Recommit r373598 "[yaml2obj/obj2yaml] - Add support for SHT_LLVM_ADDRSIG sections."
[llvm-complete.git] / lib / CodeGen / TwoAddressInstructionPass.cpp
blobea971809d4e4dc96188c5718c7a22ae554483a42
1 //===- TwoAddressInstructionPass.cpp - Two-Address instruction pass -------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the TwoAddress instruction pass which is used
10 // by most register allocators. Two-Address instructions are rewritten
11 // from:
13 // A = B op C
15 // to:
17 // A = B
18 // A op= C
20 // Note that if a register allocator chooses to use this pass, that it
21 // has to be capable of handling the non-SSA nature of these rewritten
22 // virtual registers.
24 // It is also worth noting that the duplicate operand of the two
25 // address instruction is removed.
27 //===----------------------------------------------------------------------===//
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/iterator_range.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/CodeGen/LiveInterval.h"
37 #include "llvm/CodeGen/LiveIntervals.h"
38 #include "llvm/CodeGen/LiveVariables.h"
39 #include "llvm/CodeGen/MachineBasicBlock.h"
40 #include "llvm/CodeGen/MachineFunction.h"
41 #include "llvm/CodeGen/MachineFunctionPass.h"
42 #include "llvm/CodeGen/MachineInstr.h"
43 #include "llvm/CodeGen/MachineInstrBuilder.h"
44 #include "llvm/CodeGen/MachineOperand.h"
45 #include "llvm/CodeGen/MachineRegisterInfo.h"
46 #include "llvm/CodeGen/Passes.h"
47 #include "llvm/CodeGen/SlotIndexes.h"
48 #include "llvm/CodeGen/TargetInstrInfo.h"
49 #include "llvm/CodeGen/TargetOpcodes.h"
50 #include "llvm/CodeGen/TargetRegisterInfo.h"
51 #include "llvm/CodeGen/TargetSubtargetInfo.h"
52 #include "llvm/MC/MCInstrDesc.h"
53 #include "llvm/MC/MCInstrItineraries.h"
54 #include "llvm/Pass.h"
55 #include "llvm/Support/CodeGen.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/ErrorHandling.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include "llvm/Target/TargetMachine.h"
61 #include <cassert>
62 #include <iterator>
63 #include <utility>
65 using namespace llvm;
67 #define DEBUG_TYPE "twoaddressinstruction"
69 STATISTIC(NumTwoAddressInstrs, "Number of two-address instructions");
70 STATISTIC(NumCommuted , "Number of instructions commuted to coalesce");
71 STATISTIC(NumAggrCommuted , "Number of instructions aggressively commuted");
72 STATISTIC(NumConvertedTo3Addr, "Number of instructions promoted to 3-address");
73 STATISTIC(Num3AddrSunk, "Number of 3-address instructions sunk");
74 STATISTIC(NumReSchedUps, "Number of instructions re-scheduled up");
75 STATISTIC(NumReSchedDowns, "Number of instructions re-scheduled down");
77 // Temporary flag to disable rescheduling.
78 static cl::opt<bool>
79 EnableRescheduling("twoaddr-reschedule",
80 cl::desc("Coalesce copies by rescheduling (default=true)"),
81 cl::init(true), cl::Hidden);
83 // Limit the number of dataflow edges to traverse when evaluating the benefit
84 // of commuting operands.
85 static cl::opt<unsigned> MaxDataFlowEdge(
86 "dataflow-edge-limit", cl::Hidden, cl::init(3),
87 cl::desc("Maximum number of dataflow edges to traverse when evaluating "
88 "the benefit of commuting operands"));
90 namespace {
92 class TwoAddressInstructionPass : public MachineFunctionPass {
93 MachineFunction *MF;
94 const TargetInstrInfo *TII;
95 const TargetRegisterInfo *TRI;
96 const InstrItineraryData *InstrItins;
97 MachineRegisterInfo *MRI;
98 LiveVariables *LV;
99 LiveIntervals *LIS;
100 AliasAnalysis *AA;
101 CodeGenOpt::Level OptLevel;
103 // The current basic block being processed.
104 MachineBasicBlock *MBB;
106 // Keep track the distance of a MI from the start of the current basic block.
107 DenseMap<MachineInstr*, unsigned> DistanceMap;
109 // Set of already processed instructions in the current block.
110 SmallPtrSet<MachineInstr*, 8> Processed;
112 // Set of instructions converted to three-address by target and then sunk
113 // down current basic block.
114 SmallPtrSet<MachineInstr*, 8> SunkInstrs;
116 // A map from virtual registers to physical registers which are likely targets
117 // to be coalesced to due to copies from physical registers to virtual
118 // registers. e.g. v1024 = move r0.
119 DenseMap<unsigned, unsigned> SrcRegMap;
121 // A map from virtual registers to physical registers which are likely targets
122 // to be coalesced to due to copies to physical registers from virtual
123 // registers. e.g. r1 = move v1024.
124 DenseMap<unsigned, unsigned> DstRegMap;
126 bool sink3AddrInstruction(MachineInstr *MI, unsigned Reg,
127 MachineBasicBlock::iterator OldPos);
129 bool isRevCopyChain(unsigned FromReg, unsigned ToReg, int Maxlen);
131 bool noUseAfterLastDef(unsigned Reg, unsigned Dist, unsigned &LastDef);
133 bool isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
134 MachineInstr *MI, unsigned Dist);
136 bool commuteInstruction(MachineInstr *MI, unsigned DstIdx,
137 unsigned RegBIdx, unsigned RegCIdx, unsigned Dist);
139 bool isProfitableToConv3Addr(unsigned RegA, unsigned RegB);
141 bool convertInstTo3Addr(MachineBasicBlock::iterator &mi,
142 MachineBasicBlock::iterator &nmi,
143 unsigned RegA, unsigned RegB, unsigned Dist);
145 bool isDefTooClose(unsigned Reg, unsigned Dist, MachineInstr *MI);
147 bool rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
148 MachineBasicBlock::iterator &nmi,
149 unsigned Reg);
150 bool rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
151 MachineBasicBlock::iterator &nmi,
152 unsigned Reg);
154 bool tryInstructionTransform(MachineBasicBlock::iterator &mi,
155 MachineBasicBlock::iterator &nmi,
156 unsigned SrcIdx, unsigned DstIdx,
157 unsigned Dist, bool shouldOnlyCommute);
159 bool tryInstructionCommute(MachineInstr *MI,
160 unsigned DstOpIdx,
161 unsigned BaseOpIdx,
162 bool BaseOpKilled,
163 unsigned Dist);
164 void scanUses(unsigned DstReg);
166 void processCopy(MachineInstr *MI);
168 using TiedPairList = SmallVector<std::pair<unsigned, unsigned>, 4>;
169 using TiedOperandMap = SmallDenseMap<unsigned, TiedPairList>;
171 bool collectTiedOperands(MachineInstr *MI, TiedOperandMap&);
172 void processTiedPairs(MachineInstr *MI, TiedPairList&, unsigned &Dist);
173 void eliminateRegSequence(MachineBasicBlock::iterator&);
175 public:
176 static char ID; // Pass identification, replacement for typeid
178 TwoAddressInstructionPass() : MachineFunctionPass(ID) {
179 initializeTwoAddressInstructionPassPass(*PassRegistry::getPassRegistry());
182 void getAnalysisUsage(AnalysisUsage &AU) const override {
183 AU.setPreservesCFG();
184 AU.addUsedIfAvailable<AAResultsWrapperPass>();
185 AU.addUsedIfAvailable<LiveVariables>();
186 AU.addPreserved<LiveVariables>();
187 AU.addPreserved<SlotIndexes>();
188 AU.addPreserved<LiveIntervals>();
189 AU.addPreservedID(MachineLoopInfoID);
190 AU.addPreservedID(MachineDominatorsID);
191 MachineFunctionPass::getAnalysisUsage(AU);
194 /// Pass entry point.
195 bool runOnMachineFunction(MachineFunction&) override;
198 } // end anonymous namespace
200 char TwoAddressInstructionPass::ID = 0;
202 char &llvm::TwoAddressInstructionPassID = TwoAddressInstructionPass::ID;
204 INITIALIZE_PASS_BEGIN(TwoAddressInstructionPass, DEBUG_TYPE,
205 "Two-Address instruction pass", false, false)
206 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
207 INITIALIZE_PASS_END(TwoAddressInstructionPass, DEBUG_TYPE,
208 "Two-Address instruction pass", false, false)
210 static bool isPlainlyKilled(MachineInstr *MI, unsigned Reg, LiveIntervals *LIS);
212 /// A two-address instruction has been converted to a three-address instruction
213 /// to avoid clobbering a register. Try to sink it past the instruction that
214 /// would kill the above mentioned register to reduce register pressure.
215 bool TwoAddressInstructionPass::
216 sink3AddrInstruction(MachineInstr *MI, unsigned SavedReg,
217 MachineBasicBlock::iterator OldPos) {
218 // FIXME: Shouldn't we be trying to do this before we three-addressify the
219 // instruction? After this transformation is done, we no longer need
220 // the instruction to be in three-address form.
222 // Check if it's safe to move this instruction.
223 bool SeenStore = true; // Be conservative.
224 if (!MI->isSafeToMove(AA, SeenStore))
225 return false;
227 unsigned DefReg = 0;
228 SmallSet<unsigned, 4> UseRegs;
230 for (const MachineOperand &MO : MI->operands()) {
231 if (!MO.isReg())
232 continue;
233 Register MOReg = MO.getReg();
234 if (!MOReg)
235 continue;
236 if (MO.isUse() && MOReg != SavedReg)
237 UseRegs.insert(MO.getReg());
238 if (!MO.isDef())
239 continue;
240 if (MO.isImplicit())
241 // Don't try to move it if it implicitly defines a register.
