Merge remote branch 'origin/master' into radeon-rewrite

19 files changed, 372 insertions, 161 deletions

diff --git a/src/gallium/drivers/r300/r300_chipset.c b/src/gallium/drivers/r300/r300_chipset.c
index 758f706c51..00fae8d26f 100644
--- a/src/gallium/drivers/r300/r300_chipset.c
+++ b/src/gallium/drivers/r300/r300_chipset.c
@@ -149,6 +149,7 @@ void r300_parse_chipset(struct r300_capabilities* caps)
             caps->num_vert_fpus = 6;
             break;
 
+        case 0x4B48:
         case 0x4B49:
         case 0x4B4A:
         case 0x4B4B:
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index 58f1fa0e2e..27bc7fd1a9 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -117,23 +117,24 @@ struct r300_viewport_state {
     uint32_t vte_control; /* R300_VAP_VTE_CNTL:      0x20b0 */
 };
 
-#define R300_NEW_BLEND           0x0000001
-#define R300_NEW_BLEND_COLOR     0x0000002
-#define R300_NEW_CONSTANTS       0x0000004
-#define R300_NEW_DSA             0x0000008
-#define R300_NEW_FRAMEBUFFERS    0x0000010
-#define R300_NEW_FRAGMENT_SHADER 0x0000020
-#define R300_NEW_RASTERIZER      0x0000040
-#define R300_NEW_RS_BLOCK        0x0000080
-#define R300_NEW_SAMPLER         0x0000100
-#define R300_ANY_NEW_SAMPLERS    0x000ff00
-#define R300_NEW_SCISSOR         0x0010000
-#define R300_NEW_TEXTURE         0x0020000
-#define R300_ANY_NEW_TEXTURES    0x1fe0000
-#define R300_NEW_VERTEX_FORMAT   0x2000000
-#define R300_NEW_VERTEX_SHADER   0x4000000
-#define R300_NEW_VIEWPORT        0x8000000
-#define R300_NEW_KITCHEN_SINK    0xfffffff
+#define R300_NEW_BLEND           0x00000001
+#define R300_NEW_BLEND_COLOR     0x00000002
+#define R300_NEW_CLIP            0x00000004
+#define R300_NEW_CONSTANTS       0x00000008
+#define R300_NEW_DSA             0x00000010
+#define R300_NEW_FRAMEBUFFERS    0x00000020
+#define R300_NEW_FRAGMENT_SHADER 0x00000040
+#define R300_NEW_RASTERIZER      0x00000080
+#define R300_NEW_RS_BLOCK        0x00000100
+#define R300_NEW_SAMPLER         0x00000200
+#define R300_ANY_NEW_SAMPLERS    0x0001fe00
+#define R300_NEW_SCISSOR         0x00020000
+#define R300_NEW_TEXTURE         0x00040000
+#define R300_ANY_NEW_TEXTURES    0x03fc0000
+#define R300_NEW_VERTEX_FORMAT   0x04000000
+#define R300_NEW_VERTEX_SHADER   0x08000000
+#define R300_NEW_VIEWPORT        0x10000000
+#define R300_NEW_KITCHEN_SINK    0x1fffffff
 
 /* The next several objects are not pure Radeon state; they inherit from
  * various Gallium classes. */
@@ -141,11 +142,11 @@ struct r300_viewport_state {
 struct r300_constant_buffer {
     /* Buffer of constants */
     /* XXX first number should be raised */
-    float constants[8][4];
+    float constants[32][4];
     /* Number of user-defined constants */
-    int user_count;
+    unsigned user_count;
     /* Total number of constants */
-    int count;
+    unsigned count;
 };
 
 struct r3xx_fragment_shader {
@@ -158,6 +159,10 @@ struct r3xx_fragment_shader {
 
     /* Pixel stack size */
     int stack_size;
+
+    /* Are there immediates in this shader?
+     * If not, we can heavily optimize recompilation. */
+    boolean uses_imms;
 };
 
 struct r300_fragment_shader {
@@ -248,6 +253,10 @@ struct r300_vertex_shader {
     /* Has this shader been translated yet? */
     boolean translated;
 
+    /* Are there immediates in this shader?
+     * If not, we can heavily optimize recompilation. */
+    boolean uses_imms;
+
     /* Number of used instructions */
     int instruction_count;
 
@@ -284,6 +293,8 @@ struct r300_context {
     struct r300_blend_state* blend_state;
     /* Blend color state. */
     struct r300_blend_color_state* blend_color_state;
+    /* User clip planes. */
+    struct pipe_clip_state clip_state;
     /* Shader constants. */
     struct r300_constant_buffer shader_constants[PIPE_SHADER_TYPES];
     /* Depth, stencil, and alpha state. */
diff --git a/src/gallium/drivers/r300/r300_cs.h b/src/gallium/drivers/r300/r300_cs.h
index 82a3942248..71b142c0db 100644
--- a/src/gallium/drivers/r300/r300_cs.h
+++ b/src/gallium/drivers/r300/r300_cs.h
@@ -34,6 +34,7 @@
 
 #define MAX_CS_SIZE 64 * 1024 / 4
 
+#define VERY_VERBOSE_CS 0
 #define VERY_VERBOSE_REGISTERS 0
 
 /* XXX stolen from radeon_drm.h */
@@ -56,8 +57,10 @@
 
 #define BEGIN_CS(size) do { \
     CHECK_CS(size); \
-    debug_printf("r300: BEGIN_CS, count %d, in %s (%s:%d)\n", \
-        size, __FUNCTION__, __FILE__, __LINE__); \
+    if (VERY_VERBOSE_CS) { \
+        debug_printf("r300: BEGIN_CS, count %d, in %s (%s:%d)\n", \
+                size, __FUNCTION__, __FILE__, __LINE__); \
+    } \
     cs_winsys->begin_cs(cs_winsys, (size), \
             __FILE__, __FUNCTION__, __LINE__); \
     cs_count = size; \
@@ -93,8 +96,9 @@
 } while (0)
 
 #define OUT_CS_RELOC(bo, offset, rd, wd, flags) do { \
-    debug_printf("r300: writing relocation for buffer %p, offset %d\n", \
-        bo, offset); \
+    debug_printf("r300: writing relocation for buffer %p, offset %d, " \
+            "domains (%d, %d, %d)\n", \
+        bo, offset, rd, wd, flags); \
     assert(bo); \
     OUT_CS(offset); \
     cs_winsys->write_cs_reloc(cs_winsys, bo, rd, wd, flags); \
@@ -102,16 +106,20 @@
 } while (0)
 
 #define END_CS do { \
-    debug_printf("r300: END_CS in %s (%s:%d)\n", __FUNCTION__, __FILE__, \
-        __LINE__); \
+    if (VERY_VERBOSE_CS) { \
+        debug_printf("r300: END_CS in %s (%s:%d)\n", __FUNCTION__, \
+                __FILE__, __LINE__); \
+    } \
     if (cs_count != 0) \
         debug_printf("r300: Warning: cs_count off by %d\n", cs_count); \
     cs_winsys->end_cs(cs_winsys, __FILE__, __FUNCTION__, __LINE__); \
 } while (0)
 
 #define FLUSH_CS do { \
-    debug_printf("r300: FLUSH_CS in %s (%s:%d)\n\n", __FUNCTION__, __FILE__, \
-        __LINE__); \
+    if (VERY_VERBOSE_CS) { \
+        debug_printf("r300: FLUSH_CS in %s (%s:%d)\n\n", __FUNCTION__, \
+                __FILE__, __LINE__); \
+    } \
     cs_winsys->flush_cs(cs_winsys); \
 } while (0)
 
diff --git a/src/gallium/drivers/r300/r300_debug.c b/src/gallium/drivers/r300/r300_debug.c
index ffc93eb591..678cd2b812 100644
--- a/src/gallium/drivers/r300/r300_debug.c
+++ b/src/gallium/drivers/r300/r300_debug.c
@@ -152,6 +152,8 @@ void r500_fs_dump(struct r500_fragment_shader* fs)
 
 static void r300_vs_op_dump(uint32_t op)
 {
+    debug_printf(" dst: %d%s op: ",
+            (op >> 13) & 0x7f, r300_vs_dst_debug[(op >> 8) & 0x7]);
     if (op & 0x80) {
         if (op & 0x1) {
             debug_printf("PVS_MACRO_OP_2CLK_M2X_ADD\n");
@@ -165,14 +167,32 @@ static void r300_vs_op_dump(uint32_t op)
     }
 }
 
+void r300_vs_src_dump(uint32_t src)
+{
+    debug_printf(" reg: %d%s swiz: %s%s/%s%s/%s%s/%s%s\n",
+            (src >> 5) & 0x7f, r300_vs_src_debug[src & 0x3],
+            src & (1 << 25) ? "-" : " ",
+            r300_vs_swiz_debug[(src >> 13) & 0x7],
+            src & (1 << 26) ? "-" : " ",
+            r300_vs_swiz_debug[(src >> 16) & 0x7],
+            src & (1 << 27) ? "-" : " ",
+            r300_vs_swiz_debug[(src >> 19) & 0x7],
+            src & (1 << 28) ? "-" : " ",
+            r300_vs_swiz_debug[(src >> 22) & 0x7]);
+}
+
 void r300_vs_dump(struct r300_vertex_shader* vs)
 {
     int i;
 
     for (i = 0; i < vs->instruction_count; i++) {
+        debug_printf("%d: op: 0x%08x", i, vs->instructions[i].inst0);
         r300_vs_op_dump(vs->instructions[i].inst0);
-        debug_printf("inst1: 0x%x\n", vs->instructions[i].inst1);
-        debug_printf("inst2: 0x%x\n", vs->instructions[i].inst2);
-        debug_printf("inst3: 0x%x\n", vs->instructions[i].inst3);
+        debug_printf(" src0: 0x%08x", vs->instructions[i].inst1);
+        r300_vs_src_dump(vs->instructions[i].inst1);
+        debug_printf(" src1: 0x%08x", vs->instructions[i].inst2);
+        r300_vs_src_dump(vs->instructions[i].inst2);
+        debug_printf(" src2: 0x%08x", vs->instructions[i].inst3);
+        r300_vs_src_dump(vs->instructions[i].inst3);
     }
 }
diff --git a/src/gallium/drivers/r300/r300_debug.h b/src/gallium/drivers/r300/r300_debug.h
index 6306594099..c86410ec0a 100644
--- a/src/gallium/drivers/r300/r300_debug.h
+++ b/src/gallium/drivers/r300/r300_debug.h
@@ -173,6 +173,36 @@ static char* r300_vs_me_ops[] = {
     "               (reserved)",
 };
 
+/* XXX refactor to avoid clashing symbols */
+static char* r300_vs_src_debug[] = {
+    "t",
+    "i",
+    "c",
+    "a",
+};
+
+static char* r300_vs_dst_debug[] = {
+    "t",
+    "a0",
+    "o",
+    "ox",
+    "a",
+    "i",
+    "u",
+    "u",
+};
+
+static char* r300_vs_swiz_debug[] = {
+    "X",
+    "Y",
+    "Z",
+    "W",
+    "0",
+    "1",
+    "U",
+    "U",
+};
+
 void r500_fs_dump(struct r500_fragment_shader* fs);
 
 void r300_vs_dump(struct r300_vertex_shader* vs);
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index 0cb0507fc8..d81abe4d0b 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -56,6 +56,30 @@ void r300_emit_blend_color_state(struct r300_context* r300,
     }
 }
 
+void r300_emit_clip_state(struct r300_context* r300,
+                          struct pipe_clip_state* clip)
+{
+    int i;
+    struct r300_screen* r300screen = r300_screen(r300->context.screen);
+    CS_LOCALS(r300);
+
+    BEGIN_CS(5 + (6 * 4));
+    OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG,
+            (r300screen->caps->is_r500 ?
+             R500_PVS_UCP_START : R300_PVS_UCP_START));
+    OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, 6 * 4);
+    for (i = 0; i < 6; i++) {
+        OUT_CS_32F(clip->ucp[i][0]);
+        OUT_CS_32F(clip->ucp[i][1]);
+        OUT_CS_32F(clip->ucp[i][2]);
+        OUT_CS_32F(clip->ucp[i][3]);
+    }
+
+    OUT_CS_REG(R300_VAP_CLIP_CNTL, ((1 << clip->nr) - 1) |
+            R300_PS_UCP_MODE_CLIP_AS_TRIFAN);
+    END_CS;
+}
+
 void r300_emit_dsa_state(struct r300_context* r300,
                            struct r300_dsa_state* dsa)
 {
@@ -372,17 +396,22 @@ void r300_emit_vertex_shader(struct r300_context* r300,
     }
 
     if (constants->count) {
-        BEGIN_CS(16 + (vs->instruction_count * 4) + (constants->count * 4));
+        BEGIN_CS(14 + (vs->instruction_count * 4) + (constants->count * 4));
     } else {
-        BEGIN_CS(13 + (vs->instruction_count * 4) + (constants->count * 4));
+        BEGIN_CS(11 + (vs->instruction_count * 4));
     }
 
