llvmpipe: Temporary workaround to prevent segfault on array textures.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_setup.c
index 28d202bd6555a78a9680fe59b97b0263987a72fa..8c9f0c646539b113b3b76e300bbd9bd23bec6602 100644 (file)
@@ -39,6 +39,7 @@
 #include "util/u_inlines.h"
 #include "util/u_memory.h"
 #include "util/u_pack_color.h"
+#include "draw/draw_pipe.h"
 #include "lp_context.h"
 #include "lp_memory.h"
 #include "lp_scene.h"
@@ -56,7 +57,7 @@
 #include "draw/draw_vbuf.h"
 
 
-static void set_scene_state( struct lp_setup_context *, enum setup_state,
+static boolean set_scene_state( struct lp_setup_context *, enum setup_state,
                              const char *reason);
 static boolean try_update_scene_state( struct lp_setup_context *setup );
 
@@ -64,6 +65,9 @@ static boolean try_update_scene_state( struct lp_setup_context *setup );
 static void
 lp_setup_get_empty_scene(struct lp_setup_context *setup)
 {
+   struct llvmpipe_context *lp = llvmpipe_context(setup->pipe);
+   boolean discard = lp->rasterizer ? lp->rasterizer->rasterizer_discard : FALSE;
+
    assert(setup->scene == NULL);
 
    setup->scene_idx++;
@@ -79,7 +83,7 @@ lp_setup_get_empty_scene(struct lp_setup_context *setup)
       lp_fence_wait(setup->scene->fence);
    }
 
-   lp_scene_begin_binning(setup->scene, &setup->fb);
+   lp_scene_begin_binning(setup->scene, &setup->fb, discard);
    
 }
 
@@ -114,13 +118,17 @@ first_point( struct lp_setup_context *setup,
    setup->point( setup, v0 );
 }
 
-static void lp_setup_reset( struct lp_setup_context *setup )
+void lp_setup_reset( struct lp_setup_context *setup )
 {
+   unsigned i;
+
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
    /* Reset derived state */
-   setup->constants.stored_size = 0;
-   setup->constants.stored_data = NULL;
+   for (i = 0; i < Elements(setup->constants); ++i) {
+      setup->constants[i].stored_size = 0;
+      setup->constants[i].stored_data = NULL;
+   }
    setup->fs.stored = NULL;
    setup->dirty = ~0;
 
@@ -167,7 +175,7 @@ lp_setup_rasterize_scene( struct lp_setup_context *setup )
 
 
 
-static void
+static boolean
 begin_binning( struct lp_setup_context *setup )
 {
    struct lp_scene *scene = setup->scene;
@@ -181,6 +189,8 @@ begin_binning( struct lp_setup_context *setup )
    /* Always create a fence:
     */
    scene->fence = lp_fence_create(MAX2(1, setup->num_threads));
+   if (!scene->fence)
+      return FALSE;
 
    /* Initialize the bin flags and x/y coords:
     */
@@ -192,7 +202,8 @@ begin_binning( struct lp_setup_context *setup )
    }
 
    ok = try_update_scene_state(setup);
-   assert(ok);
+   if (!ok)
+      return FALSE;
 
    if (setup->fb.zsbuf &&
        ((setup->clear.flags & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL) &&
@@ -208,7 +219,8 @@ begin_binning( struct lp_setup_context *setup )
          ok = lp_scene_bin_everywhere( scene, 
                                        LP_RAST_OP_CLEAR_COLOR, 
                                        setup->clear.color );
-         assert(ok);
+         if (!ok)
+            return FALSE;
       }
    }
 
@@ -216,28 +228,33 @@ begin_binning( struct lp_setup_context *setup )
       if (setup->clear.flags & PIPE_CLEAR_DEPTHSTENCIL) {
          if (!need_zsload)
             scene->has_depthstencil_clear = TRUE;
+
          ok = lp_scene_bin_everywhere( scene,
                                        LP_RAST_OP_CLEAR_ZSTENCIL,
                                        lp_rast_arg_clearzs(
                                           setup->clear.zsvalue,
                                           setup->clear.zsmask));
-         assert(ok);
+         if (!ok)
+            return FALSE;
       }
    }
 
-   if (setup->active_query) {
-      ok = lp_scene_bin_everywhere( scene,
-                                    LP_RAST_OP_BEGIN_QUERY,
-                                    lp_rast_arg_query(setup->active_query) );
-      assert(ok);
+   for (i = 0; i < PIPE_QUERY_TYPES; ++i) {
+      if (setup->active_query[i]) {
+         ok = lp_scene_bin_everywhere( scene,
+                                       LP_RAST_OP_BEGIN_QUERY,
+                                       lp_rast_arg_query(setup->active_query[i]) );
+         if (!ok)
+            return FALSE;
+      }
    }
-      
 
    setup->clear.flags = 0;
    setup->clear.zsmask = 0;
    setup->clear.zsvalue = 0;
 
