1 //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===//
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 implements the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Target/TargetInstrInfo.h"
15 #include "llvm/CodeGen/MachineFrameInfo.h"
16 #include "llvm/CodeGen/MachineInstrBuilder.h"
17 #include "llvm/CodeGen/MachineMemOperand.h"
18 #include "llvm/CodeGen/MachineRegisterInfo.h"
19 #include "llvm/CodeGen/PseudoSourceValue.h"
20 #include "llvm/CodeGen/ScoreboardHazardRecognizer.h"
21 #include "llvm/CodeGen/StackMaps.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/MC/MCAsmInfo.h"
24 #include "llvm/MC/MCInstrItineraries.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetFrameLowering.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetRegisterInfo.h"
35 static cl::opt<bool> DisableHazardRecognizer(
36 "disable-sched-hazard", cl::Hidden, cl::init(false),
37 cl::desc("Disable hazard detection during preRA scheduling"));
39 TargetInstrInfo::~TargetInstrInfo() {
42 const TargetRegisterClass*
43 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
44 const TargetRegisterInfo *TRI,
45 const MachineFunction &MF) const {
46 if (OpNum >= MCID.getNumOperands())
49 short RegClass = MCID.OpInfo[OpNum].RegClass;
50 if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
51 return TRI->getPointerRegClass(MF, RegClass);
53 // Instructions like INSERT_SUBREG do not have fixed register classes.
57 // Otherwise just look it up normally.
58 return TRI->getRegClass(RegClass);
61 /// insertNoop - Insert a noop into the instruction stream at the specified
63 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
64 MachineBasicBlock::iterator MI) const {
65 llvm_unreachable("Target didn't implement insertNoop!");
68 /// Measure the specified inline asm to determine an approximation of its
70 /// Comments (which run till the next SeparatorString or newline) do not
71 /// count as an instruction.
72 /// Any other non-whitespace text is considered an instruction, with
73 /// multiple instructions separated by SeparatorString or newlines.
74 /// Variable-length instructions are not handled here; this function
75 /// may be overloaded in the target code to do that.
76 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str,
77 const MCAsmInfo &MAI) const {
80 // Count the number of instructions in the asm.
81 bool atInsnStart = true;
84 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
85 strlen(MAI.getSeparatorString())) == 0)
87 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
88 Length += MAI.getMaxInstLength();
91 if (atInsnStart && strncmp(Str, MAI.getCommentString(),
92 strlen(MAI.getCommentString())) == 0)
99 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
100 /// after it, replacing it with an unconditional branch to NewDest.
102 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
103 MachineBasicBlock *NewDest) const {
104 MachineBasicBlock *MBB = Tail->getParent();
106 // Remove all the old successors of MBB from the CFG.
107 while (!MBB->succ_empty())
108 MBB->removeSuccessor(MBB->succ_begin());
110 // Remove all the dead instructions from the end of MBB.
111 MBB->erase(Tail, MBB->end());
113 // If MBB isn't immediately before MBB, insert a branch to it.
114 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
115 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(),
116 Tail->getDebugLoc());
117 MBB->addSuccessor(NewDest);
120 // commuteInstruction - The default implementation of this method just exchanges
121 // the two operands returned by findCommutedOpIndices.
122 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI,
124 const MCInstrDesc &MCID = MI->getDesc();
125 bool HasDef = MCID.getNumDefs();
126 if (HasDef && !MI->getOperand(0).isReg())
127 // No idea how to commute this instruction. Target should implement its own.
130 if (!findCommutedOpIndices(MI, Idx1, Idx2)) {
131 assert(MI->isCommutable() && "Precondition violation: MI must be commutable.");
135 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() &&
136 "This only knows how to commute register operands so far");
137 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
138 unsigned Reg1 = MI->getOperand(Idx1).getReg();
139 unsigned Reg2 = MI->getOperand(Idx2).getReg();
140 unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0;
141 unsigned SubReg1 = MI->getOperand(Idx1).getSubReg();
142 unsigned SubReg2 = MI->getOperand(Idx2).getSubReg();
143 bool Reg1IsKill = MI->getOperand(Idx1).isKill();
144 bool Reg2IsKill = MI->getOperand(Idx2).isKill();
145 bool Reg1IsUndef = MI->getOperand(Idx1).isUndef();
146 bool Reg2IsUndef = MI->getOperand(Idx2).isUndef();
147 // If destination is tied to either of the commuted source register, then
148 // it must be updated.
