEX_ID(x)				((x) << 16)
 #   define R500_TEX_INST_NOP				(0 << 22)
 #   define R500_TEX_INST_LD				(1 << 22)
 #   define R500_TEX_INST_TEXKILL			(2 << 22)
@@ -3187,9 +3187,9 @@ enum {
 #define R300_PACKET3_3D_LOAD_VBPNTR         0x00002F00
 
 #define R300_PACKET3_INDX_BUFFER            0x00003300
-#    define R300_EB_UNK1_SHIFT                      24
-#    define R300_EB_UNK1                    (0x80<<24)
-#    define R300_EB_UNK2                        0x0810
+#    define R300_INDX_BUFFER_DST_SHIFT          0
+#    define R300_INDX_BUFFER_SKIP_SHIFT         16
+#    define R300_INDX_BUFFER_ONE_REG_WR		(1<<31)
 
 /* Same as R300_PACKET3_3D_DRAW_VBUF but without VAP_VTX_FMT */
 #define R300_PACKET3_3D_DRAW_VBUF_2         0x00003400
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index eae1a5698d..da99a3be6b 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -23,7 +23,7 @@
 #include "util/u_math.h"
 #include "util/u_pack_color.h"
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/internal/p_winsys_screen.h"
 
 #include "r300_context.h"
@@ -435,6 +435,8 @@ static void* r300_create_fs_state(struct pipe_context* pipe,
     /* Copy state directly into shader. */
     fs->state = *shader;
 
+    tgsi_scan_shader(shader->tokens, &fs->info);
+
     return (void*)fs;
 }
 
diff --git a/src/gallium/drivers/r300/r300_swtcl_emit.c b/src/gallium/drivers/r300/r300_swtcl_emit.c
index 76ef48962b..fb009247a5 100644
--- a/src/gallium/drivers/r300/r300_swtcl_emit.c
+++ b/src/gallium/drivers/r300/r300_swtcl_emit.c
@@ -21,144 +21,295 @@
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
 #include "draw/draw_pipe.h"
+#include "draw/draw_vbuf.h"
 #include "util/u_memory.h"
 
 #include "r300_cs.h"
 #include "r300_context.h"
 #include "r300_reg.h"
 
-/* r300_swtcl_emit: Primitive vertex emission using an immediate
- * vertex buffer and no HW TCL. */
+/* r300_swtcl_emit: Vertex and index buffer primitive emission. No HW TCL. */
 
-struct swtcl_stage {
+struct r300_swtcl_render {
     /* Parent class */
-    struct draw_stage draw;
-
+    struct vbuf_render base;
+    
+    /* Pipe context */
     struct r300_context* r300;
+
+    /* Vertex information */
+    size_t vertex_size;
+    unsigned prim;
+    unsigned hwprim;
+
+    /* VBO */
+    struct pipe_buffer* vbo;
+    size_t vbo_size;
+    size_t vbo_offset;
+    void* vbo_map;
+    size_t vbo_alloc_size;
+    size_t vbo_max_used;
 };
 
-static INLINE struct swtcl_stage* swtcl_stage(struct draw_stage* draw) {
-    return (struct swtcl_stage*)draw;
+static INLINE struct r300_swtcl_render*
+r300_swtcl_render(struct vbuf_render* render)
+{
+    return (struct r300_swtcl_render*)render;
 }
 
-static INLINE void r300_emit_vertex(struct r300_context* r300,
-                                    const struct vertex_header* vertex)
+static const struct vertex_info*
+r300_swtcl_render_get_vertex_info(struct vbuf_render* render)
 {
-    struct vertex_info* vinfo = &r300->vertex_info;
-    CS_LOCALS(r300);
-    uint i, j;
-
-    for (i = 0; i < vinfo->num_attribs; i++) {
-        j = vinfo->attrib[i].src_index;
-        switch (vinfo->attrib[i].emit) {
-            case EMIT_1F:
-                OUT_CS_32F(vertex->data[j][0]);
-                break;
-            case EMIT_2F:
-                OUT_CS_32F(vertex->data[j][0]);
-                OUT_CS_32F(vertex->data[j][1]);
-                break;
-            case EMIT_3F:
-                OUT_CS_32F(vertex->data[j][0]);
-                OUT_CS_32F(vertex->data[j][1]);
-                OUT_CS_32F(vertex->data[j][2]);
-                break;
-            case EMIT_4F:
-                OUT_CS_32F(vertex->data[j][0]);
-                OUT_CS_32F(vertex->data[j][1]);
-                OUT_CS_32F(vertex->data[j][2]);
-                OUT_CS_32F(vertex->data[j][3]);
-                break;
-            default:
-                debug_printf("r300: Unknown emit value %d\n",
-                    vinfo->attrib[i].emit);
-                break;
-        }
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct r300_context* r300 = r300render->r300;
+
+    r300_update_derived_state(r300);
+
+    return &r300->vertex_info;
+}
+
+static boolean r300_swtcl_render_allocate_vertices(struct vbuf_render* render,
+                                                   ushort vertex_size,
+                                                   ushort count)
+{
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct r300_context* r300 = r300render->r300;
+    struct pipe_screen* screen = r300->context.screen;
+    size_t size = (size_t)vertex_size * (size_t)count;
+
+    if (r300render->vbo) {
+        pipe_buffer_reference(screen, r300render->vbo, NULL);
+    }
+
+    r300render->vbo_size = MAX2(size, r300render->vbo_alloc_size);
+    r300render->vbo_offset = 0;
+    r300render->vbo = pipe_buffer_create(screen,
+                                         64,
+                                         PIPE_BUFFER_USAGE_VERTEX,
+                                         r300render->vbo_size);
+
+    r300render->vertex_size = vertex_size;
+
+    if (r300render->vbo) {
+        return true;
+    } else {
+        return false;
     }
 }
 
