1 //===- llvm/CodeGen/TargetInstrInfo.h - Instruction Info --------*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file describes the target machine instruction set to the code generator.
11 //===----------------------------------------------------------------------===//
13 #ifndef LLVM_TARGET_TARGETINSTRINFO_H
14 #define LLVM_TARGET_TARGETINSTRINFO_H
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/DenseMapInfo.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/CodeGen/LiveRegUnits.h"
21 #include "llvm/CodeGen/MachineBasicBlock.h"
22 #include "llvm/CodeGen/MachineCombinerPattern.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstr.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.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/CodeGen/VirtRegMap.h"
31 #include "llvm/MC/MCInstrInfo.h"
32 #include "llvm/Support/BranchProbability.h"
33 #include "llvm/Support/ErrorHandling.h"
43 class InstrItineraryData
;
46 class MachineMemOperand
;
47 class MachineRegisterInfo
;
53 class ScheduleHazardRecognizer
;
57 class TargetRegisterClass
;
58 class TargetRegisterInfo
;
59 class TargetSchedModel
;
60 class TargetSubtargetInfo
;
62 template <class T
> class SmallVectorImpl
;
64 using ParamLoadedValue
= std::pair
<MachineOperand
, DIExpression
*>;
66 //---------------------------------------------------------------------------
68 /// TargetInstrInfo - Interface to description of machine instruction set
70 class TargetInstrInfo
: public MCInstrInfo
{
72 TargetInstrInfo(unsigned CFSetupOpcode
= ~0u, unsigned CFDestroyOpcode
= ~0u,
73 unsigned CatchRetOpcode
= ~0u, unsigned ReturnOpcode
= ~0u)
74 : CallFrameSetupOpcode(CFSetupOpcode
),
75 CallFrameDestroyOpcode(CFDestroyOpcode
), CatchRetOpcode(CatchRetOpcode
),
76 ReturnOpcode(ReturnOpcode
) {}
77 TargetInstrInfo(const TargetInstrInfo
&) = delete;
78 TargetInstrInfo
&operator=(const TargetInstrInfo
&) = delete;
79 virtual ~TargetInstrInfo();
81 static bool isGenericOpcode(unsigned Opc
) {
82 return Opc
<= TargetOpcode::GENERIC_OP_END
;
85 /// Given a machine instruction descriptor, returns the register
86 /// class constraint for OpNum, or NULL.
88 const TargetRegisterClass
*getRegClass(const MCInstrDesc
&MCID
, unsigned OpNum
,
89 const TargetRegisterInfo
*TRI
,
90 const MachineFunction
&MF
) const;
92 /// Return true if the instruction is trivially rematerializable, meaning it
93 /// has no side effects and requires no operands that aren't always available.
94 /// This means the only allowed uses are constants and unallocatable physical
95 /// registers so that the instructions result is independent of the place
97 bool isTriviallyReMaterializable(const MachineInstr
&MI
,
98 AliasAnalysis
*AA
= nullptr) const {
99 return MI
.getOpcode() == TargetOpcode::IMPLICIT_DEF
||
100 (MI
.getDesc().isRematerializable() &&
101 (isReallyTriviallyReMaterializable(MI
, AA
) ||
102 isReallyTriviallyReMaterializableGeneric(MI
, AA
)));
106 /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
107 /// set, this hook lets the target specify whether the instruction is actually
108 /// trivially rematerializable, taking into consideration its operands. This
109 /// predicate must return false if the instruction has any side effects other
110 /// than producing a value, or if it requres any address registers that are
111 /// not always available.
112 /// Requirements must be check as stated in isTriviallyReMaterializable() .
113 virtual bool isReallyTriviallyReMaterializable(const MachineInstr
&MI
,
114 AliasAnalysis
*AA
) const {
118 /// This method commutes the operands of the given machine instruction MI.
119 /// The operands to be commuted are specified by their indices OpIdx1 and
122 /// If a target has any instructions that are commutable but require
123 /// converting to different instructions or making non-trivial changes
124 /// to commute them, this method can be overloaded to do that.
125 /// The default implementation simply swaps the commutable operands.
127 /// If NewMI is false, MI is modified in place and returned; otherwise, a
128 /// new machine instruction is created and returned.
130 /// Do not call this method for a non-commutable instruction.
131 /// Even though the instruction is commutable, the method may still
132 /// fail to commute the operands, null pointer is returned in such cases.
133 virtual MachineInstr
*commuteInstructionImpl(MachineInstr
&MI
, bool NewMI
,
135 unsigned OpIdx2
) const;
137 /// Assigns the (CommutableOpIdx1, CommutableOpIdx2) pair of commutable
138 /// operand indices to (ResultIdx1, ResultIdx2).
139 /// One or both input values of the pair: (ResultIdx1, ResultIdx2) may be
140 /// predefined to some indices or be undefined (designated by the special
141 /// value 'CommuteAnyOperandIndex').
142 /// The predefined result indices cannot be re-defined.
143 /// The function returns true iff after the result pair redefinition
144 /// the fixed result pair is equal to or equivalent to the source pair of
145 /// indices: (CommutableOpIdx1, CommutableOpIdx2). It is assumed here that
146 /// the pairs (x,y) and (y,x) are equivalent.
147 static bool fixCommutedOpIndices(unsigned &ResultIdx1
, unsigned &ResultIdx2
,
148 unsigned CommutableOpIdx1
,
149 unsigned CommutableOpIdx2
);
152 /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is
153 /// set and the target hook isReallyTriviallyReMaterializable returns false,
154 /// this function does target-independent tests to determine if the
155 /// instruction is really trivially rematerializable.
156 bool isReallyTriviallyReMaterializableGeneric(const MachineInstr
&MI
,
157 AliasAnalysis
*AA
) const;
160 /// These methods return the opcode of the frame setup/destroy instructions
161 /// if they exist (-1 otherwise). Some targets use pseudo instructions in
162 /// order to abstract away the difference between operating with a frame
163 /// pointer and operating without, through the use of these two instructions.
165 unsigned getCallFrameSetupOpcode() const { return CallFrameSetupOpcode
; }
166 unsigned getCallFrameDestroyOpcode() const { return CallFrameDestroyOpcode
; }
168 /// Returns true if the argument is a frame pseudo instruction.
169 bool isFrameInstr(const MachineInstr
&I
) const {
170 return I
.getOpcode() == getCallFrameSetupOpcode() ||
171 I
.getOpcode() == getCallFrameDestroyOpcode();
174 /// Returns true if the argument is a frame setup pseudo instruction.
175 bool isFrameSetup(const MachineInstr
&I
) const {
176 return I
.getOpcode() == getCallFrameSetupOpcode();
179 /// Returns size of the frame associated with the given frame instruction.
180 /// For frame setup instruction this is frame that is set up space set up
181 /// after the instruction. For frame destroy instruction this is the frame
182 /// freed by the caller.
183 /// Note, in some cases a call frame (or a part of it) may be prepared prior
184 /// to the frame setup instruction. It occurs in the calls that involve
185 /// inalloca arguments. This function reports only the size of the frame part
186 /// that is set up between the frame setup and destroy pseudo instructions.
187 int64_t getFrameSize(const MachineInstr
&I
) const {
188 assert(isFrameInstr(I
) && "Not a frame instruction");
189 assert(I
.getOperand(0).getImm() >= 0);
190 return I
.getOperand(0).getImm();
193 /// Returns the total frame size, which is made up of the space set up inside
194 /// the pair of frame start-stop instructions and the space that is set up
195 /// prior to the pair.
196 int64_t getFrameTotalSize(const MachineInstr
&I
) const {
197 if (isFrameSetup(I
)) {
198 assert(I
.getOperand(1).getImm() >= 0 &&
199 "Frame size must not be negative");
200 return getFrameSize(I
) + I
.getOperand(1).getImm();
202 return getFrameSize(I
);
205 unsigned getCatchReturnOpcode() const { return CatchRetOpcode
; }
206 unsigned getReturnOpcode() const { return ReturnOpcode
; }
208 /// Returns the actual stack pointer adjustment made by an instruction
209 /// as part of a call sequence. By default, only call frame setup/destroy
210 /// instructions adjust the stack, but targets may want to override this
211 /// to enable more fine-grained adjustment, or adjust by a different value.
212 virtual int getSPAdjust(const MachineInstr
&MI
) const;
214 /// Return true if the instruction is a "coalescable" extension instruction.
215 /// That is, it's like a copy where it's legal for the source to overlap the
216 /// destination. e.g. X86::MOVSX64rr32. If this returns true, then it's
217 /// expected the pre-extension value is available as a subreg of the result
218 /// register. This also returns the sub-register index in SubIdx.
219 virtual bool isCoalescableExtInstr(const MachineInstr
&MI
, unsigned &SrcReg
,
220 unsigned &DstReg
, unsigned &SubIdx
) const {
224 /// If the specified machine instruction is a direct
225 /// load from a stack slot, return the virtual or physical register number of
226 /// the destination along with the FrameIndex of the loaded stack slot. If
227 /// not, return 0. This predicate must return 0 if the instruction has
228 /// any side effects other than loading from the stack slot.
229 virtual unsigned isLoadFromStackSlot(const MachineInstr
&MI
,
230 int &FrameIndex
) const {
234 /// Optional extension of isLoadFromStackSlot that returns the number of
235 /// bytes loaded from the stack. This must be implemented if a backend
236 /// supports partial stack slot spills/loads to further disambiguate
237 /// what the load does.