242 return false;
243 if (DefReg)
244 // For now, don't move any instructions that define multiple registers.
245 return false;
246 DefReg = MO.getReg();
249 // Find the instruction that kills SavedReg.
250 MachineInstr *KillMI = nullptr;
251 if (LIS) {
252 LiveInterval &LI = LIS->getInterval(SavedReg);
253 assert(LI.end() != LI.begin() &&
254 "Reg should not have empty live interval.");
256 SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
257 LiveInterval::const_iterator I = LI.find(MBBEndIdx);
258 if (I != LI.end() && I->start < MBBEndIdx)
259 return false;
261 --I;
262 KillMI = LIS->getInstructionFromIndex(I->end);
264 if (!KillMI) {
265 for (MachineOperand &UseMO : MRI->use_nodbg_operands(SavedReg)) {
266 if (!UseMO.isKill())
267 continue;
268 KillMI = UseMO.getParent();
269 break;
273 // If we find the instruction that kills SavedReg, and it is in an
274 // appropriate location, we can try to sink the current instruction
275 // past it.
276 if (!KillMI || KillMI->getParent() != MBB || KillMI == MI ||
277 MachineBasicBlock::iterator(KillMI) == OldPos || KillMI->isTerminator())
278 return false;
280 // If any of the definitions are used by another instruction between the
281 // position and the kill use, then it's not safe to sink it.
283 // FIXME: This can be sped up if there is an easy way to query whether an
284 // instruction is before or after another instruction. Then we can use
285 // MachineRegisterInfo def / use instead.
286 MachineOperand *KillMO = nullptr;
287 MachineBasicBlock::iterator KillPos = KillMI;
288 ++KillPos;
290 unsigned NumVisited = 0;
291 for (MachineInstr &OtherMI : make_range(std::next(OldPos), KillPos)) {
292 // Debug instructions cannot be counted against the limit.
293 if (OtherMI.isDebugInstr())
294 continue;
295 if (NumVisited > 30) // FIXME: Arbitrary limit to reduce compile time cost.
296 return false;
297 ++NumVisited;
298 for (unsigned i = 0, e = OtherMI.getNumOperands(); i != e; ++i) {
299 MachineOperand &MO = OtherMI.getOperand(i);
300 if (!MO.isReg())
301 continue;
302 Register MOReg = MO.getReg();
303 if (!MOReg)
304 continue;
305 if (DefReg == MOReg)
306 return false;
308 if (MO.isKill() || (LIS && isPlainlyKilled(&OtherMI, MOReg, LIS))) {
309 if (&OtherMI == KillMI && MOReg == SavedReg)
310 // Save the operand that kills the register. We want to unset the kill
311 // marker if we can sink MI past it.
312 KillMO = &MO;
313 else if (UseRegs.count(MOReg))
314 // One of the uses is killed before the destination.
315 return false;
319 assert(KillMO && "Didn't find kill");
321 if (!LIS) {
322 // Update kill and LV information.
323 KillMO->setIsKill(false);
324 KillMO = MI->findRegisterUseOperand(SavedReg, false, TRI);
325 KillMO->setIsKill(true);
327 if (LV)
328 LV->replaceKillInstruction(SavedReg, *KillMI, *MI);
331 // Move instruction to its destination.
332 MBB->remove(MI);
333 MBB->insert(KillPos, MI);
335 if (LIS)
336 LIS->handleMove(*MI);
338 ++Num3AddrSunk;
339 return true;
342 /// Return the MachineInstr* if it is the single def of the Reg in current BB.
343 static MachineInstr *getSingleDef(unsigned Reg, MachineBasicBlock *BB,
344 const MachineRegisterInfo *MRI) {
345 MachineInstr *Ret = nullptr;
346 for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
347 if (DefMI.getParent() != BB || DefMI.isDebugValue())
348 continue;
349 if (!Ret)
350 Ret = &DefMI;
351 else if (Ret != &DefMI)
352 return nullptr;
354 return Ret;
357 /// Check if there is a reversed copy chain from FromReg to ToReg:
358 /// %Tmp1 = copy %Tmp2;
359 /// %FromReg = copy %Tmp1;
360 /// %ToReg = add %FromReg ...
361 /// %Tmp2 = copy %ToReg;
362 /// MaxLen specifies the maximum length of the copy chain the func
363 /// can walk through.
364 bool TwoAddressInstructionPass::isRevCopyChain(unsigned FromReg, unsigned ToReg,
365 int Maxlen) {
366 unsigned TmpReg = FromReg;
367 for (int i = 0; i < Maxlen; i++) {
368 MachineInstr *Def = getSingleDef(TmpReg, MBB, MRI);
369 if (!Def || !Def->isCopy())
370 return false;
372 TmpReg = Def->getOperand(1).getReg();
374 if (TmpReg == ToReg)
375 return true;
377 return false;
380 /// Return true if there are no intervening uses between the last instruction
381 /// in the MBB that defines the specified register and the two-address
382 /// instruction which is being processed. It also returns the last def location
383 /// by reference.
384 bool TwoAddressInstructionPass::noUseAfterLastDef(unsigned Reg, unsigned Dist,
385 unsigned &LastDef) {
386 LastDef = 0;
387 unsigned LastUse = Dist;
388 for (MachineOperand &MO : MRI->reg_operands(Reg)) {
389 MachineInstr *MI = MO.getParent();
390 if (MI->getParent() != MBB || MI->isDebugValue())
391 continue;
392 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
393 if (DI == DistanceMap.end())
394 continue;
395 if (MO.isUse() && DI->second < LastUse)
396 LastUse = DI->second;
397 if (MO.isDef() && DI->second > LastDef)
398 LastDef = DI->second;
401 return !(LastUse > LastDef && LastUse < Dist);
404 /// Return true if the specified MI is a copy instruction or an extract_subreg
405 /// instruction. It also returns the source and destination registers and
406 /// whether they are physical registers by reference.
407 static bool isCopyToReg(MachineInstr &MI, const TargetInstrInfo *TII,
408 unsigned &SrcReg, unsigned &DstReg,
409 bool &IsSrcPhys, bool &IsDstPhys) {
410 SrcReg = 0;
411 DstReg = 0;
412 if (MI.isCopy()) {
413 DstReg = MI.getOperand(0).getReg();
414 SrcReg = MI.getOperand(1).getReg();
415 } else if (MI.isInsertSubreg() || MI.isSubregToReg()) {
416 DstReg = MI.getOperand(0).getReg();
417 SrcReg = MI.getOperand(2).getReg();
418 } else
419 return false;
421 IsSrcPhys = Register::isPhysicalRegister(SrcReg);
422 IsDstPhys = Register::isPhysicalRegister(DstReg);
423 return true;
426 /// Test if the given register value, which is used by the
427 /// given instruction, is killed by the given instruction.
428 static bool isPlainlyKilled(MachineInstr *MI, unsigned Reg,
429 LiveIntervals *LIS) {
430 if (LIS && Register::isVirtualRegister(Reg) && !LIS->isNotInMIMap(*MI)) {
431 // FIXME: Sometimes tryInstructionTransform() will add instructions and
432 // test whether they can be folded before keeping them. In this case it
433 // sets a kill before recursively calling tryInstructionTransform() again.
434 // If there is no interval available, we assume that this instruction is
435 // one of those. A kill flag is manually inserted on the operand so the
436 // check below will handle it.
437 LiveInterval &LI = LIS->getInterval(Reg);
438 // This is to match the kill flag version where undefs don't have kill
439 // flags.
440 if (!LI.hasAtLeastOneValue())
441 return false;
443 SlotIndex useIdx = LIS->getInstructionIndex(*MI);
444 LiveInterval::const_iterator I = LI.find(useIdx);
445 assert(I != LI.end() && "Reg must be live-in to use.");
446 return !I->end.isBlock() && SlotIndex::isSameInstr(I->end, useIdx);
449 return MI->killsRegister(Reg);
452 /// Test if the given register value, which is used by the given
453 /// instruction, is killed by the given instruction. This looks through
454 /// coalescable copies to see if the original value is potentially not killed.
456 /// For example, in this code:
458 /// %reg1034 = copy %reg1024
459 /// %reg1035 = copy killed %reg1025
460 /// %reg1036 = add killed %reg1034, killed %reg1035
462 /// %reg1034 is not considered to be killed, since it is copied from a
463 /// register which is not killed. Treating it as not killed lets the
464 /// normal heuristics commute the (two-address) add, which lets
465 /// coalescing eliminate the extra copy.
467 /// If allowFalsePositives is true then likely kills are treated as kills even
468 /// if it can't be proven that they are kills.
469 static bool isKilled(MachineInstr &MI, unsigned Reg,
470 const MachineRegisterInfo *MRI,
471 const TargetInstrInfo *TII,
472 LiveIntervals *LIS,
473 bool allowFalsePositives) {
474 MachineInstr *DefMI = &MI;
475 while (true) {
476 // All uses of physical registers are likely to be kills.
477 if (Register::isPhysicalRegister(Reg) &&
478 (allowFalsePositives || MRI->hasOneUse(Reg)))
479 return true;
480 if (!isPlainlyKilled(DefMI, Reg, LIS))
481 return false;
482 if (Register::isPhysicalRegister(Reg))
483 return true;
484 MachineRegisterInfo::def_iterator Begin = MRI->def_begin(Reg);
485 // If there are multiple defs, we can't do a simple analysis, so just
486 // go with what the kill flag says.
