xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64Subtarget.h (revision 8bcb0991864975618c09697b1aca10683346d9f0) 
1 //===--- AArch64Subtarget.h - Define Subtarget for the AArch64 -*- C++ -*--===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file declares the AArch64 specific subclass of TargetSubtarget.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef LLVM_LIB_TARGET_AARCH64_AARCH64SUBTARGET_H
14 #define LLVM_LIB_TARGET_AARCH64_AARCH64SUBTARGET_H
15 
16 #include "AArch64FrameLowering.h"
17 #include "AArch64ISelLowering.h"
18 #include "AArch64InstrInfo.h"
19 #include "AArch64RegisterInfo.h"
20 #include "AArch64SelectionDAGInfo.h"
21 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
22 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
23 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
24 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h"
25 #include "llvm/CodeGen/TargetSubtargetInfo.h"
26 #include "llvm/IR/DataLayout.h"
27 #include <string>
28 
29 #define GET_SUBTARGETINFO_HEADER
30 #include "AArch64GenSubtargetInfo.inc"
31 
32 namespace llvm {
33 class GlobalValue;
34 class StringRef;
35 class Triple;
36 
37 class AArch64Subtarget final : public AArch64GenSubtargetInfo {
38 public:
39   enum ARMProcFamilyEnum : uint8_t {
40     Others,
41     CortexA35,
42     CortexA53,
43     CortexA55,
44     CortexA57,
45     CortexA65,
46     CortexA72,
47     CortexA73,
48     CortexA75,
49     CortexA76,
50     Cyclone,
51     ExynosM1,
52     ExynosM3,
53     Falkor,
54     Kryo,
55     NeoverseE1,
56     NeoverseN1,
57     Saphira,
58     ThunderX2T99,
59     ThunderX,
60     ThunderXT81,
61     ThunderXT83,
62     ThunderXT88,
63     TSV110
64   };
65 
66 protected:
67   /// ARMProcFamily - ARM processor family: Cortex-A53, Cortex-A57, and others.
68   ARMProcFamilyEnum ARMProcFamily = Others;
69 
70   bool HasV8_1aOps = false;
71   bool HasV8_2aOps = false;
72   bool HasV8_3aOps = false;
73   bool HasV8_4aOps = false;
74   bool HasV8_5aOps = false;
75 
76   bool HasFPARMv8 = false;
77   bool HasNEON = false;
78   bool HasCrypto = false;
79   bool HasDotProd = false;
80   bool HasCRC = false;
81   bool HasLSE = false;
82   bool HasRAS = false;
83   bool HasRDM = false;
84   bool HasPerfMon = false;
85   bool HasFullFP16 = false;
86   bool HasFP16FML = false;
87   bool HasSPE = false;
88 
89   // ARMv8.1 extensions
90   bool HasVH = false;
91   bool HasPAN = false;
92   bool HasLOR = false;
93 
94   // ARMv8.2 extensions
95   bool HasPsUAO = false;
96   bool HasPAN_RWV = false;
97   bool HasCCPP = false;
98 
99   // Armv8.2 Crypto extensions
100   bool HasSM4 = false;
101   bool HasSHA3 = false;
102   bool HasSHA2 = false;
103   bool HasAES = false;
104 
105   // ARMv8.3 extensions
106   bool HasPA = false;
107   bool HasJS = false;
108   bool HasCCIDX = false;
109   bool HasComplxNum = false;
110 
111   // ARMv8.4 extensions
112   bool HasNV = false;
113   bool HasRASv8_4 = false;
114   bool HasMPAM = false;
115   bool HasDIT = false;
116   bool HasTRACEV8_4 = false;
117   bool HasAM = false;
118   bool HasSEL2 = false;
119   bool HasPMU = false;
120   bool HasTLB_RMI = false;
121   bool HasFMI = false;
122   bool HasRCPC_IMMO = false;
123 
124   bool HasLSLFast = false;
125   bool HasSVE = false;
126   bool HasSVE2 = false;
127   bool HasRCPC = false;
128   bool HasAggressiveFMA = false;
129 
130   // Armv8.5-A Extensions
131   bool HasAlternativeNZCV = false;
132   bool HasFRInt3264 = false;
133   bool HasSpecRestrict = false;
134   bool HasSSBS = false;
135   bool HasSB = false;
136   bool HasPredRes = false;
137   bool HasCCDP = false;
138   bool HasBTI = false;
139   bool HasRandGen = false;
140   bool HasMTE = false;
141   bool HasTME = false;
142 
143   // Arm SVE2 extensions
144   bool HasSVE2AES = false;
145   bool HasSVE2SM4 = false;
146   bool HasSVE2SHA3 = false;
147   bool HasSVE2BitPerm = false;
148 
149   // Future architecture extensions.
