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 "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/IntrinsicsARM.h"
26 #include <array>
27 #include <cstdint>
28 
29 #define GET_INSTRINFO_HEADER
30 #include "ARMGenInstrInfo.inc"
31 
32 namespace llvm {
33 
34 class ARMBaseRegisterInfo;
35 class ARMSubtarget;
36 
37 class ARMBaseInstrInfo : public ARMGenInstrInfo {
38   const ARMSubtarget &Subtarget;
39 
40 protected:
41   // Can be only subclassed.
42   explicit ARMBaseInstrInfo(const ARMSubtarget &STI);
43 
44   void expandLoadStackGuardBase(MachineBasicBlock::iterator MI,
45                                 unsigned LoadImmOpc, unsigned LoadOpc) const;
46 
47   /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI
48   /// and \p DefIdx.
49   /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of
50   /// the list is modeled as <Reg:SubReg, SubIdx>.
51   /// E.g., REG_SEQUENCE %1:sub1, sub0, %2, sub1 would produce
52   /// two elements:
53   /// - %1:sub1, sub0
54   /// - %2<:0>, sub1
55   ///
56   /// \returns true if it is possible to build such an input sequence
57   /// with the pair \p MI, \p DefIdx. False otherwise.
58   ///
59   /// \pre MI.isRegSequenceLike().
60   bool getRegSequenceLikeInputs(
61       const MachineInstr &MI, unsigned DefIdx,
62       SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const override;
63 
64   /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI
65   /// and \p DefIdx.
66   /// \p [out] InputReg of the equivalent EXTRACT_SUBREG.
67   /// E.g., EXTRACT_SUBREG %1:sub1, sub0, sub1 would produce:
68   /// - %1:sub1, sub0
69   ///
70   /// \returns true if it is possible to build such an input sequence
71   /// with the pair \p MI, \p DefIdx. False otherwise.
72   ///
73   /// \pre MI.isExtractSubregLike().
74   bool getExtractSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
75                                   RegSubRegPairAndIdx &InputReg) const override;
76 
77   /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI
78   /// and \p DefIdx.
79   /// \p [out] BaseReg and \p [out] InsertedReg contain
80   /// the equivalent inputs of INSERT_SUBREG.
81   /// E.g., INSERT_SUBREG %0:sub0, %1:sub1, sub3 would produce:
82   /// - BaseReg: %0:sub0
83   /// - InsertedReg: %1:sub1, sub3
84   ///
85   /// \returns true if it is possible to build such an input sequence
86   /// with the pair \p MI, \p DefIdx. False otherwise.
87   ///
88   /// \pre MI.isInsertSubregLike().
89   bool
90   getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
91                             RegSubRegPair &BaseReg,
92                             RegSubRegPairAndIdx &InsertedReg) const override;
93 
94   /// Commutes the operands in the given instruction.
95   /// The commutable operands are specified by their indices OpIdx1 and OpIdx2.
96   ///
97   /// Do not call this method for a non-commutable instruction or for
98   /// non-commutable pair of operand indices OpIdx1 and OpIdx2.
99   /// Even though the instruction is commutable, the method may still
100   /// fail to commute the operands, null pointer is returned in such cases.
101   MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
102                                        unsigned OpIdx1,
103                                        unsigned OpIdx2) const override;
104   /// If the specific machine instruction is an instruction that moves/copies
105   /// value from one register to another register return destination and source
106   /// registers as machine operands.
107   Optional<DestSourcePair>
108   isCopyInstrImpl(const MachineInstr &MI) const override;
109 
110   /// Specialization of \ref TargetInstrInfo::describeLoadedValue, used to
111   /// enhance debug entry value descriptions for ARM targets.
112   Optional<ParamLoadedValue> describeLoadedValue(const MachineInstr &MI,
113                                                  Register Reg) const override;
114 
115 public:
116   // Return whether the target has an explicit NOP encoding.
117   bool hasNOP() const;
118 
119   // Return the non-pre/post incrementing version of 'Opc'. Return 0
120   // if there is not such an opcode.
121   virtual unsigned getUnindexedOpcode(unsigned Opc) const = 0;
122 
123   MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
124                                       MachineInstr &MI,
125                                       LiveVariables *LV) const override;
126 
127   virtual const ARMBaseRegisterInfo &getRegisterInfo() const = 0;
getSubtarget()128   const ARMSubtarget &getSubtarget() const { return Subtarget; }
129 
130   ScheduleHazardRecognizer *
131   CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
132                                const ScheduleDAG *DAG) const override;
133 
134   ScheduleHazardRecognizer *
135   CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
136                                  const ScheduleDAGMI *DAG) const override;
137 
138   ScheduleHazardRecognizer *
139   CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
140                                      const ScheduleDAG *DAG) const override;
141 
142   // Branch analysis.
143   bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
144                      MachineBasicBlock *&FBB,
145                      SmallVectorImpl<MachineOperand> &Cond,
146                      bool AllowModify = false) const override;
147   unsigned removeBranch(MachineBasicBlock &MBB,
148                         int *BytesRemoved = nullptr) const override;
149   unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
150                         MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
151                         const DebugLoc &DL,
152                         int *BytesAdded = nullptr) const override;
153 
154   bool
155   reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override;
156 
157   // Predication support.
158   bool isPredicated(const MachineInstr &MI) const override;
159 
160   // MIR printer helper function to annotate Operands with a comment.