487 if (std::next(Begin) != MRI->def_end())
488 return true;
489 DefMI = Begin->getParent();
490 bool IsSrcPhys, IsDstPhys;
491 unsigned SrcReg, DstReg;
492 // If the def is something other than a copy, then it isn't going to
493 // be coalesced, so follow the kill flag.
494 if (!isCopyToReg(*DefMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
495 return true;
496 Reg = SrcReg;
500 /// Return true if the specified MI uses the specified register as a two-address
501 /// use. If so, return the destination register by reference.
502 static bool isTwoAddrUse(MachineInstr &MI, unsigned Reg, unsigned &DstReg) {
503 for (unsigned i = 0, NumOps = MI.getNumOperands(); i != NumOps; ++i) {
504 const MachineOperand &MO = MI.getOperand(i);
505 if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
506 continue;
507 unsigned ti;
508 if (MI.isRegTiedToDefOperand(i, &ti)) {
509 DstReg = MI.getOperand(ti).getReg();
510 return true;
513 return false;
516 /// Given a register, if has a single in-basic block use, return the use
517 /// instruction if it's a copy or a two-address use.
518 static
519 MachineInstr *findOnlyInterestingUse(unsigned Reg, MachineBasicBlock *MBB,
520 MachineRegisterInfo *MRI,
521 const TargetInstrInfo *TII,
522 bool &IsCopy,
523 unsigned &DstReg, bool &IsDstPhys) {
524 if (!MRI->hasOneNonDBGUse(Reg))
525 // None or more than one use.
526 return nullptr;
527 MachineInstr &UseMI = *MRI->use_instr_nodbg_begin(Reg);
528 if (UseMI.getParent() != MBB)
529 return nullptr;
530 unsigned SrcReg;
531 bool IsSrcPhys;
532 if (isCopyToReg(UseMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys)) {
533 IsCopy = true;
534 return &UseMI;
536 IsDstPhys = false;
537 if (isTwoAddrUse(UseMI, Reg, DstReg)) {
538 IsDstPhys = Register::isPhysicalRegister(DstReg);
539 return &UseMI;
541 return nullptr;
544 /// Return the physical register the specified virtual register might be mapped
545 /// to.
546 static unsigned
547 getMappedReg(unsigned Reg, DenseMap<unsigned, unsigned> &RegMap) {
548 while (Register::isVirtualRegister(Reg)) {
549 DenseMap<unsigned, unsigned>::iterator SI = RegMap.find(Reg);
550 if (SI == RegMap.end())
551 return 0;
552 Reg = SI->second;
554 if (Register::isPhysicalRegister(Reg))
555 return Reg;
556 return 0;
559 /// Return true if the two registers are equal or aliased.
560 static bool
561 regsAreCompatible(unsigned RegA, unsigned RegB, const TargetRegisterInfo *TRI) {
562 if (RegA == RegB)
563 return true;
564 if (!RegA || !RegB)
565 return false;
566 return TRI->regsOverlap(RegA, RegB);
569 // Returns true if Reg is equal or aliased to at least one register in Set.
570 static bool regOverlapsSet(const SmallVectorImpl<unsigned> &Set, unsigned Reg,
571 const TargetRegisterInfo *TRI) {
572 for (unsigned R : Set)
573 if (TRI->regsOverlap(R, Reg))
574 return true;
576 return false;
579 /// Return true if it's potentially profitable to commute the two-address
580 /// instruction that's being processed.
581 bool
582 TwoAddressInstructionPass::
583 isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
584 MachineInstr *MI, unsigned Dist) {
585 if (OptLevel == CodeGenOpt::None)
586 return false;
588 // Determine if it's profitable to commute this two address instruction. In
589 // general, we want no uses between this instruction and the definition of
590 // the two-address register.
591 // e.g.
592 // %reg1028 = EXTRACT_SUBREG killed %reg1027, 1
593 // %reg1029 = COPY %reg1028
594 // %reg1029 = SHR8ri %reg1029, 7, implicit dead %eflags
595 // insert => %reg1030 = COPY %reg1028
596 // %reg1030 = ADD8rr killed %reg1028, killed %reg1029, implicit dead %eflags
597 // In this case, it might not be possible to coalesce the second COPY
598 // instruction if the first one is coalesced. So it would be profitable to
599 // commute it:
600 // %reg1028 = EXTRACT_SUBREG killed %reg1027, 1
601 // %reg1029 = COPY %reg1028
602 // %reg1029 = SHR8ri %reg1029, 7, implicit dead %eflags
603 // insert => %reg1030 = COPY %reg1029
604 // %reg1030 = ADD8rr killed %reg1029, killed %reg1028, implicit dead %eflags
606 if (!isPlainlyKilled(MI, regC, LIS))
607 return false;
609 // Ok, we have something like:
610 // %reg1030 = ADD8rr killed %reg1028, killed %reg1029, implicit dead %eflags
611 // let's see if it's worth commuting it.
613 // Look for situations like this:
614 // %reg1024 = MOV r1
615 // %reg1025 = MOV r0
616 // %reg1026 = ADD %reg1024, %reg1025
617 // r0 = MOV %reg1026
618 // Commute the ADD to hopefully eliminate an otherwise unavoidable copy.
619 unsigned ToRegA = getMappedReg(regA, DstRegMap);
620 if (ToRegA) {
621 unsigned FromRegB = getMappedReg(regB, SrcRegMap);
622 unsigned FromRegC = getMappedReg(regC, SrcRegMap);
623 bool CompB = FromRegB && regsAreCompatible(FromRegB, ToRegA, TRI);
624 bool CompC = FromRegC && regsAreCompatible(FromRegC, ToRegA, TRI);
626 // Compute if any of the following are true:
627 // -RegB is not tied to a register and RegC is compatible with RegA.
628 // -RegB is tied to the wrong physical register, but RegC is.
629 // -RegB is tied to the wrong physical register, and RegC isn't tied.
630 if ((!FromRegB && CompC) || (FromRegB && !CompB && (!FromRegC || CompC)))
631 return true;
632 // Don't compute if any of the following are true:
633 // -RegC is not tied to a register and RegB is compatible with RegA.
634 // -RegC is tied to the wrong physical register, but RegB is.
635 // -RegC is tied to the wrong physical register, and RegB isn't tied.
636 if ((!FromRegC && CompB) || (FromRegC && !CompC && (!FromRegB || CompB)))
637 return false;
640 // If there is a use of regC between its last def (could be livein) and this
641 // instruction, then bail.
642 unsigned LastDefC = 0;
643 if (!noUseAfterLastDef(regC, Dist, LastDefC))
644 return false;
646 // If there is a use of regB between its last def (could be livein) and this
647 // instruction, then go ahead and make this transformation.
648 unsigned LastDefB = 0;
649 if (!noUseAfterLastDef(regB, Dist, LastDefB))
650 return true;
652 // Look for situation like this:
653 // %reg101 = MOV %reg100
654 // %reg102 = ...
655 // %reg103 = ADD %reg102, %reg101
656 // ... = %reg103 ...
657 // %reg100 = MOV %reg103
658 // If there is a reversed copy chain from reg101 to reg103, commute the ADD
659 // to eliminate an otherwise unavoidable copy.
660 // FIXME:
661 // We can extend the logic further: If an pair of operands in an insn has
662 // been merged, the insn could be regarded as a virtual copy, and the virtual
663 // copy could also be used to construct a copy chain.
664 // To more generally minimize register copies, ideally the logic of two addr
665 // instruction pass should be integrated with register allocation pass where
666 // interference graph is available.
667 if (isRevCopyChain(regC, regA, MaxDataFlowEdge))
668 return true;
670 if (isRevCopyChain(regB, regA, MaxDataFlowEdge))
671 return false;
673 // Since there are no intervening uses for both registers, then commute
674 // if the def of regC is closer. Its live interval is shorter.
675 return LastDefB && LastDefC && LastDefC > LastDefB;
678 /// Commute a two-address instruction and update the basic block, distance map,
679 /// and live variables if needed. Return true if it is successful.
680 bool TwoAddressInstructionPass::commuteInstruction(MachineInstr *MI,
681 unsigned DstIdx,
682 unsigned RegBIdx,
683 unsigned RegCIdx,
684 unsigned Dist) {
685 Register RegC = MI->getOperand(RegCIdx).getReg();
686 LLVM_DEBUG(dbgs() << "2addr: COMMUTING : " << *MI);
687 MachineInstr *NewMI = TII->commuteInstruction(*MI, false, RegBIdx, RegCIdx);
689 if (NewMI == nullptr) {
690 LLVM_DEBUG(dbgs() << "2addr: COMMUTING FAILED!\n");
691 return false;
694 LLVM_DEBUG(dbgs() << "2addr: COMMUTED TO: " << *NewMI);
695 assert(NewMI == MI &&
696 "TargetInstrInfo::commuteInstruction() should not return a new "
697 "instruction unless it was requested.");
699 // Update source register map.
700 unsigned FromRegC = getMappedReg(RegC, SrcRegMap);
701 if (FromRegC) {
702 Register RegA = MI->getOperand(DstIdx).getReg();
703 SrcRegMap[RegA] = FromRegC;
706 return true;
709 /// Return true if it is profitable to convert the given 2-address instruction
710 /// to a 3-address one.
