19 files changed, 367 insertions, 313 deletions

diff --git a/src/gallium/drivers/identity/id_objects.c b/src/gallium/drivers/identity/id_objects.c
index e893e59940..bc9bc7121d 100644
--- a/src/gallium/drivers/identity/id_objects.c
+++ b/src/gallium/drivers/identity/id_objects.c
@@ -180,3 +180,42 @@ identity_transfer_destroy(struct identity_transfer *id_transfer)
    screen->tex_transfer_destroy(id_transfer->transfer);
    FREE(id_transfer);
 }
+
+struct pipe_video_surface *
+identity_video_surface_create(struct identity_screen *id_screen,
+                              struct pipe_video_surface *video_surface)
+{
+   struct identity_video_surface *id_video_surface;
+
+   if (!video_surface) {
+      goto error;
+   }
+
+   assert(video_surface->screen == id_screen->screen);
+
+   id_video_surface = CALLOC_STRUCT(identity_video_surface);
+   if (!id_video_surface) {
+      goto error;
+   }
+
+   memcpy(&id_video_surface->base,
+          video_surface,
+          sizeof(struct pipe_video_surface));
+
+   pipe_reference_init(&id_video_surface->base.reference, 1);
+   id_video_surface->base.screen = &id_screen->base;
+   id_video_surface->video_surface = video_surface;
+
+   return &id_video_surface->base;
+
+error:
+   pipe_video_surface_reference(&video_surface, NULL);
+   return NULL;
+}
+
+void
+identity_video_surface_destroy(struct identity_video_surface *id_video_surface)
+{
+   pipe_video_surface_reference(&id_video_surface->video_surface, NULL);
+   FREE(id_video_surface);
+}
diff --git a/src/gallium/drivers/identity/id_objects.h b/src/gallium/drivers/identity/id_objects.h
index ce58faa3c7..77cc719079 100644
--- a/src/gallium/drivers/identity/id_objects.h
+++ b/src/gallium/drivers/identity/id_objects.h
@@ -31,6 +31,7 @@
 
 #include "pipe/p_compiler.h"
 #include "pipe/p_state.h"
+#include "pipe/p_video_state.h"
 
 #include "id_screen.h"
 
@@ -67,6 +68,14 @@ struct identity_transfer
 };
 
 
+struct identity_video_surface
+{
+   struct pipe_video_surface base;
+
+   struct pipe_video_surface *video_surface;
+};
+
+
 static INLINE struct identity_buffer *
 identity_buffer(struct pipe_buffer *_buffer)
 {
@@ -103,6 +112,15 @@ identity_transfer(struct pipe_transfer *_transfer)
    return (struct identity_transfer *)_transfer;
 }
 
+static INLINE struct identity_video_surface *
+identity_video_surface(struct pipe_video_surface *_video_surface)
+{
+   if (!_video_surface) {
+      return NULL;
+   }
+   (void)identity_screen(_video_surface->screen);
+   return (struct identity_video_surface *)_video_surface;
+}
 
 static INLINE struct pipe_buffer *
 identity_buffer_unwrap(struct pipe_buffer *_buffer)
@@ -165,5 +183,12 @@ identity_transfer_create(struct identity_texture *id_texture,
 void
 identity_transfer_destroy(struct identity_transfer *id_transfer);
 
+struct pipe_video_surface *
+identity_video_surface_create(struct identity_screen *id_screen,
+                              struct pipe_video_surface *video_surface);
+
+void
+identity_video_surface_destroy(struct identity_video_surface *id_video_surface);
+
 
 #endif /* ID_OBJECTS_H */
diff --git a/src/gallium/drivers/identity/id_public.h b/src/gallium/drivers/identity/id_public.h
index cac14cfd60..3d2862eaa0 100644
--- a/src/gallium/drivers/identity/id_public.h
+++ b/src/gallium/drivers/identity/id_public.h
@@ -37,4 +37,4 @@ identity_screen_create(struct pipe_screen *screen);
 struct pipe_context *
 identity_context_create(struct pipe_screen *screen, struct pipe_context *pipe);
 
-#endif /* PT_PUBLIC_H */
+#endif /* ID_PUBLIC_H */
diff --git a/src/gallium/drivers/identity/id_screen.c b/src/gallium/drivers/identity/id_screen.c
index 26439637d0..53eae3ef54 100644
--- a/src/gallium/drivers/identity/id_screen.c
+++ b/src/gallium/drivers/identity/id_screen.c
@@ -379,6 +379,33 @@ identity_screen_buffer_destroy(struct pipe_buffer *_buffer)
    identity_buffer_destroy(identity_buffer(_buffer));
 }
 
+static struct pipe_video_surface *
+identity_screen_video_surface_create(struct pipe_screen *_screen,
+                                     enum pipe_video_chroma_format chroma_format,
+                                     unsigned width,
+                                     unsigned height)
+{
+   struct identity_screen *id_screen = identity_screen(_screen);
+   struct pipe_screen *screen = id_screen->screen;
+   struct pipe_video_surface *result;
+
+   result = screen->video_surface_create(screen,
+                                         chroma_format,
+                                         width,
+                                         height);
+
+   if (result) {
+      return identity_video_surface_create(id_screen, result);
+   }
+   return NULL;
+}
+
+static void
+identity_screen_video_surface_destroy(struct pipe_video_surface *_vsfc)
+{
+   identity_video_surface_destroy(identity_video_surface(_vsfc));
+}
+
 static void
 identity_screen_flush_frontbuffer(struct pipe_screen *_screen,
                                   struct pipe_surface *_surface,
@@ -472,6 +499,12 @@ identity_screen_create(struct pipe_screen *screen)
    if (screen->buffer_unmap)
       id_screen->base.buffer_unmap = identity_screen_buffer_unmap;
    id_screen->base.buffer_destroy = identity_screen_buffer_destroy;
+   if (screen->video_surface_create) {
+      id_screen->base.video_surface_create = identity_screen_video_surface_create;
+   }
+   if (screen->video_surface_destroy) {
+      id_screen->base.video_surface_destroy = identity_screen_video_surface_destroy;
+   }
    id_screen->base.flush_frontbuffer = identity_screen_flush_frontbuffer;
    id_screen->base.fence_reference = identity_screen_fence_reference;
    id_screen->base.fence_signalled = identity_screen_fence_signalled;
diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
index ae23329b83..769733b6dd 100644
--- a/src/gallium/drivers/r300/r300_context.c
+++ b/src/gallium/drivers/r300/r300_context.c
@@ -69,6 +69,7 @@ static void r300_destroy_context(struct pipe_context* context)
     FREE(r300->blend_color_state);
     FREE(r300->rs_block);
     FREE(r300->scissor_state);
+    FREE(r300->vertex_info);
     FREE(r300->viewport_state);
     FREE(r300);
 }
@@ -123,15 +124,24 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     r300->context.clear = r300_clear;
 
