xref: /freebsd-src/contrib/llvm-project/llvm/lib/Target/ARM/ARMBaseInstrInfo.h (revision 8bcb0991864975618c09697b1aca10683346d9f0)
1 //===-- ARMBaseInstrInfo.h - ARM Base Instruction Information ---*- 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 contains the Base ARM implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H
14 #define LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H
15 
16 #include "MCTargetDesc/ARMBaseInfo.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/CodeGen/MachineBasicBlock.h"
20 #include "llvm/CodeGen/MachineInstr.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineOperand.h"
23 #include "llvm/CodeGen/TargetInstrInfo.h"
24 #include <array>
25 #include <cstdint>
26 
27 #define GET_INSTRINFO_HEADER
28 #include "ARMGenInstrInfo.inc"
29 
30 namespace llvm {
31 
32 class ARMBaseRegisterInfo;
33 class ARMSubtarget;
34 
35 class ARMBaseInstrInfo : public ARMGenInstrInfo {
36   const ARMSubtarget &Subtarget;
37 
38 protected:
39   // Can be only subclassed.
40   explicit ARMBaseInstrInfo(const ARMSubtarget &STI);
41 
42   void expandLoadStackGuardBase(MachineBasicBlock::iterator MI,
43                                 unsigned LoadImmOpc, unsigned LoadOpc) const;
44 
45   /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI
46   /// and \p DefIdx.
47   /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of
48   /// the list is modeled as <Reg:SubReg, SubIdx>.
49   /// E.g., REG_SEQUENCE %1:sub1, sub0, %2, sub1 would produce
50   /// two elements:
51   /// - %1:sub1, sub0
52   /// - %2<:0>, sub1
53   ///
54   /// \returns true if it is possible to build such an input sequence
55   /// with the pair \p MI, \p DefIdx. False otherwise.
56   ///
57   /// \pre MI.isRegSequenceLike().
58   bool getRegSequenceLikeInputs(
59       const MachineInstr &MI, unsigned DefIdx,
60       SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const override;
61 
62   /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI
63   /// and \p DefIdx.
64   /// \p [out] InputReg of the equivalent EXTRACT_SUBREG.
65   /// E.g., EXTRACT_SUBREG %1:sub1, sub0, sub1 would produce:
66   /// - %1:sub1, sub0
67   ///
68   /// \returns true if it is possible to build such an input sequence
69   /// with the pair \p MI, \p DefIdx. False otherwise.
70   ///
71   /// \pre MI.isExtractSubregLike().
72   bool getExtractSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
73                                   RegSubRegPairAndIdx &InputReg) const override;
74 
75   /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI
76   /// and \p DefIdx.
77   /// \p [out] BaseReg and \p [out] InsertedReg contain
78   /// the equivalent inputs of INSERT_SUBREG.
79   /// E.g., INSERT_SUBREG %0:sub0, %1:sub1, sub3 would produce:
80   /// - BaseReg: %0:sub0
81   /// - InsertedReg: %1:sub1, sub3
82   ///
83   /// \returns true if it is possible to build such an input sequence
84   /// with the pair \p MI, \p DefIdx. False otherwise.
85   ///
86   /// \pre MI.isInsertSubregLike().
87   bool
88   getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
89                             RegSubRegPair &BaseReg,
90                             RegSubRegPairAndIdx &InsertedReg) const override;
91 
92   /// Commutes the operands in the given instruction.
93   /// The commutable operands are specified by their indices OpIdx1 and OpIdx2.
94   ///
95   /// Do not call this method for a non-commutable instruction or for
96   /// non-commutable pair of operand indices OpIdx1 and OpIdx2.
97   /// Even though the instruction is commutable, the method may still
98   /// fail to commute the operands, null pointer is returned in such cases.
99   MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
100                                        unsigned OpIdx1,
101                                        unsigned OpIdx2) const override;
102 
103   /// If the specific machine instruction is a instruction that moves/copies
104   /// value from one register to another register return true along with
105   /// @Source machine operand and @Destination machine operand.
106   bool isCopyInstrImpl(const MachineInstr &MI, const MachineOperand *&Source,
107                        const MachineOperand *&Destination) const override;
108 
109 public:
110   // Return whether the target has an explicit NOP encoding.
