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 #include "ARMBaseInstrInfo.h"
18 #include "ARMBaseRegisterInfo.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMFrameLowering.h"
21 #include "ARMISelLowering.h"
22 #include "ARMSelectionDAGInfo.h"
23 #include "llvm/ADT/Triple.h"
24 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
25 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
26 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
27 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/TargetSubtargetInfo.h"
30 #include "llvm/MC/MCInstrItineraries.h"
31 #include "llvm/MC/MCSchedule.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include <memory>
34 #include <string>
35 
36 #define GET_SUBTARGETINFO_HEADER
37 #include "ARMGenSubtargetInfo.inc"
38 
39 namespace llvm {
40 
41 class ARMBaseTargetMachine;
42 class GlobalValue;
43 class StringRef;
44 
45 class ARMSubtarget : public ARMGenSubtargetInfo {
46 protected:
47   enum ARMProcFamilyEnum {
48     Others,
49 
50     CortexA12,
51     CortexA15,
52     CortexA17,
53     CortexA32,
54     CortexA35,
55     CortexA5,
56     CortexA53,
57     CortexA55,
58     CortexA57,
59     CortexA7,
60     CortexA72,
61     CortexA73,
62     CortexA75,
63     CortexA8,
64     CortexA9,
65     CortexM3,
66     CortexR4,
67     CortexR4F,
68     CortexR5,
69     CortexR52,
70     CortexR7,
71     Exynos,
72     Krait,
73     Kryo,
74     Swift
75   };
76   enum ARMProcClassEnum {
77     None,
78 
79     AClass,
80     MClass,
81     RClass
82   };
83   enum ARMArchEnum {
84     ARMv2,
85     ARMv2a,
86     ARMv3,
87     ARMv3m,
88     ARMv4,
89     ARMv4t,
90     ARMv5,
91     ARMv5t,
92     ARMv5te,
93     ARMv5tej,
94     ARMv6,
95     ARMv6k,
96     ARMv6kz,
97     ARMv6m,
98     ARMv6sm,
99     ARMv6t2,
100     ARMv7a,
101     ARMv7em,
102     ARMv7m,
103     ARMv7r,
104     ARMv7ve,
105     ARMv81a,
106     ARMv82a,
107     ARMv83a,
108     ARMv84a,
109     ARMv85a,
110     ARMv8a,
111     ARMv8mBaseline,
112     ARMv8mMainline,
113     ARMv8r
114   };
115 
116 public:
117   /// What kind of timing do load multiple/store multiple instructions have.
118   enum ARMLdStMultipleTiming {
119     /// Can load/store 2 registers/cycle.
120     DoubleIssue,
121     /// Can load/store 2 registers/cycle, but needs an extra cycle if the access
122     /// is not 64-bit aligned.
123     DoubleIssueCheckUnalignedAccess,
124     /// Can load/store 1 register/cycle.
125     SingleIssue,
126     /// Can load/store 1 register/cycle, but needs an extra cycle for address
127     /// computation and potentially also for register writeback.
128     SingleIssuePlusExtras,
129   };
130 
131 protected:
132   /// ARMProcFamily - ARM processor family: Cortex-A8, Cortex-A9, and others.
133   ARMProcFamilyEnum ARMProcFamily = Others;
134 
135   /// ARMProcClass - ARM processor class: None, AClass, RClass or MClass.
136   ARMProcClassEnum ARMProcClass = None;
137 
138   /// ARMArch - ARM architecture
139   ARMArchEnum ARMArch = ARMv4t;
140 
141   /// HasV4TOps, HasV5TOps, HasV5TEOps,
142   /// HasV6Ops, HasV6MOps, HasV6KOps, HasV6T2Ops, HasV7Ops, HasV8Ops -
143   /// Specify whether target support specific ARM ISA variants.
144   bool HasV4TOps = false;
145   bool HasV5TOps = false;
146   bool HasV5TEOps = false;
147   bool HasV6Ops = false;
148   bool HasV6MOps = false;
149   bool HasV6KOps = false;
150   bool HasV6T2Ops = false;
151   bool HasV7Ops = false;
152   bool HasV8Ops = false;
153   bool HasV8_1aOps = false;
154   bool HasV8_2aOps = false;
155   bool HasV8_3aOps = false;
156   bool HasV8_4aOps = false;
157   bool HasV8_5aOps = false;
158   bool HasV8MBaselineOps = false;
159   bool HasV8MMainlineOps = false;
160 
161   /// HasVFPv2, HasVFPv3, HasVFPv4, HasFPARMv8, HasNEON - Specify what
162   /// floating point ISAs are supported.
163   bool HasVFPv2 = false;
164   bool HasVFPv3 = false;
165   bool HasVFPv4 = false;
166   bool HasFPARMv8 = false;
167   bool HasNEON = false;
168 
169   /// HasDotProd - True if the ARMv8.2A dot product instructions are supported.