-    if (r300screen->caps->has_tcl)
-    {
+    if (r300screen->caps->has_tcl) {
         r300->context.draw_arrays = r300_draw_arrays;
         r300->context.draw_elements = r300_draw_elements;
         r300->context.draw_range_elements = r300_draw_range_elements;
-    }
-    else
-    {
-        assert(0);
+    } else {
+        r300->context.draw_arrays = r300_swtcl_draw_arrays;
+        r300->context.draw_elements = r300_draw_elements;
+        r300->context.draw_range_elements = r300_swtcl_draw_range_elements;
+
+        /* Create a Draw. This is used for SW TCL. */
+        r300->draw = draw_create();
+        /* Enable our renderer. */
+        draw_set_rasterize_stage(r300->draw, r300_draw_stage(r300));
+        /* Enable Draw's clipping. */
+        draw_set_driver_clipping(r300->draw, FALSE);
+        /* Force Draw to never do viewport transform, since we can do
+         * transform in hardware, always. */
+        draw_set_viewport_state(r300->draw, &r300_viewport_identity);
     }
 
     r300->context.is_texture_referenced = r300_is_texture_referenced;
@@ -143,18 +153,9 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
     r300->blend_color_state = CALLOC_STRUCT(r300_blend_color_state);
     r300->rs_block = CALLOC_STRUCT(r300_rs_block);
     r300->scissor_state = CALLOC_STRUCT(r300_scissor_state);
+    r300->vertex_info = CALLOC_STRUCT(r300_vertex_info);
     r300->viewport_state = CALLOC_STRUCT(r300_viewport_state);
 
-    /* Create a Draw. This is used for vert collation and SW TCL. */
-    r300->draw = draw_create();
-    /* Enable our renderer. */
-    draw_set_rasterize_stage(r300->draw, r300_draw_stage(r300));
-    /* Disable Draw's clipping if TCL is present. */
-    draw_set_driver_clipping(r300->draw, r300_screen(screen)->caps->has_tcl);
-    /* Force Draw to never do viewport transform, since (again) we can do
-     * transform in hardware, always. */
-    draw_set_viewport_state(r300->draw, &r300_viewport_identity);
-
     /* Open up the OQ BO. */
     r300->oqbo = screen->buffer_create(screen, 4096,
             PIPE_BUFFER_USAGE_VERTEX, 4096);
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index f954ba7f9a..39c0914cff 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -92,6 +92,10 @@ struct r300_sampler_state {
     uint32_t filter0;      /* R300_TX_FILTER0: 0x4400 */
     uint32_t filter1;      /* R300_TX_FILTER1: 0x4440 */
     uint32_t border_color; /* R300_TX_BORDER_COLOR: 0x45c0 */
+
+    /* Min/max LOD must be clamped to [0, last_level], thus
+     * it's dependent on a currently bound texture */
+    unsigned min_lod, max_lod;
 };
 
 struct r300_scissor_state {
@@ -219,11 +223,6 @@ struct r300_texture {
 struct r300_vertex_info {
     /* Parent class */
     struct vertex_info vinfo;
-    /* Map of vertex attributes into PVS memory for HW TCL,
-     * or GA memory for SW TCL. */
-    int vs_tab[16];
-    /* Map of rasterizer attributes from GB through RS to US. */
-    int fs_tab[16];
 
     /* R300_VAP_PROG_STREAK_CNTL_[0-7] */
     uint32_t vap_prog_stream_cntl[8];
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index eeb97a2d37..ad7dff36be 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -129,7 +129,9 @@ static const float * get_shader_constant(
     struct rc_constant * constant,
     struct r300_constant_buffer * externals)
 {
-    static const float zero[4] = { 0.0, 0.0, 0.0, 0.0 };
+    static float vec[4] = { 0.0, 0.0, 0.0, 1.0 };
+    struct pipe_texture *tex;
+
     switch(constant->Type) {
         case RC_CONSTANT_EXTERNAL:
             return externals->constants[constant->u.External];
@@ -137,11 +139,31 @@ static const float * get_shader_constant(
         case RC_CONSTANT_IMMEDIATE:
             return constant->u.Immediate;
 
+        case RC_CONSTANT_STATE:
+            switch (constant->u.State[0]) {
+                /* Factor for converting rectangle coords to
+                 * normalized coords. Should only show up on non-r500. */
+                case RC_STATE_R300_TEXRECT_FACTOR:
+                    tex = &r300->textures[constant->u.State[1]]->tex;
+                    vec[0] = 1.0 / tex->width[0];
+                    vec[1] = 1.0 / tex->height[0];
+                    break;
+
+                default:
+                    debug_printf("r300: Implementation error: "
+                        "Unknown RC_CONSTANT type %d\n", constant->u.State[0]);
+            }
+            break;
+
         default:
-            debug_printf("r300: Implementation error: Unhandled constant type %i\n",
-                constant->Type);
-            return zero;
+            debug_printf("r300: Implementation error: "
+                "Unhandled constant type %d\n", constant->Type);
     }
+
+    /* This should either be (0, 0, 0, 1), which should be a relatively safe
+     * RGBA or STRQ value, or it could be one of the RC_CONSTANT_STATE
+     * state factors. */
+    return vec;
 }
 
 /* Convert a normal single-precision float into the 7.16 format
@@ -561,6 +583,8 @@ void r300_emit_texture(struct r300_context* r300,
                        unsigned offset)
 {
     uint32_t filter0 = sampler->filter0;
+    uint32_t format0 = tex->state.format0;
+    unsigned min_level, max_level;
     CS_LOCALS(r300);
 
     /* to emulate 1D textures through 2D ones correctly */
@@ -569,13 +593,20 @@ void r300_emit_texture(struct r300_context* r300,
         filter0 |= R300_TX_WRAP_T(R300_TX_CLAMP_TO_EDGE);
     }
 
+    /* determine min/max levels */
+    /* the MAX_MIP level is the largest (finest) one */
+    max_level = MIN2(sampler->max_lod, tex->tex.last_level);
+    min_level = MIN2(sampler->min_lod, max_level);
+    format0 |= R300_TX_NUM_LEVELS(max_level);
+    filter0 |= R300_TX_MAX_MIP_LEVEL(min_level);
+
     BEGIN_CS(16);
     OUT_CS_REG(R300_TX_FILTER0_0 + (offset * 4), filter0 |
         (offset << 28));
     OUT_CS_REG(R300_TX_FILTER1_0 + (offset * 4), sampler->filter1);
     OUT_CS_REG(R300_TX_BORDER_COLOR_0 + (offset * 4), sampler->border_color);
 