711 bool
712 TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA,unsigned RegB){
713 // Look for situations like this:
714 // %reg1024 = MOV r1
715 // %reg1025 = MOV r0
716 // %reg1026 = ADD %reg1024, %reg1025
717 // r2 = MOV %reg1026
718 // Turn ADD into a 3-address instruction to avoid a copy.
719 unsigned FromRegB = getMappedReg(RegB, SrcRegMap);
720 if (!FromRegB)
721 return false;
722 unsigned ToRegA = getMappedReg(RegA, DstRegMap);
723 return (ToRegA && !regsAreCompatible(FromRegB, ToRegA, TRI));
726 /// Convert the specified two-address instruction into a three address one.
727 /// Return true if this transformation was successful.
728 bool
729 TwoAddressInstructionPass::convertInstTo3Addr(MachineBasicBlock::iterator &mi,
730 MachineBasicBlock::iterator &nmi,
731 unsigned RegA, unsigned RegB,
732 unsigned Dist) {
733 // FIXME: Why does convertToThreeAddress() need an iterator reference?
734 MachineFunction::iterator MFI = MBB->getIterator();
735 MachineInstr *NewMI = TII->convertToThreeAddress(MFI, *mi, LV);
736 assert(MBB->getIterator() == MFI &&
737 "convertToThreeAddress changed iterator reference");
738 if (!NewMI)
739 return false;
741 LLVM_DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi);
742 LLVM_DEBUG(dbgs() << "2addr: TO 3-ADDR: " << *NewMI);
743 bool Sunk = false;
745 if (LIS)
746 LIS->ReplaceMachineInstrInMaps(*mi, *NewMI);
748 if (NewMI->findRegisterUseOperand(RegB, false, TRI))
749 // FIXME: Temporary workaround. If the new instruction doesn't
750 // uses RegB, convertToThreeAddress must have created more
751 // then one instruction.
752 Sunk = sink3AddrInstruction(NewMI, RegB, mi);
754 MBB->erase(mi); // Nuke the old inst.
756 if (!Sunk) {
757 DistanceMap.insert(std::make_pair(NewMI, Dist));
758 mi = NewMI;
759 nmi = std::next(mi);
761 else
762 SunkInstrs.insert(NewMI);
764 // Update source and destination register maps.
765 SrcRegMap.erase(RegA);
766 DstRegMap.erase(RegB);
767 return true;
770 /// Scan forward recursively for only uses, update maps if the use is a copy or
771 /// a two-address instruction.
772 void
773 TwoAddressInstructionPass::scanUses(unsigned DstReg) {
774 SmallVector<unsigned, 4> VirtRegPairs;
775 bool IsDstPhys;
776 bool IsCopy = false;
777 unsigned NewReg = 0;
778 unsigned Reg = DstReg;
779 while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, MRI, TII,IsCopy,
780 NewReg, IsDstPhys)) {
781 if (IsCopy && !Processed.insert(UseMI).second)
782 break;
784 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
785 if (DI != DistanceMap.end())
786 // Earlier in the same MBB.Reached via a back edge.
787 break;
789 if (IsDstPhys) {
790 VirtRegPairs.push_back(NewReg);
791 break;
793 bool isNew = SrcRegMap.insert(std::make_pair(NewReg, Reg)).second;
794 if (!isNew)
795 assert(SrcRegMap[NewReg] == Reg && "Can't map to two src registers!");
796 VirtRegPairs.push_back(NewReg);
797 Reg = NewReg;
800 if (!VirtRegPairs.empty()) {
801 unsigned ToReg = VirtRegPairs.back();
802 VirtRegPairs.pop_back();
803 while (!VirtRegPairs.empty()) {
804 unsigned FromReg = VirtRegPairs.back();
805 VirtRegPairs.pop_back();
806 bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
807 if (!isNew)
808 assert(DstRegMap[FromReg] == ToReg &&"Can't map to two dst registers!");
809 ToReg = FromReg;
811 bool isNew = DstRegMap.insert(std::make_pair(DstReg, ToReg)).second;
812 if (!isNew)
813 assert(DstRegMap[DstReg] == ToReg && "Can't map to two dst registers!");
817 /// If the specified instruction is not yet processed, process it if it's a
818 /// copy. For a copy instruction, we find the physical registers the
819 /// source and destination registers might be mapped to. These are kept in
820 /// point-to maps used to determine future optimizations. e.g.
821 /// v1024 = mov r0
822 /// v1025 = mov r1
823 /// v1026 = add v1024, v1025
824 /// r1 = mov r1026
825 /// If 'add' is a two-address instruction, v1024, v1026 are both potentially
826 /// coalesced to r0 (from the input side). v1025 is mapped to r1. v1026 is
827 /// potentially joined with r1 on the output side. It's worthwhile to commute
828 /// 'add' to eliminate a copy.
829 void TwoAddressInstructionPass::processCopy(MachineInstr *MI) {
830 if (Processed.count(MI))
831 return;
833 bool IsSrcPhys, IsDstPhys;
834 unsigned SrcReg, DstReg;
835 if (!isCopyToReg(*MI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
836 return;
838 if (IsDstPhys && !IsSrcPhys)
839 DstRegMap.insert(std::make_pair(SrcReg, DstReg));
840 else if (!IsDstPhys && IsSrcPhys) {
841 bool isNew = SrcRegMap.insert(std::make_pair(DstReg, SrcReg)).second;
842 if (!isNew)
843 assert(SrcRegMap[DstReg] == SrcReg &&
844 "Can't map to two src physical registers!");
846 scanUses(DstReg);
849 Processed.insert(MI);
852 /// If there is one more local instruction that reads 'Reg' and it kills 'Reg,
853 /// consider moving the instruction below the kill instruction in order to
854 /// eliminate the need for the copy.
855 bool TwoAddressInstructionPass::
856 rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
857 MachineBasicBlock::iterator &nmi,
858 unsigned Reg) {
859 // Bail immediately if we don't have LV or LIS available. We use them to find
860 // kills efficiently.
861 if (!LV && !LIS)
862 return false;
864 MachineInstr *MI = &*mi;
865 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
866 if (DI == DistanceMap.end())
867 // Must be created from unfolded load. Don't waste time trying this.
868 return false;
870 MachineInstr *KillMI = nullptr;
871 if (LIS) {
872 LiveInterval &LI = LIS->getInterval(Reg);
873 assert(LI.end() != LI.begin() &&
874 "Reg should not have empty live interval.");
876 SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
877 LiveInterval::const_iterator I = LI.find(MBBEndIdx);
878 if (I != LI.end() && I->start < MBBEndIdx)
879 return false;
881 --I;
882 KillMI = LIS->getInstructionFromIndex(I->end);
883 } else {
884 KillMI = LV->getVarInfo(Reg).findKill(MBB);
886 if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
887 // Don't mess with copies, they may be coalesced later.
888 return false;
890 if (KillMI->hasUnmodeledSideEffects() || KillMI->isCall() ||
891 KillMI->isBranch() || KillMI->isTerminator())
892 // Don't move pass calls, etc.
893 return false;
895 unsigned DstReg;
896 if (isTwoAddrUse(*KillMI, Reg, DstReg))
897 return false;
899 bool SeenStore = true;
900 if (!MI->isSafeToMove(AA, SeenStore))
901 return false;
903 if (TII->getInstrLatency(InstrItins, *MI) > 1)
904 // FIXME: Needs more sophisticated heuristics.
905 return false;
907 SmallVector<unsigned, 2> Uses;
908 SmallVector<unsigned, 2> Kills;
909 SmallVector<unsigned, 2> Defs;
910 for (const MachineOperand &MO : MI->operands()) {
911 if (!MO.isReg())
912 continue;
913 Register MOReg = MO.getReg();
914 if (!MOReg)
915 continue;
916 if (MO.isDef())
917 Defs.push_back(MOReg);
918 else {
919 Uses.push_back(MOReg);
920 if (MOReg != Reg && (MO.isKill() ||
921 (LIS && isPlainlyKilled(MI, MOReg, LIS))))
922 Kills.push_back(MOReg);
926 // Move the copies connected to MI down as well.
927 MachineBasicBlock::iterator Begin = MI;
928 MachineBasicBlock::iterator AfterMI = std::next(Begin);
929 MachineBasicBlock::iterator End = AfterMI;
930 while (End != MBB->end()) {
931 End = skipDebugInstructionsForward(End, MBB->end());
932 if (End->isCopy() && regOverlapsSet(Defs, End->getOperand(1).getReg(), TRI))
933 Defs.push_back(End->getOperand(0).getReg());
934 else
935 break;
936 ++End;
939 // Check if the reschedule will not break dependencies.
940 unsigned NumVisited = 0;
941 MachineBasicBlock::iterator KillPos = KillMI;
942 ++KillPos;
943 for (MachineInstr &OtherMI : make_range(End, KillPos)) {
944 // Debug instructions cannot be counted against the limit.
945 if (OtherMI.isDebugInstr())
946 continue;
947 if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
948 return false;
949 ++NumVisited;
950 if (OtherMI.hasUnmodeledSideEffects() || OtherMI.isCall() ||
951 OtherMI.isBranch() || OtherMI.isTerminator())
952 // Don't move pass calls, etc.
953 return false;
954 for (const MachineOperand &MO : OtherMI.operands()) {
955 if (!MO.isReg())
956 continue;
957 Register MOReg = MO.getReg();
958 if (!MOReg)
959 continue;
960 if (MO.isDef()) {
961 if (regOverlapsSet(Uses, MOReg, TRI))
962 // Physical register use would be clobbered.
