Home | History | Annotate | Download | only in ARM
      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 LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H
     15 #define LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H
     16 
     17 
     18 #include "ARMFrameLowering.h"
     19 #include "ARMISelLowering.h"
     20 #include "ARMInstrInfo.h"
     21 #include "ARMSelectionDAGInfo.h"
     22 #include "ARMSubtarget.h"
     23 #include "MCTargetDesc/ARMMCTargetDesc.h"
     24 #include "Thumb1FrameLowering.h"
     25 #include "Thumb1InstrInfo.h"
     26 #include "Thumb2InstrInfo.h"
     27 #include "llvm/ADT/Triple.h"
     28 #include "llvm/IR/DataLayout.h"
     29 #include "llvm/MC/MCInstrItineraries.h"
     30 #include "llvm/Target/TargetSubtargetInfo.h"
     31 #include <string>
     32 
     33 #define GET_SUBTARGETINFO_HEADER
     34 #include "ARMGenSubtargetInfo.inc"
     35 
     36 namespace llvm {
     37 class GlobalValue;
     38 class StringRef;
     39 class TargetOptions;
     40 class ARMBaseTargetMachine;
     41 
     42 class ARMSubtarget : public ARMGenSubtargetInfo {
     43 protected:
     44   enum ARMProcFamilyEnum {
     45     Others, CortexA5, CortexA7, CortexA8, CortexA9, CortexA12, CortexA15,
     46     CortexA17, CortexR4, CortexR4F, CortexR5, Swift, CortexA53, CortexA57, Krait,
     47   };
     48   enum ARMProcClassEnum {
     49     None, AClass, RClass, MClass
     50   };
     51 
     52   /// ARMProcFamily - ARM processor family: Cortex-A8, Cortex-A9, and others.
     53   ARMProcFamilyEnum ARMProcFamily;
     54 
     55   /// ARMProcClass - ARM processor class: None, AClass, RClass or MClass.
     56   ARMProcClassEnum ARMProcClass;
     57 
     58   /// HasV4TOps, HasV5TOps, HasV5TEOps,
     59   /// HasV6Ops, HasV6MOps, HasV6KOps, HasV6T2Ops, HasV7Ops, HasV8Ops -
     60   /// Specify whether target support specific ARM ISA variants.
     61   bool HasV4TOps;
     62   bool HasV5TOps;
     63   bool HasV5TEOps;
     64   bool HasV6Ops;
     65   bool HasV6MOps;
     66   bool HasV6KOps;
     67   bool HasV6T2Ops;
     68   bool HasV7Ops;
     69   bool HasV8Ops;
     70   bool HasV8_1aOps;
     71 
     72   /// HasVFPv2, HasVFPv3, HasVFPv4, HasFPARMv8, HasNEON - Specify what
     73   /// floating point ISAs are supported.
     74   bool HasVFPv2;
     75   bool HasVFPv3;
     76   bool HasVFPv4;
     77   bool HasFPARMv8;
     78   bool HasNEON;
     79 
     80   /// UseNEONForSinglePrecisionFP - if the NEONFP attribute has been
     81   /// specified. Use the method useNEONForSinglePrecisionFP() to
     82   /// determine if NEON should actually be used.
     83   bool UseNEONForSinglePrecisionFP;
     84 
     85   /// UseMulOps - True if non-microcoded fused integer multiply-add and
     86   /// multiply-subtract instructions should be used.
     87   bool UseMulOps;
     88 
     89   /// SlowFPVMLx - If the VFP2 / NEON instructions are available, indicates
     90   /// whether the FP VML[AS] instructions are slow (if so, don't use them).
     91   bool SlowFPVMLx;
     92 
     93   /// HasVMLxForwarding - If true, NEON has special multiplier accumulator
     94   /// forwarding to allow mul + mla being issued back to back.
     95   bool HasVMLxForwarding;
     96 
     97   /// SlowFPBrcc - True if floating point compare + branch is slow.
     98   bool SlowFPBrcc;
     99 
    100   /// InThumbMode - True if compiling for Thumb, false for ARM.
    101   bool InThumbMode;
    102 
    103   /// HasThumb2 - True if Thumb2 instructions are supported.
    104   bool HasThumb2;
    105 
    106   /// NoARM - True if subtarget does not support ARM mode execution.