    LP_DBG(DEBUG_SETUP, "%s done\n", __FUNCTION__);
+   return TRUE;
 }
 
 
@@ -246,23 +263,22 @@ begin_binning( struct lp_setup_context *setup )
  *
  * TODO: fast path for fullscreen clears and no triangles.
  */
-static void
+static boolean
 execute_clears( struct lp_setup_context *setup )
 {
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
-   begin_binning( setup );
+   return begin_binning( setup );
 }
 
 const char *states[] = {
    "FLUSHED",
-   "EMPTY  ",
    "CLEARED",
    "ACTIVE "
 };
 
 
-static void
+static boolean
 set_scene_state( struct lp_setup_context *setup,
                  enum setup_state new_state,
                  const char *reason)
@@ -270,7 +286,7 @@ set_scene_state( struct lp_setup_context *setup,
    unsigned old_state = setup->state;
 
    if (old_state == new_state)
-      return;
+      return TRUE;
    
    if (LP_DEBUG & DEBUG_SCENE) {
       debug_printf("%s old %s new %s%s%s\n",
@@ -294,12 +310,14 @@ set_scene_state( struct lp_setup_context *setup,
       break;
 
    case SETUP_ACTIVE:
-      begin_binning( setup );
+      if (!begin_binning( setup ))
+         goto fail;
       break;
 
    case SETUP_FLUSHED:
       if (old_state == SETUP_CLEARED)
-         execute_clears( setup );
+         if (!execute_clears( setup ))
+            goto fail;
 
       lp_setup_rasterize_scene( setup );
       assert(setup->scene == NULL);
@@ -307,18 +325,26 @@ set_scene_state( struct lp_setup_context *setup,
 
    default:
       assert(0 && "invalid setup state mode");
+      goto fail;
    }
 
    setup->state = new_state;
+   return TRUE;
+
+fail:
+   if (setup->scene) {
+      lp_scene_end_rasterization(setup->scene);
+      setup->scene = NULL;
+   }
+
+   setup->state = SETUP_FLUSHED;
+   lp_setup_reset( setup );
+   return FALSE;
 }
 
 
-/**
- * \param flags  bitmask of PIPE_FLUSH_x flags
- */
 void
 lp_setup_flush( struct lp_setup_context *setup,
-                unsigned flags,
                 struct pipe_fence_handle **fence,
                 const char *reason)
 {
@@ -359,7 +385,7 @@ lp_setup_bind_framebuffer( struct lp_setup_context *setup,
 
 static boolean
 lp_setup_try_clear( struct lp_setup_context *setup,
-                    const float *color,
+                    const union pipe_color_union *color,
                     double depth,
                     unsigned stencil,
                     unsigned flags )
@@ -373,20 +399,23 @@ lp_setup_try_clear( struct lp_setup_context *setup,
 
    if (flags & PIPE_CLEAR_COLOR) {
       for (i = 0; i < 4; i++)
-         color_arg.clear_color[i] = float_to_ubyte(color[i]);
+         color_arg.clear_color.i[i] = color->i[i];
    }
 
    if (flags & PIPE_CLEAR_DEPTHSTENCIL) {
-      unsigned zmask = (flags & PIPE_CLEAR_DEPTH) ? ~0 : 0;
-      unsigned smask = (flags & PIPE_CLEAR_STENCIL) ? ~0 : 0;
+      uint32_t zmask = (flags & PIPE_CLEAR_DEPTH) ? ~0 : 0;
+      uint32_t smask = (flags & PIPE_CLEAR_STENCIL) ? ~0 : 0;
 
       zsvalue = util_pack_z_stencil(setup->fb.zsbuf->format,
                                     depth,
                                     stencil);
 
-      zsmask = util_pack_uint_z_stencil(setup->fb.zsbuf->format,
+
+      zsmask = util_pack_mask_z_stencil(setup->fb.zsbuf->format,
                                         zmask,
                                         smask);
+
+      zsvalue &= zsmask;
    }
 
    if (setup->state == SETUP_ACTIVE) {
@@ -429,9 +458,9 @@ lp_setup_try_clear( struct lp_setup_context *setup,
       }
 
       if (flags & PIPE_CLEAR_COLOR) {
-         memcpy(setup->clear.color.clear_color,
+         memcpy(&setup->clear.color.clear_color,
                 &color_arg,
-                sizeof color_arg);
+                sizeof setup->clear.color.clear_color);
       }
    }
    