170   bool HasDotProd = false;
171 
172   /// UseNEONForSinglePrecisionFP - if the NEONFP attribute has been
173   /// specified. Use the method useNEONForSinglePrecisionFP() to
174   /// determine if NEON should actually be used.
175   bool UseNEONForSinglePrecisionFP = false;
176 
177   /// UseMulOps - True if non-microcoded fused integer multiply-add and
178   /// multiply-subtract instructions should be used.
179   bool UseMulOps = false;
180 
181   /// SlowFPVMLx - If the VFP2 / NEON instructions are available, indicates
182   /// whether the FP VML[AS] instructions are slow (if so, don't use them).
183   bool SlowFPVMLx = false;
184 
185   /// HasVMLxForwarding - If true, NEON has special multiplier accumulator
186   /// forwarding to allow mul + mla being issued back to back.
187   bool HasVMLxForwarding = false;
188 
189   /// SlowFPBrcc - True if floating point compare + branch is slow.
190   bool SlowFPBrcc = false;
191 
192   /// InThumbMode - True if compiling for Thumb, false for ARM.
193   bool InThumbMode = false;
194 
195   /// UseSoftFloat - True if we're using software floating point features.
196   bool UseSoftFloat = false;
197 
198   /// UseMISched - True if MachineScheduler should be used for this subtarget.
199   bool UseMISched = false;
200 
201   /// DisablePostRAScheduler - False if scheduling should happen again after
202   /// register allocation.
203   bool DisablePostRAScheduler = false;
204 
205   /// UseAA - True if using AA during codegen (DAGCombine, MISched, etc)
206   bool UseAA = false;
207 
208   /// HasThumb2 - True if Thumb2 instructions are supported.
209   bool HasThumb2 = false;
210 
211   /// NoARM - True if subtarget does not support ARM mode execution.
212   bool NoARM = false;
213 
214   /// ReserveR9 - True if R9 is not available as a general purpose register.
215   bool ReserveR9 = false;
216 
217   /// NoMovt - True if MOVT / MOVW pairs are not used for materialization of
218   /// 32-bit imms (including global addresses).
219   bool NoMovt = false;
220 
221   /// SupportsTailCall - True if the OS supports tail call. The dynamic linker
222   /// must be able to synthesize call stubs for interworking between ARM and
223   /// Thumb.
224   bool SupportsTailCall = false;
225 
226   /// HasFP16 - True if subtarget supports half-precision FP conversions
227   bool HasFP16 = false;
228 
229   /// HasFullFP16 - True if subtarget supports half-precision FP operations
230   bool HasFullFP16 = false;
231 
232   /// HasFP16FML - True if subtarget supports half-precision FP fml operations
233   bool HasFP16FML = false;
234 
235   /// HasD16 - True if subtarget is limited to 16 double precision
236   /// FP registers for VFPv3.
237   bool HasD16 = false;
238 
239   /// HasHardwareDivide - True if subtarget supports [su]div in Thumb mode
240   bool HasHardwareDivideInThumb = false;
241 
242   /// HasHardwareDivideInARM - True if subtarget supports [su]div in ARM mode
243   bool HasHardwareDivideInARM = false;
244 
245   /// HasDataBarrier - True if the subtarget supports DMB / DSB data barrier
246   /// instructions.
247   bool HasDataBarrier = false;
248 
249   /// HasFullDataBarrier - True if the subtarget supports DFB data barrier
250   /// instruction.
251   bool HasFullDataBarrier = false;
252 
253   /// HasV7Clrex - True if the subtarget supports CLREX instructions
254   bool HasV7Clrex = false;
255 
256   /// HasAcquireRelease - True if the subtarget supports v8 atomics (LDA/LDAEX etc)
257   /// instructions
258   bool HasAcquireRelease = false;
259 
260   /// Pref32BitThumb - If true, codegen would prefer 32-bit Thumb instructions
261   /// over 16-bit ones.
262   bool Pref32BitThumb = false;
263 
264   /// AvoidCPSRPartialUpdate - If true, codegen would avoid using instructions
265   /// that partially update CPSR and add false dependency on the previous
266   /// CPSR setting instruction.
267   bool AvoidCPSRPartialUpdate = false;
268 
269   /// CheapPredicableCPSRDef - If true, disable +1 predication cost
270   /// for instructions updating CPSR. Enabled for Cortex-A57.
271   bool CheapPredicableCPSRDef = false;
272 
273   /// AvoidMOVsShifterOperand - If true, codegen should avoid using flag setting
274   /// movs with shifter operand (i.e. asr, lsl, lsr).
275   bool AvoidMOVsShifterOperand = false;
276 
277   /// HasRetAddrStack - Some processors perform return stack prediction. CodeGen should
278   /// avoid issue "normal" call instructions to callees which do not return.