    107   bool NoARM;
    108 
    109   /// IsR9Reserved - True if R9 is a not available as general purpose register.
    110   bool IsR9Reserved;
    111 
    112   /// UseMovt - True if MOVT / MOVW pairs are used for materialization of 32-bit
    113   /// imms (including global addresses).
    114   bool UseMovt;
    115 
    116   /// SupportsTailCall - True if the OS supports tail call. The dynamic linker
    117   /// must be able to synthesize call stubs for interworking between ARM and
    118   /// Thumb.
    119   bool SupportsTailCall;
    120 
    121   /// HasFP16 - True if subtarget supports half-precision FP (We support VFP+HF
    122   /// only so far)
    123   bool HasFP16;
    124 
    125   /// HasD16 - True if subtarget is limited to 16 double precision
    126   /// FP registers for VFPv3.
    127   bool HasD16;
    128 
    129   /// HasHardwareDivide - True if subtarget supports [su]div
    130   bool HasHardwareDivide;
    131 
    132   /// HasHardwareDivideInARM - True if subtarget supports [su]div in ARM mode
    133   bool HasHardwareDivideInARM;
    134 
    135   /// HasT2ExtractPack - True if subtarget supports thumb2 extract/pack
    136   /// instructions.
    137   bool HasT2ExtractPack;
    138 
    139   /// HasDataBarrier - True if the subtarget supports DMB / DSB data barrier
    140   /// instructions.
    141   bool HasDataBarrier;
    142 
    143   /// Pref32BitThumb - If true, codegen would prefer 32-bit Thumb instructions
    144   /// over 16-bit ones.
    145   bool Pref32BitThumb;
    146 
    147   /// AvoidCPSRPartialUpdate - If true, codegen would avoid using instructions
    148   /// that partially update CPSR and add false dependency on the previous
    149   /// CPSR setting instruction.
    150   bool AvoidCPSRPartialUpdate;
    151 
    152   /// AvoidMOVsShifterOperand - If true, codegen should avoid using flag setting
    153   /// movs with shifter operand (i.e. asr, lsl, lsr).
    154   bool AvoidMOVsShifterOperand;
    155 
    156   /// HasRAS - Some processors perform return stack prediction. CodeGen should
    157   /// avoid issue "normal" call instructions to callees which do not return.
    158   bool HasRAS;
    159 
    160   /// HasMPExtension - True if the subtarget supports Multiprocessing
    161   /// extension (ARMv7 only).
    162   bool HasMPExtension;
    163 
    164   /// HasVirtualization - True if the subtarget supports the Virtualization
    165   /// extension.
    166   bool HasVirtualization;
    167 
    168   /// FPOnlySP - If true, the floating point unit only supports single
    169   /// precision.
    170   bool FPOnlySP;
    171 
    172   /// If true, the processor supports the Performance Monitor Extensions. These
    173   /// include a generic cycle-counter as well as more fine-grained (often
    174   /// implementation-specific) events.
    175   bool HasPerfMon;
    176 
    177   /// HasTrustZone - if true, processor supports TrustZone security extensions
    178   bool HasTrustZone;
    179 
    180   /// HasCrypto - if true, processor supports Cryptography extensions
    181   bool HasCrypto;
    182 
    183   /// HasCRC - if true, processor supports CRC instructions
    184   bool HasCRC;
    185 
    186   /// If true, the instructions "vmov.i32 d0, #0" and "vmov.i32 q0, #0" are
    187   /// particularly effective at zeroing a VFP register.
    188   bool HasZeroCycleZeroing;
    189 
    190   /// AllowsUnalignedMem - If true, the subtarget allows unaligned memory
    191   /// accesses for some types.  For details, see
    192   /// ARMTargetLowering::allowsMisalignedMemoryAccesses().
    193   bool AllowsUnalignedMem;
    194 
    195   /// RestrictIT - If true, the subtarget disallows generation of deprecated IT
    196   ///  blocks to conform to ARMv8 rule.
    197   bool RestrictIT;
    198 
    199   /// Thumb2DSP - If true, the subtarget supports the v7 DSP (saturating arith
    200   /// and such) instructions in Thumb2 code.
