[compiler-rt] [AMDGPU] Change the representation of double literals in operands (PR #68740)
Stanislav Mekhanoshin via llvm-commits
llvm-commits at lists.llvm.org
Thu Oct 12 12:28:48 PDT 2023
https://github.com/rampitec updated https://github.com/llvm/llvm-project/pull/68740
>From cc9e065a9218eb36750a2c2a4a4d08fae3f329fa Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Wed, 4 Oct 2023 13:36:25 -0700
Subject: [PATCH 1/6] [AMDGPU] Change the representation of double literals in
operands
A 64-bit literal can be used as a 32-bit zero or sign extended
operand. In case of double zeroes are added to the low 32 bits.
Currently asm parser stores only high 32 bits of a double into
an operand. To support codegen as requested by the
https://github.com/llvm/llvm-project/issues/67781 we need to
change the representation to store a full 64-bit value so that
codegen can simply add immediates to an instruction.
There is some code to support compatibility with existing tests
and asm kernels. We allow to use short hex strings to represent
only a high 32 bit of a double value as a valid literal.
---
.../AMDGPU/AsmParser/AMDGPUAsmParser.cpp | 21 ++++++++++++--
.../Disassembler/AMDGPUDisassembler.cpp | 28 ++++++++++++++-----
.../AMDGPU/Disassembler/AMDGPUDisassembler.h | 9 ++++--
.../AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp | 12 +++++---
.../AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h | 2 +-
.../MCTargetDesc/AMDGPUMCCodeEmitter.cpp | 3 ++
llvm/lib/Target/AMDGPU/SIRegisterInfo.td | 4 ++-
.../Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | 7 +++++
llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h | 3 ++
9 files changed, 70 insertions(+), 19 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 35656bcaea1af7f..0553d3f20b21c56 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -2140,9 +2140,10 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
const_cast<AMDGPUAsmParser *>(AsmParser)->Warning(Inst.getLoc(),
"Can't encode literal as exact 64-bit floating-point operand. "
"Low 32-bits will be set to zero");
+ Val &= 0xffffffff00000000u;
}
- Inst.addOperand(MCOperand::createImm(Literal.lshr(32).getZExtValue()));
+ Inst.addOperand(MCOperand::createImm(Val));
setImmKindLiteral();
return;
}
@@ -2241,7 +2242,10 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
return;
}
- Inst.addOperand(MCOperand::createImm(Lo_32(Val)));
+ if (isInt<32>(Val) || isUInt<32>(Val))
+ Val = AMDGPU::isSISrcFPOperand(InstDesc, OpNum) ? Val << 32 : Lo_32(Val);
+
+ Inst.addOperand(MCOperand::createImm(Val));
setImmKindLiteral();
return;
@@ -4297,7 +4301,18 @@ bool AMDGPUAsmParser::validateVOPLiteral(const MCInst &Inst,
continue;
if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
- uint32_t Value = static_cast<uint32_t>(MO.getImm());
+ uint64_t Value = static_cast<uint64_t>(MO.getImm());
+ bool IsFP = AMDGPU::isSISrcFPOperand(Desc, OpIdx);
+ bool IsValid32Op = AMDGPU::isValid32BitLiteral(Value, IsFP);
+
+ if (!IsValid32Op && !isInt<32>(Value) && !isUInt<32>(Value)) {
+ Error(getLitLoc(Operands), "invalid operand for instruction");
+ return false;
+ }
+
+ if (IsFP && IsValid32Op)
+ Value = Hi_32(Value);
+
if (NumLiterals == 0 || LiteralValue != Value) {
LiteralValue = Value;
++NumLiterals;
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
index 439762bc6caf786..8c49c9a9c87772e 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
@@ -378,6 +378,15 @@ static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
}
+static DecodeStatus
+decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm, uint64_t Addr,
+ const MCDisassembler *Decoder) {
+ assert(Imm < (1 << 9) && "9-bit encoding");
+ auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
+ return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm,
+ false, 64, true));
+}
+
static DecodeStatus
DecodeAVLdSt_32RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
const MCDisassembler *Decoder) {
@@ -1218,7 +1227,7 @@ AMDGPUDisassembler::decodeMandatoryLiteralConstant(unsigned Val) const {
return MCOperand::createImm(Literal);
}
-MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
+MCOperand AMDGPUDisassembler::decodeLiteralConstant(bool ExtendFP64) const {
// For now all literal constants are supposed to be unsigned integer
// ToDo: deal with signed/unsigned 64-bit integer constants
// ToDo: deal with float/double constants
@@ -1228,9 +1237,11 @@ MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
Twine(Bytes.size()));
}
HasLiteral = true;
- Literal = eatBytes<uint32_t>(Bytes);
+ Literal = Literal64 = eatBytes<uint32_t>(Bytes);
+ if (ExtendFP64)
+ Literal64 <<= 32;
}
- return MCOperand::createImm(Literal);
+ return MCOperand::createImm(ExtendFP64 ? Literal64 : Literal);
}
MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
@@ -1447,7 +1458,8 @@ int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth) const {
+ unsigned ImmWidth,
+ bool IsFP) const {
using namespace AMDGPU::EncValues;
assert(Val < 1024); // enum10
@@ -1459,13 +1471,15 @@ MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
return createRegOperand(IsAGPR ? getAgprClassId(Width)
: getVgprClassId(Width), Val - VGPR_MIN);
}
- return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth);
+ return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth,
+ IsFP);
}
MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width,
unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth) const {
+ unsigned ImmWidth,
+ bool IsFP) const {
// Cases when Val{8} is 1 (vgpr, agpr or true 16 vgpr) should have been
// decoded earlier.
