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/CodeGen/TargetSchedule.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/MC/MCAsmInfo.h"
25 #include "llvm/MC/MCInstrItineraries.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Target/TargetFrameLowering.h"
30 #include "llvm/Target/TargetLowering.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
36 static cl::opt<bool> DisableHazardRecognizer(
37 "disable-sched-hazard", cl::Hidden, cl::init(false),
38 cl::desc("Disable hazard detection during preRA scheduling"));
40 TargetInstrInfo::~TargetInstrInfo() {
43 const TargetRegisterClass*
44 TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum,
45 const TargetRegisterInfo *TRI,
46 const MachineFunction &MF) const {
47 if (OpNum >= MCID.getNumOperands())
50 short RegClass = MCID.OpInfo[OpNum].RegClass;
51 if (MCID.OpInfo[OpNum].isLookupPtrRegClass())
52 return TRI->getPointerRegClass(MF, RegClass);
54 // Instructions like INSERT_SUBREG do not have fixed register classes.
58 // Otherwise just look it up normally.
59 return TRI->getRegClass(RegClass);
62 /// insertNoop - Insert a noop into the instruction stream at the specified
64 void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB,
65 MachineBasicBlock::iterator MI) const {
66 llvm_unreachable("Target didn't implement insertNoop!");
69 /// Measure the specified inline asm to determine an approximation of its
71 /// Comments (which run till the next SeparatorString or newline) do not
72 /// count as an instruction.
73 /// Any other non-whitespace text is considered an instruction, with
74 /// multiple instructions separated by SeparatorString or newlines.
75 /// Variable-length instructions are not handled here; this function
76 /// may be overloaded in the target code to do that.
77 unsigned TargetInstrInfo::getInlineAsmLength(const char *Str,
78 const MCAsmInfo &MAI) const {
81 // Count the number of instructions in the asm.
82 bool atInsnStart = true;
85 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
86 strlen(MAI.getSeparatorString())) == 0)
88 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
89 Length += MAI.getMaxInstLength();
92 if (atInsnStart && strncmp(Str, MAI.getCommentString(),
93 strlen(MAI.getCommentString())) == 0)
100 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
101 /// after it, replacing it with an unconditional branch to NewDest.
103 TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
104 MachineBasicBlock *NewDest) const {
105 MachineBasicBlock *MBB = Tail->getParent();
107 // Remove all the old successors of MBB from the CFG.
108 while (!MBB->succ_empty())
109 MBB->removeSuccessor(MBB->succ_begin());
111 // Remove all the dead instructions from the end of MBB.
112 MBB->erase(Tail, MBB->end());
114 // If MBB isn't immediately before MBB, insert a branch to it.
115 if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
116 InsertBranch(*MBB, NewDest, nullptr, SmallVector<MachineOperand, 0>(),
117 Tail->getDebugLoc());
118 MBB->addSuccessor(NewDest);
121 // commuteInstruction - The default implementation of this method just exchanges
122 // the two operands returned by findCommutedOpIndices.
123 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI,
125 const MCInstrDesc &MCID = MI->getDesc();
126 bool HasDef = MCID.getNumDefs();
127 if (HasDef && !MI->getOperand(0).isReg())
128 // No idea how to commute this instruction. Target should implement its own.
131 if (!findCommutedOpIndices(MI, Idx1, Idx2)) {
132 assert(MI->isCommutable() && "Precondition violation: MI must be commutable.");
136 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() &&
137 "This only knows how to commute register operands so far");
138 unsigned Reg0 = HasDef ? MI->getOperand(0).getReg() : 0;
139 unsigned Reg1 = MI->getOperand(Idx1).getReg();
140 unsigned Reg2 = MI->getOperand(Idx2).getReg();
141 unsigned SubReg0 = HasDef ? MI->getOperand(0).getSubReg() : 0;
142 unsigned SubReg1 = MI->getOperand(Idx1).getSubReg();
143 unsigned SubReg2 = MI->getOperand(Idx2).getSubReg();
144 bool Reg1IsKill = MI->getOperand(Idx1).isKill();
145 bool Reg2IsKill = MI->getOperand(Idx2).isKill();
146 bool Reg1IsUndef = MI->getOperand(Idx1).isUndef();
147 bool Reg2IsUndef = MI->getOperand(Idx2).isUndef();
148 bool Reg1IsInternal = MI->getOperand(Idx1).isInternalRead();
149 bool Reg2IsInternal = MI->getOperand(Idx2).isInternalRead();
150 // If destination is tied to either of the commuted source register, then
151 // it must be updated.