238 virtual unsigned isLoadFromStackSlot(const MachineInstr
&MI
,
240 unsigned &MemBytes
) const {
242 return isLoadFromStackSlot(MI
, FrameIndex
);
245 /// Check for post-frame ptr elimination stack locations as well.
246 /// This uses a heuristic so it isn't reliable for correctness.
247 virtual unsigned isLoadFromStackSlotPostFE(const MachineInstr
&MI
,
248 int &FrameIndex
) const {
252 /// If the specified machine instruction has a load from a stack slot,
253 /// return true along with the FrameIndices of the loaded stack slot and the
254 /// machine mem operands containing the reference.
255 /// If not, return false. Unlike isLoadFromStackSlot, this returns true for
256 /// any instructions that loads from the stack. This is just a hint, as some
257 /// cases may be missed.
258 virtual bool hasLoadFromStackSlot(
259 const MachineInstr
&MI
,
260 SmallVectorImpl
<const MachineMemOperand
*> &Accesses
) const;
262 /// If the specified machine instruction is a direct
263 /// store to a stack slot, return the virtual or physical register number of
264 /// the source reg along with the FrameIndex of the loaded stack slot. If
265 /// not, return 0. This predicate must return 0 if the instruction has
266 /// any side effects other than storing to the stack slot.
267 virtual unsigned isStoreToStackSlot(const MachineInstr
&MI
,
268 int &FrameIndex
) const {
272 /// Optional extension of isStoreToStackSlot that returns the number of
273 /// bytes stored to the stack. This must be implemented if a backend
274 /// supports partial stack slot spills/loads to further disambiguate
275 /// what the store does.
276 virtual unsigned isStoreToStackSlot(const MachineInstr
&MI
,
278 unsigned &MemBytes
) const {
280 return isStoreToStackSlot(MI
, FrameIndex
);
283 /// Check for post-frame ptr elimination stack locations as well.
284 /// This uses a heuristic, so it isn't reliable for correctness.
285 virtual unsigned isStoreToStackSlotPostFE(const MachineInstr
&MI
,
286 int &FrameIndex
) const {
290 /// If the specified machine instruction has a store to a stack slot,
291 /// return true along with the FrameIndices of the loaded stack slot and the
292 /// machine mem operands containing the reference.
293 /// If not, return false. Unlike isStoreToStackSlot,
294 /// this returns true for any instructions that stores to the
295 /// stack. This is just a hint, as some cases may be missed.
296 virtual bool hasStoreToStackSlot(
297 const MachineInstr
&MI
,
298 SmallVectorImpl
<const MachineMemOperand
*> &Accesses
) const;
300 /// Return true if the specified machine instruction
301 /// is a copy of one stack slot to another and has no other effect.
302 /// Provide the identity of the two frame indices.
303 virtual bool isStackSlotCopy(const MachineInstr
&MI
, int &DestFrameIndex
,
304 int &SrcFrameIndex
) const {
308 /// Compute the size in bytes and offset within a stack slot of a spilled
309 /// register or subregister.
311 /// \param [out] Size in bytes of the spilled value.
312 /// \param [out] Offset in bytes within the stack slot.
313 /// \returns true if both Size and Offset are successfully computed.
315 /// Not all subregisters have computable spill slots. For example,
316 /// subregisters registers may not be byte-sized, and a pair of discontiguous
317 /// subregisters has no single offset.
319 /// Targets with nontrivial bigendian implementations may need to override
320 /// this, particularly to support spilled vector registers.
321 virtual bool getStackSlotRange(const TargetRegisterClass
*RC
, unsigned SubIdx
,
322 unsigned &Size
, unsigned &Offset
,
323 const MachineFunction
&MF
) const;
325 /// Returns the size in bytes of the specified MachineInstr, or ~0U
326 /// when this function is not implemented by a target.
327 virtual unsigned getInstSizeInBytes(const MachineInstr
&MI
) const {
331 /// Return true if the instruction is as cheap as a move instruction.
333 /// Targets for different archs need to override this, and different
334 /// micro-architectures can also be finely tuned inside.
335 virtual bool isAsCheapAsAMove(const MachineInstr
&MI
) const {
336 return MI
.isAsCheapAsAMove();
339 /// Return true if the instruction should be sunk by MachineSink.
341 /// MachineSink determines on its own whether the instruction is safe to sink;
342 /// this gives the target a hook to override the default behavior with regards
343 /// to which instructions should be sunk.
344 virtual bool shouldSink(const MachineInstr
&MI
) const { return true; }
346 /// Re-issue the specified 'original' instruction at the
347 /// specific location targeting a new destination register.
348 /// The register in Orig->getOperand(0).getReg() will be substituted by
349 /// DestReg:SubIdx. Any existing subreg index is preserved or composed with
351 virtual void reMaterialize(MachineBasicBlock
&MBB
,
352 MachineBasicBlock::iterator MI
, unsigned DestReg
,
353 unsigned SubIdx
, const MachineInstr
&Orig
,
354 const TargetRegisterInfo
&TRI
) const;
356 /// Clones instruction or the whole instruction bundle \p Orig and
357 /// insert into \p MBB before \p InsertBefore. The target may update operands
358 /// that are required to be unique.
360 /// \p Orig must not return true for MachineInstr::isNotDuplicable().
361 virtual MachineInstr
&duplicate(MachineBasicBlock
&MBB
,
362 MachineBasicBlock::iterator InsertBefore
,
363 const MachineInstr
&Orig
) const;
365 /// This method must be implemented by targets that
366 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
367 /// may be able to convert a two-address instruction into one or more true
368 /// three-address instructions on demand. This allows the X86 target (for
369 /// example) to convert ADD and SHL instructions into LEA instructions if they
370 /// would require register copies due to two-addressness.
372 /// This method returns a null pointer if the transformation cannot be
373 /// performed, otherwise it returns the last new instruction.
375 virtual MachineInstr
*convertToThreeAddress(MachineFunction::iterator
&MFI
,
377 LiveVariables
*LV
) const {
381 // This constant can be used as an input value of operand index passed to
382 // the method findCommutedOpIndices() to tell the method that the
383 // corresponding operand index is not pre-defined and that the method
384 // can pick any commutable operand.
385 static const unsigned CommuteAnyOperandIndex
= ~0U;
387 /// This method commutes the operands of the given machine instruction MI.
389 /// The operands to be commuted are specified by their indices OpIdx1 and
390 /// OpIdx2. OpIdx1 and OpIdx2 arguments may be set to a special value
391 /// 'CommuteAnyOperandIndex', which means that the method is free to choose
392 /// any arbitrarily chosen commutable operand. If both arguments are set to
393 /// 'CommuteAnyOperandIndex' then the method looks for 2 different commutable
394 /// operands; then commutes them if such operands could be found.
396 /// If NewMI is false, MI is modified in place and returned; otherwise, a
397 /// new machine instruction is created and returned.
399 /// Do not call this method for a non-commutable instruction or
400 /// for non-commuable operands.
401 /// Even though the instruction is commutable, the method may still
402 /// fail to commute the operands, null pointer is returned in such cases.
404 commuteInstruction(MachineInstr
&MI
, bool NewMI
= false,
405 unsigned OpIdx1
= CommuteAnyOperandIndex
,
406 unsigned OpIdx2
= CommuteAnyOperandIndex
) const;
408 /// Returns true iff the routine could find two commutable operands in the
409 /// given machine instruction.
410 /// The 'SrcOpIdx1' and 'SrcOpIdx2' are INPUT and OUTPUT arguments.
411 /// If any of the INPUT values is set to the special value
412 /// 'CommuteAnyOperandIndex' then the method arbitrarily picks a commutable
413 /// operand, then returns its index in the corresponding argument.
414 /// If both of INPUT values are set to 'CommuteAnyOperandIndex' then method
415 /// looks for 2 commutable operands.
416 /// If INPUT values refer to some operands of MI, then the method simply
417 /// returns true if the corresponding operands are commutable and returns
420 /// For example, calling this method this way:
421 /// unsigned Op1 = 1, Op2 = CommuteAnyOperandIndex;
422 /// findCommutedOpIndices(MI, Op1, Op2);
423 /// can be interpreted as a query asking to find an operand that would be
424 /// commutable with the operand#1.
425 virtual bool findCommutedOpIndices(MachineInstr
&MI
, unsigned &SrcOpIdx1
,
426 unsigned &SrcOpIdx2
) const;
428 /// A pair composed of a register and a sub-register index.
429 /// Used to give some type checking when modeling Reg:SubReg.
430 struct RegSubRegPair
{
434 RegSubRegPair(unsigned Reg
= 0, unsigned SubReg
= 0)
435 : Reg(Reg
), SubReg(SubReg
) {}
437 bool operator==(const RegSubRegPair
& P
) const {
438 return Reg
== P
.Reg
&& SubReg
== P
.SubReg
;
440 bool operator!=(const RegSubRegPair
& P
) const {
441 return !(*this == P
);
445 /// A pair composed of a pair of a register and a sub-register index,
446 /// and another sub-register index.
447 /// Used to give some type checking when modeling Reg:SubReg1, SubReg2.
448 struct RegSubRegPairAndIdx
: RegSubRegPair
{
451 RegSubRegPairAndIdx(unsigned Reg
= 0, unsigned SubReg
= 0,
453 : RegSubRegPair(Reg
, SubReg
), SubIdx(SubIdx
) {}
456 /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI
458 /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of
459 /// the list is modeled as <Reg:SubReg, SubIdx>. Operands with the undef
460 /// flag are not added to this list.
461 /// E.g., REG_SEQUENCE %1:sub1, sub0, %2, sub1 would produce
466 /// \returns true if it is possible to build such an input sequence
467 /// with the pair \p MI, \p DefIdx. False otherwise.