279   bool HasRetAddrStack = false;
280 
281   /// HasBranchPredictor - True if the subtarget has a branch predictor. Having
282   /// a branch predictor or not changes the expected cost of taking a branch
283   /// which affects the choice of whether to use predicated instructions.
284   bool HasBranchPredictor = true;
285 
286   /// HasMPExtension - True if the subtarget supports Multiprocessing
287   /// extension (ARMv7 only).
288   bool HasMPExtension = false;
289 
290   /// HasVirtualization - True if the subtarget supports the Virtualization
291   /// extension.
292   bool HasVirtualization = false;
293 
294   /// FPOnlySP - If true, the floating point unit only supports single
295   /// precision.
296   bool FPOnlySP = false;
297 
298   /// If true, the processor supports the Performance Monitor Extensions. These
299   /// include a generic cycle-counter as well as more fine-grained (often
300   /// implementation-specific) events.
301   bool HasPerfMon = false;
302 
303   /// HasTrustZone - if true, processor supports TrustZone security extensions
304   bool HasTrustZone = false;
305 
306   /// Has8MSecExt - if true, processor supports ARMv8-M Security Extensions
307   bool Has8MSecExt = false;
308 
309   /// HasSHA2 - if true, processor supports SHA1 and SHA256
310   bool HasSHA2 = false;
311 
312   /// HasAES - if true, processor supports AES
313   bool HasAES = false;
314 
315   /// HasCrypto - if true, processor supports Cryptography extensions
316   bool HasCrypto = false;
317 
318   /// HasCRC - if true, processor supports CRC instructions
319   bool HasCRC = false;
320 
321   /// HasRAS - if true, the processor supports RAS extensions
322   bool HasRAS = false;
323 
324   /// If true, the instructions "vmov.i32 d0, #0" and "vmov.i32 q0, #0" are
325   /// particularly effective at zeroing a VFP register.
326   bool HasZeroCycleZeroing = false;
327 
328   /// HasFPAO - if true, processor  does positive address offset computation faster
329   bool HasFPAO = false;
330 
331   /// HasFuseAES - if true, processor executes back to back AES instruction
332   /// pairs faster.
333   bool HasFuseAES = false;
334 
335   /// HasFuseLiterals - if true, processor executes back to back
336   /// bottom and top halves of literal generation faster.
337   bool HasFuseLiterals = false;
338 
339   /// If true, if conversion may decide to leave some instructions unpredicated.
340   bool IsProfitableToUnpredicate = false;
341 
342   /// If true, VMOV will be favored over VGETLNi32.
343   bool HasSlowVGETLNi32 = false;
344 
345   /// If true, VMOV will be favored over VDUP.
346   bool HasSlowVDUP32 = false;
347 
348   /// If true, VMOVSR will be favored over VMOVDRR.
349   bool PreferVMOVSR = false;
350 
351   /// If true, ISHST barriers will be used for Release semantics.
352   bool PreferISHST = false;
353 
354   /// If true, a VLDM/VSTM starting with an odd register number is considered to
355   /// take more microops than single VLDRS/VSTRS.
356   bool SlowOddRegister = false;
357 
358   /// If true, loading into a D subregister will be penalized.
359   bool SlowLoadDSubregister = false;
360 
361   /// If true, use a wider stride when allocating VFP registers.
362   bool UseWideStrideVFP = false;
363 
364   /// If true, the AGU and NEON/FPU units are multiplexed.
365   bool HasMuxedUnits = false;
366 
367   /// If true, VMOVS will never be widened to VMOVD.
368   bool DontWidenVMOVS = false;
369 
370   /// If true, splat a register between VFP and NEON instructions.
371   bool SplatVFPToNeon = false;
372 
373   /// If true, run the MLx expansion pass.
374   bool ExpandMLx = false;
375 
376   /// If true, VFP/NEON VMLA/VMLS have special RAW hazards.
377   bool HasVMLxHazards = false;
378 
379   // If true, read thread pointer from coprocessor register.
380   bool ReadTPHard = false;
381 
382   /// If true, VMOVRS, VMOVSR and VMOVS will be converted from VFP to NEON.
383   bool UseNEONForFPMovs = false;
384 
385   /// If true, VLDn instructions take an extra cycle for unaligned accesses.
386   bool CheckVLDnAlign = false;
387 
388   /// If true, VFP instructions are not pipelined.
389   bool NonpipelinedVFP = false;
390 
391   /// StrictAlign - If true, the subtarget disallows unaligned memory
392   /// accesses for some types.  For details, see
393   /// ARMTargetLowering::allowsMisalignedMemoryAccesses().
394   bool StrictAlign = false;
395 
396   /// RestrictIT - If true, the subtarget disallows generation of deprecated IT
397   ///  blocks to conform to ARMv8 rule.