150   bool HasETE = false;
151   bool HasTRBE = false;
152 
153   // HasZeroCycleRegMove - Has zero-cycle register mov instructions.
154   bool HasZeroCycleRegMove = false;
155 
156   // HasZeroCycleZeroing - Has zero-cycle zeroing instructions.
157   bool HasZeroCycleZeroing = false;
158   bool HasZeroCycleZeroingGP = false;
159   bool HasZeroCycleZeroingFP = false;
160   bool HasZeroCycleZeroingFPWorkaround = false;
161 
162   // StrictAlign - Disallow unaligned memory accesses.
163   bool StrictAlign = false;
164 
165   // NegativeImmediates - transform instructions with negative immediates
166   bool NegativeImmediates = true;
167 
168   // Enable 64-bit vectorization in SLP.
169   unsigned MinVectorRegisterBitWidth = 64;
170 
171   bool UseAA = false;
172   bool PredictableSelectIsExpensive = false;
173   bool BalanceFPOps = false;
174   bool CustomAsCheapAsMove = false;
175   bool ExynosAsCheapAsMove = false;
176   bool UsePostRAScheduler = false;
177   bool Misaligned128StoreIsSlow = false;
178   bool Paired128IsSlow = false;
179   bool STRQroIsSlow = false;
180   bool UseAlternateSExtLoadCVTF32Pattern = false;
181   bool HasArithmeticBccFusion = false;
182   bool HasArithmeticCbzFusion = false;
183   bool HasFuseAddress = false;
184   bool HasFuseAES = false;
185   bool HasFuseArithmeticLogic = false;
186   bool HasFuseCCSelect = false;
187   bool HasFuseCryptoEOR = false;
188   bool HasFuseLiterals = false;
189   bool DisableLatencySchedHeuristic = false;
190   bool UseRSqrt = false;
191   bool Force32BitJumpTables = false;
192   bool UseEL1ForTP = false;
193   bool UseEL2ForTP = false;
194   bool UseEL3ForTP = false;
195   bool AllowTaggedGlobals = false;
196   uint8_t MaxInterleaveFactor = 2;
197   uint8_t VectorInsertExtractBaseCost = 3;
198   uint16_t CacheLineSize = 0;
199   uint16_t PrefetchDistance = 0;
200   uint16_t MinPrefetchStride = 1;
201   unsigned MaxPrefetchIterationsAhead = UINT_MAX;
202   unsigned PrefFunctionLogAlignment = 0;
203   unsigned PrefLoopLogAlignment = 0;
204   unsigned MaxJumpTableSize = 0;
205   unsigned WideningBaseCost = 0;
206 
207   // ReserveXRegister[i] - X#i is not available as a general purpose register.
208   BitVector ReserveXRegister;
209 
210   // CustomCallUsedXRegister[i] - X#i call saved.
211   BitVector CustomCallSavedXRegs;
212 
213   bool IsLittle;
214 
215   /// TargetTriple - What processor and OS we're targeting.
216   Triple TargetTriple;
217 
218   AArch64FrameLowering FrameLowering;
219   AArch64InstrInfo InstrInfo;
220   AArch64SelectionDAGInfo TSInfo;
221   AArch64TargetLowering TLInfo;
222 
223   /// GlobalISel related APIs.
224   std::unique_ptr<CallLowering> CallLoweringInfo;
225   std::unique_ptr<InstructionSelector> InstSelector;
226   std::unique_ptr<LegalizerInfo> Legalizer;
227   std::unique_ptr<RegisterBankInfo> RegBankInfo;
228 
229 private:
230   /// initializeSubtargetDependencies - Initializes using CPUString and the
231   /// passed in feature string so that we can use initializer lists for
232   /// subtarget initialization.
233   AArch64Subtarget &initializeSubtargetDependencies(StringRef FS,
234                                                     StringRef CPUString);
235 
236   /// Initialize properties based on the selected processor family.