161   std::string
162   createMIROperandComment(const MachineInstr &MI, const MachineOperand &Op,
163                           unsigned OpIdx,
164                           const TargetRegisterInfo *TRI) const override;
165 
getPredicate(const MachineInstr & MI)166   ARMCC::CondCodes getPredicate(const MachineInstr &MI) const {
167     int PIdx = MI.findFirstPredOperandIdx();
168     return PIdx != -1 ? (ARMCC::CondCodes)MI.getOperand(PIdx).getImm()
169                       : ARMCC::AL;
170   }
171 
172   bool PredicateInstruction(MachineInstr &MI,
173                             ArrayRef<MachineOperand> Pred) const override;
174 
175   bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
176                          ArrayRef<MachineOperand> Pred2) const override;
177 
178   bool ClobbersPredicate(MachineInstr &MI, std::vector<MachineOperand> &Pred,
179                          bool SkipDead) const override;
180 
181   bool isPredicable(const MachineInstr &MI) const override;
182 
183   // CPSR defined in instruction
184   static bool isCPSRDefined(const MachineInstr &MI);
185 
186   /// GetInstSize - Returns the size of the specified MachineInstr.
187   ///
188   unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
189 
190   unsigned isLoadFromStackSlot(const MachineInstr &MI,
191                                int &FrameIndex) const override;
192   unsigned isStoreToStackSlot(const MachineInstr &MI,
193                               int &FrameIndex) const override;
194   unsigned isLoadFromStackSlotPostFE(const MachineInstr &MI,
195                                      int &FrameIndex) const override;
196   unsigned isStoreToStackSlotPostFE(const MachineInstr &MI,
197                                     int &FrameIndex) const override;
198 
199   void copyToCPSR(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
200                   unsigned SrcReg, bool KillSrc,
201                   const ARMSubtarget &Subtarget) const;
202   void copyFromCPSR(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
203                     unsigned DestReg, bool KillSrc,
204                     const ARMSubtarget &Subtarget) const;
205 
206   void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
207                    const DebugLoc &DL, MCRegister DestReg, MCRegister SrcReg,
208                    bool KillSrc) const override;
209 
210   void storeRegToStackSlot(MachineBasicBlock &MBB,
211                            MachineBasicBlock::iterator MBBI,
212                            Register SrcReg, bool isKill, int FrameIndex,
213                            const TargetRegisterClass *RC,
214                            const TargetRegisterInfo *TRI) const override;
215 
216   void loadRegFromStackSlot(MachineBasicBlock &MBB,
217                             MachineBasicBlock::iterator MBBI,
218                             Register DestReg, int FrameIndex,
219                             const TargetRegisterClass *RC,
220                             const TargetRegisterInfo *TRI) const override;
221 
222   bool expandPostRAPseudo(MachineInstr &MI) const override;
223 
224   bool shouldSink(const MachineInstr &MI) const override;
225 
226   void reMaterialize(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
227                      Register DestReg, unsigned SubIdx,
228                      const MachineInstr &Orig,
229                      const TargetRegisterInfo &TRI) const override;
230 
231   MachineInstr &
232   duplicate(MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
233             const MachineInstr &Orig) const override;
234 
235   const MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB, unsigned Reg,
236                                      unsigned SubIdx, unsigned State,
237                                      const TargetRegisterInfo *TRI) const;
238 
239   bool produceSameValue(const MachineInstr &MI0, const MachineInstr &MI1,
240                         const MachineRegisterInfo *MRI) const override;
241 
242   /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to
243   /// determine if two loads are loading from the same base address. It should
244   /// only return true if the base pointers are the same and the only
245   /// differences between the two addresses is the offset. It also returns the
246   /// offsets by reference.
247   bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, int64_t &Offset1,
248                                int64_t &Offset2) const override;
249 
250   /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
251   /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads
252   /// should be scheduled togther. On some targets if two loads are loading from
253   /// addresses in the same cache line, it's better if they are scheduled
254   /// together. This function takes two integers that represent the load offsets
255   /// from the common base address. It returns true if it decides it's desirable
256   /// to schedule the two loads together. "NumLoads" is the number of loads that
257   /// have already been scheduled after Load1.
258   bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
259                                int64_t Offset1, int64_t Offset2,
260                                unsigned NumLoads) const override;
261 
262   bool isSchedulingBoundary(const MachineInstr &MI,
263                             const MachineBasicBlock *MBB,
264                             const MachineFunction &MF) const override;
265 
266   bool isProfitableToIfCvt(MachineBasicBlock &MBB,
267                            unsigned NumCycles, unsigned ExtraPredCycles,
268                            BranchProbability Probability) const override;
269 
270   bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT,
271                            unsigned ExtraT, MachineBasicBlock &FMBB,
272                            unsigned NumF, unsigned ExtraF,
273                            BranchProbability Probability) const override;
274 
isProfitableToDupForIfCvt(MachineBasicBlock & MBB,unsigned NumCycles,BranchProbability Probability)275   bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
276                                  BranchProbability Probability) const override {
277     return NumCycles == 1;
278   }
279 
280   unsigned extraSizeToPredicateInstructions(const MachineFunction &MF,
281                                             unsigned NumInsts) const override;
282   unsigned predictBranchSizeForIfCvt(MachineInstr &MI) const override;
283 
284   bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
285                                  MachineBasicBlock &FMBB) const override;
286 
287   /// analyzeCompare - For a comparison instruction, return the source registers
288   /// in SrcReg and SrcReg2 if having two register operands, and the value it
289   /// compares against in CmpValue. Return true if the comparison instruction
290   /// can be analyzed.
291   bool analyzeCompare(const MachineInstr &MI, Register &SrcReg,
292                       Register &SrcReg2, int &CmpMask,
293                       int &CmpValue) const override;
294 
295   /// optimizeCompareInstr - Convert the instruction to set the zero flag so
296   /// that we can remove a "comparison with zero"; Remove a redundant CMP
297   /// instruction if the flags can be updated in the same way by an earlier
298   /// instruction such as SUB.