469 /// \pre MI.isRegSequence() or MI.isRegSequenceLike().
471 /// \note The generic implementation does not provide any support for
472 /// MI.isRegSequenceLike(). In other words, one has to override
473 /// getRegSequenceLikeInputs for target specific instructions.
475 getRegSequenceInputs(const MachineInstr
&MI
, unsigned DefIdx
,
476 SmallVectorImpl
<RegSubRegPairAndIdx
> &InputRegs
) const;
478 /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI
480 /// \p [out] InputReg of the equivalent EXTRACT_SUBREG.
481 /// E.g., EXTRACT_SUBREG %1:sub1, sub0, sub1 would produce:
484 /// \returns true if it is possible to build such an input sequence
485 /// with the pair \p MI, \p DefIdx and the operand has no undef flag set.
488 /// \pre MI.isExtractSubreg() or MI.isExtractSubregLike().
490 /// \note The generic implementation does not provide any support for
491 /// MI.isExtractSubregLike(). In other words, one has to override
492 /// getExtractSubregLikeInputs for target specific instructions.
493 bool getExtractSubregInputs(const MachineInstr
&MI
, unsigned DefIdx
,
494 RegSubRegPairAndIdx
&InputReg
) const;
496 /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI
498 /// \p [out] BaseReg and \p [out] InsertedReg contain
499 /// the equivalent inputs of INSERT_SUBREG.
500 /// E.g., INSERT_SUBREG %0:sub0, %1:sub1, sub3 would produce:
501 /// - BaseReg: %0:sub0
502 /// - InsertedReg: %1:sub1, sub3
504 /// \returns true if it is possible to build such an input sequence
505 /// with the pair \p MI, \p DefIdx and the operand has no undef flag set.
508 /// \pre MI.isInsertSubreg() or MI.isInsertSubregLike().
510 /// \note The generic implementation does not provide any support for
511 /// MI.isInsertSubregLike(). In other words, one has to override
512 /// getInsertSubregLikeInputs for target specific instructions.
513 bool getInsertSubregInputs(const MachineInstr
&MI
, unsigned DefIdx
,
514 RegSubRegPair
&BaseReg
,
515 RegSubRegPairAndIdx
&InsertedReg
) const;
517 /// Return true if two machine instructions would produce identical values.
518 /// By default, this is only true when the two instructions
519 /// are deemed identical except for defs. If this function is called when the
520 /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for
521 /// aggressive checks.
522 virtual bool produceSameValue(const MachineInstr
&MI0
,
523 const MachineInstr
&MI1
,
524 const MachineRegisterInfo
*MRI
= nullptr) const;
526 /// \returns true if a branch from an instruction with opcode \p BranchOpc
527 /// bytes is capable of jumping to a position \p BrOffset bytes away.
528 virtual bool isBranchOffsetInRange(unsigned BranchOpc
,
529 int64_t BrOffset
) const {
530 llvm_unreachable("target did not implement");
533 /// \returns The block that branch instruction \p MI jumps to.
534 virtual MachineBasicBlock
*getBranchDestBlock(const MachineInstr
&MI
) const {
535 llvm_unreachable("target did not implement");
538 /// Insert an unconditional indirect branch at the end of \p MBB to \p
539 /// NewDestBB. \p BrOffset indicates the offset of \p NewDestBB relative to
540 /// the offset of the position to insert the new branch.
542 /// \returns The number of bytes added to the block.
543 virtual unsigned insertIndirectBranch(MachineBasicBlock
&MBB
,
544 MachineBasicBlock
&NewDestBB
,
546 int64_t BrOffset
= 0,
547 RegScavenger
*RS
= nullptr) const {
548 llvm_unreachable("target did not implement");
551 /// Analyze the branching code at the end of MBB, returning
552 /// true if it cannot be understood (e.g. it's a switch dispatch or isn't
553 /// implemented for a target). Upon success, this returns false and returns
554 /// with the following information in various cases:
556 /// 1. If this block ends with no branches (it just falls through to its succ)
557 /// just return false, leaving TBB/FBB null.
558 /// 2. If this block ends with only an unconditional branch, it sets TBB to be
559 /// the destination block.
560 /// 3. If this block ends with a conditional branch and it falls through to a
561 /// successor block, it sets TBB to be the branch destination block and a
562 /// list of operands that evaluate the condition. These operands can be
563 /// passed to other TargetInstrInfo methods to create new branches.
564 /// 4. If this block ends with a conditional branch followed by an
565 /// unconditional branch, it returns the 'true' destination in TBB, the
566 /// 'false' destination in FBB, and a list of operands that evaluate the
567 /// condition. These operands can be passed to other TargetInstrInfo
568 /// methods to create new branches.
570 /// Note that removeBranch and insertBranch must be implemented to support
571 /// cases where this method returns success.
573 /// If AllowModify is true, then this routine is allowed to modify the basic
574 /// block (e.g. delete instructions after the unconditional branch).
576 /// The CFG information in MBB.Predecessors and MBB.Successors must be valid
577 /// before calling this function.
578 virtual bool analyzeBranch(MachineBasicBlock
&MBB
, MachineBasicBlock
*&TBB
,
579 MachineBasicBlock
*&FBB
,
580 SmallVectorImpl
<MachineOperand
> &Cond
,
581 bool AllowModify
= false) const {
585 /// Represents a predicate at the MachineFunction level. The control flow a
586 /// MachineBranchPredicate represents is:
588 /// Reg = LHS `Predicate` RHS == ConditionDef
589 /// if Reg then goto TrueDest else goto FalseDest
591 struct MachineBranchPredicate
{
592 enum ComparePredicate
{
593 PRED_EQ
, // True if two values are equal
594 PRED_NE
, // True if two values are not equal
595 PRED_INVALID
// Sentinel value
598 ComparePredicate Predicate
= PRED_INVALID
;
599 MachineOperand LHS
= MachineOperand::CreateImm(0);
600 MachineOperand RHS
= MachineOperand::CreateImm(0);
601 MachineBasicBlock
*TrueDest
= nullptr;
602 MachineBasicBlock
*FalseDest
= nullptr;
603 MachineInstr
*ConditionDef
= nullptr;
605 /// SingleUseCondition is true if ConditionDef is dead except for the
606 /// branch(es) at the end of the basic block.
608 bool SingleUseCondition
= false;
610 explicit MachineBranchPredicate() = default;
613 /// Analyze the branching code at the end of MBB and parse it into the
614 /// MachineBranchPredicate structure if possible. Returns false on success
615 /// and true on failure.
617 /// If AllowModify is true, then this routine is allowed to modify the basic
618 /// block (e.g. delete instructions after the unconditional branch).
620 virtual bool analyzeBranchPredicate(MachineBasicBlock
&MBB
,
621 MachineBranchPredicate
&MBP
,
622 bool AllowModify
= false) const {
626 /// Remove the branching code at the end of the specific MBB.
627 /// This is only invoked in cases where AnalyzeBranch returns success. It
628 /// returns the number of instructions that were removed.
629 /// If \p BytesRemoved is non-null, report the change in code size from the
630 /// removed instructions.
631 virtual unsigned removeBranch(MachineBasicBlock
&MBB
,
632 int *BytesRemoved
= nullptr) const {
633 llvm_unreachable("Target didn't implement TargetInstrInfo::removeBranch!");
636 /// Insert branch code into the end of the specified MachineBasicBlock. The
637 /// operands to this method are the same as those returned by AnalyzeBranch.
638 /// This is only invoked in cases where AnalyzeBranch returns success. It
639 /// returns the number of instructions inserted. If \p BytesAdded is non-null,
640 /// report the change in code size from the added instructions.
642 /// It is also invoked by tail merging to add unconditional branches in
643 /// cases where AnalyzeBranch doesn't apply because there was no original
644 /// branch to analyze. At least this much must be implemented, else tail
645 /// merging needs to be disabled.
647 /// The CFG information in MBB.Predecessors and MBB.Successors must be valid
648 /// before calling this function.
649 virtual unsigned insertBranch(MachineBasicBlock
&MBB
, MachineBasicBlock
*TBB
,
650 MachineBasicBlock
*FBB
,
651 ArrayRef
<MachineOperand
> Cond
,
653 int *BytesAdded
= nullptr) const {
654 llvm_unreachable("Target didn't implement TargetInstrInfo::insertBranch!");
657 unsigned insertUnconditionalBranch(MachineBasicBlock
&MBB
,
658 MachineBasicBlock
*DestBB
,
660 int *BytesAdded
= nullptr) const {
661 return insertBranch(MBB
, DestBB
, nullptr, ArrayRef
<MachineOperand
>(), DL
,
665 /// Object returned by analyzeLoopForPipelining. Allows software pipelining
666 /// implementations to query attributes of the loop being pipelined and to
667 /// apply target-specific updates to the loop once pipelining is complete.
668 class PipelinerLoopInfo
{
670 virtual ~PipelinerLoopInfo();
671 /// Return true if the given instruction should not be pipelined and should
672 /// be ignored. An example could be a loop comparison, or induction variable
673 /// update with no users being pipelined.
674 virtual bool shouldIgnoreForPipelining(const MachineInstr
*MI
) const = 0;
676 /// Create a condition to determine if the trip count of the loop is greater
679 /// If the trip count is statically known to be greater than TC, return
680 /// true. If the trip count is statically known to be not greater than TC,
681 /// return false. Otherwise return nullopt and fill out Cond with the test
683 virtual Optional
<bool>
684 createTripCountGreaterCondition(int TC
, MachineBasicBlock
&MBB
,
685 SmallVectorImpl
<MachineOperand
> &Cond
) = 0;
687 /// Modify the loop such that the trip count is
688 /// OriginalTC + TripCountAdjust.
