1 //===-- ARMSubtarget.h - Define Subtarget for the ARM ----------*- C++ -*--===// 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 declares the ARM specific subclass of TargetSubtargetInfo. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef ARMSUBTARGET_H 15 #define ARMSUBTARGET_H 16 17 #include "MCTargetDesc/ARMMCTargetDesc.h" 18 #include "llvm/ADT/Triple.h" 19 #include "llvm/MC/MCInstrItineraries.h" 20 #include "llvm/Target/TargetSubtargetInfo.h" 21 #include <string> 22 23 #define GET_SUBTARGETINFO_HEADER 24 #include "ARMGenSubtargetInfo.inc" 25 26 namespace llvm { 27 class GlobalValue; 28 class StringRef; 29 class TargetOptions; 30 31 class ARMSubtarget : public ARMGenSubtargetInfo { 32 protected: 33 enum ARMProcFamilyEnum { 34 Others, CortexA5, CortexA8, CortexA9, CortexA15, CortexR5, Swift 35 }; 36 37 /// ARMProcFamily - ARM processor family: Cortex-A8, Cortex-A9, and others. 38 ARMProcFamilyEnum ARMProcFamily; 39 40 /// HasV4TOps, HasV5TOps, HasV5TEOps, 41 /// HasV6Ops, HasV6T2Ops, HasV7Ops, HasV8Ops - 42 /// Specify whether target support specific ARM ISA variants. 43 bool HasV4TOps; 44 bool HasV5TOps; 45 bool HasV5TEOps; 46 bool HasV6Ops; 47 bool HasV6T2Ops; 48 bool HasV7Ops; 49 bool HasV8Ops; 50 51 /// HasVFPv2, HasVFPv3, HasVFPv4, HasV8FP, HasNEON - Specify what 52 /// floating point ISAs are supported. 53 bool HasVFPv2; 54 bool HasVFPv3; 55 bool HasVFPv4; 56 bool HasV8FP; 57 bool HasNEON; 58 59 /// UseNEONForSinglePrecisionFP - if the NEONFP attribute has been 60 /// specified. Use the method useNEONForSinglePrecisionFP() to 61 /// determine if NEON should actually be used. 62 bool UseNEONForSinglePrecisionFP; 63 64 /// UseMulOps - True if non-microcoded fused integer multiply-add and 65 /// multiply-subtract instructions should be used. 66 bool UseMulOps; 67 68 /// SlowFPVMLx - If the VFP2 / NEON instructions are available, indicates 69 /// whether the FP VML[AS] instructions are slow (if so, don't use them). 70 bool SlowFPVMLx; 71 72 /// HasVMLxForwarding - If true, NEON has special multiplier accumulator 73 /// forwarding to allow mul + mla being issued back to back. 74 bool HasVMLxForwarding; 75 76 /// SlowFPBrcc - True if floating point compare + branch is slow. 77 bool SlowFPBrcc; 78 79 /// InThumbMode - True if compiling for Thumb, false for ARM. 80 bool InThumbMode; 81 82 /// HasThumb2 - True if Thumb2 instructions are supported. 83 bool HasThumb2; 84 85 /// IsMClass - True if the subtarget belongs to the 'M' profile of CPUs - 86 /// v6m, v7m for example. 87 bool IsMClass; 88 89 /// NoARM - True if subtarget does not support ARM mode execution. 90 bool NoARM; 91 92 /// PostRAScheduler - True if using post-register-allocation scheduler. 93 bool PostRAScheduler; 94 95 /// IsR9Reserved - True if R9 is a not available as general purpose register. 96 bool IsR9Reserved; 97 98 /// UseMovt - True if MOVT / MOVW pairs are used for materialization of 32-bit 99 /// imms (including global addresses). 100 bool UseMovt; 101 102 /// SupportsTailCall - True if the OS supports tail call. The dynamic linker 103 /// must be able to synthesize call stubs for interworking between ARM and 104 /// Thumb. 105 bool SupportsTailCall; 106 107 /// HasFP16 - True if subtarget supports half-precision FP (We support VFP+HF 108 /// only so far) 109 bool HasFP16; 110 111 /// HasD16 - True if subtarget is limited to 16 double precision 112 /// FP registers for VFPv3. 