1 //===-- llvm/Target/TargetInstrInfo.h - Instruction Info --------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the target machine instruction set to the code generator.
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_TARGET_TARGETINSTRINFO_H
15 #define LLVM_TARGET_TARGETINSTRINFO_H
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/CodeGen/MachineCombinerPattern.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/MC/MCInstrInfo.h"
22 #include "llvm/Support/BranchProbability.h"
23 #include "llvm/Target/TargetRegisterInfo.h"
27 class InstrItineraryData;
30 class MachineMemOperand;
31 class MachineRegisterInfo;
35 class MCSymbolRefExpr;
37 class ScheduleHazardRecognizer;
40 class TargetRegisterClass;
41 class TargetRegisterInfo;
42 class TargetSubtargetInfo;
43 class TargetSchedModel;
46 template<class T> class SmallVectorImpl;
49 //---------------------------------------------------------------------------
51 /// TargetInstrInfo - Interface to description of machine instruction set
53 class TargetInstrInfo : public MCInstrInfo {
54 TargetInstrInfo(const TargetInstrInfo &) = delete;
55 void operator=(const TargetInstrInfo &) = delete;
57 TargetInstrInfo(unsigned CFSetupOpcode = ~0u, unsigned CFDestroyOpcode = ~0u)
58 : CallFrameSetupOpcode(CFSetupOpcode),
59 CallFrameDestroyOpcode(CFDestroyOpcode) {
62 virtual ~TargetInstrInfo();
64 /// Given a machine instruction descriptor, returns the register
65 /// class constraint for OpNum, or NULL.
66 const TargetRegisterClass *getRegClass(const MCInstrDesc &TID,
68 const TargetRegisterInfo *TRI,
69 const MachineFunction &MF) const;
71 /// Return true if the instruction is trivially rematerializable, meaning it
72 /// has no side effects and requires no operands that aren't always available.
73 /// This means the only allowed uses are constants and unallocatable physical
74 /// registers so that the instructions result is independent of the place
76 bool isTriviallyReMaterializable(const MachineInstr *MI,
77 AliasAnalysis *AA = nullptr) const {
78 return MI->getOpcode() == TargetOpcode::IMPLICIT_DEF ||
79 (MI->getDesc().isRematerializable() &&
80 (isReallyTriviallyReMaterializable(MI, AA) ||
81 isReallyTriviallyReMaterializableGeneric(MI, AA)));
85 /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
86 /// set, this hook lets the target specify whether the instruction is actually
87 /// trivially rematerializable, taking into consideration its operands. This
88 /// predicate must return false if the instruction has any side effects other
89 /// than producing a value, or if it requres any address registers that are
90 /// not always available.
91 /// Requirements must be check as stated in isTriviallyReMaterializable() .
92 virtual bool isReallyTriviallyReMaterializable(const MachineInstr *MI,
93 AliasAnalysis *AA) const {
98 /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
99 /// set and the target hook isReallyTriviallyReMaterializable returns false,
100 /// this function does target-independent tests to determine if the
101 /// instruction is really trivially rematerializable.
102 bool isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
103 AliasAnalysis *AA) const;
106 /// These methods return the opcode of the frame setup/destroy instructions
107 /// if they exist (-1 otherwise). Some targets use pseudo instructions in
108 /// order to abstract away the difference between operating with a frame
109 /// pointer and operating without, through the use of these two instructions.
111 unsigned getCallFrameSetupOpcode() const { return CallFrameSetupOpcode; }
112 unsigned getCallFrameDestroyOpcode() const { return CallFrameDestroyOpcode; }
114 /// Returns the actual stack pointer adjustment made by an instruction
115 /// as part of a call sequence. By default, only call frame setup/destroy
116 /// instructions adjust the stack, but targets may want to override this
117 /// to enable more fine-grained adjustment, or adjust by a different value.
118 virtual int getSPAdjust(const MachineInstr *MI) const;
120 /// Return true if the instruction is a "coalescable" extension instruction.
121 /// That is, it's like a copy where it's legal for the source to overlap the
122 /// destination. e.g. X86::MOVSX64rr32. If this returns true, then it's
123 /// expected the pre-extension value is available as a subreg of the result
124 /// register. This also returns the sub-register index in SubIdx.
125 virtual bool isCoalescableExtInstr(const MachineInstr &MI,
126 unsigned &SrcReg, unsigned &DstReg,
127 unsigned &SubIdx) const {
131 /// If the specified machine instruction is a direct
132 /// load from a stack slot, return the virtual or physical register number of
133 /// the destination along with the FrameIndex of the loaded stack slot. If
134 /// not, return 0. This predicate must return 0 if the instruction has
135 /// any side effects other than loading from the stack slot.
136 virtual unsigned isLoadFromStackSlot(const MachineInstr *MI,
137 int &FrameIndex) const {
141 /// Check for post-frame ptr elimination stack locations as well.
142 /// This uses a heuristic so it isn't reliable for correctness.
143 virtual unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI,
144 int &FrameIndex) const {
148 /// If the specified machine instruction has a load from a stack slot,
149 /// return true along with the FrameIndex of the loaded stack slot and the
150 /// machine mem operand containing the reference.
151 /// If not, return false. Unlike isLoadFromStackSlot, this returns true for
152 /// any instructions that loads from the stack. This is just a hint, as some
153 /// cases may be missed.
154 virtual bool hasLoadFromStackSlot(const MachineInstr *MI,
155 const MachineMemOperand *&MMO,
156 int &FrameIndex) const;
158 /// If the specified machine instruction is a direct
159 /// store to a stack slot, return the virtual or physical register number of
160 /// the source reg along with the FrameIndex of the loaded stack slot. If
161 /// not, return 0. This predicate must return 0 if the instruction has
162 /// any side effects other than storing to the stack slot.
163 virtual unsigned isStoreToStackSlot(const MachineInstr *MI,
164 int &FrameIndex) const {
168 /// Check for post-frame ptr elimination stack locations as well.
169 /// This uses a heuristic, so it isn't reliable for correctness.
170 virtual unsigned isStoreToStackSlotPostFE(const MachineInstr *MI,
171 int &FrameIndex) const {
175 /// If the specified machine instruction has a store to a stack slot,
176 /// return true along with the FrameIndex of the loaded stack slot and the
177 /// machine mem operand containing the reference.
178 /// If not, return false. Unlike isStoreToStackSlot,
179 /// this returns true for any instructions that stores to the
180 /// stack. This is just a hint, as some cases may be missed.
181 virtual bool hasStoreToStackSlot(const MachineInstr *MI,
182 const MachineMemOperand *&MMO,
183 int &FrameIndex) const;
185 /// Return true if the specified machine instruction
186 /// is a copy of one stack slot to another and has no other effect.
