Coverage Report

Created: 2018-07-19 03:59

/Users/buildslave/jenkins/workspace/clang-stage2-coverage-R/llvm/include/llvm/CodeGen/TargetInstrInfo.h
Line
Count
Source (jump to first uncovered line)
1
//===- llvm/CodeGen/TargetInstrInfo.h - Instruction Info --------*- C++ -*-===//
2
//
3
//                     The LLVM Compiler Infrastructure
4
//
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
7
//
8
//===----------------------------------------------------------------------===//
9
//
10
// This file describes the target machine instruction set to the code generator.
11
//
12
//===----------------------------------------------------------------------===//
13
14
#ifndef LLVM_TARGET_TARGETINSTRINFO_H
15
#define LLVM_TARGET_TARGETINSTRINFO_H
16
17
#include "llvm/ADT/ArrayRef.h"
18
#include "llvm/ADT/DenseMap.h"
19
#include "llvm/ADT/DenseMapInfo.h"
20
#include "llvm/ADT/None.h"
21
#include "llvm/CodeGen/LiveRegUnits.h"
22
#include "llvm/CodeGen/MachineBasicBlock.h"
23
#include "llvm/CodeGen/MachineCombinerPattern.h"
24
#include "llvm/CodeGen/MachineFunction.h"
25
#include "llvm/CodeGen/MachineInstr.h"
26
#include "llvm/CodeGen/MachineLoopInfo.h"
27
#include "llvm/CodeGen/MachineOperand.h"
28
#include "llvm/CodeGen/MachineOutliner.h"
29
#include "llvm/CodeGen/PseudoSourceValue.h"
30
#include "llvm/MC/MCInstrInfo.h"
31
#include "llvm/Support/BranchProbability.h"
32
#include "llvm/Support/ErrorHandling.h"
33
#include <cassert>
34
#include <cstddef>
35
#include <cstdint>
36
#include <utility>
37
#include <vector>
38
39
namespace llvm {
40
41
class DFAPacketizer;
42
class InstrItineraryData;
43
class LiveIntervals;
44
class LiveVariables;
45
class MachineMemOperand;
46
class MachineRegisterInfo;
47
class MCAsmInfo;
48
class MCInst;
49
struct MCSchedModel;
50
class Module;
51
class ScheduleDAG;
52
class ScheduleHazardRecognizer;
53
class SDNode;
54
class SelectionDAG;
55
class RegScavenger;
56
class TargetRegisterClass;
57
class TargetRegisterInfo;
58
class TargetSchedModel;
59
class TargetSubtargetInfo;
60
61
template <class T> class SmallVectorImpl;
62
63
//---------------------------------------------------------------------------
64
///
65
/// TargetInstrInfo - Interface to description of machine instruction set
66
///
67
class TargetInstrInfo : public MCInstrInfo {
68
public:
69
  TargetInstrInfo(unsigned CFSetupOpcode = ~0u, unsigned CFDestroyOpcode = ~0u,
70
                  unsigned CatchRetOpcode = ~0u, unsigned ReturnOpcode = ~0u)
71
      : CallFrameSetupOpcode(CFSetupOpcode),
72
        CallFrameDestroyOpcode(CFDestroyOpcode), CatchRetOpcode(CatchRetOpcode),
73
44.4k
        ReturnOpcode(ReturnOpcode) {}
74
  TargetInstrInfo(const TargetInstrInfo &) = delete;
75
  TargetInstrInfo &operator=(const TargetInstrInfo &) = delete;
76
  virtual ~TargetInstrInfo();
77
78
22.1M
  static bool isGenericOpcode(unsigned Opc) {
79
22.1M
    return Opc <= TargetOpcode::GENERIC_OP_END;
80
22.1M
  }
81
82
  /// Given a machine instruction descriptor, returns the register
83
  /// class constraint for OpNum, or NULL.
84
  const TargetRegisterClass *getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
85
                                         const TargetRegisterInfo *TRI,
86
                                         const MachineFunction &MF) const;
87
88
  /// Return true if the instruction is trivially rematerializable, meaning it
89
  /// has no side effects and requires no operands that aren't always available.
90
  /// This means the only allowed uses are constants and unallocatable physical
91
  /// registers so that the instructions result is independent of the place
92
  /// in the function.
93
  bool isTriviallyReMaterializable(const MachineInstr &MI,
94
6.11M
                                   AliasAnalysis *AA = nullptr) const {
95
6.11M
    return MI.getOpcode() == TargetOpcode::IMPLICIT_DEF ||
96
6.11M
           
(6.11M
MI.getDesc().isRematerializable()6.11M
&&
97
6.11M
            
(3.75M
isReallyTriviallyReMaterializable(MI, AA)3.75M
||
98
3.75M
             
isReallyTriviallyReMaterializableGeneric(MI, AA)3.54M
));
99
6.11M
  }
100
101
protected:
102
  /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
103
  /// set, this hook lets the target specify whether the instruction is actually
104
  /// trivially rematerializable, taking into consideration its operands. This
105
  /// predicate must return false if the instruction has any side effects other
106
  /// than producing a value, or if it requres any address registers that are
107
  /// not always available.
108
  /// Requirements must be check as stated in isTriviallyReMaterializable() .
109
  virtual bool isReallyTriviallyReMaterializable(const MachineInstr &MI,
110
3.30M
                                                 AliasAnalysis *AA) const {
111
3.30M
    return false;
112
3.30M
  }
113
114
  /// This method commutes the operands of the given machine instruction MI.
115
  /// The operands to be commuted are specified by their indices OpIdx1 and
116
  /// OpIdx2.
117
  ///
118
  /// If a target has any instructions that are commutable but require
119
  /// converting to different instructions or making non-trivial changes
120
  /// to commute them, this method can be overloaded to do that.
121
  /// The default implementation simply swaps the commutable operands.
122
  ///
123
  /// If NewMI is false, MI is modified in place and returned; otherwise, a
124
  /// new machine instruction is created and returned.
125
  ///
126
  /// Do not call this method for a non-commutable instruction.
127
  /// Even though the instruction is commutable, the method may still
128
  /// fail to commute the operands, null pointer is returned in such cases.
129
  virtual MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
130
                                               unsigned OpIdx1,
131
                                               unsigned OpIdx2) const;
132
133
  /// Assigns the (CommutableOpIdx1, CommutableOpIdx2) pair of commutable
134
  /// operand indices to (ResultIdx1, ResultIdx2).
135
  /// One or both input values of the pair: (ResultIdx1, ResultIdx2) may be
136
  /// predefined to some indices or be undefined (designated by the special
137
  /// value 'CommuteAnyOperandIndex').
138
  /// The predefined result indices cannot be re-defined.
139
  /// The function returns true iff after the result pair redefinition
140
  /// the fixed result pair is equal to or equivalent to the source pair of
141
  /// indices: (CommutableOpIdx1, CommutableOpIdx2). It is assumed here that
142
  /// the pairs (x,y) and (y,x) are equivalent.
143
  static bool fixCommutedOpIndices(unsigned &ResultIdx1, unsigned &ResultIdx2,
144
                                   unsigned CommutableOpIdx1,
145
                                   unsigned CommutableOpIdx2);
146
147
private:
148
  /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
149
  /// set and the target hook isReallyTriviallyReMaterializable returns false,
150
  /// this function does target-independent tests to determine if the
151
  /// instruction is really trivially rematerializable.
152
  bool isReallyTriviallyReMaterializableGeneric(const MachineInstr &MI,
153
                                                AliasAnalysis *AA) const;
154
155
public:
156
  /// These methods return the opcode of the frame setup/destroy instructions
157
  /// if they exist (-1 otherwise).  Some targets use pseudo instructions in
158
  /// order to abstract away the difference between operating with a frame
159
  /// pointer and operating without, through the use of these two instructions.
160
  ///
161
43.4M
  unsigned getCallFrameSetupOpcode() const { return CallFrameSetupOpcode; }
162
55.5M
  unsigned getCallFrameDestroyOpcode() const { return CallFrameDestroyOpcode; }
163
164
  /// Returns true if the argument is a frame pseudo instruction.
165
37.8M
  bool isFrameInstr(const MachineInstr &I) const {
166
37.8M
    return I.getOpcode() == getCallFrameSetupOpcode() ||
167
37.8M
           
I.getOpcode() == getCallFrameDestroyOpcode()36.4M
;
168
37.8M
  }
169
170
  /// Returns true if the argument is a frame setup pseudo instruction.
171
459k
  bool isFrameSetup(const MachineInstr &I) const {
172
459k
    return I.getOpcode() == getCallFrameSetupOpcode();
173
459k
  }
174
175
  /// Returns size of the frame associated with the given frame instruction.
176
  /// For frame setup instruction this is frame that is set up space set up
177
  /// after the instruction. For frame destroy instruction this is the frame
178
  /// freed by the caller.
179
  /// Note, in some cases a call frame (or a part of it) may be prepared prior
180
  /// to the frame setup instruction. It occurs in the calls that involve
181
  /// inalloca arguments. This function reports only the size of the frame part
182
  /// that is set up between the frame setup and destroy pseudo instructions.
183
5.89M
  int64_t getFrameSize(const MachineInstr &I) const {
184
5.89M
    assert(isFrameInstr(I) && "Not a frame instruction");
185
5.89M
    assert(I.getOperand(0).getImm() >= 0);
186
5.89M
    return I.getOperand(0).getImm();
187
5.89M
  }
188
189
  /// Returns the total frame size, which is made up of the space set up inside
190
  /// the pair of frame start-stop instructions and the space that is set up
191
  /// prior to the pair.
192
231k
  int64_t getFrameTotalSize(const MachineInstr &I) const {
193
231k
    if (isFrameSetup(I)) {
194
115k
      assert(I.getOperand(1).getImm() >= 0 &&
195
115k
             "Frame size must not be negative");
196
115k
      return getFrameSize(I) + I.getOperand(1).getImm();
197
115k
    }
198
115k
    return getFrameSize(I);
199
115k
  }
200
201
1.23k
  unsigned getCatchReturnOpcode() const { return CatchRetOpcode; }
202
30
  unsigned getReturnOpcode() const { return ReturnOpcode; }
203
204
  /// Returns the actual stack pointer adjustment made by an instruction
205
  /// as part of a call sequence. By default, only call frame setup/destroy
206
  /// instructions adjust the stack, but targets may want to override this
207
  /// to enable more fine-grained adjustment, or adjust by a different value.
