Merge branch 'const-buffer-changes'

Conflicts:

	src/mesa/drivers/dri/i965/brw_curbe.c
	src/mesa/drivers/dri/i965/brw_vs_emit.c
	src/mesa/drivers/dri/i965/brw_wm_glsl.c

19 files changed, 505 insertions, 232 deletions

diff --git a/src/mesa/drivers/dri/i915/i915_context.c b/src/mesa/drivers/dri/i915/i915_context.c
index 1f9f363df92..367d2a3b648 100644
--- a/src/mesa/drivers/dri/i915/i915_context.c
+++ b/src/mesa/drivers/dri/i915/i915_context.c
@@ -73,7 +73,7 @@ i915InvalidateState(GLcontext * ctx, GLuint new_state)
          p->params_uptodate = 0;
    }
 
-   if (new_state & (_NEW_FOG | _NEW_HINT | _NEW_PROGRAM))
+   if (new_state & (_NEW_FOG | _NEW_HINT | _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS))
       i915_update_fog(ctx);
 }
 
diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index aef2ff5f86f..838e718d0d5 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -245,6 +245,9 @@ struct brw_vs_ouput_sizes {
 };
 
 
+/** Number of general purpose registers (VS, WM, etc) */
+#define BRW_MAX_GRF 128
+
 /** Number of texture sampler units */
 #define BRW_MAX_TEX_UNIT 16
 
@@ -450,8 +453,6 @@ struct brw_context
 
    struct {
       struct brw_state_flags dirty;
-      struct brw_tracked_state **atoms;
-      GLuint nr_atoms;
 
       GLuint nr_color_regions;
       struct intel_region *color_regions[MAX_DRAW_BUFFERS];
@@ -471,7 +472,8 @@ struct brw_context
       int validated_bo_count;
    } state;
 
-   struct brw_cache cache;
+   struct brw_cache cache;  /** non-surface items */
+   struct brw_cache surface_cache;  /* surface items */
    struct brw_cached_batch_item *cached_batch_items;
 
    struct {
@@ -555,11 +557,6 @@ struct brw_context
       GLuint vs_size;
       GLuint total_size;
 
-      /* Dynamic tracker which changes to reflect the state referenced
-       * by active fp and vp program parameters:
-       */
-      struct brw_tracked_state tracked_state;
-
       dri_bo *curbe_bo;
       /** Offset within curbe_bo of space for current curbe entry */
       GLuint curbe_offset;
diff --git a/src/mesa/drivers/dri/i965/brw_curbe.c b/src/mesa/drivers/dri/i965/brw_curbe.c
index 9197fede2d8..05a685af3da 100644
--- a/src/mesa/drivers/dri/i965/brw_curbe.c
+++ b/src/mesa/drivers/dri/i965/brw_curbe.c
@@ -36,6 +36,7 @@
 #include "main/macros.h"
 #include "main/enums.h"
 #include "shader/prog_parameter.h"
+#include "shader/prog_print.h"
 #include "shader/prog_statevars.h"
 #include "intel_batchbuffer.h"
 #include "intel_regions.h"
@@ -188,13 +189,6 @@ static void prepare_constant_buffer(struct brw_context *brw)
    GLfloat *buf;
    GLuint i;
 
-   /* Update our own dependency flags.  This works because this
-    * function will also be called whenever fp or vp changes.
-    */
-   brw->curbe.tracked_state.dirty.mesa = (_NEW_TRANSFORM|_NEW_PROJECTION);
-   brw->curbe.tracked_state.dirty.mesa |= vp->program.Base.Parameters->StateFlags;
-   brw->curbe.tracked_state.dirty.mesa |= fp->program.Base.Parameters->StateFlags;
-
    if (sz == 0) {
       if (brw->curbe.last_buf) {
 	 free(brw->curbe.last_buf);
@@ -363,11 +357,7 @@ update_constant_buffer(struct brw_context *brw,
       }
 
       if (0) {
-         int i;
-         for (i = 0; i < params->NumParameters; i++) {
-            float *p = params->ParameterValues[i];
-            printf("%d: %f %f %f %f\n", i, p[0], p[1], p[2], p[3]);
-         }
+         _mesa_print_parameter_list(params);
       }
    }
 }
@@ -380,7 +370,7 @@ update_vertex_constant_buffer(struct brw_context *brw)
    struct brw_vertex_program *vp =
       (struct brw_vertex_program *) brw->vertex_program;
    if (0) {
-      printf("update VS constants in buffer %p\n", vp->const_buffer);
+      printf("update VS constants in buffer %p  vp = %p\n", vp->const_buffer, vp);
       printf("program %u\n", vp->program.Base.Id);
    }
    if (vp->use_const_buffer)
@@ -394,6 +384,10 @@ update_fragment_constant_buffer(struct brw_context *brw)
 {
    struct brw_fragment_program *fp =
       (struct brw_fragment_program *) brw->fragment_program;
+   if (0) {
+      printf("update WM constants in buffer %p\n", fp->const_buffer);
+      printf("program %u\n", fp->program.Base.Id);
+   }
    if (fp->use_const_buffer)
       update_constant_buffer(brw, fp->program.Base.Parameters, fp->const_buffer);
 }
@@ -428,7 +422,7 @@ static void emit_constant_buffer(struct brw_context *brw)
  */
 const struct brw_tracked_state brw_constant_buffer = {
    .dirty = {
-      .mesa = (_NEW_TRANSFORM|_NEW_PROJECTION),      /* plus fp and vp flags */
+      .mesa = _NEW_PROGRAM_CONSTANTS,
       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 	       BRW_NEW_VERTEX_PROGRAM |
 	       BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
diff --git a/src/mesa/drivers/dri/i965/brw_state.h b/src/mesa/drivers/dri/i965/brw_state.h
index 81b0a45998f..7ea2fc113ca 100644
--- a/src/mesa/drivers/dri/i965/brw_state.h
+++ b/src/mesa/drivers/dri/i965/brw_state.h
@@ -135,8 +135,8 @@ dri_bo *brw_search_cache( struct brw_cache *cache,
 			  void *aux_return);
 void brw_state_cache_check_size( struct brw_context *brw );
 
-void brw_init_cache( struct brw_context *brw );
-void brw_destroy_cache( struct brw_context *brw );
+void brw_init_caches( struct brw_context *brw );
+void brw_destroy_caches( struct brw_context *brw );
 
 /***********************************************************************
  * brw_state_batch.c
diff --git a/src/mesa/drivers/dri/i965/brw_state_cache.c b/src/mesa/drivers/dri/i965/brw_state_cache.c
index d5b51664066..320d886c992 100644
--- a/src/mesa/drivers/dri/i965/brw_state_cache.c
+++ b/src/mesa/drivers/dri/i965/brw_state_cache.c
@@ -56,9 +56,9 @@
  * incorrect program is run for the other instance.
  */
 
+#include "main/imports.h"
 #include "brw_state.h"
 #include "intel_batchbuffer.h"
-#include "main/imports.h"
 
 /* XXX: Fixme - have to include these to get the sizes of the prog_key
  * structs:
@@ -69,8 +69,10 @@
 #include "brw_sf.h"
 #include "brw_gs.h"
 
-static GLuint hash_key( const void *key, GLuint key_size,
-			dri_bo **reloc_bufs, GLuint nr_reloc_bufs)
+
+static GLuint
+hash_key(const void *key, GLuint key_size,
+         dri_bo **reloc_bufs, GLuint nr_reloc_bufs)
 {
    GLuint *ikey = (GLuint *)key;
    GLuint hash = 0, i;
@@ -95,6 +97,7 @@ static GLuint hash_key( const void *key, GLuint key_size,
    return hash;
 }
 
