13 files changed, 157 insertions, 296 deletions

diff --git a/src/gallium/drivers/r300/r300_blit.c b/src/gallium/drivers/r300/r300_blit.c
index 928ad300eee..85c2c149016 100644
--- a/src/gallium/drivers/r300/r300_blit.c
+++ b/src/gallium/drivers/r300/r300_blit.c
@@ -129,15 +129,16 @@ void r300_surface_copy(struct pipe_context* pipe,
     if (dst->texture->format != src->texture->format) {
         debug_printf("r300: Implementation error: Format mismatch in %s\n"
             "    : src: %s dst: %s\n", __FUNCTION__,
-            util_format_name(src->texture->format),
-            util_format_name(dst->texture->format));
+            util_format_short_name(src->texture->format),
+            util_format_short_name(dst->texture->format));
         debug_assert(0);
     }
 
     if (!pipe->screen->is_format_supported(pipe->screen,
                                            old_format, src->texture->target,
                                            PIPE_BIND_RENDER_TARGET |
-                                           PIPE_BIND_SAMPLER_VIEW, 0)) {
+                                           PIPE_BIND_SAMPLER_VIEW, 0) &&
+        util_format_is_plain(old_format)) {
         switch (util_format_get_blocksize(old_format)) {
             case 1:
                 new_format = PIPE_FORMAT_I8_UNORM;
@@ -154,7 +155,7 @@ void r300_surface_copy(struct pipe_context* pipe,
             default:
                 debug_printf("r300: surface_copy: Unhandled format: %s. Falling back to software.\n"
                              "r300: surface_copy: Software fallback doesn't work for tiled textures.\n",
-                             util_format_name(old_format));
+                             util_format_short_name(old_format));
         }
     }
 
diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
index 4f721ebb599..e84bce0010f 100644
--- a/src/gallium/drivers/r300/r300_context.c
+++ b/src/gallium/drivers/r300/r300_context.c
@@ -76,7 +76,6 @@ static void r300_destroy_context(struct pipe_context* context)
     FREE(r300->rs_block_state.state);
     FREE(r300->scissor_state.state);
     FREE(r300->textures_state.state);
-    FREE(r300->vap_output_state.state);
     FREE(r300->viewport_state.state);
     FREE(r300->ztop_state.state);
     FREE(r300->fs_constants.state);
@@ -125,7 +124,6 @@ static void r300_setup_atoms(struct r300_context* r300)
     R300_INIT_ATOM(viewport_state, 9);
     R300_INIT_ATOM(rs_block_state, 0);
     R300_INIT_ATOM(vertex_stream_state, 0);
-    R300_INIT_ATOM(vap_output_state, 6);
     R300_INIT_ATOM(pvs_flush, 2);
     R300_INIT_ATOM(vs_state, 0);
     R300_INIT_ATOM(vs_constants, 0);
@@ -149,7 +147,6 @@ static void r300_setup_atoms(struct r300_context* r300)
     r300->rs_block_state.state = CALLOC_STRUCT(r300_rs_block);
     r300->scissor_state.state = CALLOC_STRUCT(pipe_scissor_state);
     r300->textures_state.state = CALLOC_STRUCT(r300_textures_state);
-    r300->vap_output_state.state = CALLOC_STRUCT(r300_vap_output_state);
     r300->viewport_state.state = CALLOC_STRUCT(r300_viewport_state);
     r300->ztop_state.state = CALLOC_STRUCT(r300_ztop_state);
     r300->fs_constants.state = CALLOC_STRUCT(r300_constant_buffer);
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index a05bf3ce099..e9c8fcdc157 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -119,6 +119,10 @@ struct r300_rs_state {
 };
 
 struct r300_rs_block {
+    uint32_t vap_vtx_state_cntl;  /* R300_VAP_VTX_STATE_CNTL: 0x2180 */
+    uint32_t vap_vsm_vtx_assm;    /* R300_VAP_VSM_VTX_ASSM: 0x2184 */
+    uint32_t vap_out_vtx_fmt[2];  /* R300_VAP_OUTPUT_VTX_FMT_[0-1]: 0x2090 */
+
     uint32_t ip[8]; /* R300_RS_IP_[0-7], R500_RS_IP_[0-7] */
     uint32_t count; /* R300_RS_COUNT */
     uint32_t inst_count; /* R300_RS_INST_COUNT */
@@ -190,12 +194,6 @@ struct r300_vertex_stream_state {
     unsigned count;
 };
 
-struct r300_vap_output_state {
-    uint32_t vap_vtx_state_cntl;  /* R300_VAP_VTX_STATE_CNTL: 0x2180 */
-    uint32_t vap_vsm_vtx_assm;    /* R300_VAP_VSM_VTX_ASSM: 0x2184 */
-    uint32_t vap_out_vtx_fmt[2];  /* R300_VAP_OUTPUT_VTX_FMT_[0-1]: 0x2090 */
-};
-
 struct r300_viewport_state {
     float xscale;         /* R300_VAP_VPORT_XSCALE:  0x2098 */
     float xoffset;        /* R300_VAP_VPORT_XOFFSET: 0x209c */
@@ -381,7 +379,7 @@ struct r300_context {
     struct r300_atom query_start;
     /* Rasterizer state. */
     struct r300_atom rs_state;
-    /* RS block state. */
+    /* RS block state + VAP (vertex shader) output mapping state. */
     struct r300_atom rs_block_state;
     /* Scissor state. */
     struct r300_atom scissor_state;
@@ -389,8 +387,6 @@ struct r300_context {
     struct r300_atom textures_state;
     /* Vertex stream formatting state. */
     struct r300_atom vertex_stream_state;
-    /* VAP (vertex shader) output mapping state. */
-    struct r300_atom vap_output_state;
     /* Vertex shader. */
     struct r300_atom vs_state;
     /* Vertex shader constant buffer. */
@@ -424,6 +420,9 @@ struct r300_context {
 
     struct pipe_viewport_state viewport;
 