689 virtual void adjustTripCount(int TripCountAdjust
) = 0;
691 /// Called when the loop's preheader has been modified to NewPreheader.
692 virtual void setPreheader(MachineBasicBlock
*NewPreheader
) = 0;
694 /// Called when the loop is being removed. Any instructions in the preheader
695 /// should be removed.
697 /// Once this function is called, no other functions on this object are
698 /// valid; the loop has been removed.
699 virtual void disposed() = 0;
702 /// Analyze loop L, which must be a single-basic-block loop, and if the
703 /// conditions can be understood enough produce a PipelinerLoopInfo object.
704 virtual std::unique_ptr
<PipelinerLoopInfo
>
705 analyzeLoopForPipelining(MachineBasicBlock
*LoopBB
) const {
709 /// Analyze the loop code, return true if it cannot be understoo. Upon
710 /// success, this function returns false and returns information about the
711 /// induction variable and compare instruction used at the end.
712 virtual bool analyzeLoop(MachineLoop
&L
, MachineInstr
*&IndVarInst
,
713 MachineInstr
*&CmpInst
) const {
717 /// Generate code to reduce the loop iteration by one and check if the loop
718 /// is finished. Return the value/register of the new loop count. We need
719 /// this function when peeling off one or more iterations of a loop. This
720 /// function assumes the nth iteration is peeled first.
721 virtual unsigned reduceLoopCount(MachineBasicBlock
&MBB
,
722 MachineBasicBlock
&PreHeader
,
723 MachineInstr
*IndVar
, MachineInstr
&Cmp
,
724 SmallVectorImpl
<MachineOperand
> &Cond
,
725 SmallVectorImpl
<MachineInstr
*> &PrevInsts
,
726 unsigned Iter
, unsigned MaxIter
) const {
727 llvm_unreachable("Target didn't implement ReduceLoopCount");
730 /// Delete the instruction OldInst and everything after it, replacing it with
731 /// an unconditional branch to NewDest. This is used by the tail merging pass.
732 virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail
,
733 MachineBasicBlock
*NewDest
) const;
735 /// Return true if it's legal to split the given basic
736 /// block at the specified instruction (i.e. instruction would be the start
737 /// of a new basic block).
738 virtual bool isLegalToSplitMBBAt(MachineBasicBlock
&MBB
,
739 MachineBasicBlock::iterator MBBI
) const {
743 /// Return true if it's profitable to predicate
744 /// instructions with accumulated instruction latency of "NumCycles"
745 /// of the specified basic block, where the probability of the instructions
746 /// being executed is given by Probability, and Confidence is a measure
747 /// of our confidence that it will be properly predicted.
748 virtual bool isProfitableToIfCvt(MachineBasicBlock
&MBB
, unsigned NumCycles
,
749 unsigned ExtraPredCycles
,
750 BranchProbability Probability
) const {
754 /// Second variant of isProfitableToIfCvt. This one
755 /// checks for the case where two basic blocks from true and false path
756 /// of a if-then-else (diamond) are predicated on mutally exclusive
757 /// predicates, where the probability of the true path being taken is given
758 /// by Probability, and Confidence is a measure of our confidence that it
759 /// will be properly predicted.
760 virtual bool isProfitableToIfCvt(MachineBasicBlock
&TMBB
, unsigned NumTCycles
,
761 unsigned ExtraTCycles
,
762 MachineBasicBlock
&FMBB
, unsigned NumFCycles
,
763 unsigned ExtraFCycles
,
764 BranchProbability Probability
) const {
768 /// Return true if it's profitable for if-converter to duplicate instructions
769 /// of specified accumulated instruction latencies in the specified MBB to
770 /// enable if-conversion.
771 /// The probability of the instructions being executed is given by
772 /// Probability, and Confidence is a measure of our confidence that it
773 /// will be properly predicted.
774 virtual bool isProfitableToDupForIfCvt(MachineBasicBlock
&MBB
,
776 BranchProbability Probability
) const {
780 /// Return true if it's profitable to unpredicate
781 /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually
782 /// exclusive predicates.
790 /// This may be profitable is conditional instructions are always executed.
791 virtual bool isProfitableToUnpredicate(MachineBasicBlock
&TMBB
,
792 MachineBasicBlock
&FMBB
) const {
796 /// Return true if it is possible to insert a select
797 /// instruction that chooses between TrueReg and FalseReg based on the
798 /// condition code in Cond.
800 /// When successful, also return the latency in cycles from TrueReg,
801 /// FalseReg, and Cond to the destination register. In most cases, a select
802 /// instruction will be 1 cycle, so CondCycles = TrueCycles = FalseCycles = 1
804 /// Some x86 implementations have 2-cycle cmov instructions.
806 /// @param MBB Block where select instruction would be inserted.
807 /// @param Cond Condition returned by AnalyzeBranch.
808 /// @param TrueReg Virtual register to select when Cond is true.
809 /// @param FalseReg Virtual register to select when Cond is false.
810 /// @param CondCycles Latency from Cond+Branch to select output.
811 /// @param TrueCycles Latency from TrueReg to select output.
812 /// @param FalseCycles Latency from FalseReg to select output.
813 virtual bool canInsertSelect(const MachineBasicBlock
&MBB
,
814 ArrayRef
<MachineOperand
> Cond
, unsigned TrueReg
,
815 unsigned FalseReg
, int &CondCycles
,
816 int &TrueCycles
, int &FalseCycles
) const {
820 /// Insert a select instruction into MBB before I that will copy TrueReg to
821 /// DstReg when Cond is true, and FalseReg to DstReg when Cond is false.
823 /// This function can only be called after canInsertSelect() returned true.
824 /// The condition in Cond comes from AnalyzeBranch, and it can be assumed
825 /// that the same flags or registers required by Cond are available at the
828 /// @param MBB Block where select instruction should be inserted.
829 /// @param I Insertion point.
830 /// @param DL Source location for debugging.
831 /// @param DstReg Virtual register to be defined by select instruction.
832 /// @param Cond Condition as computed by AnalyzeBranch.
833 /// @param TrueReg Virtual register to copy when Cond is true.
834 /// @param FalseReg Virtual register to copy when Cons is false.
835 virtual void insertSelect(MachineBasicBlock
&MBB
,
836 MachineBasicBlock::iterator I
, const DebugLoc
&DL
,
837 unsigned DstReg
, ArrayRef
<MachineOperand
> Cond
,
838 unsigned TrueReg
, unsigned FalseReg
) const {
839 llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!");
842 /// Analyze the given select instruction, returning true if
843 /// it cannot be understood. It is assumed that MI->isSelect() is true.
845 /// When successful, return the controlling condition and the operands that
846 /// determine the true and false result values.
848 /// Result = SELECT Cond, TrueOp, FalseOp
850 /// Some targets can optimize select instructions, for example by predicating
851 /// the instruction defining one of the operands. Such targets should set
854 /// @param MI Select instruction to analyze.
855 /// @param Cond Condition controlling the select.
856 /// @param TrueOp Operand number of the value selected when Cond is true.
857 /// @param FalseOp Operand number of the value selected when Cond is false.
858 /// @param Optimizable Returned as true if MI is optimizable.
859 /// @returns False on success.
860 virtual bool analyzeSelect(const MachineInstr
&MI
,
861 SmallVectorImpl
<MachineOperand
> &Cond
,
862 unsigned &TrueOp
, unsigned &FalseOp
,
863 bool &Optimizable
) const {
864 assert(MI
.getDesc().isSelect() && "MI must be a select instruction");
868 /// Given a select instruction that was understood by
869 /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by
870 /// merging it with one of its operands. Returns NULL on failure.
872 /// When successful, returns the new select instruction. The client is
873 /// responsible for deleting MI.
875 /// If both sides of the select can be optimized, PreferFalse is used to pick
878 /// @param MI Optimizable select instruction.
879 /// @param NewMIs Set that record all MIs in the basic block up to \p
880 /// MI. Has to be updated with any newly created MI or deleted ones.
881 /// @param PreferFalse Try to optimize FalseOp instead of TrueOp.
882 /// @returns Optimized instruction or NULL.
883 virtual MachineInstr
*optimizeSelect(MachineInstr
&MI
,
884 SmallPtrSetImpl
<MachineInstr
*> &NewMIs
,
885 bool PreferFalse
= false) const {
886 // This function must be implemented if Optimizable is ever set.
887 llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!");
890 /// Emit instructions to copy a pair of physical registers.
892 /// This function should support copies within any legal register class as
893 /// well as any cross-class copies created during instruction selection.
895 /// The source and destination registers may overlap, which may require a
896 /// careful implementation when multiple copy instructions are required for
897 /// large registers. See for example the ARM target.
898 virtual void copyPhysReg(MachineBasicBlock
&MBB
,
899 MachineBasicBlock::iterator MI
, const DebugLoc
&DL
,
900 unsigned DestReg
, unsigned SrcReg
,
901 bool KillSrc
) const {
902 llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!");
906 /// Target-dependent implemenation for IsCopyInstr.
907 /// If the specific machine instruction is a instruction that moves/copies
908 /// value from one register to another register return true along with
909 /// @Source machine operand and @Destination machine operand.
910 virtual bool isCopyInstrImpl(const MachineInstr
&MI
,
911 const MachineOperand
*&Source
,
912 const MachineOperand
*&Destination
) const {
917 /// If the specific machine instruction is a instruction that moves/copies
918 /// value from one register to another register return true along with
919 /// @Source machine operand and @Destination machine operand.