398   bool RestrictIT = false;
399 
400   /// HasDSP - If true, the subtarget supports the DSP (saturating arith
401   /// and such) instructions.
402   bool HasDSP = false;
403 
404   /// NaCl TRAP instruction is generated instead of the regular TRAP.
405   bool UseNaClTrap = false;
406 
407   /// Generate calls via indirect call instructions.
408   bool GenLongCalls = false;
409 
410   /// Generate code that does not contain data access to code sections.
411   bool GenExecuteOnly = false;
412 
413   /// Target machine allowed unsafe FP math (such as use of NEON fp)
414   bool UnsafeFPMath = false;
415 
416   /// UseSjLjEH - If true, the target uses SjLj exception handling (e.g. iOS).
417   bool UseSjLjEH = false;
418 
419   /// Has speculation barrier
420   bool HasSB = false;
421 
422   /// Implicitly convert an instruction to a different one if its immediates
423   /// cannot be encoded. For example, ADD r0, r1, #FFFFFFFF -> SUB r0, r1, #1.
424   bool NegativeImmediates = true;
425 
426   /// stackAlignment - The minimum alignment known to hold of the stack frame on
427   /// entry to the function and which must be maintained by every function.
428   unsigned stackAlignment = 4;
429 
430   /// CPUString - String name of used CPU.
431   std::string CPUString;
432 
433   unsigned MaxInterleaveFactor = 1;
434 
435   /// Clearance before partial register updates (in number of instructions)
436   unsigned PartialUpdateClearance = 0;
437 
438   /// What kind of timing do load multiple/store multiple have (double issue,
439   /// single issue etc).
440   ARMLdStMultipleTiming LdStMultipleTiming = SingleIssue;
441 
442   /// The adjustment that we need to apply to get the operand latency from the
443   /// operand cycle returned by the itinerary data for pre-ISel operands.
444   int PreISelOperandLatencyAdjustment = 2;
445 
446   /// What alignment is preferred for loop bodies, in log2(bytes).
447   unsigned PrefLoopAlignment = 0;
448 
449   /// IsLittle - The target is Little Endian
450   bool IsLittle;
451 
452   /// TargetTriple - What processor and OS we're targeting.
453   Triple TargetTriple;
454 
455   /// SchedModel - Processor specific instruction costs.
456   MCSchedModel SchedModel;
457 
458   /// Selected instruction itineraries (one entry per itinerary class.)
459   InstrItineraryData InstrItins;
460 
461   /// Options passed via command line that could influence the target
462   const TargetOptions &Options;
463 
464   const ARMBaseTargetMachine &TM;
465 
466 public:
467   /// This constructor initializes the data members to match that
468   /// of the specified triple.
469   ///
470   ARMSubtarget(const Triple &TT, const std::string &CPU, const std::string &FS,
471                const ARMBaseTargetMachine &TM, bool IsLittle);
472 
473   /// getMaxInlineSizeThreshold - Returns the maximum memset / memcpy size
474   /// that still makes it profitable to inline the call.
getMaxInlineSizeThreshold()475   unsigned getMaxInlineSizeThreshold() const {
476     return 64;
477   }
478 
479   /// ParseSubtargetFeatures - Parses features string setting specified
480   /// subtarget options.  Definition of function is auto generated by tblgen.
481   void ParseSubtargetFeatures(StringRef CPU, StringRef FS);
482 
483   /// initializeSubtargetDependencies - Initializes using a CPU and feature string
484   /// so that we can use initializer lists for subtarget initialization.
485   ARMSubtarget &initializeSubtargetDependencies(StringRef CPU, StringRef FS);
486 
getSelectionDAGInfo()487   const ARMSelectionDAGInfo *getSelectionDAGInfo() const override {
488     return &TSInfo;
489   }
490 
getInstrInfo()491   const ARMBaseInstrInfo *getInstrInfo() const override {
492     return InstrInfo.get();
493   }
494 
getTargetLowering()495   const ARMTargetLowering *getTargetLowering() const override {
496     return &TLInfo;
497   }
498 
getFrameLowering()499   const ARMFrameLowering *getFrameLowering() const override {
500     return FrameLowering.get();
501   }
502 
getRegisterInfo()503   const ARMBaseRegisterInfo *getRegisterInfo() const override {
504     return &InstrInfo->getRegisterInfo();
505   }
506 
507   const CallLowering *getCallLowering() const override;
508   const InstructionSelector *getInstructionSelector() const override;
509   const LegalizerInfo *getLegalizerInfo() const override;
510   const RegisterBankInfo *getRegBankInfo() const override;
511 
512 private:
513   ARMSelectionDAGInfo TSInfo;
514   // Either Thumb1FrameLowering or ARMFrameLowering.