+    /* Stream locations for SWTCL. */
+    int stream_loc_notcl[16];
+
     /* Flag indicating whether or not the HW is dirty. */
     uint32_t dirty_hw;
     /* Whether polygon offset is enabled. */
diff --git a/src/gallium/drivers/r300/r300_debug.c b/src/gallium/drivers/r300/r300_debug.c
index 4c2836f36a8..8eb321fa08a 100644
--- a/src/gallium/drivers/r300/r300_debug.c
+++ b/src/gallium/drivers/r300/r300_debug.c
@@ -37,6 +37,7 @@ static struct debug_option debug_options[] = {
     { "cs", DBG_CS, "Command submissions (for debugging)" },
     { "draw", DBG_DRAW, "Draw and emit (for debugging)" },
     { "tex", DBG_TEX, "Textures (for debugging)" },
+    { "texalloc", DBG_TEXALLOC, "Texture allocation (for debugging)" },
     { "fall", DBG_FALL, "Fallbacks (for debugging)" },
     { "rs", DBG_RS, "Rasterizer (for debugging)" },
     { "anisohq", DBG_ANISOHQ, "High quality anisotropic filtering (for benchmarking)" },
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index 92b7517b8d6..23bbc6a99c8 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -717,6 +717,13 @@ void r300_emit_rs_block_state(struct r300_context* r300,
     DBG(r300, DBG_DRAW, "r300: RS emit:\n");
 
     BEGIN_CS(size);
+    OUT_CS_REG_SEQ(R300_VAP_VTX_STATE_CNTL, 2);
+    OUT_CS(rs->vap_vtx_state_cntl);
+    OUT_CS(rs->vap_vsm_vtx_assm);
+    OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
+    OUT_CS(rs->vap_out_vtx_fmt[0]);
+    OUT_CS(rs->vap_out_vtx_fmt[1]);
+
     if (r300->screen->caps.is_r500) {
         OUT_CS_REG_SEQ(R500_RS_IP_0, count);
     } else {
@@ -901,25 +908,6 @@ void r300_emit_vertex_stream_state(struct r300_context* r300,
     END_CS;
 }
 
-void r300_emit_vap_output_state(struct r300_context* r300,
-                               unsigned size, void* state)
-{
-    struct r300_vap_output_state *vap_out_state =
-        (struct r300_vap_output_state*)state;
-    CS_LOCALS(r300);
-
-    DBG(r300, DBG_DRAW, "r300: VAP emit:\n");
-
-    BEGIN_CS(size);
-    OUT_CS_REG_SEQ(R300_VAP_VTX_STATE_CNTL, 2);
-    OUT_CS(vap_out_state->vap_vtx_state_cntl);
-    OUT_CS(vap_out_state->vap_vsm_vtx_assm);
-    OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
-    OUT_CS(vap_out_state->vap_out_vtx_fmt[0]);
-    OUT_CS(vap_out_state->vap_out_vtx_fmt[1]);
-    END_CS;
-}
-
 void r300_emit_pvs_flush(struct r300_context* r300, unsigned size, void* state)
 {
     CS_LOCALS(r300);
diff --git a/src/gallium/drivers/r300/r300_emit.h b/src/gallium/drivers/r300/r300_emit.h
index c3eb195d4e7..3c0edf6fdca 100644
--- a/src/gallium/drivers/r300/r300_emit.h
+++ b/src/gallium/drivers/r300/r300_emit.h
@@ -81,9 +81,6 @@ void r300_emit_vertex_buffer(struct r300_context* r300);
 void r300_emit_vertex_stream_state(struct r300_context* r300,
                                    unsigned size, void* state);
 
-void r300_emit_vap_output_state(struct r300_context* r300,
-                               unsigned size, void* state);
-
 void r300_emit_vs_constants(struct r300_context* r300,
                             unsigned size, void *state);
 
diff --git a/src/gallium/drivers/r300/r300_screen.h b/src/gallium/drivers/r300/r300_screen.h
index 735c233c9e2..d58aa138a70 100644
--- a/src/gallium/drivers/r300/r300_screen.h
+++ b/src/gallium/drivers/r300/r300_screen.h
@@ -73,6 +73,7 @@ static INLINE struct r300_screen* r300_screen(struct pipe_screen* screen) {
 #define DBG_NO_IMMD 0x0000200
 #define DBG_STATS   0x0000400
 #define DBG_RS      0x0000800
+#define DBG_TEXALLOC 0x0001000
 /*@}*/
 
 static INLINE boolean SCREEN_DBG_ON(struct r300_screen * screen, unsigned flags)
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index 9319dadfd1e..446422ca0f0 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -538,46 +538,12 @@ static void r300_set_stencil_ref(struct pipe_context* pipe,
 }
 
