aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/Target/AMDGPU/AsmParser
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib/Target/AMDGPU/AsmParser')
-rw-r--r--llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp93
1 files changed, 59 insertions, 34 deletions
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index a67a7be..a8140c3 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -1347,6 +1347,7 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
bool ForcedDPP = false;
bool ForcedSDWA = false;
KernelScopeInfo KernelScope;
+ const unsigned HwMode;
/// @name Auto-generated Match Functions
/// {
@@ -1356,6 +1357,13 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
/// }
+ /// Get size of register operand
+ unsigned getRegOperandSize(const MCInstrDesc &Desc, unsigned OpNo) const {
+ assert(OpNo < Desc.NumOperands);
+ int16_t RCID = MII.getOpRegClassID(Desc.operands()[OpNo], HwMode);
+ return getRegBitWidth(RCID) / 8;
+ }
+
private:
void createConstantSymbol(StringRef Id, int64_t Val);
@@ -1442,9 +1450,9 @@ public:
using OptionalImmIndexMap = std::map<AMDGPUOperand::ImmTy, unsigned>;
AMDGPUAsmParser(const MCSubtargetInfo &STI, MCAsmParser &_Parser,
- const MCInstrInfo &MII,
- const MCTargetOptions &Options)
- : MCTargetAsmParser(Options, STI, MII), Parser(_Parser) {
+ const MCInstrInfo &MII, const MCTargetOptions &Options)
+ : MCTargetAsmParser(Options, STI, MII), Parser(_Parser),
+ HwMode(STI.getHwMode(MCSubtargetInfo::HwMode_RegInfo)) {
MCAsmParserExtension::Initialize(Parser);
setAvailableFeatures(ComputeAvailableFeatures(getFeatureBits()));
@@ -1944,6 +1952,7 @@ public:
void cvtVOP3Interp(MCInst &Inst, const OperandVector &Operands);
void cvtVINTERP(MCInst &Inst, const OperandVector &Operands);
+ void cvtOpSelHelper(MCInst &Inst, unsigned OpSel);
bool parseDimId(unsigned &Encoding);
ParseStatus parseDim(OperandVector &Operands);
@@ -4106,7 +4115,7 @@ bool AMDGPUAsmParser::validateMIMGDataSize(const MCInst &Inst, SMLoc IDLoc) {
if ((DMaskIdx == -1 || TFEIdx == -1) && isGFX10_AEncoding()) // intersect_ray
return true;
- unsigned VDataSize = AMDGPU::getRegOperandSize(getMRI(), Desc, VDataIdx);
+ unsigned VDataSize = getRegOperandSize(Desc, VDataIdx);
unsigned TFESize = (TFEIdx != -1 && Inst.getOperand(TFEIdx).getImm()) ? 1 : 0;
unsigned DMask = Inst.getOperand(DMaskIdx).getImm() & 0xf;
if (DMask == 0)
@@ -4170,8 +4179,7 @@ bool AMDGPUAsmParser::validateMIMGAddrSize(const MCInst &Inst, SMLoc IDLoc) {
const AMDGPU::MIMGDimInfo *DimInfo = AMDGPU::getMIMGDimInfoByEncoding(Dim);
bool IsNSA = SrsrcIdx - VAddr0Idx > 1;
unsigned ActualAddrSize =
- IsNSA ? SrsrcIdx - VAddr0Idx
- : AMDGPU::getRegOperandSize(getMRI(), Desc, VAddr0Idx) / 4;
+ IsNSA ? SrsrcIdx - VAddr0Idx : getRegOperandSize(Desc, VAddr0Idx) / 4;
unsigned ExpectedAddrSize =
AMDGPU::getAddrSizeMIMGOp(BaseOpcode, DimInfo, IsA16, hasG16());
@@ -4181,8 +4189,7 @@ bool AMDGPUAsmParser::validateMIMGAddrSize(const MCInst &Inst, SMLoc IDLoc) {
ExpectedAddrSize >
getNSAMaxSize(Desc.TSFlags & SIInstrFlags::VSAMPLE)) {
int VAddrLastIdx = SrsrcIdx - 1;
- unsigned VAddrLastSize =
- AMDGPU::getRegOperandSize(getMRI(), Desc, VAddrLastIdx) / 4;
+ unsigned VAddrLastSize = getRegOperandSize(Desc, VAddrLastIdx) / 4;
ActualAddrSize = VAddrLastIdx - VAddr0Idx + VAddrLastSize;
}
@@ -4428,7 +4435,8 @@ bool AMDGPUAsmParser::validateMFMA(const MCInst &Inst,
return true;
const MCRegisterInfo *TRI = getContext().getRegisterInfo();
- if (TRI->getRegClass(Desc.operands()[0].RegClass).getSizeInBits() <= 128)
+ if (TRI->getRegClass(MII.getOpRegClassID(Desc.operands()[0], HwMode))
+ .getSizeInBits() <= 128)
return true;
if (TRI->regsOverlap(Src2Reg, DstReg)) {
@@ -4999,7 +5007,7 @@ bool AMDGPUAsmParser::validateDPP(const MCInst &Inst,
unsigned DppCtrl = Inst.getOperand(DppCtrlIdx).getImm();
if (!AMDGPU::isLegalDPALU_DPPControl(getSTI(), DppCtrl) &&
- AMDGPU::isDPALU_DPP(MII.get(Opc), getSTI())) {
+ AMDGPU::isDPALU_DPP(MII.get(Opc), MII, getSTI())) {
// DP ALU DPP is supported for row_newbcast only on GFX9* and row_share
// only on GFX12.
