1 //===-- TwoAddressInstructionPass.cpp - Two-Address instruction pass ------===//
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 TwoAddress instruction pass which is used
11 // by most register allocators. Two-Address instructions are rewritten
21 // Note that if a register allocator chooses to use this pass, that it
22 // has to be capable of handling the non-SSA nature of these rewritten
25 // It is also worth noting that the duplicate operand of the two
26 // address instruction is removed.
28 //===----------------------------------------------------------------------===//
30 #include "llvm/CodeGen/Passes.h"
31 #include "llvm/ADT/BitVector.h"
32 #include "llvm/ADT/DenseMap.h"
33 #include "llvm/ADT/STLExtras.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
38 #include "llvm/CodeGen/LiveVariables.h"
39 #include "llvm/CodeGen/MachineFunctionPass.h"
40 #include "llvm/CodeGen/MachineInstr.h"
41 #include "llvm/CodeGen/MachineInstrBuilder.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/MC/MCInstrItineraries.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetMachine.h"
51 #include "llvm/Target/TargetRegisterInfo.h"
52 #include "llvm/Target/TargetSubtargetInfo.h"
55 #define DEBUG_TYPE "twoaddrinstr"
57 STATISTIC(NumTwoAddressInstrs, "Number of two-address instructions");
58 STATISTIC(NumCommuted , "Number of instructions commuted to coalesce");
59 STATISTIC(NumAggrCommuted , "Number of instructions aggressively commuted");
60 STATISTIC(NumConvertedTo3Addr, "Number of instructions promoted to 3-address");
61 STATISTIC(Num3AddrSunk, "Number of 3-address instructions sunk");
62 STATISTIC(NumReSchedUps, "Number of instructions re-scheduled up");
63 STATISTIC(NumReSchedDowns, "Number of instructions re-scheduled down");
65 // Temporary flag to disable rescheduling.
67 EnableRescheduling("twoaddr-reschedule",
68 cl::desc("Coalesce copies by rescheduling (default=true)"),
69 cl::init(true), cl::Hidden);
72 class TwoAddressInstructionPass : public MachineFunctionPass {
74 const TargetInstrInfo *TII;
75 const TargetRegisterInfo *TRI;
76 const InstrItineraryData *InstrItins;
77 MachineRegisterInfo *MRI;
81 CodeGenOpt::Level OptLevel;
83 // The current basic block being processed.
84 MachineBasicBlock *MBB;
86 // DistanceMap - Keep track the distance of a MI from the start of the
87 // current basic block.
88 DenseMap<MachineInstr*, unsigned> DistanceMap;
90 // Set of already processed instructions in the current block.
91 SmallPtrSet<MachineInstr*, 8> Processed;
93 // SrcRegMap - A map from virtual registers to physical registers which are
94 // likely targets to be coalesced to due to copies from physical registers to
95 // virtual registers. e.g. v1024 = move r0.
96 DenseMap<unsigned, unsigned> SrcRegMap;
98 // DstRegMap - A map from virtual registers to physical registers which are
99 // likely targets to be coalesced to due to copies to physical registers from
100 // virtual registers. e.g. r1 = move v1024.
101 DenseMap<unsigned, unsigned> DstRegMap;
103 bool sink3AddrInstruction(MachineInstr *MI, unsigned Reg,
104 MachineBasicBlock::iterator OldPos);
106 bool isRevCopyChain(unsigned FromReg, unsigned ToReg, int Maxlen);
108 bool noUseAfterLastDef(unsigned Reg, unsigned Dist, unsigned &LastDef);
110 bool isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
111 MachineInstr *MI, unsigned Dist);
113 bool commuteInstruction(MachineInstr *MI,
114 unsigned RegBIdx, unsigned RegCIdx, unsigned Dist);
116 bool isProfitableToConv3Addr(unsigned RegA, unsigned RegB);
118 bool convertInstTo3Addr(MachineBasicBlock::iterator &mi,
119 MachineBasicBlock::iterator &nmi,
120 unsigned RegA, unsigned RegB, unsigned Dist);
122 bool isDefTooClose(unsigned Reg, unsigned Dist, MachineInstr *MI);
124 bool rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
125 MachineBasicBlock::iterator &nmi,
127 bool rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
128 MachineBasicBlock::iterator &nmi,
131 bool tryInstructionTransform(MachineBasicBlock::iterator &mi,
132 MachineBasicBlock::iterator &nmi,
133 unsigned SrcIdx, unsigned DstIdx,
134 unsigned Dist, bool shouldOnlyCommute);
136 bool tryInstructionCommute(MachineInstr *MI,
141 void scanUses(unsigned DstReg);
143 void processCopy(MachineInstr *MI);
145 typedef SmallVector<std::pair<unsigned, unsigned>, 4> TiedPairList;
146 typedef SmallDenseMap<unsigned, TiedPairList> TiedOperandMap;
147 bool collectTiedOperands(MachineInstr *MI, TiedOperandMap&);
148 void processTiedPairs(MachineInstr *MI, TiedPairList&, unsigned &Dist);
149 void eliminateRegSequence(MachineBasicBlock::iterator&);
152 static char ID; // Pass identification, replacement for typeid
153 TwoAddressInstructionPass() : MachineFunctionPass(ID) {
154 initializeTwoAddressInstructionPassPass(*PassRegistry::getPassRegistry());
157 void getAnalysisUsage(AnalysisUsage &AU) const override {
158 AU.setPreservesCFG();
159 AU.addRequired<AAResultsWrapperPass>();
160 AU.addPreserved<LiveVariables>();
161 AU.addPreserved<SlotIndexes>();
162 AU.addPreserved<LiveIntervals>();
163 AU.addPreservedID(MachineLoopInfoID);
164 AU.addPreservedID(MachineDominatorsID);
165 MachineFunctionPass::getAnalysisUsage(AU);
168 /// runOnMachineFunction - Pass entry point.
169 bool runOnMachineFunction(MachineFunction&) override;
171 } // end anonymous namespace
173 char TwoAddressInstructionPass::ID = 0;
174 INITIALIZE_PASS_BEGIN(TwoAddressInstructionPass, "twoaddressinstruction",
175 "Two-Address instruction pass", false, false)
176 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
177 INITIALIZE_PASS_END(TwoAddressInstructionPass, "twoaddressinstruction",
178 "Two-Address instruction pass", false, false)
180 char &llvm::TwoAddressInstructionPassID = TwoAddressInstructionPass::ID;
182 static bool isPlainlyKilled(MachineInstr *MI, unsigned Reg, LiveIntervals *LIS);
184 /// sink3AddrInstruction - A two-address instruction has been converted to a
185 /// three-address instruction to avoid clobbering a register. Try to sink it
186 /// past the instruction that would kill the above mentioned register to reduce
187 /// register pressure.
188 bool TwoAddressInstructionPass::
189 sink3AddrInstruction(MachineInstr *MI, unsigned SavedReg,
190 MachineBasicBlock::iterator OldPos) {
191 // FIXME: Shouldn't we be trying to do this before we three-addressify the
192 // instruction? After this transformation is done, we no longer need
193 // the instruction to be in three-address form.
195 // Check if it's safe to move this instruction.
196 bool SeenStore = true; // Be conservative.
197 if (!MI->isSafeToMove(AA, SeenStore))
201 SmallSet<unsigned, 4> UseRegs;
203 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
204 const MachineOperand &MO = MI->getOperand(i);
207 unsigned MOReg = MO.getReg();
210 if (MO.isUse() && MOReg != SavedReg)
211 UseRegs.insert(MO.getReg());
215 // Don't try to move it if it implicitly defines a register.
218 // For now, don't move any instructions that define multiple registers.
220 DefReg = MO.getReg();
223 // Find the instruction that kills SavedReg.
