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/Analysis.h" 26 #include "llvm/CodeGen/MachineRegisterInfo.h" 27 #include "llvm/IR/Attributes.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/GlobalValue.h" 30 #include "llvm/MC/MCAsmInfo.h" 31 #include "llvm/Support/CommandLine.h" 32 #include "llvm/Target/TargetInstrInfo.h" 33 #include "llvm/Target/TargetOptions.h" 34 #include "llvm/Target/TargetRegisterInfo.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 clEnumValEnd)); 64 65 /// ForceFastISel - Use the fast-isel, even for subtargets where it is not 66 /// currently supported (for testing only). 67 static cl::opt<bool> 68 ForceFastISel("arm-force-fast-isel", 69 cl::init(false), cl::Hidden); 70 71 /// initializeSubtargetDependencies - Initializes using a CPU and feature string 72 /// so that we can use initializer lists for subtarget initialization. 73 ARMSubtarget &ARMSubtarget::initializeSubtargetDependencies(StringRef CPU, 74 StringRef FS) { 75 initializeEnvironment(); 76 initSubtargetFeatures(CPU, FS); 77 return *this; 78 } 79 80 ARMFrameLowering *ARMSubtarget::initializeFrameLowering(StringRef CPU, 81 StringRef FS) { 82 ARMSubtarget &STI = initializeSubtargetDependencies(CPU, FS); 83 if (STI.isThumb1Only()) 84 return (ARMFrameLowering *)new Thumb1FrameLowering(STI); 85 86 return new ARMFrameLowering(STI); 87 } 88 89 ARMSubtarget::ARMSubtarget(const Triple &TT, const std::string &CPU, 90 const std::string &FS, 91 const ARMBaseTargetMachine &TM, bool IsLittle) 92 : ARMGenSubtargetInfo(TT, CPU, FS), ARMProcFamily(Others), 93 ARMProcClass(None), ARMArch(ARMv4t), stackAlignment(4), CPUString(CPU), 94 IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), TM(TM), 95 FrameLowering(initializeFrameLowering(CPU, FS)), 96 // At this point initializeSubtargetDependencies has been called so 97 // we can query directly. 98 InstrInfo(isThumb1Only() 99 ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this) 100 : !isThumb() 101 ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this) 102 : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)), 103 TLInfo(TM, *this) {} 104 105 void ARMSubtarget::initializeEnvironment() { 106 HasV4TOps = false; 107 HasV5TOps = false; 108 HasV5TEOps = false; 109 HasV6Ops = false; 110 HasV6MOps = false; 111 HasV6KOps = false; 112 HasV6T2Ops = false; 113 HasV7Ops = false; 114 HasV8Ops = false; 115 HasV8_1aOps = false; 116 HasV8_2aOps = false; 117 HasV8MBaselineOps = false; 118 HasV8MMainlineOps = false; 119 HasVFPv2 = false; 120 HasVFPv3 = false; 121 HasVFPv4 = false; 122 HasFPARMv8 = false; 123 HasNEON = false; 124 UseNEONForSinglePrecisionFP = false; 125 UseMulOps = UseFusedMulOps; 126 SlowFPVMLx = false; 127 HasVMLxForwarding = false; 128 SlowFPBrcc = false; 129 InThumbMode = false; 130 UseSoftFloat = false; 131 HasThumb2 = false; 132 NoARM = false; 133 ReserveR9 = false; 134 NoMovt = false; 135 SupportsTailCall = false; 136 HasFP16 = false; 137 HasFullFP16 = false; 138 HasD16 = false; 139 HasHardwareDivide = false; 140 HasHardwareDivideInARM = false; 141 HasT2ExtractPack = false; 142 HasDataBarrier = false; 143 Pref32BitThumb = false; 144 AvoidCPSRPartialUpdate = false; 145 AvoidMOVsShifterOperand = false; 146 HasRAS = false; 147 HasMPExtension = false; 148 HasVirtualization = false; 149 FPOnlySP = false; 150 HasPerfMon = false; 151 HasTrustZone = false; 152 Has8MSecExt = false; 153 HasCrypto = false; 154 HasCRC = false; 155 HasZeroCycleZeroing = false; 156 StrictAlign = false; 157 HasDSP = false; 158 UseNaClTrap = false; 159 GenLongCalls = false; 160 UnsafeFPMath = false; 161 HasV7Clrex = false; 162 HasAcquireRelease = false; 163 164 // MCAsmInfo isn't always present (e.g. in opt) so we can't initialize this 165 // directly from it, but we can try to make sure they're consistent when both 166 // available. 167 UseSjLjEH = isTargetDarwin() && !isTargetWatchABI(); 168 assert((!TM.getMCAsmInfo() || 169 (TM.getMCAsmInfo()->getExceptionHandlingType() == 170 ExceptionHandling::SjLj) == UseSjLjEH) && 171 "inconsistent sjlj choice between CodeGen and MC"); 172 } 173 174 void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 175 if (CPUString.empty()) { 176 CPUString = "generic"; 177 178 if (isTargetDarwin()) { 179 StringRef ArchName = TargetTriple.getArchName(); 180 if (ArchName.endswith("v7s")) 181 // Default to the Swift CPU when targeting armv7s/thumbv7s. 182 CPUString = "swift"; 183 else if (ArchName.endswith("v7k")) 184 // Default to the Cortex-a7 CPU when targeting armv7k/thumbv7k. 185 // ARMv7k does not use SjLj exception handling. 186 CPUString = "cortex-a7"; 187 } 188 } 189 190 // Insert the architecture feature derived from the target triple into the 191 // feature string. This is important for setting features that are implied 192 // based on the architecture version. 193 std::string ArchFS = ARM_MC::ParseARMTriple(TargetTriple, CPUString); 194 if (!FS.empty()) { 195 if (!ArchFS.empty()) 196 ArchFS = (Twine(ArchFS) + "," + FS).str(); 197 else 198 ArchFS = FS; 199 } 200 ParseSubtargetFeatures(CPUString, ArchFS); 201 202 // FIXME: This used enable V6T2 support implicitly for Thumb2 mode. 203 // Assert this for now to make the change obvious. 