299   bool optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
300                             Register SrcReg2, int CmpMask, int CmpValue,
301                             const MachineRegisterInfo *MRI) const override;
302 
303   bool analyzeSelect(const MachineInstr &MI,
304                      SmallVectorImpl<MachineOperand> &Cond, unsigned &TrueOp,
305                      unsigned &FalseOp, bool &Optimizable) const override;
306 
307   MachineInstr *optimizeSelect(MachineInstr &MI,
308                                SmallPtrSetImpl<MachineInstr *> &SeenMIs,
309                                bool) const override;
310 
311   /// FoldImmediate - 'Reg' is known to be defined by a move immediate
312   /// instruction, try to fold the immediate into the use instruction.
313   bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, Register Reg,
314                      MachineRegisterInfo *MRI) const override;
315 
316   unsigned getNumMicroOps(const InstrItineraryData *ItinData,
317                           const MachineInstr &MI) const override;
318 
319   int getOperandLatency(const InstrItineraryData *ItinData,
320                         const MachineInstr &DefMI, unsigned DefIdx,
321                         const MachineInstr &UseMI,
322                         unsigned UseIdx) const override;
323   int getOperandLatency(const InstrItineraryData *ItinData,
324                         SDNode *DefNode, unsigned DefIdx,
325                         SDNode *UseNode, unsigned UseIdx) const override;
326 
327   /// VFP/NEON execution domains.
328   std::pair<uint16_t, uint16_t>
329   getExecutionDomain(const MachineInstr &MI) const override;
330   void setExecutionDomain(MachineInstr &MI, unsigned Domain) const override;
331 
332   unsigned
333   getPartialRegUpdateClearance(const MachineInstr &, unsigned,
334                                const TargetRegisterInfo *) const override;
335   void breakPartialRegDependency(MachineInstr &, unsigned,
336                                  const TargetRegisterInfo *TRI) const override;
337 
338   /// Get the number of addresses by LDM or VLDM or zero for unknown.
339   unsigned getNumLDMAddresses(const MachineInstr &MI) const;
340 
341   std::pair<unsigned, unsigned>
342   decomposeMachineOperandsTargetFlags(unsigned TF) const override;
343   ArrayRef<std::pair<unsigned, const char *>>
344   getSerializableDirectMachineOperandTargetFlags() const override;
345   ArrayRef<std::pair<unsigned, const char *>>
346   getSerializableBitmaskMachineOperandTargetFlags() const override;
347 
348   /// ARM supports the MachineOutliner.
349   bool isFunctionSafeToOutlineFrom(MachineFunction &MF,
350                                    bool OutlineFromLinkOnceODRs) const override;
351   outliner::OutlinedFunction getOutliningCandidateInfo(
352       std::vector<outliner::Candidate> &RepeatedSequenceLocs) const override;
353   outliner::InstrType getOutliningType(MachineBasicBlock::iterator &MIT,
354                                        unsigned Flags) const override;
355   bool isMBBSafeToOutlineFrom(MachineBasicBlock &MBB,
356                               unsigned &Flags) const override;
357   void buildOutlinedFrame(MachineBasicBlock &MBB, MachineFunction &MF,
358                           const outliner::OutlinedFunction &OF) const override;
359   MachineBasicBlock::iterator
360   insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
361                      MachineBasicBlock::iterator &It, MachineFunction &MF,
362                      const outliner::Candidate &C) const override;
363 
364   /// Enable outlining by default at -Oz.
365   bool shouldOutlineFromFunctionByDefault(MachineFunction &MF) const override;
366 
isUnspillableTerminatorImpl(const MachineInstr * MI)367   bool isUnspillableTerminatorImpl(const MachineInstr *MI) const override {
368     return MI->getOpcode() == ARM::t2LoopEndDec ||
369            MI->getOpcode() == ARM::t2DoLoopStartTP ||
370            MI->getOpcode() == ARM::t2WhileLoopStartLR ||
371            MI->getOpcode() == ARM::t2WhileLoopStartTP;
372   }
373 
374 private:
375   /// Returns an unused general-purpose register which can be used for
376   /// constructing an outlined call if one exists. Returns 0 otherwise.
377   unsigned findRegisterToSaveLRTo(const outliner::Candidate &C) const;
378 
379   // Adds an instruction which saves the link register on top of the stack into
380   /// the MachineBasicBlock \p MBB at position \p It.
381   void saveLROnStack(MachineBasicBlock &MBB,
382                      MachineBasicBlock::iterator It) const;
383 
384   /// Adds an instruction which restores the link register from the top the
385   /// stack into the MachineBasicBlock \p MBB at position \p It.
386   void restoreLRFromStack(MachineBasicBlock &MBB,
387                           MachineBasicBlock::iterator It) const;
388 
389   /// Emit CFI instructions into the MachineBasicBlock \p MBB at position \p It,
390   /// for the case when the LR is saved on the stack.
391   void emitCFIForLRSaveOnStack(MachineBasicBlock &MBB,
392                                MachineBasicBlock::iterator It) const;
393 
394   /// Emit CFI instructions into the MachineBasicBlock \p MBB at position \p It,
395   /// for the case when the LR is saved in the register \p Reg.
396   void emitCFIForLRSaveToReg(MachineBasicBlock &MBB,
397                              MachineBasicBlock::iterator It,
398                              Register Reg) const;
399 
400   /// Emit CFI instructions into the MachineBasicBlock \p MBB at position \p It,
401   /// after the LR is was restored from the stack.