111   bool hasNOP() const;
112 
113   // Return the non-pre/post incrementing version of 'Opc'. Return 0
114   // if there is not such an opcode.
115   virtual unsigned getUnindexedOpcode(unsigned Opc) const = 0;
116 
117   MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
118                                       MachineInstr &MI,
119                                       LiveVariables *LV) const override;
120 
121   virtual const ARMBaseRegisterInfo &getRegisterInfo() const = 0;
122   const ARMSubtarget &getSubtarget() const { return Subtarget; }
123 
124   ScheduleHazardRecognizer *
125   CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
126                                const ScheduleDAG *DAG) const override;
127 
128   ScheduleHazardRecognizer *
129   CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
130                                      const ScheduleDAG *DAG) const override;
131 
132   // Branch analysis.
133   bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
134                      MachineBasicBlock *&FBB,
135                      SmallVectorImpl<MachineOperand> &Cond,
136                      bool AllowModify = false) const override;
137   unsigned removeBranch(MachineBasicBlock &MBB,
138                         int *BytesRemoved = nullptr) const override;
139   unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
140                         MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
141                         const DebugLoc &DL,
142                         int *BytesAdded = nullptr) const override;
143 
144   bool
145   reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override;
146 
147   // Predication support.
148   bool isPredicated(const MachineInstr &MI) const override;
149 
150   ARMCC::CondCodes getPredicate(const MachineInstr &MI) const {
151     int PIdx = MI.findFirstPredOperandIdx();
152     return PIdx != -1 ? (ARMCC::CondCodes)MI.getOperand(PIdx).getImm()
153                       : ARMCC::AL;
154   }
155 
156   bool PredicateInstruction(MachineInstr &MI,
157                             ArrayRef<MachineOperand> Pred) const override;
158 
159   bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
160                          ArrayRef<MachineOperand> Pred2) const override;
161 
162   bool DefinesPredicate(MachineInstr &MI,
163                         std::vector<MachineOperand> &Pred) const override;
164 
165   bool isPredicable(const MachineInstr &MI) const override;
166 
167   // CPSR defined in instruction
168   static bool isCPSRDefined(const MachineInstr &MI);
169   bool isAddrMode3OpImm(const MachineInstr &MI, unsigned Op) const;
170   bool isAddrMode3OpMinusReg(const MachineInstr &MI, unsigned Op) const;
171 
172   // Load, scaled register offset
173   bool isLdstScaledReg(const MachineInstr &MI, unsigned Op) const;
174   // Load, scaled register offset, not plus LSL2
175   bool isLdstScaledRegNotPlusLsl2(const MachineInstr &MI, unsigned Op) const;
176   // Minus reg for ldstso addr mode
177   bool isLdstSoMinusReg(const MachineInstr &MI, unsigned Op) const;
178   // Scaled register offset in address mode 2
179   bool isAm2ScaledReg(const MachineInstr &MI, unsigned Op) const;
180   // Load multiple, base reg in list
181   bool isLDMBaseRegInList(const MachineInstr &MI) const;
182   // get LDM variable defs size
183   unsigned getLDMVariableDefsSize(const MachineInstr &MI) const;
184 
185   /// GetInstSize - Returns the size of the specified MachineInstr.