149 if (HasDef && Reg0 == Reg1 &&
150 MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
154 } else if (HasDef && Reg0 == Reg2 &&
155 MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
162 // Create a new instruction.
163 MachineFunction &MF = *MI->getParent()->getParent();
164 MI = MF.CloneMachineInstr(MI);
168 MI->getOperand(0).setReg(Reg0);
169 MI->getOperand(0).setSubReg(SubReg0);
171 MI->getOperand(Idx2).setReg(Reg1);
172 MI->getOperand(Idx1).setReg(Reg2);
173 MI->getOperand(Idx2).setSubReg(SubReg1);
174 MI->getOperand(Idx1).setSubReg(SubReg2);
175 MI->getOperand(Idx2).setIsKill(Reg1IsKill);
176 MI->getOperand(Idx1).setIsKill(Reg2IsKill);
177 MI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
178 MI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
182 /// findCommutedOpIndices - If specified MI is commutable, return the two
183 /// operand indices that would swap value. Return true if the instruction
184 /// is not in a form which this routine understands.
185 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI,
187 unsigned &SrcOpIdx2) const {
188 assert(!MI->isBundle() &&
189 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
191 const MCInstrDesc &MCID = MI->getDesc();
192 if (!MCID.isCommutable())
194 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
195 // is not true, then the target must implement this.
196 SrcOpIdx1 = MCID.getNumDefs();
197 SrcOpIdx2 = SrcOpIdx1 + 1;
198 if (!MI->getOperand(SrcOpIdx1).isReg() ||
199 !MI->getOperand(SrcOpIdx2).isReg())
207 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
208 if (!MI->isTerminator()) return false;
210 // Conditional branch is a special case.
211 if (MI->isBranch() && !MI->isBarrier())
213 if (!MI->isPredicable())
215 return !isPredicated(MI);
219 bool TargetInstrInfo::PredicateInstruction(MachineInstr *MI,
220 const SmallVectorImpl<MachineOperand> &Pred) const {
221 bool MadeChange = false;
223 assert(!MI->isBundle() &&
224 "TargetInstrInfo::PredicateInstruction() can't handle bundles");
226 const MCInstrDesc &MCID = MI->getDesc();
227 if (!MI->isPredicable())
230 for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) {
231 if (MCID.OpInfo[i].isPredicate()) {
232 MachineOperand &MO = MI->getOperand(i);
234 MO.setReg(Pred[j].getReg());
236 } else if (MO.isImm()) {
237 MO.setImm(Pred[j].getImm());
239 } else if (MO.isMBB()) {
240 MO.setMBB(Pred[j].getMBB());
249 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
250 const MachineMemOperand *&MMO,
251 int &FrameIndex) const {
252 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
253 oe = MI->memoperands_end();
256 if ((*o)->isLoad()) {
257 if (const FixedStackPseudoSourceValue *Value =
258 dyn_cast_or_null<FixedStackPseudoSourceValue>(
259 (*o)->getPseudoValue())) {
260 FrameIndex = Value->getFrameIndex();
269 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
270 const MachineMemOperand *&MMO,
271 int &FrameIndex) const {
272 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
273 oe = MI->memoperands_end();
276 if ((*o)->isStore()) {
277 if (const FixedStackPseudoSourceValue *Value =
278 dyn_cast_or_null<FixedStackPseudoSourceValue>(
279 (*o)->getPseudoValue())) {
280 FrameIndex = Value->getFrameIndex();
289 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
290 unsigned SubIdx, unsigned &Size,
292 const MachineFunction &MF) const {
294 Size = RC->getSize();
298 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
299 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
300 // Convert bit size to byte size to be consistent with
301 // MCRegisterClass::getSize().
305 int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
306 if (BitOffset < 0 || BitOffset % 8)
310 Offset = (unsigned)BitOffset / 8;
312 assert(RC->getSize() >= (Offset + Size) && "bad subregister range");
314 if (!MF.getTarget().getDataLayout()->isLittleEndian()) {
315 Offset = RC->getSize() - (Offset + Size);
320 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
321 MachineBasicBlock::iterator I,
324 const MachineInstr *Orig,
325 const TargetRegisterInfo &TRI) const {
326 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
327 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
332 TargetInstrInfo::produceSameValue(const MachineInstr *MI0,
333 const MachineInstr *MI1,
334 const MachineRegisterInfo *MRI) const {
335 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
338 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig,
339 MachineFunction &MF) const {
340 assert(!Orig->isNotDuplicable() &&
341 "Instruction cannot be duplicated");
342 return MF.CloneMachineInstr(Orig);
345 // If the COPY instruction in MI can be folded to a stack operation, return
346 // the register class to use.
