1 files changed, 40 insertions, 21 deletions

diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index 29bc701a86..192846411b 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -71,9 +71,9 @@ static void r300_draw_emit_attrib(struct r300_context* r300,
     struct tgsi_shader_info* info = &r300->vs->info;
     int output;
 
-    output = draw_find_vs_output(r300->draw,
-                                 info->output_semantic_name[index],
-                                 info->output_semantic_index[index]);
+    output = draw_find_shader_output(r300->draw,
+                                     info->output_semantic_name[index],
+                                     info->output_semantic_index[index]);
     draw_emit_vertex_attr(&r300->vertex_info->vinfo, emit, interp, output);
 }
 
@@ -139,10 +139,10 @@ static void r300_vertex_psc(struct r300_context* r300)
 
     /* If TCL is bypassed, map vertex streams to equivalent VS output
      * locations. */
-    if (r300->rs_state->enable_vte) {
-        stream_tab = identity;
-    } else {
+    if (r300->tcl_bypass) {
         stream_tab = r300->vs->stream_loc_notcl;
+    } else {
+        stream_tab = identity;
     }
 
     /* Vertex shaders have no semantics on their inputs,
@@ -333,6 +333,8 @@ static void r300_update_rs_block(struct r300_context* r300,
     void (*rX00_rs_col_write)(struct r300_rs_block*, int, int);
     void (*rX00_rs_tex)(struct r300_rs_block*, int, int, boolean);
     void (*rX00_rs_tex_write)(struct r300_rs_block*, int, int);
+    boolean any_bcolor_used = vs_outputs->bcolor[0] != ATTR_UNUSED ||
+                              vs_outputs->bcolor[1] != ATTR_UNUSED;
 
     if (r300_screen(r300->context.screen)->caps->is_r500) {
         rX00_rs_col       = r500_rs_col;
@@ -348,7 +350,7 @@ static void r300_update_rs_block(struct r300_context* r300,
 
     /* Rasterize colors. */
     for (i = 0; i < ATTR_COLOR_COUNT; i++) {
-        if (vs_outputs->color[i] != ATTR_UNUSED) {
+        if (vs_outputs->color[i] != ATTR_UNUSED || any_bcolor_used) {
             /* Always rasterize if it's written by the VS,
              * otherwise it locks up. */
             rX00_rs_col(rs, col_count, i, FALSE);
@@ -410,6 +412,16 @@ static void r300_update_rs_block(struct r300_context* r300,
         }
     }
 
+    /* Rasterize WPOS. */
+    /* If the FS doesn't need it, it's not written by the VS. */
+    if (fs_inputs->wpos != ATTR_UNUSED) {
+        rX00_rs_tex(rs, tex_count, tex_count, FALSE);
+        rX00_rs_tex_write(rs, tex_count, fp_offset);
+
+        fp_offset++;
+        tex_count++;
+    }
+
     /* Rasterize at least one color, or bad things happen. */
     if (col_count == 0 && tex_count == 0) {
         rX00_rs_col(rs, 0, 0, TRUE);
@@ -496,7 +508,8 @@ static boolean r300_dsa_alpha_test_enabled(struct r300_dsa_state* dsa)
 
 static void r300_update_ztop(struct r300_context* r300)
 {
-    r300->ztop_state.z_buffer_top = R300_ZTOP_ENABLE;
+    struct r300_ztop_state* ztop_state =
+        (struct r300_ztop_state*)r300->ztop_state.state;
 
     /* This is important enough that I felt it warranted a comment.
      *
@@ -518,31 +531,37 @@ static void r300_update_ztop(struct r300_context* r300)
      * 5) Depth writes in fragment shader
      * 6) Outstanding occlusion queries
      *
+     * This register causes stalls all the way from SC to CB when changed,
+     * but it is buffered on-chip so it does not hurt to write it if it has
+     * not changed.
+     *
      * ~C.
      */
 
     /* ZS writes */
-    if (r300_dsa_writes_depth_stencil(r300->dsa_state) &&
-           (r300_dsa_alpha_test_enabled(r300->dsa_state) ||   /* (1) */
-            r300->fs->info.uses_kill)) {                      /* (2) */
-        r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
-    } else if (r300_fragment_shader_writes_depth(r300->fs)) { /* (5) */
-        r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
-    } else if (r300->query_current) {                         /* (6) */
-        r300->ztop_state.z_buffer_top = R300_ZTOP_DISABLE;
+    if (r300_dsa_writes_depth_stencil(r300->dsa_state.state) &&
+           (r300_dsa_alpha_test_enabled(r300->dsa_state.state) ||/* (1) */
+            r300->fs->info.uses_kill)) {                         /* (2) */
+        ztop_state->z_buffer_top = R300_ZTOP_DISABLE;
+    } else if (r300_fragment_shader_writes_depth(r300->fs)) {    /* (5) */
+        ztop_state->z_buffer_top = R300_ZTOP_DISABLE;
+    } else if (r300->query_current) {                            /* (6) */
+        ztop_state->z_buffer_top = R300_ZTOP_DISABLE;
+    } else {
+        ztop_state->z_buffer_top = R300_ZTOP_ENABLE;
     }
+
+    r300->ztop_state.dirty = TRUE;
 }
 
 void r300_update_derived_state(struct r300_context* r300)
 {
+    /* XXX */
     if (r300->dirty_state &
         (R300_NEW_FRAGMENT_SHADER | R300_NEW_VERTEX_SHADER |
-         R300_NEW_VERTEX_FORMAT)) {
+         R300_NEW_VERTEX_FORMAT) || r300->rs_state.dirty) {
         r300_update_derived_shader_state(r300);
     }
 
-    if (r300->dirty_state &
-            (R300_NEW_DSA | R300_NEW_FRAGMENT_SHADER | R300_NEW_QUERY)) {
-        r300_update_ztop(r300);
-    }
+    r300_update_ztop(r300);
 }