-    OUT_CS_REG(R300_VAP_PVS_CODE_CNTL_0, R300_PVS_FIRST_INST(0) |
+    /* 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... */
+    OUT_CS_REG_SEQ(R300_VAP_PVS_CODE_CNTL_0, 3);
+    OUT_CS(R300_PVS_FIRST_INST(0) |
+            R300_PVS_XYZW_VALID_INST(vs->instruction_count - 1) |
             R300_PVS_LAST_INST(vs->instruction_count - 1));
-    OUT_CS_REG(R300_VAP_PVS_CODE_CNTL_1, vs->instruction_count - 1);
-
-    /* XXX */
-    OUT_CS_REG(R300_VAP_PVS_CONST_CNTL, 0x0);
+    OUT_CS(R300_PVS_MAX_CONST_ADDR(constants->count - 1));
+    OUT_CS(vs->instruction_count - 1);
 
     OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
     OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, vs->instruction_count * 4);
@@ -412,7 +441,6 @@ void r300_emit_vertex_shader(struct r300_context* r300,
             R300_PVS_VF_MAX_VTX_NUM(12));
     OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x0);
     END_CS;
-
 }
 
 void r300_emit_viewport_state(struct r300_context* r300,
@@ -452,8 +480,8 @@ void r300_emit_dirty_state(struct r300_context* r300)
 {
     struct r300_screen* r300screen = r300_screen(r300->context.screen);
     struct r300_texture* tex;
-    int i;
-    int dirty_tex = 0;
+    int i, dirty_tex = 0;
+    boolean invalid = FALSE;
 
     if (!(r300->dirty_state)) {
         return;
@@ -462,38 +490,55 @@ void r300_emit_dirty_state(struct r300_context* r300)
     r300_update_derived_state(r300);
 
     /* XXX check size */
+validate:
     /* Color buffers... */
     for (i = 0; i < r300->framebuffer_state.nr_cbufs; i++) {
         tex = (struct r300_texture*)r300->framebuffer_state.cbufs[i]->texture;
         assert(tex && tex->buffer && "cbuf is marked, but NULL!");
-        r300->winsys->add_buffer(r300->winsys, tex->buffer,
-                0, RADEON_GEM_DOMAIN_VRAM);
+        if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
+                    0, RADEON_GEM_DOMAIN_VRAM)) {
+            r300->context.flush(&r300->context, 0, NULL);
+            goto validate;
+        }
     }
     /* ...depth buffer... */
     if (r300->framebuffer_state.zsbuf) {
         tex = (struct r300_texture*)r300->framebuffer_state.zsbuf->texture;
         assert(tex && tex->buffer && "zsbuf is marked, but NULL!");
-        r300->winsys->add_buffer(r300->winsys, tex->buffer,
-                0, RADEON_GEM_DOMAIN_VRAM);
+        if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
+                    0, RADEON_GEM_DOMAIN_VRAM)) {
+            r300->context.flush(&r300->context, 0, NULL);
+            goto validate;
+        }
     }
     /* ...textures... */
     for (i = 0; i < r300->texture_count; i++) {
         tex = r300->textures[i];
         assert(tex && tex->buffer && "texture is marked, but NULL!");
-        r300->winsys->add_buffer(r300->winsys, tex->buffer,
-                RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0);
+        if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
+                    RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0)) {
+            r300->context.flush(&r300->context, 0, NULL);
+            goto validate;
+        }
     }
     /* ...and vertex buffer. */
     if (r300->vbo) {
-        r300->winsys->add_buffer(r300->winsys, r300->vbo,
-                RADEON_GEM_DOMAIN_GTT, 0);
+        if (!r300->winsys->add_buffer(r300->winsys, r300->vbo,
+                    RADEON_GEM_DOMAIN_GTT, 0)) {
+            r300->context.flush(&r300->context, 0, NULL);
+            goto validate;
+        }
     } else {
         debug_printf("No VBO while emitting dirty state!\n");
     }
-
     if (r300->winsys->validate(r300->winsys)) {
-        /* XXX */
         r300->context.flush(&r300->context, 0, NULL);
+        if (invalid) {
+            /* Well, hell. */
+            exit(1);
+        }
+        invalid = TRUE;
+        goto validate;
     }
 
     if (r300->dirty_state & R300_NEW_BLEND) {
@@ -506,6 +551,11 @@ void r300_emit_dirty_state(struct r300_context* r300)
         r300->dirty_state &= ~R300_NEW_BLEND_COLOR;
     }
 
+    if (r300->dirty_state & R300_NEW_CLIP) {
+        r300_emit_clip_state(r300, &r300->clip_state);
+        r300->dirty_state &= ~R300_NEW_CLIP;
+    }
+
     if (r300->dirty_state & R300_NEW_DSA) {
         r300_emit_dsa_state(r300, r300->dsa_state);
         r300->dirty_state &= ~R300_NEW_DSA;
diff --git a/src/gallium/drivers/r300/r300_emit.h b/src/gallium/drivers/r300/r300_emit.h
index 36e14f69f7..946f625bd8 100644
--- a/src/gallium/drivers/r300/r300_emit.h
+++ b/src/gallium/drivers/r300/r300_emit.h
@@ -36,6 +36,9 @@ void r300_emit_blend_state(struct r300_context* r300,
 void r300_emit_blend_color_state(struct r300_context* r300,
                                  struct r300_blend_color_state* bc);
 
+void r300_emit_clip_state(struct r300_context* r300,
+                          struct pipe_clip_state* clip);
+
 void r300_emit_dsa_state(struct r300_context* r300,
                          struct r300_dsa_state* dsa);
 
diff --git a/src/gallium/drivers/r300/r300_reg.h b/src/gallium/drivers/r300/r300_reg.h
index 660816e1da..3bb9bc47b5 100644
--- a/src/gallium/drivers/r300/r300_reg.h
+++ b/src/gallium/drivers/r300/r300_reg.h
@@ -511,11 +511,13 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #       define R300_PVS_XYZW_VALID_INST_SHIFT    10
 #       define R300_PVS_LAST_INST_SHIFT          20
 #       define R300_PVS_FIRST_INST(x)            ((x) << 0)
+#       define R300_PVS_XYZW_VALID_INST(x)       ((x) << 10)
 #       define R300_PVS_LAST_INST(x)             ((x) << 20)
 /* Addresses are relative the the vertex program parameters area. */
 #define R300_VAP_PVS_CONST_CNTL             0x22D4
 #       define R300_PVS_CONST_BASE_OFFSET_SHIFT  0
 #       define R300_PVS_MAX_CONST_ADDR_SHIFT     16
+#       define R300_PVS_MAX_CONST_ADDR(x)        ((x) << 16)
 #define R300_VAP_PVS_CODE_CNTL_1	    0x22D8
 #       define R300_PVS_LAST_VTX_SRC_INST_SHIFT  0
 #define R300_VAP_PVS_FLOW_CNTL_OPC          0x22DC
@@ -3040,6 +3042,7 @@ enum {
 #   define R500_INST_RGB_WMASK_R			(1 << 11)
 #   define R500_INST_RGB_WMASK_G			(1 << 12)
 #   define R500_INST_RGB_WMASK_B			(1 << 13)
+#   define R500_INST_RGB_WMASK_RGB			(7 << 11)
 #   define R500_INST_ALPHA_WMASK			(1 << 14)
 #   define R500_INST_RGB_OMASK_R			(1 << 15)
 #   define R500_INST_RGB_OMASK_G			(1 << 16)
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index 29b66cee7e..cd458d019a 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -45,11 +45,7 @@ struct r300_render {
 