204 assert(hasV6T2Ops() || !hasThumb2()); 205 206 // Keep a pointer to static instruction cost data for the specified CPU. 207 SchedModel = getSchedModelForCPU(CPUString); 208 209 // Initialize scheduling itinerary for the specified CPU. 210 InstrItins = getInstrItineraryForCPU(CPUString); 211 212 // FIXME: this is invalid for WindowsCE 213 if (isTargetWindows()) 214 NoARM = true; 215 216 if (isAAPCS_ABI()) 217 stackAlignment = 8; 218 if (isTargetNaCl() || isAAPCS16_ABI()) 219 stackAlignment = 16; 220 221 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo:: 222 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 223 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 224 // support in the assembler and linker to be used. This would need to be 225 // fixed to fully support tail calls in Thumb1. 226 // 227 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 228 // LR. This means if we need to reload LR, it takes an extra instructions, 229 // which outweighs the value of the tail call; but here we don't know yet 230 // whether LR is going to be used. Probably the right approach is to 231 // generate the tail call here and turn it back into CALL/RET in 232 // emitEpilogue if LR is used. 233 234 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 235 // but we need to make sure there are enough registers; the only valid 236 // registers are the 4 used for parameters. We don't currently do this 237 // case. 238 239 SupportsTailCall = !isThumb() || hasV8MBaselineOps(); 240 241 if (isTargetMachO() && isTargetIOS() && getTargetTriple().isOSVersionLT(5, 0)) 242 SupportsTailCall = false; 243 244 switch (IT) { 245 case DefaultIT: 246 RestrictIT = hasV8Ops(); 247 break; 248 case RestrictedIT: 249 RestrictIT = true; 250 break; 251 case NoRestrictedIT: 252 RestrictIT = false; 253 break; 254 } 255 256 // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default. 257 const FeatureBitset &Bits = getFeatureBits(); 258 if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters 259 (Options.UnsafeFPMath || isTargetDarwin())) 260 UseNEONForSinglePrecisionFP = true; 261 } 262 263 bool ARMSubtarget::isAPCS_ABI() const { 264 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 265 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS; 266 } 267 bool ARMSubtarget::isAAPCS_ABI() const { 268 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 269 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS || 270 TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 271 } 272 bool ARMSubtarget::isAAPCS16_ABI() const { 273 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 274 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 275 } 276 277 /// true if the GV will be accessed via an indirect symbol. 278 bool 279 ARMSubtarget::GVIsIndirectSymbol(const GlobalValue *GV, 280 Reloc::Model RelocM) const { 281 if (!shouldAssumeDSOLocal(RelocM, TargetTriple, *GV->getParent(), GV)) 282 return true; 283 284 // 32 bit macho has no relocation for a-b if a is undefined, even if b is in 285 // the section that is being relocated. This means we have to use o load even 286 // for GVs that are known to be local to the dso. 287 if (isTargetDarwin() && RelocM == Reloc::PIC_ && 288 (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) 289 return true; 290 291 return false; 292 } 293 294 unsigned ARMSubtarget::getMispredictionPenalty() const { 295 return SchedModel.MispredictPenalty; 296 } 297 298 bool ARMSubtarget::hasSinCos() const { 299 return isTargetWatchOS() || 300 (isTargetIOS() && !getTargetTriple().isOSVersionLT(7, 0)); 301 } 302 303 bool ARMSubtarget::enableMachineScheduler() const { 304 // Enable the MachineScheduler before register allocation for out-of-order 305 // architectures where we do not use the PostRA scheduler anymore (for now 306 // restricted to swift). 307 return getSchedModel().isOutOfOrder() && isSwift(); 308 } 309 310 // This overrides the PostRAScheduler bit in the SchedModel for any CPU. 311 bool ARMSubtarget::enablePostRAScheduler() const { 312 // No need for PostRA scheduling on out of order CPUs (for now restricted to 313 // swift). 314 if (getSchedModel().isOutOfOrder() && isSwift()) 315 return false; 316 return (!isThumb() || hasThumb2()); 317 } 318 319 bool ARMSubtarget::enableAtomicExpand() const { 320 return hasAnyDataBarrier() && (!isThumb() || hasV8MBaselineOps()); 321 } 322 323 bool ARMSubtarget::useStride4VFPs(const MachineFunction &MF) const { 324 // For general targets, the prologue can grow when VFPs are allocated with 325 // stride 4 (more vpush instructions). But WatchOS uses a compact unwind 326 // format which it's more important to get right. 327 return isTargetWatchABI() || (isSwift() && !MF.getFunction()->optForMinSize()); 328 } 329 330 bool ARMSubtarget::useMovt(const MachineFunction &MF) const { 331 // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit 332 // immediates as it is inherently position independent, and may be out of 333 // range otherwise. 334 return !NoMovt && hasV8MBaselineOps() && 335 (isTargetWindows() || !MF.getFunction()->optForMinSize()); 336 } 337 338 bool ARMSubtarget::useFastISel() const { 339 // Enable fast-isel for any target, for testing only. 340 if (ForceFastISel) 341 return true; 342 343 // Limit fast-isel to the targets that are or have been tested. 344 if (!hasV6Ops()) 345 return false; 346 347 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl. 348 return TM.Options.EnableFastISel && 349 ((isTargetMachO() && !isThumb1Only()) || 350 (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb())); 351 } 352