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 "ARMSubtarget.h" 15 #include "ARMFrameLowering.h" 16 #include "ARMISelLowering.h" 17 #include "ARMInstrInfo.h" 18 #include "ARMMachineFunctionInfo.h" 19 #include "ARMSelectionDAGInfo.h" 20 #include "ARMSubtarget.h" 21 #include "ARMTargetMachine.h" 22 #include "Thumb1FrameLowering.h" 23 #include "Thumb1InstrInfo.h" 24 #include "Thumb2InstrInfo.h" 25 #include "llvm/CodeGen/MachineRegisterInfo.h" 26 #include "llvm/IR/Attributes.h" 27 #include "llvm/IR/Function.h" 28 #include "llvm/IR/GlobalValue.h" 29 #include "llvm/MC/MCAsmInfo.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Target/TargetInstrInfo.h" 32 #include "llvm/Target/TargetOptions.h" 33 #include "llvm/Target/TargetRegisterInfo.h" 34 #include "llvm/Support/TargetParser.h" 35 36 using namespace llvm; 37 38 #define DEBUG_TYPE "arm-subtarget" 39 40 #define GET_SUBTARGETINFO_TARGET_DESC 41 #define GET_SUBTARGETINFO_CTOR 42 #include "ARMGenSubtargetInfo.inc" 43 44 static cl::opt<bool> 45 UseFusedMulOps("arm-use-mulops", 46 cl::init(true), cl::Hidden); 47 48 enum ITMode { 49 DefaultIT, 50 RestrictedIT, 51 NoRestrictedIT 52 }; 53 54 static cl::opt<ITMode> 55 IT(cl::desc("IT block support"), cl::Hidden, cl::init(DefaultIT), 56 cl::ZeroOrMore, 57 cl::values(clEnumValN(DefaultIT, "arm-default-it", 58 "Generate IT block based on arch"), 59 clEnumValN(RestrictedIT, "arm-restrict-it", 60 "Disallow deprecated IT based on ARMv8"), 61 clEnumValN(NoRestrictedIT, "arm-no-restrict-it", 62 "Allow IT blocks based on ARMv7"))); 63 64 /// ForceFastISel - Use the fast-isel, even for subtargets where it is not 65 /// currently supported (for testing only). 66 static cl::opt<bool> 67 ForceFastISel("arm-force-fast-isel", 68 cl::init(false), cl::Hidden); 69 70 /// initializeSubtargetDependencies - Initializes using a CPU and feature string 71 /// so that we can use initializer lists for subtarget initialization. 72 ARMSubtarget &ARMSubtarget::initializeSubtargetDependencies(StringRef CPU, 73 StringRef FS) { 74 initializeEnvironment(); 75 initSubtargetFeatures(CPU, FS); 76 return *this; 77 } 78 79 ARMFrameLowering *ARMSubtarget::initializeFrameLowering(StringRef CPU, 80 StringRef FS) { 81 ARMSubtarget &STI = initializeSubtargetDependencies(CPU, FS); 82 if (STI.isThumb1Only()) 83 return (ARMFrameLowering *)new Thumb1FrameLowering(STI); 84 85 return new ARMFrameLowering(STI); 86 } 87 88 ARMSubtarget::ARMSubtarget(const Triple &TT, const std::string &CPU, 89 const std::string &FS, 90 const ARMBaseTargetMachine &TM, bool IsLittle) 91 : ARMGenSubtargetInfo(TT, CPU, FS), UseMulOps(UseFusedMulOps), 92 CPUString(CPU), IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), 93 TM(TM), FrameLowering(initializeFrameLowering(CPU, FS)), 94 // At this point initializeSubtargetDependencies has been called so 95 // we can query directly. 96 InstrInfo(isThumb1Only() 97 ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this) 98 : !isThumb() 99 ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this) 100 : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)), 101 TLInfo(TM, *this) {} 102 103 bool ARMSubtarget::isXRaySupported() const { 104 // We don't currently suppport Thumb, but Windows requires Thumb. 105 return hasV6Ops() && hasARMOps() && !isTargetWindows(); 106 } 107 108 void ARMSubtarget::initializeEnvironment() { 109 // MCAsmInfo isn't always present (e.g. in opt) so we can't initialize this 110 // directly from it, but we can try to make sure they're consistent when both 111 // available. 112 UseSjLjEH = isTargetDarwin() && !isTargetWatchABI(); 113 assert((!TM.getMCAsmInfo() || 114 (TM.getMCAsmInfo()->getExceptionHandlingType() == 115 ExceptionHandling::SjLj) == UseSjLjEH) && 116 "inconsistent sjlj choice between CodeGen and MC"); 117 } 118 119 void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 120 if (CPUString.empty()) { 121 CPUString = "generic"; 122 123 if (isTargetDarwin()) { 124 StringRef ArchName = TargetTriple.getArchName(); 125 unsigned ArchKind = llvm::ARM::parseArch(ArchName); 126 if (ArchKind == llvm::ARM::AK_ARMV7S) 127 // Default to the Swift CPU when targeting armv7s/thumbv7s. 128 CPUString = "swift"; 129 else if (ArchKind == llvm::ARM::AK_ARMV7K) 130 // Default to the Cortex-a7 CPU when targeting armv7k/thumbv7k. 131 // ARMv7k does not use SjLj exception handling. 