     /* VBO */
     struct pipe_buffer* vbo;
-    size_t vbo_size;
-    size_t vbo_offset;
-    void* vbo_map;
     size_t vbo_alloc_size;
-    size_t vbo_max_used;
 };
 
 static INLINE struct r300_render*
@@ -78,24 +74,21 @@ static boolean r300_render_allocate_vertices(struct vbuf_render* render,
     struct pipe_screen* screen = r300->context.screen;
     size_t size = (size_t)vertex_size * (size_t)count;
 
-    if (r300render->vbo) {
+    if (r300render->vbo && (size > r300render->vbo_alloc_size)) {
         pipe_buffer_reference(&r300render->vbo, NULL);
     }
+    
+    if (!r300render->vbo) {
+        r300render->vbo = pipe_buffer_create(screen,
+                                             64,
+                                             PIPE_BUFFER_USAGE_VERTEX,
+                                             size);
+    }
 
-    r300render->vbo_size = MAX2(size, r300render->vbo_alloc_size);
-    r300render->vbo_offset = 0;
-    r300render->vbo = pipe_buffer_create(screen,
-                                         64,
-                                         PIPE_BUFFER_USAGE_VERTEX,
-                                         r300render->vbo_size);
-
+    r300render->vbo_alloc_size = MAX2(size, r300render->vbo_alloc_size);
     r300render->vertex_size = vertex_size;
 
-    if (r300render->vbo) {
-        return TRUE;
-    } else {
-        return FALSE;
-    }
+    return (r300render->vbo) ? TRUE : FALSE;
 }
 
 static void* r300_render_map_vertices(struct vbuf_render* render)
@@ -103,10 +96,8 @@ static void* r300_render_map_vertices(struct vbuf_render* render)
     struct r300_render* r300render = r300_render(render);
     struct pipe_screen* screen = r300render->r300->context.screen;
 
-    r300render->vbo_map = pipe_buffer_map(screen, r300render->vbo,
-                                          PIPE_BUFFER_USAGE_CPU_WRITE);
-
-    return (unsigned char*)r300render->vbo_map + r300render->vbo_offset;
+    return (unsigned char*)pipe_buffer_map(screen, r300render->vbo,
+                                           PIPE_BUFFER_USAGE_CPU_WRITE);
 }
 
 static void r300_render_unmap_vertices(struct vbuf_render* render,
@@ -116,9 +107,6 @@ static void r300_render_unmap_vertices(struct vbuf_render* render,
     struct r300_render* r300render = r300_render(render);
     struct pipe_screen* screen = r300render->r300->context.screen;
 
-    r300render->vbo_max_used = MAX2(r300render->vbo_max_used,
-             r300render->vertex_size * (max + 1));
-
     pipe_buffer_unmap(screen, r300render->vbo);
 }
 
@@ -181,7 +169,6 @@ static void prepare_render(struct r300_render* render, unsigned count)
     CS_LOCALS(r300);
 
     r300->vbo = render->vbo;
-    r300->vbo_offset = render->vbo_offset;
 
     r300_emit_dirty_state(r300);
 }
@@ -195,8 +182,6 @@ static void r300_render_draw_arrays(struct vbuf_render* render,
 
     CS_LOCALS(r300);
 
-    r300render->vbo_offset = start;
-
     prepare_render(r300render, count);
 
     debug_printf("r300: Doing vbuf render, count %d\n", count);
@@ -231,13 +216,14 @@ static void r300_render_draw(struct vbuf_render* render,
         return;
     }
 
+/*
     index_map = pipe_buffer_map(screen, index_buffer,
                                 PIPE_BUFFER_USAGE_CPU_WRITE);
     memcpy(index_map, indices, count);
     pipe_buffer_unmap(screen, index_buffer);
 
     debug_printf("r300: Doing indexbuf render, count %d\n", count);
-/*
+
     BEGIN_CS(8);
     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
@@ -247,13 +233,15 @@ static void r300_render_draw(struct vbuf_render* render,
     OUT_CS_INDEX_RELOC(index_buffer, 0, count, RADEON_GEM_DOMAIN_GTT, 0, 0);
     END_CS; */
 
-    BEGIN_CS(2 + count);
-    OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, count);
+    BEGIN_CS(2 + (count+1)/2);
+    OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, (count+1)/2);
     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
-           r300render->hwprim | R300_VAP_VF_CNTL__INDEX_SIZE_32bit);
-    for (i = 0; i < count; i++) {
-        index = indices[i];
-        OUT_CS(index);
+           r300render->hwprim);
+    for (i = 0; i < count-1; i += 2) {
+        OUT_CS(indices[i+1] << 16 | indices[i]);
+    }
+    if (count % 2) {
+        OUT_CS(indices[count-1]);
     }
     END_CS;
 }
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index 4e65fbbabe..29e721984f 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -119,9 +119,10 @@ static void r300_set_clip_state(struct pipe_context* pipe,
                                 const struct pipe_clip_state* state)
 {
     struct r300_context* r300 = r300_context(pipe);
-    /* XXX add HW TCL clipping setup */
-    draw_flush(r300->draw);
-    draw_set_clip_state(r300->draw, state);
+
+    r300->clip_state = *state;
+
+    r300->dirty_state |= R300_NEW_CLIP;
 }
 
 static void
@@ -151,10 +152,12 @@ static void
 
     /* If the number of constants have changed, invalidate the shader. */
     if (r300->shader_constants[shader].user_count != i) {
-        if (shader == PIPE_SHADER_FRAGMENT && r300->fs) {
+        if (shader == PIPE_SHADER_FRAGMENT && r300->fs &&
+                r300->fs->uses_imms) {
             r300->fs->translated = FALSE;
             r300_translate_fragment_shader(r300, r300->fs);
-        } else if (shader == PIPE_SHADER_VERTEX && r300->vs) {
+        } else if (shader == PIPE_SHADER_VERTEX && r300->vs &&
+                r300->vs->uses_imms) {
             r300->vs->translated = FALSE;
             r300_translate_vertex_shader(r300, r300->vs);
         }
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index caa5f3b543..2477b30822 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -64,6 +64,7 @@ static void r300_vs_tab_routes(struct r300_context* r300,
                     break;
                 case TGSI_SEMANTIC_FOG:
                     fog = TRUE;
+                    /* Fall through */
                 case TGSI_SEMANTIC_GENERIC:
                     texs++;
                     break;
@@ -103,6 +104,9 @@ static void r300_vs_tab_routes(struct r300_context* r300,
         }
     }
 
