1 //===-- SimpleRegisterCoalescing.cpp - Register Coalescing ----------------===//
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 a simple register coalescing pass that attempts to
11 // aggressively coalesce every register copy that it can.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "regcoalescing"
16 #include "SimpleRegisterCoalescing.h"
17 #include "VirtRegMap.h"
18 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
19 #include "llvm/Value.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineInstr.h"
23 #include "llvm/CodeGen/MachineLoopInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/Passes.h"
26 #include "llvm/CodeGen/RegisterCoalescer.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/OwningPtr.h"
35 #include "llvm/ADT/SmallSet.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/STLExtras.h"
42 STATISTIC(numJoins , "Number of interval joins performed");
43 STATISTIC(numCrossRCs , "Number of cross class joins performed");
44 STATISTIC(numCommutes , "Number of instruction commuting performed");
45 STATISTIC(numExtends , "Number of copies extended");
46 STATISTIC(NumReMats , "Number of instructions re-materialized");
47 STATISTIC(numPeep , "Number of identity moves eliminated after coalescing");
48 STATISTIC(numAborts , "Number of times interval joining aborted");
49 STATISTIC(numDeadValNo, "Number of valno def marked dead");
51 char SimpleRegisterCoalescing::ID = 0;
53 EnableJoining("join-liveintervals",
54 cl::desc("Coalesce copies (default=true)"),
58 DisableCrossClassJoin("disable-cross-class-join",
59 cl::desc("Avoid coalescing cross register class copies"),
60 cl::init(false), cl::Hidden);
62 static RegisterPass<SimpleRegisterCoalescing>
63 X("simple-register-coalescing", "Simple Register Coalescing");
65 // Declare that we implement the RegisterCoalescer interface
66 static RegisterAnalysisGroup<RegisterCoalescer, true/*The Default*/> V(X);
68 const PassInfo *const llvm::SimpleRegisterCoalescingID = &X;
70 void SimpleRegisterCoalescing::getAnalysisUsage(AnalysisUsage &AU) const {
72 AU.addRequired<AliasAnalysis>();
73 AU.addRequired<LiveIntervals>();
74 AU.addPreserved<LiveIntervals>();
75 AU.addPreserved<SlotIndexes>();
76 AU.addRequired<MachineLoopInfo>();
77 AU.addPreserved<MachineLoopInfo>();
78 AU.addPreservedID(MachineDominatorsID);
80 AU.addPreservedID(StrongPHIEliminationID);
82 AU.addPreservedID(PHIEliminationID);
83 AU.addPreservedID(TwoAddressInstructionPassID);
84 MachineFunctionPass::getAnalysisUsage(AU);
87 /// AdjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA
88 /// being the source and IntB being the dest, thus this defines a value number
89 /// in IntB. If the source value number (in IntA) is defined by a copy from B,
90 /// see if we can merge these two pieces of B into a single value number,
91 /// eliminating a copy. For example:
95 /// B1 = A3 <- this copy
97 /// In this case, B0 can be extended to where the B1 copy lives, allowing the B1
98 /// value number to be replaced with B0 (which simplifies the B liveinterval).
100 /// This returns true if an interval was modified.
102 bool SimpleRegisterCoalescing::AdjustCopiesBackFrom(LiveInterval &IntA,
104 MachineInstr *CopyMI) {
105 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI).getDefIndex();
107 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
108 // the example above.
109 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
110 assert(BLR != IntB.end() && "Live range not found!");
111 VNInfo *BValNo = BLR->valno;
113 // Get the location that B is defined at. Two options: either this value has
114 // an unknown definition point or it is defined at CopyIdx. If unknown, we
116 if (!BValNo->getCopy()) return false;
117 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
119 // AValNo is the value number in A that defines the copy, A3 in the example.
120 SlotIndex CopyUseIdx = CopyIdx.getUseIndex();
121 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyUseIdx);
122 assert(ALR != IntA.end() && "Live range not found!");
123 VNInfo *AValNo = ALR->valno;
124 // If it's re-defined by an early clobber somewhere in the live range, then
125 // it's not safe to eliminate the copy. FIXME: This is a temporary workaround.
127 // 172 %ECX<def> = MOV32rr %reg1039<kill>
128 // 180 INLINEASM <es:subl $5,$1
129 // sbbl $3,$0>, 10, %EAX<def>, 14, %ECX<earlyclobber,def>, 9,
131 // 36, <fi#0>, 1, %reg0, 0, 9, %ECX<kill>, 36, <fi#1>, 1, %reg0, 0
132 // 188 %EAX<def> = MOV32rr %EAX<kill>
133 // 196 %ECX<def> = MOV32rr %ECX<kill>
134 // 204 %ECX<def> = MOV32rr %ECX<kill>
135 // 212 %EAX<def> = MOV32rr %EAX<kill>
136 // 220 %EAX<def> = MOV32rr %EAX
137 // 228 %reg1039<def> = MOV32rr %ECX<kill>
138 // The early clobber operand ties ECX input to the ECX def.
140 // The live interval of ECX is represented as this:
141 // %reg20,inf = [46,47:1)[174,230:0) 0@174-(230) 1@46-(47)
142 // The coalescer has no idea there was a def in the middle of [174,230].
143 if (AValNo->hasRedefByEC())
146 // If AValNo is defined as a copy from IntB, we can potentially process this.
147 // Get the instruction that defines this value number.
148 unsigned SrcReg = li_->getVNInfoSourceReg(AValNo);
149 if (!SrcReg) return false; // Not defined by a copy.
151 // If the value number is not defined by a copy instruction, ignore it.
153 // If the source register comes from an interval other than IntB, we can't
155 if (SrcReg != IntB.reg) return false;
157 // Get the LiveRange in IntB that this value number starts with.
158 LiveInterval::iterator ValLR =
159 IntB.FindLiveRangeContaining(AValNo->def.getPrevSlot());
160 assert(ValLR != IntB.end() && "Live range not found!");
162 // Make sure that the end of the live range is inside the same block as
164 MachineInstr *ValLREndInst =
165 li_->getInstructionFromIndex(ValLR->end.getPrevSlot());
167 ValLREndInst->getParent() != CopyMI->getParent()) return false;
169 // Okay, we now know that ValLR ends in the same block that the CopyMI
170 // live-range starts. If there are no intervening live ranges between them in
171 // IntB, we can merge them.
172 if (ValLR+1 != BLR) return false;
174 // If a live interval is a physical register, conservatively check if any
175 // of its sub-registers is overlapping the live interval of the virtual
176 // register. If so, do not coalesce.
177 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg) &&
178 *tri_->getSubRegisters(IntB.reg)) {
179 for (const unsigned* SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR)
180 if (li_->hasInterval(*SR) && IntA.overlaps(li_->getInterval(*SR))) {
182 dbgs() << "\t\tInterfere with sub-register ";
183 li_->getInterval(*SR).print(dbgs(), tri_);
190 dbgs() << "Extending: ";
191 IntB.print(dbgs(), tri_);
194 SlotIndex FillerStart = ValLR->end, FillerEnd = BLR->start;
195 // We are about to delete CopyMI, so need to remove it as the 'instruction
196 // that defines this value #'. Update the valnum with the new defining
198 BValNo->def = FillerStart;
201 // Okay, we can merge them. We need to insert a new liverange:
202 // [ValLR.end, BLR.begin) of either value number, then we merge the
203 // two value numbers.
204 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
206 // If the IntB live range is assigned to a physical register, and if that
207 // physreg has sub-registers, update their live intervals as well.
208 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
209 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
210 LiveInterval &SRLI = li_->getInterval(*SR);
211 SRLI.addRange(LiveRange(FillerStart, FillerEnd,
212 SRLI.getNextValue(FillerStart, 0, true,
213 li_->getVNInfoAllocator())));
217 // Okay, merge "B1" into the same value number as "B0".
218 if (BValNo != ValLR->valno) {
219 IntB.addKills(ValLR->valno, BValNo->kills);
220 IntB.MergeValueNumberInto(BValNo, ValLR->valno);
223 dbgs() << " result = ";
224 IntB.print(dbgs(), tri_);
228 // If the source instruction was killing the source register before the
229 // merge, unset the isKill marker given the live range has been extended.
230 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true);
232 ValLREndInst->getOperand(UIdx).setIsKill(false);
233 ValLR->valno->removeKill(FillerStart);
236 // If the copy instruction was killing the destination register before the
237 // merge, find the last use and trim the live range. That will also add the
239 if (ALR->valno->isKill(CopyIdx))
240 TrimLiveIntervalToLastUse(CopyUseIdx, CopyMI->getParent(), IntA, ALR);
246 /// HasOtherReachingDefs - Return true if there are definitions of IntB
247 /// other than BValNo val# that can reach uses of AValno val# of IntA.
248 bool SimpleRegisterCoalescing::HasOtherReachingDefs(LiveInterval &IntA,
252 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
254 if (AI->valno != AValNo) continue;
255 LiveInterval::Ranges::iterator BI =
256 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start);
257 if (BI != IntB.ranges.begin())
259 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) {
260 if (BI->valno == BValNo)
262 // When BValNo is null, we're looking for a dummy clobber-value for a subreg.
263 if (!BValNo && !BI->valno->isDefAccurate() && !BI->valno->getCopy())
265 if (BI->start <= AI->start && BI->end > AI->start)
267 if (BI->start > AI->start && BI->start < AI->end)
275 TransferImplicitOps(MachineInstr *MI, MachineInstr *NewMI) {
276 for (unsigned i = MI->getDesc().getNumOperands(), e = MI->getNumOperands();
278 MachineOperand &MO = MI->getOperand(i);
279 if (MO.isReg() && MO.isImplicit())
280 NewMI->addOperand(MO);
284 /// RemoveCopyByCommutingDef - We found a non-trivially-coalescable copy with
285 /// IntA being the source and IntB being the dest, thus this defines a value
286 /// number in IntB. If the source value number (in IntA) is defined by a
287 /// commutable instruction and its other operand is coalesced to the copy dest
288 /// register, see if we can transform the copy into a noop by commuting the
289 /// definition. For example,
291 /// A3 = op A2 B0<kill>
293 /// B1 = A3 <- this copy
295 /// = op A3 <- more uses
299 /// B2 = op B0 A2<kill>
301 /// B1 = B2 <- now an identify copy
303 /// = op B2 <- more uses
305 /// This returns true if an interval was modified.
307 bool SimpleRegisterCoalescing::RemoveCopyByCommutingDef(LiveInterval &IntA,
309 MachineInstr *CopyMI) {
311 li_->getInstructionIndex(CopyMI).getDefIndex();
313 // FIXME: For now, only eliminate the copy by commuting its def when the
314 // source register is a virtual register. We want to guard against cases
315 // where the copy is a back edge copy and commuting the def lengthen the
316 // live interval of the source register to the entire loop.
317 if (TargetRegisterInfo::isPhysicalRegister(IntA.reg))
320 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
321 // the example above.
322 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
323 assert(BLR != IntB.end() && "Live range not found!");
324 VNInfo *BValNo = BLR->valno;
326 // Get the location that B is defined at. Two options: either this value has
327 // an unknown definition point or it is defined at CopyIdx. If unknown, we
329 if (!BValNo->getCopy()) return false;
330 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
332 // AValNo is the value number in A that defines the copy, A3 in the example.
333 LiveInterval::iterator ALR =
334 IntA.FindLiveRangeContaining(CopyIdx.getUseIndex()); //
336 assert(ALR != IntA.end() && "Live range not found!");
337 VNInfo *AValNo = ALR->valno;
338 // If other defs can reach uses of this def, then it's not safe to perform
339 // the optimization. FIXME: Do isPHIDef and isDefAccurate both need to be
341 if (AValNo->isPHIDef() || !AValNo->isDefAccurate() ||
342 AValNo->isUnused() || AValNo->hasPHIKill())
344 MachineInstr *DefMI = li_->getInstructionFromIndex(AValNo->def);
345 const TargetInstrDesc &TID = DefMI->getDesc();
346 if (!TID.isCommutable())
348 // If DefMI is a two-address instruction then commuting it will change the
349 // destination register.
350 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
351 assert(DefIdx != -1);
353 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
355 unsigned Op1, Op2, NewDstIdx;
356 if (!tii_->findCommutedOpIndices(DefMI, Op1, Op2))
360 else if (Op2 == UseOpIdx)
365 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
366 unsigned NewReg = NewDstMO.getReg();
367 if (NewReg != IntB.reg || !NewDstMO.isKill())
370 // Make sure there are no other definitions of IntB that would reach the
371 // uses which the new definition can reach.
372 if (HasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
375 bool BHasSubRegs = false;
376 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg))
377 BHasSubRegs = *tri_->getSubRegisters(IntB.reg);
379 // Abort if the subregisters of IntB.reg have values that are not simply the
380 // clobbers from the superreg.
382 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR)
383 if (HasOtherReachingDefs(IntA, li_->getInterval(*SR), AValNo, 0))
386 // If some of the uses of IntA.reg is already coalesced away, return false.
387 // It's not possible to determine whether it's safe to perform the coalescing.
