Merge commit 'origin/gallium-0.1' into gallium-0.2

Conflicts:

	src/gallium/auxiliary/gallivm/instructionssoa.cpp
	src/gallium/auxiliary/gallivm/soabuiltins.c
	src/gallium/auxiliary/rtasm/rtasm_x86sse.c
	src/gallium/auxiliary/rtasm/rtasm_x86sse.h
	src/mesa/main/texenvprogram.c
	src/mesa/shader/arbprogparse.c
	src/mesa/shader/prog_statevars.c
	src/mesa/state_tracker/st_draw.c
	src/mesa/vbo/vbo_exec_draw.c

4 files changed, 92 insertions, 33 deletions

diff --git a/src/mesa/vbo/vbo_exec_api.c b/src/mesa/vbo/vbo_exec_api.c
index fdb0c5a9a4d..a6ce26ffed8 100644
--- a/src/mesa/vbo/vbo_exec_api.c
+++ b/src/mesa/vbo/vbo_exec_api.c
@@ -143,29 +143,37 @@ static void vbo_exec_copy_to_current( struct vbo_exec_context *exec )
 
    for (i = VBO_ATTRIB_POS+1 ; i < VBO_ATTRIB_MAX ; i++) {
       if (exec->vtx.attrsz[i]) {
-	 GLfloat *current = (GLfloat *)vbo->currval[i].Ptr;
-
          /* Note: the exec->vtx.current[i] pointers point into the
           * ctx->Current.Attrib and ctx->Light.Material.Attrib arrays.
           */
-	 COPY_CLEAN_4V(current, 
-		       exec->vtx.attrsz[i], 
-		       exec->vtx.attrptr[i]);
+	 GLfloat *current = (GLfloat *)vbo->currval[i].Ptr;
+         GLfloat tmp[4];
+
+         COPY_CLEAN_4V(tmp, 
+                       exec->vtx.attrsz[i], 
+                       exec->vtx.attrptr[i]);
+         
+         if (memcmp(current, tmp, sizeof(tmp)) != 0)
+         { 
+            memcpy(current, tmp, sizeof(tmp));
 
 	 
-	 /* Given that we explicitly state size here, there is no need
-	  * for the COPY_CLEAN above, could just copy 16 bytes and be
-	  * done.  The only problem is when Mesa accesses ctx->Current
-	  * directly.
-	  */
-	 vbo->currval[i].Size = exec->vtx.attrsz[i];
-
-	 /* This triggers rather too much recalculation of Mesa state
-	  * that doesn't get used (eg light positions).
-	  */
-	 if (i >= VBO_ATTRIB_MAT_FRONT_AMBIENT &&
-	     i <= VBO_ATTRIB_MAT_BACK_INDEXES)
-	    ctx->NewState |= _NEW_LIGHT;
+            /* Given that we explicitly state size here, there is no need
+             * for the COPY_CLEAN above, could just copy 16 bytes and be
+             * done.  The only problem is when Mesa accesses ctx->Current
+             * directly.
+             */
+            vbo->currval[i].Size = exec->vtx.attrsz[i];
+
+            /* This triggers rather too much recalculation of Mesa state
+             * that doesn't get used (eg light positions).
+             */
+            if (i >= VBO_ATTRIB_MAT_FRONT_AMBIENT &&
+                i <= VBO_ATTRIB_MAT_BACK_INDEXES)
+               ctx->NewState |= _NEW_LIGHT;
+            
+            ctx->NewState |= _NEW_CURRENT_ATTRIB;
+         }
       }
    }
 
diff --git a/src/mesa/vbo/vbo_exec_array.c b/src/mesa/vbo/vbo_exec_array.c
index 0f9d8da3568..8871e10cf60 100644
--- a/src/mesa/vbo/vbo_exec_array.c
+++ b/src/mesa/vbo/vbo_exec_array.c
@@ -127,6 +127,7 @@ static void recalculate_input_bindings( GLcontext *ctx )
    struct vbo_context *vbo = vbo_context(ctx);
    struct vbo_exec_context *exec = &vbo->exec;
    const struct gl_client_array **inputs = &exec->array.inputs[0];
+   GLbitfield const_inputs = 0x0;
    GLuint i;
 
    exec->array.program_mode = get_program_mode(ctx);
@@ -141,19 +142,24 @@ static void recalculate_input_bindings( GLcontext *ctx )
       for (i = 0; i <= VERT_ATTRIB_TEX7; i++) {
 	 if (exec->array.legacy_array[i]->Enabled)
 	    inputs[i] = exec->array.legacy_array[i];
-	 else
+	 else {
 	    inputs[i] = &vbo->legacy_currval[i];
+            const_inputs |= 1 << i;
+         }
       }
 
       for (i = 0; i < MAT_ATTRIB_MAX; i++) {
 	 inputs[VERT_ATTRIB_GENERIC0 + i] = &vbo->mat_currval[i];
+         const_inputs |= 1 << (VERT_ATTRIB_GENERIC0 + i);
       }
 
