[Mesa-dev] [PATCH 2/2] radv: run nir_opt_shrink_load

[Samuel Pitoiset]

LLVM can't shrink loads.

Totals from affected shaders:
SGPRS: 62528 -> 59955 (-4.11 %)
VGPRS: 44708 -> 44616 (-0.21 %)
Spilled SGPRs: 16 -> 8 (-50.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 1355528 -> 1355172 (-0.03 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 11710 -> 11670 (-0.34 %)
Wait states: 0 -> 0 (0.00 %)

This reduces SGPRs spilling in MadMax, and it also reduces
number of SGPRs in DOW3 and F12017. The number of waves slightly
decreases in F1 but I don't see any performance changes after
benchmarking it. Talos and Serious Sam are not affected because
they don't use any push constants.

Note that we could just do the same optimization directly in
visit_load_push_constant(), but I think it's better to move that
pass in common code if other drivers want to use it.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
---
 src/amd/vulkan/radv_shader.c                                             | 1 +
 .../nir/{nir_opt_shrink_load_constant.c => nir_opt_shrink_load.c}        | 0
 2 files changed, 1 insertion(+)
 rename src/compiler/nir/{nir_opt_shrink_load_constant.c => nir_opt_shrink_load.c} (100%)

diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index af094e6220..ad68873055 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -147,6 +147,7 @@ radv_optimize_nir(struct nir_shader *shader)
                 if (shader->options->max_unroll_iterations) {
                         NIR_PASS(progress, shader, nir_opt_loop_unroll, 0);
                 }
+                NIR_PASS(progress, shader, nir_opt_shrink_load);
         } while (progress);
 }
 
diff --git a/src/compiler/nir/nir_opt_shrink_load_constant.c b/src/compiler/nir/nir_opt_shrink_load.c
similarity index 100%
rename from src/compiler/nir/nir_opt_shrink_load_constant.c
rename to src/compiler/nir/nir_opt_shrink_load.c
-- 
2.16.1