    201   bool Thumb2DSP;
    202 
    203   /// NaCl TRAP instruction is generated instead of the regular TRAP.
    204   bool UseNaClTrap;
    205 
    206   /// Force long to be a 64-bit type (RenderScript-specific)
    207   bool UseLong64;
    208 
    209   /// Target machine allowed unsafe FP math (such as use of NEON fp)
    210   bool UnsafeFPMath;
    211 
    212   /// stackAlignment - The minimum alignment known to hold of the stack frame on
    213   /// entry to the function and which must be maintained by every function.
    214   unsigned stackAlignment;
    215 
    216   /// CPUString - String name of used CPU.
    217   std::string CPUString;
    218 
    219   /// IsLittle - The target is Little Endian
    220   bool IsLittle;
    221 
    222   /// TargetTriple - What processor and OS we're targeting.
    223   Triple TargetTriple;
    224 
    225   /// SchedModel - Processor specific instruction costs.
    226   MCSchedModel SchedModel;
    227 
    228   /// Selected instruction itineraries (one entry per itinerary class.)
    229   InstrItineraryData InstrItins;
    230 
    231   /// Options passed via command line that could influence the target
    232   const TargetOptions &Options;
    233 
    234   const ARMBaseTargetMachine &TM;
    235 
    236 public:
    237   /// This constructor initializes the data members to match that
    238   /// of the specified triple.
    239   ///
    240   ARMSubtarget(const std::string &TT, const std::string &CPU,
    241                const std::string &FS, const ARMBaseTargetMachine &TM, bool IsLittle);
    242 
    243   /// getMaxInlineSizeThreshold - Returns the maximum memset / memcpy size
    244   /// that still makes it profitable to inline the call.
    245   unsigned getMaxInlineSizeThreshold() const {
    246     return 64;
    247   }
    248   /// ParseSubtargetFeatures - Parses features string setting specified
    249   /// subtarget options.  Definition of function is auto generated by tblgen.
    250   void ParseSubtargetFeatures(StringRef CPU, StringRef FS);
    251 
    252   /// initializeSubtargetDependencies - Initializes using a CPU and feature string
    253   /// so that we can use initializer lists for subtarget initialization.
    254   ARMSubtarget &initializeSubtargetDependencies(StringRef CPU, StringRef FS);
    255 
    256   const ARMSelectionDAGInfo *getSelectionDAGInfo() const override {
    257     return &TSInfo;
    258   }
    259   const ARMBaseInstrInfo *getInstrInfo() const override {
    260     return InstrInfo.get();
    261   }
    262   const ARMTargetLowering *getTargetLowering() const override {
    263     return &TLInfo;
    264   }
    265   const ARMFrameLowering *getFrameLowering() const override {
    266     return FrameLowering.get();
    267   }
    268   const ARMBaseRegisterInfo *getRegisterInfo() const override {
    269     return &InstrInfo->getRegisterInfo();
    270   }
    271 
    272 private:
    273   ARMSelectionDAGInfo TSInfo;
    274   // Either Thumb1FrameLowering or ARMFrameLowering.
    275   std::unique_ptr<ARMFrameLowering> FrameLowering;
    276   // Either Thumb1InstrInfo or Thumb2InstrInfo.