       /* Could use just about anything, just to fill in the empty
        * slots:
        */
-      for (i = MAT_ATTRIB_MAX; i < VERT_ATTRIB_MAX - VERT_ATTRIB_GENERIC0; i++)
+      for (i = MAT_ATTRIB_MAX; i < VERT_ATTRIB_MAX - VERT_ATTRIB_GENERIC0; i++) {
 	 inputs[VERT_ATTRIB_GENERIC0 + i] = &vbo->generic_currval[i];
+         const_inputs |= 1 << (VERT_ATTRIB_GENERIC0 + i);
+      }
 
       break;
    case VP_NV:
@@ -166,15 +172,19 @@ static void recalculate_input_bindings( GLcontext *ctx )
 	    inputs[i] = exec->array.generic_array[i];
 	 else if (exec->array.legacy_array[i]->Enabled)
 	    inputs[i] = exec->array.legacy_array[i];
-	 else
+	 else {
 	    inputs[i] = &vbo->legacy_currval[i];
+            const_inputs |= 1 << i;
+         }
       }
 
       /* Could use just about anything, just to fill in the empty
        * slots:
        */
-      for (i = VERT_ATTRIB_GENERIC0; i < VERT_ATTRIB_MAX; i++)
+      for (i = VERT_ATTRIB_GENERIC0; i < VERT_ATTRIB_MAX; i++) {
 	 inputs[i] = &vbo->generic_currval[i - VERT_ATTRIB_GENERIC0];
+         const_inputs |= 1 << i;
+      }
 
       break;
    case VP_ARB:
@@ -189,25 +199,34 @@ static void recalculate_input_bindings( GLcontext *ctx )
 	 inputs[0] = exec->array.generic_array[0];
       else if (exec->array.legacy_array[0]->Enabled)
 	 inputs[0] = exec->array.legacy_array[0];
-      else
+      else {
 	 inputs[0] = &vbo->legacy_currval[0];
+         const_inputs |= 1 << 0;
+      }
 
 
       for (i = 1; i <= VERT_ATTRIB_TEX7; i++) {
 	 if (exec->array.legacy_array[i]->Enabled)
 	    inputs[i] = exec->array.legacy_array[i];
-	 else
+	 else {
 	    inputs[i] = &vbo->legacy_currval[i];
+            const_inputs |= 1 << i;
+         }
       }
 
       for (i = 0; i < 16; i++) {
 	 if (exec->array.generic_array[i]->Enabled)
 	    inputs[VERT_ATTRIB_GENERIC0 + i] = exec->array.generic_array[i];
-	 else
+	 else {
 	    inputs[VERT_ATTRIB_GENERIC0 + i] = &vbo->generic_currval[i];
+            const_inputs |= 1 << (VERT_ATTRIB_GENERIC0 + i);
+         }
+
       }
       break;
    }
+
+   _mesa_set_varying_vp_inputs( ctx, ~const_inputs );
 }
 
 static void bind_arrays( GLcontext *ctx )
@@ -257,6 +276,11 @@ vbo_exec_DrawArrays(GLenum mode, GLint start, GLsizei count)
 
    bind_arrays( ctx );
 
+   /* Again...
+    */
+   if (ctx->NewState)
+      _mesa_update_state( ctx );
+
    prim[0].begin = 1;
    prim[0].end = 1;
    prim[0].weak = 0;
@@ -297,6 +321,9 @@ vbo_exec_DrawRangeElements(GLenum mode,
 
    bind_arrays( ctx );
 