152 if (HasDef && Reg0 == Reg1 &&
153 MI->getDesc().getOperandConstraint(Idx1, MCOI::TIED_TO) == 0) {
157 } else if (HasDef && Reg0 == Reg2 &&
158 MI->getDesc().getOperandConstraint(Idx2, MCOI::TIED_TO) == 0) {
165 // Create a new instruction.
166 MachineFunction &MF = *MI->getParent()->getParent();
167 MI = MF.CloneMachineInstr(MI);
171 MI->getOperand(0).setReg(Reg0);
172 MI->getOperand(0).setSubReg(SubReg0);
174 MI->getOperand(Idx2).setReg(Reg1);
175 MI->getOperand(Idx1).setReg(Reg2);
176 MI->getOperand(Idx2).setSubReg(SubReg1);
177 MI->getOperand(Idx1).setSubReg(SubReg2);
178 MI->getOperand(Idx2).setIsKill(Reg1IsKill);
179 MI->getOperand(Idx1).setIsKill(Reg2IsKill);
180 MI->getOperand(Idx2).setIsUndef(Reg1IsUndef);
181 MI->getOperand(Idx1).setIsUndef(Reg2IsUndef);
182 MI->getOperand(Idx2).setIsInternalRead(Reg1IsInternal);
183 MI->getOperand(Idx1).setIsInternalRead(Reg2IsInternal);
187 /// findCommutedOpIndices - If specified MI is commutable, return the two
188 /// operand indices that would swap value. Return true if the instruction
189 /// is not in a form which this routine understands.
190 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI,
192 unsigned &SrcOpIdx2) const {
193 assert(!MI->isBundle() &&
194 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
196 const MCInstrDesc &MCID = MI->getDesc();
197 if (!MCID.isCommutable())
199 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
200 // is not true, then the target must implement this.
201 SrcOpIdx1 = MCID.getNumDefs();
202 SrcOpIdx2 = SrcOpIdx1 + 1;
203 if (!MI->getOperand(SrcOpIdx1).isReg() ||
204 !MI->getOperand(SrcOpIdx2).isReg())
212 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
213 if (!MI->isTerminator()) return false;
215 // Conditional branch is a special case.
216 if (MI->isBranch() && !MI->isBarrier())
218 if (!MI->isPredicable())
220 return !isPredicated(MI);
223 bool TargetInstrInfo::PredicateInstruction(
224 MachineInstr *MI, ArrayRef<MachineOperand> Pred) const {
225 bool MadeChange = false;
227 assert(!MI->isBundle() &&
228 "TargetInstrInfo::PredicateInstruction() can't handle bundles");
230 const MCInstrDesc &MCID = MI->getDesc();
231 if (!MI->isPredicable())
234 for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) {
235 if (MCID.OpInfo[i].isPredicate()) {
236 MachineOperand &MO = MI->getOperand(i);
238 MO.setReg(Pred[j].getReg());
240 } else if (MO.isImm()) {
241 MO.setImm(Pred[j].getImm());
243 } else if (MO.isMBB()) {
244 MO.setMBB(Pred[j].getMBB());
253 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
254 const MachineMemOperand *&MMO,
255 int &FrameIndex) const {
256 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
257 oe = MI->memoperands_end();
260 if ((*o)->isLoad()) {
261 if (const FixedStackPseudoSourceValue *Value =
262 dyn_cast_or_null<FixedStackPseudoSourceValue>(
263 (*o)->getPseudoValue())) {
264 FrameIndex = Value->getFrameIndex();
273 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
274 const MachineMemOperand *&MMO,
275 int &FrameIndex) const {
276 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
277 oe = MI->memoperands_end();
280 if ((*o)->isStore()) {
281 if (const FixedStackPseudoSourceValue *Value =
282 dyn_cast_or_null<FixedStackPseudoSourceValue>(
283 (*o)->getPseudoValue())) {
284 FrameIndex = Value->getFrameIndex();
293 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
294 unsigned SubIdx, unsigned &Size,
296 const MachineFunction &MF) const {
298 Size = RC->getSize();
302 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
303 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
304 // Convert bit size to byte size to be consistent with
305 // MCRegisterClass::getSize().