208
  virtual int getSPAdjust(const MachineInstr &MI) const;
209
210
  /// Return true if the instruction is a "coalescable" extension instruction.
211
  /// That is, it's like a copy where it's legal for the source to overlap the
212
  /// destination. e.g. X86::MOVSX64rr32. If this returns true, then it's
213
  /// expected the pre-extension value is available as a subreg of the result
214
  /// register. This also returns the sub-register index in SubIdx.
215
  virtual bool isCoalescableExtInstr(const MachineInstr &MI, unsigned &SrcReg,
216
1.81M
                                     unsigned &DstReg, unsigned &SubIdx) const {
217
1.81M
    return false;
218
1.81M
  }
219
220
  /// If the specified machine instruction is a direct
221
  /// load from a stack slot, return the virtual or physical register number of
222
  /// the destination along with the FrameIndex of the loaded stack slot.  If
223
  /// not, return 0.  This predicate must return 0 if the instruction has
224
  /// any side effects other than loading from the stack slot.
225
  virtual unsigned isLoadFromStackSlot(const MachineInstr &MI,
226
1.70k
                                       int &FrameIndex) const {
227
1.70k
    return 0;
228
1.70k
  }
229
230
  /// Optional extension of isLoadFromStackSlot that returns the number of
231
  /// bytes loaded from the stack. This must be implemented if a backend
232
  /// supports partial stack slot spills/loads to further disambiguate
233
  /// what the load does.
234
  virtual unsigned isLoadFromStackSlot(const MachineInstr &MI,
235
                                       int &FrameIndex,
236
4.22M
                                       unsigned &MemBytes) const {
237
4.22M
    MemBytes = 0;
238
4.22M
    return isLoadFromStackSlot(MI, FrameIndex);
239
4.22M
  }
240
241
  /// Check for post-frame ptr elimination stack locations as well.
242
  /// This uses a heuristic so it isn't reliable for correctness.
243
  virtual unsigned isLoadFromStackSlotPostFE(const MachineInstr &MI,
244
778k
                                             int &FrameIndex) const {
245
778k
    return 0;
246
778k
  }
247
248
  /// If the specified machine instruction has a load from a stack slot,
249
  /// return true along with the FrameIndex of the loaded stack slot and the
250
  /// machine mem operand containing the reference.
251
  /// If not, return false.  Unlike isLoadFromStackSlot, this returns true for
252
  /// any instructions that loads from the stack.  This is just a hint, as some
253
  /// cases may be missed.
254
  virtual bool hasLoadFromStackSlot(const MachineInstr &MI,
255
                                    const MachineMemOperand *&MMO,
256
                                    int &FrameIndex) const;
257
258
  /// If the specified machine instruction is a direct
259
  /// store to a stack slot, return the virtual or physical register number of
260
  /// the source reg along with the FrameIndex of the loaded stack slot.  If
261
  /// not, return 0.  This predicate must return 0 if the instruction has
262
  /// any side effects other than storing to the stack slot.
263
  virtual unsigned isStoreToStackSlot(const MachineInstr &MI,
264
1.03k
                                      int &FrameIndex) const {
265
1.03k
    return 0;
266
1.03k
  }
267
268
  /// Optional extension of isStoreToStackSlot that returns the number of
269
  /// bytes stored to the stack. This must be implemented if a backend
270
  /// supports partial stack slot spills/loads to further disambiguate
271
  /// what the store does.
272
  virtual unsigned isStoreToStackSlot(const MachineInstr &MI,
273
                                      int &FrameIndex,
274
150k
                                      unsigned &MemBytes) const {
275
150k
    MemBytes = 0;
276
150k
    return isStoreToStackSlot(MI, FrameIndex);
277
150k
  }
278
279
  /// Check for post-frame ptr elimination stack locations as well.
280
  /// This uses a heuristic, so it isn't reliable for correctness.
281
  virtual unsigned isStoreToStackSlotPostFE(const MachineInstr &MI,
282
758k
                                            int &FrameIndex) const {
283
758k
    return 0;
284
758k
  }
285
286
  /// If the specified machine instruction has a store to a stack slot,
287
  /// return true along with the FrameIndex of the loaded stack slot and the
288
  /// machine mem operand containing the reference.
289
  /// If not, return false.  Unlike isStoreToStackSlot,
290
  /// this returns true for any instructions that stores to the
291
  /// stack.  This is just a hint, as some cases may be missed.
292
  virtual bool hasStoreToStackSlot(const MachineInstr &MI,
293
                                   const MachineMemOperand *&MMO,
294
                                   int &FrameIndex) const;
295
296
  /// Return true if the specified machine instruction
297
  /// is a copy of one stack slot to another and has no other effect.
298
  /// Provide the identity of the two frame indices.
299
  virtual bool isStackSlotCopy(const MachineInstr &MI, int &DestFrameIndex,
300
5.09M
                               int &SrcFrameIndex) const {
301
5.09M
    return false;
302
5.09M
  }
303
304
  /// Compute the size in bytes and offset within a stack slot of a spilled
305
  /// register or subregister.
306
  ///
307
  /// \param [out] Size in bytes of the spilled value.
308
  /// \param [out] Offset in bytes within the stack slot.
309
  /// \returns true if both Size and Offset are successfully computed.
310
  ///
311
  /// Not all subregisters have computable spill slots. For example,
312
  /// subregisters registers may not be byte-sized, and a pair of discontiguous
313
  /// subregisters has no single offset.
314
  ///
315
  /// Targets with nontrivial bigendian implementations may need to override
316
  /// this, particularly to support spilled vector registers.
317
  virtual bool getStackSlotRange(const TargetRegisterClass *RC, unsigned SubIdx,
318
                                 unsigned &Size, unsigned &Offset,
319
                                 const MachineFunction &MF) const;
320
321
  /// Returns the size in bytes of the specified MachineInstr, or ~0U
322
  /// when this function is not implemented by a target.
323
0
  virtual unsigned getInstSizeInBytes(const MachineInstr &MI) const {
324
0
    return ~0U;
325
0
  }
326
327
  /// Return true if the instruction is as cheap as a move instruction.
328
  ///
329
  /// Targets for different archs need to override this, and different
330
  /// micro-architectures can also be finely tuned inside.
331
675k
  virtual bool isAsCheapAsAMove(const MachineInstr &MI) const {
332
675k
    return MI.isAsCheapAsAMove();
333
675k
  }
334
335
  /// Return true if the instruction should be sunk by MachineSink.
336
  ///
337
  /// MachineSink determines on its own whether the instruction is safe to sink;
338
  /// this gives the target a hook to override the default behavior with regards
339
  /// to which instructions should be sunk.
340
29.3M
  virtual bool shouldSink(const MachineInstr &MI) const { return true; }
341
342
  /// Re-issue the specified 'original' instruction at the
343
  /// specific location targeting a new destination register.
344
  /// The register in Orig->getOperand(0).getReg() will be substituted by
345
  /// DestReg:SubIdx. Any existing subreg index is preserved or composed with
346
  /// SubIdx.
347
  virtual void reMaterialize(MachineBasicBlock &MBB,
348
                             MachineBasicBlock::iterator MI, unsigned DestReg,
349
                             unsigned SubIdx, const MachineInstr &Orig,
350
                             const TargetRegisterInfo &TRI) const;
351
352
  /// Clones instruction or the whole instruction bundle \p Orig and
353
  /// insert into \p MBB before \p InsertBefore. The target may update operands
354
  /// that are required to be unique.
355
  ///
356
  /// \p Orig must not return true for MachineInstr::isNotDuplicable().
357
  virtual MachineInstr &duplicate(MachineBasicBlock &MBB,
358
                                  MachineBasicBlock::iterator InsertBefore,
359
                                  const MachineInstr &Orig) const;
360
361
  /// This method must be implemented by targets that
362
  /// set the M_CONVERTIBLE_TO_3_ADDR flag.  When this flag is set, the target
363
  /// may be able to convert a two-address instruction into one or more true
364
  /// three-address instructions on demand.  This allows the X86 target (for
365
  /// example) to convert ADD and SHL instructions into LEA instructions if they
366
  /// would require register copies due to two-addressness.
367
  ///
368
  /// This method returns a null pointer if the transformation cannot be
369
  /// performed, otherwise it returns the last new instruction.
370
  ///
371
  virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI,
372
                                              MachineInstr &MI,
373
0
                                              LiveVariables *LV) const {
374
0
    return nullptr;
375
0
  }
376
377
  // This constant can be used as an input value of operand index passed to
378
  // the method findCommutedOpIndices() to tell the method that the
379
  // corresponding operand index is not pre-defined and that the method
380
  // can pick any commutable operand.
381
  static const unsigned CommuteAnyOperandIndex = ~0U;
382
383
  /// This method commutes the operands of the given machine instruction MI.
384
  ///
385
  /// The operands to be commuted are specified by their indices OpIdx1 and
386
  /// OpIdx2. OpIdx1 and OpIdx2 arguments may be set to a special value
387
  /// 'CommuteAnyOperandIndex', which means that the method is free to choose
388
  /// any arbitrarily chosen commutable operand. If both arguments are set to
389
  /// 'CommuteAnyOperandIndex' then the method looks for 2 different commutable
390
  /// operands; then commutes them if such operands could be found.
391
  ///
392
  /// If NewMI is false, MI is modified in place and returned; otherwise, a
393
  /// new machine instruction is created and returned.
394
  ///
395
  /// Do not call this method for a non-commutable instruction or
396
  /// for non-commuable operands.
397
  /// Even though the instruction is commutable, the method may still
398
  /// fail to commute the operands, null pointer is returned in such cases.
399
  MachineInstr *
400
  commuteInstruction(MachineInstr &MI, bool NewMI = false,
401
                     unsigned OpIdx1 = CommuteAnyOperandIndex,
402
                     unsigned OpIdx2 = CommuteAnyOperandIndex) const;
403
404
  /// Returns true iff the routine could find two commutable operands in the
405
  /// given machine instruction.
406
  /// The 'SrcOpIdx1' and 'SrcOpIdx2' are INPUT and OUTPUT arguments.
407
  /// If any of the INPUT values is set to the special value
408
  /// 'CommuteAnyOperandIndex' then the method arbitrarily picks a commutable
409
  /// operand, then returns its index in the corresponding argument.