187 /// Provide the identity of the two frame indices.
188 virtual bool isStackSlotCopy(const MachineInstr *MI, int &DestFrameIndex,
189 int &SrcFrameIndex) const {
193 /// Compute the size in bytes and offset within a stack slot of a spilled
194 /// register or subregister.
196 /// \param [out] Size in bytes of the spilled value.
197 /// \param [out] Offset in bytes within the stack slot.
198 /// \returns true if both Size and Offset are successfully computed.
200 /// Not all subregisters have computable spill slots. For example,
201 /// subregisters registers may not be byte-sized, and a pair of discontiguous
202 /// subregisters has no single offset.
204 /// Targets with nontrivial bigendian implementations may need to override
205 /// this, particularly to support spilled vector registers.
206 virtual bool getStackSlotRange(const TargetRegisterClass *RC, unsigned SubIdx,
207 unsigned &Size, unsigned &Offset,
208 const MachineFunction &MF) const;
210 /// Return true if the instruction is as cheap as a move instruction.
212 /// Targets for different archs need to override this, and different
213 /// micro-architectures can also be finely tuned inside.
214 virtual bool isAsCheapAsAMove(const MachineInstr *MI) const {
215 return MI->isAsCheapAsAMove();
218 /// Re-issue the specified 'original' instruction at the
219 /// specific location targeting a new destination register.
220 /// The register in Orig->getOperand(0).getReg() will be substituted by
221 /// DestReg:SubIdx. Any existing subreg index is preserved or composed with
223 virtual void reMaterialize(MachineBasicBlock &MBB,
224 MachineBasicBlock::iterator MI,
225 unsigned DestReg, unsigned SubIdx,
226 const MachineInstr *Orig,
227 const TargetRegisterInfo &TRI) const;
229 /// Create a duplicate of the Orig instruction in MF. This is like
230 /// MachineFunction::CloneMachineInstr(), but the target may update operands
231 /// that are required to be unique.
233 /// The instruction must be duplicable as indicated by isNotDuplicable().
234 virtual MachineInstr *duplicate(MachineInstr *Orig,
235 MachineFunction &MF) const;
237 /// This method must be implemented by targets that
238 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
239 /// may be able to convert a two-address instruction into one or more true
240 /// three-address instructions on demand. This allows the X86 target (for
241 /// example) to convert ADD and SHL instructions into LEA instructions if they
242 /// would require register copies due to two-addressness.
244 /// This method returns a null pointer if the transformation cannot be
245 /// performed, otherwise it returns the last new instruction.
247 virtual MachineInstr *
248 convertToThreeAddress(MachineFunction::iterator &MFI,
249 MachineBasicBlock::iterator &MBBI, LiveVariables *LV) const {
253 /// If a target has any instructions that are commutable but require
254 /// converting to different instructions or making non-trivial changes to
255 /// commute them, this method can overloaded to do that.
256 /// The default implementation simply swaps the commutable operands.
257 /// If NewMI is false, MI is modified in place and returned; otherwise, a
258 /// new machine instruction is created and returned. Do not call this
259 /// method for a non-commutable instruction, but there may be some cases
260 /// where this method fails and returns null.
261 virtual MachineInstr *commuteInstruction(MachineInstr *MI,
262 bool NewMI = false) const;
264 /// If specified MI is commutable, return the two operand indices that would
265 /// swap value. Return false if the instruction
266 /// is not in a form which this routine understands.
267 virtual bool findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
268 unsigned &SrcOpIdx2) const;
270 /// A pair composed of a register and a sub-register index.
271 /// Used to give some type checking when modeling Reg:SubReg.
272 struct RegSubRegPair {
275 RegSubRegPair(unsigned Reg = 0, unsigned SubReg = 0)
276 : Reg(Reg), SubReg(SubReg) {}
278 /// A pair composed of a pair of a register and a sub-register index,
279 /// and another sub-register index.
280 /// Used to give some type checking when modeling Reg:SubReg1, SubReg2.
281 struct RegSubRegPairAndIdx : RegSubRegPair {
283 RegSubRegPairAndIdx(unsigned Reg = 0, unsigned SubReg = 0,
285 : RegSubRegPair(Reg, SubReg), SubIdx(SubIdx) {}
288 /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI
290 /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of
291 /// the list is modeled as <Reg:SubReg, SubIdx>.
292 /// E.g., REG_SEQUENCE vreg1:sub1, sub0, vreg2, sub1 would produce
294 /// - vreg1:sub1, sub0
295 /// - vreg2<:0>, sub1
297 /// \returns true if it is possible to build such an input sequence
298 /// with the pair \p MI, \p DefIdx. False otherwise.
300 /// \pre MI.isRegSequence() or MI.isRegSequenceLike().
302 /// \note The generic implementation does not provide any support for
303 /// MI.isRegSequenceLike(). In other words, one has to override
304 /// getRegSequenceLikeInputs for target specific instructions.
306 getRegSequenceInputs(const MachineInstr &MI, unsigned DefIdx,
307 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const;
309 /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI
311 /// \p [out] InputReg of the equivalent EXTRACT_SUBREG.
312 /// E.g., EXTRACT_SUBREG vreg1:sub1, sub0, sub1 would produce:
313 /// - vreg1:sub1, sub0
315 /// \returns true if it is possible to build such an input sequence
316 /// with the pair \p MI, \p DefIdx. False otherwise.
318 /// \pre MI.isExtractSubreg() or MI.isExtractSubregLike().
320 /// \note The generic implementation does not provide any support for
321 /// MI.isExtractSubregLike(). In other words, one has to override
322 /// getExtractSubregLikeInputs for target specific instructions.
324 getExtractSubregInputs(const MachineInstr &MI, unsigned DefIdx,
325 RegSubRegPairAndIdx &InputReg) const;
327 /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI
329 /// \p [out] BaseReg and \p [out] InsertedReg contain
330 /// the equivalent inputs of INSERT_SUBREG.
331 /// E.g., INSERT_SUBREG vreg0:sub0, vreg1:sub1, sub3 would produce:
332 /// - BaseReg: vreg0:sub0
333 /// - InsertedReg: vreg1:sub1, sub3
335 /// \returns true if it is possible to build such an input sequence
336 /// with the pair \p MI, \p DefIdx. False otherwise.
338 /// \pre MI.isInsertSubreg() or MI.isInsertSubregLike().
340 /// \note The generic implementation does not provide any support for
341 /// MI.isInsertSubregLike(). In other words, one has to override
342 /// getInsertSubregLikeInputs for target specific instructions.