388 for (MachineRegisterInfo::use_nodbg_iterator UI =
389 mri_->use_nodbg_begin(IntA.reg),
390 UE = mri_->use_nodbg_end(); UI != UE; ++UI) {
391 MachineInstr *UseMI = &*UI;
392 SlotIndex UseIdx = li_->getInstructionIndex(UseMI);
393 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
394 if (ULR == IntA.end())
396 if (ULR->valno == AValNo && JoinedCopies.count(UseMI))
400 // At this point we have decided that it is legal to do this
401 // transformation. Start by commuting the instruction.
402 MachineBasicBlock *MBB = DefMI->getParent();
403 MachineInstr *NewMI = tii_->commuteInstruction(DefMI);
406 if (NewMI != DefMI) {
407 li_->ReplaceMachineInstrInMaps(DefMI, NewMI);
408 MBB->insert(DefMI, NewMI);
411 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false);
412 NewMI->getOperand(OpIdx).setIsKill();
414 bool BHasPHIKill = BValNo->hasPHIKill();
415 SmallVector<VNInfo*, 4> BDeadValNos;
416 VNInfo::KillSet BKills;
417 std::map<SlotIndex, SlotIndex> BExtend;
419 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
428 // then do not add kills of A to the newly created B interval.
429 bool Extended = BLR->end > ALR->end && ALR->end != ALR->start;
431 BExtend[ALR->end] = BLR->end;
433 // Update uses of IntA of the specific Val# with IntB.
434 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(IntA.reg),
435 UE = mri_->use_end(); UI != UE;) {
436 MachineOperand &UseMO = UI.getOperand();
437 MachineInstr *UseMI = &*UI;
439 if (JoinedCopies.count(UseMI))
441 if (UseMI->isDebugValue()) {
442 // FIXME These don't have an instruction index. Not clear we have enough
443 // info to decide whether to do this replacement or not. For now do it.
444 UseMO.setReg(NewReg);
447 SlotIndex UseIdx = li_->getInstructionIndex(UseMI).getUseIndex();
448 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
449 if (ULR == IntA.end() || ULR->valno != AValNo)
451 UseMO.setReg(NewReg);
454 if (UseMO.isKill()) {
456 UseMO.setIsKill(false);
458 BKills.push_back(UseIdx.getDefIndex());
460 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
461 if (!tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
463 if (DstReg == IntB.reg && DstSubIdx == 0) {
464 // This copy will become a noop. If it's defining a new val#,
465 // remove that val# as well. However this live range is being
466 // extended to the end of the existing live range defined by the copy.
467 SlotIndex DefIdx = UseIdx.getDefIndex();
468 const LiveRange *DLR = IntB.getLiveRangeContaining(DefIdx);
469 BHasPHIKill |= DLR->valno->hasPHIKill();
470 assert(DLR->valno->def == DefIdx);
471 BDeadValNos.push_back(DLR->valno);
472 BExtend[DLR->start] = DLR->end;
473 JoinedCopies.insert(UseMI);
474 // If this is a kill but it's going to be removed, the last use
475 // of the same val# is the new kill.
481 // We need to insert a new liverange: [ALR.start, LastUse). It may be we can
482 // simply extend BLR if CopyMI doesn't end the range.
484 dbgs() << "Extending: ";
485 IntB.print(dbgs(), tri_);
488 // Remove val#'s defined by copies that will be coalesced away.
489 for (unsigned i = 0, e = BDeadValNos.size(); i != e; ++i) {
490 VNInfo *DeadVNI = BDeadValNos[i];
492 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
493 LiveInterval &SRLI = li_->getInterval(*SR);
494 const LiveRange *SRLR = SRLI.getLiveRangeContaining(DeadVNI->def);
495 SRLI.removeValNo(SRLR->valno);
498 IntB.removeValNo(BDeadValNos[i]);
501 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition
502 // is updated. Kills are also updated.
503 VNInfo *ValNo = BValNo;
504 ValNo->def = AValNo->def;
506 for (unsigned j = 0, ee = ValNo->kills.size(); j != ee; ++j) {
507 if (ValNo->kills[j] != BLR->end)
508 BKills.push_back(ValNo->kills[j]);
510 ValNo->kills.clear();
511 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
513 if (AI->valno != AValNo) continue;
514 SlotIndex End = AI->end;
515 std::map<SlotIndex, SlotIndex>::iterator
516 EI = BExtend.find(End);
517 if (EI != BExtend.end())
519 IntB.addRange(LiveRange(AI->start, End, ValNo));
521 // If the IntB live range is assigned to a physical register, and if that
522 // physreg has sub-registers, update their live intervals as well.
524 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
525 LiveInterval &SRLI = li_->getInterval(*SR);
526 SRLI.MergeInClobberRange(*li_, AI->start, End,
527 li_->getVNInfoAllocator());
531 IntB.addKills(ValNo, BKills);
532 ValNo->setHasPHIKill(BHasPHIKill);
535 dbgs() << " result = ";
536 IntB.print(dbgs(), tri_);
537 dbgs() << "\nShortening: ";
538 IntA.print(dbgs(), tri_);
541 IntA.removeValNo(AValNo);
544 dbgs() << " result = ";
545 IntA.print(dbgs(), tri_);
553 /// isSameOrFallThroughBB - Return true if MBB == SuccMBB or MBB simply
554 /// fallthoughs to SuccMBB.
555 static bool isSameOrFallThroughBB(MachineBasicBlock *MBB,
556 MachineBasicBlock *SuccMBB,
557 const TargetInstrInfo *tii_) {
560 MachineBasicBlock *TBB = 0, *FBB = 0;
561 SmallVector<MachineOperand, 4> Cond;
562 return !tii_->AnalyzeBranch(*MBB, TBB, FBB, Cond) && !TBB && !FBB &&
563 MBB->isSuccessor(SuccMBB);
566 /// removeRange - Wrapper for LiveInterval::removeRange. This removes a range
567 /// from a physical register live interval as well as from the live intervals
568 /// of its sub-registers.
569 static void removeRange(LiveInterval &li,
570 SlotIndex Start, SlotIndex End,
571 LiveIntervals *li_, const TargetRegisterInfo *tri_) {
572 li.removeRange(Start, End, true);
573 if (TargetRegisterInfo::isPhysicalRegister(li.reg)) {
574 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
575 if (!li_->hasInterval(*SR))
577 LiveInterval &sli = li_->getInterval(*SR);
578 SlotIndex RemoveStart = Start;
579 SlotIndex RemoveEnd = Start;
581 while (RemoveEnd != End) {
582 LiveInterval::iterator LR = sli.FindLiveRangeContaining(RemoveStart);
585 RemoveEnd = (LR->end < End) ? LR->end : End;
586 sli.removeRange(RemoveStart, RemoveEnd, true);
587 RemoveStart = RemoveEnd;
593 /// TrimLiveIntervalToLastUse - If there is a last use in the same basic block
594 /// as the copy instruction, trim the live interval to the last use and return
597 SimpleRegisterCoalescing::TrimLiveIntervalToLastUse(SlotIndex CopyIdx,
598 MachineBasicBlock *CopyMBB,
600 const LiveRange *LR) {
601 SlotIndex MBBStart = li_->getMBBStartIdx(CopyMBB);
602 SlotIndex LastUseIdx;
603 MachineOperand *LastUse =
604 lastRegisterUse(LR->start, CopyIdx.getPrevSlot(), li.reg, LastUseIdx);
606 MachineInstr *LastUseMI = LastUse->getParent();
607 if (!isSameOrFallThroughBB(LastUseMI->getParent(), CopyMBB, tii_)) {
614 // r1025<dead> = r1024<kill>
615 if (MBBStart < LR->end)
616 removeRange(li, MBBStart, LR->end, li_, tri_);
620 // There are uses before the copy, just shorten the live range to the end
622 LastUse->setIsKill();
623 removeRange(li, LastUseIdx.getDefIndex(), LR->end, li_, tri_);
624 LR->valno->addKill(LastUseIdx.getDefIndex());
625 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
626 if (tii_->isMoveInstr(*LastUseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
627 DstReg == li.reg && DstSubIdx == 0) {
628 // Last use is itself an identity code.
629 int DeadIdx = LastUseMI->findRegisterDefOperandIdx(li.reg,
631 LastUseMI->getOperand(DeadIdx).setIsDead();
637 if (LR->start <= MBBStart && LR->end > MBBStart) {
638 if (LR->start == li_->getZeroIndex()) {
639 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
640 // Live-in to the function but dead. Remove it from entry live-in set.
641 mf_->begin()->removeLiveIn(li.reg);
643 // FIXME: Shorten intervals in BBs that reaches this BB.
649 /// ReMaterializeTrivialDef - If the source of a copy is defined by a trivial
650 /// computation, replace the copy by rematerialize the definition.
651 bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
654 MachineInstr *CopyMI) {
655 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI).getUseIndex();
656 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx);
657 assert(SrcLR != SrcInt.end() && "Live range not found!");
658 VNInfo *ValNo = SrcLR->valno;
659 // If other defs can reach uses of this def, then it's not safe to perform
660 // the optimization. FIXME: Do isPHIDef and isDefAccurate both need to be
662 if (ValNo->isPHIDef() || !ValNo->isDefAccurate() ||
663 ValNo->isUnused() || ValNo->hasPHIKill())
665 MachineInstr *DefMI = li_->getInstructionFromIndex(ValNo->def);
666 const TargetInstrDesc &TID = DefMI->getDesc();
667 if (!TID.isAsCheapAsAMove())
669 if (!tii_->isTriviallyReMaterializable(DefMI, AA))
671 bool SawStore = false;
672 if (!DefMI->isSafeToMove(tii_, AA, SawStore))
674 if (TID.getNumDefs() != 1)
676 if (!DefMI->isImplicitDef()) {
677 // Make sure the copy destination register class fits the instruction
678 // definition register class. The mismatch can happen as a result of earlier
679 // extract_subreg, insert_subreg, subreg_to_reg coalescing.
680 const TargetRegisterClass *RC = TID.OpInfo[0].getRegClass(tri_);
681 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
682 if (mri_->getRegClass(DstReg) != RC)
684 } else if (!RC->contains(DstReg))
688 // If destination register has a sub-register index on it, make sure it mtches
689 // the instruction register class.
691 const TargetInstrDesc &TID = DefMI->getDesc();
692 if (TID.getNumDefs() != 1)
694 const TargetRegisterClass *DstRC = mri_->getRegClass(DstReg);
695 const TargetRegisterClass *DstSubRC =
696 DstRC->getSubRegisterRegClass(DstSubIdx);
697 const TargetRegisterClass *DefRC = TID.OpInfo[0].getRegClass(tri_);
700 else if (DefRC != DstSubRC)
704 SlotIndex DefIdx = CopyIdx.getDefIndex();
705 const LiveRange *DLR= li_->getInterval(DstReg).getLiveRangeContaining(DefIdx);
706 DLR->valno->setCopy(0);
707 // Don't forget to update sub-register intervals.
708 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
709 for (const unsigned* SR = tri_->getSubRegisters(DstReg); *SR; ++SR) {
710 if (!li_->hasInterval(*SR))
712 const LiveRange *DLR =
713 li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
714 if (DLR && DLR->valno->getCopy() == CopyMI)
715 DLR->valno->setCopy(0);
719 // If copy kills the source register, find the last use and propagate
721 bool checkForDeadDef = false;
722 MachineBasicBlock *MBB = CopyMI->getParent();
723 if (SrcLR->valno->isKill(DefIdx))
724 if (!TrimLiveIntervalToLastUse(CopyIdx, MBB, SrcInt, SrcLR)) {
725 checkForDeadDef = true;
728 MachineBasicBlock::iterator MII =
729 llvm::next(MachineBasicBlock::iterator(CopyMI));
730 tii_->reMaterialize(*MBB, MII, DstReg, DstSubIdx, DefMI, *tri_);
731 MachineInstr *NewMI = prior(MII);
733 if (checkForDeadDef) {
734 // PR4090 fix: Trim interval failed because there was no use of the
735 // source interval in this MBB. If the def is in this MBB too then we
736 // should mark it dead:
737 if (DefMI->getParent() == MBB) {
738 DefMI->addRegisterDead(SrcInt.reg, tri_);
739 SrcLR->end = SrcLR->start.getNextSlot();
743 // CopyMI may have implicit operands, transfer them over to the newly
744 // rematerialized instruction. And update implicit def interval valnos.
