llvmpipe: pass mask into fragment shader

Move this code back out to C for now, will generate separately.

Shader now takes a mask parameter instead of C0/C1/C2/etc.

Shader does not currently use that parameter and rasterizes whole
pixel stamps always.

14 files changed, 901 insertions, 790 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c
index 23aa34ddec..8e6dfb293d 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.c
+++ b/src/gallium/drivers/llvmpipe/lp_jit.c
@@ -103,10 +103,6 @@ lp_jit_init_globals(struct llvmpipe_screen *screen)
       elem_types[LP_JIT_CTX_ALPHA_REF] = LLVMFloatType();
       elem_types[LP_JIT_CTX_STENCIL_REF_FRONT] = LLVMInt32Type();
       elem_types[LP_JIT_CTX_STENCIL_REF_BACK] = LLVMInt32Type();
-      elem_types[LP_JIT_CTX_SCISSOR_XMIN] = LLVMFloatType();
-      elem_types[LP_JIT_CTX_SCISSOR_YMIN] = LLVMFloatType();
-      elem_types[LP_JIT_CTX_SCISSOR_XMAX] = LLVMFloatType();
-      elem_types[LP_JIT_CTX_SCISSOR_YMAX] = LLVMFloatType();
       elem_types[LP_JIT_CTX_BLEND_COLOR] = LLVMPointerType(LLVMInt8Type(), 0);
       elem_types[LP_JIT_CTX_TEXTURES] = LLVMArrayType(texture_type,
                                                       PIPE_MAX_SAMPLERS);
@@ -125,18 +121,6 @@ lp_jit_init_globals(struct llvmpipe_screen *screen)
       LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, stencil_ref_back,
                              screen->target, context_type,
                              LP_JIT_CTX_STENCIL_REF_BACK);
-      LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, scissor_xmin,
-                             screen->target, context_type,
-                             LP_JIT_CTX_SCISSOR_XMIN);
-      LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, scissor_ymin,
-                             screen->target, context_type,
-                             LP_JIT_CTX_SCISSOR_YMIN);
-      LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, scissor_xmax,
-                             screen->target, context_type,
-                             LP_JIT_CTX_SCISSOR_XMAX);
-      LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, scissor_ymax,
-                             screen->target, context_type,
-                             LP_JIT_CTX_SCISSOR_YMAX);
       LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, blend_color,
                              screen->target, context_type,
                              LP_JIT_CTX_BLEND_COLOR);
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h
index 8d06e65725..c94189413a 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.h
+++ b/src/gallium/drivers/llvmpipe/lp_jit.h
@@ -89,9 +89,6 @@ struct lp_jit_context
 
    uint32_t stencil_ref_front, stencil_ref_back;
 
-   /** floats, not ints */
-   float scissor_xmin, scissor_ymin, scissor_xmax, scissor_ymax;
-
    /* FIXME: store (also?) in floats */
    uint8_t *blend_color;
 
@@ -108,10 +105,6 @@ enum {
    LP_JIT_CTX_ALPHA_REF,
    LP_JIT_CTX_STENCIL_REF_FRONT,
    LP_JIT_CTX_STENCIL_REF_BACK,
-   LP_JIT_CTX_SCISSOR_XMIN,
-   LP_JIT_CTX_SCISSOR_YMIN,
-   LP_JIT_CTX_SCISSOR_XMAX,
-   LP_JIT_CTX_SCISSOR_YMAX,
    LP_JIT_CTX_BLEND_COLOR,
    LP_JIT_CTX_TEXTURES,
    LP_JIT_CTX_COUNT
@@ -130,18 +123,6 @@ enum {
 #define lp_jit_context_stencil_ref_back_value(_builder, _ptr) \
    lp_build_struct_get(_builder, _ptr, LP_JIT_CTX_STENCIL_REF_BACK, "stencil_ref_back")
 
-#define lp_jit_context_scissor_xmin_value(_builder, _ptr) \
-   lp_build_struct_get(_builder, _ptr, LP_JIT_CTX_SCISSOR_XMIN, "scissor_xmin")
-
-#define lp_jit_context_scissor_ymin_value(_builder, _ptr) \
-   lp_build_struct_get(_builder, _ptr, LP_JIT_CTX_SCISSOR_YMIN, "scissor_ymin")
-
-#define lp_jit_context_scissor_xmax_value(_builder, _ptr) \
-   lp_build_struct_get(_builder, _ptr, LP_JIT_CTX_SCISSOR_XMAX, "scissor_xmax")
-
-#define lp_jit_context_scissor_ymax_value(_builder, _ptr) \
-   lp_build_struct_get(_builder, _ptr, LP_JIT_CTX_SCISSOR_YMAX, "scissor_ymax")
-
 #define lp_jit_context_blend_color(_builder, _ptr) \
    lp_build_struct_get(_builder, _ptr, LP_JIT_CTX_BLEND_COLOR, "blend_color")
 
@@ -160,12 +141,7 @@ typedef void
                     const void *dady,
                     uint8_t **color,
                     void *depth,
-                    const int32_t c1,
-                    const int32_t c2,
-                    const int32_t c3,
-                    const int32_t *step1,
-                    const int32_t *step2,
-                    const int32_t *step3,
+                    uint32_t mask,
                     uint32_t *counter);
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_perf.c b/src/gallium/drivers/llvmpipe/lp_perf.c
index a316597675..083e7e30a5 100644
--- a/src/gallium/drivers/llvmpipe/lp_perf.c
+++ b/src/gallium/drivers/llvmpipe/lp_perf.c
@@ -46,10 +46,10 @@ lp_print_counters(void)
 {
    if (LP_DEBUG & DEBUG_COUNTERS) {
       unsigned total_64, total_16, total_4;
-      float p1, p2, p3;
+      float p1, p2, p3, p4;
 
-      debug_printf("llvmpipe: nr_triangles:               %9u\n", lp_count.nr_tris);
-      debug_printf("llvmpipe: nr_culled_triangles:        %9u\n", lp_count.nr_culled_tris);
+      debug_printf("llvmpipe: nr_triangles:                 %9u\n", lp_count.nr_tris);
+      debug_printf("llvmpipe: nr_culled_triangles:          %9u\n", lp_count.nr_culled_tris);
 
       total_64 = (lp_count.nr_empty_64 + 
                   lp_count.nr_fully_covered_64 +
@@ -58,10 +58,13 @@ lp_print_counters(void)
       p1 = 100.0 * (float) lp_count.nr_empty_64 / (float) total_64;
       p2 = 100.0 * (float) lp_count.nr_fully_covered_64 / (float) total_64;
       p3 = 100.0 * (float) lp_count.nr_partially_covered_64 / (float) total_64;
+      p4 = 100.0 * (float) lp_count.nr_shade_opaque_64 / (float) total_64;
 
-      debug_printf("llvmpipe: nr_empty_64x64:             %9u (%2.0f%% of %u)\n", lp_count.nr_empty_64, p1, total_64);
-      debug_printf("llvmpipe: nr_fully_covered_64x64:     %9u (%2.0f%% of %u)\n", lp_count.nr_fully_covered_64, p2, total_64);
-      debug_printf("llvmpipe: nr_partially_covered_64x64: %9u (%2.0f%% of %u)\n", lp_count.nr_partially_covered_64, p3, total_64);
+      debug_printf("llvmpipe: nr_64x64:                     %9u\n", total_64);
+      debug_printf("llvmpipe:   nr_fully_covered_64x64:     %9u (%3.0f%% of %u)\n", lp_count.nr_fully_covered_64, p2, total_64);
+      debug_printf("llvmpipe:     nr_shade_opaque_64x64:    %9u (%3.0f%% of %u)\n", lp_count.nr_shade_opaque_64, p4, total_64);
+      debug_printf("llvmpipe:   nr_partially_covered_64x64: %9u (%3.0f%% of %u)\n", lp_count.nr_partially_covered_64, p3, total_64);
+      debug_printf("llvmpipe:   nr_empty_64x64:             %9u (%3.0f%% of %u)\n", lp_count.nr_empty_64, p1, total_64);
 
       total_16 = (lp_count.nr_empty_16 + 
                   lp_count.nr_fully_covered_16 +
@@ -71,25 +74,27 @@ lp_print_counters(void)
       p2 = 100.0 * (float) lp_count.nr_fully_covered_16 / (float) total_16;
       p3 = 100.0 * (float) lp_count.nr_partially_covered_16 / (float) total_16;
 
-      debug_printf("llvmpipe: nr_empty_16x16:             %9u (%2.0f%% of %u)\n", lp_count.nr_empty_16, p1, total_16);
-      debug_printf("llvmpipe: nr_fully_covered_16x16:     %9u (%2.0f%% of %u)\n", lp_count.nr_fully_covered_16, p2, total_16);
-      debug_printf("llvmpipe: nr_partially_covered_16x16: %9u (%2.0f%% of %u)\n", lp_count.nr_partially_covered_16, p3, total_16);
+      debug_printf("llvmpipe: nr_16x16:                     %9u\n", total_16);
+      debug_printf("llvmpipe:   nr_fully_covered_16x16:     %9u (%3.0f%% of %u)\n", lp_count.nr_fully_covered_16, p2, total_16);
+      debug_printf("llvmpipe:   nr_partially_covered_16x16: %9u (%3.0f%% of %u)\n", lp_count.nr_partially_covered_16, p3, total_16);
+      debug_printf("llvmpipe:   nr_empty_16x16:             %9u (%3.0f%% of %u)\n", lp_count.nr_empty_16, p1, total_16);
 
       total_4 = (lp_count.nr_empty_4 + lp_count.nr_non_empty_4);
 
       p1 = 100.0 * (float) lp_count.nr_empty_4 / (float) total_4;
       p2 = 100.0 * (float) lp_count.nr_non_empty_4 / (float) total_4;
 
-      debug_printf("llvmpipe: nr_empty_4x4:               %9u (%2.0f%% of %u)\n", lp_count.nr_empty_4, p1, total_4);
-      debug_printf("llvmpipe: nr_non_empty_4x4:           %9u (%2.0f%% of %u)\n", lp_count.nr_non_empty_4, p2, total_4);
+      debug_printf("llvmpipe: nr_4x4:                       %9u\n", total_4);
+      debug_printf("llvmpipe:   nr_empty_4x4:               %9u (%3.0f%% of %u)\n", lp_count.nr_empty_4, p1, total_4);
+      debug_printf("llvmpipe:   nr_non_empty_4x4:           %9u (%3.0f%% of %u)\n", lp_count.nr_non_empty_4, p2, total_4);
 
-      debug_printf("llvmpipe: nr_color_tile_clear:        %9u\n", lp_count.nr_color_tile_clear);
-      debug_printf("llvmpipe: nr_color_tile_load:         %9u\n", lp_count.nr_color_tile_load);
-      debug_printf("llvmpipe: nr_color_tile_store:        %9u\n", lp_count.nr_color_tile_store);
+      debug_printf("llvmpipe: nr_color_tile_clear:          %9u\n", lp_count.nr_color_tile_clear);
+      debug_printf("llvmpipe: nr_color_tile_load:           %9u\n", lp_count.nr_color_tile_load);
+      debug_printf("llvmpipe: nr_color_tile_store:          %9u\n", lp_count.nr_color_tile_store);
 
-      debug_printf("llvmpipe: nr_llvm_compiles:           %u\n", lp_count.nr_llvm_compiles);
-      debug_printf("llvmpipe: total LLVM compile time:    %.2f sec\n", lp_count.llvm_compile_time / 1000000.0);
-      debug_printf("llvmpipe: average LLVM compile time:  %.2f sec\n", lp_count.llvm_compile_time / 1000000.0 / lp_count.nr_llvm_compiles);
+      debug_printf("llvmpipe: nr_llvm_compiles:             %u\n", lp_count.nr_llvm_compiles);
+      debug_printf("llvmpipe: total LLVM compile time:      %.2f sec\n", lp_count.llvm_compile_time / 1000000.0);
+      debug_printf("llvmpipe: average LLVM compile time:    %.2f sec\n", lp_count.llvm_compile_time / 1000000.0 / lp_count.nr_llvm_compiles);
 
