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 MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr *MI,
124 unsigned Idx2) const {
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 unsigned CommutableOpIdx1 = Idx1, CommutableOpIdx2 = Idx2;
132 assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) &&
133 CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 &&
134 "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands.");
135 assert(MI->getOperand(Idx1).isReg() && MI->getOperand(Idx2).isReg() &&
136 "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 MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr *MI,
190 unsigned OpIdx2) const {
191 // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose
192 // any commutable operand, which is done in findCommutedOpIndices() method
194 if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) &&
195 !findCommutedOpIndices(MI, OpIdx1, OpIdx2)) {
196 assert(MI->isCommutable() &&
197 "Precondition violation: MI must be commutable.");
200 return commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
203 bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1,
204 unsigned &ResultIdx2,
205 unsigned CommutableOpIdx1,
206 unsigned CommutableOpIdx2) {
207 if (ResultIdx1 == CommuteAnyOperandIndex &&
208 ResultIdx2 == CommuteAnyOperandIndex) {
209 ResultIdx1 = CommutableOpIdx1;
210 ResultIdx2 = CommutableOpIdx2;
211 } else if (ResultIdx1 == CommuteAnyOperandIndex) {
212 if (ResultIdx2 == CommutableOpIdx1)
213 ResultIdx1 = CommutableOpIdx2;
214 else if (ResultIdx2 == CommutableOpIdx2)
215 ResultIdx1 = CommutableOpIdx1;
218 } else if (ResultIdx2 == CommuteAnyOperandIndex) {
219 if (ResultIdx1 == CommutableOpIdx1)
220 ResultIdx2 = CommutableOpIdx2;
221 else if (ResultIdx1 == CommutableOpIdx2)
222 ResultIdx2 = CommutableOpIdx1;
226 // Check that the result operand indices match the given commutable
228 return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) ||
229 (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1);
234 bool TargetInstrInfo::findCommutedOpIndices(MachineInstr *MI,
236 unsigned &SrcOpIdx2) const {
237 assert(!MI->isBundle() &&
238 "TargetInstrInfo::findCommutedOpIndices() can't handle bundles");
240 const MCInstrDesc &MCID = MI->getDesc();
241 if (!MCID.isCommutable())
244 // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this
245 // is not true, then the target must implement this.
246 unsigned CommutableOpIdx1 = MCID.getNumDefs();
247 unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1;
248 if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2,
249 CommutableOpIdx1, CommutableOpIdx2))
252 if (!MI->getOperand(SrcOpIdx1).isReg() ||
253 !MI->getOperand(SrcOpIdx2).isReg())
260 TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const {
261 if (!MI->isTerminator()) return false;
263 // Conditional branch is a special case.
264 if (MI->isBranch() && !MI->isBarrier())
266 if (!MI->isPredicable())
268 return !isPredicated(MI);
271 bool TargetInstrInfo::PredicateInstruction(
272 MachineInstr *MI, ArrayRef<MachineOperand> Pred) const {
273 bool MadeChange = false;
275 assert(!MI->isBundle() &&
276 "TargetInstrInfo::PredicateInstruction() can't handle bundles");
278 const MCInstrDesc &MCID = MI->getDesc();
279 if (!MI->isPredicable())
282 for (unsigned j = 0, i = 0, e = MI->getNumOperands(); i != e; ++i) {
283 if (MCID.OpInfo[i].isPredicate()) {
284 MachineOperand &MO = MI->getOperand(i);
286 MO.setReg(Pred[j].getReg());
288 } else if (MO.isImm()) {
289 MO.setImm(Pred[j].getImm());
291 } else if (MO.isMBB()) {
292 MO.setMBB(Pred[j].getMBB());
301 bool TargetInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
302 const MachineMemOperand *&MMO,
303 int &FrameIndex) const {
304 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
305 oe = MI->memoperands_end();
308 if ((*o)->isLoad()) {
309 if (const FixedStackPseudoSourceValue *Value =
310 dyn_cast_or_null<FixedStackPseudoSourceValue>(
311 (*o)->getPseudoValue())) {
312 FrameIndex = Value->getFrameIndex();
321 bool TargetInstrInfo::hasStoreToStackSlot(const MachineInstr *MI,
322 const MachineMemOperand *&MMO,
323 int &FrameIndex) const {
324 for (MachineInstr::mmo_iterator o = MI->memoperands_begin(),
325 oe = MI->memoperands_end();
328 if ((*o)->isStore()) {
329 if (const FixedStackPseudoSourceValue *Value =
330 dyn_cast_or_null<FixedStackPseudoSourceValue>(
331 (*o)->getPseudoValue())) {
332 FrameIndex = Value->getFrameIndex();
341 bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC,
342 unsigned SubIdx, unsigned &Size,
344 const MachineFunction &MF) const {
346 Size = RC->getSize();
350 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
351 unsigned BitSize = TRI->getSubRegIdxSize(SubIdx);
352 // Convert bit size to byte size to be consistent with
353 // MCRegisterClass::getSize().