186   ///
187   unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
188 
189   unsigned isLoadFromStackSlot(const MachineInstr &MI,
190                                int &FrameIndex) const override;
191   unsigned isStoreToStackSlot(const MachineInstr &MI,
192                               int &FrameIndex) const override;
193   unsigned isLoadFromStackSlotPostFE(const MachineInstr &MI,
194                                      int &FrameIndex) const override;
195   unsigned isStoreToStackSlotPostFE(const MachineInstr &MI,
196                                     int &FrameIndex) const override;
197 
198   void copyToCPSR(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
199                   unsigned SrcReg, bool KillSrc,
200                   const ARMSubtarget &Subtarget) const;
201   void copyFromCPSR(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
202                     unsigned DestReg, bool KillSrc,
203                     const ARMSubtarget &Subtarget) const;
204 
205   void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
206                    const DebugLoc &DL, unsigned DestReg, unsigned SrcReg,
207                    bool KillSrc) const override;
208 
209   void storeRegToStackSlot(MachineBasicBlock &MBB,
210                            MachineBasicBlock::iterator MBBI,
211                            unsigned SrcReg, bool isKill, int FrameIndex,
212                            const TargetRegisterClass *RC,
213                            const TargetRegisterInfo *TRI) const override;
214 
215   void loadRegFromStackSlot(MachineBasicBlock &MBB,
216                             MachineBasicBlock::iterator MBBI,
217                             unsigned DestReg, int FrameIndex,
218                             const TargetRegisterClass *RC,
219                             const TargetRegisterInfo *TRI) const override;
220 
221   bool expandPostRAPseudo(MachineInstr &MI) const override;
222 
223   bool shouldSink(const MachineInstr &MI) const override;
224 
225   void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
226                      unsigned DestReg, unsigned SubIdx,
227                      const MachineInstr &Orig,
228                      const TargetRegisterInfo &TRI) const override;
229 
230   MachineInstr &
231   duplicate(MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
232             const MachineInstr &Orig) const override;
233 
234   const MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
235                                      unsigned SubIdx, unsigned State,
236                                      const TargetRegisterInfo *TRI) const;
237 
238   bool produceSameValue(const MachineInstr &MI0, const MachineInstr &MI1,
239                         const MachineRegisterInfo *MRI) const override;
240 
241   /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
242   /// determine if two loads are loading from the same base address. It should
243   /// only return true if the base pointers are the same and the only
244   /// differences between the two addresses is the offset. It also returns the
245   /// offsets by reference.
246   bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, int64_t &Offset1,
247                                int64_t &Offset2) const override;
248 
249   /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
250   /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads
251   /// should be scheduled togther. On some targets if two loads are loading from
252   /// addresses in the same cache line, it's better if they are scheduled
253   /// together. This function takes two integers that represent the load offsets
254   /// from the common base address. It returns true if it decides it's desirable
255   /// to schedule the two loads together. "NumLoads" is the number of loads that
256   /// have already been scheduled after Load1.
257   bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
258                                int64_t Offset1, int64_t Offset2,
259                                unsigned NumLoads) const override;
260 
261   bool isSchedulingBoundary(const MachineInstr &MI,
262                             const MachineBasicBlock *MBB,
263                             const MachineFunction &MF) const override;
264 
265   bool isProfitableToIfCvt(MachineBasicBlock &MBB,
266                            unsigned NumCycles, unsigned ExtraPredCycles,
267                            BranchProbability Probability) const override;
268 
269   bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
270                            unsigned ExtraT, MachineBasicBlock &FMBB,
271                            unsigned NumF, unsigned ExtraF,
272                            BranchProbability Probability) const override;
273 
274   bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
275                                  BranchProbability Probability) const override {
276     return NumCycles == 1;
277   }
278 
279   unsigned extraSizeToPredicateInstructions(const MachineFunction &MF,
280                                             unsigned NumInsts) const override;
281   unsigned predictBranchSizeForIfCvt(MachineInstr &MI) const override;
282 
283   bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
284                                  MachineBasicBlock &FMBB) const override;
285 
286   /// analyzeCompare - For a comparison instruction, return the source registers
287   /// in SrcReg and SrcReg2 if having two register operands, and the value it
288   /// compares against in CmpValue. Return true if the comparison instruction
289   /// can be analyzed.
290   bool analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
291                       unsigned &SrcReg2, int &CmpMask,
292                       int &CmpValue) const override;
293 
294   /// optimizeCompareInstr - Convert the instruction to set the zero flag so
295   /// that we can remove a "comparison with zero"; Remove a redundant CMP
296   /// instruction if the flags can be updated in the same way by an earlier
297   /// instruction such as SUB.
298   bool optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
299                             unsigned SrcReg2, int CmpMask, int CmpValue,
300                             const MachineRegisterInfo *MRI) const override;
301 
302   bool analyzeSelect(const MachineInstr &MI,
303                      SmallVectorImpl<MachineOperand> &Cond, unsigned &TrueOp,
304                      unsigned &FalseOp, bool &Optimizable) const override;
305 
306   MachineInstr *optimizeSelect(MachineInstr &MI,
307                                SmallPtrSetImpl<MachineInstr *> &SeenMIs,
308                                bool) const override;
309 
310   /// FoldImmediate - 'Reg' is known to be defined by a move immediate
311   /// instruction, try to fold the immediate into the use instruction.