@@ -440,13 +469,13 @@ lp_setup_try_clear( struct lp_setup_context *setup,
 
 void
 lp_setup_clear( struct lp_setup_context *setup,
-                const float *color,
+                const union pipe_color_union *color,
                 double depth,
                 unsigned stencil,
                 unsigned flags )
 {
    if (!lp_setup_try_clear( setup, color, depth, stencil, flags )) {
-      lp_setup_flush(setup, 0, NULL, __FUNCTION__);
+      lp_setup_flush(setup, NULL, __FUNCTION__);
 
       if (!lp_setup_try_clear( setup, color, depth, stencil, flags ))
          assert(0);
@@ -462,14 +491,16 @@ lp_setup_set_triangle_state( struct lp_setup_context *setup,
                              unsigned cull_mode,
                              boolean ccw_is_frontface,
                              boolean scissor,
-                             boolean gl_rasterization_rules)
+                             boolean half_pixel_center,
+                             boolean bottom_edge_rule)
 {
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
    setup->ccw_is_frontface = ccw_is_frontface;
    setup->cullmode = cull_mode;
    setup->triangle = first_triangle;
-   setup->pixel_offset = gl_rasterization_rules ? 0.5f : 0.0f;
+   setup->pixel_offset = half_pixel_center ? 0.5f : 0.0f;
+   setup->bottom_edge_rule = bottom_edge_rule;
 
    if (setup->scissor_test != scissor) {
       setup->dirty |= LP_SETUP_NEW_SCISSOR;
@@ -502,14 +533,12 @@ lp_setup_set_point_state( struct lp_setup_context *setup,
 }
 
 void
-lp_setup_set_fs_inputs( struct lp_setup_context *setup,
-                        const struct lp_shader_input *input,
-                        unsigned nr )
+lp_setup_set_setup_variant( struct lp_setup_context *setup,
+                           const struct lp_setup_variant *variant)
 {
-   LP_DBG(DEBUG_SETUP, "%s %p %u\n", __FUNCTION__, (void *) input, nr);
-
-   memcpy( setup->fs.input, input, nr * sizeof input[0] );
-   setup->fs.nr_inputs = nr;
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+   
+   setup->setup.variant = variant;
 }
 
 void
@@ -526,12 +555,21 @@ lp_setup_set_fs_variant( struct lp_setup_context *setup,
 
 void
 lp_setup_set_fs_constants(struct lp_setup_context *setup,
-                          struct pipe_resource *buffer)
+                          unsigned num,
+                          struct pipe_constant_buffer *buffers)
 {
-   LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *) buffer);
+   unsigned i;
 
-   pipe_resource_reference(&setup->constants.current, buffer);
+   LP_DBG(DEBUG_SETUP, "%s %p\n", __FUNCTION__, (void *) buffers);
 
+   assert(num <= Elements(setup->constants));
+
+   for (i = 0; i < num; ++i) {
+      util_copy_constant_buffer(&setup->constants[i].current, &buffers[i]);
+   }
+   for (; i < Elements(setup->constants); i++) {
+      util_copy_constant_buffer(&setup->constants[i].current, NULL);
+   }
    setup->dirty |= LP_SETUP_NEW_CONSTANTS;
 }
 
@@ -600,6 +638,15 @@ lp_setup_set_flatshade_first( struct lp_setup_context *setup,
    setup->flatshade_first = flatshade_first;
 }
 
+void
+lp_setup_set_rasterizer_discard( struct lp_setup_context *setup,
+                                 boolean rasterizer_discard )
+{
+   if (setup->rasterizer_discard != rasterizer_discard) {
+      setup->rasterizer_discard = rasterizer_discard;
+      set_scene_state( setup, SETUP_FLUSHED, __FUNCTION__ );
+   }
+}
 
 void 
 lp_setup_set_vertex_info( struct lp_setup_context *setup,
@@ -617,58 +664,123 @@ lp_setup_set_vertex_info( struct lp_setup_context *setup,
 void
 lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
                                     unsigned num,
-                                    struct pipe_sampler_view **views,
-                                    const struct pipe_sampler_state **samplers)
+                                    struct pipe_sampler_view **views)
 {
    unsigned i;
 
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
 
-   assert(num <= PIPE_MAX_SAMPLERS);
+   assert(num <= PIPE_MAX_SHADER_SAMPLER_VIEWS);
 
-   for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
+   for (i = 0; i < PIPE_MAX_SHADER_SAMPLER_VIEWS; i++) {
       struct pipe_sampler_view *view = i < num ? views[i] : NULL;
 