224 MachineInstr *KillMI = nullptr;
226 LiveInterval &LI = LIS->getInterval(SavedReg);
227 assert(LI.end() != LI.begin() &&
228 "Reg should not have empty live interval.");
230 SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
231 LiveInterval::const_iterator I = LI.find(MBBEndIdx);
232 if (I != LI.end() && I->start < MBBEndIdx)
236 KillMI = LIS->getInstructionFromIndex(I->end);
239 for (MachineRegisterInfo::use_nodbg_iterator
240 UI = MRI->use_nodbg_begin(SavedReg),
241 UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
242 MachineOperand &UseMO = *UI;
245 KillMI = UseMO.getParent();
250 // If we find the instruction that kills SavedReg, and it is in an
251 // appropriate location, we can try to sink the current instruction
253 if (!KillMI || KillMI->getParent() != MBB || KillMI == MI ||
254 KillMI == OldPos || KillMI->isTerminator())
257 // If any of the definitions are used by another instruction between the
258 // position and the kill use, then it's not safe to sink it.
260 // FIXME: This can be sped up if there is an easy way to query whether an
261 // instruction is before or after another instruction. Then we can use
262 // MachineRegisterInfo def / use instead.
263 MachineOperand *KillMO = nullptr;
264 MachineBasicBlock::iterator KillPos = KillMI;
267 unsigned NumVisited = 0;
268 for (MachineBasicBlock::iterator I = std::next(OldPos); I != KillPos; ++I) {
269 MachineInstr *OtherMI = I;
270 // DBG_VALUE cannot be counted against the limit.
271 if (OtherMI->isDebugValue())
273 if (NumVisited > 30) // FIXME: Arbitrary limit to reduce compile time cost.
276 for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
277 MachineOperand &MO = OtherMI->getOperand(i);
280 unsigned MOReg = MO.getReg();
286 if (MO.isKill() || (LIS && isPlainlyKilled(OtherMI, MOReg, LIS))) {
287 if (OtherMI == KillMI && MOReg == SavedReg)
288 // Save the operand that kills the register. We want to unset the kill
289 // marker if we can sink MI past it.
291 else if (UseRegs.count(MOReg))
292 // One of the uses is killed before the destination.
297 assert(KillMO && "Didn't find kill");
300 // Update kill and LV information.
301 KillMO->setIsKill(false);
302 KillMO = MI->findRegisterUseOperand(SavedReg, false, TRI);
303 KillMO->setIsKill(true);
306 LV->replaceKillInstruction(SavedReg, KillMI, MI);
309 // Move instruction to its destination.
311 MBB->insert(KillPos, MI);
320 /// getSingleDef -- return the MachineInstr* if it is the single def of the Reg
322 static MachineInstr *getSingleDef(unsigned Reg, MachineBasicBlock *BB,
323 const MachineRegisterInfo *MRI) {
324 MachineInstr *Ret = nullptr;
325 for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
326 if (DefMI.getParent() != BB || DefMI.isDebugValue())
330 else if (Ret != &DefMI)
336 /// Check if there is a reversed copy chain from FromReg to ToReg:
337 /// %Tmp1 = copy %Tmp2;
338 /// %FromReg = copy %Tmp1;
339 /// %ToReg = add %FromReg ...
340 /// %Tmp2 = copy %ToReg;
341 /// MaxLen specifies the maximum length of the copy chain the func
342 /// can walk through.
343 bool TwoAddressInstructionPass::isRevCopyChain(unsigned FromReg, unsigned ToReg,
345 unsigned TmpReg = FromReg;
346 for (int i = 0; i < Maxlen; i++) {
347 MachineInstr *Def = getSingleDef(TmpReg, MBB, MRI);
348 if (!Def || !Def->isCopy())
351 TmpReg = Def->getOperand(1).getReg();
359 /// noUseAfterLastDef - Return true if there are no intervening uses between the
360 /// last instruction in the MBB that defines the specified register and the
361 /// two-address instruction which is being processed. It also returns the last
362 /// def location by reference
363 bool TwoAddressInstructionPass::noUseAfterLastDef(unsigned Reg, unsigned Dist,
366 unsigned LastUse = Dist;
367 for (MachineOperand &MO : MRI->reg_operands(Reg)) {
368 MachineInstr *MI = MO.getParent();
369 if (MI->getParent() != MBB || MI->isDebugValue())
371 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
372 if (DI == DistanceMap.end())
374 if (MO.isUse() && DI->second < LastUse)
375 LastUse = DI->second;
376 if (MO.isDef() && DI->second > LastDef)
377 LastDef = DI->second;
380 return !(LastUse > LastDef && LastUse < Dist);
383 /// isCopyToReg - Return true if the specified MI is a copy instruction or
384 /// a extract_subreg instruction. It also returns the source and destination
385 /// registers and whether they are physical registers by reference.
386 static bool isCopyToReg(MachineInstr &MI, const TargetInstrInfo *TII,
387 unsigned &SrcReg, unsigned &DstReg,
388 bool &IsSrcPhys, bool &IsDstPhys) {
392 DstReg = MI.getOperand(0).getReg();
393 SrcReg = MI.getOperand(1).getReg();
394 } else if (MI.isInsertSubreg() || MI.isSubregToReg()) {
395 DstReg = MI.getOperand(0).getReg();
396 SrcReg = MI.getOperand(2).getReg();
400 IsSrcPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
401 IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
405 /// isPLainlyKilled - Test if the given register value, which is used by the
406 // given instruction, is killed by the given instruction.
407 static bool isPlainlyKilled(MachineInstr *MI, unsigned Reg,
408 LiveIntervals *LIS) {
409 if (LIS && TargetRegisterInfo::isVirtualRegister(Reg) &&
410 !LIS->isNotInMIMap(MI)) {
411 // FIXME: Sometimes tryInstructionTransform() will add instructions and
412 // test whether they can be folded before keeping them. In this case it
413 // sets a kill before recursively calling tryInstructionTransform() again.
414 // If there is no interval available, we assume that this instruction is
415 // one of those. A kill flag is manually inserted on the operand so the
416 // check below will handle it.
417 LiveInterval &LI = LIS->getInterval(Reg);
418 // This is to match the kill flag version where undefs don't have kill
420 if (!LI.hasAtLeastOneValue())
423 SlotIndex useIdx = LIS->getInstructionIndex(MI);
424 LiveInterval::const_iterator I = LI.find(useIdx);
425 assert(I != LI.end() && "Reg must be live-in to use.");
426 return !I->end.isBlock() && SlotIndex::isSameInstr(I->end, useIdx);
429 return MI->killsRegister(Reg);
432 /// isKilled - Test if the given register value, which is used by the given
433 /// instruction, is killed by the given instruction. This looks through
434 /// coalescable copies to see if the original value is potentially not killed.
436 /// For example, in this code:
438 /// %reg1034 = copy %reg1024
439 /// %reg1035 = copy %reg1025<kill>
440 /// %reg1036 = add %reg1034<kill>, %reg1035<kill>
442 /// %reg1034 is not considered to be killed, since it is copied from a
443 /// register which is not killed. Treating it as not killed lets the
444 /// normal heuristics commute the (two-address) add, which lets
445 /// coalescing eliminate the extra copy.
447 /// If allowFalsePositives is true then likely kills are treated as kills even
448 /// if it can't be proven that they are kills.
449 static bool isKilled(MachineInstr &MI, unsigned Reg,
450 const MachineRegisterInfo *MRI,
451 const TargetInstrInfo *TII,
453 bool allowFalsePositives) {
454 MachineInstr *DefMI = &MI;
456 // All uses of physical registers are likely to be kills.
457 if (TargetRegisterInfo::isPhysicalRegister(Reg) &&
458 (allowFalsePositives || MRI->hasOneUse(Reg)))
460 if (!isPlainlyKilled(DefMI, Reg, LIS))
462 if (TargetRegisterInfo::isPhysicalRegister(Reg))
464 MachineRegisterInfo::def_iterator Begin = MRI->def_begin(Reg);
465 // If there are multiple defs, we can't do a simple analysis, so just
466 // go with what the kill flag says.
467 if (std::next(Begin) != MRI->def_end())
469 DefMI = Begin->getParent();
470 bool IsSrcPhys, IsDstPhys;
471 unsigned SrcReg, DstReg;
472 // If the def is something other than a copy, then it isn't going to
473 // be coalesced, so follow the kill flag.