+
 /**
  * Marks a new buffer as being chosen for the given cache id.
  */
@@ -111,6 +114,7 @@ update_cache_last(struct brw_cache *cache, enum brw_cache_id cache_id,
    cache->brw->state.dirty.cache |= 1 << cache_id;
 }
 
+
 static struct brw_cache_item *
 search_cache(struct brw_cache *cache, enum brw_cache_id cache_id,
 	     GLuint hash, const void *key, GLuint key_size,
@@ -143,7 +147,8 @@ search_cache(struct brw_cache *cache, enum brw_cache_id cache_id,
 }
 
 
-static void rehash( struct brw_cache *cache )
+static void
+rehash(struct brw_cache *cache)
 {
    struct brw_cache_item **items;
    struct brw_cache_item *c, *next;
@@ -164,15 +169,17 @@ static void rehash( struct brw_cache *cache )
    cache->size = size;
 }
 
+
 /**
  * Returns the buffer object matching cache_id and key, or NULL.
  */
-dri_bo *brw_search_cache( struct brw_cache *cache,
-			  enum brw_cache_id cache_id,
-			  const void *key,
-			  GLuint key_size,
-			  dri_bo **reloc_bufs, GLuint nr_reloc_bufs,
-			  void *aux_return )
+dri_bo *
+brw_search_cache(struct brw_cache *cache,
+                 enum brw_cache_id cache_id,
+                 const void *key,
+                 GLuint key_size,
+                 dri_bo **reloc_bufs, GLuint nr_reloc_bufs,
+                 void *aux_return)
 {
    struct brw_cache_item *item;
    GLuint hash = hash_key(key, key_size, reloc_bufs, nr_reloc_bufs);
@@ -192,6 +199,7 @@ dri_bo *brw_search_cache( struct brw_cache *cache,
    return item->bo;
 }
 
+
 dri_bo *
 brw_upload_cache( struct brw_cache *cache,
 		  enum brw_cache_id cache_id,
@@ -265,7 +273,9 @@ brw_upload_cache( struct brw_cache *cache,
    return bo;
 }
 
-/* This doesn't really work with aux data.  Use search/upload instead
+
+/**
+ * This doesn't really work with aux data.  Use search/upload instead
  */
 dri_bo *
 brw_cache_data_sz(struct brw_cache *cache,
@@ -296,6 +306,7 @@ brw_cache_data_sz(struct brw_cache *cache,
    return bo;
 }
 
+
 /**
  * Wrapper around brw_cache_data_sz using the cache_id's canonical key size.
  *
@@ -319,21 +330,22 @@ enum pool_type {
    DW_GENERAL_STATE
 };
 
+
 static void
-brw_init_cache_id( struct brw_context *brw,
-		const char *name,
-		enum brw_cache_id id,
-		GLuint key_size,
-		GLuint aux_size)
+brw_init_cache_id(struct brw_cache *cache,
+                  const char *name,
+                  enum brw_cache_id id,
+                  GLuint key_size,
+                  GLuint aux_size)
 {
-   struct brw_cache *cache = &brw->cache;
-
    cache->name[id] = strdup(name);
    cache->key_size[id] = key_size;
    cache->aux_size[id] = aux_size;
 }
 
-void brw_init_cache( struct brw_context *brw )
+
+static void
+brw_init_non_surface_cache(struct brw_context *brw)
 {
    struct brw_cache *cache = &brw->cache;
 
@@ -342,114 +354,136 @@ void brw_init_cache( struct brw_context *brw )
    cache->size = 7;
    cache->n_items = 0;
    cache->items = (struct brw_cache_item **)
-      _mesa_calloc(cache->size * 
-		   sizeof(struct brw_cache_item));
+      _mesa_calloc(cache->size * sizeof(struct brw_cache_item));
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "CC_VP",
 		     BRW_CC_VP,
 		     sizeof(struct brw_cc_viewport),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "CC_UNIT",
 		     BRW_CC_UNIT,
 		     sizeof(struct brw_cc_unit_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "WM_PROG",
 		     BRW_WM_PROG,
 		     sizeof(struct brw_wm_prog_key),
 		     sizeof(struct brw_wm_prog_data));
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "SAMPLER_DEFAULT_COLOR",
 		     BRW_SAMPLER_DEFAULT_COLOR,
 		     sizeof(struct brw_sampler_default_color),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "SAMPLER",
 		     BRW_SAMPLER,
 		     0,		/* variable key/data size */
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "WM_UNIT",
 		     BRW_WM_UNIT,
 		     sizeof(struct brw_wm_unit_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "SF_PROG",
 		     BRW_SF_PROG,
 		     sizeof(struct brw_sf_prog_key),
 		     sizeof(struct brw_sf_prog_data));
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "SF_VP",
 		     BRW_SF_VP,
 		     sizeof(struct brw_sf_viewport),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "SF_UNIT",
 		     BRW_SF_UNIT,
 		     sizeof(struct brw_sf_unit_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "VS_UNIT",
 		     BRW_VS_UNIT,
 		     sizeof(struct brw_vs_unit_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "VS_PROG",
 		     BRW_VS_PROG,
 		     sizeof(struct brw_vs_prog_key),
 		     sizeof(struct brw_vs_prog_data));
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "CLIP_UNIT",
 		     BRW_CLIP_UNIT,
 		     sizeof(struct brw_clip_unit_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "CLIP_PROG",
 		     BRW_CLIP_PROG,
 		     sizeof(struct brw_clip_prog_key),
 		     sizeof(struct brw_clip_prog_data));
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "GS_UNIT",
 		     BRW_GS_UNIT,
 		     sizeof(struct brw_gs_unit_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "GS_PROG",
 		     BRW_GS_PROG,
 		     sizeof(struct brw_gs_prog_key),
 		     sizeof(struct brw_gs_prog_data));
+}
+
+
+static void
+brw_init_surface_cache(struct brw_context *brw)
+{
+   struct brw_cache *cache = &brw->surface_cache;
 
-   brw_init_cache_id(brw,
+   cache->brw = brw;
+
+   cache->size = 7;
+   cache->n_items = 0;
+   cache->items = (struct brw_cache_item **)
+      _mesa_calloc(cache->size * sizeof(struct brw_cache_item));
+
+   brw_init_cache_id(cache,
 		     "SS_SURFACE",
 		     BRW_SS_SURFACE,
 		     sizeof(struct brw_surface_state),
 		     0);
 
-   brw_init_cache_id(brw,
+   brw_init_cache_id(cache,
 		     "SS_SURF_BIND",
 		     BRW_SS_SURF_BIND,
 		     0,
 		     0);
 }
 
+
+void
+brw_init_caches(struct brw_context *brw)
+{
+   brw_init_non_surface_cache(brw);
+   brw_init_surface_cache(brw);
+}
+
+
 static void
-brw_clear_cache( struct brw_context *brw )
+brw_clear_cache(struct brw_context *brw, struct brw_cache *cache)
 {
    struct brw_cache_item *c, *next;
    GLuint i;
@@ -457,8 +491,8 @@ brw_clear_cache( struct brw_context *brw )
    if (INTEL_DEBUG & DEBUG_STATE)
       _mesa_printf("%s\n", __FUNCTION__);
 
-   for (i = 0; i < brw->cache.size; i++) {
-      for (c = brw->cache.items[i]; c; c = next) {
+   for (i = 0; i < cache->size; i++) {
+      for (c = cache->items[i]; c; c = next) {
 	 int j;
 
 	 next = c->next;
@@ -468,10 +502,10 @@ brw_clear_cache( struct brw_context *brw )
 	 free((void *)c->key);
 	 free(c);
       }
-      brw->cache.items[i] = NULL;
+      cache->items[i] = NULL;
    }
 
-   brw->cache.n_items = 0;
+   cache->n_items = 0;
 
    if (brw->curbe.last_buf) {
       _mesa_free(brw->curbe.last_buf);
@@ -483,25 +517,40 @@ brw_clear_cache( struct brw_context *brw )
    brw->state.dirty.cache |= ~0;
 }
 