-      if(view) {
-         struct pipe_resource *tex = view->texture;
-         struct llvmpipe_resource *lp_tex = llvmpipe_resource(tex);
+      if (view) {
+         struct pipe_resource *res = view->texture;
+         struct llvmpipe_resource *lp_tex = llvmpipe_resource(res);
          struct lp_jit_texture *jit_tex;
          jit_tex = &setup->fs.current.jit_context.textures[i];
-         jit_tex->width = tex->width0;
-         jit_tex->height = tex->height0;
-         jit_tex->depth = tex->depth0;
-         jit_tex->last_level = tex->last_level;
-
-         /* sampler state */
-         jit_tex->min_lod = samplers[i]->min_lod;
-         jit_tex->max_lod = samplers[i]->max_lod;
-         jit_tex->lod_bias = samplers[i]->lod_bias;
 
          /* We're referencing the texture's internal data, so save a
           * reference to it.
           */
-         pipe_resource_reference(&setup->fs.current_tex[i], tex);
+         pipe_resource_reference(&setup->fs.current_tex[i], res);
 
          if (!lp_tex->dt) {
-            /* regular texture - setup array of mipmap level pointers */
+            /* regular texture - setup array of mipmap level offsets */
+            void *mip_ptr;
             int j;
-            for (j = 0; j <= tex->last_level; j++) {
-               jit_tex->data[j] =
-                  llvmpipe_get_texture_image_all(lp_tex, j, LP_TEX_USAGE_READ,
-                                                 LP_TEX_LAYOUT_LINEAR);
-               jit_tex->row_stride[j] = lp_tex->row_stride[j];
-               jit_tex->img_stride[j] = lp_tex->img_stride[j];
-
-               if ((LP_PERF & PERF_TEX_MEM) ||
-                  !jit_tex->data[j]) {
-                  /* out of memory - use dummy tile memory */
-                  jit_tex->data[j] = lp_dummy_tile;
-                  jit_tex->width = TILE_SIZE/8;
-                  jit_tex->height = TILE_SIZE/8;
-                  jit_tex->depth = 1;
-                  jit_tex->last_level = 0;
-                  jit_tex->row_stride[j] = 0;
-                  jit_tex->img_stride[j] = 0;
+            unsigned first_level = 0;
+            unsigned last_level = 0;
+
+            if (llvmpipe_resource_is_texture(res)) {
+               first_level = view->u.tex.first_level;
+               last_level = view->u.tex.last_level;
+               assert(first_level <= last_level);
+               assert(last_level <= res->last_level);
+
+               /*
+                * The complexity here is only necessary for depth textures which
+                * still are tiled.
+                */
+               mip_ptr = llvmpipe_get_texture_image_all(lp_tex, first_level,
+                                                        LP_TEX_USAGE_READ,
+                                                        LP_TEX_LAYOUT_LINEAR);
+               jit_tex->base = lp_tex->linear_img.data;
+            }
+            else {
+               mip_ptr = lp_tex->data;
+               jit_tex->base = mip_ptr;
+            }
+
+            if ((LP_PERF & PERF_TEX_MEM) || !mip_ptr) {
+               /* out of memory - use dummy tile memory */
+               /* Note if using PERF_TEX_MEM will also skip tile conversion */
+               jit_tex->base = lp_dummy_tile;
+               jit_tex->width = TILE_SIZE/8;
+               jit_tex->height = TILE_SIZE/8;
+               jit_tex->depth = 1;
+               jit_tex->first_level = 0;
+               jit_tex->last_level = 0;
+               jit_tex->mip_offsets[0] = 0;
+               jit_tex->row_stride[0] = 0;
+               jit_tex->img_stride[0] = 0;
+            }
+            else {
+               jit_tex->width = res->width0;
+               jit_tex->height = res->height0;
+               jit_tex->depth = res->depth0;
+               jit_tex->first_level = first_level;
+               jit_tex->last_level = last_level;
+
+               if (llvmpipe_resource_is_texture(res)) {
+                  for (j = first_level; j <= last_level; j++) {
+                     mip_ptr = llvmpipe_get_texture_image_all(lp_tex, j,
+                                                              LP_TEX_USAGE_READ,
+                                                              LP_TEX_LAYOUT_LINEAR);
+                     jit_tex->mip_offsets[j] = (uint8_t *)mip_ptr - (uint8_t *)jit_tex->base;
+                     /*
+                      * could get mip offset directly but need call above to
+                      * invoke tiled->linear conversion.
+                      */
+                     assert(lp_tex->linear_mip_offsets[j] == jit_tex->mip_offsets[j]);
+                     jit_tex->row_stride[j] = lp_tex->row_stride[j];
+                     jit_tex->img_stride[j] = lp_tex->img_stride[j];
+                  }
+
+                 /* FIXME: This is incorrect, as currently layer stride depends on miplevel */
+#if 0
+                  /*
+                   * We don't use anything like first_element (for buffers) or
+                   * first_layer (for arrays), instead adjust the last_element
+                   * (width) or last_layer (depth) plus the base pointer.
+                   * Less parameters and faster at shader execution.
+                   * XXX Could do the same for mip levels.
+                   */
+                  if (res->target == PIPE_TEXTURE_1D_ARRAY ||
+                      res->target == PIPE_TEXTURE_2D_ARRAY) {
+                     jit_tex->depth = view->u.tex.last_layer - view->u.tex.first_layer + 1;
+                     jit_tex->base = (uint8_t *)jit_tex->base +
+                                     view->u.tex.first_layer * lp_tex->img_stride[0];
+                     assert(view->u.tex.first_layer <= view->u.tex.last_layer);
+                     assert(view->u.tex.last_layer < res->array_size);
+                  }
+#endif
+               }
+               else {
+                  unsigned view_blocksize = util_format_get_blocksize(view->format);
+                  /* probably don't really need to fill that out */
+                  jit_tex->mip_offsets[0] = 0;
+                  jit_tex->row_stride[0] = 0;
+                  jit_tex->row_stride[0] = 0;
+
+                  /* everything specified in number of elements here. */
+                  jit_tex->width = view->u.buf.last_element - view->u.buf.first_element + 1;
+                  jit_tex->base = (uint8_t *)jit_tex->base + view->u.buf.first_element *
+                                  view_blocksize;
+                  /* XXX Unsure if we need to sanitize parameters? */
+                  assert(view->u.buf.first_element <= view->u.buf.last_element);
+                  assert(view->u.buf.last_element * view_blocksize < res->width0);
                }
             }
          }
@@ -677,13 +789,18 @@ lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
             /*
              * XXX: Where should this be unmapped?
              */
-            struct llvmpipe_screen *screen = llvmpipe_screen(tex->screen);
+            struct llvmpipe_screen *screen = llvmpipe_screen(res->screen);
             struct sw_winsys *winsys = screen->winsys;
-            jit_tex->data[0] = winsys->displaytarget_map(winsys, lp_tex->dt,
-                                                        PIPE_TRANSFER_READ);
+            jit_tex->base = winsys->displaytarget_map(winsys, lp_tex->dt,
+                                                         PIPE_TRANSFER_READ);
             jit_tex->row_stride[0] = lp_tex->row_stride[0];
             jit_tex->img_stride[0] = lp_tex->img_stride[0];
-            assert(jit_tex->data[0]);
+            jit_tex->mip_offsets[0] = 0;
+            jit_tex->width = res->width0;
+            jit_tex->height = res->height0;
+            jit_tex->depth = res->depth0;
+            jit_tex->first_level = jit_tex->last_level = 0;
+            assert(jit_tex->base);
          }
       }
    }
@@ -692,6 +809,38 @@ lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
 }
 
