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/CodeGen/DFAPacketizer.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/MC/MCInstrInfo.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 &) LLVM_DELETED_FUNCTION;
49 void operator=(const TargetInstrInfo &) LLVM_DELETED_FUNCTION;
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;
148 /// isStoreToStackSlot - If the specified machine instruction is a direct
149 /// store to a stack slot, return the virtual or physical register number of
150 /// the source reg along with the FrameIndex of the loaded stack slot. If
151 /// not, return 0. This predicate must return 0 if the instruction has
152 /// any side effects other than storing to the stack slot.
153 virtual unsigned isStoreToStackSlot(const MachineInstr *MI,
154 int &FrameIndex) const {
158 /// isStoreToStackSlotPostFE - Check for post-frame ptr elimination
159 /// stack locations as well. This uses a heuristic so it isn't
160 /// reliable for correctness.
161 virtual unsigned isStoreToStackSlotPostFE(const MachineInstr *MI,
162 int &FrameIndex) const {
166 /// hasStoreToStackSlot - If the specified machine instruction has a
167 /// store to a stack slot, return true along with the FrameIndex of
168 /// the loaded stack slot and the machine mem operand containing the
169 /// reference. If not, return false. Unlike isStoreToStackSlot,
170 /// this returns true for any instructions that stores to the
171 /// stack. This is just a hint, as some cases may be missed.
172 virtual bool hasStoreToStackSlot(const MachineInstr *MI,
173 const MachineMemOperand *&MMO,
174 int &FrameIndex) const;
176 /// reMaterialize - Re-issue the specified 'original' instruction at the
177 /// specific location targeting a new destination register.
178 /// The register in Orig->getOperand(0).getReg() will be substituted by
179 /// DestReg:SubIdx. Any existing subreg index is preserved or composed with
181 virtual void reMaterialize(MachineBasicBlock &MBB,
182 MachineBasicBlock::iterator MI,
183 unsigned DestReg, unsigned SubIdx,
184 const MachineInstr *Orig,
185 const TargetRegisterInfo &TRI) const;
187 /// duplicate - Create a duplicate of the Orig instruction in MF. This is like
188 /// MachineFunction::CloneMachineInstr(), but the target may update operands
189 /// that are required to be unique.
191 /// The instruction must be duplicable as indicated by isNotDuplicable().
192 virtual MachineInstr *duplicate(MachineInstr *Orig,
193 MachineFunction &MF) const;
195 /// convertToThreeAddress - This method must be implemented by targets that
196 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target
197 /// may be able to convert a two-address instruction into one or more true
198 /// three-address instructions on demand. This allows the X86 target (for
199 /// example) to convert ADD and SHL instructions into LEA instructions if they
200 /// would require register copies due to two-addressness.
202 /// This method returns a null pointer if the transformation cannot be
203 /// performed, otherwise it returns the last new instruction.
205 virtual MachineInstr *
206 convertToThreeAddress(MachineFunction::iterator &MFI,
207 MachineBasicBlock::iterator &MBBI, LiveVariables *LV) const {
211 /// commuteInstruction - If a target has any instructions that are
212 /// commutable but require converting to different instructions or making
213 /// non-trivial changes to commute them, this method can overloaded to do
214 /// that. The default implementation simply swaps the commutable operands.
215 /// If NewMI is false, MI is modified in place and returned; otherwise, a
216 /// new machine instruction is created and returned. Do not call this
217 /// method for a non-commutable instruction, but there may be some cases
218 /// where this method fails and returns null.
219 virtual MachineInstr *commuteInstruction(MachineInstr *MI,
220 bool NewMI = false) const;
222 /// findCommutedOpIndices - If specified MI is commutable, return the two
223 /// operand indices that would swap value. Return false if the instruction
224 /// is not in a form which this routine understands.
225 virtual bool findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
226 unsigned &SrcOpIdx2) const;
228 /// produceSameValue - Return true if two machine instructions would produce
229 /// identical values. By default, this is only true when the two instructions
230 /// are deemed identical except for defs. If this function is called when the
231 /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for
232 /// aggressive checks.