474 if (!isCopyToReg(*DefMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
480 /// isTwoAddrUse - Return true if the specified MI uses the specified register
481 /// as a two-address use. If so, return the destination register by reference.
482 static bool isTwoAddrUse(MachineInstr &MI, unsigned Reg, unsigned &DstReg) {
483 for (unsigned i = 0, NumOps = MI.getNumOperands(); i != NumOps; ++i) {
484 const MachineOperand &MO = MI.getOperand(i);
485 if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
488 if (MI.isRegTiedToDefOperand(i, &ti)) {
489 DstReg = MI.getOperand(ti).getReg();
496 /// findOnlyInterestingUse - Given a register, if has a single in-basic block
497 /// use, return the use instruction if it's a copy or a two-address use.
499 MachineInstr *findOnlyInterestingUse(unsigned Reg, MachineBasicBlock *MBB,
500 MachineRegisterInfo *MRI,
501 const TargetInstrInfo *TII,
503 unsigned &DstReg, bool &IsDstPhys) {
504 if (!MRI->hasOneNonDBGUse(Reg))
505 // None or more than one use.
507 MachineInstr &UseMI = *MRI->use_instr_nodbg_begin(Reg);
508 if (UseMI.getParent() != MBB)
512 if (isCopyToReg(UseMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys)) {
517 if (isTwoAddrUse(UseMI, Reg, DstReg)) {
518 IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
524 /// getMappedReg - Return the physical register the specified virtual register
525 /// might be mapped to.
527 getMappedReg(unsigned Reg, DenseMap<unsigned, unsigned> &RegMap) {
528 while (TargetRegisterInfo::isVirtualRegister(Reg)) {
529 DenseMap<unsigned, unsigned>::iterator SI = RegMap.find(Reg);
530 if (SI == RegMap.end())
534 if (TargetRegisterInfo::isPhysicalRegister(Reg))
539 /// regsAreCompatible - Return true if the two registers are equal or aliased.
542 regsAreCompatible(unsigned RegA, unsigned RegB, const TargetRegisterInfo *TRI) {
547 return TRI->regsOverlap(RegA, RegB);
551 /// isProfitableToCommute - Return true if it's potentially profitable to commute
552 /// the two-address instruction that's being processed.
554 TwoAddressInstructionPass::
555 isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
556 MachineInstr *MI, unsigned Dist) {
557 if (OptLevel == CodeGenOpt::None)
560 // Determine if it's profitable to commute this two address instruction. In
561 // general, we want no uses between this instruction and the definition of
562 // the two-address register.
564 // %reg1028<def> = EXTRACT_SUBREG %reg1027<kill>, 1
565 // %reg1029<def> = MOV8rr %reg1028
566 // %reg1029<def> = SHR8ri %reg1029, 7, %EFLAGS<imp-def,dead>
567 // insert => %reg1030<def> = MOV8rr %reg1028
568 // %reg1030<def> = ADD8rr %reg1028<kill>, %reg1029<kill>, %EFLAGS<imp-def,dead>
569 // In this case, it might not be possible to coalesce the second MOV8rr
570 // instruction if the first one is coalesced. So it would be profitable to
572 // %reg1028<def> = EXTRACT_SUBREG %reg1027<kill>, 1
573 // %reg1029<def> = MOV8rr %reg1028
574 // %reg1029<def> = SHR8ri %reg1029, 7, %EFLAGS<imp-def,dead>
575 // insert => %reg1030<def> = MOV8rr %reg1029
576 // %reg1030<def> = ADD8rr %reg1029<kill>, %reg1028<kill>, %EFLAGS<imp-def,dead>
578 if (!isPlainlyKilled(MI, regC, LIS))
581 // Ok, we have something like:
582 // %reg1030<def> = ADD8rr %reg1028<kill>, %reg1029<kill>, %EFLAGS<imp-def,dead>
583 // let's see if it's worth commuting it.
585 // Look for situations like this:
586 // %reg1024<def> = MOV r1
587 // %reg1025<def> = MOV r0
588 // %reg1026<def> = ADD %reg1024, %reg1025
590 // Commute the ADD to hopefully eliminate an otherwise unavoidable copy.
591 unsigned ToRegA = getMappedReg(regA, DstRegMap);
593 unsigned FromRegB = getMappedReg(regB, SrcRegMap);
594 unsigned FromRegC = getMappedReg(regC, SrcRegMap);
595 bool CompB = FromRegB && regsAreCompatible(FromRegB, ToRegA, TRI);
596 bool CompC = FromRegC && regsAreCompatible(FromRegC, ToRegA, TRI);
598 // Compute if any of the following are true:
599 // -RegB is not tied to a register and RegC is compatible with RegA.
600 // -RegB is tied to the wrong physical register, but RegC is.
601 // -RegB is tied to the wrong physical register, and RegC isn't tied.
602 if ((!FromRegB && CompC) || (FromRegB && !CompB && (!FromRegC || CompC)))
604 // Don't compute if any of the following are true:
605 // -RegC is not tied to a register and RegB is compatible with RegA.
606 // -RegC is tied to the wrong physical register, but RegB is.
607 // -RegC is tied to the wrong physical register, and RegB isn't tied.
608 if ((!FromRegC && CompB) || (FromRegC && !CompC && (!FromRegB || CompB)))
612 // If there is a use of regC between its last def (could be livein) and this
613 // instruction, then bail.
614 unsigned LastDefC = 0;
615 if (!noUseAfterLastDef(regC, Dist, LastDefC))
618 // If there is a use of regB between its last def (could be livein) and this
619 // instruction, then go ahead and make this transformation.
620 unsigned LastDefB = 0;
621 if (!noUseAfterLastDef(regB, Dist, LastDefB))
624 // Look for situation like this:
625 // %reg101 = MOV %reg100
627 // %reg103 = ADD %reg102, %reg101
629 // %reg100 = MOV %reg103
630 // If there is a reversed copy chain from reg101 to reg103, commute the ADD
631 // to eliminate an otherwise unavoidable copy.
633 // We can extend the logic further: If an pair of operands in an insn has
634 // been merged, the insn could be regarded as a virtual copy, and the virtual
635 // copy could also be used to construct a copy chain.
636 // To more generally minimize register copies, ideally the logic of two addr
637 // instruction pass should be integrated with register allocation pass where
638 // interference graph is available.
639 if (isRevCopyChain(regC, regA, 3))
642 if (isRevCopyChain(regB, regA, 3))
645 // Since there are no intervening uses for both registers, then commute
646 // if the def of regC is closer. Its live interval is shorter.
647 return LastDefB && LastDefC && LastDefC > LastDefB;
650 /// commuteInstruction - Commute a two-address instruction and update the basic
651 /// block, distance map, and live variables if needed. Return true if it is
653 bool TwoAddressInstructionPass::commuteInstruction(MachineInstr *MI,
657 unsigned RegC = MI->getOperand(RegCIdx).getReg();
658 DEBUG(dbgs() << "2addr: COMMUTING : " << *MI);
659 MachineInstr *NewMI = TII->commuteInstruction(MI, false, RegBIdx, RegCIdx);
661 if (NewMI == nullptr) {
662 DEBUG(dbgs() << "2addr: COMMUTING FAILED!\n");
666 DEBUG(dbgs() << "2addr: COMMUTED TO: " << *NewMI);
667 assert(NewMI == MI &&
668 "TargetInstrInfo::commuteInstruction() should not return a new "
669 "instruction unless it was requested.");
671 // Update source register map.
672 unsigned FromRegC = getMappedReg(RegC, SrcRegMap);
674 unsigned RegA = MI->getOperand(0).getReg();
675 SrcRegMap[RegA] = FromRegC;
681 /// isProfitableToConv3Addr - Return true if it is profitable to convert the
682 /// given 2-address instruction to a 3-address one.