963 return false;
964 if (!MO.isDead() && regOverlapsSet(Defs, MOReg, TRI))
965 // May clobber a physical register def.
966 // FIXME: This may be too conservative. It's ok if the instruction
967 // is sunken completely below the use.
968 return false;
969 } else {
970 if (regOverlapsSet(Defs, MOReg, TRI))
971 return false;
972 bool isKill =
973 MO.isKill() || (LIS && isPlainlyKilled(&OtherMI, MOReg, LIS));
974 if (MOReg != Reg && ((isKill && regOverlapsSet(Uses, MOReg, TRI)) ||
975 regOverlapsSet(Kills, MOReg, TRI)))
976 // Don't want to extend other live ranges and update kills.
977 return false;
978 if (MOReg == Reg && !isKill)
979 // We can't schedule across a use of the register in question.
980 return false;
981 // Ensure that if this is register in question, its the kill we expect.
982 assert((MOReg != Reg || &OtherMI == KillMI) &&
983 "Found multiple kills of a register in a basic block");
988 // Move debug info as well.
989 while (Begin != MBB->begin() && std::prev(Begin)->isDebugInstr())
990 --Begin;
992 nmi = End;
993 MachineBasicBlock::iterator InsertPos = KillPos;
994 if (LIS) {
995 // We have to move the copies first so that the MBB is still well-formed
996 // when calling handleMove().
997 for (MachineBasicBlock::iterator MBBI = AfterMI; MBBI != End;) {
998 auto CopyMI = MBBI++;
999 MBB->splice(InsertPos, MBB, CopyMI);
1000 LIS->handleMove(*CopyMI);
1001 InsertPos = CopyMI;
1003 End = std::next(MachineBasicBlock::iterator(MI));
1006 // Copies following MI may have been moved as well.
1007 MBB->splice(InsertPos, MBB, Begin, End);
1008 DistanceMap.erase(DI);
1010 // Update live variables
1011 if (LIS) {
1012 LIS->handleMove(*MI);
1013 } else {
1014 LV->removeVirtualRegisterKilled(Reg, *KillMI);
1015 LV->addVirtualRegisterKilled(Reg, *MI);
1018 LLVM_DEBUG(dbgs() << "\trescheduled below kill: " << *KillMI);
1019 return true;
1022 /// Return true if the re-scheduling will put the given instruction too close
1023 /// to the defs of its register dependencies.
1024 bool TwoAddressInstructionPass::isDefTooClose(unsigned Reg, unsigned Dist,
1025 MachineInstr *MI) {
1026 for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
1027 if (DefMI.getParent() != MBB || DefMI.isCopy() || DefMI.isCopyLike())
1028 continue;
1029 if (&DefMI == MI)
1030 return true; // MI is defining something KillMI uses
1031 DenseMap<MachineInstr*, unsigned>::iterator DDI = DistanceMap.find(&DefMI);
1032 if (DDI == DistanceMap.end())
1033 return true; // Below MI
1034 unsigned DefDist = DDI->second;
1035 assert(Dist > DefDist && "Visited def already?");
1036 if (TII->getInstrLatency(InstrItins, DefMI) > (Dist - DefDist))
1037 return true;
1039 return false;
1042 /// If there is one more local instruction that reads 'Reg' and it kills 'Reg,
1043 /// consider moving the kill instruction above the current two-address
1044 /// instruction in order to eliminate the need for the copy.
1045 bool TwoAddressInstructionPass::
1046 rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
1047 MachineBasicBlock::iterator &nmi,
1048 unsigned Reg) {
1049 // Bail immediately if we don't have LV or LIS available. We use them to find
1050 // kills efficiently.
1051 if (!LV && !LIS)
1052 return false;
1054 MachineInstr *MI = &*mi;
1055 DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
1056 if (DI == DistanceMap.end())
1057 // Must be created from unfolded load. Don't waste time trying this.
1058 return false;
1060 MachineInstr *KillMI = nullptr;
1061 if (LIS) {
1062 LiveInterval &LI = LIS->getInterval(Reg);
1063 assert(LI.end() != LI.begin() &&
1064 "Reg should not have empty live interval.");
1066 SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
1067 LiveInterval::const_iterator I = LI.find(MBBEndIdx);
1068 if (I != LI.end() && I->start < MBBEndIdx)
1069 return false;
1071 --I;
1072 KillMI = LIS->getInstructionFromIndex(I->end);
1073 } else {
1074 KillMI = LV->getVarInfo(Reg).findKill(MBB);
1076 if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
1077 // Don't mess with copies, they may be coalesced later.
1078 return false;
1080 unsigned DstReg;
1081 if (isTwoAddrUse(*KillMI, Reg, DstReg))
1082 return false;
1084 bool SeenStore = true;
1085 if (!KillMI->isSafeToMove(AA, SeenStore))
1086 return false;
1088 SmallSet<unsigned, 2> Uses;
1089 SmallSet<unsigned, 2> Kills;
1090 SmallSet<unsigned, 2> Defs;
1091 SmallSet<unsigned, 2> LiveDefs;
1092 for (const MachineOperand &MO : KillMI->operands()) {
1093 if (!MO.isReg())
1094 continue;
1095 Register MOReg = MO.getReg();
1096 if (MO.isUse()) {
1097 if (!MOReg)
1098 continue;
1099 if (isDefTooClose(MOReg, DI->second, MI))
1100 return false;
1101 bool isKill = MO.isKill() || (LIS && isPlainlyKilled(KillMI, MOReg, LIS));
1102 if (MOReg == Reg && !isKill)
1103 return false;
1104 Uses.insert(MOReg);
1105 if (isKill && MOReg != Reg)
1106 Kills.insert(MOReg);
1107 } else if (Register::isPhysicalRegister(MOReg)) {
1108 Defs.insert(MOReg);
1109 if (!MO.isDead())
1110 LiveDefs.insert(MOReg);
1114 // Check if the reschedule will not break depedencies.
1115 unsigned NumVisited = 0;
1116 for (MachineInstr &OtherMI :
1117 make_range(mi, MachineBasicBlock::iterator(KillMI))) {
1118 // Debug instructions cannot be counted against the limit.
1119 if (OtherMI.isDebugInstr())
1120 continue;
1121 if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
1122 return false;
1123 ++NumVisited;
1124 if (OtherMI.hasUnmodeledSideEffects() || OtherMI.isCall() ||
1125 OtherMI.isBranch() || OtherMI.isTerminator())
1126 // Don't move pass calls, etc.
1127 return false;
1128 SmallVector<unsigned, 2> OtherDefs;
1129 for (const MachineOperand &MO : OtherMI.operands()) {
1130 if (!MO.isReg())
1131 continue;
1132 Register MOReg = MO.getReg();
1133 if (!MOReg)
1134 continue;
1135 if (MO.isUse()) {
1136 if (Defs.count(MOReg))
1137 // Moving KillMI can clobber the physical register if the def has
1138 // not been seen.
1139 return false;
1140 if (Kills.count(MOReg))
1141 // Don't want to extend other live ranges and update kills.
1142 return false;
1143 if (&OtherMI != MI && MOReg == Reg &&
1144 !(MO.isKill() || (LIS && isPlainlyKilled(&OtherMI, MOReg, LIS))))
1145 // We can't schedule across a use of the register in question.
1146 return false;
1147 } else {
1148 OtherDefs.push_back(MOReg);
1152 for (unsigned i = 0, e = OtherDefs.size(); i != e; ++i) {
1153 unsigned MOReg = OtherDefs[i];
1154 if (Uses.count(MOReg))
1155 return false;
1156 if (Register::isPhysicalRegister(MOReg) && LiveDefs.count(MOReg))
1157 return false;
1158 // Physical register def is seen.
1159 Defs.erase(MOReg);
1163 // Move the old kill above MI, don't forget to move debug info as well.
1164 MachineBasicBlock::iterator InsertPos = mi;
1165 while (InsertPos != MBB->begin() && std::prev(InsertPos)->isDebugInstr())
1166 --InsertPos;
1167 MachineBasicBlock::iterator From = KillMI;
1168 MachineBasicBlock::iterator To = std::next(From);
1169 while (std::prev(From)->isDebugInstr())
1170 --From;
1171 MBB->splice(InsertPos, MBB, From, To);
1173 nmi = std::prev(InsertPos); // Backtrack so we process the moved instr.
1174 DistanceMap.erase(DI);
1176 // Update live variables
1177 if (LIS) {
1178 LIS->handleMove(*KillMI);
1179 } else {
1180 LV->removeVirtualRegisterKilled(Reg, *KillMI);
1181 LV->addVirtualRegisterKilled(Reg, *MI);
1184 LLVM_DEBUG(dbgs() << "\trescheduled kill: " << *KillMI);
1185 return true;
1188 /// Tries to commute the operand 'BaseOpIdx' and some other operand in the
1189 /// given machine instruction to improve opportunities for coalescing and
1190 /// elimination of a register to register copy.
1192 /// 'DstOpIdx' specifies the index of MI def operand.
1193 /// 'BaseOpKilled' specifies if the register associated with 'BaseOpIdx'
1194 /// operand is killed by the given instruction.
1195 /// The 'Dist' arguments provides the distance of MI from the start of the
1196 /// current basic block and it is used to determine if it is profitable
1197 /// to commute operands in the instruction.
1199 /// Returns true if the transformation happened. Otherwise, returns false.