237   void initializeProperties();
238 
239 public:
240   /// This constructor initializes the data members to match that
241   /// of the specified triple.
242   AArch64Subtarget(const Triple &TT, const std::string &CPU,
243                    const std::string &FS, const TargetMachine &TM,
244                    bool LittleEndian);
245 
246   const AArch64SelectionDAGInfo *getSelectionDAGInfo() const override {
247     return &TSInfo;
248   }
249   const AArch64FrameLowering *getFrameLowering() const override {
250     return &FrameLowering;
251   }
252   const AArch64TargetLowering *getTargetLowering() const override {
253     return &TLInfo;
254   }
255   const AArch64InstrInfo *getInstrInfo() const override { return &InstrInfo; }
256   const AArch64RegisterInfo *getRegisterInfo() const override {
257     return &getInstrInfo()->getRegisterInfo();
258   }
259   const CallLowering *getCallLowering() const override;
260   InstructionSelector *getInstructionSelector() const override;
261   const LegalizerInfo *getLegalizerInfo() const override;
262   const RegisterBankInfo *getRegBankInfo() const override;
263   const Triple &getTargetTriple() const { return TargetTriple; }
264   bool enableMachineScheduler() const override { return true; }
265   bool enablePostRAScheduler() const override {
266     return UsePostRAScheduler;
267   }
268 
269   /// Returns ARM processor family.
270   /// Avoid this function! CPU specifics should be kept local to this class
271   /// and preferably modeled with SubtargetFeatures or properties in
272   /// initializeProperties().
273   ARMProcFamilyEnum getProcFamily() const {
274     return ARMProcFamily;
275   }
276 
277   bool hasV8_1aOps() const { return HasV8_1aOps; }
278   bool hasV8_2aOps() const { return HasV8_2aOps; }
279   bool hasV8_3aOps() const { return HasV8_3aOps; }
280   bool hasV8_4aOps() const { return HasV8_4aOps; }
281   bool hasV8_5aOps() const { return HasV8_5aOps; }
282 
283   bool hasZeroCycleRegMove() const { return HasZeroCycleRegMove; }
284 
285   bool hasZeroCycleZeroingGP() const { return HasZeroCycleZeroingGP; }
286 
287   bool hasZeroCycleZeroingFP() const { return HasZeroCycleZeroingFP; }
288 
289   bool hasZeroCycleZeroingFPWorkaround() const {
290     return HasZeroCycleZeroingFPWorkaround;
291   }
292 
293   bool requiresStrictAlign() const { return StrictAlign; }
294 
295   bool isXRaySupported() const override { return true; }
296 
297   unsigned getMinVectorRegisterBitWidth() const {
298     return MinVectorRegisterBitWidth;
299   }
300 
301   bool isXRegisterReserved(size_t i) const { return ReserveXRegister[i]; }
302   unsigned getNumXRegisterReserved() const { return ReserveXRegister.count(); }
303   bool isXRegCustomCalleeSaved(size_t i) const {
304     return CustomCallSavedXRegs[i];
305   }
306   bool hasCustomCallingConv() const { return CustomCallSavedXRegs.any(); }
307   bool hasFPARMv8() const { return HasFPARMv8; }
308   bool hasNEON() const { return HasNEON; }
309   bool hasCrypto() const { return HasCrypto; }
310   bool hasDotProd() const { return HasDotProd; }
311   bool hasCRC() const { return HasCRC; }
312   bool hasLSE() const { return HasLSE; }
313   bool hasRAS() const { return HasRAS; }
314   bool hasRDM() const { return HasRDM; }
315   bool hasSM4() const { return HasSM4; }
316   bool hasSHA3() const { return HasSHA3; }
317   bool hasSHA2() const { return HasSHA2; }
318   bool hasAES() const { return HasAES; }
319   bool balanceFPOps() const { return BalanceFPOps; }
320   bool predictableSelectIsExpensive() const {
321     return PredictableSelectIsExpensive;
322   }
323   bool hasCustomCheapAsMoveHandling() const { return CustomAsCheapAsMove; }
324   bool hasExynosCheapAsMoveHandling() const { return ExynosAsCheapAsMove; }
325   bool isMisaligned128StoreSlow() const { return Misaligned128StoreIsSlow; }
326   bool isPaired128Slow() const { return Paired128IsSlow; }
327   bool isSTRQroSlow() const { return STRQroIsSlow; }
328   bool useAlternateSExtLoadCVTF32Pattern() const {
329     return UseAlternateSExtLoadCVTF32Pattern;
330   }
331   bool hasArithmeticBccFusion() const { return HasArithmeticBccFusion; }
332   bool hasArithmeticCbzFusion() const { return HasArithmeticCbzFusion; }
333   bool hasFuseAddress() const { return HasFuseAddress; }
334   bool hasFuseAES() const { return HasFuseAES; }
335   bool hasFuseArithmeticLogic() const { return HasFuseArithmeticLogic; }
336   bool hasFuseCCSelect() const { return HasFuseCCSelect; }
337   bool hasFuseCryptoEOR() const { return HasFuseCryptoEOR; }
338   bool hasFuseLiterals() const { return HasFuseLiterals; }
339 
340   /// Return true if the CPU supports any kind of instruction fusion.