-    OUT_CS_REG(R300_TX_FORMAT0_0 + (offset * 4), tex->state.format0);
+    OUT_CS_REG(R300_TX_FORMAT0_0 + (offset * 4), format0);
     OUT_CS_REG(R300_TX_FORMAT1_0 + (offset * 4), tex->state.format1);
     OUT_CS_REG(R300_TX_FORMAT2_0 + (offset * 4), tex->state.format2);
     OUT_CS_REG_SEQ(R300_TX_OFFSET_0 + (offset * 4), 1);
@@ -690,12 +721,37 @@ void r300_emit_vertex_format_state(struct r300_context* r300)
     END_CS;
 }
 
+/* XXX This should go to util ... */
+/* Return the number of bits set in the given number. */
+static unsigned bitcount(unsigned n)
+{
+    unsigned bits = 0;
+
+    while (n) {
+        if (n & 1) {
+            bits++;
+        }
+        n >>= 1;
+    }
+
+    return bits;
+}
+
 void r300_emit_vertex_program_code(struct r300_context* r300,
                                    struct r300_vertex_program_code* code)
 {
     int i;
     struct r300_screen* r300screen = r300_screen(r300->context.screen);
     unsigned instruction_count = code->length / 4;
+
+    int vtx_mem_size = r300screen->caps->is_r500 ? 128 : 72;
+    int input_count = MAX2(bitcount(code->InputsRead), 1);
+    int output_count = MAX2(bitcount(code->OutputsWritten), 1);
+    int temp_count = MAX2(code->num_temporaries, 1);
+    int pvs_num_slots = MIN3(vtx_mem_size / input_count,
+                             vtx_mem_size / output_count, 10);
+    int pvs_num_controllers = MIN2(vtx_mem_size / temp_count, 6);
+
     CS_LOCALS(r300);
 
     if (!r300screen->caps->has_tcl) {
@@ -708,8 +764,7 @@ void r300_emit_vertex_program_code(struct r300_context* r300,
     /* R300_VAP_PVS_CODE_CNTL_0
      * R300_VAP_PVS_CONST_CNTL
      * R300_VAP_PVS_CODE_CNTL_1
-     * See the r5xx docs for instructions on how to use these.
-     * XXX these could be optimized to select better values... */
+     * See the r5xx docs for instructions on how to use these. */
     OUT_CS_REG_SEQ(R300_VAP_PVS_CODE_CNTL_0, 3);
     OUT_CS(R300_PVS_FIRST_INST(0) |
             R300_PVS_XYZW_VALID_INST(instruction_count - 1) |
@@ -722,10 +777,11 @@ void r300_emit_vertex_program_code(struct r300_context* r300,
     for (i = 0; i < code->length; i++)
         OUT_CS(code->body.d[i]);
 
-    OUT_CS_REG(R300_VAP_CNTL, R300_PVS_NUM_SLOTS(10) |
-            R300_PVS_NUM_CNTLRS(5) |
+    OUT_CS_REG(R300_VAP_CNTL, R300_PVS_NUM_SLOTS(pvs_num_slots) |
+            R300_PVS_NUM_CNTLRS(pvs_num_controllers) |
             R300_PVS_NUM_FPUS(r300screen->caps->num_vert_fpus) |
-            R300_PVS_VF_MAX_VTX_NUM(12));
+            R300_PVS_VF_MAX_VTX_NUM(12) |
+            (r300screen->caps->is_r500 ? R500_TCL_STATE_OPTIMIZATION : 0));
     END_CS;
 }
 
@@ -790,13 +846,22 @@ void r300_emit_viewport_state(struct r300_context* r300,
     END_CS;
 }
 
+void r300_emit_texture_count(struct r300_context* r300)
+{
+    CS_LOCALS(r300);
+
+    BEGIN_CS(2);
+    OUT_CS_REG(R300_TX_ENABLE, (1 << r300->texture_count) - 1);
+    END_CS;
+
+}
+
 void r300_flush_textures(struct r300_context* r300)
 {
     CS_LOCALS(r300);
 
-    BEGIN_CS(4);
+    BEGIN_CS(2);
     OUT_CS_REG(R300_TX_INVALTAGS, 0);
-    OUT_CS_REG(R300_TX_ENABLE, (1 << r300->texture_count) - 1);
     END_CS;
 }
 
@@ -950,6 +1015,8 @@ validate:
     /* Samplers and textures are tracked separately but emitted together. */
     if (r300->dirty_state &
             (R300_ANY_NEW_SAMPLERS | R300_ANY_NEW_TEXTURES)) {
+        r300_emit_texture_count(r300);
+
         for (i = 0; i < MIN2(r300->sampler_count, r300->texture_count); i++) {
   	    if (r300->dirty_state &
 		((R300_NEW_SAMPLER << i) | (R300_NEW_TEXTURE << i))) {
diff --git a/src/gallium/drivers/r300/r300_emit.h b/src/gallium/drivers/r300/r300_emit.h
index 7c83c5166d..3797d3d332 100644
--- a/src/gallium/drivers/r300/r300_emit.h
+++ b/src/gallium/drivers/r300/r300_emit.h
@@ -92,6 +92,8 @@ void r300_emit_vertex_shader(struct r300_context* r300,
 void r300_emit_viewport_state(struct r300_context* r300,
                               struct r300_viewport_state* viewport);
 
+void r300_emit_texture_count(struct r300_context* r300);
+
 void r300_flush_textures(struct r300_context* r300);
 
 /* Emit all dirty state. */
diff --git a/src/gallium/drivers/r300/r300_reg.h b/src/gallium/drivers/r300/r300_reg.h
index 8ca785cb58..3a419b24b0 100644
--- a/src/gallium/drivers/r300/r300_reg.h
+++ b/src/gallium/drivers/r300/r300_reg.h
@@ -1293,7 +1293,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #        define R500_RS_INST_TEX_ID(x)                  ((x) << 0)
 #define R500_RS_INST_TEX_CN_WRITE			(1 << 4)
 #define R500_RS_INST_TEX_ADDR_SHIFT			5
-#        define R500_RS_INST_TEX_ADDR(x)                ((x) << 0)
+#        define R500_RS_INST_TEX_ADDR(x)                ((x) << 5)
 #define R500_RS_INST_COL_ID_SHIFT			12
 #        define R500_RS_INST_COL_ID(x)                  ((x) << 12)
 #define R500_RS_INST_COL_CN_NO_WRITE			(0 << 16)
@@ -1463,6 +1463,8 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #	define R300_TX_MIN_FILTER_MIP_NEAREST    (1 << 13)
 #	define R300_TX_MIN_FILTER_MIP_LINEAR     (2 << 13)
 #	define R300_TX_MIN_FILTER_MIP_MASK       (3 << 13)
+#       define R300_TX_MAX_MIP_LEVEL_SHIFT       17
+#       define R300_TX_MAX_MIP_LEVEL_MASK        (0xf << 17)
 #	define R300_TX_MAX_ANISO_1_TO_1          (0 << 21)
 #	define R300_TX_MAX_ANISO_2_TO_1          (1 << 21)
 #	define R300_TX_MAX_ANISO_4_TO_1          (2 << 21)
@@ -1471,6 +1473,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #	define R300_TX_MAX_ANISO_MASK            (7 << 21)
 #       define R300_TX_WRAP_S(x)                 ((x) << 0)
 #       define R300_TX_WRAP_T(x)                 ((x) << 3)
+#       define R300_TX_MAX_MIP_LEVEL(x)          ((x) << 17)
 