309 int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
310 if (BitOffset < 0 || BitOffset % 8)
314 Offset = (unsigned)BitOffset / 8;
316 assert(RC->getSize() >= (Offset + Size) && "bad subregister range");
318 if (!MF.getDataLayout().isLittleEndian()) {
319 Offset = RC->getSize() - (Offset + Size);
324 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
325 MachineBasicBlock::iterator I,
328 const MachineInstr *Orig,
329 const TargetRegisterInfo &TRI) const {
330 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
331 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
336 TargetInstrInfo::produceSameValue(const MachineInstr *MI0,
337 const MachineInstr *MI1,
338 const MachineRegisterInfo *MRI) const {
339 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
342 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig,
343 MachineFunction &MF) const {
344 assert(!Orig->isNotDuplicable() &&
345 "Instruction cannot be duplicated");
346 return MF.CloneMachineInstr(Orig);
349 // If the COPY instruction in MI can be folded to a stack operation, return
350 // the register class to use.
351 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI,
353 assert(MI->isCopy() && "MI must be a COPY instruction");
354 if (MI->getNumOperands() != 2)
356 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
358 const MachineOperand &FoldOp = MI->getOperand(FoldIdx);
359 const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx);
361 if (FoldOp.getSubReg() || LiveOp.getSubReg())
364 unsigned FoldReg = FoldOp.getReg();
365 unsigned LiveReg = LiveOp.getReg();
367 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
368 "Cannot fold physregs");
370 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
371 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
373 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
374 return RC->contains(LiveOp.getReg()) ? RC : nullptr;
376 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
379 // FIXME: Allow folding when register classes are memory compatible.
383 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
384 llvm_unreachable("Not a MachO target");
387 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI,
388 ArrayRef<unsigned> Ops, int FrameIndex,
389 const TargetInstrInfo &TII) {
390 unsigned StartIdx = 0;
391 switch (MI->getOpcode()) {
392 case TargetOpcode::STACKMAP:
393 StartIdx = 2; // Skip ID, nShadowBytes.
395 case TargetOpcode::PATCHPOINT: {
396 // For PatchPoint, the call args are not foldable.
397 PatchPointOpers opers(MI);
398 StartIdx = opers.getVarIdx();
402 llvm_unreachable("unexpected stackmap opcode");
405 // Return false if any operands requested for folding are not foldable (not
406 // part of the stackmap's live values).
407 for (unsigned Op : Ops) {
412 MachineInstr *NewMI =
413 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true);
414 MachineInstrBuilder MIB(MF, NewMI);
416 // No need to fold return, the meta data, and function arguments
417 for (unsigned i = 0; i < StartIdx; ++i)
418 MIB.addOperand(MI->getOperand(i));
420 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) {
421 MachineOperand &MO = MI->getOperand(i);
422 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
424 unsigned SpillOffset;
425 // Compute the spill slot size and offset.
426 const TargetRegisterClass *RC =
427 MF.getRegInfo().getRegClass(MO.getReg());
429 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
431 report_fatal_error("cannot spill patchpoint subregister operand");
432 MIB.addImm(StackMaps::IndirectMemRefOp);
433 MIB.addImm(SpillSize);
434 MIB.addFrameIndex(FrameIndex);
435 MIB.addImm(SpillOffset);
443 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
444 /// slot into the specified machine instruction for the specified operand(s).
445 /// If this is possible, a new instruction is returned with the specified
446 /// operand folded, otherwise NULL is returned. The client is responsible for
447 /// removing the old instruction and adding the new one in the instruction
449 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
450 ArrayRef<unsigned> Ops,
453 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
454 if (MI->getOperand(Ops[i]).isDef())
455 Flags |= MachineMemOperand::MOStore;
457 Flags |= MachineMemOperand::MOLoad;
459 MachineBasicBlock *MBB = MI->getParent();
460 assert(MBB && "foldMemoryOperand needs an inserted instruction");
461 MachineFunction &MF = *MBB->getParent();
463 MachineInstr *NewMI = nullptr;
465 if (MI->getOpcode() == TargetOpcode::STACKMAP ||
466 MI->getOpcode() == TargetOpcode::PATCHPOINT) {
467 // Fold stackmap/patchpoint.