 
+/**
+ * Called during state validation when LP_NEW_SAMPLER is set.
+ */
+void
+lp_setup_set_fragment_sampler_state(struct lp_setup_context *setup,
+                                    unsigned num,
+                                    struct pipe_sampler_state **samplers)
+{
+   unsigned i;
+
+   LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
+
+   assert(num <= PIPE_MAX_SAMPLERS);
+
+   for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
+      const struct pipe_sampler_state *sampler = i < num ? samplers[i] : NULL;
+
+      if (sampler) {
+         struct lp_jit_sampler *jit_sam;
+         jit_sam = &setup->fs.current.jit_context.samplers[i];
+
+         jit_sam->min_lod = sampler->min_lod;
+         jit_sam->max_lod = sampler->max_lod;
+         jit_sam->lod_bias = sampler->lod_bias;
+         COPY_4V(jit_sam->border_color, sampler->border_color.f);
+      }
+   }
+
+   setup->dirty |= LP_SETUP_NEW_FS;
+}
+
+
 /**
  * Is the given texture referenced by any scene?
  * Note: we have to check all scenes including any scenes currently
@@ -706,20 +855,20 @@ lp_setup_is_resource_referenced( const struct lp_setup_context *setup,
    /* check the render targets */
    for (i = 0; i < setup->fb.nr_cbufs; i++) {
       if (setup->fb.cbufs[i]->texture == texture)
-         return PIPE_REFERENCED_FOR_READ | PIPE_REFERENCED_FOR_WRITE;
+         return LP_REFERENCED_FOR_READ | LP_REFERENCED_FOR_WRITE;
    }
    if (setup->fb.zsbuf && setup->fb.zsbuf->texture == texture) {
-      return PIPE_REFERENCED_FOR_READ | PIPE_REFERENCED_FOR_WRITE;
+      return LP_REFERENCED_FOR_READ | LP_REFERENCED_FOR_WRITE;
    }
 
    /* check textures referenced by the scene */
    for (i = 0; i < Elements(setup->scenes); i++) {
       if (lp_scene_is_resource_referenced(setup->scenes[i], texture)) {
-         return PIPE_REFERENCED_FOR_READ;
+         return LP_REFERENCED_FOR_READ;
       }
    }
 