assert(Val < (1 << 8) && "9-bit Src encoding when Val{8} is 0");
@@ -1493,7 +1507,7 @@ MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width,
// Keep a sentinel value for deferred setting
return MCOperand::createImm(LITERAL_CONST);
else
- return decodeLiteralConstant();
+ return decodeLiteralConstant(IsFP && ImmWidth == 64);
}
switch (Width) {
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
index 5f3b277d577ff7c..865db2b26307b43 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
@@ -97,6 +97,7 @@ class AMDGPUDisassembler : public MCDisassembler {
const unsigned TargetMaxInstBytes;
mutable ArrayRef<uint8_t> Bytes;
mutable uint32_t Literal;
+ mutable uint64_t Literal64;
mutable bool HasLiteral;
mutable std::optional<bool> EnableWavefrontSize32;
@@ -229,15 +230,17 @@ class AMDGPUDisassembler : public MCDisassembler {
static MCOperand decodeFPImmed(unsigned ImmWidth, unsigned Imm);
MCOperand decodeMandatoryLiteralConstant(unsigned Imm) const;
- MCOperand decodeLiteralConstant() const;
+ MCOperand decodeLiteralConstant(bool ExtendFP64) const;
MCOperand decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0) const;
+ unsigned ImmWidth = 0,
+ bool IsFP = false) const;
MCOperand decodeNonVGPRSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0) const;
+ unsigned ImmWidth = 0,
+ bool IsFP = false) const;
MCOperand decodeVOPDDstYOp(MCInst &Inst, unsigned Val) const;
MCOperand decodeSpecialReg32(unsigned Val) const;
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
index ad4c48a8d65581a..40e92f00a9e52a6 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
@@ -426,7 +426,7 @@ void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
const MCSubtargetInfo &STI,
- raw_ostream &O) {
+ raw_ostream &O, bool IsFP) {
int64_t SImm = static_cast<int64_t>(Imm);
if (SImm >= -16 && SImm <= 64) {
O << SImm;
@@ -454,6 +454,8 @@ void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
else if (Imm == 0x3fc45f306dc9c882 &&
STI.hasFeature(AMDGPU::FeatureInv2PiInlineImm))
O << "0.15915494309189532";
+ else if (IsFP && AMDGPU::isValid32BitLiteral(Imm, true))
+ O << formatHex(static_cast<uint64_t>(Hi_32(Imm)));
else {
assert(isUInt<32>(Imm) || isInt<32>(Imm));
@@ -605,11 +607,13 @@ void AMDGPUInstPrinter::printRegularOperand(const MCInst *MI, unsigned OpNo,
printImmediate32(Op.getImm(), STI, O);
break;
case AMDGPU::OPERAND_REG_IMM_INT64:
- case AMDGPU::OPERAND_REG_IMM_FP64:
case AMDGPU::OPERAND_REG_INLINE_C_INT64:
+ printImmediate64(Op.getImm(), STI, O, false);
+ break;
+ case AMDGPU::OPERAND_REG_IMM_FP64:
case AMDGPU::OPERAND_REG_INLINE_C_FP64:
case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
- printImmediate64(Op.getImm(), STI, O);
+ printImmediate64(Op.getImm(), STI, O, true);
break;
case AMDGPU::OPERAND_REG_INLINE_C_INT16:
case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
@@ -671,7 +675,7 @@ void AMDGPUInstPrinter::printRegularOperand(const MCInst *MI, unsigned OpNo,
if (RCBits == 32)
printImmediate32(llvm::bit_cast<uint32_t>((float)Value), STI, O);
else if (RCBits == 64)
- printImmediate64(llvm::bit_cast<uint64_t>(Value), STI, O);
+ printImmediate64(llvm::bit_cast<uint64_t>(Value), STI, O, true);
else
llvm_unreachable("Invalid register class size");
}
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
index 3b14faab136b35a..dc83547a4afe049 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
@@ -91,7 +91,7 @@ class AMDGPUInstPrinter : public MCInstPrinter {
void printImmediate32(uint32_t Imm, const MCSubtargetInfo &STI,
raw_ostream &O);
void printImmediate64(uint64_t Imm, const MCSubtargetInfo &STI,
- raw_ostream &O);
+ raw_ostream &O, bool IsFP);
void printOperand(const MCInst *MI, unsigned OpNo, const MCSubtargetInfo &STI,
raw_ostream &O);
void printRegularOperand(const MCInst *MI, unsigned OpNo,
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
index 57ccb523c70eee6..d93f747bf6f0a64 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
@@ -411,6 +411,9 @@ void AMDGPUMCCodeEmitter::encodeInstruction(const MCInst &MI,
} else if (!Op.isExpr()) // Exprs will be replaced with a fixup value.