468 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
470 MBB->insert(MI, NewMI);
472 // Ask the target to do the actual folding.
473 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI);
477 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
478 // Add a memory operand, foldMemoryOperandImpl doesn't do that.
479 assert((!(Flags & MachineMemOperand::MOStore) ||
480 NewMI->mayStore()) &&
481 "Folded a def to a non-store!");
482 assert((!(Flags & MachineMemOperand::MOLoad) ||
484 "Folded a use to a non-load!");
485 const MachineFrameInfo &MFI = *MF.getFrameInfo();
486 assert(MFI.getObjectOffset(FI) != -1);
487 MachineMemOperand *MMO = MF.getMachineMemOperand(
488 MachinePointerInfo::getFixedStack(MF, FI), Flags, MFI.getObjectSize(FI),
489 MFI.getObjectAlignment(FI));
490 NewMI->addMemOperand(MF, MMO);
495 // Straight COPY may fold as load/store.
496 if (!MI->isCopy() || Ops.size() != 1)
499 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
503 const MachineOperand &MO = MI->getOperand(1-Ops[0]);
504 MachineBasicBlock::iterator Pos = MI;
505 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
507 if (Flags == MachineMemOperand::MOStore)
508 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
510 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
514 bool TargetInstrInfo::hasReassociableOperands(
515 const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
516 const MachineOperand &Op1 = Inst.getOperand(1);
517 const MachineOperand &Op2 = Inst.getOperand(2);
518 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
520 // We need virtual register definitions for the operands that we will
522 MachineInstr *MI1 = nullptr;
523 MachineInstr *MI2 = nullptr;
524 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
525 MI1 = MRI.getUniqueVRegDef(Op1.getReg());
526 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
527 MI2 = MRI.getUniqueVRegDef(Op2.getReg());
529 // And they need to be in the trace (otherwise, they won't have a depth).
530 if (MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB)
536 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
537 bool &Commuted) const {
538 const MachineBasicBlock *MBB = Inst.getParent();
539 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
540 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
541 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
542 unsigned AssocOpcode = Inst.getOpcode();
544 // If only one operand has the same opcode and it's the second source operand,
545 // the operands must be commuted.
546 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
550 // 1. The previous instruction must be the same type as Inst.
551 // 2. The previous instruction must have virtual register definitions for its
552 // operands in the same basic block as Inst.
553 // 3. The previous instruction's result must only be used by Inst.
554 if (MI1->getOpcode() == AssocOpcode && hasReassociableOperands(*MI1, MBB) &&
555 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()))
561 // 1. The operation must be associative and commutative.
562 // 2. The instruction must have virtual register definitions for its
563 // operands in the same basic block.
564 // 3. The instruction must have a reassociable sibling.
565 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
566 bool &Commuted) const {
567 if (isAssociativeAndCommutative(Inst) &&
568 hasReassociableOperands(Inst, Inst.getParent()) &&
569 hasReassociableSibling(Inst, Commuted))
575 // The concept of the reassociation pass is that these operations can benefit
576 // from this kind of transformation:
586 // breaking the dependency between A and B, allowing them to be executed in
587 // parallel (or back-to-back in a pipeline) instead of depending on each other.
589 // FIXME: This has the potential to be expensive (compile time) while not
590 // improving the code at all. Some ways to limit the overhead:
591 // 1. Track successful transforms; bail out if hit rate gets too low.
592 // 2. Only enable at -O3 or some other non-default optimization level.
593 // 3. Pre-screen pattern candidates here: if an operand of the previous
594 // instruction is known to not increase the critical path, then don't match
596 bool TargetInstrInfo::getMachineCombinerPatterns(
598 SmallVectorImpl<MachineCombinerPattern::MC_PATTERN> &Patterns) const {
601 if (isReassociationCandidate(Root, Commute)) {
602 // We found a sequence of instructions that may be suitable for a
603 // reassociation of operands to increase ILP. Specify each commutation
604 // possibility for the Prev instruction in the sequence and let the
605 // machine combiner decide if changing the operands is worthwhile.
607 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_YB);
608 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_YB);
610 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_BY);
611 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_BY);
619 /// Attempt the reassociation transformation to reduce critical path length.