    }
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_perf.h b/src/gallium/drivers/llvmpipe/lp_perf.h
index a9629dae3c..4774f64550 100644
--- a/src/gallium/drivers/llvmpipe/lp_perf.h
+++ b/src/gallium/drivers/llvmpipe/lp_perf.h
@@ -44,6 +44,7 @@ struct lp_counters
    unsigned nr_empty_64;
    unsigned nr_fully_covered_64;
    unsigned nr_partially_covered_64;
+   unsigned nr_shade_opaque_64;
    unsigned nr_empty_16;
    unsigned nr_fully_covered_16;
    unsigned nr_partially_covered_16;
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index 1dde327836..0130e39fd8 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -28,6 +28,7 @@
 #include <limits.h>
 #include "util/u_memory.h"
 #include "util/u_math.h"
+#include "util/u_rect.h"
 #include "util/u_surface.h"
 
 #include "lp_scene_queue.h"
@@ -136,7 +137,6 @@ lp_rast_tile_begin(struct lp_rasterizer_task *task,
    struct lp_rasterizer *rast = task->rast;
    struct lp_scene *scene = rast->curr_scene;
    enum lp_texture_usage usage;
-   unsigned buf;
 
    LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
 
@@ -146,24 +146,8 @@ lp_rast_tile_begin(struct lp_rasterizer_task *task,
    task->x = x;
    task->y = y;
 
-   if (scene->has_color_clear)
-      usage = LP_TEX_USAGE_WRITE_ALL;
-   else
-      usage = LP_TEX_USAGE_READ_WRITE;
-
-   /* get pointers to color tile(s) */
-   for (buf = 0; buf < rast->state.nr_cbufs; buf++) {
-      struct pipe_surface *cbuf = rast->curr_scene->fb.cbufs[buf];
-      struct llvmpipe_resource *lpt;
-      assert(cbuf);
-      lpt = llvmpipe_resource(cbuf->texture);
-      task->color_tiles[buf] = llvmpipe_get_texture_tile(lpt,
-                                                         cbuf->face + cbuf->zslice,
-                                                         cbuf->level,
-                                                         usage,
-                                                         x, y);
-      assert(task->color_tiles[buf]);
-   }
+   /* reset pointers to color tile(s) */
+   memset(task->color_tiles, 0, sizeof(task->color_tiles));
 
    /* get pointer to depth/stencil tile */
    {
@@ -222,7 +206,8 @@ lp_rast_clear_color(struct lp_rasterizer_task *task,
        clear_color[2] == clear_color[3]) {
       /* clear to grayscale value {x, x, x, x} */
       for (i = 0; i < rast->state.nr_cbufs; i++) {
-         uint8_t *ptr = task->color_tiles[i];
+         uint8_t *ptr =
+            lp_rast_get_color_tile_pointer(task, i, LP_TEX_USAGE_WRITE_ALL);
 	 memset(ptr, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
       }
    }
@@ -234,7 +219,8 @@ lp_rast_clear_color(struct lp_rasterizer_task *task,
        */
       const unsigned chunk = TILE_SIZE / 4;
       for (i = 0; i < rast->state.nr_cbufs; i++) {
-         uint8_t *c = task->color_tiles[i];
+         uint8_t *c =
+            lp_rast_get_color_tile_pointer(task, i, LP_TEX_USAGE_WRITE_ALL);
          unsigned j;
 
          for (j = 0; j < 4 * TILE_SIZE; j++) {
@@ -378,8 +364,8 @@ lp_rast_load_color(struct lp_rasterizer_task *task,
  * This is a bin command which is stored in all bins.
  */
 void
-lp_rast_store_color( struct lp_rasterizer_task *task,
-                     const union lp_rast_cmd_arg arg)
+lp_rast_store_linear_color( struct lp_rasterizer_task *task,
+                            const union lp_rast_cmd_arg arg)
 {
    struct lp_rasterizer *rast = task->rast;
    struct lp_scene *scene = rast->curr_scene;
@@ -448,30 +434,54 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task,
 
          /* run shader on 4x4 block */
          variant->jit_function[RAST_WHOLE]( &state->jit_context,
-                                          tile_x + x, tile_y + y,
-                                          inputs->facing,
-                                          inputs->a0,
-                                          inputs->dadx,
-                                          inputs->dady,
-                                          color,
-                                          depth,
-                                          INT_MIN, INT_MIN, INT_MIN,
-                                          NULL, NULL, NULL, &task->vis_counter);
+                                            tile_x + x, tile_y + y,
+                                            inputs->facing,
+                                            inputs->a0,
+                                            inputs->dadx,
+                                            inputs->dady,
+                                            color,
+                                            depth,
+                                            0xffff,
+                                            &task->vis_counter);
       }
    }
 }
 
 
 /**
- * Compute shading for a 4x4 block of pixels.
+ * Run the shader on all blocks in a tile.  This is used when a tile is
+ * completely contained inside a triangle, and the shader is opaque.
+ * This is a bin command called during bin processing.
+ */
+void
+lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
+                          const union lp_rast_cmd_arg arg)
+{
+   struct lp_rasterizer *rast = task->rast;
+   unsigned i;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   /* this will prevent converting the layout from tiled to linear */
+   for (i = 0; i < rast->state.nr_cbufs; i++) {
+      (void)lp_rast_get_color_tile_pointer(task, i, LP_TEX_USAGE_WRITE_ALL);
+   }
+
+   lp_rast_shade_tile(task, arg);
+}
+
+
+/**
+ * Compute shading for a 4x4 block of pixels inside a triangle.
  * This is a bin command called during bin processing.
  * \param x  X position of quad in window coords
  * \param y  Y position of quad in window coords
  */
-void lp_rast_shade_quads( struct lp_rasterizer_task *task,
-                          const struct lp_rast_shader_inputs *inputs,
-                          unsigned x, unsigned y,
-                          int32_t c1, int32_t c2, int32_t c3)
+void
+lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
+                         const struct lp_rast_shader_inputs *inputs,
+                         unsigned x, unsigned y,
+                         unsigned mask)
 {
    const struct lp_rast_state *state = task->current_state;
    struct lp_fragment_shader_variant *variant = state->variant;
@@ -501,27 +511,21 @@ void lp_rast_shade_quads( struct lp_rasterizer_task *task,
 
    assert(lp_check_alignment(state->jit_context.blend_color, 16));
 
-   assert(lp_check_alignment(inputs->step[0], 16));
-   assert(lp_check_alignment(inputs->step[1], 16));
-   assert(lp_check_alignment(inputs->step[2], 16));
-
    /* run shader on 4x4 block */
-   variant->jit_function[RAST_EDGE_TEST]( &state->jit_context,
-                                        x, y,
-                                        inputs->facing,
-                                        inputs->a0,
-                                        inputs->dadx,
-                                        inputs->dady,
-                                        color,
-                                        depth,
-                                        c1, c2, c3,
-                                        inputs->step[0],
-                                        inputs->step[1],
-                                        inputs->step[2],
-					&task->vis_counter);
+   variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
+                                         x, y,
+                                         inputs->facing,
+                                         inputs->a0,
+                                         inputs->dadx,
+                                         inputs->dady,
+                                         color,
+                                         depth,
+                                         mask,
+                                         &task->vis_counter);
 }
 
 
+
 /**
  * Set top row and left column of the tile's pixels to white.  For debugging.
  */
@@ -717,10 +721,17 @@ static struct {
 {
    RAST(clear_color),
    RAST(clear_zstencil),
-   RAST(triangle),
+   RAST(triangle_1),
+   RAST(triangle_2),
+   RAST(triangle_3),
+   RAST(triangle_4),
+   RAST(triangle_5),
+   RAST(triangle_6),
+   RAST(triangle_7),
    RAST(shade_tile),
+   RAST(shade_tile_opaque),
    RAST(set_state),
-   RAST(store_color),
+   RAST(store_linear_color),
    RAST(fence),
    RAST(begin_query),
    RAST(end_query),
@@ -775,7 +786,8 @@ is_empty_bin( const struct cmd_bin *bin )
    }
 
    for (i = 0; i < head->count; i++)
-      if (head->cmd[i] != lp_rast_set_state) {
+      if (head->cmd[i] != lp_rast_set_state &&
+          head->cmd[i] != lp_rast_store_linear_color) {
          return FALSE;
       }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.h b/src/gallium/drivers/llvmpipe/lp_rast.h
index 80ca68f5a2..ae73e6d8c9 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast.h
@@ -83,9 +83,6 @@ struct lp_rast_shader_inputs {
    float (*a0)[4];
    float (*dadx)[4];
    float (*dady)[4];
-
-   /* edge/step info for 3 edges and 4x4 block of pixels */
-   PIPE_ALIGN_VAR(16) int step[3][16];
 };
 
 struct lp_rast_clearzs {
@@ -93,6 +90,22 @@ struct lp_rast_clearzs {
    unsigned clearzs_mask;
 };
 
+struct lp_rast_plane {
+   /* one-pixel sized trivial accept offsets for each plane */
+   int ei;
+
+   /* one-pixel sized trivial reject offsets for each plane */
+   int eo;
+
+   /* edge function values at minx,miny ?? */
+   int c;
+
+   int dcdx;
+   int dcdy;
+   
+   /* edge/step info for 3 edges and 4x4 block of pixels */
+   const int *step;
+};
 
 /**
  * Rasterization information for a triangle known to be in this bin,
@@ -101,35 +114,16 @@ struct lp_rast_clearzs {
  * Objects of this type are put into the lp_setup_context::data buffer.
  */
 struct lp_rast_triangle {
+   /* inputs for the shader */
+   PIPE_ALIGN_VAR(16) struct lp_rast_shader_inputs inputs;
+
+   int step[3][16];
+
 #ifdef DEBUG
    float v[3][2];
 #endif
 
-   /* one-pixel sized trivial accept offsets for each plane */
-   int ei1;                   
-   int ei2;
-   int ei3;
-
-   /* one-pixel sized trivial reject offsets for each plane */
-   int eo1;                   
-   int eo2;
-   int eo3;
-
-   /* y deltas for vertex pairs (in fixed pt) */
-   int dy12;
-   int dy23;
-   int dy31;
-
-   /* x deltas for vertex pairs (in fixed pt) */
-   int dx12;
-   int dx23;
-   int dx31;
-
-   /* edge function values at minx,miny ?? */
-   int c1, c2, c3;
-
-   /* inputs for the shader */
-   PIPE_ALIGN_VAR(16) struct lp_rast_shader_inputs inputs;
+   struct lp_rast_plane plane[7]; /* NOTE: may allocate fewer planes */
 };
 
 
@@ -153,7 +147,10 @@ lp_rast_finish( struct lp_rasterizer *rast );
 
 union lp_rast_cmd_arg {
    const struct lp_rast_shader_inputs *shade_tile;
-   const struct lp_rast_triangle *triangle;
+   struct {
+      const struct lp_rast_triangle *tri;
+      unsigned plane_mask;
+   } triangle;
    const struct lp_rast_state *set_state;
    uint8_t clear_color[4];
    const struct lp_rast_clearzs *clear_zstencil;
@@ -173,10 +170,12 @@ lp_rast_arg_inputs( const struct lp_rast_shader_inputs *shade_tile )
 }
 
 static INLINE union lp_rast_cmd_arg
-lp_rast_arg_triangle( const struct lp_rast_triangle *triangle )
+lp_rast_arg_triangle( const struct lp_rast_triangle *triangle,
+                      unsigned plane_mask)
 {
    union lp_rast_cmd_arg arg;
-   arg.triangle = triangle;
+   arg.triangle.tri = triangle;
+   arg.triangle.plane_mask = plane_mask;
    return arg;
 }
 
@@ -229,16 +228,31 @@ void lp_rast_clear_zstencil( struct lp_rasterizer_task *,
 void lp_rast_set_state( struct lp_rasterizer_task *, 
                         const union lp_rast_cmd_arg );
 