 /* This switcheroo is needed just because of goddamned MACRO_SWITCH. */
-static void r300_fb_update_tiling_flags(struct r300_context *r300,
+static void r300_fb_set_tiling_flags(struct r300_context *r300,
                                const struct pipe_framebuffer_state *old_state,
                                const struct pipe_framebuffer_state *new_state)
 {
     struct r300_texture *tex;
-    unsigned i, j, level;
-
-    /* Reset tiling flags for old surfaces to default values. */
-    for (i = 0; i < old_state->nr_cbufs; i++) {
-        for (j = 0; j < new_state->nr_cbufs; j++) {
-            if (old_state->cbufs[i]->texture == new_state->cbufs[j]->texture) {
-                break;
-            }
-        }
-        /* If not binding the surface again... */
-        if (j != new_state->nr_cbufs) {
-            continue;
-        }
-
-        tex = r300_texture(old_state->cbufs[i]->texture);
-
-        if (tex) {
-            r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
-                                            tex->pitch[0],
-                                            tex->microtile,
-                                            tex->macrotile);
-        }
-    }
-    if (old_state->zsbuf &&
-        (!new_state->zsbuf ||
-         old_state->zsbuf->texture != new_state->zsbuf->texture)) {
-        tex = r300_texture(old_state->zsbuf->texture);
-
-        if (tex) {
-            r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
-                                            tex->pitch[0],
-                                            tex->microtile,
-                                            tex->macrotile);
-        }
-    }
+    unsigned i, level;
 
     /* Set tiling flags for new surfaces. */
     for (i = 0; i < new_state->nr_cbufs; i++) {
@@ -585,7 +551,7 @@ static void r300_fb_update_tiling_flags(struct r300_context *r300,
         level = new_state->cbufs[i]->level;
 
         r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
-                                        tex->pitch[level],
+                                        tex->pitch[0],
                                         tex->microtile,
                                         tex->mip_macrotile[level]);
     }
@@ -594,7 +560,7 @@ static void r300_fb_update_tiling_flags(struct r300_context *r300,
         level = new_state->zsbuf->level;
 
         r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
-                                        tex->pitch[level],
+                                        tex->pitch[0],
                                         tex->microtile,
                                         tex->mip_macrotile[level]);
     }
@@ -644,7 +610,8 @@ static void
         r300->dsa_state.dirty = TRUE;
     }
 
-    r300_fb_update_tiling_flags(r300, r300->fb_state.state, state);
+    /* The tiling flags are dependent on the surface miplevel, unfortunately. */
+    r300_fb_set_tiling_flags(r300, r300->fb_state.state, state);
 
     memcpy(r300->fb_state.state, state, sizeof(struct pipe_framebuffer_state));
 
@@ -719,10 +686,6 @@ static void r300_bind_fs_state(struct pipe_context* pipe, void* shader)
     r300_mark_fs_code_dirty(r300);
 
     r300->rs_block_state.dirty = TRUE; /* Will be updated before the emission. */
-
-    if (r300->vs_state.state && r300_vertex_shader_setup_wpos(r300)) {
-        r300->vap_output_state.dirty = TRUE;
-    }
 }
 
 /* Delete fragment shader state. */
@@ -853,11 +816,8 @@ static void* r300_create_rs_state(struct pipe_context* pipe,
 
     rs->clip_rule = state->scissor ? 0xAAAA : 0xFFFF;
 
-    /* XXX Disable point sprites until we know what's wrong with them. */
-    rs->rs.sprite_coord_enable = 0;
-
     /* Point sprites */
-    if (rs->rs.sprite_coord_enable) {
+    if (state->sprite_coord_enable) {
         rs->stuffing_enable = R300_GB_POINT_STUFF_ENABLE;
 	for (i = 0; i < 8; i++) {
 	    if (state->sprite_coord_enable & (1 << i))
@@ -1357,7 +1317,7 @@ static void* r300_create_vertex_elements_state(struct pipe_context* pipe,
                     /* XXX Shouldn't we align the format? */
                     fprintf(stderr, "r300_create_vertex_elements_state: "
                             "Unaligned format %s:%i isn't supported\n",
-                            util_format_name(*format), size);
+                            util_format_short_name(*format), size);
                     assert(0);
                     abort();
                 }
@@ -1428,14 +1388,6 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
     }
     r300->vs_state.state = vs;
 
-    // VS output mapping for HWTCL or stream mapping for SWTCL to the RS block
-    if (r300->fs.state) {
-        r300_vertex_shader_setup_wpos(r300);
-    }
-    memcpy(r300->vap_output_state.state, &vs->vap_out,
-           sizeof(struct r300_vap_output_state));
-    r300->vap_output_state.dirty = TRUE;
-
     /* The majority of the RS block bits is dependent on the vertex shader. */
     r300->rs_block_state.dirty = TRUE; /* Will be updated before the emission. */
 
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index ccc4b583a6a..e3adace0faa 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -115,12 +115,11 @@ static void r300_draw_emit_all_attribs(struct r300_context* r300)
 static void r300_swtcl_vertex_psc(struct r300_context *r300)
 {
     struct r300_vertex_stream_state *vstream = r300->vertex_stream_state.state;
-    struct r300_vertex_shader* vs = r300->vs_state.state;
-    struct vertex_info* vinfo = &r300->vertex_info;
+        struct vertex_info* vinfo = &r300->vertex_info;
     uint16_t type, swizzle;
     enum pipe_format format;
     unsigned i, attrib_count;
-    int* vs_output_tab = vs->stream_loc_notcl;
+    int* vs_output_tab = r300->stream_loc_notcl;
 
     /* XXX hax */
     memset(vstream, 0, sizeof(struct r300_vertex_stream_state));
@@ -269,21 +268,29 @@ static void r500_rs_tex_write(struct r300_rs_block* rs, int id, int fp_offset)
 