620 /// See the above comments before getMachineCombinerPatterns().
621 void TargetInstrInfo::reassociateOps(
622 MachineInstr &Root, MachineInstr &Prev,
623 MachineCombinerPattern::MC_PATTERN Pattern,
624 SmallVectorImpl<MachineInstr *> &InsInstrs,
625 SmallVectorImpl<MachineInstr *> &DelInstrs,
626 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
627 MachineFunction *MF = Root.getParent()->getParent();
628 MachineRegisterInfo &MRI = MF->getRegInfo();
629 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
630 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
631 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
633 // This array encodes the operand index for each parameter because the
634 // operands may be commuted. Each row corresponds to a pattern value,
635 // and each column specifies the index of A, B, X, Y.
636 unsigned OpIdx[4][4] = {
643 MachineOperand &OpA = Prev.getOperand(OpIdx[Pattern][0]);
644 MachineOperand &OpB = Root.getOperand(OpIdx[Pattern][1]);
645 MachineOperand &OpX = Prev.getOperand(OpIdx[Pattern][2]);
646 MachineOperand &OpY = Root.getOperand(OpIdx[Pattern][3]);
647 MachineOperand &OpC = Root.getOperand(0);
649 unsigned RegA = OpA.getReg();
650 unsigned RegB = OpB.getReg();
651 unsigned RegX = OpX.getReg();
652 unsigned RegY = OpY.getReg();
653 unsigned RegC = OpC.getReg();
655 if (TargetRegisterInfo::isVirtualRegister(RegA))
656 MRI.constrainRegClass(RegA, RC);
657 if (TargetRegisterInfo::isVirtualRegister(RegB))
658 MRI.constrainRegClass(RegB, RC);
659 if (TargetRegisterInfo::isVirtualRegister(RegX))
660 MRI.constrainRegClass(RegX, RC);
661 if (TargetRegisterInfo::isVirtualRegister(RegY))
662 MRI.constrainRegClass(RegY, RC);
663 if (TargetRegisterInfo::isVirtualRegister(RegC))
664 MRI.constrainRegClass(RegC, RC);
666 // Create a new virtual register for the result of (X op Y) instead of
667 // recycling RegB because the MachineCombiner's computation of the critical
668 // path requires a new register definition rather than an existing one.
669 unsigned NewVR = MRI.createVirtualRegister(RC);
670 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
672 unsigned Opcode = Root.getOpcode();
673 bool KillA = OpA.isKill();
674 bool KillX = OpX.isKill();
675 bool KillY = OpY.isKill();
677 // Create new instructions for insertion.
678 MachineInstrBuilder MIB1 =
679 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
680 .addReg(RegX, getKillRegState(KillX))
681 .addReg(RegY, getKillRegState(KillY));
682 MachineInstrBuilder MIB2 =
683 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
684 .addReg(RegA, getKillRegState(KillA))
685 .addReg(NewVR, getKillRegState(true));
687 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
689 // Record new instructions for insertion and old instructions for deletion.
690 InsInstrs.push_back(MIB1);
691 InsInstrs.push_back(MIB2);
692 DelInstrs.push_back(&Prev);
693 DelInstrs.push_back(&Root);
696 void TargetInstrInfo::genAlternativeCodeSequence(
697 MachineInstr &Root, MachineCombinerPattern::MC_PATTERN Pattern,
698 SmallVectorImpl<MachineInstr *> &InsInstrs,
699 SmallVectorImpl<MachineInstr *> &DelInstrs,
700 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
701 MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo();
703 // Select the previous instruction in the sequence based on the input pattern.
704 MachineInstr *Prev = nullptr;
706 case MachineCombinerPattern::MC_REASSOC_AX_BY:
707 case MachineCombinerPattern::MC_REASSOC_XA_BY:
708 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
710 case MachineCombinerPattern::MC_REASSOC_AX_YB:
711 case MachineCombinerPattern::MC_REASSOC_XA_YB:
712 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
718 assert(Prev && "Unknown pattern for machine combiner");
720 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
724 /// foldMemoryOperand - Same as the previous version except it allows folding
725 /// of any load and store from / to any address, not just from a specific
727 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
728 ArrayRef<unsigned> Ops,
729 MachineInstr *LoadMI) const {
730 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!");
732 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
733 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!");
735 MachineBasicBlock &MBB = *MI->getParent();
736 MachineFunction &MF = *MBB.getParent();
738 // Ask the target to do the actual folding.