745 for (unsigned i = CopyMI->getDesc().getNumOperands(),
746 e = CopyMI->getNumOperands(); i != e; ++i) {
747 MachineOperand &MO = CopyMI->getOperand(i);
748 if (MO.isReg() && MO.isImplicit())
749 NewMI->addOperand(MO);
750 if (MO.isDef() && li_->hasInterval(MO.getReg())) {
751 unsigned Reg = MO.getReg();
752 const LiveRange *DLR =
753 li_->getInterval(Reg).getLiveRangeContaining(DefIdx);
754 if (DLR && DLR->valno->getCopy() == CopyMI)
755 DLR->valno->setCopy(0);
756 // Handle subregs as well
757 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
758 for (const unsigned* SR = tri_->getSubRegisters(Reg); *SR; ++SR) {
759 if (!li_->hasInterval(*SR))
761 const LiveRange *DLR =
762 li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
763 if (DLR && DLR->valno->getCopy() == CopyMI)
764 DLR->valno->setCopy(0);
770 TransferImplicitOps(CopyMI, NewMI);
771 li_->ReplaceMachineInstrInMaps(CopyMI, NewMI);
772 CopyMI->eraseFromParent();
773 ReMatCopies.insert(CopyMI);
774 ReMatDefs.insert(DefMI);
775 DEBUG(dbgs() << "Remat: " << *NewMI);
780 /// UpdateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
781 /// update the subregister number if it is not zero. If DstReg is a
782 /// physical register and the existing subregister number of the def / use
783 /// being updated is not zero, make sure to set it to the correct physical
786 SimpleRegisterCoalescing::UpdateRegDefsUses(unsigned SrcReg, unsigned DstReg,
788 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
789 if (DstIsPhys && SubIdx) {
790 // Figure out the real physical register we are updating with.
791 DstReg = tri_->getSubReg(DstReg, SubIdx);
795 // Collect all the instructions using SrcReg.
796 SmallPtrSet<MachineInstr*, 32> Instrs;
797 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(SrcReg),
798 E = mri_->reg_end(); I != E; ++I)
801 for (SmallPtrSet<MachineInstr*, 32>::const_iterator I = Instrs.begin(),
802 E = Instrs.end(); I != E; ++I) {
803 MachineInstr *UseMI = *I;
805 // A PhysReg copy that won't be coalesced can perhaps be rematerialized
808 unsigned CopySrcReg, CopyDstReg, CopySrcSubIdx, CopyDstSubIdx;
809 if (tii_->isMoveInstr(*UseMI, CopySrcReg, CopyDstReg,
810 CopySrcSubIdx, CopyDstSubIdx) &&
811 CopySrcSubIdx == 0 && CopyDstSubIdx == 0 &&
812 CopySrcReg != CopyDstReg && CopySrcReg == SrcReg &&
813 CopyDstReg != DstReg && !JoinedCopies.count(UseMI) &&
814 ReMaterializeTrivialDef(li_->getInterval(SrcReg), CopyDstReg, 0,
819 SmallVector<unsigned,8> Ops;
821 tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
822 bool Kills = false, Deads = false;
824 // Replace SrcReg with DstReg in all UseMI operands.
825 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
826 MachineOperand &MO = UseMI->getOperand(Ops[i]);
827 Kills |= MO.isKill();
828 Deads |= MO.isDead();
831 MO.substPhysReg(DstReg, *tri_);
833 MO.substVirtReg(DstReg, SubIdx, *tri_);
836 // This instruction is a copy that will be removed.
837 if (JoinedCopies.count(UseMI))
841 // If UseMI was a simple SrcReg def, make sure we didn't turn it into a
842 // read-modify-write of DstReg.
844 UseMI->addRegisterDead(DstReg, tri_);
845 else if (!Reads && Writes)
846 UseMI->addRegisterDefined(DstReg, tri_);
848 // Kill flags apply to the whole physical register.
849 if (DstIsPhys && Kills)
850 UseMI->addRegisterKilled(DstReg, tri_);
854 dbgs() << "\t\tupdated: ";
855 if (!UseMI->isDebugValue())
856 dbgs() << li_->getInstructionIndex(UseMI) << "\t";
861 // After updating the operand, check if the machine instruction has
862 // become a copy. If so, update its val# information.
863 const TargetInstrDesc &TID = UseMI->getDesc();
864 if (DstIsPhys || TID.getNumDefs() != 1 || TID.getNumOperands() <= 2)
867 unsigned CopySrcReg, CopyDstReg, CopySrcSubIdx, CopyDstSubIdx;
868 if (tii_->isMoveInstr(*UseMI, CopySrcReg, CopyDstReg,
869 CopySrcSubIdx, CopyDstSubIdx) &&
870 CopySrcReg != CopyDstReg &&
871 (TargetRegisterInfo::isVirtualRegister(CopyDstReg) ||
872 allocatableRegs_[CopyDstReg])) {
873 LiveInterval &LI = li_->getInterval(CopyDstReg);
875 li_->getInstructionIndex(UseMI).getDefIndex();
876 if (const LiveRange *DLR = LI.getLiveRangeContaining(DefIdx)) {
877 if (DLR->valno->def == DefIdx)
878 DLR->valno->setCopy(UseMI);
884 /// removeIntervalIfEmpty - Check if the live interval of a physical register
885 /// is empty, if so remove it and also remove the empty intervals of its
886 /// sub-registers. Return true if live interval is removed.
887 static bool removeIntervalIfEmpty(LiveInterval &li, LiveIntervals *li_,
888 const TargetRegisterInfo *tri_) {
890 if (TargetRegisterInfo::isPhysicalRegister(li.reg))
891 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
892 if (!li_->hasInterval(*SR))
894 LiveInterval &sli = li_->getInterval(*SR);
896 li_->removeInterval(*SR);
898 li_->removeInterval(li.reg);
904 /// ShortenDeadCopyLiveRange - Shorten a live range defined by a dead copy.
905 /// Return true if live interval is removed.
906 bool SimpleRegisterCoalescing::ShortenDeadCopyLiveRange(LiveInterval &li,
907 MachineInstr *CopyMI) {
908 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI);
909 LiveInterval::iterator MLR =
910 li.FindLiveRangeContaining(CopyIdx.getDefIndex());
912 return false; // Already removed by ShortenDeadCopySrcLiveRange.
913 SlotIndex RemoveStart = MLR->start;
914 SlotIndex RemoveEnd = MLR->end;
915 SlotIndex DefIdx = CopyIdx.getDefIndex();
916 // Remove the liverange that's defined by this.
917 if (RemoveStart == DefIdx && RemoveEnd == DefIdx.getStoreIndex()) {
918 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
919 return removeIntervalIfEmpty(li, li_, tri_);
924 /// RemoveDeadDef - If a def of a live interval is now determined dead, remove
925 /// the val# it defines. If the live interval becomes empty, remove it as well.
926 bool SimpleRegisterCoalescing::RemoveDeadDef(LiveInterval &li,
927 MachineInstr *DefMI) {
928 SlotIndex DefIdx = li_->getInstructionIndex(DefMI).getDefIndex();
929 LiveInterval::iterator MLR = li.FindLiveRangeContaining(DefIdx);
930 if (DefIdx != MLR->valno->def)
932 li.removeValNo(MLR->valno);
933 return removeIntervalIfEmpty(li, li_, tri_);
936 /// PropagateDeadness - Propagate the dead marker to the instruction which
937 /// defines the val#.
938 static void PropagateDeadness(LiveInterval &li, MachineInstr *CopyMI,
939 SlotIndex &LRStart, LiveIntervals *li_,
940 const TargetRegisterInfo* tri_) {
941 MachineInstr *DefMI =
942 li_->getInstructionFromIndex(LRStart.getDefIndex());
943 if (DefMI && DefMI != CopyMI) {
944 int DeadIdx = DefMI->findRegisterDefOperandIdx(li.reg);
946 DefMI->getOperand(DeadIdx).setIsDead();
948 DefMI->addOperand(MachineOperand::CreateReg(li.reg,
949 /*def*/true, /*implicit*/true, /*kill*/false, /*dead*/true));
950 LRStart = LRStart.getNextSlot();
954 /// ShortenDeadCopySrcLiveRange - Shorten a live range as it's artificially
955 /// extended by a dead copy. Mark the last use (if any) of the val# as kill as
956 /// ends the live range there. If there isn't another use, then this live range
957 /// is dead. Return true if live interval is removed.
959 SimpleRegisterCoalescing::ShortenDeadCopySrcLiveRange(LiveInterval &li,
960 MachineInstr *CopyMI) {
961 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI);
962 if (CopyIdx == SlotIndex()) {
963 // FIXME: special case: function live in. It can be a general case if the
964 // first instruction index starts at > 0 value.
965 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
966 // Live-in to the function but dead. Remove it from entry live-in set.
967 if (mf_->begin()->isLiveIn(li.reg))
968 mf_->begin()->removeLiveIn(li.reg);
969 const LiveRange *LR = li.getLiveRangeContaining(CopyIdx);
970 removeRange(li, LR->start, LR->end, li_, tri_);
971 return removeIntervalIfEmpty(li, li_, tri_);
974 LiveInterval::iterator LR =
975 li.FindLiveRangeContaining(CopyIdx.getPrevIndex().getStoreIndex());
977 // Livein but defined by a phi.
980 SlotIndex RemoveStart = LR->start;
981 SlotIndex RemoveEnd = CopyIdx.getStoreIndex();
982 if (LR->end > RemoveEnd)
983 // More uses past this copy? Nothing to do.
986 // If there is a last use in the same bb, we can't remove the live range.
987 // Shorten the live interval and return.
988 MachineBasicBlock *CopyMBB = CopyMI->getParent();
989 if (TrimLiveIntervalToLastUse(CopyIdx, CopyMBB, li, LR))
992 // There are other kills of the val#. Nothing to do.
993 if (!li.isOnlyLROfValNo(LR))
996 MachineBasicBlock *StartMBB = li_->getMBBFromIndex(RemoveStart);
997 if (!isSameOrFallThroughBB(StartMBB, CopyMBB, tii_))
998 // If the live range starts in another mbb and the copy mbb is not a fall
999 // through mbb, then we can only cut the range from the beginning of the
1001 RemoveStart = li_->getMBBStartIdx(CopyMBB).getNextIndex().getBaseIndex();
1003 if (LR->valno->def == RemoveStart) {
1004 // If the def MI defines the val# and this copy is the only kill of the
1005 // val#, then propagate the dead marker.
1006 PropagateDeadness(li, CopyMI, RemoveStart, li_, tri_);
1009 if (LR->valno->isKill(RemoveEnd))
1010 LR->valno->removeKill(RemoveEnd);
1013 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
1014 return removeIntervalIfEmpty(li, li_, tri_);
1017 /// CanCoalesceWithImpDef - Returns true if the specified copy instruction
1018 /// from an implicit def to another register can be coalesced away.
1019 bool SimpleRegisterCoalescing::CanCoalesceWithImpDef(MachineInstr *CopyMI,
1021 LiveInterval &ImpLi) const{
1022 if (!CopyMI->killsRegister(ImpLi.reg))
1024 // Make sure this is the only use.
1025 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(ImpLi.reg),
1026 UE = mri_->use_end(); UI != UE;) {
1027 MachineInstr *UseMI = &*UI;
1029 if (CopyMI == UseMI || JoinedCopies.count(UseMI))
1037 /// isWinToJoinVRWithSrcPhysReg - Return true if it's worth while to join a
1038 /// a virtual destination register with physical source register.
1040 SimpleRegisterCoalescing::isWinToJoinVRWithSrcPhysReg(MachineInstr *CopyMI,
1041 MachineBasicBlock *CopyMBB,
1042 LiveInterval &DstInt,
1043 LiveInterval &SrcInt) {
1044 // If the virtual register live interval is long but it has low use desity,
1045 // do not join them, instead mark the physical register as its allocation
1047 const TargetRegisterClass *RC = mri_->getRegClass(DstInt.reg);
1048 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1049 unsigned Length = li_->getApproximateInstructionCount(DstInt);
1050 if (Length > Threshold &&
1051 std::distance(mri_->use_nodbg_begin(DstInt.reg),
1052 mri_->use_nodbg_end()) * Threshold < Length)
1055 // If the virtual register live interval extends into a loop, turn down
1058 li_->getInstructionIndex(CopyMI).getDefIndex();
1059 const MachineLoop *L = loopInfo->getLoopFor(CopyMBB);
1061 // Let's see if the virtual register live interval extends into the loop.
1062 LiveInterval::iterator DLR = DstInt.FindLiveRangeContaining(CopyIdx);
1063 assert(DLR != DstInt.end() && "Live range not found!");
1064 DLR = DstInt.FindLiveRangeContaining(DLR->end.getNextSlot());
1065 if (DLR != DstInt.end()) {
1066 CopyMBB = li_->getMBBFromIndex(DLR->start);
1067 L = loopInfo->getLoopFor(CopyMBB);
1071 if (!L || Length <= Threshold)
1074 SlotIndex UseIdx = CopyIdx.getUseIndex();
1075 LiveInterval::iterator SLR = SrcInt.FindLiveRangeContaining(UseIdx);
1076 MachineBasicBlock *SMBB = li_->getMBBFromIndex(SLR->start);
1077 if (loopInfo->getLoopFor(SMBB) != L) {
1078 if (!loopInfo->isLoopHeader(CopyMBB))
1080 // If vr's live interval extends pass the loop header, do not join.
1081 for (MachineBasicBlock::succ_iterator SI = CopyMBB->succ_begin(),
1082 SE = CopyMBB->succ_end(); SI != SE; ++SI) {
1083 MachineBasicBlock *SuccMBB = *SI;
1084 if (SuccMBB == CopyMBB)
1086 if (DstInt.overlaps(li_->getMBBStartIdx(SuccMBB),
1087 li_->getMBBEndIdx(SuccMBB)))
1094 /// isWinToJoinVRWithDstPhysReg - Return true if it's worth while to join a
1095 /// copy from a virtual source register to a physical destination register.