515   std::unique_ptr<ARMFrameLowering> FrameLowering;
516   // Either Thumb1InstrInfo or Thumb2InstrInfo.
517   std::unique_ptr<ARMBaseInstrInfo> InstrInfo;
518   ARMTargetLowering   TLInfo;
519 
520   /// GlobalISel related APIs.
521   std::unique_ptr<CallLowering> CallLoweringInfo;
522   std::unique_ptr<InstructionSelector> InstSelector;
523   std::unique_ptr<LegalizerInfo> Legalizer;
524   std::unique_ptr<RegisterBankInfo> RegBankInfo;
525 
526   void initializeEnvironment();
527   void initSubtargetFeatures(StringRef CPU, StringRef FS);
528   ARMFrameLowering *initializeFrameLowering(StringRef CPU, StringRef FS);
529 
530 public:
531   void computeIssueWidth();
532 
hasV4TOps()533   bool hasV4TOps()  const { return HasV4TOps;  }
hasV5TOps()534   bool hasV5TOps()  const { return HasV5TOps;  }
hasV5TEOps()535   bool hasV5TEOps() const { return HasV5TEOps; }
hasV6Ops()536   bool hasV6Ops()   const { return HasV6Ops;   }
hasV6MOps()537   bool hasV6MOps()  const { return HasV6MOps;  }
hasV6KOps()538   bool hasV6KOps()  const { return HasV6KOps; }
hasV6T2Ops()539   bool hasV6T2Ops() const { return HasV6T2Ops; }
hasV7Ops()540   bool hasV7Ops()   const { return HasV7Ops;  }
hasV8Ops()541   bool hasV8Ops()   const { return HasV8Ops;  }
hasV8_1aOps()542   bool hasV8_1aOps() const { return HasV8_1aOps; }
hasV8_2aOps()543   bool hasV8_2aOps() const { return HasV8_2aOps; }
hasV8_3aOps()544   bool hasV8_3aOps() const { return HasV8_3aOps; }
hasV8_4aOps()545   bool hasV8_4aOps() const { return HasV8_4aOps; }
hasV8_5aOps()546   bool hasV8_5aOps() const { return HasV8_5aOps; }
hasV8MBaselineOps()547   bool hasV8MBaselineOps() const { return HasV8MBaselineOps; }
hasV8MMainlineOps()548   bool hasV8MMainlineOps() const { return HasV8MMainlineOps; }
549 
550   /// @{
551   /// These functions are obsolete, please consider adding subtarget features
552   /// or properties instead of calling them.
isCortexA5()553   bool isCortexA5() const { return ARMProcFamily == CortexA5; }
isCortexA7()554   bool isCortexA7() const { return ARMProcFamily == CortexA7; }
isCortexA8()555   bool isCortexA8() const { return ARMProcFamily == CortexA8; }
isCortexA9()556   bool isCortexA9() const { return ARMProcFamily == CortexA9; }
isCortexA15()557   bool isCortexA15() const { return ARMProcFamily == CortexA15; }
isSwift()558   bool isSwift()    const { return ARMProcFamily == Swift; }
isCortexM3()559   bool isCortexM3() const { return ARMProcFamily == CortexM3; }
isLikeA9()560   bool isLikeA9() const { return isCortexA9() || isCortexA15() || isKrait(); }
isCortexR5()561   bool isCortexR5() const { return ARMProcFamily == CortexR5; }
isKrait()562   bool isKrait() const { return ARMProcFamily == Krait; }
563   /// @}
564 
hasARMOps()565   bool hasARMOps() const { return !NoARM; }
566 
hasVFP2()567   bool hasVFP2() const { return HasVFPv2; }
hasVFP3()568   bool hasVFP3() const { return HasVFPv3; }
hasVFP4()569   bool hasVFP4() const { return HasVFPv4; }
hasFPARMv8()570   bool hasFPARMv8() const { return HasFPARMv8; }
hasNEON()571   bool hasNEON() const { return HasNEON;  }
hasSHA2()572   bool hasSHA2() const { return HasSHA2; }
hasAES()573   bool hasAES() const { return HasAES; }
hasCrypto()574   bool hasCrypto() const { return HasCrypto; }
hasDotProd()575   bool hasDotProd() const { return HasDotProd; }
hasCRC()576   bool hasCRC() const { return HasCRC; }
hasRAS()577   bool hasRAS() const { return HasRAS; }
hasVirtualization()578   bool hasVirtualization() const { return HasVirtualization; }
579 
useNEONForSinglePrecisionFP()580   bool useNEONForSinglePrecisionFP() const {
581     return hasNEON() && UseNEONForSinglePrecisionFP;
582   }
583 
hasDivideInThumbMode()584   bool hasDivideInThumbMode() const { return HasHardwareDivideInThumb; }
hasDivideInARMMode()585   bool hasDivideInARMMode() const { return HasHardwareDivideInARM; }
hasDataBarrier()586   bool hasDataBarrier() const { return HasDataBarrier; }