233 virtual bool produceSameValue(const MachineInstr *MI0,
234 const MachineInstr *MI1,
235 const MachineRegisterInfo *MRI = 0) const;
237 /// AnalyzeBranch - Analyze the branching code at the end of MBB, returning
238 /// true if it cannot be understood (e.g. it's a switch dispatch or isn't
239 /// implemented for a target). Upon success, this returns false and returns
240 /// with the following information in various cases:
242 /// 1. If this block ends with no branches (it just falls through to its succ)
243 /// just return false, leaving TBB/FBB null.
244 /// 2. If this block ends with only an unconditional branch, it sets TBB to be
245 /// the destination block.
246 /// 3. If this block ends with a conditional branch and it falls through to a
247 /// successor block, it sets TBB to be the branch destination block and a
248 /// list of operands that evaluate the condition. These operands can be
249 /// passed to other TargetInstrInfo methods to create new branches.
250 /// 4. If this block ends with a conditional branch followed by an
251 /// unconditional branch, it returns the 'true' destination in TBB, the
252 /// 'false' destination in FBB, and a list of operands that evaluate the
253 /// condition. These operands can be passed to other TargetInstrInfo
254 /// methods to create new branches.
256 /// Note that RemoveBranch and InsertBranch must be implemented to support
257 /// cases where this method returns success.
259 /// If AllowModify is true, then this routine is allowed to modify the basic
260 /// block (e.g. delete instructions after the unconditional branch).
262 virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
263 MachineBasicBlock *&FBB,
264 SmallVectorImpl<MachineOperand> &Cond,
265 bool AllowModify = false) const {
269 /// RemoveBranch - Remove the branching code at the end of the specific MBB.
270 /// This is only invoked in cases where AnalyzeBranch returns success. It
271 /// returns the number of instructions that were removed.
272 virtual unsigned RemoveBranch(MachineBasicBlock &MBB) const {
273 llvm_unreachable("Target didn't implement TargetInstrInfo::RemoveBranch!");
276 /// InsertBranch - Insert branch code into the end of the specified
277 /// MachineBasicBlock. The operands to this method are the same as those
278 /// returned by AnalyzeBranch. This is only invoked in cases where
279 /// AnalyzeBranch returns success. It returns the number of instructions
282 /// It is also invoked by tail merging to add unconditional branches in
283 /// cases where AnalyzeBranch doesn't apply because there was no original
284 /// branch to analyze. At least this much must be implemented, else tail
285 /// merging needs to be disabled.
286 virtual unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
287 MachineBasicBlock *FBB,
288 const SmallVectorImpl<MachineOperand> &Cond,
290 llvm_unreachable("Target didn't implement TargetInstrInfo::InsertBranch!");
293 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
294 /// after it, replacing it with an unconditional branch to NewDest. This is
295 /// used by the tail merging pass.
296 virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
297 MachineBasicBlock *NewDest) const;
299 /// isLegalToSplitMBBAt - Return true if it's legal to split the given basic
300 /// block at the specified instruction (i.e. instruction would be the start
301 /// of a new basic block).
302 virtual bool isLegalToSplitMBBAt(MachineBasicBlock &MBB,
303 MachineBasicBlock::iterator MBBI) const {
307 /// isProfitableToIfCvt - Return true if it's profitable to predicate
308 /// instructions with accumulated instruction latency of "NumCycles"
309 /// of the specified basic block, where the probability of the instructions
310 /// being executed is given by Probability, and Confidence is a measure
311 /// of our confidence that it will be properly predicted.
313 bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
314 unsigned ExtraPredCycles,
315 const BranchProbability &Probability) const {
319 /// isProfitableToIfCvt - Second variant of isProfitableToIfCvt, this one
320 /// checks for the case where two basic blocks from true and false path
321 /// of a if-then-else (diamond) are predicated on mutally exclusive
322 /// predicates, where the probability of the true path being taken is given
323 /// by Probability, and Confidence is a measure of our confidence that it
324 /// will be properly predicted.