-void lp_rast_triangle( struct lp_rasterizer_task *, 
-                       const union lp_rast_cmd_arg );
+void lp_rast_triangle_1( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
+void lp_rast_triangle_2( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
+void lp_rast_triangle_3( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
+void lp_rast_triangle_4( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
+void lp_rast_triangle_5( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
+void lp_rast_triangle_6( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
+void lp_rast_triangle_7( struct lp_rasterizer_task *, 
+                         const union lp_rast_cmd_arg );
 
 void lp_rast_shade_tile( struct lp_rasterizer_task *,
                          const union lp_rast_cmd_arg );
 
+void lp_rast_shade_tile_opaque( struct lp_rasterizer_task *,
+                                const union lp_rast_cmd_arg );
+
 void lp_rast_fence( struct lp_rasterizer_task *,
                     const union lp_rast_cmd_arg );
 
-void lp_rast_store_color( struct lp_rasterizer_task *,
+void lp_rast_store_linear_color( struct lp_rasterizer_task *,
                           const union lp_rast_cmd_arg );
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_priv.h b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
index eb4175dfa6..024a28be59 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_priv.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
@@ -119,10 +119,12 @@ struct lp_rasterizer
 };
 
 
-void lp_rast_shade_quads( struct lp_rasterizer_task *task,
-                          const struct lp_rast_shader_inputs *inputs,
-                          unsigned x, unsigned y,
-                          int32_t c1, int32_t c2, int32_t c3);
+void
+lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
+                         const struct lp_rast_shader_inputs *inputs,
+                         unsigned x, unsigned y,
+                         unsigned mask);
+
 
 
 /**
@@ -158,6 +160,40 @@ lp_rast_get_depth_block_pointer(struct lp_rasterizer_task *task,
 
 
 /**
+ * Get pointer to the swizzled color tile
+ */
+static INLINE uint8_t *
+lp_rast_get_color_tile_pointer(struct lp_rasterizer_task *task,
+                               unsigned buf, enum lp_texture_usage usage)
+{
+   struct lp_rasterizer *rast = task->rast;
+
+   assert(task->x % TILE_SIZE == 0);
+   assert(task->y % TILE_SIZE == 0);
+   assert(buf < rast->state.nr_cbufs);
+
+   if (!task->color_tiles[buf]) {
+      struct pipe_surface *cbuf = rast->curr_scene->fb.cbufs[buf];
+      struct llvmpipe_resource *lpt;
+      assert(cbuf);
+      lpt = llvmpipe_resource(cbuf->texture);
+      task->color_tiles[buf] = llvmpipe_get_texture_tile(lpt,
+                                                         cbuf->face + cbuf->zslice,
+                                                         cbuf->level,
+                                                         usage,
+                                                         task->x,
+                                                         task->y);
+      if (!task->color_tiles[buf]) {
+         /* out of memory - use dummy tile memory */
+         return lp_get_dummy_tile();
+      }
+   }
+
+   return task->color_tiles[buf];
+}
+
+
+/**
  * Get the pointer to a 4x4 color block (within a 64x64 tile).
  * We'll map the color buffer on demand here.
  * Note that this may be called even when there's no color buffers - return
@@ -174,6 +210,7 @@ lp_rast_get_color_block_pointer(struct lp_rasterizer_task *task,
    assert((x % TILE_VECTOR_WIDTH) == 0);
    assert((y % TILE_VECTOR_HEIGHT) == 0);
 
+   color = lp_rast_get_color_tile_pointer(task, buf, LP_TEX_USAGE_READ_WRITE);
    color = task->color_tiles[buf];
    if (!color) {
       /* out of memory - use dummy tile memory */
@@ -217,15 +254,15 @@ lp_rast_shade_quads_all( struct lp_rasterizer_task *task,
 
    /* run shader on 4x4 block */
    variant->jit_function[RAST_WHOLE]( &state->jit_context,
-                                    x, y,
-                                    inputs->facing,
-                                    inputs->a0,
-                                    inputs->dadx,
-                                    inputs->dady,
-                                    color,
-                                    depth,
-                                    INT_MIN, INT_MIN, INT_MIN,
-                                    NULL, NULL, NULL, &task->vis_counter );
+                                      x, y,
+                                      inputs->facing,
+                                      inputs->a0,
+                                      inputs->dadx,
+                                      inputs->dady,
+                                      color,
+                                      depth,
+                                      0xffff,
+                                      &task->vis_counter );
 }
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index a5f0d14c95..ebe9a8e92b 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -113,168 +113,31 @@ block_full_16(struct lp_rasterizer_task *task,
 	 block_full_4(task, tri, x + ix, y + iy);
 }
 
+#define TAG(x) x##_1
+#define NR_PLANES 1
+#include "lp_rast_tri_tmp.h"
 
-/**
- * Pass the 4x4 pixel block to the shader function.
- * Determination of which of the 16 pixels lies inside the triangle
- * will be done as part of the fragment shader.
- */
-static void
-do_block_4(struct lp_rasterizer_task *task,
-           const struct lp_rast_triangle *tri,
-           int x, int y,
-           int c1, int c2, int c3)
-{
-   assert(x >= 0);
-   assert(y >= 0);
-
-   lp_rast_shade_quads(task, &tri->inputs, x, y, -c1, -c2, -c3);
-}
-
-
-/**
- * Evaluate a 16x16 block of pixels to determine which 4x4 subblocks are in/out
- * of the triangle's bounds.
- */
-static void
-do_block_16(struct lp_rasterizer_task *task,
-            const struct lp_rast_triangle *tri,
-            int x, int y,
-            int c0, int c1, int c2)
-{
-   unsigned mask = 0;
-   int eo[3];
-   int c[3];
-   int i, j;
-
-   assert(x >= 0);
-   assert(y >= 0);
-   assert(x % 16 == 0);
-   assert(y % 16 == 0);
-
-   eo[0] = tri->eo1 * 4;
-   eo[1] = tri->eo2 * 4;
-   eo[2] = tri->eo3 * 4;
-
-   c[0] = c0;
-   c[1] = c1;
-   c[2] = c2;
-
-   for (j = 0; j < 3; j++) {
-      const int *step = tri->inputs.step[j];
-      const int cx = c[j] + eo[j];
-
-      /* Mask has bits set whenever we are outside any of the edges.
-       */
-      for (i = 0; i < 16; i++) {
-         int out = cx + step[i] * 4;
-         mask |= (out >> 31) & (1 << i);
-      }
-   }
+#define TAG(x) x##_2
+#define NR_PLANES 2
+#include "lp_rast_tri_tmp.h"
 
-   mask = ~mask & 0xffff;
-   while (mask) {
-      int i = ffs(mask) - 1;
-      int px = x + pos_table4[i][0];
-      int py = y + pos_table4[i][1];
-      int cx1 = c0 + tri->inputs.step[0][i] * 4;
-      int cx2 = c1 + tri->inputs.step[1][i] * 4;
-      int cx3 = c2 + tri->inputs.step[2][i] * 4;
+#define TAG(x) x##_3
+#define NR_PLANES 3
+#include "lp_rast_tri_tmp.h"
 
-      mask &= ~(1 << i);
+#define TAG(x) x##_4
+#define NR_PLANES 4
+#include "lp_rast_tri_tmp.h"
 
-      /* Don't bother testing if the 4x4 block is entirely in/out of
-       * the triangle.  It's a little faster to do it in the jit code.
-       */
-      LP_COUNT(nr_non_empty_4);
-      do_block_4(task, tri, px, py, cx1, cx2, cx3);
-   }
-}
-
-
-/**
- * Scan the tile in chunks and figure out which pixels to rasterize
- * for this triangle.
- */
-void
-lp_rast_triangle(struct lp_rasterizer_task *task,
-                 const union lp_rast_cmd_arg arg)
-{
-   const struct lp_rast_triangle *tri = arg.triangle;
-   const int x = task->x, y = task->y;
-   int ei[3], eo[3], c[3];
-   unsigned outmask, inmask, partial_mask;
-   unsigned i, j;
-
-   c[0] = tri->c1 + tri->dx12 * y - tri->dy12 * x;
-   c[1] = tri->c2 + tri->dx23 * y - tri->dy23 * x;
-   c[2] = tri->c3 + tri->dx31 * y - tri->dy31 * x;
-
-   eo[0] = tri->eo1 * 16;
-   eo[1] = tri->eo2 * 16;
-   eo[2] = tri->eo3 * 16;
-
-   ei[0] = tri->ei1 * 16;
-   ei[1] = tri->ei2 * 16;
-   ei[2] = tri->ei3 * 16;
-
-   outmask = 0;
-   inmask = 0xffff;
+#define TAG(x) x##_5
+#define NR_PLANES 5
+#include "lp_rast_tri_tmp.h"
 
-   for (j = 0; j < 3; j++) {
-      const int *step = tri->inputs.step[j];
-      const int cox = c[j] + eo[j];
-      const int cio = ei[j]- eo[j];
+#define TAG(x) x##_6
+#define NR_PLANES 6
+#include "lp_rast_tri_tmp.h"
 
-      /* Outmask has bits set whenever we are outside any of the
-       * edges.
-       */
-      /* Inmask has bits set whenever we are inside all of the edges.
-       */
-      for (i = 0; i < 16; i++) {
-         int out = cox + step[i] * 16;
-         int in = out + cio;
-         outmask |= (out >> 31) & (1 << i);
-         inmask &= ~((in >> 31) & (1 << i));
-      }
-   }
+#define TAG(x) x##_7
+#define NR_PLANES 7
+#include "lp_rast_tri_tmp.h"
 