341   bool hasFusion() const {
342     return hasArithmeticBccFusion() || hasArithmeticCbzFusion() ||
343            hasFuseAES() || hasFuseArithmeticLogic() ||
344            hasFuseCCSelect() || hasFuseLiterals();
345   }
346 
347   bool useEL1ForTP() const { return UseEL1ForTP; }
348   bool useEL2ForTP() const { return UseEL2ForTP; }
349   bool useEL3ForTP() const { return UseEL3ForTP; }
350 
351   bool useRSqrt() const { return UseRSqrt; }
352   bool force32BitJumpTables() const { return Force32BitJumpTables; }
353   unsigned getMaxInterleaveFactor() const { return MaxInterleaveFactor; }
354   unsigned getVectorInsertExtractBaseCost() const {
355     return VectorInsertExtractBaseCost;
356   }
357   unsigned getCacheLineSize() const override { return CacheLineSize; }
358   unsigned getPrefetchDistance() const override { return PrefetchDistance; }
359   unsigned getMinPrefetchStride() const override { return MinPrefetchStride; }
360   unsigned getMaxPrefetchIterationsAhead() const override {
361     return MaxPrefetchIterationsAhead;
362   }
363   unsigned getPrefFunctionLogAlignment() const {
364     return PrefFunctionLogAlignment;
365   }
366   unsigned getPrefLoopLogAlignment() const { return PrefLoopLogAlignment; }
367 
368   unsigned getMaximumJumpTableSize() const { return MaxJumpTableSize; }
369 
370   unsigned getWideningBaseCost() const { return WideningBaseCost; }
371 
372   /// CPU has TBI (top byte of addresses is ignored during HW address
373   /// translation) and OS enables it.
374   bool supportsAddressTopByteIgnored() const;
375 
376   bool hasPerfMon() const { return HasPerfMon; }
377   bool hasFullFP16() const { return HasFullFP16; }
378   bool hasFP16FML() const { return HasFP16FML; }
379   bool hasSPE() const { return HasSPE; }
380   bool hasLSLFast() const { return HasLSLFast; }
381   bool hasSVE() const { return HasSVE; }
382   bool hasSVE2() const { return HasSVE2; }
383   bool hasRCPC() const { return HasRCPC; }
384   bool hasAggressiveFMA() const { return HasAggressiveFMA; }
385   bool hasAlternativeNZCV() const { return HasAlternativeNZCV; }
386   bool hasFRInt3264() const { return HasFRInt3264; }
387   bool hasSpecRestrict() const { return HasSpecRestrict; }
388   bool hasSSBS() const { return HasSSBS; }
389   bool hasSB() const { return HasSB; }
390   bool hasPredRes() const { return HasPredRes; }
391   bool hasCCDP() const { return HasCCDP; }
392   bool hasBTI() const { return HasBTI; }
393   bool hasRandGen() const { return HasRandGen; }
394   bool hasMTE() const { return HasMTE; }
395   bool hasTME() const { return HasTME; }
396   // Arm SVE2 extensions
397   bool hasSVE2AES() const { return HasSVE2AES; }
398   bool hasSVE2SM4() const { return HasSVE2SM4; }
399   bool hasSVE2SHA3() const { return HasSVE2SHA3; }
400   bool hasSVE2BitPerm() const { return HasSVE2BitPerm; }
401 
402   bool isLittleEndian() const { return IsLittle; }
403 
404   bool isTargetDarwin() const { return TargetTriple.isOSDarwin(); }
405   bool isTargetIOS() const { return TargetTriple.isiOS(); }
406   bool isTargetLinux() const { return TargetTriple.isOSLinux(); }
407   bool isTargetWindows() const { return TargetTriple.isOSWindows(); }
408   bool isTargetAndroid() const { return TargetTriple.isAndroid(); }
409   bool isTargetFuchsia() const { return TargetTriple.isOSFuchsia(); }
410 
