Merge remote branch 'upstream/gallium-0.1' into nouveau-gallium-0.1

Conflicts:
	configs/default

39 files changed, 272 insertions, 148 deletions

diff --git a/src/gallium/auxiliary/draw/draw_context.c b/src/gallium/auxiliary/draw/draw_context.c
index 678c70e478..78249054f2 100644
--- a/src/gallium/auxiliary/draw/draw_context.c
+++ b/src/gallium/auxiliary/draw/draw_context.c
@@ -31,7 +31,8 @@
   */
 
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
 #include "draw_context.h"
 #include "draw_vbuf.h"
 #include "draw_vs.h"
@@ -289,7 +290,7 @@ draw_enable_point_sprites(struct draw_context *draw, boolean enable)
  * work for the drivers.
  */
 int
-draw_find_vs_output(struct draw_context *draw,
+draw_find_vs_output(const struct draw_context *draw,
                     uint semantic_name, uint semantic_index)
 {
    const struct draw_vertex_shader *vs = draw->vs.vertex_shader;
@@ -315,7 +316,7 @@ draw_find_vs_output(struct draw_context *draw,
  * Return number of vertex shader outputs.
  */
 uint
-draw_num_vs_outputs(struct draw_context *draw)
+draw_num_vs_outputs(const struct draw_context *draw)
 {
    uint count = draw->vs.vertex_shader->info.num_outputs;
    if (draw->extra_vp_outputs.slot > 0)
diff --git a/src/gallium/auxiliary/draw/draw_context.h b/src/gallium/auxiliary/draw/draw_context.h
index b8f2bfa332..0ab3681b64 100644
--- a/src/gallium/auxiliary/draw/draw_context.h
+++ b/src/gallium/auxiliary/draw/draw_context.h
@@ -84,11 +84,11 @@ draw_install_pstipple_stage(struct draw_context *draw, struct pipe_context *pipe
 
 
 int
-draw_find_vs_output(struct draw_context *draw,
+draw_find_vs_output(const struct draw_context *draw,
                     uint semantic_name, uint semantic_index);
 
 uint
-draw_num_vs_outputs(struct draw_context *draw);
+draw_num_vs_outputs(const struct draw_context *draw);
 
 
 
diff --git a/src/gallium/auxiliary/draw/draw_pipe.c b/src/gallium/auxiliary/draw/draw_pipe.c
index 3355c871ee..3cde9d36d3 100644
--- a/src/gallium/auxiliary/draw/draw_pipe.c
+++ b/src/gallium/auxiliary/draw/draw_pipe.c
@@ -30,7 +30,6 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pipe.h"
 
@@ -238,7 +237,7 @@ void draw_pipeline_run( struct draw_context *draw,
    do_line( draw,                                           \
             flags,                                          \
             verts + stride * ((i0) & ~DRAW_PIPE_FLAG_MASK), \
-            verts + stride * (i+1))
+            verts + stride * (i1))
 
 #define POINT(i0)                               \
    do_point( draw,                              \
diff --git a/src/gallium/auxiliary/draw/draw_pipe_aaline.c b/src/gallium/auxiliary/draw/draw_pipe_aaline.c
index 991304b2c8..20841bb5d6 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_aaline.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_aaline.c
@@ -32,11 +32,12 @@
  */
 
 
-#include "pipe/p_util.h"
 #include "pipe/p_inlines.h"
 #include "pipe/p_context.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_shader_tokens.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 
 #include "tgsi/tgsi_transform.h"
 #include "tgsi/tgsi_dump.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_aapoint.c b/src/gallium/auxiliary/draw/draw_pipe_aapoint.c
index c7f4349cb3..2c1cacbdb4 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_aapoint.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_aapoint.c
@@ -38,7 +38,6 @@
  */
 
 
-#include "pipe/p_util.h"
 #include "pipe/p_inlines.h"
 #include "pipe/p_context.h"
 #include "pipe/p_defines.h"
@@ -47,6 +46,9 @@
 #include "tgsi/tgsi_transform.h"
 #include "tgsi/tgsi_dump.h"
 
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
 #include "draw_context.h"
 #include "draw_vs.h"
 #include "draw_pipe.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_clip.c b/src/gallium/auxiliary/draw/draw_pipe_clip.c
index fa10f8efca..3265dcd154 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_clip.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_clip.c
@@ -32,7 +32,9 @@
  */
 
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
+
 #include "pipe/p_shader_tokens.h"
 
 #include "draw_vs.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_cull.c b/src/gallium/auxiliary/draw/draw_pipe_cull.c
index d0d22a38e0..053be5f050 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_cull.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_cull.c
@@ -33,7 +33,7 @@
  */
 
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "pipe/p_defines.h"
 #include "draw_pipe.h"
 
diff --git a/src/gallium/auxiliary/draw/draw_pipe_flatshade.c b/src/gallium/auxiliary/draw/draw_pipe_flatshade.c
index 4741b22d02..43d1fecc4d 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_flatshade.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_flatshade.c
@@ -28,7 +28,9 @@
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
 #include "pipe/p_shader_tokens.h"
 #include "draw_vs.h"
 #include "draw_pipe.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_offset.c b/src/gallium/auxiliary/draw/draw_pipe_offset.c
index 8f1650e55c..1fea5e6dcb 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_offset.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_offset.c
@@ -32,7 +32,8 @@
  * \author  Brian Paul
  */
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "draw_pipe.h"
 