344 getInsertSubregInputs(const MachineInstr &MI, unsigned DefIdx,
345 RegSubRegPair &BaseReg,
346 RegSubRegPairAndIdx &InsertedReg) const;
349 /// Return true if two machine instructions would produce identical values.
350 /// By default, this is only true when the two instructions
351 /// are deemed identical except for defs. If this function is called when the
352 /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for
353 /// aggressive checks.
354 virtual bool produceSameValue(const MachineInstr *MI0,
355 const MachineInstr *MI1,
356 const MachineRegisterInfo *MRI = nullptr) const;
358 /// Analyze the branching code at the end of MBB, returning
359 /// true if it cannot be understood (e.g. it's a switch dispatch or isn't
360 /// implemented for a target). Upon success, this returns false and returns
361 /// with the following information in various cases:
363 /// 1. If this block ends with no branches (it just falls through to its succ)
364 /// just return false, leaving TBB/FBB null.
365 /// 2. If this block ends with only an unconditional branch, it sets TBB to be
366 /// the destination block.
367 /// 3. If this block ends with a conditional branch and it falls through to a
368 /// successor block, it sets TBB to be the branch destination block and a
369 /// list of operands that evaluate the condition. These operands can be
370 /// passed to other TargetInstrInfo methods to create new branches.
371 /// 4. If this block ends with a conditional branch followed by an
372 /// unconditional branch, it returns the 'true' destination in TBB, the
373 /// 'false' destination in FBB, and a list of operands that evaluate the
374 /// condition. These operands can be passed to other TargetInstrInfo
375 /// methods to create new branches.
377 /// Note that RemoveBranch and InsertBranch must be implemented to support
378 /// cases where this method returns success.
380 /// If AllowModify is true, then this routine is allowed to modify the basic
381 /// block (e.g. delete instructions after the unconditional branch).
383 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
384 MachineBasicBlock *&FBB,
385 SmallVectorImpl<MachineOperand> &Cond,
386 bool AllowModify = false) const {
390 /// Represents a predicate at the MachineFunction level. The control flow a
391 /// MachineBranchPredicate represents is:
393 /// Reg <def>= LHS `Predicate` RHS == ConditionDef
394 /// if Reg then goto TrueDest else goto FalseDest
396 struct MachineBranchPredicate {
397 enum ComparePredicate {
398 PRED_EQ, // True if two values are equal
399 PRED_NE, // True if two values are not equal
400 PRED_INVALID // Sentinel value
403 ComparePredicate Predicate;
406 MachineBasicBlock *TrueDest;
407 MachineBasicBlock *FalseDest;
408 MachineInstr *ConditionDef;
410 /// SingleUseCondition is true if ConditionDef is dead except for the
411 /// branch(es) at the end of the basic block.
413 bool SingleUseCondition;
415 explicit MachineBranchPredicate()
416 : Predicate(PRED_INVALID), LHS(MachineOperand::CreateImm(0)),
417 RHS(MachineOperand::CreateImm(0)), TrueDest(nullptr),
418 FalseDest(nullptr), ConditionDef(nullptr), SingleUseCondition(false) {
422 /// Analyze the branching code at the end of MBB and parse it into the
423 /// MachineBranchPredicate structure if possible. Returns false on success
424 /// and true on failure.
426 /// If AllowModify is true, then this routine is allowed to modify the basic
427 /// block (e.g. delete instructions after the unconditional branch).
429 virtual bool AnalyzeBranchPredicate(MachineBasicBlock &MBB,
430 MachineBranchPredicate &MBP,
431 bool AllowModify = false) const {
435 /// Remove the branching code at the end of the specific MBB.
436 /// This is only invoked in cases where AnalyzeBranch returns success. It
437 /// returns the number of instructions that were removed.
438 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const {
439 llvm_unreachable("Target didn't implement TargetInstrInfo::RemoveBranch!");
442 /// Insert branch code into the end of the specified MachineBasicBlock.
443 /// The operands to this method are the same as those
444 /// returned by AnalyzeBranch. This is only invoked in cases where
445 /// AnalyzeBranch returns success. It returns the number of instructions
448 /// It is also invoked by tail merging to add unconditional branches in
449 /// cases where AnalyzeBranch doesn't apply because there was no original
450 /// branch to analyze. At least this much must be implemented, else tail
451 /// merging needs to be disabled.
452 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
453 MachineBasicBlock *FBB,
454 ArrayRef<MachineOperand> Cond,
456 llvm_unreachable("Target didn't implement TargetInstrInfo::InsertBranch!");
459 /// Delete the instruction OldInst and everything after it, replacing it with
460 /// an unconditional branch to NewDest. This is used by the tail merging pass.
461 virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
462 MachineBasicBlock *NewDest) const;
464 /// Get an instruction that performs an unconditional branch to the given
467 getUnconditionalBranch(MCInst &MI,
468 const MCSymbolRefExpr *BranchTarget) const {
469 llvm_unreachable("Target didn't implement "
470 "TargetInstrInfo::getUnconditionalBranch!");
473 /// Get a machine trap instruction.
474 virtual void getTrap(MCInst &MI) const {
475 llvm_unreachable("Target didn't implement TargetInstrInfo::getTrap!");
478 /// Get a number of bytes that suffices to hold
479 /// either the instruction returned by getUnconditionalBranch or the
480 /// instruction returned by getTrap. This only makes sense because
481 /// getUnconditionalBranch returns a single, specific instruction. This
482 /// information is needed by the jumptable construction code, since it must
483 /// decide how many bytes to use for a jumptable entry so it can generate the
486 /// Note that if the jumptable instruction requires alignment, then that
487 /// alignment should be factored into this required bound so that the
488 /// resulting bound gives the right alignment for the instruction.
489 virtual unsigned getJumpInstrTableEntryBound() const {
490 // This method gets called by LLVMTargetMachine always, so it can't fail
491 // just because there happens to be no implementation for this target.
492 // Any code that tries to use a jumptable annotation without defining
493 // getUnconditionalBranch on the appropriate Target will fail anyway, and
494 // the value returned here won't matter in that case.
498 /// Return true if it's legal to split the given basic
499 /// block at the specified instruction (i.e. instruction would be the start
500 /// of a new basic block).
501 virtual bool isLegalToSplitMBBAt(MachineBasicBlock &MBB,
502 MachineBasicBlock::iterator MBBI) const {
506 /// Return true if it's profitable to predicate
507 /// instructions with accumulated instruction latency of "NumCycles"
508 /// of the specified basic block, where the probability of the instructions
509 /// being executed is given by Probability, and Confidence is a measure
510 /// of our confidence that it will be properly predicted.
