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/SmallSet.h"
18 #include "llvm/MC/MCInstrInfo.h"
19 #include "llvm/CodeGen/DFAPacketizer.h"
20 #include "llvm/CodeGen/MachineFunction.h"
24 class InstrItineraryData;
27 class MachineMemOperand;
28 class MachineRegisterInfo;
33 class ScheduleHazardRecognizer;
36 class TargetRegisterClass;
37 class TargetRegisterInfo;
38 class BranchProbability;
40 template<class T> class SmallVectorImpl;
43 //---------------------------------------------------------------------------
45 /// TargetInstrInfo - Interface to description of machine instruction set
47 class TargetInstrInfo : public MCInstrInfo {
48 TargetInstrInfo(const TargetInstrInfo &); // DO NOT IMPLEMENT
49 void operator=(const TargetInstrInfo &); // DO NOT IMPLEMENT
51 TargetInstrInfo(int CFSetupOpcode = -1, int CFDestroyOpcode = -1)
52 : CallFrameSetupOpcode(CFSetupOpcode),
53 CallFrameDestroyOpcode(CFDestroyOpcode) {
56 virtual ~TargetInstrInfo();
58 /// getRegClass - Givem a machine instruction descriptor, returns the register
59 /// class constraint for OpNum, or NULL.
60 const TargetRegisterClass *getRegClass(const MCInstrDesc &TID,
62 const TargetRegisterInfo *TRI,
63 const MachineFunction &MF) const;
65 /// isTriviallyReMaterializable - Return true if the instruction is trivially
66 /// rematerializable, meaning it has no side effects and requires no operands
67 /// that aren't always available.
68 bool isTriviallyReMaterializable(const MachineInstr *MI,
69 AliasAnalysis *AA = 0) const {
70 return MI->getOpcode() == TargetOpcode::IMPLICIT_DEF ||
71 (MI->getDesc().isRematerializable() &&
72 (isReallyTriviallyReMaterializable(MI, AA) ||
73 isReallyTriviallyReMaterializableGeneric(MI, AA)));
77 /// isReallyTriviallyReMaterializable - For instructions with opcodes for
78 /// which the M_REMATERIALIZABLE flag is set, this hook lets the target
79 /// specify whether the instruction is actually trivially rematerializable,
80 /// taking into consideration its operands. This predicate must return false
81 /// if the instruction has any side effects other than producing a value, or
82 /// if it requres any address registers that are not always available.
83 virtual bool isReallyTriviallyReMaterializable(const MachineInstr *MI,
84 AliasAnalysis *AA) const {
89 /// isReallyTriviallyReMaterializableGeneric - For instructions with opcodes
90 /// for which the M_REMATERIALIZABLE flag is set and the target hook
91 /// isReallyTriviallyReMaterializable returns false, this function does
92 /// target-independent tests to determine if the instruction is really
93 /// trivially rematerializable.
94 bool isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
95 AliasAnalysis *AA) const;
98 /// getCallFrameSetup/DestroyOpcode - These methods return the opcode of the
99 /// frame setup/destroy instructions if they exist (-1 otherwise). Some
100 /// targets use pseudo instructions in order to abstract away the difference
101 /// between operating with a frame pointer and operating without, through the
102 /// use of these two instructions.
104 int getCallFrameSetupOpcode() const { return CallFrameSetupOpcode; }
105 int getCallFrameDestroyOpcode() const { return CallFrameDestroyOpcode; }
107 /// isCoalescableExtInstr - Return true if the instruction is a "coalescable"
108 /// extension instruction. That is, it's like a copy where it's legal for the
109 /// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns
110 /// true, then it's expected the pre-extension value is available as a subreg
111 /// of the result register. This also returns the sub-register index in
113 virtual bool isCoalescableExtInstr(const MachineInstr &MI,
114 unsigned &SrcReg, unsigned &DstReg,
115 unsigned &SubIdx) const {
119 /// isLoadFromStackSlot - If the specified machine instruction is a direct
120 /// load from a stack slot, return the virtual or physical register number of
121 /// the destination along with the FrameIndex of the loaded stack slot. If
122 /// not, return 0. This predicate must return 0 if the instruction has
123 /// any side effects other than loading from the stack slot.
124 virtual unsigned isLoadFromStackSlot(const MachineInstr *MI,
125 int &FrameIndex) const {
129 /// isLoadFromStackSlotPostFE - Check for post-frame ptr elimination
130 /// stack locations as well. This uses a heuristic so it isn't
131 /// reliable for correctness.
132 virtual unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI,
133 int &FrameIndex) const {
137 /// hasLoadFromStackSlot - If the specified machine instruction has
138 /// a load from a stack slot, return true along with the FrameIndex
139 /// of the loaded stack slot and the machine mem operand containing
140 /// the reference. If not, return false. Unlike
141 /// isLoadFromStackSlot, this returns true for any instructions that
142 /// loads from the stack. This is just a hint, as some cases may be
144 virtual bool hasLoadFromStackSlot(const MachineInstr *MI,
145 const MachineMemOperand *&MMO,
146 int &FrameIndex) const {
150 /// isStoreToStackSlot - If the specified machine instruction is a direct
151 /// store to a stack slot, return the virtual or physical register number of
152 /// the source reg along with the FrameIndex of the loaded stack slot. If
153 /// not, return 0. This predicate must return 0 if the instruction has
154 /// any side effects other than storing to the stack slot.