920 /// For COPY-instruction the method naturally returns true, for all other
921 /// instructions the method calls target-dependent implementation.
922 bool isCopyInstr(const MachineInstr
&MI
, const MachineOperand
*&Source
,
923 const MachineOperand
*&Destination
) const {
925 Destination
= &MI
.getOperand(0);
926 Source
= &MI
.getOperand(1);
929 return isCopyInstrImpl(MI
, Source
, Destination
);
932 /// Store the specified register of the given register class to the specified
933 /// stack frame index. The store instruction is to be added to the given
934 /// machine basic block before the specified machine instruction. If isKill
935 /// is true, the register operand is the last use and must be marked kill.
936 virtual void storeRegToStackSlot(MachineBasicBlock
&MBB
,
937 MachineBasicBlock::iterator MI
,
938 unsigned SrcReg
, bool isKill
, int FrameIndex
,
939 const TargetRegisterClass
*RC
,
940 const TargetRegisterInfo
*TRI
) const {
941 llvm_unreachable("Target didn't implement "
942 "TargetInstrInfo::storeRegToStackSlot!");
945 /// Load the specified register of the given register class from the specified
946 /// stack frame index. The load instruction is to be added to the given
947 /// machine basic block before the specified machine instruction.
948 virtual void loadRegFromStackSlot(MachineBasicBlock
&MBB
,
949 MachineBasicBlock::iterator MI
,
950 unsigned DestReg
, int FrameIndex
,
951 const TargetRegisterClass
*RC
,
952 const TargetRegisterInfo
*TRI
) const {
953 llvm_unreachable("Target didn't implement "
954 "TargetInstrInfo::loadRegFromStackSlot!");
957 /// This function is called for all pseudo instructions
958 /// that remain after register allocation. Many pseudo instructions are
959 /// created to help register allocation. This is the place to convert them
960 /// into real instructions. The target can edit MI in place, or it can insert
961 /// new instructions and erase MI. The function should return true if
962 /// anything was changed.
963 virtual bool expandPostRAPseudo(MachineInstr
&MI
) const { return false; }
965 /// Check whether the target can fold a load that feeds a subreg operand
966 /// (or a subreg operand that feeds a store).
967 /// For example, X86 may want to return true if it can fold
968 /// movl (%esp), %eax
973 /// Ideally, we'd like the target implementation of foldMemoryOperand() to
974 /// reject subregs - but since this behavior used to be enforced in the
975 /// target-independent code, moving this responsibility to the targets
976 /// has the potential of causing nasty silent breakage in out-of-tree targets.
977 virtual bool isSubregFoldable() const { return false; }
979 /// Attempt to fold a load or store of the specified stack
980 /// slot into the specified machine instruction for the specified operand(s).
981 /// If this is possible, a new instruction is returned with the specified
982 /// operand folded, otherwise NULL is returned.
983 /// The new instruction is inserted before MI, and the client is responsible
984 /// for removing the old instruction.
985 /// If VRM is passed, the assigned physregs can be inspected by target to
986 /// decide on using an opcode (note that those assignments can still change).
987 MachineInstr
*foldMemoryOperand(MachineInstr
&MI
, ArrayRef
<unsigned> Ops
,
989 LiveIntervals
*LIS
= nullptr,
990 VirtRegMap
*VRM
= nullptr) const;
992 /// Same as the previous version except it allows folding of any load and
993 /// store from / to any address, not just from a specific stack slot.
994 MachineInstr
*foldMemoryOperand(MachineInstr
&MI
, ArrayRef
<unsigned> Ops
,
995 MachineInstr
&LoadMI
,
996 LiveIntervals
*LIS
= nullptr) const;
998 /// Return true when there is potentially a faster code sequence
999 /// for an instruction chain ending in \p Root. All potential patterns are
1000 /// returned in the \p Pattern vector. Pattern should be sorted in priority
1001 /// order since the pattern evaluator stops checking as soon as it finds a
1002 /// faster sequence.
1003 /// \param Root - Instruction that could be combined with one of its operands
1004 /// \param Patterns - Vector of possible combination patterns
1005 virtual bool getMachineCombinerPatterns(
1007 SmallVectorImpl
<MachineCombinerPattern
> &Patterns
) const;
1009 /// Return true when a code sequence can improve throughput. It
1010 /// should be called only for instructions in loops.
1011 /// \param Pattern - combiner pattern
1012 virtual bool isThroughputPattern(MachineCombinerPattern Pattern
) const;
1014 /// Return true if the input \P Inst is part of a chain of dependent ops
1015 /// that are suitable for reassociation, otherwise return false.
1016 /// If the instruction's operands must be commuted to have a previous
1017 /// instruction of the same type define the first source operand, \P Commuted
1018 /// will be set to true.
1019 bool isReassociationCandidate(const MachineInstr
&Inst
, bool &Commuted
) const;
1021 /// Return true when \P Inst is both associative and commutative.
1022 virtual bool isAssociativeAndCommutative(const MachineInstr
&Inst
) const {
1026 /// Return true when \P Inst has reassociable operands in the same \P MBB.
1027 virtual bool hasReassociableOperands(const MachineInstr
&Inst
,
1028 const MachineBasicBlock
*MBB
) const;
1030 /// Return true when \P Inst has reassociable sibling.
1031 bool hasReassociableSibling(const MachineInstr
&Inst
, bool &Commuted
) const;
1033 /// When getMachineCombinerPatterns() finds patterns, this function generates
1034 /// the instructions that could replace the original code sequence. The client
1035 /// has to decide whether the actual replacement is beneficial or not.
1036 /// \param Root - Instruction that could be combined with one of its operands
1037 /// \param Pattern - Combination pattern for Root
1038 /// \param InsInstrs - Vector of new instructions that implement P
1039 /// \param DelInstrs - Old instructions, including Root, that could be
1040 /// replaced by InsInstr
1041 /// \param InstIdxForVirtReg - map of virtual register to instruction in
1042 /// InsInstr that defines it
1043 virtual void genAlternativeCodeSequence(
1044 MachineInstr
&Root
, MachineCombinerPattern Pattern
,
1045 SmallVectorImpl
<MachineInstr
*> &InsInstrs
,
1046 SmallVectorImpl
<MachineInstr
*> &DelInstrs
,
1047 DenseMap
<unsigned, unsigned> &InstIdxForVirtReg
) const;
1049 /// Attempt to reassociate \P Root and \P Prev according to \P Pattern to
1050 /// reduce critical path length.
1051 void reassociateOps(MachineInstr
&Root
, MachineInstr
&Prev
,
1052 MachineCombinerPattern Pattern
,
1053 SmallVectorImpl
<MachineInstr
*> &InsInstrs
,
1054 SmallVectorImpl
<MachineInstr
*> &DelInstrs
,
1055 DenseMap
<unsigned, unsigned> &InstrIdxForVirtReg
) const;
1057 /// This is an architecture-specific helper function of reassociateOps.
1058 /// Set special operand attributes for new instructions after reassociation.
1059 virtual void setSpecialOperandAttr(MachineInstr
&OldMI1
, MachineInstr
&OldMI2
,
1060 MachineInstr
&NewMI1
,
1061 MachineInstr
&NewMI2
) const {}
1063 /// Return true when a target supports MachineCombiner.
1064 virtual bool useMachineCombiner() const { return false; }
1066 /// Return true if the given SDNode can be copied during scheduling
1067 /// even if it has glue.
1068 virtual bool canCopyGluedNodeDuringSchedule(SDNode
*N
) const { return false; }
1071 /// Target-dependent implementation for foldMemoryOperand.
1072 /// Target-independent code in foldMemoryOperand will
1073 /// take care of adding a MachineMemOperand to the newly created instruction.
1074 /// The instruction and any auxiliary instructions necessary will be inserted
1076 virtual MachineInstr
*
1077 foldMemoryOperandImpl(MachineFunction
&MF
, MachineInstr
&MI
,
1078 ArrayRef
<unsigned> Ops
,
1079 MachineBasicBlock::iterator InsertPt
, int FrameIndex
,
1080 LiveIntervals
*LIS
= nullptr,
1081 VirtRegMap
*VRM
= nullptr) const {
1085 /// Target-dependent implementation for foldMemoryOperand.
1086 /// Target-independent code in foldMemoryOperand will
1087 /// take care of adding a MachineMemOperand to the newly created instruction.
1088 /// The instruction and any auxiliary instructions necessary will be inserted
1090 virtual MachineInstr
*foldMemoryOperandImpl(
1091 MachineFunction
&MF
, MachineInstr
&MI
, ArrayRef
<unsigned> Ops
,
1092 MachineBasicBlock::iterator InsertPt
, MachineInstr
&LoadMI
,
1093 LiveIntervals
*LIS
= nullptr) const {
1097 /// Target-dependent implementation of getRegSequenceInputs.
1099 /// \returns true if it is possible to build the equivalent
1100 /// REG_SEQUENCE inputs with the pair \p MI, \p DefIdx. False otherwise.
1102 /// \pre MI.isRegSequenceLike().
1104 /// \see TargetInstrInfo::getRegSequenceInputs.
1105 virtual bool getRegSequenceLikeInputs(
1106 const MachineInstr
&MI
, unsigned DefIdx
,
1107 SmallVectorImpl
<RegSubRegPairAndIdx
> &InputRegs
) const {
1111 /// Target-dependent implementation of getExtractSubregInputs.
