Revert "shader: preserve raw VCC bits in S_MOV_B64 (#418)"

This reverts commit 6758d67942.

An EXEC pair loses its mask provenance and falls into the raw high-word path. The saved all-ones high word then keeps waterfall branches true and wedges GPU work.
This commit is contained in:
nmzik
2026-09-01 07:32:18 +04:00
parent b8a459e83d
commit e591a66b67
13 changed files with 185 additions and 444 deletions
@@ -76,14 +76,6 @@ void IREmitter::SetExec(U1 value) {
Emit(ValueOpcode::SetExec, {value});
}
U1 IREmitter::GetExecMaskTag() {
return U1(Emit(ValueOpcode::GetExecMaskTag));
}
void IREmitter::SetExecMaskTag(U1 value) {
Emit(ValueOpcode::SetExecMaskTag, {value});
}
U32 IREmitter::GetExecLo() {
return U32(Emit(ValueOpcode::GetExecLo));
}
@@ -108,14 +100,6 @@ void IREmitter::SetVcc(U1 value) {
Emit(ValueOpcode::SetVcc, {value});
}
U1 IREmitter::GetVccMaskValidTag() {
return U1(Emit(ValueOpcode::GetVccMaskValidTag));
}
void IREmitter::SetVccMaskValidTag(U1 value) {
Emit(ValueOpcode::SetVccMaskValidTag, {value});
}
U32 IREmitter::GetVccLo() {
return U32(Emit(ValueOpcode::GetVccLo));
}
@@ -39,16 +39,12 @@ public:
void SetScc(U1 value);
U1 GetExec();
void SetExec(U1 value);
U1 GetExecMaskTag();
void SetExecMaskTag(U1 value);
U32 GetExecLo();
void SetExecLo(U32 value);
U32 GetExecHi();
void SetExecHi(U32 value);
U1 GetVcc();
void SetVcc(U1 value);
U1 GetVccMaskValidTag();
void SetVccMaskValidTag(U1 value);
U32 GetVccLo();
void SetVccLo(U32 value);
U32 GetVccHi();
@@ -33,16 +33,12 @@ bool IsRegisterStatePseudo(ValueOpcode opcode) {
case ValueOpcode::SetScc:
case ValueOpcode::GetExec:
case ValueOpcode::SetExec:
case ValueOpcode::GetExecMaskTag:
case ValueOpcode::SetExecMaskTag:
case ValueOpcode::GetExecLo:
case ValueOpcode::SetExecLo:
case ValueOpcode::GetExecHi:
case ValueOpcode::SetExecHi:
case ValueOpcode::GetVcc:
case ValueOpcode::SetVcc:
case ValueOpcode::GetVccMaskValidTag:
case ValueOpcode::SetVccMaskValidTag:
case ValueOpcode::GetVccLo:
case ValueOpcode::SetVccLo:
case ValueOpcode::GetVccHi:
@@ -23,16 +23,12 @@ VALUE_OPCODE(GetScc, U1)
VALUE_OPCODE(SetScc, Void, U1)
VALUE_OPCODE(GetExec, U1)
VALUE_OPCODE(SetExec, Void, U1)
VALUE_OPCODE(GetExecMaskTag, U1)
VALUE_OPCODE(SetExecMaskTag, Void, U1)
VALUE_OPCODE(GetExecLo, U32)
VALUE_OPCODE(SetExecLo, Void, U32)
VALUE_OPCODE(GetExecHi, U32)
VALUE_OPCODE(SetExecHi, Void, U32)
VALUE_OPCODE(GetVcc, U1)
VALUE_OPCODE(SetVcc, Void, U1)
VALUE_OPCODE(GetVccMaskValidTag, U1)
VALUE_OPCODE(SetVccMaskValidTag, Void, U1)
VALUE_OPCODE(GetVccLo, U32)
VALUE_OPCODE(SetVccLo, Void, U32)
VALUE_OPCODE(GetVccHi, U32)
@@ -22,9 +22,6 @@ struct ScalarMaskTag {
struct ExecTag {
auto operator<=>(const ExecTag&) const = default;
};
struct ExecMaskTag {
auto operator<=>(const ExecMaskTag&) const = default;
};
struct ExecLoTag {
auto operator<=>(const ExecLoTag&) const = default;
};
@@ -34,9 +31,6 @@ struct ExecHiTag {
struct VccTag {
auto operator<=>(const VccTag&) const = default;
};
struct VccMaskValidTag {
auto operator<=>(const VccMaskValidTag&) const = default;
};
struct VccLoTag {
auto operator<=>(const VccLoTag&) const = default;
};
@@ -51,9 +45,9 @@ struct GotoVariable {
auto operator<=>(const GotoVariable&) const = default;
};
using Variable = std::variant<ScalarReg, ThreadBitScalarReg, ScalarMaskTag, VectorReg, GotoVariable,
SccTag, ExecTag, ExecMaskTag, ExecLoTag, ExecHiTag, VccTag,
VccMaskValidTag, VccLoTag, VccHiTag, M0Tag>;
using Variable =
std::variant<ScalarReg, ThreadBitScalarReg, ScalarMaskTag, VectorReg, GotoVariable, SccTag,
ExecTag, ExecLoTag, ExecHiTag, VccTag, VccLoTag, VccHiTag, M0Tag>;
using ValueMap = std::unordered_map<Block*, Value>;
struct DefTable {
@@ -105,16 +99,12 @@ struct DefTable {
void Set(Block* block, SccTag, Value value) { scc[block] = value; }
const Value& Get(Block* block, ExecTag) { return exec[block]; }
void Set(Block* block, ExecTag, Value value) { exec[block] = value; }
const Value& Get(Block* block, ExecMaskTag) { return exec_mask_tag[block]; }