 
@@ -84,8 +85,8 @@ static void do_offset_tri( struct draw_stage *stage,
    float a = ey*fz - ez*fy;
    float b = ez*fx - ex*fz;
 
-   float dzdx = FABSF(a * inv_det);
-   float dzdy = FABSF(b * inv_det);
+   float dzdx = fabsf(a * inv_det);
+   float dzdy = fabsf(b * inv_det);
 
    float zoffset = offset->units + MAX2(dzdx, dzdy) * offset->scale;
 
diff --git a/src/gallium/auxiliary/draw/draw_pipe_pstipple.c b/src/gallium/auxiliary/draw/draw_pipe_pstipple.c
index e97136fa1f..b764d9c518 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_pstipple.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_pstipple.c
@@ -34,12 +34,14 @@
  */
 
 
-#include "pipe/p_util.h"
 #include "pipe/p_inlines.h"
 #include "pipe/p_context.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_shader_tokens.h"
 
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
 #include "tgsi/tgsi_transform.h"
 #include "tgsi/tgsi_dump.h"
 
diff --git a/src/gallium/auxiliary/draw/draw_pipe_stipple.c b/src/gallium/auxiliary/draw/draw_pipe_stipple.c
index bf0db18a68..b65e2aa102 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_stipple.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_stipple.c
@@ -36,10 +36,12 @@
  */
 
 
-#include "pipe/p_util.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_shader_tokens.h"
-#include "draw_pipe.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
+#include "draw/draw_pipe.h"
 
 
 /** Subclass of draw_stage */
diff --git a/src/gallium/auxiliary/draw/draw_pipe_twoside.c b/src/gallium/auxiliary/draw/draw_pipe_twoside.c
index 3ac825f565..c329d92339 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_twoside.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_twoside.c
@@ -28,7 +28,8 @@
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_shader_tokens.h"
 #include "draw_vs.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_unfilled.c b/src/gallium/auxiliary/draw/draw_pipe_unfilled.c
index 8f97fdedaa..68835fd1a5 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_unfilled.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_unfilled.c
@@ -33,7 +33,7 @@
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "pipe/p_defines.h"
 #include "draw_private.h"
 #include "draw_pipe.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_util.c b/src/gallium/auxiliary/draw/draw_pipe_util.c
index 04438f4dd0..e22e5fed0c 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_util.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_util.c
@@ -30,7 +30,7 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pipe.h"
 
diff --git a/src/gallium/auxiliary/draw/draw_pipe_validate.c b/src/gallium/auxiliary/draw/draw_pipe_validate.c
index 6be1d369c3..f34c68728e 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_validate.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_validate.c
@@ -28,7 +28,7 @@
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "pipe/p_defines.h"
 #include "draw_private.h"
 #include "draw_pipe.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
index a6fde77a0e..c0cf4269db 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
@@ -35,7 +35,8 @@
 
 
 #include "pipe/p_debug.h"
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 
 #include "draw_vbuf.h"
 #include "draw_private.h"
diff --git a/src/gallium/auxiliary/draw/draw_pipe_wide_line.c b/src/gallium/auxiliary/draw/draw_pipe_wide_line.c
index 29649f5787..184e363594 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_wide_line.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_wide_line.c
@@ -28,9 +28,10 @@
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
 
-#include "pipe/p_util.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_shader_tokens.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "draw_private.h"
 #include "draw_pipe.h"
 
@@ -73,8 +74,8 @@ static void wideline_line( struct draw_stage *stage,
    float *pos2 = v2->data[pos];
    float *pos3 = v3->data[pos];
 
-   const float dx = FABSF(pos0[0] - pos2[0]);
-   const float dy = FABSF(pos0[1] - pos2[1]);
+   const float dx = fabsf(pos0[0] - pos2[0]);
+   const float dy = fabsf(pos0[1] - pos2[1]);
 
    /* small tweak to meet GL specification */
    const float bias = 0.125f;
diff --git a/src/gallium/auxiliary/draw/draw_pipe_wide_point.c b/src/gallium/auxiliary/draw/draw_pipe_wide_point.c
index 54590984c6..4f1326053d 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_wide_point.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_wide_point.c
@@ -28,7 +28,8 @@
 /* Authors:  Keith Whitwell <keith@tungstengraphics.com>
  */
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "pipe/p_defines.h"
 #include "pipe/p_shader_tokens.h"
 #include "draw_vs.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt.c b/src/gallium/auxiliary/draw/draw_pt.c
index 85a75525c8..669c11c993 100644
--- a/src/gallium/auxiliary/draw/draw_pt.c
+++ b/src/gallium/auxiliary/draw/draw_pt.c
@@ -30,7 +30,6 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pt.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt.h b/src/gallium/auxiliary/draw/draw_pt.h
index 3d2a9c78b7..c02f229110 100644
--- a/src/gallium/auxiliary/draw/draw_pt.h
+++ b/src/gallium/auxiliary/draw/draw_pt.h
@@ -98,9 +98,9 @@ struct draw_pt_middle_end {
                       unsigned count);
 