155 virtual unsigned isStoreToStackSlot(const MachineInstr *MI,
156 int &FrameIndex) const {
160 /// isStoreToStackSlotPostFE - Check for post-frame ptr elimination
161 /// stack locations as well. This uses a heuristic so it isn't
162 /// reliable for correctness.
163 virtual unsigned isStoreToStackSlotPostFE(const MachineInstr *MI,
164 int &FrameIndex) const {
168 /// hasStoreToStackSlot - If the specified machine instruction has a
169 /// store to a stack slot, return true along with the FrameIndex of
170 /// the loaded stack slot and the machine mem operand containing the
171 /// reference. If not, return false. Unlike isStoreToStackSlot,
172 /// this returns true for any instructions that stores to the
173 /// stack. This is just a hint, as some cases may be missed.
174 virtual bool hasStoreToStackSlot(const MachineInstr *MI,
175 const MachineMemOperand *&MMO,
176 int &FrameIndex) const {
180 /// reMaterialize - Re-issue the specified 'original' instruction at the
181 /// specific location targeting a new destination register.
182 /// The register in Orig->getOperand(0).getReg() will be substituted by
183 /// DestReg:SubIdx. Any existing subreg index is preserved or composed with
185 virtual void reMaterialize(MachineBasicBlock &MBB,
186 MachineBasicBlock::iterator MI,
187 unsigned DestReg, unsigned SubIdx,
188 const MachineInstr *Orig,
189 const TargetRegisterInfo &TRI) const = 0;
191 /// duplicate - Create a duplicate of the Orig instruction in MF. This is like
192 /// MachineFunction::CloneMachineInstr(), but the target may update operands
193 /// that are required to be unique.
195 /// The instruction must be duplicable as indicated by isNotDuplicable().
196 virtual MachineInstr *duplicate(MachineInstr *Orig,
197 MachineFunction &MF) const = 0;
199 /// convertToThreeAddress - This method must be implemented by targets that
200 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
201 /// may be able to convert a two-address instruction into one or more true
202 /// three-address instructions on demand. This allows the X86 target (for
203 /// example) to convert ADD and SHL instructions into LEA instructions if they
204 /// would require register copies due to two-addressness.
206 /// This method returns a null pointer if the transformation cannot be
207 /// performed, otherwise it returns the last new instruction.
209 virtual MachineInstr *
210 convertToThreeAddress(MachineFunction::iterator &MFI,
211 MachineBasicBlock::iterator &MBBI, LiveVariables *LV) const {
215 /// commuteInstruction - If a target has any instructions that are
216 /// commutable but require converting to different instructions or making
217 /// non-trivial changes to commute them, this method can overloaded to do
218 /// that. The default implementation simply swaps the commutable operands.
219 /// If NewMI is false, MI is modified in place and returned; otherwise, a
220 /// new machine instruction is created and returned. Do not call this
221 /// method for a non-commutable instruction, but there may be some cases
222 /// where this method fails and returns null.
223 virtual MachineInstr *commuteInstruction(MachineInstr *MI,
224 bool NewMI = false) const = 0;
226 /// findCommutedOpIndices - If specified MI is commutable, return the two
227 /// operand indices that would swap value. Return false if the instruction
228 /// is not in a form which this routine understands.
229 virtual bool findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
230 unsigned &SrcOpIdx2) const = 0;
232 /// produceSameValue - Return true if two machine instructions would produce
233 /// identical values. By default, this is only true when the two instructions
234 /// are deemed identical except for defs. If this function is called when the
235 /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for
236 /// aggressive checks.
237 virtual bool produceSameValue(const MachineInstr *MI0,
238 const MachineInstr *MI1,
239 const MachineRegisterInfo *MRI = 0) const = 0;
241 /// AnalyzeBranch - Analyze the branching code at the end of MBB, returning
242 /// true if it cannot be understood (e.g. it's a switch dispatch or isn't
243 /// implemented for a target). Upon success, this returns false and returns
244 /// with the following information in various cases:
246 /// 1. If this block ends with no branches (it just falls through to its succ)
247 /// just return false, leaving TBB/FBB null.
248 /// 2. If this block ends with only an unconditional branch, it sets TBB to be
249 /// the destination block.
250 /// 3. If this block ends with a conditional branch and it falls through to a
251 /// successor block, it sets TBB to be the branch destination block and a
252 /// list of operands that evaluate the condition. These operands can be
253 /// passed to other TargetInstrInfo methods to create new branches.
254 /// 4. If this block ends with a conditional branch followed by an
255 /// unconditional branch, it returns the 'true' destination in TBB, the
256 /// 'false' destination in FBB, and a list of operands that evaluate the
257 /// condition. These operands can be passed to other TargetInstrInfo
258 /// methods to create new branches.