 #define R300_TX_FILTER1_0                      0x4440
 #	define R300_CHROMA_KEY_MODE_DISABLE    0
@@ -1500,8 +1503,6 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #       define R300_TX_HEIGHTMASK_MASK           (2047 << 11)
 #	define R300_TX_DEPTHMASK_SHIFT           22
 #	define R300_TX_DEPTHMASK_MASK            (0xf << 22)
-#       define R300_TX_MAX_MIP_LEVEL_SHIFT       26
-#       define R300_TX_MAX_MIP_LEVEL_MASK        (0xf << 26)
 #       define R300_TX_SIZE_PROJECTED            (1 << 30)
 #       define R300_TX_PITCH_EN                  (1 << 31)
 #       define R300_TX_WIDTH(x)                  ((x) << 0)
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index 62e1456ed3..4c5fb405c6 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -183,8 +183,6 @@ boolean r300_draw_range_elements(struct pipe_context* pipe,
         return FALSE;
     }
 
-    setup_vertex_attributes(r300);
-
     setup_index_buffer(r300, indexBuffer, indexSize);
 
     r300_emit_dirty_state(r300);
@@ -226,8 +224,6 @@ boolean r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
         return FALSE;
     }
 
-    setup_vertex_attributes(r300);
-
     r300_emit_dirty_state(r300);
 
     r300_emit_aos(r300, start);
diff --git a/src/gallium/drivers/r300/r300_screen.h b/src/gallium/drivers/r300/r300_screen.h
index 41df31f670..1ce5ff3904 100644
--- a/src/gallium/drivers/r300/r300_screen.h
+++ b/src/gallium/drivers/r300/r300_screen.h
@@ -27,6 +27,8 @@
 
 #include "r300_chipset.h"
 
+struct r300_winsys;
+
 struct r300_screen {
     /* Parent class */
     struct pipe_screen screen;
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index d1eced61db..a88d66db24 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -523,6 +523,11 @@ static void*
                                                    state->mag_img_filter,
                                                    state->min_mip_filter);
 
+    /* Unfortunately, r300-r500 don't support floating-point mipmap lods. */
+    /* We must pass these to the emit function to clamp them properly. */
+    sampler->min_lod = MAX2((unsigned)state->min_lod, 0);
+    sampler->max_lod = MAX2((unsigned)ceilf(state->max_lod), 0);
+
     lod_bias = CLAMP((int)(state->lod_bias * 32), -(1 << 9), (1 << 9) - 1);
 
     sampler->filter1 |= lod_bias << R300_LOD_BIAS_SHIFT;
@@ -571,6 +576,7 @@ static void r300_set_sampler_textures(struct pipe_context* pipe,
                                       struct pipe_texture** texture)
 {
     struct r300_context* r300 = r300_context(pipe);
+    boolean is_r500 = r300_screen(r300->context.screen)->caps->is_r500;
     int i;
 
     /* XXX magic num */
@@ -585,6 +591,13 @@ static void r300_set_sampler_textures(struct pipe_context* pipe,
             pipe_texture_reference((struct pipe_texture**)&r300->textures[i],
                 texture[i]);
             r300->dirty_state |= (R300_NEW_TEXTURE << i);
+
+            /* R300-specific - set the texrect factor in a fragment shader */
+            if (!is_r500 && r300->textures[i]->is_npot) {
+                /* XXX It would be nice to re-emit just 1 constant,
+                 * XXX not all of them */
+                r300->dirty_state |= R300_NEW_FRAGMENT_SHADER_CONSTANTS;
+            }
         }
     }
 
@@ -674,6 +687,8 @@ static void r300_set_vertex_buffers(struct pipe_context* pipe,
         draw_flush(r300->draw);
         draw_set_vertex_buffers(r300->draw, count, buffers);
     }
+
+    r300->dirty_state |= R300_NEW_VERTEX_FORMAT;
 }
 
 static void r300_set_vertex_elements(struct pipe_context* pipe,
@@ -706,9 +721,6 @@ static void* r300_create_vs_state(struct pipe_context* pipe,
 
         tgsi_scan_shader(shader->tokens, &vs->info);
 
-        /* Appease Draw. */
-        vs->draw = draw_create_vertex_shader(r300->draw, shader);
-
         return (void*)vs;
     } else {
         return draw_create_vertex_shader(r300->draw, shader);
@@ -719,8 +731,6 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
 {
     struct r300_context* r300 = r300_context(pipe);
 
-    draw_flush(r300->draw);
-
     if (r300_screen(pipe->screen)->caps->has_tcl) {
         struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader;
 
@@ -731,10 +741,10 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
             r300_translate_vertex_shader(r300, vs);
         }
 
-        draw_bind_vertex_shader(r300->draw, vs->draw);
         r300->vs = vs;
         r300->dirty_state |= R300_NEW_VERTEX_SHADER | R300_NEW_VERTEX_SHADER_CONSTANTS;
     } else {
+        draw_flush(r300->draw);
         draw_bind_vertex_shader(r300->draw,
                 (struct draw_vertex_shader*)shader);
     }
@@ -748,7 +758,6 @@ static void r300_delete_vs_state(struct pipe_context* pipe, void* shader)
         struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader;
 
         rc_constants_destroy(&vs->code.constants);
-        draw_delete_vertex_shader(r300->draw, vs->draw);
         FREE((void*)vs->state.tokens);
         FREE(shader);
     } else {
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index 7166694edf..6fb780cb29 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -47,8 +47,8 @@ struct r300_shader_derived_value {
 
 unsigned r300_shader_key_hash(void* key) {
     struct r300_shader_key* shader_key = (struct r300_shader_key*)key;
-    unsigned vs = (unsigned)shader_key->vs;
-    unsigned fs = (unsigned)shader_key->fs;
+    unsigned vs = (intptr_t)shader_key->vs;
+    unsigned fs = (intptr_t)shader_key->fs;
 
     return (vs << 16) | (fs & 0xffff);
 }
@@ -61,88 +61,46 @@ int r300_shader_key_compare(void* key1, void* key2) {
         (shader_key1->fs == shader_key2->fs);
 }
 