llvm_unreachable("Must be immediate or expr");
+ if (Desc.operands()[i].OperandType == AMDGPU::OPERAND_REG_IMM_FP64)
+ Imm = Hi_32(Imm);
+
support::endian::write<uint32_t>(CB, Imm, support::endianness::little);
// Only one literal value allowed
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
index c3c5bfae405aa45..ea06e85fb400c1b 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
@@ -1263,7 +1263,9 @@ def VSrc_f32 : RegOrF32 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_v2b16 : RegOrV2B16 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_v2f16 : RegOrV2F16 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_b64 : RegOrB64 <"VS_64", "OPERAND_REG_IMM">;
-def VSrc_f64 : RegOrF64 <"VS_64", "OPERAND_REG_IMM">;
+def VSrc_f64 : RegOrF64 <"VS_64", "OPERAND_REG_IMM"> {
+ let DecoderMethod = "decodeOperand_VSrc_f64";
+}
def VSrc_v2b32 : RegOrV2B32 <"VS_64", "OPERAND_REG_IMM">;
def VSrc_v2f32 : RegOrV2F32 <"VS_64", "OPERAND_REG_IMM">;
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 6d0ad763d9e6cc1..e7907b28abedf9d 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -2519,6 +2519,13 @@ bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi) {
return Lo16 == Hi16;
}
+bool isValid32BitLiteral(uint64_t Val, bool IsFP) {
+ if (IsFP)
+ return !(Val & 0xffffffffu);
+
+ return isUInt<32>(Val) || isInt<32>(Val);
+}
+
bool isArgPassedInSGPR(const Argument *A) {
const Function *F = A->getParent();
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index 297a69f54d63721..fbe9adfd74fa9c6 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -1290,6 +1290,9 @@ bool isInlinableIntLiteralV216(int32_t Literal);
LLVM_READNONE
bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi);
+LLVM_READNONE
+bool isValid32BitLiteral(uint64_t Val, bool IsFP);
+
bool isArgPassedInSGPR(const Argument *Arg);
bool isArgPassedInSGPR(const CallBase *CB, unsigned ArgNo);
>From b74ad67d36f6832fe582ab33080fdda4e5da5408 Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Wed, 4 Oct 2023 13:36:25 -0700
Subject: [PATCH 2/6] [AMDGPU] Change the representation of double literals in
operands
A 64-bit literal can be used as a 32-bit zero or sign extended
operand. In case of double zeroes are added to the low 32 bits.
Currently asm parser stores only high 32 bits of a double into
an operand. To support codegen as requested by the
https://github.com/llvm/llvm-project/issues/67781 we need to
change the representation to store a full 64-bit value so that
codegen can simply add immediates to an instruction.
There is some code to support compatibility with existing tests
and asm kernels. We allow to use short hex strings to represent
only a high 32 bit of a double value as a valid literal.
---
.../AMDGPU/AsmParser/AMDGPUAsmParser.cpp | 21 ++++++++++++--
.../Disassembler/AMDGPUDisassembler.cpp | 28 ++++++++++++++-----
.../AMDGPU/Disassembler/AMDGPUDisassembler.h | 9 ++++--
.../AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp | 12 +++++---
.../AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h | 2 +-
.../MCTargetDesc/AMDGPUMCCodeEmitter.cpp | 5 +++-
llvm/lib/Target/AMDGPU/SIRegisterInfo.td | 4 ++-
.../Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | 7 +++++
llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h | 3 ++
9 files changed, 71 insertions(+), 20 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 1e07e8deb560fcb..253a2e98f0cb685 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -2141,9 +2141,10 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
const_cast<AMDGPUAsmParser *>(AsmParser)->Warning(Inst.getLoc(),
"Can't encode literal as exact 64-bit floating-point operand. "
"Low 32-bits will be set to zero");
+ Val &= 0xffffffff00000000u;
}
- Inst.addOperand(MCOperand::createImm(Literal.lshr(32).getZExtValue()));
+ Inst.addOperand(MCOperand::createImm(Val));
setImmKindLiteral();
return;
}
@@ -2242,7 +2243,10 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
return;
}
- Inst.addOperand(MCOperand::createImm(Lo_32(Val)));
+ if (isInt<32>(Val) || isUInt<32>(Val))
+ Val = AMDGPU::isSISrcFPOperand(InstDesc, OpNum) ? Val << 32 : Lo_32(Val);
+
+ Inst.addOperand(MCOperand::createImm(Val));
setImmKindLiteral();
return;
@@ -4309,7 +4313,18 @@ bool AMDGPUAsmParser::validateVOPLiteral(const MCInst &Inst,
continue;
if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
- uint32_t Value = static_cast<uint32_t>(MO.getImm());
+ uint64_t Value = static_cast<uint64_t>(MO.getImm());
+ bool IsFP = AMDGPU::isSISrcFPOperand(Desc, OpIdx);
+ bool IsValid32Op = AMDGPU::isValid32BitLiteral(Value, IsFP);
+
+ if (!IsValid32Op && !isInt<32>(Value) && !isUInt<32>(Value)) {
+ Error(getLitLoc(Operands), "invalid operand for instruction");
+ return false;
+ }
+
+ if (IsFP && IsValid32Op)
+ Value = Hi_32(Value);
+
if (NumLiterals == 0 || LiteralValue != Value) {
LiteralValue = Value;
++NumLiterals;
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
index a504a5e86760bd6..83d973dc62e7770 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
@@ -378,6 +378,15 @@ static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
}
+static DecodeStatus
+decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm, uint64_t Addr,
+ const MCDisassembler *Decoder) {
+ assert(Imm < (1 << 9) && "9-bit encoding");
+ auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
+ return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm,
+ false, 64, true));
+}
+
static DecodeStatus
DecodeAVLdSt_32RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
const MCDisassembler *Decoder) {
@@ -1219,7 +1228,7 @@ AMDGPUDisassembler::decodeMandatoryLiteralConstant(unsigned Val) const {
return MCOperand::createImm(Literal);
}
-MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
+MCOperand AMDGPUDisassembler::decodeLiteralConstant(bool ExtendFP64) const {
// For now all literal constants are supposed to be unsigned integer
// ToDo: deal with signed/unsigned 64-bit integer constants
// ToDo: deal with float/double constants
@@ -1229,9 +1238,11 @@ MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
Twine(Bytes.size()));
}
HasLiteral = true;
- Literal = eatBytes<uint32_t>(Bytes);
+ Literal = Literal64 = eatBytes<uint32_t>(Bytes);
+ if (ExtendFP64)
+ Literal64 <<= 32;
}
- return MCOperand::createImm(Literal);
+ return MCOperand::createImm(ExtendFP64 ? Literal64 : Literal);
}
MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
@@ -1448,7 +1459,8 @@ int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth) const {
+ unsigned ImmWidth,
+ bool IsFP) const {
using namespace AMDGPU::EncValues;
assert(Val < 1024); // enum10
@@ -1460,13 +1472,15 @@ MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
return createRegOperand(IsAGPR ? getAgprClassId(Width)
: getVgprClassId(Width), Val - VGPR_MIN);
}
- return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth);
+ return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth,
+ IsFP);
}
MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width,
unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth) const {
+ unsigned ImmWidth,
+ bool IsFP) const {
// Cases when Val{8} is 1 (vgpr, agpr or true 16 vgpr) should have been
// decoded earlier.