1113 /// \returns true if it is possible to build the equivalent
1114 /// EXTRACT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise.
1116 /// \pre MI.isExtractSubregLike().
1118 /// \see TargetInstrInfo::getExtractSubregInputs.
1119 virtual bool getExtractSubregLikeInputs(const MachineInstr
&MI
,
1121 RegSubRegPairAndIdx
&InputReg
) const {
1125 /// Target-dependent implementation of getInsertSubregInputs.
1127 /// \returns true if it is possible to build the equivalent
1128 /// INSERT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise.
1130 /// \pre MI.isInsertSubregLike().
1132 /// \see TargetInstrInfo::getInsertSubregInputs.
1134 getInsertSubregLikeInputs(const MachineInstr
&MI
, unsigned DefIdx
,
1135 RegSubRegPair
&BaseReg
,
1136 RegSubRegPairAndIdx
&InsertedReg
) const {
1141 /// getAddressSpaceForPseudoSourceKind - Given the kind of memory
1142 /// (e.g. stack) the target returns the corresponding address space.
1144 getAddressSpaceForPseudoSourceKind(unsigned Kind
) const {
1148 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
1149 /// a store or a load and a store into two or more instruction. If this is
1150 /// possible, returns true as well as the new instructions by reference.
1152 unfoldMemoryOperand(MachineFunction
&MF
, MachineInstr
&MI
, unsigned Reg
,
1153 bool UnfoldLoad
, bool UnfoldStore
,
1154 SmallVectorImpl
<MachineInstr
*> &NewMIs
) const {
1158 virtual bool unfoldMemoryOperand(SelectionDAG
&DAG
, SDNode
*N
,
1159 SmallVectorImpl
<SDNode
*> &NewNodes
) const {
1163 /// Returns the opcode of the would be new
1164 /// instruction after load / store are unfolded from an instruction of the
1165 /// specified opcode. It returns zero if the specified unfolding is not
1166 /// possible. If LoadRegIndex is non-null, it is filled in with the operand
1167 /// index of the operand which will hold the register holding the loaded
1170 getOpcodeAfterMemoryUnfold(unsigned Opc
, bool UnfoldLoad
, bool UnfoldStore
,
1171 unsigned *LoadRegIndex
= nullptr) const {
1175 /// This is used by the pre-regalloc scheduler to determine if two loads are
1176 /// loading from the same base address. It should only return true if the base
1177 /// pointers are the same and the only differences between the two addresses
1178 /// are the offset. It also returns the offsets by reference.
1179 virtual bool areLoadsFromSameBasePtr(SDNode
*Load1
, SDNode
*Load2
,
1181 int64_t &Offset2
) const {
1185 /// This is a used by the pre-regalloc scheduler to determine (in conjunction
1186 /// with areLoadsFromSameBasePtr) if two loads should be scheduled together.
1187 /// On some targets if two loads are loading from
1188 /// addresses in the same cache line, it's better if they are scheduled
1189 /// together. This function takes two integers that represent the load offsets
1190 /// from the common base address. It returns true if it decides it's desirable
1191 /// to schedule the two loads together. "NumLoads" is the number of loads that
1192 /// have already been scheduled after Load1.
1193 virtual bool shouldScheduleLoadsNear(SDNode
*Load1
, SDNode
*Load2
,
1194 int64_t Offset1
, int64_t Offset2
,
1195 unsigned NumLoads
) const {
1199 /// Get the base operand and byte offset of an instruction that reads/writes
1201 virtual bool getMemOperandWithOffset(const MachineInstr
&MI
,
1202 const MachineOperand
*&BaseOp
,
1204 const TargetRegisterInfo
*TRI
) const {
1208 /// Return true if the instruction contains a base register and offset. If
1209 /// true, the function also sets the operand position in the instruction
1210 /// for the base register and offset.
1211 virtual bool getBaseAndOffsetPosition(const MachineInstr
&MI
,
1213 unsigned &OffsetPos
) const {
1217 /// If the instruction is an increment of a constant value, return the amount.
1218 virtual bool getIncrementValue(const MachineInstr
&MI
, int &Value
) const {
1222 /// Returns true if the two given memory operations should be scheduled
1223 /// adjacent. Note that you have to add:
1224 /// DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
1226 /// DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
1227 /// to TargetPassConfig::createMachineScheduler() to have an effect.
1228 virtual bool shouldClusterMemOps(const MachineOperand
&BaseOp1
,
1229 const MachineOperand
&BaseOp2
,
1230 unsigned NumLoads
) const {
1231 llvm_unreachable("target did not implement shouldClusterMemOps()");
1234 /// Reverses the branch condition of the specified condition list,
1235 /// returning false on success and true if it cannot be reversed.
1237 reverseBranchCondition(SmallVectorImpl
<MachineOperand
> &Cond
) const {
1241 /// Insert a noop into the instruction stream at the specified point.
1242 virtual void insertNoop(MachineBasicBlock
&MBB
,
1243 MachineBasicBlock::iterator MI
) const;
1245 /// Return the noop instruction to use for a noop.
1246 virtual void getNoop(MCInst
&NopInst
) const;
1248 /// Return true for post-incremented instructions.
1249 virtual bool isPostIncrement(const MachineInstr
&MI
) const { return false; }
1251 /// Returns true if the instruction is already predicated.
1252 virtual bool isPredicated(const MachineInstr
&MI
) const { return false; }
1254 /// Returns true if the instruction is a
1255 /// terminator instruction that has not been predicated.
1256 virtual bool isUnpredicatedTerminator(const MachineInstr
&MI
) const;
1258 /// Returns true if MI is an unconditional tail call.
1259 virtual bool isUnconditionalTailCall(const MachineInstr
&MI
) const {
1263 /// Returns true if the tail call can be made conditional on BranchCond.
1264 virtual bool canMakeTailCallConditional(SmallVectorImpl
<MachineOperand
> &Cond
,
1265 const MachineInstr
&TailCall
) const {
1269 /// Replace the conditional branch in MBB with a conditional tail call.
1270 virtual void replaceBranchWithTailCall(MachineBasicBlock
&MBB
,
1271 SmallVectorImpl
<MachineOperand
> &Cond
,
1272 const MachineInstr
&TailCall
) const {
1273 llvm_unreachable("Target didn't implement replaceBranchWithTailCall!");
1276 /// Convert the instruction into a predicated instruction.
1277 /// It returns true if the operation was successful.
1278 virtual bool PredicateInstruction(MachineInstr
&MI
,
1279 ArrayRef
<MachineOperand
> Pred
) const;
1281 /// Returns true if the first specified predicate
1282 /// subsumes the second, e.g. GE subsumes GT.
1283 virtual bool SubsumesPredicate(ArrayRef
<MachineOperand
> Pred1
,
1284 ArrayRef
<MachineOperand
> Pred2
) const {
1288 /// If the specified instruction defines any predicate
1289 /// or condition code register(s) used for predication, returns true as well
1290 /// as the definition predicate(s) by reference.
1291 virtual bool DefinesPredicate(MachineInstr
&MI
,
1292 std::vector
<MachineOperand
> &Pred
) const {
1296 /// Return true if the specified instruction can be predicated.
1297 /// By default, this returns true for every instruction with a
1298 /// PredicateOperand.
1299 virtual bool isPredicable(const MachineInstr
&MI
) const {
1300 return MI
.getDesc().isPredicable();
1303 /// Return true if it's safe to move a machine
1304 /// instruction that defines the specified register class.
1305 virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass
*RC
) const {
1309 /// Test if the given instruction should be considered a scheduling boundary.
1310 /// This primarily includes labels and terminators.
1311 virtual bool isSchedulingBoundary(const MachineInstr
&MI
,
1312 const MachineBasicBlock
*MBB
,
1313 const MachineFunction
&MF
) const;
1315 /// Measure the specified inline asm to determine an approximation of its
1317 virtual unsigned getInlineAsmLength(
1318 const char *Str
, const MCAsmInfo
&MAI
,
1319 const TargetSubtargetInfo
*STI
= nullptr) const;
1321 /// Allocate and return a hazard recognizer to use for this target when
1322 /// scheduling the machine instructions before register allocation.
1323 virtual ScheduleHazardRecognizer
*
1324 CreateTargetHazardRecognizer(const TargetSubtargetInfo
*STI
,
1325 const ScheduleDAG
*DAG
) const;
1327 /// Allocate and return a hazard recognizer to use for this target when
1328 /// scheduling the machine instructions before register allocation.
1329 virtual ScheduleHazardRecognizer
*
1330 CreateTargetMIHazardRecognizer(const InstrItineraryData
*,
1331 const ScheduleDAG
*DAG
) const;
1333 /// Allocate and return a hazard recognizer to use for this target when
1334 /// scheduling the machine instructions after register allocation.
1335 virtual ScheduleHazardRecognizer
*
1336 CreateTargetPostRAHazardRecognizer(const InstrItineraryData
*,
1337 const ScheduleDAG
*DAG
) const;
1339 /// Allocate and return a hazard recognizer to use for by non-scheduling
1341 virtual ScheduleHazardRecognizer
*
1342 CreateTargetPostRAHazardRecognizer(const MachineFunction
&MF
) const {
1346 /// Provide a global flag for disabling the PreRA hazard recognizer that
1347 /// targets may choose to honor.
1348 bool usePreRAHazardRecognizer() const;
1350 /// For a comparison instruction, return the source registers
1351 /// in SrcReg and SrcReg2 if having two register operands, and the value it
1352 /// compares against in CmpValue. Return true if the comparison instruction
1353 /// can be analyzed.