410
  /// If both of INPUT values are set to 'CommuteAnyOperandIndex' then method
411
  /// looks for 2 commutable operands.
412
  /// If INPUT values refer to some operands of MI, then the method simply
413
  /// returns true if the corresponding operands are commutable and returns
414
  /// false otherwise.
415
  ///
416
  /// For example, calling this method this way:
417
  ///     unsigned Op1 = 1, Op2 = CommuteAnyOperandIndex;
418
  ///     findCommutedOpIndices(MI, Op1, Op2);
419
  /// can be interpreted as a query asking to find an operand that would be
420
  /// commutable with the operand#1.
421
  virtual bool findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
422
                                     unsigned &SrcOpIdx2) const;
423
424
  /// A pair composed of a register and a sub-register index.
425
  /// Used to give some type checking when modeling Reg:SubReg.
426
  struct RegSubRegPair {
427
    unsigned Reg;
428
    unsigned SubReg;
429
430
    RegSubRegPair(unsigned Reg = 0, unsigned SubReg = 0)
431
59.6M
        : Reg(Reg), SubReg(SubReg) {}
432
  };
433
434
  /// A pair composed of a pair of a register and a sub-register index,
435
  /// and another sub-register index.
436
  /// Used to give some type checking when modeling Reg:SubReg1, SubReg2.
437
  struct RegSubRegPairAndIdx : RegSubRegPair {
438
    unsigned SubIdx;
439
440
    RegSubRegPairAndIdx(unsigned Reg = 0, unsigned SubReg = 0,
441
                        unsigned SubIdx = 0)
442
803k
        : RegSubRegPair(Reg, SubReg), SubIdx(SubIdx) {}
443
  };
444
445
  /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI
446
  /// and \p DefIdx.
447
  /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of
448
  /// the list is modeled as <Reg:SubReg, SubIdx>. Operands with the undef
449
  /// flag are not added to this list.
450
  /// E.g., REG_SEQUENCE %1:sub1, sub0, %2, sub1 would produce
451
  /// two elements:
452
  /// - %1:sub1, sub0
453
  /// - %2<:0>, sub1
454
  ///
455
  /// \returns true if it is possible to build such an input sequence
456
  /// with the pair \p MI, \p DefIdx. False otherwise.
457
  ///
458
  /// \pre MI.isRegSequence() or MI.isRegSequenceLike().
459
  ///
460
  /// \note The generic implementation does not provide any support for
461
  /// MI.isRegSequenceLike(). In other words, one has to override
462
  /// getRegSequenceLikeInputs for target specific instructions.
463
  bool
464
  getRegSequenceInputs(const MachineInstr &MI, unsigned DefIdx,
465
                       SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const;
466
467
  /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI
468
  /// and \p DefIdx.
469
  /// \p [out] InputReg of the equivalent EXTRACT_SUBREG.
470
  /// E.g., EXTRACT_SUBREG %1:sub1, sub0, sub1 would produce:
471
  /// - %1:sub1, sub0
472
  ///
473
  /// \returns true if it is possible to build such an input sequence
474
  /// with the pair \p MI, \p DefIdx and the operand has no undef flag set.
475
  /// False otherwise.
476
  ///
477
  /// \pre MI.isExtractSubreg() or MI.isExtractSubregLike().
478
  ///
479
  /// \note The generic implementation does not provide any support for
480
  /// MI.isExtractSubregLike(). In other words, one has to override
481
  /// getExtractSubregLikeInputs for target specific instructions.
482
  bool getExtractSubregInputs(const MachineInstr &MI, unsigned DefIdx,
483
                              RegSubRegPairAndIdx &InputReg) const;
484
485
  /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI
486
  /// and \p DefIdx.
487
  /// \p [out] BaseReg and \p [out] InsertedReg contain
488
  /// the equivalent inputs of INSERT_SUBREG.
489
  /// E.g., INSERT_SUBREG %0:sub0, %1:sub1, sub3 would produce:
490
  /// - BaseReg: %0:sub0
491
  /// - InsertedReg: %1:sub1, sub3
492
  ///
493
  /// \returns true if it is possible to build such an input sequence
494
  /// with the pair \p MI, \p DefIdx and the operand has no undef flag set.
495
  /// False otherwise.
496
  ///
497
  /// \pre MI.isInsertSubreg() or MI.isInsertSubregLike().
498
  ///
499
  /// \note The generic implementation does not provide any support for
500
  /// MI.isInsertSubregLike(). In other words, one has to override
501
  /// getInsertSubregLikeInputs for target specific instructions.
502
  bool getInsertSubregInputs(const MachineInstr &MI, unsigned DefIdx,
503
                             RegSubRegPair &BaseReg,
504
                             RegSubRegPairAndIdx &InsertedReg) const;
505
506
  /// Return true if two machine instructions would produce identical values.
507
  /// By default, this is only true when the two instructions
508
  /// are deemed identical except for defs. If this function is called when the
509
  /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for
510
  /// aggressive checks.
511
  virtual bool produceSameValue(const MachineInstr &MI0,
512
                                const MachineInstr &MI1,
513
                                const MachineRegisterInfo *MRI = nullptr) const;
514
515
  /// \returns true if a branch from an instruction with opcode \p BranchOpc
516
  ///  bytes is capable of jumping to a position \p BrOffset bytes away.
517
  virtual bool isBranchOffsetInRange(unsigned BranchOpc,
518
0
                                     int64_t BrOffset) const {
519
0
    llvm_unreachable("target did not implement");
520
0
  }
521
522
  /// \returns The block that branch instruction \p MI jumps to.
523
0
  virtual MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const {
524
0
    llvm_unreachable("target did not implement");
525
0
  }
526
527
  /// Insert an unconditional indirect branch at the end of \p MBB to \p
528
  /// NewDestBB.  \p BrOffset indicates the offset of \p NewDestBB relative to
529
  /// the offset of the position to insert the new branch.
530
  ///
531
  /// \returns The number of bytes added to the block.
532
  virtual unsigned insertIndirectBranch(MachineBasicBlock &MBB,
533
                                        MachineBasicBlock &NewDestBB,
534
                                        const DebugLoc &DL,
535
                                        int64_t BrOffset = 0,
536
0
                                        RegScavenger *RS = nullptr) const {
537
0
    llvm_unreachable("target did not implement");
538
0
  }
539
540
  /// Analyze the branching code at the end of MBB, returning
541
  /// true if it cannot be understood (e.g. it's a switch dispatch or isn't
542
  /// implemented for a target).  Upon success, this returns false and returns
543
  /// with the following information in various cases:
544
  ///
545
  /// 1. If this block ends with no branches (it just falls through to its succ)
546
  ///    just return false, leaving TBB/FBB null.
547
  /// 2. If this block ends with only an unconditional branch, it sets TBB to be
548
  ///    the destination block.
549
  /// 3. If this block ends with a conditional branch and it falls through to a
550
  ///    successor block, it sets TBB to be the branch destination block and a
551
  ///    list of operands that evaluate the condition. These operands can be
552
  ///    passed to other TargetInstrInfo methods to create new branches.
553
  /// 4. If this block ends with a conditional branch followed by an
554
  ///    unconditional branch, it returns the 'true' destination in TBB, the
555
  ///    'false' destination in FBB, and a list of operands that evaluate the
556
  ///    condition.  These operands can be passed to other TargetInstrInfo
557
  ///    methods to create new branches.
558
  ///
559
  /// Note that removeBranch and insertBranch must be implemented to support
560
  /// cases where this method returns success.
561
  ///
562
  /// If AllowModify is true, then this routine is allowed to modify the basic
563
  /// block (e.g. delete instructions after the unconditional branch).
564
  ///
565
  /// The CFG information in MBB.Predecessors and MBB.Successors must be valid
566
  /// before calling this function.
567
  virtual bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
568
                             MachineBasicBlock *&FBB,
569
                             SmallVectorImpl<MachineOperand> &Cond,
570
0
                             bool AllowModify = false) const {
571
0
    return true;
572
0
  }
573
574
  /// Represents a predicate at the MachineFunction level.  The control flow a
575
  /// MachineBranchPredicate represents is:
576
  ///
577
  ///  Reg = LHS `Predicate` RHS         == ConditionDef
578
  ///  if Reg then goto TrueDest else goto FalseDest
579
  ///
580
  struct MachineBranchPredicate {
581
    enum ComparePredicate {
582
      PRED_EQ,     // True if two values are equal
583
      PRED_NE,     // True if two values are not equal
584
      PRED_INVALID // Sentinel value
585
    };
586
587
    ComparePredicate Predicate = PRED_INVALID;
588
    MachineOperand LHS = MachineOperand::CreateImm(0);
589
    MachineOperand RHS = MachineOperand::CreateImm(0);
590
    MachineBasicBlock *TrueDest = nullptr;
591
    MachineBasicBlock *FalseDest = nullptr;
592
    MachineInstr *ConditionDef = nullptr;
593
594
    /// SingleUseCondition is true if ConditionDef is dead except for the
595
    /// branch(es) at the end of the basic block.
596
    ///
597
    bool SingleUseCondition = false;
598
599
61
    explicit MachineBranchPredicate() = default;
600
  };
601
602
  /// Analyze the branching code at the end of MBB and parse it into the
603
  /// MachineBranchPredicate structure if possible.  Returns false on success
604
  /// and true on failure.
605
  ///
606
  /// If AllowModify is true, then this routine is allowed to modify the basic
607
  /// block (e.g. delete instructions after the unconditional branch).
608
  ///
609
  virtual bool analyzeBranchPredicate(MachineBasicBlock &MBB,
610
                                      MachineBranchPredicate &MBP,
611
0
                                      bool AllowModify = false) const {
612
0
    return true;
613
0
  }
614
615
  /// Remove the branching code at the end of the specific MBB.
616
  /// This is only invoked in cases where AnalyzeBranch returns success. It
617
  /// returns the number of instructions that were removed.
618
  /// If \p BytesRemoved is non-null, report the change in code size from the
619
  /// removed instructions.
620
  virtual unsigned removeBranch(MachineBasicBlock &MBB,
621
0
                                int *BytesRemoved = nullptr) const {
622
0
    llvm_unreachable("Target didn't implement TargetInstrInfo::removeBranch!");
623
0
  }
624
625
  /// Insert branch code into the end of the specified MachineBasicBlock. The
626
  /// operands to this method are the same as those returned by AnalyzeBranch.