1200 bool TwoAddressInstructionPass::tryInstructionCommute(MachineInstr *MI,
1201 unsigned DstOpIdx,
1202 unsigned BaseOpIdx,
1203 bool BaseOpKilled,
1204 unsigned Dist) {
1205 if (!MI->isCommutable())
1206 return false;
1208 bool MadeChange = false;
1209 Register DstOpReg = MI->getOperand(DstOpIdx).getReg();
1210 Register BaseOpReg = MI->getOperand(BaseOpIdx).getReg();
1211 unsigned OpsNum = MI->getDesc().getNumOperands();
1212 unsigned OtherOpIdx = MI->getDesc().getNumDefs();
1213 for (; OtherOpIdx < OpsNum; OtherOpIdx++) {
1214 // The call of findCommutedOpIndices below only checks if BaseOpIdx
1215 // and OtherOpIdx are commutable, it does not really search for
1216 // other commutable operands and does not change the values of passed
1217 // variables.
1218 if (OtherOpIdx == BaseOpIdx || !MI->getOperand(OtherOpIdx).isReg() ||
1219 !TII->findCommutedOpIndices(*MI, BaseOpIdx, OtherOpIdx))
1220 continue;
1222 Register OtherOpReg = MI->getOperand(OtherOpIdx).getReg();
1223 bool AggressiveCommute = false;
1225 // If OtherOp dies but BaseOp does not, swap the OtherOp and BaseOp
1226 // operands. This makes the live ranges of DstOp and OtherOp joinable.
1227 bool OtherOpKilled = isKilled(*MI, OtherOpReg, MRI, TII, LIS, false);
1228 bool DoCommute = !BaseOpKilled && OtherOpKilled;
1230 if (!DoCommute &&
1231 isProfitableToCommute(DstOpReg, BaseOpReg, OtherOpReg, MI, Dist)) {
1232 DoCommute = true;
1233 AggressiveCommute = true;
1236 // If it's profitable to commute, try to do so.
1237 if (DoCommute && commuteInstruction(MI, DstOpIdx, BaseOpIdx, OtherOpIdx,
1238 Dist)) {
1239 MadeChange = true;
1240 ++NumCommuted;
1241 if (AggressiveCommute) {
1242 ++NumAggrCommuted;
1243 // There might be more than two commutable operands, update BaseOp and
1244 // continue scanning.
1245 // FIXME: This assumes that the new instruction's operands are in the
1246 // same positions and were simply swapped.
1247 BaseOpReg = OtherOpReg;
1248 BaseOpKilled = OtherOpKilled;
1249 // Resamples OpsNum in case the number of operands was reduced. This
1250 // happens with X86.
1251 OpsNum = MI->getDesc().getNumOperands();
1252 continue;
1254 // If this was a commute based on kill, we won't do better continuing.
1255 return MadeChange;
1258 return MadeChange;
1261 /// For the case where an instruction has a single pair of tied register
1262 /// operands, attempt some transformations that may either eliminate the tied
1263 /// operands or improve the opportunities for coalescing away the register copy.
1264 /// Returns true if no copy needs to be inserted to untie mi's operands
1265 /// (either because they were untied, or because mi was rescheduled, and will
1266 /// be visited again later). If the shouldOnlyCommute flag is true, only
1267 /// instruction commutation is attempted.
1268 bool TwoAddressInstructionPass::
1269 tryInstructionTransform(MachineBasicBlock::iterator &mi,
1270 MachineBasicBlock::iterator &nmi,
1271 unsigned SrcIdx, unsigned DstIdx,
1272 unsigned Dist, bool shouldOnlyCommute) {
1273 if (OptLevel == CodeGenOpt::None)
1274 return false;
1276 MachineInstr &MI = *mi;
1277 Register regA = MI.getOperand(DstIdx).getReg();
1278 Register regB = MI.getOperand(SrcIdx).getReg();
1280 assert(Register::isVirtualRegister(regB) &&
1281 "cannot make instruction into two-address form");
1282 bool regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
1284 if (Register::isVirtualRegister(regA))
1285 scanUses(regA);
1287 bool Commuted = tryInstructionCommute(&MI, DstIdx, SrcIdx, regBKilled, Dist);
1289 // If the instruction is convertible to 3 Addr, instead
1290 // of returning try 3 Addr transformation aggresively and
1291 // use this variable to check later. Because it might be better.
1292 // For example, we can just use `leal (%rsi,%rdi), %eax` and `ret`
1293 // instead of the following code.
1294 // addl %esi, %edi
1295 // movl %edi, %eax
1296 // ret
1297 if (Commuted && !MI.isConvertibleTo3Addr())
1298 return false;
1300 if (shouldOnlyCommute)
1301 return false;
1303 // If there is one more use of regB later in the same MBB, consider
1304 // re-schedule this MI below it.
1305 if (!Commuted && EnableRescheduling && rescheduleMIBelowKill(mi, nmi, regB)) {
1306 ++NumReSchedDowns;
1307 return true;
1310 // If we commuted, regB may have changed so we should re-sample it to avoid
1311 // confusing the three address conversion below.
1312 if (Commuted) {
1313 regB = MI.getOperand(SrcIdx).getReg();
1314 regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
1317 if (MI.isConvertibleTo3Addr()) {
1318 // This instruction is potentially convertible to a true
1319 // three-address instruction. Check if it is profitable.
1320 if (!regBKilled || isProfitableToConv3Addr(regA, regB)) {
1321 // Try to convert it.
1322 if (convertInstTo3Addr(mi, nmi, regA, regB, Dist)) {
1323 ++NumConvertedTo3Addr;
1324 return true; // Done with this instruction.
1329 // Return if it is commuted but 3 addr conversion is failed.
1330 if (Commuted)
1331 return false;
1333 // If there is one more use of regB later in the same MBB, consider
1334 // re-schedule it before this MI if it's legal.
1335 if (EnableRescheduling && rescheduleKillAboveMI(mi, nmi, regB)) {
1336 ++NumReSchedUps;
1337 return true;
1340 // If this is an instruction with a load folded into it, try unfolding
1341 // the load, e.g. avoid this:
1342 // movq %rdx, %rcx
1343 // addq (%rax), %rcx
1344 // in favor of this:
1345 // movq (%rax), %rcx
1346 // addq %rdx, %rcx
1347 // because it's preferable to schedule a load than a register copy.
1348 if (MI.mayLoad() && !regBKilled) {
1349 // Determine if a load can be unfolded.
1350 unsigned LoadRegIndex;
1351 unsigned NewOpc =
1352 TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(),
1353 /*UnfoldLoad=*/true,
1354 /*UnfoldStore=*/false,
1355 &LoadRegIndex);
1356 if (NewOpc != 0) {
1357 const MCInstrDesc &UnfoldMCID = TII->get(NewOpc);
1358 if (UnfoldMCID.getNumDefs() == 1) {
1359 // Unfold the load.
1360 LLVM_DEBUG(dbgs() << "2addr: UNFOLDING: " << MI);
1361 const TargetRegisterClass *RC =
1362 TRI->getAllocatableClass(
1363 TII->getRegClass(UnfoldMCID, LoadRegIndex, TRI, *MF));
1364 Register Reg = MRI->createVirtualRegister(RC);
1365 SmallVector<MachineInstr *, 2> NewMIs;
1366 if (!TII->unfoldMemoryOperand(*MF, MI, Reg,
1367 /*UnfoldLoad=*/true,
1368 /*UnfoldStore=*/false, NewMIs)) {
1369 LLVM_DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
1370 return false;
1372 assert(NewMIs.size() == 2 &&
1373 "Unfolded a load into multiple instructions!");
1374 // The load was previously folded, so this is the only use.
1375 NewMIs[1]->addRegisterKilled(Reg, TRI);
1377 // Tentatively insert the instructions into the block so that they
1378 // look "normal" to the transformation logic.
1379 MBB->insert(mi, NewMIs[0]);
1380 MBB->insert(mi, NewMIs[1]);
1382 LLVM_DEBUG(dbgs() << "2addr: NEW LOAD: " << *NewMIs[0]
1383 << "2addr: NEW INST: " << *NewMIs[1]);
1385 // Transform the instruction, now that it no longer has a load.
1386 unsigned NewDstIdx = NewMIs[1]->findRegisterDefOperandIdx(regA);
1387 unsigned NewSrcIdx = NewMIs[1]->findRegisterUseOperandIdx(regB);
1388 MachineBasicBlock::iterator NewMI = NewMIs[1];
1389 bool TransformResult =
1390 tryInstructionTransform(NewMI, mi, NewSrcIdx, NewDstIdx, Dist, true);
1391 (void)TransformResult;
1392 assert(!TransformResult &&
1393 "tryInstructionTransform() should return false.");
1394 if (NewMIs[1]->getOperand(NewSrcIdx).isKill()) {
1395 // Success, or at least we made an improvement. Keep the unfolded
1396 // instructions and discard the original.
1397 if (LV) {
1398 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1399 MachineOperand &MO = MI.getOperand(i);
1400 if (MO.isReg() && Register::isVirtualRegister(MO.getReg())) {
1401 if (MO.isUse()) {
1402 if (MO.isKill()) {
1403 if (NewMIs[0]->killsRegister(MO.getReg()))
1404 LV->replaceKillInstruction(MO.getReg(), MI, *NewMIs[0]);
1405 else {
1406 assert(NewMIs[1]->killsRegister(MO.getReg()) &&
1407 "Kill missing after load unfold!");
1408 LV->replaceKillInstruction(MO.getReg(), MI, *NewMIs[1]);
1411 } else if (LV->removeVirtualRegisterDead(MO.getReg(), MI)) {
1412 if (NewMIs[1]->registerDefIsDead(MO.getReg()))
1413 LV->addVirtualRegisterDead(MO.getReg(), *NewMIs[1]);
1414 else {
1415 assert(NewMIs[0]->registerDefIsDead(MO.getReg()) &&
1416 "Dead flag missing after load unfold!");
1417 LV->addVirtualRegisterDead(MO.getReg(), *NewMIs[0]);
1422 LV->addVirtualRegisterKilled(Reg, *NewMIs[1]);
1425 SmallVector<unsigned, 4> OrigRegs;
1426 if (LIS) {
1427 for (const MachineOperand &MO : MI.operands()) {
1428 if (MO.isReg())
1429 OrigRegs.push_back(MO.getReg());
1433 MI.eraseFromParent();
1435 // Update LiveIntervals.
