From cb81ec7a61fb41dd51bc3e1a22efffb4c1492086 Mon Sep 17 00:00:00 2001
From: Rhys Perry <pendingchaos02@gmail.com>
Date: Fri, 26 Apr 2024 11:52:13 +0100
Subject: [PATCH] aco: don't count certain pseudo towards
 VMEM_STORE_CLAUSE_MAX_GRAB_DIST
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

fossil-db (navi31):
Totals from 1023 (1.29% of 79395) affected shaders:
MaxWaves: 29258 -> 29240 (-0.06%)
Instrs: 1134024 -> 1133163 (-0.08%); split: -0.10%, +0.02%
CodeSize: 5682108 -> 5678696 (-0.06%); split: -0.08%, +0.02%
VGPRs: 60248 -> 60272 (+0.04%); split: -0.08%, +0.12%
Latency: 3797510 -> 3792797 (-0.12%); split: -0.18%, +0.05%
InvThroughput: 781270 -> 781239 (-0.00%); split: -0.03%, +0.03%
VClause: 24701 -> 23976 (-2.94%); split: -3.55%, +0.61%
Copies: 75177 -> 75169 (-0.01%); split: -0.20%, +0.19%
VALU: 659939 -> 659962 (+0.00%); split: -0.02%, +0.02%
VOPD: 2040 -> 2009 (-1.52%); split: +0.29%, -1.81%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28948>
---
 src/amd/compiler/aco_scheduler.cpp | 31 +++++++++++++++++++++++++++++-
 1 file changed, 30 insertions(+), 1 deletion(-)

diff --git a/src/amd/compiler/aco_scheduler.cpp b/src/amd/compiler/aco_scheduler.cpp
index 004d442b6f9..3bb119f5e43 100644
--- a/src/amd/compiler/aco_scheduler.cpp
+++ b/src/amd/compiler/aco_scheduler.cpp
@@ -665,6 +665,34 @@ perform_hazard_query(hazard_query* query, Instruction* instr, bool upwards)
    return hazard_success;
 }
 
+unsigned
+get_likely_cost(Instruction* instr)
+{
+   if (instr->opcode == aco_opcode::p_split_vector ||
+       instr->opcode == aco_opcode::p_extract_vector) {
+      unsigned cost = 0;
+      for (Definition def : instr->definitions) {
+         if (instr->operands[0].isKill() &&
+             def.regClass().type() == instr->operands[0].regClass().type())
+            continue;
+         cost += def.size();
+      }
+      return cost;
+   } else if (instr->opcode == aco_opcode::p_create_vector) {
+      unsigned cost = 0;
+      for (Operand op : instr->operands) {
+         if (op.isTemp() && op.isFirstKill() &&
+             op.regClass().type() == instr->definitions[0].regClass().type())
+            continue;
+         cost += op.size();
+      }
+      return cost;
+   } else {
+      /* For the moment, just assume the same cost for all other instructions. */
+      return 1;
+   }
+}
+
 void
 schedule_SMEM(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& register_demand,
               Instruction* current, int idx)
@@ -1118,7 +1146,7 @@ schedule_VMEM_store(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& r
    DownwardsCursor cursor = ctx.mv.downwards_init(idx, true, true);
    int skip = 0;
 
-   for (int i = 0; (i - skip) < VMEM_STORE_CLAUSE_MAX_GRAB_DIST; i++) {
+   for (int16_t k = 0; k < VMEM_STORE_CLAUSE_MAX_GRAB_DIST;) {
       aco_ptr<Instruction>& candidate = block->instructions[cursor.source_idx];
       if (candidate->opcode == aco_opcode::p_logical_start)
          break;
@@ -1126,6 +1154,7 @@ schedule_VMEM_store(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& r
       if (!should_form_clause(current, candidate.get())) {
          add_to_hazard_query(&hq, candidate.get());
          ctx.mv.downwards_skip(cursor);
+         k += get_likely_cost(candidate.get());
          continue;
       }