260 /// Note that RemoveBranch and InsertBranch must be implemented to support
261 /// cases where this method returns success.
263 /// If AllowModify is true, then this routine is allowed to modify the basic
264 /// block (e.g. delete instructions after the unconditional branch).
266 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
267 MachineBasicBlock *&FBB,
268 SmallVectorImpl<MachineOperand> &Cond,
269 bool AllowModify = false) const {
273 /// RemoveBranch - Remove the branching code at the end of the specific MBB.
274 /// This is only invoked in cases where AnalyzeBranch returns success. It
275 /// returns the number of instructions that were removed.
276 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const {
277 llvm_unreachable("Target didn't implement TargetInstrInfo::RemoveBranch!");
280 /// InsertBranch - Insert branch code into the end of the specified
281 /// MachineBasicBlock. The operands to this method are the same as those
282 /// returned by AnalyzeBranch. This is only invoked in cases where
283 /// AnalyzeBranch returns success. It returns the number of instructions
286 /// It is also invoked by tail merging to add unconditional branches in
287 /// cases where AnalyzeBranch doesn't apply because there was no original
288 /// branch to analyze. At least this much must be implemented, else tail
289 /// merging needs to be disabled.
290 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
291 MachineBasicBlock *FBB,
292 const SmallVectorImpl<MachineOperand> &Cond,
294 llvm_unreachable("Target didn't implement TargetInstrInfo::InsertBranch!");
297 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
298 /// after it, replacing it with an unconditional branch to NewDest. This is
299 /// used by the tail merging pass.
300 virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
301 MachineBasicBlock *NewDest) const = 0;
303 /// isLegalToSplitMBBAt - Return true if it's legal to split the given basic
304 /// block at the specified instruction (i.e. instruction would be the start
305 /// of a new basic block).
306 virtual bool isLegalToSplitMBBAt(MachineBasicBlock &MBB,
307 MachineBasicBlock::iterator MBBI) const {
311 /// isProfitableToIfCvt - Return true if it's profitable to predicate
312 /// instructions with accumulated instruction latency of "NumCycles"
313 /// of the specified basic block, where the probability of the instructions
314 /// being executed is given by Probability, and Confidence is a measure
315 /// of our confidence that it will be properly predicted.
317 bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
318 unsigned ExtraPredCycles,
319 const BranchProbability &Probability) const {
323 /// isProfitableToIfCvt - Second variant of isProfitableToIfCvt, this one
324 /// checks for the case where two basic blocks from true and false path
325 /// of a if-then-else (diamond) are predicated on mutally exclusive
326 /// predicates, where the probability of the true path being taken is given
327 /// by Probability, and Confidence is a measure of our confidence that it
328 /// will be properly predicted.
330 isProfitableToIfCvt(MachineBasicBlock &TMBB,
331 unsigned NumTCycles, unsigned ExtraTCycles,
332 MachineBasicBlock &FMBB,
333 unsigned NumFCycles, unsigned ExtraFCycles,
334 const BranchProbability &Probability) const {
338 /// isProfitableToDupForIfCvt - Return true if it's profitable for
339 /// if-converter to duplicate instructions of specified accumulated
340 /// instruction latencies in the specified MBB to enable if-conversion.
341 /// The probability of the instructions being executed is given by
342 /// Probability, and Confidence is a measure of our confidence that it
343 /// will be properly predicted.
345 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
346 const BranchProbability &Probability) const {
350 /// isProfitableToUnpredicate - Return true if it's profitable to unpredicate
351 /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually
352 /// exclusive predicates.
360 /// This may be profitable is conditional instructions are always executed.
361 virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
362 MachineBasicBlock &FMBB) const {
366 /// canInsertSelect - Return true if it is possible to insert a select
367 /// instruction that chooses between TrueReg and FalseReg based on the
368 /// condition code in Cond.
370 /// When successful, also return the latency in cycles from TrueReg,
371 /// FalseReg, and Cond to the destination register. The Cond latency should
372 /// compensate for a conditional branch being removed. For example, if a
373 /// conditional branch has a 3 cycle latency from the condition code read,
374 /// and a cmov instruction has a 2 cycle latency from the condition code
375 /// read, CondCycles should be returned as -1.
377 /// @param MBB Block where select instruction would be inserted.
378 /// @param Cond Condition returned by AnalyzeBranch.
379 /// @param TrueReg Virtual register to select when Cond is true.
380 /// @param FalseReg Virtual register to select when Cond is false.
381 /// @param CondCycles Latency from Cond+Branch to select output.
382 /// @param TrueCycles Latency from TrueReg to select output.
383 /// @param FalseCycles Latency from FalseReg to select output.
384 virtual bool canInsertSelect(const MachineBasicBlock &MBB,
385 const SmallVectorImpl<MachineOperand> &Cond,
386 unsigned TrueReg, unsigned FalseReg,
388 int &TrueCycles, int &FalseCycles) const {
392 /// insertSelect - Insert a select instruction into MBB before I that will
393 /// copy TrueReg to DstReg when Cond is true, and FalseReg to DstReg when
396 /// This function can only be called after canInsertSelect() returned true.
