17 files changed, 370 insertions, 143 deletions

diff --git a/src/gallium/drivers/r300/r300_blit.c b/src/gallium/drivers/r300/r300_blit.c
index 6f8d9abfc8..47ffc0cb3c 100644
--- a/src/gallium/drivers/r300/r300_blit.c
+++ b/src/gallium/drivers/r300/r300_blit.c
@@ -21,6 +21,8 @@
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
 #include "r300_context.h"
+#include "r300_emit.h"
+#include "r300_hyperz.h"
 #include "r300_texture.h"
 #include "r300_winsys.h"
 
@@ -99,9 +101,6 @@ static boolean r300_cbzb_clear_allowed(struct r300_context *r300,
     struct pipe_framebuffer_state *fb =
         (struct pipe_framebuffer_state*)r300->fb_state.state;
 
-    if (r300->z_fastfill)
-        clear_buffers &= ~(PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL);
-
     /* Only color clear allowed, and only one colorbuffer. */
     if (clear_buffers != PIPE_CLEAR_COLOR || fb->nr_cbufs != 1)
         return FALSE;
@@ -173,22 +172,25 @@ static void r300_clear(struct pipe_context* pipe,
         (struct pipe_framebuffer_state*)r300->fb_state.state;
     struct r300_hyperz_state *hyperz =
         (struct r300_hyperz_state*)r300->hyperz_state.state;
+    struct r300_texture *zstex =
+            fb->zsbuf ? r300_texture(fb->zsbuf->texture) : NULL;
     uint32_t width = fb->width;
     uint32_t height = fb->height;
     boolean has_hyperz = r300->rws->get_value(r300->rws, R300_CAN_HYPERZ);
-    uint32_t hyperz_dcv = 0;
+    uint32_t hyperz_dcv = hyperz->zb_depthclearvalue;
 
     /* Enable fast Z clear.
      * The zbuffer must be in micro-tiled mode, otherwise it locks up. */
-    if ((buffers & (PIPE_CLEAR_DEPTH|PIPE_CLEAR_STENCIL)) && has_hyperz) {
-      
+    if ((buffers & PIPE_CLEAR_DEPTHSTENCIL) && has_hyperz) {
         hyperz_dcv = hyperz->zb_depthclearvalue =
             r300_depth_clear_value(fb->zsbuf->format, depth, stencil);
 
         r300_mark_fb_state_dirty(r300, R300_CHANGED_ZCLEAR_FLAG);
-        if (r300->z_compression || r300->z_fastfill)
+        if (zstex->zmask_mem[fb->zsbuf->level]) {
             r300->zmask_clear.dirty = TRUE;
-        if (r300->hiz_enable)
+            buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
+        }
+        if (zstex->hiz_mem[fb->zsbuf->level])
             r300->hiz_clear.dirty = TRUE;
     }
 
@@ -207,13 +209,43 @@ static void r300_clear(struct pipe_context* pipe,
     }
 
     /* Clear. */
-    r300_blitter_begin(r300, R300_CLEAR);
-    util_blitter_clear(r300->blitter,
-                       width,
-                       height,
-                       fb->nr_cbufs,
-                       buffers, rgba, depth, stencil);
-    r300_blitter_end(r300);
+    if (buffers) {
+        /* Clear using the blitter. */
+        r300_blitter_begin(r300, R300_CLEAR);
+        util_blitter_clear(r300->blitter,
+                           width,
+                           height,
+                           fb->nr_cbufs,
+                           buffers, rgba, depth, stencil);
+        r300_blitter_end(r300);
+    } else if (r300->zmask_clear.dirty) {
+        /* Just clear zmask and hiz now, this does not use a standard draw
+         * procedure. */
+        unsigned dwords;
+
+        /* Calculate zmask_clear and hiz_clear atom sizes. */
+        r300_update_hyperz_state(r300);
+        dwords = r300->zmask_clear.size +
+                 (r300->hiz_clear.dirty ? r300->hiz_clear.size : 0) +
+                 r300_get_num_cs_end_dwords(r300);
+
+        /* Reserve CS space. */
+        if (dwords > (r300->cs->ndw - r300->cs->cdw)) {
+            r300->context.flush(&r300->context, 0, NULL);
+        }
+
+        /* Emit clear packets. */
+        r300_emit_zmask_clear(r300, r300->zmask_clear.size,
+                              r300->zmask_clear.state);
+        r300->zmask_clear.dirty = FALSE;
+        if (r300->hiz_clear.dirty) {
+            r300_emit_hiz_clear(r300, r300->hiz_clear.size,
+                                r300->hiz_clear.state);
+            r300->hiz_clear.dirty = FALSE;
+        }
+    } else {
+        assert(0);
+    }
 
     /* Disable CBZB clear. */
     if (r300->cbzb_clear) {
@@ -222,6 +254,16 @@ static void r300_clear(struct pipe_context* pipe,
         r300_mark_fb_state_dirty(r300, R300_CHANGED_CBZB_FLAG);
     }
 
+    /* Enable fastfill and/or hiz.
+     *
+     * If we cleared zmask/hiz, it's in use now. The Hyper-Z state update
+     * looks if zmask/hiz is in use and enables fastfill accordingly. */
+    if (zstex &&
+        (zstex->zmask_in_use[fb->zsbuf->level] ||
+         zstex->hiz_in_use[fb->zsbuf->level])) {
+        r300->hyperz_state.dirty = TRUE;
+    }
+
     /* XXX this flush "fixes" a hardlock in the cubestorm xscreensaver */
     if (r300->flush_counter == 0)
         pipe->flush(pipe, 0, NULL);
@@ -259,27 +301,31 @@ static void r300_clear_depth_stencil(struct pipe_context *pipe,
     r300_blitter_end(r300);
 }
 
-/* Clear a region of a depth stencil surface. */
-static void r300_flush_depth_stencil(struct pipe_context *pipe,
-                                     struct pipe_resource *dst,
-                                     struct pipe_subresource subdst)
+/* Flush a depth stencil buffer. */
+void r300_flush_depth_stencil(struct pipe_context *pipe,
+                              struct pipe_resource *dst,
+                              struct pipe_subresource subdst,
+                              unsigned zslice)
 {
     struct r300_context *r300 = r300_context(pipe);
     struct pipe_surface *dstsurf;
     struct r300_texture *tex = r300_texture(dst);
 
-    /* only flush the zmask if we have one attached to this texture */
     if (!tex->zmask_mem[subdst.level])
         return;
+    if (!tex->zmask_in_use[subdst.level])
+        return;
 
     dstsurf = pipe->screen->get_tex_surface(pipe->screen, dst,
-                                            subdst.face, subdst.level, 0,
+                                            subdst.face, subdst.level, zslice,
                                             PIPE_BIND_DEPTH_STENCIL);
     r300->z_decomp_rd = TRUE;
     r300_blitter_begin(r300, R300_CLEAR_SURFACE);
     util_blitter_flush_depth_stencil(r300->blitter, dstsurf);
     r300_blitter_end(r300);
     r300->z_decomp_rd = FALSE;
+
+    tex->zmask_in_use[subdst.level] = FALSE;
 }
 
