1 //===-- ARMSubtarget.cpp - ARM Subtarget Information ----------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the ARM specific subclass of TargetSubtargetInfo. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "ARM.h" 15 16 #include "ARMCallLowering.h" 17 #include "ARMLegalizerInfo.h" 18 #include "ARMRegisterBankInfo.h" 19 #include "ARMSubtarget.h" 20 #include "ARMFrameLowering.h" 21 #include "ARMInstrInfo.h" 22 #include "ARMSubtarget.h" 23 #include "ARMTargetMachine.h" 24 #include "MCTargetDesc/ARMMCTargetDesc.h" 25 #include "Thumb1FrameLowering.h" 26 #include "Thumb1InstrInfo.h" 27 #include "Thumb2InstrInfo.h" 28 #include "llvm/ADT/StringRef.h" 29 #include "llvm/ADT/Triple.h" 30 #include "llvm/ADT/Twine.h" 31 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 32 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 33 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 34 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 35 #include "llvm/CodeGen/MachineFunction.h" 36 #include "llvm/IR/Function.h" 37 #include "llvm/IR/GlobalValue.h" 38 #include "llvm/MC/MCAsmInfo.h" 39 #include "llvm/MC/MCTargetOptions.h" 40 #include "llvm/Support/CodeGen.h" 41 #include "llvm/Support/CommandLine.h" 42 #include "llvm/Support/TargetParser.h" 43 #include "llvm/Target/TargetOptions.h" 44 #include <cassert> 45 #include <string> 46 47 using namespace llvm; 48 49 #define DEBUG_TYPE "arm-subtarget" 50 51 #define GET_SUBTARGETINFO_TARGET_DESC 52 #define GET_SUBTARGETINFO_CTOR 53 #include "ARMGenSubtargetInfo.inc" 54 55 static cl::opt<bool> 56 UseFusedMulOps("arm-use-mulops", 57 cl::init(true), cl::Hidden); 58 59 enum ITMode { 60 DefaultIT, 61 RestrictedIT, 62 NoRestrictedIT 63 }; 64 65 static cl::opt<ITMode> 66 IT(cl::desc("IT block support"), cl::Hidden, cl::init(DefaultIT), 67 cl::ZeroOrMore, 68 cl::values(clEnumValN(DefaultIT, "arm-default-it", 69 "Generate IT block based on arch"), 70 clEnumValN(RestrictedIT, "arm-restrict-it", 71 "Disallow deprecated IT based on ARMv8"), 72 clEnumValN(NoRestrictedIT, "arm-no-restrict-it", 73 "Allow IT blocks based on ARMv7"))); 74 75 /// ForceFastISel - Use the fast-isel, even for subtargets where it is not 76 /// currently supported (for testing only). 77 static cl::opt<bool> 78 ForceFastISel("arm-force-fast-isel", 79 cl::init(false), cl::Hidden); 80 81 /// initializeSubtargetDependencies - Initializes using a CPU and feature string 82 /// so that we can use initializer lists for subtarget initialization. 83 ARMSubtarget &ARMSubtarget::initializeSubtargetDependencies(StringRef CPU, 84 StringRef FS) { 85 initializeEnvironment(); 86 initSubtargetFeatures(CPU, FS); 87 return *this; 88 } 89 90 ARMFrameLowering *ARMSubtarget::initializeFrameLowering(StringRef CPU, 91 StringRef FS) { 92 ARMSubtarget &STI = initializeSubtargetDependencies(CPU, FS); 93 if (STI.isThumb1Only()) 94 return (ARMFrameLowering *)new Thumb1FrameLowering(STI); 95 96 return new ARMFrameLowering(STI); 97 } 98 99 ARMSubtarget::ARMSubtarget(const Triple &TT, const std::string &CPU, 100 const std::string &FS, 101 const ARMBaseTargetMachine &TM, bool IsLittle) 102 : ARMGenSubtargetInfo(TT, CPU, FS), UseMulOps(UseFusedMulOps), 103 CPUString(CPU), IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), 104 TM(TM), FrameLowering(initializeFrameLowering(CPU, FS)), 105 // At this point initializeSubtargetDependencies has been called so 106 // we can query directly. 107 InstrInfo(isThumb1Only() 108 ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this) 109 : !isThumb() 110 ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this) 111 : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)), 112 TLInfo(TM, *this) { 113 114 CallLoweringInfo.reset(new ARMCallLowering(*getTargetLowering())); 115 Legalizer.reset(new ARMLegalizerInfo(*this)); 116 117 auto *RBI = new ARMRegisterBankInfo(*getRegisterInfo()); 118 119 // FIXME: At this point, we can't rely on Subtarget having RBI. 120 // It's awkward to mix passing RBI and the Subtarget; should we pass 121 // TII/TRI as well? 122 InstSelector.reset(createARMInstructionSelector( 123 *static_cast<const ARMBaseTargetMachine *>(&TM), *this, *RBI)); 124 125 RegBankInfo.reset(RBI); 126 } 127 128 const CallLowering *ARMSubtarget::getCallLowering() const { 129 return CallLoweringInfo.get(); 130 } 131 132 const InstructionSelector *ARMSubtarget::getInstructionSelector() const { 133 return InstSelector.get(); 134 } 135 136 const LegalizerInfo *ARMSubtarget::getLegalizerInfo() const { 137 return Legalizer.get(); 138 } 139 140 const RegisterBankInfo *ARMSubtarget::getRegBankInfo() const { 141 return RegBankInfo.get(); 142 } 143 144 bool ARMSubtarget::isXRaySupported() const { 145 // We don't currently suppport Thumb, but Windows requires Thumb. 