-static INLINE void r300_emit_prim(struct draw_stage* draw,
-                                  struct prim_header* prim,
-                                  unsigned hwprim,
-                                  unsigned count)
+static void* r300_swtcl_render_map_vertices(struct vbuf_render* render)
+{
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct pipe_screen* screen = r300render->r300->context.screen;
+
+    r300render->vbo_map = pipe_buffer_map(screen, r300render->vbo,
+                                          PIPE_BUFFER_USAGE_CPU_WRITE);
+
+    return (unsigned char*)r300render->vbo_map + r300render->vbo_offset;
+}
+
+static void* r300_swtcl_render_unmap_vertices(struct vbuf_render* render,
+                                              ushort min,
+                                              ushort max)
 {
-    struct r300_context* r300 = swtcl_stage(draw)->r300;
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct pipe_screen* screen = r300render->r300->context.screen;
+
+    r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
+             r300render->vertex_size * (max + 1));
+
+    pipe_buffer_unmap(screen, r300render->vbo);
+}
+
+static void r300_swtcl_render_release_vertices(struct vbuf_render* render)
+{
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct pipe_screen* screen = r300render->r300->context.screen;
+
+    pipe_buffer_reference(screen, r300render->vbo, NULL);
+}
+
+static boolean r300_swtcl_render_set_primitive(struct vbuf_render* render,
+                                               unsigned prim)
+{
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    r300render->prim = prim;
+
+    switch (prim) {
+        case PIPE_PRIM_POINTS:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_POINTS;
+            break;
+        case PIPE_PRIM_LINES:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_LINES;
+            break;
+        case PIPE_PRIM_LINE_LOOP:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_LINE_LOOP;
+            break;
+        case PIPE_PRIM_LINE_STRIP:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_LINE_STRIP;
+            break;
+        case PIPE_PRIM_TRIANGLES:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_TRIANGLES;
+            break;
+        case PIPE_PRIM_TRIANGLE_STRIP:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_TRIANGLE_STRIP;
+            break;
+        case PIPE_PRIM_TRIANGLE_FAN:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_TRIANGLE_FAN;
+            break;
+        case PIPE_PRIM_QUADS:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_QUADS;
+            break;
+        case PIPE_PRIM_QUAD_STRIP:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_QUAD_STRIP;
+            break;
+        case PIPE_PRIM_POLYGON:
+            r300render->hwprim = R300_VAP_VF_CNTL__PRIM_POLYGON;
+            break;
+        default:
+            return false;
+            break;
+    }
+
+    return true;
+}
+
+static void r300_swtcl_render_draw_arrays(struct vbuf_render* render,
+                                          unsigned start,
+                                          unsigned count)
+{
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct r300_context* r300 = r300render->r300;
+    struct pipe_screen* screen = r300->context.screen;
+
     CS_LOCALS(r300);
-    int i;
 
+    /* Make sure that all possible state is emitted. */
+    r300_update_derived_state(r300);
     r300_emit_dirty_state(r300);
 
+    /* Take care of vertex formats and routes */
     BEGIN_CS(3);
     OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
     OUT_CS(r300->vertex_info.hwfmt[0]);
     OUT_CS(r300->vertex_info.hwfmt[1]);
     END_CS;
 
-    BEGIN_CS(2 + (count * r300->vertex_info.size) + 2);
-    OUT_CS(CP_PACKET3(R200_3D_DRAW_IMMD_2, count));
-    OUT_CS(hwprim | R300_PRIM_WALK_RING |
-        (count << R300_PRIM_NUM_VERTICES_SHIFT));
+    /* Draw stuff! */
+    BEGIN_CS(2);
+    OUT_CS(CP_PACKET3(R300_PACKET3_3D_DRAW_VBUF_2, 0));
+    OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
+           r300render->hwprim | R300_VAP_VF_CNTL__INDEX_SIZE_32bit);
+}
+
+static void r300_swtcl_render_draw(struct vbuf_render* render,
+                                   const ushort* indices,
+                                   uint count)
+{
+    struct r300_swtcl_render* r300render = r300_swtcl_render(render);
+    struct r300_context* r300 = r300render->r300;
+    struct pipe_screen* screen = r300->context.screen;
+    struct pipe_buffer* index_buffer;
+    void* index_map;
+
+    CS_LOCALS(r300);
+
+    /* Make sure that all possible state is emitted. */
+    r300_update_derived_state(r300);
+    r300_emit_dirty_state(r300);
 
-    for (i = 0; i < count; i++) {
-        r300_emit_vertex(r300, prim->v[i]);
+    /* Send our indices into an index buffer. */
+    index_buffer = pipe_buffer_create(screen, 64, PIPE_BUFFER_USAGE_VERTEX,
+                                      count * 4);
+    if (!index_buffer) {
+        return;
     }
 
+    index_map = pipe_buffer_map(screen, index_buffer,
+                                PIPE_BUFFER_USAGE_CPU_WRITE);
+    memcpy(index_map, indices, count * 4);
+    pipe_buffer_unmap(screen, index_buffer);
+
+    /* Take care of vertex formats and routes */
+    BEGIN_CS(3);
+    OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
+    OUT_CS(r300->vertex_info.hwfmt[0]);
+    OUT_CS(r300->vertex_info.hwfmt[1]);
     END_CS;
-}
 
-/* Just as an aside...
- *
- * Radeons can do many more primitives:
- * - Line strip
- * - Triangle fan
- * - Triangle strip
- * - Line loop
- * - Quads
- * - Quad strip
- * - Polygons
- *
- * The following were just the only ones in Draw. */
+    /* Draw stuff! */
+    BEGIN_CS(10);
+    OUT_CS(CP_PACKET3(R300_PACKET3_3D_DRAW_INDX_2, 0));
+    OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
+           r300render->hwprim | R300_VAP_VF_CNTL__INDEX_SIZE_32bit);
 
-static void r300_emit_point(struct draw_stage* draw, struct prim_header* prim)
-{
-    r300_emit_prim(draw, prim, R300_PRIM_TYPE_POINT, 1);
+    OUT_CS(CP_PACKET3(R300_PACKET3_INDX_BUFFER, 2));
+    OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2));
+    OUT_CS_RELOC(index_buffer, 0, RADEON_GEM_DOMAIN_GTT, 0, 0);
+    END_CS;
 }
 
-static void r300_emit_line(struct draw_stage* draw, struct prim_header* prim)
+static void r300_swtcl_render_destroy(struct vbuf_render* render)
 {
-    r300_emit_prim(draw, prim, R300_PRIM_TYPE_LINE, 2);
+    FREE(render);
 }
 
-static void r300_emit_tri(struct draw_stage* draw, struct prim_header* prim)
+static struct vbuf_render* r300_swtcl_render_create(struct r300_context* r300)
 {
-    r300_emit_prim(draw, prim, R300_PRIM_TYPE_TRI_LIST, 3);
-}
+    struct r300_swtcl_render* r300render = CALLOC_STRUCT(r300_swtcl_render);
+    struct pipe_screen* screen = r300->context.screen;
 