void Set(Block* block, ExecMaskTag, Value value) { exec_mask_tag[block] = value; }
const Value& Get(Block* block, ExecLoTag) { return exec_lo[block]; }
void Set(Block* block, ExecLoTag, Value value) { exec_lo[block] = value; }
const Value& Get(Block* block, ExecHiTag) { return exec_hi[block]; }
void Set(Block* block, ExecHiTag, Value value) { exec_hi[block] = value; }
const Value& Get(Block* block, VccTag) { return vcc[block]; }
void Set(Block* block, VccTag, Value value) { vcc[block] = value; }
const Value& Get(Block* block, VccMaskValidTag) { return vcc_mask_valid_tag[block]; }
void Set(Block* block, VccMaskValidTag, Value value) { vcc_mask_valid_tag[block] = value; }
const Value& Get(Block* block, VccLoTag) { return vcc_lo[block]; }
void Set(Block* block, VccLoTag, Value value) { vcc_lo[block] = value; }
const Value& Get(Block* block, VccHiTag) { return vcc_hi[block]; }
@@ -124,11 +114,9 @@ struct DefTable {
ValueMap scc;
ValueMap exec;
ValueMap exec_mask_tag;
ValueMap exec_lo;
ValueMap exec_hi;
ValueMap vcc;
ValueMap vcc_mask_valid_tag;
ValueMap vcc_lo;
ValueMap vcc_hi;
ValueMap m0;
@@ -156,9 +144,6 @@ ValueOpcode UndefOpcode(SccTag) {
ValueOpcode UndefOpcode(ExecTag) {
return ValueOpcode::UndefU1;
}
ValueOpcode UndefOpcode(ExecMaskTag) {
return ValueOpcode::UndefU1;
}
ValueOpcode UndefOpcode(ExecLoTag) {
return ValueOpcode::UndefU32;
}
@@ -168,9 +153,6 @@ ValueOpcode UndefOpcode(ExecHiTag) {
ValueOpcode UndefOpcode(VccTag) {
return ValueOpcode::UndefU1;
}
ValueOpcode UndefOpcode(VccMaskValidTag) {
return ValueOpcode::UndefU1;
}
ValueOpcode UndefOpcode(VccLoTag) {
return ValueOpcode::UndefU32;
}
@@ -202,9 +184,6 @@ Value InitialValue(SccTag) {
Value InitialValue(ExecTag) {
return Value(false);
}
Value InitialValue(ExecMaskTag) {
return Value(false);
}
Value InitialValue(ExecLoTag) {
return Value(0u);
}
@@ -214,9 +193,6 @@ Value InitialValue(ExecHiTag) {
Value InitialValue(VccTag) {
return Value(false);
}
Value InitialValue(VccMaskValidTag) {
return Value(false);
}
Value InitialValue(VccLoTag) {
return Value(0u);
}
@@ -375,15 +351,9 @@ void VisitInstruction(Pass& pass, Block* block, Inst& inst) {
break;
case ValueOpcode::SetScc: pass.Write(SccTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetExec: pass.Write(ExecTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetExecMaskTag:
pass.Write(ExecMaskTag {}, block, inst.Arg(0));
break;
case ValueOpcode::SetExecLo: pass.Write(ExecLoTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetExecHi: pass.Write(ExecHiTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetVcc: pass.Write(VccTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetVccMaskValidTag:
pass.Write(VccMaskValidTag {}, block, inst.Arg(0));
break;
case ValueOpcode::SetVccLo: pass.Write(VccLoTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetVccHi: pass.Write(VccHiTag {}, block, inst.Arg(0)); break;
case ValueOpcode::SetM0: pass.Write(M0Tag {}, block, inst.Arg(0)); break;
@@ -405,15 +375,9 @@ void VisitInstruction(Pass& pass, Block* block, Inst& inst) {
break;
case ValueOpcode::GetScc: inst.ReplaceUsesWith(pass.Read(SccTag {}, block)); break;
case ValueOpcode::GetExec: inst.ReplaceUsesWith(pass.Read(ExecTag {}, block)); break;
case ValueOpcode::GetExecMaskTag:
inst.ReplaceUsesWith(pass.Read(ExecMaskTag {}, block));
break;
case ValueOpcode::GetExecLo: inst.ReplaceUsesWith(pass.Read(ExecLoTag {}, block)); break;
case ValueOpcode::GetExecHi: inst.ReplaceUsesWith(pass.Read(ExecHiTag {}, block)); break;
case ValueOpcode::GetVcc: inst.ReplaceUsesWith(pass.Read(VccTag {}, block)); break;
case ValueOpcode::GetVccMaskValidTag:
inst.ReplaceUsesWith(pass.Read(VccMaskValidTag {}, block));
break;
case ValueOpcode::GetVccLo: inst.ReplaceUsesWith(pass.Read(VccLoTag {}, block)); break;
case ValueOpcode::GetVccHi: inst.ReplaceUsesWith(pass.Read(VccHiTag {}, block)); break;
case ValueOpcode::GetM0: inst.ReplaceUsesWith(pass.Read(M0Tag {}, block)); break;