402   void emitCFIForLRRestoreFromStack(MachineBasicBlock &MBB,
403                                     MachineBasicBlock::iterator It) const;
404 
405   /// Emit CFI instructions into the MachineBasicBlock \p MBB at position \p It,
406   /// after the LR is was restored from a register.
407   void emitCFIForLRRestoreFromReg(MachineBasicBlock &MBB,
408                                   MachineBasicBlock::iterator It) const;
409   /// \brief Sets the offsets on outlined instructions in \p MBB which use SP
410   /// so that they will be valid post-outlining.
411   ///
412   /// \param MBB A \p MachineBasicBlock in an outlined function.
413   void fixupPostOutline(MachineBasicBlock &MBB) const;
414 
415   /// Returns true if the machine instruction offset can handle the stack fixup
416   /// and updates it if requested.
417   bool checkAndUpdateStackOffset(MachineInstr *MI, int64_t Fixup,
418                                  bool Updt) const;
419 
420   unsigned getInstBundleLength(const MachineInstr &MI) const;
421 
422   int getVLDMDefCycle(const InstrItineraryData *ItinData,
423                       const MCInstrDesc &DefMCID,
424                       unsigned DefClass,
425                       unsigned DefIdx, unsigned DefAlign) const;
426   int getLDMDefCycle(const InstrItineraryData *ItinData,
427                      const MCInstrDesc &DefMCID,
428                      unsigned DefClass,
429                      unsigned DefIdx, unsigned DefAlign) const;
430   int getVSTMUseCycle(const InstrItineraryData *ItinData,
431                       const MCInstrDesc &UseMCID,
432                       unsigned UseClass,
433                       unsigned UseIdx, unsigned UseAlign) const;
434   int getSTMUseCycle(const InstrItineraryData *ItinData,
435                      const MCInstrDesc &UseMCID,
436                      unsigned UseClass,
437                      unsigned UseIdx, unsigned UseAlign) const;
438   int getOperandLatency(const InstrItineraryData *ItinData,
439                         const MCInstrDesc &DefMCID,
440                         unsigned DefIdx, unsigned DefAlign,
441                         const MCInstrDesc &UseMCID,
442                         unsigned UseIdx, unsigned UseAlign) const;
443 
444   int getOperandLatencyImpl(const InstrItineraryData *ItinData,
445                             const MachineInstr &DefMI, unsigned DefIdx,
446                             const MCInstrDesc &DefMCID, unsigned DefAdj,
447                             const MachineOperand &DefMO, unsigned Reg,
448                             const MachineInstr &UseMI, unsigned UseIdx,
449                             const MCInstrDesc &UseMCID, unsigned UseAdj) const;
450 
451   unsigned getPredicationCost(const MachineInstr &MI) const override;
452 
453   unsigned getInstrLatency(const InstrItineraryData *ItinData,
454                            const MachineInstr &MI,
455                            unsigned *PredCost = nullptr) const override;
456 
457   int getInstrLatency(const InstrItineraryData *ItinData,
458                       SDNode *Node) const override;
459 
460   bool hasHighOperandLatency(const TargetSchedModel &SchedModel,
461                              const MachineRegisterInfo *MRI,
462                              const MachineInstr &DefMI, unsigned DefIdx,
463                              const MachineInstr &UseMI,
464                              unsigned UseIdx) const override;
465   bool hasLowDefLatency(const TargetSchedModel &SchedModel,
466                         const MachineInstr &DefMI,
467                         unsigned DefIdx) const override;
468 
469   /// verifyInstruction - Perform target specific instruction verification.
470   bool verifyInstruction(const MachineInstr &MI,
471                          StringRef &ErrInfo) const override;
472 
473   virtual void expandLoadStackGuard(MachineBasicBlock::iterator MI) const = 0;
474 
475   void expandMEMCPY(MachineBasicBlock::iterator) const;
476 
477   /// Identify instructions that can be folded into a MOVCC instruction, and
478   /// return the defining instruction.
479   MachineInstr *canFoldIntoMOVCC(Register Reg, const MachineRegisterInfo &MRI,
480                                  const TargetInstrInfo *TII) const;
481 
482   bool isReallyTriviallyReMaterializable(const MachineInstr &MI,
483                                          AAResults *AA) const override;
484 
485 private:
486   /// Modeling special VFP / NEON fp MLA / MLS hazards.
487 
488   /// MLxEntryMap - Map fp MLA / MLS to the corresponding entry in the internal
489   /// MLx table.
490   DenseMap<unsigned, unsigned> MLxEntryMap;
491 
492   /// MLxHazardOpcodes - Set of add / sub and multiply opcodes that would cause
493   /// stalls when scheduled together with fp MLA / MLS opcodes.
494   SmallSet<unsigned, 16> MLxHazardOpcodes;
495 
496 public:
497   /// isFpMLxInstruction - Return true if the specified opcode is a fp MLA / MLS
498   /// instruction.
isFpMLxInstruction(unsigned Opcode)499   bool isFpMLxInstruction(unsigned Opcode) const {
500     return MLxEntryMap.count(Opcode);
501   }
502 
503   /// isFpMLxInstruction - This version also returns the multiply opcode and the
504   /// addition / subtraction opcode to expand to. Return true for 'HasLane' for
505   /// the MLX instructions with an extra lane operand.
506   bool isFpMLxInstruction(unsigned Opcode, unsigned &MulOpc,
507                           unsigned &AddSubOpc, bool &NegAcc,
508                           bool &HasLane) const;
509 
510   /// canCauseFpMLxStall - Return true if an instruction of the specified opcode
511   /// will cause stalls when scheduled after (within 4-cycle window) a fp
512   /// MLA / MLS instruction.
canCauseFpMLxStall(unsigned Opcode)513   bool canCauseFpMLxStall(unsigned Opcode) const {
514     return MLxHazardOpcodes.count(Opcode);
515   }
516 
517   /// Returns true if the instruction has a shift by immediate that can be
518   /// executed in one cycle less.