357 int BitOffset = TRI->getSubRegIdxOffset(SubIdx);
358 if (BitOffset < 0 || BitOffset % 8)
362 Offset = (unsigned)BitOffset / 8;
364 assert(RC->getSize() >= (Offset + Size) && "bad subregister range");
366 if (!MF.getDataLayout().isLittleEndian()) {
367 Offset = RC->getSize() - (Offset + Size);
372 void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB,
373 MachineBasicBlock::iterator I,
376 const MachineInstr *Orig,
377 const TargetRegisterInfo &TRI) const {
378 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig);
379 MI->substituteRegister(MI->getOperand(0).getReg(), DestReg, SubIdx, TRI);
384 TargetInstrInfo::produceSameValue(const MachineInstr *MI0,
385 const MachineInstr *MI1,
386 const MachineRegisterInfo *MRI) const {
387 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
390 MachineInstr *TargetInstrInfo::duplicate(MachineInstr *Orig,
391 MachineFunction &MF) const {
392 assert(!Orig->isNotDuplicable() &&
393 "Instruction cannot be duplicated");
394 return MF.CloneMachineInstr(Orig);
397 // If the COPY instruction in MI can be folded to a stack operation, return
398 // the register class to use.
399 static const TargetRegisterClass *canFoldCopy(const MachineInstr *MI,
401 assert(MI->isCopy() && "MI must be a COPY instruction");
402 if (MI->getNumOperands() != 2)
404 assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
406 const MachineOperand &FoldOp = MI->getOperand(FoldIdx);
407 const MachineOperand &LiveOp = MI->getOperand(1-FoldIdx);
409 if (FoldOp.getSubReg() || LiveOp.getSubReg())
412 unsigned FoldReg = FoldOp.getReg();
413 unsigned LiveReg = LiveOp.getReg();
415 assert(TargetRegisterInfo::isVirtualRegister(FoldReg) &&
416 "Cannot fold physregs");
418 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
419 const TargetRegisterClass *RC = MRI.getRegClass(FoldReg);
421 if (TargetRegisterInfo::isPhysicalRegister(LiveOp.getReg()))
422 return RC->contains(LiveOp.getReg()) ? RC : nullptr;
424 if (RC->hasSubClassEq(MRI.getRegClass(LiveReg)))
427 // FIXME: Allow folding when register classes are memory compatible.
431 void TargetInstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
432 llvm_unreachable("Not a MachO target");
435 static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr *MI,
436 ArrayRef<unsigned> Ops, int FrameIndex,
437 const TargetInstrInfo &TII) {
438 unsigned StartIdx = 0;
439 switch (MI->getOpcode()) {
440 case TargetOpcode::STACKMAP:
441 StartIdx = 2; // Skip ID, nShadowBytes.
443 case TargetOpcode::PATCHPOINT: {
444 // For PatchPoint, the call args are not foldable.
445 PatchPointOpers opers(MI);
446 StartIdx = opers.getVarIdx();
450 llvm_unreachable("unexpected stackmap opcode");
453 // Return false if any operands requested for folding are not foldable (not
454 // part of the stackmap's live values).
455 for (unsigned Op : Ops) {
460 MachineInstr *NewMI =
461 MF.CreateMachineInstr(TII.get(MI->getOpcode()), MI->getDebugLoc(), true);
462 MachineInstrBuilder MIB(MF, NewMI);
464 // No need to fold return, the meta data, and function arguments
465 for (unsigned i = 0; i < StartIdx; ++i)
466 MIB.addOperand(MI->getOperand(i));
468 for (unsigned i = StartIdx; i < MI->getNumOperands(); ++i) {
469 MachineOperand &MO = MI->getOperand(i);
470 if (std::find(Ops.begin(), Ops.end(), i) != Ops.end()) {
472 unsigned SpillOffset;
473 // Compute the spill slot size and offset.