1354 virtual bool analyzeCompare(const MachineInstr
&MI
, unsigned &SrcReg
,
1355 unsigned &SrcReg2
, int &Mask
, int &Value
) const {
1359 /// See if the comparison instruction can be converted
1360 /// into something more efficient. E.g., on ARM most instructions can set the
1361 /// flags register, obviating the need for a separate CMP.
1362 virtual bool optimizeCompareInstr(MachineInstr
&CmpInstr
, unsigned SrcReg
,
1363 unsigned SrcReg2
, int Mask
, int Value
,
1364 const MachineRegisterInfo
*MRI
) const {
1367 virtual bool optimizeCondBranch(MachineInstr
&MI
) const { return false; }
1369 /// Try to remove the load by folding it to a register operand at the use.
1370 /// We fold the load instructions if and only if the
1371 /// def and use are in the same BB. We only look at one load and see
1372 /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
1373 /// defined by the load we are trying to fold. DefMI returns the machine
1374 /// instruction that defines FoldAsLoadDefReg, and the function returns
1375 /// the machine instruction generated due to folding.
1376 virtual MachineInstr
*optimizeLoadInstr(MachineInstr
&MI
,
1377 const MachineRegisterInfo
*MRI
,
1378 unsigned &FoldAsLoadDefReg
,
1379 MachineInstr
*&DefMI
) const {
1383 /// 'Reg' is known to be defined by a move immediate instruction,
1384 /// try to fold the immediate into the use instruction.
1385 /// If MRI->hasOneNonDBGUse(Reg) is true, and this function returns true,
1386 /// then the caller may assume that DefMI has been erased from its parent
1387 /// block. The caller may assume that it will not be erased by this
1388 /// function otherwise.
1389 virtual bool FoldImmediate(MachineInstr
&UseMI
, MachineInstr
&DefMI
,
1390 unsigned Reg
, MachineRegisterInfo
*MRI
) const {
1394 /// Return the number of u-operations the given machine
1395 /// instruction will be decoded to on the target cpu. The itinerary's
1396 /// IssueWidth is the number of microops that can be dispatched each
1397 /// cycle. An instruction with zero microops takes no dispatch resources.
1398 virtual unsigned getNumMicroOps(const InstrItineraryData
*ItinData
,
1399 const MachineInstr
&MI
) const;
1401 /// Return true for pseudo instructions that don't consume any
1402 /// machine resources in their current form. These are common cases that the
1403 /// scheduler should consider free, rather than conservatively handling them
1404 /// as instructions with no itinerary.
1405 bool isZeroCost(unsigned Opcode
) const {
1406 return Opcode
<= TargetOpcode::COPY
;
1409 virtual int getOperandLatency(const InstrItineraryData
*ItinData
,
1410 SDNode
*DefNode
, unsigned DefIdx
,
1411 SDNode
*UseNode
, unsigned UseIdx
) const;
1413 /// Compute and return the use operand latency of a given pair of def and use.
1414 /// In most cases, the static scheduling itinerary was enough to determine the
1415 /// operand latency. But it may not be possible for instructions with variable
1416 /// number of defs / uses.
1418 /// This is a raw interface to the itinerary that may be directly overridden
1419 /// by a target. Use computeOperandLatency to get the best estimate of
1421 virtual int getOperandLatency(const InstrItineraryData
*ItinData
,
1422 const MachineInstr
&DefMI
, unsigned DefIdx
,
1423 const MachineInstr
&UseMI
,
1424 unsigned UseIdx
) const;
1426 /// Compute the instruction latency of a given instruction.
1427 /// If the instruction has higher cost when predicated, it's returned via
1429 virtual unsigned getInstrLatency(const InstrItineraryData
*ItinData
,
1430 const MachineInstr
&MI
,
1431 unsigned *PredCost
= nullptr) const;
1433 virtual unsigned getPredicationCost(const MachineInstr
&MI
) const;
1435 virtual int getInstrLatency(const InstrItineraryData
*ItinData
,
1436 SDNode
*Node
) const;
1438 /// Return the default expected latency for a def based on its opcode.
1439 unsigned defaultDefLatency(const MCSchedModel
&SchedModel
,
1440 const MachineInstr
&DefMI
) const;
1442 int computeDefOperandLatency(const InstrItineraryData
*ItinData
,
1443 const MachineInstr
&DefMI
) const;
1445 /// Return true if this opcode has high latency to its result.
1446 virtual bool isHighLatencyDef(int opc
) const { return false; }
1448 /// Compute operand latency between a def of 'Reg'
1449 /// and a use in the current loop. Return true if the target considered
1450 /// it 'high'. This is used by optimization passes such as machine LICM to
1451 /// determine whether it makes sense to hoist an instruction out even in a
1452 /// high register pressure situation.
1453 virtual bool hasHighOperandLatency(const TargetSchedModel
&SchedModel
,
1454 const MachineRegisterInfo
*MRI
,
1455 const MachineInstr
&DefMI
, unsigned DefIdx
,
1456 const MachineInstr
&UseMI
,
1457 unsigned UseIdx
) const {
1461 /// Compute operand latency of a def of 'Reg'. Return true
1462 /// if the target considered it 'low'.
1463 virtual bool hasLowDefLatency(const TargetSchedModel
&SchedModel
,
1464 const MachineInstr
&DefMI
,
1465 unsigned DefIdx
) const;
1467 /// Perform target-specific instruction verification.
1468 virtual bool verifyInstruction(const MachineInstr
&MI
,
1469 StringRef
&ErrInfo
) const {
1473 /// Return the current execution domain and bit mask of
1474 /// possible domains for instruction.
1476 /// Some micro-architectures have multiple execution domains, and multiple
1477 /// opcodes that perform the same operation in different domains. For
1478 /// example, the x86 architecture provides the por, orps, and orpd
1479 /// instructions that all do the same thing. There is a latency penalty if a
1480 /// register is written in one domain and read in another.
1482 /// This function returns a pair (domain, mask) containing the execution
1483 /// domain of MI, and a bit mask of possible domains. The setExecutionDomain
1484 /// function can be used to change the opcode to one of the domains in the
1485 /// bit mask. Instructions whose execution domain can't be changed should
1486 /// return a 0 mask.
1488 /// The execution domain numbers don't have any special meaning except domain
1489 /// 0 is used for instructions that are not associated with any interesting
1490 /// execution domain.
1492 virtual std::pair
<uint16_t, uint16_t>
1493 getExecutionDomain(const MachineInstr
&MI
) const {
1494 return std::make_pair(0, 0);
1497 /// Change the opcode of MI to execute in Domain.
1499 /// The bit (1 << Domain) must be set in the mask returned from
1500 /// getExecutionDomain(MI).
1501 virtual void setExecutionDomain(MachineInstr
&MI
, unsigned Domain
) const {}
1503 /// Returns the preferred minimum clearance
1504 /// before an instruction with an unwanted partial register update.
1506 /// Some instructions only write part of a register, and implicitly need to
1507 /// read the other parts of the register. This may cause unwanted stalls
1508 /// preventing otherwise unrelated instructions from executing in parallel in
1509 /// an out-of-order CPU.
1511 /// For example, the x86 instruction cvtsi2ss writes its result to bits
1512 /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so
1513 /// the instruction needs to wait for the old value of the register to become
1516 /// addps %xmm1, %xmm0
1517 /// movaps %xmm0, (%rax)
1518 /// cvtsi2ss %rbx, %xmm0
1520 /// In the code above, the cvtsi2ss instruction needs to wait for the addps
1521 /// instruction before it can issue, even though the high bits of %xmm0
1522 /// probably aren't needed.
1524 /// This hook returns the preferred clearance before MI, measured in
1525 /// instructions. Other defs of MI's operand OpNum are avoided in the last N
1526 /// instructions before MI. It should only return a positive value for
1527 /// unwanted dependencies. If the old bits of the defined register have
1528 /// useful values, or if MI is determined to otherwise read the dependency,
1529 /// the hook should return 0.
1531 /// The unwanted dependency may be handled by:
1533 /// 1. Allocating the same register for an MI def and use. That makes the
1534 /// unwanted dependency identical to a required dependency.
1536 /// 2. Allocating a register for the def that has no defs in the previous N
1539 /// 3. Calling breakPartialRegDependency() with the same arguments. This
1540 /// allows the target to insert a dependency breaking instruction.
1543 getPartialRegUpdateClearance(const MachineInstr
&MI
, unsigned OpNum
,
1544 const TargetRegisterInfo
*TRI
) const {
1545 // The default implementation returns 0 for no partial register dependency.
1549 /// Return the minimum clearance before an instruction that reads an
1550 /// unused register.
1552 /// For example, AVX instructions may copy part of a register operand into
1553 /// the unused high bits of the destination register.
1555 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14
1557 /// In the code above, vcvtsi2sdq copies %xmm0[127:64] into %xmm14 creating a
1558 /// false dependence on any previous write to %xmm0.
1560 /// This hook works similarly to getPartialRegUpdateClearance, except that it
1561 /// does not take an operand index. Instead sets \p OpNum to the index of the
1562 /// unused register.
1563 virtual unsigned getUndefRegClearance(const MachineInstr
&MI
, unsigned &OpNum
,
1564 const TargetRegisterInfo
*TRI
) const {
1565 // The default implementation returns 0 for no undef register dependency.
1569 /// Insert a dependency-breaking instruction
1570 /// before MI to eliminate an unwanted dependency on OpNum.
1572 /// If it wasn't possible to avoid a def in the last N instructions before MI
1573 /// (see getPartialRegUpdateClearance), this hook will be called to break the
1574 /// unwanted dependency.