-void brw_state_cache_check_size( struct brw_context *brw )
+
+void
+brw_state_cache_check_size(struct brw_context *brw)
 {
    /* un-tuned guess.  We've got around 20 state objects for a total of around
     * 32k, so 1000 of them is around 1.5MB.
     */
    if (brw->cache.n_items > 1000)
-      brw_clear_cache(brw);
+      brw_clear_cache(brw, &brw->cache);
+
+   if (brw->surface_cache.n_items > 1000)
+      brw_clear_cache(brw, &brw->surface_cache);
 }
 
-void brw_destroy_cache( struct brw_context *brw )
+
+static void
+brw_destroy_cache(struct brw_context *brw, struct brw_cache *cache)
 {
    GLuint i;
 
-   brw_clear_cache(brw);
+   brw_clear_cache(brw, cache);
    for (i = 0; i < BRW_MAX_CACHE; i++) {
-      dri_bo_unreference(brw->cache.last_bo[i]);
-      free(brw->cache.name[i]);
+      dri_bo_unreference(cache->last_bo[i]);
+      free(cache->name[i]);
    }
-   free(brw->cache.items);
-   brw->cache.items = NULL;
-   brw->cache.size = 0;
+   free(cache->items);
+   cache->items = NULL;
+   cache->size = 0;
+}
+
+
+void
+brw_destroy_caches(struct brw_context *brw)
+{
+   brw_destroy_cache(brw, &brw->cache);
+   brw_destroy_cache(brw, &brw->surface_cache);
 }
diff --git a/src/mesa/drivers/dri/i965/brw_state_upload.c b/src/mesa/drivers/dri/i965/brw_state_upload.c
index 5de1450e612..2641bcb2aa8 100644
--- a/src/mesa/drivers/dri/i965/brw_state_upload.c
+++ b/src/mesa/drivers/dri/i965/brw_state_upload.c
@@ -59,7 +59,6 @@ const struct brw_tracked_state *atoms[] =
    &brw_curbe_offsets,
    &brw_recalculate_urb_fence,
 
-
    &brw_cc_vp,
    &brw_cc_unit,
 
@@ -88,54 +87,26 @@ const struct brw_tracked_state *atoms[] =
 
    &brw_line_stipple,
    &brw_aa_line_parameters,
-   /* Ordering of the commands below is documented as fixed.  
-    */
-#if 0
-   &brw_pipelined_state_pointers,
-   &brw_urb_fence,
-   &brw_constant_buffer_state,
-#else
+
    &brw_psp_urb_cbs,
-#endif
 
    &brw_drawing_rect,
    &brw_indices,
    &brw_vertices,
 
-   NULL,			/* brw_constant_buffer */
+   &brw_constant_buffer
 };
 
 
 void brw_init_state( struct brw_context *brw )
 {
-   GLuint i;
-
-   brw_init_cache(brw);
-
-   brw->state.atoms = _mesa_malloc(sizeof(atoms));
-   brw->state.nr_atoms = sizeof(atoms)/sizeof(*atoms);
-   _mesa_memcpy(brw->state.atoms, atoms, sizeof(atoms));
-
-   /* Patch in a pointer to the dynamic state atom:
-    */
-   for (i = 0; i < brw->state.nr_atoms; i++)
-      if (brw->state.atoms[i] == NULL)
-	 brw->state.atoms[i] = &brw->curbe.tracked_state;
-
-   _mesa_memcpy(&brw->curbe.tracked_state, 
-		&brw_constant_buffer,
-		sizeof(brw_constant_buffer));
+   brw_init_caches(brw);
 }
 
 
 void brw_destroy_state( struct brw_context *brw )
 {
-   if (brw->state.atoms) {
-      _mesa_free(brw->state.atoms);
-      brw->state.atoms = NULL;
-   }
-
-   brw_destroy_cache(brw);
+   brw_destroy_caches(brw);
    brw_destroy_batch_cache(brw);
 }
 
@@ -218,6 +189,7 @@ static struct dirty_bit_map mesa_bits[] = {
    DEFINE_BIT(_NEW_MULTISAMPLE),
    DEFINE_BIT(_NEW_TRACK_MATRIX),
    DEFINE_BIT(_NEW_PROGRAM),
+   DEFINE_BIT(_NEW_PROGRAM_CONSTANTS),
    {0, 0, 0}
 };
 
@@ -336,7 +308,7 @@ void brw_validate_state( struct brw_context *brw )
 
    /* do prepare stage for all atoms */
    for (i = 0; i < Elements(atoms); i++) {
-      const struct brw_tracked_state *atom = brw->state.atoms[i];
+      const struct brw_tracked_state *atom = atoms[i];
 
       if (brw->intel.Fallback)
          break;
@@ -367,8 +339,8 @@ void brw_upload_state(struct brw_context *brw)
       _mesa_memset(&examined, 0, sizeof(examined));
       prev = *state;
 
-      for (i = 0; i < brw->state.nr_atoms; i++) {	 
-	 const struct brw_tracked_state *atom = brw->state.atoms[i];
+      for (i = 0; i < Elements(atoms); i++) {	 
+	 const struct brw_tracked_state *atom = atoms[i];
 	 struct brw_state_flags generated;
 
 	 assert(atom->dirty.mesa ||
@@ -397,7 +369,7 @@ void brw_upload_state(struct brw_context *brw)
    }
    else {
       for (i = 0; i < Elements(atoms); i++) {	 
-	 const struct brw_tracked_state *atom = brw->state.atoms[i];
+	 const struct brw_tracked_state *atom = atoms[i];
 
 	 if (brw->intel.Fallback)
 	    break;
diff --git a/src/mesa/drivers/dri/i965/brw_vs_emit.c b/src/mesa/drivers/dri/i965/brw_vs_emit.c
index b69616d6e52..3fdc48583bf 100644
--- a/src/mesa/drivers/dri/i965/brw_vs_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_vs_emit.c
@@ -69,13 +69,18 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c )
 {
    GLuint i, reg = 0, mrf;
 
-#if 0
-   if (c->vp->program.Base.Parameters->NumParameters >= 6)
-      c->vp->use_const_buffer = 1;
+   /* Determine whether to use a real constant buffer or use a block
+    * of GRF registers for constants.  The later is faster but only
+    * works if everything fits in the GRF.
+    * XXX this heuristic/check may need some fine tuning...
+    */
+   if (c->vp->program.Base.Parameters->NumParameters +
+       c->vp->program.Base.NumTemporaries + 20 > BRW_MAX_GRF)
+      c->vp->use_const_buffer = GL_TRUE;
    else
-#endif
       c->vp->use_const_buffer = GL_FALSE;
-   /*printf("use_const_buffer = %d\n", c->use_const_buffer);*/
+
+   /*printf("use_const_buffer = %d\n", c->vp->use_const_buffer);*/
 
    /* r0 -- reserved as usual
     */
diff --git a/src/mesa/drivers/dri/i965/brw_wm.h b/src/mesa/drivers/dri/i965/brw_wm.h
index f0d31fc1ddc..2f80a60c12b 100644
--- a/src/mesa/drivers/dri/i965/brw_wm.h
+++ b/src/mesa/drivers/dri/i965/brw_wm.h
@@ -240,15 +240,18 @@ struct brw_wm_compile {
    GLuint max_wm_grf;
    GLuint last_scratch;
 
+   GLuint cur_inst;  /**< index of current instruction */
+
    /** Mapping from Mesa registers to hardware registers */
    struct {
       GLboolean inited;
       struct brw_reg reg;
    } wm_regs[PROGRAM_PAYLOAD+1][256][4];
 
