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), UseMulOps(UseFusedMulOps), 93 CPUString(CPU), IsLittle(IsLittle), TargetTriple(TT), Options(TM.Options), 94 TM(TM), 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 // MCAsmInfo isn't always present (e.g. in opt) so we can't initialize this 106 // directly from it, but we can try to make sure they're consistent when both 107 // available. 108 UseSjLjEH = isTargetDarwin() && !isTargetWatchABI(); 109 assert((!TM.getMCAsmInfo() || 110 (TM.getMCAsmInfo()->getExceptionHandlingType() == 111 ExceptionHandling::SjLj) == UseSjLjEH) && 112 "inconsistent sjlj choice between CodeGen and MC"); 113 } 114 115 void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) { 116 if (CPUString.empty()) { 117 CPUString = "generic"; 118 119 if (isTargetDarwin()) { 120 StringRef ArchName = TargetTriple.getArchName(); 121 if (ArchName.endswith("v7s")) 122 // Default to the Swift CPU when targeting armv7s/thumbv7s. 123 CPUString = "swift"; 124 else if (ArchName.endswith("v7k")) 125 // Default to the Cortex-a7 CPU when targeting armv7k/thumbv7k. 126 // ARMv7k does not use SjLj exception handling. 127 CPUString = "cortex-a7"; 128 } 129 } 130 131 // Insert the architecture feature derived from the target triple into the 132 // feature string. This is important for setting features that are implied 133 // based on the architecture version. 134 std::string ArchFS = ARM_MC::ParseARMTriple(TargetTriple, CPUString); 135 if (!FS.empty()) { 136 if (!ArchFS.empty()) 137 ArchFS = (Twine(ArchFS) + "," + FS).str(); 138 else 139 ArchFS = FS; 140 } 141 ParseSubtargetFeatures(CPUString, ArchFS); 142 143 // FIXME: This used enable V6T2 support implicitly for Thumb2 mode. 144 // Assert this for now to make the change obvious. 145 assert(hasV6T2Ops() || !hasThumb2()); 146 147 // Keep a pointer to static instruction cost data for the specified CPU. 148 SchedModel = getSchedModelForCPU(CPUString); 149 150 // Initialize scheduling itinerary for the specified CPU. 151 InstrItins = getInstrItineraryForCPU(CPUString); 152 153 // FIXME: this is invalid for WindowsCE 154 if (isTargetWindows()) 155 NoARM = true; 156 157 if (isAAPCS_ABI()) 158 stackAlignment = 8; 159 if (isTargetNaCl() || isAAPCS16_ABI()) 160 stackAlignment = 16; 161 162 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo:: 163 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as 164 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation 165 // support in the assembler and linker to be used. This would need to be 166 // fixed to fully support tail calls in Thumb1. 167 // 168 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take 169 // LR. This means if we need to reload LR, it takes an extra instructions, 170 // which outweighs the value of the tail call; but here we don't know yet 171 // whether LR is going to be used. Probably the right approach is to 172 // generate the tail call here and turn it back into CALL/RET in 173 // emitEpilogue if LR is used. 174 175 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls, 176 // but we need to make sure there are enough registers; the only valid 177 // registers are the 4 used for parameters. We don't currently do this 178 // case. 179 180 SupportsTailCall = !isThumb() || hasV8MBaselineOps(); 181 182 if (isTargetMachO() && isTargetIOS() && getTargetTriple().isOSVersionLT(5, 0)) 183 SupportsTailCall = false; 184 185 switch (IT) { 186 case DefaultIT: 187 RestrictIT = hasV8Ops(); 188 break; 189 case RestrictedIT: 190 RestrictIT = true; 191 break; 192 case NoRestrictedIT: 193 RestrictIT = false; 194 break; 195 } 196 197 // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default. 