113 bool HasD16; 114 115 /// HasHardwareDivide - True if subtarget supports [su]div 116 bool HasHardwareDivide; 117 118 /// HasHardwareDivideInARM - True if subtarget supports [su]div in ARM mode 119 bool HasHardwareDivideInARM; 120 121 /// HasT2ExtractPack - True if subtarget supports thumb2 extract/pack 122 /// instructions. 123 bool HasT2ExtractPack; 124 125 /// HasDataBarrier - True if the subtarget supports DMB / DSB data barrier 126 /// instructions. 127 bool HasDataBarrier; 128 129 /// Pref32BitThumb - If true, codegen would prefer 32-bit Thumb instructions 130 /// over 16-bit ones. 131 bool Pref32BitThumb; 132 133 /// AvoidCPSRPartialUpdate - If true, codegen would avoid using instructions 134 /// that partially update CPSR and add false dependency on the previous 135 /// CPSR setting instruction. 136 bool AvoidCPSRPartialUpdate; 137 138 /// AvoidMOVsShifterOperand - If true, codegen should avoid using flag setting 139 /// movs with shifter operand (i.e. asr, lsl, lsr). 140 bool AvoidMOVsShifterOperand; 141 142 /// HasRAS - Some processors perform return stack prediction. CodeGen should 143 /// avoid issue "normal" call instructions to callees which do not return. 144 bool HasRAS; 145 146 /// HasMPExtension - True if the subtarget supports Multiprocessing 147 /// extension (ARMv7 only). 148 bool HasMPExtension; 149 150 /// FPOnlySP - If true, the floating point unit only supports single 151 /// precision. 152 bool FPOnlySP; 153 154 /// If true, the processor supports the Performance Monitor Extensions. These 155 /// include a generic cycle-counter as well as more fine-grained (often 156 /// implementation-specific) events. 157 bool HasPerfMon; 158 159 /// HasTrustZone - if true, processor supports TrustZone security extensions 160 bool HasTrustZone; 161 162 /// AllowsUnalignedMem - If true, the subtarget allows unaligned memory 163 /// accesses for some types. For details, see 164 /// ARMTargetLowering::allowsUnalignedMemoryAccesses(). 165 bool AllowsUnalignedMem; 166 167 /// Thumb2DSP - If true, the subtarget supports the v7 DSP (saturating arith 168 /// and such) instructions in Thumb2 code. 169 bool Thumb2DSP; 170 171 /// NaCl TRAP instruction is generated instead of the regular TRAP. 172 bool UseNaClTrap; 173 174 /// Target machine allowed unsafe FP math (such as use of NEON fp) 175 bool UnsafeFPMath; 176 177 /// stackAlignment - The minimum alignment known to hold of the stack frame on 178 /// entry to the function and which must be maintained by every function. 179 unsigned stackAlignment; 180 181 /// CPUString - String name of used CPU. 182 std::string CPUString; 183 184 /// TargetTriple - What processor and OS we're targeting. 185 Triple TargetTriple; 186 187 /// SchedModel - Processor specific instruction costs. 188 const MCSchedModel *SchedModel; 189 190 /// Selected instruction itineraries (one entry per itinerary class.) 191 InstrItineraryData InstrItins; 192 193 /// Options passed via command line that could influence the target 194 const TargetOptions &Options; 195 196 public: 197 enum { 198 ARM_ABI_APCS, 199 ARM_ABI_AAPCS // ARM EABI 200 } TargetABI; 201 202 /// This constructor initializes the data members to match that 203 /// of the specified triple. 