684 TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA,unsigned RegB){
685 // Look for situations like this:
686 // %reg1024<def> = MOV r1
687 // %reg1025<def> = MOV r0
688 // %reg1026<def> = ADD %reg1024, %reg1025
690 // Turn ADD into a 3-address instruction to avoid a copy.
691 unsigned FromRegB = getMappedReg(RegB, SrcRegMap);
694 unsigned ToRegA = getMappedReg(RegA, DstRegMap);
695 return (ToRegA && !regsAreCompatible(FromRegB, ToRegA, TRI));
698 /// convertInstTo3Addr - Convert the specified two-address instruction into a
699 /// three address one. Return true if this transformation was successful.
701 TwoAddressInstructionPass::convertInstTo3Addr(MachineBasicBlock::iterator &mi,
702 MachineBasicBlock::iterator &nmi,
703 unsigned RegA, unsigned RegB,
705 // FIXME: Why does convertToThreeAddress() need an iterator reference?
706 MachineFunction::iterator MFI = MBB;
707 MachineInstr *NewMI = TII->convertToThreeAddress(MFI, mi, LV);
708 assert(MBB == MFI && "convertToThreeAddress changed iterator reference");
712 DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi);
713 DEBUG(dbgs() << "2addr: TO 3-ADDR: " << *NewMI);
717 LIS->ReplaceMachineInstrInMaps(mi, NewMI);
719 if (NewMI->findRegisterUseOperand(RegB, false, TRI))
720 // FIXME: Temporary workaround. If the new instruction doesn't
721 // uses RegB, convertToThreeAddress must have created more
722 // then one instruction.
723 Sunk = sink3AddrInstruction(NewMI, RegB, mi);
725 MBB->erase(mi); // Nuke the old inst.
728 DistanceMap.insert(std::make_pair(NewMI, Dist));
733 // Update source and destination register maps.
734 SrcRegMap.erase(RegA);
735 DstRegMap.erase(RegB);
739 /// scanUses - Scan forward recursively for only uses, update maps if the use
740 /// is a copy or a two-address instruction.
742 TwoAddressInstructionPass::scanUses(unsigned DstReg) {
743 SmallVector<unsigned, 4> VirtRegPairs;
747 unsigned Reg = DstReg;
748 while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, MRI, TII,IsCopy,
749 NewReg, IsDstPhys)) {
750 if (IsCopy && !Processed.insert(UseMI).second)
753 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
754 if (DI != DistanceMap.end())
755 // Earlier in the same MBB.Reached via a back edge.
759 VirtRegPairs.push_back(NewReg);
762 bool isNew = SrcRegMap.insert(std::make_pair(NewReg, Reg)).second;
764 assert(SrcRegMap[NewReg] == Reg && "Can't map to two src registers!");
765 VirtRegPairs.push_back(NewReg);
769 if (!VirtRegPairs.empty()) {
770 unsigned ToReg = VirtRegPairs.back();
771 VirtRegPairs.pop_back();
772 while (!VirtRegPairs.empty()) {
773 unsigned FromReg = VirtRegPairs.back();
774 VirtRegPairs.pop_back();
775 bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
777 assert(DstRegMap[FromReg] == ToReg &&"Can't map to two dst registers!");
780 bool isNew = DstRegMap.insert(std::make_pair(DstReg, ToReg)).second;
782 assert(DstRegMap[DstReg] == ToReg && "Can't map to two dst registers!");
786 /// processCopy - If the specified instruction is not yet processed, process it
787 /// if it's a copy. For a copy instruction, we find the physical registers the
788 /// source and destination registers might be mapped to. These are kept in
789 /// point-to maps used to determine future optimizations. e.g.
792 /// v1026 = add v1024, v1025
794 /// If 'add' is a two-address instruction, v1024, v1026 are both potentially
795 /// coalesced to r0 (from the input side). v1025 is mapped to r1. v1026 is
796 /// potentially joined with r1 on the output side. It's worthwhile to commute
797 /// 'add' to eliminate a copy.
798 void TwoAddressInstructionPass::processCopy(MachineInstr *MI) {
799 if (Processed.count(MI))
802 bool IsSrcPhys, IsDstPhys;
803 unsigned SrcReg, DstReg;
804 if (!isCopyToReg(*MI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
807 if (IsDstPhys && !IsSrcPhys)
808 DstRegMap.insert(std::make_pair(SrcReg, DstReg));
809 else if (!IsDstPhys && IsSrcPhys) {
810 bool isNew = SrcRegMap.insert(std::make_pair(DstReg, SrcReg)).second;
812 assert(SrcRegMap[DstReg] == SrcReg &&
813 "Can't map to two src physical registers!");
818 Processed.insert(MI);
822 /// rescheduleMIBelowKill - If there is one more local instruction that reads
823 /// 'Reg' and it kills 'Reg, consider moving the instruction below the kill
824 /// instruction in order to eliminate the need for the copy.
825 bool TwoAddressInstructionPass::
826 rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
827 MachineBasicBlock::iterator &nmi,
829 // Bail immediately if we don't have LV or LIS available. We use them to find
830 // kills efficiently.
834 MachineInstr *MI = &*mi;
835 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
836 if (DI == DistanceMap.end())
837 // Must be created from unfolded load. Don't waste time trying this.
840 MachineInstr *KillMI = nullptr;
842 LiveInterval &LI = LIS->getInterval(Reg);
843 assert(LI.end() != LI.begin() &&
844 "Reg should not have empty live interval.");
846 SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
847 LiveInterval::const_iterator I = LI.find(MBBEndIdx);
848 if (I != LI.end() && I->start < MBBEndIdx)
852 KillMI = LIS->getInstructionFromIndex(I->end);
854 KillMI = LV->getVarInfo(Reg).findKill(MBB);
856 if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
857 // Don't mess with copies, they may be coalesced later.
860 if (KillMI->hasUnmodeledSideEffects() || KillMI->isCall() ||
861 KillMI->isBranch() || KillMI->isTerminator())
862 // Don't move pass calls, etc.
866 if (isTwoAddrUse(*KillMI, Reg, DstReg))
869 bool SeenStore = true;
870 if (!MI->isSafeToMove(AA, SeenStore))
873 if (TII->getInstrLatency(InstrItins, MI) > 1)
874 // FIXME: Needs more sophisticated heuristics.
877 SmallSet<unsigned, 2> Uses;
878 SmallSet<unsigned, 2> Kills;
879 SmallSet<unsigned, 2> Defs;
880 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
881 const MachineOperand &MO = MI->getOperand(i);
884 unsigned MOReg = MO.getReg();
891 if (MOReg != Reg && (MO.isKill() ||
892 (LIS && isPlainlyKilled(MI, MOReg, LIS))))
897 // Move the copies connected to MI down as well.
898 MachineBasicBlock::iterator Begin = MI;
899 MachineBasicBlock::iterator AfterMI = std::next(Begin);
901 MachineBasicBlock::iterator End = AfterMI;
902 while (End->isCopy() && Defs.count(End->getOperand(1).getReg())) {
903 Defs.insert(End->getOperand(0).getReg());
907 // Check if the reschedule will not break depedencies.
908 unsigned NumVisited = 0;
909 MachineBasicBlock::iterator KillPos = KillMI;
911 for (MachineBasicBlock::iterator I = End; I != KillPos; ++I) {
912 MachineInstr *OtherMI = I;
913 // DBG_VALUE cannot be counted against the limit.
914 if (OtherMI->isDebugValue())
916 if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
919 if (OtherMI->hasUnmodeledSideEffects() || OtherMI->isCall() ||
920 OtherMI->isBranch() || OtherMI->isTerminator())
921 // Don't move pass calls, etc.
923 for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
924 const MachineOperand &MO = OtherMI->getOperand(i);
927 unsigned MOReg = MO.getReg();
931 if (Uses.count(MOReg))
932 // Physical register use would be clobbered.
934 if (!MO.isDead() && Defs.count(MOReg))
935 // May clobber a physical register def.
936 // FIXME: This may be too conservative. It's ok if the instruction
937 // is sunken completely below the use.