 /* 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_shader_semantics* vs_outputs,
-                                 struct r300_shader_semantics* fs_inputs)
+ * This is the part of the chipset that is responsible for linking vertex
+ * and fragment shaders and stuffed texture coordinates.
+ *
+ * The rasterizer reads data from VAP, which produces vertex shader outputs,
+ * and GA, which produces stuffed texture coordinates. VAP outputs have
+ * precedence over GA. All outputs must be rasterized otherwise it locks up.
+ * If there are more outputs rasterized than is set in VAP/GA, it locks up
+ * too. The funky part is that this info has been pretty much obtained by trial
+ * and error. */
+static void r300_update_rs_block(struct r300_context *r300)
 {
-    struct r300_rs_block rs = { { 0 } };
-    int i, col_count = 0, tex_count = 0, fp_offset = 0, count;
+    struct r300_vertex_shader *vs = r300->vs_state.state;
+    struct r300_shader_semantics *vs_outputs = &vs->outputs;
+    struct r300_shader_semantics *fs_inputs = &r300_fs(r300)->shader->inputs;
+    struct r300_rs_block rs = {0};
+    int i, col_count = 0, tex_count = 0, fp_offset = 0, count, loc = 0;
     void (*rX00_rs_col)(struct r300_rs_block*, int, int, enum r300_rs_swizzle);
     void (*rX00_rs_col_write)(struct r300_rs_block*, int, int);
     void (*rX00_rs_tex)(struct r300_rs_block*, int, int, enum r300_rs_swizzle);
     void (*rX00_rs_tex_write)(struct r300_rs_block*, int, int);
     boolean any_bcolor_used = vs_outputs->bcolor[0] != ATTR_UNUSED ||
                               vs_outputs->bcolor[1] != ATTR_UNUSED;
+    int *stream_loc_notcl = r300->stream_loc_notcl;
 
     if (r300->screen->caps.is_r500) {
         rX00_rs_col       = r500_rs_col;
@@ -297,15 +304,31 @@ static void r300_update_rs_block(struct r300_context* r300,
         rX00_rs_tex_write = r300_rs_tex_write;
     }
 
-    /* Rasterize colors. */
+    /* The position is always present in VAP. */
+    rs.vap_vsm_vtx_assm |= R300_INPUT_CNTL_POS;
+    rs.vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT;
+    stream_loc_notcl[loc++] = 0;
+
+    /* Set up the point size in VAP. */
+    if (vs_outputs->psize != ATTR_UNUSED) {
+        rs.vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__PT_SIZE_PRESENT;
+        stream_loc_notcl[loc++] = 1;
+    }
+
+    /* Set up and rasterize colors. */
     for (i = 0; i < ATTR_COLOR_COUNT; i++) {
         if (vs_outputs->color[i] != ATTR_UNUSED || any_bcolor_used ||
             vs_outputs->color[1] != ATTR_UNUSED) {
-            /* Always rasterize if it's written by the VS,
-             * otherwise it locks up. */
+            /* Set up the color in VAP. */
+            rs.vap_vsm_vtx_assm |= R300_INPUT_CNTL_COLOR;
+            rs.vap_out_vtx_fmt[0] |=
+                    R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << i;
+            stream_loc_notcl[loc++] = 2 + i;
+
+            /* Rasterize it. */
             rX00_rs_col(&rs, col_count, col_count, SWIZ_XYZW);
 
-            /* Write it to the FS input register if it's used by the FS. */
+            /* Write it to the FS input register if it's needed by the FS. */
             if (fs_inputs->color[i] != ATTR_UNUSED) {
                 rX00_rs_col_write(&rs, col_count, fp_offset);
                 fp_offset++;
@@ -328,17 +351,33 @@ static void r300_update_rs_block(struct r300_context* r300,
         }
     }
 
+    /* Set up back-face colors. The rasterizer will do the color selection
+     * automatically. */
+    if (any_bcolor_used) {
+        for (i = 0; i < ATTR_COLOR_COUNT; i++) {
+            rs.vap_vsm_vtx_assm |= R300_INPUT_CNTL_COLOR;
+            rs.vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << (2+i);
+            stream_loc_notcl[loc++] = 4 + i;
+        }
+    }
+
     /* Rasterize texture coordinates. */
-    for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
+    for (i = 0; i < ATTR_GENERIC_COUNT && tex_count < 8; i++) {
 	bool sprite_coord = !!(r300->sprite_coord_enable & (1 << i));
 
         if (vs_outputs->generic[i] != ATTR_UNUSED || sprite_coord) {
-            /* Always rasterize if it's written by the VS,
-             * otherwise it locks up. */
+            if (!sprite_coord) {
+                /* Set up the texture coordinates in VAP. */
+                rs.vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << tex_count);
+                rs.vap_out_vtx_fmt[1] |= (4 << (3 * tex_count));
+                stream_loc_notcl[loc++] = 6 + tex_count;
+            }
+
+            /* Rasterize it. */
             rX00_rs_tex(&rs, tex_count, tex_count,
 			sprite_coord ? SWIZ_XY01 : SWIZ_XYZW);
 
-            /* Write it to the FS input register if it's used by the FS. */
+            /* Write it to the FS input register if it's needed by the FS. */
             if (fs_inputs->generic[i] != ATTR_UNUSED) {
                 rX00_rs_tex_write(&rs, tex_count, fp_offset);
                 fp_offset++;
@@ -365,12 +404,16 @@ static void r300_update_rs_block(struct r300_context* r300,
     }
 