474 const TargetRegisterClass *RC =
475 MF.getRegInfo().getRegClass(MO.getReg());
477 TII.getStackSlotRange(RC, MO.getSubReg(), SpillSize, SpillOffset, MF);
479 report_fatal_error("cannot spill patchpoint subregister operand");
480 MIB.addImm(StackMaps::IndirectMemRefOp);
481 MIB.addImm(SpillSize);
482 MIB.addFrameIndex(FrameIndex);
483 MIB.addImm(SpillOffset);
491 /// foldMemoryOperand - Attempt to fold a load or store of the specified stack
492 /// slot into the specified machine instruction for the specified operand(s).
493 /// If this is possible, a new instruction is returned with the specified
494 /// operand folded, otherwise NULL is returned. The client is responsible for
495 /// removing the old instruction and adding the new one in the instruction
497 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
498 ArrayRef<unsigned> Ops,
501 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
502 if (MI->getOperand(Ops[i]).isDef())
503 Flags |= MachineMemOperand::MOStore;
505 Flags |= MachineMemOperand::MOLoad;
507 MachineBasicBlock *MBB = MI->getParent();
508 assert(MBB && "foldMemoryOperand needs an inserted instruction");
509 MachineFunction &MF = *MBB->getParent();
511 MachineInstr *NewMI = nullptr;
513 if (MI->getOpcode() == TargetOpcode::STACKMAP ||
514 MI->getOpcode() == TargetOpcode::PATCHPOINT) {
515 // Fold stackmap/patchpoint.
516 NewMI = foldPatchpoint(MF, MI, Ops, FI, *this);
518 MBB->insert(MI, NewMI);
520 // Ask the target to do the actual folding.
521 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, FI);
525 NewMI->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
526 // Add a memory operand, foldMemoryOperandImpl doesn't do that.
527 assert((!(Flags & MachineMemOperand::MOStore) ||
528 NewMI->mayStore()) &&
529 "Folded a def to a non-store!");
530 assert((!(Flags & MachineMemOperand::MOLoad) ||
532 "Folded a use to a non-load!");
533 const MachineFrameInfo &MFI = *MF.getFrameInfo();
534 assert(MFI.getObjectOffset(FI) != -1);
535 MachineMemOperand *MMO = MF.getMachineMemOperand(
536 MachinePointerInfo::getFixedStack(MF, FI), Flags, MFI.getObjectSize(FI),
537 MFI.getObjectAlignment(FI));
538 NewMI->addMemOperand(MF, MMO);
543 // Straight COPY may fold as load/store.
544 if (!MI->isCopy() || Ops.size() != 1)
547 const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
551 const MachineOperand &MO = MI->getOperand(1-Ops[0]);
552 MachineBasicBlock::iterator Pos = MI;
553 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
555 if (Flags == MachineMemOperand::MOStore)
556 storeRegToStackSlot(*MBB, Pos, MO.getReg(), MO.isKill(), FI, RC, TRI);
558 loadRegFromStackSlot(*MBB, Pos, MO.getReg(), FI, RC, TRI);
562 bool TargetInstrInfo::hasReassociableOperands(
563 const MachineInstr &Inst, const MachineBasicBlock *MBB) const {
564 const MachineOperand &Op1 = Inst.getOperand(1);
565 const MachineOperand &Op2 = Inst.getOperand(2);
566 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
568 // We need virtual register definitions for the operands that we will
570 MachineInstr *MI1 = nullptr;
571 MachineInstr *MI2 = nullptr;
572 if (Op1.isReg() && TargetRegisterInfo::isVirtualRegister(Op1.getReg()))
573 MI1 = MRI.getUniqueVRegDef(Op1.getReg());
574 if (Op2.isReg() && TargetRegisterInfo::isVirtualRegister(Op2.getReg()))
575 MI2 = MRI.getUniqueVRegDef(Op2.getReg());
577 // And they need to be in the trace (otherwise, they won't have a depth).
578 if (MI1 && MI2 && MI1->getParent() == MBB && MI2->getParent() == MBB)
584 bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst,
585 bool &Commuted) const {
586 const MachineBasicBlock *MBB = Inst.getParent();
587 const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
588 MachineInstr *MI1 = MRI.getUniqueVRegDef(Inst.getOperand(1).getReg());
589 MachineInstr *MI2 = MRI.getUniqueVRegDef(Inst.getOperand(2).getReg());
590 unsigned AssocOpcode = Inst.getOpcode();
592 // If only one operand has the same opcode and it's the second source operand,
593 // the operands must be commuted.