347 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI,
349 assert(MI->isCopy() && "MI must be a COPY instruction");
350 if (MI->getNumOperands() != 2)
352 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
354 const MachineOperand &FoldOp = MI->getOperand(FoldIdx);
355 const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx);
357 if (FoldOp.getSubReg() || LiveOp.getSubReg())
360 unsigned FoldReg = FoldOp.getReg();
361 unsigned LiveReg = LiveOp.getReg();
363 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
364 "Cannot fold physregs");
366 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
367 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
369 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
370 return RC->contains(LiveOp.getReg()) ? RC : nullptr;
372 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
375 // FIXME: Allow folding when register classes are memory compatible.
379 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
380 llvm_unreachable("Not a MachO target");
383 bool TargetInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
384 ArrayRef<unsigned> Ops) const {
385 return MI->isCopy() && Ops.size() == 1 && canFoldCopy(MI, Ops[0]);
388 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI,
389 ArrayRef<unsigned> Ops, int FrameIndex,
390 const TargetInstrInfo &TII) {
391 unsigned StartIdx = 0;
392 switch (MI->getOpcode()) {
393 case TargetOpcode::STACKMAP:
394 StartIdx = 2; // Skip ID, nShadowBytes.
396 case TargetOpcode::PATCHPOINT: {
397 // For PatchPoint, the call args are not foldable.
398 PatchPointOpers opers(MI);
399 StartIdx = opers.getVarIdx();
403 llvm_unreachable("unexpected stackmap opcode");
406 // Return false if any operands requested for folding are not foldable (not
407 // part of the stackmap's live values).
408 for (unsigned Op : Ops) {
413 MachineInstr *NewMI =
414 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true);
415 MachineInstrBuilder MIB(MF, NewMI);
417 // No need to fold return, the meta data, and function arguments
418 for (unsigned i = 0; i < StartIdx; ++i)
419 MIB.addOperand(MI->getOperand(i));
421 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) {
422 MachineOperand &MO = MI->getOperand(i);
423 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
425 unsigned SpillOffset;
426 // Compute the spill slot size and offset.
427 const TargetRegisterClass *RC =
428 MF.getRegInfo().getRegClass(MO.getReg());
430 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
432 report_fatal_error("cannot spill patchpoint subregister operand");
433 MIB.addImm(StackMaps::IndirectMemRefOp);
434 MIB.addImm(SpillSize);
435 MIB.addFrameIndex(FrameIndex);
436 MIB.addImm(SpillOffset);
444 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
445 /// slot into the specified machine instruction for the specified operand(s).
446 /// If this is possible, a new instruction is returned with the specified
447 /// operand folded, otherwise NULL is returned. The client is responsible for
448 /// removing the old instruction and adding the new one in the instruction
450 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
451 ArrayRef<unsigned> Ops,
454 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
455 if (MI->getOperand(Ops[i]).isDef())
456 Flags |= MachineMemOperand::MOStore;
458 Flags |= MachineMemOperand::MOLoad;
460 MachineBasicBlock *MBB = MI->getParent();
461 assert(MBB && "foldMemoryOperand needs an inserted instruction");
462 MachineFunction &MF = *MBB->getParent();
464 MachineInstr *NewMI = nullptr;
466 if (MI->getOpcode() == TargetOpcode::STACKMAP ||
467 MI->getOpcode() == TargetOpcode::PATCHPOINT) {
468 // Fold stackmap/patchpoint.
469 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
471 // Ask the target to do the actual folding.
472 NewMI =foldMemoryOperandImpl(MF, MI, Ops, FI);
476 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
477 // Add a memory operand, foldMemoryOperandImpl doesn't do that.
478 assert((!(Flags & MachineMemOperand::MOStore) ||
479 NewMI->mayStore()) &&
480 "Folded a def to a non-store!");
481 assert((!(Flags & MachineMemOperand::MOLoad) ||
483 "Folded a use to a non-load!");
484 const MachineFrameInfo &MFI = *MF.getFrameInfo();
485 assert(MFI.getObjectOffset(FI) != -1);
486 MachineMemOperand *MMO =
487 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
488 Flags, MFI.getObjectSize(FI),
489 MFI.getObjectAlignment(FI));
490 NewMI->addMemOperand(MF, MMO);
492 // FIXME: change foldMemoryOperandImpl semantics to also insert NewMI.