 /* Copy a block of pixels from one surface to another using HW. */
@@ -342,7 +388,7 @@ static void r300_resource_copy_region(struct pipe_context *pipe,
 
     is_depth = util_format_get_component_bits(src->format, UTIL_FORMAT_COLORSPACE_ZS, 0) != 0;
     if (is_depth) {
-        r300_flush_depth_stencil(pipe, src, subsrc);
+        r300_flush_depth_stencil(pipe, src, subsrc, srcz);
     }
     if (old_format != new_format) {
         dst->format = new_format;
diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
index e8b6c4f7af..a83ad892ea 100644
--- a/src/gallium/drivers/r300/r300_context.c
+++ b/src/gallium/drivers/r300/r300_context.c
@@ -99,8 +99,10 @@ static void r300_destroy_context(struct pipe_context* context)
     struct r300_context* r300 = r300_context(context);
     struct r300_atom *atom;
 
-    util_blitter_destroy(r300->blitter);
-    draw_destroy(r300->draw);
+    if (r300->blitter)
+        util_blitter_destroy(r300->blitter);
+    if (r300->draw)
+        draw_destroy(r300->draw);
 
     /* Print stats, if enabled. */
     if (SCREEN_DBG_ON(r300->screen, DBG_STATS)) {
@@ -112,40 +114,48 @@ static void r300_destroy_context(struct pipe_context* context)
         }
     }
 
-    u_upload_destroy(r300->upload_vb);
-    u_upload_destroy(r300->upload_ib);
+    if (r300->upload_vb)
+        u_upload_destroy(r300->upload_vb);
+    if (r300->upload_ib)
+        u_upload_destroy(r300->upload_ib);
 
-    /* setup hyper-z mm */
-    if (r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
+    if (r300->zmask_mm)
         r300_hyperz_destroy_mm(r300);
 
-    translate_cache_destroy(r300->tran.translate_cache);
+    if (r300->tran.translate_cache)
+        translate_cache_destroy(r300->tran.translate_cache);
 
+    /* XXX: This function assumes r300->query_list was initialized */
     r300_release_referenced_objects(r300);
 
-    r300->rws->cs_destroy(r300->cs);
+    if (r300->cs)
+        r300->rws->cs_destroy(r300->cs);
 
+    /* XXX: No way to tell if this was initialized or not? */
     util_mempool_destroy(&r300->pool_transfers);
 
     r300_update_num_contexts(r300->screen, -1);
 
-    FREE(r300->aa_state.state);
-    FREE(r300->blend_color_state.state);
-    FREE(r300->clip_state.state);
-    FREE(r300->fb_state.state);
-    FREE(r300->gpu_flush.state);
-    FREE(r300->hyperz_state.state);
-    FREE(r300->invariant_state.state);
-    FREE(r300->rs_block_state.state);
-    FREE(r300->scissor_state.state);
-    FREE(r300->textures_state.state);
-    FREE(r300->vap_invariant_state.state);
-    FREE(r300->viewport_state.state);
-    FREE(r300->ztop_state.state);
-    FREE(r300->fs_constants.state);
-    FREE(r300->vs_constants.state);
-    if (!r300->screen->caps.has_tcl) {
-        FREE(r300->vertex_stream_state.state);
+    /* Free the structs allocated in r300_setup_atoms() */
+    if (r300->aa_state.state) {
+        FREE(r300->aa_state.state);
+        FREE(r300->blend_color_state.state);
+        FREE(r300->clip_state.state);
+        FREE(r300->fb_state.state);
+        FREE(r300->gpu_flush.state);
+        FREE(r300->hyperz_state.state);
+        FREE(r300->invariant_state.state);
+        FREE(r300->rs_block_state.state);
+        FREE(r300->scissor_state.state);
+        FREE(r300->textures_state.state);
+        FREE(r300->vap_invariant_state.state);
+        FREE(r300->viewport_state.state);
+        FREE(r300->ztop_state.state);
+        FREE(r300->fs_constants.state);
+        FREE(r300->vs_constants.state);
+        if (!r300->screen->caps.has_tcl) {
+            FREE(r300->vertex_stream_state.state);
+        }
     }
     FREE(r300);
 }
@@ -158,12 +168,14 @@ void r300_flush_cb(void *data)
 }
 
 #define R300_INIT_ATOM(atomname, atomsize) \
+ do { \
     r300->atomname.name = #atomname; \
     r300->atomname.state = NULL; \
     r300->atomname.size = atomsize; \
     r300->atomname.emit = r300_emit_##atomname; \
     r300->atomname.dirty = FALSE; \
-    insert_at_tail(&r300->atom_list, &r300->atomname);
+    insert_at_tail(&r300->atom_list, &r300->atomname); \
+ } while (0)
 
 static void r300_setup_atoms(struct r300_context* r300)
 {
@@ -404,12 +416,16 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     r300->context.destroy = r300_destroy_context;
 
-    r300->cs = rws->cs_create(rws);
+    make_empty_list(&r300->query_list);
 
     util_mempool_create(&r300->pool_transfers,
                         sizeof(struct pipe_transfer), 64,
                         UTIL_MEMPOOL_SINGLETHREADED);
 
+    r300->cs = rws->cs_create(rws);
+    if (r300->cs == NULL)
+        goto fail;
+
     if (!r300screen->caps.has_tcl) {
         /* Create a Draw. This is used for SW TCL. */
         r300->draw = draw_create(&r300->context);
@@ -424,8 +440,6 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     r300_setup_atoms(r300);
 
-    make_empty_list(&r300->query_list);
-
     r300_init_blit_functions(r300);
     r300_init_flush_functions(r300);
     r300_init_query_functions(r300);
@@ -433,6 +447,8 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
     r300_init_resource_functions(r300);
 
     r300->blitter = util_blitter_create(&r300->context);
+    if (r300->blitter == NULL)
+        goto fail;
 
     /* Render functions must be initialized after blitter. */
     r300_init_render_functions(r300);
@@ -441,22 +457,25 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     /* setup hyper-z mm */
     if (r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
-        r300_hyperz_init_mm(r300);
+        if (!r300_hyperz_init_mm(r300))
+            goto fail;
 
     r300->upload_ib = u_upload_create(&r300->context,
 				      32 * 1024, 16,
 				      PIPE_BIND_INDEX_BUFFER);
 
     if (r300->upload_ib == NULL)
-        goto no_upload_ib;
+        goto fail;
 
     r300->upload_vb = u_upload_create(&r300->context,
 				      128 * 1024, 16,
 				      PIPE_BIND_VERTEX_BUFFER);
     if (r300->upload_vb == NULL)
-        goto no_upload_vb;
+        goto fail;
 
     r300->tran.translate_cache = translate_cache_create();
+    if (r300->tran.translate_cache == NULL)
+        goto fail;
 
     r300_init_states(&r300->context);
 
@@ -486,10 +505,8 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     return &r300->context;
 
- no_upload_ib:
-    u_upload_destroy(r300->upload_ib);
- no_upload_vb:
-    FREE(r300);
+ fail:
+    r300_destroy_context(&r300->context);
     return NULL;
 }
 