940 if (Defs.count(MOReg))
942 bool isKill = MO.isKill() ||
943 (LIS && isPlainlyKilled(OtherMI, MOReg, LIS));
945 ((isKill && Uses.count(MOReg)) || Kills.count(MOReg)))
946 // Don't want to extend other live ranges and update kills.
948 if (MOReg == Reg && !isKill)
949 // We can't schedule across a use of the register in question.
951 // Ensure that if this is register in question, its the kill we expect.
952 assert((MOReg != Reg || OtherMI == KillMI) &&
953 "Found multiple kills of a register in a basic block");
958 // Move debug info as well.
959 while (Begin != MBB->begin() && std::prev(Begin)->isDebugValue())
963 MachineBasicBlock::iterator InsertPos = KillPos;
965 // We have to move the copies first so that the MBB is still well-formed
966 // when calling handleMove().
967 for (MachineBasicBlock::iterator MBBI = AfterMI; MBBI != End;) {
968 MachineInstr *CopyMI = MBBI;
970 MBB->splice(InsertPos, MBB, CopyMI);
971 LIS->handleMove(CopyMI);
974 End = std::next(MachineBasicBlock::iterator(MI));
977 // Copies following MI may have been moved as well.
978 MBB->splice(InsertPos, MBB, Begin, End);
979 DistanceMap.erase(DI);
981 // Update live variables
985 LV->removeVirtualRegisterKilled(Reg, KillMI);
986 LV->addVirtualRegisterKilled(Reg, MI);
989 DEBUG(dbgs() << "\trescheduled below kill: " << *KillMI);
993 /// isDefTooClose - Return true if the re-scheduling will put the given
994 /// instruction too close to the defs of its register dependencies.
995 bool TwoAddressInstructionPass::isDefTooClose(unsigned Reg, unsigned Dist,
997 for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
998 if (DefMI.getParent() != MBB || DefMI.isCopy() || DefMI.isCopyLike())
1001 return true; // MI is defining something KillMI uses
1002 DenseMap<MachineInstr*, unsigned>::iterator DDI = DistanceMap.find(&DefMI);
1003 if (DDI == DistanceMap.end())
1004 return true; // Below MI
1005 unsigned DefDist = DDI->second;
1006 assert(Dist > DefDist && "Visited def already?");
1007 if (TII->getInstrLatency(InstrItins, &DefMI) > (Dist - DefDist))
1013 /// rescheduleKillAboveMI - If there is one more local instruction that reads
1014 /// 'Reg' and it kills 'Reg, consider moving the kill instruction above the
1015 /// current two-address instruction in order to eliminate the need for the
1017 bool TwoAddressInstructionPass::
1018 rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
1019 MachineBasicBlock::iterator &nmi,
1021 // Bail immediately if we don't have LV or LIS available. We use them to find
1022 // kills efficiently.
1026 MachineInstr *MI = &*mi;
1027 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
1028 if (DI == DistanceMap.end())
1029 // Must be created from unfolded load. Don't waste time trying this.
1032 MachineInstr *KillMI = nullptr;
1034 LiveInterval &LI = LIS->getInterval(Reg);
1035 assert(LI.end() != LI.begin() &&
1036 "Reg should not have empty live interval.");
1038 SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
1039 LiveInterval::const_iterator I = LI.find(MBBEndIdx);
1040 if (I != LI.end() && I->start < MBBEndIdx)
1044 KillMI = LIS->getInstructionFromIndex(I->end);
1046 KillMI = LV->getVarInfo(Reg).findKill(MBB);
1048 if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
1049 // Don't mess with copies, they may be coalesced later.
1053 if (isTwoAddrUse(*KillMI, Reg, DstReg))
1056 bool SeenStore = true;
1057 if (!KillMI->isSafeToMove(AA, SeenStore))
1060 SmallSet<unsigned, 2> Uses;
1061 SmallSet<unsigned, 2> Kills;
1062 SmallSet<unsigned, 2> Defs;
1063 SmallSet<unsigned, 2> LiveDefs;
1064 for (unsigned i = 0, e = KillMI->getNumOperands(); i != e; ++i) {
1065 const MachineOperand &MO = KillMI->getOperand(i);
1068 unsigned MOReg = MO.getReg();
1072 if (isDefTooClose(MOReg, DI->second, MI))
1074 bool isKill = MO.isKill() || (LIS && isPlainlyKilled(KillMI, MOReg, LIS));
1075 if (MOReg == Reg && !isKill)
1078 if (isKill && MOReg != Reg)
1079 Kills.insert(MOReg);
1080 } else if (TargetRegisterInfo::isPhysicalRegister(MOReg)) {
1083 LiveDefs.insert(MOReg);
1087 // Check if the reschedule will not break depedencies.
1088 unsigned NumVisited = 0;
1089 MachineBasicBlock::iterator KillPos = KillMI;
1090 for (MachineBasicBlock::iterator I = mi; I != KillPos; ++I) {
1091 MachineInstr *OtherMI = I;
1092 // DBG_VALUE cannot be counted against the limit.
1093 if (OtherMI->isDebugValue())
1095 if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
1098 if (OtherMI->hasUnmodeledSideEffects() || OtherMI->isCall() ||
1099 OtherMI->isBranch() || OtherMI->isTerminator())
1100 // Don't move pass calls, etc.
1102 SmallVector<unsigned, 2> OtherDefs;
1103 for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
1104 const MachineOperand &MO = OtherMI->getOperand(i);
1107 unsigned MOReg = MO.getReg();
1111 if (Defs.count(MOReg))
1112 // Moving KillMI can clobber the physical register if the def has
1115 if (Kills.count(MOReg))
1116 // Don't want to extend other live ranges and update kills.
1118 if (OtherMI != MI && MOReg == Reg &&
1119 !(MO.isKill() || (LIS && isPlainlyKilled(OtherMI, MOReg, LIS))))
1120 // We can't schedule across a use of the register in question.
1123 OtherDefs.push_back(MOReg);
1127 for (unsigned i = 0, e = OtherDefs.size(); i != e; ++i) {
1128 unsigned MOReg = OtherDefs[i];
1129 if (Uses.count(MOReg))
1131 if (TargetRegisterInfo::isPhysicalRegister(MOReg) &&
1132 LiveDefs.count(MOReg))
1134 // Physical register def is seen.
1139 // Move the old kill above MI, don't forget to move debug info as well.
1140 MachineBasicBlock::iterator InsertPos = mi;
1141 while (InsertPos != MBB->begin() && std::prev(InsertPos)->isDebugValue())
1143 MachineBasicBlock::iterator From = KillMI;
1144 MachineBasicBlock::iterator To = std::next(From);
1145 while (std::prev(From)->isDebugValue())
1147 MBB->splice(InsertPos, MBB, From, To);
1149 nmi = std::prev(InsertPos); // Backtrack so we process the moved instr.
1150 DistanceMap.erase(DI);
1152 // Update live variables
1154 LIS->handleMove(KillMI);
1156 LV->removeVirtualRegisterKilled(Reg, KillMI);
1157 LV->addVirtualRegisterKilled(Reg, MI);
1160 DEBUG(dbgs() << "\trescheduled kill: " << *KillMI);
1164 /// Tries to commute the operand 'BaseOpIdx' and some other operand in the
1165 /// given machine instruction to improve opportunities for coalescing and
1166 /// elimination of a register to register copy.
1168 /// 'DstOpIdx' specifies the index of MI def operand.
1169 /// 'BaseOpKilled' specifies if the register associated with 'BaseOpIdx'
1170 /// operand is killed by the given instruction.
1171 /// The 'Dist' arguments provides the distance of MI from the start of the
1172 /// current basic block and it is used to determine if it is profitable
1173 /// to commute operands in the instruction.
1175 /// Returns true if the transformation happened. Otherwise, returns false.