627
  /// This is only invoked in cases where AnalyzeBranch returns success. It
628
  /// returns the number of instructions inserted. If \p BytesAdded is non-null,
629
  /// report the change in code size from the added instructions.
630
  ///
631
  /// It is also invoked by tail merging to add unconditional branches in
632
  /// cases where AnalyzeBranch doesn't apply because there was no original
633
  /// branch to analyze.  At least this much must be implemented, else tail
634
  /// merging needs to be disabled.
635
  ///
636
  /// The CFG information in MBB.Predecessors and MBB.Successors must be valid
637
  /// before calling this function.
638
  virtual unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
639
                                MachineBasicBlock *FBB,
640
                                ArrayRef<MachineOperand> Cond,
641
                                const DebugLoc &DL,
642
0
                                int *BytesAdded = nullptr) const {
643
0
    llvm_unreachable("Target didn't implement TargetInstrInfo::insertBranch!");
644
0
  }
645
646
  unsigned insertUnconditionalBranch(MachineBasicBlock &MBB,
647
                                     MachineBasicBlock *DestBB,
648
                                     const DebugLoc &DL,
649
3
                                     int *BytesAdded = nullptr) const {
650
3
    return insertBranch(MBB, DestBB, nullptr, ArrayRef<MachineOperand>(), DL,
651
3
                        BytesAdded);
652
3
  }
653
654
  /// Analyze the loop code, return true if it cannot be understoo. Upon
655
  /// success, this function returns false and returns information about the
656
  /// induction variable and compare instruction used at the end.
657
  virtual bool analyzeLoop(MachineLoop &L, MachineInstr *&IndVarInst,
658
0
                           MachineInstr *&CmpInst) const {
659
0
    return true;
660
0
  }
661
662
  /// Generate code to reduce the loop iteration by one and check if the loop
663
  /// is finished.  Return the value/register of the new loop count.  We need
664
  /// this function when peeling off one or more iterations of a loop. This
665
  /// function assumes the nth iteration is peeled first.
666
  virtual unsigned reduceLoopCount(MachineBasicBlock &MBB, MachineInstr *IndVar,
667
                                   MachineInstr &Cmp,
668
                                   SmallVectorImpl<MachineOperand> &Cond,
669
                                   SmallVectorImpl<MachineInstr *> &PrevInsts,
670
0
                                   unsigned Iter, unsigned MaxIter) const {
671
0
    llvm_unreachable("Target didn't implement ReduceLoopCount");
672
0
  }
673
674
  /// Delete the instruction OldInst and everything after it, replacing it with
675
  /// an unconditional branch to NewDest. This is used by the tail merging pass.
676
  virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
677
                                       MachineBasicBlock *NewDest) const;
678
679
  /// Return true if it's legal to split the given basic
680
  /// block at the specified instruction (i.e. instruction would be the start
681
  /// of a new basic block).
682
  virtual bool isLegalToSplitMBBAt(MachineBasicBlock &MBB,
683
62.8k
                                   MachineBasicBlock::iterator MBBI) const {
684
62.8k
    return true;
685
62.8k
  }
686
687
  /// Return true if it's profitable to predicate
688
  /// instructions with accumulated instruction latency of "NumCycles"
689
  /// of the specified basic block, where the probability of the instructions
690
  /// being executed is given by Probability, and Confidence is a measure
691
  /// of our confidence that it will be properly predicted.
692
  virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
693
                                   unsigned ExtraPredCycles,
694
0
                                   BranchProbability Probability) const {
695
0
    return false;
696
0
  }
697
698
  /// Second variant of isProfitableToIfCvt. This one
699
  /// checks for the case where two basic blocks from true and false path
700
  /// of a if-then-else (diamond) are predicated on mutally exclusive
701
  /// predicates, where the probability of the true path being taken is given
702
  /// by Probability, and Confidence is a measure of our confidence that it
703
  /// will be properly predicted.
704
  virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumTCycles,
705
                                   unsigned ExtraTCycles,
706
                                   MachineBasicBlock &FMBB, unsigned NumFCycles,
707
                                   unsigned ExtraFCycles,
708
0
                                   BranchProbability Probability) const {
709
0
    return false;
710
0
  }
711
712
  /// Return true if it's profitable for if-converter to duplicate instructions
713
  /// of specified accumulated instruction latencies in the specified MBB to
714
  /// enable if-conversion.
715
  /// The probability of the instructions being executed is given by
716
  /// Probability, and Confidence is a measure of our confidence that it
717
  /// will be properly predicted.
718
  virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
719
                                         unsigned NumCycles,
720
0
                                         BranchProbability Probability) const {
721
0
    return false;
722
0
  }
723
724
  /// Return true if it's profitable to unpredicate
725
  /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually
726
  /// exclusive predicates.
727
  /// e.g.
728
  ///   subeq  r0, r1, #1
729
  ///   addne  r0, r1, #1
730
  /// =>
731
  ///   sub    r0, r1, #1
732
  ///   addne  r0, r1, #1
733
  ///
734
  /// This may be profitable is conditional instructions are always executed.
735
  virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
736
18
                                         MachineBasicBlock &FMBB) const {
737
18
    return false;
738
18
  }
739
740
  /// Return true if it is possible to insert a select
741
  /// instruction that chooses between TrueReg and FalseReg based on the
742
  /// condition code in Cond.
743
  ///
744
  /// When successful, also return the latency in cycles from TrueReg,
745
  /// FalseReg, and Cond to the destination register. In most cases, a select
746
  /// instruction will be 1 cycle, so CondCycles = TrueCycles = FalseCycles = 1
747
  ///
748
  /// Some x86 implementations have 2-cycle cmov instructions.
749
  ///
750
  /// @param MBB         Block where select instruction would be inserted.
751
  /// @param Cond        Condition returned by AnalyzeBranch.
752
  /// @param TrueReg     Virtual register to select when Cond is true.
753
  /// @param FalseReg    Virtual register to select when Cond is false.
754
  /// @param CondCycles  Latency from Cond+Branch to select output.
755
  /// @param TrueCycles  Latency from TrueReg to select output.
756
  /// @param FalseCycles Latency from FalseReg to select output.
757
  virtual bool canInsertSelect(const MachineBasicBlock &MBB,
758
                               ArrayRef<MachineOperand> Cond, unsigned TrueReg,
759
                               unsigned FalseReg, int &CondCycles,
760
0
                               int &TrueCycles, int &FalseCycles) const {
761
0
    return false;
762
0
  }
763
764
  /// Insert a select instruction into MBB before I that will copy TrueReg to
765
  /// DstReg when Cond is true, and FalseReg to DstReg when Cond is false.
766
  ///
767
  /// This function can only be called after canInsertSelect() returned true.
768
  /// The condition in Cond comes from AnalyzeBranch, and it can be assumed
769
  /// that the same flags or registers required by Cond are available at the
770
  /// insertion point.
771
  ///
772
  /// @param MBB      Block where select instruction should be inserted.
773
  /// @param I        Insertion point.
774
  /// @param DL       Source location for debugging.
775
  /// @param DstReg   Virtual register to be defined by select instruction.
776
  /// @param Cond     Condition as computed by AnalyzeBranch.
777
  /// @param TrueReg  Virtual register to copy when Cond is true.
778
  /// @param FalseReg Virtual register to copy when Cons is false.
779
  virtual void insertSelect(MachineBasicBlock &MBB,
780
                            MachineBasicBlock::iterator I, const DebugLoc &DL,
781
                            unsigned DstReg, ArrayRef<MachineOperand> Cond,
782
0
                            unsigned TrueReg, unsigned FalseReg) const {
783
0
    llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!");
784
0
  }
785
786
  /// Analyze the given select instruction, returning true if
787
  /// it cannot be understood. It is assumed that MI->isSelect() is true.
788
  ///
789
  /// When successful, return the controlling condition and the operands that
790
  /// determine the true and false result values.
791
  ///
792
  ///   Result = SELECT Cond, TrueOp, FalseOp
793
  ///
794
  /// Some targets can optimize select instructions, for example by predicating
795
  /// the instruction defining one of the operands. Such targets should set
796
  /// Optimizable.
797
  ///
798
  /// @param         MI Select instruction to analyze.
799
  /// @param Cond    Condition controlling the select.
800
  /// @param TrueOp  Operand number of the value selected when Cond is true.
801
  /// @param FalseOp Operand number of the value selected when Cond is false.
802
  /// @param Optimizable Returned as true if MI is optimizable.
803
  /// @returns False on success.
804
  virtual bool analyzeSelect(const MachineInstr &MI,
805
                             SmallVectorImpl<MachineOperand> &Cond,
806
                             unsigned &TrueOp, unsigned &FalseOp,
807
397
                             bool &Optimizable) const {
808
397
    assert(MI.getDesc().isSelect() && "MI must be a select instruction");
809
397
    return true;
810
397
  }
811
812
  /// Given a select instruction that was understood by
813
  /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by
814
  /// merging it with one of its operands. Returns NULL on failure.
815
  ///
816
  /// When successful, returns the new select instruction. The client is
817
  /// responsible for deleting MI.
818
  ///
819
  /// If both sides of the select can be optimized, PreferFalse is used to pick
820
  /// a side.
821
  ///
822
  /// @param MI          Optimizable select instruction.
823
  /// @param NewMIs     Set that record all MIs in the basic block up to \p
824
  /// MI. Has to be updated with any newly created MI or deleted ones.
825
  /// @param PreferFalse Try to optimize FalseOp instead of TrueOp.
826
  /// @returns Optimized instruction or NULL.
827
  virtual MachineInstr *optimizeSelect(MachineInstr &MI,
828
                                       SmallPtrSetImpl<MachineInstr *> &NewMIs,
829
0
                                       bool PreferFalse = false) const {
830
0
    // This function must be implemented if Optimizable is ever set.
831
0
    llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!");
832
0
  }
833
834
  /// Emit instructions to copy a pair of physical registers.
835
  ///
836
  /// This function should support copies within any legal register class as
837
  /// well as any cross-class copies created during instruction selection.
838
  ///
839
  /// The source and destination registers may overlap, which may require a
840
  /// careful implementation when multiple copy instructions are required for
841
  /// large registers. See for example the ARM target.