    277   std::unique_ptr<ARMBaseInstrInfo> InstrInfo;
    278   ARMTargetLowering   TLInfo;
    279 
    280   void initializeEnvironment();
    281   void initSubtargetFeatures(StringRef CPU, StringRef FS);
    282   ARMFrameLowering *initializeFrameLowering(StringRef CPU, StringRef FS);
    283 
    284 public:
    285   void computeIssueWidth();
    286 
    287   bool hasV4TOps()  const { return HasV4TOps;  }
    288   bool hasV5TOps()  const { return HasV5TOps;  }
    289   bool hasV5TEOps() const { return HasV5TEOps; }
    290   bool hasV6Ops()   const { return HasV6Ops;   }
    291   bool hasV6MOps()  const { return HasV6MOps;  }
    292   bool hasV6KOps()  const { return HasV6KOps; }
    293   bool hasV6T2Ops() const { return HasV6T2Ops; }
    294   bool hasV7Ops()   const { return HasV7Ops;  }
    295   bool hasV8Ops()   const { return HasV8Ops;  }
    296   bool hasV8_1aOps() const { return HasV8_1aOps; }
    297 
    298   bool isCortexA5() const { return ARMProcFamily == CortexA5; }
    299   bool isCortexA7() const { return ARMProcFamily == CortexA7; }
    300   bool isCortexA8() const { return ARMProcFamily == CortexA8; }
    301   bool isCortexA9() const { return ARMProcFamily == CortexA9; }
    302   bool isCortexA15() const { return ARMProcFamily == CortexA15; }
    303   bool isSwift()    const { return ARMProcFamily == Swift; }
    304   bool isCortexM3() const { return CPUString == "cortex-m3"; }
    305   bool isLikeA9() const { return isCortexA9() || isCortexA15() || isKrait(); }
    306   bool isCortexR5() const { return ARMProcFamily == CortexR5; }
    307   bool isKrait() const { return ARMProcFamily == Krait; }
    308 
    309   bool hasARMOps() const { return !NoARM; }
    310 
    311   bool hasVFP2() const { return HasVFPv2; }
    312   bool hasVFP3() const { return HasVFPv3; }
    313   bool hasVFP4() const { return HasVFPv4; }
    314   bool hasFPARMv8() const { return HasFPARMv8; }
    315   bool hasNEON() const { return HasNEON;  }
    316   bool hasCrypto() const { return HasCrypto; }
    317   bool hasCRC() const { return HasCRC; }
    318   bool hasVirtualization() const { return HasVirtualization; }
    319   bool useNEONForSinglePrecisionFP() const {
    320     return hasNEON() && UseNEONForSinglePrecisionFP;
    321   }
    322 
    323   bool hasDivide() const { return HasHardwareDivide; }
    324   bool hasDivideInARMMode() const { return HasHardwareDivideInARM; }
    325   bool hasT2ExtractPack() const { return HasT2ExtractPack; }
    326   bool hasDataBarrier() const { return HasDataBarrier; }
    327   bool hasAnyDataBarrier() const {
    328     return HasDataBarrier || (hasV6Ops() && !isThumb());
    329   }
    330   bool useMulOps() const { return UseMulOps; }
    331   bool useFPVMLx() const { return !SlowFPVMLx; }
    332   bool hasVMLxForwarding() const { return HasVMLxForwarding; }
    333   bool isFPBrccSlow() const { return SlowFPBrcc; }
    334   bool isFPOnlySP() const { return FPOnlySP; }
    335   bool hasPerfMon() const { return HasPerfMon; }
    336   bool hasTrustZone() const { return HasTrustZone; }
    337   bool hasZeroCycleZeroing() const { return HasZeroCycleZeroing; }
    338   bool prefers32BitThumb() const { return Pref32BitThumb; }
    339   bool avoidCPSRPartialUpdate() const { return AvoidCPSRPartialUpdate; }
    340   bool avoidMOVsShifterOperand() const { return AvoidMOVsShifterOperand; }
    341   bool hasRAS() const { return HasRAS; }
    342   bool hasMPExtension() const { return HasMPExtension; }
    343   bool hasThumb2DSP() const { return Thumb2DSP; }
    344   bool useNaClTrap() const { return UseNaClTrap; }
    345 
    346   bool hasFP16() const { return HasFP16; }
    347   bool hasD16() const { return HasD16; }
    348 
    349   const Triple &getTargetTriple() const { return TargetTriple; }
    350 
    351   bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); }
    352   bool isTargetIOS() const { return TargetTriple.isiOS(); }
    353   bool isTargetLinux() const { return TargetTriple.isOSLinux(); }
    354   bool isTargetNaCl() const { return TargetTriple.isOSNaCl(); }
    355   bool isTargetNetBSD() const { return TargetTriple.isOSNetBSD(); }
    356   bool isTargetWindows() const { return TargetTriple.isOSWindows(); }
    357 
    358   bool isTargetCOFF() const { return TargetTriple.isOSBinFormatCOFF(); }
    359   bool isTargetELF() const { return TargetTriple.isOSBinFormatELF(); }
    360   bool isTargetMachO() const { return TargetTriple.isOSBinFormatMachO(); }
    361 
    362   // ARM EABI is the bare-metal EABI described in ARM ABI documents and
    363   // can be accessed via -target arm-none-eabi. This is NOT GNUEABI.