146 return hasV6Ops() && hasARMOps() && !isTargetWindows(); 147 } 148 149 void ARMSubtarget::initializeEnvironment() { 150 // MCAsmInfo isn't always present (e.g. in opt) so we can't initialize this 151 // directly from it, but we can try to make sure they're consistent when both 152 // available. 153 UseSjLjEH = isTargetDarwin() && !isTargetWatchABI(); 154 assert((!TM.getMCAsmInfo() || 155 (TM.getMCAsmInfo()->getExceptionHandlingType() == 156 ExceptionHandling::SjLj) == UseSjLjEH) && 157 "inconsistent sjlj choice between CodeGen and MC"); 158 } 159 160 void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 161 if (CPUString.empty()) { 162 CPUString = "generic"; 163 164 if (isTargetDarwin()) { 165 StringRef ArchName = TargetTriple.getArchName(); 166 ARM::ArchKind AK = ARM::parseArch(ArchName); 167 if (AK == ARM::ArchKind::ARMV7S) 168 // Default to the Swift CPU when targeting armv7s/thumbv7s. 169 CPUString = "swift"; 170 else if (AK == ARM::ArchKind::ARMV7K) 171 // Default to the Cortex-a7 CPU when targeting armv7k/thumbv7k. 172 // ARMv7k does not use SjLj exception handling. 173 CPUString = "cortex-a7"; 174 } 175 } 176 177 // Insert the architecture feature derived from the target triple into the 178 // feature string. This is important for setting features that are implied 179 // based on the architecture version. 180 std::string ArchFS = ARM_MC::ParseARMTriple(TargetTriple, CPUString); 181 if (!FS.empty()) { 182 if (!ArchFS.empty()) 183 ArchFS = (Twine(ArchFS) + "," + FS).str(); 184 else 185 ArchFS = FS; 186 } 187 ParseSubtargetFeatures(CPUString, ArchFS); 188 189 // FIXME: This used enable V6T2 support implicitly for Thumb2 mode. 190 // Assert this for now to make the change obvious. 191 assert(hasV6T2Ops() || !hasThumb2()); 192 193 // Execute only support requires movt support 194 if (genExecuteOnly()) 195 assert(hasV8MBaselineOps() && !NoMovt && "Cannot generate execute-only code for this target"); 196 197 // Keep a pointer to static instruction cost data for the specified CPU. 198 SchedModel = getSchedModelForCPU(CPUString); 199 200 // Initialize scheduling itinerary for the specified CPU. 201 InstrItins = getInstrItineraryForCPU(CPUString); 202 203 // FIXME: this is invalid for WindowsCE 204 if (isTargetWindows()) 205 NoARM = true; 206 207 if (isAAPCS_ABI()) 208 stackAlignment = 8; 209 if (isTargetNaCl() || isAAPCS16_ABI()) 210 stackAlignment = 16; 211 212 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo:: 213 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 214 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 215 // support in the assembler and linker to be used. This would need to be 216 // fixed to fully support tail calls in Thumb1. 217 // 218 // For ARMv8-M, we /do/ implement tail calls. Doing this is tricky for v8-M 219 // baseline, since the LDM/POP instruction on Thumb doesn't take LR. This 220 // means if we need to reload LR, it takes extra instructions, which outweighs 221 // the value of the tail call; but here we don't know yet whether LR is going 222 // to be used. We generate the tail call here and turn it back into CALL/RET 223 // in emitEpilogue if LR is used. 224 225 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 226 // but we need to make sure there are enough registers; the only valid 227 // registers are the 4 used for parameters. We don't currently do this 228 // case. 229 230 SupportsTailCall = !isThumb() || hasV8MBaselineOps(); 231 232 if (isTargetMachO() && isTargetIOS() && getTargetTriple().isOSVersionLT(5, 0)) 233 SupportsTailCall = false; 234 235 switch (IT) { 236 case DefaultIT: 237 RestrictIT = hasV8Ops(); 238 break; 239 case RestrictedIT: 240 RestrictIT = true; 241 break; 242 case NoRestrictedIT: 243 RestrictIT = false; 244 break; 245 } 246 247 // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default. 