aco: only apply extract if not used more than 4 times

Totals from 61 (0.04% of 150170) affected shaders: (GFX10.3)
CodeSize: 1087732 -> 1087380 (-0.03%); split: -0.22%, +0.18%
Instrs: 192343 -> 192205 (-0.07%); split: -0.16%, +0.09%
Latency: 7231670 -> 7148073 (-1.16%); split: -1.19%, +0.04%
InvThroughput: 3436715 -> 3394926 (-1.22%); split: -1.25%, +0.04%
VClause: 4831 -> 4833 (+0.04%)
Copies: 50130 -> 49934 (-0.39%); split: -0.67%, +0.28%
Branches: 5945 -> 5948 (+0.05%)
PreSGPRs: 3486 -> 3472 (-0.40%)
PreVGPRs: 5154 -> 5152 (-0.04%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11678>

diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index acfdfd1accd4b37d01d2e225f5ada8378bcf1d0b..e9f8514befbdfc3bed0f27d81726728fbeb01872 100644 (file)
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -3149,6 +3149,14 @@ combine_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
          if (!op.isTemp())
             continue;
          ssa_info& info = ctx.info[op.tempId()];
+         if (!info.is_extract())
+            continue;
+         /* if there are that many uses, there are likely better combinations */
+         // TODO: delay applying extract to a point where we know better
+         if (ctx.uses[op.tempId()] > 4) {
+            info.label &= ~label_extract;
+            continue;
+         }
          if (info.is_extract() &&
              (info.instr->operands[0].getTemp().type() == RegType::vgpr ||
               instr->operands[i].getTemp().type() == RegType::sgpr) &&