     /* Rasterize fog coordinates. */
-    if (vs_outputs->fog != ATTR_UNUSED) {
-        /* Always rasterize if it's written by the VS,
-         * otherwise it locks up. */
+    if (vs_outputs->fog != ATTR_UNUSED && tex_count < 8) {
+        /* Set up the fog coordinates in VAP. */
+        rs.vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << tex_count);
+        rs.vap_out_vtx_fmt[1] |= (4 << (3 * tex_count));
+        stream_loc_notcl[loc++] = 6 + tex_count;
+
+        /* Rasterize it. */
         rX00_rs_tex(&rs, tex_count, tex_count, SWIZ_X001);
 
-        /* Write it to the FS input register if it's used by the FS. */
+        /* Write it to the FS input register if it's needed by the FS. */
         if (fs_inputs->fog != ATTR_UNUSED) {
             rX00_rs_tex_write(&rs, tex_count, fp_offset);
             fp_offset++;
@@ -391,9 +434,17 @@ static void r300_update_rs_block(struct r300_context* r300,
     }
 
     /* Rasterize WPOS. */
-    /* If the FS doesn't need it, it's not written by the VS. */
-    if (vs_outputs->wpos != ATTR_UNUSED && fs_inputs->wpos != ATTR_UNUSED) {
+    /* Don't set it in VAP if the FS doesn't need it. */
+    if (fs_inputs->wpos != ATTR_UNUSED && tex_count < 8) {
+        /* Set up the WPOS coordinates in VAP. */
+        rs.vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << tex_count);
+        rs.vap_out_vtx_fmt[1] |= (4 << (3 * tex_count));
+        stream_loc_notcl[loc++] = 6 + tex_count;
+
+        /* Rasterize it. */
         rX00_rs_tex(&rs, tex_count, tex_count, SWIZ_XYZW);
+
+        /* Write it to the FS input register. */
         rX00_rs_tex_write(&rs, tex_count, fp_offset);
 
         DBG(r300, DBG_RS, "r300: Rasterized WPOS written to FS.\n");
@@ -402,6 +453,11 @@ static void r300_update_rs_block(struct r300_context* r300,
         tex_count++;
     }
 
+    /* Invalidate the rest of the no-TCL (GA) stream locations. */
+    for (; loc < 16;) {
+        stream_loc_notcl[loc++] = -1;
+    }
+
     /* Rasterize at least one color, or bad things happen. */
     if (col_count == 0 && tex_count == 0) {
         rX00_rs_col(&rs, 0, 0, SWIZ_0001);
@@ -422,18 +478,10 @@ static void r300_update_rs_block(struct r300_context* r300,
     /* Now, after all that, see if we actually need to update the state. */
     if (memcmp(r300->rs_block_state.state, &rs, sizeof(struct r300_rs_block))) {
         memcpy(r300->rs_block_state.state, &rs, sizeof(struct r300_rs_block));
-        r300->rs_block_state.size = 5 + count*2;
+        r300->rs_block_state.size = 11 + count*2;
     }
 }
 
-/* Update the shader-dependant states. */
-static void r300_update_derived_shader_state(struct r300_context* r300)
-{
-    struct r300_vertex_shader* vs = r300->vs_state.state;
-
-    r300_update_rs_block(r300, &vs->outputs, &r300_fs(r300)->shader->inputs);
-}
-
 static void r300_merge_textures_and_samplers(struct r300_context* r300)
 {
     struct r300_textures_state *state =
@@ -529,7 +577,7 @@ void r300_update_derived_state(struct r300_context* r300)
     }
 
     if (r300->rs_block_state.dirty) {
-        r300_update_derived_shader_state(r300);
+        r300_update_rs_block(r300);
     }
 
     if (r300->draw) {
diff --git a/src/gallium/drivers/r300/r300_state_inlines.h b/src/gallium/drivers/r300/r300_state_inlines.h
index fcbdb91b67e..715601042fa 100644
--- a/src/gallium/drivers/r300/r300_state_inlines.h
+++ b/src/gallium/drivers/r300/r300_state_inlines.h
@@ -371,7 +371,7 @@ r300_translate_vertex_data_type(enum pipe_format format) {
     desc = util_format_description(format);
 
     if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN) {
-        fprintf(stderr, "r300: Bad format %s in %s:%d\n", util_format_name(format),
+        fprintf(stderr, "r300: Bad format %s in %s:%d\n", util_format_short_name(format),
             __FUNCTION__, __LINE__);
         assert(0);
         abort();
@@ -394,7 +394,7 @@ r300_translate_vertex_data_type(enum pipe_format format) {
                     break;
                 default:
                     fprintf(stderr, "r300: Bad format %s in %s:%d\n",
-                        util_format_name(format), __FUNCTION__, __LINE__);
+                        util_format_short_name(format), __FUNCTION__, __LINE__);
                     assert(0);
                     abort();
             }
@@ -416,7 +416,7 @@ r300_translate_vertex_data_type(enum pipe_format format) {
                     break;
                 default:
                     fprintf(stderr, "r300: Bad format %s in %s:%d\n",
-                        util_format_name(format), __FUNCTION__, __LINE__);
+                        util_format_short_name(format), __FUNCTION__, __LINE__);
                     fprintf(stderr, "r300: desc->channel[0].size == %d\n",
                         desc->channel[0].size);
                     assert(0);
@@ -425,7 +425,7 @@ r300_translate_vertex_data_type(enum pipe_format format) {
             break;
         default:
             fprintf(stderr, "r300: Bad format %s in %s:%d\n",
-                util_format_name(format), __FUNCTION__, __LINE__);
+                util_format_short_name(format), __FUNCTION__, __LINE__);
             assert(0);
             abort();
     }
@@ -449,7 +449,7 @@ r300_translate_vertex_data_swizzle(enum pipe_format format) {
 