248 const FeatureBitset &Bits = getFeatureBits(); 249 if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters 250 (Options.UnsafeFPMath || isTargetDarwin())) 251 UseNEONForSinglePrecisionFP = true; 252 253 if (isRWPI()) 254 ReserveR9 = true; 255 256 // FIXME: Teach TableGen to deal with these instead of doing it manually here. 257 switch (ARMProcFamily) { 258 case Others: 259 case CortexA5: 260 break; 261 case CortexA7: 262 LdStMultipleTiming = DoubleIssue; 263 break; 264 case CortexA8: 265 LdStMultipleTiming = DoubleIssue; 266 break; 267 case CortexA9: 268 LdStMultipleTiming = DoubleIssueCheckUnalignedAccess; 269 PreISelOperandLatencyAdjustment = 1; 270 break; 271 case CortexA12: 272 break; 273 case CortexA15: 274 MaxInterleaveFactor = 2; 275 PreISelOperandLatencyAdjustment = 1; 276 PartialUpdateClearance = 12; 277 break; 278 case CortexA17: 279 case CortexA32: 280 case CortexA35: 281 case CortexA53: 282 case CortexA55: 283 case CortexA57: 284 case CortexA72: 285 case CortexA73: 286 case CortexA75: 287 case CortexR4: 288 case CortexR4F: 289 case CortexR5: 290 case CortexR7: 291 case CortexM3: 292 case CortexR52: 293 case ExynosM1: 294 case Kryo: 295 break; 296 case Krait: 297 PreISelOperandLatencyAdjustment = 1; 298 break; 299 case Swift: 300 MaxInterleaveFactor = 2; 301 LdStMultipleTiming = SingleIssuePlusExtras; 302 PreISelOperandLatencyAdjustment = 1; 303 PartialUpdateClearance = 12; 304 break; 305 } 306 } 307 308 bool ARMSubtarget::isAPCS_ABI() const { 309 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 310 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS; 311 } 312 bool ARMSubtarget::isAAPCS_ABI() const { 313 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 314 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS || 315 TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 316 } 317 bool ARMSubtarget::isAAPCS16_ABI() const { 318 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 319 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 320 } 321 322 bool ARMSubtarget::isROPI() const { 323 return TM.getRelocationModel() == Reloc::ROPI || 324 TM.getRelocationModel() == Reloc::ROPI_RWPI; 325 } 326 bool ARMSubtarget::isRWPI() const { 327 return TM.getRelocationModel() == Reloc::RWPI || 328 TM.getRelocationModel() == Reloc::ROPI_RWPI; 329 } 330 331 bool ARMSubtarget::isGVIndirectSymbol(const GlobalValue *GV) const { 332 if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) 333 return true; 334 335 // 32 bit macho has no relocation for a-b if a is undefined, even if b is in 336 // the section that is being relocated. This means we have to use o load even 337 // for GVs that are known to be local to the dso. 338 if (isTargetMachO() && TM.isPositionIndependent() && 339 (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) 340 return true; 341 342 return false; 343 } 344 345 ARMCP::ARMCPModifier ARMSubtarget::getCPModifier(const GlobalValue *GV) const { 346 if (isTargetELF() && TM.isPositionIndependent() && 347 !TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) 348 return ARMCP::GOT_PREL; 349 return ARMCP::no_modifier; 350 } 351 352 unsigned ARMSubtarget::getMispredictionPenalty() const { 353 return SchedModel.MispredictPenalty; 354 } 355 356 bool ARMSubtarget::hasSinCos() const { 357 return isTargetWatchOS() || 358 (isTargetIOS() && !getTargetTriple().isOSVersionLT(7, 0)); 359 } 360 361 bool ARMSubtarget::enableMachineScheduler() const { 362 // Enable the MachineScheduler before register allocation for subtargets 363 // with the use-misched feature. 364 return useMachineScheduler(); 365 } 366 367 // This overrides the PostRAScheduler bit in the SchedModel for any CPU. 368 bool ARMSubtarget::enablePostRAScheduler() const { 369 if (disablePostRAScheduler()) 370 return false; 371 // Don't reschedule potential IT blocks. 372 return !isThumb1Only(); 373 } 374 375 bool ARMSubtarget::enableAtomicExpand() const { return hasAnyDataBarrier(); } 376 377 bool ARMSubtarget::useStride4VFPs(const MachineFunction &MF) const { 378 // For general targets, the prologue can grow when VFPs are allocated with 379 // stride 4 (more vpush instructions). But WatchOS uses a compact unwind 380 // format which it's more important to get right. 381 return isTargetWatchABI() || (isSwift() && !MF.getFunction()->optForMinSize()); 382 } 383 384 bool ARMSubtarget::useMovt(const MachineFunction &MF) const { 385 // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit 386 // immediates as it is inherently position independent, and may be out of 387 // range otherwise. 388 return !NoMovt && hasV8MBaselineOps() && 389 (isTargetWindows() || !MF.getFunction()->optForMinSize() || genExecuteOnly()); 390 } 391 392 bool ARMSubtarget::useFastISel() const { 393 // Enable fast-isel for any target, for testing only. 394 if (ForceFastISel) 395 return true; 396 397 // Limit fast-isel to the targets that are or have been tested. 398 if (!hasV6Ops()) 399 return false; 400 401 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl. 402 return TM.Options.EnableFastISel && 403 ((isTargetMachO() && !isThumb1Only()) || 404 (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb())); 405 } 406