+    /* XXX magic */
+    assert(texs <= 8);
+
     /* Do the actual vertex_info setup.
      *
      * vertex_info has four uints of hardware-specific data in it.
@@ -140,21 +144,32 @@ static void r300_vs_tab_routes(struct r300_context* r300,
         vinfo->hwfmt[2] |= (R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << i);
     }
 
-    for (i = 0; i < texs; i++) {
+    /* Init i right here, increment it if fog is enabled.
+     * This gets around a double-increment problem. */
+    i = 0;
+
+    if (fog) {
+        i++;
         draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE,
-            draw_find_vs_output(r300->draw, TGSI_SEMANTIC_GENERIC, i));
+            draw_find_vs_output(r300->draw, TGSI_SEMANTIC_FOG, 0));
         vinfo->hwfmt[1] |= (R300_INPUT_CNTL_TC0 << i);
         vinfo->hwfmt[3] |= (4 << (3 * i));
     }
 
-    if (fog) {
-        i++;
+    for (i; i < texs; i++) {
         draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE,
-            draw_find_vs_output(r300->draw, TGSI_SEMANTIC_FOG, 0));
+            draw_find_vs_output(r300->draw, TGSI_SEMANTIC_GENERIC, i));
         vinfo->hwfmt[1] |= (R300_INPUT_CNTL_TC0 << i);
         vinfo->hwfmt[3] |= (4 << (3 * i));
     }
 
+    /* Handle the case where the vertex shader will be generating some of
+     * the attribs based on its inputs. */
+    if (r300screen->caps->has_tcl &&
+            info->num_inputs < info->num_outputs) {
+        vinfo->num_attribs = info->num_inputs;
+    }
+
     draw_compute_vertex_size(vinfo);
 }
 
diff --git a/src/gallium/drivers/r300/r300_state_invariant.c b/src/gallium/drivers/r300/r300_state_invariant.c
index 9dde662802..60eff08f2e 100644
--- a/src/gallium/drivers/r300/r300_state_invariant.c
+++ b/src/gallium/drivers/r300/r300_state_invariant.c
@@ -69,7 +69,7 @@ void r300_emit_invariant_state(struct r300_context* r300)
     END_CS;
 
     /* XXX unsorted stuff from surface_fill */
-    BEGIN_CS(77 + (caps->has_tcl ? 7 : 0));
+    BEGIN_CS(75 + (caps->has_tcl ? 7 : 0));
     /* Flush PVS. */
     OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x0);
 
@@ -80,8 +80,6 @@ void r300_emit_invariant_state(struct r300_context* r300)
     /* XXX endian */
     if (caps->has_tcl) {
         OUT_CS_REG(R300_VAP_CNTL_STATUS, R300_VC_NO_SWAP);
-        OUT_CS_REG(R300_VAP_CLIP_CNTL, R300_CLIP_DISABLE |
-            R300_PS_UCP_MODE_CLIP_AS_TRIFAN);
         OUT_CS_REG_SEQ(R300_VAP_GB_VERT_CLIP_ADJ, 4);
         OUT_CS_32F(1.0);
         OUT_CS_32F(1.0);
diff --git a/src/gallium/drivers/r300/r300_state_shader.c b/src/gallium/drivers/r300/r300_state_shader.c
index ed99c76c15..cc7f6a7c4b 100644
--- a/src/gallium/drivers/r300/r300_state_shader.c
+++ b/src/gallium/drivers/r300/r300_state_shader.c
@@ -22,24 +22,6 @@
 
 #include "r300_state_shader.h"
 
-static void r300_copy_passthrough_shader(struct r300_fragment_shader* fs)
-{
-    struct r300_fragment_shader* pt = &r300_passthrough_fragment_shader;
-    fs->shader.stack_size = pt->shader.stack_size;
-    fs->alu_instruction_count = pt->alu_instruction_count;
-    fs->tex_instruction_count = pt->tex_instruction_count;
-    fs->indirections = pt->indirections;
-    fs->instructions[0] = pt->instructions[0];
-}
-
-static void r500_copy_passthrough_shader(struct r500_fragment_shader* fs)
-{
-    struct r500_fragment_shader* pt = &r500_passthrough_fragment_shader;
-    fs->shader.stack_size = pt->shader.stack_size;
-    fs->instruction_count = pt->instruction_count;
-    fs->instructions[0] = pt->instructions[0];
-}
-
 static void r300_fs_declare(struct r300_fs_asm* assembler,
                             struct tgsi_full_declaration* decl)
 {
@@ -49,6 +31,7 @@ static void r300_fs_declare(struct r300_fs_asm* assembler,
                 case TGSI_SEMANTIC_COLOR:
                     assembler->color_count++;
                     break;
+                case TGSI_SEMANTIC_FOG:
                 case TGSI_SEMANTIC_GENERIC:
                     assembler->tex_count++;
                     break;
@@ -323,9 +306,12 @@ static INLINE void r300_emit_maths(struct r300_fragment_shader* fs,
 }
 
 /* Setup an ALU operation. */
-static INLINE void r500_emit_alu(struct r500_fragment_shader* fs,
-                                 struct r300_fs_asm* assembler,
-                                 struct tgsi_full_dst_register* dst)
+static INLINE void r500_emit_maths(struct r500_fragment_shader* fs,
+                                   struct r300_fs_asm* assembler,
+                                   struct tgsi_full_src_register* src,
+                                   struct tgsi_full_dst_register* dst,
+                                   unsigned op,
+                                   unsigned count)
 {
     int i = fs->instruction_count;
 