132 CPUString = "cortex-a7"; 133 } 134 } 135 136 // Insert the architecture feature derived from the target triple into the 137 // feature string. This is important for setting features that are implied 138 // based on the architecture version. 139 std::string ArchFS = ARM_MC::ParseARMTriple(TargetTriple, CPUString); 140 if (!FS.empty()) { 141 if (!ArchFS.empty()) 142 ArchFS = (Twine(ArchFS) + "," + FS).str(); 143 else 144 ArchFS = FS; 145 } 146 ParseSubtargetFeatures(CPUString, ArchFS); 147 148 // FIXME: This used enable V6T2 support implicitly for Thumb2 mode. 149 // Assert this for now to make the change obvious. 150 assert(hasV6T2Ops() || !hasThumb2()); 151 152 // Keep a pointer to static instruction cost data for the specified CPU. 153 SchedModel = getSchedModelForCPU(CPUString); 154 155 // Initialize scheduling itinerary for the specified CPU. 156 InstrItins = getInstrItineraryForCPU(CPUString); 157 158 // FIXME: this is invalid for WindowsCE 159 if (isTargetWindows()) 160 NoARM = true; 161 162 if (isAAPCS_ABI()) 163 stackAlignment = 8; 164 if (isTargetNaCl() || isAAPCS16_ABI()) 165 stackAlignment = 16; 166 167 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo:: 168 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 169 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 170 // support in the assembler and linker to be used. This would need to be 171 // fixed to fully support tail calls in Thumb1. 172 // 173 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 174 // LR. This means if we need to reload LR, it takes an extra instructions, 175 // which outweighs the value of the tail call; but here we don't know yet 176 // whether LR is going to be used. Probably the right approach is to 177 // generate the tail call here and turn it back into CALL/RET in 178 // emitEpilogue if LR is used. 179 180 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 181 // but we need to make sure there are enough registers; the only valid 182 // registers are the 4 used for parameters. We don't currently do this 183 // case. 184 185 SupportsTailCall = !isThumb() || hasV8MBaselineOps(); 186 187 if (isTargetMachO() && isTargetIOS() && getTargetTriple().isOSVersionLT(5, 0)) 188 SupportsTailCall = false; 189 190 switch (IT) { 191 case DefaultIT: 192 RestrictIT = hasV8Ops(); 193 break; 194 case RestrictedIT: 195 RestrictIT = true; 196 break; 197 case NoRestrictedIT: 198 RestrictIT = false; 199 break; 200 } 201 202 // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default. 203 const FeatureBitset &Bits = getFeatureBits(); 204 if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters 205 (Options.UnsafeFPMath || isTargetDarwin())) 206 UseNEONForSinglePrecisionFP = true; 207 208 if (isRWPI()) 209 ReserveR9 = true; 210 211 // FIXME: Teach TableGen to deal with these instead of doing it manually here. 212 switch (ARMProcFamily) { 213 case Others: 214 case CortexA5: 215 break; 216 case CortexA7: 217 LdStMultipleTiming = DoubleIssue; 218 break; 219 case CortexA8: 220 LdStMultipleTiming = DoubleIssue; 221 break; 222 case CortexA9: 223 LdStMultipleTiming = DoubleIssueCheckUnalignedAccess; 224 PreISelOperandLatencyAdjustment = 1; 225 break; 226 case CortexA12: 227 break; 228 case CortexA15: 229 MaxInterleaveFactor = 2; 230 PreISelOperandLatencyAdjustment = 1; 231 PartialUpdateClearance = 12; 232 break; 233 case CortexA17: 234 case CortexA32: 235 case CortexA35: 236 case CortexA53: 237 case CortexA57: 238 case CortexA72: 239 case CortexA73: 240 case CortexR4: 241 case CortexR4F: 242 case CortexR5: 243 case CortexR7: 244 case CortexM3: 