SMLoc S = getImmLoc(AMDGPUOperand::ImmTyDppCtrl, Operands);
@@ -5522,7 +5530,8 @@ bool AMDGPUAsmParser::validateWMMA(const MCInst &Inst,
unsigned Fmt = Inst.getOperand(FmtIdx).getImm();
int SrcIdx = AMDGPU::getNamedOperandIdx(Opc, SrcOp);
unsigned RegSize =
- TRI->getRegClass(Desc.operands()[SrcIdx].RegClass).getSizeInBits();
+ TRI->getRegClass(MII.getOpRegClassID(Desc.operands()[SrcIdx], HwMode))
+ .getSizeInBits();
if (RegSize == AMDGPU::wmmaScaleF8F6F4FormatToNumRegs(Fmt) * 32)
return true;
@@ -9239,6 +9248,33 @@ static bool isRegOrImmWithInputMods(const MCInstrDesc &Desc, unsigned OpNum) {
MCOI::OperandConstraint::TIED_TO) == -1;
}
+void AMDGPUAsmParser::cvtOpSelHelper(MCInst &Inst, unsigned OpSel) {
+ unsigned Opc = Inst.getOpcode();
+ constexpr AMDGPU::OpName Ops[] = {AMDGPU::OpName::src0, AMDGPU::OpName::src1,
+ AMDGPU::OpName::src2};
+ constexpr AMDGPU::OpName ModOps[] = {AMDGPU::OpName::src0_modifiers,
+ AMDGPU::OpName::src1_modifiers,
+ AMDGPU::OpName::src2_modifiers};
+ for (int J = 0; J < 3; ++J) {
+ int OpIdx = AMDGPU::getNamedOperandIdx(Opc, Ops[J]);
+ if (OpIdx == -1)
+ // Some instructions, e.g. v_interp_p2_f16 in GFX9, have src0, src2, but
+ // no src1. So continue instead of break.
+ continue;
+
+ int ModIdx = AMDGPU::getNamedOperandIdx(Opc, ModOps[J]);
+ uint32_t ModVal = Inst.getOperand(ModIdx).getImm();
+
+ if ((OpSel & (1 << J)) != 0)
+ ModVal |= SISrcMods::OP_SEL_0;
+ // op_sel[3] is encoded in src0_modifiers.
+ if (ModOps[J] == AMDGPU::OpName::src0_modifiers && (OpSel & (1 << 3)) != 0)
+ ModVal |= SISrcMods::DST_OP_SEL;
+
+ Inst.getOperand(ModIdx).setImm(ModVal);
+ }
+}
+
void AMDGPUAsmParser::cvtVOP3Interp(MCInst &Inst, const OperandVector &Operands)
{
OptionalImmIndexMap OptionalIdx;
@@ -9275,6 +9311,16 @@ void AMDGPUAsmParser::cvtVOP3Interp(MCInst &Inst, const OperandVector &Operands)
if (AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::omod))
addOptionalImmOperand(Inst, Operands, OptionalIdx,
AMDGPUOperand::ImmTyOModSI);
+
+ // Some v_interp instructions use op_sel[3] for dst.
+ if (AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel)) {
+ addOptionalImmOperand(Inst, Operands, OptionalIdx,
+ AMDGPUOperand::ImmTyOpSel);
+ int OpSelIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel);
+ unsigned OpSel = Inst.getOperand(OpSelIdx).getImm();
+
+ cvtOpSelHelper(Inst, OpSel);
+ }
}
void AMDGPUAsmParser::cvtVINTERP(MCInst &Inst, const OperandVector &Operands)
@@ -9310,31 +9356,10 @@ void AMDGPUAsmParser::cvtVINTERP(MCInst &Inst, const OperandVector &Operands)
if (OpSelIdx == -1)
return;
- const AMDGPU::OpName Ops[] = {AMDGPU::OpName::src0, AMDGPU::OpName::src1,
- AMDGPU::OpName::src2};
- const AMDGPU::OpName ModOps[] = {AMDGPU::OpName::src0_modifiers,
- AMDGPU::OpName::src1_modifiers,
- AMDGPU::OpName::src2_modifiers};
-
unsigned OpSel = Inst.getOperand(OpSelIdx).getImm();
-
- for (int J = 0; J < 3; ++J) {
- int OpIdx = AMDGPU::getNamedOperandIdx(Opc, Ops[J]);
- if (OpIdx == -1)
- break;
-
- int ModIdx = AMDGPU::getNamedOperandIdx(Opc, ModOps[J]);
- uint32_t ModVal = Inst.getOperand(ModIdx).getImm();
-
- if ((OpSel & (1 << J)) != 0)
- ModVal |= SISrcMods::OP_SEL_0;
- if (ModOps[J] == AMDGPU::OpName::src0_modifiers &&
- (OpSel & (1 << 3)) != 0)
- ModVal |= SISrcMods::DST_OP_SEL;
-
- Inst.getOperand(ModIdx).setImm(ModVal);
- }
+ cvtOpSelHelper(Inst, OpSel);
}
+
void AMDGPUAsmParser::cvtScaledMFMA(MCInst &Inst,
const OperandVector &Operands) {
OptionalImmIndexMap OptionalIdx;
generated by cgit v1.1 at 2025-10-08 07:22:46 +0000