diff options
| -rw-r--r-- | src/shader_recompiler/ir_opt/constant_propagation_pass.cpp | 32 |
1 files changed, 32 insertions, 0 deletions
diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp index 6285fef16f..0c0d4fb1a2 100644 --- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp +++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp @@ -953,6 +953,36 @@ void FoldDriverConstBuffer(Environment& env, IR::Block& block, IR::Inst& inst, u } } +void FoldSMax32(IR::Block& block, IR::Inst& inst) { + for (size_t index = 0; index < 2; ++index) { + const IR::Value value{inst.Arg(index).Resolve()}; + const IR::Inst* const negation{inst.Arg(1 - index).TryInstRecursive()}; + if (!negation || negation->GetOpcode() != IR::Opcode::INeg32 || + negation->Arg(0).Resolve() != value) { + continue; + } + const IR::Inst* const select{value.TryInstRecursive()}; + if (!select || select->GetOpcode() != IR::Opcode::SelectU32) { + continue; + } + const IR::Value true_value{select->Arg(1).Resolve()}; + const IR::Value false_value{select->Arg(2).Resolve()}; + const auto is_mask = [](const IR::Value& candidate) { + return candidate.IsImmediate() && + (candidate.U32() == 0 || candidate.U32() == 1 || + candidate.U32() == 0xffffffffU); + }; + if (!is_mask(true_value) || !is_mask(false_value)) { + continue; + } + // max(x, -x) is 0 or 1 for boolean and predicate-mask constants. + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + inst.ReplaceUsesWith(ir.Select(IR::U1{select->Arg(0)}, + ir.Imm32(true_value.U32() != 0 ? 1U : 0U), + ir.Imm32(false_value.U32() != 0 ? 1U : 0U))); + return; + } +} void ConstantPropagation(Environment& env, IR::Block& block, IR::Inst& inst) { switch (inst.GetOpcode()) { case IR::Opcode::GetRegister: @@ -963,6 +993,8 @@ void ConstantPropagation(Environment& env, IR::Block& block, IR::Inst& inst) { return FoldAdd<u32>(block, inst); case IR::Opcode::ISub32: return FoldISub32(inst); + case IR::Opcode::SMax32: + return FoldSMax32(block, inst); case IR::Opcode::IMul32: FoldWhenAllImmediates(inst, [](u32 a, u32 b) { return a * b; }); return; |