739 MachineInstr *NewMI = nullptr;
742 if ((MI->getOpcode() == TargetOpcode::STACKMAP ||
743 MI->getOpcode() == TargetOpcode::PATCHPOINT) &&
744 isLoadFromStackSlot(LoadMI, FrameIndex)) {
745 // Fold stackmap/patchpoint.
746 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
748 NewMI = MBB.insert(MI, NewMI);
750 // Ask the target to do the actual folding.
751 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI);
754 if (!NewMI) return nullptr;
756 // Copy the memoperands from the load to the folded instruction.
757 if (MI->memoperands_empty()) {
758 NewMI->setMemRefs(LoadMI->memoperands_begin(),
759 LoadMI->memoperands_end());
762 // Handle the rare case of folding multiple loads.
763 NewMI->setMemRefs(MI->memoperands_begin(),
764 MI->memoperands_end());
765 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(),
766 E = LoadMI->memoperands_end(); I != E; ++I) {
767 NewMI->addMemOperand(MF, *I);
773 bool TargetInstrInfo::
774 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
775 AliasAnalysis *AA) const {
776 const MachineFunction &MF = *MI->getParent()->getParent();
777 const MachineRegisterInfo &MRI = MF.getRegInfo();
779 // Remat clients assume operand 0 is the defined register.
780 if (!MI->getNumOperands() || !MI->getOperand(0).isReg())
782 unsigned DefReg = MI->getOperand(0).getReg();
784 // A sub-register definition can only be rematerialized if the instruction
785 // doesn't read the other parts of the register. Otherwise it is really a
786 // read-modify-write operation on the full virtual register which cannot be
788 if (TargetRegisterInfo::isVirtualRegister(DefReg) &&
789 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg))
792 // A load from a fixed stack slot can be rematerialized. This may be
793 // redundant with subsequent checks, but it's target-independent,
794 // simple, and a common case.
796 if (isLoadFromStackSlot(MI, FrameIdx) &&
797 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx))
800 // Avoid instructions obviously unsafe for remat.
801 if (MI->isNotDuplicable() || MI->mayStore() ||
802 MI->hasUnmodeledSideEffects())
805 // Don't remat inline asm. We have no idea how expensive it is
806 // even if it's side effect free.
807 if (MI->isInlineAsm())
810 // Avoid instructions which load from potentially varying memory.
811 if (MI->mayLoad() && !MI->isInvariantLoad(AA))
814 // If any of the registers accessed are non-constant, conservatively assume
815 // the instruction is not rematerializable.
816 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
817 const MachineOperand &MO = MI->getOperand(i);
818 if (!MO.isReg()) continue;
819 unsigned Reg = MO.getReg();
823 // Check for a well-behaved physical register.
824 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
826 // If the physreg has no defs anywhere, it's just an ambient register
827 // and we can freely move its uses. Alternatively, if it's allocatable,
828 // it could get allocated to something with a def during allocation.
829 if (!MRI.isConstantPhysReg(Reg, MF))
832 // A physreg def. We can't remat it.
838 // Only allow one virtual-register def. There may be multiple defs of the
839 // same virtual register, though.
840 if (MO.isDef() && Reg != DefReg)
843 // Don't allow any virtual-register uses. Rematting an instruction with
844 // virtual register uses would length the live ranges of the uses, which
845 // is not necessarily a good idea, certainly not "trivial".
850 // Everything checked out.
854 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const {
855 const MachineFunction *MF = MI->getParent()->getParent();
856 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
857 bool StackGrowsDown =
858 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
860 unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
861 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
863 if (MI->getOpcode() != FrameSetupOpcode &&
864 MI->getOpcode() != FrameDestroyOpcode)
867 int SPAdj = MI->getOperand(0).getImm();
868 SPAdj = TFI->alignSPAdjust(SPAdj);
870 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) ||
871 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode))
877 /// isSchedulingBoundary - Test if the given instruction should be
878 /// considered a scheduling boundary. This primarily includes labels
880 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
881 const MachineBasicBlock *MBB,
882 const MachineFunction &MF) const {
883 // Terminators and labels can't be scheduled around.