512 bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
513 unsigned ExtraPredCycles,
514 BranchProbability Probability) const {
518 /// Second variant of isProfitableToIfCvt. This one
519 /// checks for the case where two basic blocks from true and false path
520 /// of a if-then-else (diamond) are predicated on mutally exclusive
521 /// predicates, where the probability of the true path being taken is given
522 /// by Probability, and Confidence is a measure of our confidence that it
523 /// will be properly predicted.
525 isProfitableToIfCvt(MachineBasicBlock &TMBB,
526 unsigned NumTCycles, unsigned ExtraTCycles,
527 MachineBasicBlock &FMBB,
528 unsigned NumFCycles, unsigned ExtraFCycles,
529 BranchProbability Probability) const {
533 /// Return true if it's profitable for if-converter to duplicate instructions
534 /// of specified accumulated instruction latencies in the specified MBB to
535 /// enable if-conversion.
536 /// The probability of the instructions being executed is given by
537 /// Probability, and Confidence is a measure of our confidence that it
538 /// will be properly predicted.
540 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
541 BranchProbability Probability) const {
545 /// Return true if it's profitable to unpredicate
546 /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually
547 /// exclusive predicates.
555 /// This may be profitable is conditional instructions are always executed.
556 virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
557 MachineBasicBlock &FMBB) const {
561 /// Return true if it is possible to insert a select
562 /// instruction that chooses between TrueReg and FalseReg based on the
563 /// condition code in Cond.
565 /// When successful, also return the latency in cycles from TrueReg,
566 /// FalseReg, and Cond to the destination register. In most cases, a select
567 /// instruction will be 1 cycle, so CondCycles = TrueCycles = FalseCycles = 1
569 /// Some x86 implementations have 2-cycle cmov instructions.
571 /// @param MBB Block where select instruction would be inserted.
572 /// @param Cond Condition returned by AnalyzeBranch.
573 /// @param TrueReg Virtual register to select when Cond is true.
574 /// @param FalseReg Virtual register to select when Cond is false.
575 /// @param CondCycles Latency from Cond+Branch to select output.
576 /// @param TrueCycles Latency from TrueReg to select output.
577 /// @param FalseCycles Latency from FalseReg to select output.
578 virtual bool canInsertSelect(const MachineBasicBlock &MBB,
579 ArrayRef<MachineOperand> Cond,
580 unsigned TrueReg, unsigned FalseReg,
582 int &TrueCycles, int &FalseCycles) const {
586 /// Insert a select instruction into MBB before I that will copy TrueReg to
587 /// DstReg when Cond is true, and FalseReg to DstReg when Cond is false.
589 /// This function can only be called after canInsertSelect() returned true.
590 /// The condition in Cond comes from AnalyzeBranch, and it can be assumed
591 /// that the same flags or registers required by Cond are available at the
594 /// @param MBB Block where select instruction should be inserted.
595 /// @param I Insertion point.
596 /// @param DL Source location for debugging.
597 /// @param DstReg Virtual register to be defined by select instruction.
598 /// @param Cond Condition as computed by AnalyzeBranch.
599 /// @param TrueReg Virtual register to copy when Cond is true.
600 /// @param FalseReg Virtual register to copy when Cons is false.
601 virtual void insertSelect(MachineBasicBlock &MBB,
602 MachineBasicBlock::iterator I, DebugLoc DL,
603 unsigned DstReg, ArrayRef<MachineOperand> Cond,
604 unsigned TrueReg, unsigned FalseReg) const {
605 llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!");
608 /// Analyze the given select instruction, returning true if
609 /// it cannot be understood. It is assumed that MI->isSelect() is true.
611 /// When successful, return the controlling condition and the operands that
612 /// determine the true and false result values.
614 /// Result = SELECT Cond, TrueOp, FalseOp
616 /// Some targets can optimize select instructions, for example by predicating
617 /// the instruction defining one of the operands. Such targets should set
620 /// @param MI Select instruction to analyze.
621 /// @param Cond Condition controlling the select.
622 /// @param TrueOp Operand number of the value selected when Cond is true.
623 /// @param FalseOp Operand number of the value selected when Cond is false.
624 /// @param Optimizable Returned as true if MI is optimizable.
625 /// @returns False on success.
626 virtual bool analyzeSelect(const MachineInstr *MI,
627 SmallVectorImpl<MachineOperand> &Cond,
628 unsigned &TrueOp, unsigned &FalseOp,
629 bool &Optimizable) const {
630 assert(MI && MI->getDesc().isSelect() && "MI must be a select instruction");
634 /// Given a select instruction that was understood by
635 /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by
636 /// merging it with one of its operands. Returns NULL on failure.
638 /// When successful, returns the new select instruction. The client is
639 /// responsible for deleting MI.
641 /// If both sides of the select can be optimized, PreferFalse is used to pick
644 /// @param MI Optimizable select instruction.
645 /// @param NewMIs Set that record all MIs in the basic block up to \p
646 /// MI. Has to be updated with any newly created MI or deleted ones.
647 /// @param PreferFalse Try to optimize FalseOp instead of TrueOp.
648 /// @returns Optimized instruction or NULL.
649 virtual MachineInstr *optimizeSelect(MachineInstr *MI,
650 SmallPtrSetImpl<MachineInstr *> &NewMIs,
651 bool PreferFalse = false) const {
652 // This function must be implemented if Optimizable is ever set.
653 llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!");
656 /// Emit instructions to copy a pair of physical registers.
658 /// This function should support copies within any legal register class as
659 /// well as any cross-class copies created during instruction selection.
661 /// The source and destination registers may overlap, which may require a
662 /// careful implementation when multiple copy instructions are required for
663 /// large registers. See for example the ARM target.
664 virtual void copyPhysReg(MachineBasicBlock &MBB,
665 MachineBasicBlock::iterator MI, DebugLoc DL,
666 unsigned DestReg, unsigned SrcReg,
667 bool KillSrc) const {
668 llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!");
671 /// Store the specified register of the given register class to the specified
672 /// stack frame index. The store instruction is to be added to the given
673 /// machine basic block before the specified machine instruction. If isKill
674 /// is true, the register operand is the last use and must be marked kill.
675 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
676 MachineBasicBlock::iterator MI,
677 unsigned SrcReg, bool isKill, int FrameIndex,
678 const TargetRegisterClass *RC,
679 const TargetRegisterInfo *TRI) const {
680 llvm_unreachable("Target didn't implement "
681 "TargetInstrInfo::storeRegToStackSlot!");
684 /// Load the specified register of the given register class from the specified
685 /// stack frame index. The load instruction is to be added to the given
686 /// machine basic block before the specified machine instruction.