397 /// The condition in Cond comes from AnalyzeBranch, and it can be assumed
398 /// that the same flags or registers required by Cond are available at the
401 /// @param MBB Block where select instruction should be inserted.
402 /// @param I Insertion point.
403 /// @param DL Source location for debugging.
404 /// @param DstReg Virtual register to be defined by select instruction.
405 /// @param Cond Condition as computed by AnalyzeBranch.
406 /// @param TrueReg Virtual register to copy when Cond is true.
407 /// @param FalseReg Virtual register to copy when Cons is false.
408 virtual void insertSelect(MachineBasicBlock &MBB,
409 MachineBasicBlock::iterator I, DebugLoc DL,
411 const SmallVectorImpl<MachineOperand> &Cond,
412 unsigned TrueReg, unsigned FalseReg) const {
413 llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!");
416 /// analyzeSelect - Analyze the given select instruction, returning true if
417 /// it cannot be understood. It is assumed that MI->isSelect() is true.
419 /// When successful, return the controlling condition and the operands that
420 /// determine the true and false result values.
422 /// Result = SELECT Cond, TrueOp, FalseOp
424 /// Some targets can optimize select instructions, for example by predicating
425 /// the instruction defining one of the operands. Such targets should set
428 /// @param MI Select instruction to analyze.
429 /// @param Cond Condition controlling the select.
430 /// @param TrueOp Operand number of the value selected when Cond is true.
431 /// @param FalseOp Operand number of the value selected when Cond is false.
432 /// @param Optimizable Returned as true if MI is optimizable.
433 /// @returns False on success.
434 virtual bool analyzeSelect(const MachineInstr *MI,
435 SmallVectorImpl<MachineOperand> &Cond,
436 unsigned &TrueOp, unsigned &FalseOp,
437 bool &Optimizable) const {
438 assert(MI && MI->isSelect() && "MI must be a select instruction");
442 /// optimizeSelect - Given a select instruction that was understood by
443 /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by
444 /// merging it with one of its operands. Returns NULL on failure.
446 /// When successful, returns the new select instruction. The client is
447 /// responsible for deleting MI.
449 /// If both sides of the select can be optimized, PreferFalse is used to pick
452 /// @param MI Optimizable select instruction.
453 /// @param PreferFalse Try to optimize FalseOp instead of TrueOp.
454 /// @returns Optimized instruction or NULL.
455 virtual MachineInstr *optimizeSelect(MachineInstr *MI,
456 bool PreferFalse = false) const {
457 // This function must be implemented if Optimizable is ever set.
458 llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!");
461 /// copyPhysReg - Emit instructions to copy a pair of physical registers.
462 virtual void copyPhysReg(MachineBasicBlock &MBB,
463 MachineBasicBlock::iterator MI, DebugLoc DL,
464 unsigned DestReg, unsigned SrcReg,
465 bool KillSrc) const {
466 llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!");
469 /// storeRegToStackSlot - Store the specified register of the given register
470 /// class to the specified stack frame index. The store instruction is to be
471 /// added to the given machine basic block before the specified machine
472 /// instruction. If isKill is true, the register operand is the last use and
473 /// must be marked kill.
474 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
475 MachineBasicBlock::iterator MI,
476 unsigned SrcReg, bool isKill, int FrameIndex,
477 const TargetRegisterClass *RC,
478 const TargetRegisterInfo *TRI) const {
479 llvm_unreachable("Target didn't implement "
480 "TargetInstrInfo::storeRegToStackSlot!");
483 /// loadRegFromStackSlot - Load the specified register of the given register
484 /// class from the specified stack frame index. The load instruction is to be
485 /// added to the given machine basic block before the specified machine
487 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
488 MachineBasicBlock::iterator MI,
489 unsigned DestReg, int FrameIndex,
490 const TargetRegisterClass *RC,
491 const TargetRegisterInfo *TRI) const {
492 llvm_unreachable("Target didn't implement "
493 "TargetInstrInfo::loadRegFromStackSlot!");
496 /// expandPostRAPseudo - This function is called for all pseudo instructions
497 /// that remain after register allocation. Many pseudo instructions are
498 /// created to help register allocation. This is the place to convert them
499 /// into real instructions. The target can edit MI in place, or it can insert
500 /// new instructions and erase MI. The function should return true if
501 /// anything was changed.
502 virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
506 /// emitFrameIndexDebugValue - Emit a target-dependent form of
507 /// DBG_VALUE encoding the address of a frame index. Addresses would
508 /// normally be lowered the same way as other addresses on the target,
509 /// e.g. in load instructions. For targets that do not support this
510 /// the debug info is simply lost.
511 /// If you add this for a target you should handle this DBG_VALUE in the
512 /// target-specific AsmPrinter code as well; you will probably get invalid
513 /// assembly output if you don't.
514 virtual MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF,
522 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
523 /// slot into the specified machine instruction for the specified operand(s).
524 /// If this is possible, a new instruction is returned with the specified
525 /// operand folded, otherwise NULL is returned.