1436 if (LIS) {
1437 MachineBasicBlock::iterator Begin(NewMIs[0]);
1438 MachineBasicBlock::iterator End(NewMIs[1]);
1439 LIS->repairIntervalsInRange(MBB, Begin, End, OrigRegs);
1442 mi = NewMIs[1];
1443 } else {
1444 // Transforming didn't eliminate the tie and didn't lead to an
1445 // improvement. Clean up the unfolded instructions and keep the
1446 // original.
1447 LLVM_DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
1448 NewMIs[0]->eraseFromParent();
1449 NewMIs[1]->eraseFromParent();
1455 return false;
1458 // Collect tied operands of MI that need to be handled.
1459 // Rewrite trivial cases immediately.
1460 // Return true if any tied operands where found, including the trivial ones.
1461 bool TwoAddressInstructionPass::
1462 collectTiedOperands(MachineInstr *MI, TiedOperandMap &TiedOperands) {
1463 const MCInstrDesc &MCID = MI->getDesc();
1464 bool AnyOps = false;
1465 unsigned NumOps = MI->getNumOperands();
1467 for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) {
1468 unsigned DstIdx = 0;
1469 if (!MI->isRegTiedToDefOperand(SrcIdx, &DstIdx))
1470 continue;
1471 AnyOps = true;
1472 MachineOperand &SrcMO = MI->getOperand(SrcIdx);
1473 MachineOperand &DstMO = MI->getOperand(DstIdx);
1474 Register SrcReg = SrcMO.getReg();
1475 Register DstReg = DstMO.getReg();
1476 // Tied constraint already satisfied?
1477 if (SrcReg == DstReg)
1478 continue;
1480 assert(SrcReg && SrcMO.isUse() && "two address instruction invalid");
1482 // Deal with undef uses immediately - simply rewrite the src operand.
1483 if (SrcMO.isUndef() && !DstMO.getSubReg()) {
1484 // Constrain the DstReg register class if required.
1485 if (Register::isVirtualRegister(DstReg))
1486 if (const TargetRegisterClass *RC = TII->getRegClass(MCID, SrcIdx,
1487 TRI, *MF))
1488 MRI->constrainRegClass(DstReg, RC);
1489 SrcMO.setReg(DstReg);
1490 SrcMO.setSubReg(0);
1491 LLVM_DEBUG(dbgs() << "\t\trewrite undef:\t" << *MI);
1492 continue;
1494 TiedOperands[SrcReg].push_back(std::make_pair(SrcIdx, DstIdx));
1496 return AnyOps;
1499 // Process a list of tied MI operands that all use the same source register.
1500 // The tied pairs are of the form (SrcIdx, DstIdx).
1501 void
1502 TwoAddressInstructionPass::processTiedPairs(MachineInstr *MI,
1503 TiedPairList &TiedPairs,
1504 unsigned &Dist) {
1505 bool IsEarlyClobber = false;
1506 for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
1507 const MachineOperand &DstMO = MI->getOperand(TiedPairs[tpi].second);
1508 IsEarlyClobber |= DstMO.isEarlyClobber();
1511 bool RemovedKillFlag = false;
1512 bool AllUsesCopied = true;
1513 unsigned LastCopiedReg = 0;
1514 SlotIndex LastCopyIdx;
1515 unsigned RegB = 0;
1516 unsigned SubRegB = 0;
1517 for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
1518 unsigned SrcIdx = TiedPairs[tpi].first;
1519 unsigned DstIdx = TiedPairs[tpi].second;
1521 const MachineOperand &DstMO = MI->getOperand(DstIdx);
1522 Register RegA = DstMO.getReg();
1524 // Grab RegB from the instruction because it may have changed if the
1525 // instruction was commuted.
1526 RegB = MI->getOperand(SrcIdx).getReg();
1527 SubRegB = MI->getOperand(SrcIdx).getSubReg();
1529 if (RegA == RegB) {
1530 // The register is tied to multiple destinations (or else we would
1531 // not have continued this far), but this use of the register
1532 // already matches the tied destination. Leave it.
1533 AllUsesCopied = false;
1534 continue;
1536 LastCopiedReg = RegA;
1538 assert(Register::isVirtualRegister(RegB) &&
1539 "cannot make instruction into two-address form");
1541 #ifndef NDEBUG
1542 // First, verify that we don't have a use of "a" in the instruction
1543 // (a = b + a for example) because our transformation will not
1544 // work. This should never occur because we are in SSA form.
1545 for (unsigned i = 0; i != MI->getNumOperands(); ++i)
1546 assert(i == DstIdx ||
1547 !MI->getOperand(i).isReg() ||
1548 MI->getOperand(i).getReg() != RegA);
1549 #endif
1551 // Emit a copy.
1552 MachineInstrBuilder MIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1553 TII->get(TargetOpcode::COPY), RegA);
1554 // If this operand is folding a truncation, the truncation now moves to the
1555 // copy so that the register classes remain valid for the operands.
1556 MIB.addReg(RegB, 0, SubRegB);
1557 const TargetRegisterClass *RC = MRI->getRegClass(RegB);
1558 if (SubRegB) {
1559 if (Register::isVirtualRegister(RegA)) {
1560 assert(TRI->getMatchingSuperRegClass(RC, MRI->getRegClass(RegA),
1561 SubRegB) &&
1562 "tied subregister must be a truncation");
1563 // The superreg class will not be used to constrain the subreg class.
1564 RC = nullptr;
1565 } else {
1566 assert(TRI->getMatchingSuperReg(RegA, SubRegB, MRI->getRegClass(RegB))
1567 && "tied subregister must be a truncation");
1571 // Update DistanceMap.
1572 MachineBasicBlock::iterator PrevMI = MI;
1573 --PrevMI;
1574 DistanceMap.insert(std::make_pair(&*PrevMI, Dist));
1575 DistanceMap[MI] = ++Dist;
1577 if (LIS) {
1578 LastCopyIdx = LIS->InsertMachineInstrInMaps(*PrevMI).getRegSlot();
1580 if (Register::isVirtualRegister(RegA)) {
1581 LiveInterval &LI = LIS->getInterval(RegA);
1582 VNInfo *VNI = LI.getNextValue(LastCopyIdx, LIS->getVNInfoAllocator());
1583 SlotIndex endIdx =
1584 LIS->getInstructionIndex(*MI).getRegSlot(IsEarlyClobber);
1585 LI.addSegment(LiveInterval::Segment(LastCopyIdx, endIdx, VNI));
1589 LLVM_DEBUG(dbgs() << "\t\tprepend:\t" << *MIB);
1591 MachineOperand &MO = MI->getOperand(SrcIdx);
1592 assert(MO.isReg() && MO.getReg() == RegB && MO.isUse() &&
1593 "inconsistent operand info for 2-reg pass");
1594 if (MO.isKill()) {
1595 MO.setIsKill(false);
1596 RemovedKillFlag = true;
1599 // Make sure regA is a legal regclass for the SrcIdx operand.
1600 if (Register::isVirtualRegister(RegA) && Register::isVirtualRegister(RegB))
1601 MRI->constrainRegClass(RegA, RC);
1602 MO.setReg(RegA);
1603 // The getMatchingSuper asserts guarantee that the register class projected
1604 // by SubRegB is compatible with RegA with no subregister. So regardless of
1605 // whether the dest oper writes a subreg, the source oper should not.
1606 MO.setSubReg(0);
1608 // Propagate SrcRegMap.
1609 SrcRegMap[RegA] = RegB;
1612 if (AllUsesCopied) {
1613 bool ReplacedAllUntiedUses = true;
1614 if (!IsEarlyClobber) {
1615 // Replace other (un-tied) uses of regB with LastCopiedReg.
1616 for (MachineOperand &MO : MI->operands()) {
1617 if (MO.isReg() && MO.getReg() == RegB && MO.isUse()) {
1618 if (MO.getSubReg() == SubRegB) {
1619 if (MO.isKill()) {
1620 MO.setIsKill(false);
1621 RemovedKillFlag = true;
1623 MO.setReg(LastCopiedReg);
1624 MO.setSubReg(0);
1625 } else {
1626 ReplacedAllUntiedUses = false;
1632 // Update live variables for regB.
1633 if (RemovedKillFlag && ReplacedAllUntiedUses &&
1634 LV && LV->getVarInfo(RegB).removeKill(*MI)) {
1635 MachineBasicBlock::iterator PrevMI = MI;
1636 --PrevMI;
1637 LV->addVirtualRegisterKilled(RegB, *PrevMI);
1640 // Update LiveIntervals.