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index d86a5c8fc9..6fa7f470f9 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -397,6 +397,8 @@ struct r300_texture {
     /* hyper-z memory allocs */
     struct mem_block *hiz_mem[R300_MAX_TEXTURE_LEVELS];
     struct mem_block *zmask_mem[R300_MAX_TEXTURE_LEVELS];
+    boolean zmask_in_use[R300_MAX_TEXTURE_LEVELS];
+    boolean hiz_in_use[R300_MAX_TEXTURE_LEVELS];
 
     /* This is the level tiling flags were last time set for.
      * It's used to prevent redundant tiling-flags changes from happening.*/
@@ -564,12 +566,9 @@ struct r300_context {
     boolean two_sided_color;
     /* Incompatible vertex buffer layout? (misaligned stride or buffer_offset) */
     boolean incompatible_vb_layout;
-    /* Whether fast zclear is enabled. */
-    boolean z_fastfill;
 #define R300_Z_COMPRESS_44 1
 #define RV350_Z_COMPRESS_88 2
     int z_compression;
-    boolean hiz_enable;
     boolean cbzb_clear;
     boolean z_decomp_rd;
 
@@ -628,6 +627,12 @@ void r300_init_render_functions(struct r300_context *r300);
 void r300_init_state_functions(struct r300_context* r300);
 void r300_init_resource_functions(struct r300_context* r300);
 
+/* r300_blit.c */
+void r300_flush_depth_stencil(struct pipe_context *pipe,
+                              struct pipe_resource *dst,
+                              struct pipe_subresource subdst,
+                              unsigned zslice);
+
 /* r300_query.c */
 void r300_resume_query(struct r300_context *r300,
                        struct r300_query *query);
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index 17e180a79a..d0fd45349e 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -393,7 +393,7 @@ void r300_emit_fb_state(struct r300_context* r300, unsigned size, void* state)
             /* HiZ RAM. */
             if (r300->screen->caps.hiz_ram) {
                 if (tex->hiz_mem[level]) {
-                    OUT_CS_REG(R300_ZB_HIZ_OFFSET, tex->hiz_mem[level]->ofs);
+                    OUT_CS_REG(R300_ZB_HIZ_OFFSET, tex->hiz_mem[level]->ofs << 2);
                     OUT_CS_REG(R300_ZB_HIZ_PITCH, surf_pitch);
                 } else {
                     OUT_CS_REG(R300_ZB_HIZ_OFFSET, 0);
@@ -402,7 +402,7 @@ void r300_emit_fb_state(struct r300_context* r300, unsigned size, void* state)
             }
             /* Z Mask RAM. (compressed zbuffer) */
             if (tex->zmask_mem[level]) {
-                OUT_CS_REG(R300_ZB_ZMASK_OFFSET, tex->zmask_mem[level]->ofs);
+                OUT_CS_REG(R300_ZB_ZMASK_OFFSET, tex->zmask_mem[level]->ofs << 2);
                 OUT_CS_REG(R300_ZB_ZMASK_PITCH, surf_pitch);
             } else {
                 OUT_CS_REG(R300_ZB_ZMASK_OFFSET, 0);
@@ -936,6 +936,22 @@ void r300_emit_vs_state(struct r300_context* r300, unsigned size, void* state)
             OUT_CS_TABLE(data, 4);
         }
     }
+
+    /* Emit flow control instructions. */
+    if (code->num_fc_ops) {
+
+        OUT_CS_REG(R300_VAP_PVS_FLOW_CNTL_OPC, code->fc_ops);
+        if (r300screen->caps.is_r500) {
+            OUT_CS_REG_SEQ(R500_VAP_PVS_FLOW_CNTL_ADDRS_LW_0, code->num_fc_ops * 2);
+            OUT_CS_TABLE(code->fc_op_addrs.r500, code->num_fc_ops * 2);
+        } else {
+            OUT_CS_REG_SEQ(R300_VAP_PVS_FLOW_CNTL_ADDRS_0, code->num_fc_ops);
+            OUT_CS_TABLE(code->fc_op_addrs.r300, code->num_fc_ops);
+        }
+        OUT_CS_REG_SEQ(R300_VAP_PVS_FLOW_CNTL_LOOP_INDEX_0, code->num_fc_ops);
+        OUT_CS_TABLE(code->fc_loop_index, code->num_fc_ops);
+    }
+
     END_CS;
 }
 
@@ -1008,6 +1024,8 @@ void r300_emit_hiz_clear(struct r300_context *r300, unsigned size, void *state)
     int i;
 
     tex = r300_texture(fb->zsbuf->texture);
+
+    offset = tex->hiz_mem[fb->zsbuf->level]->ofs;
     stride = tex->desc.stride_in_pixels[fb->zsbuf->level];
 
     /* convert from pixels to 4x4 blocks */
@@ -1028,6 +1046,9 @@ void r300_emit_hiz_clear(struct r300_context *r300, unsigned size, void *state)
         r300_emit_hiz_line_clear(r300, offset, stride, 0xffffffff);
     }
     z->current_func = -1;
+
+    /* Mark the current zbuffer's hiz ram as in use. */
+    tex->hiz_in_use[fb->zsbuf->level] = TRUE;
 }
 
 void r300_emit_zmask_clear(struct r300_context *r300, unsigned size, void *state)
@@ -1043,6 +1064,8 @@ void r300_emit_zmask_clear(struct r300_context *r300, unsigned size, void *state
     tex = r300_texture(fb->zsbuf->texture);
     stride = tex->desc.stride_in_pixels[fb->zsbuf->level];
 
+    offset = tex->zmask_mem[fb->zsbuf->level]->ofs;
+
     if (r300->z_compression == RV350_Z_COMPRESS_88)
         mult = 8;
     else
@@ -1065,6 +1088,9 @@ void r300_emit_zmask_clear(struct r300_context *r300, unsigned size, void *state
         offset <<= offset_shift;
         r300_emit_zmask_line_clear(r300, offset, stride, 0x0);//0xffffffff);
     }
+
+    /* Mark the current zbuffer's zmask as in use. */
+    tex->zmask_in_use[fb->zsbuf->level] = TRUE;
 }
 
 void r300_emit_ztop_state(struct r300_context* r300,
@@ -1186,6 +1212,17 @@ unsigned r300_get_num_dirty_dwords(struct r300_context *r300)
     return dwords;
 }
 
+unsigned r300_get_num_cs_end_dwords(struct r300_context *r300)
+{
+    unsigned dwords = 0;
+
+    /* Emitted in flush. */
+    dwords += 26; /* emit_query_end */
+    dwords += r300->hyperz_state.size + 2; /* emit_hyperz_end + zcache flush */
+
+    return dwords;
+}
+
 /* Emit all dirty state. */
 void r300_emit_dirty_state(struct r300_context* r300)
 {
diff --git a/src/gallium/drivers/r300/r300_emit.h b/src/gallium/drivers/r300/r300_emit.h
index 2f2c2f2dcb..bae2525634 100644
--- a/src/gallium/drivers/r300/r300_emit.h
+++ b/src/gallium/drivers/r300/r300_emit.h
@@ -116,6 +116,7 @@ void r300_emit_hiz_clear(struct r300_context *r300, unsigned size, void *state);
 void r300_emit_zmask_clear(struct r300_context *r300, unsigned size, void *state);
 
 unsigned r300_get_num_dirty_dwords(struct r300_context *r300);
+unsigned r300_get_num_cs_end_dwords(struct r300_context *r300);
 