+   GLboolean used_grf[BRW_WM_MAX_GRF];
+   GLuint first_free_grf;
    struct brw_reg stack;
    struct brw_reg emit_mask_reg;
-   GLuint reg_index;  /**< Index of next free GRF register */
    GLuint tmp_regs[BRW_WM_MAX_GRF];
    GLuint tmp_index;
    GLuint tmp_max;
diff --git a/src/mesa/drivers/dri/i965/brw_wm_glsl.c b/src/mesa/drivers/dri/i965/brw_wm_glsl.c
index 117460842a3..39ea95f6fc2 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_glsl.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_glsl.c
@@ -1,5 +1,7 @@
 #include "main/macros.h"
 #include "shader/prog_parameter.h"
+#include "shader/prog_print.h"
+#include "shader/prog_optimize.h"
 #include "brw_context.h"
 #include "brw_eu.h"
 #include "brw_wm.h"
@@ -42,6 +44,76 @@ GLboolean brw_wm_is_glsl(const struct gl_fragment_program *fp)
 }
 
 
+
+static void
+reclaim_temps(struct brw_wm_compile *c);
+
+
+/** Mark GRF register as used. */
+static void
+prealloc_grf(struct brw_wm_compile *c, int r)
+{
+   c->used_grf[r] = GL_TRUE;
+}
+
+
+/** Mark given GRF register as not in use. */
+static void
+release_grf(struct brw_wm_compile *c, int r)
+{
+   /*assert(c->used_grf[r]);*/
+   c->used_grf[r] = GL_FALSE;
+   c->first_free_grf = MIN2(c->first_free_grf, r);
+}
+
+
+/** Return index of a free GRF, mark it as used. */
+static int
+alloc_grf(struct brw_wm_compile *c)
+{
+   GLuint r;
+   for (r = c->first_free_grf; r < BRW_WM_MAX_GRF; r++) {
+      if (!c->used_grf[r]) {
+         c->used_grf[r] = GL_TRUE;
+         c->first_free_grf = r + 1;  /* a guess */
+         return r;
+      }
+   }
+
+   /* no free temps, try to reclaim some */
+   reclaim_temps(c);
+   c->first_free_grf = 0;
+
+   /* try alloc again */
+   for (r = c->first_free_grf; r < BRW_WM_MAX_GRF; r++) {
+      if (!c->used_grf[r]) {
+         c->used_grf[r] = GL_TRUE;
+         c->first_free_grf = r + 1;  /* a guess */
+         return r;
+      }
+   }
+
+   for (r = 0; r < BRW_WM_MAX_GRF; r++) {
+      assert(c->used_grf[r]);
+   }
+   /*printf("Really out of temp regs!\n");*/
+   return 60;
+}
+
+
+/** Return number of GRF registers used */
+static int
+num_grf_used(const struct brw_wm_compile *c)
+{
+   int r;
+   for (r = BRW_WM_MAX_GRF - 1; r >= 0; r--)
+      if (c->used_grf[r])
+         return r + 1;
+   return 0;
+}
+
+
+
 /**
  * Record the mapping of a Mesa register to a hardware register.
  */
@@ -68,11 +140,18 @@ static int get_scalar_dst_index(const struct prog_instruction *inst)
 static struct brw_reg alloc_tmp(struct brw_wm_compile *c)
 {
     struct brw_reg reg;
-    if(c->tmp_index == c->tmp_max)
-	c->tmp_regs[ c->tmp_max++ ] = c->reg_index++;
-    
+
+    /* if we need to allocate another temp, grow the tmp_regs[] array */
+    if (c->tmp_index == c->tmp_max) {
+       c->tmp_regs[ c->tmp_max++ ] = alloc_grf(c);
+    }
+
+    /* form the GRF register */
     reg = brw_vec8_grf(c->tmp_regs[ c->tmp_index++ ], 0);
+    /*printf("alloc_temp %d\n", reg.nr);*/
+    assert(reg.nr < BRW_WM_MAX_GRF);
     return reg;
+
 }
 
 /**
@@ -130,35 +209,26 @@ get_reg(struct brw_wm_compile *c, int file, int index, int component,
 	    return brw_null_reg();
     }
 
+    assert(index < 256);
     /* see if we've already allocated a HW register for this Mesa register */
     if (c->wm_regs[file][index][component].inited) {
-	/* yes, re-use */
-	reg = c->wm_regs[file][index][component].reg;
+       /* yes, re-use */
+       reg = c->wm_regs[file][index][component].reg;
     }
     else {
 	/* no, allocate new register */
-	reg = brw_vec8_grf(c->reg_index, 0);
-    }
+       int grf = alloc_grf(c);
+       if (grf < 0) {
+          /* totally out of temps */
+          grf = 70; /* XXX !!!! */
+       }
 
-    /* if this is a new register allocation, record it in the table */
-    if (!c->wm_regs[file][index][component].inited) {
-	set_reg(c, file, index, component, reg);
-	c->reg_index++;
-    }
+       reg = brw_vec8_grf(grf, 0);
+       /*printf("Alloc new grf %d for %d.%d\n", reg.nr, index, component);*/
 
-    if (c->reg_index >= BRW_WM_MAX_GRF - 12) {
-	/* ran out of temporary registers! */
-#if 1
-        /* This is a big hack for now.
-         * Return bad register index, just don't hang the GPU.
-         */
-        _mesa_fprintf(stderr, "out of regs %d\n", c->reg_index);
-        c->reg_index = BRW_WM_MAX_GRF - 13;
-#else
-	return brw_null_reg();
-#endif
+       set_reg(c, file, index, component, reg);
     }
- 
+
     if (neg & (1 << component)) {
 	reg = negate(reg);
     }
@@ -168,6 +238,46 @@ get_reg(struct brw_wm_compile *c, int file, int index, int component,
 }
 
 
+
+/**
+ * This is called if we run out of GRF registers.  Examine the live intervals
+ * of temp regs in the program and free those which won't be used again.
+ */
+static void
+reclaim_temps(struct brw_wm_compile *c)
+{
+   GLint intBegin[MAX_PROGRAM_TEMPS];
+   GLint intEnd[MAX_PROGRAM_TEMPS];
+   int index;
+
+   /*printf("Reclaim temps:\n");*/
+
+   _mesa_find_temp_intervals(c->prog_instructions, c->nr_fp_insns,
+                             intBegin, intEnd);
+
+   for (index = 0; index < MAX_PROGRAM_TEMPS; index++) {
+      if (intEnd[index] != -1 && intEnd[index] < c->cur_inst) {
+         /* program temp[i] can be freed */
+         int component;
+         /*printf("  temp[%d] is dead\n", index);*/
+         for (component = 0; component < 4; component++) {
+            if (c->wm_regs[PROGRAM_TEMPORARY][index][component].inited) {
+               int r = c->wm_regs[PROGRAM_TEMPORARY][index][component].reg.nr;
+               release_grf(c, r);
+               /*
+               printf("  Reclaim temp %d, reg %d at inst %d\n",
+                      index, r, c->cur_inst);
+               */
+               c->wm_regs[PROGRAM_TEMPORARY][index][component].inited = GL_FALSE;
+            }
+         }
+      }
+   }
+}
+
+
+
+
 /**
  * Preallocate registers.  This sets up the Mesa to hardware register
  * mapping for certain registers, such as constants (uniforms/state vars)
@@ -179,6 +289,10 @@ static void prealloc_reg(struct brw_wm_compile *c)
     struct brw_reg reg;
     int nr_interp_regs = 0;
     GLuint inputs = FRAG_BIT_WPOS | c->fp_interp_emitted | c->fp_deriv_emitted;
+    GLuint reg_index = 0;
+
+    memset(c->used_grf, GL_FALSE, sizeof(c->used_grf));
+    c->first_free_grf = 0;
 
     for (i = 0; i < 4; i++) {
         if (i < c->key.nr_depth_regs) 
@@ -187,14 +301,20 @@ static void prealloc_reg(struct brw_wm_compile *c)
             reg = brw_vec8_grf(0, 0);
 	set_reg(c, PROGRAM_PAYLOAD, PAYLOAD_DEPTH, i, reg);
     }
-    c->reg_index += 2 * c->key.nr_depth_regs;
+    reg_index += 2 * c->key.nr_depth_regs;
 