assert(Val < (1 << 8) && "9-bit Src encoding when Val{8} is 0");
@@ -1494,7 +1508,7 @@ MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width,
// Keep a sentinel value for deferred setting
return MCOperand::createImm(LITERAL_CONST);
else
- return decodeLiteralConstant();
+ return decodeLiteralConstant(IsFP && ImmWidth == 64);
}
switch (Width) {
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
index 5f3b277d577ff7c..865db2b26307b43 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
@@ -97,6 +97,7 @@ class AMDGPUDisassembler : public MCDisassembler {
const unsigned TargetMaxInstBytes;
mutable ArrayRef<uint8_t> Bytes;
mutable uint32_t Literal;
+ mutable uint64_t Literal64;
mutable bool HasLiteral;
mutable std::optional<bool> EnableWavefrontSize32;
@@ -229,15 +230,17 @@ class AMDGPUDisassembler : public MCDisassembler {
static MCOperand decodeFPImmed(unsigned ImmWidth, unsigned Imm);
MCOperand decodeMandatoryLiteralConstant(unsigned Imm) const;
- MCOperand decodeLiteralConstant() const;
+ MCOperand decodeLiteralConstant(bool ExtendFP64) const;
MCOperand decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0) const;
+ unsigned ImmWidth = 0,
+ bool IsFP = false) const;
MCOperand decodeNonVGPRSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0) const;
+ unsigned ImmWidth = 0,
+ bool IsFP = false) const;
MCOperand decodeVOPDDstYOp(MCInst &Inst, unsigned Val) const;
MCOperand decodeSpecialReg32(unsigned Val) const;
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
index ad4c48a8d65581a..40e92f00a9e52a6 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
@@ -426,7 +426,7 @@ void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
const MCSubtargetInfo &STI,
- raw_ostream &O) {
+ raw_ostream &O, bool IsFP) {
int64_t SImm = static_cast<int64_t>(Imm);
if (SImm >= -16 && SImm <= 64) {
O << SImm;
@@ -454,6 +454,8 @@ void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
else if (Imm == 0x3fc45f306dc9c882 &&
STI.hasFeature(AMDGPU::FeatureInv2PiInlineImm))
O << "0.15915494309189532";
+ else if (IsFP && AMDGPU::isValid32BitLiteral(Imm, true))
+ O << formatHex(static_cast<uint64_t>(Hi_32(Imm)));
else {
assert(isUInt<32>(Imm) || isInt<32>(Imm));
@@ -605,11 +607,13 @@ void AMDGPUInstPrinter::printRegularOperand(const MCInst *MI, unsigned OpNo,
printImmediate32(Op.getImm(), STI, O);
break;
case AMDGPU::OPERAND_REG_IMM_INT64:
- case AMDGPU::OPERAND_REG_IMM_FP64:
case AMDGPU::OPERAND_REG_INLINE_C_INT64:
+ printImmediate64(Op.getImm(), STI, O, false);
+ break;
+ case AMDGPU::OPERAND_REG_IMM_FP64:
case AMDGPU::OPERAND_REG_INLINE_C_FP64:
case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
- printImmediate64(Op.getImm(), STI, O);
+ printImmediate64(Op.getImm(), STI, O, true);
break;
case AMDGPU::OPERAND_REG_INLINE_C_INT16:
case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
@@ -671,7 +675,7 @@ void AMDGPUInstPrinter::printRegularOperand(const MCInst *MI, unsigned OpNo,
if (RCBits == 32)
printImmediate32(llvm::bit_cast<uint32_t>((float)Value), STI, O);
else if (RCBits == 64)
- printImmediate64(llvm::bit_cast<uint64_t>(Value), STI, O);
+ printImmediate64(llvm::bit_cast<uint64_t>(Value), STI, O, true);
else
llvm_unreachable("Invalid register class size");
}
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
index 3b14faab136b35a..dc83547a4afe049 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
@@ -91,7 +91,7 @@ class AMDGPUInstPrinter : public MCInstPrinter {
void printImmediate32(uint32_t Imm, const MCSubtargetInfo &STI,
raw_ostream &O);
void printImmediate64(uint64_t Imm, const MCSubtargetInfo &STI,
- raw_ostream &O);
+ raw_ostream &O, bool IsFP);
void printOperand(const MCInst *MI, unsigned OpNo, const MCSubtargetInfo &STI,
raw_ostream &O);
void printRegularOperand(const MCInst *MI, unsigned OpNo,
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
index 21243f80e055499..d93f747bf6f0a64 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
@@ -411,7 +411,10 @@ void AMDGPUMCCodeEmitter::encodeInstruction(const MCInst &MI,
} else if (!Op.isExpr()) // Exprs will be replaced with a fixup value.
llvm_unreachable("Must be immediate or expr");
- support::endian::write<uint32_t>(CB, Imm, llvm::endianness::little);
+ if (Desc.operands()[i].OperandType == AMDGPU::OPERAND_REG_IMM_FP64)
+ Imm = Hi_32(Imm);
+
+ support::endian::write<uint32_t>(CB, Imm, support::endianness::little);
// Only one literal value allowed
break;
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
index c3c5bfae405aa45..ea06e85fb400c1b 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
@@ -1263,7 +1263,9 @@ def VSrc_f32 : RegOrF32 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_v2b16 : RegOrV2B16 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_v2f16 : RegOrV2F16 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_b64 : RegOrB64 <"VS_64", "OPERAND_REG_IMM">;
-def VSrc_f64 : RegOrF64 <"VS_64", "OPERAND_REG_IMM">;
+def VSrc_f64 : RegOrF64 <"VS_64", "OPERAND_REG_IMM"> {
+ let DecoderMethod = "decodeOperand_VSrc_f64";
+}
def VSrc_v2b32 : RegOrV2B32 <"VS_64", "OPERAND_REG_IMM">;
def VSrc_v2f32 : RegOrV2F32 <"VS_64", "OPERAND_REG_IMM">;
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 6d0ad763d9e6cc1..e7907b28abedf9d 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -2519,6 +2519,13 @@ bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi) {
return Lo16 == Hi16;
}
+bool isValid32BitLiteral(uint64_t Val, bool IsFP) {
+ if (IsFP)
+ return !(Val & 0xffffffffu);
+
+ return isUInt<32>(Val) || isInt<32>(Val);
+}
+
bool isArgPassedInSGPR(const Argument *A) {
const Function *F = A->getParent();
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index 297a69f54d63721..fbe9adfd74fa9c6 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -1290,6 +1290,9 @@ bool isInlinableIntLiteralV216(int32_t Literal);
LLVM_READNONE
bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi);
+LLVM_READNONE
+bool isValid32BitLiteral(uint64_t Val, bool IsFP);
+
bool isArgPassedInSGPR(const Argument *Arg);
bool isArgPassedInSGPR(const CallBase *CB, unsigned ArgNo);
>From 6e7dffa80693b714a15c644c3f8df71d2ac14be8 Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Thu, 12 Oct 2023 01:16:56 -0700
Subject: [PATCH 3/6] [AMDGPU] Make clang-format happy with disasm cganges
---
.../AMDGPU/Disassembler/AMDGPUDisassembler.cpp | 13 ++++++-------
.../Target/AMDGPU/Disassembler/AMDGPUDisassembler.h | 6 ++----
2 files changed, 8 insertions(+), 11 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
index 83d973dc62e7770..d74fd0b3a9ea74e 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
@@ -378,13 +378,13 @@ static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
}
-static DecodeStatus
-decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm, uint64_t Addr,
- const MCDisassembler *Decoder) {
+static DecodeStatus decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm,
+ uint64_t Addr,
+ const MCDisassembler *Decoder) {
assert(Imm < (1 << 9) && "9-bit encoding");
auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
- return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm,
- false, 64, true));
+ return addOperand(
+ Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm, false, 64, true));
}
static DecodeStatus
@@ -1459,8 +1459,7 @@ int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth,
- bool IsFP) const {
+ unsigned ImmWidth, bool IsFP) const {
using namespace AMDGPU::EncValues;
assert(Val < 1024); // enum10
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
index 865db2b26307b43..91b73b593d61617 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
@@ -233,14 +233,12 @@ class AMDGPUDisassembler : public MCDisassembler {
MCOperand decodeLiteralConstant(bool ExtendFP64) const;
MCOperand decodeSrcOp(const OpWidthTy Width, unsigned Val,
- bool MandatoryLiteral = false,
- unsigned ImmWidth = 0,
+ bool MandatoryLiteral = false, unsigned ImmWidth = 0,
bool IsFP = false) const;
MCOperand decodeNonVGPRSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0,
- bool IsFP = false) const;
+ unsigned ImmWidth = 0, bool IsFP = false) const;
MCOperand decodeVOPDDstYOp(MCInst &Inst, unsigned Val) const;
MCOperand decodeSpecialReg32(unsigned Val) const;
>From 88e9ea98cd49c20b188466bcf518218940405b11 Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Wed, 4 Oct 2023 13:36:25 -0700
Subject: [PATCH 4/6] [AMDGPU] Change the representation of double literals in
operands
A 64-bit literal can be used as a 32-bit zero or sign extended
operand. In case of double zeroes are added to the low 32 bits.
Currently asm parser stores only high 32 bits of a double into
an operand. To support codegen as requested by the
https://github.com/llvm/llvm-project/issues/67781 we need to
change the representation to store a full 64-bit value so that
codegen can simply add immediates to an instruction.