 /* Emit all dirty state. */
 void r300_emit_dirty_state(struct r300_context* r300);
diff --git a/src/gallium/drivers/r300/r300_flush.c b/src/gallium/drivers/r300/r300_flush.c
index 7fed9b5d07..fe182b6615 100644
--- a/src/gallium/drivers/r300/r300_flush.c
+++ b/src/gallium/drivers/r300/r300_flush.c
@@ -44,8 +44,7 @@ static void r300_flush(struct pipe_context* pipe,
     u_upload_flush(r300->upload_ib);
 
     if (r300->dirty_hw) {
-        if (r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
-            r300_emit_hyperz_end(r300);
+        r300_emit_hyperz_end(r300);
         r300_emit_query_end(r300);
 
         r300->flush_counter++;
diff --git a/src/gallium/drivers/r300/r300_fs.c b/src/gallium/drivers/r300/r300_fs.c
index 87ff49a90c..2a0c30620a 100644
--- a/src/gallium/drivers/r300/r300_fs.c
+++ b/src/gallium/drivers/r300/r300_fs.c
@@ -72,6 +72,11 @@ void r300_shader_read_fs_inputs(struct tgsi_shader_info* info,
                 fs_inputs->wpos = i;
                 break;
 
+            case TGSI_SEMANTIC_FACE:
+                assert(index == 0);
+                fs_inputs->face = i;
+                break;
+
             default:
                 fprintf(stderr, "r300: FP: Unknown input semantic: %i\n",
                         info->input_semantic_name[i]);
@@ -120,6 +125,9 @@ static void allocate_hardware_inputs(
             allocate(mydata, inputs->color[i], reg++);
         }
     }
+    if (inputs->face != ATTR_UNUSED) {
+        allocate(mydata, inputs->face, reg++);
+    }
     for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
         if (inputs->generic[i] != ATTR_UNUSED) {
             allocate(mydata, inputs->generic[i], reg++);
@@ -360,13 +368,14 @@ static void r300_translate_fragment_shader(
 {
     struct r300_fragment_program_compiler compiler;
     struct tgsi_to_rc ttr;
-    int wpos;
+    int wpos, face;
     unsigned i;
 
     tgsi_scan_shader(tokens, &shader->info);
     r300_shader_read_fs_inputs(&shader->info, &shader->inputs);
 
     wpos = shader->inputs.wpos;
+    face = shader->inputs.face;
 
     /* Setup the compiler. */
     memset(&compiler, 0, sizeof(compiler));
@@ -383,7 +392,7 @@ static void r300_translate_fragment_shader(
     find_output_registers(&compiler, shader);
 
     if (compiler.Base.Debug) {
-        debug_printf("r300: Initial fragment program\n");
+        DBG(r300, DBG_FP, "r300: Initial fragment program\n");
         tgsi_dump(tokens, 0);
     }
 
@@ -406,6 +415,10 @@ static void r300_translate_fragment_shader(
         rc_transform_fragment_wpos(&compiler.Base, wpos, wpos, TRUE);
     }
 
+    if (face != ATTR_UNUSED) {
+        rc_transform_fragment_face(&compiler.Base, face);
+    }
+
     /* Invoke the compiler */
     r3xx_compile_fragment_program(&compiler);
 
@@ -418,7 +431,7 @@ static void r300_translate_fragment_shader(
     }
 
     if (compiler.Base.Error) {
-        fprintf(stderr, "r300 FP: Compiler Error:\n%sUsing a dummy shader"
+        DBG(r300, DBG_FP, "r300 FP: Compiler Error:\n%sUsing a dummy shader"
                 " instead.\nIf there's an 'unknown opcode' message, please"
                 " file a bug report and attach this log.\n", compiler.Base.ErrorMsg);
 
diff --git a/src/gallium/drivers/r300/r300_hyperz.c b/src/gallium/drivers/r300/r300_hyperz.c
index 10e440ce30..a471b7353b 100644
--- a/src/gallium/drivers/r300/r300_hyperz.c
+++ b/src/gallium/drivers/r300/r300_hyperz.c
@@ -21,12 +21,14 @@
  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
 
-#include "util/u_format.h"
-#include "util/u_mm.h"
 #include "r300_context.h"
 #include "r300_hyperz.h"
 #include "r300_reg.h"
 #include "r300_fs.h"
+#include "r300_winsys.h"
+
+#include "util/u_format.h"
+#include "util/u_mm.h"
 
 /*
   HiZ rules - taken from various docs 
@@ -127,6 +129,12 @@ static void r300_update_hyperz(struct r300_context* r300)
 {
     struct r300_hyperz_state *z =
         (struct r300_hyperz_state*)r300->hyperz_state.state;
+    struct pipe_framebuffer_state *fb =
+        (struct pipe_framebuffer_state*)r300->fb_state.state;
+    struct r300_texture *zstex =
+            fb->zsbuf ? r300_texture(fb->zsbuf->texture) : NULL;
+    boolean zmask_in_use = FALSE;
+    boolean hiz_in_use = FALSE;
 
     z->gb_z_peq_config = 0;
     z->zb_bw_cntl = 0;
@@ -138,22 +146,32 @@ static void r300_update_hyperz(struct r300_context* r300)
         return;
     }
 
+    if (!zstex)
+        return;
+
+    if (!r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
+        return;
+
+    zmask_in_use = zstex->zmask_in_use[fb->zsbuf->level];
+    hiz_in_use = zstex->hiz_in_use[fb->zsbuf->level];
+
+    /* Z fastfill. */
+    if (zmask_in_use) {
+        z->zb_bw_cntl |= R300_FAST_FILL_ENABLE; /*  | R300_FORCE_COMPRESSED_STENCIL_VALUE_ENABLE;*/
+    }
+
     /* Zbuffer compression. */
-    if (r300->z_compression) {
+    if (zmask_in_use && r300->z_compression) {
         z->zb_bw_cntl |= R300_RD_COMP_ENABLE;
         if (r300->z_decomp_rd == false)
             z->zb_bw_cntl |= R300_WR_COMP_ENABLE;
-       /* RV350 and up optimizations. */
-       if (r300->z_compression == RV350_Z_COMPRESS_88)
-           z->gb_z_peq_config |= R300_GB_Z_PEQ_CONFIG_Z_PEQ_SIZE_8_8;
-    }
-
-    /* Z fastfill. */
-    if (r300->z_fastfill) {
-        z->zb_bw_cntl |= R300_FAST_FILL_ENABLE; /*  | R300_FORCE_COMPRESSED_STENCIL_VALUE_ENABLE;*/
     }
+    /* RV350 and up optimizations. */
+    /* The section 10.4.9 in the docs is a lie. */
+    if (r300->z_compression == RV350_Z_COMPRESS_88)
+        z->gb_z_peq_config |= R300_GB_Z_PEQ_CONFIG_Z_PEQ_SIZE_8_8;
 
