aco: keep loop live-through variables spilled

fossil-db (Navi):
Totals from 3149 (2.32% of 135946) affected shaders:
VGPRs: 280928 -> 280932 (+0.00%)
SpillSGPRs: 51133 -> 30042 (-41.25%)
CodeSize: 43063076 -> 41377252 (-3.91%); split: -3.92%, +0.00%
Instrs: 8278435 -> 8037133 (-2.91%); split: -2.92%, +0.00%
Cycles: 709575456 -> 683366172 (-3.69%); split: -3.69%, +0.00%
VMEM: 542887 -> 542937 (+0.01%); split: +0.05%, -0.04%
SMEM: 210255 -> 206368 (-1.85%); split: +0.12%, -1.97%
SClause: 258847 -> 258019 (-0.32%); split: -0.52%, +0.20%
Copies: 731836 -> 684784 (-6.43%); split: -6.44%, +0.01%
Branches: 305422 -> 292844 (-4.12%); split: -4.12%, +0.00%
PreSGPRs: 333103 -> 332701 (-0.12%)
PreVGPRs: 280086 -> 280089 (+0.00%)

Helps mostly Detroit: Become Human and the single spilling Doom Eternal
shader.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Cc: 20.2 <mesa-stable>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6212>

diff --git a/src/amd/compiler/aco_spill.cpp b/src/amd/compiler/aco_spill.cpp
index 94364bfa54ec1de365ce7fe404985b62844d1591..7d3055e33e097bc5917342af3161c4498a1bf0ae 100644 (file)
--- a/src/amd/compiler/aco_spill.cpp
+++ b/src/amd/compiler/aco_spill.cpp
@@ -383,6 +383,20 @@ RegisterDemand init_live_in_vars(spill_ctx& ctx, Block* block, unsigned block_id
       }
       unsigned loop_end = i;
 
+      /* keep live-through spilled */
+      for (std::pair<Temp, std::pair<uint32_t, uint32_t>> pair : ctx.next_use_distances_end[block_idx - 1]) {
+         if (pair.second.first < loop_end)
+            continue;
+
+         Temp to_spill = pair.first;
+         auto it = ctx.spills_exit[block_idx - 1].find(to_spill);
+         if (it == ctx.spills_exit[block_idx - 1].end())
+            continue;
+
+         ctx.spills_entry[block_idx][to_spill] = it->second;
+         spilled_registers += to_spill;
+      }
+
       /* select live-through vgpr variables */
       while (new_demand.vgpr - spilled_registers.vgpr > ctx.target_pressure.vgpr) {
          unsigned distance = 0;