@@ -348,18 +334,6 @@ static INLINE void r500_emit_alu(struct r500_fragment_shader* fs,
         R500_ALPHA_ADDRD(r300_fs_dst(assembler, &dst->DstRegister));
     fs->instructions[i].inst5 =
         R500_ALU_RGBA_ADDRD(r300_fs_dst(assembler, &dst->DstRegister));
-}
-
-static INLINE void r500_emit_maths(struct r500_fragment_shader* fs,
-                                   struct r300_fs_asm* assembler,
-                                   struct tgsi_full_src_register* src,
-                                   struct tgsi_full_dst_register* dst,
-                                   unsigned op,
-                                   unsigned count)
-{
-    int i = fs->instruction_count;
-
-    r500_emit_alu(fs, assembler, dst);
 
     switch (count) {
         case 3:
@@ -381,8 +355,8 @@ static INLINE void r500_emit_maths(struct r500_fragment_shader* fs,
                 R500_ALU_RGB_SEL_B_SRC1 |
                 R500_SWIZ_RGB_B(r500_rgb_swiz(&src[1]));
             fs->instructions[i].inst4 |=
-                R500_SWIZ_ALPHA_B(r500_alpha_swiz(&src[1])) |
-                R500_ALPHA_SEL_B_SRC1;
+                R500_ALPHA_SEL_B_SRC1 |
+                R500_SWIZ_ALPHA_B(r500_alpha_swiz(&src[1]));
         case 1:
         case 0:
         default:
@@ -394,8 +368,8 @@ static INLINE void r500_emit_maths(struct r500_fragment_shader* fs,
                 R500_ALU_RGB_SEL_A_SRC0 |
                 R500_SWIZ_RGB_A(r500_rgb_swiz(&src[0]));
             fs->instructions[i].inst4 |=
-                R500_SWIZ_ALPHA_A(r500_alpha_swiz(&src[0])) |
-                R500_ALPHA_SEL_A_SRC0;
+                R500_ALPHA_SEL_A_SRC0 |
+                R500_SWIZ_ALPHA_A(r500_alpha_swiz(&src[0]));
             break;
     }
 
@@ -565,6 +539,60 @@ static void r500_fs_instruction(struct r500_fragment_shader* fs,
                     &inst->FullDstRegisters[0], inst->Instruction.Opcode, 3);
             break;
 
+        /* The compound and hybrid insts. */
+        case TGSI_OPCODE_LRP:
+            /* LRP DST A, B, C -> MAD TMP -A, C, C; MAD DST A, B, TMP */
+            inst->FullSrcRegisters[3] = inst->FullSrcRegisters[1];
+            inst->FullSrcRegisters[1] = inst->FullSrcRegisters[2];
+            inst->FullSrcRegisters[0].SrcRegister.Negate =
+                !(inst->FullSrcRegisters[0].SrcRegister.Negate);
+            inst->FullDstRegisters[1] = inst->FullDstRegisters[0];
+            inst->FullDstRegisters[0].DstRegister.Index =
+                assembler->temp_count;
+            inst->FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
+            r500_emit_maths(fs, assembler, inst->FullSrcRegisters,
+                    &inst->FullDstRegisters[0], TGSI_OPCODE_MAD, 3);
+            inst->FullSrcRegisters[2].SrcRegister.Index =
+                assembler->temp_count;
+            inst->FullSrcRegisters[2].SrcRegister.File = TGSI_FILE_TEMPORARY;
+            inst->FullSrcRegisters[2].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
+            inst->FullSrcRegisters[2].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
+            inst->FullSrcRegisters[2].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Z;
+            inst->FullSrcRegisters[2].SrcRegister.SwizzleW = TGSI_SWIZZLE_W;
+            inst->FullSrcRegisters[1] = inst->FullSrcRegisters[3];
+            inst->FullSrcRegisters[0].SrcRegister.Negate =
+                !(inst->FullSrcRegisters[0].SrcRegister.Negate);
+            inst->FullDstRegisters[0] = inst->FullDstRegisters[1];
+            r500_emit_maths(fs, assembler, inst->FullSrcRegisters,
+                    &inst->FullDstRegisters[0], TGSI_OPCODE_MAD, 3);
+            break;
+        case TGSI_OPCODE_POW:
+            /* POW DST A, B -> LG2 TMP A; MUL TMP TMP, B; EX2 DST TMP */
+            inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtSwizzleW =
+                inst->FullSrcRegisters[0].SrcRegisterExtSwz.ExtSwizzleX;
+            inst->FullSrcRegisters[0].SrcRegister.SwizzleW =
+                inst->FullSrcRegisters[0].SrcRegister.SwizzleX;
+            inst->FullDstRegisters[1] = inst->FullDstRegisters[0];
+            inst->FullDstRegisters[0].DstRegister.Index =
+                assembler->temp_count;
+            inst->FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
+            r500_emit_maths(fs, assembler, inst->FullSrcRegisters,
+                    &inst->FullDstRegisters[0], TGSI_OPCODE_LG2, 1);
+            inst->FullSrcRegisters[0].SrcRegister.Index =
+                assembler->temp_count;
+            inst->FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
+            inst->FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
+            inst->FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
+            inst->FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Z;
+            inst->FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_W;
+            inst->FullSrcRegisters[2] = r500_constant_zero;
+            r500_emit_maths(fs, assembler, inst->FullSrcRegisters,
+                    &inst->FullDstRegisters[0], TGSI_OPCODE_MUL, 3);
+            inst->FullDstRegisters[0] = inst->FullDstRegisters[1];
+            r500_emit_maths(fs, assembler, inst->FullSrcRegisters,
+                    &inst->FullDstRegisters[0], TGSI_OPCODE_EX2, 1);
+            break;
+
         /* The texture instruction set. */
         case TGSI_OPCODE_KIL:
         case TGSI_OPCODE_TEX:
@@ -593,7 +621,7 @@ static void r500_fs_instruction(struct r500_fragment_shader* fs,
 static void r300_fs_finalize(struct r3xx_fragment_shader* fs,
                              struct r300_fs_asm* assembler)
 {
-    fs->stack_size = assembler->temp_count + assembler->temp_offset;
+    fs->stack_size = assembler->temp_count + assembler->temp_offset + 1;
 }
 
 static void r500_fs_finalize(struct r500_fragment_shader* fs,
@@ -670,6 +698,7 @@ void r300_translate_fragment_shader(struct r300_context* r300,
             assembler->tex_count + assembler->color_count);
 