    /* Transform all vertices in a linear range and then draw them with
-    * the supplied element list.
+    * the supplied element list.  May fail and return FALSE.
     */
-   void (*run_linear_elts)( struct draw_pt_middle_end *,
+   boolean (*run_linear_elts)( struct draw_pt_middle_end *,
                             unsigned fetch_start,
                             unsigned fetch_count,
                             const ushort *draw_elts,
diff --git a/src/gallium/auxiliary/draw/draw_pt_emit.c b/src/gallium/auxiliary/draw/draw_pt_emit.c
index 40f05cb9e0..d4eca80588 100644
--- a/src/gallium/auxiliary/draw/draw_pt_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_emit.c
@@ -25,7 +25,7 @@
  *
  **************************************************************************/
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_vbuf.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch.c b/src/gallium/auxiliary/draw/draw_pt_fetch.c
index 07f4c99164..6377f896fb 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch.c
@@ -25,7 +25,7 @@
  *
  **************************************************************************/
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_vbuf.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
index 4a1f3b0953..5a4db6cfe5 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
@@ -30,7 +30,7 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_vbuf.h"
@@ -324,7 +324,7 @@ static void fetch_emit_run_linear( struct draw_pt_middle_end *middle,
 }
 
 
-static void fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
+static boolean fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
                                         unsigned start,
                                         unsigned count,
                                         const ushort *draw_elts,
@@ -341,10 +341,8 @@ static void fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
    hw_verts = draw->render->allocate_vertices( draw->render,
                                                (ushort)feme->translate->key.output_stride,
                                                (ushort)count );
-   if (!hw_verts) {
-      assert(0);
-      return;
-   }
+   if (!hw_verts) 
+      return FALSE;
 
    /* Single routine to fetch vertices and emit HW verts.
     */
@@ -367,6 +365,7 @@ static void fetch_emit_run_linear_elts( struct draw_pt_middle_end *middle,
                                    feme->translate->key.output_stride,
                                    count );
 
+   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 fdf9b6fe6a..73fc70c1bc 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_emit.c
@@ -31,7 +31,8 @@
   */
 
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_vbuf.h"
@@ -331,7 +332,7 @@ fse_run(struct draw_pt_middle_end *middle,
 
 
 
-static void fse_run_linear_elts( struct draw_pt_middle_end *middle, 
+static boolean fse_run_linear_elts( struct draw_pt_middle_end *middle, 
                                  unsigned start, 
                                  unsigned count,
                                  const ushort *draw_elts,
@@ -350,8 +351,7 @@ static void fse_run_linear_elts( struct draw_pt_middle_end *middle,
                                                (ushort)count );
 
    if (!hw_verts) {
-      assert(0);
-      return;
+      return FALSE;
    }
 
    /* Single routine to fetch vertices, run shader and emit HW verts.
@@ -373,6 +373,8 @@ static void fse_run_linear_elts( struct draw_pt_middle_end *middle,
 				   hw_verts, 
 				   fse->key.output_stride, 
 				   count );
+
+   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 0aec4b71ba..ec3b41c320 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_shade_pipeline.c
@@ -25,7 +25,8 @@
  *
  **************************************************************************/
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_vbuf.h"
 #include "draw/draw_vertex.h"
@@ -119,7 +120,7 @@ static void fetch_pipeline_run( struct draw_pt_middle_end *middle,
    struct draw_context *draw = fpme->draw;
    struct draw_vertex_shader *shader = draw->vs.vertex_shader;
    unsigned opt = fpme->opt;
-   unsigned alloc_count = align_int( fetch_count, 4 );
+   unsigned alloc_count = align( fetch_count, 4 );
 
    struct vertex_header *pipeline_verts = 
       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
@@ -195,7 +196,7 @@ static void fetch_pipeline_linear_run( struct draw_pt_middle_end *middle,
    struct draw_context *draw = fpme->draw;
    struct draw_vertex_shader *shader = draw->vs.vertex_shader;
    unsigned opt = fpme->opt;
-   unsigned alloc_count = align_int( count, 4 );
+   unsigned alloc_count = align( count, 4 );
 
    struct vertex_header *pipeline_verts =
       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
@@ -261,7 +262,7 @@ static void fetch_pipeline_linear_run( struct draw_pt_middle_end *middle,
 
 
 
-static void fetch_pipeline_linear_run_elts( struct draw_pt_middle_end *middle,
+static boolean fetch_pipeline_linear_run_elts( struct draw_pt_middle_end *middle,
                                             unsigned start,
                                             unsigned count,
                                             const ushort *draw_elts,
@@ -271,17 +272,13 @@ static void fetch_pipeline_linear_run_elts( struct draw_pt_middle_end *middle,
    struct draw_context *draw = fpme->draw;
    struct draw_vertex_shader *shader = draw->vs.vertex_shader;
    unsigned opt = fpme->opt;
-   unsigned alloc_count = align_int( count, 4 );
+   unsigned alloc_count = align( count, 4 );
 
    struct vertex_header *pipeline_verts =
       (struct vertex_header *)MALLOC(fpme->vertex_size * alloc_count);
 