-    if (r300->hiz_enable) {
+    if (hiz_in_use) {
         bool can_hiz = r300_can_hiz(r300);
         if (can_hiz) {
             z->zb_bw_cntl |= R300_HIZ_ENABLE;
@@ -163,8 +181,8 @@ static void r300_update_hyperz(struct r300_context* r300)
         }
     }
 
+    /* R500-specific features and optimizations. */
     if (r300->screen->caps.is_r500) {
-        /* XXX Are these bits really available on RV350? */
         z->zb_bw_cntl |= R500_HIZ_FP_EXP_BITS_3;
         z->zb_bw_cntl |=
                 R500_HIZ_EQUAL_REJECT_ENABLE |
@@ -333,6 +351,12 @@ void r300_zmask_alloc_block(struct r300_context *r300, struct r300_surface *surf
 
     tex = r300_texture(surf->base.texture);
 
+    /* We currently don't handle decompression for 3D textures and cubemaps
+     * correctly. */
+    if (tex->desc.b.b.target != PIPE_TEXTURE_1D &&
+        tex->desc.b.b.target != PIPE_TEXTURE_2D)
+        return;
+
     if (tex->zmask_mem[level])
         return;
 
@@ -349,23 +373,36 @@ void r300_zmask_alloc_block(struct r300_context *r300, struct r300_surface *surf
     return;
 }
 
-void r300_hyperz_init_mm(struct r300_context *r300)
+boolean r300_hyperz_init_mm(struct r300_context *r300)
 {
     struct r300_screen* r300screen = r300->screen;
     int frag_pipes = r300screen->caps.num_frag_pipes;
 
-    if (r300screen->caps.hiz_ram)
+    r300->zmask_mm = u_mmInit(0, r300screen->caps.zmask_ram * frag_pipes);
+    if (!r300->zmask_mm)
+      return FALSE;
+
+    if (r300screen->caps.hiz_ram) {
       r300->hiz_mm = u_mmInit(0, r300screen->caps.hiz_ram * frag_pipes);
+      if (!r300->hiz_mm) {
+        u_mmDestroy(r300->zmask_mm);
+        r300->zmask_mm = NULL;
+        return FALSE;
+      }
+    }
 
-    r300->zmask_mm = u_mmInit(0, r300screen->caps.zmask_ram * frag_pipes);
+    return TRUE;
 }
 
 void r300_hyperz_destroy_mm(struct r300_context *r300)
 {
     struct r300_screen* r300screen = r300->screen;
 
-    if (r300screen->caps.hiz_ram)
+    if (r300screen->caps.hiz_ram) {
       u_mmDestroy(r300->hiz_mm);
+      r300->hiz_mm = NULL;
+    }
 
     u_mmDestroy(r300->zmask_mm);
+    r300->zmask_mm = NULL;
 }
diff --git a/src/gallium/drivers/r300/r300_hyperz.h b/src/gallium/drivers/r300/r300_hyperz.h
index 09e1ff6625..30a23ec649 100644
--- a/src/gallium/drivers/r300/r300_hyperz.h
+++ b/src/gallium/drivers/r300/r300_hyperz.h
@@ -30,6 +30,6 @@ void r300_update_hyperz_state(struct r300_context* r300);
 void r300_hiz_alloc_block(struct r300_context *r300, struct r300_surface *surf);
 void r300_zmask_alloc_block(struct r300_context *r300, struct r300_surface *surf, int compress);
 
-void r300_hyperz_init_mm(struct r300_context *r300);
+boolean r300_hyperz_init_mm(struct r300_context *r300);
 void r300_hyperz_destroy_mm(struct r300_context *r300);
 #endif
diff --git a/src/gallium/drivers/r300/r300_reg.h b/src/gallium/drivers/r300/r300_reg.h
index 99a9d65055..60d3b600cb 100644
--- a/src/gallium/drivers/r300/r300_reg.h
+++ b/src/gallium/drivers/r300/r300_reg.h
@@ -496,6 +496,12 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define R300_VAP_GB_HORZ_CLIP_ADJ                   0x2228
 #define R300_VAP_GB_HORZ_DISC_ADJ                   0x222c
 
+#define R300_VAP_PVS_FLOW_CNTL_ADDRS_0      0x2230
+#define R300_PVS_FC_ACT_ADRS(x)             ((x) << 0)
+#define R300_PVS_FC_LOOP_CNT_JMP_INST(x)    ((x) << 8)
+#define R300_PVS_FC_LAST_INST(x)            ((x) << 16)
+#define R300_PVS_FC_RTN_INST(x)             ((x) << 24)
+
 /* gap */
 
 /* Sometimes, END_OF_PKT and 0x2284=0 are the only commands sent between
@@ -514,6 +520,10 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #       define R300_2288_R300                    0x00750000 /* -- nh */
 #       define R300_2288_RV350                   0x0000FFFF /* -- Vladimir */
 
+#define R300_VAP_PVS_FLOW_CNTL_LOOP_INDEX_0 0x2290
+#define R300_PVS_FC_LOOP_INIT_VAL(x)        ((x) << 0)
+#define R300_PVS_FC_LOOP_STEP_VAL(x)        ((x) << 8)
+
 /* gap */
 
 /* Addresses are relative to the vertex program instruction area of the
@@ -548,6 +558,9 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #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
+#define R300_VAP_PVS_FC_OPC_JUMP(x)         (1 << (2 * (x)))
+#define R300_VAP_PVS_FC_OPC_LOOP(x)         (2 << (2 * (x)))
+#define R300_VAP_PVS_FC_OPC_JSR(x)          (3 << (2 * (x)))
 
 /* The entire range from 0x2300 to 0x2AC inclusive seems to be used for
  * immediate vertices
@@ -564,6 +577,14 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 /* write 0 to indicate end of packet? */
 #define R300_VAP_VTX_END_OF_PKT             0x24AC
 
+#define R500_VAP_PVS_FLOW_CNTL_ADDRS_LW_0   0x2500
+#define R500_PVS_FC_ACT_ADRS(x)             ((x) << 0)
+#define R500_PVS_FC_LOOP_CNT_JMP_INST(x)    ((x) << 16)
+
+#define R500_VAP_PVS_FLOW_CNTL_ADDRS_UW_0   0x2504
+#define R500_PVS_FC_LAST_INST(x)            ((x) << 0)
+#define R500_PVS_FC_RTN_INST(x)             ((x) << 16)
+
 /* gap */
 