hasFullDataBarrier()587   bool hasFullDataBarrier() const { return HasFullDataBarrier; }
hasV7Clrex()588   bool hasV7Clrex() const { return HasV7Clrex; }
hasAcquireRelease()589   bool hasAcquireRelease() const { return HasAcquireRelease; }
590 
hasAnyDataBarrier()591   bool hasAnyDataBarrier() const {
592     return HasDataBarrier || (hasV6Ops() && !isThumb());
593   }
594 
useMulOps()595   bool useMulOps() const { return UseMulOps; }
useFPVMLx()596   bool useFPVMLx() const { return !SlowFPVMLx; }
hasVMLxForwarding()597   bool hasVMLxForwarding() const { return HasVMLxForwarding; }
isFPBrccSlow()598   bool isFPBrccSlow() const { return SlowFPBrcc; }
isFPOnlySP()599   bool isFPOnlySP() const { return FPOnlySP; }
hasPerfMon()600   bool hasPerfMon() const { return HasPerfMon; }
hasTrustZone()601   bool hasTrustZone() const { return HasTrustZone; }
has8MSecExt()602   bool has8MSecExt() const { return Has8MSecExt; }
hasZeroCycleZeroing()603   bool hasZeroCycleZeroing() const { return HasZeroCycleZeroing; }
hasFPAO()604   bool hasFPAO() const { return HasFPAO; }
isProfitableToUnpredicate()605   bool isProfitableToUnpredicate() const { return IsProfitableToUnpredicate; }
hasSlowVGETLNi32()606   bool hasSlowVGETLNi32() const { return HasSlowVGETLNi32; }
hasSlowVDUP32()607   bool hasSlowVDUP32() const { return HasSlowVDUP32; }
preferVMOVSR()608   bool preferVMOVSR() const { return PreferVMOVSR; }
preferISHSTBarriers()609   bool preferISHSTBarriers() const { return PreferISHST; }
expandMLx()610   bool expandMLx() const { return ExpandMLx; }
hasVMLxHazards()611   bool hasVMLxHazards() const { return HasVMLxHazards; }
hasSlowOddRegister()612   bool hasSlowOddRegister() const { return SlowOddRegister; }
hasSlowLoadDSubregister()613   bool hasSlowLoadDSubregister() const { return SlowLoadDSubregister; }
useWideStrideVFP()614   bool useWideStrideVFP() const { return UseWideStrideVFP; }
hasMuxedUnits()615   bool hasMuxedUnits() const { return HasMuxedUnits; }
dontWidenVMOVS()616   bool dontWidenVMOVS() const { return DontWidenVMOVS; }
useSplatVFPToNeon()617   bool useSplatVFPToNeon() const { return SplatVFPToNeon; }
useNEONForFPMovs()618   bool useNEONForFPMovs() const { return UseNEONForFPMovs; }
checkVLDnAccessAlignment()619   bool checkVLDnAccessAlignment() const { return CheckVLDnAlign; }
nonpipelinedVFP()620   bool nonpipelinedVFP() const { return NonpipelinedVFP; }
prefers32BitThumb()621   bool prefers32BitThumb() const { return Pref32BitThumb; }
avoidCPSRPartialUpdate()622   bool avoidCPSRPartialUpdate() const { return AvoidCPSRPartialUpdate; }
cheapPredicableCPSRDef()623   bool cheapPredicableCPSRDef() const { return CheapPredicableCPSRDef; }
avoidMOVsShifterOperand()624   bool avoidMOVsShifterOperand() const { return AvoidMOVsShifterOperand; }
hasRetAddrStack()625   bool hasRetAddrStack() const { return HasRetAddrStack; }
hasBranchPredictor()626   bool hasBranchPredictor() const { return HasBranchPredictor; }
hasMPExtension()627   bool hasMPExtension() const { return HasMPExtension; }
hasDSP()628   bool hasDSP() const { return HasDSP; }
useNaClTrap()629   bool useNaClTrap() const { return UseNaClTrap; }
useSjLjEH()630   bool useSjLjEH() const { return UseSjLjEH; }
hasSB()631   bool hasSB() const { return HasSB; }
genLongCalls()632   bool genLongCalls() const { return GenLongCalls; }
genExecuteOnly()633   bool genExecuteOnly() const { return GenExecuteOnly; }
634 
hasFP16()635   bool hasFP16() const { return HasFP16; }
hasD16()636   bool hasD16() const { return HasD16; }
hasFullFP16()637   bool hasFullFP16() const { return HasFullFP16; }
hasFP16FML()638   bool hasFP16FML() const { return HasFP16FML; }
639 
hasFuseAES()640   bool hasFuseAES() const { return HasFuseAES; }
hasFuseLiterals()641   bool hasFuseLiterals() const { return HasFuseLiterals; }
642   /// Return true if the CPU supports any kind of instruction fusion.