     /* constants */
     {
-        const int nr_params = c->fp->program.Base.Parameters->NumParameters;
+        const GLuint nr_params = c->fp->program.Base.Parameters->NumParameters;
+        const GLuint nr_temps = c->fp->program.Base.NumTemporaries;
 
         /* use a real constant buffer, or just use a section of the GRF? */
-        c->fp->use_const_buffer = GL_FALSE; /* (nr_params > 8);*/
+        /* XXX this heuristic may need adjustment... */
+        if ((nr_params + nr_temps) * 4 + reg_index > 80)
+           c->fp->use_const_buffer = GL_TRUE;
+        else
+           c->fp->use_const_buffer = GL_FALSE;
+        /*printf("WM use_const_buffer = %d\n", c->fp->use_const_buffer);*/
 
         if (c->fp->use_const_buffer) {
            /* We'll use a real constant buffer and fetch constants from
@@ -216,7 +336,7 @@ static void prealloc_reg(struct brw_wm_compile *c)
            for (i = 0; i < nr_params; i++) {
               /* loop over XYZW channels */
               for (j = 0; j < 4; j++, index++) {
-                 reg = brw_vec1_grf(c->reg_index + index / 8, index % 8);
+                 reg = brw_vec1_grf(reg_index + index / 8, index % 8);
                  /* Save pointer to parameter/constant value.
                   * Constants will be copied in prepare_constant_buffer()
                   */
@@ -226,7 +346,7 @@ static void prealloc_reg(struct brw_wm_compile *c)
            }
            /* number of constant regs used (each reg is float[8]) */
            c->nr_creg = 2 * ((4 * nr_params + 15) / 16);
-           c->reg_index += c->nr_creg;
+           reg_index += c->nr_creg;
         }
     }
 
@@ -234,20 +354,24 @@ static void prealloc_reg(struct brw_wm_compile *c)
     for (i = 0; i < FRAG_ATTRIB_MAX; i++) {
 	if (inputs & (1<<i)) {
 	    nr_interp_regs++;
-	    reg = brw_vec8_grf(c->reg_index, 0);
+	    reg = brw_vec8_grf(reg_index, 0);
 	    for (j = 0; j < 4; j++)
 		set_reg(c, PROGRAM_PAYLOAD, i, j, reg);
-	    c->reg_index += 2;
+	    reg_index += 2;
 	}
     }
 
     c->prog_data.first_curbe_grf = c->key.nr_depth_regs * 2;
     c->prog_data.urb_read_length = nr_interp_regs * 2;
     c->prog_data.curb_read_length = c->nr_creg;
-    c->emit_mask_reg = brw_uw1_reg(BRW_GENERAL_REGISTER_FILE, c->reg_index, 0);
-    c->reg_index++;
-    c->stack =  brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, c->reg_index, 0);
-    c->reg_index += 2;
+    c->emit_mask_reg = brw_uw1_reg(BRW_GENERAL_REGISTER_FILE, reg_index, 0);
+    reg_index++;
+    c->stack =  brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, reg_index, 0);
+    reg_index += 2;
+
+    /* mark GRF regs [0..reg_index-1] as in-use */
+    for (i = 0; i < reg_index; i++)
+       prealloc_grf(c, i);
 
     /* An instruction may reference up to three constants.
      * They'll be found in these registers.
@@ -256,7 +380,7 @@ static void prealloc_reg(struct brw_wm_compile *c)
     if (c->fp->use_const_buffer) {
        for (i = 0; i < 3; i++) {
           c->current_const[i].index = -1;
-          c->current_const[i].reg = alloc_tmp(c);
+          c->current_const[i].reg = brw_vec8_grf(alloc_grf(c), 0);
        }
     }
 #if 0
@@ -2595,7 +2719,6 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c)
     struct brw_compile *p = &c->func;
     struct brw_indirect stack_index = brw_indirect(0, 0);
 
-    c->reg_index = 0;
     prealloc_reg(c);
     brw_set_compression_control(p, BRW_COMPRESSION_NONE);
     brw_MOV(p, get_addr_reg(stack_index), brw_address(c->stack));
@@ -2603,6 +2726,8 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c)
     for (i = 0; i < c->nr_fp_insns; i++) {
         const struct prog_instruction *inst = &c->prog_instructions[i];
 
+        c->cur_inst = i;
+
 #if 0
         _mesa_printf("Inst %d: ", i);
         _mesa_print_instruction(inst);
@@ -2833,17 +2958,13 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c)
 		_mesa_printf("unsupported IR in fragment shader %d\n",
 			inst->Opcode);
 	}
+
 	if (inst->CondUpdate)
 	    brw_set_predicate_control(p, BRW_PREDICATE_NORMAL);
 	else
 	    brw_set_predicate_control(p, BRW_PREDICATE_NONE);
     }
     post_wm_emit(c);
-
-    if (c->reg_index >= BRW_WM_MAX_GRF) {
-        _mesa_problem(NULL, "Ran out of registers in brw_wm_emit_glsl()");
-        /* XXX we need to do some proper error recovery here */
-    }
 }
 
 
@@ -2867,6 +2988,6 @@ void brw_wm_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c)
         brw_wm_print_program(c, "brw_wm_glsl_emit done");
     }
 
-    c->prog_data.total_grf = c->reg_index;
+    c->prog_data.total_grf = num_grf_used(c);
     c->prog_data.total_scratch = 0;
 }
diff --git a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
index 805df8a4af4..f646ee7fc3a 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
@@ -268,7 +268,7 @@ brw_create_texture_surface( struct brw_context *brw,
       surf.ss0.cube_neg_z = 1;
    }
 
-   bo = brw_upload_cache(&brw->cache, BRW_SS_SURFACE,
+   bo = brw_upload_cache(&brw->surface_cache, BRW_SS_SURFACE,
 			 key, sizeof(*key),
 			 &key->bo, key->bo ? 1 : 0,
 			 &surf, sizeof(surf),
@@ -321,10 +321,11 @@ brw_update_texture_surface( GLcontext *ctx, GLuint unit )
    key.tiling = intelObj->mt->region->tiling;
 
    dri_bo_unreference(brw->wm.surf_bo[surf]);
-   brw->wm.surf_bo[surf] = brw_search_cache(&brw->cache, BRW_SS_SURFACE,
-                                         &key, sizeof(key),
-                                         &key.bo, key.bo ? 1 : 0,
-                                         NULL);
+   brw->wm.surf_bo[surf] = brw_search_cache(&brw->surface_cache,
+                                            BRW_SS_SURFACE,
+                                            &key, sizeof(key),
+                                            &key.bo, key.bo ? 1 : 0,
+                                            NULL);
    if (brw->wm.surf_bo[surf] == NULL) {
       brw->wm.surf_bo[surf] = brw_create_texture_surface(brw, &key);
    }
@@ -362,7 +363,7 @@ brw_create_constant_surface( struct brw_context *brw,
    surf.ss3.pitch = (key->pitch * key->cpp) - 1; /* ignored?? */
    brw_set_surface_tiling(&surf, key->tiling); /* tiling now allowed */
  