326 isProfitableToIfCvt(MachineBasicBlock &TMBB,
327 unsigned NumTCycles, unsigned ExtraTCycles,
328 MachineBasicBlock &FMBB,
329 unsigned NumFCycles, unsigned ExtraFCycles,
330 const BranchProbability &Probability) const {
334 /// isProfitableToDupForIfCvt - Return true if it's profitable for
335 /// if-converter to duplicate instructions of specified accumulated
336 /// instruction latencies in the specified MBB to enable if-conversion.
337 /// The probability of the instructions being executed is given by
338 /// Probability, and Confidence is a measure of our confidence that it
339 /// will be properly predicted.
341 isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
342 const BranchProbability &Probability) const {
346 /// isProfitableToUnpredicate - Return true if it's profitable to unpredicate
347 /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually
348 /// exclusive predicates.
356 /// This may be profitable is conditional instructions are always executed.
357 virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
358 MachineBasicBlock &FMBB) const {
362 /// canInsertSelect - Return true if it is possible to insert a select
363 /// instruction that chooses between TrueReg and FalseReg based on the
364 /// condition code in Cond.
366 /// When successful, also return the latency in cycles from TrueReg,
367 /// FalseReg, and Cond to the destination register. In most cases, a select
368 /// instruction will be 1 cycle, so CondCycles = TrueCycles = FalseCycles = 1
370 /// Some x86 implementations have 2-cycle cmov instructions.
372 /// @param MBB Block where select instruction would be inserted.
373 /// @param Cond Condition returned by AnalyzeBranch.
374 /// @param TrueReg Virtual register to select when Cond is true.
375 /// @param FalseReg Virtual register to select when Cond is false.
376 /// @param CondCycles Latency from Cond+Branch to select output.
377 /// @param TrueCycles Latency from TrueReg to select output.
378 /// @param FalseCycles Latency from FalseReg to select output.
379 virtual bool canInsertSelect(const MachineBasicBlock &MBB,
380 const SmallVectorImpl<MachineOperand> &Cond,
381 unsigned TrueReg, unsigned FalseReg,
383 int &TrueCycles, int &FalseCycles) const {
387 /// insertSelect - Insert a select instruction into MBB before I that will
388 /// copy TrueReg to DstReg when Cond is true, and FalseReg to DstReg when
391 /// This function can only be called after canInsertSelect() returned true.
392 /// The condition in Cond comes from AnalyzeBranch, and it can be assumed
393 /// that the same flags or registers required by Cond are available at the
396 /// @param MBB Block where select instruction should be inserted.
397 /// @param I Insertion point.
398 /// @param DL Source location for debugging.
399 /// @param DstReg Virtual register to be defined by select instruction.
400 /// @param Cond Condition as computed by AnalyzeBranch.
401 /// @param TrueReg Virtual register to copy when Cond is true.
402 /// @param FalseReg Virtual register to copy when Cons is false.
403 virtual void insertSelect(MachineBasicBlock &MBB,
404 MachineBasicBlock::iterator I, DebugLoc DL,
406 const SmallVectorImpl<MachineOperand> &Cond,
407 unsigned TrueReg, unsigned FalseReg) const {
408 llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!");
411 /// analyzeSelect - Analyze the given select instruction, returning true if
412 /// it cannot be understood. It is assumed that MI->isSelect() is true.
414 /// When successful, return the controlling condition and the operands that
415 /// determine the true and false result values.
417 /// Result = SELECT Cond, TrueOp, FalseOp
419 /// Some targets can optimize select instructions, for example by predicating
420 /// the instruction defining one of the operands. Such targets should set
423 /// @param MI Select instruction to analyze.
424 /// @param Cond Condition controlling the select.
425 /// @param TrueOp Operand number of the value selected when Cond is true.
426 /// @param FalseOp Operand number of the value selected when Cond is false.
427 /// @param Optimizable Returned as true if MI is optimizable.
428 /// @returns False on success.