519   bool isSwiftFastImmShift(const MachineInstr *MI) const;
520 
521   /// Returns predicate register associated with the given frame instruction.
getFramePred(const MachineInstr & MI)522   unsigned getFramePred(const MachineInstr &MI) const {
523     assert(isFrameInstr(MI));
524     // Operands of ADJCALLSTACKDOWN/ADJCALLSTACKUP:
525     // - argument declared in the pattern:
526     // 0 - frame size
527     // 1 - arg of CALLSEQ_START/CALLSEQ_END
528     // 2 - predicate code (like ARMCC::AL)
529     // - added by predOps:
530     // 3 - predicate reg
531     return MI.getOperand(3).getReg();
532   }
533 
534   Optional<RegImmPair> isAddImmediate(const MachineInstr &MI,
535                                       Register Reg) const override;
536 };
537 
538 /// Get the operands corresponding to the given \p Pred value. By default, the
539 /// predicate register is assumed to be 0 (no register), but you can pass in a
540 /// \p PredReg if that is not the case.
541 static inline std::array<MachineOperand, 2> predOps(ARMCC::CondCodes Pred,
542                                                     unsigned PredReg = 0) {
543   return {{MachineOperand::CreateImm(static_cast<int64_t>(Pred)),
544            MachineOperand::CreateReg(PredReg, false)}};
545 }
546 
547 /// Get the operand corresponding to the conditional code result. By default,
548 /// this is 0 (no register).
549 static inline MachineOperand condCodeOp(unsigned CCReg = 0) {
550   return MachineOperand::CreateReg(CCReg, false);
551 }
552 
553 /// Get the operand corresponding to the conditional code result for Thumb1.
554 /// This operand will always refer to CPSR and it will have the Define flag set.
555 /// You can optionally set the Dead flag by means of \p isDead.
556 static inline MachineOperand t1CondCodeOp(bool isDead = false) {
557   return MachineOperand::CreateReg(ARM::CPSR,
558                                    /*Define*/ true, /*Implicit*/ false,
559                                    /*Kill*/ false, isDead);
560 }
561 
562 static inline
isUncondBranchOpcode(int Opc)563 bool isUncondBranchOpcode(int Opc) {
564   return Opc == ARM::B || Opc == ARM::tB || Opc == ARM::t2B;
565 }
566 
567 // This table shows the VPT instruction variants, i.e. the different
568 // mask field encodings, see also B5.6. Predication/conditional execution in
569 // the ArmARM.
isVPTOpcode(int Opc)570 static inline bool isVPTOpcode(int Opc) {
571   return Opc == ARM::MVE_VPTv16i8 || Opc == ARM::MVE_VPTv16u8 ||
572          Opc == ARM::MVE_VPTv16s8 || Opc == ARM::MVE_VPTv8i16 ||
573          Opc == ARM::MVE_VPTv8u16 || Opc == ARM::MVE_VPTv8s16 ||
574          Opc == ARM::MVE_VPTv4i32 || Opc == ARM::MVE_VPTv4u32 ||
575          Opc == ARM::MVE_VPTv4s32 || Opc == ARM::MVE_VPTv4f32 ||
576          Opc == ARM::MVE_VPTv8f16 || Opc == ARM::MVE_VPTv16i8r ||
577          Opc == ARM::MVE_VPTv16u8r || Opc == ARM::MVE_VPTv16s8r ||
578          Opc == ARM::MVE_VPTv8i16r || Opc == ARM::MVE_VPTv8u16r ||
579          Opc == ARM::MVE_VPTv8s16r || Opc == ARM::MVE_VPTv4i32r ||
580          Opc == ARM::MVE_VPTv4u32r || Opc == ARM::MVE_VPTv4s32r ||
581          Opc == ARM::MVE_VPTv4f32r || Opc == ARM::MVE_VPTv8f16r ||
582          Opc == ARM::MVE_VPST;
583 }
584 
585 static inline
VCMPOpcodeToVPT(unsigned Opcode)586 unsigned VCMPOpcodeToVPT(unsigned Opcode) {
587   switch (Opcode) {
588   default:
589     return 0;
590   case ARM::MVE_VCMPf32:
591     return ARM::MVE_VPTv4f32;
592   case ARM::MVE_VCMPf16:
593     return ARM::MVE_VPTv8f16;
594   case ARM::MVE_VCMPi8:
595     return ARM::MVE_VPTv16i8;
596   case ARM::MVE_VCMPi16:
597     return ARM::MVE_VPTv8i16;
598   case ARM::MVE_VCMPi32:
599     return ARM::MVE_VPTv4i32;
600   case ARM::MVE_VCMPu8:
601     return ARM::MVE_VPTv16u8;
602   case ARM::MVE_VCMPu16:
603     return ARM::MVE_VPTv8u16;
604   case ARM::MVE_VCMPu32:
605     return ARM::MVE_VPTv4u32;
606   case ARM::MVE_VCMPs8:
607     return ARM::MVE_VPTv16s8;
608   case ARM::MVE_VCMPs16:
609     return ARM::MVE_VPTv8s16;
610   case ARM::MVE_VCMPs32:
611     return ARM::MVE_VPTv4s32;
612 
613   case ARM::MVE_VCMPf32r:
614     return ARM::MVE_VPTv4f32r;
615   case ARM::MVE_VCMPf16r:
616     return ARM::MVE_VPTv8f16r;
617   case ARM::MVE_VCMPi8r:
618     return ARM::MVE_VPTv16i8r;
