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 35 using namespace llvm; 36 37 #define DEBUG_TYPE "arm-subtarget" 38 39 #define GET_SUBTARGETINFO_TARGET_DESC 40 #define GET_SUBTARGETINFO_CTOR 41 #include "ARMGenSubtargetInfo.inc" 42 43 static cl::opt<bool> 44 UseFusedMulOps("arm-use-mulops", 45 cl::init(true), cl::Hidden); 46 47 enum ITMode { 48 DefaultIT, 49 RestrictedIT, 50 NoRestrictedIT 51 }; 52 53 static cl::opt<ITMode> 54 IT(cl::desc("IT block support"), cl::Hidden, cl::init(DefaultIT), 55 cl::ZeroOrMore, 56 cl::values(clEnumValN(DefaultIT, "arm-default-it", 57 "Generate IT block based on arch"), 58 clEnumValN(RestrictedIT, "arm-restrict-it", 59 "Disallow deprecated IT based on ARMv8"), 60 clEnumValN(NoRestrictedIT, "arm-no-restrict-it", 61 "Allow IT blocks based on ARMv7"), 62 clEnumValEnd)); 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), ARMProcFamily(Others), 92 ARMProcClass(None), ARMArch(ARMv4t), stackAlignment(4), CPUString(CPU), 93 IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), TM(TM), 94 FrameLowering(initializeFrameLowering(CPU, FS)), 95 // At this point initializeSubtargetDependencies has been called so 96 // we can query directly. 97 InstrInfo(isThumb1Only() 98 ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this) 99 : !isThumb() 100 ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this) 101 : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)), 102 TLInfo(TM, *this) {} 103 104 void ARMSubtarget::initializeEnvironment() { 105 HasV4TOps = false; 106 HasV5TOps = false; 107 HasV5TEOps = false; 108 HasV6Ops = false; 109 HasV6MOps = false; 110 HasV6KOps = false; 111 HasV6T2Ops = false; 112 HasV7Ops = false; 113 HasV8Ops = false; 114 HasV8_1aOps = false; 115 HasVFPv2 = false; 116 HasVFPv3 = false; 117 HasVFPv4 = false; 118 HasFPARMv8 = false; 119 HasNEON = false; 120 UseNEONForSinglePrecisionFP = false; 121 UseMulOps = UseFusedMulOps; 122 SlowFPVMLx = false; 123 HasVMLxForwarding = false; 124 SlowFPBrcc = false; 125 InThumbMode = false; 126 UseSoftFloat = false; 127 HasThumb2 = false; 128 NoARM = false; 129 ReserveR9 = false; 130 NoMovt = false; 131 SupportsTailCall = false; 132 HasFP16 = false; 133 HasD16 = false; 134 HasHardwareDivide = false; 135 HasHardwareDivideInARM = false; 136 HasT2ExtractPack = false; 137 HasDataBarrier = false; 138 Pref32BitThumb = false; 139 AvoidCPSRPartialUpdate = false; 140 AvoidMOVsShifterOperand = false; 141 HasRAS = false; 142 HasMPExtension = false; 143 HasVirtualization = false; 144 FPOnlySP = false; 145 HasPerfMon = false; 146 HasTrustZone = false; 147 HasCrypto = false; 148 HasCRC = false; 149 HasZeroCycleZeroing = false; 150 StrictAlign = false; 151 HasDSP = false; 152 UseNaClTrap = false; 153 GenLongCalls = false; 154 UnsafeFPMath = false; 155 156 // MCAsmInfo isn't always present (e.g. in opt) so we can't initialize this 157 // directly from it, but we can try to make sure they're consistent when both 158 // available. 159 UseSjLjEH = isTargetDarwin() && !isTargetWatchOS(); 160 assert((!TM.getMCAsmInfo() || 161 (TM.getMCAsmInfo()->getExceptionHandlingType() == 162 ExceptionHandling::SjLj) == UseSjLjEH) && 163 "inconsistent sjlj choice between CodeGen and MC"); 164 } 165 166 void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 167 if (CPUString.empty()) { 168 CPUString = "generic"; 169 170 if (isTargetDarwin()) { 171 StringRef ArchName = TargetTriple.getArchName(); 172 if (ArchName.endswith("v7s")) 173 // Default to the Swift CPU when targeting armv7s/thumbv7s. 174 CPUString = "swift"; 175 else if (ArchName.endswith("v7k")) 176 // Default to the Cortex-a7 CPU when targeting armv7k/thumbv7k. 177 // ARMv7k does not use SjLj exception handling. 178 CPUString = "cortex-a7"; 179 } 180 } 181 182 // Insert the architecture feature derived from the target triple into the 183 // feature string. This is important for setting features that are implied 184 // based on the architecture version. 185 std::string ArchFS = ARM_MC::ParseARMTriple(TargetTriple, CPUString); 186 if (!FS.empty()) { 187 if (!ArchFS.empty()) 188 ArchFS = (Twine(ArchFS) + "," + FS).str(); 189 else 190 ArchFS = FS; 191 } 192 ParseSubtargetFeatures(CPUString, ArchFS); 193 194 // FIXME: This used enable V6T2 support implicitly for Thumb2 mode. 195 // Assert this for now to make the change obvious. 196 assert(hasV6T2Ops() || !hasThumb2()); 197 198 // Keep a pointer to static instruction cost data for the specified CPU. 199 SchedModel = getSchedModelForCPU(CPUString); 200 201 // Initialize scheduling itinerary for the specified CPU. 202 InstrItins = getInstrItineraryForCPU(CPUString); 203 204 // FIXME: this is invalid for WindowsCE 205 if (isTargetWindows()) 206 NoARM = true; 207 208 if (isAAPCS_ABI()) 209 stackAlignment = 8; 210 if (isTargetNaCl() || isAAPCS16_ABI()) 211 stackAlignment = 16; 212 213 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo:: 214 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 215 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 216 // support in the assembler and linker to be used. This would need to be 217 // fixed to fully support tail calls in Thumb1. 