842
  virtual void copyPhysReg(MachineBasicBlock &MBB,
843
                           MachineBasicBlock::iterator MI, const DebugLoc &DL,
844
                           unsigned DestReg, unsigned SrcReg,
845
0
                           bool KillSrc) const {
846
0
    llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!");
847
0
  }
848
849
  /// If the specific machine instruction is a instruction that moves/copies
850
  /// value from one register to another register return true along with
851
  /// @Source machine operand and @Destination machine operand.
852
  virtual bool isCopyInstr(const MachineInstr &MI,
853
                           const MachineOperand *&SourceOpNum,
854
333
                           const MachineOperand *&Destination) const {
855
333
    return false;
856
333
  }
857
858
  /// Store the specified register of the given register class to the specified
859
  /// stack frame index. The store instruction is to be added to the given
860
  /// machine basic block before the specified machine instruction. If isKill
861
  /// is true, the register operand is the last use and must be marked kill.
862
  virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
863
                                   MachineBasicBlock::iterator MI,
864
                                   unsigned SrcReg, bool isKill, int FrameIndex,
865
                                   const TargetRegisterClass *RC,
866
0
                                   const TargetRegisterInfo *TRI) const {
867
0
    llvm_unreachable("Target didn't implement "
868
0
                     "TargetInstrInfo::storeRegToStackSlot!");
869
0
  }
870
871
  /// Load the specified register of the given register class from the specified
872
  /// stack frame index. The load instruction is to be added to the given
873
  /// machine basic block before the specified machine instruction.
874
  virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
875
                                    MachineBasicBlock::iterator MI,
876
                                    unsigned DestReg, int FrameIndex,
877
                                    const TargetRegisterClass *RC,
878
0
                                    const TargetRegisterInfo *TRI) const {
879
0
    llvm_unreachable("Target didn't implement "
880
0
                     "TargetInstrInfo::loadRegFromStackSlot!");
881
0
  }
882
883
  /// This function is called for all pseudo instructions
884
  /// that remain after register allocation. Many pseudo instructions are
885
  /// created to help register allocation. This is the place to convert them
886
  /// into real instructions. The target can edit MI in place, or it can insert
887
  /// new instructions and erase MI. The function should return true if
888
  /// anything was changed.
889
266k
  virtual bool expandPostRAPseudo(MachineInstr &MI) const { return false; }
890
891
  /// Check whether the target can fold a load that feeds a subreg operand
892
  /// (or a subreg operand that feeds a store).
893
  /// For example, X86 may want to return true if it can fold
894
  /// movl (%esp), %eax
895
  /// subb, %al, ...
896
  /// Into:
897
  /// subb (%esp), ...
898
  ///
899
  /// Ideally, we'd like the target implementation of foldMemoryOperand() to
900
  /// reject subregs - but since this behavior used to be enforced in the
901
  /// target-independent code, moving this responsibility to the targets
902
  /// has the potential of causing nasty silent breakage in out-of-tree targets.
903
32.5k
  virtual bool isSubregFoldable() const { return false; }
904
905
  /// Attempt to fold a load or store of the specified stack
906
  /// slot into the specified machine instruction for the specified operand(s).
907
  /// If this is possible, a new instruction is returned with the specified
908
  /// operand folded, otherwise NULL is returned.
909
  /// The new instruction is inserted before MI, and the client is responsible
910
  /// for removing the old instruction.
911
  MachineInstr *foldMemoryOperand(MachineInstr &MI, ArrayRef<unsigned> Ops,
912
                                  int FI,
913
                                  LiveIntervals *LIS = nullptr) const;
914
915
  /// Same as the previous version except it allows folding of any load and
916
  /// store from / to any address, not just from a specific stack slot.
917
  MachineInstr *foldMemoryOperand(MachineInstr &MI, ArrayRef<unsigned> Ops,
918
                                  MachineInstr &LoadMI,
919
                                  LiveIntervals *LIS = nullptr) const;
920
921
  /// Return true when there is potentially a faster code sequence
922
  /// for an instruction chain ending in \p Root. All potential patterns are
923
  /// returned in the \p Pattern vector. Pattern should be sorted in priority
924
  /// order since the pattern evaluator stops checking as soon as it finds a
925
  /// faster sequence.
926
  /// \param Root - Instruction that could be combined with one of its operands
927
  /// \param Patterns - Vector of possible combination patterns
928
  virtual bool getMachineCombinerPatterns(
929
      MachineInstr &Root,
930
      SmallVectorImpl<MachineCombinerPattern> &Patterns) const;
931
932
  /// Return true when a code sequence can improve throughput. It
933
  /// should be called only for instructions in loops.
934
  /// \param Pattern - combiner pattern
935
  virtual bool isThroughputPattern(MachineCombinerPattern Pattern) const;
936
937
  /// Return true if the input \P Inst is part of a chain of dependent ops
938
  /// that are suitable for reassociation, otherwise return false.
939
  /// If the instruction's operands must be commuted to have a previous
940
  /// instruction of the same type define the first source operand, \P Commuted
941
  /// will be set to true.
942
  bool isReassociationCandidate(const MachineInstr &Inst, bool &Commuted) const;
943
944
  /// Return true when \P Inst is both associative and commutative.
945
0
  virtual bool isAssociativeAndCommutative(const MachineInstr &Inst) const {
946
0
    return false;
947
0
  }
948
949
  /// Return true when \P Inst has reassociable operands in the same \P MBB.
950
  virtual bool hasReassociableOperands(const MachineInstr &Inst,
951
                                       const MachineBasicBlock *MBB) const;
952
953
  /// Return true when \P Inst has reassociable sibling.
954
  bool hasReassociableSibling(const MachineInstr &Inst, bool &Commuted) const;
955
956
  /// When getMachineCombinerPatterns() finds patterns, this function generates
957
  /// the instructions that could replace the original code sequence. The client
958
  /// has to decide whether the actual replacement is beneficial or not.
959
  /// \param Root - Instruction that could be combined with one of its operands
960
  /// \param Pattern - Combination pattern for Root
961
  /// \param InsInstrs - Vector of new instructions that implement P
962
  /// \param DelInstrs - Old instructions, including Root, that could be
963
  /// replaced by InsInstr
964
  /// \param InstIdxForVirtReg - map of virtual register to instruction in
965
  /// InsInstr that defines it
966
  virtual void genAlternativeCodeSequence(
967
      MachineInstr &Root, MachineCombinerPattern Pattern,
968
      SmallVectorImpl<MachineInstr *> &InsInstrs,
969
      SmallVectorImpl<MachineInstr *> &DelInstrs,
970
      DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const;
971
972
  /// Attempt to reassociate \P Root and \P Prev according to \P Pattern to
973
  /// reduce critical path length.
974
  void reassociateOps(MachineInstr &Root, MachineInstr &Prev,
975
                      MachineCombinerPattern Pattern,
976
                      SmallVectorImpl<MachineInstr *> &InsInstrs,
977
                      SmallVectorImpl<MachineInstr *> &DelInstrs,
978
                      DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const;
979
980
  /// This is an architecture-specific helper function of reassociateOps.
981
  /// Set special operand attributes for new instructions after reassociation.
982
  virtual void setSpecialOperandAttr(MachineInstr &OldMI1, MachineInstr &OldMI2,
983
                                     MachineInstr &NewMI1,
984
336
                                     MachineInstr &NewMI2) const {}
985
986
  /// Return true when a target supports MachineCombiner.
987
0
  virtual bool useMachineCombiner() const { return false; }
988
989
  /// Return true if the given SDNode can be copied during scheduling
990
  /// even if it has glue.
991
43
  virtual bool canCopyGluedNodeDuringSchedule(SDNode *N) const { return false; }
992
993
protected:
994
  /// Target-dependent implementation for foldMemoryOperand.
995
  /// Target-independent code in foldMemoryOperand will
996
  /// take care of adding a MachineMemOperand to the newly created instruction.
997
  /// The instruction and any auxiliary instructions necessary will be inserted
998
  /// at InsertPt.
999
  virtual MachineInstr *
1000
  foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI,
1001
                        ArrayRef<unsigned> Ops,
1002
                        MachineBasicBlock::iterator InsertPt, int FrameIndex,
1003
30.6k
                        LiveIntervals *LIS = nullptr) const {
1004
30.6k
    return nullptr;
1005
30.6k
  }
1006
1007
  /// Target-dependent implementation for foldMemoryOperand.
1008
  /// Target-independent code in foldMemoryOperand will
1009
  /// take care of adding a MachineMemOperand to the newly created instruction.
1010
  /// The instruction and any auxiliary instructions necessary will be inserted
1011
  /// at InsertPt.
1012
  virtual MachineInstr *foldMemoryOperandImpl(
1013
      MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
1014
      MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI,
1015
851
      LiveIntervals *LIS = nullptr) const {
1016
851
    return nullptr;
1017
851
  }
1018
1019
  /// Target-dependent implementation of getRegSequenceInputs.
1020
  ///
1021
  /// \returns true if it is possible to build the equivalent
1022
  /// REG_SEQUENCE inputs with the pair \p MI, \p DefIdx. False otherwise.
1023
  ///
1024
  /// \pre MI.isRegSequenceLike().
1025
  ///
1026
  /// \see TargetInstrInfo::getRegSequenceInputs.
1027
  virtual bool getRegSequenceLikeInputs(
1028
      const MachineInstr &MI, unsigned DefIdx,
1029
0
      SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1030
0
    return false;
1031
0
  }
1032
1033
  /// Target-dependent implementation of getExtractSubregInputs.
1034
  ///
1035
  /// \returns true if it is possible to build the equivalent
1036
  /// EXTRACT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise.
1037
  ///
1038
  /// \pre MI.isExtractSubregLike().
1039
  ///
1040
  /// \see TargetInstrInfo::getExtractSubregInputs.
1041
  virtual bool getExtractSubregLikeInputs(const MachineInstr &MI,
1042
                                          unsigned DefIdx,
1043
0
                                          RegSubRegPairAndIdx &InputReg) const {
1044
0
    return false;
1045
0
  }
1046
1047
  /// Target-dependent implementation of getInsertSubregInputs.
1048
  ///
1049
  /// \returns true if it is possible to build the equivalent
1050
  /// INSERT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise.
1051
  ///
1052
  /// \pre MI.isInsertSubregLike().