429 virtual bool analyzeSelect(const MachineInstr *MI,
430 SmallVectorImpl<MachineOperand> &Cond,
431 unsigned &TrueOp, unsigned &FalseOp,
432 bool &Optimizable) const {
433 assert(MI && MI->getDesc().isSelect() && "MI must be a select instruction");
437 /// optimizeSelect - Given a select instruction that was understood by
438 /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by
439 /// merging it with one of its operands. Returns NULL on failure.
441 /// When successful, returns the new select instruction. The client is
442 /// responsible for deleting MI.
444 /// If both sides of the select can be optimized, PreferFalse is used to pick
447 /// @param MI Optimizable select instruction.
448 /// @param PreferFalse Try to optimize FalseOp instead of TrueOp.
449 /// @returns Optimized instruction or NULL.
450 virtual MachineInstr *optimizeSelect(MachineInstr *MI,
451 bool PreferFalse = false) const {
452 // This function must be implemented if Optimizable is ever set.
453 llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!");
456 /// copyPhysReg - Emit instructions to copy a pair of physical registers.
458 /// This function should support copies within any legal register class as
459 /// well as any cross-class copies created during instruction selection.
461 /// The source and destination registers may overlap, which may require a
462 /// careful implementation when multiple copy instructions are required for
463 /// large registers. See for example the ARM target.
464 virtual void copyPhysReg(MachineBasicBlock &MBB,
465 MachineBasicBlock::iterator MI, DebugLoc DL,
466 unsigned DestReg, unsigned SrcReg,
467 bool KillSrc) const {
468 llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!");
471 /// storeRegToStackSlot - Store the specified register of the given register
472 /// class to the specified stack frame index. The store instruction is to be
473 /// added to the given machine basic block before the specified machine
474 /// instruction. If isKill is true, the register operand is the last use and
475 /// must be marked kill.
476 virtual void storeRegToStackSlot(MachineBasicBlock &MBB,
477 MachineBasicBlock::iterator MI,
478 unsigned SrcReg, bool isKill, int FrameIndex,
479 const TargetRegisterClass *RC,
480 const TargetRegisterInfo *TRI) const {
481 llvm_unreachable("Target didn't implement "
482 "TargetInstrInfo::storeRegToStackSlot!");
485 /// loadRegFromStackSlot - Load the specified register of the given register
486 /// class from the specified stack frame index. The load instruction is to be
487 /// added to the given machine basic block before the specified machine
489 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB,
490 MachineBasicBlock::iterator MI,
491 unsigned DestReg, int FrameIndex,
492 const TargetRegisterClass *RC,
493 const TargetRegisterInfo *TRI) const {
494 llvm_unreachable("Target didn't implement "
495 "TargetInstrInfo::loadRegFromStackSlot!");
498 /// expandPostRAPseudo - This function is called for all pseudo instructions
499 /// that remain after register allocation. Many pseudo instructions are
500 /// created to help register allocation. This is the place to convert them
501 /// into real instructions. The target can edit MI in place, or it can insert
502 /// new instructions and erase MI. The function should return true if
503 /// anything was changed.
504 virtual bool expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
508 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
509 /// slot into the specified machine instruction for the specified operand(s).
510 /// If this is possible, a new instruction is returned with the specified
511 /// operand folded, otherwise NULL is returned.
512 /// The new instruction is inserted before MI, and the client is responsible
513 /// for removing the old instruction.
514 MachineInstr* foldMemoryOperand(MachineBasicBlock::iterator MI,
515 const SmallVectorImpl<unsigned> &Ops,
516 int FrameIndex) const;
518 /// foldMemoryOperand - Same as the previous version except it allows folding
519 /// of any load and store from / to any address, not just from a specific
521 MachineInstr* foldMemoryOperand(MachineBasicBlock::iterator MI,
522 const SmallVectorImpl<unsigned> &Ops,
523 MachineInstr* LoadMI) const;
526 /// foldMemoryOperandImpl - Target-dependent implementation for
527 /// foldMemoryOperand. Target-independent code in foldMemoryOperand will
528 /// take care of adding a MachineMemOperand to the newly created instruction.
529 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
531 const SmallVectorImpl<unsigned> &Ops,
532 int FrameIndex) const {
536 /// foldMemoryOperandImpl - Target-dependent implementation for
537 /// foldMemoryOperand. Target-independent code in foldMemoryOperand will
538 /// take care of adding a MachineMemOperand to the newly created instruction.