884 if (MI->isTerminator() || MI->isPosition())
887 // Don't attempt to schedule around any instruction that defines
888 // a stack-oriented pointer, as it's unlikely to be profitable. This
889 // saves compile time, because it doesn't require every single
890 // stack slot reference to depend on the instruction that does the
892 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
893 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
894 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI))
900 // Provide a global flag for disabling the PreRA hazard recognizer that targets
901 // may choose to honor.
902 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
903 return !DisableHazardRecognizer;
906 // Default implementation of CreateTargetRAHazardRecognizer.
907 ScheduleHazardRecognizer *TargetInstrInfo::
908 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
909 const ScheduleDAG *DAG) const {
910 // Dummy hazard recognizer allows all instructions to issue.
911 return new ScheduleHazardRecognizer();
914 // Default implementation of CreateTargetMIHazardRecognizer.
915 ScheduleHazardRecognizer *TargetInstrInfo::
916 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
917 const ScheduleDAG *DAG) const {
918 return (ScheduleHazardRecognizer *)
919 new ScoreboardHazardRecognizer(II, DAG, "misched");
922 // Default implementation of CreateTargetPostRAHazardRecognizer.
923 ScheduleHazardRecognizer *TargetInstrInfo::
924 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
925 const ScheduleDAG *DAG) const {
926 return (ScheduleHazardRecognizer *)
927 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
930 //===----------------------------------------------------------------------===//
931 // SelectionDAG latency interface.
932 //===----------------------------------------------------------------------===//
935 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
936 SDNode *DefNode, unsigned DefIdx,
937 SDNode *UseNode, unsigned UseIdx) const {
938 if (!ItinData || ItinData->isEmpty())
941 if (!DefNode->isMachineOpcode())
944 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
945 if (!UseNode->isMachineOpcode())
946 return ItinData->getOperandCycle(DefClass, DefIdx);
947 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
948 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
951 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
953 if (!ItinData || ItinData->isEmpty())
956 if (!N->isMachineOpcode())
959 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
962 //===----------------------------------------------------------------------===//
963 // MachineInstr latency interface.
964 //===----------------------------------------------------------------------===//
967 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
968 const MachineInstr *MI) const {
969 if (!ItinData || ItinData->isEmpty())
972 unsigned Class = MI->getDesc().getSchedClass();
973 int UOps = ItinData->Itineraries[Class].NumMicroOps;
977 // The # of u-ops is dynamically determined. The specific target should
978 // override this function to return the right number.
982 /// Return the default expected latency for a def based on it's opcode.
983 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
984 const MachineInstr *DefMI) const {
985 if (DefMI->isTransient())
987 if (DefMI->mayLoad())
988 return SchedModel.LoadLatency;
989 if (isHighLatencyDef(DefMI->getOpcode()))
990 return SchedModel.HighLatency;
994 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const {
998 unsigned TargetInstrInfo::
999 getInstrLatency(const InstrItineraryData *ItinData,
1000 const MachineInstr *MI,
1001 unsigned *PredCost) const {
1002 // Default to one cycle for no itinerary. However, an "empty" itinerary may
1003 // still have a MinLatency property, which getStageLatency checks.
1005 return MI->mayLoad() ? 2 : 1;
1007 return ItinData->getStageLatency(MI->getDesc().getSchedClass());
1010 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1011 const MachineInstr *DefMI,
1012 unsigned DefIdx) const {
1013 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1014 if (!ItinData || ItinData->isEmpty())
1017 unsigned DefClass = DefMI->getDesc().getSchedClass();
1018 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1019 return (DefCycle != -1 && DefCycle <= 1);
1022 /// Both DefMI and UseMI must be valid. By default, call directly to the
1023 /// itinerary. This may be overriden by the target.
1024 int TargetInstrInfo::
1025 getOperandLatency(const InstrItineraryData *ItinData,
1026 const MachineInstr *DefMI, unsigned DefIdx,
1027 const MachineInstr *UseMI, unsigned UseIdx) const {
1028 unsigned DefClass = DefMI->getDesc().getSchedClass();
1029 unsigned UseClass = UseMI->getDesc().getSchedClass();
1030 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1033 /// If we can determine the operand latency from the def only, without itinerary
1034 /// lookup, do so. Otherwise return -1.