     consts->count = consts->user_count + assembler->imm_count;
+    fs->uses_imms = assembler->imm_count;
     debug_printf("r300: fs: %d total constants, "
             "%d from user and %d from immediates\n", consts->count,
             consts->user_count, assembler->imm_count);
diff --git a/src/gallium/drivers/r300/r300_state_shader.h b/src/gallium/drivers/r300/r300_state_shader.h
index 06260e61fe..b6087404ce 100644
--- a/src/gallium/drivers/r300/r300_state_shader.h
+++ b/src/gallium/drivers/r300/r300_state_shader.h
@@ -181,7 +181,7 @@ static struct r500_fragment_shader r500_texture_fragment_shader = {
     .instruction_count = 2,
     .instructions[0].inst0 = R500_INST_TYPE_TEX |
         R500_INST_TEX_SEM_WAIT |
-        R500_INST_RGB_OMASK_RGB | R500_INST_ALPHA_OMASK |
+        R500_INST_RGB_WMASK_RGB | R500_INST_ALPHA_WMASK |
         R500_INST_RGB_CLAMP | R500_INST_ALPHA_CLAMP,
     .instructions[0].inst1 = R500_TEX_ID(0) | R500_TEX_INST_LD |
         R500_TEX_SEM_ACQUIRE | R500_TEX_IGNORE_UNCOVERED,
diff --git a/src/gallium/drivers/r300/r300_state_tcl.c b/src/gallium/drivers/r300/r300_state_tcl.c
index fdbcbf3db8..8cf8250425 100644
--- a/src/gallium/drivers/r300/r300_state_tcl.c
+++ b/src/gallium/drivers/r300/r300_state_tcl.c
@@ -144,6 +144,7 @@ static uint32_t r300_vs_op(unsigned op)
             return R300_VE_MULTIPLY;
         case TGSI_OPCODE_ADD:
         case TGSI_OPCODE_MOV:
+        case TGSI_OPCODE_SUB:
         case TGSI_OPCODE_SWZ:
             return R300_VE_ADD;
         case TGSI_OPCODE_MAX:
@@ -163,12 +164,14 @@ static uint32_t r300_vs_op(unsigned op)
 static uint32_t r300_vs_swiz(struct tgsi_full_src_register* reg)
 {
     if (reg->SrcRegister.Extended) {
-        return reg->SrcRegisterExtSwz.ExtSwizzleX |
+        return (reg->SrcRegister.Negate ? (0xf << 12) : 0) |
+            reg->SrcRegisterExtSwz.ExtSwizzleX |
             (reg->SrcRegisterExtSwz.ExtSwizzleY << 3) |
             (reg->SrcRegisterExtSwz.ExtSwizzleZ << 6) |
             (reg->SrcRegisterExtSwz.ExtSwizzleW << 9);
     } else {
-        return reg->SrcRegister.SwizzleX |
+        return (reg->SrcRegister.Negate ? (0xf << 12) : 0) |
+            reg->SrcRegister.SwizzleX |
             (reg->SrcRegister.SwizzleY << 3) |
             (reg->SrcRegister.SwizzleZ << 6) |
             (reg->SrcRegister.SwizzleW << 9);
@@ -179,12 +182,14 @@ static uint32_t r300_vs_swiz(struct tgsi_full_src_register* reg)
 static uint32_t r300_vs_scalar_swiz(struct tgsi_full_src_register* reg)
 {
     if (reg->SrcRegister.Extended) {
-        return reg->SrcRegisterExtSwz.ExtSwizzleX |
+        return (reg->SrcRegister.Negate ? (0xf << 12) : 0) |
+            reg->SrcRegisterExtSwz.ExtSwizzleX |
             (reg->SrcRegisterExtSwz.ExtSwizzleX << 3) |
             (reg->SrcRegisterExtSwz.ExtSwizzleX << 6) |
             (reg->SrcRegisterExtSwz.ExtSwizzleX << 9);
     } else {
-        return reg->SrcRegister.SwizzleX |
+        return (reg->SrcRegister.Negate ? (0xf << 12) : 0) |
+            reg->SrcRegister.SwizzleX |
             (reg->SrcRegister.SwizzleX << 3) |
             (reg->SrcRegister.SwizzleX << 6) |
             (reg->SrcRegister.SwizzleX << 9);
@@ -246,6 +251,10 @@ static void r300_vs_instruction(struct r300_vertex_shader* vs,
                     &inst->FullDstRegisters[0], inst->Instruction.Opcode,
                     1, TRUE);
             break;
+        case TGSI_OPCODE_SUB:
+            inst->FullSrcRegisters[1].SrcRegister.Negate =
+                !inst->FullSrcRegisters[1].SrcRegister.Negate;
+            /* Fall through */
         case TGSI_OPCODE_ADD:
         case TGSI_OPCODE_MUL:
         case TGSI_OPCODE_MAX:
@@ -386,6 +395,7 @@ void r300_translate_vertex_shader(struct r300_context* r300,
             assembler->tex_count + assembler->color_count);
 
     consts->count = consts->user_count + assembler->imm_count;
+    vs->uses_imms = assembler->imm_count;
     debug_printf("r300: vs: %d total constants, "
             "%d from user and %d from immediates\n", consts->count,
             consts->user_count, assembler->imm_count);
diff --git a/src/gallium/drivers/r300/r300_state_tcl.h b/src/gallium/drivers/r300/r300_state_tcl.h
index d5d425e9d6..2c8b586c2f 100644
--- a/src/gallium/drivers/r300/r300_state_tcl.h
+++ b/src/gallium/drivers/r300/r300_state_tcl.h
@@ -76,6 +76,13 @@
     ((R300_PVS_SRC_SELECT_FORCE_1 | (R300_PVS_SRC_SELECT_FORCE_1 << 3) | \
      (R300_PVS_SRC_SELECT_FORCE_1 << 6) | \
       (R300_PVS_SRC_SELECT_FORCE_1 << 9)) << 13)
+#define R300_PVS_MODIFIER_X        (1 << 25)
+#define R300_PVS_MODIFIER_Y        (1 << 26)
+#define R300_PVS_MODIFIER_Z        (1 << 27)
+#define R300_PVS_MODIFIER_W        (1 << 28)
+#define R300_PVS_NEGATE_XYZW \
+    (R300_PVS_MODIFIER_X | R300_PVS_MODIFIER_Y | \
+     R300_PVS_MODIFIER_Z | R300_PVS_MODIFIER_W)
 
 static const struct tgsi_full_src_register r300_constant_zero = {
     .SrcRegister.Extended = TRUE,
diff --git a/src/gallium/drivers/r300/r300_surface.c b/src/gallium/drivers/r300/r300_surface.c
index acb6192492..c9e2dff14e 100644
--- a/src/gallium/drivers/r300/r300_surface.c
+++ b/src/gallium/drivers/r300/r300_surface.c
@@ -32,13 +32,6 @@ static void r300_surface_setup(struct r300_context* r300,
     unsigned pixpitch = dest->stride / dest->tex.block.size;
     CS_LOCALS(r300);
 
-    /* Make sure our target BO is okay. */
-    r300->winsys->add_buffer(r300->winsys, dest->buffer,
-            0, RADEON_GEM_DOMAIN_VRAM);
-    if (r300->winsys->validate(r300->winsys)) {
-        r300->context.flush(&r300->context, 0, NULL);
-    }
-
     r300_emit_blend_state(r300, &blend_clear_state);
     r300_emit_blend_color_state(r300, &blend_color_clear_state);
     r300_emit_dsa_state(r300, &dsa_clear_state);
@@ -106,6 +99,7 @@ static void r300_surface_fill(struct pipe_context* pipe,
     struct r300_capabilities* caps = r300_screen(pipe->screen)->caps;
     struct r300_texture* tex = (struct r300_texture*)dest->texture;
     unsigned pixpitch = tex->stride / tex->tex.block.size;
+    boolean invalid = FALSE;
     CS_LOCALS(r300);
 
     a = (float)((color >> 24) & 0xff) / 255.0f;
@@ -118,11 +112,28 @@ static void r300_surface_fill(struct pipe_context* pipe,
 