-   assert((outmask & inmask) == 0);
-
-   if (outmask == 0xffff)
-      return;
-
-   /* Invert mask, so that bits are set whenever we are at least
-    * partially inside all of the edges:
-    */
-   partial_mask = ~inmask & ~outmask & 0xffff;
-
-   /* Iterate over partials:
-    */
-   while (partial_mask) {
-      int i = ffs(partial_mask) - 1;
-      int px = x + pos_table16[i][0];
-      int py = y + pos_table16[i][1];
-      int cx1 = c[0] + tri->inputs.step[0][i] * 16;
-      int cx2 = c[1] + tri->inputs.step[1][i] * 16;
-      int cx3 = c[2] + tri->inputs.step[2][i] * 16;
-
-      partial_mask &= ~(1 << i);
-
-      LP_COUNT(nr_partially_covered_16);
-      do_block_16(task, tri, px, py, cx1, cx2, cx3);
-   }
-
-   /* Iterate over fulls: 
-    */
-   while (inmask) {
-      int i = ffs(inmask) - 1;
-      int px = x + pos_table16[i][0];
-      int py = y + pos_table16[i][1];
-
-      inmask &= ~(1 << i);
-
-      LP_COUNT(nr_fully_covered_16);
-      block_full_16(task, tri, px, py);
-   }
-}
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
new file mode 100644
index 0000000000..a410c611a3
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
@@ -0,0 +1,238 @@
+/**************************************************************************
+ *
+ * Copyright 2007-2010 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+/*
+ * Rasterization for binned triangles within a tile
+ */
+
+
+
+/**
+ * Prototype for a 7 plane rasterizer function.  Will codegenerate
+ * several of these.
+ *
+ * XXX: Varients for more/fewer planes.
+ * XXX: Need ways of dropping planes as we descend.
+ * XXX: SIMD
+ */
+static void
+TAG(do_block_4)(struct lp_rasterizer_task *task,
+                const struct lp_rast_triangle *tri,
+                const struct lp_rast_plane *plane,
+                int x, int y,
+                const int *c)
+{
+   unsigned mask = 0;
+   int i;
+
+   for (i = 0; i < 16; i++) {
+      int any_negative = 0;
+      int j;
+
+      for (j = 0; j < NR_PLANES; j++) 
+         any_negative |= (c[j] - 1 + plane[j].step[i]);
+         
+      any_negative >>= 31;
+
+      mask |= (~any_negative) & (1 << i);
+   }
+
+   /* Now pass to the shader:
+    */
+   if (mask)
+      lp_rast_shade_quads_mask(task, &tri->inputs, x, y, mask);
+}
+
+/**
+ * Evaluate a 16x16 block of pixels to determine which 4x4 subblocks are in/out
+ * of the triangle's bounds.
+ */
+static void
+TAG(do_block_16)(struct lp_rasterizer_task *task,
+                 const struct lp_rast_triangle *tri,
+                 const struct lp_rast_plane *plane,
+                 int x, int y,
+                 const int *c)
+{
+   unsigned outmask, inmask, partmask, partial_mask;
+   unsigned i, j;
+
+   outmask = 0;                 /* outside one or more trivial reject planes */
+   partmask = 0;                /* outside one or more trivial accept planes */
+
+   for (j = 0; j < NR_PLANES; j++) {
+      const int *step = plane[j].step;
+      const int eo = plane[j].eo * 4;
+      const int ei = plane[j].ei * 4;
+      const int cox = c[j] + eo;
+      const int cio = ei - 1 - eo;
+
+      for (i = 0; i < 16; i++) {
+         int out = cox + step[i] * 4;
+         int part = out + cio;
+         outmask  |= (out >> 31) & (1 << i);
+         partmask |= (part >> 31) & (1 << i);
+      }
+   }
+
+   if (outmask == 0xffff)
+      return;
+
+   /* Mask of sub-blocks which are inside all trivial accept planes:
+    */
+   inmask = ~partmask & 0xffff;
+
+   /* Mask of sub-blocks which are inside all trivial reject planes,
+    * but outside at least one trivial accept plane:
+    */
+   partial_mask = partmask & ~outmask;
+
+   assert((partial_mask & inmask) == 0);
+
+   /* Iterate over partials:
+    */
+   while (partial_mask) {
+      int i = ffs(partial_mask) - 1;
+      int px = x + pos_table4[i][0];
+      int py = y + pos_table4[i][1];
+      int cx[NR_PLANES];
+
+      for (j = 0; j < NR_PLANES; j++)
+         cx[j] = c[j] + plane[j].step[i] * 4;
+
+      partial_mask &= ~(1 << i);
+
+      TAG(do_block_4)(task, tri, plane, px, py, cx);
+   }
+
+   /* Iterate over fulls: 
+    */
+   while (inmask) {
+      int i = ffs(inmask) - 1;
+      int px = x + pos_table4[i][0];
+      int py = y + pos_table4[i][1];
+
+      inmask &= ~(1 << i);
+
+      block_full_4(task, tri, px, py);
+   }
+}
+
+
+/**
+ * Scan the tile in chunks and figure out which pixels to rasterize
+ * for this triangle.
+ */
+void
+TAG(lp_rast_triangle)(struct lp_rasterizer_task *task,
+                      const union lp_rast_cmd_arg arg)
+{
+   const struct lp_rast_triangle *tri = arg.triangle.tri;
+   unsigned plane_mask = arg.triangle.plane_mask;
+   const int x = task->x, y = task->y;
+   struct lp_rast_plane plane[NR_PLANES];
+   int c[NR_PLANES];
+   unsigned outmask, inmask, partmask, partial_mask;
+   unsigned i, j, nr_planes = 0;
+
+   while (plane_mask) {
+      int i = ffs(plane_mask) - 1;
+      plane[nr_planes] = tri->plane[i];
+      plane_mask &= ~(1 << i);
+      nr_planes++;
+   };
+
+   assert(nr_planes == NR_PLANES);
+   outmask = 0;                 /* outside one or more trivial reject planes */
+   partmask = 0;                /* outside one or more trivial accept planes */
+
+   for (j = 0; j < NR_PLANES; j++) {
+      const int *step = plane[j].step;
+      const int eo = plane[j].eo * 16;
+      const int ei = plane[j].ei * 16;
+      int cox, cio;
+
+      c[j] = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
+      cox = c[j] + eo;
+      cio = ei - 1 - eo;
+
+      for (i = 0; i < 16; i++) {
+         int out = cox + step[i] * 16;
+         int part = out + cio;
+         outmask  |= (out >> 31) & (1 << i);
+         partmask |= (part >> 31) & (1 << i);
+      }
+   }
+
+   if (outmask == 0xffff)
+      return;
+
+   /* Mask of sub-blocks which are inside all trivial accept planes:
+    */
+   inmask = ~partmask & 0xffff;
+
+   /* Mask of sub-blocks which are inside all trivial reject planes,
+    * but outside at least one trivial accept plane:
+    */
+   partial_mask = partmask & ~outmask;
+
+   assert((partial_mask & inmask) == 0);
+
+   /* Iterate over partials:
+    */
+   while (partial_mask) {
+      int i = ffs(partial_mask) - 1;
+      int px = x + pos_table16[i][0];
+      int py = y + pos_table16[i][1];
+      int cx[NR_PLANES];
+
+      for (j = 0; j < NR_PLANES; j++)
+         cx[j] = c[j] + plane[j].step[i] * 16;
+
+      partial_mask &= ~(1 << i);
+
+      LP_COUNT(nr_partially_covered_16);
+      TAG(do_block_16)(task, tri, plane, px, py, cx);
+   }
+
+   /* Iterate over fulls: 
+    */
+   while (inmask) {
+      int i = ffs(inmask) - 1;
+      int px = x + pos_table16[i][0];
+      int py = y + pos_table16[i][1];
+
+      inmask &= ~(1 << i);
+
+      LP_COUNT(nr_fully_covered_16);
+      block_full_16(task, tri, px, py);
+   }
+}
+
+#undef TAG
+#undef NR_PLANES
+
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c
index 3b83f4e742..40959e6208 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup.c
@@ -287,7 +287,7 @@ lp_setup_flush( struct lp_setup_context *setup,
           * data to linear in the texture_unmap() function, which will
           * not be a parallel/threaded operation as here.
           */
-         lp_scene_bin_everywhere(scene, lp_rast_store_color, dummy);
+         lp_scene_bin_everywhere(scene, lp_rast_store_linear_color, dummy);
       }
 
 
@@ -752,28 +752,6 @@ lp_setup_update_state( struct lp_setup_context *setup )
       setup->dirty |= LP_SETUP_NEW_FS;
    }
 
-   if (setup->dirty & LP_SETUP_NEW_SCISSOR) {
-      float *stored;
-
-      stored = lp_scene_alloc_aligned(scene, 4 * sizeof(int32_t), 16);
-
-      if (stored) {
-         stored[0] = (float) setup->scissor.current.minx;
-         stored[1] = (float) setup->scissor.current.miny;
-         stored[2] = (float) setup->scissor.current.maxx;
-         stored[3] = (float) setup->scissor.current.maxy;
-
-         setup->scissor.stored = stored;
-
-         setup->fs.current.jit_context.scissor_xmin = stored[0];
-         setup->fs.current.jit_context.scissor_ymin = stored[1];
-         setup->fs.current.jit_context.scissor_xmax = stored[2];
-         setup->fs.current.jit_context.scissor_ymax = stored[3];
-      }
-
-      setup->dirty |= LP_SETUP_NEW_FS;
-   }
-
    if(setup->dirty & LP_SETUP_NEW_CONSTANTS) {
       struct pipe_resource *buffer = setup->constants.current;
 
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_context.h b/src/gallium/drivers/llvmpipe/lp_setup_context.h
index 8f4e00f073..0cea7791f5 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_setup_context.h
@@ -130,7 +130,6 @@ struct lp_setup_context
 
    struct {
       struct pipe_scissor_state current;
-      const void *stored;
    } scissor;
 
    unsigned dirty;   /**< bitmask of LP_SETUP_NEW_x bits */
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_tri.c b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
index 4e2e17f77b..036b5497fa 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
@@ -38,12 +38,78 @@
 
 #define NUM_CHANNELS 4
 
+struct tri_info {
+
+   float pixel_offset;
+
+   /* fixed point vertex coordinates */
+   int x[3];
+   int y[3];
+
+   /* float x,y deltas - all from the original coordinates
+    */
+   float dy01, dy20;
+   float dx01, dx20;
+   float oneoverarea;
+
+   const float (*v0)[4];
+   const float (*v1)[4];
+   const float (*v2)[4];
+
+   boolean frontfacing;
+};
+
+
+
+static const int step_scissor_minx[16] = {
+   0, 1, 0, 1,
+   2, 3, 2, 3,
+   0, 1, 0, 1,
+   2, 3, 2, 3
+};
+
+static const int step_scissor_maxx[16] = {
+    0, -1,  0, -1,
+   -2, -3, -2, -3,
+    0, -1,  0, -1,
+   -2, -3, -2, -3
+};
+
+static const int step_scissor_miny[16] = {
+   0, 0, 1, 1,
+   0, 0, 1, 1,
+   2, 2, 3, 3,
+   2, 2, 3, 3
+};
+
+static const int step_scissor_maxy[16] = {
+    0,  0, -1, -1,
+    0,  0, -1, -1,
+   -2, -2, -3, -3,
+   -2, -2, -3, -3
+};
+
+
+
+   
+static INLINE int
+subpixel_snap(float a)
+{
+   return util_iround(FIXED_ONE * a);
+}
+
+static INLINE float
+fixed_to_float(int a)
+{
+   return a * (1.0 / FIXED_ONE);
+}
+
+
 
 /**
  * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
  */
-static void constant_coef( struct lp_setup_context *setup,
-                           struct lp_rast_triangle *tri,
+static void constant_coef( struct lp_rast_triangle *tri,
                            unsigned slot,
 			   const float value,
                            unsigned i )
@@ -54,28 +120,21 @@ static void constant_coef( struct lp_setup_context *setup,
 }
 
 
-/**
- * Compute a0, dadx and dady for a linearly interpolated coefficient,
- * for a triangle.
- */
-static void linear_coef( struct lp_setup_context *setup,
-                         struct lp_rast_triangle *tri,
-                         float oneoverarea,
+
+static void linear_coef( struct lp_rast_triangle *tri,
+                         const struct tri_info *info,
                          unsigned slot,
-                         const float (*v1)[4],
-                         const float (*v2)[4],
-                         const float (*v3)[4],
                          unsigned vert_attr,
                          unsigned i)
 {
-   float a1 = v1[vert_attr][i];
-   float a2 = v2[vert_attr][i];
-   float a3 = v3[vert_attr][i];
+   float a0 = info->v0[vert_attr][i];
+   float a1 = info->v1[vert_attr][i];
+   float a2 = info->v2[vert_attr][i];
 
-   float da12 = a1 - a2;
-   float da31 = a3 - a1;
-   float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * oneoverarea;
-   float dady = (da31 * tri->dx12 - tri->dx31 * da12) * oneoverarea;
+   float da01 = a0 - a1;
+   float da20 = a2 - a0;
+   float dadx = (da01 * info->dy20 - info->dy01 * da20) * info->oneoverarea;
+   float dady = (da20 * info->dx01 - info->dx20 * da01) * info->oneoverarea;
 
    tri->inputs.dadx[slot][i] = dadx;
    tri->inputs.dady[slot][i] = dady;
@@ -92,9 +151,9 @@ static void linear_coef( struct lp_setup_context *setup,
     * to define a0 as the sample at a pixel center somewhere near vmin
     * instead - i'll switch to this later.
     */
-   tri->inputs.a0[slot][i] = (a1 -
-                              (dadx * (v1[0][0] - setup->pixel_offset) +
-                               dady * (v1[0][1] - setup->pixel_offset)));
+   tri->inputs.a0[slot][i] = (a0 -
+                              (dadx * (info->v0[0][0] - info->pixel_offset) +
+                               dady * (info->v0[0][1] - info->pixel_offset)));
 }
 