-static void r300_swtcl_flush(struct draw_stage* draw, unsigned flags)
-{
-}
+    r300render->r300 = r300;
 
-static void r300_reset_stipple(struct draw_stage* draw)
-{
-    /* XXX */
-}
+    /* XXX find real numbers plz */
+    r300render->base.max_vertex_buffer_bytes = 128 * 1024;
+    r300render->base.max_indices = 16 * 1024;
 
-static void r300_swtcl_destroy(struct draw_stage* draw)
-{
-    FREE(draw);
+    r300render->base.get_vertex_info = r300_swtcl_render_get_vertex_info;
+    r300render->base.allocate_vertices = r300_swtcl_render_allocate_vertices;
+    r300render->base.map_vertices = r300_swtcl_render_map_vertices;
+    r300render->base.unmap_vertices = r300_swtcl_render_unmap_vertices;
+    r300render->base.set_primitive = r300_swtcl_render_set_primitive;
+    r300render->base.draw = r300_swtcl_render_draw;
+    r300render->base.draw_arrays = r300_swtcl_render_draw_arrays;
+    r300render->base.release_vertices = r300_swtcl_render_release_vertices;
+    r300render->base.destroy = r300_swtcl_render_destroy;
+
+    /* XXX bonghits ahead
+    r300render->vbo_alloc_size = 128 * 4096;
+    r300render->vbo_size = r300render->vbo_alloc_size;
+    r300render->vbo_offset = 0;
+    r300render->vbo = pipe_buffer_create(screen,
+                                         64,
+                                         PIPE_BUFFER_USAGE_VERTEX,
+                                         r300render->vbo_size);
+    r300render->vbo_map = pipe_buffer_map(screen,
+                                          r300render->vbo,
+                                          PIPE_BUFFER_USAGE_CPU_WRITE);
+    pipe_buffer_unmap(screen, r300render->vbo); */
+
+    return &r300render->base;
 }
 
 struct draw_stage* r300_draw_swtcl_stage(struct r300_context* r300)
 {
-    struct swtcl_stage* swtcl = CALLOC_STRUCT(swtcl_stage);
+    struct vbuf_render* render;
+    struct draw_stage* stage;
+
+    render = r300_swtcl_render_create(r300);
+
+    if (!render) {
+        return NULL;
+    }
+
+    stage = draw_vbuf_stage(r300->draw, render);
+
+    if (!stage) {
+        render->destroy(render);
+        return NULL;
+    }
 
-    swtcl->r300 = r300;
-    swtcl->draw.point = r300_emit_point;
-    swtcl->draw.line = r300_emit_line;
-    swtcl->draw.tri = r300_emit_tri;
-    swtcl->draw.flush = r300_swtcl_flush;
-    swtcl->draw.reset_stipple_counter = r300_reset_stipple;
-    swtcl->draw.destroy = r300_swtcl_destroy;
+    draw_set_render(r300->draw, render);
 
-    return &swtcl->draw;
+    return stage;
 }
diff --git a/src/gallium/drivers/softpipe/sp_prim_vbuf.c b/src/gallium/drivers/softpipe/sp_prim_vbuf.c
index 9cd5784e5b..d56eed80a4 100644
--- a/src/gallium/drivers/softpipe/sp_prim_vbuf.c
+++ b/src/gallium/drivers/softpipe/sp_prim_vbuf.c
@@ -60,6 +60,7 @@ struct softpipe_vbuf_render
    struct softpipe_context *softpipe;
    uint prim;
    uint vertex_size;
+   uint vertex_buffer_size;
    void *vertex_buffer;
 };
 
@@ -80,26 +81,44 @@ sp_vbuf_get_vertex_info(struct vbuf_render *vbr)
 }
 
 
-static void *
+static boolean
 sp_vbuf_allocate_vertices(struct vbuf_render *vbr,
-                            ushort vertex_size, ushort nr_vertices)
+                          ushort vertex_size, ushort nr_vertices)
 {
    struct softpipe_vbuf_render *cvbr = softpipe_vbuf_render(vbr);
-   assert(!cvbr->vertex_buffer);
-   cvbr->vertex_buffer = align_malloc(vertex_size * nr_vertices, 16);
+   unsigned size = vertex_size * nr_vertices;
+
+   if (cvbr->vertex_buffer_size < size) {
+      align_free(cvbr->vertex_buffer);
+      cvbr->vertex_buffer = align_malloc(size, 16);
+      cvbr->vertex_buffer_size = size;
+   }
+
    cvbr->vertex_size = vertex_size;
-   return cvbr->vertex_buffer;
+   return cvbr->vertex_buffer != NULL;
 }
 
-
 static void
-sp_vbuf_release_vertices(struct vbuf_render *vbr, void *vertices,
-                           unsigned vertex_size, unsigned vertices_used)
+sp_vbuf_release_vertices(struct vbuf_render *vbr)
+{
+   /* keep the old allocation for next time */
+}
+
+static void *
+sp_vbuf_map_vertices(struct vbuf_render *vbr)
 {
    struct softpipe_vbuf_render *cvbr = softpipe_vbuf_render(vbr);
-   align_free(vertices);
-   assert(vertices == cvbr->vertex_buffer);
-   cvbr->vertex_buffer = NULL;
+   return cvbr->vertex_buffer;
+}
+
+static void 
+sp_vbuf_unmap_vertices(struct vbuf_render *vbr, 
+                       ushort min_index,
+                       ushort max_index )
+{
+   struct softpipe_vbuf_render *cvbr = softpipe_vbuf_render(vbr);
+   assert( cvbr->vertex_buffer_size >= (max_index+1) * cvbr->vertex_size );
+   /* do nothing */
 }
 
 
@@ -115,8 +134,6 @@ sp_vbuf_set_primitive(struct vbuf_render *vbr, unsigned prim)
    
    setup_prepare( setup_ctx );
 
-
-
    cvbr->prim = prim;
    return TRUE;
 