1053
  ///
1054
  /// \see TargetInstrInfo::getInsertSubregInputs.
1055
  virtual bool
1056
  getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
1057
                            RegSubRegPair &BaseReg,
1058
0
                            RegSubRegPairAndIdx &InsertedReg) const {
1059
0
    return false;
1060
0
  }
1061
1062
public:
1063
  /// getAddressSpaceForPseudoSourceKind - Given the kind of memory
1064
  /// (e.g. stack) the target returns the corresponding address space.
1065
  virtual unsigned
1066
3.37M
  getAddressSpaceForPseudoSourceKind(PseudoSourceValue::PSVKind Kind) const {
1067
3.37M
    return 0;
1068
3.37M
  }
1069
1070
  /// unfoldMemoryOperand - Separate a single instruction which folded a load or
1071
  /// a store or a load and a store into two or more instruction. If this is
1072
  /// possible, returns true as well as the new instructions by reference.
1073
  virtual bool
1074
  unfoldMemoryOperand(MachineFunction &MF, MachineInstr &MI, unsigned Reg,
1075
                      bool UnfoldLoad, bool UnfoldStore,
1076
0
                      SmallVectorImpl<MachineInstr *> &NewMIs) const {
1077
0
    return false;
1078
0
  }
1079
1080
  virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
1081
0
                                   SmallVectorImpl<SDNode *> &NewNodes) const {
1082
0
    return false;
1083
0
  }
1084
1085
  /// Returns the opcode of the would be new
1086
  /// instruction after load / store are unfolded from an instruction of the
1087
  /// specified opcode. It returns zero if the specified unfolding is not
1088
  /// possible. If LoadRegIndex is non-null, it is filled in with the operand
1089
  /// index of the operand which will hold the register holding the loaded
1090
  /// value.
1091
  virtual unsigned
1092
  getOpcodeAfterMemoryUnfold(unsigned Opc, bool UnfoldLoad, bool UnfoldStore,
1093
8.00k
                             unsigned *LoadRegIndex = nullptr) const {
1094
8.00k
    return 0;
1095
8.00k
  }
1096
1097
  /// This is used by the pre-regalloc scheduler to determine if two loads are
1098
  /// loading from the same base address. It should only return true if the base
1099
  /// pointers are the same and the only differences between the two addresses
1100
  /// are the offset. It also returns the offsets by reference.
1101
  virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
1102
                                       int64_t &Offset1,
1103
5.14M
                                       int64_t &Offset2) const {
1104
5.14M
    return false;
1105
5.14M
  }
1106
1107
  /// This is a used by the pre-regalloc scheduler to determine (in conjunction
1108
  /// with areLoadsFromSameBasePtr) if two loads should be scheduled together.
1109
  /// On some targets if two loads are loading from
1110
  /// addresses in the same cache line, it's better if they are scheduled
1111
  /// together. This function takes two integers that represent the load offsets
1112
  /// from the common base address. It returns true if it decides it's desirable
1113
  /// to schedule the two loads together. "NumLoads" is the number of loads that
1114
  /// have already been scheduled after Load1.
1115
  virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
1116
                                       int64_t Offset1, int64_t Offset2,
1117
0
                                       unsigned NumLoads) const {
1118
0
    return false;
1119
0
  }
1120
1121
  /// Get the base register and byte offset of an instruction that reads/writes
1122
  /// memory.
1123
  virtual bool getMemOpBaseRegImmOfs(MachineInstr &MemOp, unsigned &BaseReg,
1124
                                     int64_t &Offset,
1125
0
                                     const TargetRegisterInfo *TRI) const {
1126
0
    return false;
1127
0
  }
1128
1129
  /// Return true if the instruction contains a base register and offset. If
1130
  /// true, the function also sets the operand position in the instruction
1131
  /// for the base register and offset.
1132
  virtual bool getBaseAndOffsetPosition(const MachineInstr &MI,
1133
                                        unsigned &BasePos,
1134
0
                                        unsigned &OffsetPos) const {
1135
0
    return false;
1136
0
  }
1137
1138
  /// If the instruction is an increment of a constant value, return the amount.
1139
0
  virtual bool getIncrementValue(const MachineInstr &MI, int &Value) const {
1140
0
    return false;
1141
0
  }
1142
1143
  /// Returns true if the two given memory operations should be scheduled
1144
  /// adjacent. Note that you have to add:
1145
  ///   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
1146
  /// or
1147
  ///   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
1148
  /// to TargetPassConfig::createMachineScheduler() to have an effect.
1149
  virtual bool shouldClusterMemOps(MachineInstr &FirstLdSt, unsigned BaseReg1,
1150
                                   MachineInstr &SecondLdSt, unsigned BaseReg2,
1151
0
                                   unsigned NumLoads) const {
1152
0
    llvm_unreachable("target did not implement shouldClusterMemOps()");
1153
0
  }
1154
1155
  /// Reverses the branch condition of the specified condition list,
1156
  /// returning false on success and true if it cannot be reversed.
1157
  virtual bool
1158
19
  reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
1159
19
    return true;
1160
19
  }
1161
1162
  /// Insert a noop into the instruction stream at the specified point.
1163
  virtual void insertNoop(MachineBasicBlock &MBB,
1164
                          MachineBasicBlock::iterator MI) const;
1165
1166
  /// Return the noop instruction to use for a noop.
1167
  virtual void getNoop(MCInst &NopInst) const;
1168
1169
  /// Return true for post-incremented instructions.
1170
0
  virtual bool isPostIncrement(const MachineInstr &MI) const { return false; }
1171
1172
  /// Returns true if the instruction is already predicated.
1173
2.27M
  virtual bool isPredicated(const MachineInstr &MI) const { return false; }
1174
1175
  /// Returns true if the instruction is a
1176
  /// terminator instruction that has not been predicated.
1177
  virtual bool isUnpredicatedTerminator(const MachineInstr &MI) const;
1178
1179
  /// Returns true if MI is an unconditional tail call.
1180
23.9k
  virtual bool isUnconditionalTailCall(const MachineInstr &MI) const {
1181
23.9k
    return false;
1182
23.9k
  }
1183
1184
  /// Returns true if the tail call can be made conditional on BranchCond.
1185
  virtual bool canMakeTailCallConditional(SmallVectorImpl<MachineOperand> &Cond,
1186
0
                                          const MachineInstr &TailCall) const {
1187
0
    return false;
1188
0
  }
1189
1190
  /// Replace the conditional branch in MBB with a conditional tail call.
1191
  virtual void replaceBranchWithTailCall(MachineBasicBlock &MBB,
1192
                                         SmallVectorImpl<MachineOperand> &Cond,
1193
0
                                         const MachineInstr &TailCall) const {
1194
0
    llvm_unreachable("Target didn't implement replaceBranchWithTailCall!");
1195
0
  }
1196
1197
  /// Convert the instruction into a predicated instruction.
1198
  /// It returns true if the operation was successful.
1199
  virtual bool PredicateInstruction(MachineInstr &MI,
1200
                                    ArrayRef<MachineOperand> Pred) const;
1201
1202
  /// Returns true if the first specified predicate
1203
  /// subsumes the second, e.g. GE subsumes GT.
1204
  virtual bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1205
0
                                 ArrayRef<MachineOperand> Pred2) const {
1206
0
    return false;
1207
0
  }
1208
1209
  /// If the specified instruction defines any predicate
1210
  /// or condition code register(s) used for predication, returns true as well
1211
  /// as the definition predicate(s) by reference.
1212
  virtual bool DefinesPredicate(MachineInstr &MI,
1213
11.1k
                                std::vector<MachineOperand> &Pred) const {
1214
11.1k
    return false;
1215
11.1k
  }
1216
1217
  /// Return true if the specified instruction can be predicated.
1218
  /// By default, this returns true for every instruction with a
1219
  /// PredicateOperand.
1220
914
  virtual bool isPredicable(const MachineInstr &MI) const {
1221
914
    return MI.getDesc().isPredicable();
1222
914
  }
1223
1224
  /// Return true if it's safe to move a machine
1225
  /// instruction that defines the specified register class.
1226
12.8M
  virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
1227
12.8M
    return true;
1228
12.8M
  }
1229
1230
  /// Test if the given instruction should be considered a scheduling boundary.
1231
  /// This primarily includes labels and terminators.
1232
  virtual bool isSchedulingBoundary(const MachineInstr &MI,
1233
                                    const MachineBasicBlock *MBB,
1234
                                    const MachineFunction &MF) const;
1235
1236
  /// Measure the specified inline asm to determine an approximation of its
1237
  /// length.
1238
  virtual unsigned getInlineAsmLength(const char *Str,
1239
                                      const MCAsmInfo &MAI) const;
1240
1241
  /// Allocate and return a hazard recognizer to use for this target when
1242
  /// scheduling the machine instructions before register allocation.
1243
  virtual ScheduleHazardRecognizer *
1244
  CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
1245
                               const ScheduleDAG *DAG) const;
1246
1247
  /// Allocate and return a hazard recognizer to use for this target when
1248
  /// scheduling the machine instructions before register allocation.
1249
  virtual ScheduleHazardRecognizer *
1250
  CreateTargetMIHazardRecognizer(const InstrItineraryData *,
1251
                                 const ScheduleDAG *DAG) const;
1252
1253
  /// Allocate and return a hazard recognizer to use for this target when
1254
  /// scheduling the machine instructions after register allocation.
1255
  virtual ScheduleHazardRecognizer *
1256
  CreateTargetPostRAHazardRecognizer(const InstrItineraryData *,
1257
                                     const ScheduleDAG *DAG) const;
1258
1259
  /// Allocate and return a hazard recognizer to use for by non-scheduling
1260
  /// passes.
1261
  virtual ScheduleHazardRecognizer *
1262
0
  CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const {
1263
0
    return nullptr;
1264
0
  }
1265
1266
  /// Provide a global flag for disabling the PreRA hazard recognizer that
1267
  /// targets may choose to honor.
1268
  bool usePreRAHazardRecognizer() const;
1269
1270
  /// For a comparison instruction, return the source registers
1271
  /// in SrcReg and SrcReg2 if having two register operands, and the value it
1272
  /// compares against in CmpValue. Return true if the comparison instruction
1273
  /// can be analyzed.
