raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: propagate late constants through affected IR users

Browse Source
This commit is contained in:
nmzik
2026-10-12 18:17:24 +02:00
parent fa68e104ea
commit df08b83287
2 changed files with 196 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/shaderCfgTests.cpp
+81
View File
@@ -12534,6 +12534,86 @@ void TestLaneMaskProjection() {
}
}
void TestLateConstantPropagation() {
using namespace ShaderRecompiler::IR;
for (const uint32_t wave_size : {32u, 64u}) {
for (const bool reverse : {false, true}) {
Program program;
program.stage = ShaderType::Compute;
program.wave_size = wave_size;
for (uint32_t id = 0; id < 4; ++id) {
program.block_storage.push_back(std::make_unique<Block>());
program.blocks.push_back(program.block_storage.back().get());
program.blocks.back()->id = id;
}
auto *entry = program.blocks[0];
auto *header = program.blocks[1];
auto *latch = program.blocks[2];
auto *exit = program.blocks[3];
entry->AddBranch(header);
header->AddBranch(latch);
header->AddBranch(exit);
latch->AddBranch(header);
entry->terminator.kind = ShaderRecompiler::CFG::TerminatorKind::Branch;
entry->terminator.true_block = header;
header->terminator.kind = ShaderRecompiler::CFG::TerminatorKind::ConditionalBranch;
header->terminator.true_block = latch;
header->terminator.false_block = exit;
latch->terminator.kind = ShaderRecompiler::CFG::TerminatorKind::Branch;
latch->terminator.true_block = header;
IREmitter e(entry);
const auto raw = e.GetUserData(static_cast<ScalarReg>(0));
const auto active = e.INotEqual(raw, U32(Value(0u)));
header->condition = active;
auto &predicate = header->AppendNewInst(ValueOpcode::Phi);
predicate.SetFlags(Type::U1);
auto &mask = header->AppendNewInst(ValueOpcode::Phi);
mask.SetFlags(Type::U32);
const auto thread_bit = [](IREmitter &ir, U32 word) {
const auto lane = ir.BitwiseAnd(U32(ir.Emit(ValueOpcode::LaneId)), U32(Value(31u)));
return ir.INotEqual(ir.BitwiseAnd(ir.ShiftRightLogical(word, lane), U32(Value(1u))),
U32(Value(0u)));
};
IREmitter body(header);
const auto live = thread_bit(body, U32(Value(&mask)));
const auto selected = body.Select(live, U32(Value(42u)), raw);
const auto projected = body.Emit(ValueOpcode::SelectU1,
{Value(&predicate), live, Value(false)});
const auto absorbed = body.LogicalAnd(active,
body.LogicalOr(body.LogicalNot(active), U1(Value(&predicate))));
IREmitter tail(latch);
const auto full = thread_bit(tail, U32(Value(UINT32_MAX)));
const auto full_word = tail.Select(full, U32(Value(UINT32_MAX)), U32(Value(0u)));
predicate.AddPhiOperand(entry, Value(true));
predicate.AddPhiOperand(latch, full);
mask.AddPhiOperand(entry, Value(UINT32_MAX));
mask.AddPhiOperand(latch, full_word);
exit->condition = projected;
if (reverse) std::ranges::reverse(program.blocks);
ConstantPropagationPass(program.blocks, wave_size);
if (reverse) std::ranges::reverse(program.blocks);
Check(Value(&predicate).Resolve() == Value(true) &&
Value(&mask).Resolve() == Value(UINT32_MAX) && live.Resolve() == Value(true) &&
selected.Resolve() == Value(42u) && projected.Resolve() == Value(true) &&
absorbed.Resolve() == active.Resolve(),
"late constants did not reach Phi, nested mask patterns or new logical nodes");
ResolveControlFlowIdentities(program);
Check(exit->condition == Value(true), "late constant lost its control-flow use");
RemoveIdentities(program.blocks);
ValidateProgram(program, true);
}
}
Block cycle;
auto &phi = cycle.AppendNewInst(ValueOpcode::Phi);
phi.SetFlags(Type::U32);
phi.AddPhiOperand(&cycle, Value(&phi));
ConstantPropagationPass({&cycle});
Check(phi.GetOpcode() == ValueOpcode::Phi && phi.Arg(0).TryInstruction() == &phi,
"an ungrounded Phi cycle acquired a constant");
}
void TestU64ShiftConstantPropagation() {
using namespace ShaderRecompiler::IR;
@@ -15722,6 +15802,7 @@ int main() {
#endif
TestValuePhiValidation();
TestLaneMaskProjection();
TestLateConstantPropagation();
TestU64ShiftConstantPropagation();
TestU64RegisterPairRoundTrip();
#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS