fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-03 18:00:10 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

aco/spill: Replace vector<map> with vector<vector> for local_next_use

Browse source 
While adding/removing elements is faster with std::map, the cost of container
copies (and the involved memory allocations) vastly outweigh that benefit in
this usage pattern.

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11925>
This commit is contained in:
Tony Wasserka 2021-07-20 11:48:15 +02:00 • committed by Marge Bot
parent 92d7a6ab1c
commit 4453bce770
1 changed files with 17 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

23
src/amd/compiler/aco_spill.cpp
View file

@ -28,6 +28,7 @@
#include "common/sid.h"
#include <algorithm>
#include <map>
#include <set>
#include <stack>
@ -60,7 +61,7 @@ struct spill_ctx {
std::stack<Block*, std::vector<Block*>> loop_header;
std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_start;
std::vector<std::map<Temp, std::pair<uint32_t, uint32_t>>> next_use_distances_end;
std::vector<std::map<Temp, uint32_t>> local_next_use_distance; /* Working buffer */
std::vector<std::vector<std::pair<Temp, uint32_t>>> local_next_use_distance; /* Working buffer */
std::vector<std::pair<RegClass, std::unordered_set<uint32_t>>> interferences;
std::vector<std::vector<uint32_t>> affinities;
std::vector<bool> is_reloaded;
@ -358,7 +359,7 @@ get_rematerialize_info(spill_ctx& ctx)
void
update_local_next_uses(spill_ctx& ctx, Block* block,
std::vector<std::map<Temp, uint32_t>>& local_next_uses)
std::vector<std::vector<std::pair<Temp, uint32_t>>>& local_next_uses)
{
if (local_next_uses.size() < block->instructions.size()) {
/* Allocate more next-use-maps. Note that by never reducing the vector size, we enable
@ -369,8 +370,8 @@ update_local_next_uses(spill_ctx& ctx, Block* block,
local_next_uses[block->instructions.size() - 1].clear();
for (std::pair<const Temp, std::pair<uint32_t, uint32_t>>& pair :
ctx.next_use_distances_end[block->index]) {
local_next_uses[block->instructions.size() - 1].insert(
{pair.first, pair.second.second + block->instructions.size()});
local_next_uses[block->instructions.size() - 1].push_back(std::make_pair<Temp, uint32_t>(
(Temp)pair.first, pair.second.second + block->instructions.size()));
}
for (int idx = block->instructions.size() - 1; idx >= 0; idx--) {
@ -390,12 +391,22 @@ update_local_next_uses(spill_ctx& ctx, Block* block,
if (op.regClass().type() == RegType::vgpr && op.regClass().is_linear())
continue;
if (op.isTemp()) {
local_next_uses[idx][op.getTemp()] = idx;
auto it = std::find_if(local_next_uses[idx].begin(), local_next_uses[idx].end(),
[op](auto& pair) { return pair.first == op.getTemp(); });
if (it == local_next_uses[idx].end()) {
local_next_uses[idx].push_back(std::make_pair<Temp, uint32_t>(op.getTemp(), idx));
} else {
it->second = idx;
}
}
}
for (const Definition& def : instr->definitions) {
if (def.isTemp()) {
local_next_uses[idx].erase(def.getTemp());
auto it = std::find_if(local_next_uses[idx].begin(), local_next_uses[idx].end(),
[def](auto& pair) { return pair.first == def.getTemp(); });
if (it != local_next_uses[idx].end()) {
local_next_uses[idx].erase(it);
}
}
}
}