     if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN) {
         fprintf(stderr, "r300: Bad format %s in %s:%d\n",
-            util_format_name(format), __FUNCTION__, __LINE__);
+            util_format_short_name(format), __FUNCTION__, __LINE__);
         return 0;
     }
 
diff --git a/src/gallium/drivers/r300/r300_texture.c b/src/gallium/drivers/r300/r300_texture.c
index 69e6a124458..d6f629cf9c6 100644
--- a/src/gallium/drivers/r300/r300_texture.c
+++ b/src/gallium/drivers/r300/r300_texture.c
@@ -46,18 +46,6 @@ static const unsigned microblock_table[5][3][2] = {
     {{ 2, 1}, {0, 0}, {0, 0}}  /* 128 bits per pixel */
 };
 
-/* Return true for non-compressed and non-YUV formats. */
-static boolean r300_format_is_plain(enum pipe_format format)
-{
-    const struct util_format_description *desc = util_format_description(format);
-
-    if (!format) {
-        return FALSE;
-    }
-
-    return desc->layout == UTIL_FORMAT_LAYOUT_PLAIN;
-}
-
 /* Translate a pipe_format into a useful texture format for sampling.
  *
  * Some special formats are translated directly using R300_EASY_TX_FORMAT,
@@ -585,9 +573,6 @@ static void r300_texture_setup_immutable_state(struct r300_screen* screen,
 
     f->tile_config = R300_TXO_MACRO_TILE(tex->macrotile) |
                      R300_TXO_MICRO_TILE(tex->microtile);
-
-    SCREEN_DBG(screen, DBG_TEX, "r300: Set texture state (%dx%d, %d levels)\n",
-               pt->width0, pt->height0, pt->last_level);
 }
 
 static void r300_texture_setup_fb_state(struct r300_screen* screen,
@@ -622,8 +607,8 @@ void r300_texture_reinterpret_format(struct pipe_screen *screen,
 {
     struct r300_screen *r300screen = r300_screen(screen);
 
-    SCREEN_DBG(r300screen, DBG_TEX, "r300: Reinterpreting format: %s -> %s\n",
-               util_format_name(tex->format), util_format_name(new_format));
+    SCREEN_DBG(r300screen, DBG_TEX, "r300: texture_reinterpret_format: %s -> %s\n",
+               util_format_short_name(tex->format), util_format_short_name(new_format));
 
     tex->format = new_format;
 
@@ -714,7 +699,7 @@ unsigned r300_texture_get_stride(struct r300_screen* screen,
 
     width = u_minify(tex->b.b.width0, level);
 
-    if (r300_format_is_plain(tex->b.b.format)) {
+    if (util_format_is_plain(tex->b.b.format)) {
         tile_width = r300_texture_get_tile_size(tex, TILE_WIDTH,
                                                 tex->mip_macrotile[level]);
         width = align(width, tile_width);
@@ -732,7 +717,7 @@ static unsigned r300_texture_get_nblocksy(struct r300_texture* tex,
 
     height = u_minify(tex->b.b.height0, level);
 
-    if (r300_format_is_plain(tex->b.b.format)) {
+    if (util_format_is_plain(tex->b.b.format)) {
         tile_height = r300_texture_get_tile_size(tex, TILE_HEIGHT,
                                                  tex->mip_macrotile[level]);
         height = align(height, tile_height);
@@ -774,8 +759,8 @@ static void r300_setup_miptree(struct r300_screen* screen,
     unsigned stride, size, layer_size, nblocksy, i;
     boolean rv350_mode = screen->caps.is_rv350;
 
-    SCREEN_DBG(screen, DBG_TEX, "r300: Making miptree for texture, format %s\n",
-               util_format_name(base->format));
+    SCREEN_DBG(screen, DBG_TEXALLOC, "r300: Making miptree for texture, format %s\n",
+               util_format_short_name(base->format));
 
     for (i = 0; i <= base->last_level; i++) {
         /* Let's see if this miplevel can be macrotiled. */
@@ -801,7 +786,7 @@ static void r300_setup_miptree(struct r300_screen* screen,
         tex->hwpitch[i] =
                 tex->pitch[i] * util_format_get_blockwidth(base->format);
 
-        SCREEN_DBG(screen, DBG_TEX, "r300: Texture miptree: Level %d "
+        SCREEN_DBG(screen, DBG_TEXALLOC, "r300: Texture miptree: Level %d "
                 "(%dx%dx%d px, pitch %d bytes) %d bytes total, macrotiled %s\n",
                 i, u_minify(base->width0, i), u_minify(base->height0, i),
                 u_minify(base->depth0, i), stride, tex->size,
@@ -825,7 +810,7 @@ static void r300_setup_tiling(struct pipe_screen *screen,
     boolean is_zb = util_format_is_depth_or_stencil(format);
     boolean dbg_no_tiling = SCREEN_DBG_ON(r300_screen(screen), DBG_NO_TILING);
 
-    if (!r300_format_is_plain(format)) {
+    if (!util_format_is_plain(format)) {
         return;
     }
 
@@ -927,6 +912,17 @@ struct pipe_resource* r300_texture_create(struct pipe_screen* screen,
         return NULL;
     }
 