 /* These are values from r300_reg/r300_reg.h - they are known to be correct
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index 910f5f7113..86b11ca045 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -186,20 +186,14 @@ enum r300_prepare_flags {
  * \param cs_dwords     The number of dwords to reserve in CS.
  * \param aos_offset    The offset passed to emit_aos.
  * \param index_bias    The index bias to emit.
- * \param end_cs_dwords The number of free dwords which must be available
- *                      at the end of CS after drawing in case the CS space
- *                      management is performed by a draw_* function manually.
- *                      The parameter may be NULL.
  */
 static void r300_prepare_for_rendering(struct r300_context *r300,
                                        enum r300_prepare_flags flags,
                                        struct pipe_resource *index_buffer,
                                        unsigned cs_dwords,
                                        int aos_offset,
-                                       int index_bias,
-                                       unsigned *end_cs_dwords)
+                                       int index_bias)
 {
-    unsigned end_dwords    = 0;
     boolean flushed        = FALSE;
     boolean first_draw     = flags & PREP_FIRST_DRAW;
     boolean emit_aos       = flags & PREP_EMIT_AOS;
@@ -221,12 +215,7 @@ static void r300_prepare_for_rendering(struct r300_context *r300,
             cs_dwords += 7; /* emit_aos_swtcl */
     }
 
-    /* Emitted in flush. */
-    end_dwords += 26; /* emit_query_end */
-    if (r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
-        end_dwords += r300->hyperz_state.size + 2; /* emit_hyperz_end + zcache flush */
-
-    cs_dwords += end_dwords;
+    cs_dwords += r300_get_num_cs_end_dwords(r300);
 
     /* Reserve requested CS space. */
     if (cs_dwords > (r300->cs->ndw - r300->cs->cdw)) {
@@ -251,9 +240,6 @@ static void r300_prepare_for_rendering(struct r300_context *r300,
         if (emit_aos_swtcl)
             r300_emit_aos_swtcl(r300, indexed);
     }
-
-    if (end_cs_dwords)
-        *end_cs_dwords = end_dwords;
 }
 
 static boolean immd_is_good_idea(struct r300_context *r300,
@@ -354,7 +340,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 
     dwords = 9 + count * vertex_size;
 
-    r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0, NULL);
+    r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0);
 
     BEGIN_CS(dwords);
     OUT_CS_REG(R300_GA_COLOR_CONTROL,
@@ -534,7 +520,7 @@ static void r300_draw_range_elements(struct pipe_context* pipe,
     /* 15 dwords for emit_draw_elements */
     r300_prepare_for_rendering(r300,
         PREP_FIRST_DRAW | PREP_VALIDATE_VBOS | PREP_EMIT_AOS | PREP_INDEXED,
-        indexBuffer, 15, buffer_offset, indexBias, NULL);
+        indexBuffer, 15, buffer_offset, indexBias);
 
     if (alt_num_verts || count <= 65535) {
         r300_emit_draw_elements(r300, indexBuffer, indexSize,
@@ -553,7 +539,7 @@ static void r300_draw_range_elements(struct pipe_context* pipe,
             if (count) {
                 r300_prepare_for_rendering(r300,
                     PREP_VALIDATE_VBOS | PREP_EMIT_AOS | PREP_INDEXED,
-                    indexBuffer, 15, buffer_offset, indexBias, NULL);
+                    indexBuffer, 15, buffer_offset, indexBias);
             }
         } while (count);
     }
@@ -598,7 +584,7 @@ static void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
     } else {
         /* 9 spare dwords for emit_draw_arrays. */
         r300_prepare_for_rendering(r300, PREP_FIRST_DRAW | PREP_VALIDATE_VBOS | PREP_EMIT_AOS,
-                               NULL, 9, start, 0, NULL);
+                               NULL, 9, start, 0);
 
         if (alt_num_verts || count <= 65535) {
             r300_emit_draw_arrays(r300, mode, count);
@@ -614,7 +600,7 @@ static void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
                 if (count) {
                     r300_prepare_for_rendering(r300,
                         PREP_VALIDATE_VBOS | PREP_EMIT_AOS, NULL, 9,
-                        start, 0, NULL);
+                        start, 0);
                 }
             } while (count);
         }
@@ -855,7 +841,7 @@ static void r300_render_draw_arrays(struct vbuf_render* render,
     (void) i; (void) ptr;
 
     r300_prepare_for_rendering(r300, PREP_FIRST_DRAW | PREP_EMIT_AOS_SWTCL,
-                               NULL, dwords, 0, 0, NULL);
+                               NULL, dwords, 0, 0);
 
     DBG(r300, DBG_DRAW, "r300: render_draw_arrays (count: %d)\n", count);
 
@@ -908,7 +894,8 @@ static void r300_render_draw_elements(struct vbuf_render* render,
      * indices than it can fit in CS. */
     r300_prepare_for_rendering(r300,
         PREP_FIRST_DRAW | PREP_EMIT_AOS_SWTCL | PREP_INDEXED,
-        NULL, 256, 0, 0, &end_cs_dwords);
+        NULL, 256, 0, 0);
+    end_cs_dwords = r300_get_num_cs_end_dwords(r300);
 
     while (count) {
         free_dwords = r300->cs->ndw - r300->cs->cdw;
@@ -938,7 +925,8 @@ static void r300_render_draw_elements(struct vbuf_render* render,
         if (count) {
             r300_prepare_for_rendering(r300,
                 PREP_EMIT_AOS_SWTCL | PREP_INDEXED,
-                NULL, 256, 0, 0, &end_cs_dwords);
+                NULL, 256, 0, 0);
+            end_cs_dwords = r300_get_num_cs_end_dwords(r300);
         }
     }
 }
@@ -1032,7 +1020,7 @@ static void r300_blitter_draw_rectangle(struct blitter_context *blitter,
     r300->clip_state.dirty = FALSE;
     r300->viewport_state.dirty = FALSE;
 
-    r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0, NULL);
+    r300_prepare_for_rendering(r300, PREP_FIRST_DRAW, NULL, dwords, 0, 0);
 
     DBG(r300, DBG_DRAW, "r300: draw_rectangle\n");
 
diff --git a/src/gallium/drivers/r300/r300_screen.c b/src/gallium/drivers/r300/r300_screen.c
index 6268001054..1e4edcdbc3 100644
--- a/src/gallium/drivers/r300/r300_screen.c
+++ b/src/gallium/drivers/r300/r300_screen.c
@@ -115,7 +115,6 @@ static int r300_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
         case PIPE_CAP_TEXTURE_MIRROR_REPEAT:
         case PIPE_CAP_BLEND_EQUATION_SEPARATE:
         case PIPE_CAP_TEXTURE_SWIZZLE:
-        case PIPE_CAP_DEPTH_CLAMP:
             return 1;
 
         /* Unsupported features (boolean caps). */
@@ -124,6 +123,8 @@ static int r300_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
         case PIPE_CAP_TGSI_CONT_SUPPORTED:
         case PIPE_CAP_INDEP_BLEND_ENABLE:
         case PIPE_CAP_INDEP_BLEND_FUNC:
+        case PIPE_CAP_DEPTH_CLAMP: /* XXX implemented, but breaks Regnum Online */
+        case PIPE_CAP_DEPTHSTENCIL_CLEAR_SEPARATE:
             return 0;
 
         /* Texturing. */
@@ -150,9 +151,6 @@ static int r300_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
         case PIPE_CAP_MAX_CONST_BUFFER_SIZE:
             return 256;
 