539 virtual MachineInstr* foldMemoryOperandImpl(MachineFunction &MF,
541 const SmallVectorImpl<unsigned> &Ops,
542 MachineInstr* LoadMI) const {
547 /// canFoldMemoryOperand - Returns true for the specified load / store if
548 /// folding is possible.
550 bool canFoldMemoryOperand(const MachineInstr *MI,
551 const SmallVectorImpl<unsigned> &Ops) const;
553 /// unfoldMemoryOperand - Separate a single instruction which folded a load or
554 /// a store or a load and a store into two or more instruction. If this is
555 /// possible, returns true as well as the new instructions by reference.
556 virtual bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
557 unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
558 SmallVectorImpl<MachineInstr*> &NewMIs) const{
562 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
563 SmallVectorImpl<SDNode*> &NewNodes) const {
567 /// getOpcodeAfterMemoryUnfold - Returns the opcode of the would be new
568 /// instruction after load / store are unfolded from an instruction of the
569 /// specified opcode. It returns zero if the specified unfolding is not
570 /// possible. If LoadRegIndex is non-null, it is filled in with the operand
571 /// index of the operand which will hold the register holding the loaded
573 virtual unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
574 bool UnfoldLoad, bool UnfoldStore,
575 unsigned *LoadRegIndex = 0) const {
579 /// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler
580 /// to determine if two loads are loading from the same base address. It
581 /// should only return true if the base pointers are the same and the
582 /// only differences between the two addresses are the offset. It also returns
583 /// the offsets by reference.
584 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2,
585 int64_t &Offset1, int64_t &Offset2) const {
589 /// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to
590 /// determine (in conjunction with areLoadsFromSameBasePtr) if two loads should
591 /// be scheduled togther. On some targets if two loads are loading from
592 /// addresses in the same cache line, it's better if they are scheduled
593 /// together. This function takes two integers that represent the load offsets
594 /// from the common base address. It returns true if it decides it's desirable
595 /// to schedule the two loads together. "NumLoads" is the number of loads that
596 /// have already been scheduled after Load1.
597 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
598 int64_t Offset1, int64_t Offset2,
599 unsigned NumLoads) const {
603 /// \brief Get the base register and byte offset of a load/store instr.
604 virtual bool getLdStBaseRegImmOfs(MachineInstr *LdSt,
605 unsigned &BaseReg, unsigned &Offset,
606 const TargetRegisterInfo *TRI) const {
610 virtual bool shouldClusterLoads(MachineInstr *FirstLdSt,
611 MachineInstr *SecondLdSt,
612 unsigned NumLoads) const {
616 /// \brief Can this target fuse the given instructions if they are scheduled
618 virtual bool shouldScheduleAdjacent(MachineInstr* First,
619 MachineInstr *Second) const {
623 /// ReverseBranchCondition - Reverses the branch condition of the specified
624 /// condition list, returning false on success and true if it cannot be
627 bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
631 /// insertNoop - Insert a noop into the instruction stream at the specified
633 virtual void insertNoop(MachineBasicBlock &MBB,
634 MachineBasicBlock::iterator MI) const;
637 /// getNoopForMachoTarget - Return the noop instruction to use for a noop.
638 virtual void getNoopForMachoTarget(MCInst &NopInst) const {
639 // Default to just using 'nop' string.
643 /// isPredicated - Returns true if the instruction is already predicated.
645 virtual bool isPredicated(const MachineInstr *MI) const {
649 /// isUnpredicatedTerminator - Returns true if the instruction is a
650 /// terminator instruction that has not been predicated.
651 virtual bool isUnpredicatedTerminator(const MachineInstr *MI) const;
653 /// PredicateInstruction - Convert the instruction into a predicated
654 /// instruction. It returns true if the operation was successful.
656 bool PredicateInstruction(MachineInstr *MI,
657 const SmallVectorImpl<MachineOperand> &Pred) const;
659 /// SubsumesPredicate - Returns true if the first specified predicate
660 /// subsumes the second, e.g. GE subsumes GT.