     /* Fallback? */
     if (FALSE) {
+fallback:
         debug_printf("r300: Falling back on surface clear...");
         util_surface_fill(pipe, dest, x, y, w, h, color);
         return;
     }
 
+    /* Make sure our target BO is okay. */
+validate:
+    if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
+                0, RADEON_GEM_DOMAIN_VRAM)) {
+        r300->context.flush(&r300->context, 0, NULL);
+        goto validate;
+    }
+    if (r300->winsys->validate(r300->winsys)) {
+        r300->context.flush(&r300->context, 0, NULL);
+        if (invalid) {
+            goto fallback;
+        }
+        invalid = TRUE;
+        goto validate;
+    }
+
     r300_surface_setup(r300, tex, x, y, w, h);
 
     /* Vertex shader setup */
@@ -216,6 +227,7 @@ static void r300_surface_copy(struct pipe_context* pipe,
     struct r300_texture* srctex = (struct r300_texture*)src->texture;
     struct r300_texture* desttex = (struct r300_texture*)dest->texture;
     unsigned pixpitch = srctex->stride / srctex->tex.block.size;
+    boolean invalid = FALSE;
     CS_LOCALS(r300);
 
     debug_printf("r300: Copying surface %p at (%d,%d) to %p at (%d, %d),"
@@ -225,21 +237,44 @@ static void r300_surface_copy(struct pipe_context* pipe,
     if ((srctex == desttex) &&
             ((destx < srcx + w) || (srcx < destx + w)) &&
             ((desty < srcy + h) || (srcy < desty + h))) {
+fallback:
         debug_printf("r300: Falling back on surface_copy\n");
         util_surface_copy(pipe, FALSE, dest, destx, desty, src,
                 srcx, srcy, w, h);
     }
 
-    /* Add our source texture to the BO list before emitting anything.
-     * r300_surface_setup will flush if needed for us. */
-    r300->winsys->add_buffer(r300->winsys, srctex->buffer,
-            RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0);
+    /* Add our target BOs to the list. */
+validate:
+    if (!r300->winsys->add_buffer(r300->winsys, srctex->buffer,
+                RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0)) {
+        r300->context.flush(&r300->context, 0, NULL);
+        goto validate;
+    }
+    if (!r300->winsys->add_buffer(r300->winsys, desttex->buffer,
+                0, RADEON_GEM_DOMAIN_VRAM)) {
+        r300->context.flush(&r300->context, 0, NULL);
+        goto validate;
+    }
+    if (r300->winsys->validate(r300->winsys)) {
+        r300->context.flush(&r300->context, 0, NULL);
+        if (invalid) {
+            goto fallback;
+        }
+        invalid = TRUE;
+        goto validate;
+    }
 
     r300_surface_setup(r300, desttex, destx, desty, w, h);
 
+    /* Setup the texture. */
     r300_emit_sampler(r300, &r300_sampler_copy_state, 0);
     r300_emit_texture(r300, srctex, 0);
-    r300_flush_textures(r300);
+
+    /* Flush and enable. */
+    BEGIN_CS(4);
+    OUT_CS_REG(R300_TX_INVALTAGS, 0);
+    OUT_CS_REG(R300_TX_ENABLE, 0x1);
+    END_CS;
 
     /* Vertex shader setup */
     if (caps->has_tcl) {
@@ -263,7 +298,7 @@ static void r300_surface_copy(struct pipe_context* pipe,
         r300_emit_rs_block_state(r300, &r300_rs_block_copy_state);
     }
 
-    BEGIN_CS(28);
+    BEGIN_CS(30);
     /* VAP stream control, mapping from input memory to PVS/RS memory */
     if (caps->has_tcl) {
         OUT_CS_REG(R300_VAP_PROG_STREAM_CNTL_0,
@@ -287,7 +322,7 @@ static void r300_surface_copy(struct pipe_context* pipe,
     OUT_CS_REG(R300_VAP_OUTPUT_VTX_FMT_1, 0x2);
 
     /* Vertex size. */
-    OUT_CS_REG(R300_VAP_VTX_SIZE, 0x8);
+    OUT_CS_REG(R300_VAP_VTX_SIZE, 0x4);
 
     /* Packet3 with our texcoords */
     OUT_CS_PKT3(R200_3D_DRAW_IMMD_2, 16);
diff --git a/src/gallium/drivers/r300/r300_surface.h b/src/gallium/drivers/r300/r300_surface.h
index 894def07aa..9a4c39f58b 100644
--- a/src/gallium/drivers/r300/r300_surface.h
+++ b/src/gallium/drivers/r300/r300_surface.h
@@ -101,7 +101,7 @@ static struct r300_rs_block r300_rs_block_copy_state = {
         R500_RS_SEL_Q(R300_RS_SEL_K1),
     .inst[0] = R300_RS_INST_COL_CN_WRITE,
     .count = R300_IT_COUNT(2) | R300_IC_COUNT(0) | R300_HIRES_EN,
-    .inst_count = R300_RS_TX_OFFSET(6),
+    .inst_count = R300_RS_TX_OFFSET(0),
 };
 
 static struct r300_rs_block r500_rs_block_copy_state = {
@@ -111,7 +111,7 @@ static struct r300_rs_block r500_rs_block_copy_state = {
         R500_RS_SEL_Q(R500_RS_IP_PTR_K1),
     .inst[0] = R500_RS_INST_TEX_CN_WRITE,
     .count = R300_IT_COUNT(2) | R300_IC_COUNT(0) | R300_HIRES_EN,
-    .inst_count = R300_RS_TX_OFFSET(6),
+    .inst_count = R300_RS_TX_OFFSET(0),
 };
 
 static struct r300_sampler_state r300_sampler_copy_state = {
diff --git a/src/gallium/drivers/r300/r300_winsys.h b/src/gallium/drivers/r300/r300_winsys.h
index a5ced8041c..d2893c3b9d 100644
--- a/src/gallium/drivers/r300/r300_winsys.h
+++ b/src/gallium/drivers/r300/r300_winsys.h
@@ -52,10 +52,10 @@ struct r300_winsys {
     uint32_t vram_size;
 
     /* Add a pipe_buffer to the list of buffer objects to validate. */
-    void (*add_buffer)(struct r300_winsys* winsys,
-                       struct pipe_buffer* pbuffer,
-                       uint32_t rd,
-                       uint32_t wd);
+    boolean (*add_buffer)(struct r300_winsys* winsys,
+                          struct pipe_buffer* pbuffer,
+                          uint32_t rd,
+                          uint32_t wd);
 
     /* Revalidate all currently setup pipe_buffers.
      * Returns TRUE if a flush is required. */