-   if (!pipeline_verts) {
-      /* Not much we can do here - just skip the rendering.
-       */
-      assert(0);
-      return;
-   }
+   if (!pipeline_verts) 
+      return FALSE;
 
    /* Fetch into our vertex buffer
     */
@@ -335,6 +332,7 @@ static void fetch_pipeline_linear_run_elts( struct draw_pt_middle_end *middle,
    }
 
    FREE(pipeline_verts);
+   return TRUE;
 }
 
 
diff --git a/src/gallium/auxiliary/draw/draw_pt_post_vs.c b/src/gallium/auxiliary/draw/draw_pt_post_vs.c
index af6306b1c6..96dc706b99 100644
--- a/src/gallium/auxiliary/draw/draw_pt_post_vs.c
+++ b/src/gallium/auxiliary/draw/draw_pt_post_vs.c
@@ -25,7 +25,7 @@
  *
  **************************************************************************/
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "pipe/p_context.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt_util.c b/src/gallium/auxiliary/draw/draw_pt_util.c
index 32c8a9632c..3bc7939c55 100644
--- a/src/gallium/auxiliary/draw/draw_pt_util.c
+++ b/src/gallium/auxiliary/draw/draw_pt_util.c
@@ -30,7 +30,6 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pt.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt_varray.c b/src/gallium/auxiliary/draw/draw_pt_varray.c
index 46e722a154..c15afe65f1 100644
--- a/src/gallium/auxiliary/draw/draw_pt_varray.c
+++ b/src/gallium/auxiliary/draw/draw_pt_varray.c
@@ -25,7 +25,9 @@
  *
  **************************************************************************/
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pt.h"
diff --git a/src/gallium/auxiliary/draw/draw_pt_vcache.c b/src/gallium/auxiliary/draw/draw_pt_vcache.c
index cda2987c9e..80d7200ca6 100644
--- a/src/gallium/auxiliary/draw/draw_pt_vcache.c
+++ b/src/gallium/auxiliary/draw/draw_pt_vcache.c
@@ -30,7 +30,7 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pt.h"
@@ -63,7 +63,8 @@ struct vcache_frontend {
    unsigned opt;
 };
 
-static void vcache_flush( struct vcache_frontend *vcache )
+static INLINE void 
+vcache_flush( struct vcache_frontend *vcache )
 {
    if (vcache->middle_prim != vcache->output_prim) {
       vcache->middle_prim = vcache->output_prim;
@@ -86,7 +87,8 @@ static void vcache_flush( struct vcache_frontend *vcache )
    vcache->draw_count = 0;
 }
 
-static void vcache_check_flush( struct vcache_frontend *vcache )
+static INLINE void 
+vcache_check_flush( struct vcache_frontend *vcache )
 {
    if ( vcache->draw_count + 6 >= DRAW_MAX ||
         vcache->fetch_count + 4 >= FETCH_MAX )
@@ -96,9 +98,10 @@ static void vcache_check_flush( struct vcache_frontend *vcache )
 }
 
 
-static INLINE void vcache_elt( struct vcache_frontend *vcache,
-                               unsigned felt,
-                               ushort flags )
+static INLINE void 
+vcache_elt( struct vcache_frontend *vcache,
+            unsigned felt,
+            ushort flags )
 {
    unsigned idx = felt % CACHE_MAX;
 
@@ -115,10 +118,11 @@ static INLINE void vcache_elt( struct vcache_frontend *vcache,
 
 
                    
-static void vcache_triangle( struct vcache_frontend *vcache,
-                             unsigned i0,
-                             unsigned i1,
-                             unsigned i2 )
+static INLINE void 
+vcache_triangle( struct vcache_frontend *vcache,
+                 unsigned i0,
+                 unsigned i1,
+                 unsigned i2 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_elt(vcache, i1, 0);
@@ -127,11 +131,12 @@ static void vcache_triangle( struct vcache_frontend *vcache,
 }
 
 			  
-static void vcache_triangle_flags( struct vcache_frontend *vcache,
-                                   ushort flags,
-                                   unsigned i0,
-                                   unsigned i1,
-                                   unsigned i2 )
+static INLINE void 
+vcache_triangle_flags( struct vcache_frontend *vcache,
+                       ushort flags,
+                       unsigned i0,
+                       unsigned i1,
+                       unsigned i2 )
 {
    vcache_elt(vcache, i0, flags);
    vcache_elt(vcache, i1, 0);
@@ -139,9 +144,10 @@ static void vcache_triangle_flags( struct vcache_frontend *vcache,
    vcache_check_flush(vcache);
 }
 
-static void vcache_line( struct vcache_frontend *vcache,
-                         unsigned i0,
-                         unsigned i1 )
+static INLINE void 
+vcache_line( struct vcache_frontend *vcache,
+             unsigned i0,
+             unsigned i1 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_elt(vcache, i1, 0);
@@ -149,10 +155,11 @@ static void vcache_line( struct vcache_frontend *vcache,
 }
 