1097 SimpleRegisterCoalescing::isWinToJoinVRWithDstPhysReg(MachineInstr *CopyMI,
1098 MachineBasicBlock *CopyMBB,
1099 LiveInterval &DstInt,
1100 LiveInterval &SrcInt) {
1101 // If the virtual register live interval is long but it has low use density,
1102 // do not join them, instead mark the physical register as its allocation
1104 const TargetRegisterClass *RC = mri_->getRegClass(SrcInt.reg);
1105 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1106 unsigned Length = li_->getApproximateInstructionCount(SrcInt);
1107 if (Length > Threshold &&
1108 std::distance(mri_->use_nodbg_begin(SrcInt.reg),
1109 mri_->use_nodbg_end()) * Threshold < Length)
1113 // Must be implicit_def.
1116 // If the virtual register live interval is defined or cross a loop, turn
1117 // down aggressiveness.
1119 li_->getInstructionIndex(CopyMI).getDefIndex();
1120 SlotIndex UseIdx = CopyIdx.getUseIndex();
1121 LiveInterval::iterator SLR = SrcInt.FindLiveRangeContaining(UseIdx);
1122 assert(SLR != SrcInt.end() && "Live range not found!");
1123 SLR = SrcInt.FindLiveRangeContaining(SLR->start.getPrevSlot());
1124 if (SLR == SrcInt.end())
1126 MachineBasicBlock *SMBB = li_->getMBBFromIndex(SLR->start);
1127 const MachineLoop *L = loopInfo->getLoopFor(SMBB);
1129 if (!L || Length <= Threshold)
1132 if (loopInfo->getLoopFor(CopyMBB) != L) {
1133 if (SMBB != L->getLoopLatch())
1135 // If vr's live interval is extended from before the loop latch, do not
1137 for (MachineBasicBlock::pred_iterator PI = SMBB->pred_begin(),
1138 PE = SMBB->pred_end(); PI != PE; ++PI) {
1139 MachineBasicBlock *PredMBB = *PI;
1140 if (PredMBB == SMBB)
1142 if (SrcInt.overlaps(li_->getMBBStartIdx(PredMBB),
1143 li_->getMBBEndIdx(PredMBB)))
1150 /// isWinToJoinCrossClass - Return true if it's profitable to coalesce
1151 /// two virtual registers from different register classes.
1153 SimpleRegisterCoalescing::isWinToJoinCrossClass(unsigned SrcReg,
1155 const TargetRegisterClass *SrcRC,
1156 const TargetRegisterClass *DstRC,
1157 const TargetRegisterClass *NewRC) {
1158 unsigned NewRCCount = allocatableRCRegs_[NewRC].count();
1159 // This heuristics is good enough in practice, but it's obviously not *right*.
1160 // 4 is a magic number that works well enough for x86, ARM, etc. It filter
1161 // out all but the most restrictive register classes.
1162 if (NewRCCount > 4 ||
1163 // Early exit if the function is fairly small, coalesce aggressively if
1164 // that's the case. For really special register classes with 3 or
1165 // fewer registers, be a bit more careful.
1166 (li_->getFuncInstructionCount() / NewRCCount) < 8)
1168 LiveInterval &SrcInt = li_->getInterval(SrcReg);
1169 LiveInterval &DstInt = li_->getInterval(DstReg);
1170 unsigned SrcSize = li_->getApproximateInstructionCount(SrcInt);
1171 unsigned DstSize = li_->getApproximateInstructionCount(DstInt);
1172 if (SrcSize <= NewRCCount && DstSize <= NewRCCount)
1174 // Estimate *register use density*. If it doubles or more, abort.
1175 unsigned SrcUses = std::distance(mri_->use_nodbg_begin(SrcReg),
1176 mri_->use_nodbg_end());
1177 unsigned DstUses = std::distance(mri_->use_nodbg_begin(DstReg),
1178 mri_->use_nodbg_end());
1179 unsigned NewUses = SrcUses + DstUses;
1180 unsigned NewSize = SrcSize + DstSize;
1181 if (SrcRC != NewRC && SrcSize > NewRCCount) {
1182 unsigned SrcRCCount = allocatableRCRegs_[SrcRC].count();
1183 if (NewUses*SrcSize*SrcRCCount > 2*SrcUses*NewSize*NewRCCount)
1186 if (DstRC != NewRC && DstSize > NewRCCount) {
1187 unsigned DstRCCount = allocatableRCRegs_[DstRC].count();
1188 if (NewUses*DstSize*DstRCCount > 2*DstUses*NewSize*NewRCCount)
1194 /// HasIncompatibleSubRegDefUse - If we are trying to coalesce a virtual
1195 /// register with a physical register, check if any of the virtual register
1196 /// operand is a sub-register use or def. If so, make sure it won't result
1197 /// in an illegal extract_subreg or insert_subreg instruction. e.g.
1198 /// vr1024 = extract_subreg vr1025, 1
1200 /// vr1024 = mov8rr AH
1201 /// If vr1024 is coalesced with AH, the extract_subreg is now illegal since
1202 /// AH does not have a super-reg whose sub-register 1 is AH.
1204 SimpleRegisterCoalescing::HasIncompatibleSubRegDefUse(MachineInstr *CopyMI,
1207 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(VirtReg),
1208 E = mri_->reg_end(); I != E; ++I) {
1209 MachineOperand &O = I.getOperand();
1212 MachineInstr *MI = &*I;
1213 if (MI == CopyMI || JoinedCopies.count(MI))
1215 unsigned SubIdx = O.getSubReg();
1216 if (SubIdx && !tri_->getSubReg(PhysReg, SubIdx))
1218 if (MI->isExtractSubreg()) {
1219 SubIdx = MI->getOperand(2).getImm();
1220 if (O.isUse() && !tri_->getSubReg(PhysReg, SubIdx))
1223 unsigned SrcReg = MI->getOperand(1).getReg();
1224 const TargetRegisterClass *RC =
1225 TargetRegisterInfo::isPhysicalRegister(SrcReg)
1226 ? tri_->getPhysicalRegisterRegClass(SrcReg)
1227 : mri_->getRegClass(SrcReg);
1228 if (!tri_->getMatchingSuperReg(PhysReg, SubIdx, RC))
1232 if (MI->isInsertSubreg() || MI->isSubregToReg()) {
1233 SubIdx = MI->getOperand(3).getImm();
1234 if (VirtReg == MI->getOperand(0).getReg()) {
1235 if (!tri_->getSubReg(PhysReg, SubIdx))
1238 unsigned DstReg = MI->getOperand(0).getReg();
1239 const TargetRegisterClass *RC =
1240 TargetRegisterInfo::isPhysicalRegister(DstReg)
1241 ? tri_->getPhysicalRegisterRegClass(DstReg)
1242 : mri_->getRegClass(DstReg);
1243 if (!tri_->getMatchingSuperReg(PhysReg, SubIdx, RC))
1252 /// CanJoinExtractSubRegToPhysReg - Return true if it's possible to coalesce
1253 /// an extract_subreg where dst is a physical register, e.g.
1254 /// cl = EXTRACT_SUBREG reg1024, 1
1256 SimpleRegisterCoalescing::CanJoinExtractSubRegToPhysReg(unsigned DstReg,
1257 unsigned SrcReg, unsigned SubIdx,
1258 unsigned &RealDstReg) {
1259 const TargetRegisterClass *RC = mri_->getRegClass(SrcReg);
1260 RealDstReg = tri_->getMatchingSuperReg(DstReg, SubIdx, RC);
1262 DEBUG(dbgs() << "\tIncompatible source regclass: "
1263 << "none of the super-registers of " << tri_->getName(DstReg)
1264 << " are in " << RC->getName() << ".\n");
1268 LiveInterval &RHS = li_->getInterval(SrcReg);
1269 // For this type of EXTRACT_SUBREG, conservatively
1270 // check if the live interval of the source register interfere with the
1271 // actual super physical register we are trying to coalesce with.
1272 if (li_->hasInterval(RealDstReg) &&
1273 RHS.overlaps(li_->getInterval(RealDstReg))) {
1275 dbgs() << "\t\tInterfere with register ";
1276 li_->getInterval(RealDstReg).print(dbgs(), tri_);
1278 return false; // Not coalescable
1280 for (const unsigned* SR = tri_->getSubRegisters(RealDstReg); *SR; ++SR)
1281 // Do not check DstReg or its sub-register. JoinIntervals() will take care
1283 if (*SR != DstReg &&
1284 !tri_->isSubRegister(DstReg, *SR) &&
1285 li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
1287 dbgs() << "\t\tInterfere with sub-register ";
1288 li_->getInterval(*SR).print(dbgs(), tri_);
1290 return false; // Not coalescable
1295 /// CanJoinInsertSubRegToPhysReg - Return true if it's possible to coalesce
1296 /// an insert_subreg where src is a physical register, e.g.
1297 /// reg1024 = INSERT_SUBREG reg1024, c1, 0
1299 SimpleRegisterCoalescing::CanJoinInsertSubRegToPhysReg(unsigned DstReg,
1300 unsigned SrcReg, unsigned SubIdx,
1301 unsigned &RealSrcReg) {
1302 const TargetRegisterClass *RC = mri_->getRegClass(DstReg);
1303 RealSrcReg = tri_->getMatchingSuperReg(SrcReg, SubIdx, RC);
1305 DEBUG(dbgs() << "\tIncompatible destination regclass: "
1306 << "none of the super-registers of " << tri_->getName(SrcReg)
1307 << " are in " << RC->getName() << ".\n");
1311 LiveInterval &LHS = li_->getInterval(DstReg);
1312 if (li_->hasInterval(RealSrcReg) &&
1313 LHS.overlaps(li_->getInterval(RealSrcReg))) {
1315 dbgs() << "\t\tInterfere with register ";
1316 li_->getInterval(RealSrcReg).print(dbgs(), tri_);
1318 return false; // Not coalescable
1320 for (const unsigned* SR = tri_->getSubRegisters(RealSrcReg); *SR; ++SR)
1321 // Do not check SrcReg or its sub-register. JoinIntervals() will take care
1323 if (*SR != SrcReg &&
1324 !tri_->isSubRegister(SrcReg, *SR) &&
1325 li_->hasInterval(*SR) && LHS.overlaps(li_->getInterval(*SR))) {
1327 dbgs() << "\t\tInterfere with sub-register ";
1328 li_->getInterval(*SR).print(dbgs(), tri_);
1330 return false; // Not coalescable
1335 /// getRegAllocPreference - Return register allocation preference register.
1337 static unsigned getRegAllocPreference(unsigned Reg, MachineFunction &MF,
1338 MachineRegisterInfo *MRI,
1339 const TargetRegisterInfo *TRI) {
1340 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1342 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
1343 return TRI->ResolveRegAllocHint(Hint.first, Hint.second, MF);
1346 /// JoinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
1347 /// which are the src/dst of the copy instruction CopyMI. This returns true
1348 /// if the copy was successfully coalesced away. If it is not currently
1349 /// possible to coalesce this interval, but it may be possible if other
1350 /// things get coalesced, then it returns true by reference in 'Again'.
1351 bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) {
1352 MachineInstr *CopyMI = TheCopy.MI;
1355 if (JoinedCopies.count(CopyMI) || ReMatCopies.count(CopyMI))
1356 return false; // Already done.
1358 DEBUG(dbgs() << li_->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
1360 unsigned SrcReg, DstReg, SrcSubIdx = 0, DstSubIdx = 0;
1361 bool isExtSubReg = CopyMI->isExtractSubreg();
1362 bool isInsSubReg = CopyMI->isInsertSubreg();
1363 bool isSubRegToReg = CopyMI->isSubregToReg();
1364 unsigned SubIdx = 0;
1366 DstReg = CopyMI->getOperand(0).getReg();
1367 DstSubIdx = CopyMI->getOperand(0).getSubReg();
1368 SrcReg = CopyMI->getOperand(1).getReg();
1369 SrcSubIdx = CopyMI->getOperand(2).getImm();
1370 } else if (isInsSubReg || isSubRegToReg) {
1371 DstReg = CopyMI->getOperand(0).getReg();
1372 DstSubIdx = CopyMI->getOperand(3).getImm();
1373 SrcReg = CopyMI->getOperand(2).getReg();
1374 SrcSubIdx = CopyMI->getOperand(2).getSubReg();
1375 if (SrcSubIdx && SrcSubIdx != DstSubIdx) {
1376 // r1025 = INSERT_SUBREG r1025, r1024<2>, 2 Then r1024 has already been
1377 // coalesced to a larger register so the subreg indices cancel out.
1378 DEBUG(dbgs() << "\tSource of insert_subreg or subreg_to_reg is already "
1379 "coalesced to another register.\n");
1380 return false; // Not coalescable.
1382 } else if (tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
1383 if (SrcSubIdx && DstSubIdx && SrcSubIdx != DstSubIdx) {
1384 // e.g. %reg16404:1<def> = MOV8rr %reg16412:2<kill>
1386 return false; // Not coalescable.
1389 llvm_unreachable("Unrecognized copy instruction!");
1392 // If they are already joined we continue.