493 return MBB->insert(MI, NewMI);
496 // Straight COPY may fold as load/store.
497 if (!MI->isCopy() || Ops.size() != 1)
500 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
504 const MachineOperand &MO = MI->getOperand(1-Ops[0]);
505 MachineBasicBlock::iterator Pos = MI;
506 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
508 if (Flags == MachineMemOperand::MOStore)
509 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
511 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
515 /// foldMemoryOperand - Same as the previous version except it allows folding
516 /// of any load and store from / to any address, not just from a specific
518 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
519 ArrayRef<unsigned> Ops,
520 MachineInstr *LoadMI) const {
521 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!");
523 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
524 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!");
526 MachineBasicBlock &MBB = *MI->getParent();
527 MachineFunction &MF = *MBB.getParent();
529 // Ask the target to do the actual folding.
530 MachineInstr *NewMI = nullptr;
533 if ((MI->getOpcode() == TargetOpcode::STACKMAP ||
534 MI->getOpcode() == TargetOpcode::PATCHPOINT) &&
535 isLoadFromStackSlot(LoadMI, FrameIndex)) {
536 // Fold stackmap/patchpoint.
537 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
539 // Ask the target to do the actual folding.
540 NewMI = foldMemoryOperandImpl(MF, MI, Ops, LoadMI);
543 if (!NewMI) return nullptr;
545 NewMI = MBB.insert(MI, NewMI);
547 // Copy the memoperands from the load to the folded instruction.
548 if (MI->memoperands_empty()) {
549 NewMI->setMemRefs(LoadMI->memoperands_begin(),
550 LoadMI->memoperands_end());
553 // Handle the rare case of folding multiple loads.
554 NewMI->setMemRefs(MI->memoperands_begin(),
555 MI->memoperands_end());
556 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(),
557 E = LoadMI->memoperands_end(); I != E; ++I) {
558 NewMI->addMemOperand(MF, *I);
564 bool TargetInstrInfo::
565 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
566 AliasAnalysis *AA) const {
567 const MachineFunction &MF = *MI->getParent()->getParent();
568 const MachineRegisterInfo &MRI = MF.getRegInfo();
570 // Remat clients assume operand 0 is the defined register.
571 if (!MI->getNumOperands() || !MI->getOperand(0).isReg())
573 unsigned DefReg = MI->getOperand(0).getReg();
575 // A sub-register definition can only be rematerialized if the instruction
576 // doesn't read the other parts of the register. Otherwise it is really a
577 // read-modify-write operation on the full virtual register which cannot be
579 if (TargetRegisterInfo::isVirtualRegister(DefReg) &&
580 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg))
583 // A load from a fixed stack slot can be rematerialized. This may be
584 // redundant with subsequent checks, but it's target-independent,
585 // simple, and a common case.
587 if (isLoadFromStackSlot(MI, FrameIdx) &&
588 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx))
591 // Avoid instructions obviously unsafe for remat.
592 if (MI->isNotDuplicable() || MI->mayStore() ||
593 MI->hasUnmodeledSideEffects())
596 // Don't remat inline asm. We have no idea how expensive it is
597 // even if it's side effect free.
598 if (MI->isInlineAsm())
601 // Avoid instructions which load from potentially varying memory.
602 if (MI->mayLoad() && !MI->isInvariantLoad(AA))
605 // If any of the registers accessed are non-constant, conservatively assume
606 // the instruction is not rematerializable.
607 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
608 const MachineOperand &MO = MI->getOperand(i);
609 if (!MO.isReg()) continue;
610 unsigned Reg = MO.getReg();
614 // Check for a well-behaved physical register.
615 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
617 // If the physreg has no defs anywhere, it's just an ambient register
618 // and we can freely move its uses. Alternatively, if it's allocatable,
619 // it could get allocated to something with a def during allocation.
620 if (!MRI.isConstantPhysReg(Reg, MF))
623 // A physreg def. We can't remat it.
629 // Only allow one virtual-register def. There may be multiple defs of the
630 // same virtual register, though.
631 if (MO.isDef() && Reg != DefReg)
634 // Don't allow any virtual-register uses. Rematting an instruction with
635 // virtual register uses would length the live ranges of the uses, which
636 // is not necessarily a good idea, certainly not "trivial".
641 // Everything checked out.