204 /// 205 ARMSubtarget(const std::string &TT, const std::string &CPU, 206 const std::string &FS, const TargetOptions &Options); 207 208 /// getMaxInlineSizeThreshold - Returns the maximum memset / memcpy size 209 /// that still makes it profitable to inline the call. getMaxInlineSizeThreshold()210 unsigned getMaxInlineSizeThreshold() const { 211 // FIXME: For now, we don't lower memcpy's to loads / stores for Thumb1. 212 // Change this once Thumb1 ldmia / stmia support is added. 213 return isThumb1Only() ? 0 : 64; 214 } 215 /// ParseSubtargetFeatures - Parses features string setting specified 216 /// subtarget options. Definition of function is auto generated by tblgen. 217 void ParseSubtargetFeatures(StringRef CPU, StringRef FS); 218 219 /// \brief Reset the features for the ARM target. 220 virtual void resetSubtargetFeatures(const MachineFunction *MF); 221 private: 222 void initializeEnvironment(); 223 void resetSubtargetFeatures(StringRef CPU, StringRef FS); 224 public: 225 void computeIssueWidth(); 226 hasV4TOps()227 bool hasV4TOps() const { return HasV4TOps; } hasV5TOps()228 bool hasV5TOps() const { return HasV5TOps; } hasV5TEOps()229 bool hasV5TEOps() const { return HasV5TEOps; } hasV6Ops()230 bool hasV6Ops() const { return HasV6Ops; } hasV6T2Ops()231 bool hasV6T2Ops() const { return HasV6T2Ops; } hasV7Ops()232 bool hasV7Ops() const { return HasV7Ops; } hasV8Ops()233 bool hasV8Ops() const { return HasV8Ops; } 234 isCortexA5()235 bool isCortexA5() const { return ARMProcFamily == CortexA5; } isCortexA8()236 bool isCortexA8() const { return ARMProcFamily == CortexA8; } isCortexA9()237 bool isCortexA9() const { return ARMProcFamily == CortexA9; } isCortexA15()238 bool isCortexA15() const { return ARMProcFamily == CortexA15; } isSwift()239 bool isSwift() const { return ARMProcFamily == Swift; } isCortexM3()240 bool isCortexM3() const { return CPUString == "cortex-m3"; } isLikeA9()241 bool isLikeA9() const { return isCortexA9() || isCortexA15(); } isCortexR5()242 bool isCortexR5() const { return ARMProcFamily == CortexR5; } 243 hasARMOps()244 bool hasARMOps() const { return !NoARM; } 245 hasVFP2()246 bool hasVFP2() const { return HasVFPv2; } hasVFP3()247 bool hasVFP3() const { return HasVFPv3; } hasVFP4()248 bool hasVFP4() const { return HasVFPv4; } hasV8FP()249 bool hasV8FP() const { return HasV8FP; } hasNEON()250 bool hasNEON() const { return HasNEON; } useNEONForSinglePrecisionFP()251 bool useNEONForSinglePrecisionFP() const { 252 return hasNEON() && UseNEONForSinglePrecisionFP; } 253 hasDivide()254 bool hasDivide() const { return HasHardwareDivide; } hasDivideInARMMode()255 bool hasDivideInARMMode() const { return HasHardwareDivideInARM; } hasT2ExtractPack()256 bool hasT2ExtractPack() const { return HasT2ExtractPack; } hasDataBarrier()257 bool hasDataBarrier() const { return HasDataBarrier; } useMulOps()258 bool useMulOps() const { return UseMulOps; } useFPVMLx()259 bool useFPVMLx() const { return !SlowFPVMLx; } hasVMLxForwarding()260 bool hasVMLxForwarding() const { return HasVMLxForwarding; } isFPBrccSlow()261 bool isFPBrccSlow() const { return SlowFPBrcc; } isFPOnlySP()262 bool isFPOnlySP() const { return FPOnlySP; } hasPerfMon()263 bool hasPerfMon() const { return HasPerfMon; } hasTrustZone()264 bool hasTrustZone() const { return HasTrustZone; } prefers32BitThumb()265 bool prefers32BitThumb() const { return Pref32BitThumb; } avoidCPSRPartialUpdate()266 bool avoidCPSRPartialUpdate() const { return AvoidCPSRPartialUpdate; } avoidMOVsShifterOperand()267 bool