1393 if (SrcReg == DstReg) {
1394 DEBUG(dbgs() << "\tCopy already coalesced.\n");
1395 return false; // Not coalescable.
1398 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
1399 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1401 // If they are both physical registers, we cannot join them.
1402 if (SrcIsPhys && DstIsPhys) {
1403 DEBUG(dbgs() << "\tCan not coalesce physregs.\n");
1404 return false; // Not coalescable.
1407 // We only join virtual registers with allocatable physical registers.
1408 if (SrcIsPhys && !allocatableRegs_[SrcReg]) {
1409 DEBUG(dbgs() << "\tSrc reg is unallocatable physreg.\n");
1410 return false; // Not coalescable.
1412 if (DstIsPhys && !allocatableRegs_[DstReg]) {
1413 DEBUG(dbgs() << "\tDst reg is unallocatable physreg.\n");
1414 return false; // Not coalescable.
1417 // We cannot handle dual subreg indices and mismatched classes at the same
1419 if (SrcSubIdx && DstSubIdx && differingRegisterClasses(SrcReg, DstReg)) {
1420 DEBUG(dbgs() << "\tCannot handle subreg indices and mismatched classes.\n");
1424 // Check that a physical source register is compatible with dst regclass
1426 unsigned SrcSubReg = SrcSubIdx ?
1427 tri_->getSubReg(SrcReg, SrcSubIdx) : SrcReg;
1428 const TargetRegisterClass *DstRC = mri_->getRegClass(DstReg);
1429 const TargetRegisterClass *DstSubRC = DstRC;
1431 DstSubRC = DstRC->getSubRegisterRegClass(DstSubIdx);
1432 assert(DstSubRC && "Illegal subregister index");
1433 if (!DstSubRC->contains(SrcSubReg)) {
1434 DEBUG(dbgs() << "\tIncompatible destination regclass: "
1435 << "none of the super-registers of "
1436 << tri_->getName(SrcSubReg) << " are in "
1437 << DstSubRC->getName() << ".\n");
1438 return false; // Not coalescable.
1442 // Check that a physical dst register is compatible with source regclass
1444 unsigned DstSubReg = DstSubIdx ?
1445 tri_->getSubReg(DstReg, DstSubIdx) : DstReg;
1446 const TargetRegisterClass *SrcRC = mri_->getRegClass(SrcReg);
1447 const TargetRegisterClass *SrcSubRC = SrcRC;
1449 SrcSubRC = SrcRC->getSubRegisterRegClass(SrcSubIdx);
1450 assert(SrcSubRC && "Illegal subregister index");
1451 if (!SrcSubRC->contains(DstSubReg)) {
1452 DEBUG(dbgs() << "\tIncompatible source regclass: "
1453 << "none of the super-registers of "
1454 << tri_->getName(DstSubReg) << " are in "
1455 << SrcSubRC->getName() << ".\n");
1457 return false; // Not coalescable.
1461 // Should be non-null only when coalescing to a sub-register class.
1462 bool CrossRC = false;
1463 const TargetRegisterClass *SrcRC= SrcIsPhys ? 0 : mri_->getRegClass(SrcReg);
1464 const TargetRegisterClass *DstRC= DstIsPhys ? 0 : mri_->getRegClass(DstReg);
1465 const TargetRegisterClass *NewRC = NULL;
1466 unsigned RealDstReg = 0;
1467 unsigned RealSrcReg = 0;
1468 if (isExtSubReg || isInsSubReg || isSubRegToReg) {
1469 SubIdx = CopyMI->getOperand(isExtSubReg ? 2 : 3).getImm();
1470 if (SrcIsPhys && isExtSubReg) {
1471 // r1024 = EXTRACT_SUBREG EAX, 0 then r1024 is really going to be
1472 // coalesced with AX.
1473 unsigned DstSubIdx = CopyMI->getOperand(0).getSubReg();
1475 // r1024<2> = EXTRACT_SUBREG EAX, 2. Then r1024 has already been
1476 // coalesced to a larger register so the subreg indices cancel out.
1477 if (DstSubIdx != SubIdx) {
1478 DEBUG(dbgs() << "\t Sub-register indices mismatch.\n");
1479 return false; // Not coalescable.
1482 SrcReg = tri_->getSubReg(SrcReg, SubIdx);
1484 } else if (DstIsPhys && (isInsSubReg || isSubRegToReg)) {
1485 // EAX = INSERT_SUBREG EAX, r1024, 0
1486 unsigned SrcSubIdx = CopyMI->getOperand(2).getSubReg();
1488 // EAX = INSERT_SUBREG EAX, r1024<2>, 2 Then r1024 has already been
1489 // coalesced to a larger register so the subreg indices cancel out.
1490 if (SrcSubIdx != SubIdx) {
1491 DEBUG(dbgs() << "\t Sub-register indices mismatch.\n");
1492 return false; // Not coalescable.
1495 DstReg = tri_->getSubReg(DstReg, SubIdx);
1497 } else if ((DstIsPhys && isExtSubReg) ||
1498 (SrcIsPhys && (isInsSubReg || isSubRegToReg))) {
1499 if (!isSubRegToReg && CopyMI->getOperand(1).getSubReg()) {
1500 DEBUG(dbgs() << "\tSrc of extract_subreg already coalesced with reg"
1501 << " of a super-class.\n");
1502 return false; // Not coalescable.
1505 // FIXME: The following checks are somewhat conservative. Perhaps a better
1506 // way to implement this is to treat this as coalescing a vr with the
1507 // super physical register.
1509 if (!CanJoinExtractSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealDstReg))
1510 return false; // Not coalescable
1512 if (!CanJoinInsertSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealSrcReg))
1513 return false; // Not coalescable
1517 unsigned OldSubIdx = isExtSubReg ? CopyMI->getOperand(0).getSubReg()
1518 : CopyMI->getOperand(2).getSubReg();
1520 if (OldSubIdx == SubIdx && !differingRegisterClasses(SrcReg, DstReg))
1521 // r1024<2> = EXTRACT_SUBREG r1025, 2. Then r1024 has already been
1522 // coalesced to a larger register so the subreg indices cancel out.
1523 // Also check if the other larger register is of the same register
1524 // class as the would be resulting register.
1527 DEBUG(dbgs() << "\t Sub-register indices mismatch.\n");
1528 return false; // Not coalescable.
1532 if (!DstIsPhys && !SrcIsPhys) {
1533 if (isInsSubReg || isSubRegToReg) {
1534 NewRC = tri_->getMatchingSuperRegClass(DstRC, SrcRC, SubIdx);
1535 } else // extract_subreg {
1536 NewRC = tri_->getMatchingSuperRegClass(SrcRC, DstRC, SubIdx);
1539 DEBUG(dbgs() << "\t Conflicting sub-register indices.\n");
1540 return false; // Not coalescable
1543 if (!isWinToJoinCrossClass(SrcReg, DstReg, SrcRC, DstRC, NewRC)) {
1544 DEBUG(dbgs() << "\tAvoid coalescing to constrained register class: "
1545 << SrcRC->getName() << "/"
1546 << DstRC->getName() << " -> "
1547 << NewRC->getName() << ".\n");
1548 Again = true; // May be possible to coalesce later.
1553 } else if (differingRegisterClasses(SrcReg, DstReg)) {
1554 if (DisableCrossClassJoin)
1558 // FIXME: What if the result of a EXTRACT_SUBREG is then coalesced
1559 // with another? If it's the resulting destination register, then
1560 // the subidx must be propagated to uses (but only those defined
1561 // by the EXTRACT_SUBREG). If it's being coalesced into another
1562 // register, it should be safe because register is assumed to have
1563 // the register class of the super-register.
1565 // Process moves where one of the registers have a sub-register index.
1566 MachineOperand *DstMO = CopyMI->findRegisterDefOperand(DstReg);
1567 MachineOperand *SrcMO = CopyMI->findRegisterUseOperand(SrcReg);
1568 SubIdx = DstMO->getSubReg();
1570 if (SrcMO->getSubReg())
1571 // FIXME: can we handle this?
1573 // This is not an insert_subreg but it looks like one.
1574 // e.g. %reg1024:4 = MOV32rr %EAX
1577 if (!CanJoinInsertSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealSrcReg))
1578 return false; // Not coalescable
1582 SubIdx = SrcMO->getSubReg();
1584 // This is not a extract_subreg but it looks like one.
1585 // e.g. %cl = MOV16rr %reg1024:1
1588 if (!CanJoinExtractSubRegToPhysReg(DstReg, SrcReg, SubIdx,RealDstReg))
1589 return false; // Not coalescable
1595 // Now determine the register class of the joined register.
1596 if (!SrcIsPhys && !DstIsPhys) {
1599 SubIdx ? tri_->getMatchingSuperRegClass(SrcRC, DstRC, SubIdx) : SrcRC;
1600 } else if (isInsSubReg) {
1602 SubIdx ? tri_->getMatchingSuperRegClass(DstRC, SrcRC, SubIdx) : DstRC;
1604 NewRC = getCommonSubClass(SrcRC, DstRC);
1608 DEBUG(dbgs() << "\tDisjoint regclasses: "
1609 << SrcRC->getName() << ", "
1610 << DstRC->getName() << ".\n");
1611 return false; // Not coalescable.
1614 // If we are joining two virtual registers and the resulting register
1615 // class is more restrictive (fewer register, smaller size). Check if it's
1616 // worth doing the merge.
1617 if (!isWinToJoinCrossClass(SrcReg, DstReg, SrcRC, DstRC, NewRC)) {
1618 DEBUG(dbgs() << "\tAvoid coalescing to constrained register class: "
1619 << SrcRC->getName() << "/"
1620 << DstRC->getName() << " -> "
1621 << NewRC->getName() << ".\n");
1622 // Allow the coalescer to try again in case either side gets coalesced to
1623 // a physical register that's compatible with the other side. e.g.
1624 // r1024 = MOV32to32_ r1025
1625 // But later r1024 is assigned EAX then r1025 may be coalesced with EAX.
1626 Again = true; // May be possible to coalesce later.
1632 // Will it create illegal extract_subreg / insert_subreg?
1633 if (SrcIsPhys && HasIncompatibleSubRegDefUse(CopyMI, DstReg, SrcReg))
1635 if (DstIsPhys && HasIncompatibleSubRegDefUse(CopyMI, SrcReg, DstReg))
1638 LiveInterval &SrcInt = li_->getInterval(SrcReg);
1639 LiveInterval &DstInt = li_->getInterval(DstReg);
1640 assert(SrcInt.reg == SrcReg && DstInt.reg == DstReg &&
1641 "Register mapping is horribly broken!");
1644 dbgs() << "\t\tInspecting ";
1645 if (SrcRC) dbgs() << SrcRC->getName() << ": ";
1646 SrcInt.print(dbgs(), tri_);
1647 dbgs() << "\n\t\t and ";
1648 if (DstRC) dbgs() << DstRC->getName() << ": ";
1649 DstInt.print(dbgs(), tri_);
1653 // Save a copy of the virtual register live interval. We'll manually
1654 // merge this into the "real" physical register live interval this is
1656 OwningPtr<LiveInterval> SavedLI;
1658 SavedLI.reset(li_->dupInterval(&SrcInt));
1659 else if (RealSrcReg)
1660 SavedLI.reset(li_->dupInterval(&DstInt));
1662 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg) {
1663 // Check if it is necessary to propagate "isDead" property.
1664 MachineOperand *mopd = CopyMI->findRegisterDefOperand(DstReg, false);
1665 bool isDead = mopd->isDead();
1667 // We need to be careful about coalescing a source physical register with a
1668 // virtual register. Once the coalescing is done, it cannot be broken and
1669 // these are not spillable! If the destination interval uses are far away,
1670 // think twice about coalescing them!
1671 if (!isDead && (SrcIsPhys || DstIsPhys)) {
1672 // If the virtual register live interval is long but it has low use
1673 // density, do not join them, instead mark the physical register as its
1674 // allocation preference.
1675 LiveInterval &JoinVInt = SrcIsPhys ? DstInt : SrcInt;
1676 LiveInterval &JoinPInt = SrcIsPhys ? SrcInt : DstInt;
1677 unsigned JoinVReg = SrcIsPhys ? DstReg : SrcReg;
1678 unsigned JoinPReg = SrcIsPhys ? SrcReg : DstReg;
1680 // Don't join with physregs that have a ridiculous number of live
1681 // ranges. The data structure performance is really bad when that
1683 if (JoinPInt.ranges.size() > 1000) {
1684 mri_->setRegAllocationHint(JoinVInt.reg, 0, JoinPReg);
1687 << "\tPhysical register live interval too complicated, abort!\n");
1691 const TargetRegisterClass *RC = mri_->getRegClass(JoinVReg);
1692 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1693 unsigned Length = li_->getApproximateInstructionCount(JoinVInt);
1694 if (Length > Threshold &&
1695 std::distance(mri_->use_nodbg_begin(JoinVReg),
1696 mri_->use_nodbg_end()) * Threshold < Length) {
1697 // Before giving up coalescing, if definition of source is defined by
1698 // trivial computation, try rematerializing it.