-   bo = brw_upload_cache(&brw->cache, BRW_SS_SURFACE,
+   bo = brw_upload_cache(&brw->surface_cache, BRW_SS_SURFACE,
 			 key, sizeof(*key),
 			 &key->bo, key->bo ? 1 : 0,
 			 &surf, sizeof(surf),
@@ -427,7 +428,8 @@ brw_update_wm_constant_surface( GLcontext *ctx,
    */
 
    dri_bo_unreference(brw->wm.surf_bo[surf]);
-   brw->wm.surf_bo[surf] = brw_search_cache(&brw->cache, BRW_SS_SURFACE,
+   brw->wm.surf_bo[surf] = brw_search_cache(&brw->surface_cache,
+                                            BRW_SS_SURFACE,
                                             &key, sizeof(key),
                                             &key.bo, key.bo ? 1 : 0,
                                             NULL);
@@ -456,17 +458,14 @@ brw_update_vs_constant_surface( GLcontext *ctx,
 
    assert(surf == 0);
 
-   /* free old const buffer if too small */
-   if (const_buffer && const_buffer->size < size) {
-      dri_bo_unreference(const_buffer);
-      const_buffer = NULL;
-   }
+   /* We always create a new VS constant buffer so that several can be
+    * in flight at a time.  Free the old one first...
+    */
+   dri_bo_unreference(const_buffer);
 
-   /* alloc new buffer if needed */
-   if (!const_buffer) {
-      const_buffer =
-         drm_intel_bo_alloc(intel->bufmgr, "vp_const_buffer", size, 64);
-   }
+   /* alloc new buffer */
+   const_buffer =
+      drm_intel_bo_alloc(intel->bufmgr, "vp_const_buffer", size, 64);
 
    memset(&key, 0, sizeof(key));
 
@@ -487,7 +486,8 @@ brw_update_vs_constant_surface( GLcontext *ctx,
    */
 
    dri_bo_unreference(brw->vs.surf_bo[surf]);
-   brw->vs.surf_bo[surf] = brw_search_cache(&brw->cache, BRW_SS_SURFACE,
+   brw->vs.surf_bo[surf] = brw_search_cache(&brw->surface_cache,
+                                            BRW_SS_SURFACE,
                                             &key, sizeof(key),
                                             &key.bo, key.bo ? 1 : 0,
                                             NULL);
@@ -569,10 +569,11 @@ brw_update_renderbuffer_surface(struct brw_context *brw,
    dri_bo_unreference(brw->wm.surf_bo[unit]);
    brw->wm.surf_bo[unit] = NULL;
    if (cached) 
-       brw->wm.surf_bo[unit] = brw_search_cache(&brw->cache, BRW_SS_SURFACE,
-	       &key, sizeof(key),
-	       &region_bo, 1,
-	       NULL);
+       brw->wm.surf_bo[unit] = brw_search_cache(&brw->surface_cache,
+                                                BRW_SS_SURFACE,
+                                                &key, sizeof(key),
+                                                &region_bo, 1,
+                                                NULL);
 
    if (brw->wm.surf_bo[unit] == NULL) {
       struct brw_surface_state surf;
@@ -598,7 +599,8 @@ brw_update_renderbuffer_surface(struct brw_context *brw,
       surf.ss0.writedisable_alpha = !key.color_mask[3];
 
       /* Key size will never match key size for textures, so we're safe. */
-      brw->wm.surf_bo[unit] = brw_upload_cache(&brw->cache, BRW_SS_SURFACE,
+      brw->wm.surf_bo[unit] = brw_upload_cache(&brw->surface_cache,
+                                               BRW_SS_SURFACE,
                                                &key, sizeof(key),
 					       &region_bo, 1,
 					       &surf, sizeof(surf),
@@ -630,7 +632,7 @@ brw_wm_get_binding_table(struct brw_context *brw)
 
    assert(brw->wm.nr_surfaces <= BRW_WM_MAX_SURF);
 
-   bind_bo = brw_search_cache(&brw->cache, BRW_SS_SURF_BIND,
+   bind_bo = brw_search_cache(&brw->surface_cache, BRW_SS_SURF_BIND,
 			      NULL, 0,
 			      brw->wm.surf_bo, brw->wm.nr_surfaces,
 			      NULL);
@@ -646,7 +648,7 @@ brw_wm_get_binding_table(struct brw_context *brw)
          else
             data[i] = 0;
 
-      bind_bo = brw_upload_cache( &brw->cache, BRW_SS_SURF_BIND,
+      bind_bo = brw_upload_cache( &brw->surface_cache, BRW_SS_SURF_BIND,
 				  NULL, 0,
 				  brw->wm.surf_bo, brw->wm.nr_surfaces,
 				  data, data_size,
@@ -746,7 +748,7 @@ brw_vs_get_binding_table(struct brw_context *brw)
 
    assert(brw->vs.nr_surfaces <= BRW_VS_MAX_SURF);
 
-   bind_bo = brw_search_cache(&brw->cache, BRW_SS_SURF_BIND,
+   bind_bo = brw_search_cache(&brw->surface_cache, BRW_SS_SURF_BIND,
 			      NULL, 0,
 			      brw->vs.surf_bo, brw->vs.nr_surfaces,
 			      NULL);
@@ -762,7 +764,7 @@ brw_vs_get_binding_table(struct brw_context *brw)
          else
             data[i] = 0;
 
-      bind_bo = brw_upload_cache( &brw->cache, BRW_SS_SURF_BIND,
+      bind_bo = brw_upload_cache( &brw->surface_cache, BRW_SS_SURF_BIND,
 				  NULL, 0,
 				  brw->vs.surf_bo, brw->vs.nr_surfaces,
 				  data, data_size,
@@ -787,8 +789,7 @@ brw_vs_get_binding_table(struct brw_context *brw)
 
 
 /**
- * Vertex shader surfaces.  Just constant buffer for now.  Could add vertex 
- * shader textures in the future.
+ * Vertex shader surfaces (constant buffer).
  */
 static void prepare_vs_surfaces(struct brw_context *brw )
 {
@@ -824,8 +825,12 @@ prepare_surfaces(struct brw_context *brw)
 
 const struct brw_tracked_state brw_wm_surfaces = {
    .dirty = {
-      .mesa = _NEW_COLOR | _NEW_TEXTURE | _NEW_BUFFERS | _NEW_PROGRAM,
-      .brw = BRW_NEW_CONTEXT,
+      .mesa = (_NEW_COLOR |
+               _NEW_TEXTURE |
+               _NEW_BUFFERS |
+               _NEW_PROGRAM |
+               _NEW_PROGRAM_CONSTANTS),
+      .brw = (BRW_NEW_CONTEXT),
       .cache = 0
    },
    .prepare = prepare_surfaces,
diff --git a/src/mesa/drivers/dri/r200/r200_state.c b/src/mesa/drivers/dri/r200/r200_state.c
index 2fcc87c0f5a..28ba5f49bc0 100644
--- a/src/mesa/drivers/dri/r200/r200_state.c
+++ b/src/mesa/drivers/dri/r200/r200_state.c
@@ -2484,7 +2484,7 @@ void r200ValidateState( GLcontext *ctx )
      r200UpdateDrawBuffer(ctx);
    }
 
