[Mesa-dev] [PATCH] :( fail. (i965 TXD state dependent recompiles)

[Kenneth Graunke]

On first compile...
- brw_prepare_wm_prog populates the key with depth_compare_func = 0
  and searches the cache.  No hit.
- It then sets depth_compare_func and searches the cache again.  No hit.
  (This is an additional cost.)
- It resets depth_compare_func to 0 and compiles.
- During compilation, emit_texture_gen5 will set key->depth_compare_func
  only if manual shadow comparisons are required (otherwise it stays 0).
- The newly compiled shader is added to the cache.

Subsequently...
- For most shaders:
  - brw_prepare_wm_prog searches the cache with depth_compare_func = 0.
    The shader is immediately found (no additional cost).
- If manual compares are required:
  - brw_prepare_wm_prog searches the cache with depth_compare_func = 0.
    No hit.  It then depth_compare_func and searches again.  If we've
    seen this DepthFunc before, we find the shader.  Total cost is two
    cache searches.  If we haven't seen DepthFunc before, compile.
---
 src/mesa/drivers/dri/i965/brw_wm.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_wm.c b/src/mesa/drivers/dri/i965/brw_wm.c
index 1aebd12..ffbc3b3 100644
--- a/src/mesa/drivers/dri/i965/brw_wm.c
+++ b/src/mesa/drivers/dri/i965/brw_wm.c
@@ -484,6 +484,14 @@ static void brw_prepare_wm_prog(struct brw_context *brw)
 				      &key, sizeof(key),
 				      &brw->wm.prog_data);
    if (brw->wm.prog_bo == NULL) {
+      key->depth_compare_func = ctx->Depth.Func;
+      brw->wm.prog_bo = brw_search_cache(&brw->cache, BRW_WM_PROG,
+				         &key, sizeof(key),
+					 &brw->wm.prog_data);
+   }
+
+   if (brw->wm.prog_bo == NULL) {
+      key->depth_compare_func = 0;
       bool success = do_wm_prog(brw, ctx->Shader.CurrentFragmentProgram, fp,
 				&key);
       assert(success);
-- 
1.7.5.4