 
-static void vcache_line_flags( struct vcache_frontend *vcache,
-                               ushort flags,
-                               unsigned i0,
-                               unsigned i1 )
+static INLINE void 
+vcache_line_flags( struct vcache_frontend *vcache,
+                   ushort flags,
+                   unsigned i0,
+                   unsigned i1 )
 {
    vcache_elt(vcache, i0, flags);
    vcache_elt(vcache, i1, 0);
@@ -160,28 +167,31 @@ static void vcache_line_flags( struct vcache_frontend *vcache,
 }
 
 
-static void vcache_point( struct vcache_frontend *vcache,
-                          unsigned i0 )
+static INLINE void 
+vcache_point( struct vcache_frontend *vcache,
+              unsigned i0 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_check_flush(vcache);
 }
 
-static void vcache_quad( struct vcache_frontend *vcache,
-                         unsigned i0,
-                         unsigned i1,
-                         unsigned i2,
-                         unsigned i3 )
+static INLINE void 
+vcache_quad( struct vcache_frontend *vcache,
+             unsigned i0,
+             unsigned i1,
+             unsigned i2,
+             unsigned i3 )
 {
    vcache_triangle( vcache, i0, i1, i3 );
    vcache_triangle( vcache, i1, i2, i3 );
 }
 
-static void vcache_ef_quad( struct vcache_frontend *vcache,
-                            unsigned i0,
-                            unsigned i1,
-                            unsigned i2,
-                            unsigned i3 )
+static INLINE void 
+vcache_ef_quad( struct vcache_frontend *vcache,
+                unsigned i0,
+                unsigned i1,
+                unsigned i2,
+                unsigned i3 )
 {
    vcache_triangle_flags( vcache,
                           ( DRAW_PIPE_RESET_STIPPLE |
@@ -213,10 +223,11 @@ static void vcache_ef_quad( struct vcache_frontend *vcache,
 #define FUNC vcache_run
 #include "draw_pt_vcache_tmp.h"
 
-static void rebase_uint_elts( const unsigned *src,
-                              unsigned count,
-                              int delta,
-                              ushort *dest )
+static INLINE void 
+rebase_uint_elts( const unsigned *src,
+                  unsigned count,
+                  int delta,
+                  ushort *dest )
 {
    unsigned i;
 
@@ -224,9 +235,10 @@ static void rebase_uint_elts( const unsigned *src,
       dest[i] = (ushort)(src[i] + delta);
 }
 
-static void rebase_ushort_elts( const ushort *src,
-                                unsigned count,
-                                int delta,
+static INLINE void 
+rebase_ushort_elts( const ushort *src,
+                    unsigned count,
+                    int delta,
                                 ushort *dest )
 {
    unsigned i;
@@ -235,10 +247,11 @@ static void rebase_ushort_elts( const ushort *src,
       dest[i] = (ushort)(src[i] + delta);
 }
 
-static void rebase_ubyte_elts( const ubyte *src,
-                               unsigned count,
-                               int delta,
-                               ushort *dest )
+static INLINE void 
+rebase_ubyte_elts( const ubyte *src,
+                   unsigned count,
+                   int delta,
+                   ushort *dest )
 {
    unsigned i;
 
@@ -248,9 +261,10 @@ static void rebase_ubyte_elts( const ubyte *src,
 
 
 
-static void translate_uint_elts( const unsigned *src,
-                                 unsigned count,
-                                 ushort *dest )
+static INLINE void 
+translate_uint_elts( const unsigned *src,
+                     unsigned count,
+                     ushort *dest )
 {
    unsigned i;
 
@@ -258,9 +272,10 @@ static void translate_uint_elts( const unsigned *src,
       dest[i] = (ushort)(src[i]);
 }
 
-static void translate_ushort_elts( const ushort *src,
-                                   unsigned count,
-                                   ushort *dest )
+static INLINE void 
+translate_ushort_elts( const ushort *src,
+                       unsigned count,
+                       ushort *dest )
 {
    unsigned i;
 
@@ -268,9 +283,10 @@ static void translate_ushort_elts( const ushort *src,
       dest[i] = (ushort)(src[i]);
 }
 
-static void translate_ubyte_elts( const ubyte *src,
-                                  unsigned count,
-                                  ushort *dest )
+static INLINE void 
+translate_ubyte_elts( const ubyte *src,
+                      unsigned count,
+                      ushort *dest )
 {
    unsigned i;
 
@@ -282,7 +298,8 @@ static void translate_ubyte_elts( const ubyte *src,
 
 
 #if 0
-static enum pipe_format format_from_get_elt( pt_elt_func get_elt )
+static INLINE enum pipe_format 
+format_from_get_elt( pt_elt_func get_elt )
 {
    switch (draw->pt.user.eltSize) {
    case 1: return PIPE_FORMAT_R8_UNORM;
@@ -293,10 +310,11 @@ static enum pipe_format format_from_get_elt( pt_elt_func get_elt )
 }
 #endif
 