1035 int TargetInstrInfo::computeDefOperandLatency(
1036 const InstrItineraryData *ItinData,
1037 const MachineInstr *DefMI) const {
1039 // Let the target hook getInstrLatency handle missing itineraries.
1041 return getInstrLatency(ItinData, DefMI);
1043 if(ItinData->isEmpty())
1044 return defaultDefLatency(ItinData->SchedModel, DefMI);
1046 // ...operand lookup required
1050 /// computeOperandLatency - Compute and return the latency of the given data
1051 /// dependent def and use when the operand indices are already known. UseMI may
1052 /// be NULL for an unknown use.
1054 /// FindMin may be set to get the minimum vs. expected latency. Minimum
1055 /// latency is used for scheduling groups, while expected latency is for
1056 /// instruction cost and critical path.
1058 /// Depending on the subtarget's itinerary properties, this may or may not need
1059 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or
1060 /// UseIdx to compute min latency.
1061 unsigned TargetInstrInfo::
1062 computeOperandLatency(const InstrItineraryData *ItinData,
1063 const MachineInstr *DefMI, unsigned DefIdx,
1064 const MachineInstr *UseMI, unsigned UseIdx) const {
1066 int DefLatency = computeDefOperandLatency(ItinData, DefMI);
1067 if (DefLatency >= 0)
1070 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail");
1072 int OperLatency = 0;
1074 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
1076 unsigned DefClass = DefMI->getDesc().getSchedClass();
1077 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx);
1079 if (OperLatency >= 0)
1082 // No operand latency was found.
1083 unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
1085 // Expected latency is the max of the stage latency and itinerary props.
1086 InstrLatency = std::max(InstrLatency,
1087 defaultDefLatency(ItinData->SchedModel, DefMI));
1088 return InstrLatency;
1091 bool TargetInstrInfo::getRegSequenceInputs(
1092 const MachineInstr &MI, unsigned DefIdx,
1093 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1094 assert((MI.isRegSequence() ||
1095 MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1097 if (!MI.isRegSequence())
1098 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1100 // We are looking at:
1101 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1102 assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1103 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1105 const MachineOperand &MOReg = MI.getOperand(OpIdx);
1106 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1107 assert(MOSubIdx.isImm() &&
1108 "One of the subindex of the reg_sequence is not an immediate");
1109 // Record Reg:SubReg, SubIdx.
1110 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1111 (unsigned)MOSubIdx.getImm()));
1116 bool TargetInstrInfo::getExtractSubregInputs(
1117 const MachineInstr &MI, unsigned DefIdx,
1118 RegSubRegPairAndIdx &InputReg) const {
1119 assert((MI.isExtractSubreg() ||
1120 MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1122 if (!MI.isExtractSubreg())
1123 return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1125 // We are looking at:
1126 // Def = EXTRACT_SUBREG v0.sub1, sub0.
1127 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1128 const MachineOperand &MOReg = MI.getOperand(1);
1129 const MachineOperand &MOSubIdx = MI.getOperand(2);
1130 assert(MOSubIdx.isImm() &&
1131 "The subindex of the extract_subreg is not an immediate");
1133 InputReg.Reg = MOReg.getReg();
1134 InputReg.SubReg = MOReg.getSubReg();
1135 InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1139 bool TargetInstrInfo::getInsertSubregInputs(
1140 const MachineInstr &MI, unsigned DefIdx,
1141 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1142 assert((MI.isInsertSubreg() ||
1143 MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1145 if (!MI.isInsertSubreg())
1146 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1148 // We are looking at:
1149 // Def = INSERT_SEQUENCE v0, v1, sub0.
1150 assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1151 const MachineOperand &MOBaseReg = MI.getOperand(1);
1152 const MachineOperand &MOInsertedReg = MI.getOperand(2);
1153 const MachineOperand &MOSubIdx = MI.getOperand(3);
1154 assert(MOSubIdx.isImm() &&
1155 "One of the subindex of the reg_sequence is not an immediate");
1156 BaseReg.Reg = MOBaseReg.getReg();
1157 BaseReg.SubReg = MOBaseReg.getSubReg();
1159 InsertedReg.Reg = MOInsertedReg.getReg();
1160 InsertedReg.SubReg = MOInsertedReg.getSubReg();
1161 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();