218 // 219 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 220 // LR. This means if we need to reload LR, it takes an extra instructions, 221 // which outweighs the value of the tail call; but here we don't know yet 222 // whether LR is going to be used. Probably the right approach is to 223 // generate the tail call here and turn it back into CALL/RET in 224 // emitEpilogue if LR is used. 225 226 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 227 // but we need to make sure there are enough registers; the only valid 228 // registers are the 4 used for parameters. We don't currently do this 229 // case. 230 231 SupportsTailCall = !isThumb1Only(); 232 233 if (isTargetMachO() && isTargetIOS() && getTargetTriple().isOSVersionLT(5, 0)) 234 SupportsTailCall = false; 235 236 switch (IT) { 237 case DefaultIT: 238 RestrictIT = hasV8Ops(); 239 break; 240 case RestrictedIT: 241 RestrictIT = true; 242 break; 243 case NoRestrictedIT: 244 RestrictIT = false; 245 break; 246 } 247 248 // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default. 249 const FeatureBitset &Bits = getFeatureBits(); 250 if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters 251 (Options.UnsafeFPMath || isTargetDarwin())) 252 UseNEONForSinglePrecisionFP = true; 253 } 254 255 bool ARMSubtarget::isAPCS_ABI() const { 256 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 257 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS; 258 } 259 bool ARMSubtarget::isAAPCS_ABI() const { 260 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 261 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS || 262 TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 263 } 264 bool ARMSubtarget::isAAPCS16_ABI() const { 265 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 266 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 267 } 268 269 270 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol. 271 bool 272 ARMSubtarget::GVIsIndirectSymbol(const GlobalValue *GV, 273 Reloc::Model RelocM) const { 274 if (RelocM == Reloc::Static) 275 return false; 276 277 bool isDef = GV->isStrongDefinitionForLinker(); 278 279 if (!isTargetMachO()) { 280 // Extra load is needed for all externally visible. 281 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility()) 282 return false; 283 return true; 284 } else { 285 // If this is a strong reference to a definition, it is definitely not 286 // through a stub. 287 if (isDef) 288 return false; 289 290 // Unless we have a symbol with hidden visibility, we have to go through a 291 // normal $non_lazy_ptr stub because this symbol might be resolved late. 292 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference. 293 return true; 294 295 if (RelocM == Reloc::PIC_) { 296 // If symbol visibility is hidden, we have a stub for common symbol 297 // references and external declarations. 298 if (GV->isDeclarationForLinker() || GV->hasCommonLinkage()) 299 // Hidden $non_lazy_ptr reference. 300 return true; 301 } 302 } 303 304 return false; 305 } 306 307 unsigned ARMSubtarget::getMispredictionPenalty() const { 308 return SchedModel.MispredictPenalty; 309 } 310 311 bool ARMSubtarget::hasSinCos() const { 312 return isTargetWatchOS() || 313 (isTargetIOS() && !getTargetTriple().isOSVersionLT(7, 0)); 314 } 315 316 bool ARMSubtarget::enableMachineScheduler() const { 317 // Enable the MachineScheduler before register allocation for out-of-order 318 // architectures where we do not use the PostRA scheduler anymore (for now 319 // restricted to swift). 320 return getSchedModel().isOutOfOrder() && isSwift(); 321 } 322 323 // This overrides the PostRAScheduler bit in the SchedModel for any CPU. 324 bool ARMSubtarget::enablePostRAScheduler() const { 325 // No need for PostRA scheduling on out of order CPUs (for now restricted to 326 // swift). 327 if (getSchedModel().isOutOfOrder() && isSwift()) 328 return false; 329 return (!isThumb() || hasThumb2()); 330 } 331 332 bool ARMSubtarget::enableAtomicExpand() const { 333 return hasAnyDataBarrier() && !isThumb1Only(); 334 } 335 336 bool ARMSubtarget::useStride4VFPs(const MachineFunction &MF) const { 337 // For general targets, the prologue can grow when VFPs are allocated with 338 // stride 4 (more vpush instructions). But WatchOS uses a compact unwind 339 // format which it's more important to get right. 340 return isTargetWatchOS() || (isSwift() && !MF.getFunction()->optForMinSize()); 341 } 342 343 bool ARMSubtarget::useMovt(const MachineFunction &MF) const { 344 // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit 345 // immediates as it is inherently position independent, and may be out of 346 // range otherwise. 347 return !NoMovt && hasV6T2Ops() && 348 (isTargetWindows() || !MF.getFunction()->optForMinSize()); 349 } 350 351 bool ARMSubtarget::useFastISel() const { 352 // Enable fast-isel for any target, for testing only. 353 if (ForceFastISel) 354 return true; 355 356 // Limit fast-isel to the targets that are or have been tested. 357 if (!hasV6Ops()) 358 return false; 359 360 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl. 361 return TM.Options.EnableFastISel && 362 ((isTargetMachO() && !isThumb1Only()) || 363 (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb())); 364 } 365