645 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const {
646 const MachineFunction *MF = MI->getParent()->getParent();
647 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
648 bool StackGrowsDown =
649 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
651 int FrameSetupOpcode = getCallFrameSetupOpcode();
652 int FrameDestroyOpcode = getCallFrameDestroyOpcode();
654 if (MI->getOpcode() != FrameSetupOpcode &&
655 MI->getOpcode() != FrameDestroyOpcode)
658 int SPAdj = MI->getOperand(0).getImm();
660 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) ||
661 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode))
667 /// isSchedulingBoundary - Test if the given instruction should be
668 /// considered a scheduling boundary. This primarily includes labels
670 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
671 const MachineBasicBlock *MBB,
672 const MachineFunction &MF) const {
673 // Terminators and labels can't be scheduled around.
674 if (MI->isTerminator() || MI->isPosition())
677 // Don't attempt to schedule around any instruction that defines
678 // a stack-oriented pointer, as it's unlikely to be profitable. This
679 // saves compile time, because it doesn't require every single
680 // stack slot reference to depend on the instruction that does the
682 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
683 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
684 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI))
690 // Provide a global flag for disabling the PreRA hazard recognizer that targets
691 // may choose to honor.
692 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
693 return !DisableHazardRecognizer;
696 // Default implementation of CreateTargetRAHazardRecognizer.
697 ScheduleHazardRecognizer *TargetInstrInfo::
698 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
699 const ScheduleDAG *DAG) const {
700 // Dummy hazard recognizer allows all instructions to issue.
701 return new ScheduleHazardRecognizer();
704 // Default implementation of CreateTargetMIHazardRecognizer.
705 ScheduleHazardRecognizer *TargetInstrInfo::
706 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
707 const ScheduleDAG *DAG) const {
708 return (ScheduleHazardRecognizer *)
709 new ScoreboardHazardRecognizer(II, DAG, "misched");
712 // Default implementation of CreateTargetPostRAHazardRecognizer.
713 ScheduleHazardRecognizer *TargetInstrInfo::
714 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
715 const ScheduleDAG *DAG) const {
716 return (ScheduleHazardRecognizer *)
717 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
720 //===----------------------------------------------------------------------===//
721 // SelectionDAG latency interface.
722 //===----------------------------------------------------------------------===//
725 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
726 SDNode *DefNode, unsigned DefIdx,
727 SDNode *UseNode, unsigned UseIdx) const {
728 if (!ItinData || ItinData->isEmpty())
731 if (!DefNode->isMachineOpcode())
734 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
735 if (!UseNode->isMachineOpcode())
736 return ItinData->getOperandCycle(DefClass, DefIdx);
737 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
738 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
741 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
743 if (!ItinData || ItinData->isEmpty())
746 if (!N->isMachineOpcode())
749 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
752 //===----------------------------------------------------------------------===//
753 // MachineInstr latency interface.
754 //===----------------------------------------------------------------------===//
757 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
758 const MachineInstr *MI) const {
759 if (!ItinData || ItinData->isEmpty())
762 unsigned Class = MI->getDesc().getSchedClass();
763 int UOps = ItinData->Itineraries[Class].NumMicroOps;
767 // The # of u-ops is dynamically determined. The specific target should
768 // override this function to return the right number.
772 /// Return the default expected latency for a def based on it's opcode.
773 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
774 const MachineInstr *DefMI) const {
775 if (DefMI->isTransient())
777 if (DefMI->mayLoad())
778 return SchedModel.LoadLatency;
779 if (isHighLatencyDef(DefMI->getOpcode()))
780 return SchedModel.HighLatency;
784 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const {
788 unsigned TargetInstrInfo::
789 getInstrLatency(const InstrItineraryData *ItinData,
790 const MachineInstr *MI,
791 unsigned *PredCost) const {
792 // Default to one cycle for no itinerary. However, an "empty" itinerary may
793 // still have a MinLatency property, which getStageLatency checks.
795 return MI->mayLoad() ? 2 : 1;
797 return ItinData->getStageLatency(MI->getDesc().getSchedClass());
800 bool TargetInstrInfo::hasLowDefLatency(const InstrItineraryData *ItinData,
801 const MachineInstr *DefMI,
802 unsigned DefIdx) const {
803 if (!ItinData || ItinData->isEmpty())
806 unsigned DefClass = DefMI->getDesc().getSchedClass();
807 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
808 return (DefCycle != -1 && DefCycle <= 1);
811 /// Both DefMI and UseMI must be valid. By default, call directly to the
812 /// itinerary. This may be overriden by the target.
