ARMSubtarget.h revision 929bdb23794b615dc6b0cc59db21f0450c3ce33b
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 26namespace llvm { 27class GlobalValue; 28class StringRef; 29class TargetOptions; 30 31class ARMSubtarget : public ARMGenSubtargetInfo { 32protected: 33 enum ARMProcFamilyEnum { 34 Others, CortexA5, CortexA8, CortexA9, CortexA15, CortexR5, Swift, CortexA53, CortexA57 35 }; 36 enum ARMProcClassEnum { 37 None, AClass, RClass, MClass 38 }; 39 40 /// ARMProcFamily - ARM processor family: Cortex-A8, Cortex-A9, and others. 41 ARMProcFamilyEnum ARMProcFamily; 42 43 /// ARMProcClass - ARM processor class: None, AClass, RClass or MClass. 44 ARMProcClassEnum ARMProcClass; 45 46 /// HasV4TOps, HasV5TOps, HasV5TEOps, 47 /// HasV6Ops, HasV6MOps, HasV6T2Ops, HasV7Ops, HasV8Ops - 48 /// Specify whether target support specific ARM ISA variants. 49 bool HasV4TOps; 50 bool HasV5TOps; 51 bool HasV5TEOps; 52 bool HasV6Ops; 53 bool HasV6MOps; 54 bool HasV6T2Ops; 55 bool HasV7Ops; 56 bool HasV8Ops; 57 58 /// HasVFPv2, HasVFPv3, HasVFPv4, HasFPARMv8, HasNEON - Specify what 59 /// floating point ISAs are supported. 60 bool HasVFPv2; 61 bool HasVFPv3; 62 bool HasVFPv4; 63 bool HasFPARMv8; 64 bool HasNEON; 65 66 /// UseNEONForSinglePrecisionFP - if the NEONFP attribute has been 67 /// specified. Use the method useNEONForSinglePrecisionFP() to 68 /// determine if NEON should actually be used. 69 bool UseNEONForSinglePrecisionFP; 70 71 /// UseMulOps - True if non-microcoded fused integer multiply-add and 72 /// multiply-subtract instructions should be used. 73 bool UseMulOps; 74 75 /// SlowFPVMLx - If the VFP2 / NEON instructions are available, indicates 76 /// whether the FP VML[AS] instructions are slow (if so, don't use them). 77 bool SlowFPVMLx; 78 79 /// HasVMLxForwarding - If true, NEON has special multiplier accumulator 80 /// forwarding to allow mul + mla being issued back to back. 81 bool HasVMLxForwarding; 82 83 /// SlowFPBrcc - True if floating point compare + branch is slow. 84 bool SlowFPBrcc; 85 86 /// InThumbMode - True if compiling for Thumb, false for ARM. 87 bool InThumbMode; 88 89 /// HasThumb2 - True if Thumb2 instructions are supported. 90 bool HasThumb2; 91 92 /// NoARM - True if subtarget does not support ARM mode execution. 93 bool NoARM; 94 95 /// PostRAScheduler - True if using post-register-allocation scheduler. 96 bool PostRAScheduler; 97 98 /// IsR9Reserved - True if R9 is a not available as general purpose register. 99 bool IsR9Reserved; 100 101 /// UseMovt - True if MOVT / MOVW pairs are used for materialization of 32-bit 102 /// imms (including global addresses). 103 bool UseMovt; 104 105 /// SupportsTailCall - True if the OS supports tail call. The dynamic linker 106 /// must be able to synthesize call stubs for interworking between ARM and 107 /// Thumb. 108 bool SupportsTailCall; 109 110 /// HasFP16 - True if subtarget supports half-precision FP (We support VFP+HF 111 /// only so far) 112 bool HasFP16; 113 114 /// HasD16 - True if subtarget is limited to 16 double precision 115 /// FP registers for VFPv3. 116 bool HasD16; 117 118 /// HasHardwareDivide - True if subtarget supports [su]div 119 bool HasHardwareDivide; 120 121 /// HasHardwareDivideInARM - True if subtarget supports [su]div in ARM mode 122 bool HasHardwareDivideInARM; 123 124 /// HasT2ExtractPack - True if subtarget supports thumb2 extract/pack 125 /// instructions. 126 bool HasT2ExtractPack; 127 128 /// HasDataBarrier - True if the subtarget supports DMB / DSB data barrier 129 /// instructions. 130 bool HasDataBarrier; 131 132 /// Pref32BitThumb - If true, codegen would prefer 32-bit Thumb instructions 133 /// over 16-bit ones. 134 bool Pref32BitThumb; 135 136 /// AvoidCPSRPartialUpdate - If true, codegen would avoid using instructions 137 /// that partially update CPSR and add false dependency on the previous 138 /// CPSR setting instruction. 139 bool AvoidCPSRPartialUpdate; 140 141 /// AvoidMOVsShifterOperand - If true, codegen should avoid using flag setting 142 /// movs with shifter operand (i.e. asr, lsl, lsr). 143 bool AvoidMOVsShifterOperand; 144 145 /// HasRAS - Some processors perform return stack prediction. CodeGen should 146 /// avoid issue "normal" call instructions to callees which do not return. 147 bool HasRAS; 148 149 /// HasMPExtension - True if the subtarget supports Multiprocessing 150 /// extension (ARMv7 only). 