 
@@ -106,31 +165,27 @@ static void linear_coef( struct lp_setup_context *setup,
  * Later, when we compute the value at a particular fragment position we'll
  * divide the interpolated value by the interpolated W at that fragment.
  */
-static void perspective_coef( struct lp_setup_context *setup,
-                              struct lp_rast_triangle *tri,
-                              float oneoverarea,
+static void perspective_coef( struct lp_rast_triangle *tri,
+                              const struct tri_info *info,
                               unsigned slot,
-			      const float (*v1)[4],
-			      const float (*v2)[4],
-			      const float (*v3)[4],
 			      unsigned vert_attr,
                               unsigned i)
 {
    /* premultiply by 1/w  (v[0][3] is always 1/w):
     */
-   float a1 = v1[vert_attr][i] * v1[0][3];
-   float a2 = v2[vert_attr][i] * v2[0][3];
-   float a3 = v3[vert_attr][i] * v3[0][3];
-   float da12 = a1 - a2;
-   float da31 = a3 - a1;
-   float dadx = (da12 * tri->dy31 - tri->dy12 * da31) * oneoverarea;
-   float dady = (da31 * tri->dx12 - tri->dx31 * da12) * oneoverarea;
+   float a0 = info->v0[vert_attr][i] * info->v0[0][3];
+   float a1 = info->v1[vert_attr][i] * info->v1[0][3];
+   float a2 = info->v2[vert_attr][i] * info->v2[0][3];
+   float da01 = a0 - a1;
+   float da20 = a2 - a0;
+   float dadx = (da01 * info->dy20 - info->dy01 * da20) * info->oneoverarea;
+   float dady = (da20 * info->dx01 - info->dx20 * da01) * info->oneoverarea;
 
    tri->inputs.dadx[slot][i] = dadx;
    tri->inputs.dady[slot][i] = dady;
-   tri->inputs.a0[slot][i] = (a1 -
-                              (dadx * (v1[0][0] - setup->pixel_offset) +
-                               dady * (v1[0][1] - setup->pixel_offset)));
+   tri->inputs.a0[slot][i] = (a0 -
+                              (dadx * (info->v0[0][0] - info->pixel_offset) +
+                               dady * (info->v0[0][1] - info->pixel_offset)));
 }
 
 
@@ -141,13 +196,9 @@ static void perspective_coef( struct lp_setup_context *setup,
  * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
  */
 static void
-setup_fragcoord_coef(struct lp_setup_context *setup,
-                     struct lp_rast_triangle *tri,
-                     float oneoverarea,
+setup_fragcoord_coef(struct lp_rast_triangle *tri,
+                     const struct tri_info *info,
                      unsigned slot,
-                     const float (*v1)[4],
-                     const float (*v2)[4],
-                     const float (*v3)[4],
                      unsigned usage_mask)
 {
    /*X*/
@@ -166,12 +217,12 @@ setup_fragcoord_coef(struct lp_setup_context *setup,
 
    /*Z*/
    if (usage_mask & TGSI_WRITEMASK_Z) {
-      linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 2);
+      linear_coef(tri, info, slot, 0, 2);
    }
 
    /*W*/
    if (usage_mask & TGSI_WRITEMASK_W) {
-      linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 3);
+      linear_coef(tri, info, slot, 0, 3);
    }
 }
 
@@ -180,24 +231,23 @@ setup_fragcoord_coef(struct lp_setup_context *setup,
  * Setup the fragment input attribute with the front-facing value.
  * \param frontface  is the triangle front facing?
  */
-static void setup_facing_coef( struct lp_setup_context *setup,
-                               struct lp_rast_triangle *tri,
+static void setup_facing_coef( struct lp_rast_triangle *tri,
                                unsigned slot,
                                boolean frontface,
                                unsigned usage_mask)
 {
    /* convert TRUE to 1.0 and FALSE to -1.0 */
    if (usage_mask & TGSI_WRITEMASK_X)
-      constant_coef( setup, tri, slot, 2.0f * frontface - 1.0f, 0 );
+      constant_coef( tri, slot, 2.0f * frontface - 1.0f, 0 );
 
    if (usage_mask & TGSI_WRITEMASK_Y)
-      constant_coef( setup, tri, slot, 0.0f, 1 ); /* wasted */
+      constant_coef( tri, slot, 0.0f, 1 ); /* wasted */
 
    if (usage_mask & TGSI_WRITEMASK_Z)
-      constant_coef( setup, tri, slot, 0.0f, 2 ); /* wasted */
+      constant_coef( tri, slot, 0.0f, 2 ); /* wasted */
 
    if (usage_mask & TGSI_WRITEMASK_W)
-      constant_coef( setup, tri, slot, 0.0f, 3 ); /* wasted */
+      constant_coef( tri, slot, 0.0f, 3 ); /* wasted */
 }
 
 
@@ -206,11 +256,7 @@ static void setup_facing_coef( struct lp_setup_context *setup,
  */
 static void setup_tri_coefficients( struct lp_setup_context *setup,
 				    struct lp_rast_triangle *tri,
-                                    float oneoverarea,
-				    const float (*v1)[4],
-				    const float (*v2)[4],
-				    const float (*v3)[4],
-				    boolean frontface)
+                                    const struct tri_info *info)
 {
    unsigned fragcoord_usage_mask = TGSI_WRITEMASK_XYZ;
    unsigned slot;
@@ -227,25 +273,25 @@ static void setup_tri_coefficients( struct lp_setup_context *setup,
          if (setup->flatshade_first) {
             for (i = 0; i < NUM_CHANNELS; i++)
                if (usage_mask & (1 << i))
-                  constant_coef(setup, tri, slot+1, v1[vert_attr][i], i);
+                  constant_coef(tri, slot+1, info->v0[vert_attr][i], i);
          }
          else {
             for (i = 0; i < NUM_CHANNELS; i++)
                if (usage_mask & (1 << i))
-                  constant_coef(setup, tri, slot+1, v3[vert_attr][i], i);
+                  constant_coef(tri, slot+1, info->v2[vert_attr][i], i);
          }
          break;
 
       case LP_INTERP_LINEAR:
          for (i = 0; i < NUM_CHANNELS; i++)
             if (usage_mask & (1 << i))
-               linear_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
+               linear_coef(tri, info, slot+1, vert_attr, i);
          break;
 
       case LP_INTERP_PERSPECTIVE:
          for (i = 0; i < NUM_CHANNELS; i++)
             if (usage_mask & (1 << i))
-               perspective_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
+               perspective_coef(tri, info, slot+1, vert_attr, i);
          fragcoord_usage_mask |= TGSI_WRITEMASK_W;
          break;
 
@@ -259,7 +305,7 @@ static void setup_tri_coefficients( struct lp_setup_context *setup,
          break;
 
       case LP_INTERP_FACING:
-         setup_facing_coef(setup, tri, slot+1, frontface, usage_mask);
+         setup_facing_coef(tri, slot+1, info->frontfacing, usage_mask);
          break;
 
       default:
@@ -269,16 +315,11 @@ static void setup_tri_coefficients( struct lp_setup_context *setup,
 
    /* The internal position input is in slot zero:
     */
-   setup_fragcoord_coef(setup, tri, oneoverarea, 0, v1, v2, v3,
-                        fragcoord_usage_mask);
+   setup_fragcoord_coef(tri, info, 0, fragcoord_usage_mask);
 }
 
 
 
-static INLINE int subpixel_snap( float a )
-{
-   return util_iround(FIXED_ONE * a - (FIXED_ONE / 2));
-}
 
 
 
@@ -291,21 +332,25 @@ static INLINE int subpixel_snap( float a )
  * \return pointer to triangle space
  */
 static INLINE struct lp_rast_triangle *
-alloc_triangle(struct lp_scene *scene, unsigned nr_inputs, unsigned *tri_size)
+alloc_triangle(struct lp_scene *scene,
+               unsigned nr_inputs,
+               unsigned nr_planes,
+               unsigned *tri_size)
 {
    unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
    struct lp_rast_triangle *tri;
-   unsigned bytes;
+   unsigned tri_bytes, bytes;
    char *inputs;
 
    assert(sizeof(*tri) % 16 == 0);
 
-   bytes = sizeof(*tri) + (3 * input_array_sz);
+   tri_bytes = align(Offset(struct lp_rast_triangle, plane[nr_planes]), 16);
+   bytes = tri_bytes + (3 * input_array_sz);
 
    tri = lp_scene_alloc_aligned( scene, bytes, 16 );
 
    if (tri) {
-      inputs = (char *) (tri + 1);
+      inputs = ((char *)tri) + tri_bytes;
       tri->inputs.a0   = (float (*)[4]) inputs;
       tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
       tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
@@ -329,52 +374,71 @@ print_triangle(struct lp_setup_context *setup,
    uint i;
 
    debug_printf("llvmpipe triangle\n");
-   for (i = 0; i < setup->fs.nr_inputs; i++) {
+   for (i = 0; i < 1 + setup->fs.nr_inputs; i++) {
       debug_printf("  v1[%d]:  %f %f %f %f\n", i,
                    v1[i][0], v1[i][1], v1[i][2], v1[i][3]);
    }
-   for (i = 0; i < setup->fs.nr_inputs; i++) {
+   for (i = 0; i < 1 + setup->fs.nr_inputs; i++) {
       debug_printf("  v2[%d]:  %f %f %f %f\n", i,
                    v2[i][0], v2[i][1], v2[i][2], v2[i][3]);
    }
-   for (i = 0; i < setup->fs.nr_inputs; i++) {
+   for (i = 0; i < 1 + setup->fs.nr_inputs; i++) {
       debug_printf("  v3[%d]:  %f %f %f %f\n", i,
                    v3[i][0], v3[i][1], v3[i][2], v3[i][3]);
    }
 }
 
 
+lp_rast_cmd lp_rast_tri_tab[8] = {
+   NULL,               /* should be impossible */
+   lp_rast_triangle_1,
+   lp_rast_triangle_2,
+   lp_rast_triangle_3,
+   lp_rast_triangle_4,
+   lp_rast_triangle_5,
+   lp_rast_triangle_6,
+   lp_rast_triangle_7
+};
+
 /**
  * Do basic setup for triangle rasterization and determine which
  * framebuffer tiles are touched.  Put the triangle in the scene's
  * bins for the tiles which we overlap.
  */
-static void 
+static void
 do_triangle_ccw(struct lp_setup_context *setup,
 		const float (*v1)[4],
 		const float (*v2)[4],
 		const float (*v3)[4],
 		boolean frontfacing )
 {
-   /* x/y positions in fixed point */
-   const int x1 = subpixel_snap(v1[0][0] + 0.5 - setup->pixel_offset);
-   const int x2 = subpixel_snap(v2[0][0] + 0.5 - setup->pixel_offset);
-   const int x3 = subpixel_snap(v3[0][0] + 0.5 - setup->pixel_offset);
-   const int y1 = subpixel_snap(v1[0][1] + 0.5 - setup->pixel_offset);
-   const int y2 = subpixel_snap(v2[0][1] + 0.5 - setup->pixel_offset);
-   const int y3 = subpixel_snap(v3[0][1] + 0.5 - setup->pixel_offset);
 
    struct lp_scene *scene = lp_setup_get_current_scene(setup);
+   struct lp_fragment_shader_variant *variant = setup->fs.current.variant;
    struct lp_rast_triangle *tri;
+   struct tri_info info;
    int area;
-   float oneoverarea;
    int minx, maxx, miny, maxy;
+   int ix0, ix1, iy0, iy1;
    unsigned tri_bytes;
-
+   int i;
+   int nr_planes = 3;
+      
    if (0)
       print_triangle(setup, v1, v2, v3);
 
-   tri = alloc_triangle(scene, setup->fs.nr_inputs, &tri_bytes);
+   if (setup->scissor_test) {
+      nr_planes = 7;
+   }
+   else {
+      nr_planes = 3;
+   }
+
+
+   tri = alloc_triangle(scene,
+                        setup->fs.nr_inputs,
+                        nr_planes,
+                        &tri_bytes);
    if (!tri)
       return;
 
@@ -387,15 +451,24 @@ do_triangle_ccw(struct lp_setup_context *setup,
    tri->v[2][1] = v3[0][1];
 #endif
 
