Mesa (master): aco: Spill more optimally before loops.

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: b75d8052a7767af86d7abf85fca3db371f05d362
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=b75d8052a7767af86d7abf85fca3db371f05d362

Author: Timur Kristóf <timur.kristof at gmail.com>
Date:   Wed Nov 25 20:05:25 2020 +0100

aco: Spill more optimally before loops.

This further reduces the dead code emitted by the spiller.
Some minimal amount of dead IR is still emitted sometimes,
but that doesn't generate any compiled code at the end.

Totals from 1953 (1.40% of 139391) affected shaders:
VGPRs: 206980 -> 206588 (-0.19%)
SpillSGPRs: 24719 -> 16423 (-33.56%); split: -33.58%, +0.02%
CodeSize: 28448516 -> 28343836 (-0.37%); split: -0.38%, +0.01%
MaxWaves: 8960 -> 8992 (+0.36%)
Instrs: 5422049 -> 5408334 (-0.25%); split: -0.26%, +0.01%
Cycles: 511240864 -> 512460764 (+0.24%); split: -0.02%, +0.26%
VMEM: 346681 -> 346468 (-0.06%); split: +0.27%, -0.33%
SMEM: 124160 -> 122802 (-1.09%); split: +0.33%, -1.42%
VClause: 81102 -> 81163 (+0.08%); split: -0.01%, +0.09%
SClause: 174404 -> 174237 (-0.10%); split: -0.23%, +0.13%
Copies: 530216 -> 532961 (+0.52%); split: -0.90%, +1.42%
Branches: 189114 -> 189221 (+0.06%); split: -0.13%, +0.18%
PreSGPRs: 206017 -> 206526 (+0.25%); split: -0.08%, +0.33%
PreVGPRs: 183103 -> 182964 (-0.08%)

Co-authored-by: Daniel Schürmann <daniel at schuermann.dev>
Signed-off-by: Timur Kristóf <timur.kristof at gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8026>

---

 src/amd/compiler/aco_spill.cpp | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/src/amd/compiler/aco_spill.cpp b/src/amd/compiler/aco_spill.cpp
index 68af30c1016..0a6e8bec49e 100644
--- a/src/amd/compiler/aco_spill.cpp
+++ b/src/amd/compiler/aco_spill.cpp
@@ -399,12 +399,13 @@ RegisterDemand init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_id
          }
       }
 
-      /* select live-through vgpr variables */
+      /* select live-through vgpr variables and constants */
       while (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
          unsigned distance = 0;
          Temp to_spill;
          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
             if (pair.first.type() == RegType::vgpr &&
+                (pair.second.first >= loop_end || ctx.remat.count(pair.first)) &&
                 pair.second.first >= loop_end &&
                 pair.second.second > distance &&
                 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
@@ -426,13 +427,13 @@ RegisterDemand init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_id
          spilled_registers.vgpr += to_spill.size();
       }
 
-      /* select live-through sgpr variables */
+      /* select live-through sgpr variables and constants */
       while (new_demand.sgpr - spilled_registers.sgpr > ctx.target_pressure.sgpr) {
          unsigned distance = 0;
          Temp to_spill;
          for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
             if (pair.first.type() == RegType::sgpr &&
-                pair.second.first >= loop_end &&
+                (pair.second.first >= loop_end || ctx.remat.count(pair.first)) &&
                 pair.second.second > distance &&
                 ctx.spills_entry[block_idx].find(pair.first) == ctx.spills_entry[block_idx].end()) {
                to_spill = pair.first;