-static void vcache_check_run( struct draw_pt_front_end *frontend, 
-                              pt_elt_func get_elt,
-                              const void *elts,
-                              unsigned draw_count )
+static INLINE void 
+vcache_check_run( struct draw_pt_front_end *frontend, 
+                  pt_elt_func get_elt,
+                  const void *elts,
+                  unsigned draw_count )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend; 
    struct draw_context *draw = vcache->draw;
@@ -306,6 +324,7 @@ static void vcache_check_run( struct draw_pt_front_end *frontend,
    unsigned fetch_count = max_index + 1 - min_index;
    const ushort *transformed_elts;
    ushort *storage = NULL;
+   boolean ok;
 
 
    if (0) debug_printf("fetch_count %d fetch_max %d draw_count %d\n", fetch_count, 
@@ -313,7 +332,6 @@ static void vcache_check_run( struct draw_pt_front_end *frontend,
                        draw_count);
       
    if (max_index == 0xffffffff ||
-       fetch_count >= vcache->fetch_max ||
        fetch_count > draw_count) {
       if (0) debug_printf("fail\n");
       goto fail;
@@ -395,14 +413,19 @@ static void vcache_check_run( struct draw_pt_front_end *frontend,
       transformed_elts = storage;
    }
 
-   vcache->middle->run_linear_elts( vcache->middle,
-                                    min_index, /* start */
-                                    fetch_count,
-                                    transformed_elts,
-                                    draw_count );
-
+   ok = vcache->middle->run_linear_elts( vcache->middle,
+                                         min_index, /* start */
+                                         fetch_count,
+                                         transformed_elts,
+                                         draw_count );
+   
    FREE(storage);
-   return;
+
+   if (ok)
+      return;
+
+   debug_printf("failed to execute atomic draw elts for %d/%d, splitting up\n",
+                fetch_count, draw_count);
 
  fail:
    vcache_run( frontend, get_elt, elts, draw_count );
@@ -411,10 +434,11 @@ static void vcache_check_run( struct draw_pt_front_end *frontend,
 
 
 
-static void vcache_prepare( struct draw_pt_front_end *frontend,
-                            unsigned prim,
-                            struct draw_pt_middle_end *middle,
-			    unsigned opt )
+static void 
+vcache_prepare( struct draw_pt_front_end *frontend,
+                unsigned prim,
+                struct draw_pt_middle_end *middle,
+                unsigned opt )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
 
@@ -443,14 +467,16 @@ static void vcache_prepare( struct draw_pt_front_end *frontend,
 
 
 
-static void vcache_finish( struct draw_pt_front_end *frontend )
+static void 
+vcache_finish( struct draw_pt_front_end *frontend )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
    vcache->middle->finish( vcache->middle );
    vcache->middle = NULL;
 }
 
-static void vcache_destroy( struct draw_pt_front_end *frontend )
+static void 
+vcache_destroy( struct draw_pt_front_end *frontend )
 {
    FREE(frontend);
 }
diff --git a/src/gallium/auxiliary/draw/draw_vbuf.h b/src/gallium/auxiliary/draw/draw_vbuf.h
index e90f37872a..b0aa2df309 100644
--- a/src/gallium/auxiliary/draw/draw_vbuf.h
+++ b/src/gallium/auxiliary/draw/draw_vbuf.h
@@ -37,8 +37,6 @@
 #define DRAW_VBUF_H_
 
 
-#include "pipe/p_util.h"
-
 
 struct draw_context;
 struct vertex_info;
@@ -81,7 +79,8 @@ struct vbuf_render {
    boolean (*set_primitive)( struct vbuf_render *, unsigned prim );
 
    /**
-    * DrawElements, note indices are ushort:
+    * DrawElements, note indices are ushort.  The driver must complete
+    * this call, if necessary splitting the index list itself.
     */
    void (*draw)( struct vbuf_render *,
 		 const ushort *indices,
diff --git a/src/gallium/auxiliary/draw/draw_vs.c b/src/gallium/auxiliary/draw/draw_vs.c
index f798b20492..34adbd49b0 100644
--- a/src/gallium/auxiliary/draw/draw_vs.c
+++ b/src/gallium/auxiliary/draw/draw_vs.c
@@ -31,11 +31,15 @@
   *   Brian Paul
   */
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
+
 #include "pipe/p_shader_tokens.h"
+
 #include "draw_private.h"
 #include "draw_context.h"
 #include "draw_vs.h"
+
 #include "translate/translate.h"
 #include "translate/translate_cache.h"
 
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index 441877d46f..a556477a76 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -29,8 +29,10 @@
  */
 
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
 #include "pipe/p_shader_tokens.h"
+#include "pipe/p_debug.h"
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_util.h"
 #include "tgsi/tgsi_exec.h"
@@ -43,6 +45,7 @@
 
