llvmpipe: reorganize block4 loop, nice speedup

isosurf 95->115 fps just by exchanging the two inner loops in this
function...

2 files changed, 19 insertions, 12 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index ebe9a8e92b4..c1f2680ddca 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -113,6 +113,21 @@ block_full_16(struct lp_rasterizer_task *task,
 	 block_full_4(task, tri, x + ix, y + iy);
 }
 
+
+static INLINE unsigned
+build_mask(int c, const int *step)
+{
+   int mask = 0;
+   int i;
+
+   for (i = 0; i < 16; i++) {
+      mask |= ((c + step[i]) >> 31) & (1 << i);
+   }
+   
+   return mask;
+}
+
+
 #define TAG(x) x##_1
 #define NR_PLANES 1
 #include "lp_rast_tri_tmp.h"
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
index a410c611a3f..fcb8e2b05d2 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
@@ -46,19 +46,11 @@ TAG(do_block_4)(struct lp_rasterizer_task *task,
                 int x, int y,
                 const int *c)
 {
-   unsigned mask = 0;
-   int i;
+   unsigned mask = 0xffff;
+   int j;
 
-   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);
+   for (j = 0; j < NR_PLANES; j++) {
+      mask &= ~build_mask(c[j] - 1, plane[j].step);
    }
 
    /* Now pass to the shader: