1 //===-- AArch64Subtarget.cpp - AArch64 Subtarget Information ----*- 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 // 9 // This file implements the AArch64 specific subclass of TargetSubtarget. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "AArch64Subtarget.h" 14 15 #include "AArch64.h" 16 #include "AArch64CallLowering.h" 17 #include "AArch64InstrInfo.h" 18 #include "AArch64LegalizerInfo.h" 19 #include "AArch64PBQPRegAlloc.h" 20 #include "AArch64RegisterBankInfo.h" 21 #include "AArch64TargetMachine.h" 22 #include "MCTargetDesc/AArch64AddressingModes.h" 23 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 24 #include "llvm/CodeGen/MachineScheduler.h" 25 #include "llvm/IR/GlobalValue.h" 26 #include "llvm/Support/TargetParser.h" 27 28 using namespace llvm; 29 30 #define DEBUG_TYPE "aarch64-subtarget" 31 32 #define GET_SUBTARGETINFO_CTOR 33 #define GET_SUBTARGETINFO_TARGET_DESC 34 #include "AArch64GenSubtargetInfo.inc" 35 36 static cl::opt<bool> 37 EnableEarlyIfConvert("aarch64-early-ifcvt", cl::desc("Enable the early if " 38 "converter pass"), cl::init(true), cl::Hidden); 39 40 // If OS supports TBI, use this flag to enable it. 41 static cl::opt<bool> 42 UseAddressTopByteIgnored("aarch64-use-tbi", cl::desc("Assume that top byte of " 43 "an address is ignored"), cl::init(false), cl::Hidden); 44 45 static cl::opt<bool> 46 UseNonLazyBind("aarch64-enable-nonlazybind", 47 cl::desc("Call nonlazybind functions via direct GOT load"), 48 cl::init(false), cl::Hidden); 49 50 AArch64Subtarget & 51 AArch64Subtarget::initializeSubtargetDependencies(StringRef FS, 52 StringRef CPUString) { 53 // Determine default and user-specified characteristics 54 55 if (CPUString.empty()) 56 CPUString = "generic"; 57 58 ParseSubtargetFeatures(CPUString, FS); 59 initializeProperties(); 60 61 return *this; 62 } 63 64 void AArch64Subtarget::initializeProperties() { 65 // Initialize CPU specific properties. We should add a tablegen feature for 66 // this in the future so we can specify it together with the subtarget 67 // features. 68 switch (ARMProcFamily) { 69 case Others: 70 break; 71 case CortexA35: 72 break; 73 case CortexA53: 74 PrefFunctionLogAlignment = 3; 75 break; 76 case CortexA55: 77 break; 78 case CortexA57: 79 MaxInterleaveFactor = 4; 80 PrefFunctionLogAlignment = 4; 81 break; 82 case CortexA65: 83 PrefFunctionLogAlignment = 3; 84 break; 85 case CortexA72: 86 case CortexA73: 87 case CortexA75: 88 case CortexA76: 89 PrefFunctionLogAlignment = 4; 90 break; 91 case Cyclone: 92 CacheLineSize = 64; 93 PrefetchDistance = 280; 94 MinPrefetchStride = 2048; 95 MaxPrefetchIterationsAhead = 3; 96 break; 97 case ExynosM3: 98 MaxInterleaveFactor = 4; 99 MaxJumpTableSize = 20; 100 PrefFunctionLogAlignment = 5; 101 PrefLoopLogAlignment = 4; 102 break; 103 case Falkor: 104 MaxInterleaveFactor = 4; 105 // FIXME: remove this to enable 64-bit SLP if performance looks good. 106 MinVectorRegisterBitWidth = 128; 107 CacheLineSize = 128; 108 PrefetchDistance = 820; 109 MinPrefetchStride = 2048; 110 MaxPrefetchIterationsAhead = 8; 111 break; 112 case Kryo: 113 MaxInterleaveFactor = 4; 114 VectorInsertExtractBaseCost = 2; 115 CacheLineSize = 128; 116 PrefetchDistance = 740; 117 MinPrefetchStride = 1024; 118 MaxPrefetchIterationsAhead = 11; 119 // FIXME: remove this to enable 64-bit SLP if performance looks good. 120 MinVectorRegisterBitWidth = 128; 121 break; 122 case NeoverseE1: 123 PrefFunctionLogAlignment = 3; 124 break; 125 case NeoverseN1: 126 PrefFunctionLogAlignment = 4; 127 break; 128 case Saphira: 129 MaxInterleaveFactor = 4; 130 // FIXME: remove this to enable 64-bit SLP if performance looks good. 131 MinVectorRegisterBitWidth = 128; 132 break; 133 case ThunderX2T99: 134 CacheLineSize = 64; 135 PrefFunctionLogAlignment = 3; 136 PrefLoopLogAlignment = 2; 137 MaxInterleaveFactor = 4; 138 PrefetchDistance = 128; 139 MinPrefetchStride = 1024; 140 MaxPrefetchIterationsAhead = 4; 141 // FIXME: remove this to enable 64-bit SLP if performance looks good. 142 MinVectorRegisterBitWidth = 128; 143 break; 144 case ThunderX: 145 case ThunderXT88: 146 case ThunderXT81: 147 case ThunderXT83: 148 CacheLineSize = 128; 149 PrefFunctionLogAlignment = 3; 150 PrefLoopLogAlignment = 2; 151 // FIXME: remove this to enable 64-bit SLP if performance looks good. 152 MinVectorRegisterBitWidth = 128; 153 break; 154 case TSV110: 155 CacheLineSize = 64; 156 PrefFunctionLogAlignment = 4; 157 PrefLoopLogAlignment = 2; 158 break; 159 } 160 } 161 162 AArch64Subtarget::AArch64Subtarget(const Triple &TT, const std::string &CPU, 163 const std::string &FS, 164 const TargetMachine &TM, bool LittleEndian) 165 : AArch64GenSubtargetInfo(TT, CPU, FS), 166 ReserveXRegister(AArch64::GPR64commonRegClass.getNumRegs()), 167 CustomCallSavedXRegs(AArch64::GPR64commonRegClass.getNumRegs()), 168 IsLittle(LittleEndian), 169 TargetTriple(TT), FrameLowering(), 170 InstrInfo(initializeSubtargetDependencies(FS, CPU)), TSInfo(), 171 TLInfo(TM, *this) { 172 if (AArch64::isX18ReservedByDefault(TT)) 173 ReserveXRegister.set(18); 174 175 CallLoweringInfo.reset(new AArch64CallLowering(*getTargetLowering())); 176 Legalizer.reset(new AArch64LegalizerInfo(*this)); 177 178 auto *RBI = new AArch64RegisterBankInfo(*getRegisterInfo()); 179 180 // FIXME: At this point, we can't rely on Subtarget having RBI. 181 // It's awkward to mix passing RBI and the Subtarget; should we pass 182 // TII/TRI as well? 183 InstSelector.reset(createAArch64InstructionSelector( 184 *static_cast<const AArch64TargetMachine *>(&TM), *this, *RBI)); 185 186 RegBankInfo.reset(RBI); 187 } 188 189 const CallLowering *AArch64Subtarget::getCallLowering() const { 190 return CallLoweringInfo.get(); 191 } 192 193 InstructionSelector *AArch64Subtarget::getInstructionSelector() const { 194 return InstSelector.get(); 195 } 196 197 const LegalizerInfo *AArch64Subtarget::getLegalizerInfo() const { 198 return Legalizer.get(); 199 } 200 201 const RegisterBankInfo *AArch64Subtarget::getRegBankInfo() const { 202 return RegBankInfo.get(); 203 } 204 205 /// Find the target operand flags that describe how a global value should be 206 /// referenced for the current subtarget. 207 unsigned 208 AArch64Subtarget::ClassifyGlobalReference(const GlobalValue *GV, 209 const TargetMachine &TM) const { 210 // MachO large model always goes via a GOT, simply to get a single 8-byte 211 // absolute relocation on all global addresses. 212 if (TM.getCodeModel() == CodeModel::Large && isTargetMachO()) 213 return AArch64II::MO_GOT; 214 215 if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) { 216 if (GV->hasDLLImportStorageClass()) 217 return AArch64II::MO_GOT | AArch64II::MO_DLLIMPORT; 218 if (getTargetTriple().isOSWindows()) 219 return AArch64II::MO_GOT | AArch64II::MO_COFFSTUB; 220 return AArch64II::MO_GOT; 221 } 222 223 // The small code model's direct accesses use ADRP, which cannot 224 // necessarily produce the value 0 (if the code is above 4GB). 