-        case PIPE_CAP_DEPTHSTENCIL_CLEAR_SEPARATE:
-            return 1;
-
         /* Fragment coordinate conventions. */
         case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
         case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
diff --git a/src/gallium/drivers/r300/r300_shader_semantics.h b/src/gallium/drivers/r300/r300_shader_semantics.h
index cb7a37033f..4be23e64ce 100644
--- a/src/gallium/drivers/r300/r300_shader_semantics.h
+++ b/src/gallium/drivers/r300/r300_shader_semantics.h
@@ -38,6 +38,7 @@ struct r300_shader_semantics {
     int psize;
     int color[ATTR_COLOR_COUNT];
     int bcolor[ATTR_COLOR_COUNT];
+    int face;
     int generic[ATTR_GENERIC_COUNT];
     int fog;
     int wpos;
@@ -50,6 +51,7 @@ static INLINE void r300_shader_semantics_reset(
 
     info->pos = ATTR_UNUSED;
     info->psize = ATTR_UNUSED;
+    info->face = ATTR_UNUSED;
     info->fog = ATTR_UNUSED;
     info->wpos = ATTR_UNUSED;
 
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index 1e6b81d798..239edd98e3 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -689,8 +689,7 @@ void r300_mark_fb_state_dirty(struct r300_context *r300,
     /* What is marked as dirty depends on the enum r300_fb_state_change. */
     r300->gpu_flush.dirty = TRUE;
     r300->fb_state.dirty = TRUE;
-    if (r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
-        r300->hyperz_state.dirty = TRUE;
+    r300->hyperz_state.dirty = TRUE;
 
     if (change == R300_CHANGED_FB_STATE) {
         r300->aa_state.dirty = TRUE;
@@ -753,8 +752,6 @@ static void
 
     r300_mark_fb_state_dirty(r300, R300_CHANGED_FB_STATE);
 
-    r300->hiz_enable = false;
-    r300->z_fastfill = false;
     r300->z_compression = false;
     
     if (state->zsbuf) {
@@ -781,23 +778,18 @@ static void
             /* work out whether we can support zmask features on this buffer */
             r300_zmask_alloc_block(r300, zs_surf, compress);
 
-            if (tex->hiz_mem[level]) {
-                r300->hiz_enable = 1;
-            }
-
             if (tex->zmask_mem[level]) {
-                r300->z_fastfill = 1;
                 /* compression causes hangs on 16-bit */
                 if (zbuffer_bpp == 24)
                     r300->z_compression = compress;
             }
             DBG(r300, DBG_HYPERZ,
-                "hyper-z features: hiz: %d @ %08x z-compression: %d z-fastfill: %d @ %08x\n", r300->hiz_enable,
+                "hyper-z features: hiz: %d @ %08x z-compression: %d z-fastfill: %d @ %08x\n", tex->hiz_mem[level] ? 1 : 0,
                 tex->hiz_mem[level] ? tex->hiz_mem[level]->ofs : 0xdeadbeef,
-                r300->z_compression, r300->z_fastfill,
+                r300->z_compression, tex->zmask_mem[level] ? 1 : 0,
                 tex->zmask_mem[level] ? tex->zmask_mem[level]->ofs : 0xdeadbeef);
         }
-            
+
         /* Polygon offset depends on the zbuffer bit depth. */
         if (r300->zbuffer_bpp != zbuffer_bpp) {
             r300->zbuffer_bpp = zbuffer_bpp;
@@ -1759,10 +1751,12 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
     r300->rs_block_state.dirty = TRUE; /* Will be updated before the emission. */
 
     if (r300->screen->caps.has_tcl) {
+        unsigned fc_op_dwords = r300->screen->caps.is_r500 ? 3 : 2;
         r300->vs_state.dirty = TRUE;
         r300->vs_state.size =
                 vs->code.length + 9 +
-                (vs->immediates_count ? vs->immediates_count * 4 + 3 : 0);
+                (vs->immediates_count ? vs->immediates_count * 4 + 3 : 0) +
+        (vs->code.num_fc_ops ? vs->code.num_fc_ops * fc_op_dwords + 4 : 0);
 
         if (vs->externals_count) {
             r300->vs_constants.dirty = TRUE;
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index f3dad4c292..c8de3e1c52 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -35,7 +35,6 @@
 #include "r300_state_inlines.h"
 #include "r300_texture.h"
 #include "r300_vs.h"
-#include "r300_winsys.h"
 
 /* r300_state_derived: Various bits of state which are dependent upon
  * currently bound CSO data. */
@@ -47,6 +46,11 @@ enum r300_rs_swizzle {
     SWIZ_0001,
 };
 
+enum r300_rs_col_write_type {
+    WRITE_COLOR = 0,
+    WRITE_FACE
+};
+
 static void r300_draw_emit_attrib(struct r300_context* r300,
                                   enum attrib_emit emit,
                                   enum interp_mode interp,
@@ -204,8 +208,10 @@ static void r300_rs_col(struct r300_rs_block* rs, int id, int ptr,
     rs->inst[id] |= R300_RS_INST_COL_ID(id);
 }
 
-static void r300_rs_col_write(struct r300_rs_block* rs, int id, int fp_offset)
+static void r300_rs_col_write(struct r300_rs_block* rs, int id, int fp_offset,
+                              enum r300_rs_col_write_type type)
 {
+    assert(type == WRITE_COLOR);
     rs->inst[id] |= R300_RS_INST_COL_CN_WRITE |
                     R300_RS_INST_COL_ADDR(fp_offset);
 }
@@ -253,10 +259,16 @@ static void r500_rs_col(struct r300_rs_block* rs, int id, int ptr,
     rs->inst[id] |= R500_RS_INST_COL_ID(id);
 }
 
-static void r500_rs_col_write(struct r300_rs_block* rs, int id, int fp_offset)
+static void r500_rs_col_write(struct r300_rs_block* rs, int id, int fp_offset,
+                              enum r300_rs_col_write_type type)
 {
-    rs->inst[id] |= R500_RS_INST_COL_CN_WRITE |
-                    R500_RS_INST_COL_ADDR(fp_offset);
+    if (type == WRITE_FACE)
+        rs->inst[id] |= R500_RS_INST_COL_CN_WRITE_BACKFACE |
+                        R500_RS_INST_COL_ADDR(fp_offset);
+    else
+        rs->inst[id] |= R500_RS_INST_COL_CN_WRITE |
+                        R500_RS_INST_COL_ADDR(fp_offset);
+
 }
 
 static void r500_rs_tex(struct r300_rs_block* rs, int id, int ptr,
@@ -306,7 +318,7 @@ 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, loc = 0, tex_ptr = 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_col_write)(struct r300_rs_block*, int, int, enum r300_rs_col_write_type);
     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 ||
@@ -325,6 +337,11 @@ static void r300_update_rs_block(struct r300_context *r300)
         rX00_rs_tex_write = r300_rs_tex_write;
     }
 
+    /* 0x5555 copied from classic, which means:
+     * Select user color 0 for COLOR0 up to COLOR7.
+     * What the hell does that mean? */
+    rs.vap_vtx_state_cntl = 0x5555;
+
     /* 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;
@@ -351,7 +368,7 @@ static void r300_update_rs_block(struct r300_context *r300)
 
             /* 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);
+                rX00_rs_col_write(&rs, col_count, fp_offset, WRITE_COLOR);
                 fp_offset++;
 