-   return PIPE_UNREFERENCED;
+   return LP_UNREFERENCED;
 }
 
 
@@ -731,19 +880,32 @@ try_update_scene_state( struct lp_setup_context *setup )
 {
    boolean new_scene = (setup->fs.stored == NULL);
    struct lp_scene *scene = setup->scene;
+   unsigned i;
 
    assert(scene);
 
    if(setup->dirty & LP_SETUP_NEW_BLEND_COLOR) {
       uint8_t *stored;
+      float* fstored;
       unsigned i, j;
+      unsigned size;
+
+      /* Alloc u8_blend_color (16 x i8) and f_blend_color (4 or 8 x f32) */
+      size  = 4 * 16 * sizeof(uint8_t);
+      size += (LP_MAX_VECTOR_LENGTH / 4) * sizeof(float);
+      stored = lp_scene_alloc_aligned(scene, size, LP_MAX_VECTOR_LENGTH);
 
-      stored = lp_scene_alloc_aligned(scene, 4 * 16, 16);
       if (!stored) {
          assert(!new_scene);
          return FALSE;
       }
 
+      /* Store floating point colour */
+      fstored = (float*)(stored + 4*16);
+      for (i = 0; i < (LP_MAX_VECTOR_LENGTH / 4); ++i) {
+         fstored[i] = setup->blend_color.current.color[i % 4];
+      }
+
       /* smear each blend color component across 16 ubyte elements */
       for (i = 0; i < 4; ++i) {
          uint8_t c = float_to_ubyte(setup->blend_color.current.color[i]);
@@ -752,46 +914,59 @@ try_update_scene_state( struct lp_setup_context *setup )
       }
 
       setup->blend_color.stored = stored;
-      setup->fs.current.jit_context.blend_color = setup->blend_color.stored;
+      setup->fs.current.jit_context.u8_blend_color = stored;
+      setup->fs.current.jit_context.f_blend_color = fstored;
       setup->dirty |= LP_SETUP_NEW_FS;
    }
 
-   if(setup->dirty & LP_SETUP_NEW_CONSTANTS) {
-      struct pipe_resource *buffer = setup->constants.current;
+   if (setup->dirty & LP_SETUP_NEW_CONSTANTS) {
+      for (i = 0; i < Elements(setup->constants); ++i) {
+         struct pipe_resource *buffer = setup->constants[i].current.buffer;
+         const unsigned current_size = setup->constants[i].current.buffer_size;
+         const ubyte *current_data = NULL;
 
-      if(buffer) {
-         unsigned current_size = buffer->width0;
-         const void *current_data = llvmpipe_resource_data(buffer);
+         if (buffer) {
+            /* resource buffer */
+            current_data = (ubyte *) llvmpipe_resource_data(buffer);
+         }
+         else if (setup->constants[i].current.user_buffer) {
+            /* user-space buffer */
+            current_data = (ubyte *) setup->constants[i].current.user_buffer;
+         }
 
-         /* TODO: copy only the actually used constants? */
+         if (current_data) {
+            current_data += setup->constants[i].current.buffer_offset;
 
-         if(setup->constants.stored_size != current_size ||
-            !setup->constants.stored_data ||
-            memcmp(setup->constants.stored_data,
-                   current_data,
-                   current_size) != 0) {
-            void *stored;
+            /* TODO: copy only the actually used constants? */
 
-            stored = lp_scene_alloc(scene, current_size);
-            if (!stored) {
-               assert(!new_scene);
-               return FALSE;
-            }
+            if (setup->constants[i].stored_size != current_size ||
+               !setup->constants[i].stored_data ||
+               memcmp(setup->constants[i].stored_data,
+                      current_data,
+                      current_size) != 0) {
+               void *stored;
+
+               stored = lp_scene_alloc(scene, current_size);
+               if (!stored) {
+                  assert(!new_scene);
+                  return FALSE;
+               }
 
-            memcpy(stored,
-                   current_data,
-                   current_size);
-            setup->constants.stored_size = current_size;
-            setup->constants.stored_data = stored;
+               memcpy(stored,
+                      current_data,
+                      current_size);
+               setup->constants[i].stored_size = current_size;
+               setup->constants[i].stored_data = stored;
+            }
+         }
+         else {
+            setup->constants[i].stored_size = 0;
+            setup->constants[i].stored_data = NULL;
          }
-      }
-      else {
-         setup->constants.stored_size = 0;
-         setup->constants.stored_data = NULL;
-      }
 
-      setup->fs.current.jit_context.constants = setup->constants.stored_data;
-      setup->dirty |= LP_SETUP_NEW_FS;
+         setup->fs.current.jit_context.constants[i] = setup->constants[i].stored_data;
+         setup->dirty |= LP_SETUP_NEW_FS;
+      }
    }
 