@@ -394,6 +411,8 @@ sp_init_vbuf(struct softpipe_context *sp)
 
    sp->vbuf_render->base.get_vertex_info = sp_vbuf_get_vertex_info;
    sp->vbuf_render->base.allocate_vertices = sp_vbuf_allocate_vertices;
+   sp->vbuf_render->base.map_vertices = sp_vbuf_map_vertices;
+   sp->vbuf_render->base.unmap_vertices = sp_vbuf_unmap_vertices;
    sp->vbuf_render->base.set_primitive = sp_vbuf_set_primitive;
    sp->vbuf_render->base.draw = sp_vbuf_draw;
    sp->vbuf_render->base.draw_arrays = sp_vbuf_draw_arrays;
diff --git a/src/gallium/drivers/softpipe/sp_tile_cache.c b/src/gallium/drivers/softpipe/sp_tile_cache.c
index ccc518eb61..593360aab0 100644
--- a/src/gallium/drivers/softpipe/sp_tile_cache.c
+++ b/src/gallium/drivers/softpipe/sp_tile_cache.c
@@ -40,7 +40,7 @@
 #include "sp_texture.h"
 #include "sp_tile_cache.h"
 
-#define NUM_ENTRIES 32
+#define NUM_ENTRIES 50
 
 
 /** XXX move these */
@@ -501,7 +501,7 @@ sp_get_cached_tile(struct softpipe_context *softpipe,
 static INLINE uint
 tex_cache_pos(int x, int y, int z, int face, int level)
 {
-   uint entry = x + y * 5 + z * 4 + face + level;
+   uint entry = x + y * 9 + z * 3 + face + level * 7;
    return entry % NUM_ENTRIES;
 }
 
@@ -527,8 +527,12 @@ sp_get_cached_tile_tex(struct softpipe_context *sp,
    if (tc->texture) {
       struct softpipe_texture *spt = softpipe_texture(tc->texture);
       if (spt->modified) {
-         /* texture was modified, force a cache reload */
-         tile->x = -1;
+         /* texture was modified, invalidate all cached tiles */
+         uint p;
+         for (p = 0; p < NUM_ENTRIES; p++) {
+            tile = tc->entries + p;
+            tile->x = -1;
+         }
          spt->modified = FALSE;
       }
    }
@@ -540,6 +544,10 @@ sp_get_cached_tile_tex(struct softpipe_context *sp,
        level != tile->level) {
       /* cache miss */
 
+#if 0
+      printf("miss at %u  x=%d y=%d z=%d face=%d level=%d\n", pos,
+             x/TILE_SIZE, y/TILE_SIZE, z, face, level);
+#endif
       /* check if we need to get a new transfer */
       if (!tc->tex_trans ||
           tc->tex_face != face ||
diff --git a/src/gallium/drivers/trace/tr_context.h b/src/gallium/drivers/trace/tr_context.h
index 7831900ec2..6704175964 100644
--- a/src/gallium/drivers/trace/tr_context.h
+++ b/src/gallium/drivers/trace/tr_context.h
@@ -30,7 +30,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_context.h"
 
 
diff --git a/src/gallium/drivers/trace/tr_dump.c b/src/gallium/drivers/trace/tr_dump.c
index a0ead0ded3..d98cef221b 100644
--- a/src/gallium/drivers/trace/tr_dump.c
+++ b/src/gallium/drivers/trace/tr_dump.c
@@ -45,7 +45,7 @@
 #endif
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 #include "util/u_string.h"
 #include "util/u_stream.h"
diff --git a/src/gallium/drivers/trace/tr_winsys.h b/src/gallium/drivers/trace/tr_winsys.h
index 0fd2a40556..3670cb915e 100644
--- a/src/gallium/drivers/trace/tr_winsys.h
+++ b/src/gallium/drivers/trace/tr_winsys.h
@@ -30,7 +30,7 @@
 
 
 #include "pipe/p_compiler.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/internal/p_winsys_screen.h"
 
 
diff --git a/src/gallium/include/pipe/p_format.h b/src/gallium/include/pipe/p_format.h
index b42f98ceba..3f65a60436 100644
--- a/src/gallium/include/pipe/p_format.h
+++ b/src/gallium/include/pipe/p_format.h
@@ -30,8 +30,9 @@
 #define PIPE_FORMAT_H
 
 #include "p_compiler.h"
-#include "p_debug.h"
 
+/* FIXME: remove these header dependencies */
+#include "util/u_debug.h"
 #include "util/u_string.h"
 
 #ifdef __cplusplus
diff --git a/src/gallium/state_trackers/wgl/icd/stw_icd.c b/src/gallium/state_trackers/wgl/icd/stw_icd.c
index 1aa4b8a6e2..5ac871da81 100644
--- a/src/gallium/state_trackers/wgl/icd/stw_icd.c
+++ b/src/gallium/state_trackers/wgl/icd/stw_icd.c
@@ -30,7 +30,7 @@
 
 #include "GL/gl.h"
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_thread.h"
 
 #include "shared/stw_public.h"
diff --git a/src/gallium/state_trackers/wgl/shared/stw_device.c b/src/gallium/state_trackers/wgl/shared/stw_device.c
index 6873e813ee..4b3cf51a53 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_device.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_device.c
@@ -27,7 +27,7 @@
 
 #include <windows.h>
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "pipe/p_screen.h"
 
 #include "shared/stw_device.h"
@@ -66,6 +66,10 @@ stw_shared_init(const struct stw_winsys *stw_winsys)
    stw_dev = &stw_dev_storage;
    memset(stw_dev, 0, sizeof(*stw_dev));
 
+#ifdef DEBUG
+   stw_dev->memdbg_no = debug_memory_begin();
+#endif
+   
    stw_dev->stw_winsys = stw_winsys;
 
    stw_dev->screen = stw_winsys->create_screen();
@@ -87,5 +91,11 @@ error1:
 void
 stw_shared_cleanup(void)
 {
+   if(stw_dev) {
+#ifdef DEBUG
+      debug_memory_end(stw_dev->memdbg_no);
+#endif
+   }
+
    stw_dev = NULL;
 }
diff --git a/src/gallium/state_trackers/wgl/shared/stw_device.h b/src/gallium/state_trackers/wgl/shared/stw_device.h
index bc0bce37c6..c1e041212b 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_device.h
+++ b/src/gallium/state_trackers/wgl/shared/stw_device.h
@@ -35,6 +35,10 @@ struct stw_device
 {
    const struct stw_winsys *stw_winsys;
    struct pipe_screen *screen;
+   
+#ifdef DEBUG
+   unsigned long memdbg_no;
+#endif
 };
 
 
diff --git a/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c b/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c
index 84b7b287b9..5cfdd41597 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c
@@ -25,7 +25,7 @@
  * 
  **************************************************************************/
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "stw_pixelformat.h"
 #include "stw_public.h"
 
diff --git a/src/gallium/state_trackers/wgl/wgl/stw_wgl.c b/src/gallium/state_trackers/wgl/wgl/stw_wgl.c
index f50b79b4e1..d7077ca5d4 100644
--- a/src/gallium/state_trackers/wgl/wgl/stw_wgl.c
+++ b/src/gallium/state_trackers/wgl/wgl/stw_wgl.c
@@ -27,7 +27,7 @@
 