594 Commuted = MI1->getOpcode() != AssocOpcode && MI2->getOpcode() == AssocOpcode;
598 // 1. The previous instruction must be the same type as Inst.
599 // 2. The previous instruction must have virtual register definitions for its
600 // operands in the same basic block as Inst.
601 // 3. The previous instruction's result must only be used by Inst.
602 if (MI1->getOpcode() == AssocOpcode && hasReassociableOperands(*MI1, MBB) &&
603 MRI.hasOneNonDBGUse(MI1->getOperand(0).getReg()))
609 // 1. The operation must be associative and commutative.
610 // 2. The instruction must have virtual register definitions for its
611 // operands in the same basic block.
612 // 3. The instruction must have a reassociable sibling.
613 bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst,
614 bool &Commuted) const {
615 if (isAssociativeAndCommutative(Inst) &&
616 hasReassociableOperands(Inst, Inst.getParent()) &&
617 hasReassociableSibling(Inst, Commuted))
623 // The concept of the reassociation pass is that these operations can benefit
624 // from this kind of transformation:
634 // breaking the dependency between A and B, allowing them to be executed in
635 // parallel (or back-to-back in a pipeline) instead of depending on each other.
637 // FIXME: This has the potential to be expensive (compile time) while not
638 // improving the code at all. Some ways to limit the overhead:
639 // 1. Track successful transforms; bail out if hit rate gets too low.
640 // 2. Only enable at -O3 or some other non-default optimization level.
641 // 3. Pre-screen pattern candidates here: if an operand of the previous
642 // instruction is known to not increase the critical path, then don't match
644 bool TargetInstrInfo::getMachineCombinerPatterns(
646 SmallVectorImpl<MachineCombinerPattern::MC_PATTERN> &Patterns) const {
649 if (isReassociationCandidate(Root, Commute)) {
650 // We found a sequence of instructions that may be suitable for a
651 // reassociation of operands to increase ILP. Specify each commutation
652 // possibility for the Prev instruction in the sequence and let the
653 // machine combiner decide if changing the operands is worthwhile.
655 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_YB);
656 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_YB);
658 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_AX_BY);
659 Patterns.push_back(MachineCombinerPattern::MC_REASSOC_XA_BY);
667 /// Attempt the reassociation transformation to reduce critical path length.
668 /// See the above comments before getMachineCombinerPatterns().
669 void TargetInstrInfo::reassociateOps(
670 MachineInstr &Root, MachineInstr &Prev,
671 MachineCombinerPattern::MC_PATTERN Pattern,
672 SmallVectorImpl<MachineInstr *> &InsInstrs,
673 SmallVectorImpl<MachineInstr *> &DelInstrs,
674 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const {
675 MachineFunction *MF = Root.getParent()->getParent();
676 MachineRegisterInfo &MRI = MF->getRegInfo();
677 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
678 const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo();
679 const TargetRegisterClass *RC = Root.getRegClassConstraint(0, TII, TRI);
681 // This array encodes the operand index for each parameter because the
682 // operands may be commuted. Each row corresponds to a pattern value,
683 // and each column specifies the index of A, B, X, Y.
684 unsigned OpIdx[4][4] = {
691 MachineOperand &OpA = Prev.getOperand(OpIdx[Pattern][0]);
692 MachineOperand &OpB = Root.getOperand(OpIdx[Pattern][1]);
693 MachineOperand &OpX = Prev.getOperand(OpIdx[Pattern][2]);
694 MachineOperand &OpY = Root.getOperand(OpIdx[Pattern][3]);
695 MachineOperand &OpC = Root.getOperand(0);
697 unsigned RegA = OpA.getReg();
698 unsigned RegB = OpB.getReg();
699 unsigned RegX = OpX.getReg();
700 unsigned RegY = OpY.getReg();
701 unsigned RegC = OpC.getReg();
703 if (TargetRegisterInfo::isVirtualRegister(RegA))
704 MRI.constrainRegClass(RegA, RC);
705 if (TargetRegisterInfo::isVirtualRegister(RegB))
706 MRI.constrainRegClass(RegB, RC);
707 if (TargetRegisterInfo::isVirtualRegister(RegX))
708 MRI.constrainRegClass(RegX, RC);
709 if (TargetRegisterInfo::isVirtualRegister(RegY))
710 MRI.constrainRegClass(RegY, RC);
711 if (TargetRegisterInfo::isVirtualRegister(RegC))
712 MRI.constrainRegClass(RegC, RC);
714 // Create a new virtual register for the result of (X op Y) instead of
715 // recycling RegB because the MachineCombiner's computation of the critical
716 // path requires a new register definition rather than an existing one.