526 /// The new instruction is inserted before MI, and the client is responsible
527 /// for removing the old instruction.
528 MachineInstr* foldMemoryOperand(MachineBasicBlock::iterator MI,
529 const SmallVectorImpl<unsigned> &Ops,
530 int FrameIndex) const;
532 /// foldMemoryOperand - Same as the previous version except it allows folding
533 /// of any load and store from / to any address, not just from a specific
535 MachineInstr* foldMemoryOperand(MachineBasicBlock::iterator MI,
536 const SmallVectorImpl<unsigned> &Ops,
537 MachineInstr* LoadMI) const;
540 /// foldMemoryOperandImpl - Target-dependent implementation for
541 /// foldMemoryOperand. Target-independent code in foldMemoryOperand will
542 /// take care of adding a MachineMemOperand to the newly created instruction.
543 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
545 const SmallVectorImpl<unsigned> &Ops,
546 int FrameIndex) const {
550 /// foldMemoryOperandImpl - Target-dependent implementation for
551 /// foldMemoryOperand. Target-independent code in foldMemoryOperand will
552 /// take care of adding a MachineMemOperand to the newly created instruction.
553 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
555 const SmallVectorImpl<unsigned> &Ops,
556 MachineInstr* LoadMI) const {
561 /// canFoldMemoryOperand - Returns true for the specified load / store if
562 /// folding is possible.
564 bool canFoldMemoryOperand(const MachineInstr *MI,
565 const SmallVectorImpl<unsigned> &Ops) const =0;
567 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
568 /// a store or a load and a store into two or more instruction. If this is
569 /// possible, returns true as well as the new instructions by reference.
570 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
571 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
572 SmallVectorImpl<MachineInstr*> &NewMIs) const{
576 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
577 SmallVectorImpl<SDNode*> &NewNodes) const {
581 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
582 /// instruction after load / store are unfolded from an instruction of the
583 /// specified opcode. It returns zero if the specified unfolding is not
584 /// possible. If LoadRegIndex is non-null, it is filled in with the operand
585 /// index of the operand which will hold the register holding the loaded
587 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
588 bool UnfoldLoad, bool UnfoldStore,
589 unsigned *LoadRegIndex = 0) const {
593 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler
594 /// to determine if two loads are loading from the same base address. It
595 /// should only return true if the base pointers are the same and the
596 /// only differences between the two addresses are the offset. It also returns
597 /// the offsets by reference.
598 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
599 int64_t &Offset1, int64_t &Offset2) const {
603 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
604 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
605 /// be scheduled togther. On some targets if two loads are loading from
606 /// addresses in the same cache line, it's better if they are scheduled
607 /// together. This function takes two integers that represent the load offsets
608 /// from the common base address. It returns true if it decides it's desirable
609 /// to schedule the two loads together. "NumLoads" is the number of loads that
610 /// have already been scheduled after Load1.
611 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
612 int64_t Offset1, int64_t Offset2,
613 unsigned NumLoads) const {
617 /// ReverseBranchCondition - Reverses the branch condition of the specified
618 /// condition list, returning false on success and true if it cannot be
621 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
625 /// insertNoop - Insert a noop into the instruction stream at the specified
627 virtual void insertNoop(MachineBasicBlock &MBB,
628 MachineBasicBlock::iterator MI) const;
631 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
632 virtual void getNoopForMachoTarget(MCInst &NopInst) const {
633 // Default to just using 'nop' string.
637 /// isPredicated - Returns true if the instruction is already predicated.
639 virtual bool isPredicated(const MachineInstr *MI) const {
643 /// isUnpredicatedTerminator - Returns true if the instruction is a
644 /// terminator instruction that has not been predicated.
645 virtual bool isUnpredicatedTerminator(const MachineInstr *MI) const = 0;
647 /// PredicateInstruction - Convert the instruction into a predicated
648 /// instruction. It returns true if the operation was successful.
650 bool PredicateInstruction(MachineInstr *MI,
651 const SmallVectorImpl<MachineOperand> &Pred) const = 0;
653 /// SubsumesPredicate - Returns true if the first specified predicate
654 /// subsumes the second, e.g. GE subsumes GT.
656 bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
657 const SmallVectorImpl<MachineOperand> &Pred2) const {
661 /// DefinesPredicate - If the specified instruction defines any predicate
662 /// or condition code register(s) used for predication, returns true as well
663 /// as the definition predicate(s) by reference.
664 virtual bool DefinesPredicate(MachineInstr *MI,
665 std::vector<MachineOperand> &Pred) const {
669 /// isPredicable - Return true if the specified instruction can be predicated.
670 /// By default, this returns true for every instruction with a
671 /// PredicateOperand.
672 virtual bool isPredicable(MachineInstr *MI) const {
673 return MI->getDesc().isPredicable();
676 /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
677 /// instruction that defines the specified register class.