-/* Set up the vs_tab and routes. */
-static void r300_vs_tab_routes(struct r300_context* r300,
-                               struct r300_vertex_info* vformat)
+/* Set up the vs_output_tab and routes. */
+static void r300_vs_output_tab_routes(struct r300_context* r300,
+                                      int* vs_output_tab)
 {
-    struct r300_screen* r300screen = r300_screen(r300->context.screen);
-    struct vertex_info* vinfo = &vformat->vinfo;
-    int* tab = vformat->vs_tab;
+    struct vertex_info* vinfo = &r300->vertex_info->vinfo;
     boolean pos = FALSE, psize = FALSE, fog = FALSE;
     int i, texs = 0, cols = 0;
-    struct tgsi_shader_info* info;
-
-    if (r300screen->caps->has_tcl) {
-        /* Use vertex shader to determine required routes. */
-        info = &r300->vs->info;
-    } else {
-        /* Use fragment shader to determine required routes. */
-        info = &r300->fs->info;
-    }
+    struct tgsi_shader_info* info = &r300->fs->info;
 
-    assert(info->num_inputs <= 16);
+    /* XXX One day we should figure out how to handle a different number of
+     * VS outputs and FS inputs, as well as a different number of vertex streams
+     * and VS inputs. It's definitely one of the sources of hardlocks. */
 
-    if (!r300screen->caps->has_tcl || !r300->rs_state->enable_vte)
-    {
-        for (i = 0; i < info->num_inputs; i++) {
-            switch (r300->vs->code.inputs[i]) {
-                case TGSI_SEMANTIC_POSITION:
-                    pos = TRUE;
-                    tab[i] = 0;
-                    break;
-                case TGSI_SEMANTIC_COLOR:
-                    tab[i] = 2 + cols;
-                    cols++;
-                    break;
-                case TGSI_SEMANTIC_PSIZE:
-                    assert(psize == FALSE);
-                    psize = TRUE;
-                    tab[i] = 15;
-                    break;
-                case TGSI_SEMANTIC_FOG:
-                    assert(fog == FALSE);
-                    fog = TRUE;
-                    /* Fall through */
-                case TGSI_SEMANTIC_GENERIC:
-                    tab[i] = 6 + texs;
-                    texs++;
-                    break;
-                default:
-                    debug_printf("r300: Unknown vertex input %d\n",
-                        info->input_semantic_name[i]);
-                    break;
-            }
-        }
-    }
-    else
-    {
-        /* Just copy vert attribs over as-is. */
-        for (i = 0; i < info->num_inputs; i++) {
-            tab[i] = i;
-        }
-
-        for (i = 0; i < info->num_outputs; i++) {
-            switch (info->output_semantic_name[i]) {
-                case TGSI_SEMANTIC_POSITION:
-                    pos = TRUE;
-                    break;
-                case TGSI_SEMANTIC_COLOR:
-                    cols++;
-                    break;
-                case TGSI_SEMANTIC_PSIZE:
-                    psize = TRUE;
-                    break;
-                case TGSI_SEMANTIC_FOG:
-                    fog = TRUE;
-                    /* Fall through */
-                case TGSI_SEMANTIC_GENERIC:
-                    texs++;
-                    break;
-                default:
-                    debug_printf("r300: Unknown vertex output %d\n",
-                        info->output_semantic_name[i]);
-                    break;
-            }
+    for (i = 0; i < info->num_inputs; i++) {
+        switch (info->input_semantic_name[i]) {
+            case TGSI_SEMANTIC_POSITION:
+                pos = TRUE;
+                vs_output_tab[i] = 0;
+                break;
+            case TGSI_SEMANTIC_COLOR:
+                vs_output_tab[i] = 2 + cols;
+                cols++;
+                break;
+            case TGSI_SEMANTIC_PSIZE:
+                assert(psize == FALSE);
+                psize = TRUE;
+                vs_output_tab[i] = 15;
+                break;
+            case TGSI_SEMANTIC_FOG:
+                assert(fog == FALSE);
+                fog = TRUE;
+                /* Fall through */
+            case TGSI_SEMANTIC_GENERIC:
+                vs_output_tab[i] = 6 + texs;
+                texs++;
+                break;
+            default:
+                debug_printf("r300: Unknown vertex input %d\n",
+                    info->input_semantic_name[i]);
+                break;
         }
     }
 
@@ -161,14 +119,13 @@ static void r300_vs_tab_routes(struct r300_context* r300,
 
     /* We need to add vertex position attribute only for SW TCL case,
      * for HW TCL case it could be generated by vertex shader */
-    if (!pos && !r300screen->caps->has_tcl) {
-        debug_printf("r300: Forcing vertex position attribute emit...\n");
+    if (!pos) {
         /* Make room for the position attribute
-         * at the beginning of the tab. */
+         * at the beginning of the vs_output_tab. */
         for (i = 15; i > 0; i--) {
-            tab[i] = tab[i-1];
+            vs_output_tab[i] = vs_output_tab[i-1];
         }
-        tab[0] = 0;
+        vs_output_tab[0] = 0;
     }
 
     /* Position. */
@@ -227,43 +184,78 @@ static void r300_vs_tab_routes(struct r300_context* r300,
 }
 
 /* Update the PSC tables. */
-static void r300_vertex_psc(struct r300_context* r300,
-                            struct r300_vertex_info* vformat)
+static void r300_vertex_psc(struct r300_context* r300)
 {
-    struct r300_screen* r300screen = r300_screen(r300->context.screen);
-    struct vertex_info* vinfo = &vformat->vinfo;
-    int* tab = vformat->vs_tab;
+    struct r300_vertex_info *vformat = r300->vertex_info;
     uint16_t type, swizzle;
     enum pipe_format format;
-    unsigned i, attrib_count;
+    unsigned i;
 