1576 /// On x86, an xorps instruction can be used as a dependency breaker:
1578 /// addps %xmm1, %xmm0
1579 /// movaps %xmm0, (%rax)
1580 /// xorps %xmm0, %xmm0
1581 /// cvtsi2ss %rbx, %xmm0
1583 /// An <imp-kill> operand should be added to MI if an instruction was
1584 /// inserted. This ties the instructions together in the post-ra scheduler.
1586 virtual void breakPartialRegDependency(MachineInstr
&MI
, unsigned OpNum
,
1587 const TargetRegisterInfo
*TRI
) const {}
1589 /// Create machine specific model for scheduling.
1590 virtual DFAPacketizer
*
1591 CreateTargetScheduleState(const TargetSubtargetInfo
&) const {
1595 /// Sometimes, it is possible for the target
1596 /// to tell, even without aliasing information, that two MIs access different
1597 /// memory addresses. This function returns true if two MIs access different
1598 /// memory addresses and false otherwise.
1600 /// Assumes any physical registers used to compute addresses have the same
1601 /// value for both instructions. (This is the most useful assumption for
1602 /// post-RA scheduling.)
1604 /// See also MachineInstr::mayAlias, which is implemented on top of this
1607 areMemAccessesTriviallyDisjoint(const MachineInstr
&MIa
,
1608 const MachineInstr
&MIb
,
1609 AliasAnalysis
*AA
= nullptr) const {
1610 assert((MIa
.mayLoad() || MIa
.mayStore()) &&
1611 "MIa must load from or modify a memory location");
1612 assert((MIb
.mayLoad() || MIb
.mayStore()) &&
1613 "MIb must load from or modify a memory location");
1617 /// Return the value to use for the MachineCSE's LookAheadLimit,
1618 /// which is a heuristic used for CSE'ing phys reg defs.
1619 virtual unsigned getMachineCSELookAheadLimit() const {
1620 // The default lookahead is small to prevent unprofitable quadratic
1625 /// Return an array that contains the ids of the target indices (used for the
1626 /// TargetIndex machine operand) and their names.
1628 /// MIR Serialization is able to serialize only the target indices that are
1629 /// defined by this method.
1630 virtual ArrayRef
<std::pair
<int, const char *>>
1631 getSerializableTargetIndices() const {
1635 /// Decompose the machine operand's target flags into two values - the direct
1636 /// target flag value and any of bit flags that are applied.
1637 virtual std::pair
<unsigned, unsigned>
1638 decomposeMachineOperandsTargetFlags(unsigned /*TF*/) const {
1639 return std::make_pair(0u, 0u);
1642 /// Return an array that contains the direct target flag values and their
1645 /// MIR Serialization is able to serialize only the target flags that are
1646 /// defined by this method.
1647 virtual ArrayRef
<std::pair
<unsigned, const char *>>
1648 getSerializableDirectMachineOperandTargetFlags() const {
1652 /// Return an array that contains the bitmask target flag values and their
1655 /// MIR Serialization is able to serialize only the target flags that are
1656 /// defined by this method.
1657 virtual ArrayRef
<std::pair
<unsigned, const char *>>
1658 getSerializableBitmaskMachineOperandTargetFlags() const {
1662 /// Return an array that contains the MMO target flag values and their
1665 /// MIR Serialization is able to serialize only the MMO target flags that are
1666 /// defined by this method.
1667 virtual ArrayRef
<std::pair
<MachineMemOperand::Flags
, const char *>>
1668 getSerializableMachineMemOperandTargetFlags() const {
1672 /// Determines whether \p Inst is a tail call instruction. Override this
1673 /// method on targets that do not properly set MCID::Return and MCID::Call on
1674 /// tail call instructions."
1675 virtual bool isTailCall(const MachineInstr
&Inst
) const {
1676 return Inst
.isReturn() && Inst
.isCall();
1679 /// True if the instruction is bound to the top of its basic block and no
1680 /// other instructions shall be inserted before it. This can be implemented
1681 /// to prevent register allocator to insert spills before such instructions.
1682 virtual bool isBasicBlockPrologue(const MachineInstr
&MI
) const {
1686 /// During PHI eleimination lets target to make necessary checks and
1687 /// insert the copy to the PHI destination register in a target specific
1689 virtual MachineInstr
*createPHIDestinationCopy(
1690 MachineBasicBlock
&MBB
, MachineBasicBlock::iterator InsPt
,
1691 const DebugLoc
&DL
, Register Src
, Register Dst
) const {
1692 return BuildMI(MBB
, InsPt
, DL
, get(TargetOpcode::COPY
), Dst
)
1696 /// During PHI eleimination lets target to make necessary checks and
1697 /// insert the copy to the PHI destination register in a target specific
1699 virtual MachineInstr
*createPHISourceCopy(MachineBasicBlock
&MBB
,
1700 MachineBasicBlock::iterator InsPt
,
1701 const DebugLoc
&DL
, Register Src
,
1703 Register Dst
) const {
1704 return BuildMI(MBB
, InsPt
, DL
, get(TargetOpcode::COPY
), Dst
)
1705 .addReg(Src
, 0, SrcSubReg
);
1708 /// Returns a \p outliner::OutlinedFunction struct containing target-specific
1709 /// information for a set of outlining candidates.
1710 virtual outliner::OutlinedFunction
getOutliningCandidateInfo(
1711 std::vector
<outliner::Candidate
> &RepeatedSequenceLocs
) const {
1713 "Target didn't implement TargetInstrInfo::getOutliningCandidateInfo!");
1716 /// Returns how or if \p MI should be outlined.
1717 virtual outliner::InstrType
1718 getOutliningType(MachineBasicBlock::iterator
&MIT
, unsigned Flags
) const {
1720 "Target didn't implement TargetInstrInfo::getOutliningType!");
1723 /// Optional target hook that returns true if \p MBB is safe to outline from,
1724 /// and returns any target-specific information in \p Flags.
1725 virtual bool isMBBSafeToOutlineFrom(MachineBasicBlock
&MBB
,
1726 unsigned &Flags
) const {
1730 /// Insert a custom frame for outlined functions.
1731 virtual void buildOutlinedFrame(MachineBasicBlock
&MBB
, MachineFunction
&MF
,
1732 const outliner::OutlinedFunction
&OF
) const {
1734 "Target didn't implement TargetInstrInfo::buildOutlinedFrame!");
1737 /// Insert a call to an outlined function into the program.
1738 /// Returns an iterator to the spot where we inserted the call. This must be
1739 /// implemented by the target.
1740 virtual MachineBasicBlock::iterator
1741 insertOutlinedCall(Module
&M
, MachineBasicBlock
&MBB
,
1742 MachineBasicBlock::iterator
&It
, MachineFunction
&MF
,
1743 const outliner::Candidate
&C
) const {
1745 "Target didn't implement TargetInstrInfo::insertOutlinedCall!");
1748 /// Return true if the function can safely be outlined from.
1749 /// A function \p MF is considered safe for outlining if an outlined function
1750 /// produced from instructions in F will produce a program which produces the
1751 /// same output for any set of given inputs.
1752 virtual bool isFunctionSafeToOutlineFrom(MachineFunction
&MF
,
1753 bool OutlineFromLinkOnceODRs
) const {
1754 llvm_unreachable("Target didn't implement "
1755 "TargetInstrInfo::isFunctionSafeToOutlineFrom!");
1758 /// Return true if the function should be outlined from by default.
1759 virtual bool shouldOutlineFromFunctionByDefault(MachineFunction
&MF
) const {
1763 /// Produce the expression describing the \p MI loading a value into
1764 /// the parameter's forwarding register.
1765 virtual Optional
<ParamLoadedValue
>
1766 describeLoadedValue(const MachineInstr
&MI
) const;
1769 unsigned CallFrameSetupOpcode
, CallFrameDestroyOpcode
;
1770 unsigned CatchRetOpcode
;
1771 unsigned ReturnOpcode
;
1774 /// Provide DenseMapInfo for TargetInstrInfo::RegSubRegPair.
1775 template <> struct DenseMapInfo
<TargetInstrInfo::RegSubRegPair
> {
1776 using RegInfo
= DenseMapInfo
<unsigned>;
1778 static inline TargetInstrInfo::RegSubRegPair
getEmptyKey() {
1779 return TargetInstrInfo::RegSubRegPair(RegInfo::getEmptyKey(),
1780 RegInfo::getEmptyKey());
1783 static inline TargetInstrInfo::RegSubRegPair
getTombstoneKey() {
1784 return TargetInstrInfo::RegSubRegPair(RegInfo::getTombstoneKey(),
1785 RegInfo::getTombstoneKey());
1788 /// Reuse getHashValue implementation from
1789 /// std::pair<unsigned, unsigned>.
1790 static unsigned getHashValue(const TargetInstrInfo::RegSubRegPair
&Val
) {
1791 std::pair
<unsigned, unsigned> PairVal
= std::make_pair(Val
.Reg
, Val
.SubReg
);
1792 return DenseMapInfo
<std::pair
<unsigned, unsigned>>::getHashValue(PairVal
);
1795 static bool isEqual(const TargetInstrInfo::RegSubRegPair
&LHS
,
1796 const TargetInstrInfo::RegSubRegPair
&RHS
) {
1797 return RegInfo::isEqual(LHS
.Reg
, RHS
.Reg
) &&
1798 RegInfo::isEqual(LHS
.SubReg
, RHS
.SubReg
);
1802 } // end namespace llvm
1804 #endif // LLVM_TARGET_TARGETINSTRINFO_H