1274
  virtual bool analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
1275
4.26k
                              unsigned &SrcReg2, int &Mask, int &Value) const {
1276
4.26k
    return false;
1277
4.26k
  }
1278
1279
  /// See if the comparison instruction can be converted
1280
  /// into something more efficient. E.g., on ARM most instructions can set the
1281
  /// flags register, obviating the need for a separate CMP.
1282
  virtual bool optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
1283
                                    unsigned SrcReg2, int Mask, int Value,
1284
905
                                    const MachineRegisterInfo *MRI) const {
1285
905
    return false;
1286
905
  }
1287
221k
  virtual bool optimizeCondBranch(MachineInstr &MI) const { return false; }
1288
1289
  /// Try to remove the load by folding it to a register operand at the use.
1290
  /// We fold the load instructions if and only if the
1291
  /// def and use are in the same BB. We only look at one load and see
1292
  /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
1293
  /// defined by the load we are trying to fold. DefMI returns the machine
1294
  /// instruction that defines FoldAsLoadDefReg, and the function returns
1295
  /// the machine instruction generated due to folding.
1296
  virtual MachineInstr *optimizeLoadInstr(MachineInstr &MI,
1297
                                          const MachineRegisterInfo *MRI,
1298
                                          unsigned &FoldAsLoadDefReg,
1299
27.0k
                                          MachineInstr *&DefMI) const {
1300
27.0k
    return nullptr;
1301
27.0k
  }
1302
1303
  /// 'Reg' is known to be defined by a move immediate instruction,
1304
  /// try to fold the immediate into the use instruction.
1305
  /// If MRI->hasOneNonDBGUse(Reg) is true, and this function returns true,
1306
  /// then the caller may assume that DefMI has been erased from its parent
1307
  /// block. The caller may assume that it will not be erased by this
1308
  /// function otherwise.
1309
  virtual bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
1310
810k
                             unsigned Reg, MachineRegisterInfo *MRI) const {
1311
810k
    return false;
1312
810k
  }
1313
1314
  /// Return the number of u-operations the given machine
1315
  /// instruction will be decoded to on the target cpu. The itinerary's
1316
  /// IssueWidth is the number of microops that can be dispatched each
1317
  /// cycle. An instruction with zero microops takes no dispatch resources.
1318
  virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
1319
                                  const MachineInstr &MI) const;
1320
1321
  /// Return true for pseudo instructions that don't consume any
1322
  /// machine resources in their current form. These are common cases that the
1323
  /// scheduler should consider free, rather than conservatively handling them
1324
  /// as instructions with no itinerary.
1325
1.14M
  bool isZeroCost(unsigned Opcode) const {
1326
1.14M
    return Opcode <= TargetOpcode::COPY;
1327
1.14M
  }
1328
1329
  virtual int getOperandLatency(const InstrItineraryData *ItinData,
1330
                                SDNode *DefNode, unsigned DefIdx,
1331
                                SDNode *UseNode, unsigned UseIdx) const;
1332
1333
  /// Compute and return the use operand latency of a given pair of def and use.
1334
  /// In most cases, the static scheduling itinerary was enough to determine the
1335
  /// operand latency. But it may not be possible for instructions with variable
1336
  /// number of defs / uses.
1337
  ///
1338
  /// This is a raw interface to the itinerary that may be directly overridden
1339
  /// by a target. Use computeOperandLatency to get the best estimate of
1340
  /// latency.
1341
  virtual int getOperandLatency(const InstrItineraryData *ItinData,
1342
                                const MachineInstr &DefMI, unsigned DefIdx,
1343
                                const MachineInstr &UseMI,
1344
                                unsigned UseIdx) const;
1345
1346
  /// Compute the instruction latency of a given instruction.
1347
  /// If the instruction has higher cost when predicated, it's returned via
1348
  /// PredCost.
1349
  virtual unsigned getInstrLatency(const InstrItineraryData *ItinData,
1350
                                   const MachineInstr &MI,
1351
                                   unsigned *PredCost = nullptr) const;
1352
1353
  virtual unsigned getPredicationCost(const MachineInstr &MI) const;
1354
1355
  virtual int getInstrLatency(const InstrItineraryData *ItinData,
1356
                              SDNode *Node) const;
1357
1358
  /// Return the default expected latency for a def based on its opcode.
1359
  unsigned defaultDefLatency(const MCSchedModel &SchedModel,
1360
                             const MachineInstr &DefMI) const;
1361
1362
  int computeDefOperandLatency(const InstrItineraryData *ItinData,
1363
                               const MachineInstr &DefMI) const;
1364
1365
  /// Return true if this opcode has high latency to its result.
1366
2.15M
  virtual bool isHighLatencyDef(int opc) const { return false; }
1367
1368
  /// Compute operand latency between a def of 'Reg'
1369
  /// and a use in the current loop. Return true if the target considered
1370
  /// it 'high'. This is used by optimization passes such as machine LICM to
1371
  /// determine whether it makes sense to hoist an instruction out even in a
1372
  /// high register pressure situation.
1373
  virtual bool hasHighOperandLatency(const TargetSchedModel &SchedModel,
1374
                                     const MachineRegisterInfo *MRI,
1375
                                     const MachineInstr &DefMI, unsigned DefIdx,
1376
                                     const MachineInstr &UseMI,
1377
94.1k
                                     unsigned UseIdx) const {
1378
94.1k
    return false;
1379
94.1k
  }
1380
1381
  /// Compute operand latency of a def of 'Reg'. Return true
1382
  /// if the target considered it 'low'.
1383
  virtual bool hasLowDefLatency(const TargetSchedModel &SchedModel,
1384
                                const MachineInstr &DefMI,
1385
                                unsigned DefIdx) const;
1386
1387
  /// Perform target-specific instruction verification.
1388
  virtual bool verifyInstruction(const MachineInstr &MI,
1389
5.92M
                                 StringRef &ErrInfo) const {
1390
5.92M
    return true;
1391
5.92M
  }
1392
1393
  /// Return the current execution domain and bit mask of
1394
  /// possible domains for instruction.
1395
  ///
1396
  /// Some micro-architectures have multiple execution domains, and multiple
1397
  /// opcodes that perform the same operation in different domains.  For
1398
  /// example, the x86 architecture provides the por, orps, and orpd
1399
  /// instructions that all do the same thing.  There is a latency penalty if a
1400
  /// register is written in one domain and read in another.
1401
  ///
1402
  /// This function returns a pair (domain, mask) containing the execution
1403
  /// domain of MI, and a bit mask of possible domains.  The setExecutionDomain
1404
  /// function can be used to change the opcode to one of the domains in the
1405
  /// bit mask.  Instructions whose execution domain can't be changed should
1406
  /// return a 0 mask.
1407
  ///
1408
  /// The execution domain numbers don't have any special meaning except domain
1409
  /// 0 is used for instructions that are not associated with any interesting
1410
  /// execution domain.
1411
  ///
1412
  virtual std::pair<uint16_t, uint16_t>
1413
0
  getExecutionDomain(const MachineInstr &MI) const {
1414
0
    return std::make_pair(0, 0);
1415
0
  }
1416
1417
  /// Change the opcode of MI to execute in Domain.
1418
  ///
1419
  /// The bit (1 << Domain) must be set in the mask returned from
1420
  /// getExecutionDomain(MI).
1421
0
  virtual void setExecutionDomain(MachineInstr &MI, unsigned Domain) const {}
1422
1423
  /// Returns the preferred minimum clearance
1424
  /// before an instruction with an unwanted partial register update.
1425
  ///
1426
  /// Some instructions only write part of a register, and implicitly need to
1427
  /// read the other parts of the register.  This may cause unwanted stalls
1428
  /// preventing otherwise unrelated instructions from executing in parallel in
1429
  /// an out-of-order CPU.
1430
  ///
1431
  /// For example, the x86 instruction cvtsi2ss writes its result to bits
1432
  /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so
1433
  /// the instruction needs to wait for the old value of the register to become
1434
  /// available:
1435
  ///
1436
  ///   addps %xmm1, %xmm0
1437
  ///   movaps %xmm0, (%rax)
1438
  ///   cvtsi2ss %rbx, %xmm0
1439
  ///
1440
  /// In the code above, the cvtsi2ss instruction needs to wait for the addps
1441
  /// instruction before it can issue, even though the high bits of %xmm0
1442
  /// probably aren't needed.
1443
  ///
1444
  /// This hook returns the preferred clearance before MI, measured in
1445
  /// instructions.  Other defs of MI's operand OpNum are avoided in the last N
1446
  /// instructions before MI.  It should only return a positive value for
1447
  /// unwanted dependencies.  If the old bits of the defined register have
1448
  /// useful values, or if MI is determined to otherwise read the dependency,
1449
  /// the hook should return 0.
1450
  ///
1451
  /// The unwanted dependency may be handled by:
1452
  ///
1453
  /// 1. Allocating the same register for an MI def and use.  That makes the
1454
  ///    unwanted dependency identical to a required dependency.
1455
  ///
1456
  /// 2. Allocating a register for the def that has no defs in the previous N
1457
  ///    instructions.
1458
  ///
1459
  /// 3. Calling breakPartialRegDependency() with the same arguments.  This
1460
  ///    allows the target to insert a dependency breaking instruction.
1461
  ///
1462
  virtual unsigned
1463
  getPartialRegUpdateClearance(const MachineInstr &MI, unsigned OpNum,
1464
0
                               const TargetRegisterInfo *TRI) const {
1465
0
    // The default implementation returns 0 for no partial register dependency.
1466
0
    return 0;
1467
0
  }
1468
1469
  /// Return the minimum clearance before an instruction that reads an
1470
  /// unused register.
1471
  ///
1472
  /// For example, AVX instructions may copy part of a register operand into
1473
  /// the unused high bits of the destination register.
1474
  ///
1475
  /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
1476
  ///
1477
  /// In the code above, vcvtsi2sdq copies %xmm0[127:64] into %xmm14 creating a
1478
  /// false dependence on any previous write to %xmm0.
1479
  ///
1480
  /// This hook works similarly to getPartialRegUpdateClearance, except that it
1481
  /// does not take an operand index. Instead sets \p OpNum to the index of the
1482
  /// unused register.
1483
  virtual unsigned getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum,
1484
761k
                                        const TargetRegisterInfo *TRI) const {
1485
761k
    // The default implementation returns 0 for no undef register dependency.