     /* Vertex shaders have no semantics on their inputs,
-     * so PSC should just route stuff based on their info,
+     * so PSC should just route stuff based on the vertex elements,
      * and not on attrib information. */
-    if (r300screen->caps->has_tcl) {
-        attrib_count = r300->vs->info.num_inputs;
-        DBG(r300, DBG_DRAW, "r300: routing %d attribs in psc for vs\n",
-                attrib_count);
-    } else {
-        attrib_count = vinfo->num_attribs;
-        DBG(r300, DBG_DRAW, "r300: attrib count: %d\n", attrib_count);
-        for (i = 0; i < attrib_count; i++) {
-            DBG(r300, DBG_DRAW, "r300: attrib: offset %d, interp %d, size %d,"
-                   " tab %d\n", vinfo->attrib[i].src_index,
-                   vinfo->attrib[i].interp_mode, vinfo->attrib[i].emit,
-                   tab[i]);
+    DBG(r300, DBG_DRAW, "r300: vs expects %d attribs, routing %d elements"
+            " in psc\n",
+            r300->vs->info.num_inputs,
+            r300->vertex_element_count);
+
+    for (i = 0; i < r300->vertex_element_count; i++) {
+        format = r300->vertex_element[i].src_format;
+
+        type = r300_translate_vertex_data_type(format) |
+            (i << R300_DST_VEC_LOC_SHIFT);
+        swizzle = r300_translate_vertex_data_swizzle(format);
+
+        if (i % 2) {
+            vformat->vap_prog_stream_cntl[i >> 1] |= type << 16;
+            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
+        } else {
+            vformat->vap_prog_stream_cntl[i >> 1] |= type;
+            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
         }
     }
 
+
+    assert(i <= 15);
+
+    /* Set the last vector in the PSC. */
+    if (i) {
+        i -= 1;
+    }
+    vformat->vap_prog_stream_cntl[i >> 1] |=
+        (R300_LAST_VEC << (i & 1 ? 16 : 0));
+}
+
+/* Update the PSC tables for SW TCL, using Draw. */
+static void r300_swtcl_vertex_psc(struct r300_context* r300,
+                                  int* vs_output_tab)
+{
+    struct r300_vertex_info *vformat = r300->vertex_info;
+    struct vertex_info* vinfo = &vformat->vinfo;
+    uint16_t type, swizzle;
+    enum pipe_format format;
+    unsigned i, attrib_count;
+
+    /* For each Draw attribute, route it to the fragment shader according
+     * to the vs_output_tab. */
+    attrib_count = vinfo->num_attribs;
+    DBG(r300, DBG_DRAW, "r300: attrib count: %d\n", attrib_count);
+    for (i = 0; i < attrib_count; i++) {
+        DBG(r300, DBG_DRAW, "r300: attrib: offset %d, interp %d, size %d,"
+               " vs_output_tab %d\n", vinfo->attrib[i].src_index,
+               vinfo->attrib[i].interp_mode, vinfo->attrib[i].emit,
+               vs_output_tab[i]);
+    }
+
     for (i = 0; i < attrib_count; i++) {
         /* Make sure we have a proper destination for our attribute. */
-        assert(tab[i] != -1);
+        assert(vs_output_tab[i] != -1);
 
         format = draw_translate_vinfo_format(vinfo->attrib[i].emit);
 
         /* Obtain the type of data in this attribute. */
         type = r300_translate_vertex_data_type(format) |
-            tab[i] << R300_DST_VEC_LOC_SHIFT;
+            vs_output_tab[i] << R300_DST_VEC_LOC_SHIFT;
 
         /* Obtain the swizzle for this attribute. Note that the default
          * swizzle in the hardware is not XYZW! */
@@ -272,12 +264,10 @@ static void r300_vertex_psc(struct r300_context* r300,
         /* Add the attribute to the PSC table. */
         if (i & 1) {
             vformat->vap_prog_stream_cntl[i >> 1] |= type << 16;
-
             vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
         } else {
-            vformat->vap_prog_stream_cntl[i >> 1] |= type <<  0;
-
-            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 0;
+            vformat->vap_prog_stream_cntl[i >> 1] |= type;
+            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
         }
     }
 
@@ -289,74 +279,12 @@ static void r300_vertex_psc(struct r300_context* r300,
         (R300_LAST_VEC << (i & 1 ? 16 : 0));
 }
 
-/* Set up the mappings from GB to US, for RS block. */
-static void r300_update_fs_tab(struct r300_context* r300,
-                               struct r300_vertex_info* vformat)
-{
-    struct tgsi_shader_info* info = &r300->fs->info;
-    int i, cols = 0, texs = 0, cols_emitted = 0;
-    int* tab = vformat->fs_tab;
-
-    for (i = 0; i < 16; i++) {
-        tab[i] = -1;
-    }
-
-    assert(info->num_inputs <= 16);
-    for (i = 0; i < info->num_inputs; i++) {
-        switch (info->input_semantic_name[i]) {
-            case TGSI_SEMANTIC_COLOR:
-                tab[i] = INTERP_LINEAR;
-                cols++;
-                break;
-            case TGSI_SEMANTIC_POSITION:
-            case TGSI_SEMANTIC_PSIZE:
-                debug_printf("r300: Implementation error: Can't use "
-                        "pos attribs in fragshader yet!\n");
-                /* Pass through for now */
-            case TGSI_SEMANTIC_FOG:
-            case TGSI_SEMANTIC_GENERIC:
-                tab[i] = INTERP_PERSPECTIVE;
-                break;
-            default:
-                debug_printf("r300: Unknown vertex input %d\n",
-                    info->input_semantic_name[i]);
-                break;
-        }
-    }
-
-    /* Now that we know where everything is... */
-    DBG(r300, DBG_DRAW, "r300: fp input count: %d\n", info->num_inputs);
-    for (i = 0; i < info->num_inputs; i++) {
-        switch (tab[i]) {
-            case INTERP_LINEAR:
-                DBG(r300, DBG_DRAW, "r300: attrib: "
-                        "stack offset %d, color,    tab %d\n",
-                        i, cols_emitted);
-                tab[i] = cols_emitted;
-                cols_emitted++;
-                break;
-            case INTERP_PERSPECTIVE:
-                DBG(r300, DBG_DRAW, "r300: attrib: "
-                        "stack offset %d, texcoord, tab %d\n",
-                        i, cols + texs);
-                tab[i] = cols + texs;
-                texs++;
-                break;
-            case -1:
-                debug_printf("r300: Implementation error: Bad fp interp!\n");
-            default:
-                break;
-        }
-    }
-
-}
-
 /* Set up the RS block. This is the part of the chipset that actually does
  * the rasterization of vertices into fragments. This is also the part of the
  * chipset that locks up if any part of it is even slightly wrong. */
-static void r300_update_rs_block(struct r300_context* r300,
-                                 struct r300_rs_block* rs)
+static void r300_update_rs_block(struct r300_context* r300)
 {
+    struct r300_rs_block* rs = r300->rs_block;
     struct tgsi_shader_info* info = &r300->fs->info;
     int col_count = 0, fp_offset = 0, i, tex_count = 0;
     int rs_tex_comp = 0;
@@ -390,17 +318,18 @@ static void r300_update_rs_block(struct r300_context* r300,
             col_count++;
         }
 