717 unsigned NewVR = MRI.createVirtualRegister(RC);
718 InstrIdxForVirtReg.insert(std::make_pair(NewVR, 0));
720 unsigned Opcode = Root.getOpcode();
721 bool KillA = OpA.isKill();
722 bool KillX = OpX.isKill();
723 bool KillY = OpY.isKill();
725 // Create new instructions for insertion.
726 MachineInstrBuilder MIB1 =
727 BuildMI(*MF, Prev.getDebugLoc(), TII->get(Opcode), NewVR)
728 .addReg(RegX, getKillRegState(KillX))
729 .addReg(RegY, getKillRegState(KillY));
730 MachineInstrBuilder MIB2 =
731 BuildMI(*MF, Root.getDebugLoc(), TII->get(Opcode), RegC)
732 .addReg(RegA, getKillRegState(KillA))
733 .addReg(NewVR, getKillRegState(true));
735 setSpecialOperandAttr(Root, Prev, *MIB1, *MIB2);
737 // Record new instructions for insertion and old instructions for deletion.
738 InsInstrs.push_back(MIB1);
739 InsInstrs.push_back(MIB2);
740 DelInstrs.push_back(&Prev);
741 DelInstrs.push_back(&Root);
744 void TargetInstrInfo::genAlternativeCodeSequence(
745 MachineInstr &Root, MachineCombinerPattern::MC_PATTERN Pattern,
746 SmallVectorImpl<MachineInstr *> &InsInstrs,
747 SmallVectorImpl<MachineInstr *> &DelInstrs,
748 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const {
749 MachineRegisterInfo &MRI = Root.getParent()->getParent()->getRegInfo();
751 // Select the previous instruction in the sequence based on the input pattern.
752 MachineInstr *Prev = nullptr;
754 case MachineCombinerPattern::MC_REASSOC_AX_BY:
755 case MachineCombinerPattern::MC_REASSOC_XA_BY:
756 Prev = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
758 case MachineCombinerPattern::MC_REASSOC_AX_YB:
759 case MachineCombinerPattern::MC_REASSOC_XA_YB:
760 Prev = MRI.getUniqueVRegDef(Root.getOperand(2).getReg());
766 assert(Prev && "Unknown pattern for machine combiner");
768 reassociateOps(Root, *Prev, Pattern, InsInstrs, DelInstrs, InstIdxForVirtReg);
772 /// foldMemoryOperand - Same as the previous version except it allows folding
773 /// of any load and store from / to any address, not just from a specific
775 MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineBasicBlock::iterator MI,
776 ArrayRef<unsigned> Ops,
777 MachineInstr *LoadMI) const {
778 assert(LoadMI->canFoldAsLoad() && "LoadMI isn't foldable!");
780 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
781 assert(MI->getOperand(Ops[i]).isUse() && "Folding load into def!");
783 MachineBasicBlock &MBB = *MI->getParent();
784 MachineFunction &MF = *MBB.getParent();
786 // Ask the target to do the actual folding.
787 MachineInstr *NewMI = nullptr;
790 if ((MI->getOpcode() == TargetOpcode::STACKMAP ||
791 MI->getOpcode() == TargetOpcode::PATCHPOINT) &&
792 isLoadFromStackSlot(LoadMI, FrameIndex)) {
793 // Fold stackmap/patchpoint.
794 NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, *this);
796 NewMI = MBB.insert(MI, NewMI);
798 // Ask the target to do the actual folding.
799 NewMI = foldMemoryOperandImpl(MF, MI, Ops, MI, LoadMI);
802 if (!NewMI) return nullptr;
804 // Copy the memoperands from the load to the folded instruction.
805 if (MI->memoperands_empty()) {
806 NewMI->setMemRefs(LoadMI->memoperands_begin(),
807 LoadMI->memoperands_end());
810 // Handle the rare case of folding multiple loads.
811 NewMI->setMemRefs(MI->memoperands_begin(),
812 MI->memoperands_end());
813 for (MachineInstr::mmo_iterator I = LoadMI->memoperands_begin(),
814 E = LoadMI->memoperands_end(); I != E; ++I) {
815 NewMI->addMemOperand(MF, *I);
821 bool TargetInstrInfo::
822 isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
823 AliasAnalysis *AA) const {
824 const MachineFunction &MF = *MI->getParent()->getParent();
825 const MachineRegisterInfo &MRI = MF.getRegInfo();
827 // Remat clients assume operand 0 is the defined register.
