Mesa (master): aco/cssa: don't create parallelcopies for constants and exec

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: b6a28aaa8bacebf0ff78f2d811a80d3e77a8ef3a
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=b6a28aaa8bacebf0ff78f2d811a80d3e77a8ef3a

Author: Daniel Schürmann <daniel at schuermann.dev>
Date:   Mon Feb 22 14:58:37 2021 +0100

aco/cssa: don't create parallelcopies for constants and exec

if we are able to spill these directly.

Totals from 4913 (3.60% of 136546) affected shaders (Raven):
SpillSGPRs: 16021 -> 15451 (-3.56%); split: -3.87%, +0.31%
CodeSize: 58102020 -> 57371464 (-1.26%); split: -1.26%, +0.00%
Instrs: 11411454 -> 11230105 (-1.59%); split: -1.59%, +0.00%
Latency: 555706331 -> 550058635 (-1.02%); split: -1.07%, +0.05%
InvThroughput: 273023354 -> 271854469 (-0.43%); split: -0.44%, +0.01%
SClause: 385168 -> 385371 (+0.05%); split: -0.01%, +0.06%
Copies: 1342084 -> 1175762 (-12.39%); split: -12.40%, +0.01%
Branches: 392619 -> 378662 (-3.55%); split: -3.56%, +0.00%

Reviewed-by: Rhys Perry <pendingchaos02 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9196>

---

 src/amd/compiler/aco_lower_to_cssa.cpp | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/src/amd/compiler/aco_lower_to_cssa.cpp b/src/amd/compiler/aco_lower_to_cssa.cpp
index 3613df1c8c3..e1eec37c20d 100644
--- a/src/amd/compiler/aco_lower_to_cssa.cpp
+++ b/src/amd/compiler/aco_lower_to_cssa.cpp
@@ -78,10 +78,17 @@ void collect_parallelcopies(cssa_ctx& ctx)
              phi->opcode != aco_opcode::p_linear_phi)
             break;
 
+         const Definition& def = phi->definitions[0];
+
+         /* if the definition is not temp, it is the exec mask.
+          * We can reload the exec mask directly from the spill slot.
+          */
+         if (!def.isTemp())
+            continue;
+
          std::vector<unsigned>& preds = phi->opcode == aco_opcode::p_phi ?
                                         block.logical_preds :
                                         block.linear_preds;
-         const Definition& def = phi->definitions[0];
          uint32_t index = ctx.merge_sets.size();
          merge_set set;
 
@@ -91,6 +98,20 @@ void collect_parallelcopies(cssa_ctx& ctx)
             if (op.isUndefined())
                continue;
 
+            if (def.regClass().type() == RegType::sgpr && !op.isTemp()) {
+               /* SGPR inline constants and literals on GFX10+ can be spilled
+                * and reloaded directly (without intermediate register) */
+               if (op.isConstant()) {
+                  if (ctx.program->chip_class >= GFX10)
+                     continue;
+                  if (op.size() == 1 && !op.isLiteral())
+                     continue;
+               } else {
+                  assert(op.isFixed() && op.physReg() == exec);
+                  continue;
+               }
+            }
+
             /* create new temporary and rename operands */
             Temp tmp = bld.tmp(def.regClass());
             ctx.parallelcopies[preds[i]].emplace_back(copy{Definition(tmp), op});
@@ -107,6 +128,10 @@ void collect_parallelcopies(cssa_ctx& ctx)
 
             has_preheader_copy |= i == 0 && block.kind & block_kind_loop_header;
          }
+
+         if (set.empty())
+            continue;
+
          /* place the definition in dominance-order */
          if (def.isTemp()) {
             if (has_preheader_copy)