There is some code to support compatibility with existing tests
and asm kernels. We allow to use short hex strings to represent
only a high 32 bit of a double value as a valid literal.
---
.../AMDGPU/AsmParser/AMDGPUAsmParser.cpp | 21 ++++++++++++--
.../Disassembler/AMDGPUDisassembler.cpp | 28 ++++++++++++++-----
.../AMDGPU/Disassembler/AMDGPUDisassembler.h | 9 ++++--
.../AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp | 12 +++++---
.../AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h | 2 +-
.../MCTargetDesc/AMDGPUMCCodeEmitter.cpp | 5 +++-
llvm/lib/Target/AMDGPU/SIRegisterInfo.td | 4 ++-
.../Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | 7 +++++
llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h | 3 ++
9 files changed, 71 insertions(+), 20 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 1e07e8deb560fcb..253a2e98f0cb685 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -2141,9 +2141,10 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
const_cast<AMDGPUAsmParser *>(AsmParser)->Warning(Inst.getLoc(),
"Can't encode literal as exact 64-bit floating-point operand. "
"Low 32-bits will be set to zero");
+ Val &= 0xffffffff00000000u;
}
- Inst.addOperand(MCOperand::createImm(Literal.lshr(32).getZExtValue()));
+ Inst.addOperand(MCOperand::createImm(Val));
setImmKindLiteral();
return;
}
@@ -2242,7 +2243,10 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
return;
}
- Inst.addOperand(MCOperand::createImm(Lo_32(Val)));
+ if (isInt<32>(Val) || isUInt<32>(Val))
+ Val = AMDGPU::isSISrcFPOperand(InstDesc, OpNum) ? Val << 32 : Lo_32(Val);
+
+ Inst.addOperand(MCOperand::createImm(Val));
setImmKindLiteral();
return;
@@ -4309,7 +4313,18 @@ bool AMDGPUAsmParser::validateVOPLiteral(const MCInst &Inst,
continue;
if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
- uint32_t Value = static_cast<uint32_t>(MO.getImm());
+ uint64_t Value = static_cast<uint64_t>(MO.getImm());
+ bool IsFP = AMDGPU::isSISrcFPOperand(Desc, OpIdx);
+ bool IsValid32Op = AMDGPU::isValid32BitLiteral(Value, IsFP);
+
+ if (!IsValid32Op && !isInt<32>(Value) && !isUInt<32>(Value)) {
+ Error(getLitLoc(Operands), "invalid operand for instruction");
+ return false;
+ }
+
+ if (IsFP && IsValid32Op)
+ Value = Hi_32(Value);
+
if (NumLiterals == 0 || LiteralValue != Value) {
LiteralValue = Value;
++NumLiterals;
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
index a504a5e86760bd6..83d973dc62e7770 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
@@ -378,6 +378,15 @@ static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
}
+static DecodeStatus
+decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm, uint64_t Addr,
+ const MCDisassembler *Decoder) {
+ assert(Imm < (1 << 9) && "9-bit encoding");
+ auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
+ return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm,
+ false, 64, true));
+}
+
static DecodeStatus
DecodeAVLdSt_32RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
const MCDisassembler *Decoder) {
@@ -1219,7 +1228,7 @@ AMDGPUDisassembler::decodeMandatoryLiteralConstant(unsigned Val) const {
return MCOperand::createImm(Literal);
}
-MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
+MCOperand AMDGPUDisassembler::decodeLiteralConstant(bool ExtendFP64) const {
// For now all literal constants are supposed to be unsigned integer
// ToDo: deal with signed/unsigned 64-bit integer constants
// ToDo: deal with float/double constants
@@ -1229,9 +1238,11 @@ MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
Twine(Bytes.size()));
}
HasLiteral = true;
- Literal = eatBytes<uint32_t>(Bytes);
+ Literal = Literal64 = eatBytes<uint32_t>(Bytes);
+ if (ExtendFP64)
+ Literal64 <<= 32;
}
- return MCOperand::createImm(Literal);
+ return MCOperand::createImm(ExtendFP64 ? Literal64 : Literal);
}
MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
@@ -1448,7 +1459,8 @@ int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth) const {
+ unsigned ImmWidth,
+ bool IsFP) const {
using namespace AMDGPU::EncValues;
assert(Val < 1024); // enum10
@@ -1460,13 +1472,15 @@ MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
return createRegOperand(IsAGPR ? getAgprClassId(Width)
: getVgprClassId(Width), Val - VGPR_MIN);
}
- return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth);
+ return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth,
+ IsFP);
}
MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width,
unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth) const {
+ unsigned ImmWidth,
+ bool IsFP) const {
// Cases when Val{8} is 1 (vgpr, agpr or true 16 vgpr) should have been
// decoded earlier.
