r300g: Begin separating HW TCL and SW TCL state and setup.

This patch removes draw_context entirely from the HW TCL path and cleans up
a few other things along the way. Hopefully, nothing got broken.

Thanks to Marek Olšák for testing, review, and pointing out my bugs. :3

6 files changed, 110 insertions, 174 deletions

diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
index ae23329b83f..26db536248e 100644
--- a/src/gallium/drivers/r300/r300_context.c
+++ b/src/gallium/drivers/r300/r300_context.c
@@ -123,15 +123,24 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     r300->context.clear = r300_clear;
 
-    if (r300screen->caps->has_tcl)
-    {
+    if (r300screen->caps->has_tcl) {
         r300->context.draw_arrays = r300_draw_arrays;
         r300->context.draw_elements = r300_draw_elements;
         r300->context.draw_range_elements = r300_draw_range_elements;
-    }
-    else
-    {
-        assert(0);
+    } else {
+        r300->context.draw_arrays = r300_swtcl_draw_arrays;
+        r300->context.draw_elements = r300_draw_elements;
+        r300->context.draw_range_elements = r300_swtcl_draw_range_elements;
+
+        /* Create a Draw. This is used for SW TCL. */
+        r300->draw = draw_create();
+        /* Enable our renderer. */
+        draw_set_rasterize_stage(r300->draw, r300_draw_stage(r300));
+        /* Enable Draw's clipping. */
+        draw_set_driver_clipping(r300->draw, FALSE);
+        /* Force Draw to never do viewport transform, since we can do
+         * transform in hardware, always. */
+        draw_set_viewport_state(r300->draw, &r300_viewport_identity);
     }
 
     r300->context.is_texture_referenced = r300_is_texture_referenced;
@@ -145,16 +154,6 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
     r300->scissor_state = CALLOC_STRUCT(r300_scissor_state);
     r300->viewport_state = CALLOC_STRUCT(r300_viewport_state);
 
-    /* Create a Draw. This is used for vert collation and SW TCL. */
-    r300->draw = draw_create();
-    /* Enable our renderer. */
-    draw_set_rasterize_stage(r300->draw, r300_draw_stage(r300));
-    /* Disable Draw's clipping if TCL is present. */
-    draw_set_driver_clipping(r300->draw, r300_screen(screen)->caps->has_tcl);
-    /* Force Draw to never do viewport transform, since (again) we can do
-     * transform in hardware, always. */
-    draw_set_viewport_state(r300->draw, &r300_viewport_identity);
-
     /* Open up the OQ BO. */
     r300->oqbo = screen->buffer_create(screen, 4096,
             PIPE_BUFFER_USAGE_VERTEX, 4096);
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index 62e1456ed36..4c5fb405c6a 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -183,8 +183,6 @@ boolean r300_draw_range_elements(struct pipe_context* pipe,
         return FALSE;
     }
 
-    setup_vertex_attributes(r300);
-
     setup_index_buffer(r300, indexBuffer, indexSize);
 
     r300_emit_dirty_state(r300);
@@ -226,8 +224,6 @@ boolean r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
         return FALSE;
     }
 
-    setup_vertex_attributes(r300);
-
     r300_emit_dirty_state(r300);
 
     r300_emit_aos(r300, start);
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index 00f10ffd73c..5422a2cc9c8 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -714,9 +714,6 @@ static void* r300_create_vs_state(struct pipe_context* pipe,
 
         tgsi_scan_shader(shader->tokens, &vs->info);
 
-        /* Appease Draw. */
-        vs->draw = draw_create_vertex_shader(r300->draw, shader);
-
         return (void*)vs;
     } else {
         return draw_create_vertex_shader(r300->draw, shader);
@@ -727,8 +724,6 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
 {
     struct r300_context* r300 = r300_context(pipe);
 
-    draw_flush(r300->draw);
-
     if (r300_screen(pipe->screen)->caps->has_tcl) {
         struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader;
 
@@ -739,10 +734,10 @@ static void r300_bind_vs_state(struct pipe_context* pipe, void* shader)
             r300_translate_vertex_shader(r300, vs);
         }
 