1176 bool TwoAddressInstructionPass::tryInstructionCommute(MachineInstr *MI,
1181 unsigned DstOpReg = MI->getOperand(DstOpIdx).getReg();
1182 unsigned BaseOpReg = MI->getOperand(BaseOpIdx).getReg();
1183 unsigned OpsNum = MI->getDesc().getNumOperands();
1184 unsigned OtherOpIdx = MI->getDesc().getNumDefs();
1185 for (; OtherOpIdx < OpsNum; OtherOpIdx++) {
1186 // The call of findCommutedOpIndices below only checks if BaseOpIdx
1187 // and OtherOpIdx are commutable, it does not really searches for
1188 // other commutable operands and does not change the values of passed
1190 if (OtherOpIdx == BaseOpIdx ||
1191 !TII->findCommutedOpIndices(MI, BaseOpIdx, OtherOpIdx))
1194 unsigned OtherOpReg = MI->getOperand(OtherOpIdx).getReg();
1195 bool AggressiveCommute = false;
1197 // If OtherOp dies but BaseOp does not, swap the OtherOp and BaseOp
1198 // operands. This makes the live ranges of DstOp and OtherOp joinable.
1200 !BaseOpKilled && isKilled(*MI, OtherOpReg, MRI, TII, LIS, false);
1203 isProfitableToCommute(DstOpReg, BaseOpReg, OtherOpReg, MI, Dist)) {
1205 AggressiveCommute = true;
1208 // If it's profitable to commute, try to do so.
1209 if (DoCommute && commuteInstruction(MI, BaseOpIdx, OtherOpIdx, Dist)) {
1211 if (AggressiveCommute)
1219 /// tryInstructionTransform - For the case where an instruction has a single
1220 /// pair of tied register operands, attempt some transformations that may
1221 /// either eliminate the tied operands or improve the opportunities for
1222 /// coalescing away the register copy. Returns true if no copy needs to be
1223 /// inserted to untie mi's operands (either because they were untied, or
1224 /// because mi was rescheduled, and will be visited again later). If the
1225 /// shouldOnlyCommute flag is true, only instruction commutation is attempted.
1226 bool TwoAddressInstructionPass::
1227 tryInstructionTransform(MachineBasicBlock::iterator &mi,
1228 MachineBasicBlock::iterator &nmi,
1229 unsigned SrcIdx, unsigned DstIdx,
1230 unsigned Dist, bool shouldOnlyCommute) {
1231 if (OptLevel == CodeGenOpt::None)
1234 MachineInstr &MI = *mi;
1235 unsigned regA = MI.getOperand(DstIdx).getReg();
1236 unsigned regB = MI.getOperand(SrcIdx).getReg();
1238 assert(TargetRegisterInfo::isVirtualRegister(regB) &&
1239 "cannot make instruction into two-address form");
1240 bool regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
1242 if (TargetRegisterInfo::isVirtualRegister(regA))
1245 bool Commuted = tryInstructionCommute(&MI, DstIdx, SrcIdx, regBKilled, Dist);
1247 // If the instruction is convertible to 3 Addr, instead
1248 // of returning try 3 Addr transformation aggresively and
1249 // use this variable to check later. Because it might be better.
1250 // For example, we can just use `leal (%rsi,%rdi), %eax` and `ret`
1251 // instead of the following code.
1255 if (Commuted && !MI.isConvertibleTo3Addr())
1258 if (shouldOnlyCommute)
1261 // If there is one more use of regB later in the same MBB, consider
1262 // re-schedule this MI below it.
1263 if (!Commuted && EnableRescheduling && rescheduleMIBelowKill(mi, nmi, regB)) {
1268 if (MI.isConvertibleTo3Addr()) {
1269 // This instruction is potentially convertible to a true
1270 // three-address instruction. Check if it is profitable.
1271 if (!regBKilled || isProfitableToConv3Addr(regA, regB)) {
1272 // Try to convert it.
1273 if (convertInstTo3Addr(mi, nmi, regA, regB, Dist)) {
1274 ++NumConvertedTo3Addr;
1275 return true; // Done with this instruction.
1280 // Return if it is commuted but 3 addr conversion is failed.
1284 // If there is one more use of regB later in the same MBB, consider
1285 // re-schedule it before this MI if it's legal.
1286 if (EnableRescheduling && rescheduleKillAboveMI(mi, nmi, regB)) {
1291 // If this is an instruction with a load folded into it, try unfolding
1292 // the load, e.g. avoid this:
1294 // addq (%rax), %rcx
1295 // in favor of this:
1296 // movq (%rax), %rcx
1298 // because it's preferable to schedule a load than a register copy.
1299 if (MI.mayLoad() && !regBKilled) {
1300 // Determine if a load can be unfolded.
1301 unsigned LoadRegIndex;
1303 TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(),
1304 /*UnfoldLoad=*/true,
1305 /*UnfoldStore=*/false,
1308 const MCInstrDesc &UnfoldMCID = TII->get(NewOpc);
1309 if (UnfoldMCID.getNumDefs() == 1) {
1311 DEBUG(dbgs() << "2addr: UNFOLDING: " << MI);
1312 const TargetRegisterClass *RC =
1313 TRI->getAllocatableClass(
1314 TII->getRegClass(UnfoldMCID, LoadRegIndex, TRI, *MF));
1315 unsigned Reg = MRI->createVirtualRegister(RC);
1316 SmallVector<MachineInstr *, 2> NewMIs;
1317 if (!TII->unfoldMemoryOperand(*MF, &MI, Reg,
1318 /*UnfoldLoad=*/true,/*UnfoldStore=*/false,
1320 DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
1323 assert(NewMIs.size() == 2 &&
1324 "Unfolded a load into multiple instructions!");
1325 // The load was previously folded, so this is the only use.
1326 NewMIs[1]->addRegisterKilled(Reg, TRI);
1328 // Tentatively insert the instructions into the block so that they
1329 // look "normal" to the transformation logic.
1330 MBB->insert(mi, NewMIs[0]);
1331 MBB->insert(mi, NewMIs[1]);
1333 DEBUG(dbgs() << "2addr: NEW LOAD: " << *NewMIs[0]
1334 << "2addr: NEW INST: " << *NewMIs[1]);
1336 // Transform the instruction, now that it no longer has a load.
1337 unsigned NewDstIdx = NewMIs[1]->findRegisterDefOperandIdx(regA);
1338 unsigned NewSrcIdx = NewMIs[1]->findRegisterUseOperandIdx(regB);
1339 MachineBasicBlock::iterator NewMI = NewMIs[1];
1340 bool TransformResult =
1341 tryInstructionTransform(NewMI, mi, NewSrcIdx, NewDstIdx, Dist, true);
1342 (void)TransformResult;
1343 assert(!TransformResult &&
1344 "tryInstructionTransform() should return false.");
1345 if (NewMIs[1]->getOperand(NewSrcIdx).isKill()) {
1346 // Success, or at least we made an improvement. Keep the unfolded
1347 // instructions and discard the original.
1349 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1350 MachineOperand &MO = MI.getOperand(i);
1352 TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
1355 if (NewMIs[0]->killsRegister(MO.getReg()))
1356 LV->replaceKillInstruction(MO.getReg(), &MI, NewMIs[0]);
1358 assert(NewMIs[1]->killsRegister(MO.getReg()) &&
1359 "Kill missing after load unfold!");
1360 LV->replaceKillInstruction(MO.getReg(), &MI, NewMIs[1]);
1363 } else if (LV->removeVirtualRegisterDead(MO.getReg(), &MI)) {
1364 if (NewMIs[1]->registerDefIsDead(MO.getReg()))
1365 LV->addVirtualRegisterDead(MO.getReg(), NewMIs[1]);
1367 assert(NewMIs[0]->registerDefIsDead(MO.getReg()) &&
1368 "Dead flag missing after load unfold!");
1369 LV->addVirtualRegisterDead(MO.getReg(), NewMIs[0]);
1374 LV->addVirtualRegisterKilled(Reg, NewMIs[1]);
1377 SmallVector<unsigned, 4> OrigRegs;
1379 for (MachineInstr::const_mop_iterator MOI = MI.operands_begin(),
1380 MOE = MI.operands_end(); MOI != MOE; ++MOI) {
1382 OrigRegs.push_back(MOI->getReg());
1386 MI.eraseFromParent();
1388 // Update LiveIntervals.