+        for (i = 0; i < col_count; i++) {
+            rs->inst[i] |= R500_RS_INST_COL_ID(i) |
+                R500_RS_INST_COL_CN_WRITE | R500_RS_INST_COL_ADDR(fp_offset);
+            fp_offset++;
+        }
+
         for (i = 0; i < tex_count; i++) {
             rs->inst[i] |= R500_RS_INST_TEX_ID(i) |
                 R500_RS_INST_TEX_CN_WRITE | R500_RS_INST_TEX_ADDR(fp_offset);
             fp_offset++;
         }
 
-        for (i = 0; i < col_count; i++) {
-            rs->inst[i] |= R500_RS_INST_COL_ID(i) |
-                R500_RS_INST_COL_CN_WRITE | R500_RS_INST_COL_ADDR(fp_offset);
-            fp_offset++;
-        }
     } else {
         for (i = 0; i < info->num_inputs; i++) {
             switch (info->input_semantic_name[i]) {
@@ -425,8 +354,10 @@ static void r300_update_rs_block(struct r300_context* r300,
             }
         }
 
+        /* Rasterize at least one color, or bad things happen. */
         if (col_count == 0) {
             rs->ip[0] |= R300_RS_COL_FMT(R300_RS_COL_FMT_0001);
+            col_count++;
         }
 
         if (tex_count == 0) {
@@ -437,9 +368,10 @@ static void r300_update_rs_block(struct r300_context* r300,
                 R300_RS_SEL_Q(R300_RS_SEL_K1);
         }
 
-        /* Rasterize at least one color, or bad things happen. */
-        if ((col_count == 0) && (tex_count == 0)) {
-            col_count++;
+        for (i = 0; i < col_count; i++) {
+            rs->inst[i] |= R300_RS_INST_COL_ID(i) |
+                R300_RS_INST_COL_CN_WRITE | R300_RS_INST_COL_ADDR(fp_offset);
+            fp_offset++;
         }
 
         for (i = 0; i < tex_count; i++) {
@@ -447,28 +379,22 @@ static void r300_update_rs_block(struct r300_context* r300,
                 R300_RS_INST_TEX_CN_WRITE | R300_RS_INST_TEX_ADDR(fp_offset);
             fp_offset++;
         }
-
-        for (i = 0; i < col_count; i++) {
-            rs->inst[i] |= R300_RS_INST_COL_ID(i) |
-                R300_RS_INST_COL_CN_WRITE | R300_RS_INST_COL_ADDR(fp_offset);
-            fp_offset++;
-        }
     }
 
     rs->count = (rs_tex_comp) | (col_count << R300_IC_COUNT_SHIFT) |
         R300_HIRES_EN;
 
-    rs->inst_count = MAX2(MAX2(col_count - 1, tex_count - 1), 0);
+    rs->inst_count = MAX3(col_count - 1, tex_count - 1, 0);
 }
 
 /* Update the vertex format. */
 static void r300_update_derived_shader_state(struct r300_context* r300)
 {
     struct r300_screen* r300screen = r300_screen(r300->context.screen);
-    struct r300_vertex_info* vformat;
-    struct r300_rs_block* rs_block;
+    int vs_output_tab[16];
     int i;
 
+
     /*
     struct r300_shader_key* key;
     struct r300_shader_derived_value* value;
@@ -495,27 +421,26 @@ static void r300_update_derived_shader_state(struct r300_context* r300)
             (void*)key, (void*)value);
     } */
 
-    /* XXX This will be refactored ASAP. */
-    vformat = CALLOC_STRUCT(r300_vertex_info);
-    rs_block = CALLOC_STRUCT(r300_rs_block);
+    /* Reset structures */
+    memset(r300->rs_block, 0, sizeof(struct r300_rs_block));
+    memset(r300->vertex_info, 0, sizeof(struct r300_vertex_info));
 
     for (i = 0; i < 16; i++) {
-        vformat->vs_tab[i] = -1;
-        vformat->fs_tab[i] = -1;
+        vs_output_tab[i] = -1;
     }
 
-    r300_vs_tab_routes(r300, vformat);
-    r300_vertex_psc(r300, vformat);
-    r300_update_fs_tab(r300, vformat);
+    /* Update states */
+    r300_vs_output_tab_routes(r300, vs_output_tab);
 
-    r300_update_rs_block(r300, rs_block);
+    if (r300screen->caps->has_tcl) {
+        r300_vertex_psc(r300);
+    } else {
+        r300_swtcl_vertex_psc(r300, vs_output_tab);
+    }
 
-    FREE(r300->vertex_info);
-    FREE(r300->rs_block);
+    r300_update_rs_block(r300);
 
-    r300->vertex_info = vformat;
-    r300->rs_block = rs_block;
-    r300->dirty_state |= (R300_NEW_VERTEX_FORMAT | R300_NEW_RS_BLOCK);
+    r300->dirty_state |= R300_NEW_RS_BLOCK;
 }
 
 static void r300_update_ztop(struct r300_context* r300)
@@ -553,9 +478,9 @@ static void r300_update_ztop(struct r300_context* r300)
 
 void r300_update_derived_state(struct r300_context* r300)
 {
-    /* XXX */
-    if (TRUE || r300->dirty_state &
-        (R300_NEW_FRAGMENT_SHADER | R300_NEW_VERTEX_SHADER)) {
+    if (r300->dirty_state &
+        (R300_NEW_FRAGMENT_SHADER | R300_NEW_VERTEX_SHADER |
+         R300_NEW_VERTEX_FORMAT)) {
         r300_update_derived_shader_state(r300);
     }
 
diff --git a/src/gallium/drivers/r300/r300_state_invariant.c b/src/gallium/drivers/r300/r300_state_invariant.c
index c07e6ae676..46d1cb39b5 100644
--- a/src/gallium/drivers/r300/r300_state_invariant.c
+++ b/src/gallium/drivers/r300/r300_state_invariant.c
@@ -84,7 +84,7 @@ void r300_emit_invariant_state(struct r300_context* r300)
     END_CS;
 
     /* XXX unsorted stuff from surface_fill */
-    BEGIN_CS(60 + (caps->has_tcl ? 5 : 0) + (caps->is_r500 ? 4 : 0));
+    BEGIN_CS(56 + (caps->has_tcl ? 5 : 0) + (caps->is_r500 ? 4 : 0));
     /* Flush PVS. */
     OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x0);
 