-   tri->dx12 = x1 - x2;
-   tri->dx23 = x2 - x3;
-   tri->dx31 = x3 - x1;
+   /* x/y positions in fixed point */
+   info.x[0] = subpixel_snap(v1[0][0] - setup->pixel_offset);
+   info.x[1] = subpixel_snap(v2[0][0] - setup->pixel_offset);
+   info.x[2] = subpixel_snap(v3[0][0] - setup->pixel_offset);
+   info.y[0] = subpixel_snap(v1[0][1] - setup->pixel_offset);
+   info.y[1] = subpixel_snap(v2[0][1] - setup->pixel_offset);
+   info.y[2] = subpixel_snap(v3[0][1] - setup->pixel_offset);
+
+   tri->plane[0].dcdy = info.x[0] - info.x[1];
+   tri->plane[1].dcdy = info.x[1] - info.x[2];
+   tri->plane[2].dcdy = info.x[2] - info.x[0];
 
-   tri->dy12 = y1 - y2;
-   tri->dy23 = y2 - y3;
-   tri->dy31 = y3 - y1;
+   tri->plane[0].dcdx = info.y[0] - info.y[1];
+   tri->plane[1].dcdx = info.y[1] - info.y[2];
+   tri->plane[2].dcdx = info.y[2] - info.y[0];
 
-   area = (tri->dx12 * tri->dy31 - tri->dx31 * tri->dy12);
+   area = (tri->plane[0].dcdy * tri->plane[2].dcdx -
+           tri->plane[2].dcdy * tri->plane[0].dcdx);
 
    LP_COUNT(nr_tris);
 
@@ -410,20 +483,35 @@ do_triangle_ccw(struct lp_setup_context *setup,
    }
 
    /* Bounding rectangle (in pixels) */
-   minx = (MIN3(x1, x2, x3) + (FIXED_ONE-1)) >> FIXED_ORDER;
-   maxx = (MAX3(x1, x2, x3) + (FIXED_ONE-1)) >> FIXED_ORDER;
-   miny = (MIN3(y1, y2, y3) + (FIXED_ONE-1)) >> FIXED_ORDER;
-   maxy = (MAX3(y1, y2, y3) + (FIXED_ONE-1)) >> FIXED_ORDER;
-   
+   {
+      /* Yes this is necessary to accurately calculate bounding boxes
+       * with the two fill-conventions we support.  GL (normally) ends
+       * up needing a bottom-left fill convention, which requires
+       * slightly different rounding.
+       */
+      int adj = (setup->pixel_offset != 0) ? 1 : 0;
+
+      minx = (MIN3(info.x[0], info.x[1], info.x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
+      maxx = (MAX3(info.x[0], info.x[1], info.x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
+      miny = (MIN3(info.y[0], info.y[1], info.y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
+      maxy = (MAX3(info.y[0], info.y[1], info.y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
+   }
+
    if (setup->scissor_test) {
       minx = MAX2(minx, setup->scissor.current.minx);
       maxx = MIN2(maxx, setup->scissor.current.maxx);
       miny = MAX2(miny, setup->scissor.current.miny);
       maxy = MIN2(maxy, setup->scissor.current.maxy);
    }
+   else {
+      minx = MAX2(minx, 0);
+      miny = MAX2(miny, 0);
+      maxx = MIN2(maxx, scene->fb.width);
+      maxy = MIN2(maxy, scene->fb.height);
+   }
+
 
-   if (miny == maxy || 
-       minx == maxx) {
+   if (miny >= maxy || minx >= maxx) {
       lp_scene_putback_data( scene, tri_bytes );
       LP_COUNT(nr_culled_tris);
       return;
@@ -431,75 +519,87 @@ do_triangle_ccw(struct lp_setup_context *setup,
 
    /* 
     */
-   oneoverarea = ((float)FIXED_ONE) / (float)area;
+   info.pixel_offset = setup->pixel_offset;
+   info.v0 = v1;
+   info.v1 = v2;
+   info.v2 = v3;
+   info.dx01 = info.v0[0][0] - info.v1[0][0];
+   info.dx20 = info.v2[0][0] - info.v0[0][0];
+   info.dy01 = info.v0[0][1] - info.v1[0][1];
+   info.dy20 = info.v2[0][1] - info.v0[0][1];
+   info.oneoverarea = 1.0 / (info.dx01 * info.dy20 - info.dx20 * info.dy01);
+   info.frontfacing = frontfacing;
 
    /* Setup parameter interpolants:
     */
-   setup_tri_coefficients( setup, tri, oneoverarea, v1, v2, v3, frontfacing );
+   setup_tri_coefficients( setup, tri, &info );
 
    tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
 
-   /* half-edge constants, will be interated over the whole render target.
-    */
-   tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
-   tri->c2 = tri->dy23 * x2 - tri->dx23 * y2;
-   tri->c3 = tri->dy31 * x3 - tri->dx31 * y3;
 
-   /* correct for top-left fill convention:
-    */
-   if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1++;
-   if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2++;
-   if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3++;
-
-   tri->dy12 *= FIXED_ONE;
-   tri->dy23 *= FIXED_ONE;
-   tri->dy31 *= FIXED_ONE;
-
-   tri->dx12 *= FIXED_ONE;
-   tri->dx23 *= FIXED_ONE;
-   tri->dx31 *= FIXED_ONE;
-
-   /* find trivial reject offsets for each edge for a single-pixel
-    * sized block.  These will be scaled up at each recursive level to
-    * match the active blocksize.  Scaling in this way works best if
-    * the blocks are square.
-    */
-   tri->eo1 = 0;
-   if (tri->dy12 < 0) tri->eo1 -= tri->dy12;
-   if (tri->dx12 > 0) tri->eo1 += tri->dx12;
+  
+   for (i = 0; i < 3; i++) {
+      struct lp_rast_plane *plane = &tri->plane[i];
 
-   tri->eo2 = 0;
-   if (tri->dy23 < 0) tri->eo2 -= tri->dy23;
-   if (tri->dx23 > 0) tri->eo2 += tri->dx23;
+      /* half-edge constants, will be interated over the whole render
+       * target.
+       */
+      plane->c = plane->dcdx * info.x[i] - plane->dcdy * info.y[i];
+
+      /* correct for top-left vs. bottom-left fill convention.  
+       *
+       * note that we're overloading gl_rasterization_rules to mean
+       * both (0.5,0.5) pixel centers *and* bottom-left filling
+       * convention.
+       *
+       * GL actually has a top-left filling convention, but GL's
+       * notion of "top" differs from gallium's...
+       *
+       * Also, sometimes (in FBO cases) GL will render upside down
+       * to its usual method, in which case it will probably want
+       * to use the opposite, top-left convention.
+       */         
+      if (plane->dcdx < 0) {
+         /* both fill conventions want this - adjust for left edges */
+         plane->c++;            
+      }
+      else if (plane->dcdx == 0) {
+         if (setup->pixel_offset == 0) {
+            /* correct for top-left fill convention:
+             */
+            if (plane->dcdy > 0) plane->c++;
+         }
+         else {
+            /* correct for bottom-left fill convention:
+             */
+            if (plane->dcdy < 0) plane->c++;
+         }
+      }
 
-   tri->eo3 = 0;
-   if (tri->dy31 < 0) tri->eo3 -= tri->dy31;
-   if (tri->dx31 > 0) tri->eo3 += tri->dx31;
+      plane->dcdx *= FIXED_ONE;
+      plane->dcdy *= FIXED_ONE;
 
-   /* Calculate trivial accept offsets from the above.
-    */
-   tri->ei1 = tri->dx12 - tri->dy12 - tri->eo1;
-   tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
-   tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
+      /* find trivial reject offsets for each edge for a single-pixel
+       * sized block.  These will be scaled up at each recursive level to
+       * match the active blocksize.  Scaling in this way works best if
+       * the blocks are square.
+       */
+      plane->eo = 0;
+      if (plane->dcdx < 0) plane->eo -= plane->dcdx;
+      if (plane->dcdy > 0) plane->eo += plane->dcdy;
 
-   /* Fill in the inputs.step[][] arrays.
-    * We've manually unrolled some loops here.
-    */
-   {
-      const int xstep1 = -tri->dy12;
-      const int xstep2 = -tri->dy23;
-      const int xstep3 = -tri->dy31;
-      const int ystep1 = tri->dx12;
-      const int ystep2 = tri->dx23;
-      const int ystep3 = tri->dx31;
-
-#define SETUP_STEP(i, x, y)                                \
-      do {                                                 \
-         tri->inputs.step[0][i] = x * xstep1 + y * ystep1; \
-         tri->inputs.step[1][i] = x * xstep2 + y * ystep2; \
-         tri->inputs.step[2][i] = x * xstep3 + y * ystep3; \
-      } while (0)
+      /* Calculate trivial accept offsets from the above.
+       */
+      plane->ei = plane->dcdy - plane->dcdx - plane->eo;
 
+      plane->step = tri->step[i];
+
+      /* Fill in the inputs.step[][] arrays.
+       * We've manually unrolled some loops here.
+       */
+#define SETUP_STEP(j, x, y) \
+      tri->step[i][j] = y * plane->dcdy - x * plane->dcdx
+      
       SETUP_STEP(0, 0, 0);
       SETUP_STEP(1, 1, 0);
       SETUP_STEP(2, 0, 1);
@@ -522,63 +622,106 @@ do_triangle_ccw(struct lp_setup_context *setup,
 #undef STEP
    }
 
+
+   /* 
+    * When rasterizing scissored tris, use the intersection of the
+    * triangle bounding box and the scissor rect to generate the
+    * scissor planes.
+    *
+    * This permits us to cut off the triangle "tails" that are present
+    * in the intermediate recursive levels caused when two of the
+    * triangles edges don't diverge quickly enough to trivially reject
+    * exterior blocks from the triangle.
+    *
+    * It's not really clear if it's worth worrying about these tails,
+    * but since we generate the planes for each scissored tri, it's
+    * free to trim them in this case.
+    * 
+    * Note that otherwise, the scissor planes only vary in 'C' value,
+    * and even then only on state-changes.  Could alternatively store
+    * these planes elsewhere.
+    */
+   if (nr_planes == 7) {
+      tri->plane[3].step = step_scissor_minx;
+      tri->plane[3].dcdx = -1;
+      tri->plane[3].dcdy = 0;
+      tri->plane[3].c = 1-minx;
+      tri->plane[3].ei = 0;
+      tri->plane[3].eo = 1;
+
+      tri->plane[4].step = step_scissor_maxx;
+      tri->plane[4].dcdx = 1;
+      tri->plane[4].dcdy = 0;
+      tri->plane[4].c = maxx;
+      tri->plane[4].ei = -1;
+      tri->plane[4].eo = 0;
+
+      tri->plane[5].step = step_scissor_miny;
+      tri->plane[5].dcdx = 0;
+      tri->plane[5].dcdy = 1;
+      tri->plane[5].c = 1-miny;
+      tri->plane[5].ei = 0;
+      tri->plane[5].eo = 1;
+
+      tri->plane[6].step = step_scissor_maxy;
+      tri->plane[6].dcdx = 0;
+      tri->plane[6].dcdy = -1;
+      tri->plane[6].c = maxy;
+      tri->plane[6].ei = -1;
+      tri->plane[6].eo = 0;
+   }
+
+
    /*
     * All fields of 'tri' are now set.  The remaining code here is
     * concerned with binning.
     */
 
-   /* Convert to tile coordinates:
+   /* Convert to tile coordinates, and inclusive ranges:
     */
-   minx = minx / TILE_SIZE;
-   miny = miny / TILE_SIZE;
-   maxx = maxx / TILE_SIZE;
-   maxy = maxy / TILE_SIZE;
+   ix0 = minx / TILE_SIZE;
+   iy0 = miny / TILE_SIZE;
+   ix1 = (maxx-1) / TILE_SIZE;
+   iy1 = (maxy-1) / TILE_SIZE;
 