619   case ARM::MVE_VCMPi16r:
620     return ARM::MVE_VPTv8i16r;
621   case ARM::MVE_VCMPi32r:
622     return ARM::MVE_VPTv4i32r;
623   case ARM::MVE_VCMPu8r:
624     return ARM::MVE_VPTv16u8r;
625   case ARM::MVE_VCMPu16r:
626     return ARM::MVE_VPTv8u16r;
627   case ARM::MVE_VCMPu32r:
628     return ARM::MVE_VPTv4u32r;
629   case ARM::MVE_VCMPs8r:
630     return ARM::MVE_VPTv16s8r;
631   case ARM::MVE_VCMPs16r:
632     return ARM::MVE_VPTv8s16r;
633   case ARM::MVE_VCMPs32r:
634     return ARM::MVE_VPTv4s32r;
635   }
636 }
637 
638 static inline
isCondBranchOpcode(int Opc)639 bool isCondBranchOpcode(int Opc) {
640   return Opc == ARM::Bcc || Opc == ARM::tBcc || Opc == ARM::t2Bcc;
641 }
642 
isJumpTableBranchOpcode(int Opc)643 static inline bool isJumpTableBranchOpcode(int Opc) {
644   return Opc == ARM::BR_JTr || Opc == ARM::BR_JTm_i12 ||
645          Opc == ARM::BR_JTm_rs || Opc == ARM::BR_JTadd || Opc == ARM::tBR_JTr ||
646          Opc == ARM::t2BR_JT;
647 }
648 
649 static inline
isIndirectBranchOpcode(int Opc)650 bool isIndirectBranchOpcode(int Opc) {
651   return Opc == ARM::BX || Opc == ARM::MOVPCRX || Opc == ARM::tBRIND;
652 }
653 
isIndirectCall(const MachineInstr & MI)654 static inline bool isIndirectCall(const MachineInstr &MI) {
655   int Opc = MI.getOpcode();
656   switch (Opc) {
657     // indirect calls:
658   case ARM::BLX:
659   case ARM::BLX_noip:
660   case ARM::BLX_pred:
661   case ARM::BLX_pred_noip:
662   case ARM::BX_CALL:
663   case ARM::BMOVPCRX_CALL:
664   case ARM::TCRETURNri:
665   case ARM::TAILJMPr:
666   case ARM::TAILJMPr4:
667   case ARM::tBLXr:
668   case ARM::tBLXr_noip:
669   case ARM::tBLXNSr:
670   case ARM::tBLXNS_CALL:
671   case ARM::tBX_CALL:
672   case ARM::tTAILJMPr:
673     assert(MI.isCall(MachineInstr::IgnoreBundle));
674     return true;
675     // direct calls:
676   case ARM::BL:
677   case ARM::BL_pred:
678   case ARM::BMOVPCB_CALL:
679   case ARM::BL_PUSHLR:
680   case ARM::BLXi:
681   case ARM::TCRETURNdi:
682   case ARM::TAILJMPd:
683   case ARM::SVC:
684   case ARM::HVC:
685   case ARM::TPsoft:
686   case ARM::tTAILJMPd:
687   case ARM::t2SMC:
688   case ARM::t2HVC:
689   case ARM::tBL:
690   case ARM::tBLXi:
691   case ARM::tBL_PUSHLR:
692   case ARM::tTAILJMPdND:
693   case ARM::tSVC:
694   case ARM::tTPsoft:
695     assert(MI.isCall(MachineInstr::IgnoreBundle));
696     return false;
697   }
698   assert(!MI.isCall(MachineInstr::IgnoreBundle));
699   return false;
700 }
701 
isIndirectControlFlowNotComingBack(const MachineInstr & MI)702 static inline bool isIndirectControlFlowNotComingBack(const MachineInstr &MI) {
703   int opc = MI.getOpcode();
704   return MI.isReturn() || isIndirectBranchOpcode(MI.getOpcode()) ||
705          isJumpTableBranchOpcode(opc);
706 }
707 
isSpeculationBarrierEndBBOpcode(int Opc)708 static inline bool isSpeculationBarrierEndBBOpcode(int Opc) {
709   return Opc == ARM::SpeculationBarrierISBDSBEndBB ||
710          Opc == ARM::SpeculationBarrierSBEndBB ||
711          Opc == ARM::t2SpeculationBarrierISBDSBEndBB ||
712          Opc == ARM::t2SpeculationBarrierSBEndBB;
713 }
714 
isPopOpcode(int Opc)715 static inline bool isPopOpcode(int Opc) {
716   return Opc == ARM::tPOP_RET || Opc == ARM::LDMIA_RET ||
717          Opc == ARM::t2LDMIA_RET || Opc == ARM::tPOP || Opc == ARM::LDMIA_UPD ||
718          Opc == ARM::t2LDMIA_UPD || Opc == ARM::VLDMDIA_UPD;
719 }
720 
isPushOpcode(int Opc)721 static inline bool isPushOpcode(int Opc) {
722   return Opc == ARM::tPUSH || Opc == ARM::t2STMDB_UPD ||
723          Opc == ARM::STMDB_UPD || Opc == ARM::VSTMDDB_UPD;
724 }
725 
isSubImmOpcode(int Opc)726 static inline bool isSubImmOpcode(int Opc) {
727   return Opc == ARM::SUBri ||
728          Opc == ARM::tSUBi3 || Opc == ARM::tSUBi8 ||
729          Opc == ARM::tSUBSi3 || Opc == ARM::tSUBSi8 ||
730          Opc == ARM::t2SUBri || Opc == ARM::t2SUBri12 || Opc == ARM::t2SUBSri;
731 }
732 
isMovRegOpcode(int Opc)733 static inline bool isMovRegOpcode(int Opc) {
734   return Opc == ARM::MOVr || Opc == ARM::tMOVr || Opc == ARM::t2MOVr;
735 }
736 /// isValidCoprocessorNumber - decide whether an explicit coprocessor
737 /// number is legal in generic instructions like CDP. The answer can
738 /// vary with the subtarget.