312   bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, unsigned Reg,
313                      MachineRegisterInfo *MRI) const override;
314 
315   unsigned getNumMicroOps(const InstrItineraryData *ItinData,
316                           const MachineInstr &MI) const override;
317 
318   int getOperandLatency(const InstrItineraryData *ItinData,
319                         const MachineInstr &DefMI, unsigned DefIdx,
320                         const MachineInstr &UseMI,
321                         unsigned UseIdx) const override;
322   int getOperandLatency(const InstrItineraryData *ItinData,
323                         SDNode *DefNode, unsigned DefIdx,
324                         SDNode *UseNode, unsigned UseIdx) const override;
325 
326   /// VFP/NEON execution domains.
327   std::pair<uint16_t, uint16_t>
328   getExecutionDomain(const MachineInstr &MI) const override;
329   void setExecutionDomain(MachineInstr &MI, unsigned Domain) const override;
330 
331   unsigned
332   getPartialRegUpdateClearance(const MachineInstr &, unsigned,
333                                const TargetRegisterInfo *) const override;
334   void breakPartialRegDependency(MachineInstr &, unsigned,
335                                  const TargetRegisterInfo *TRI) const override;
336 
337   /// Get the number of addresses by LDM or VLDM or zero for unknown.
338   unsigned getNumLDMAddresses(const MachineInstr &MI) const;
339 
340   std::pair<unsigned, unsigned>
341   decomposeMachineOperandsTargetFlags(unsigned TF) const override;
342   ArrayRef<std::pair<unsigned, const char *>>
343   getSerializableDirectMachineOperandTargetFlags() const override;
344   ArrayRef<std::pair<unsigned, const char *>>
345   getSerializableBitmaskMachineOperandTargetFlags() const override;
346 
347 private:
348   unsigned getInstBundleLength(const MachineInstr &MI) const;
349 
350   int getVLDMDefCycle(const InstrItineraryData *ItinData,
351                       const MCInstrDesc &DefMCID,
352                       unsigned DefClass,
353                       unsigned DefIdx, unsigned DefAlign) const;
354   int getLDMDefCycle(const InstrItineraryData *ItinData,
355                      const MCInstrDesc &DefMCID,
356                      unsigned DefClass,
357                      unsigned DefIdx, unsigned DefAlign) const;
358   int getVSTMUseCycle(const InstrItineraryData *ItinData,
359                       const MCInstrDesc &UseMCID,
360                       unsigned UseClass,
361                       unsigned UseIdx, unsigned UseAlign) const;
362   int getSTMUseCycle(const InstrItineraryData *ItinData,
363                      const MCInstrDesc &UseMCID,
364                      unsigned UseClass,
365                      unsigned UseIdx, unsigned UseAlign) const;
366   int getOperandLatency(const InstrItineraryData *ItinData,
367                         const MCInstrDesc &DefMCID,
368                         unsigned DefIdx, unsigned DefAlign,
369                         const MCInstrDesc &UseMCID,
370                         unsigned UseIdx, unsigned UseAlign) const;
371 
372   int getOperandLatencyImpl(const InstrItineraryData *ItinData,
373                             const MachineInstr &DefMI, unsigned DefIdx,
374                             const MCInstrDesc &DefMCID, unsigned DefAdj,
375                             const MachineOperand &DefMO, unsigned Reg,
376                             const MachineInstr &UseMI, unsigned UseIdx,
377                             const MCInstrDesc &UseMCID, unsigned UseAdj) const;
378 
379   unsigned getPredicationCost(const MachineInstr &MI) const override;
380 
381   unsigned getInstrLatency(const InstrItineraryData *ItinData,
382                            const MachineInstr &MI,
383                            unsigned *PredCost = nullptr) const override;
384 
385   int getInstrLatency(const InstrItineraryData *ItinData,
386                       SDNode *Node) const override;
387 
388   bool hasHighOperandLatency(const TargetSchedModel &SchedModel,
389                              const MachineRegisterInfo *MRI,
390                              const MachineInstr &DefMI, unsigned DefIdx,
391                              const MachineInstr &UseMI,
392                              unsigned UseIdx) const override;
393   bool hasLowDefLatency(const TargetSchedModel &SchedModel,
394                         const MachineInstr &DefMI,
395                         unsigned DefIdx) const override;
396 
397   /// verifyInstruction - Perform target specific instruction verification.