 
@@ -802,7 +977,6 @@ try_update_scene_state( struct lp_setup_context *setup )
                  sizeof setup->fs.current) != 0)
       {
          struct lp_rast_state *stored;
-         uint i;
          
          /* The fs state that's been stored in the scene is different from
           * the new, current state.  So allocate a new lp_rast_state object
@@ -841,6 +1015,16 @@ try_update_scene_state( struct lp_setup_context *setup )
          u_rect_possible_intersection(&setup->scissor,
                                       &setup->draw_region);
       }
+      /* If the framebuffer is large we have to think about fixed-point
+       * integer overflow.  For 2K by 2K images, coordinates need 15 bits
+       * (2^11 + 4 subpixel bits).  The product of two such numbers would
+       * use 30 bits.  Any larger and we could overflow a 32-bit int.
+       *
+       * To cope with this problem we check if triangles are large and
+       * subdivide them if needed.
+       */
+      setup->subdivide_large_triangles = (setup->fb.width > 2048 &&
+                                          setup->fb.height > 2048);
    }
                                       
    setup->dirty = 0;
@@ -849,7 +1033,7 @@ try_update_scene_state( struct lp_setup_context *setup )
    return TRUE;
 }
 
-void
+boolean
 lp_setup_update_state( struct lp_setup_context *setup,
                        boolean update_scene )
 {
@@ -865,28 +1049,59 @@ lp_setup_update_state( struct lp_setup_context *setup,
          llvmpipe_update_derived(lp);
       }
 
+      if (lp->setup->dirty) {
+         llvmpipe_update_setup(lp);
+      }
+
+      assert(setup->setup.variant);
+
       /* Will probably need to move this somewhere else, just need  
        * to know about vertex shader point size attribute.
        */
       setup->psize = lp->psize_slot;
 
       assert(lp->dirty == 0);
+
+      assert(lp->setup_variant.key.size == 
+            setup->setup.variant->key.size);
+
+      assert(memcmp(&lp->setup_variant.key,
+                   &setup->setup.variant->key,
+                   setup->setup.variant->key.size) == 0);
    }
 
-   if (update_scene)
-      set_scene_state( setup, SETUP_ACTIVE, __FUNCTION__ );
+   if (update_scene && setup->state != SETUP_ACTIVE) {
+      if (!set_scene_state( setup, SETUP_ACTIVE, __FUNCTION__ ))
+         return FALSE;
+   }
 
    /* Only call into update_scene_state() if we already have a
     * scene:
     */
    if (update_scene && setup->scene) {
       assert(setup->state == SETUP_ACTIVE);
-      if (!try_update_scene_state(setup)) {
-         lp_setup_flush_and_restart(setup);
-         if (!try_update_scene_state(setup))
-            assert(0);
-      }
+
+      if (try_update_scene_state(setup))
+         return TRUE;
+
+      /* Update failed, try to restart the scene.
+       *
+       * Cannot call lp_setup_flush_and_restart() directly here
+       * because of potential recursion.
+       */
+      if (!set_scene_state(setup, SETUP_FLUSHED, __FUNCTION__))
+         return FALSE;
+
+      if (!set_scene_state(setup, SETUP_ACTIVE, __FUNCTION__))
+         return FALSE;
+
+      if (!setup->scene)
+         return FALSE;
+
+      return try_update_scene_state(setup);
    }
+
+   return TRUE;
 }
 
 
@@ -906,7 +1121,9 @@ lp_setup_destroy( struct lp_setup_context *setup )
       pipe_resource_reference(&setup->fs.current_tex[i], NULL);
    }
 
-   pipe_resource_reference(&setup->constants.current, NULL);
+   for (i = 0; i < Elements(setup->constants); i++) {
+      pipe_resource_reference(&setup->constants[i].current.buffer, NULL);
+   }
 
    /* free the scenes in the 'empty' queue */
    for (i = 0; i < Elements(setup->scenes); i++) {
@@ -918,6 +1135,8 @@ lp_setup_destroy( struct lp_setup_context *setup )
       lp_scene_destroy(scene);
    }
 
+   lp_fence_reference(&setup->last_fence, NULL);
+
    FREE( setup );
 }
 
@@ -932,11 +1151,13 @@ lp_setup_create( struct pipe_context *pipe,
                  struct draw_context *draw )
 {
    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
-   struct lp_setup_context *setup = CALLOC_STRUCT(lp_setup_context);
+   struct lp_setup_context *setup;
    unsigned i;
 
-   if (!setup)
-      return NULL;
+   setup = CALLOC_STRUCT(lp_setup_context);
+   if (!setup) {
+      goto no_setup;
+   }
 
    lp_setup_init_vbuf(setup);
    