1390 MachineBasicBlock::iterator Begin(NewMIs[0]);
1391 MachineBasicBlock::iterator End(NewMIs[1]);
1392 LIS->repairIntervalsInRange(MBB, Begin, End, OrigRegs);
1397 // Transforming didn't eliminate the tie and didn't lead to an
1398 // improvement. Clean up the unfolded instructions and keep the
1400 DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
1401 NewMIs[0]->eraseFromParent();
1402 NewMIs[1]->eraseFromParent();
1411 // Collect tied operands of MI that need to be handled.
1412 // Rewrite trivial cases immediately.
1413 // Return true if any tied operands where found, including the trivial ones.
1414 bool TwoAddressInstructionPass::
1415 collectTiedOperands(MachineInstr *MI, TiedOperandMap &TiedOperands) {
1416 const MCInstrDesc &MCID = MI->getDesc();
1417 bool AnyOps = false;
1418 unsigned NumOps = MI->getNumOperands();
1420 for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) {
1421 unsigned DstIdx = 0;
1422 if (!MI->isRegTiedToDefOperand(SrcIdx, &DstIdx))
1425 MachineOperand &SrcMO = MI->getOperand(SrcIdx);
1426 MachineOperand &DstMO = MI->getOperand(DstIdx);
1427 unsigned SrcReg = SrcMO.getReg();
1428 unsigned DstReg = DstMO.getReg();
1429 // Tied constraint already satisfied?
1430 if (SrcReg == DstReg)
1433 assert(SrcReg && SrcMO.isUse() && "two address instruction invalid");
1435 // Deal with <undef> uses immediately - simply rewrite the src operand.
1436 if (SrcMO.isUndef() && !DstMO.getSubReg()) {
1437 // Constrain the DstReg register class if required.
1438 if (TargetRegisterInfo::isVirtualRegister(DstReg))
1439 if (const TargetRegisterClass *RC = TII->getRegClass(MCID, SrcIdx,
1441 MRI->constrainRegClass(DstReg, RC);
1442 SrcMO.setReg(DstReg);
1444 DEBUG(dbgs() << "\t\trewrite undef:\t" << *MI);
1447 TiedOperands[SrcReg].push_back(std::make_pair(SrcIdx, DstIdx));
1452 // Process a list of tied MI operands that all use the same source register.
1453 // The tied pairs are of the form (SrcIdx, DstIdx).
1455 TwoAddressInstructionPass::processTiedPairs(MachineInstr *MI,
1456 TiedPairList &TiedPairs,
1458 bool IsEarlyClobber = false;
1459 for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
1460 const MachineOperand &DstMO = MI->getOperand(TiedPairs[tpi].second);
1461 IsEarlyClobber |= DstMO.isEarlyClobber();
1464 bool RemovedKillFlag = false;
1465 bool AllUsesCopied = true;
1466 unsigned LastCopiedReg = 0;
1467 SlotIndex LastCopyIdx;
1469 unsigned SubRegB = 0;
1470 for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
1471 unsigned SrcIdx = TiedPairs[tpi].first;
1472 unsigned DstIdx = TiedPairs[tpi].second;
1474 const MachineOperand &DstMO = MI->getOperand(DstIdx);
1475 unsigned RegA = DstMO.getReg();
1477 // Grab RegB from the instruction because it may have changed if the
1478 // instruction was commuted.
1479 RegB = MI->getOperand(SrcIdx).getReg();
1480 SubRegB = MI->getOperand(SrcIdx).getSubReg();
1483 // The register is tied to multiple destinations (or else we would
1484 // not have continued this far), but this use of the register
1485 // already matches the tied destination. Leave it.
1486 AllUsesCopied = false;
1489 LastCopiedReg = RegA;
1491 assert(TargetRegisterInfo::isVirtualRegister(RegB) &&
1492 "cannot make instruction into two-address form");
1495 // First, verify that we don't have a use of "a" in the instruction
1496 // (a = b + a for example) because our transformation will not
1497 // work. This should never occur because we are in SSA form.
1498 for (unsigned i = 0; i != MI->getNumOperands(); ++i)
1499 assert(i == DstIdx ||
1500 !MI->getOperand(i).isReg() ||
1501 MI->getOperand(i).getReg() != RegA);
1505 MachineInstrBuilder MIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1506 TII->get(TargetOpcode::COPY), RegA);
1507 // If this operand is folding a truncation, the truncation now moves to the
1508 // copy so that the register classes remain valid for the operands.
1509 MIB.addReg(RegB, 0, SubRegB);
1510 const TargetRegisterClass *RC = MRI->getRegClass(RegB);
1512 if (TargetRegisterInfo::isVirtualRegister(RegA)) {
1513 assert(TRI->getMatchingSuperRegClass(RC, MRI->getRegClass(RegA),
1515 "tied subregister must be a truncation");
1516 // The superreg class will not be used to constrain the subreg class.
1520 assert(TRI->getMatchingSuperReg(RegA, SubRegB, MRI->getRegClass(RegB))
1521 && "tied subregister must be a truncation");
1525 // Update DistanceMap.
1526 MachineBasicBlock::iterator PrevMI = MI;
1528 DistanceMap.insert(std::make_pair(PrevMI, Dist));
1529 DistanceMap[MI] = ++Dist;
1532 LastCopyIdx = LIS->InsertMachineInstrInMaps(PrevMI).getRegSlot();
1534 if (TargetRegisterInfo::isVirtualRegister(RegA)) {
1535 LiveInterval &LI = LIS->getInterval(RegA);
1536 VNInfo *VNI = LI.getNextValue(LastCopyIdx, LIS->getVNInfoAllocator());
1538 LIS->getInstructionIndex(MI).getRegSlot(IsEarlyClobber);
1539 LI.addSegment(LiveInterval::Segment(LastCopyIdx, endIdx, VNI));
1543 DEBUG(dbgs() << "\t\tprepend:\t" << *MIB);
1545 MachineOperand &MO = MI->getOperand(SrcIdx);
1546 assert(MO.isReg() && MO.getReg() == RegB && MO.isUse() &&
1547 "inconsistent operand info for 2-reg pass");
1549 MO.setIsKill(false);
1550 RemovedKillFlag = true;
1553 // Make sure regA is a legal regclass for the SrcIdx operand.
1554 if (TargetRegisterInfo::isVirtualRegister(RegA) &&
1555 TargetRegisterInfo::isVirtualRegister(RegB))
1556 MRI->constrainRegClass(RegA, RC);
1558 // The getMatchingSuper asserts guarantee that the register class projected
1559 // by SubRegB is compatible with RegA with no subregister. So regardless of
1560 // whether the dest oper writes a subreg, the source oper should not.
1563 // Propagate SrcRegMap.
1564 SrcRegMap[RegA] = RegB;
1567 if (AllUsesCopied) {
1568 if (!IsEarlyClobber) {
1569 // Replace other (un-tied) uses of regB with LastCopiedReg.
1570 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1571 MachineOperand &MO = MI->getOperand(i);
1572 if (MO.isReg() && MO.getReg() == RegB && MO.getSubReg() == SubRegB &&
1575 MO.setIsKill(false);
1576 RemovedKillFlag = true;
1578 MO.setReg(LastCopiedReg);
1584 // Update live variables for regB.
1585 if (RemovedKillFlag && LV && LV->getVarInfo(RegB).removeKill(MI)) {
1586 MachineBasicBlock::iterator PrevMI = MI;
1588 LV->addVirtualRegisterKilled(RegB, PrevMI);
1591 // Update LiveIntervals.