411   bool isTargetCOFF() const { return TargetTriple.isOSBinFormatCOFF(); }
412   bool isTargetELF() const { return TargetTriple.isOSBinFormatELF(); }
413   bool isTargetMachO() const { return TargetTriple.isOSBinFormatMachO(); }
414 
415   bool isTargetILP32() const { return TargetTriple.isArch32Bit(); }
416 
417   bool useAA() const override { return UseAA; }
418 
419   bool hasVH() const { return HasVH; }
420   bool hasPAN() const { return HasPAN; }
421   bool hasLOR() const { return HasLOR; }
422 
423   bool hasPsUAO() const { return HasPsUAO; }
424   bool hasPAN_RWV() const { return HasPAN_RWV; }
425   bool hasCCPP() const { return HasCCPP; }
426 
427   bool hasPA() const { return HasPA; }
428   bool hasJS() const { return HasJS; }
429   bool hasCCIDX() const { return HasCCIDX; }
430   bool hasComplxNum() const { return HasComplxNum; }
431 
432   bool hasNV() const { return HasNV; }
433   bool hasRASv8_4() const { return HasRASv8_4; }
434   bool hasMPAM() const { return HasMPAM; }
435   bool hasDIT() const { return HasDIT; }
436   bool hasTRACEV8_4() const { return HasTRACEV8_4; }
437   bool hasAM() const { return HasAM; }
438   bool hasSEL2() const { return HasSEL2; }
439   bool hasPMU() const { return HasPMU; }
440   bool hasTLB_RMI() const { return HasTLB_RMI; }
441   bool hasFMI() const { return HasFMI; }
442   bool hasRCPC_IMMO() const { return HasRCPC_IMMO; }
443 
444   bool addrSinkUsingGEPs() const override {
445     // Keeping GEPs inbounds is important for exploiting AArch64
446     // addressing-modes in ILP32 mode.
447     return useAA() || isTargetILP32();
448   }
449 
450   bool useSmallAddressing() const {
451     switch (TLInfo.getTargetMachine().getCodeModel()) {
452       case CodeModel::Kernel:
453         // Kernel is currently allowed only for Fuchsia targets,
454         // where it is the same as Small for almost all purposes.
455       case CodeModel::Small:
456         return true;
457       default:
458         return false;
459     }
460   }
461 
462   /// ParseSubtargetFeatures - Parses features string setting specified
463   /// subtarget options.  Definition of function is auto generated by tblgen.
464   void ParseSubtargetFeatures(StringRef CPU, StringRef FS);
465 
466   /// ClassifyGlobalReference - Find the target operand flags that describe
467   /// how a global value should be referenced for the current subtarget.
468   unsigned ClassifyGlobalReference(const GlobalValue *GV,
469                                    const TargetMachine &TM) const;
470 
471   unsigned classifyGlobalFunctionReference(const GlobalValue *GV,
472                                            const TargetMachine &TM) const;
473 
474   void overrideSchedPolicy(MachineSchedPolicy &Policy,
475                            unsigned NumRegionInstrs) const override;
476 
477   bool enableEarlyIfConversion() const override;
478 
479   std::unique_ptr<PBQPRAConstraint> getCustomPBQPConstraints() const override;
480 
481   bool isCallingConvWin64(CallingConv::ID CC) const {
482     switch (CC) {
483     case CallingConv::C:
484     case CallingConv::Fast:
485     case CallingConv::Swift:
486       return isTargetWindows();
487     case CallingConv::Win64:
488       return true;
489     default:
490       return false;
491     }
492   }
493 
494   void mirFileLoaded(MachineFunction &MF) const override;
495 };
496 } // End llvm namespace
497 
498 #endif
499