828 if (!MI->getNumOperands() || !MI->getOperand(0).isReg())
830 unsigned DefReg = MI->getOperand(0).getReg();
832 // A sub-register definition can only be rematerialized if the instruction
833 // doesn't read the other parts of the register. Otherwise it is really a
834 // read-modify-write operation on the full virtual register which cannot be
836 if (TargetRegisterInfo::isVirtualRegister(DefReg) &&
837 MI->getOperand(0).getSubReg() && MI->readsVirtualRegister(DefReg))
840 // A load from a fixed stack slot can be rematerialized. This may be
841 // redundant with subsequent checks, but it's target-independent,
842 // simple, and a common case.
844 if (isLoadFromStackSlot(MI, FrameIdx) &&
845 MF.getFrameInfo()->isImmutableObjectIndex(FrameIdx))
848 // Avoid instructions obviously unsafe for remat.
849 if (MI->isNotDuplicable() || MI->mayStore() ||
850 MI->hasUnmodeledSideEffects())
853 // Don't remat inline asm. We have no idea how expensive it is
854 // even if it's side effect free.
855 if (MI->isInlineAsm())
858 // Avoid instructions which load from potentially varying memory.
859 if (MI->mayLoad() && !MI->isInvariantLoad(AA))
862 // If any of the registers accessed are non-constant, conservatively assume
863 // the instruction is not rematerializable.
864 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
865 const MachineOperand &MO = MI->getOperand(i);
866 if (!MO.isReg()) continue;
867 unsigned Reg = MO.getReg();
871 // Check for a well-behaved physical register.
872 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
874 // If the physreg has no defs anywhere, it's just an ambient register
875 // and we can freely move its uses. Alternatively, if it's allocatable,
876 // it could get allocated to something with a def during allocation.
877 if (!MRI.isConstantPhysReg(Reg, MF))
880 // A physreg def. We can't remat it.
886 // Only allow one virtual-register def. There may be multiple defs of the
887 // same virtual register, though.
888 if (MO.isDef() && Reg != DefReg)
891 // Don't allow any virtual-register uses. Rematting an instruction with
892 // virtual register uses would length the live ranges of the uses, which
893 // is not necessarily a good idea, certainly not "trivial".
898 // Everything checked out.
902 int TargetInstrInfo::getSPAdjust(const MachineInstr *MI) const {
903 const MachineFunction *MF = MI->getParent()->getParent();
904 const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
905 bool StackGrowsDown =
906 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
908 unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
909 unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
911 if (MI->getOpcode() != FrameSetupOpcode &&
912 MI->getOpcode() != FrameDestroyOpcode)
915 int SPAdj = MI->getOperand(0).getImm();
916 SPAdj = TFI->alignSPAdjust(SPAdj);
918 if ((!StackGrowsDown && MI->getOpcode() == FrameSetupOpcode) ||
919 (StackGrowsDown && MI->getOpcode() == FrameDestroyOpcode))
925 /// isSchedulingBoundary - Test if the given instruction should be
926 /// considered a scheduling boundary. This primarily includes labels
928 bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
929 const MachineBasicBlock *MBB,
930 const MachineFunction &MF) const {
931 // Terminators and labels can't be scheduled around.
932 if (MI->isTerminator() || MI->isPosition())
935 // Don't attempt to schedule around any instruction that defines
936 // a stack-oriented pointer, as it's unlikely to be profitable. This
937 // saves compile time, because it doesn't require every single
938 // stack slot reference to depend on the instruction that does the
940 const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering();
941 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
942 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore(), TRI))
948 // Provide a global flag for disabling the PreRA hazard recognizer that targets
949 // may choose to honor.
950 bool TargetInstrInfo::usePreRAHazardRecognizer() const {
951 return !DisableHazardRecognizer;
954 // Default implementation of CreateTargetRAHazardRecognizer.
955 ScheduleHazardRecognizer *TargetInstrInfo::
956 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
957 const ScheduleDAG *DAG) const {
958 // Dummy hazard recognizer allows all instructions to issue.
959 return new ScheduleHazardRecognizer();
962 // Default implementation of CreateTargetMIHazardRecognizer.
963 ScheduleHazardRecognizer *TargetInstrInfo::
964 CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
965 const ScheduleDAG *DAG) const {
966 return (ScheduleHazardRecognizer *)
967 new ScoreboardHazardRecognizer(II, DAG, "misched");
970 // Default implementation of CreateTargetPostRAHazardRecognizer.