isValidCoprocessorNumber(unsigned Num,const FeatureBitset & featureBits)739 static inline bool isValidCoprocessorNumber(unsigned Num,
740                                             const FeatureBitset& featureBits) {
741   // In Armv7 and Armv8-M CP10 and CP11 clash with VFP/NEON, however, the
742   // coprocessor is still valid for CDP/MCR/MRC and friends. Allowing it is
743   // useful for code which is shared with older architectures which do not know
744   // the new VFP/NEON mnemonics.
745 
746   // Armv8-A disallows everything *other* than 111x (CP14 and CP15).
747   if (featureBits[ARM::HasV8Ops] && (Num & 0xE) != 0xE)
748     return false;
749 
750   // Armv8.1-M disallows 100x (CP8,CP9) and 111x (CP14,CP15)
751   // which clash with MVE.
752   if (featureBits[ARM::HasV8_1MMainlineOps] &&
753       ((Num & 0xE) == 0x8 || (Num & 0xE) == 0xE))
754     return false;
755 
756   return true;
757 }
758 
759 /// getInstrPredicate - If instruction is predicated, returns its predicate
760 /// condition, otherwise returns AL. It also returns the condition code
761 /// register by reference.
762 ARMCC::CondCodes getInstrPredicate(const MachineInstr &MI, Register &PredReg);
763 
764 unsigned getMatchingCondBranchOpcode(unsigned Opc);
765 
766 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether
767 /// the instruction is encoded with an 'S' bit is determined by the optional
768 /// CPSR def operand.
769 unsigned convertAddSubFlagsOpcode(unsigned OldOpc);
770 
771 /// emitARMRegPlusImmediate / emitT2RegPlusImmediate - Emits a series of
772 /// instructions to materializea destreg = basereg + immediate in ARM / Thumb2
773 /// code.
774 void emitARMRegPlusImmediate(MachineBasicBlock &MBB,
775                              MachineBasicBlock::iterator &MBBI,
776                              const DebugLoc &dl, Register DestReg,
777                              Register BaseReg, int NumBytes,
778                              ARMCC::CondCodes Pred, Register PredReg,
779                              const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
780 
781 void emitT2RegPlusImmediate(MachineBasicBlock &MBB,
782                             MachineBasicBlock::iterator &MBBI,
783                             const DebugLoc &dl, Register DestReg,
784                             Register BaseReg, int NumBytes,
785                             ARMCC::CondCodes Pred, Register PredReg,
786                             const ARMBaseInstrInfo &TII, unsigned MIFlags = 0);
787 void emitThumbRegPlusImmediate(MachineBasicBlock &MBB,
788                                MachineBasicBlock::iterator &MBBI,
789                                const DebugLoc &dl, Register DestReg,
790                                Register BaseReg, int NumBytes,
791                                const TargetInstrInfo &TII,
792                                const ARMBaseRegisterInfo &MRI,
793                                unsigned MIFlags = 0);
794 
795 /// Tries to add registers to the reglist of a given base-updating
796 /// push/pop instruction to adjust the stack by an additional
797 /// NumBytes. This can save a few bytes per function in code-size, but
798 /// obviously generates more memory traffic. As such, it only takes
799 /// effect in functions being optimised for size.
800 bool tryFoldSPUpdateIntoPushPop(const ARMSubtarget &Subtarget,
801                                 MachineFunction &MF, MachineInstr *MI,
802                                 unsigned NumBytes);
803 
804 /// rewriteARMFrameIndex / rewriteT2FrameIndex -
805 /// Rewrite MI to access 'Offset' bytes from the FP. Return false if the
806 /// offset could not be handled directly in MI, and return the left-over
807 /// portion by reference.
808 bool rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
809                           Register FrameReg, int &Offset,
810                           const ARMBaseInstrInfo &TII);
811 
812 bool rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
813                          Register FrameReg, int &Offset,
814                          const ARMBaseInstrInfo &TII,
815                          const TargetRegisterInfo *TRI);
816 
817 /// Return true if Reg is defd between From and To
818 bool registerDefinedBetween(unsigned Reg, MachineBasicBlock::iterator From,
819                             MachineBasicBlock::iterator To,
820                             const TargetRegisterInfo *TRI);
821 
822 /// Search backwards from a tBcc to find a tCMPi8 against 0, meaning
823 /// we can convert them to a tCBZ or tCBNZ. Return nullptr if not found.
824 MachineInstr *findCMPToFoldIntoCBZ(MachineInstr *Br,
825                                    const TargetRegisterInfo *TRI);
826 
827 void addUnpredicatedMveVpredNOp(MachineInstrBuilder &MIB);
828 void addUnpredicatedMveVpredROp(MachineInstrBuilder &MIB, Register DestReg);
829 
830 void addPredicatedMveVpredNOp(MachineInstrBuilder &MIB, unsigned Cond);
831 void addPredicatedMveVpredROp(MachineInstrBuilder &MIB, unsigned Cond,
832                               unsigned Inactive);
833 
834 /// Returns the number of instructions required to materialize the given
835 /// constant in a register, or 3 if a literal pool load is needed.