    364   // FIXME: Add a flag for bare-metal for that target and set Triple::EABI
    365   // even for GNUEABI, so we can make a distinction here and still conform to
    366   // the EABI on GNU (and Android) mode. This requires change in Clang, too.
    367   // FIXME: The Darwin exception is temporary, while we move users to
    368   // "*-*-*-macho" triples as quickly as possible.
    369   bool isTargetAEABI() const {
    370     return (TargetTriple.getEnvironment() == Triple::EABI ||
    371             TargetTriple.getEnvironment() == Triple::EABIHF) &&
    372            !isTargetDarwin() && !isTargetWindows();
    373   }
    374 
    375   // ARM Targets that support EHABI exception handling standard
    376   // Darwin uses SjLj. Other targets might need more checks.
    377   bool isTargetEHABICompatible() const {
    378     return (TargetTriple.getEnvironment() == Triple::EABI ||
    379             TargetTriple.getEnvironment() == Triple::GNUEABI ||
    380             TargetTriple.getEnvironment() == Triple::EABIHF ||
    381             TargetTriple.getEnvironment() == Triple::GNUEABIHF ||
    382             TargetTriple.getEnvironment() == Triple::Android) &&
    383            !isTargetDarwin() && !isTargetWindows();
    384   }
    385 
    386   bool isTargetHardFloat() const {
    387     // FIXME: this is invalid for WindowsCE
    388     return TargetTriple.getEnvironment() == Triple::GNUEABIHF ||
    389            TargetTriple.getEnvironment() == Triple::EABIHF ||
    390            isTargetWindows();
    391   }
    392   bool isTargetAndroid() const {
    393     return TargetTriple.getEnvironment() == Triple::Android;
    394   }
    395 
    396   bool isAPCS_ABI() const;
    397   bool isAAPCS_ABI() const;
    398 
    399   bool isThumb() const { return InThumbMode; }
    400   bool isThumb1Only() const { return InThumbMode && !HasThumb2; }
    401   bool isThumb2() const { return InThumbMode && HasThumb2; }
    402   bool hasThumb2() const { return HasThumb2; }
    403   bool isMClass() const { return ARMProcClass == MClass; }
    404   bool isRClass() const { return ARMProcClass == RClass; }
    405   bool isAClass() const { return ARMProcClass == AClass; }
    406 
    407   bool isV6M() const {
    408     return isThumb1Only() && isMClass();
    409   }
    410 
    411   bool isR9Reserved() const { return IsR9Reserved; }
    412 
    413   bool useMovt(const MachineFunction &MF) const;
    414 
    415   bool supportsTailCall() const { return SupportsTailCall; }
    416 
    417   bool allowsUnalignedMem() const { return AllowsUnalignedMem; }
    418 
    419   bool restrictIT() const { return RestrictIT; }
    420 
    421   const std::string & getCPUString() const { return CPUString; }
    422 
    423   bool isLittle() const { return IsLittle; }
    424 
    425   unsigned getMispredictionPenalty() const;
    426 
    427   /// This function returns true if the target has sincos() routine in its
    428   /// compiler runtime or math libraries.
    429   bool hasSinCos() const;
    430 
    431   /// True for some subtargets at > -O0.
    432   bool enablePostMachineScheduler() const override;
    433 
    434   // enableAtomicExpand- True if we need to expand our atomics.
    435   bool enableAtomicExpand() const override;
    436 
    437   /// getInstrItins - Return the instruction itineraries based on subtarget
    438   /// selection.
    439   const InstrItineraryData *getInstrItineraryData() const override {
    440     return &InstrItins;
    441   }
    442 
    443   /// getStackAlignment - Returns the minimum alignment known to hold of the
    444   /// stack frame on entry to the function and which must be maintained by every
    445   /// function for this subtarget.
    446   unsigned getStackAlignment() const { return stackAlignment; }
    447 
    448   /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect
    449   /// symbol.
    450   bool GVIsIndirectSymbol(const GlobalValue *GV, Reloc::Model RelocM) const;
    451 
    452 };
    453 } // End llvm namespace
    454 
    455 #endif  // ARMSUBTARGET_H
    456