 #ifdef PIPE_ARCH_X86
 #define DISASSEM 0
+#define FAST_MATH 1
 
 static const char *files[] =
 {
@@ -117,21 +120,26 @@ static struct x86_reg get_reg_ptr(struct aos_compilation *cp,
 
    switch (file) {
    case TGSI_FILE_INPUT:
+      assert(idx < MAX_INPUTS);
       return x86_make_disp(ptr, Offset(struct aos_machine, input[idx]));
 
    case TGSI_FILE_OUTPUT:
       return x86_make_disp(ptr, Offset(struct aos_machine, output[idx]));
 
    case TGSI_FILE_TEMPORARY:
+      assert(idx < MAX_TEMPS);
       return x86_make_disp(ptr, Offset(struct aos_machine, temp[idx]));
 
    case AOS_FILE_INTERNAL:
+      assert(idx < MAX_INTERNALS);
       return x86_make_disp(ptr, Offset(struct aos_machine, internal[idx]));
 
    case TGSI_FILE_IMMEDIATE: 
+      assert(idx < MAX_IMMEDIATES);  /* just a sanity check */
       return x86_make_disp(aos_get_x86(cp, 0, X86_IMMEDIATES), idx * 4 * sizeof(float));
 
    case TGSI_FILE_CONSTANT: 
+      assert(idx < MAX_CONSTANTS);  /* just a sanity check */
       return x86_make_disp(aos_get_x86(cp, 1, X86_CONSTANTS), idx * 4 * sizeof(float));
 
    default:
@@ -1380,14 +1388,28 @@ static boolean emit_MAD( struct aos_compilation *cp, const struct tgsi_full_inst
    return TRUE;
 }
 
+
+
 /* A wrapper for powf().
  * Makes sure it is cdecl and operates on floats.
  */
 static float PIPE_CDECL _powerf( float x, float y )
 {
+#if FAST_MATH
+   return util_fast_pow(x, y);
+#else
    return powf( x, y );
+#endif
 }
 
+#if FAST_MATH
+static float PIPE_CDECL _exp2(float x)
+{
+   return util_fast_exp2(x);
+}
+#endif
+
+
 /* Really not sufficient -- need to check for conditions that could
  * generate inf/nan values, which will slow things down hugely.
  */
@@ -1442,6 +1464,48 @@ static boolean emit_POW( struct aos_compilation *cp, const struct tgsi_full_inst
 }
 
 
+#if FAST_MATH
+static boolean emit_EXPBASE2( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
+{
+   uint i;
+
+   /* For absolute correctness, need to spill/invalidate all XMM regs
+    * too.  
+    */
+   for (i = 0; i < 8; i++) {
+      if (cp->xmm[i].dirty) 
+         spill(cp, i);
+      aos_release_xmm_reg(cp, i);
+   }
+
+   /* Push caller-save (ie scratch) regs.  
+    */
+   x86_cdecl_caller_push_regs( cp->func );
+
+   x86_lea( cp->func, cp->stack_ESP, x86_make_disp(cp->stack_ESP, -4) );
+
+   x87_fld_src( cp, &op->FullSrcRegisters[0], 0 );
+   x87_fstp( cp->func, x86_make_disp( cp->stack_ESP, 0 ) );
+
+   /* tmp_EAX has been pushed & will be restored below */
+   x86_mov_reg_imm( cp->func, cp->tmp_EAX, (unsigned long) _exp2 );
+   x86_call( cp->func, cp->tmp_EAX );
+
+   x86_lea( cp->func, cp->stack_ESP, x86_make_disp(cp->stack_ESP, 4) );
+
+   x86_cdecl_caller_pop_regs( cp->func );
+
+   /* Note retval on x87 stack:
+    */
+   cp->func->x87_stack++;
+
+   x87_fstp_dest4( cp, &op->FullDstRegisters[0] );
+
+   return TRUE;
+}
+#endif
+
+
 static boolean emit_RCP( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
 {
    struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
@@ -1662,7 +1726,9 @@ emit_instruction( struct aos_compilation *cp,
       return emit_RND(cp, inst);
 
    case TGSI_OPCODE_EXPBASE2:
-#if 0
+#if FAST_MATH
+      return emit_EXPBASE2(cp, inst);
+#elif 0
       /* this seems to fail for "larger" exponents.
        * See glean tvertProg1's EX2 test.
        */
@@ -1827,6 +1893,8 @@ static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient,
    struct aos_compilation cp;
    unsigned fixup, label;
 
+   util_init_math();
+
    tgsi_parse_init( &parse, varient->base.vs->state.tokens );
 
    memset(&cp, 0, sizeof(cp));
@@ -2046,6 +2114,11 @@ static void PIPE_CDECL vaos_run_linear( struct draw_vs_varient *varient,
                          start,
                          count,
                          output_buffer );
+
+   /* Sanity spot checks to make sure we didn't trash our constants */
+   assert(machine->internal[IMM_ONES][0] == 1.0f);
+   assert(machine->internal[IMM_IDENTITY][0] == 0.0f);
+   assert(machine->internal[IMM_NEGS][0] == -1.0f);
 }
 
 
@@ -2135,4 +2208,4 @@ struct draw_vs_varient *draw_vs_varient_aos_sse( struct draw_vertex_shader *vs,
 
 
 
-#endif
+#endif /* PIPE_ARCH_X86 */
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h
index 64e021ff6b..7fe6f79db0 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.h
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.h
@@ -51,11 +51,11 @@ struct x86_function;
 #define W    3
 