-        draw_bind_vertex_shader(r300->draw, vs->draw);
         r300->vs = vs;
         r300->dirty_state |= R300_NEW_VERTEX_SHADER | R300_NEW_VERTEX_SHADER_CONSTANTS;
     } else {
+        draw_flush(r300->draw);
         draw_bind_vertex_shader(r300->draw,
                 (struct draw_vertex_shader*)shader);
     }
@@ -756,7 +751,6 @@ static void r300_delete_vs_state(struct pipe_context* pipe, void* shader)
         struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader;
 
         rc_constants_destroy(&vs->code.constants);
-        draw_delete_vertex_shader(r300->draw, vs->draw);
         FREE((void*)vs->state.tokens);
         FREE(shader);
     } else {
diff --git a/src/gallium/drivers/r300/r300_state_derived.c b/src/gallium/drivers/r300/r300_state_derived.c
index b4d0eeaf8c5..5aa4166d939 100644
--- a/src/gallium/drivers/r300/r300_state_derived.c
+++ b/src/gallium/drivers/r300/r300_state_derived.c
@@ -65,84 +65,43 @@ int r300_shader_key_compare(void* key1, void* key2) {
 static void r300_vs_tab_routes(struct r300_context* r300,
                                struct r300_vertex_info* vformat)
 {
-    struct r300_screen* r300screen = r300_screen(r300->context.screen);
     struct vertex_info* vinfo = &vformat->vinfo;
     int* tab = vformat->vs_tab;
     boolean pos = FALSE, psize = FALSE, fog = FALSE;
     int i, texs = 0, cols = 0;
-    struct tgsi_shader_info* info;
-
-    if (r300screen->caps->has_tcl) {
-        /* Use vertex shader to determine required routes. */
-        info = &r300->vs->info;
-    } else {
-        /* Use fragment shader to determine required routes. */
-        info = &r300->fs->info;
-    }
+    struct tgsi_shader_info* info = &r300->fs->info;
 
-    assert(info->num_inputs <= 16);
+    /* XXX One day we should figure out how to handle a different number of
+     * VS outputs and FS inputs, as well as a different number of vertex streams
+     * and VS inputs. It's definitely one of the sources of hardlocks. */
 
-    if (!r300screen->caps->has_tcl || !r300->rs_state->enable_vte)
-    {
-        for (i = 0; i < info->num_inputs; i++) {
-            switch (r300->vs->code.inputs[i]) {
-                case TGSI_SEMANTIC_POSITION:
-                    pos = TRUE;
-                    tab[i] = 0;
-                    break;
-                case TGSI_SEMANTIC_COLOR:
-                    tab[i] = 2 + cols;
-                    cols++;
-                    break;
-                case TGSI_SEMANTIC_PSIZE:
-                    assert(psize == FALSE);
-                    psize = TRUE;
-                    tab[i] = 15;
-                    break;
-                case TGSI_SEMANTIC_FOG:
-                    assert(fog == FALSE);
-                    fog = TRUE;
-                    /* Fall through */
-                case TGSI_SEMANTIC_GENERIC:
-                    tab[i] = 6 + texs;
-                    texs++;
-                    break;
-                default:
-                    debug_printf("r300: Unknown vertex input %d\n",
-                        info->input_semantic_name[i]);
-                    break;
-            }
-        }
-    }
-    else
-    {
-        /* Just copy vert attribs over as-is. */
-        for (i = 0; i < info->num_inputs; i++) {
-            tab[i] = i;
-        }
-
-        for (i = 0; i < info->num_outputs; i++) {
-            switch (info->output_semantic_name[i]) {
-                case TGSI_SEMANTIC_POSITION:
-                    pos = TRUE;
-                    break;
-                case TGSI_SEMANTIC_COLOR:
-                    cols++;
-                    break;
-                case TGSI_SEMANTIC_PSIZE:
-                    psize = TRUE;
-                    break;
-                case TGSI_SEMANTIC_FOG:
-                    fog = TRUE;
-                    /* Fall through */
-                case TGSI_SEMANTIC_GENERIC:
-                    texs++;
-                    break;
-                default:
-                    debug_printf("r300: Unknown vertex output %d\n",
-                        info->output_semantic_name[i]);
-                    break;
-            }
+    for (i = 0; i < info->num_inputs; i++) {
+        switch (info->input_semantic_name[i]) {
+            case TGSI_SEMANTIC_POSITION:
+                pos = TRUE;
+                tab[i] = 0;
+                break;
+            case TGSI_SEMANTIC_COLOR:
+                tab[i] = 2 + cols;
+                cols++;
+                break;
+            case TGSI_SEMANTIC_PSIZE:
+                assert(psize == FALSE);
+                psize = TRUE;
+                tab[i] = 15;
+                break;
+            case TGSI_SEMANTIC_FOG:
+                assert(fog == FALSE);
+                fog = TRUE;
+                /* Fall through */
+            case TGSI_SEMANTIC_GENERIC:
+                tab[i] = 6 + texs;
+                texs++;
+                break;
+            default:
+                debug_printf("r300: Unknown vertex input %d\n",
+                    info->input_semantic_name[i]);
+                break;
         }
     }
 
