Mesa (lp-binning): llvmpipe: improve the in/out test a little

[Brian Paul]

Module: Mesa
Branch: lp-binning
Commit: b9d33db0a4cb818154b713a27834f66025b14672
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=b9d33db0a4cb818154b713a27834f66025b14672

Author: Brian Paul <brianp at vmware.com>
Date:   Wed Dec 16 17:08:13 2009 -0700

llvmpipe: improve the in/out test a little

Instead of:
  s = c + step
  m = s > 0
Do:
  m = step > c  (with negated c)

---

 src/gallium/drivers/llvmpipe/lp_rast.c     |    3 +--
 src/gallium/drivers/llvmpipe/lp_rast_tri.c |    4 ++--
 src/gallium/drivers/llvmpipe/lp_state_fs.c |   24 ++++++------------------
 3 files changed, 9 insertions(+), 22 deletions(-)

diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index b1bd27d..015865a 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -281,11 +281,10 @@ void lp_rast_shade_tile( struct lp_rasterizer *rast,
                          const union lp_rast_cmd_arg arg )
 {
    /* Set c1,c2,c3 to large values so the in/out test always passes */
-   const int32_t c1 = INT_MAX/2, c2 = INT_MAX/2, c3 = INT_MAX/2;
+   const int32_t c1 = INT_MIN/2, c2 = INT_MIN/2, c3 = INT_MIN/2;
    const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
    const unsigned tile_x = rast->tasks[thread_index].x;
    const unsigned tile_y = rast->tasks[thread_index].y;
-   const unsigned mask = ~0;
    unsigned x, y;
 
    LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index 9b18612..d6e8d6d 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -90,7 +90,7 @@ block_full_4( struct lp_rasterizer_task *rast_task,
               int x, int y )
 {
    /* Set c1,c2,c3 to large values so the in/out test always passes */
-   const int32_t c1 = INT_MAX/2, c2 = INT_MAX/2, c3 = INT_MAX/2;
+   const int32_t c1 = INT_MIN/2, c2 = INT_MIN/2, c3 = INT_MIN/2;
    lp_rast_shade_quads(rast_task->rast,
                        rast_task->thread_index,
                        &tri->inputs, 
@@ -133,7 +133,7 @@ do_block_4( struct lp_rasterizer_task *rast_task,
                        rast_task->thread_index,
                        &tri->inputs, 
                        x, y,
-                       c1, c2, c3);
+                       -c1, -c2, -c3);
 }
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index 4af37e3..15b175a 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -196,12 +196,9 @@ generate_tri_edge_mask(LLVMBuilderRef builder,
      c0_vec = splat(c0)
      c1_vec = splat(c1)
      c2_vec = splat(c2)
-     s0_vec = c0_vec + step0_ptr[i]
-     s1_vec = c1_vec + step1_ptr[i]
-     s2_vec = c2_vec + step2_ptr[i]
-     m0_vec = s0_vec > {0,0,0,0}
-     m1_vec = s1_vec > {0,0,0,0}
-     m2_vec = s2_vec > {0,0,0,0}
+     m0_vec = step0_ptr[i] > c0_vec
+     m1_vec = step1_ptr[i] > c1_vec
+     m2_vec = step2_ptr[i] > c2_vec
      mask = m0_vec & m1_vec & m2_vec
     */
    struct lp_type i32_type;
@@ -211,7 +208,6 @@ generate_tri_edge_mask(LLVMBuilderRef builder,
    LLVMValueRef c0_vec, c1_vec, c2_vec;
    LLVMValueRef step0_vec, step1_vec, step2_vec;
    LLVMValueRef m0_vec, m1_vec, m2_vec;
-   LLVMValueRef s0_vec, s1_vec, s2_vec;
    LLVMValueRef m;
 
    LLVMValueRef zeros;
@@ -240,21 +236,13 @@ generate_tri_edge_mask(LLVMBuilderRef builder,
    step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
    step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 
-   /** XXX with a little work, we could remove the add here and just
-    * compare c0_vec > step0_vec.
-    */
-   s0_vec = LLVMBuildAdd(builder, c0_vec, step0_vec, "");
-   s1_vec = LLVMBuildAdd(builder, c1_vec, step1_vec, "");
-   s2_vec = LLVMBuildAdd(builder, c2_vec, step2_vec, "");
-   m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, s0_vec, zeros);
-   m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, s1_vec, zeros);
-   m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, s2_vec, zeros);
+   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);
 
    m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
    m = LLVMBuildAnd(builder, m, m2_vec, "");
 
-   lp_build_name(m, "m");
-
    *mask = m;
 }