    /*
     * Clamp to framebuffer size
     */
-   minx = MAX2(minx, 0);
-   miny = MAX2(miny, 0);
-   maxx = MIN2(maxx, scene->tiles_x - 1);
-   maxy = MIN2(maxy, scene->tiles_y - 1);
+   assert(ix0 == MAX2(ix0, 0));
+   assert(iy0 == MAX2(iy0, 0));
+   assert(ix1 == MIN2(ix1, scene->tiles_x - 1));
+   assert(iy1 == MIN2(iy1, scene->tiles_y - 1));
 
    /* Determine which tile(s) intersect the triangle's bounding box
     */
-   if (miny == maxy && minx == maxx)
+   if (iy0 == iy1 && ix0 == ix1)
    {
       /* Triangle is contained in a single tile:
        */
-      lp_scene_bin_command( scene, minx, miny, lp_rast_triangle, 
-			    lp_rast_arg_triangle(tri) );
+      lp_scene_bin_command( scene, ix0, iy0,
+                            lp_rast_tri_tab[nr_planes], 
+			    lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
    }
-   else 
+   else
    {
-      int c1 = (tri->c1 + 
-                tri->dx12 * miny * TILE_SIZE - 
-                tri->dy12 * minx * TILE_SIZE);
-      int c2 = (tri->c2 + 
-                tri->dx23 * miny * TILE_SIZE -
-                tri->dy23 * minx * TILE_SIZE);
-      int c3 = (tri->c3 +
-                tri->dx31 * miny * TILE_SIZE -
-                tri->dy31 * minx * TILE_SIZE);
-
-      int ei1 = tri->ei1 << TILE_ORDER;
-      int ei2 = tri->ei2 << TILE_ORDER;
-      int ei3 = tri->ei3 << TILE_ORDER;
-
-      int eo1 = tri->eo1 << TILE_ORDER;
-      int eo2 = tri->eo2 << TILE_ORDER;
-      int eo3 = tri->eo3 << TILE_ORDER;
-
-      int xstep1 = -(tri->dy12 << TILE_ORDER);
-      int xstep2 = -(tri->dy23 << TILE_ORDER);
-      int xstep3 = -(tri->dy31 << TILE_ORDER);
-
-      int ystep1 = tri->dx12 << TILE_ORDER;
-      int ystep2 = tri->dx23 << TILE_ORDER;
-      int ystep3 = tri->dx31 << TILE_ORDER;
+      int c[7];
+      int ei[7];
+      int eo[7];
+      int xstep[7];
+      int ystep[7];
       int x, y;
+      
+      for (i = 0; i < nr_planes; i++) {
+         c[i] = (tri->plane[i].c + 
+                 tri->plane[i].dcdy * iy0 * TILE_SIZE - 
+                 tri->plane[i].dcdx * ix0 * TILE_SIZE);
+
+         ei[i] = tri->plane[i].ei << TILE_ORDER;
+         eo[i] = tri->plane[i].eo << TILE_ORDER;
+         xstep[i] = -(tri->plane[i].dcdx << TILE_ORDER);
+         ystep[i] = tri->plane[i].dcdy << TILE_ORDER;
+      }
+
 
 
       /* Test tile-sized blocks against the triangle.
@@ -586,32 +729,49 @@ do_triangle_ccw(struct lp_setup_context *setup,
        * contained inside the tri, bin an lp_rast_shade_tile command.
        * Else, bin a lp_rast_triangle command.
        */
-      for (y = miny; y <= maxy; y++)
+      for (y = iy0; y <= iy1; y++)
       {
-	 int cx1 = c1;
-	 int cx2 = c2;
-	 int cx3 = c3;
 	 boolean in = FALSE;  /* are we inside the triangle? */
+	 int cx[7];
+
+         for (i = 0; i < nr_planes; i++)
+            cx[i] = c[i];
 
-	 for (x = minx; x <= maxx; x++)
+	 for (x = ix0; x <= ix1; x++)
 	 {
-	    if (cx1 + eo1 < 0 || 
-		cx2 + eo2 < 0 ||
-		cx3 + eo3 < 0) 
-	    {
-	       /* do nothing */
+            int out = 0;
+            int partial = 0;
+
+            for (i = 0; i < nr_planes; i++) {
+               int planeout = cx[i] + eo[i];
+               int planepartial = cx[i] + ei[i] - 1;
+               out |= (planeout >> 31);
+               partial |= (planepartial >> 31) & (1<<i);
+            }
+
+            if (out) {
+               /* do nothing */
+               if (in)
+                  break;  /* exiting triangle, all done with this row */
                LP_COUNT(nr_empty_64);
-	       if (in)
-		  break;  /* exiting triangle, all done with this row */
-	    }
-	    else if (cx1 + ei1 > 0 &&
-		     cx2 + ei2 > 0 &&
-		     cx3 + ei3 > 0) 
-	    {
+            }
+            else if (partial) {
+               /* Not trivially accepted by at least one plane - 
+                * rasterize/shade partial tile
+                */
+               int count = util_bitcount(partial);
+               in = TRUE;
+               lp_scene_bin_command( scene, x, y,
+                                     lp_rast_tri_tab[count], 
+                                     lp_rast_arg_triangle(tri, partial) );
+
+               LP_COUNT(nr_partially_covered_64);
+            }
+            else {
                /* triangle covers the whole tile- shade whole tile */
                LP_COUNT(nr_fully_covered_64);
-	       in = TRUE;
-	       if (setup->fs.current.variant->opaque &&
+               in = TRUE;
+	       if (variant->opaque &&
 	           !setup->fb.zsbuf) {
 	          lp_scene_bin_reset( scene, x, y );
 	          lp_scene_bin_command( scene, x, y,
@@ -621,29 +781,18 @@ do_triangle_ccw(struct lp_setup_context *setup,
                lp_scene_bin_command( scene, x, y,
 				     lp_rast_shade_tile,
 				     lp_rast_arg_inputs(&tri->inputs) );
-	    }
-	    else 
-	    { 
-               /* rasterizer/shade partial tile */
-               LP_COUNT(nr_partially_covered_64);
-	       in = TRUE;
-               lp_scene_bin_command( scene, x, y,
-				     lp_rast_triangle, 
-				     lp_rast_arg_triangle(tri) );
-	    }
+            }
 
 	    /* Iterate cx values across the region:
 	     */
-	    cx1 += xstep1;
-	    cx2 += xstep2;
-	    cx3 += xstep3;
+            for (i = 0; i < nr_planes; i++)
+               cx[i] += xstep[i];
 	 }
       
 	 /* Iterate c values down the region:
 	  */
-	 c1 += ystep1;
-	 c2 += ystep2;
-	 c3 += ystep3;    
+         for (i = 0; i < nr_planes; i++)
+            c[i] += ystep[i];
       }
    }
 }
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index 65115052cd..5953d690a4 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -31,9 +31,6 @@
  * Code generate the whole fragment pipeline.
  *
  * The fragment pipeline consists of the following stages:
- * - triangle edge in/out testing
- * - scissor test
- * - stipple (TBI)
  * - early depth test
  * - fragment shader
  * - alpha test
@@ -97,6 +94,7 @@
 #include "lp_state.h"
 #include "lp_tex_sample.h"
 #include "lp_flush.h"
+#include "lp_state_fs.h"
 
 
 #include <llvm-c/Analysis.h>
@@ -170,177 +168,63 @@ generate_depth_stencil(LLVMBuilderRef builder,
 
 
 /**
- * Generate the code to do inside/outside triangle testing for the
+ * Expand the relevent bits of mask_input to a 4-dword mask for the 
  * four pixels in a 2x2 quad.  This will set the four elements of the
  * quad mask vector to 0 or ~0.
- * \param i  which quad of the quad group to test, in [0,3]
+ *
+ * \param quad  which quad of the quad group to test, in [0,3]
+ * \param mask_input  bitwise mask for the whole 4x4 stamp
  */
-static void
-generate_tri_edge_mask(LLVMBuilderRef builder,
-                       unsigned i,
-                       LLVMValueRef *mask,      /* ivec4, out */
-                       LLVMValueRef c0,         /* int32 */
-                       LLVMValueRef c1,         /* int32 */
-                       LLVMValueRef c2,         /* int32 */
-                       LLVMValueRef step0_ptr,  /* ivec4 */
-                       LLVMValueRef step1_ptr,  /* ivec4 */
-                       LLVMValueRef step2_ptr)  /* ivec4 */
+static LLVMValueRef
+generate_quad_mask(LLVMBuilderRef builder,
+                   struct lp_type fs_type,
+                   unsigned quad,
+                   LLVMValueRef mask_input) /* int32 */
 {
-#define OPTIMIZE_IN_OUT_TEST 0
-#if OPTIMIZE_IN_OUT_TEST
-   struct lp_build_if_state ifctx;
-   LLVMValueRef not_draw_all;
-#endif
-   struct lp_build_flow_context *flow;
-   struct lp_type i32_type;
-   LLVMTypeRef i32vec4_type;
-   LLVMValueRef c0_vec, c1_vec, c2_vec;
-   LLVMValueRef in_out_mask;
-
-   assert(i < 4);
-   
-   /* int32 vector type */
-   memset(&i32_type, 0, sizeof i32_type);
-   i32_type.floating = FALSE; /* values are integers */
-   i32_type.sign = TRUE;      /* values are signed */
-   i32_type.norm = FALSE;     /* values are not normalized */
-   i32_type.width = 32;       /* 32-bit int values */
-   i32_type.length = 4;       /* 4 elements per vector */
-
-   i32vec4_type = lp_build_int32_vec4_type();
+   struct lp_type mask_type;
+   LLVMTypeRef i32t = LLVMInt32Type();
+   LLVMValueRef bits[4];
+   LLVMValueRef mask;
 
    /*
-    * Use a conditional here to do detailed pixel in/out testing.
-    * We only have to do this if c0 != INT_MIN.
+    * XXX: We'll need a different path for 16 x u8
     */
-   flow = lp_build_flow_create(builder);
-   lp_build_flow_scope_begin(flow);
-
-   {
-#if OPTIMIZE_IN_OUT_TEST
-      /* not_draw_all = (c0 != INT_MIN) */
-      not_draw_all = LLVMBuildICmp(builder,
-                                   LLVMIntNE,
-                                   c0,
-                                   LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
-                                   "");
-
-      in_out_mask = lp_build_const_int_vec(i32_type, ~0);
-
-
-      lp_build_flow_scope_declare(flow, &in_out_mask);
-
-      /* if (not_draw_all) {... */
-      lp_build_if(&ifctx, flow, builder, not_draw_all);
-#endif
-      {
-         LLVMValueRef step0_vec, step1_vec, step2_vec;
-         LLVMValueRef m0_vec, m1_vec, m2_vec;
-         LLVMValueRef index, m;
-
-         /* c0_vec = {c0, c0, c0, c0}
-          * Note that we emit this code four times but LLVM optimizes away
-          * three instances of it.
-          */
-         c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
-         c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
-         c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
-         lp_build_name(c0_vec, "edgeconst0vec");
-         lp_build_name(c1_vec, "edgeconst1vec");
-         lp_build_name(c2_vec, "edgeconst2vec");
-
-         /* load step0vec, step1, step2 vec from memory */
-         index = LLVMConstInt(LLVMInt32Type(), i, 0);
-         step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
-         step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
-         step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
-         lp_build_name(step0_vec, "step0vec");
-         lp_build_name(step1_vec, "step1vec");
-         lp_build_name(step2_vec, "step2vec");
-
-         /* m0_vec = step0_ptr[i] > c0_vec */
-         m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
-         m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
-         m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
-
-         /* in_out_mask = m0_vec & m1_vec & m2_vec */
-         m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
-         in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
-         lp_build_name(in_out_mask, "inoutmaskvec");
-      }
-#if OPTIMIZE_IN_OUT_TEST
-      lp_build_endif(&ifctx);
-#endif
-
-   }
-   lp_build_flow_scope_end(flow);
-   lp_build_flow_destroy(flow);
+   assert(fs_type.width == 32);
+   assert(fs_type.length == 4);
+   mask_type = lp_int_type(fs_type);
 