662 bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
663 const SmallVectorImpl<MachineOperand> &Pred2) const {
667 /// DefinesPredicate - If the specified instruction defines any predicate
668 /// or condition code register(s) used for predication, returns true as well
669 /// as the definition predicate(s) by reference.
670 virtual bool DefinesPredicate(MachineInstr *MI,
671 std::vector<MachineOperand> &Pred) const {
675 /// isPredicable - Return true if the specified instruction can be predicated.
676 /// By default, this returns true for every instruction with a
677 /// PredicateOperand.
678 virtual bool isPredicable(MachineInstr *MI) const {
679 return MI->getDesc().isPredicable();
682 /// isSafeToMoveRegClassDefs - Return true if it's safe to move a machine
683 /// instruction that defines the specified register class.
684 virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
688 /// isSchedulingBoundary - Test if the given instruction should be
689 /// considered a scheduling boundary. This primarily includes labels and
691 virtual bool isSchedulingBoundary(const MachineInstr *MI,
692 const MachineBasicBlock *MBB,
693 const MachineFunction &MF) const;
695 /// Measure the specified inline asm to determine an approximation of its
697 virtual unsigned getInlineAsmLength(const char *Str,
698 const MCAsmInfo &MAI) const;
700 /// CreateTargetHazardRecognizer - Allocate and return a hazard recognizer to
701 /// use for this target when scheduling the machine instructions before
702 /// register allocation.
703 virtual ScheduleHazardRecognizer*
704 CreateTargetHazardRecognizer(const TargetMachine *TM,
705 const ScheduleDAG *DAG) const;
707 /// CreateTargetMIHazardRecognizer - Allocate and return a hazard recognizer
708 /// to use for this target when scheduling the machine instructions before
709 /// register allocation.
710 virtual ScheduleHazardRecognizer*
711 CreateTargetMIHazardRecognizer(const InstrItineraryData*,
712 const ScheduleDAG *DAG) const;
714 /// CreateTargetPostRAHazardRecognizer - Allocate and return a hazard
715 /// recognizer to use for this target when scheduling the machine instructions
716 /// after register allocation.
717 virtual ScheduleHazardRecognizer*
718 CreateTargetPostRAHazardRecognizer(const InstrItineraryData*,
719 const ScheduleDAG *DAG) const;
721 /// Provide a global flag for disabling the PreRA hazard recognizer that
722 /// targets may choose to honor.
723 bool usePreRAHazardRecognizer() const;
725 /// analyzeCompare - For a comparison instruction, return the source registers
726 /// in SrcReg and SrcReg2 if having two register operands, and the value it
727 /// compares against in CmpValue. Return true if the comparison instruction
729 virtual bool analyzeCompare(const MachineInstr *MI,
730 unsigned &SrcReg, unsigned &SrcReg2,
731 int &Mask, int &Value) const {
735 /// optimizeCompareInstr - See if the comparison instruction can be converted
736 /// into something more efficient. E.g., on ARM most instructions can set the
737 /// flags register, obviating the need for a separate CMP.
738 virtual bool optimizeCompareInstr(MachineInstr *CmpInstr,
739 unsigned SrcReg, unsigned SrcReg2,
741 const MachineRegisterInfo *MRI) const {
745 /// optimizeLoadInstr - Try to remove the load by folding it to a register
746 /// operand at the use. We fold the load instructions if and only if the
747 /// def and use are in the same BB. We only look at one load and see
748 /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register
749 /// defined by the load we are trying to fold. DefMI returns the machine
750 /// instruction that defines FoldAsLoadDefReg, and the function returns
751 /// the machine instruction generated due to folding.
752 virtual MachineInstr* optimizeLoadInstr(MachineInstr *MI,
753 const MachineRegisterInfo *MRI,
754 unsigned &FoldAsLoadDefReg,
755 MachineInstr *&DefMI) const {
759 /// FoldImmediate - 'Reg' is known to be defined by a move immediate
760 /// instruction, try to fold the immediate into the use instruction.