687 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
688 MachineBasicBlock::iterator MI,
689 unsigned DestReg, int FrameIndex,
690 const TargetRegisterClass *RC,
691 const TargetRegisterInfo *TRI) const {
692 llvm_unreachable("Target didn't implement "
693 "TargetInstrInfo::loadRegFromStackSlot!");
696 /// This function is called for all pseudo instructions
697 /// that remain after register allocation. Many pseudo instructions are
698 /// created to help register allocation. This is the place to convert them
699 /// into real instructions. The target can edit MI in place, or it can insert
700 /// new instructions and erase MI. The function should return true if
701 /// anything was changed.
702 virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
706 /// Attempt to fold a load or store of the specified stack
707 /// slot into the specified machine instruction for the specified operand(s).
708 /// If this is possible, a new instruction is returned with the specified
709 /// operand folded, otherwise NULL is returned.
710 /// The new instruction is inserted before MI, and the client is responsible
711 /// for removing the old instruction.
712 MachineInstr *foldMemoryOperand(MachineBasicBlock::iterator MI,
713 ArrayRef<unsigned> Ops, int FrameIndex) const;
715 /// Same as the previous version except it allows folding of any load and
716 /// store from / to any address, not just from a specific stack slot.
717 MachineInstr *foldMemoryOperand(MachineBasicBlock::iterator MI,
718 ArrayRef<unsigned> Ops,
719 MachineInstr *LoadMI) const;
721 /// Return true when there is potentially a faster code sequence
722 /// for an instruction chain ending in \p Root. All potential patterns are
723 /// returned in the \p Pattern vector. Pattern should be sorted in priority
724 /// order since the pattern evaluator stops checking as soon as it finds a
726 /// \param Root - Instruction that could be combined with one of its operands
727 /// \param Patterns - Vector of possible combination patterns
728 virtual bool getMachineCombinerPatterns(
730 SmallVectorImpl<MachineCombinerPattern::MC_PATTERN> &Patterns) const;
732 /// Return true if the input \P Inst is part of a chain of dependent ops
733 /// that are suitable for reassociation, otherwise return false.
734 /// If the instruction's operands must be commuted to have a previous
735 /// instruction of the same type define the first source operand, \P Commuted
736 /// will be set to true.
737 bool isReassociationCandidate(const MachineInstr &Inst, bool &Commuted) const;
739 /// Return true when \P Inst is both associative and commutative.
740 virtual bool isAssociativeAndCommutative(const MachineInstr &Inst) const {
744 /// Return true when \P Inst has reassociable operands in the same \P MBB.
745 virtual bool hasReassociableOperands(const MachineInstr &Inst,
746 const MachineBasicBlock *MBB) const;
748 /// Return true when \P Inst has reassociable sibling.
749 bool hasReassociableSibling(const MachineInstr &Inst, bool &Commuted) const;
751 /// When getMachineCombinerPatterns() finds patterns, this function generates
752 /// the instructions that could replace the original code sequence. The client
753 /// has to decide whether the actual replacement is beneficial or not.
754 /// \param Root - Instruction that could be combined with one of its operands
755 /// \param Pattern - Combination pattern for Root
756 /// \param InsInstrs - Vector of new instructions that implement P
757 /// \param DelInstrs - Old instructions, including Root, that could be
758 /// replaced by InsInstr
759 /// \param InstrIdxForVirtReg - map of virtual register to instruction in
760 /// InsInstr that defines it
761 virtual void genAlternativeCodeSequence(
762 MachineInstr &Root, MachineCombinerPattern::MC_PATTERN Pattern,
763 SmallVectorImpl<MachineInstr *> &InsInstrs,
764 SmallVectorImpl<MachineInstr *> &DelInstrs,
765 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const;
767 /// Attempt to reassociate \P Root and \P Prev according to \P Pattern to
768 /// reduce critical path length.
769 void reassociateOps(MachineInstr &Root, MachineInstr &Prev,
770 MachineCombinerPattern::MC_PATTERN Pattern,
771 SmallVectorImpl<MachineInstr *> &InsInstrs,
772 SmallVectorImpl<MachineInstr *> &DelInstrs,
773 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const;
775 /// This is an architecture-specific helper function of reassociateOps.
776 /// Set special operand attributes for new instructions after reassociation.
777 virtual void setSpecialOperandAttr(MachineInstr &OldMI1, MachineInstr &OldMI2,
778 MachineInstr &NewMI1,
779 MachineInstr &NewMI2) const {
783 /// Return true when a target supports MachineCombiner.
784 virtual bool useMachineCombiner() const { return false; }
787 /// Target-dependent implementation for foldMemoryOperand.
788 /// Target-independent code in foldMemoryOperand will
789 /// take care of adding a MachineMemOperand to the newly created instruction.
790 /// The instruction and any auxiliary instructions necessary will be inserted
792 virtual MachineInstr *foldMemoryOperandImpl(
793 MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops,
794 MachineBasicBlock::iterator InsertPt, int FrameIndex) const {
798 /// Target-dependent implementation for foldMemoryOperand.
799 /// Target-independent code in foldMemoryOperand will
800 /// take care of adding a MachineMemOperand to the newly created instruction.
801 /// The instruction and any auxiliary instructions necessary will be inserted
803 virtual MachineInstr *foldMemoryOperandImpl(
804 MachineFunction &MF, MachineInstr *MI, ArrayRef<unsigned> Ops,
805 MachineBasicBlock::iterator InsertPt, MachineInstr *LoadMI) const {
809 /// \brief Target-dependent implementation of getRegSequenceInputs.
811 /// \returns true if it is possible to build the equivalent
812 /// REG_SEQUENCE inputs with the pair \p MI, \p DefIdx. False otherwise.
814 /// \pre MI.isRegSequenceLike().
816 /// \see TargetInstrInfo::getRegSequenceInputs.
817 virtual bool getRegSequenceLikeInputs(
818 const MachineInstr &MI, unsigned DefIdx,
819 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
823 /// \brief Target-dependent implementation of getExtractSubregInputs.
825 /// \returns true if it is possible to build the equivalent
826 /// EXTRACT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise.
828 /// \pre MI.isExtractSubregLike().
830 /// \see TargetInstrInfo::getExtractSubregInputs.
831 virtual bool getExtractSubregLikeInputs(
832 const MachineInstr &MI, unsigned DefIdx,
833 RegSubRegPairAndIdx &InputReg) const {
837 /// \brief Target-dependent implementation of getInsertSubregInputs.
839 /// \returns true if it is possible to build the equivalent
840 /// INSERT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise.