398   bool verifyInstruction(const MachineInstr &MI,
399                          StringRef &ErrInfo) const override;
400 
401   virtual void expandLoadStackGuard(MachineBasicBlock::iterator MI) const = 0;
402 
403   void expandMEMCPY(MachineBasicBlock::iterator) const;
404 
405   /// Identify instructions that can be folded into a MOVCC instruction, and
406   /// return the defining instruction.
407   MachineInstr *canFoldIntoMOVCC(unsigned Reg, const MachineRegisterInfo &MRI,
408                                  const TargetInstrInfo *TII) const;
409 
410 private:
411   /// Modeling special VFP / NEON fp MLA / MLS hazards.
412 
413   /// MLxEntryMap - Map fp MLA / MLS to the corresponding entry in the internal
414   /// MLx table.
415   DenseMap<unsigned, unsigned> MLxEntryMap;
416 
417   /// MLxHazardOpcodes - Set of add / sub and multiply opcodes that would cause
418   /// stalls when scheduled together with fp MLA / MLS opcodes.
419   SmallSet<unsigned, 16> MLxHazardOpcodes;
420 
421 public:
422   /// isFpMLxInstruction - Return true if the specified opcode is a fp MLA / MLS
423   /// instruction.
424   bool isFpMLxInstruction(unsigned Opcode) const {
425     return MLxEntryMap.count(Opcode);
426   }
427 
428   /// isFpMLxInstruction - This version also returns the multiply opcode and the
429   /// addition / subtraction opcode to expand to. Return true for 'HasLane' for
430   /// the MLX instructions with an extra lane operand.
431   bool isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
432                           unsigned &AddSubOpc, bool &NegAcc,
433                           bool &HasLane) const;
434 
435   /// canCauseFpMLxStall - Return true if an instruction of the specified opcode
436   /// will cause stalls when scheduled after (within 4-cycle window) a fp
437   /// MLA / MLS instruction.
438   bool canCauseFpMLxStall(unsigned Opcode) const {
439     return MLxHazardOpcodes.count(Opcode);
440   }
441 
442   /// Returns true if the instruction has a shift by immediate that can be
443   /// executed in one cycle less.
444   bool isSwiftFastImmShift(const MachineInstr *MI) const;
445 
446   /// Returns predicate register associated with the given frame instruction.
447   unsigned getFramePred(const MachineInstr &MI) const {
448     assert(isFrameInstr(MI));
449     // Operands of ADJCALLSTACKDOWN/ADJCALLSTACKUP:
450     // - argument declared in the pattern:
451     // 0 - frame size
452     // 1 - arg of CALLSEQ_START/CALLSEQ_END
453     // 2 - predicate code (like ARMCC::AL)
454     // - added by predOps:
455     // 3 - predicate reg
456     return MI.getOperand(3).getReg();
457   }
458 };
459 
460 /// Get the operands corresponding to the given \p Pred value. By default, the
461 /// predicate register is assumed to be 0 (no register), but you can pass in a
462 /// \p PredReg if that is not the case.
463 static inline std::array<MachineOperand, 2> predOps(ARMCC::CondCodes Pred,
464                                                     unsigned PredReg = 0) {
465   return {{MachineOperand::CreateImm(static_cast<int64_t>(Pred)),
466            MachineOperand::CreateReg(PredReg, false)}};
467 }
468 
469 /// Get the operand corresponding to the conditional code result. By default,
470 /// this is 0 (no register).
471 static inline MachineOperand condCodeOp(unsigned CCReg = 0) {
472   return MachineOperand::CreateReg(CCReg, false);
473 }
474 
475 /// Get the operand corresponding to the conditional code result for Thumb1.
476 /// This operand will always refer to CPSR and it will have the Define flag set.
477 /// You can optionally set the Dead flag by means of \p isDead.