-   /* This is the initial alive/dead pixel mask for a quad of four pixels.
-    * It's an int[4] vector with each word set to 0 or ~0.
-    * Words will get cleared when pixels faile the Z test, etc.
+   /*
+    * mask_input >>= (quad * 4)
     */
-   *mask = in_out_mask;
-}
-
-
-static LLVMValueRef
-generate_scissor_test(LLVMBuilderRef builder,
-                      LLVMValueRef context_ptr,
-                      const struct lp_build_interp_soa_context *interp,
-                      struct lp_type type)
-{
-   LLVMTypeRef vec_type = lp_build_vec_type(type);
-   LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
-   LLVMValueRef xmin, ymin, xmax, ymax;
-   LLVMValueRef m0, m1, m2, m3, m;
-
-   /* xpos, ypos contain the window coords for the four pixels in the quad */
-   assert(xpos);
-   assert(ypos);
-
-   /* get the current scissor bounds, convert to vectors */
-   xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
-   xmin = lp_build_broadcast(builder, vec_type, xmin);
-
-   ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
-   ymin = lp_build_broadcast(builder, vec_type, ymin);
 
-   xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
-   xmax = lp_build_broadcast(builder, vec_type, xmax);
+   mask_input = LLVMBuildLShr(builder,
+                              mask_input,
+                              LLVMConstInt(i32t, quad * 4, 0),
+                              "");
 
-   ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
-   ymax = lp_build_broadcast(builder, vec_type, ymax);
+   /*
+    * mask = { mask_input & (1 << i), for i in [0,3] }
+    */
 
-   /* compare the fragment's position coordinates against the scissor bounds */
-   m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
-   m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
-   m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
-   m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
+   mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input);
 
-   /* AND all the masks together */
-   m = LLVMBuildAnd(builder, m0, m1, "");
-   m = LLVMBuildAnd(builder, m, m2, "");
-   m = LLVMBuildAnd(builder, m, m3, "");
+   bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
+   bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
+   bits[2] = LLVMConstInt(i32t, 1 << 2, 0);
+   bits[3] = LLVMConstInt(i32t, 1 << 3, 0);
 
-   lp_build_name(m, "scissormask");
+   mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
 
-   return m;
-}
+   /*
+    * mask = mask != 0 ? ~0 : 0
+    */
 
+   mask = lp_build_compare(builder,
+                           mask_type, PIPE_FUNC_NOTEQUAL,
+                           mask,
+                           lp_build_const_int_vec(mask_type, 0));
 
-static LLVMValueRef
-build_int32_vec_const(int value)
-{
-   struct lp_type i32_type;
-
-   memset(&i32_type, 0, sizeof i32_type);
-   i32_type.floating = FALSE; /* values are integers */
-   i32_type.sign = TRUE;      /* values are signed */
-   i32_type.norm = FALSE;     /* values are not normalized */
-   i32_type.width = 32;       /* 32-bit int values */
-   i32_type.length = 4;       /* 4 elements per vector */
-   return lp_build_const_int_vec(i32_type, value);
+   return mask;
 }
 
 
@@ -348,7 +232,7 @@ build_int32_vec_const(int value)
 /**
  * Generate the fragment shader, depth/stencil test, and alpha tests.
  * \param i  which quad in the tile, in range [0,3]
- * \param do_tri_test  if 1, do triangle edge in/out testing
+ * \param partial_mask  if 1, do mask_input testing
  */
 static void
 generate_fs(struct llvmpipe_context *lp,
@@ -364,13 +248,8 @@ generate_fs(struct llvmpipe_context *lp,
             LLVMValueRef (*color)[4],
             LLVMValueRef depth_ptr,
             LLVMValueRef facing,
-            unsigned do_tri_test,
-            LLVMValueRef c0,
-            LLVMValueRef c1,
-            LLVMValueRef c2,
-            LLVMValueRef step0_ptr,
-            LLVMValueRef step1_ptr,
-            LLVMValueRef step2_ptr,
+            unsigned partial_mask,
+            LLVMValueRef mask_input,
             LLVMValueRef counter)
 {
    const struct tgsi_token *tokens = shader->base.tokens;
@@ -411,23 +290,17 @@ generate_fs(struct llvmpipe_context *lp,
    lp_build_flow_scope_declare(flow, &z);
 
    /* do triangle edge testing */
-   if (do_tri_test) {
-      generate_tri_edge_mask(builder, i, pmask,
-                             c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
+   if (partial_mask) {
+      *pmask = generate_quad_mask(builder, type,
+                                  i, mask_input);
    }
    else {
-      *pmask = build_int32_vec_const(~0);
+      *pmask = lp_build_const_int_vec(type, ~0);
    }
 
    /* 'mask' will control execution based on quad's pixel alive/killed state */
    lp_build_mask_begin(&mask, flow, type, *pmask);
 
-   if (key->scissor) {
-      LLVMValueRef smask =
-         generate_scissor_test(builder, context_ptr, interp, type);
-      lp_build_mask_update(&mask, smask);
-   }
-
    early_depth_stencil_test =
       (key->depth.enabled || key->stencil[0].enabled) &&
       !key->alpha.enabled &&
@@ -579,7 +452,7 @@ static void
 generate_fragment(struct llvmpipe_context *lp,
                   struct lp_fragment_shader *shader,
                   struct lp_fragment_shader_variant *variant,
-                  unsigned do_tri_test)
+                  unsigned partial_mask)
 {
    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
    const struct lp_fragment_shader_variant_key *key = &variant->key;
@@ -589,9 +462,8 @@ generate_fragment(struct llvmpipe_context *lp,
    LLVMTypeRef fs_elem_type;
    LLVMTypeRef fs_int_vec_type;
    LLVMTypeRef blend_vec_type;
-   LLVMTypeRef arg_types[16];
+   LLVMTypeRef arg_types[11];
    LLVMTypeRef func_type;
-   LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
    LLVMValueRef context_ptr;
    LLVMValueRef x;
    LLVMValueRef y;
@@ -600,7 +472,8 @@ generate_fragment(struct llvmpipe_context *lp,
    LLVMValueRef dady_ptr;
    LLVMValueRef color_ptr_ptr;
    LLVMValueRef depth_ptr;
-   LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr, counter = NULL;
+   LLVMValueRef mask_input;
+   LLVMValueRef counter = NULL;
    LLVMBasicBlockRef block;
    LLVMBuilderRef builder;
    struct lp_build_sampler_soa *sampler;
@@ -645,7 +518,7 @@ generate_fragment(struct llvmpipe_context *lp,
    blend_vec_type = lp_build_vec_type(blend_type);
 
    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s", 
-		 shader->no, variant->no, do_tri_test ? "edge" : "whole");
+		 shader->no, variant->no, partial_mask ? "partial" : "whole");
 
    arg_types[0] = screen->context_ptr_type;            /* context */
    arg_types[1] = LLVMInt32Type();                     /* x */
@@ -656,23 +529,15 @@ generate_fragment(struct llvmpipe_context *lp,
    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
-   arg_types[9] = LLVMInt32Type();                     /* c0 */
-   arg_types[10] = LLVMInt32Type();                    /* c1 */
-   arg_types[11] = LLVMInt32Type();                    /* c2 */
-   /* Note: the step arrays are built as int32[16] but we interpret
-    * them here as int32_vec4[4].
-    */
-   arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
-   arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
-   arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
-   arg_types[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
+   arg_types[9] = LLVMInt32Type();                     /* mask_input */
+   arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 
    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 
    function = LLVMAddFunction(screen->module, func_name, func_type);
    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 
-   variant->function[do_tri_test] = function;
+   variant->function[partial_mask] = function;
 
 
    /* XXX: need to propagate noalias down into color param now we are
@@ -691,12 +556,7 @@ generate_fragment(struct llvmpipe_context *lp,
    dady_ptr     = LLVMGetParam(function, 6);
    color_ptr_ptr = LLVMGetParam(function, 7);
    depth_ptr    = LLVMGetParam(function, 8);
-   c0           = LLVMGetParam(function, 9);
-   c1           = LLVMGetParam(function, 10);
-   c2           = LLVMGetParam(function, 11);
-   step0_ptr    = LLVMGetParam(function, 12);
-   step1_ptr    = LLVMGetParam(function, 13);
-   step2_ptr    = LLVMGetParam(function, 14);
+   mask_input   = LLVMGetParam(function, 9);
 
    lp_build_name(context_ptr, "context");
    lp_build_name(x, "x");
@@ -706,15 +566,10 @@ generate_fragment(struct llvmpipe_context *lp,
    lp_build_name(dady_ptr, "dady");
    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
    lp_build_name(depth_ptr, "depth");
-   lp_build_name(c0, "c0");
-   lp_build_name(c1, "c1");
-   lp_build_name(c2, "c2");
-   lp_build_name(step0_ptr, "step0");
-   lp_build_name(step1_ptr, "step1");
-   lp_build_name(step2_ptr, "step2");
+   lp_build_name(mask_input, "mask_input");
 
    if (key->occlusion_count) {
-      counter = LLVMGetParam(function, 15);
+      counter = LLVMGetParam(function, 10);
       lp_build_name(counter, "counter");
    }
 
@@ -763,9 +618,9 @@ generate_fragment(struct llvmpipe_context *lp,
                   out_color,
                   depth_ptr_i,
                   facing,
-                  do_tri_test,
-                  c0, c1, c2,
-                  step0_ptr, step1_ptr, step2_ptr, counter);
+                  partial_mask,
+                  mask_input,
+                  counter);
 
       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 	 for(chan = 0; chan < NUM_CHANNELS; ++chan)
@@ -792,9 +647,13 @@ generate_fragment(struct llvmpipe_context *lp,
 	 lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
       }
 
-      lp_build_conv_mask(builder, fs_type, blend_type,
-			 fs_mask, num_fs,
-			 &blend_mask, 1);
+      if (partial_mask || !variant->opaque) {
+         lp_build_conv_mask(builder, fs_type, blend_type,
+                            fs_mask, num_fs,
+                            &blend_mask, 1);
+      } else {
+         blend_mask = lp_build_const_int_vec(blend_type, ~0);
+      }
 
       color_ptr = LLVMBuildLoad(builder, 
 				LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
@@ -832,8 +691,7 @@ generate_fragment(struct llvmpipe_context *lp,
 #endif
 
    /* Apply optimizations to LLVM IR */
-   if (1)
-      LLVMRunFunctionPassManager(screen->pass, function);
+   LLVMRunFunctionPassManager(screen->pass, function);
 
    if (gallivm_debug & GALLIVM_DEBUG_IR) {
       /* Print the LLVM IR to stderr */
@@ -847,7 +705,7 @@ generate_fragment(struct llvmpipe_context *lp,
    {
       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 
-      variant->jit_function[do_tri_test] = (lp_jit_frag_func)pointer_to_func(f);
+      variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 
       if (gallivm_debug & GALLIVM_DEBUG_ASM) {
          lp_disassemble(f);
@@ -963,7 +821,6 @@ generate_variant(struct llvmpipe_context *lp,
          !key->stencil[0].enabled &&
          !key->alpha.enabled &&
          !key->depth.enabled &&
-         !key->scissor &&
          !shader->info.uses_kill
          ? TRUE : FALSE;
 
@@ -1182,7 +1039,6 @@ make_variant_key(struct llvmpipe_context *lp,
    /* alpha.ref_value is passed in jit_context */
 
    key->flatshade = lp->rasterizer->flatshade;
-   key->scissor = lp->rasterizer->scissor;
    if (lp->active_query_count) {
       key->occlusion_count = TRUE;
    }
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.h b/src/gallium/drivers/llvmpipe/lp_state_fs.h
index 593cd4de6b..37900fc544 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.h
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.h
@@ -54,7 +54,6 @@ struct lp_fragment_shader_variant_key
    enum pipe_format zsbuf_format;
    unsigned nr_cbufs:8;
    unsigned flatshade:1;
-   unsigned scissor:1;
    unsigned occlusion_count:1;
 
    struct {