198 const FeatureBitset &Bits = getFeatureBits(); 199 if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters 200 (Options.UnsafeFPMath || isTargetDarwin())) 201 UseNEONForSinglePrecisionFP = true; 202 203 // FIXME: Teach TableGen to deal with these instead of doing it manually here. 204 switch (ARMProcFamily) { 205 case Others: 206 case CortexA5: 207 break; 208 case CortexA7: 209 LdStMultipleTiming = DoubleIssue; 210 break; 211 case CortexA8: 212 LdStMultipleTiming = DoubleIssue; 213 break; 214 case CortexA9: 215 LdStMultipleTiming = DoubleIssueCheckUnalignedAccess; 216 PreISelOperandLatencyAdjustment = 1; 217 break; 218 case CortexA12: 219 break; 220 case CortexA15: 221 MaxInterleaveFactor = 2; 222 PreISelOperandLatencyAdjustment = 1; 223 break; 224 case CortexA17: 225 case CortexA32: 226 case CortexA35: 227 case CortexA53: 228 case CortexA57: 229 case CortexA72: 230 case CortexA73: 231 case CortexR4: 232 case CortexR4F: 233 case CortexR5: 234 case CortexR7: 235 case CortexM3: 236 case ExynosM1: 237 break; 238 case Krait: 239 PreISelOperandLatencyAdjustment = 1; 240 break; 241 case Swift: 242 MaxInterleaveFactor = 2; 243 LdStMultipleTiming = SingleIssuePlusExtras; 244 PreISelOperandLatencyAdjustment = 1; 245 break; 246 } 247 } 248 249 bool ARMSubtarget::isAPCS_ABI() const { 250 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 251 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS; 252 } 253 bool ARMSubtarget::isAAPCS_ABI() const { 254 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 255 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS || 256 TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 257 } 258 bool ARMSubtarget::isAAPCS16_ABI() const { 259 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN); 260 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16; 261 } 262 263 bool ARMSubtarget::isGVIndirectSymbol(const GlobalValue *GV) const { 264 if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) 265 return true; 266 267 // 32 bit macho has no relocation for a-b if a is undefined, even if b is in 268 // the section that is being relocated. This means we have to use o load even 269 // for GVs that are known to be local to the dso. 270 if (isTargetDarwin() && TM.isPositionIndependent() && 271 (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) 272 return true; 273 274 return false; 275 } 276 277 unsigned ARMSubtarget::getMispredictionPenalty() const { 278 return SchedModel.MispredictPenalty; 279 } 280 281 bool ARMSubtarget::hasSinCos() const { 282 return isTargetWatchOS() || 283 (isTargetIOS() && !getTargetTriple().isOSVersionLT(7, 0)); 284 } 285 286 bool ARMSubtarget::enableMachineScheduler() const { 287 // Enable the MachineScheduler before register allocation for out-of-order 288 // architectures where we do not use the PostRA scheduler anymore (for now 289 // restricted to swift). 290 return getSchedModel().isOutOfOrder() && isSwift(); 291 } 292 293 // This overrides the PostRAScheduler bit in the SchedModel for any CPU. 294 bool ARMSubtarget::enablePostRAScheduler() const { 295 // No need for PostRA scheduling on out of order CPUs (for now restricted to 296 // swift). 297 if (getSchedModel().isOutOfOrder() && isSwift()) 298 return false; 299 return (!isThumb() || hasThumb2()); 300 } 301 302 bool ARMSubtarget::enableAtomicExpand() const { 303 return hasAnyDataBarrier() && (!isThumb() || hasV8MBaselineOps()); 304 } 305 306 bool ARMSubtarget::useStride4VFPs(const MachineFunction &MF) const { 307 // For general targets, the prologue can grow when VFPs are allocated with 308 // stride 4 (more vpush instructions). But WatchOS uses a compact unwind 309 // format which it's more important to get right. 310 return isTargetWatchABI() || (isSwift() && !MF.getFunction()->optForMinSize()); 311 } 312 313 bool ARMSubtarget::useMovt(const MachineFunction &MF) const { 314 // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit 315 // immediates as it is inherently position independent, and may be out of 316 // range otherwise. 317 return !NoMovt && hasV8MBaselineOps() && 318 (isTargetWindows() || !MF.getFunction()->optForMinSize()); 319 } 320 321 bool ARMSubtarget::useFastISel() const { 322 // Enable fast-isel for any target, for testing only. 323 if (ForceFastISel) 324 return true; 325 326 // Limit fast-isel to the targets that are or have been tested. 327 if (!hasV6Ops()) 328 return false; 329 330 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl. 331 return TM.Options.EnableFastISel && 332 ((isTargetMachO() && !isThumb1Only()) || 333 (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb())); 334 } 335