1699 if (ReMaterializeTrivialDef(SrcInt, DstReg, DstSubIdx, CopyMI))
1702 mri_->setRegAllocationHint(JoinVInt.reg, 0, JoinPReg);
1704 DEBUG(dbgs() << "\tMay tie down a physical register, abort!\n");
1705 Again = true; // May be possible to coalesce later.
1711 // Okay, attempt to join these two intervals. On failure, this returns false.
1712 // Otherwise, if one of the intervals being joined is a physreg, this method
1713 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1714 // been modified, so we can use this information below to update aliases.
1715 bool Swapped = false;
1716 // If SrcInt is implicitly defined, it's safe to coalesce.
1717 if (SrcInt.empty()) {
1718 if (!CanCoalesceWithImpDef(CopyMI, DstInt, SrcInt)) {
1719 // Only coalesce an empty interval (defined by implicit_def) with
1720 // another interval which has a valno defined by the CopyMI and the CopyMI
1721 // is a kill of the implicit def.
1722 DEBUG(dbgs() << "\tNot profitable!\n");
1725 } else if (!JoinIntervals(DstInt, SrcInt, Swapped)) {
1726 // Coalescing failed.
1728 // If definition of source is defined by trivial computation, try
1729 // rematerializing it.
1730 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg &&
1731 ReMaterializeTrivialDef(SrcInt, DstReg, DstSubIdx, CopyMI))
1734 // If we can eliminate the copy without merging the live ranges, do so now.
1735 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg &&
1736 (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI) ||
1737 RemoveCopyByCommutingDef(SrcInt, DstInt, CopyMI))) {
1738 JoinedCopies.insert(CopyMI);
1739 DEBUG(dbgs() << "\tTrivial!\n");
1743 // Otherwise, we are unable to join the intervals.
1744 DEBUG(dbgs() << "\tInterference!\n");
1745 Again = true; // May be possible to coalesce later.
1749 LiveInterval *ResSrcInt = &SrcInt;
1750 LiveInterval *ResDstInt = &DstInt;
1752 std::swap(SrcReg, DstReg);
1753 std::swap(ResSrcInt, ResDstInt);
1755 assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1756 "LiveInterval::join didn't work right!");
1758 // If we're about to merge live ranges into a physical register live interval,
1759 // we have to update any aliased register's live ranges to indicate that they
1760 // have clobbered values for this range.
1761 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
1762 // If this is a extract_subreg where dst is a physical register, e.g.
1763 // cl = EXTRACT_SUBREG reg1024, 1
1764 // then create and update the actual physical register allocated to RHS.
1765 if (RealDstReg || RealSrcReg) {
1766 LiveInterval &RealInt =
1767 li_->getOrCreateInterval(RealDstReg ? RealDstReg : RealSrcReg);
1768 for (LiveInterval::const_vni_iterator I = SavedLI->vni_begin(),
1769 E = SavedLI->vni_end(); I != E; ++I) {
1770 const VNInfo *ValNo = *I;
1771 VNInfo *NewValNo = RealInt.getNextValue(ValNo->def, ValNo->getCopy(),
1772 false, // updated at *
1773 li_->getVNInfoAllocator());
1774 NewValNo->setFlags(ValNo->getFlags()); // * updated here.
1775 RealInt.addKills(NewValNo, ValNo->kills);
1776 RealInt.MergeValueInAsValue(*SavedLI, ValNo, NewValNo);
1778 RealInt.weight += SavedLI->weight;
1779 DstReg = RealDstReg ? RealDstReg : RealSrcReg;
1782 // Update the liveintervals of sub-registers.
1783 for (const unsigned *AS = tri_->getSubRegisters(DstReg); *AS; ++AS)
1784 li_->getOrCreateInterval(*AS).MergeInClobberRanges(*li_, *ResSrcInt,
1785 li_->getVNInfoAllocator());
1788 // If this is a EXTRACT_SUBREG, make sure the result of coalescing is the
1789 // larger super-register.
1790 if ((isExtSubReg || isInsSubReg || isSubRegToReg) &&
1791 !SrcIsPhys && !DstIsPhys) {
1792 if ((isExtSubReg && !Swapped) ||
1793 ((isInsSubReg || isSubRegToReg) && Swapped)) {
1794 ResSrcInt->Copy(*ResDstInt, mri_, li_->getVNInfoAllocator());
1795 std::swap(SrcReg, DstReg);
1796 std::swap(ResSrcInt, ResDstInt);
1800 // Coalescing to a virtual register that is of a sub-register class of the
1801 // other. Make sure the resulting register is set to the right register class.
1805 // This may happen even if it's cross-rc coalescing. e.g.
1806 // %reg1026<def> = SUBREG_TO_REG 0, %reg1037<kill>, 4
1807 // reg1026 -> GR64, reg1037 -> GR32_ABCD. The resulting register will have to
1808 // be allocate a register from GR64_ABCD.
1810 mri_->setRegClass(DstReg, NewRC);
1812 // Remember to delete the copy instruction.
1813 JoinedCopies.insert(CopyMI);
1815 UpdateRegDefsUses(SrcReg, DstReg, SubIdx);
1817 // If we have extended the live range of a physical register, make sure we
1818 // update live-in lists as well.
1819 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
1820 const LiveInterval &VRegInterval = li_->getInterval(SrcReg);
1821 SmallVector<MachineBasicBlock*, 16> BlockSeq;
1822 for (LiveInterval::const_iterator I = VRegInterval.begin(),
1823 E = VRegInterval.end(); I != E; ++I ) {
1824 li_->findLiveInMBBs(I->start, I->end, BlockSeq);
1825 for (unsigned idx = 0, size = BlockSeq.size(); idx != size; ++idx) {
1826 MachineBasicBlock &block = *BlockSeq[idx];
1827 if (!block.isLiveIn(DstReg))
1828 block.addLiveIn(DstReg);
1834 // SrcReg is guarateed to be the register whose live interval that is
1836 li_->removeInterval(SrcReg);
1838 // Update regalloc hint.
1839 tri_->UpdateRegAllocHint(SrcReg, DstReg, *mf_);
1841 // Manually deleted the live interval copy.
1847 // If resulting interval has a preference that no longer fits because of subreg
1848 // coalescing, just clear the preference.
1849 unsigned Preference = getRegAllocPreference(ResDstInt->reg, *mf_, mri_, tri_);
1850 if (Preference && (isExtSubReg || isInsSubReg || isSubRegToReg) &&
1851 TargetRegisterInfo::isVirtualRegister(ResDstInt->reg)) {
1852 const TargetRegisterClass *RC = mri_->getRegClass(ResDstInt->reg);
1853 if (!RC->contains(Preference))
1854 mri_->setRegAllocationHint(ResDstInt->reg, 0, 0);
1858 dbgs() << "\t\tJoined. Result = ";
1859 ResDstInt->print(dbgs(), tri_);
1867 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
1868 /// compute what the resultant value numbers for each value in the input two
1869 /// ranges will be. This is complicated by copies between the two which can
1870 /// and will commonly cause multiple value numbers to be merged into one.
1872 /// VN is the value number that we're trying to resolve. InstDefiningValue
1873 /// keeps track of the new InstDefiningValue assignment for the result
1874 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
1875 /// whether a value in this or other is a copy from the opposite set.
1876 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
1877 /// already been assigned.
1879 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
1880 /// contains the value number the copy is from.
1882 static unsigned ComputeUltimateVN(VNInfo *VNI,
1883 SmallVector<VNInfo*, 16> &NewVNInfo,
1884 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
1885 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
1886 SmallVector<int, 16> &ThisValNoAssignments,
1887 SmallVector<int, 16> &OtherValNoAssignments) {
1888 unsigned VN = VNI->id;
1890 // If the VN has already been computed, just return it.
1891 if (ThisValNoAssignments[VN] >= 0)
1892 return ThisValNoAssignments[VN];
1893 assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers");
1895 // If this val is not a copy from the other val, then it must be a new value
1896 // number in the destination.
1897 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
1898 if (I == ThisFromOther.end()) {
1899 NewVNInfo.push_back(VNI);
1900 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
1902 VNInfo *OtherValNo = I->second;
1904 // Otherwise, this *is* a copy from the RHS. If the other side has already
1905 // been computed, return it.
1906 if (OtherValNoAssignments[OtherValNo->id] >= 0)
1907 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
1909 // Mark this value number as currently being computed, then ask what the
1910 // ultimate value # of the other value is.
1911 ThisValNoAssignments[VN] = -2;
1912 unsigned UltimateVN =
1913 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
1914 OtherValNoAssignments, ThisValNoAssignments);
1915 return ThisValNoAssignments[VN] = UltimateVN;
1918 static bool InVector(VNInfo *Val, const SmallVector<VNInfo*, 8> &V) {
1919 return std::find(V.begin(), V.end(), Val) != V.end();
1922 static bool isValNoDefMove(const MachineInstr *MI, unsigned DR, unsigned SR,
1923 const TargetInstrInfo *TII,
1924 const TargetRegisterInfo *TRI) {
1925 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
1926 if (TII->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
1928 else if (MI->isExtractSubreg()) {
1929 DstReg = MI->getOperand(0).getReg();
1930 SrcReg = MI->getOperand(1).getReg();
1931 } else if (MI->isSubregToReg() ||
1932 MI->isInsertSubreg()) {
1933 DstReg = MI->getOperand(0).getReg();
1934 SrcReg = MI->getOperand(2).getReg();
1937 return (SrcReg == SR || TRI->isSuperRegister(SR, SrcReg)) &&
1938 (DstReg == DR || TRI->isSuperRegister(DR, DstReg));
1941 /// RangeIsDefinedByCopyFromReg - Return true if the specified live range of
1942 /// the specified live interval is defined by a copy from the specified
1944 bool SimpleRegisterCoalescing::RangeIsDefinedByCopyFromReg(LiveInterval &li,
1947 unsigned SrcReg = li_->getVNInfoSourceReg(LR->valno);
1950 // FIXME: Do isPHIDef and isDefAccurate both need to be tested?
1951 if ((LR->valno->isPHIDef() || !LR->valno->isDefAccurate()) &&
1952 TargetRegisterInfo::isPhysicalRegister(li.reg) &&
1953 *tri_->getSuperRegisters(li.reg)) {
1954 // It's a sub-register live interval, we may not have precise information.
1956 MachineInstr *DefMI = li_->getInstructionFromIndex(LR->start);
1957 if (DefMI && isValNoDefMove(DefMI, li.reg, Reg, tii_, tri_)) {
1958 // Cache computed info.
1959 LR->valno->def = LR->start;
1960 LR->valno->setCopy(DefMI);
1968 /// ValueLiveAt - Return true if the LiveRange pointed to by the given
1969 /// iterator, or any subsequent range with the same value number,
1970 /// is live at the given point.
1971 bool SimpleRegisterCoalescing::ValueLiveAt(LiveInterval::iterator LRItr,
1972 LiveInterval::iterator LREnd,
1973 SlotIndex defPoint) const {
1974 for (const VNInfo *valno = LRItr->valno;
1975 (LRItr != LREnd) && (LRItr->valno == valno); ++LRItr) {
1976 if (LRItr->contains(defPoint))
1984 /// SimpleJoin - Attempt to joint the specified interval into this one. The
1985 /// caller of this method must guarantee that the RHS only contains a single
1986 /// value number and that the RHS is not defined by a copy from this
1987 /// interval. This returns false if the intervals are not joinable, or it
1988 /// joins them and returns true.
1989 bool SimpleRegisterCoalescing::SimpleJoin(LiveInterval &LHS, LiveInterval &RHS){
1990 assert(RHS.containsOneValue());
1992 // Some number (potentially more than one) value numbers in the current
1993 // interval may be defined as copies from the RHS. Scan the overlapping
1994 // portions of the LHS and RHS, keeping track of this and looking for
1995 // overlapping live ranges that are NOT defined as copies. If these exist, we
1998 LiveInterval::iterator LHSIt = LHS.begin(), LHSEnd = LHS.end();
1999 LiveInterval::iterator RHSIt = RHS.begin(), RHSEnd = RHS.end();
2001 if (LHSIt->start < RHSIt->start) {
2002 LHSIt = std::upper_bound(LHSIt, LHSEnd, RHSIt->start);
2003 if (LHSIt != LHS.begin()) --LHSIt;
2004 } else if (RHSIt->start < LHSIt->start) {
2005 RHSIt = std::upper_bound(RHSIt, RHSEnd, LHSIt->start);
2006 if (RHSIt != RHS.begin()) --RHSIt;
2009 SmallVector<VNInfo*, 8> EliminatedLHSVals;
2012 // Determine if these live intervals overlap.
2013 bool Overlaps = false;
2014 if (LHSIt->start <= RHSIt->start)
2015 Overlaps = LHSIt->end > RHSIt->start;
2017 Overlaps = RHSIt->end > LHSIt->start;
2019 // If the live intervals overlap, there are two interesting cases: if the
2020 // LHS interval is defined by a copy from the RHS, it's ok and we record
2021 // that the LHS value # is the same as the RHS. If it's not, then we cannot
2022 // coalesce these live ranges and we bail out.