 #include <windows.h>
 
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "shared/stw_public.h"
 #include "stw_wgl.h"
 #include "stw.h"
diff --git a/src/gallium/winsys/xlib/xlib_brw_aub.c b/src/gallium/winsys/xlib/xlib_brw_aub.c
index 73d1ac9161..b6bd849ef2 100644
--- a/src/gallium/winsys/xlib/xlib_brw_aub.c
+++ b/src/gallium/winsys/xlib/xlib_brw_aub.c
@@ -34,7 +34,7 @@
 #include "xlib_brw_aub.h"
 #include "pipe/p_context.h"
 #include "pipe/p_state.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 #include "util/u_memory.h"
 #include "softpipe/sp_texture.h"
 
diff --git a/src/mesa/drivers/dri/i915/i830_context.c b/src/mesa/drivers/dri/i915/i830_context.c
index fdce8af31f..9c540cb2bb 100644
--- a/src/mesa/drivers/dri/i915/i830_context.c
+++ b/src/mesa/drivers/dri/i915/i830_context.c
@@ -73,6 +73,8 @@ i830CreateContext(const __GLcontextModes * mesaVis,
       return GL_FALSE;
    }
 
+   _math_matrix_ctr(&intel->ViewportMatrix);
+
    /* Initialize swrast, tnl driver tables: */
    intelInitSpanFuncs(ctx);
    intelInitTriFuncs(ctx);
diff --git a/src/mesa/drivers/dri/i915/i915_context.c b/src/mesa/drivers/dri/i915/i915_context.c
index 6e2d41e19a..7549029a1b 100644
--- a/src/mesa/drivers/dri/i915/i915_context.c
+++ b/src/mesa/drivers/dri/i915/i915_context.c
@@ -119,6 +119,8 @@ i915CreateContext(const __GLcontextModes * mesaVis,
       return GL_FALSE;
    }
 
+   _math_matrix_ctr(&intel->ViewportMatrix);
+
    /* Initialize swrast, tnl driver tables: */
    intelInitSpanFuncs(ctx);
    intelInitTriFuncs(ctx);
diff --git a/src/mesa/drivers/dri/i965/brw_fallback.c b/src/mesa/drivers/dri/i965/brw_fallback.c
index e63098fdd4..5f4f2d515d 100644
--- a/src/mesa/drivers/dri/i965/brw_fallback.c
+++ b/src/mesa/drivers/dri/i965/brw_fallback.c
@@ -75,8 +75,8 @@ static GLboolean do_check_fallback(struct brw_context *brw)
    
    /* _NEW_STENCIL 
     */
-   if (ctx->Stencil.Enabled && 
-       !brw->intel.hw_stencil) {
+   if (ctx->Stencil.Enabled &&
+       (ctx->DrawBuffer->Name == 0 && !brw->intel.hw_stencil)) {
       DBG("FALLBACK: stencil\n");
       return GL_TRUE;
    }
diff --git a/src/mesa/drivers/dri/i965/brw_vtbl.c b/src/mesa/drivers/dri/i965/brw_vtbl.c
index b501a59ccd..15ceac6ba1 100644
--- a/src/mesa/drivers/dri/i965/brw_vtbl.c
+++ b/src/mesa/drivers/dri/i965/brw_vtbl.c
@@ -67,6 +67,8 @@ static void brw_destroy_context( struct intel_context *intel )
    brw_destroy_state(brw);
    brw_draw_destroy( brw );
 
+   _mesa_free(brw->wm.compile_data);
+
    brw_FrameBufferTexDestroy( brw );
 
    for (i = 0; i < brw->state.nr_draw_regions; i++)
@@ -90,6 +92,7 @@ static void brw_destroy_context( struct intel_context *intel )
    dri_bo_release(&brw->wm.bind_bo);
    for (i = 0; i < BRW_WM_MAX_SURF; i++)
       dri_bo_release(&brw->wm.surf_bo[i]);
+   dri_bo_release(&brw->wm.sampler_bo);
    dri_bo_release(&brw->wm.prog_bo);
    dri_bo_release(&brw->wm.state_bo);
    dri_bo_release(&brw->cc.prog_bo);
diff --git a/src/mesa/drivers/dri/intel/intel_clear.c b/src/mesa/drivers/dri/intel/intel_clear.c
index b229136316..7c297ec936 100644
--- a/src/mesa/drivers/dri/intel/intel_clear.c
+++ b/src/mesa/drivers/dri/intel/intel_clear.c
@@ -54,6 +54,17 @@
 
 #define FILE_DEBUG_FLAG DEBUG_BLIT
 
+#define TRI_CLEAR_COLOR_BITS (BUFFER_BIT_BACK_LEFT |			\
+			      BUFFER_BIT_FRONT_LEFT |			\
+			      BUFFER_BIT_COLOR0 |			\
+			      BUFFER_BIT_COLOR1 |			\
+			      BUFFER_BIT_COLOR2 |			\
+			      BUFFER_BIT_COLOR3 |			\
+			      BUFFER_BIT_COLOR4 |			\
+			      BUFFER_BIT_COLOR5 |			\
+			      BUFFER_BIT_COLOR6 |			\
+			      BUFFER_BIT_COLOR7)
+
 /**
  * Perform glClear where mask contains only color, depth, and/or stencil.
  *
@@ -75,8 +86,8 @@ intel_clear_tris(GLcontext *ctx, GLbitfield mask)
    GLboolean saved_shader_program = 0;
    unsigned int saved_active_texture;
 
-   assert((mask & ~(BUFFER_BIT_BACK_LEFT | BUFFER_BIT_FRONT_LEFT |
-		    BUFFER_BIT_DEPTH | BUFFER_BIT_STENCIL)) == 0);
+   assert((mask & ~(TRI_CLEAR_COLOR_BITS | BUFFER_BIT_DEPTH |
+		    BUFFER_BIT_STENCIL)) == 0);
 
