shader: fold exact float class masks into native NaN predicates

This commit is contained in:
nmzik
2026-10-09 02:49:49 +02:00
parent 393257eaae
commit 7b9997baea
4 changed files with 63 additions and 5 deletions
@@ -263,10 +263,6 @@ uint32_t EmitUMedTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32
return EmitUMax32(state, EmitUMin32(state, arg0, arg1), high_min);
}
uint32_t EmitFPIsNan32(EmitterState& state, uint32_t arg0) {
return EmitNative<spv::OpFUnordNotEqual, IR::Type::U1>(state, arg0, arg0);
}
uint32_t EmitFPMin32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
return EmitMinMaxF32Value(state, arg0, arg1, false);
}
@@ -157,7 +157,7 @@ inline constexpr auto EmitFPOrdLessThanEqual32 = EmitFloatCompare32<spv::OpFOrdL
inline constexpr auto EmitFPUnordLessThanEqual32 = EmitFloatCompare32<spv::OpFUnordLessThanEqual>;
inline constexpr auto EmitFPOrdGreaterThanEqual32 = EmitFloatCompare32<spv::OpFOrdGreaterThanEqual>;
inline constexpr auto EmitFPUnordGreaterThanEqual32 = EmitFloatCompare32<spv::OpFUnordGreaterThanEqual>;
uint32_t EmitFPIsNan32(EmitterState& state, uint32_t arg0);
EMIT_NATIVE(FPIsNan32, OpIsNan, U1, uint32_t)
inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
inline constexpr auto EmitFPCmpClass16 = EmitClassMaskF16;
EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
@@ -215,6 +215,24 @@ void FoldInstruction(Block& block, Block::iterator instruction,
return;
case ValueOpcode::SelectU32:
case ValueOpcode::SelectF32: FoldSelect(inst); return;
case ValueOpcode::FPCmpClass32: {
const auto mask = Arg(inst, 1);
if (!IsImmediate(mask, Type::U32)) {
return;
}
switch (mask.U32() & 0x3ffu) {
case 0u: Replace(inst, Value(false)); break;
case 0x3ffu: Replace(inst, Value(true)); break;
case 3u: {
const auto result = block.PrependNewInst(
instruction, ValueOpcode::FPIsNan32, {Arg(inst, 0)});
Replace(inst, Value(&*result));
break;
}
default: break;
}
return;
}
case ValueOpcode::BitFieldInsert: {
const auto base = Arg(inst, 0);
const auto insert = Arg(inst, 1);