@@ -947,8 +1168,9 @@ lp_setup_create( struct pipe_context *pipe,
 
    setup->num_threads = screen->num_threads;
    setup->vbuf = draw_vbuf_stage(draw, &setup->base);
-   if (!setup->vbuf)
-      goto fail;
+   if (!setup->vbuf) {
+      goto no_vbuf;
+   }
 
    draw_set_rasterize_stage(draw, setup->vbuf);
    draw_set_render(draw, &setup->base);
@@ -956,6 +1178,9 @@ lp_setup_create( struct pipe_context *pipe,
    /* create some empty scenes */
    for (i = 0; i < MAX_SCENES; i++) {
       setup->scenes[i] = lp_scene_create( pipe );
+      if (!setup->scenes[i]) {
+         goto no_scenes;
+      }
    }
 
    setup->triangle = first_triangle;
@@ -966,11 +1191,17 @@ lp_setup_create( struct pipe_context *pipe,
 
    return setup;
 
-fail:
-   if (setup->vbuf)
-      ;
+no_scenes:
+   for (i = 0; i < MAX_SCENES; i++) {
+      if (setup->scenes[i]) {
+         lp_scene_destroy(setup->scenes[i]);
+      }
+   }
 
+   setup->vbuf->destroy(setup->vbuf);
+no_vbuf:
    FREE(setup);
+no_setup:
    return NULL;
 }
 
@@ -983,25 +1214,32 @@ lp_setup_begin_query(struct lp_setup_context *setup,
                      struct llvmpipe_query *pq)
 {
    /* init the query to its beginning state */
-   assert(setup->active_query == NULL);
+   assert(setup->active_query[pq->type] == NULL);
+
+   set_scene_state(setup, SETUP_ACTIVE, "begin_query");
    
+   setup->active_query[pq->type] = pq;
+
+   /* XXX: It is possible that a query is created before the scene
+    * has been created. This means that setup->scene == NULL resulting
+    * in the query not being binned and thus is ignored.
+    */
+
    if (setup->scene) {
       if (!lp_scene_bin_everywhere(setup->scene,
                                    LP_RAST_OP_BEGIN_QUERY,
                                    lp_rast_arg_query(pq))) {
 
-         lp_setup_flush_and_restart(setup);
+         if (!lp_setup_flush_and_restart(setup))
+            return;
 
          if (!lp_scene_bin_everywhere(setup->scene,
                                       LP_RAST_OP_BEGIN_QUERY,
                                       lp_rast_arg_query(pq))) {
-            assert(0);
             return;
          }
       }
    }
-
-   setup->active_query = pq;
 }
 
 
@@ -1011,10 +1249,12 @@ lp_setup_begin_query(struct lp_setup_context *setup,
 void
 lp_setup_end_query(struct lp_setup_context *setup, struct llvmpipe_query *pq)
 {
-   union lp_rast_cmd_arg dummy = { 0 };
+   set_scene_state(setup, SETUP_ACTIVE, "end_query");
 
-   assert(setup->active_query == pq);
-   setup->active_query = NULL;
+   if (pq->type != PIPE_QUERY_TIMESTAMP) {
+      assert(setup->active_query[pq->type] == pq);
+      setup->active_query[pq->type] = NULL;
+   }
 
    /* Setup will automatically re-issue any query which carried over a
     * scene boundary, and the rasterizer automatically "ends" queries
@@ -1029,8 +1269,8 @@ lp_setup_end_query(struct lp_setup_context *setup, struct llvmpipe_query *pq)
 
       if (!lp_scene_bin_everywhere(setup->scene,
                                    LP_RAST_OP_END_QUERY,
-                                   dummy)) {
-         lp_setup_flush(setup, 0, NULL, __FUNCTION__);
+                                   lp_rast_arg_query(pq))) {
+         lp_setup_flush(setup, NULL, __FUNCTION__);
       }
    }
    else {
@@ -1039,14 +1279,20 @@ lp_setup_end_query(struct lp_setup_context *setup, struct llvmpipe_query *pq)
 }
 
 
-void
+boolean
 lp_setup_flush_and_restart(struct lp_setup_context *setup)
 {
    if (0) debug_printf("%s\n", __FUNCTION__);
 
    assert(setup->state == SETUP_ACTIVE);
-   set_scene_state(setup, SETUP_FLUSHED, __FUNCTION__);
-   lp_setup_update_state(setup, TRUE);
+
+   if (!set_scene_state(setup, SETUP_FLUSHED, __FUNCTION__))
+      return FALSE;
+   
+   if (!lp_setup_update_state(setup, TRUE))
+      return FALSE;
+
+   return TRUE;
 }