2024 // If we haven't already recorded that this value # is safe, check it.
2025 if (!InVector(LHSIt->valno, EliminatedLHSVals)) {
2026 // If it's re-defined by an early clobber somewhere in the live range,
2027 // then conservatively abort coalescing.
2028 if (LHSIt->valno->hasRedefByEC())
2030 // Copy from the RHS?
2031 if (!RangeIsDefinedByCopyFromReg(LHS, LHSIt, RHS.reg))
2032 return false; // Nope, bail out.
2034 if (ValueLiveAt(LHSIt, LHS.end(), RHSIt->valno->def))
2035 // Here is an interesting situation:
2037 // vr1025 = copy vr1024
2042 // Even though vr1025 is copied from vr1024, it's not safe to
2043 // coalesce them since the live range of vr1025 intersects the
2044 // def of vr1024. This happens because vr1025 is assigned the
2045 // value of the previous iteration of vr1024.
2047 EliminatedLHSVals.push_back(LHSIt->valno);
2050 // We know this entire LHS live range is okay, so skip it now.
2051 if (++LHSIt == LHSEnd) break;
2055 if (LHSIt->end < RHSIt->end) {
2056 if (++LHSIt == LHSEnd) break;
2058 // One interesting case to check here. It's possible that we have
2059 // something like "X3 = Y" which defines a new value number in the LHS,
2060 // and is the last use of this liverange of the RHS. In this case, we
2061 // want to notice this copy (so that it gets coalesced away) even though
2062 // the live ranges don't actually overlap.
2063 if (LHSIt->start == RHSIt->end) {
2064 if (InVector(LHSIt->valno, EliminatedLHSVals)) {
2065 // We already know that this value number is going to be merged in
2066 // if coalescing succeeds. Just skip the liverange.
2067 if (++LHSIt == LHSEnd) break;
2069 // If it's re-defined by an early clobber somewhere in the live range,
2070 // then conservatively abort coalescing.
2071 if (LHSIt->valno->hasRedefByEC())
2073 // Otherwise, if this is a copy from the RHS, mark it as being merged
2075 if (RangeIsDefinedByCopyFromReg(LHS, LHSIt, RHS.reg)) {
2076 if (ValueLiveAt(LHSIt, LHS.end(), RHSIt->valno->def))
2077 // Here is an interesting situation:
2079 // vr1025 = copy vr1024
2084 // Even though vr1025 is copied from vr1024, it's not safe to
2085 // coalesced them since live range of vr1025 intersects the
2086 // def of vr1024. This happens because vr1025 is assigned the
2087 // value of the previous iteration of vr1024.
2089 EliminatedLHSVals.push_back(LHSIt->valno);
2091 // We know this entire LHS live range is okay, so skip it now.
2092 if (++LHSIt == LHSEnd) break;
2097 if (++RHSIt == RHSEnd) break;
2101 // If we got here, we know that the coalescing will be successful and that
2102 // the value numbers in EliminatedLHSVals will all be merged together. Since
2103 // the most common case is that EliminatedLHSVals has a single number, we
2104 // optimize for it: if there is more than one value, we merge them all into
2105 // the lowest numbered one, then handle the interval as if we were merging
2106 // with one value number.
2107 VNInfo *LHSValNo = NULL;
2108 if (EliminatedLHSVals.size() > 1) {
2109 // Loop through all the equal value numbers merging them into the smallest
2111 VNInfo *Smallest = EliminatedLHSVals[0];
2112 for (unsigned i = 1, e = EliminatedLHSVals.size(); i != e; ++i) {
2113 if (EliminatedLHSVals[i]->id < Smallest->id) {
2114 // Merge the current notion of the smallest into the smaller one.
2115 LHS.MergeValueNumberInto(Smallest, EliminatedLHSVals[i]);
2116 Smallest = EliminatedLHSVals[i];
2118 // Merge into the smallest.
2119 LHS.MergeValueNumberInto(EliminatedLHSVals[i], Smallest);
2122 LHSValNo = Smallest;
2123 } else if (EliminatedLHSVals.empty()) {
2124 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
2125 *tri_->getSuperRegisters(LHS.reg))
2126 // Imprecise sub-register information. Can't handle it.
2128 llvm_unreachable("No copies from the RHS?");
2130 LHSValNo = EliminatedLHSVals[0];
2133 // Okay, now that there is a single LHS value number that we're merging the
2134 // RHS into, update the value number info for the LHS to indicate that the
2135 // value number is defined where the RHS value number was.
2136 const VNInfo *VNI = RHS.getValNumInfo(0);
2137 LHSValNo->def = VNI->def;
2138 LHSValNo->setCopy(VNI->getCopy());
2140 // Okay, the final step is to loop over the RHS live intervals, adding them to
2142 if (VNI->hasPHIKill())
2143 LHSValNo->setHasPHIKill(true);
2144 LHS.addKills(LHSValNo, VNI->kills);
2145 LHS.MergeRangesInAsValue(RHS, LHSValNo);
2147 LHS.ComputeJoinedWeight(RHS);
2149 // Update regalloc hint if both are virtual registers.
2150 if (TargetRegisterInfo::isVirtualRegister(LHS.reg) &&
2151 TargetRegisterInfo::isVirtualRegister(RHS.reg)) {
2152 std::pair<unsigned, unsigned> RHSPref = mri_->getRegAllocationHint(RHS.reg);
2153 std::pair<unsigned, unsigned> LHSPref = mri_->getRegAllocationHint(LHS.reg);
2154 if (RHSPref != LHSPref)
2155 mri_->setRegAllocationHint(LHS.reg, RHSPref.first, RHSPref.second);
2158 // Update the liveintervals of sub-registers.
2159 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg))
2160 for (const unsigned *AS = tri_->getSubRegisters(LHS.reg); *AS; ++AS)
2161 li_->getOrCreateInterval(*AS).MergeInClobberRanges(*li_, LHS,
2162 li_->getVNInfoAllocator());
2167 /// JoinIntervals - Attempt to join these two intervals. On failure, this
2168 /// returns false. Otherwise, if one of the intervals being joined is a
2169 /// physreg, this method always canonicalizes LHS to be it. The output
2170 /// "RHS" will not have been modified, so we can use this information
2171 /// below to update aliases.
2173 SimpleRegisterCoalescing::JoinIntervals(LiveInterval &LHS, LiveInterval &RHS,
2175 // Compute the final value assignment, assuming that the live ranges can be
2177 SmallVector<int, 16> LHSValNoAssignments;
2178 SmallVector<int, 16> RHSValNoAssignments;
2179 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
2180 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
2181 SmallVector<VNInfo*, 16> NewVNInfo;
2183 // If a live interval is a physical register, conservatively check if any
2184 // of its sub-registers is overlapping the live interval of the virtual
2185 // register. If so, do not coalesce.
2186 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
2187 *tri_->getSubRegisters(LHS.reg)) {
2188 // If it's coalescing a virtual register to a physical register, estimate
2189 // its live interval length. This is the *cost* of scanning an entire live
2190 // interval. If the cost is low, we'll do an exhaustive check instead.
2192 // If this is something like this:
2200 // That is, the live interval of v1024 crosses a bb. Then we can't rely on
2201 // less conservative check. It's possible a sub-register is defined before
2202 // v1024 (or live in) and live out of BB1.
2203 if (RHS.containsOneValue() &&
2204 li_->intervalIsInOneMBB(RHS) &&
2205 li_->getApproximateInstructionCount(RHS) <= 10) {
2206 // Perform a more exhaustive check for some common cases.
2207 if (li_->conflictsWithSubPhysRegRef(RHS, LHS.reg, true, JoinedCopies))
2210 for (const unsigned* SR = tri_->getSubRegisters(LHS.reg); *SR; ++SR)
2211 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
2213 dbgs() << "\tInterfere with sub-register ";
2214 li_->getInterval(*SR).print(dbgs(), tri_);
2219 } else if (TargetRegisterInfo::isPhysicalRegister(RHS.reg) &&
2220 *tri_->getSubRegisters(RHS.reg)) {
2221 if (LHS.containsOneValue() &&
2222 li_->getApproximateInstructionCount(LHS) <= 10) {
2223 // Perform a more exhaustive check for some common cases.
2224 if (li_->conflictsWithSubPhysRegRef(LHS, RHS.reg, false, JoinedCopies))
2227 for (const unsigned* SR = tri_->getSubRegisters(RHS.reg); *SR; ++SR)
2228 if (li_->hasInterval(*SR) && LHS.overlaps(li_->getInterval(*SR))) {
2230 dbgs() << "\tInterfere with sub-register ";
2231 li_->getInterval(*SR).print(dbgs(), tri_);
2238 // Compute ultimate value numbers for the LHS and RHS values.
2239 if (RHS.containsOneValue()) {
2240 // Copies from a liveinterval with a single value are simple to handle and
2241 // very common, handle the special case here. This is important, because
2242 // often RHS is small and LHS is large (e.g. a physreg).
2244 // Find out if the RHS is defined as a copy from some value in the LHS.
2245 int RHSVal0DefinedFromLHS = -1;
2247 VNInfo *RHSValNoInfo = NULL;
2248 VNInfo *RHSValNoInfo0 = RHS.getValNumInfo(0);
2249 unsigned RHSSrcReg = li_->getVNInfoSourceReg(RHSValNoInfo0);
2250 if (RHSSrcReg == 0 || RHSSrcReg != LHS.reg) {
2251 // If RHS is not defined as a copy from the LHS, we can use simpler and
2252 // faster checks to see if the live ranges are coalescable. This joiner
2253 // can't swap the LHS/RHS intervals though.
2254 if (!TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
2255 return SimpleJoin(LHS, RHS);
2257 RHSValNoInfo = RHSValNoInfo0;
2260 // It was defined as a copy from the LHS, find out what value # it is.
2262 LHS.getLiveRangeContaining(RHSValNoInfo0->def.getPrevSlot())->valno;
2263 RHSValID = RHSValNoInfo->id;
2264 RHSVal0DefinedFromLHS = RHSValID;
2267 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
2268 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
2269 NewVNInfo.resize(LHS.getNumValNums(), NULL);
2271 // Okay, *all* of the values in LHS that are defined as a copy from RHS
2272 // should now get updated.
2273 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2276 unsigned VN = VNI->id;
2277 if (unsigned LHSSrcReg = li_->getVNInfoSourceReg(VNI)) {
2278 if (LHSSrcReg != RHS.reg) {
2279 // If this is not a copy from the RHS, its value number will be
2280 // unmodified by the coalescing.
2281 NewVNInfo[VN] = VNI;
2282 LHSValNoAssignments[VN] = VN;
2283 } else if (RHSValID == -1) {
2284 // Otherwise, it is a copy from the RHS, and we don't already have a
2285 // value# for it. Keep the current value number, but remember it.
2286 LHSValNoAssignments[VN] = RHSValID = VN;
2287 NewVNInfo[VN] = RHSValNoInfo;
2288 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
2290 // Otherwise, use the specified value #.
2291 LHSValNoAssignments[VN] = RHSValID;
2292 if (VN == (unsigned)RHSValID) { // Else this val# is dead.
2293 NewVNInfo[VN] = RHSValNoInfo;
2294 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
2298 NewVNInfo[VN] = VNI;
2299 LHSValNoAssignments[VN] = VN;
2303 assert(RHSValID != -1 && "Didn't find value #?");
2304 RHSValNoAssignments[0] = RHSValID;
2305 if (RHSVal0DefinedFromLHS != -1) {
2306 // This path doesn't go through ComputeUltimateVN so just set
2308 RHSValsDefinedFromLHS[RHSValNoInfo0] = (VNInfo*)1;
2311 // Loop over the value numbers of the LHS, seeing if any are defined from
2313 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2316 if (VNI->isUnused() || VNI->getCopy() == 0) // Src not defined by a copy?
2319 // DstReg is known to be a register in the LHS interval. If the src is
2320 // from the RHS interval, we can use its value #.
2321 if (li_->getVNInfoSourceReg(VNI) != RHS.reg)
2324 // Figure out the value # from the RHS.
2325 LiveRange *lr = RHS.getLiveRangeContaining(VNI->def.getPrevSlot());
2326 assert(lr && "Cannot find live range");
2327 LHSValsDefinedFromRHS[VNI] = lr->valno;
2330 // Loop over the value numbers of the RHS, seeing if any are defined from
2332 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
2335 if (VNI->isUnused() || VNI->getCopy() == 0) // Src not defined by a copy?
2338 // DstReg is known to be a register in the RHS interval. If the src is
2339 // from the LHS interval, we can use its value #.
2340 if (li_->getVNInfoSourceReg(VNI) != LHS.reg)
2343 // Figure out the value # from the LHS.
2344 LiveRange *lr = LHS.getLiveRangeContaining(VNI->def.getPrevSlot());
2345 assert(lr && "Cannot find live range");
2346 RHSValsDefinedFromLHS[VNI] = lr->valno;
2349 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
2350 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
2351 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
2353 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2356 unsigned VN = VNI->id;
2357 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused())
2359 ComputeUltimateVN(VNI, NewVNInfo,
2360 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
2361 LHSValNoAssignments, RHSValNoAssignments);
2363 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
2366 unsigned VN = VNI->id;
2367 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused())
2369 // If this value number isn't a copy from the LHS, it's a new number.