813 int TargetInstrInfo::
814 getOperandLatency(const InstrItineraryData *ItinData,
815 const MachineInstr *DefMI, unsigned DefIdx,
816 const MachineInstr *UseMI, unsigned UseIdx) const {
817 unsigned DefClass = DefMI->getDesc().getSchedClass();
818 unsigned UseClass = UseMI->getDesc().getSchedClass();
819 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
822 /// If we can determine the operand latency from the def only, without itinerary
823 /// lookup, do so. Otherwise return -1.
824 int TargetInstrInfo::computeDefOperandLatency(
825 const InstrItineraryData *ItinData,
826 const MachineInstr *DefMI) const {
828 // Let the target hook getInstrLatency handle missing itineraries.
830 return getInstrLatency(ItinData, DefMI);
832 if(ItinData->isEmpty())
833 return defaultDefLatency(ItinData->SchedModel, DefMI);
835 // ...operand lookup required
839 /// computeOperandLatency - Compute and return the latency of the given data
840 /// dependent def and use when the operand indices are already known. UseMI may
841 /// be NULL for an unknown use.
843 /// FindMin may be set to get the minimum vs. expected latency. Minimum
844 /// latency is used for scheduling groups, while expected latency is for
845 /// instruction cost and critical path.
847 /// Depending on the subtarget's itinerary properties, this may or may not need
848 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or
849 /// UseIdx to compute min latency.
850 unsigned TargetInstrInfo::
851 computeOperandLatency(const InstrItineraryData *ItinData,
852 const MachineInstr *DefMI, unsigned DefIdx,
853 const MachineInstr *UseMI, unsigned UseIdx) const {
855 int DefLatency = computeDefOperandLatency(ItinData, DefMI);
859 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail");
863 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
865 unsigned DefClass = DefMI->getDesc().getSchedClass();
866 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx);
868 if (OperLatency >= 0)
871 // No operand latency was found.
872 unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
874 // Expected latency is the max of the stage latency and itinerary props.
875 InstrLatency = std::max(InstrLatency,
876 defaultDefLatency(ItinData->SchedModel, DefMI));
880 bool TargetInstrInfo::getRegSequenceInputs(
881 const MachineInstr &MI, unsigned DefIdx,
882 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
883 assert((MI.isRegSequence() ||
884 MI.isRegSequenceLike()) && "Instruction do not have the proper type");
886 if (!MI.isRegSequence())
887 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
889 // We are looking at:
890 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
891 assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
892 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
894 const MachineOperand &MOReg = MI.getOperand(OpIdx);
895 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
896 assert(MOSubIdx.isImm() &&
897 "One of the subindex of the reg_sequence is not an immediate");
898 // Record Reg:SubReg, SubIdx.
899 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
900 (unsigned)MOSubIdx.getImm()));
905 bool TargetInstrInfo::getExtractSubregInputs(
906 const MachineInstr &MI, unsigned DefIdx,
907 RegSubRegPairAndIdx &InputReg) const {
908 assert((MI.isExtractSubreg() ||
909 MI.isExtractSubregLike()) && "Instruction do not have the proper type");
911 if (!MI.isExtractSubreg())
912 return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
914 // We are looking at:
915 // Def = EXTRACT_SUBREG v0.sub1, sub0.
916 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
917 const MachineOperand &MOReg = MI.getOperand(1);
918 const MachineOperand &MOSubIdx = MI.getOperand(2);
919 assert(MOSubIdx.isImm() &&
920 "The subindex of the extract_subreg is not an immediate");
922 InputReg.Reg = MOReg.getReg();
923 InputReg.SubReg = MOReg.getSubReg();
924 InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
928 bool TargetInstrInfo::getInsertSubregInputs(
929 const MachineInstr &MI, unsigned DefIdx,
930 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
931 assert((MI.isInsertSubreg() ||
932 MI.isInsertSubregLike()) && "Instruction do not have the proper type");
934 if (!MI.isInsertSubreg())
935 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
937 // We are looking at:
938 // Def = INSERT_SEQUENCE v0, v1, sub0.
939 assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
940 const MachineOperand &MOBaseReg = MI.getOperand(1);
941 const MachineOperand &MOInsertedReg = MI.getOperand(2);
942 const MachineOperand &MOSubIdx = MI.getOperand(3);
943 assert(MOSubIdx.isImm() &&
944 "One of the subindex of the reg_sequence is not an immediate");
945 BaseReg.Reg = MOBaseReg.getReg();
946 BaseReg.SubReg = MOBaseReg.getSubReg();
948 InsertedReg.Reg = MOInsertedReg.getReg();
949 InsertedReg.SubReg = MOInsertedReg.getSubReg();
950 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();