842 /// \pre MI.isInsertSubregLike().
844 /// \see TargetInstrInfo::getInsertSubregInputs.
846 getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx,
847 RegSubRegPair &BaseReg,
848 RegSubRegPairAndIdx &InsertedReg) const {
853 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
854 /// a store or a load and a store into two or more instruction. If this is
855 /// possible, returns true as well as the new instructions by reference.
856 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
857 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
858 SmallVectorImpl<MachineInstr*> &NewMIs) const{
862 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
863 SmallVectorImpl<SDNode*> &NewNodes) const {
867 /// Returns the opcode of the would be new
868 /// instruction after load / store are unfolded from an instruction of the
869 /// specified opcode. It returns zero if the specified unfolding is not
870 /// possible. If LoadRegIndex is non-null, it is filled in with the operand
871 /// index of the operand which will hold the register holding the loaded
873 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
874 bool UnfoldLoad, bool UnfoldStore,
875 unsigned *LoadRegIndex = nullptr) const {
879 /// This is used by the pre-regalloc scheduler to determine if two loads are
880 /// loading from the same base address. It should only return true if the base
881 /// pointers are the same and the only differences between the two addresses
882 /// are the offset. It also returns the offsets by reference.
883 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
884 int64_t &Offset1, int64_t &Offset2) const {
888 /// This is a used by the pre-regalloc scheduler to determine (in conjunction
889 /// with areLoadsFromSameBasePtr) if two loads should be scheduled together.
890 /// On some targets if two loads are loading from
891 /// addresses in the same cache line, it's better if they are scheduled
892 /// together. This function takes two integers that represent the load offsets
893 /// from the common base address. It returns true if it decides it's desirable
894 /// to schedule the two loads together. "NumLoads" is the number of loads that
895 /// have already been scheduled after Load1.
896 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
897 int64_t Offset1, int64_t Offset2,
898 unsigned NumLoads) const {
902 /// Get the base register and byte offset of an instruction that reads/writes
904 virtual bool getMemOpBaseRegImmOfs(MachineInstr *MemOp, unsigned &BaseReg,
906 const TargetRegisterInfo *TRI) const {
910 virtual bool enableClusterLoads() const { return false; }
912 virtual bool shouldClusterLoads(MachineInstr *FirstLdSt,
913 MachineInstr *SecondLdSt,
914 unsigned NumLoads) const {
918 /// Can this target fuse the given instructions if they are scheduled
920 virtual bool shouldScheduleAdjacent(MachineInstr* First,
921 MachineInstr *Second) const {
925 /// Reverses the branch condition of the specified condition list,
926 /// returning false on success and true if it cannot be reversed.
928 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
932 /// Insert a noop into the instruction stream at the specified point.
933 virtual void insertNoop(MachineBasicBlock &MBB,
934 MachineBasicBlock::iterator MI) const;
937 /// Return the noop instruction to use for a noop.
938 virtual void getNoopForMachoTarget(MCInst &NopInst) const;
941 /// Returns true if the instruction is already predicated.
942 virtual bool isPredicated(const MachineInstr *MI) const {
946 /// Returns true if the instruction is a
947 /// terminator instruction that has not been predicated.
948 virtual bool isUnpredicatedTerminator(const MachineInstr *MI) const;
950 /// Convert the instruction into a predicated instruction.
951 /// It returns true if the operation was successful.
953 bool PredicateInstruction(MachineInstr *MI,
954 ArrayRef<MachineOperand> Pred) const;
956 /// Returns true if the first specified predicate
957 /// subsumes the second, e.g. GE subsumes GT.
959 bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
960 ArrayRef<MachineOperand> Pred2) const {
964 /// If the specified instruction defines any predicate
965 /// or condition code register(s) used for predication, returns true as well
966 /// as the definition predicate(s) by reference.
967 virtual bool DefinesPredicate(MachineInstr *MI,
968 std::vector<MachineOperand> &Pred) const {
972 /// Return true if the specified instruction can be predicated.
973 /// By default, this returns true for every instruction with a
974 /// PredicateOperand.
975 virtual bool isPredicable(MachineInstr *MI) const {
976 return MI->getDesc().isPredicable();
979 /// Return true if it's safe to move a machine
980 /// instruction that defines the specified register class.
981 virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
985 /// Test if the given instruction should be considered a scheduling boundary.
986 /// This primarily includes labels and terminators.
987 virtual bool isSchedulingBoundary(const MachineInstr *MI,
988 const MachineBasicBlock *MBB,
989 const MachineFunction &MF) const;
991 /// Measure the specified inline asm to determine an approximation of its
993 virtual unsigned getInlineAsmLength(const char *Str,
994 const MCAsmInfo &MAI) const;
996 /// Allocate and return a hazard recognizer to use for this target when
997 /// scheduling the machine instructions before register allocation.
998 virtual ScheduleHazardRecognizer*
999 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
1000 const ScheduleDAG *DAG) const;
1002 /// Allocate and return a hazard recognizer to use for this target when
1003 /// scheduling the machine instructions before register allocation.
1004 virtual ScheduleHazardRecognizer*
1005 CreateTargetMIHazardRecognizer(const InstrItineraryData*,
1006 const ScheduleDAG *DAG) const;
1008 /// Allocate and return a hazard recognizer to use for this target when
1009 /// scheduling the machine instructions after register allocation.
1010 virtual ScheduleHazardRecognizer*
1011 CreateTargetPostRAHazardRecognizer(const InstrItineraryData*,
1012 const ScheduleDAG *DAG) const;
1014 /// Provide a global flag for disabling the PreRA hazard recognizer that
1015 /// targets may choose to honor.
1016 bool usePreRAHazardRecognizer() const;
1018 /// For a comparison instruction, return the source registers
1019 /// in SrcReg and SrcReg2 if having two register operands, and the value it
1020 /// compares against in CmpValue. Return true if the comparison instruction
1021 /// can be analyzed.
1022 virtual bool analyzeCompare(const MachineInstr *MI,
1023 unsigned &SrcReg, unsigned &SrcReg2,
1024 int &Mask, int &Value) const {
1028 /// See if the comparison instruction can be converted
1029 /// into something more efficient. E.g., on ARM most instructions can set the
1030 /// flags register, obviating the need for a separate CMP.
1031 virtual bool optimizeCompareInstr(MachineInstr *CmpInstr,
1032 unsigned SrcReg, unsigned SrcReg2,
1033 int Mask, int Value,
1034 const MachineRegisterInfo *MRI) const {
1037 virtual bool optimizeCondBranch(MachineInstr *MI) const { return false; }
1039 /// Try to remove the load by folding it to a register operand at the use.