1593 LiveInterval &LI = LIS->getInterval(RegB);
1594 SlotIndex MIIdx = LIS->getInstructionIndex(MI);
1595 LiveInterval::const_iterator I = LI.find(MIIdx);
1596 assert(I != LI.end() && "RegB must be live-in to use.");
1598 SlotIndex UseIdx = MIIdx.getRegSlot(IsEarlyClobber);
1599 if (I->end == UseIdx)
1600 LI.removeSegment(LastCopyIdx, UseIdx);
1603 } else if (RemovedKillFlag) {
1604 // Some tied uses of regB matched their destination registers, so
1605 // regB is still used in this instruction, but a kill flag was
1606 // removed from a different tied use of regB, so now we need to add
1607 // a kill flag to one of the remaining uses of regB.
1608 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1609 MachineOperand &MO = MI->getOperand(i);
1610 if (MO.isReg() && MO.getReg() == RegB && MO.isUse()) {
1618 /// runOnMachineFunction - Reduce two-address instructions to two operands.
1620 bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &Func) {
1622 const TargetMachine &TM = MF->getTarget();
1623 MRI = &MF->getRegInfo();
1624 TII = MF->getSubtarget().getInstrInfo();
1625 TRI = MF->getSubtarget().getRegisterInfo();
1626 InstrItins = MF->getSubtarget().getInstrItineraryData();
1627 LV = getAnalysisIfAvailable<LiveVariables>();
1628 LIS = getAnalysisIfAvailable<LiveIntervals>();
1629 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1630 OptLevel = TM.getOptLevel();
1632 bool MadeChange = false;
1634 DEBUG(dbgs() << "********** REWRITING TWO-ADDR INSTRS **********\n");
1635 DEBUG(dbgs() << "********** Function: "
1636 << MF->getName() << '\n');
1638 // This pass takes the function out of SSA form.
1641 TiedOperandMap TiedOperands;
1642 for (MachineFunction::iterator MBBI = MF->begin(), MBBE = MF->end();
1643 MBBI != MBBE; ++MBBI) {
1646 DistanceMap.clear();
1650 for (MachineBasicBlock::iterator mi = MBB->begin(), me = MBB->end();
1652 MachineBasicBlock::iterator nmi = std::next(mi);
1653 if (mi->isDebugValue()) {
1658 // Expand REG_SEQUENCE instructions. This will position mi at the first
1659 // expanded instruction.
1660 if (mi->isRegSequence())
1661 eliminateRegSequence(mi);
1663 DistanceMap.insert(std::make_pair(mi, ++Dist));
1667 // First scan through all the tied register uses in this instruction
1668 // and record a list of pairs of tied operands for each register.
1669 if (!collectTiedOperands(mi, TiedOperands)) {
1674 ++NumTwoAddressInstrs;
1676 DEBUG(dbgs() << '\t' << *mi);
1678 // If the instruction has a single pair of tied operands, try some
1679 // transformations that may either eliminate the tied operands or
1680 // improve the opportunities for coalescing away the register copy.
1681 if (TiedOperands.size() == 1) {
1682 SmallVectorImpl<std::pair<unsigned, unsigned> > &TiedPairs
1683 = TiedOperands.begin()->second;
1684 if (TiedPairs.size() == 1) {
1685 unsigned SrcIdx = TiedPairs[0].first;
1686 unsigned DstIdx = TiedPairs[0].second;
1687 unsigned SrcReg = mi->getOperand(SrcIdx).getReg();
1688 unsigned DstReg = mi->getOperand(DstIdx).getReg();
1689 if (SrcReg != DstReg &&
1690 tryInstructionTransform(mi, nmi, SrcIdx, DstIdx, Dist, false)) {
1691 // The tied operands have been eliminated or shifted further down
1692 // the block to ease elimination. Continue processing with 'nmi'.
1693 TiedOperands.clear();
1700 // Now iterate over the information collected above.
1701 for (TiedOperandMap::iterator OI = TiedOperands.begin(),
1702 OE = TiedOperands.end(); OI != OE; ++OI) {
1703 processTiedPairs(mi, OI->second, Dist);
1704 DEBUG(dbgs() << "\t\trewrite to:\t" << *mi);
1707 // Rewrite INSERT_SUBREG as COPY now that we no longer need SSA form.
1708 if (mi->isInsertSubreg()) {
1709 // From %reg = INSERT_SUBREG %reg, %subreg, subidx
1710 // To %reg:subidx = COPY %subreg
1711 unsigned SubIdx = mi->getOperand(3).getImm();
1712 mi->RemoveOperand(3);
1713 assert(mi->getOperand(0).getSubReg() == 0 && "Unexpected subreg idx");
1714 mi->getOperand(0).setSubReg(SubIdx);
1715 mi->getOperand(0).setIsUndef(mi->getOperand(1).isUndef());
1716 mi->RemoveOperand(1);
1717 mi->setDesc(TII->get(TargetOpcode::COPY));
1718 DEBUG(dbgs() << "\t\tconvert to:\t" << *mi);
1721 // Clear TiedOperands here instead of at the top of the loop
1722 // since most instructions do not have tied operands.
1723 TiedOperands.clear();
1729 MF->verify(this, "After two-address instruction pass");
1734 /// Eliminate a REG_SEQUENCE instruction as part of the de-ssa process.
1736 /// The instruction is turned into a sequence of sub-register copies:
1738 /// %dst = REG_SEQUENCE %v1, ssub0, %v2, ssub1
1742 /// %dst:ssub0<def,undef> = COPY %v1
1743 /// %dst:ssub1<def> = COPY %v2
1745 void TwoAddressInstructionPass::
1746 eliminateRegSequence(MachineBasicBlock::iterator &MBBI) {
1747 MachineInstr *MI = MBBI;
1748 unsigned DstReg = MI->getOperand(0).getReg();
1749 if (MI->getOperand(0).getSubReg() ||
1750 TargetRegisterInfo::isPhysicalRegister(DstReg) ||
1751 !(MI->getNumOperands() & 1)) {
1752 DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << *MI);
1753 llvm_unreachable(nullptr);
1756 SmallVector<unsigned, 4> OrigRegs;
1758 OrigRegs.push_back(MI->getOperand(0).getReg());
1759 for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2)
1760 OrigRegs.push_back(MI->getOperand(i).getReg());
1763 bool DefEmitted = false;
1764 for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) {
1765 MachineOperand &UseMO = MI->getOperand(i);
1766 unsigned SrcReg = UseMO.getReg();
1767 unsigned SubIdx = MI->getOperand(i+1).getImm();
1768 // Nothing needs to be inserted for <undef> operands.
1769 if (UseMO.isUndef())
1772 // Defer any kill flag to the last operand using SrcReg. Otherwise, we
1773 // might insert a COPY that uses SrcReg after is was killed.
1774 bool isKill = UseMO.isKill();
1776 for (unsigned j = i + 2; j < e; j += 2)
1777 if (MI->getOperand(j).getReg() == SrcReg) {
1778 MI->getOperand(j).setIsKill();
1779 UseMO.setIsKill(false);
1784 // Insert the sub-register copy.
1785 MachineInstr *CopyMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1786 TII->get(TargetOpcode::COPY))
1787 .addReg(DstReg, RegState::Define, SubIdx)
1790 // The first def needs an <undef> flag because there is no live register
1793 CopyMI->getOperand(0).setIsUndef(true);
1794 // Return an iterator pointing to the first inserted instr.
1799 // Update LiveVariables' kill info.
1800 if (LV && isKill && !TargetRegisterInfo::isPhysicalRegister(SrcReg))
1801 LV->replaceKillInstruction(SrcReg, MI, CopyMI);
1803 DEBUG(dbgs() << "Inserted: " << *CopyMI);
1806 MachineBasicBlock::iterator EndMBBI =
1807 std::next(MachineBasicBlock::iterator(MI));
1810 DEBUG(dbgs() << "Turned: " << *MI << " into an IMPLICIT_DEF");
1811 MI->setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
1812 for (int j = MI->getNumOperands() - 1, ee = 0; j > ee; --j)
1813 MI->RemoveOperand(j);
1815 DEBUG(dbgs() << "Eliminated: " << *MI);
1816 MI->eraseFromParent();
1819 // Udpate LiveIntervals.
1821 LIS->repairIntervalsInRange(MBB, MBBI, EndMBBI, OrigRegs);