971 ScheduleHazardRecognizer *TargetInstrInfo::
972 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
973 const ScheduleDAG *DAG) const {
974 return (ScheduleHazardRecognizer *)
975 new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched");
978 //===----------------------------------------------------------------------===//
979 // SelectionDAG latency interface.
980 //===----------------------------------------------------------------------===//
983 TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
984 SDNode *DefNode, unsigned DefIdx,
985 SDNode *UseNode, unsigned UseIdx) const {
986 if (!ItinData || ItinData->isEmpty())
989 if (!DefNode->isMachineOpcode())
992 unsigned DefClass = get(DefNode->getMachineOpcode()).getSchedClass();
993 if (!UseNode->isMachineOpcode())
994 return ItinData->getOperandCycle(DefClass, DefIdx);
995 unsigned UseClass = get(UseNode->getMachineOpcode()).getSchedClass();
996 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
999 int TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1001 if (!ItinData || ItinData->isEmpty())
1004 if (!N->isMachineOpcode())
1007 return ItinData->getStageLatency(get(N->getMachineOpcode()).getSchedClass());
1010 //===----------------------------------------------------------------------===//
1011 // MachineInstr latency interface.
1012 //===----------------------------------------------------------------------===//
1015 TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
1016 const MachineInstr *MI) const {
1017 if (!ItinData || ItinData->isEmpty())
1020 unsigned Class = MI->getDesc().getSchedClass();
1021 int UOps = ItinData->Itineraries[Class].NumMicroOps;
1025 // The # of u-ops is dynamically determined. The specific target should
1026 // override this function to return the right number.
1030 /// Return the default expected latency for a def based on it's opcode.
1031 unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel,
1032 const MachineInstr *DefMI) const {
1033 if (DefMI->isTransient())
1035 if (DefMI->mayLoad())
1036 return SchedModel.LoadLatency;
1037 if (isHighLatencyDef(DefMI->getOpcode()))
1038 return SchedModel.HighLatency;
1042 unsigned TargetInstrInfo::getPredicationCost(const MachineInstr *) const {
1046 unsigned TargetInstrInfo::
1047 getInstrLatency(const InstrItineraryData *ItinData,
1048 const MachineInstr *MI,
1049 unsigned *PredCost) const {
1050 // Default to one cycle for no itinerary. However, an "empty" itinerary may
1051 // still have a MinLatency property, which getStageLatency checks.
1053 return MI->mayLoad() ? 2 : 1;
1055 return ItinData->getStageLatency(MI->getDesc().getSchedClass());
1058 bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel,
1059 const MachineInstr *DefMI,
1060 unsigned DefIdx) const {
1061 const InstrItineraryData *ItinData = SchedModel.getInstrItineraries();
1062 if (!ItinData || ItinData->isEmpty())
1065 unsigned DefClass = DefMI->getDesc().getSchedClass();
1066 int DefCycle = ItinData->getOperandCycle(DefClass, DefIdx);
1067 return (DefCycle != -1 && DefCycle <= 1);
1070 /// Both DefMI and UseMI must be valid. By default, call directly to the
1071 /// itinerary. This may be overriden by the target.
1072 int TargetInstrInfo::
1073 getOperandLatency(const InstrItineraryData *ItinData,
1074 const MachineInstr *DefMI, unsigned DefIdx,
1075 const MachineInstr *UseMI, unsigned UseIdx) const {
1076 unsigned DefClass = DefMI->getDesc().getSchedClass();
1077 unsigned UseClass = UseMI->getDesc().getSchedClass();
1078 return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx);
1081 /// If we can determine the operand latency from the def only, without itinerary
1082 /// lookup, do so. Otherwise return -1.
1083 int TargetInstrInfo::computeDefOperandLatency(
1084 const InstrItineraryData *ItinData,
1085 const MachineInstr *DefMI) const {
1087 // Let the target hook getInstrLatency handle missing itineraries.
1089 return getInstrLatency(ItinData, DefMI);
1091 if(ItinData->isEmpty())
1092 return defaultDefLatency(ItinData->SchedModel, DefMI);
1094 // ...operand lookup required
1098 /// computeOperandLatency - Compute and return the latency of the given data
1099 /// dependent def and use when the operand indices are already known. UseMI may
1100 /// be NULL for an unknown use.
1102 /// FindMin may be set to get the minimum vs. expected latency. Minimum
1103 /// latency is used for scheduling groups, while expected latency is for
1104 /// instruction cost and critical path.