761 /// If MRI->hasOneNonDBGUse(Reg) is true, and this function returns true,
762 /// then the caller may assume that DefMI has been erased from its parent
763 /// block. The caller may assume that it will not be erased by this
764 /// function otherwise.
765 virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
766 unsigned Reg, MachineRegisterInfo *MRI) const {
770 /// getNumMicroOps - Return the number of u-operations the given machine
771 /// instruction will be decoded to on the target cpu. The itinerary's
772 /// IssueWidth is the number of microops that can be dispatched each
773 /// cycle. An instruction with zero microops takes no dispatch resources.
774 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData,
775 const MachineInstr *MI) const;
777 /// isZeroCost - Return true for pseudo instructions that don't consume any
778 /// machine resources in their current form. These are common cases that the
779 /// scheduler should consider free, rather than conservatively handling them
780 /// as instructions with no itinerary.
781 bool isZeroCost(unsigned Opcode) const {
782 return Opcode <= TargetOpcode::COPY;
785 virtual int getOperandLatency(const InstrItineraryData *ItinData,
786 SDNode *DefNode, unsigned DefIdx,
787 SDNode *UseNode, unsigned UseIdx) const;
789 /// getOperandLatency - Compute and return the use operand latency of a given
790 /// pair of def and use.
791 /// In most cases, the static scheduling itinerary was enough to determine the
792 /// operand latency. But it may not be possible for instructions with variable
793 /// number of defs / uses.
795 /// This is a raw interface to the itinerary that may be directly overriden by
796 /// a target. Use computeOperandLatency to get the best estimate of latency.
797 virtual int getOperandLatency(const InstrItineraryData *ItinData,
798 const MachineInstr *DefMI, unsigned DefIdx,
799 const MachineInstr *UseMI,
800 unsigned UseIdx) const;
802 /// computeOperandLatency - Compute and return the latency of the given data
803 /// dependent def and use when the operand indices are already known.
804 unsigned computeOperandLatency(const InstrItineraryData *ItinData,
805 const MachineInstr *DefMI, unsigned DefIdx,
806 const MachineInstr *UseMI, unsigned UseIdx)
809 /// getInstrLatency - Compute the instruction latency of a given instruction.
810 /// If the instruction has higher cost when predicated, it's returned via
812 virtual unsigned getInstrLatency(const InstrItineraryData *ItinData,
813 const MachineInstr *MI,
814 unsigned *PredCost = 0) const;
816 virtual int getInstrLatency(const InstrItineraryData *ItinData,
819 /// Return the default expected latency for a def based on it's opcode.
820 unsigned defaultDefLatency(const MCSchedModel *SchedModel,
821 const MachineInstr *DefMI) const;
823 int computeDefOperandLatency(const InstrItineraryData *ItinData,
824 const MachineInstr *DefMI) const;
826 /// isHighLatencyDef - Return true if this opcode has high latency to its
828 virtual bool isHighLatencyDef(int opc) const { return false; }
830 /// hasHighOperandLatency - Compute operand latency between a def of 'Reg'
831 /// and an use in the current loop, return true if the target considered
832 /// it 'high'. This is used by optimization passes such as machine LICM to
833 /// determine whether it makes sense to hoist an instruction out even in
834 /// high register pressure situation.
836 bool hasHighOperandLatency(const InstrItineraryData *ItinData,
837 const MachineRegisterInfo *MRI,
838 const MachineInstr *DefMI, unsigned DefIdx,
839 const MachineInstr *UseMI, unsigned UseIdx) const {
843 /// hasLowDefLatency - Compute operand latency of a def of 'Reg', return true
844 /// if the target considered it 'low'.
846 bool hasLowDefLatency(const InstrItineraryData *ItinData,
847 const MachineInstr *DefMI, unsigned DefIdx) const;
849 /// verifyInstruction - Perform target specific instruction verification.
851 bool verifyInstruction(const MachineInstr *MI, StringRef &ErrInfo) const {
855 /// getExecutionDomain - Return the current execution domain and bit mask of
856 /// possible domains for instruction.