245 case ExynosM1: 246 case CortexR52: 247 break; 248 case Krait: 249 PreISelOperandLatencyAdjustment = 1; 250 break; 251 case Swift: 252 MaxInterleaveFactor = 2; 253 LdStMultipleTiming = SingleIssuePlusExtras; 254 PreISelOperandLatencyAdjustment = 1; 255 PartialUpdateClearance = 12; 256 break; 257 } 258 } 259 260 bool ARMSubtarget::isAPCS_ABI() const { 261 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 262 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS; 263 } 264 bool ARMSubtarget::isAAPCS_ABI() const { 265 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 266 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS || 267 TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 268 } 269 bool ARMSubtarget::isAAPCS16_ABI() const { 270 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 271 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 272 } 273 274 bool ARMSubtarget::isROPI() const { 275 return TM.getRelocationModel() == Reloc::ROPI || 276 TM.getRelocationModel() == Reloc::ROPI_RWPI; 277 } 278 bool ARMSubtarget::isRWPI() const { 279 return TM.getRelocationModel() == Reloc::RWPI || 280 TM.getRelocationModel() == Reloc::ROPI_RWPI; 281 } 282 283 bool ARMSubtarget::isGVIndirectSymbol(const GlobalValue *GV) const { 284 if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) 285 return true; 286 287 // 32 bit macho has no relocation for a-b if a is undefined, even if b is in 288 // the section that is being relocated. This means we have to use o load even 289 // for GVs that are known to be local to the dso. 290 if (isTargetMachO() && TM.isPositionIndependent() && 291 (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) 292 return true; 293 294 return false; 295 } 296 297 unsigned ARMSubtarget::getMispredictionPenalty() const { 298 return SchedModel.MispredictPenalty; 299 } 300 301 bool ARMSubtarget::hasSinCos() const { 302 return isTargetWatchOS() || 303 (isTargetIOS() && !getTargetTriple().isOSVersionLT(7, 0)); 304 } 305 306 bool ARMSubtarget::enableMachineScheduler() const { 307 // Enable the MachineScheduler before register allocation for out-of-order 308 // architectures where we do not use the PostRA scheduler anymore (for now 309 // restricted to swift). 310 return getSchedModel().isOutOfOrder() && isSwift(); 311 } 312 313 // This overrides the PostRAScheduler bit in the SchedModel for any CPU. 314 bool ARMSubtarget::enablePostRAScheduler() const { 315 // No need for PostRA scheduling on out of order CPUs (for now restricted to 316 // swift). 317 if (getSchedModel().isOutOfOrder() && isSwift()) 318 return false; 319 return (!isThumb() || hasThumb2()); 320 } 321 322 bool ARMSubtarget::enableAtomicExpand() const { 323 return hasAnyDataBarrier() && (!isThumb() || hasV8MBaselineOps()); 324 } 325 326 bool ARMSubtarget::useStride4VFPs(const MachineFunction &MF) const { 327 // For general targets, the prologue can grow when VFPs are allocated with 328 // stride 4 (more vpush instructions). But WatchOS uses a compact unwind 329 // format which it's more important to get right. 330 return isTargetWatchABI() || (isSwift() && !MF.getFunction()->optForMinSize()); 331 } 332 333 bool ARMSubtarget::useMovt(const MachineFunction &MF) const { 334 // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit 335 // immediates as it is inherently position independent, and may be out of 336 // range otherwise. 337 return !NoMovt && hasV8MBaselineOps() && 338 (isTargetWindows() || !MF.getFunction()->optForMinSize()); 339 } 340 341 bool ARMSubtarget::useFastISel() const { 342 // Enable fast-isel for any target, for testing only. 343 if (ForceFastISel) 344 return true; 345 346 // Limit fast-isel to the targets that are or have been tested. 347 if (!hasV6Ops()) 348 return false; 349 350 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl. 351 return TM.Options.EnableFastISel && 352 ((isTargetMachO() && !isThumb1Only()) || 353 (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb())); 354 } 355