1040 /// We fold the load instructions if and only if the
1041 /// def and use are in the same BB. We only look at one load and see
1042 /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
1043 /// defined by the load we are trying to fold. DefMI returns the machine
1044 /// instruction that defines FoldAsLoadDefReg, and the function returns
1045 /// the machine instruction generated due to folding.
1046 virtual MachineInstr* optimizeLoadInstr(MachineInstr *MI,
1047 const MachineRegisterInfo *MRI,
1048 unsigned &FoldAsLoadDefReg,
1049 MachineInstr *&DefMI) const {
1053 /// 'Reg' is known to be defined by a move immediate instruction,
1054 /// try to fold the immediate into the use instruction.
1055 /// If MRI->hasOneNonDBGUse(Reg) is true, and this function returns true,
1056 /// then the caller may assume that DefMI has been erased from its parent
1057 /// block. The caller may assume that it will not be erased by this
1058 /// function otherwise.
1059 virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
1060 unsigned Reg, MachineRegisterInfo *MRI) const {
1064 /// Return the number of u-operations the given machine
1065 /// instruction will be decoded to on the target cpu. The itinerary's
1066 /// IssueWidth is the number of microops that can be dispatched each
1067 /// cycle. An instruction with zero microops takes no dispatch resources.
1068 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
1069 const MachineInstr *MI) const;
1071 /// Return true for pseudo instructions that don't consume any
1072 /// machine resources in their current form. These are common cases that the
1073 /// scheduler should consider free, rather than conservatively handling them
1074 /// as instructions with no itinerary.
1075 bool isZeroCost(unsigned Opcode) const {
1076 return Opcode <= TargetOpcode::COPY;
1079 virtual int getOperandLatency(const InstrItineraryData *ItinData,
1080 SDNode *DefNode, unsigned DefIdx,
1081 SDNode *UseNode, unsigned UseIdx) const;
1083 /// Compute and return the use operand latency of a given pair of def and use.
1084 /// In most cases, the static scheduling itinerary was enough to determine the
1085 /// operand latency. But it may not be possible for instructions with variable
1086 /// number of defs / uses.
1088 /// This is a raw interface to the itinerary that may be directly overridden
1089 /// by a target. Use computeOperandLatency to get the best estimate of
1091 virtual int getOperandLatency(const InstrItineraryData *ItinData,
1092 const MachineInstr *DefMI, unsigned DefIdx,
1093 const MachineInstr *UseMI,
1094 unsigned UseIdx) const;
1096 /// Compute and return the latency of the given data
1097 /// dependent def and use when the operand indices are already known.
1098 unsigned computeOperandLatency(const InstrItineraryData *ItinData,
1099 const MachineInstr *DefMI, unsigned DefIdx,
1100 const MachineInstr *UseMI, unsigned UseIdx)
1103 /// Compute the instruction latency of a given instruction.
1104 /// If the instruction has higher cost when predicated, it's returned via
1106 virtual unsigned getInstrLatency(const InstrItineraryData *ItinData,
1107 const MachineInstr *MI,
1108 unsigned *PredCost = nullptr) const;
1110 virtual unsigned getPredicationCost(const MachineInstr *MI) const;
1112 virtual int getInstrLatency(const InstrItineraryData *ItinData,
1113 SDNode *Node) const;
1115 /// Return the default expected latency for a def based on it's opcode.
1116 unsigned defaultDefLatency(const MCSchedModel &SchedModel,
1117 const MachineInstr *DefMI) const;
1119 int computeDefOperandLatency(const InstrItineraryData *ItinData,
1120 const MachineInstr *DefMI) const;
1122 /// Return true if this opcode has high latency to its result.
1123 virtual bool isHighLatencyDef(int opc) const { return false; }
1125 /// Compute operand latency between a def of 'Reg'
1126 /// and a use in the current loop. Return true if the target considered
1127 /// it 'high'. This is used by optimization passes such as machine LICM to
1128 /// determine whether it makes sense to hoist an instruction out even in a
1129 /// high register pressure situation.
1131 bool hasHighOperandLatency(const TargetSchedModel &SchedModel,
1132 const MachineRegisterInfo *MRI,
1133 const MachineInstr *DefMI, unsigned DefIdx,
1134 const MachineInstr *UseMI, unsigned UseIdx) const {
1138 /// Compute operand latency of a def of 'Reg'. Return true
1139 /// if the target considered it 'low'.
1141 bool hasLowDefLatency(const TargetSchedModel &SchedModel,
1142 const MachineInstr *DefMI, unsigned DefIdx) const;
1144 /// Perform target-specific instruction verification.
1146 bool verifyInstruction(const MachineInstr *MI, StringRef &ErrInfo) const {
1150 /// Return the current execution domain and bit mask of
1151 /// possible domains for instruction.
1153 /// Some micro-architectures have multiple execution domains, and multiple
1154 /// opcodes that perform the same operation in different domains. For
1155 /// example, the x86 architecture provides the por, orps, and orpd
1156 /// instructions that all do the same thing. There is a latency penalty if a
1157 /// register is written in one domain and read in another.
1159 /// This function returns a pair (domain, mask) containing the execution
1160 /// domain of MI, and a bit mask of possible domains. The setExecutionDomain
1161 /// function can be used to change the opcode to one of the domains in the
1162 /// bit mask. Instructions whose execution domain can't be changed should
1163 /// return a 0 mask.
1165 /// The execution domain numbers don't have any special meaning except domain
1166 /// 0 is used for instructions that are not associated with any interesting
1167 /// execution domain.
1169 virtual std::pair<uint16_t, uint16_t>
1170 getExecutionDomain(const MachineInstr *MI) const {
1171 return std::make_pair(0, 0);
1174 /// Change the opcode of MI to execute in Domain.
1176 /// The bit (1 << Domain) must be set in the mask returned from
1177 /// getExecutionDomain(MI).
1178 virtual void setExecutionDomain(MachineInstr *MI, unsigned Domain) const {}
1181 /// Returns the preferred minimum clearance
1182 /// before an instruction with an unwanted partial register update.
1184 /// Some instructions only write part of a register, and implicitly need to
1185 /// read the other parts of the register. This may cause unwanted stalls
1186 /// preventing otherwise unrelated instructions from executing in parallel in
1187 /// an out-of-order CPU.
1189 /// For example, the x86 instruction cvtsi2ss writes its result to bits
1190 /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so
1191 /// the instruction needs to wait for the old value of the register to become
1194 /// addps %xmm1, %xmm0
1195 /// movaps %xmm0, (%rax)
1196 /// cvtsi2ss %rbx, %xmm0
1198 /// In the code above, the cvtsi2ss instruction needs to wait for the addps
1199 /// instruction before it can issue, even though the high bits of %xmm0
1200 /// probably aren't needed.