478 static inline MachineOperand t1CondCodeOp(bool isDead = false) {
479   return MachineOperand::CreateReg(ARM::CPSR,
480                                    /*Define*/ true, /*Implicit*/ false,
481                                    /*Kill*/ false, isDead);
482 }
483 
484 static inline
485 bool isUncondBranchOpcode(int Opc) {
486   return Opc == ARM::B || Opc == ARM::tB || Opc == ARM::t2B;
487 }
488 
489 static inline bool isVPTOpcode(int Opc) {
490   return Opc == ARM::MVE_VPTv16i8 || Opc == ARM::MVE_VPTv16u8 ||
491          Opc == ARM::MVE_VPTv16s8 || Opc == ARM::MVE_VPTv8i16 ||
492          Opc == ARM::MVE_VPTv8u16 || Opc == ARM::MVE_VPTv8s16 ||
493          Opc == ARM::MVE_VPTv4i32 || Opc == ARM::MVE_VPTv4u32 ||
494          Opc == ARM::MVE_VPTv4s32 || Opc == ARM::MVE_VPTv4f32 ||
495          Opc == ARM::MVE_VPTv8f16 || Opc == ARM::MVE_VPTv16i8r ||
496          Opc == ARM::MVE_VPTv16u8r || Opc == ARM::MVE_VPTv16s8r ||
497          Opc == ARM::MVE_VPTv8i16r || Opc == ARM::MVE_VPTv8u16r ||
498          Opc == ARM::MVE_VPTv8s16r || Opc == ARM::MVE_VPTv4i32r ||
499          Opc == ARM::MVE_VPTv4u32r || Opc == ARM::MVE_VPTv4s32r ||
500          Opc == ARM::MVE_VPTv4f32r || Opc == ARM::MVE_VPTv8f16r ||
501          Opc == ARM::MVE_VPST;
502 }
503 
504 static inline
505 bool isCondBranchOpcode(int Opc) {
506   return Opc == ARM::Bcc || Opc == ARM::tBcc || Opc == ARM::t2Bcc;
507 }
508 
509 static inline bool isJumpTableBranchOpcode(int Opc) {
510   return Opc == ARM::BR_JTr || Opc == ARM::BR_JTm_i12 ||
511          Opc == ARM::BR_JTm_rs || Opc == ARM::BR_JTadd || Opc == ARM::tBR_JTr ||
512          Opc == ARM::t2BR_JT;
513 }
514 
515 static inline
516 bool isIndirectBranchOpcode(int Opc) {
517   return Opc == ARM::BX || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND;
518 }
519 
520 static inline bool isPopOpcode(int Opc) {
521   return Opc == ARM::tPOP_RET || Opc == ARM::LDMIA_RET ||
522          Opc == ARM::t2LDMIA_RET || Opc == ARM::tPOP || Opc == ARM::LDMIA_UPD ||
523          Opc == ARM::t2LDMIA_UPD || Opc == ARM::VLDMDIA_UPD;
524 }
525 
526 static inline bool isPushOpcode(int Opc) {
527   return Opc == ARM::tPUSH || Opc == ARM::t2STMDB_UPD ||
528          Opc == ARM::STMDB_UPD || Opc == ARM::VSTMDDB_UPD;
529 }
530 
531 /// isValidCoprocessorNumber - decide whether an explicit coprocessor
532 /// number is legal in generic instructions like CDP. The answer can
533 /// vary with the subtarget.
534 static inline bool isValidCoprocessorNumber(unsigned Num,
535                                             const FeatureBitset& featureBits) {
536   // Armv8-A disallows everything *other* than 111x (CP14 and CP15).
537   if (featureBits[ARM::HasV8Ops] && (Num & 0xE) != 0xE)
538     return false;
539 
540   // Armv7 disallows 101x (CP10 and CP11), which clash with VFP/NEON.
541   if (featureBits[ARM::HasV7Ops] && (Num & 0xE) == 0xA)
542     return false;
543 
544   // Armv8.1-M also disallows 100x (CP8,CP9) and 111x (CP14,CP15)
545   // which clash with MVE.
546   if (featureBits[ARM::HasV8_1MMainlineOps] &&
547       ((Num & 0xE) == 0x8 || (Num & 0xE) == 0xE))
548     return false;
549 
550   return true;
551 }
552 
553 /// getInstrPredicate - If instruction is predicated, returns its predicate
554 /// condition, otherwise returns AL. It also returns the condition code
555 /// register by reference.