1641 if (LIS) {
1642 LiveInterval &LI = LIS->getInterval(RegB);
1643 SlotIndex MIIdx = LIS->getInstructionIndex(*MI);
1644 LiveInterval::const_iterator I = LI.find(MIIdx);
1645 assert(I != LI.end() && "RegB must be live-in to use.");
1647 SlotIndex UseIdx = MIIdx.getRegSlot(IsEarlyClobber);
1648 if (I->end == UseIdx)
1649 LI.removeSegment(LastCopyIdx, UseIdx);
1651 } else if (RemovedKillFlag) {
1652 // Some tied uses of regB matched their destination registers, so
1653 // regB is still used in this instruction, but a kill flag was
1654 // removed from a different tied use of regB, so now we need to add
1655 // a kill flag to one of the remaining uses of regB.
1656 for (MachineOperand &MO : MI->operands()) {
1657 if (MO.isReg() && MO.getReg() == RegB && MO.isUse()) {
1658 MO.setIsKill(true);
1659 break;
1665 /// Reduce two-address instructions to two operands.
1666 bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &Func) {
1667 MF = &Func;
1668 const TargetMachine &TM = MF->getTarget();
1669 MRI = &MF->getRegInfo();
1670 TII = MF->getSubtarget().getInstrInfo();
1671 TRI = MF->getSubtarget().getRegisterInfo();
1672 InstrItins = MF->getSubtarget().getInstrItineraryData();
1673 LV = getAnalysisIfAvailable<LiveVariables>();
1674 LIS = getAnalysisIfAvailable<LiveIntervals>();
1675 if (auto *AAPass = getAnalysisIfAvailable<AAResultsWrapperPass>())
1676 AA = &AAPass->getAAResults();
1677 else
1678 AA = nullptr;
1679 OptLevel = TM.getOptLevel();
1680 // Disable optimizations if requested. We cannot skip the whole pass as some
1681 // fixups are necessary for correctness.
1682 if (skipFunction(Func.getFunction()))
1683 OptLevel = CodeGenOpt::None;
1685 bool MadeChange = false;
1687 LLVM_DEBUG(dbgs() << "********** REWRITING TWO-ADDR INSTRS **********\n");
1688 LLVM_DEBUG(dbgs() << "********** Function: " << MF->getName() << '\n');
1690 // This pass takes the function out of SSA form.
1691 MRI->leaveSSA();
1693 TiedOperandMap TiedOperands;
1694 for (MachineFunction::iterator MBBI = MF->begin(), MBBE = MF->end();
1695 MBBI != MBBE; ++MBBI) {
1696 MBB = &*MBBI;
1697 unsigned Dist = 0;
1698 DistanceMap.clear();
1699 SrcRegMap.clear();
1700 DstRegMap.clear();
1701 Processed.clear();
1702 SunkInstrs.clear();
1703 for (MachineBasicBlock::iterator mi = MBB->begin(), me = MBB->end();
1704 mi != me; ) {
1705 MachineBasicBlock::iterator nmi = std::next(mi);
1706 // Don't revisit an instruction previously converted by target. It may
1707 // contain undef register operands (%noreg), which are not handled.
1708 if (mi->isDebugInstr() || SunkInstrs.count(&*mi)) {
1709 mi = nmi;
1710 continue;
1713 // Expand REG_SEQUENCE instructions. This will position mi at the first
1714 // expanded instruction.
1715 if (mi->isRegSequence())
1716 eliminateRegSequence(mi);
1718 DistanceMap.insert(std::make_pair(&*mi, ++Dist));
1720 processCopy(&*mi);
1722 // First scan through all the tied register uses in this instruction
1723 // and record a list of pairs of tied operands for each register.
1724 if (!collectTiedOperands(&*mi, TiedOperands)) {
1725 mi = nmi;
1726 continue;
1729 ++NumTwoAddressInstrs;
1730 MadeChange = true;
1731 LLVM_DEBUG(dbgs() << '\t' << *mi);
1733 // If the instruction has a single pair of tied operands, try some
1734 // transformations that may either eliminate the tied operands or
1735 // improve the opportunities for coalescing away the register copy.
1736 if (TiedOperands.size() == 1) {
1737 SmallVectorImpl<std::pair<unsigned, unsigned>> &TiedPairs
1738 = TiedOperands.begin()->second;
1739 if (TiedPairs.size() == 1) {
1740 unsigned SrcIdx = TiedPairs[0].first;
1741 unsigned DstIdx = TiedPairs[0].second;
1742 Register SrcReg = mi->getOperand(SrcIdx).getReg();
1743 Register DstReg = mi->getOperand(DstIdx).getReg();
1744 if (SrcReg != DstReg &&
1745 tryInstructionTransform(mi, nmi, SrcIdx, DstIdx, Dist, false)) {
1746 // The tied operands have been eliminated or shifted further down
1747 // the block to ease elimination. Continue processing with 'nmi'.
1748 TiedOperands.clear();
1749 mi = nmi;
1750 continue;
1755 // Now iterate over the information collected above.
1756 for (auto &TO : TiedOperands) {
1757 processTiedPairs(&*mi, TO.second, Dist);
1758 LLVM_DEBUG(dbgs() << "\t\trewrite to:\t" << *mi);
1761 // Rewrite INSERT_SUBREG as COPY now that we no longer need SSA form.
1762 if (mi->isInsertSubreg()) {
1763 // From %reg = INSERT_SUBREG %reg, %subreg, subidx
1764 // To %reg:subidx = COPY %subreg
1765 unsigned SubIdx = mi->getOperand(3).getImm();
1766 mi->RemoveOperand(3);
1767 assert(mi->getOperand(0).getSubReg() == 0 && "Unexpected subreg idx");
1768 mi->getOperand(0).setSubReg(SubIdx);
1769 mi->getOperand(0).setIsUndef(mi->getOperand(1).isUndef());
1770 mi->RemoveOperand(1);
1771 mi->setDesc(TII->get(TargetOpcode::COPY));
1772 LLVM_DEBUG(dbgs() << "\t\tconvert to:\t" << *mi);
1775 // Clear TiedOperands here instead of at the top of the loop
1776 // since most instructions do not have tied operands.
1777 TiedOperands.clear();
1778 mi = nmi;
1782 if (LIS)
1783 MF->verify(this, "After two-address instruction pass");
1785 return MadeChange;
1788 /// Eliminate a REG_SEQUENCE instruction as part of the de-ssa process.
1790 /// The instruction is turned into a sequence of sub-register copies:
1792 /// %dst = REG_SEQUENCE %v1, ssub0, %v2, ssub1
1794 /// Becomes:
1796 /// undef %dst:ssub0 = COPY %v1
1797 /// %dst:ssub1 = COPY %v2
1798 void TwoAddressInstructionPass::
1799 eliminateRegSequence(MachineBasicBlock::iterator &MBBI) {
1800 MachineInstr &MI = *MBBI;
1801 Register DstReg = MI.getOperand(0).getReg();
1802 if (MI.getOperand(0).getSubReg() || Register::isPhysicalRegister(DstReg) ||
1803 !(MI.getNumOperands() & 1)) {
1804 LLVM_DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << MI);
1805 llvm_unreachable(nullptr);
1808 SmallVector<unsigned, 4> OrigRegs;
1809 if (LIS) {
1810 OrigRegs.push_back(MI.getOperand(0).getReg());
1811 for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2)
1812 OrigRegs.push_back(MI.getOperand(i).getReg());
1815 bool DefEmitted = false;
1816 for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2) {
1817 MachineOperand &UseMO = MI.getOperand(i);
1818 Register SrcReg = UseMO.getReg();
1819 unsigned SubIdx = MI.getOperand(i+1).getImm();
1820 // Nothing needs to be inserted for undef operands.
1821 if (UseMO.isUndef())
1822 continue;
1824 // Defer any kill flag to the last operand using SrcReg. Otherwise, we
1825 // might insert a COPY that uses SrcReg after is was killed.
1826 bool isKill = UseMO.isKill();
1827 if (isKill)
1828 for (unsigned j = i + 2; j < e; j += 2)
1829 if (MI.getOperand(j).getReg() == SrcReg) {
1830 MI.getOperand(j).setIsKill();
1831 UseMO.setIsKill(false);
1832 isKill = false;
1833 break;
1836 // Insert the sub-register copy.
1837 MachineInstr *CopyMI = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(),
1838 TII->get(TargetOpcode::COPY))
1839 .addReg(DstReg, RegState::Define, SubIdx)
1840 .add(UseMO);
1842 // The first def needs an undef flag because there is no live register
1843 // before it.
1844 if (!DefEmitted) {
1845 CopyMI->getOperand(0).setIsUndef(true);
1846 // Return an iterator pointing to the first inserted instr.
1847 MBBI = CopyMI;
1849 DefEmitted = true;
1851 // Update LiveVariables' kill info.
1852 if (LV && isKill && !Register::isPhysicalRegister(SrcReg))
1853 LV->replaceKillInstruction(SrcReg, MI, *CopyMI);
1855 LLVM_DEBUG(dbgs() << "Inserted: " << *CopyMI);
1858 MachineBasicBlock::iterator EndMBBI =
1859 std::next(MachineBasicBlock::iterator(MI));
1861 if (!DefEmitted) {
1862 LLVM_DEBUG(dbgs() << "Turned: " << MI << " into an IMPLICIT_DEF");
1863 MI.setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
1864 for (int j = MI.getNumOperands() - 1, ee = 0; j > ee; --j)
1865 MI.RemoveOperand(j);
1866 } else {
1867 LLVM_DEBUG(dbgs() << "Eliminated: " << MI);
1868 MI.eraseFromParent();
1871 // Udpate LiveIntervals.
1872 if (LIS)
1873 LIS->repairIntervalsInRange(MBB, MBBI, EndMBBI, OrigRegs);