1202 /// This hook returns the preferred clearance before MI, measured in
1203 /// instructions. Other defs of MI's operand OpNum are avoided in the last N
1204 /// instructions before MI. It should only return a positive value for
1205 /// unwanted dependencies. If the old bits of the defined register have
1206 /// useful values, or if MI is determined to otherwise read the dependency,
1207 /// the hook should return 0.
1209 /// The unwanted dependency may be handled by:
1211 /// 1. Allocating the same register for an MI def and use. That makes the
1212 /// unwanted dependency identical to a required dependency.
1214 /// 2. Allocating a register for the def that has no defs in the previous N
1217 /// 3. Calling breakPartialRegDependency() with the same arguments. This
1218 /// allows the target to insert a dependency breaking instruction.
1221 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
1222 const TargetRegisterInfo *TRI) const {
1223 // The default implementation returns 0 for no partial register dependency.
1227 /// \brief Return the minimum clearance before an instruction that reads an
1228 /// unused register.
1230 /// For example, AVX instructions may copy part of a register operand into
1231 /// the unused high bits of the destination register.
1233 /// vcvtsi2sdq %rax, %xmm0<undef>, %xmm14
1235 /// In the code above, vcvtsi2sdq copies %xmm0[127:64] into %xmm14 creating a
1236 /// false dependence on any previous write to %xmm0.
1238 /// This hook works similarly to getPartialRegUpdateClearance, except that it
1239 /// does not take an operand index. Instead sets \p OpNum to the index of the
1240 /// unused register.
1241 virtual unsigned getUndefRegClearance(const MachineInstr *MI, unsigned &OpNum,
1242 const TargetRegisterInfo *TRI) const {
1243 // The default implementation returns 0 for no undef register dependency.
1247 /// Insert a dependency-breaking instruction
1248 /// before MI to eliminate an unwanted dependency on OpNum.
1250 /// If it wasn't possible to avoid a def in the last N instructions before MI
1251 /// (see getPartialRegUpdateClearance), this hook will be called to break the
1252 /// unwanted dependency.
1254 /// On x86, an xorps instruction can be used as a dependency breaker:
1256 /// addps %xmm1, %xmm0
1257 /// movaps %xmm0, (%rax)
1258 /// xorps %xmm0, %xmm0
1259 /// cvtsi2ss %rbx, %xmm0
1261 /// An <imp-kill> operand should be added to MI if an instruction was
1262 /// inserted. This ties the instructions together in the post-ra scheduler.
1265 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
1266 const TargetRegisterInfo *TRI) const {}
1268 /// Create machine specific model for scheduling.
1269 virtual DFAPacketizer *
1270 CreateTargetScheduleState(const TargetSubtargetInfo &) const {
1274 // Sometimes, it is possible for the target
1275 // to tell, even without aliasing information, that two MIs access different
1276 // memory addresses. This function returns true if two MIs access different
1277 // memory addresses and false otherwise.
1279 areMemAccessesTriviallyDisjoint(MachineInstr *MIa, MachineInstr *MIb,
1280 AliasAnalysis *AA = nullptr) const {
1281 assert(MIa && (MIa->mayLoad() || MIa->mayStore()) &&
1282 "MIa must load from or modify a memory location");
1283 assert(MIb && (MIb->mayLoad() || MIb->mayStore()) &&
1284 "MIb must load from or modify a memory location");
1288 /// \brief Return the value to use for the MachineCSE's LookAheadLimit,
1289 /// which is a heuristic used for CSE'ing phys reg defs.
1290 virtual unsigned getMachineCSELookAheadLimit () const {
1291 // The default lookahead is small to prevent unprofitable quadratic
1296 /// Return an array that contains the ids of the target indices (used for the
1297 /// TargetIndex machine operand) and their names.
1299 /// MIR Serialization is able to serialize only the target indices that are
1300 /// defined by this method.
1301 virtual ArrayRef<std::pair<int, const char *>>
1302 getSerializableTargetIndices() const {
1306 /// Decompose the machine operand's target flags into two values - the direct
1307 /// target flag value and any of bit flags that are applied.
1308 virtual std::pair<unsigned, unsigned>
1309 decomposeMachineOperandsTargetFlags(unsigned /*TF*/) const {
1310 return std::make_pair(0u, 0u);
1313 /// Return an array that contains the direct target flag values and their
1316 /// MIR Serialization is able to serialize only the target flags that are
1317 /// defined by this method.
1318 virtual ArrayRef<std::pair<unsigned, const char *>>
1319 getSerializableDirectMachineOperandTargetFlags() const {
1323 /// Return an array that contains the bitmask target flag values and their
1326 /// MIR Serialization is able to serialize only the target flags that are
1327 /// defined by this method.
1328 virtual ArrayRef<std::pair<unsigned, const char *>>
1329 getSerializableBitmaskMachineOperandTargetFlags() const {
1334 unsigned CallFrameSetupOpcode, CallFrameDestroyOpcode;
1337 /// \brief Provide DenseMapInfo for TargetInstrInfo::RegSubRegPair.
1339 struct DenseMapInfo<TargetInstrInfo::RegSubRegPair> {
1340 typedef DenseMapInfo<unsigned> RegInfo;
1342 static inline TargetInstrInfo::RegSubRegPair getEmptyKey() {
1343 return TargetInstrInfo::RegSubRegPair(RegInfo::getEmptyKey(),
1344 RegInfo::getEmptyKey());
1346 static inline TargetInstrInfo::RegSubRegPair getTombstoneKey() {
1347 return TargetInstrInfo::RegSubRegPair(RegInfo::getTombstoneKey(),
1348 RegInfo::getTombstoneKey());
1350 /// \brief Reuse getHashValue implementation from
1351 /// std::pair<unsigned, unsigned>.
1352 static unsigned getHashValue(const TargetInstrInfo::RegSubRegPair &Val) {
1353 std::pair<unsigned, unsigned> PairVal =
1354 std::make_pair(Val.Reg, Val.SubReg);
1355 return DenseMapInfo<std::pair<unsigned, unsigned>>::getHashValue(PairVal);
1357 static bool isEqual(const TargetInstrInfo::RegSubRegPair &LHS,
1358 const TargetInstrInfo::RegSubRegPair &RHS) {
1359 return RegInfo::isEqual(LHS.Reg, RHS.Reg) &&
1360 RegInfo::isEqual(LHS.SubReg, RHS.SubReg);
1364 } // End llvm namespace