                 DBG(r300, DBG_RS,
@@ -399,6 +416,24 @@ static void r300_update_rs_block(struct r300_context *r300)
         }
     }
 
+    /* gl_FrontFacing.
+     * Note that we can use either the two-sided color selection based on
+     * the front and back vertex shader colors, or gl_FrontFacing,
+     * but not both! It locks up otherwise.
+     *
+     * In Direct3D 9, the two-sided color selection can be used
+     * with shaders 2.0 only, while gl_FrontFacing can be used
+     * with shaders 3.0 only. The hardware apparently hasn't been designed
+     * to support both at the same time. */
+    if (r300->screen->caps.is_r500 && fs_inputs->face != ATTR_UNUSED &&
+        !(any_bcolor_used && r300->two_sided_color)) {
+        rX00_rs_col(&rs, col_count, col_count, SWIZ_XYZW);
+        rX00_rs_col_write(&rs, col_count, fp_offset, WRITE_FACE);
+        fp_offset++;
+        col_count++;
+        DBG(r300, DBG_RS, "r300: Rasterized FACE written to FS.\n");
+    }
+
     /* Rasterize texture coordinates. */
     for (i = 0; i < ATTR_GENERIC_COUNT && tex_count < 8; i++) {
 	bool sprite_coord = !!(r300->sprite_coord_enable & (1 << i));
@@ -677,8 +712,44 @@ static void r300_merge_textures_and_samplers(struct r300_context* r300)
     }
 }
 
+/* We can't use compressed zbuffers as samplers. */
+static void r300_flush_depth_textures(struct r300_context *r300)
+{
+    struct r300_textures_state *state =
+        (struct r300_textures_state*)r300->textures_state.state;
+    unsigned i, level;
+    unsigned count = MIN2(state->sampler_view_count,
+                          state->sampler_state_count);
+
+    if (r300->z_decomp_rd)
+        return;
+
+    for (i = 0; i < count; i++)
+        if (state->sampler_views[i] && state->sampler_states[i]) {
+            struct pipe_resource *tex = state->sampler_views[i]->base.texture;
+
+            if (tex->target == PIPE_TEXTURE_3D ||
+                tex->target == PIPE_TEXTURE_CUBE)
+                continue;
+
+            /* Ignore non-depth textures.
+             * Also ignore reinterpreted depth textures, e.g. resource_copy. */
+            if (!util_format_is_depth_or_stencil(tex->format))
+                continue;
+
+            for (level = 0; level <= tex->last_level; level++)
+                if (r300_texture(tex)->zmask_in_use[level]) {
+                    /* We don't handle 3D textures and cubemaps yet. */
+                    r300_flush_depth_stencil(&r300->context, tex,
+                                             u_subresource(0, level), 0);
+                }
+        }
+}
+
 void r300_update_derived_state(struct r300_context* r300)
 {
+    r300_flush_depth_textures(r300);
+
     if (r300->textures_state.dirty) {
         r300_merge_textures_and_samplers(r300);
     }
@@ -694,6 +765,5 @@ void r300_update_derived_state(struct r300_context* r300)
         }
     }
 
-    if (r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
-        r300_update_hyperz_state(r300);
+    r300_update_hyperz_state(r300);
 }
diff --git a/src/gallium/drivers/r300/r300_tgsi_to_rc.c b/src/gallium/drivers/r300/r300_tgsi_to_rc.c
index 51b2c55550..a4911b9a2a 100644
--- a/src/gallium/drivers/r300/r300_tgsi_to_rc.c
+++ b/src/gallium/drivers/r300/r300_tgsi_to_rc.c
@@ -97,13 +97,13 @@ static unsigned translate_opcode(unsigned opcode)
      /* case TGSI_OPCODE_BRA: return RC_OPCODE_BRA; */
      /* case TGSI_OPCODE_CAL: return RC_OPCODE_CAL; */
      /* case TGSI_OPCODE_RET: return RC_OPCODE_RET; */
-     /* case TGSI_OPCODE_SSG: return RC_OPCODE_SSG; */
+        case TGSI_OPCODE_SSG: return RC_OPCODE_SSG;
         case TGSI_OPCODE_CMP: return RC_OPCODE_CMP;
         case TGSI_OPCODE_SCS: return RC_OPCODE_SCS;
         case TGSI_OPCODE_TXB: return RC_OPCODE_TXB;
      /* case TGSI_OPCODE_NRM: return RC_OPCODE_NRM; */
      /* case TGSI_OPCODE_DIV: return RC_OPCODE_DIV; */
-     /* case TGSI_OPCODE_DP2: return RC_OPCODE_DP2; */
+        case TGSI_OPCODE_DP2: return RC_OPCODE_DP2;
         case TGSI_OPCODE_TXL: return RC_OPCODE_TXL;
         case TGSI_OPCODE_BRK: return RC_OPCODE_BRK;
         case TGSI_OPCODE_IF: return RC_OPCODE_IF;
@@ -126,7 +126,7 @@ static unsigned translate_opcode(unsigned opcode)
      /* case TGSI_OPCODE_SAD: return RC_OPCODE_SAD; */
      /* case TGSI_OPCODE_TXF: return RC_OPCODE_TXF; */
      /* case TGSI_OPCODE_TXQ: return RC_OPCODE_TXQ; */
-     /* case TGSI_OPCODE_CONT: return RC_OPCODE_CONT; */
+        case TGSI_OPCODE_CONT: return RC_OPCODE_CONT;
      /* case TGSI_OPCODE_EMIT: return RC_OPCODE_EMIT; */
      /* case TGSI_OPCODE_ENDPRIM: return RC_OPCODE_ENDPRIM; */
      /* case TGSI_OPCODE_BGNLOOP2: return RC_OPCODE_BGNLOOP2; */
diff --git a/src/gallium/drivers/r300/r300_vs.c b/src/gallium/drivers/r300/r300_vs.c
index b25c786d6b..54c8de1241 100644
--- a/src/gallium/drivers/r300/r300_vs.c
+++ b/src/gallium/drivers/r300/r300_vs.c
@@ -207,7 +207,7 @@ void r300_translate_vertex_shader(struct r300_context *r300,
     compiler.Base.max_temp_regs = 32;
 
     if (compiler.Base.Debug) {
-        debug_printf("r300: Initial vertex program\n");
+        DBG(r300, DBG_VP, "r300: Initial vertex program\n");
         tgsi_dump(vs->state.tokens, 0);
     }
 
@@ -227,8 +227,7 @@ void r300_translate_vertex_shader(struct r300_context *r300,
     /* Invoke the compiler */
     r3xx_compile_vertex_program(&compiler);
     if (compiler.Base.Error) {
-        /* XXX We should fallback using Draw. */
-        fprintf(stderr, "r300 VP: Compiler error:\n%sUsing a dummy shader"
+        DBG(r300, DBG_VP, "r300 VP: Compiler error:\n%sUsing a dummy shader"
                 " instead.\nIf there's an 'unknown opcode' message, please"
                 " file a bug report and attach this log.\n", compiler.Base.ErrorMsg);