678 virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
682 /// isSchedulingBoundary - Test if the given instruction should be
683 /// considered a scheduling boundary. This primarily includes labels and
685 virtual bool isSchedulingBoundary(const MachineInstr *MI,
686 const MachineBasicBlock *MBB,
687 const MachineFunction &MF) const = 0;
689 /// Measure the specified inline asm to determine an approximation of its
691 virtual unsigned getInlineAsmLength(const char *Str,
692 const MCAsmInfo &MAI) const;
694 /// CreateTargetHazardRecognizer - Allocate and return a hazard recognizer to
695 /// use for this target when scheduling the machine instructions before
696 /// register allocation.
697 virtual ScheduleHazardRecognizer*
698 CreateTargetHazardRecognizer(const TargetMachine *TM,
699 const ScheduleDAG *DAG) const = 0;
701 /// CreateTargetMIHazardRecognizer - Allocate and return a hazard recognizer
702 /// to use for this target when scheduling the machine instructions before
703 /// register allocation.
704 virtual ScheduleHazardRecognizer*
705 CreateTargetMIHazardRecognizer(const InstrItineraryData*,
706 const ScheduleDAG *DAG) const = 0;
708 /// CreateTargetPostRAHazardRecognizer - Allocate and return a hazard
709 /// recognizer to use for this target when scheduling the machine instructions
710 /// after register allocation.
711 virtual ScheduleHazardRecognizer*
712 CreateTargetPostRAHazardRecognizer(const InstrItineraryData*,
713 const ScheduleDAG *DAG) const = 0;
715 /// analyzeCompare - For a comparison instruction, return the source registers
716 /// in SrcReg and SrcReg2 if having two register operands, and the value it
717 /// compares against in CmpValue. Return true if the comparison instruction
719 virtual bool analyzeCompare(const MachineInstr *MI,
720 unsigned &SrcReg, unsigned &SrcReg2,
721 int &Mask, int &Value) const {
725 /// optimizeCompareInstr - See if the comparison instruction can be converted
726 /// into something more efficient. E.g., on ARM most instructions can set the
727 /// flags register, obviating the need for a separate CMP.
728 virtual bool optimizeCompareInstr(MachineInstr *CmpInstr,
729 unsigned SrcReg, unsigned SrcReg2,
731 const MachineRegisterInfo *MRI) const {
735 /// optimizeLoadInstr - Try to remove the load by folding it to a register
736 /// operand at the use. We fold the load instructions if and only if the
737 /// def and use are in the same BB. We only look at one load and see
738 /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
739 /// defined by the load we are trying to fold. DefMI returns the machine
740 /// instruction that defines FoldAsLoadDefReg, and the function returns
741 /// the machine instruction generated due to folding.
742 virtual MachineInstr* optimizeLoadInstr(MachineInstr *MI,
743 const MachineRegisterInfo *MRI,
744 unsigned &FoldAsLoadDefReg,
745 MachineInstr *&DefMI) const {
749 /// FoldImmediate - 'Reg' is known to be defined by a move immediate
750 /// instruction, try to fold the immediate into the use instruction.
751 virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
752 unsigned Reg, MachineRegisterInfo *MRI) const {
756 /// getNumMicroOps - Return the number of u-operations the given machine
757 /// instruction will be decoded to on the target cpu. The itinerary's
758 /// IssueWidth is the number of microops that can be dispatched each
759 /// cycle. An instruction with zero microops takes no dispatch resources.
760 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
761 const MachineInstr *MI) const = 0;
763 /// isZeroCost - Return true for pseudo instructions that don't consume any
764 /// machine resources in their current form. These are common cases that the
765 /// scheduler should consider free, rather than conservatively handling them
766 /// as instructions with no itinerary.
767 bool isZeroCost(unsigned Opcode) const {
768 return Opcode <= TargetOpcode::COPY;
771 virtual int getOperandLatency(const InstrItineraryData *ItinData,
772 SDNode *DefNode, unsigned DefIdx,
773 SDNode *UseNode, unsigned UseIdx) const = 0;
775 /// getOperandLatency - Compute and return the use operand latency of a given
776 /// pair of def and use.
777 /// In most cases, the static scheduling itinerary was enough to determine the
778 /// operand latency. But it may not be possible for instructions with variable
779 /// number of defs / uses.
781 /// This is a raw interface to the itinerary that may be directly overriden by
782 /// a target. Use computeOperandLatency to get the best estimate of latency.
783 virtual int getOperandLatency(const InstrItineraryData *ItinData,
784 const MachineInstr *DefMI, unsigned DefIdx,
785 const MachineInstr *UseMI,
786 unsigned UseIdx) const = 0;
788 /// computeOperandLatency - Compute and return the latency of the given data
789 /// dependent def and use when the operand indices are already known.
791 /// FindMin may be set to get the minimum vs. expected latency.
792 unsigned computeOperandLatency(const InstrItineraryData *ItinData,
793 const MachineInstr *DefMI, unsigned DefIdx,
794 const MachineInstr *UseMI, unsigned UseIdx,
795 bool FindMin = false) const;
797 /// computeOperandLatency - Compute and return the latency of the given data
798 /// dependent def and use. DefMI must be a valid def. UseMI may be NULL for
799 /// an unknown use. If the subtarget allows, this may or may not need to call
800 /// getOperandLatency().