avoidMOVsShifterOperand() const { return AvoidMOVsShifterOperand; } hasRAS()268 bool hasRAS() const { return HasRAS; } hasMPExtension()269 bool hasMPExtension() const { return HasMPExtension; } hasThumb2DSP()270 bool hasThumb2DSP() const { return Thumb2DSP; } useNaClTrap()271 bool useNaClTrap() const { return UseNaClTrap; } 272 hasFP16()273 bool hasFP16() const { return HasFP16; } hasD16()274 bool hasD16() const { return HasD16; } 275 getTargetTriple()276 const Triple &getTargetTriple() const { return TargetTriple; } 277 isTargetIOS()278 bool isTargetIOS() const { return TargetTriple.getOS() == Triple::IOS; } isTargetDarwin()279 bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); } isTargetNaCl()280 bool isTargetNaCl() const { return TargetTriple.getOS() == Triple::NaCl; } isTargetLinux()281 bool isTargetLinux() const { return TargetTriple.getOS() == Triple::Linux; } isTargetELF()282 bool isTargetELF() const { return !isTargetDarwin(); } 283 // ARM EABI is the bare-metal EABI described in ARM ABI documents and 284 // can be accessed via -target arm-none-eabi. This is NOT GNUEABI. 285 // FIXME: Add a flag for bare-metal for that target and set Triple::EABI 286 // even for GNUEABI, so we can make a distinction here and still conform to 287 // the EABI on GNU (and Android) mode. This requires change in Clang, too. isTargetAEABI()288 bool isTargetAEABI() const { 289 return TargetTriple.getEnvironment() == Triple::EABI; 290 } 291 isAPCS_ABI()292 bool isAPCS_ABI() const { return TargetABI == ARM_ABI_APCS; } isAAPCS_ABI()293 bool isAAPCS_ABI() const { return TargetABI == ARM_ABI_AAPCS; } 294 isThumb()295 bool isThumb() const { return InThumbMode; } isThumb1Only()296 bool isThumb1Only() const { return InThumbMode && !HasThumb2; } isThumb2()297 bool isThumb2() const { return InThumbMode && HasThumb2; } hasThumb2()298 bool hasThumb2() const { return HasThumb2; } isMClass()299 bool isMClass() const { return IsMClass; } isARClass()300 bool isARClass() const { return !IsMClass; } 301 isR9Reserved()302 bool isR9Reserved() const { return IsR9Reserved; } 303 useMovt()304 bool useMovt() const { return UseMovt && hasV6T2Ops(); } supportsTailCall()305 bool supportsTailCall() const { return SupportsTailCall; } 306 allowsUnalignedMem()307 bool allowsUnalignedMem() const { return AllowsUnalignedMem; } 308 getCPUString()309 const std::string & getCPUString() const { return CPUString; } 310 311 unsigned getMispredictionPenalty() const; 312 313 /// enablePostRAScheduler - True at 'More' optimization. 314 bool enablePostRAScheduler(CodeGenOpt::Level OptLevel, 315 TargetSubtargetInfo::AntiDepBreakMode& Mode, 316 RegClassVector& CriticalPathRCs) const; 317 318 /// getInstrItins - Return the instruction itineraies based on subtarget 319 /// selection. getInstrItineraryData()320 const InstrItineraryData &getInstrItineraryData() const { return InstrItins; } 321 322 /// getStackAlignment - Returns the minimum alignment known to hold of the 323 /// stack frame on entry to the function and which must be maintained by every 324 /// function for this subtarget. getStackAlignment()325 unsigned getStackAlignment() const { return stackAlignment; } 326 327 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect 328 /// symbol. 329 bool GVIsIndirectSymbol(const GlobalValue *GV, Reloc::Model RelocM) const; 330 }; 331 } // End llvm namespace 332 333 #endif // ARMSUBTARGET_H 334