1486
761k
    return 0;
1487
761k
  }
1488
1489
  /// Insert a dependency-breaking instruction
1490
  /// before MI to eliminate an unwanted dependency on OpNum.
1491
  ///
1492
  /// If it wasn't possible to avoid a def in the last N instructions before MI
1493
  /// (see getPartialRegUpdateClearance), this hook will be called to break the
1494
  /// unwanted dependency.
1495
  ///
1496
  /// On x86, an xorps instruction can be used as a dependency breaker:
1497
  ///
1498
  ///   addps %xmm1, %xmm0
1499
  ///   movaps %xmm0, (%rax)
1500
  ///   xorps %xmm0, %xmm0
1501
  ///   cvtsi2ss %rbx, %xmm0
1502
  ///
1503
  /// An <imp-kill> operand should be added to MI if an instruction was
1504
  /// inserted.  This ties the instructions together in the post-ra scheduler.
1505
  ///
1506
  virtual void breakPartialRegDependency(MachineInstr &MI, unsigned OpNum,
1507
0
                                         const TargetRegisterInfo *TRI) const {}
1508
1509
  /// Create machine specific model for scheduling.
1510
  virtual DFAPacketizer *
1511
0
  CreateTargetScheduleState(const TargetSubtargetInfo &) const {
1512
0
    return nullptr;
1513
0
  }
1514
1515
  /// Sometimes, it is possible for the target
1516
  /// to tell, even without aliasing information, that two MIs access different
1517
  /// memory addresses. This function returns true if two MIs access different
1518
  /// memory addresses and false otherwise.
1519
  ///
1520
  /// Assumes any physical registers used to compute addresses have the same
1521
  /// value for both instructions. (This is the most useful assumption for
1522
  /// post-RA scheduling.)
1523
  ///
1524
  /// See also MachineInstr::mayAlias, which is implemented on top of this
1525
  /// function.
1526
  virtual bool
1527
  areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb,
1528
1.31M
                                  AliasAnalysis *AA = nullptr) const {
1529
1.31M
    assert((MIa.mayLoad() || MIa.mayStore()) &&
1530
1.31M
           "MIa must load from or modify a memory location");
1531
1.31M
    assert((MIb.mayLoad() || MIb.mayStore()) &&
1532
1.31M
           "MIb must load from or modify a memory location");
1533
1.31M
    return false;
1534
1.31M
  }
1535
1536
  /// Return the value to use for the MachineCSE's LookAheadLimit,
1537
  /// which is a heuristic used for CSE'ing phys reg defs.
1538
431k
  virtual unsigned getMachineCSELookAheadLimit() const {
1539
431k
    // The default lookahead is small to prevent unprofitable quadratic
1540
431k
    // behavior.
1541
431k
    return 5;
1542
431k
  }
1543
1544
  /// Return an array that contains the ids of the target indices (used for the
1545
  /// TargetIndex machine operand) and their names.
1546
  ///
1547
  /// MIR Serialization is able to serialize only the target indices that are
1548
  /// defined by this method.
1549
  virtual ArrayRef<std::pair<int, const char *>>
1550
0
  getSerializableTargetIndices() const {
1551
0
    return None;
1552
0
  }
1553
1554
  /// Decompose the machine operand's target flags into two values - the direct
1555
  /// target flag value and any of bit flags that are applied.
1556
  virtual std::pair<unsigned, unsigned>
1557
0
  decomposeMachineOperandsTargetFlags(unsigned /*TF*/) const {
1558
0
    return std::make_pair(0u, 0u);
1559
0
  }
1560
1561
  /// Return an array that contains the direct target flag values and their
1562
  /// names.
1563
  ///
1564
  /// MIR Serialization is able to serialize only the target flags that are
1565
  /// defined by this method.
1566
  virtual ArrayRef<std::pair<unsigned, const char *>>
1567
0
  getSerializableDirectMachineOperandTargetFlags() const {
1568
0
    return None;
1569
0
  }
1570
1571
  /// Return an array that contains the bitmask target flag values and their
1572
  /// names.
1573
  ///
1574
  /// MIR Serialization is able to serialize only the target flags that are
1575
  /// defined by this method.
1576
  virtual ArrayRef<std::pair<unsigned, const char *>>
1577
1
  getSerializableBitmaskMachineOperandTargetFlags() const {
1578
1
    return None;
1579
1
  }
1580
1581
  /// Return an array that contains the MMO target flag values and their
1582
  /// names.
1583
  ///
1584
  /// MIR Serialization is able to serialize only the MMO target flags that are
1585
  /// defined by this method.
1586
  virtual ArrayRef<std::pair<MachineMemOperand::Flags, const char *>>
1587
0
  getSerializableMachineMemOperandTargetFlags() const {
1588
0
    return None;
1589
0
  }
1590
1591
  /// Determines whether \p Inst is a tail call instruction. Override this
1592
  /// method on targets that do not properly set MCID::Return and MCID::Call on
1593
  /// tail call instructions."
1594
354
  virtual bool isTailCall(const MachineInstr &Inst) const {
1595
354
    return Inst.isReturn() && 
Inst.isCall()100
;
1596
354
  }
1597
1598
  /// True if the instruction is bound to the top of its basic block and no
1599
  /// other instructions shall be inserted before it. This can be implemented
1600
  /// to prevent register allocator to insert spills before such instructions.
1601
490k
  virtual bool isBasicBlockPrologue(const MachineInstr &MI) const {
1602
490k
    return false;
1603
490k
  }
1604
1605
  /// Returns a \p outliner::TargetCostInfo struct containing target-specific
1606
  /// information for a set of outlining candidates.
1607
  virtual outliner::TargetCostInfo getOutliningCandidateInfo(
1608
0
      std::vector<outliner::Candidate> &RepeatedSequenceLocs) const {
1609
0
    llvm_unreachable(
1610
0
        "Target didn't implement TargetInstrInfo::getOutliningCandidateInfo!");
1611
0
  }
1612
1613
  /// Returns how or if \p MI should be outlined.
1614
  virtual outliner::InstrType
1615
0
  getOutliningType(MachineBasicBlock::iterator &MIT, unsigned Flags) const {
1616
0
    llvm_unreachable(
1617
0
        "Target didn't implement TargetInstrInfo::getOutliningType!");
1618
0
  }
1619
1620
  /// Returns target-defined flags defining properties of the MBB for
1621
  /// the outliner.
1622
26
  virtual unsigned getMachineOutlinerMBBFlags(MachineBasicBlock &MBB) const {
1623
26
    return 0x0;
1624
26
  }
1625
1626
  /// Insert a custom frame for outlined functions.
1627
  virtual void buildOutlinedFrame(MachineBasicBlock &MBB,
1628
                                      MachineFunction &MF,
1629
0
                                    const outliner::TargetCostInfo &TCI) const {
1630
0
    llvm_unreachable(
1631
0
        "Target didn't implement TargetInstrInfo::buildOutlinedFrame!");
1632
0
  }
1633
1634
  /// Insert a call to an outlined function into the program.
1635
  /// Returns an iterator to the spot where we inserted the call. This must be
1636
  /// implemented by the target.
1637
  virtual MachineBasicBlock::iterator
1638
  insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
1639
                     MachineBasicBlock::iterator &It, MachineFunction &MF,
1640
0
                     const outliner::TargetCostInfo &TCI) const {
1641
0
    llvm_unreachable(
1642
0
        "Target didn't implement TargetInstrInfo::insertOutlinedCall!");
1643
0
  }
1644
1645
  /// Return true if the function can safely be outlined from.
1646
  /// A function \p MF is considered safe for outlining if an outlined function
1647
  /// produced from instructions in F will produce a program which produces the
1648
  /// same output for any set of given inputs.
1649
  virtual bool isFunctionSafeToOutlineFrom(MachineFunction &MF,
1650
0
                                           bool OutlineFromLinkOnceODRs) const {
1651
0
    llvm_unreachable("Target didn't implement "
1652
0
                     "TargetInstrInfo::isFunctionSafeToOutlineFrom!");
1653
0
  }
1654
1655
  /// Return true if the function should be outlined from by default.
1656
0
  virtual bool shouldOutlineFromFunctionByDefault(MachineFunction &MF) const {
1657
0
    return false;
1658
0
  }
1659
1660
private:
1661
  unsigned CallFrameSetupOpcode, CallFrameDestroyOpcode;
1662
  unsigned CatchRetOpcode;
1663
  unsigned ReturnOpcode;
1664
};
1665
1666
/// Provide DenseMapInfo for TargetInstrInfo::RegSubRegPair.
1667
template <> struct DenseMapInfo<TargetInstrInfo::RegSubRegPair> {
1668
  using RegInfo = DenseMapInfo<unsigned>;
1669
1670
23.0M
  static inline TargetInstrInfo::RegSubRegPair getEmptyKey() {
1671
23.0M
    return TargetInstrInfo::RegSubRegPair(RegInfo::getEmptyKey(),
1672
23.0M
                                          RegInfo::getEmptyKey());
1673
23.0M
  }
1674
1675
15.1M
  static inline TargetInstrInfo::RegSubRegPair getTombstoneKey() {
1676
15.1M
    return TargetInstrInfo::RegSubRegPair(RegInfo::getTombstoneKey(),
1677
15.1M
                                          RegInfo::getTombstoneKey());
1678
15.1M
  }
1679
1680
  /// Reuse getHashValue implementation from
1681
  /// std::pair<unsigned, unsigned>.
1682
11.2M
  static unsigned getHashValue(const TargetInstrInfo::RegSubRegPair &Val) {
1683
11.2M
    std::pair<unsigned, unsigned> PairVal = std::make_pair(Val.Reg, Val.SubReg);
1684
11.2M
    return DenseMapInfo<std::pair<unsigned, unsigned>>::getHashValue(PairVal);
1685
11.2M
  }
1686
1687
  static bool isEqual(const TargetInstrInfo::RegSubRegPair &LHS,
1688
47.6M
                      const TargetInstrInfo::RegSubRegPair &RHS) {
1689
47.6M
    return RegInfo::isEqual(LHS.Reg, RHS.Reg) &&
1690
47.6M
           
RegInfo::isEqual(LHS.SubReg, RHS.SubReg)28.2M
;
1691
47.6M
  }
1692
};
1693
1694
} // end namespace llvm
1695
1696
#endif // LLVM_TARGET_TARGETINSTRINFO_H