assert(Val < (1 << 8) && "9-bit Src encoding when Val{8} is 0");
@@ -1494,7 +1508,7 @@ MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width,
// Keep a sentinel value for deferred setting
return MCOperand::createImm(LITERAL_CONST);
else
- return decodeLiteralConstant();
+ return decodeLiteralConstant(IsFP && ImmWidth == 64);
}
switch (Width) {
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
index 5f3b277d577ff7c..865db2b26307b43 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
@@ -97,6 +97,7 @@ class AMDGPUDisassembler : public MCDisassembler {
const unsigned TargetMaxInstBytes;
mutable ArrayRef<uint8_t> Bytes;
mutable uint32_t Literal;
+ mutable uint64_t Literal64;
mutable bool HasLiteral;
mutable std::optional<bool> EnableWavefrontSize32;
@@ -229,15 +230,17 @@ class AMDGPUDisassembler : public MCDisassembler {
static MCOperand decodeFPImmed(unsigned ImmWidth, unsigned Imm);
MCOperand decodeMandatoryLiteralConstant(unsigned Imm) const;
- MCOperand decodeLiteralConstant() const;
+ MCOperand decodeLiteralConstant(bool ExtendFP64) const;
MCOperand decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0) const;
+ unsigned ImmWidth = 0,
+ bool IsFP = false) const;
MCOperand decodeNonVGPRSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0) const;
+ unsigned ImmWidth = 0,
+ bool IsFP = false) const;
MCOperand decodeVOPDDstYOp(MCInst &Inst, unsigned Val) const;
MCOperand decodeSpecialReg32(unsigned Val) const;
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
index ad4c48a8d65581a..40e92f00a9e52a6 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
@@ -426,7 +426,7 @@ void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
const MCSubtargetInfo &STI,
- raw_ostream &O) {
+ raw_ostream &O, bool IsFP) {
int64_t SImm = static_cast<int64_t>(Imm);
if (SImm >= -16 && SImm <= 64) {
O << SImm;
@@ -454,6 +454,8 @@ void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
else if (Imm == 0x3fc45f306dc9c882 &&
STI.hasFeature(AMDGPU::FeatureInv2PiInlineImm))
O << "0.15915494309189532";
+ else if (IsFP && AMDGPU::isValid32BitLiteral(Imm, true))
+ O << formatHex(static_cast<uint64_t>(Hi_32(Imm)));
else {
assert(isUInt<32>(Imm) || isInt<32>(Imm));
@@ -605,11 +607,13 @@ void AMDGPUInstPrinter::printRegularOperand(const MCInst *MI, unsigned OpNo,
printImmediate32(Op.getImm(), STI, O);
break;
case AMDGPU::OPERAND_REG_IMM_INT64:
- case AMDGPU::OPERAND_REG_IMM_FP64:
case AMDGPU::OPERAND_REG_INLINE_C_INT64:
+ printImmediate64(Op.getImm(), STI, O, false);
+ break;
+ case AMDGPU::OPERAND_REG_IMM_FP64:
case AMDGPU::OPERAND_REG_INLINE_C_FP64:
case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
- printImmediate64(Op.getImm(), STI, O);
+ printImmediate64(Op.getImm(), STI, O, true);
break;
case AMDGPU::OPERAND_REG_INLINE_C_INT16:
case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
@@ -671,7 +675,7 @@ void AMDGPUInstPrinter::printRegularOperand(const MCInst *MI, unsigned OpNo,
if (RCBits == 32)
printImmediate32(llvm::bit_cast<uint32_t>((float)Value), STI, O);
else if (RCBits == 64)
- printImmediate64(llvm::bit_cast<uint64_t>(Value), STI, O);
+ printImmediate64(llvm::bit_cast<uint64_t>(Value), STI, O, true);
else
llvm_unreachable("Invalid register class size");
}
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
index 3b14faab136b35a..dc83547a4afe049 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.h
@@ -91,7 +91,7 @@ class AMDGPUInstPrinter : public MCInstPrinter {
void printImmediate32(uint32_t Imm, const MCSubtargetInfo &STI,
raw_ostream &O);
void printImmediate64(uint64_t Imm, const MCSubtargetInfo &STI,
- raw_ostream &O);
+ raw_ostream &O, bool IsFP);
void printOperand(const MCInst *MI, unsigned OpNo, const MCSubtargetInfo &STI,
raw_ostream &O);
void printRegularOperand(const MCInst *MI, unsigned OpNo,
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
index 21243f80e055499..d93f747bf6f0a64 100644
--- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
+++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCCodeEmitter.cpp
@@ -411,7 +411,10 @@ void AMDGPUMCCodeEmitter::encodeInstruction(const MCInst &MI,
} else if (!Op.isExpr()) // Exprs will be replaced with a fixup value.