802 /// FindMin may be set to get the minimum vs. expected latency. Minimum
803 /// latency is used for scheduling groups, while expected latency is for
804 /// instruction cost and critical path.
805 unsigned computeOperandLatency(const InstrItineraryData *ItinData,
806 const TargetRegisterInfo *TRI,
807 const MachineInstr *DefMI,
808 const MachineInstr *UseMI,
809 unsigned Reg, bool FindMin) const;
811 /// getOutputLatency - Compute and return the output dependency latency of a
812 /// a given pair of defs which both target the same register. This is usually
814 virtual unsigned getOutputLatency(const InstrItineraryData *ItinData,
815 const MachineInstr *DefMI, unsigned DefIdx,
816 const MachineInstr *DepMI) const {
820 /// getInstrLatency - Compute the instruction latency of a given instruction.
821 /// If the instruction has higher cost when predicated, it's returned via
823 virtual unsigned getInstrLatency(const InstrItineraryData *ItinData,
824 const MachineInstr *MI,
825 unsigned *PredCost = 0) const = 0;
827 virtual int getInstrLatency(const InstrItineraryData *ItinData,
828 SDNode *Node) const = 0;
830 /// Return the default expected latency for a def based on it's opcode.
831 unsigned defaultDefLatency(const MCSchedModel *SchedModel,
832 const MachineInstr *DefMI) const;
834 /// isHighLatencyDef - Return true if this opcode has high latency to its
836 virtual bool isHighLatencyDef(int opc) const { return false; }
838 /// hasHighOperandLatency - Compute operand latency between a def of 'Reg'
839 /// and an use in the current loop, return true if the target considered
840 /// it 'high'. This is used by optimization passes such as machine LICM to
841 /// determine whether it makes sense to hoist an instruction out even in
842 /// high register pressure situation.
844 bool hasHighOperandLatency(const InstrItineraryData *ItinData,
845 const MachineRegisterInfo *MRI,
846 const MachineInstr *DefMI, unsigned DefIdx,
847 const MachineInstr *UseMI, unsigned UseIdx) const {
851 /// hasLowDefLatency - Compute operand latency of a def of 'Reg', return true
852 /// if the target considered it 'low'.
854 bool hasLowDefLatency(const InstrItineraryData *ItinData,
855 const MachineInstr *DefMI, unsigned DefIdx) const = 0;
857 /// verifyInstruction - Perform target specific instruction verification.
859 bool verifyInstruction(const MachineInstr *MI, StringRef &ErrInfo) const {
863 /// getExecutionDomain - Return the current execution domain and bit mask of
864 /// possible domains for instruction.
866 /// Some micro-architectures have multiple execution domains, and multiple
867 /// opcodes that perform the same operation in different domains. For
868 /// example, the x86 architecture provides the por, orps, and orpd
869 /// instructions that all do the same thing. There is a latency penalty if a
870 /// register is written in one domain and read in another.
872 /// This function returns a pair (domain, mask) containing the execution
873 /// domain of MI, and a bit mask of possible domains. The setExecutionDomain
874 /// function can be used to change the opcode to one of the domains in the
875 /// bit mask. Instructions whose execution domain can't be changed should
878 /// The execution domain numbers don't have any special meaning except domain
879 /// 0 is used for instructions that are not associated with any interesting
880 /// execution domain.
882 virtual std::pair<uint16_t, uint16_t>
883 getExecutionDomain(const MachineInstr *MI) const {
884 return std::make_pair(0, 0);
887 /// setExecutionDomain - Change the opcode of MI to execute in Domain.
889 /// The bit (1 << Domain) must be set in the mask returned from
890 /// getExecutionDomain(MI).
892 virtual void setExecutionDomain(MachineInstr *MI, unsigned Domain) const {}
895 /// getPartialRegUpdateClearance - Returns the preferred minimum clearance
896 /// before an instruction with an unwanted partial register update.
898 /// Some instructions only write part of a register, and implicitly need to
899 /// read the other parts of the register. This may cause unwanted stalls
900 /// preventing otherwise unrelated instructions from executing in parallel in
901 /// an out-of-order CPU.
903 /// For example, the x86 instruction cvtsi2ss writes its result to bits
904 /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so
905 /// the instruction needs to wait for the old value of the register to become
908 /// addps %xmm1, %xmm0
909 /// movaps %xmm0, (%rax)
910 /// cvtsi2ss %rbx, %xmm0
912 /// In the code above, the cvtsi2ss instruction needs to wait for the addps
913 /// instruction before it can issue, even though the high bits of %xmm0
914 /// probably aren't needed.
916 /// This hook returns the preferred clearance before MI, measured in
917 /// instructions. Other defs of MI's operand OpNum are avoided in the last N
918 /// instructions before MI. It should only return a positive value for
919 /// unwanted dependencies. If the old bits of the defined register have
920 /// useful values, or if MI is determined to otherwise read the dependency,
921 /// the hook should return 0.
923 /// The unwanted dependency may be handled by:
925 /// 1. Allocating the same register for an MI def and use. That makes the
926 /// unwanted dependency identical to a required dependency.
