| 1 |
//===- AArch64GlobalISelUtils.cpp --------------------------------*- C++ -*-==// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
/// \file Implementations of AArch64-specific helper functions used in the |
| 9 |
/// GlobalISel pipeline. |
| 10 |
//===----------------------------------------------------------------------===// |
| 11 |
#include "AArch64GlobalISelUtils.h" |
| 12 |
#include "AArch64InstrInfo.h" |
| 13 |
#include "llvm/CodeGen/GlobalISel/Utils.h" |
| 14 |
#include "llvm/CodeGen/TargetLowering.h" |
| 15 |
#include "llvm/IR/InstrTypes.h" |
| 16 |
#include "llvm/Support/raw_ostream.h" |
| 17 |
|
| 18 |
using namespace llvm; |
| 19 |
|
| 20 |
std::optional<RegOrConstant> |
| 21 |
AArch64GISelUtils::getAArch64VectorSplat(const MachineInstr &MI, |
| 22 |
const MachineRegisterInfo &MRI) { |
| 23 |
if (auto Splat = getVectorSplat(MI, MRI)) |
| 24 |
return Splat; |
| 25 |
if (MI.getOpcode() != AArch64::G_DUP) |
| 26 |
return std::nullopt; |
| 27 |
Register Src = MI.getOperand(1).getReg(); |
| 28 |
if (auto ValAndVReg = |
| 29 |
getAnyConstantVRegValWithLookThrough(MI.getOperand(1).getReg(), MRI)) |
| 30 |
return RegOrConstant(ValAndVReg->Value.getSExtValue()); |
| 31 |
return RegOrConstant(Src); |
| 32 |
} |
| 33 |
|
| 34 |
std::optional<int64_t> |
| 35 |
AArch64GISelUtils::getAArch64VectorSplatScalar(const MachineInstr &MI, |
| 36 |
const MachineRegisterInfo &MRI) { |
| 37 |
auto Splat = getAArch64VectorSplat(MI, MRI); |
| 38 |
if (!Splat || Splat->isReg()) |
| 39 |
return std::nullopt; |
| 40 |
return Splat->getCst(); |
| 41 |
} |
| 42 |
|
| 43 |
bool AArch64GISelUtils::isCMN(const MachineInstr *MaybeSub, |
| 44 |
const CmpInst::Predicate &Pred, |
| 45 |
const MachineRegisterInfo &MRI) { |
| 46 |
// Match: |
| 47 |
// |
| 48 |
// %sub = G_SUB 0, %y |
| 49 |
// %cmp = G_ICMP eq/ne, %sub, %z |
| 50 |
// |
| 51 |
// Or |
| 52 |
// |
| 53 |
// %sub = G_SUB 0, %y |
| 54 |
// %cmp = G_ICMP eq/ne, %z, %sub |
| 55 |
if (!MaybeSub || MaybeSub->getOpcode() != TargetOpcode::G_SUB || |
| 56 |
!CmpInst::isEquality(Pred)) |
| 57 |
return false; |
| 58 |
auto MaybeZero = |
| 59 |
getIConstantVRegValWithLookThrough(MaybeSub->getOperand(1).getReg(), MRI); |
| 60 |
return MaybeZero && MaybeZero->Value.getZExtValue() == 0; |
| 61 |
} |
| 62 |
|
| 63 |
bool AArch64GISelUtils::tryEmitBZero(MachineInstr &MI, |
| 64 |
MachineIRBuilder &MIRBuilder, |
| 65 |
bool MinSize) { |
| 66 |
assert(MI.getOpcode() == TargetOpcode::G_MEMSET); |
| 67 |
MachineRegisterInfo &MRI = *MIRBuilder.getMRI(); |
| 68 |
auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); |
| 69 |
if (!TLI.getLibcallName(RTLIB::BZERO)) |
| 70 |
return false; |
| 71 |
auto Zero = |
| 72 |
getIConstantVRegValWithLookThrough(MI.getOperand(1).getReg(), MRI); |
| 73 |
if (!Zero || Zero->Value.getSExtValue() != 0) |
| 74 |
return false; |
| 75 |
|
| 76 |
// It's not faster to use bzero rather than memset for sizes <= 256. |
| 77 |
// However, it *does* save us a mov from wzr, so if we're going for |
| 78 |
// minsize, use bzero even if it's slower. |
| 79 |
if (!MinSize) { |
| 80 |
// If the size is known, check it. If it is not known, assume using bzero is |
| 81 |
// better. |
| 82 |