+    /* Refuse to create a texture with size 0. */
+    if (!base->width0 ||
+        (!base->height0 && (base->target == PIPE_TEXTURE_2D ||
+                            base->target == PIPE_TEXTURE_CUBE)) ||
+        (!base->depth0 && base->target == PIPE_TEXTURE_3D)) {
+        fprintf(stderr, "r300: texture_create: "
+                "Got invalid texture dimensions: %ix%ix%i\n",
+                base->width0, base->height0, base->depth0);
+        return NULL;
+    }
+
     tex->b.b = *base;
     tex->b.vtbl = &r300_texture_vtbl;
     pipe_reference_init(&tex->b.b.reference, 1);
@@ -942,6 +938,15 @@ struct pipe_resource* r300_texture_create(struct pipe_screen* screen,
     r300_texture_setup_immutable_state(rscreen, tex);
     r300_texture_setup_fb_state(rscreen, tex);
 
+    SCREEN_DBG(rscreen, DBG_TEX,
+               "r300: texture_create: Macro: %s, Micro: %s, Pitch: %i, "
+               "Dim: %ix%ix%i, LastLevel: %i, Format: %s\n",
+               tex->macrotile ? "YES" : " NO",
+               tex->microtile ? "YES" : " NO",
+               tex->hwpitch[0],
+               base->width0, base->height0, base->depth0, base->last_level,
+               util_format_short_name(base->format));
+
     tex->buffer = rws->buffer_create(rws, 2048,
 				     PIPE_BIND_SAMPLER_VIEW, /* XXX */
 				     tex->size);
@@ -1039,6 +1044,14 @@ r300_texture_from_handle(struct pipe_screen* screen,
 
     rws->buffer_get_tiling(rws, buffer, &tex->microtile, &tex->macrotile);
     r300_setup_flags(tex);
+    SCREEN_DBG(rscreen, DBG_TEX,
+               "r300: texture_from_handle: Macro: %s, Micro: %s, "
+               "Pitch: % 4i, Dim: %ix%i, Format: %s\n",
+               tex->macrotile ? "YES" : " NO",
+               tex->microtile ? "YES" : " NO",
+               stride / util_format_get_blocksize(base->format),
+               base->width0, base->height0,
+               util_format_short_name(base->format));
 
     /* Enforce microtiled zbuffer. */
     override_zb_flags = util_format_is_depth_or_stencil(base->format) &&
diff --git a/src/gallium/drivers/r300/r300_vs.c b/src/gallium/drivers/r300/r300_vs.c
index bfab9c3b014..f3186431e1d 100644
--- a/src/gallium/drivers/r300/r300_vs.c
+++ b/src/gallium/drivers/r300/r300_vs.c
@@ -22,7 +22,6 @@
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
 #include "r300_vs.h"
-#include "r300_fs.h"
 
 #include "r300_context.h"
 #include "r300_screen.h"
@@ -94,94 +93,6 @@ static void r300_shader_read_vs_outputs(
     vs_outputs->wpos = i;
 }
 
-/* This function sets up:
- * - VAP mapping, which maps VS registers to output semantics and
- *   at the same time it indicates which attributes are enabled and should
- *   be rasterized.
- * - Stream mapping to VS outputs if TCL is not present. */
-static void r300_init_vs_output_mapping(struct r300_vertex_shader* vs)
-{
-    struct r300_shader_semantics* vs_outputs = &vs->outputs;
-    struct r300_vap_output_state *vap_out = &vs->vap_out;
-    int *stream_loc = vs->stream_loc_notcl;
-    int i, gen_count, tabi = 0;
-    boolean any_bcolor_used = vs_outputs->bcolor[0] != ATTR_UNUSED ||
-                              vs_outputs->bcolor[1] != ATTR_UNUSED;
-
-    vap_out->vap_vtx_state_cntl = 0x5555; /* XXX this is classic Mesa bonghits */
-
-    /* Position. */
-    if (vs_outputs->pos != ATTR_UNUSED) {
-        vap_out->vap_vsm_vtx_assm |= R300_INPUT_CNTL_POS;
-        vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT;
-
-        stream_loc[tabi++] = 0;
-    } else {
-        assert(0);
-    }
-
-    /* Point size. */
-    if (vs_outputs->psize != ATTR_UNUSED) {
-        vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__PT_SIZE_PRESENT;
-
-        stream_loc[tabi++] = 1;
-    }
-
-    /* Colors. */
-    for (i = 0; i < ATTR_COLOR_COUNT; i++) {
-        if (vs_outputs->color[i] != ATTR_UNUSED || any_bcolor_used ||
-            vs_outputs->color[1] != ATTR_UNUSED) {
-            vap_out->vap_vsm_vtx_assm |= R300_INPUT_CNTL_COLOR;
-            vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << i;
-
-            stream_loc[tabi++] = 2 + i;
-        }
-    }
-
-    /* Back-face colors. */
-    if (any_bcolor_used) {
-        for (i = 0; i < ATTR_COLOR_COUNT; i++) {
-            vap_out->vap_vsm_vtx_assm |= R300_INPUT_CNTL_COLOR;
-            vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << (2+i);
-
-            stream_loc[tabi++] = 4 + i;
-        }
-    }
-
-    /* Texture coordinates. */
-    gen_count = 0;
-    for (i = 0; i < ATTR_GENERIC_COUNT && gen_count < 8; i++) {
-        if (vs_outputs->generic[i] != ATTR_UNUSED) {
-            vap_out->vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << gen_count);
-            vap_out->vap_out_vtx_fmt[1] |= (4 << (3 * gen_count));
-
-            stream_loc[tabi++] = 6 + gen_count;
-            gen_count++;
-        }
-    }
-
-    /* Fog coordinates. */
-    if (gen_count < 8 && vs_outputs->fog != ATTR_UNUSED) {
-        vap_out->vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << gen_count);
-        vap_out->vap_out_vtx_fmt[1] |= (4 << (3 * gen_count));
-
-        stream_loc[tabi++] = 6 + gen_count;
-        gen_count++;
-    }
-
-    /* WPOS. */
-    if (gen_count < 8) {
-        vs->wpos_tex_output = gen_count;
-        stream_loc[tabi++] = 6 + gen_count;
-    } else {
-        vs_outputs->wpos = ATTR_UNUSED;
-    }
-
-    for (; tabi < 16;) {
-        stream_loc[tabi++] = -1;
-    }
-}
-
 static void set_vertex_inputs_outputs(struct r300_vertex_program_compiler * c)
 {
     struct r300_vertex_shader * vs = c->UserData;
@@ -246,9 +157,7 @@ static void set_vertex_inputs_outputs(struct r300_vertex_program_compiler * c)
     }
 