836 /// If ForCodesize is specified, an approximate cost in bytes is returned.
837 unsigned ConstantMaterializationCost(unsigned Val,
838                                      const ARMSubtarget *Subtarget,
839                                      bool ForCodesize = false);
840 
841 /// Returns true if Val1 has a lower Constant Materialization Cost than Val2.
842 /// Uses the cost from ConstantMaterializationCost, first with ForCodesize as
843 /// specified. If the scores are equal, return the comparison for !ForCodesize.
844 bool HasLowerConstantMaterializationCost(unsigned Val1, unsigned Val2,
845                                          const ARMSubtarget *Subtarget,
846                                          bool ForCodesize = false);
847 
848 // Return the immediate if this is ADDri or SUBri, scaled as appropriate.
849 // Returns 0 for unknown instructions.
getAddSubImmediate(MachineInstr & MI)850 inline int getAddSubImmediate(MachineInstr &MI) {
851   int Scale = 1;
852   unsigned ImmOp;
853   switch (MI.getOpcode()) {
854   case ARM::t2ADDri:
855     ImmOp = 2;
856     break;
857   case ARM::t2SUBri:
858   case ARM::t2SUBri12:
859     ImmOp = 2;
860     Scale = -1;
861     break;
862   case ARM::tSUBi3:
863   case ARM::tSUBi8:
864     ImmOp = 3;
865     Scale = -1;
866     break;
867   default:
868     return 0;
869   }
870   return Scale * MI.getOperand(ImmOp).getImm();
871 }
872 
873 // Given a memory access Opcode, check that the give Imm would be a valid Offset
874 // for this instruction using its addressing mode.
isLegalAddressImm(unsigned Opcode,int Imm,const TargetInstrInfo * TII)875 inline bool isLegalAddressImm(unsigned Opcode, int Imm,
876                               const TargetInstrInfo *TII) {
877   const MCInstrDesc &Desc = TII->get(Opcode);
878   unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
879   switch (AddrMode) {
880   case ARMII::AddrModeT2_i7:
881     return std::abs(Imm) < (((1 << 7) * 1) - 1);
882   case ARMII::AddrModeT2_i7s2:
883     return std::abs(Imm) < (((1 << 7) * 2) - 1) && Imm % 2 == 0;
884   case ARMII::AddrModeT2_i7s4:
885     return std::abs(Imm) < (((1 << 7) * 4) - 1) && Imm % 4 == 0;
886   case ARMII::AddrModeT2_i8:
887     return std::abs(Imm) < (((1 << 8) * 1) - 1);
888   case ARMII::AddrMode2:
889     return std::abs(Imm) < (((1 << 12) * 1) - 1);
890   case ARMII::AddrModeT2_i12:
891     return Imm >= 0 && Imm < (((1 << 12) * 1) - 1);
892   case ARMII::AddrModeT2_i8s4:
893     return std::abs(Imm) < (((1 << 8) * 4) - 1) && Imm % 4 == 0;
894   default:
895     llvm_unreachable("Unhandled Addressing mode");
896   }
897 }
898 
899 // Return true if the given intrinsic is a gather
isGather(IntrinsicInst * IntInst)900 inline bool isGather(IntrinsicInst *IntInst) {
901   if (IntInst == nullptr)
902     return false;
903   unsigned IntrinsicID = IntInst->getIntrinsicID();
904   return (IntrinsicID == Intrinsic::masked_gather ||
905           IntrinsicID == Intrinsic::arm_mve_vldr_gather_base ||
906           IntrinsicID == Intrinsic::arm_mve_vldr_gather_base_predicated ||
907           IntrinsicID == Intrinsic::arm_mve_vldr_gather_base_wb ||
908           IntrinsicID == Intrinsic::arm_mve_vldr_gather_base_wb_predicated ||
909           IntrinsicID == Intrinsic::arm_mve_vldr_gather_offset ||
910           IntrinsicID == Intrinsic::arm_mve_vldr_gather_offset_predicated);
911 }
912 
913 // Return true if the given intrinsic is a scatter
isScatter(IntrinsicInst * IntInst)914 inline bool isScatter(IntrinsicInst *IntInst) {
915   if (IntInst == nullptr)
916     return false;
917   unsigned IntrinsicID = IntInst->getIntrinsicID();
918   return (IntrinsicID == Intrinsic::masked_scatter ||
919           IntrinsicID == Intrinsic::arm_mve_vstr_scatter_base ||
920           IntrinsicID == Intrinsic::arm_mve_vstr_scatter_base_predicated ||
921           IntrinsicID == Intrinsic::arm_mve_vstr_scatter_base_wb ||
922           IntrinsicID == Intrinsic::arm_mve_vstr_scatter_base_wb_predicated ||
923           IntrinsicID == Intrinsic::arm_mve_vstr_scatter_offset ||
924           IntrinsicID == Intrinsic::arm_mve_vstr_scatter_offset_predicated);
925 }
926 
927 // Return true if the given intrinsic is a gather or scatter
isGatherScatter(IntrinsicInst * IntInst)928 inline bool isGatherScatter(IntrinsicInst *IntInst) {
929   if (IntInst == nullptr)
930     return false;
931   return isGather(IntInst) || isScatter(IntInst);
932 }
933 
934 unsigned getBLXOpcode(const MachineFunction &MF);
935 unsigned gettBLXrOpcode(const MachineFunction &MF);
936 unsigned getBLXpredOpcode(const MachineFunction &MF);
937 
938 } // end namespace llvm
939 
940 #endif // LLVM_LIB_TARGET_ARM_ARMBASEINSTRINFO_H
941