hasFusion()643   bool hasFusion() const { return hasFuseAES() || hasFuseLiterals(); }
644 
getTargetTriple()645   const Triple &getTargetTriple() const { return TargetTriple; }
646 
isTargetDarwin()647   bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); }
isTargetIOS()648   bool isTargetIOS() const { return TargetTriple.isiOS(); }
isTargetWatchOS()649   bool isTargetWatchOS() const { return TargetTriple.isWatchOS(); }
isTargetWatchABI()650   bool isTargetWatchABI() const { return TargetTriple.isWatchABI(); }
isTargetLinux()651   bool isTargetLinux() const { return TargetTriple.isOSLinux(); }
isTargetNaCl()652   bool isTargetNaCl() const { return TargetTriple.isOSNaCl(); }
isTargetNetBSD()653   bool isTargetNetBSD() const { return TargetTriple.isOSNetBSD(); }
isTargetWindows()654   bool isTargetWindows() const { return TargetTriple.isOSWindows(); }
655 
isTargetCOFF()656   bool isTargetCOFF() const { return TargetTriple.isOSBinFormatCOFF(); }
isTargetELF()657   bool isTargetELF() const { return TargetTriple.isOSBinFormatELF(); }
isTargetMachO()658   bool isTargetMachO() const { return TargetTriple.isOSBinFormatMachO(); }
659 
660   // ARM EABI is the bare-metal EABI described in ARM ABI documents and
661   // can be accessed via -target arm-none-eabi. This is NOT GNUEABI.
662   // FIXME: Add a flag for bare-metal for that target and set Triple::EABI
663   // even for GNUEABI, so we can make a distinction here and still conform to
664   // the EABI on GNU (and Android) mode. This requires change in Clang, too.
665   // FIXME: The Darwin exception is temporary, while we move users to
666   // "*-*-*-macho" triples as quickly as possible.
isTargetAEABI()667   bool isTargetAEABI() const {
668     return (TargetTriple.getEnvironment() == Triple::EABI ||
669             TargetTriple.getEnvironment() == Triple::EABIHF) &&
670            !isTargetDarwin() && !isTargetWindows();
671   }
isTargetGNUAEABI()672   bool isTargetGNUAEABI() const {
673     return (TargetTriple.getEnvironment() == Triple::GNUEABI ||
674             TargetTriple.getEnvironment() == Triple::GNUEABIHF) &&
675            !isTargetDarwin() && !isTargetWindows();
676   }
isTargetMuslAEABI()677   bool isTargetMuslAEABI() const {
678     return (TargetTriple.getEnvironment() == Triple::MuslEABI ||
679             TargetTriple.getEnvironment() == Triple::MuslEABIHF) &&
680            !isTargetDarwin() && !isTargetWindows();
681   }
682 
683   // ARM Targets that support EHABI exception handling standard
684   // Darwin uses SjLj. Other targets might need more checks.
isTargetEHABICompatible()685   bool isTargetEHABICompatible() const {
686     return (TargetTriple.getEnvironment() == Triple::EABI ||
687             TargetTriple.getEnvironment() == Triple::GNUEABI ||
688             TargetTriple.getEnvironment() == Triple::MuslEABI ||
689             TargetTriple.getEnvironment() == Triple::EABIHF ||
690             TargetTriple.getEnvironment() == Triple::GNUEABIHF ||
691             TargetTriple.getEnvironment() == Triple::MuslEABIHF ||
692             isTargetAndroid()) &&
693            !isTargetDarwin() && !isTargetWindows();
694   }
695 
696   bool isTargetHardFloat() const;
697 
isTargetAndroid()698   bool isTargetAndroid() const { return TargetTriple.isAndroid(); }
699 
700   bool isXRaySupported() const override;
701 
702   bool isAPCS_ABI() const;
703   bool isAAPCS_ABI() const;
704   bool isAAPCS16_ABI() const;
705 
706   bool isROPI() const;
707   bool isRWPI() const;
708 
useMachineScheduler()709   bool useMachineScheduler() const { return UseMISched; }
disablePostRAScheduler()710   bool disablePostRAScheduler() const { return DisablePostRAScheduler; }
useSoftFloat()711   bool useSoftFloat() const { return UseSoftFloat; }
isThumb()712   bool isThumb() const { return InThumbMode; }
isThumb1Only()713   bool isThumb1Only() const { return InThumbMode && !HasThumb2; }
isThumb2()714   bool isThumb2() const { return InThumbMode && HasThumb2; }
hasThumb2()715   bool hasThumb2() const { return HasThumb2; }
isMClass()716   bool isMClass() const { return ARMProcClass == MClass; }
isRClass()717   bool isRClass() const { return ARMProcClass == RClass; }
isAClass()718   bool isAClass() const { return ARMProcClass == AClass; }
isReadTPHard()719   bool isReadTPHard() const { return ReadTPHard; }
720 
isR9Reserved()721   bool isR9Reserved() const {
722     return isTargetMachO() ? (ReserveR9 || !HasV6Ops) : ReserveR9;
723   }
724 
useR7AsFramePointer()725   bool useR7AsFramePointer() const {
726     return isTargetDarwin() || (!isTargetWindows() && isThumb());
727   }
728 
729   /// Returns true if the frame setup is split into two separate pushes (first
730   /// r0-r7,lr then r8-r11), principally so that the frame pointer is adjacent
731   /// to lr. This is always required on Thumb1-only targets, as the push and
732   /// pop instructions can't access the high registers.