1106 /// Depending on the subtarget's itinerary properties, this may or may not need
1107 /// to call getOperandLatency(). For most subtargets, we don't need DefIdx or
1108 /// UseIdx to compute min latency.
1109 unsigned TargetInstrInfo::
1110 computeOperandLatency(const InstrItineraryData *ItinData,
1111 const MachineInstr *DefMI, unsigned DefIdx,
1112 const MachineInstr *UseMI, unsigned UseIdx) const {
1114 int DefLatency = computeDefOperandLatency(ItinData, DefMI);
1115 if (DefLatency >= 0)
1118 assert(ItinData && !ItinData->isEmpty() && "computeDefOperandLatency fail");
1120 int OperLatency = 0;
1122 OperLatency = getOperandLatency(ItinData, DefMI, DefIdx, UseMI, UseIdx);
1124 unsigned DefClass = DefMI->getDesc().getSchedClass();
1125 OperLatency = ItinData->getOperandCycle(DefClass, DefIdx);
1127 if (OperLatency >= 0)
1130 // No operand latency was found.
1131 unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
1133 // Expected latency is the max of the stage latency and itinerary props.
1134 InstrLatency = std::max(InstrLatency,
1135 defaultDefLatency(ItinData->SchedModel, DefMI));
1136 return InstrLatency;
1139 bool TargetInstrInfo::getRegSequenceInputs(
1140 const MachineInstr &MI, unsigned DefIdx,
1141 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const {
1142 assert((MI.isRegSequence() ||
1143 MI.isRegSequenceLike()) && "Instruction do not have the proper type");
1145 if (!MI.isRegSequence())
1146 return getRegSequenceLikeInputs(MI, DefIdx, InputRegs);
1148 // We are looking at:
1149 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1150 assert(DefIdx == 0 && "REG_SEQUENCE only has one def");
1151 for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx;
1153 const MachineOperand &MOReg = MI.getOperand(OpIdx);
1154 const MachineOperand &MOSubIdx = MI.getOperand(OpIdx + 1);
1155 assert(MOSubIdx.isImm() &&
1156 "One of the subindex of the reg_sequence is not an immediate");
1157 // Record Reg:SubReg, SubIdx.
1158 InputRegs.push_back(RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(),
1159 (unsigned)MOSubIdx.getImm()));
1164 bool TargetInstrInfo::getExtractSubregInputs(
1165 const MachineInstr &MI, unsigned DefIdx,
1166 RegSubRegPairAndIdx &InputReg) const {
1167 assert((MI.isExtractSubreg() ||
1168 MI.isExtractSubregLike()) && "Instruction do not have the proper type");
1170 if (!MI.isExtractSubreg())
1171 return getExtractSubregLikeInputs(MI, DefIdx, InputReg);
1173 // We are looking at:
1174 // Def = EXTRACT_SUBREG v0.sub1, sub0.
1175 assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def");
1176 const MachineOperand &MOReg = MI.getOperand(1);
1177 const MachineOperand &MOSubIdx = MI.getOperand(2);
1178 assert(MOSubIdx.isImm() &&
1179 "The subindex of the extract_subreg is not an immediate");
1181 InputReg.Reg = MOReg.getReg();
1182 InputReg.SubReg = MOReg.getSubReg();
1183 InputReg.SubIdx = (unsigned)MOSubIdx.getImm();
1187 bool TargetInstrInfo::getInsertSubregInputs(
1188 const MachineInstr &MI, unsigned DefIdx,
1189 RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const {
1190 assert((MI.isInsertSubreg() ||
1191 MI.isInsertSubregLike()) && "Instruction do not have the proper type");
1193 if (!MI.isInsertSubreg())
1194 return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg);
1196 // We are looking at:
1197 // Def = INSERT_SEQUENCE v0, v1, sub0.
1198 assert(DefIdx == 0 && "INSERT_SUBREG only has one def");
1199 const MachineOperand &MOBaseReg = MI.getOperand(1);
1200 const MachineOperand &MOInsertedReg = MI.getOperand(2);
1201 const MachineOperand &MOSubIdx = MI.getOperand(3);
1202 assert(MOSubIdx.isImm() &&
1203 "One of the subindex of the reg_sequence is not an immediate");
1204 BaseReg.Reg = MOBaseReg.getReg();
1205 BaseReg.SubReg = MOBaseReg.getSubReg();
1207 InsertedReg.Reg = MOInsertedReg.getReg();
1208 InsertedReg.SubReg = MOInsertedReg.getSubReg();
1209 InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm();