AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aco: always use p_parallelcopy for pre-RA copies

Browse Source 
Most fossil-db changes are because literals are applied earlier
(in label_instruction), so use counts are more accurate and more literals
are applied.

fossil-db (Navi):
Totals from 79551 (57.89% of 137413) affected shaders:
SGPRs: 4549610 -> 4542802 (-0.15%); split: -0.19%, +0.04%
VGPRs: 3326764 -> 3324172 (-0.08%); split: -0.10%, +0.03%
SpillSGPRs: 38886 -> 34562 (-11.12%); split: -11.14%, +0.02%
CodeSize: 240143456 -> 240001008 (-0.06%); split: -0.11%, +0.05%
MaxWaves: 1078919 -> 1079281 (+0.03%); split: +0.04%, -0.01%
Instrs: 46627073 -> 46528490 (-0.21%); split: -0.22%, +0.01%

fossil-db (Polaris):
Totals from 98463 (70.90% of 138881) affected shaders:
SGPRs: 5164689 -> 5164353 (-0.01%); split: -0.02%, +0.01%
VGPRs: 3920936 -> 3921856 (+0.02%); split: -0.00%, +0.03%
SpillSGPRs: 56298 -> 52259 (-7.17%); split: -7.22%, +0.04%
CodeSize: 258680092 -> 258692712 (+0.00%); split: -0.02%, +0.03%
MaxWaves: 620863 -> 620823 (-0.01%); split: +0.00%, -0.01%
Instrs: 50776289 -> 50757577 (-0.04%); split: -0.04%, +0.00%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7216>
This commit is contained in:
Rhys Perry
2020-10-14 13:50:24 +01:00
committed by Marge Bot
parent 6db5fbf9f2
commit e54c111c45
6 changed files with 13 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/aco_optimizer.cpp
+3 -1
View File
@@ -853,7 +853,8 @@ void label_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
case aco_opcode::p_create_vector:
case aco_opcode::p_split_vector:
case aco_opcode::p_extract_vector:
case aco_opcode::p_phi: {
case aco_opcode::p_phi:
case aco_opcode::p_parallelcopy: {
const bool all_vgpr = std::none_of(instr->definitions.begin(), instr->definitions.end(),
[] (const Definition& def) { return def.getTemp().type() != RegType::vgpr;});
if (all_vgpr) {
@@ -1212,6 +1213,7 @@ void label_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
case aco_opcode::s_mov_b32: /* propagate */
case aco_opcode::s_mov_b64:
case aco_opcode::v_mov_b32:
case aco_opcode::p_parallelcopy:
if (instr->operands[0].isTemp() && ctx.info[instr->operands[0].tempId()].is_vec() &&
instr->operands[0].regClass() != instr->definitions[0].regClass()) {
/* We might not be able to copy-propagate if it's a SGPR->VGPR copy, so