-   if (new_state & (_NEW_TEXTURE | _NEW_PROGRAM)) {
+   if (new_state & (_NEW_TEXTURE | _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS)) {
       r200UpdateTextureState( ctx );
       new_state |= rmesa->NewGLState; /* may add TEXTURE_MATRIX */
       r200UpdateLocalViewer( ctx );
@@ -2523,6 +2523,7 @@ void r200ValidateState( GLcontext *ctx )
    }
 
    if (new_state & (_NEW_PROGRAM|
+                    _NEW_PROGRAM_CONSTANTS |
    /* need to test for pretty much anything due to possible parameter bindings */
 	_NEW_MODELVIEW|_NEW_PROJECTION|_NEW_TRANSFORM|
 	_NEW_LIGHT|_NEW_TEXTURE|_NEW_TEXTURE_MATRIX|
diff --git a/src/mesa/drivers/dri/r300/r300_fragprog.c b/src/mesa/drivers/dri/r300/r300_fragprog.c
index 873cde44144..2f45429cf27 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog.c
+++ b/src/mesa/drivers/dri/r300/r300_fragprog.c
@@ -470,7 +470,8 @@ void r300TranslateFragmentShader(r300ContextPtr r300,
 			fp->translated = GL_TRUE;
 		if (fp->error || (RADEON_DEBUG & DEBUG_PIXEL))
 			r300FragmentProgramDump(fp, &fp->code);
-		r300UpdateStateParameters(r300->radeon.glCtx, _NEW_PROGRAM);
+		r300UpdateStateParameters(r300->radeon.glCtx, _NEW_PROGRAM |
+                                          _NEW_PROGRAM_CONSTANTS);
 	}
 
 	update_params(r300, fp);
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index 79f0b3625ca..b96ba4ed948 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -1109,7 +1109,7 @@ void r300UpdateStateParameters(GLcontext * ctx, GLuint new_state)
 	struct gl_program_parameter_list *paramList;
 	GLuint i;
 
-	if (!(new_state & (_NEW_BUFFERS | _NEW_PROGRAM)))
+	if (!(new_state & (_NEW_BUFFERS | _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS)))
 		return;
 
 	fp = (struct r300_fragment_program *)ctx->FragmentProgram._Current;
@@ -2357,11 +2357,12 @@ void r300UpdateShaders(r300ContextPtr rmesa)
 			hw_tcl_on = future_hw_tcl_on = 0;
 			r300ResetHwState(rmesa);
 
-			r300UpdateStateParameters(ctx, _NEW_PROGRAM);
+			r300UpdateStateParameters(ctx, _NEW_PROGRAM |
+                                                  _NEW_PROGRAM_CONSTANTS);
 			return;
 		}
 	}
-	r300UpdateStateParameters(ctx, _NEW_PROGRAM);
+	r300UpdateStateParameters(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
 }
 
 static const GLfloat *get_fragmentprogram_constant(GLcontext *ctx,
diff --git a/src/mesa/drivers/dri/r300/r500_fragprog.c b/src/mesa/drivers/dri/r300/r500_fragprog.c
index 292573de893..300559d0b4d 100644
--- a/src/mesa/drivers/dri/r300/r500_fragprog.c
+++ b/src/mesa/drivers/dri/r300/r500_fragprog.c
@@ -501,7 +501,8 @@ void r500TranslateFragmentShader(r300ContextPtr r300,
 
 		_mesa_reference_program(r300->radeon.glCtx, &compiler.program, 0);
 
-		r300UpdateStateParameters(r300->radeon.glCtx, _NEW_PROGRAM);
+		r300UpdateStateParameters(r300->radeon.glCtx, _NEW_PROGRAM |
+                                          _NEW_PROGRAM_CONSTANTS);
 
 		if (RADEON_DEBUG & DEBUG_PIXEL) {
 			if (fp->translated) {
diff --git a/src/mesa/shader/arbprogram.c b/src/mesa/shader/arbprogram.c
index 981565ab8f1..317d623a228 100644
--- a/src/mesa/shader/arbprogram.c
+++ b/src/mesa/shader/arbprogram.c
@@ -74,8 +74,6 @@ _mesa_BindProgram(GLenum target, GLuint id)
    GET_CURRENT_CONTEXT(ctx);
    ASSERT_OUTSIDE_BEGIN_END(ctx);
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM);
-
    /* Error-check target and get curProg */
    if ((target == GL_VERTEX_PROGRAM_ARB) && /* == GL_VERTEX_PROGRAM_NV */
         (ctx->Extensions.NV_vertex_program ||
@@ -132,6 +130,9 @@ _mesa_BindProgram(GLenum target, GLuint id)
       return;
    }
 
+   /* signal new program (and its new constants) */
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+
    /* bind newProg */
    if (target == GL_VERTEX_PROGRAM_ARB) { /* == GL_VERTEX_PROGRAM_NV */
       _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
@@ -489,7 +490,7 @@ _mesa_ProgramEnvParameter4fARB(GLenum target, GLuint index,
    GET_CURRENT_CONTEXT(ctx);
    ASSERT_OUTSIDE_BEGIN_END(ctx);
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    if (target == GL_FRAGMENT_PROGRAM_ARB
        && ctx->Extensions.ARB_fragment_program) {
@@ -537,7 +538,7 @@ _mesa_ProgramEnvParameters4fvEXT(GLenum target, GLuint index, GLsizei count,
    GLfloat * dest;
    ASSERT_OUTSIDE_BEGIN_END(ctx);
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    if (count <= 0) {
       _mesa_error(ctx, GL_INVALID_VALUE, "glProgramEnvParameters4fv(count)");
@@ -631,7 +632,7 @@ _mesa_ProgramLocalParameter4fARB(GLenum target, GLuint index,
    struct gl_program *prog;
    ASSERT_OUTSIDE_BEGIN_END(ctx);
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    if ((target == GL_FRAGMENT_PROGRAM_NV
         && ctx->Extensions.NV_fragment_program) ||
@@ -685,7 +686,7 @@ _mesa_ProgramLocalParameters4fvEXT(GLenum target, GLuint index, GLsizei count,
    GLint i;
    ASSERT_OUTSIDE_BEGIN_END(ctx);
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    if (count <= 0) {
       _mesa_error(ctx, GL_INVALID_VALUE, "glProgramLocalParameters4fv(count)");
diff --git a/src/mesa/shader/nvprogram.c b/src/mesa/shader/nvprogram.c
index 5142c2a4a59..8ba521182b8 100644
--- a/src/mesa/shader/nvprogram.c
+++ b/src/mesa/shader/nvprogram.c
@@ -706,7 +706,7 @@ _mesa_ProgramNamedParameter4fNV(GLuint id, GLsizei len, const GLubyte *name,
    GET_CURRENT_CONTEXT(ctx);
    ASSERT_OUTSIDE_BEGIN_END(ctx);
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    prog = _mesa_lookup_program(ctx, id);
    if (!prog || prog->Target != GL_FRAGMENT_PROGRAM_NV) {
diff --git a/src/mesa/shader/prog_optimize.c b/src/mesa/shader/prog_optimize.c
index 6ba2e76ff96..a02f5efa417 100644
--- a/src/mesa/shader/prog_optimize.c
+++ b/src/mesa/shader/prog_optimize.c
@@ -547,15 +547,13 @@ update_interval(GLint intBegin[], GLint intEnd[], GLuint index, GLuint ic)
 
 
 /**
- * Find the live intervals for each temporary register in the program.
- * For register R, the interval [A,B] indicates that R is referenced
- * from instruction A through instruction B.
- * Special consideration is needed for loops and subroutines.
- * \return GL_TRUE if success, GL_FALSE if we cannot proceed for some reason
+ * Find first/last instruction that references each temporary register.
  */
-static GLboolean
-find_live_intervals(struct gl_program *prog,
-                    struct interval_list *liveIntervals)
+GLboolean
+_mesa_find_temp_intervals(const struct prog_instruction *instructions,
+                          GLuint numInstructions,
+                          GLint intBegin[MAX_PROGRAM_TEMPS],
+                          GLint intEnd[MAX_PROGRAM_TEMPS])
 {
    struct loop_info
    {
@@ -563,26 +561,15 @@ find_live_intervals(struct gl_program *prog,
    };
    struct loop_info loopStack[MAX_LOOP_NESTING];
    GLuint loopStackDepth = 0;
-   GLint intBegin[MAX_PROGRAM_TEMPS], intEnd[MAX_PROGRAM_TEMPS];
    GLuint i;
 
-   /*
-    * Note: we'll return GL_FALSE below if we find relative indexing
-    * into the TEMP register file.  We can't handle that yet.
-    * We also give up on subroutines for now.
-    */
-
-   if (dbg) {
-      _mesa_printf("Optimize: Begin find intervals\n");
-   }
-
    for (i = 0; i < MAX_PROGRAM_TEMPS; i++){
       intBegin[i] = intEnd[i] = -1;
    }
 