556 ARMCC::CondCodes getInstrPredicate(const MachineInstr &MI, unsigned &PredReg);
557 
558 unsigned getMatchingCondBranchOpcode(unsigned Opc);
559 
560 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether
561 /// the instruction is encoded with an 'S' bit is determined by the optional
562 /// CPSR def operand.
563 unsigned convertAddSubFlagsOpcode(unsigned OldOpc);
564 
565 /// emitARMRegPlusImmediate / emitT2RegPlusImmediate - Emits a series of
566 /// instructions to materializea destreg = basereg + immediate in ARM / Thumb2
567 /// code.
568 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
569                              MachineBasicBlock::iterator &MBBI,
570                              const DebugLoc &dl, unsigned DestReg,
571                              unsigned BaseReg, int NumBytes,
572                              ARMCC::CondCodes Pred, unsigned PredReg,
573                              const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
574 
575 void emitT2RegPlusImmediate(MachineBasicBlock &MBB,
576                             MachineBasicBlock::iterator &MBBI,
577                             const DebugLoc &dl, unsigned DestReg,
578                             unsigned BaseReg, int NumBytes,
579                             ARMCC::CondCodes Pred, unsigned PredReg,
580                             const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
581 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
582                                MachineBasicBlock::iterator &MBBI,
583                                const DebugLoc &dl, unsigned DestReg,
584                                unsigned BaseReg, int NumBytes,
585                                const TargetInstrInfo &TII,
586                                const ARMBaseRegisterInfo &MRI,
587                                unsigned MIFlags = 0);
588 
589 /// Tries to add registers to the reglist of a given base-updating
590 /// push/pop instruction to adjust the stack by an additional
591 /// NumBytes. This can save a few bytes per function in code-size, but
592 /// obviously generates more memory traffic. As such, it only takes
593 /// effect in functions being optimised for size.
594 bool tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
595                                 MachineFunction &MF, MachineInstr *MI,
596                                 unsigned NumBytes);
597 
598 /// rewriteARMFrameIndex / rewriteT2FrameIndex -
599 /// Rewrite MI to access 'Offset' bytes from the FP. Return false if the
600 /// offset could not be handled directly in MI, and return the left-over
601 /// portion by reference.
602 bool rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
603                           unsigned FrameReg, int &Offset,
604                           const ARMBaseInstrInfo &TII);
605 
606 bool rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
607                          unsigned FrameReg, int &Offset,
608                          const ARMBaseInstrInfo &TII,
609                          const TargetRegisterInfo *TRI);
610 
611 /// Return true if Reg is defd between From and To
612 bool registerDefinedBetween(unsigned Reg, MachineBasicBlock::iterator From,
613                             MachineBasicBlock::iterator To,
614                             const TargetRegisterInfo *TRI);
615 
616 /// Search backwards from a tBcc to find a tCMPi8 against 0, meaning
617 /// we can convert them to a tCBZ or tCBNZ. Return nullptr if not found.
618 MachineInstr *findCMPToFoldIntoCBZ(MachineInstr *Br,
619                                    const TargetRegisterInfo *TRI);
620 
621 void addUnpredicatedMveVpredNOp(MachineInstrBuilder &MIB);
622 void addUnpredicatedMveVpredROp(MachineInstrBuilder &MIB, unsigned DestReg);
623 
624 void addPredicatedMveVpredNOp(MachineInstrBuilder &MIB, unsigned Cond);
625 void addPredicatedMveVpredROp(MachineInstrBuilder &MIB, unsigned Cond,
626                               unsigned Inactive);
627 
628 /// Returns the number of instructions required to materialize the given
629 /// constant in a register, or 3 if a literal pool load is needed.
630 /// If ForCodesize is specified, an approximate cost in bytes is returned.
631 unsigned ConstantMaterializationCost(unsigned Val,
632                                      const ARMSubtarget *Subtarget,
633                                      bool ForCodesize = false);
634 
635 /// Returns true if Val1 has a lower Constant Materialization Cost than Val2.
636 /// Uses the cost from ConstantMaterializationCost, first with ForCodesize as
637 /// specified. If the scores are equal, return the comparison for !ForCodesize.
638 bool HasLowerConstantMaterializationCost(unsigned Val1, unsigned Val2,
639                                          const ARMSubtarget *Subtarget,
640                                          bool ForCodesize = false);
641 
642 } // end namespace llvm
643 
644 #endif // LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H
645