151 bool HasMPExtension; 152 153 /// HasVirtualization - True if the subtarget supports the Virtualization 154 /// extension. 155 bool HasVirtualization; 156 157 /// FPOnlySP - If true, the floating point unit only supports single 158 /// precision. 159 bool FPOnlySP; 160 161 /// If true, the processor supports the Performance Monitor Extensions. These 162 /// include a generic cycle-counter as well as more fine-grained (often 163 /// implementation-specific) events. 164 bool HasPerfMon; 165 166 /// HasTrustZone - if true, processor supports TrustZone security extensions 167 bool HasTrustZone; 168 169 /// HasCrypto - if true, processor supports Cryptography extensions 170 bool HasCrypto; 171 172 /// HasCRC - if true, processor supports CRC instructions 173 bool HasCRC; 174 175 /// AllowsUnalignedMem - If true, the subtarget allows unaligned memory 176 /// accesses for some types. For details, see 177 /// ARMTargetLowering::allowsUnalignedMemoryAccesses(). 178 bool AllowsUnalignedMem; 179 180 /// RestrictIT - If true, the subtarget disallows generation of deprecated IT 181 /// blocks to conform to ARMv8 rule. 182 bool RestrictIT; 183 184 /// Thumb2DSP - If true, the subtarget supports the v7 DSP (saturating arith 185 /// and such) instructions in Thumb2 code. 186 bool Thumb2DSP; 187 188 /// NaCl TRAP instruction is generated instead of the regular TRAP. 189 bool UseNaClTrap; 190 191 /// Target machine allowed unsafe FP math (such as use of NEON fp) 192 bool UnsafeFPMath; 193 194 /// stackAlignment - The minimum alignment known to hold of the stack frame on 195 /// entry to the function and which must be maintained by every function. 196 unsigned stackAlignment; 197 198 /// CPUString - String name of used CPU. 199 std::string CPUString; 200 201 /// TargetTriple - What processor and OS we're targeting. 202 Triple TargetTriple; 203 204 /// SchedModel - Processor specific instruction costs. 205 const MCSchedModel *SchedModel; 206 207 /// Selected instruction itineraries (one entry per itinerary class.) 208 InstrItineraryData InstrItins; 209 210 /// Options passed via command line that could influence the target 211 const TargetOptions &Options; 212 213 public: 214 enum { 215 ARM_ABI_APCS, 216 ARM_ABI_AAPCS // ARM EABI 217 } TargetABI; 218 219 /// This constructor initializes the data members to match that 220 /// of the specified triple. 221 /// 222 ARMSubtarget(const std::string &TT, const std::string &CPU, 223 const std::string &FS, const TargetOptions &Options); 224 225 /// getMaxInlineSizeThreshold - Returns the maximum memset / memcpy size 226 /// that still makes it profitable to inline the call. 227 unsigned getMaxInlineSizeThreshold() const { 228 // FIXME: For now, we don't lower memcpy's to loads / stores for Thumb1. 229 // Change this once Thumb1 ldmia / stmia support is added. 230 return isThumb1Only() ? 0 : 64; 231 } 232 /// ParseSubtargetFeatures - Parses features string setting specified 233 /// subtarget options. Definition of function is auto generated by tblgen. 234 void ParseSubtargetFeatures(StringRef CPU, StringRef FS); 235 236 /// \brief Reset the features for the ARM target. 237 virtual void resetSubtargetFeatures(const MachineFunction *MF); 238private: 239 void initializeEnvironment(); 240 void resetSubtargetFeatures(StringRef CPU, StringRef FS); 241public: 242 void computeIssueWidth(); 243 244 bool hasV4TOps() const { return HasV4TOps; } 245 bool hasV5TOps() const { return HasV5TOps; } 246 bool hasV5TEOps() const { return HasV5TEOps; } 247 bool hasV6Ops() const { return HasV6Ops; } 248 bool hasV6MOps() const { return HasV6MOps; } 249 bool hasV6T2Ops() const { return HasV6T2Ops; } 250 bool hasV7Ops() const { return HasV7Ops; } 251 bool hasV8Ops() const { return HasV8Ops; } 252 253 bool isCortexA5() const { return ARMProcFamily == CortexA5; } 254 bool isCortexA8() const { return ARMProcFamily == CortexA8; } 255 bool isCortexA9() const { return ARMProcFamily == CortexA9; } 256 bool isCortexA15() const { return ARMProcFamily == CortexA15; } 257 bool isSwift() const { return ARMProcFamily == Swift; } 258 bool isCortexM3() const { return CPUString == "cortex-m3"; } 259 bool isLikeA9() const { return isCortexA9() || isCortexA15(); } 260 bool isCortexR5() const { return ARMProcFamily == CortexR5; } 261 262 bool hasARMOps() const { return !NoARM; } 263 264 bool hasVFP2() const { return HasVFPv2; } 265 bool hasVFP3() const { return HasVFPv3; } 266 bool hasVFP4() const { return HasVFPv4; } 267 bool hasFPARMv8() const { return