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

aco/ssa_elimination: create a single parallelcopy instruction for linear and logical phis

Browse Source 
Totals from 6651 (8.38% of 79377) affected shaders: (Navi31)

Instrs: 14722896 -> 14722290 (-0.00%); split: -0.01%, +0.00%
CodeSize: 77992072 -> 77989284 (-0.00%); split: -0.01%, +0.00%
Latency: 160542885 -> 160541215 (-0.00%); split: -0.00%, +0.00%
InvThroughput: 24543177 -> 24542710 (-0.00%); split: -0.00%, +0.00%
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33527>
This commit is contained in:
Daniel Schürmann
2025-02-12 12:25:46 +01:00
committed by Marge Bot
parent 0e98388614
commit ea765162c3
1 changed files with 13 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/aco_ssa_elimination.cpp
+13 -40
View File
@@ -18,16 +18,13 @@ struct phi_info {
};
struct ssa_elimination_ctx {
/* The outer vectors should be indexed by block index. The inner vectors store phi information
* for each block. */
std::vector<phi_info> logical_phi_info;
std::vector<phi_info> linear_phi_info;
/* The outer vectors should be indexed by block index.
* The inner vectors store phi information for each block.
*/
std::vector<phi_info> phi_infos;
Program* program;
ssa_elimination_ctx(Program* program_)
: logical_phi_info(program_->blocks.size()), linear_phi_info(program_->blocks.size()),
program(program_)
{}
ssa_elimination_ctx(Program* program_) : phi_infos(program_->blocks.size()), program(program_) {}
};
void
@@ -48,9 +45,7 @@ collect_phi_info(ssa_elimination_ctx& ctx)
Block::edge_vec& preds =
phi->opcode == aco_opcode::p_phi ? block.logical_preds : block.linear_preds;
uint32_t pred_idx = preds[i];
auto& info = phi->opcode == aco_opcode::p_phi ? ctx.logical_phi_info[pred_idx]
: ctx.linear_phi_info[pred_idx];
auto& info = ctx.phi_infos[preds[i]];
info.copies.emplace_back(phi->definitions[0], phi->operands[i]);
if (phi->pseudo().needs_scratch_reg) {
info.needs_scratch_reg = true;
@@ -64,45 +59,23 @@ collect_phi_info(ssa_elimination_ctx& ctx)
void
insert_parallelcopies(ssa_elimination_ctx& ctx)
{
/* insert the parallelcopies from logical phis before branch */
/* insert parallelcopies for the phis at the end of blocks just before the branch */
for (unsigned block_idx = 0; block_idx < ctx.program->blocks.size(); ++block_idx) {
auto& logical_phi_info = ctx.logical_phi_info[block_idx];
if (logical_phi_info.copies.empty())
auto& phi_info = ctx.phi_infos[block_idx];
if (phi_info.copies.empty())
continue;
Block& block = ctx.program->blocks[block_idx];
aco_ptr<Instruction> pc{create_instruction(aco_opcode::p_parallelcopy, Format::PSEUDO,
logical_phi_info.copies.size(),
logical_phi_info.copies.size())};
phi_info.copies.size(), phi_info.copies.size())};
unsigned i = 0;
for (auto& pair : logical_phi_info.copies) {
for (auto& pair : phi_info.copies) {
pc->definitions[i] = pair.first;
pc->operands[i] = pair.second;
i++;
}
pc->pseudo().needs_scratch_reg = false;
auto it = std::prev(block.instructions.end());
block.instructions.insert(it, std::move(pc));
}
/* insert parallelcopies for the linear phis at the end of blocks just before the branch */
for (unsigned block_idx = 0; block_idx < ctx.program->blocks.size(); ++block_idx) {
auto& linear_phi_info = ctx.linear_phi_info[block_idx];
if (linear_phi_info.copies.empty())
continue;
Block& block = ctx.program->blocks[block_idx];
aco_ptr<Instruction> pc{create_instruction(aco_opcode::p_parallelcopy, Format::PSEUDO,
linear_phi_info.copies.size(),
linear_phi_info.copies.size())};
unsigned i = 0;
for (auto& pair : linear_phi_info.copies) {
pc->definitions[i] = pair.first;
pc->operands[i] = pair.second;
i++;
}
pc->pseudo().scratch_sgpr = linear_phi_info.scratch_sgpr;
pc->pseudo().needs_scratch_reg = linear_phi_info.needs_scratch_reg;
pc->pseudo().scratch_sgpr = phi_info.scratch_sgpr;
pc->pseudo().needs_scratch_reg = phi_info.needs_scratch_reg;
auto it = std::prev(block.instructions.end());
block.instructions.insert(it, std::move(pc));
}