if (auto Size = getIConstantVRegValWithLookThrough( |
| 83 |
MI.getOperand(2).getReg(), MRI)) { |
| 84 |
if (Size->Value.getSExtValue() <= 256) |
| 85 |
return false; |
| 86 |
} |
| 87 |
} |
| 88 |
|
| 89 |
MIRBuilder.setInstrAndDebugLoc(MI); |
| 90 |
MIRBuilder |
| 91 |
.buildInstr(TargetOpcode::G_BZERO, {}, |
| 92 |
{MI.getOperand(0), MI.getOperand(2)}) |
| 93 |
.addImm(MI.getOperand(3).getImm()) |
| 94 |
.addMemOperand(*MI.memoperands_begin()); |
| 95 |
MI.eraseFromParent(); |
| 96 |
return true; |
| 97 |
} |
| 98 |
|
| 99 |
void AArch64GISelUtils::changeFCMPPredToAArch64CC( |
| 100 |
const CmpInst::Predicate P, AArch64CC::CondCode &CondCode, |
| 101 |
AArch64CC::CondCode &CondCode2) { |
| 102 |
CondCode2 = AArch64CC::AL; |
| 103 |
switch (P) { |
| 104 |
default: |
| 105 |
llvm_unreachable("Unknown FP condition!"); |
| 106 |
case CmpInst::FCMP_OEQ: |
| 107 |
CondCode = AArch64CC::EQ; |
| 108 |
break; |
| 109 |
case CmpInst::FCMP_OGT: |
| 110 |
CondCode = AArch64CC::GT; |
| 111 |
break; |
| 112 |
case CmpInst::FCMP_OGE: |
| 113 |
CondCode = AArch64CC::GE; |
| 114 |
break; |
| 115 |
case CmpInst::FCMP_OLT: |
| 116 |
CondCode = AArch64CC::MI; |
| 117 |
break; |
| 118 |
case CmpInst::FCMP_OLE: |
| 119 |
CondCode = AArch64CC::LS; |
| 120 |
break; |
| 121 |
case CmpInst::FCMP_ONE: |
| 122 |
CondCode = AArch64CC::MI; |
| 123 |
CondCode2 = AArch64CC::GT; |
| 124 |
break; |
| 125 |
case CmpInst::FCMP_ORD: |
| 126 |
CondCode = AArch64CC::VC; |
| 127 |
break; |
| 128 |
case CmpInst::FCMP_UNO: |
| 129 |
CondCode = AArch64CC::VS; |
| 130 |
break; |
| 131 |
case CmpInst::FCMP_UEQ: |
| 132 |
CondCode = AArch64CC::EQ; |
| 133 |
CondCode2 = AArch64CC::VS; |
| 134 |
break; |
| 135 |
case CmpInst::FCMP_UGT: |
| 136 |
CondCode = AArch64CC::HI; |
| 137 |
break; |
| 138 |
case CmpInst::FCMP_UGE: |
| 139 |
CondCode = AArch64CC::PL; |
| 140 |
break; |
| 141 |
case CmpInst::FCMP_ULT: |
| 142 |
CondCode = AArch64CC::LT; |
| 143 |
break; |
| 144 |
case CmpInst::FCMP_ULE: |
| 145 |
CondCode = AArch64CC::LE; |
| 146 |
break; |
| 147 |
case CmpInst::FCMP_UNE: |
| 148 |
CondCode = AArch64CC::NE; |
| 149 |
break; |
| 150 |
} |
| 151 |
} |
| 152 |
|
| 153 |
void AArch64GISelUtils::changeVectorFCMPPredToAArch64CC( |
| 154 |
const CmpInst::Predicate P, AArch64CC::CondCode &CondCode, |
| 155 |
AArch64CC::CondCode &CondCode2, bool &Invert) { |
| 156 |
Invert = false; |
| 157 |
switch (P) { |
| 158 |
default: |
| 159 |
// Mostly the scalar mappings work fine. |
| 160 |
changeFCMPPredToAArch64CC(P, CondCode, CondCode2); |
| 161 |
break; |
| 162 |
case CmpInst::FCMP_UNO: |
| 163 |
Invert = true; |
| 164 |
[[fallthrough]]; |
| 165 |
case CmpInst::FCMP_ORD: |
| 166 |
CondCode = AArch64CC::MI; |
| 167 |
CondCode2 = AArch64CC::GE; |
| 168 |
break; |
| 169 |
case CmpInst::FCMP_UEQ: |
| 170 |
case CmpInst::FCMP_ULT: |
| 171 |
case CmpInst::FCMP_ULE: |
| 172 |
case CmpInst::FCMP_UGT: |
| 173 |
case CmpInst::FCMP_UGE: |
| 174 |
// All of the compare-mask comparisons are ordered, but we can switch |
| 175 |
// between the two by a double inversion. E.g. ULE == !OGT. |
| 176 |
Invert = true; |
| 177 |
changeFCMPPredToAArch64CC(CmpInst::getInversePredicate(P), CondCode, |
| 178 |
CondCode2); |
| 179 |
break; |
| 180 |
} |
| 181 |
} |