+   if (ctx->NewState)
+      _mesa_update_state( ctx );
+
    ib.count = count;
    ib.type = type; 
    ib.obj = ctx->Array.ElementArrayBufferObj;
diff --git a/src/mesa/vbo/vbo_exec_draw.c b/src/mesa/vbo/vbo_exec_draw.c
index 92356ba9773..5bf3d836db5 100644
--- a/src/mesa/vbo/vbo_exec_draw.c
+++ b/src/mesa/vbo/vbo_exec_draw.c
@@ -150,6 +150,7 @@ static void vbo_exec_bind_arrays( GLcontext *ctx )
    GLubyte *data = exec->vtx.buffer_map;
    const GLuint *map;
    GLuint attr;
+   GLbitfield varying_inputs = 0x0;
 
    /* Install the default (ie Current) attributes first, then overlay
     * all active ones.
@@ -211,8 +212,11 @@ static void vbo_exec_bind_arrays( GLcontext *ctx )
 	 arrays[attr]._MaxElement = count; /* ??? */
 
 	 data += exec->vtx.attrsz[src] * sizeof(GLfloat);
+         varying_inputs |= 1<<attr;
       }
    }
+
+   _mesa_set_varying_vp_inputs( ctx, varying_inputs );
 }
 
 
@@ -242,6 +246,9 @@ void vbo_exec_vtx_flush( struct vbo_exec_context *exec )
 	  */
 	 vbo_exec_bind_arrays( ctx );
 
+         if (ctx->NewState)
+            _mesa_update_state( ctx );
+
          /* if using a real VBO, unmap it before drawing */
          if (exec->vtx.bufferobj->Name) {
             ctx->Driver.UnmapBuffer(ctx, target, exec->vtx.bufferobj);
diff --git a/src/mesa/vbo/vbo_save_draw.c b/src/mesa/vbo/vbo_save_draw.c
index ed82f09958d..0488c5d7182 100644
--- a/src/mesa/vbo/vbo_save_draw.c
+++ b/src/mesa/vbo/vbo_save_draw.c
@@ -64,18 +64,26 @@ static void _playback_copy_to_current( GLcontext *ctx,
    for (i = VBO_ATTRIB_POS+1 ; i < VBO_ATTRIB_MAX ; i++) {
       if (node->attrsz[i]) {
 	 GLfloat *current = (GLfloat *)vbo->currval[i].Ptr;
+         GLfloat tmp[4];
 
-	 COPY_CLEAN_4V(current, 
-		       node->attrsz[i], 
-		       data);
+         COPY_CLEAN_4V(tmp, 
+                       node->attrsz[i], 
+                       data);
+         
+         if (memcmp(current, tmp, 4 * sizeof(GLfloat)) != 0)
+         {
+            memcpy(current, tmp, 4 * sizeof(GLfloat));
 
-	 vbo->currval[i].Size = node->attrsz[i];
+            vbo->currval[i].Size = node->attrsz[i];
 
-	 data += node->attrsz[i];
+            if (i >= VBO_ATTRIB_FIRST_MATERIAL &&
+                i <= VBO_ATTRIB_LAST_MATERIAL)
+               ctx->NewState |= _NEW_LIGHT;
+
+            ctx->NewState |= _NEW_CURRENT_ATTRIB;
+         }
 
-	 if (i >= VBO_ATTRIB_FIRST_MATERIAL &&
-	     i <= VBO_ATTRIB_LAST_MATERIAL)
-	    ctx->NewState |= _NEW_LIGHT;
+	 data += node->attrsz[i];
       }
    }
 
@@ -110,6 +118,7 @@ static void vbo_bind_vertex_list( GLcontext *ctx,
    GLuint data = node->buffer_offset;
    const GLuint *map;
    GLuint attr;
+   GLbitfield varying_inputs = 0x0;
 
    /* Install the default (ie Current) attributes first, then overlay
     * all active ones.
@@ -159,8 +168,11 @@ static void vbo_bind_vertex_list( GLcontext *ctx,
 	 assert(arrays[attr].BufferObj->Name);
 
 	 data += node->attrsz[src] * sizeof(GLfloat);
+         varying_inputs |= 1<<attr;
       }
    }
+
+   _mesa_set_varying_vp_inputs( ctx, varying_inputs );
 }
 
 static void vbo_save_loopback_vertex_list( GLcontext *ctx,
@@ -229,6 +241,11 @@ void vbo_save_playback_vertex_list( GLcontext *ctx, void *data )
 
       vbo_bind_vertex_list( ctx, node );
 
+      /* Again...
+       */
+      if (ctx->NewState)
+	 _mesa_update_state( ctx );
+
       vbo_context(ctx)->draw_prims( ctx, 
 				    save->inputs, 
 				    node->prim,