225 // Same for the tiny code model, where we have a pc relative LDR. 226 if ((useSmallAddressing() || TM.getCodeModel() == CodeModel::Tiny) && 227 GV->hasExternalWeakLinkage()) 228 return AArch64II::MO_GOT; 229 230 // References to tagged globals are marked with MO_NC | MO_TAGGED to indicate 231 // that their nominal addresses are tagged and outside of the code model. In 232 // AArch64ExpandPseudo::expandMI we emit an additional instruction to set the 233 // tag if necessary based on MO_TAGGED. 234 if (AllowTaggedGlobals && !isa<FunctionType>(GV->getValueType())) 235 return AArch64II::MO_NC | AArch64II::MO_TAGGED; 236 237 return AArch64II::MO_NO_FLAG; 238 } 239 240 unsigned AArch64Subtarget::classifyGlobalFunctionReference( 241 const GlobalValue *GV, const TargetMachine &TM) const { 242 // MachO large model always goes via a GOT, because we don't have the 243 // relocations available to do anything else.. 244 if (TM.getCodeModel() == CodeModel::Large && isTargetMachO() && 245 !GV->hasInternalLinkage()) 246 return AArch64II::MO_GOT; 247 248 // NonLazyBind goes via GOT unless we know it's available locally. 249 auto *F = dyn_cast<Function>(GV); 250 if (UseNonLazyBind && F && F->hasFnAttribute(Attribute::NonLazyBind) && 251 !TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) 252 return AArch64II::MO_GOT; 253 254 return AArch64II::MO_NO_FLAG; 255 } 256 257 void AArch64Subtarget::overrideSchedPolicy(MachineSchedPolicy &Policy, 258 unsigned NumRegionInstrs) const { 259 // LNT run (at least on Cyclone) showed reasonably significant gains for 260 // bi-directional scheduling. 253.perlbmk. 261 Policy.OnlyTopDown = false; 262 Policy.OnlyBottomUp = false; 263 // Enabling or Disabling the latency heuristic is a close call: It seems to 264 // help nearly no benchmark on out-of-order architectures, on the other hand 265 // it regresses register pressure on a few benchmarking. 266 Policy.DisableLatencyHeuristic = DisableLatencySchedHeuristic; 267 } 268 269 bool AArch64Subtarget::enableEarlyIfConversion() const { 270 return EnableEarlyIfConvert; 271 } 272 273 bool AArch64Subtarget::supportsAddressTopByteIgnored() const { 274 if (!UseAddressTopByteIgnored) 275 return false; 276 277 if (TargetTriple.isiOS()) { 278 unsigned Major, Minor, Micro; 279 TargetTriple.getiOSVersion(Major, Minor, Micro); 280 return Major >= 8; 281 } 282 283 return false; 284 } 285 286 std::unique_ptr<PBQPRAConstraint> 287 AArch64Subtarget::getCustomPBQPConstraints() const { 288 return balanceFPOps() ? std::make_unique<A57ChainingConstraint>() : nullptr; 289 } 290 291 void AArch64Subtarget::mirFileLoaded(MachineFunction &MF) const { 292 // We usually compute max call frame size after ISel. Do the computation now 293 // if the .mir file didn't specify it. Note that this will probably give you 294 // bogus values after PEI has eliminated the callframe setup/destroy pseudo 295 // instructions, specify explicitly if you need it to be correct. 296 MachineFrameInfo &MFI = MF.getFrameInfo(); 297 if (!MFI.isMaxCallFrameSizeComputed()) 298 MFI.computeMaxCallFrameSize(MF); 299 } 300