 #define MAX_INPUTS     PIPE_MAX_ATTRIBS
-#define MAX_OUTPUTS    PIPE_MAX_ATTRIBS
-#define MAX_TEMPS      PIPE_MAX_ATTRIBS /* say */
-#define MAX_CONSTANTS  PIPE_MAX_ATTRIBS /* say */
-#define MAX_IMMEDIATES PIPE_MAX_ATTRIBS /* say */
-#define MAX_INTERNALS  8
+#define MAX_OUTPUTS    PIPE_MAX_SHADER_OUTPUTS
+#define MAX_TEMPS      TGSI_EXEC_NUM_TEMPS
+#define MAX_CONSTANTS  1024  /** only used for sanity checking */
+#define MAX_IMMEDIATES 1024  /** only used for sanity checking */
+#define MAX_INTERNALS  8     /** see IMM_x values below */
 
 #define AOS_FILE_INTERNAL TGSI_FILE_COUNT
 
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
index eda677cc62..26297c74f8 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
@@ -26,7 +26,7 @@
  **************************************************************************/
 
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "pipe/p_shader_tokens.h"
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_util.h"
@@ -55,9 +55,13 @@ static void emit_load_R32G32B32( struct aos_compilation *cp,
 				 struct x86_reg src_ptr )
 {
    sse_movss(cp->func, data, x86_make_disp(src_ptr, 8));
+   /* data = z ? ? ? */
    sse_shufps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) );
+   /* data = z ? 0 1 */
    sse_shufps(cp->func, data, data, SHUF(Y,Z,X,W) );
+   /* data = ? 0 z 1 */
    sse_movlps(cp->func, data, src_ptr);
+   /* data = x y z 1 */
 }
 
 static void emit_load_R32G32( struct aos_compilation *cp, 
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_machine.c b/src/gallium/auxiliary/draw/draw_vs_aos_machine.c
index e029b7b4bb..b358bd2df4 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos_machine.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos_machine.c
@@ -29,8 +29,9 @@
 #include "pipe/p_config.h"
 
 
-#include "pipe/p_util.h"
 #include "pipe/p_shader_tokens.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_util.h"
 #include "tgsi/tgsi_exec.h"
diff --git a/src/gallium/auxiliary/draw/draw_vs_exec.c b/src/gallium/auxiliary/draw/draw_vs_exec.c
index e26903d8cc..44563803f9 100644
--- a/src/gallium/auxiliary/draw/draw_vs_exec.c
+++ b/src/gallium/auxiliary/draw/draw_vs_exec.c
@@ -31,7 +31,8 @@
   *   Brian Paul
   */
 
-#include "pipe/p_util.h"
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "pipe/p_shader_tokens.h"
 
 #include "draw_private.h"
diff --git a/src/gallium/auxiliary/draw/draw_vs_llvm.c b/src/gallium/auxiliary/draw/draw_vs_llvm.c
index fc03473b91..2ce30b9a02 100644
--- a/src/gallium/auxiliary/draw/draw_vs_llvm.c
+++ b/src/gallium/auxiliary/draw/draw_vs_llvm.c
@@ -32,7 +32,6 @@
   *   Brian Paul
   */
 
-#include "pipe/p_util.h"
 #include "pipe/p_shader_tokens.h"
 #include "draw_private.h"
 #include "draw_context.h"
diff --git a/src/gallium/auxiliary/draw/draw_vs_sse.c b/src/gallium/auxiliary/draw/draw_vs_sse.c
index 61f0c084c3..0efabd9de8 100644
--- a/src/gallium/auxiliary/draw/draw_vs_sse.c
+++ b/src/gallium/auxiliary/draw/draw_vs_sse.c
@@ -31,13 +31,14 @@
   *   Brian Paul
   */
 
+#include "util/u_math.h"
+#include "util/u_memory.h"
 #include "pipe/p_config.h"
 
 #include "draw_vs.h"
 
 #if defined(PIPE_ARCH_X86)
 
-#include "pipe/p_util.h"
 #include "pipe/p_shader_tokens.h"
 
 #include "draw_private.h"
diff --git a/src/gallium/auxiliary/draw/draw_vs_varient.c b/src/gallium/auxiliary/draw/draw_vs_varient.c
index ad0b829afa..4daf05dae7 100644
--- a/src/gallium/auxiliary/draw/draw_vs_varient.c
+++ b/src/gallium/auxiliary/draw/draw_vs_varient.c
@@ -30,7 +30,8 @@
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_vbuf.h"
@@ -298,7 +299,7 @@ struct draw_vs_varient *draw_vs_varient_generic( struct draw_vertex_shader *vs,
          emit.element[i].input_offset = key->element[i].out.vs_output * 4 * sizeof(float);
          emit.element[i].output_format = draw_translate_vinfo_format(key->element[i].out.format);
          emit.element[i].output_offset = key->element[i].out.offset;
-         assert(emit.element[i].input_offset < fetch.output_stride);
+         assert(emit.element[i].input_offset <= fetch.output_stride);
       }
       else {
          emit.element[i].input_format = PIPE_FORMAT_R32_FLOAT;