    _mesa_PushAttrib(GL_COLOR_BUFFER_BIT |
 		    GL_CURRENT_BIT |
@@ -170,11 +181,11 @@ intel_clear_tris(GLcontext *ctx, GLbitfield mask)
 
    while (mask != 0) {
       GLuint this_mask = 0;
+      GLuint color_bit;
 
-      if (mask & BUFFER_BIT_BACK_LEFT)
-	 this_mask = BUFFER_BIT_BACK_LEFT;
-      else if (mask & BUFFER_BIT_FRONT_LEFT)
-	 this_mask = BUFFER_BIT_FRONT_LEFT;
+      color_bit = _mesa_ffs(mask & TRI_CLEAR_COLOR_BITS);
+      if (color_bit != 0)
+	 this_mask |= (1 << (color_bit - 1));
 
       /* Clear depth/stencil in the same pass as color. */
       this_mask |= (mask & (BUFFER_BIT_DEPTH | BUFFER_BIT_STENCIL));
@@ -186,6 +197,9 @@ intel_clear_tris(GLcontext *ctx, GLbitfield mask)
 	 _mesa_DrawBuffer(GL_FRONT_LEFT);
       else if (this_mask & BUFFER_BIT_BACK_LEFT)
 	 _mesa_DrawBuffer(GL_BACK_LEFT);
+      else if (color_bit != 0)
+	 _mesa_DrawBuffer(GL_COLOR_ATTACHMENT0 +
+			  (color_bit - BUFFER_COLOR0 - 1));
       else
 	 _mesa_ColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
 
@@ -313,8 +327,11 @@ intelClear(GLcontext *ctx, GLbitfield mask)
     * buffer with it.
     */
    if (mask & (BUFFER_BIT_DEPTH | BUFFER_BIT_STENCIL)) {
-      tri_mask |= blit_mask & BUFFER_BIT_BACK_LEFT;
-      blit_mask &= ~BUFFER_BIT_BACK_LEFT;
+      int color_bit = _mesa_ffs(mask & TRI_CLEAR_COLOR_BITS);
+      if (color_bit != 0) {
+	 tri_mask |= blit_mask & (1 << (color_bit - 1));
+	 blit_mask &= ~(1 << (color_bit - 1));
+      }
    }
 
    /* SW fallback clearing */
diff --git a/src/mesa/drivers/dri/intel/intel_context.c b/src/mesa/drivers/dri/intel/intel_context.c
index d7ccfa0605..1aa173db18 100644
--- a/src/mesa/drivers/dri/intel/intel_context.c
+++ b/src/mesa/drivers/dri/intel/intel_context.c
@@ -573,8 +573,6 @@ intelInitContext(struct intel_context *intel,
 
    intel->do_usleeps = (fthrottle_mode == DRI_CONF_FTHROTTLE_USLEEPS);
 
-   _math_matrix_ctr(&intel->ViewportMatrix);
-
    if (IS_965(intelScreen->deviceID) && !intel->intelScreen->irq_active) {
       _mesa_printf("IRQs not active.  Exiting\n");
       exit(1);
diff --git a/src/mesa/drivers/dri/intel/intel_screen.c b/src/mesa/drivers/dri/intel/intel_screen.c
index a52271158c..d6110b8163 100644
--- a/src/mesa/drivers/dri/intel/intel_screen.c
+++ b/src/mesa/drivers/dri/intel/intel_screen.c
@@ -302,6 +302,7 @@ intelDestroyScreen(__DRIscreenPrivate * sPriv)
 
    dri_bufmgr_destroy(intelScreen->bufmgr);
    intelUnmapScreenRegions(intelScreen);
+   driDestroyOptionCache(&intelScreen->optionCache);
 
    FREE(intelScreen);
    sPriv->private = NULL;
diff --git a/src/mesa/drivers/glslcompiler/glslcompiler.c b/src/mesa/drivers/glslcompiler/glslcompiler.c
index 918ec41856..54a4c1a7c6 100644
--- a/src/mesa/drivers/glslcompiler/glslcompiler.c
+++ b/src/mesa/drivers/glslcompiler/glslcompiler.c
@@ -72,6 +72,7 @@ struct options {
    const char *VertFile;
    const char *FragFile;
    const char *OutputFile;
+   GLboolean Params;
 };
 
 static struct options Options;
@@ -228,6 +229,8 @@ PrintShaderInstructions(GLuint shader, FILE *f)
    struct gl_shader *sh = _mesa_lookup_shader(ctx, shader);
    struct gl_program *prog = sh->Program;
    _mesa_fprint_program_opt(stdout, prog, Options.Mode, Options.LineNumbers);
+   if (Options.Params)
+      _mesa_print_program_parameters(ctx, prog);
 }
 
 
@@ -258,6 +261,7 @@ Usage(void)
    printf("  --debug            emit debug-style instructions\n");
    printf("  --number, -n       emit line numbers\n");
    printf("  --output, -o FILE  output filename\n");
+   printf("  --params           also emit program parameter info\n");
    printf("  --help             display this information\n");
 }
 
@@ -272,6 +276,7 @@ ParseOptions(int argc, char *argv[])
    Options.VertFile = NULL;
    Options.FragFile = NULL;
    Options.OutputFile = NULL;
+   Options.Params = GL_FALSE;
 
    if (argc == 1) {
       Usage();
@@ -305,6 +310,9 @@ ParseOptions(int argc, char *argv[])
          Options.OutputFile = argv[i + 1];
          i++;
       }
+      else if (strcmp(argv[i], "--params") == 0) {
+         Options.Params = GL_TRUE;
+      }
       else if (strcmp(argv[i], "--help") == 0) {
          Usage();
          exit(0);
diff --git a/src/mesa/main/extensions.c b/src/mesa/main/extensions.c
index 1b31e283cd..9c8bd13e69 100644
--- a/src/mesa/main/extensions.c
+++ b/src/mesa/main/extensions.c
@@ -208,7 +208,7 @@ _mesa_enable_sw_extensions(GLcontext *ctx)
    ctx->Extensions.ARB_shading_language_100 = GL_TRUE;
 #endif
 #if FEATURE_ARB_shading_language_120
-   ctx->Extensions.ARB_shading_language_120 = GL_FALSE; /* not quite done */
+   ctx->Extensions.ARB_shading_language_120 = GL_TRUE;
 #endif
    ctx->Extensions.ARB_shadow = GL_TRUE;
    ctx->Extensions.ARB_shadow_ambient = GL_TRUE;
@@ -433,7 +433,7 @@ _mesa_enable_2_1_extensions(GLcontext *ctx)
    ctx->Extensions.EXT_texture_sRGB = GL_TRUE;
 #endif
 #ifdef FEATURE_ARB_shading_language_120
-   ctx->Extensions.ARB_shading_language_120 = GL_FALSE; /* not quite done */
+   ctx->Extensions.ARB_shading_language_120 = GL_TRUE;
 #endif
 }
 
diff --git a/src/mesa/main/fbobject.c b/src/mesa/main/fbobject.c
index c3cdc11037..23b3fb68fb 100644
--- a/src/mesa/main/fbobject.c
+++ b/src/mesa/main/fbobject.c
@@ -303,6 +303,20 @@ _mesa_framebuffer_renderbuffer(GLcontext *ctx, struct gl_framebuffer *fb,
 