llvm_unreachable("Must be immediate or expr");
- support::endian::write<uint32_t>(CB, Imm, llvm::endianness::little);
+ if (Desc.operands()[i].OperandType == AMDGPU::OPERAND_REG_IMM_FP64)
+ Imm = Hi_32(Imm);
+
+ support::endian::write<uint32_t>(CB, Imm, support::endianness::little);
// Only one literal value allowed
break;
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
index c3c5bfae405aa45..ea06e85fb400c1b 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.td
@@ -1263,7 +1263,9 @@ def VSrc_f32 : RegOrF32 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_v2b16 : RegOrV2B16 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_v2f16 : RegOrV2F16 <"VS_32", "OPERAND_REG_IMM">;
def VSrc_b64 : RegOrB64 <"VS_64", "OPERAND_REG_IMM">;
-def VSrc_f64 : RegOrF64 <"VS_64", "OPERAND_REG_IMM">;
+def VSrc_f64 : RegOrF64 <"VS_64", "OPERAND_REG_IMM"> {
+ let DecoderMethod = "decodeOperand_VSrc_f64";
+}
def VSrc_v2b32 : RegOrV2B32 <"VS_64", "OPERAND_REG_IMM">;
def VSrc_v2f32 : RegOrV2F32 <"VS_64", "OPERAND_REG_IMM">;
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 6d0ad763d9e6cc1..e7907b28abedf9d 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -2519,6 +2519,13 @@ bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi) {
return Lo16 == Hi16;
}
+bool isValid32BitLiteral(uint64_t Val, bool IsFP) {
+ if (IsFP)
+ return !(Val & 0xffffffffu);
+
+ return isUInt<32>(Val) || isInt<32>(Val);
+}
+
bool isArgPassedInSGPR(const Argument *A) {
const Function *F = A->getParent();
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index 297a69f54d63721..fbe9adfd74fa9c6 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -1290,6 +1290,9 @@ bool isInlinableIntLiteralV216(int32_t Literal);
LLVM_READNONE
bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi);
+LLVM_READNONE
+bool isValid32BitLiteral(uint64_t Val, bool IsFP);
+
bool isArgPassedInSGPR(const Argument *Arg);
bool isArgPassedInSGPR(const CallBase *CB, unsigned ArgNo);
>From 0bb5d0690804e371c7d506f4159d1a26617b6f5c Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Thu, 12 Oct 2023 01:16:56 -0700
Subject: [PATCH 5/6] [AMDGPU] Make clang-format happy with disasm cganges
---
.../AMDGPU/Disassembler/AMDGPUDisassembler.cpp | 13 ++++++-------
.../Target/AMDGPU/Disassembler/AMDGPUDisassembler.h | 6 ++----
2 files changed, 8 insertions(+), 11 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
index 83d973dc62e7770..d74fd0b3a9ea74e 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
@@ -378,13 +378,13 @@ static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
}
-static DecodeStatus
-decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm, uint64_t Addr,
- const MCDisassembler *Decoder) {
+static DecodeStatus decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm,
+ uint64_t Addr,
+ const MCDisassembler *Decoder) {
assert(Imm < (1 << 9) && "9-bit encoding");
auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
- return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm,
- false, 64, true));
+ return addOperand(
+ Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm, false, 64, true));
}
static DecodeStatus
@@ -1459,8 +1459,7 @@ int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral,
- unsigned ImmWidth,
- bool IsFP) const {
+ unsigned ImmWidth, bool IsFP) const {
using namespace AMDGPU::EncValues;
assert(Val < 1024); // enum10
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
index 865db2b26307b43..91b73b593d61617 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h
@@ -233,14 +233,12 @@ class AMDGPUDisassembler : public MCDisassembler {
MCOperand decodeLiteralConstant(bool ExtendFP64) const;
MCOperand decodeSrcOp(const OpWidthTy Width, unsigned Val,
- bool MandatoryLiteral = false,
- unsigned ImmWidth = 0,
+ bool MandatoryLiteral = false, unsigned ImmWidth = 0,
bool IsFP = false) const;
MCOperand decodeNonVGPRSrcOp(const OpWidthTy Width, unsigned Val,
bool MandatoryLiteral = false,
- unsigned ImmWidth = 0,
- bool IsFP = false) const;
+ unsigned ImmWidth = 0, bool IsFP = false) const;
MCOperand decodeVOPDDstYOp(MCInst &Inst, unsigned Val) const;
MCOperand decodeSpecialReg32(unsigned Val) const;
>From e2e1efa0e821bbe88b8fb9505d37a6fda8f9209b Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Thu, 12 Oct 2023 12:18:08 -0700
Subject: [PATCH 6/6] Change argument of AMDGPU::isValid32BitLiteral to IsFP64
---
llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp | 7 ++++---
llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | 4 ++--
llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h | 2 +-
3 files changed, 7 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 253a2e98f0cb685..ba967c303d0924f 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -4314,15 +4314,16 @@ bool AMDGPUAsmParser::validateVOPLiteral(const MCInst &Inst,
if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
uint64_t Value = static_cast<uint64_t>(MO.getImm());
- bool IsFP = AMDGPU::isSISrcFPOperand(Desc, OpIdx);
- bool IsValid32Op = AMDGPU::isValid32BitLiteral(Value, IsFP);
+ bool IsFP64 = AMDGPU::isSISrcFPOperand(Desc, OpIdx) &&
+ AMDGPU::getOperandSize(Desc.operands()[OpIdx]) == 8;
+ bool IsValid32Op = AMDGPU::isValid32BitLiteral(Value, IsFP64);
if (!IsValid32Op && !isInt<32>(Value) && !isUInt<32>(Value)) {
Error(getLitLoc(Operands), "invalid operand for instruction");
return false;
}
- if (IsFP && IsValid32Op)
+ if (IsFP64 && IsValid32Op)
Value = Hi_32(Value);
if (NumLiterals == 0 || LiteralValue != Value) {
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index e7907b28abedf9d..d123b384a27d4cc 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -2519,8 +2519,8 @@ bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi) {
return Lo16 == Hi16;
}
-bool isValid32BitLiteral(uint64_t Val, bool IsFP) {
- if (IsFP)
+bool isValid32BitLiteral(uint64_t Val, bool IsFP64) {
+ if (IsFP64)
return !(Val & 0xffffffffu);
return isUInt<32>(Val) || isInt<32>(Val);
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index fbe9adfd74fa9c6..bb2964f592f66bf 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -1291,7 +1291,7 @@ LLVM_READNONE
bool isFoldableLiteralV216(int32_t Literal, bool HasInv2Pi);
LLVM_READNONE
-bool isValid32BitLiteral(uint64_t Val, bool IsFP);
+bool isValid32BitLiteral(uint64_t Val, bool IsFP64);
bool isArgPassedInSGPR(const Argument *Arg);
More information about the llvm-commits
mailing list