splitFramePushPop(const MachineFunction & MF)733   bool splitFramePushPop(const MachineFunction &MF) const {
734     return (useR7AsFramePointer() &&
735             MF.getTarget().Options.DisableFramePointerElim(MF)) ||
736            isThumb1Only();
737   }
738 
739   bool useStride4VFPs(const MachineFunction &MF) const;
740 
741   bool useMovt(const MachineFunction &MF) const;
742 
supportsTailCall()743   bool supportsTailCall() const { return SupportsTailCall; }
744 
allowsUnalignedMem()745   bool allowsUnalignedMem() const { return !StrictAlign; }
746 
restrictIT()747   bool restrictIT() const { return RestrictIT; }
748 
getCPUString()749   const std::string & getCPUString() const { return CPUString; }
750 
isLittle()751   bool isLittle() const { return IsLittle; }
752 
753   unsigned getMispredictionPenalty() const;
754 
755   /// Returns true if machine scheduler should be enabled.
756   bool enableMachineScheduler() const override;
757 
758   /// True for some subtargets at > -O0.
759   bool enablePostRAScheduler() const override;
760 
761   /// Enable use of alias analysis during code generation (during MI
762   /// scheduling, DAGCombine, etc.).
useAA()763   bool useAA() const override { return UseAA; }
764 
765   // enableAtomicExpand- True if we need to expand our atomics.
766   bool enableAtomicExpand() const override;
767 
768   /// getInstrItins - Return the instruction itineraries based on subtarget
769   /// selection.
getInstrItineraryData()770   const InstrItineraryData *getInstrItineraryData() const override {
771     return &InstrItins;
772   }
773 
774   /// getStackAlignment - Returns the minimum alignment known to hold of the
775   /// stack frame on entry to the function and which must be maintained by every
776   /// function for this subtarget.
getStackAlignment()777   unsigned getStackAlignment() const { return stackAlignment; }
778 
getMaxInterleaveFactor()779   unsigned getMaxInterleaveFactor() const { return MaxInterleaveFactor; }
780 
getPartialUpdateClearance()781   unsigned getPartialUpdateClearance() const { return PartialUpdateClearance; }
782 
getLdStMultipleTiming()783   ARMLdStMultipleTiming getLdStMultipleTiming() const {
784     return LdStMultipleTiming;
785   }
786 
getPreISelOperandLatencyAdjustment()787   int getPreISelOperandLatencyAdjustment() const {
788     return PreISelOperandLatencyAdjustment;
789   }
790 
791   /// True if the GV will be accessed via an indirect symbol.
792   bool isGVIndirectSymbol(const GlobalValue *GV) const;
793 
794   /// Returns the constant pool modifier needed to access the GV.
795   bool isGVInGOT(const GlobalValue *GV) const;
796 
797   /// True if fast-isel is used.
798   bool useFastISel() const;
799 
800   /// Returns the correct return opcode for the current feature set.
801   /// Use BX if available to allow mixing thumb/arm code, but fall back
802   /// to plain mov pc,lr on ARMv4.
getReturnOpcode()803   unsigned getReturnOpcode() const {
804     if (isThumb())
805       return ARM::tBX_RET;
806     if (hasV4TOps())
807       return ARM::BX_RET;
808     return ARM::MOVPCLR;
809   }
810 
811   /// Allow movt+movw for PIC global address calculation.
812   /// ELF does not have GOT relocations for movt+movw.
813   /// ROPI does not use GOT.
allowPositionIndependentMovt()814   bool allowPositionIndependentMovt() const {
815     return isROPI() || !isTargetELF();
816   }
817 
getPrefLoopAlignment()818   unsigned getPrefLoopAlignment() const {
819     return PrefLoopAlignment;
820   }
821 };
822 
823 } // end namespace llvm
824 
825 #endif  // LLVM_LIB_TARGET_ARM_ARMSUBTARGET_H
826