 
 /**
+ * For debug only.
+ */
+static void
+att_incomplete(const char *msg)
+{
+#if 0
+   _mesa_printf("attachment incomplete: %s\n", msg);
+#else
+   (void) msg;
+#endif
+}
+
+
+/**
  * Test if an attachment point is complete and update its Complete field.
  * \param format if GL_COLOR, this is a color attachment point,
  *               if GL_DEPTH, this is a depth component attachment point,
@@ -323,20 +337,26 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
       struct gl_texture_image *texImage;
 
       if (!texObj) {
+         att_incomplete("no texobj");
          att->Complete = GL_FALSE;
          return;
       }
 
       texImage = texObj->Image[att->CubeMapFace][att->TextureLevel];
       if (!texImage) {
+         att_incomplete("no teximage");
          att->Complete = GL_FALSE;
          return;
       }
       if (texImage->Width < 1 || texImage->Height < 1) {
+         att_incomplete("teximage width/height=0");
+         _mesa_printf("texobj = %u\n", texObj->Name);
+         _mesa_printf("level = %d\n", att->TextureLevel);
          att->Complete = GL_FALSE;
          return;
       }
       if (texObj->Target == GL_TEXTURE_3D && att->Zoffset >= texImage->Depth) {
+         att_incomplete("bad z offset");
          att->Complete = GL_FALSE;
          return;
       }
@@ -344,6 +364,7 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
       if (format == GL_COLOR) {
          if (texImage->TexFormat->BaseFormat != GL_RGB &&
              texImage->TexFormat->BaseFormat != GL_RGBA) {
+            att_incomplete("bad format");
             att->Complete = GL_FALSE;
             return;
          }
@@ -358,11 +379,13 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
          }
          else {
             att->Complete = GL_FALSE;
+            att_incomplete("bad depth format");
             return;
          }
       }
       else {
          /* no such thing as stencil textures */
+         att_incomplete("illegal stencil texture");
          att->Complete = GL_FALSE;
          return;
       }
@@ -372,6 +395,7 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
       if (!att->Renderbuffer->InternalFormat ||
           att->Renderbuffer->Width < 1 ||
           att->Renderbuffer->Height < 1) {
+         att_incomplete("0x0 renderbuffer");
          att->Complete = GL_FALSE;
          return;
       }
@@ -381,6 +405,7 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
             ASSERT(att->Renderbuffer->RedBits);
             ASSERT(att->Renderbuffer->GreenBits);
             ASSERT(att->Renderbuffer->BlueBits);
+            att_incomplete("bad renderbuffer color format");
             att->Complete = GL_FALSE;
             return;
          }
@@ -395,6 +420,7 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
             /* OK */
          }
          else {
+            att_incomplete("bad renderbuffer depth format");
             att->Complete = GL_FALSE;
             return;
          }
@@ -411,6 +437,7 @@ test_attachment_completeness(const GLcontext *ctx, GLenum format,
          }
          else {
             att->Complete = GL_FALSE;
+            att_incomplete("bad renderbuffer stencil format");
             return;
          }
       }
diff --git a/src/mesa/main/getstring.c b/src/mesa/main/getstring.c
index a9e22d340a..a42c44686d 100644
--- a/src/mesa/main/getstring.c
+++ b/src/mesa/main/getstring.c
@@ -93,7 +93,7 @@ compute_version(const GLcontext *ctx)
 			      (ctx->Extensions.EXT_stencil_two_side
 			       || ctx->Extensions.ATI_separate_stencil));
    const GLboolean ver_2_1 = (ver_2_0 &&
-                              /*ctx->Extensions.ARB_shading_language_120 &&*/
+                              ctx->Extensions.ARB_shading_language_120 &&
                               ctx->Extensions.EXT_pixel_buffer_object &&
                               ctx->Extensions.EXT_texture_sRGB);
    if (ver_2_1)
diff --git a/src/mesa/shader/prog_print.c b/src/mesa/shader/prog_print.c
index ce48767a87..516ea73301 100644
--- a/src/mesa/shader/prog_print.c
+++ b/src/mesa/shader/prog_print.c
@@ -212,15 +212,13 @@ reg_string(enum register_file f, GLint index, gl_prog_print_mode mode,
            GLboolean relAddr, const struct gl_program *prog)
 {
    static char str[100];
+   const char *addr = relAddr ? "ADDR+" : "";
 
    str[0] = 0;
 
    switch (mode) {
    case PROG_PRINT_DEBUG:
-      if (relAddr)
-         _mesa_sprintf(str, "%s[ADDR+%d]", file_string(f, mode), index);
-      else
-         _mesa_sprintf(str, "%s[%d]", file_string(f, mode), index);
+      _mesa_sprintf(str, "%s[%s%d]", file_string(f, mode), addr, index);
       break;
 