@@ -161,8 +120,7 @@ static void r300_vs_tab_routes(struct r300_context* r300,
 
     /* We need to add vertex position attribute only for SW TCL case,
      * for HW TCL case it could be generated by vertex shader */
-    if (!pos && !r300screen->caps->has_tcl) {
-        debug_printf("r300: Forcing vertex position attribute emit...\n");
+    if (!pos) {
         /* Make room for the position attribute
          * at the beginning of the tab. */
         for (i = 15; i > 0; i--) {
@@ -230,31 +188,66 @@ static void r300_vs_tab_routes(struct r300_context* r300,
 static void r300_vertex_psc(struct r300_context* r300,
                             struct r300_vertex_info* vformat)
 {
-    struct r300_screen* r300screen = r300_screen(r300->context.screen);
-    struct vertex_info* vinfo = &vformat->vinfo;
-    int* tab = vformat->vs_tab;
     uint16_t type, swizzle;
     enum pipe_format format;
-    unsigned i, attrib_count;
+    unsigned i;
 
     /* Vertex shaders have no semantics on their inputs,
-     * so PSC should just route stuff based on their info,
+     * so PSC should just route stuff based on the vertex elements,
      * and not on attrib information. */
-    if (r300screen->caps->has_tcl) {
-        attrib_count = r300->vs->info.num_inputs;
-        DBG(r300, DBG_DRAW, "r300: routing %d attribs in psc for vs\n",
-                attrib_count);
-    } else {
-        attrib_count = vinfo->num_attribs;
-        DBG(r300, DBG_DRAW, "r300: attrib count: %d\n", attrib_count);
-        for (i = 0; i < attrib_count; i++) {
-            DBG(r300, DBG_DRAW, "r300: attrib: offset %d, interp %d, size %d,"
-                   " tab %d\n", vinfo->attrib[i].src_index,
-                   vinfo->attrib[i].interp_mode, vinfo->attrib[i].emit,
-                   tab[i]);
+    DBG(r300, DBG_DRAW, "r300: vs expects %d attribs, routing %d elements"
+            " in psc\n",
+            r300->vs->info.num_inputs,
+            r300->vertex_element_count);
+
+    for (i = 0; i < r300->vertex_element_count; i++) {
+        format = r300->vertex_element[i].src_format;
+
+        type = r300_translate_vertex_data_type(format) |
+            (i << R300_DST_VEC_LOC_SHIFT);
+        swizzle = r300_translate_vertex_data_swizzle(format);
+
+        if (i % 2) {
+            vformat->vap_prog_stream_cntl[i >> 1] |= type << 16;
+            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
+        } else {
+            vformat->vap_prog_stream_cntl[i >> 1] |= type;
+            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
         }
     }
 
+
+    assert(i <= 15);
+
+    /* Set the last vector in the PSC. */
+    if (i) {
+        i -= 1;
+    }
+    vformat->vap_prog_stream_cntl[i >> 1] |=
+        (R300_LAST_VEC << (i & 1 ? 16 : 0));
+}
+
+/* Update the PSC tables for SW TCL, using Draw. */
+static void r300_swtcl_vertex_psc(struct r300_context* r300,
+                                  struct r300_vertex_info* vformat)
+{
+    struct vertex_info* vinfo = &vformat->vinfo;
+    int* tab = vformat->vs_tab;
+    uint16_t type, swizzle;
+    enum pipe_format format;
+    unsigned i, attrib_count;
+
+    /* For each Draw attribute, route it to the fragment shader according
+     * to the tab. */
+    attrib_count = vinfo->num_attribs;
+    DBG(r300, DBG_DRAW, "r300: attrib count: %d\n", attrib_count);
+    for (i = 0; i < attrib_count; i++) {
+        DBG(r300, DBG_DRAW, "r300: attrib: offset %d, interp %d, size %d,"
+               " tab %d\n", vinfo->attrib[i].src_index,
+               vinfo->attrib[i].interp_mode, vinfo->attrib[i].emit,
+               tab[i]);
+    }
+
     for (i = 0; i < attrib_count; i++) {
         /* Make sure we have a proper destination for our attribute. */
         assert(tab[i] != -1);
@@ -272,12 +265,10 @@ static void r300_vertex_psc(struct r300_context* r300,
         /* Add the attribute to the PSC table. */
         if (i & 1) {
             vformat->vap_prog_stream_cntl[i >> 1] |= type << 16;
-
             vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 16;
         } else {
-            vformat->vap_prog_stream_cntl[i >> 1] |= type <<  0;
-
-            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle << 0;
+            vformat->vap_prog_stream_cntl[i >> 1] |= type;
+            vformat->vap_prog_stream_cntl_ext[i >> 1] |= swizzle;
         }
     }
 
@@ -505,7 +496,13 @@ static void r300_update_derived_shader_state(struct r300_context* r300)
     }
 
     r300_vs_tab_routes(r300, vformat);
-    r300_vertex_psc(r300, vformat);
+
+    if (r300screen->caps->has_tcl) {
+        r300_vertex_psc(r300, vformat);
+    } else {
+        r300_swtcl_vertex_psc(r300, vformat);
+    }
+
     r300_update_fs_tab(r300, vformat);
 
     r300_update_rs_block(r300, rs_block);
@@ -553,8 +550,7 @@ static void r300_update_ztop(struct r300_context* r300)
 
 void r300_update_derived_state(struct r300_context* r300)
 {
-    /* XXX */
-    if (TRUE || r300->dirty_state &
+    if (r300->dirty_state &
         (R300_NEW_FRAGMENT_SHADER | R300_NEW_VERTEX_SHADER)) {
         r300_update_derived_shader_state(r300);
     }
diff --git a/src/gallium/drivers/r300/r300_vbo.c b/src/gallium/drivers/r300/r300_vbo.c
index a6a159667a3..6ebaf715dc5 100644
--- a/src/gallium/drivers/r300/r300_vbo.c
+++ b/src/gallium/drivers/r300/r300_vbo.c
@@ -34,52 +34,6 @@
 #include "r300_reg.h"
 #include "r300_winsys.h"
 
-static INLINE void setup_vertex_attribute(struct r300_vertex_info *vinfo,
-                                          struct pipe_vertex_element *vert_elem,
-                                          unsigned attr_num)
-{
-    uint16_t hw_fmt1, hw_fmt2;
-
-    hw_fmt1 = r300_translate_vertex_data_type(vert_elem->src_format) |
-        (attr_num << R300_DST_VEC_LOC_SHIFT);
-    hw_fmt2 = r300_translate_vertex_data_swizzle(vert_elem->src_format);
-
-    if (attr_num % 2 == 0)
-    {
-        vinfo->vap_prog_stream_cntl[attr_num >> 1] = hw_fmt1;
-        vinfo->vap_prog_stream_cntl_ext[attr_num >> 1] = hw_fmt2;
-    }
-    else
-    {
-        vinfo->vap_prog_stream_cntl[attr_num >> 1] |= hw_fmt1 << 16;
-        vinfo->vap_prog_stream_cntl_ext[attr_num >> 1] |= hw_fmt2 << 16;
-    }
-}
-
-static void finish_vertex_attribs_setup(struct r300_vertex_info *vinfo,
-                                        unsigned attribs_num)
-{
-    uint32_t last_vec_bit = (attribs_num % 2 == 0) ?
-        (R300_LAST_VEC << 16) : R300_LAST_VEC;
-
-    assert(attribs_num > 0 && attribs_num <= 16);
-    vinfo->vap_prog_stream_cntl[(attribs_num - 1) >> 1] |= last_vec_bit;
-}
-
-void setup_vertex_attributes(struct r300_context *r300)
-{
-    struct pipe_vertex_element *vert_elem;
-    int i;
-
-    for (i = 0; i < r300->vertex_element_count; i++) {
-        vert_elem = &r300->vertex_element[i];
-        setup_vertex_attribute(r300->vertex_info, vert_elem, i);
-    }
-
-    finish_vertex_attribs_setup(r300->vertex_info,
-        r300->vertex_element_count);
-}
-
 static INLINE int get_buffer_offset(struct r300_context *r300,
                                     unsigned int buf_nr,
                                     unsigned int elem_offset)
diff --git a/src/gallium/drivers/r300/r300_vs.h b/src/gallium/drivers/r300/r300_vs.h
index 2a4ce315e32..00b02bf510d 100644
--- a/src/gallium/drivers/r300/r300_vs.h
+++ b/src/gallium/drivers/r300/r300_vs.h
@@ -35,9 +35,6 @@ struct r300_vertex_shader {
     struct pipe_shader_state state;
     struct tgsi_shader_info info;
 
-    /* Fallback shader, because Draw has issues */
-    struct draw_vertex_shader* draw;
-
     /* Has this shader been translated yet? */
     boolean translated;