@@ -135,8 +135,6 @@ void r300_emit_invariant_state(struct r300_context* r300)
     OUT_CS_REG(R300_ZB_DEPTHCLEARVALUE, 0x00000000);
     OUT_CS_REG(R300_ZB_HIZ_OFFSET, 0x00000000);
     OUT_CS_REG(R300_ZB_HIZ_PITCH, 0x00000000);
-    OUT_CS_REG(R300_VAP_VTX_STATE_CNTL, 0x1);
-    OUT_CS_REG(R300_VAP_VSM_VTX_ASSM, 0x405);
     OUT_CS_REG(R300_SE_VTE_CNTL, 0x0000043F);
 
     /* XXX */
diff --git a/src/gallium/drivers/r300/r300_texture.c b/src/gallium/drivers/r300/r300_texture.c
index aea25cf71d..d13aa8f036 100644
--- a/src/gallium/drivers/r300/r300_texture.c
+++ b/src/gallium/drivers/r300/r300_texture.c
@@ -43,8 +43,7 @@ static void r300_setup_texture_state(struct r300_texture* tex, boolean is_r500)
         state->format2 = (tex->pitch[0] - 1) & 0x1fff;
     } else {
         /* power of two textures (3D, mipmaps, and no pitch) */
-        state->format0 |= R300_TX_DEPTH(util_logbase2(pt->depth[0]) & 0xf) |
-                          R300_TX_NUM_LEVELS(pt->last_level & 0xf);
+        state->format0 |= R300_TX_DEPTH(util_logbase2(pt->depth[0]) & 0xf);
     }
 
     state->format1 = r300_translate_texformat(pt->format);
diff --git a/src/gallium/drivers/r300/r300_vbo.c b/src/gallium/drivers/r300/r300_vbo.c
index a6a159667a..6ebaf715dc 100644
--- a/src/gallium/drivers/r300/r300_vbo.c
+++ b/src/gallium/drivers/r300/r300_vbo.c
@@ -34,52 +34,6 @@
 #include "r300_reg.h"
 #include "r300_winsys.h"
 
-static INLINE void setup_vertex_attribute(struct r300_vertex_info *vinfo,
-                                          struct pipe_vertex_element *vert_elem,
-                                          unsigned attr_num)
-{
-    uint16_t hw_fmt1, hw_fmt2;
-
-    hw_fmt1 = r300_translate_vertex_data_type(vert_elem->src_format) |
-        (attr_num << R300_DST_VEC_LOC_SHIFT);
-    hw_fmt2 = r300_translate_vertex_data_swizzle(vert_elem->src_format);
-
-    if (attr_num % 2 == 0)
-    {
-        vinfo->vap_prog_stream_cntl[attr_num >> 1] = hw_fmt1;
-        vinfo->vap_prog_stream_cntl_ext[attr_num >> 1] = hw_fmt2;
-    }
-    else
-    {
-        vinfo->vap_prog_stream_cntl[attr_num >> 1] |= hw_fmt1 << 16;
-        vinfo->vap_prog_stream_cntl_ext[attr_num >> 1] |= hw_fmt2 << 16;
-    }
-}
-
-static void finish_vertex_attribs_setup(struct r300_vertex_info *vinfo,
-                                        unsigned attribs_num)
-{
-    uint32_t last_vec_bit = (attribs_num % 2 == 0) ?
-        (R300_LAST_VEC << 16) : R300_LAST_VEC;
-
-    assert(attribs_num > 0 && attribs_num <= 16);
-    vinfo->vap_prog_stream_cntl[(attribs_num - 1) >> 1] |= last_vec_bit;
-}
-
-void setup_vertex_attributes(struct r300_context *r300)
-{
-    struct pipe_vertex_element *vert_elem;
-    int i;
-
-    for (i = 0; i < r300->vertex_element_count; i++) {
-        vert_elem = &r300->vertex_element[i];
-        setup_vertex_attribute(r300->vertex_info, vert_elem, i);
-    }
-
-    finish_vertex_attribs_setup(r300->vertex_info,
-        r300->vertex_element_count);
-}
-
 static INLINE int get_buffer_offset(struct r300_context *r300,
                                     unsigned int buf_nr,
                                     unsigned int elem_offset)
diff --git a/src/gallium/drivers/r300/r300_vs.h b/src/gallium/drivers/r300/r300_vs.h
index 2a4ce315e3..00b02bf510 100644
--- a/src/gallium/drivers/r300/r300_vs.h
+++ b/src/gallium/drivers/r300/r300_vs.h
@@ -35,9 +35,6 @@ struct r300_vertex_shader {
     struct pipe_shader_state state;
     struct tgsi_shader_info info;
 
-    /* Fallback shader, because Draw has issues */
-    struct draw_vertex_shader* draw;
-
     /* Has this shader been translated yet? */
     boolean translated;
 
diff --git a/src/gallium/drivers/softpipe/sp_context.c b/src/gallium/drivers/softpipe/sp_context.c
index 5f60139968..bdbb7fa9b9 100644
--- a/src/gallium/drivers/softpipe/sp_context.c
+++ b/src/gallium/drivers/softpipe/sp_context.c
@@ -90,14 +90,15 @@ softpipe_destroy( struct pipe_context *pipe )
    if (softpipe->draw)
       draw_destroy( softpipe->draw );
 
-      softpipe->quad.shade->destroy( softpipe->quad.shade );
-      softpipe->quad.depth_test->destroy( softpipe->quad.depth_test );
-      softpipe->quad.blend->destroy( softpipe->quad.blend );
+   softpipe->quad.shade->destroy( softpipe->quad.shade );
+   softpipe->quad.depth_test->destroy( softpipe->quad.depth_test );
+   softpipe->quad.blend->destroy( softpipe->quad.blend );
 
    for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) {
       sp_destroy_tile_cache(softpipe->cbuf_cache[i]);
       pipe_surface_reference(&softpipe->framebuffer.cbufs[i], NULL);
    }
+
    sp_destroy_tile_cache(softpipe->zsbuf_cache);
    pipe_surface_reference(&softpipe->framebuffer.zsbuf, NULL);
 
diff --git a/src/gallium/drivers/softpipe/sp_fs_exec.c b/src/gallium/drivers/softpipe/sp_fs_exec.c
index 4076114d39..a8999ed347 100644
--- a/src/gallium/drivers/softpipe/sp_fs_exec.c
+++ b/src/gallium/drivers/softpipe/sp_fs_exec.c
@@ -126,7 +126,13 @@ exec_run( const struct sp_fragment_shader *base,
    setup_pos_vector(quad->posCoef, 
                     (float)quad->input.x0, (float)quad->input.y0, 
                     &machine->QuadPos);
-   
+
+   if (quad->input.facing) {
+      machine->Face = -1.0f;
+   } else {
+      machine->Face = 1.0f;
+   }
+
    quad->inout.mask &= tgsi_exec_machine_run( machine );
    if (quad->inout.mask == 0)
       return FALSE;