928 /// 2. Allocating a register for the def that has no defs in the previous N
931 /// 3. Calling breakPartialRegDependency() with the same arguments. This
932 /// allows the target to insert a dependency breaking instruction.
935 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
936 const TargetRegisterInfo *TRI) const {
937 // The default implementation returns 0 for no partial register dependency.
941 /// breakPartialRegDependency - Insert a dependency-breaking instruction
942 /// before MI to eliminate an unwanted dependency on OpNum.
944 /// If it wasn't possible to avoid a def in the last N instructions before MI
945 /// (see getPartialRegUpdateClearance), this hook will be called to break the
946 /// unwanted dependency.
948 /// On x86, an xorps instruction can be used as a dependency breaker:
950 /// addps %xmm1, %xmm0
951 /// movaps %xmm0, (%rax)
952 /// xorps %xmm0, %xmm0
953 /// cvtsi2ss %rbx, %xmm0
955 /// An <imp-kill> operand should be added to MI if an instruction was
956 /// inserted. This ties the instructions together in the post-ra scheduler.
959 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
960 const TargetRegisterInfo *TRI) const {}
962 /// Create machine specific model for scheduling.
963 virtual DFAPacketizer*
964 CreateTargetScheduleState(const TargetMachine*, const ScheduleDAG*) const {
969 int CallFrameSetupOpcode, CallFrameDestroyOpcode;
972 /// TargetInstrInfoImpl - This is the default implementation of
973 /// TargetInstrInfo, which just provides a couple of default implementations
974 /// for various methods. This separated out because it is implemented in
975 /// libcodegen, not in libtarget.
976 class TargetInstrInfoImpl : public TargetInstrInfo {
978 TargetInstrInfoImpl(int CallFrameSetupOpcode = -1,
979 int CallFrameDestroyOpcode = -1)
980 : TargetInstrInfo(CallFrameSetupOpcode, CallFrameDestroyOpcode) {}
982 virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
983 MachineBasicBlock *NewDest) const;
984 virtual MachineInstr *commuteInstruction(MachineInstr *MI,
985 bool NewMI = false) const;
986 virtual bool findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
987 unsigned &SrcOpIdx2) const;
988 virtual bool canFoldMemoryOperand(const MachineInstr *MI,
989 const SmallVectorImpl<unsigned> &Ops) const;
990 virtual bool hasLoadFromStackSlot(const MachineInstr *MI,
991 const MachineMemOperand *&MMO,
992 int &FrameIndex) const;
993 virtual bool hasStoreToStackSlot(const MachineInstr *MI,
994 const MachineMemOperand *&MMO,
995 int &FrameIndex) const;
996 virtual bool isUnpredicatedTerminator(const MachineInstr *MI) const;
997 virtual bool PredicateInstruction(MachineInstr *MI,
998 const SmallVectorImpl<MachineOperand> &Pred) const;
999 virtual void reMaterialize(MachineBasicBlock &MBB,
1000 MachineBasicBlock::iterator MI,
1001 unsigned DestReg, unsigned SubReg,
1002 const MachineInstr *Orig,
1003 const TargetRegisterInfo &TRI) const;
1004 virtual MachineInstr *duplicate(MachineInstr *Orig,
1005 MachineFunction &MF) const;
1006 virtual bool produceSameValue(const MachineInstr *MI0,
1007 const MachineInstr *MI1,
1008 const MachineRegisterInfo *MRI) const;
1009 virtual bool isSchedulingBoundary(const MachineInstr *MI,
1010 const MachineBasicBlock *MBB,
1011 const MachineFunction &MF) const;
1013 virtual int getOperandLatency(const InstrItineraryData *ItinData,
1014 SDNode *DefNode, unsigned DefIdx,
1015 SDNode *UseNode, unsigned UseIdx) const;
1017 virtual int getInstrLatency(const InstrItineraryData *ItinData,
1018 SDNode *Node) const;
1020 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
1021 const MachineInstr *MI) const;
1023 virtual unsigned getInstrLatency(const InstrItineraryData *ItinData,
1024 const MachineInstr *MI,
1025 unsigned *PredCost = 0) const;
1028 bool hasLowDefLatency(const InstrItineraryData *ItinData,
1029 const MachineInstr *DefMI, unsigned DefIdx) const;
1031 virtual int getOperandLatency(const InstrItineraryData *ItinData,
1032 const MachineInstr *DefMI, unsigned DefIdx,
1033 const MachineInstr *UseMI,
1034 unsigned UseIdx) const;
1036 bool usePreRAHazardRecognizer() const;
1038 virtual ScheduleHazardRecognizer *
1039 CreateTargetHazardRecognizer(const TargetMachine*, const ScheduleDAG*) const;
1041 virtual ScheduleHazardRecognizer *
1042 CreateTargetMIHazardRecognizer(const InstrItineraryData*,
1043 const ScheduleDAG*) const;
1045 virtual ScheduleHazardRecognizer *
1046 CreateTargetPostRAHazardRecognizer(const InstrItineraryData*,
1047 const ScheduleDAG*) const;
1050 } // End llvm namespace