    case PROG_PRINT_ARB:
@@ -235,19 +233,19 @@ reg_string(enum register_file f, GLint index, gl_prog_print_mode mode,
          _mesa_sprintf(str, "temp%d", index);
          break;
       case PROGRAM_ENV_PARAM:
-         _mesa_sprintf(str, "program.env[%d]", index);
+         _mesa_sprintf(str, "program.env[%s%d]", addr, index);
          break;
       case PROGRAM_LOCAL_PARAM:
-         _mesa_sprintf(str, "program.local[%d]", index);
+         _mesa_sprintf(str, "program.local[%s%d]", addr, index);
          break;
       case PROGRAM_VARYING: /* extension */
-         _mesa_sprintf(str, "varying[%d]", index);
+         _mesa_sprintf(str, "varying[%s%d]", addr, index);
          break;
       case PROGRAM_CONSTANT: /* extension */
-         _mesa_sprintf(str, "constant[%d]", index);
+         _mesa_sprintf(str, "constant[%s%d]", addr, index);
          break;
       case PROGRAM_UNIFORM: /* extension */
-         _mesa_sprintf(str, "uniform[%d]", index);
+         _mesa_sprintf(str, "uniform[%s%d]", addr, index);
          break;
       case PROGRAM_STATE_VAR:
          {
@@ -284,16 +282,16 @@ reg_string(enum register_file f, GLint index, gl_prog_print_mode mode,
          _mesa_sprintf(str, "c[%d]", index);
          break;
       case PROGRAM_VARYING: /* extension */
-         _mesa_sprintf(str, "varying[%d]", index);
+         _mesa_sprintf(str, "varying[%s%d]", addr, index);
          break;
       case PROGRAM_UNIFORM: /* extension */
-         _mesa_sprintf(str, "uniform[%d]", index);
+         _mesa_sprintf(str, "uniform[%s%d]", addr, index);
          break;
       case PROGRAM_CONSTANT: /* extension */
-         _mesa_sprintf(str, "constant[%d]", index);
+         _mesa_sprintf(str, "constant[%s%d]", addr, index);
          break;
       case PROGRAM_STATE_VAR: /* extension */
-         _mesa_sprintf(str, "state[%d]", index);
+         _mesa_sprintf(str, "state[%s%d]", addr, index);
          break;
       default:
          _mesa_problem(NULL, "bad file in reg_string()");
diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c
index 2dd122c9a5..ac01462bd5 100644
--- a/src/mesa/shader/slang/slang_emit.c
+++ b/src/mesa/shader/slang/slang_emit.c
@@ -451,7 +451,7 @@ emit_arl_load(slang_emit_info *emitInfo,
    struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_ARL);
    inst->SrcReg[0].File = file;
    inst->SrcReg[0].Index = index;
-   inst->SrcReg[0].Swizzle = swizzle;
+   inst->SrcReg[0].Swizzle = fix_swizzle(swizzle);
    inst->DstReg.File = PROGRAM_ADDRESS;
    inst->DstReg.Index = 0;
    inst->DstReg.WriteMask = WRITEMASK_X;
@@ -873,6 +873,7 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
 
    if (n->Children[0]->Store->Size != n->Children[1]->Store->Size) {
       slang_info_log_error(emitInfo->log, "invalid operands to == or !=");
+      n->Store = NULL;
       return NULL;
    }
 
@@ -902,6 +903,7 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
       slang_ir_storage tempStore;
 
       if (!alloc_local_temp(emitInfo, &tempStore, 4)) {
+         n->Store = NULL;
          return NULL;
          /* out of temps */
       }
@@ -1813,6 +1815,25 @@ emit_cont_break_if_true(slang_emit_info *emitInfo, slang_ir_node *n)
 }
 
 
+/**
+ * Return the size of a swizzle mask given that some swizzle components
+ * may be NIL/undefined.  For example:
+ *  swizzle_size(".zzxx") = 4
+ *  swizzle_size(".xy??") = 2
+ *  swizzle_size(".w???") = 1
+ */
+static GLuint
+swizzle_size(GLuint swizzle)
+{
+   GLuint i;
+   for (i = 0; i < 4; i++) {
+      if (GET_SWZ(swizzle, i) == SWIZZLE_NIL)
+         return i;
+   }
+   return 4;
+}
+
+
 static struct prog_instruction *
 emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
 {
@@ -1820,14 +1841,24 @@ emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
 
    inst = emit(emitInfo, n->Children[0]);
 
-#if 0
-   assert(n->Store->Parent);
-   /* Apply this node's swizzle to parent's storage */
-   GLuint swizzle = n->Store->Swizzle;
-   _slang_copy_ir_storage(n->Store, n->Store->Parent);
-   n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
-   assert(!n->Store->Parent);
-#endif
+   if (n->Children[0]->Opcode == IR_VAR ||
+       n->Children[0]->Opcode == IR_SWIZZLE ||
+       n->Children[0]->Opcode == IR_ELEMENT) {
+      /* We can resolve the swizzle now.  Other swizzles will be resolved
+       * in storage_to_src_reg().
+       */
+      const GLuint swizzle = n->Store->Swizzle;
+      assert(n->Store->Parent);
+      /* new storage is parent storage with updated Swizzle + Size fields */
+      _slang_copy_ir_storage(n->Store, n->Store->Parent);
+      /* Apply this node's swizzle to parent's storage */
+      n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
+      /* Update size */
+      n->Store->Size = swizzle_size(n->Store->Swizzle);
+      assert(!n->Store->Parent);
+      assert(n->Store->Index >= 0);
+   }
+
    return inst;
 }
 
@@ -2428,7 +2459,9 @@ _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
       maxUniforms = ctx->Const.VertexProgram.MaxUniformComponents / 4;
    }
    if (prog->Parameters->NumParameters > maxUniforms) {
-      slang_info_log_error(log, "Constant/uniform register limit exceeded");
+      slang_info_log_error(log, "Constant/uniform register limit exceeded "
+                           "(max=%u vec4)", maxUniforms);
+
       return GL_FALSE;
    }
 
diff --git a/src/mesa/state_tracker/st_mesa_to_tgsi.c b/src/mesa/state_tracker/st_mesa_to_tgsi.c
index 8b9b18f94d..862d29fb1b 100644
--- a/src/mesa/state_tracker/st_mesa_to_tgsi.c
+++ b/src/mesa/state_tracker/st_mesa_to_tgsi.c
@@ -41,7 +41,7 @@
 #include "shader/prog_instruction.h"
 #include "shader/prog_parameter.h"
 #include "shader/prog_print.h"
-#include "pipe/p_debug.h"
+#include "util/u_debug.h"
 
 /*
  * Map mesa register file to TGSI register file.
@@ -475,15 +475,6 @@ compile_instruction(
       break;
    case OPCODE_RSQ:
       fullinst->Instruction.Opcode = TGSI_OPCODE_RSQ;
-
-      /* KW: Don't do this here.  If particular hardware needs to do
-       * this, can do so in the driver..
-       */
-#if 0
-       tgsi_util_set_full_src_register_sign_mode(
-          &fullinst->FullSrcRegisters[0],
-          TGSI_UTIL_SIGN_CLEAR ); 
-#endif
       break;
    case OPCODE_SCS:
       fullinst->Instruction.Opcode = TGSI_OPCODE_SCS;