    /* Scan instructions looking for temporary registers */
-   for (i = 0; i < prog->NumInstructions; i++) {
-      const struct prog_instruction *inst = prog->Instructions + i;
+   for (i = 0; i < numInstructions; i++) {
+      const struct prog_instruction *inst = instructions + i;
       if (inst->Opcode == OPCODE_BGNLOOP) {
          loopStack[loopStackDepth].Start = i;
          loopStack[loopStackDepth].End = inst->BranchTarget;
@@ -595,7 +582,7 @@ find_live_intervals(struct gl_program *prog,
          return GL_FALSE;
       }
       else {
-         const GLuint numSrc = _mesa_num_inst_src_regs(inst->Opcode);
+         const GLuint numSrc = 3;/*_mesa_num_inst_src_regs(inst->Opcode);*/
          GLuint j;
          for (j = 0; j < numSrc; j++) {
             if (inst->SrcReg[j].File == PROGRAM_TEMPORARY) {
@@ -624,6 +611,39 @@ find_live_intervals(struct gl_program *prog,
       }
    }
 
+   return GL_TRUE;
+}
+
+
+/**
+ * Find the live intervals for each temporary register in the program.
+ * For register R, the interval [A,B] indicates that R is referenced
+ * from instruction A through instruction B.
+ * Special consideration is needed for loops and subroutines.
+ * \return GL_TRUE if success, GL_FALSE if we cannot proceed for some reason
+ */
+static GLboolean
+find_live_intervals(struct gl_program *prog,
+                    struct interval_list *liveIntervals)
+{
+   GLint intBegin[MAX_PROGRAM_TEMPS], intEnd[MAX_PROGRAM_TEMPS];
+   GLuint i;
+
+   /*
+    * Note: we'll return GL_FALSE below if we find relative indexing
+    * into the TEMP register file.  We can't handle that yet.
+    * We also give up on subroutines for now.
+    */
+
+   if (dbg) {
+      _mesa_printf("Optimize: Begin find intervals\n");
+   }
+
+   /* build intermediate arrays */
+   if (!_mesa_find_temp_intervals(prog->Instructions, prog->NumInstructions,
+                                  intBegin, intEnd))
+      return GL_FALSE;
+
    /* Build live intervals list from intermediate arrays */
    liveIntervals->Num = 0;
    for (i = 0; i < MAX_PROGRAM_TEMPS; i++) {
@@ -794,6 +814,96 @@ _mesa_reallocate_registers(struct gl_program *prog)
 
 
 
+
+
+
+
+#if 0
+static void
+_mesa_find_temporary_live_intervals(struct gl_program *prog,
+                                    GLint firstInst[MAX_PROGRAM_TEMPS],
+                                    GLint lastInst[MAX_PROGRAM_TEMPS])
+{
+   GLuint i;
+
+   for (i = 0; i < MAX_PROGRAM_TEMPS; i++) {
+      firstInst[i] = lastInst[i] = -1;
+   }
+
+   struct loop_info loopStack[MAX_LOOP_NESTING];
+   GLuint loopStackDepth = 0;
+   GLint intBegin[MAX_PROGRAM_TEMPS], intEnd[MAX_PROGRAM_TEMPS];
+   GLuint i;
+
+   /*
+    * Note: we'll return GL_FALSE below if we find relative indexing
+    * into the TEMP register file.  We can't handle that yet.
+    * We also give up on subroutines for now.
+    */
+
+   if (dbg) {
+      _mesa_printf("Optimize: Begin find intervals\n");
+   }
+
+   for (i = 0; i < MAX_PROGRAM_TEMPS; i++){
+      intBegin[i] = intEnd[i] = -1;
+   }
+
+   /* Scan instructions looking for temporary registers */
+   for (i = 0; i < prog->NumInstructions; i++) {
+      const struct prog_instruction *inst = prog->Instructions + i;
+      if (inst->Opcode == OPCODE_BGNLOOP) {
+         loopStack[loopStackDepth].Start = i;
+         loopStack[loopStackDepth].End = inst->BranchTarget;
+         loopStackDepth++;
+      }
+      else if (inst->Opcode == OPCODE_ENDLOOP) {
+         loopStackDepth--;
+      }
+      else if (inst->Opcode == OPCODE_CAL) {
+         return GL_FALSE;
+      }
+      else {
+         const GLuint numSrc = _mesa_num_inst_src_regs(inst->Opcode);
+         GLuint j;
+         for (j = 0; j < numSrc; j++) {
+            if (inst->SrcReg[j].File == PROGRAM_TEMPORARY) {
+               const GLuint index = inst->SrcReg[j].Index;
+               if (inst->SrcReg[j].RelAddr)
+                  return GL_FALSE;
+               update_interval(intBegin, intEnd, index, i);
+               if (loopStackDepth > 0) {
+                  /* extend temp register's interval to end of loop */
+                  GLuint loopEnd = loopStack[loopStackDepth - 1].End;
+                  update_interval(intBegin, intEnd, index, loopEnd);
+               }
+            }
+         }
+         if (inst->DstReg.File == PROGRAM_TEMPORARY) {
+            const GLuint index = inst->DstReg.Index;
+            if (inst->DstReg.RelAddr)
+               return GL_FALSE;
+            update_interval(intBegin, intEnd, index, i);
+            if (loopStackDepth > 0) {
+               /* extend temp register's interval to end of loop */
+               GLuint loopEnd = loopStack[loopStackDepth - 1].End;
+               update_interval(intBegin, intEnd, index, loopEnd);
+            }
+         }
+      }
+   }
+
+
+
+
+#endif
+
+
+
+
+
+
+
 /**
  * Apply optimizations to the given program to eliminate unnecessary
  * instructions, temp regs, etc.
diff --git a/src/mesa/shader/prog_optimize.h b/src/mesa/shader/prog_optimize.h
index d102cfd9fc1..43894a27237 100644
--- a/src/mesa/shader/prog_optimize.h
+++ b/src/mesa/shader/prog_optimize.h
@@ -25,7 +25,19 @@
 #ifndef PROG_OPT_H
 #define PROG_OPT_H
 
+
+#include "main/config.h"
+
+
 struct gl_program;
+struct prog_instruction;
+
+
+extern GLboolean
+_mesa_find_temp_intervals(const struct prog_instruction *instructions,
+                          GLuint numInstructions,
+                          GLint intBegin[MAX_PROGRAM_TEMPS],
+                          GLint intEnd[MAX_PROGRAM_TEMPS]);
 
 extern void
 _mesa_optimize_program(GLcontext *ctx, struct gl_program *program);
diff --git a/src/mesa/shader/shader_api.c b/src/mesa/shader/shader_api.c
index 644cd39185c..8f414a08891 100644
--- a/src/mesa/shader/shader_api.c
+++ b/src/mesa/shader/shader_api.c
@@ -1487,7 +1487,7 @@ _mesa_use_program(GLcontext *ctx, GLuint program)
       return;
    }
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
 
    if (program) {
       shProg = _mesa_lookup_shader_program_err(ctx, program, "glUseProgram");
@@ -1789,7 +1789,7 @@ _mesa_uniform(GLcontext *ctx, GLint location, GLsizei count,
       return;
    }
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    uniform = &shProg->Uniforms->Uniforms[location];
 
@@ -1929,7 +1929,7 @@ _mesa_uniform_matrix(GLcontext *ctx, GLint cols, GLint rows,
       return;
    }
 
-   FLUSH_VERTICES(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
+   FLUSH_VERTICES(ctx, _NEW_PROGRAM_CONSTANTS);
 
    uniform = &shProg->Uniforms->Uniforms[location];