HasFPARMv8; } 268 bool hasNEON() const { return HasNEON; } 269 bool hasCrypto() const { return HasCrypto; } 270 bool hasCRC() const { return HasCRC; } 271 bool hasVirtualization() const { return HasVirtualization; } 272 bool useNEONForSinglePrecisionFP() const { 273 return hasNEON() && UseNEONForSinglePrecisionFP; } 274 275 bool hasDivide() const { return HasHardwareDivide; } 276 bool hasDivideInARMMode() const { return HasHardwareDivideInARM; } 277 bool hasT2ExtractPack() const { return HasT2ExtractPack; } 278 bool hasDataBarrier() const { return HasDataBarrier; } 279 bool hasAnyDataBarrier() const { 280 return HasDataBarrier || (hasV6Ops() && !isThumb()); 281 } 282 bool useMulOps() const { return UseMulOps; } 283 bool useFPVMLx() const { return !SlowFPVMLx; } 284 bool hasVMLxForwarding() const { return HasVMLxForwarding; } 285 bool isFPBrccSlow() const { return SlowFPBrcc; } 286 bool isFPOnlySP() const { return FPOnlySP; } 287 bool hasPerfMon() const { return HasPerfMon; } 288 bool hasTrustZone() const { return HasTrustZone; } 289 bool prefers32BitThumb() const { return Pref32BitThumb; } 290 bool avoidCPSRPartialUpdate() const { return AvoidCPSRPartialUpdate; } 291 bool avoidMOVsShifterOperand() const { return AvoidMOVsShifterOperand; } 292 bool hasRAS() const { return HasRAS; } 293 bool hasMPExtension() const { return HasMPExtension; } 294 bool hasThumb2DSP() const { return Thumb2DSP; } 295 bool useNaClTrap() const { return UseNaClTrap; } 296 297 bool hasFP16() const { return HasFP16; } 298 bool hasD16() const { return HasD16; } 299 300 const Triple &getTargetTriple() const { return TargetTriple; } 301 302 bool isTargetIOS() const { return TargetTriple.isiOS(); } 303 bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); } 304 bool isTargetNaCl() const { return TargetTriple.isOSNaCl(); } 305 bool isTargetLinux() const { return TargetTriple.isOSLinux(); } 306 bool isTargetELF() const { return !isTargetDarwin(); } 307 // ARM EABI is the bare-metal EABI described in ARM ABI documents and 308 // can be accessed via -target arm-none-eabi. This is NOT GNUEABI. 309 // FIXME: Add a flag for bare-metal for that target and set Triple::EABI 310 // even for GNUEABI, so we can make a distinction here and still conform to 311 // the EABI on GNU (and Android) mode. This requires change in Clang, too. 312 bool isTargetAEABI() const { 313 return TargetTriple.getEnvironment() == Triple::EABI; 314 } 315 316 bool isAPCS_ABI() const { return TargetABI == ARM_ABI_APCS; } 317 bool isAAPCS_ABI() const { return TargetABI == ARM_ABI_AAPCS; } 318 319 bool isThumb() const { return InThumbMode; } 320 bool isThumb1Only() const { return InThumbMode && !HasThumb2; } 321 bool isThumb2() const { return InThumbMode && HasThumb2; } 322 bool hasThumb2() const { return HasThumb2; } 323 bool isMClass() const { return ARMProcClass == MClass; } 324 bool isRClass() const { return ARMProcClass == RClass; } 325 bool isAClass() const { return ARMProcClass == AClass; } 326 327 bool isR9Reserved() const { return IsR9Reserved; } 328 329 bool useMovt() const { return UseMovt && hasV6T2Ops(); } 330 bool supportsTailCall() const { return SupportsTailCall; } 331 332 bool allowsUnalignedMem() const { return AllowsUnalignedMem; } 333 334 bool restrictIT() const { return RestrictIT; } 335 336 const std::string & getCPUString() const { return CPUString; } 337 338 unsigned getMispredictionPenalty() const; 339 340 /// This function returns true if the target has sincos() routine in its 341 /// compiler runtime or math libraries. 342 bool hasSinCos() const; 343 344 /// enablePostRAScheduler - True at 'More' optimization. 345 bool enablePostRAScheduler(CodeGenOpt::Level OptLevel, 346 TargetSubtargetInfo::AntiDepBreakMode& Mode, 347 RegClassVector& CriticalPathRCs) const; 348 349 /// getInstrItins - Return the instruction itineraies based on subtarget 350 /// selection. 351 const InstrItineraryData &getInstrItineraryData() const { return InstrItins; } 352 353 /// getStackAlignment - Returns the minimum alignment known to hold of the 354 /// stack frame on entry to the function and which must be maintained by every 355 /// function for this subtarget. 356 unsigned getStackAlignment() const { return stackAlignment; } 357 358 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect 359 /// symbol. 360 bool GVIsIndirectSymbol(const GlobalValue *GV, Reloc::Model RelocM) const; 361}; 362} // End llvm namespace 363 364#endif // ARMSUBTARGET_H 365