858 /// Some micro-architectures have multiple execution domains, and multiple
859 /// opcodes that perform the same operation in different domains. For
860 /// example, the x86 architecture provides the por, orps, and orpd
861 /// instructions that all do the same thing. There is a latency penalty if a
862 /// register is written in one domain and read in another.
864 /// This function returns a pair (domain, mask) containing the execution
865 /// domain of MI, and a bit mask of possible domains. The setExecutionDomain
866 /// function can be used to change the opcode to one of the domains in the
867 /// bit mask. Instructions whose execution domain can't be changed should
870 /// The execution domain numbers don't have any special meaning except domain
871 /// 0 is used for instructions that are not associated with any interesting
872 /// execution domain.
874 virtual std::pair<uint16_t, uint16_t>
875 getExecutionDomain(const MachineInstr *MI) const {
876 return std::make_pair(0, 0);
879 /// setExecutionDomain - Change the opcode of MI to execute in Domain.
881 /// The bit (1 << Domain) must be set in the mask returned from
882 /// getExecutionDomain(MI).
884 virtual void setExecutionDomain(MachineInstr *MI, unsigned Domain) const {}
887 /// getPartialRegUpdateClearance - Returns the preferred minimum clearance
888 /// before an instruction with an unwanted partial register update.
890 /// Some instructions only write part of a register, and implicitly need to
891 /// read the other parts of the register. This may cause unwanted stalls
892 /// preventing otherwise unrelated instructions from executing in parallel in
893 /// an out-of-order CPU.
895 /// For example, the x86 instruction cvtsi2ss writes its result to bits
896 /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so
897 /// the instruction needs to wait for the old value of the register to become
900 /// addps %xmm1, %xmm0
901 /// movaps %xmm0, (%rax)
902 /// cvtsi2ss %rbx, %xmm0
904 /// In the code above, the cvtsi2ss instruction needs to wait for the addps
905 /// instruction before it can issue, even though the high bits of %xmm0
906 /// probably aren't needed.
908 /// This hook returns the preferred clearance before MI, measured in
909 /// instructions. Other defs of MI's operand OpNum are avoided in the last N
910 /// instructions before MI. It should only return a positive value for
911 /// unwanted dependencies. If the old bits of the defined register have
912 /// useful values, or if MI is determined to otherwise read the dependency,
913 /// the hook should return 0.
915 /// The unwanted dependency may be handled by:
917 /// 1. Allocating the same register for an MI def and use. That makes the
918 /// unwanted dependency identical to a required dependency.
920 /// 2. Allocating a register for the def that has no defs in the previous N
923 /// 3. Calling breakPartialRegDependency() with the same arguments. This
924 /// allows the target to insert a dependency breaking instruction.
927 getPartialRegUpdateClearance(const MachineInstr *MI, unsigned OpNum,
928 const TargetRegisterInfo *TRI) const {
929 // The default implementation returns 0 for no partial register dependency.
933 /// breakPartialRegDependency - Insert a dependency-breaking instruction
934 /// before MI to eliminate an unwanted dependency on OpNum.
936 /// If it wasn't possible to avoid a def in the last N instructions before MI
937 /// (see getPartialRegUpdateClearance), this hook will be called to break the
938 /// unwanted dependency.
940 /// On x86, an xorps instruction can be used as a dependency breaker:
942 /// addps %xmm1, %xmm0
943 /// movaps %xmm0, (%rax)
944 /// xorps %xmm0, %xmm0
945 /// cvtsi2ss %rbx, %xmm0
947 /// An <imp-kill> operand should be added to MI if an instruction was
948 /// inserted. This ties the instructions together in the post-ra scheduler.
951 breakPartialRegDependency(MachineBasicBlock::iterator MI, unsigned OpNum,
952 const TargetRegisterInfo *TRI) const {}
954 /// Create machine specific model for scheduling.
955 virtual DFAPacketizer*
956 CreateTargetScheduleState(const TargetMachine*, const ScheduleDAG*) const {
961 int CallFrameSetupOpcode, CallFrameDestroyOpcode;
964 } // End llvm namespace