     /* WPOS. */
-    if (outputs->wpos != ATTR_UNUSED) {
-        c->code->outputs[outputs->wpos] = reg++;
-    }
+    c->code->outputs[outputs->wpos] = reg++;
 }
 
 static void r300_dummy_vertex_shader(
@@ -286,7 +195,6 @@ void r300_translate_vertex_shader(struct r300_context* r300,
 
     tgsi_scan_shader(tokens, &vs->info);
     r300_shader_read_vs_outputs(&vs->info, &vs->outputs);
-    r300_init_vs_output_mapping(vs);
 
     /* Setup the compiler */
     rc_init(&compiler.Base);
@@ -307,16 +215,11 @@ void r300_translate_vertex_shader(struct r300_context* r300,
 
     r300_tgsi_to_rc(&ttr, tokens);
 
-    compiler.RequiredOutputs =
-        ~(~0 << (vs->info.num_outputs +
-                 (vs->outputs.wpos != ATTR_UNUSED ? 1 : 0)));
-
+    compiler.RequiredOutputs = ~(~0 << (vs->info.num_outputs + 1));
     compiler.SetHwInputOutput = &set_vertex_inputs_outputs;
 
     /* Insert the WPOS output. */
-    if (vs->outputs.wpos != ATTR_UNUSED) {
-        rc_copy_output(&compiler.Base, 0, vs->outputs.wpos);
-    }
+    rc_copy_output(&compiler.Base, 0, vs->outputs.wpos);
 
     /* Invoke the compiler */
     r3xx_compile_vertex_program(&compiler);
@@ -343,32 +246,3 @@ void r300_translate_vertex_shader(struct r300_context* r300,
     /* And, finally... */
     rc_destroy(&compiler.Base);
 }
-
-boolean r300_vertex_shader_setup_wpos(struct r300_context* r300)
-{
-    struct r300_vertex_shader* vs = r300->vs_state.state;
-    struct r300_vap_output_state *vap_out = &vs->vap_out;
-    int tex_output = vs->wpos_tex_output;
-    uint32_t tex_fmt = R300_INPUT_CNTL_TC0 << tex_output;
-
-    if (vs->outputs.wpos == ATTR_UNUSED) {
-        return FALSE;
-    }
-
-    if (r300_fs(r300)->shader->inputs.wpos != ATTR_UNUSED) {
-        /* Enable WPOS in VAP. */
-        if (!(vap_out->vap_vsm_vtx_assm & tex_fmt)) {
-            vap_out->vap_vsm_vtx_assm |= tex_fmt;
-            vap_out->vap_out_vtx_fmt[1] |= (4 << (3 * tex_output));
-            return TRUE;
-        }
-    } else {
-        /* Disable WPOS in VAP. */
-        if (vap_out->vap_vsm_vtx_assm & tex_fmt) {
-            vap_out->vap_vsm_vtx_assm &= ~tex_fmt;
-            vap_out->vap_out_vtx_fmt[1] &= ~(4 << (3 * tex_output));
-            return TRUE;
-        }
-    }
-    return FALSE;
-}
diff --git a/src/gallium/drivers/r300/r300_vs.h b/src/gallium/drivers/r300/r300_vs.h
index 56bcc3b70b8..57b3fbca0bb 100644
--- a/src/gallium/drivers/r300/r300_vs.h
+++ b/src/gallium/drivers/r300/r300_vs.h
@@ -39,7 +39,6 @@ struct r300_vertex_shader {
 
     struct tgsi_shader_info info;
     struct r300_shader_semantics outputs;
-    struct r300_vap_output_state vap_out;
 
     /* Whether the shader was replaced by a dummy one due to a shader
      * compilation failure. */
@@ -49,12 +48,6 @@ struct r300_vertex_shader {
     unsigned externals_count;
     unsigned immediates_count;
 
-    /* Stream locations for SWTCL or if TCL is bypassed. */
-    int stream_loc_notcl[16];
-
-    /* Output stream location for WPOS. */
-    int wpos_tex_output;
-
     /* HWTCL-specific.  */
     /* Machine code (if translated) */
     struct r300_vertex_program_code code;
@@ -67,7 +60,4 @@ void r300_translate_vertex_shader(struct r300_context* r300,
                                   struct r300_vertex_shader* vs,
                                   const struct tgsi_token *tokens);
 
-/* Return TRUE if VAP (hwfmt) needs to be re-emitted. */
-boolean r300_vertex_shader_setup_wpos(struct r300_context* r300);
-
 #endif /* R300_VS_H */