2370 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
2371 NewVNInfo.push_back(VNI);
2372 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
2376 ComputeUltimateVN(VNI, NewVNInfo,
2377 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
2378 RHSValNoAssignments, LHSValNoAssignments);
2382 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
2383 // interval lists to see if these intervals are coalescable.
2384 LiveInterval::const_iterator I = LHS.begin();
2385 LiveInterval::const_iterator IE = LHS.end();
2386 LiveInterval::const_iterator J = RHS.begin();
2387 LiveInterval::const_iterator JE = RHS.end();
2389 // Skip ahead until the first place of potential sharing.
2390 if (I->start < J->start) {
2391 I = std::upper_bound(I, IE, J->start);
2392 if (I != LHS.begin()) --I;
2393 } else if (J->start < I->start) {
2394 J = std::upper_bound(J, JE, I->start);
2395 if (J != RHS.begin()) --J;
2399 // Determine if these two live ranges overlap.
2401 if (I->start < J->start) {
2402 Overlaps = I->end > J->start;
2404 Overlaps = J->end > I->start;
2407 // If so, check value # info to determine if they are really different.
2409 // If the live range overlap will map to the same value number in the
2410 // result liverange, we can still coalesce them. If not, we can't.
2411 if (LHSValNoAssignments[I->valno->id] !=
2412 RHSValNoAssignments[J->valno->id])
2414 // If it's re-defined by an early clobber somewhere in the live range,
2415 // then conservatively abort coalescing.
2416 if (NewVNInfo[LHSValNoAssignments[I->valno->id]]->hasRedefByEC())
2420 if (I->end < J->end) {
2429 // Update kill info. Some live ranges are extended due to copy coalescing.
2430 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
2431 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
2432 VNInfo *VNI = I->first;
2433 unsigned LHSValID = LHSValNoAssignments[VNI->id];
2434 NewVNInfo[LHSValID]->removeKill(VNI->def);
2435 if (VNI->hasPHIKill())
2436 NewVNInfo[LHSValID]->setHasPHIKill(true);
2437 RHS.addKills(NewVNInfo[LHSValID], VNI->kills);
2440 // Update kill info. Some live ranges are extended due to copy coalescing.
2441 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
2442 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
2443 VNInfo *VNI = I->first;
2444 unsigned RHSValID = RHSValNoAssignments[VNI->id];
2445 NewVNInfo[RHSValID]->removeKill(VNI->def);
2446 if (VNI->hasPHIKill())
2447 NewVNInfo[RHSValID]->setHasPHIKill(true);
2448 LHS.addKills(NewVNInfo[RHSValID], VNI->kills);
2451 // If we get here, we know that we can coalesce the live ranges. Ask the
2452 // intervals to coalesce themselves now.
2453 if ((RHS.ranges.size() > LHS.ranges.size() &&
2454 TargetRegisterInfo::isVirtualRegister(LHS.reg)) ||
2455 TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
2456 RHS.join(LHS, &RHSValNoAssignments[0], &LHSValNoAssignments[0], NewVNInfo,
2460 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo,
2468 // DepthMBBCompare - Comparison predicate that sort first based on the loop
2469 // depth of the basic block (the unsigned), and then on the MBB number.
2470 struct DepthMBBCompare {
2471 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
2472 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
2473 // Deeper loops first
2474 if (LHS.first != RHS.first)
2475 return LHS.first > RHS.first;
2477 // Prefer blocks that are more connected in the CFG. This takes care of
2478 // the most difficult copies first while intervals are short.
2479 unsigned cl = LHS.second->pred_size() + LHS.second->succ_size();
2480 unsigned cr = RHS.second->pred_size() + RHS.second->succ_size();
2484 // As a last resort, sort by block number.
2485 return LHS.second->getNumber() < RHS.second->getNumber();
2490 void SimpleRegisterCoalescing::CopyCoalesceInMBB(MachineBasicBlock *MBB,
2491 std::vector<CopyRec> &TryAgain) {
2492 DEBUG(dbgs() << MBB->getName() << ":\n");
2494 std::vector<CopyRec> VirtCopies;
2495 std::vector<CopyRec> PhysCopies;
2496 std::vector<CopyRec> ImpDefCopies;
2497 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2499 MachineInstr *Inst = MII++;
2501 // If this isn't a copy nor a extract_subreg, we can't join intervals.
2502 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2503 bool isInsUndef = false;
2504 if (Inst->isExtractSubreg()) {
2505 DstReg = Inst->getOperand(0).getReg();
2506 SrcReg = Inst->getOperand(1).getReg();
2507 } else if (Inst->isInsertSubreg()) {
2508 DstReg = Inst->getOperand(0).getReg();
2509 SrcReg = Inst->getOperand(2).getReg();
2510 if (Inst->getOperand(1).isUndef())
2512 } else if (Inst->isInsertSubreg() || Inst->isSubregToReg()) {
2513 DstReg = Inst->getOperand(0).getReg();
2514 SrcReg = Inst->getOperand(2).getReg();
2515 } else if (!tii_->isMoveInstr(*Inst, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
2518 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
2519 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
2521 (li_->hasInterval(SrcReg) && li_->getInterval(SrcReg).empty()))
2522 ImpDefCopies.push_back(CopyRec(Inst, 0));
2523 else if (SrcIsPhys || DstIsPhys)
2524 PhysCopies.push_back(CopyRec(Inst, 0));
2526 VirtCopies.push_back(CopyRec(Inst, 0));
2529 // Try coalescing implicit copies and insert_subreg <undef> first,
2530 // followed by copies to / from physical registers, then finally copies
2531 // from virtual registers to virtual registers.
2532 for (unsigned i = 0, e = ImpDefCopies.size(); i != e; ++i) {
2533 CopyRec &TheCopy = ImpDefCopies[i];
2535 if (!JoinCopy(TheCopy, Again))
2537 TryAgain.push_back(TheCopy);
2539 for (unsigned i = 0, e = PhysCopies.size(); i != e; ++i) {
2540 CopyRec &TheCopy = PhysCopies[i];
2542 if (!JoinCopy(TheCopy, Again))
2544 TryAgain.push_back(TheCopy);
2546 for (unsigned i = 0, e = VirtCopies.size(); i != e; ++i) {
2547 CopyRec &TheCopy = VirtCopies[i];
2549 if (!JoinCopy(TheCopy, Again))
2551 TryAgain.push_back(TheCopy);
2555 void SimpleRegisterCoalescing::joinIntervals() {
2556 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
2558 std::vector<CopyRec> TryAgainList;
2559 if (loopInfo->empty()) {
2560 // If there are no loops in the function, join intervals in function order.
2561 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();
2563 CopyCoalesceInMBB(I, TryAgainList);
2565 // Otherwise, join intervals in inner loops before other intervals.
2566 // Unfortunately we can't just iterate over loop hierarchy here because
2567 // there may be more MBB's than BB's. Collect MBB's for sorting.
2569 // Join intervals in the function prolog first. We want to join physical
2570 // registers with virtual registers before the intervals got too long.
2571 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
2572 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();I != E;++I){
2573 MachineBasicBlock *MBB = I;
2574 MBBs.push_back(std::make_pair(loopInfo->getLoopDepth(MBB), I));
2577 // Sort by loop depth.
2578 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
2580 // Finally, join intervals in loop nest order.
2581 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
2582 CopyCoalesceInMBB(MBBs[i].second, TryAgainList);
2585 // Joining intervals can allow other intervals to be joined. Iteratively join
2586 // until we make no progress.
2587 bool ProgressMade = true;
2588 while (ProgressMade) {
2589 ProgressMade = false;
2591 for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
2592 CopyRec &TheCopy = TryAgainList[i];
2597 bool Success = JoinCopy(TheCopy, Again);
2598 if (Success || !Again) {
2599 TheCopy.MI = 0; // Mark this one as done.
2600 ProgressMade = true;
2606 /// Return true if the two specified registers belong to different register
2607 /// classes. The registers may be either phys or virt regs.
2609 SimpleRegisterCoalescing::differingRegisterClasses(unsigned RegA,
2610 unsigned RegB) const {
2611 // Get the register classes for the first reg.
2612 if (TargetRegisterInfo::isPhysicalRegister(RegA)) {
2613 assert(TargetRegisterInfo::isVirtualRegister(RegB) &&
2614 "Shouldn't consider two physregs!");
2615 return !mri_->getRegClass(RegB)->contains(RegA);
2618 // Compare against the regclass for the second reg.
2619 const TargetRegisterClass *RegClassA = mri_->getRegClass(RegA);
2620 if (TargetRegisterInfo::isVirtualRegister(RegB)) {
2621 const TargetRegisterClass *RegClassB = mri_->getRegClass(RegB);
2622 return RegClassA != RegClassB;
2624 return !RegClassA->contains(RegB);
2627 /// lastRegisterUse - Returns the last (non-debug) use of the specific register
2628 /// between cycles Start and End or NULL if there are no uses.
2630 SimpleRegisterCoalescing::lastRegisterUse(SlotIndex Start,
2633 SlotIndex &UseIdx) const{
2634 UseIdx = SlotIndex();
2635 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2636 MachineOperand *LastUse = NULL;
2637 for (MachineRegisterInfo::use_nodbg_iterator I = mri_->use_nodbg_begin(Reg),
2638 E = mri_->use_nodbg_end(); I != E; ++I) {
2639 MachineOperand &Use = I.getOperand();
2640 MachineInstr *UseMI = Use.getParent();
2641 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2642 if (tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
2643 SrcReg == DstReg && SrcSubIdx == DstSubIdx)
2644 // Ignore identity copies.
2646 SlotIndex Idx = li_->getInstructionIndex(UseMI);
2647 // FIXME: Should this be Idx != UseIdx? SlotIndex() will return something
2648 // that compares higher than any other interval.
2649 if (Idx >= Start && Idx < End && Idx >= UseIdx) {
2651 UseIdx = Idx.getUseIndex();
2657 SlotIndex s = Start;
2658 SlotIndex e = End.getPrevSlot().getBaseIndex();
2660 // Skip deleted instructions
2661 MachineInstr *MI = li_->getInstructionFromIndex(e);
2662 while (e != SlotIndex() && e.getPrevIndex() >= s && !MI) {
2663 e = e.getPrevIndex();
2664 MI = li_->getInstructionFromIndex(e);
2666 if (e < s || MI == NULL)
2669 // Ignore identity copies.
2670 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2671 if (!(tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
2672 SrcReg == DstReg && SrcSubIdx == DstSubIdx))
2673 for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
2674 MachineOperand &Use = MI->getOperand(i);
2675 if (Use.isReg() && Use.isUse() && Use.getReg() &&
2676 tri_->regsOverlap(Use.getReg(), Reg)) {
2677 UseIdx = e.getUseIndex();
2682 e = e.getPrevIndex();
2688 void SimpleRegisterCoalescing::releaseMemory() {
2689 JoinedCopies.clear();
2690 ReMatCopies.clear();
2694 bool SimpleRegisterCoalescing::runOnMachineFunction(MachineFunction &fn) {
2696 mri_ = &fn.getRegInfo();
2697 tm_ = &fn.getTarget();
2698 tri_ = tm_->getRegisterInfo();
2699 tii_ = tm_->getInstrInfo();
2700 li_ = &getAnalysis<LiveIntervals>();
2701 AA = &getAnalysis<AliasAnalysis>();
2702 loopInfo = &getAnalysis<MachineLoopInfo>();
2704 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
2705 << "********** Function: "
2706 << ((Value*)mf_->getFunction())->getName() << '\n');
2708 allocatableRegs_ = tri_->getAllocatableSet(fn);
2709 for (TargetRegisterInfo::regclass_iterator I = tri_->regclass_begin(),
2710 E = tri_->regclass_end(); I != E; ++I)
2711 allocatableRCRegs_.insert(std::make_pair(*I,
2712 tri_->getAllocatableSet(fn, *I)));
2714 // Join (coalesce) intervals if requested.
2715 if (EnableJoining) {
2718 dbgs() << "********** INTERVALS POST JOINING **********\n";
2719 for (LiveIntervals::iterator I = li_->begin(), E = li_->end();
2721 I->second->print(dbgs(), tri_);
2727 // Perform a final pass over the instructions and compute spill weights
2728 // and remove identity moves.
2729 SmallVector<unsigned, 4> DeadDefs;
2730 for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end();
2731 mbbi != mbbe; ++mbbi) {
2732 MachineBasicBlock* mbb = mbbi;
2733 for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
2735 MachineInstr *MI = mii;
2736 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2737 if (JoinedCopies.count(MI)) {
2738 // Delete all coalesced copies.
2739 bool DoDelete = true;
2740 if (!tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
2741 assert((MI->isExtractSubreg() || MI->isInsertSubreg() ||
2742 MI->isSubregToReg()) && "Unrecognized copy instruction");
2743 DstReg = MI->getOperand(0).getReg();
2744 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
2745 // Do not delete extract_subreg, insert_subreg of physical
2746 // registers unless the definition is dead. e.g.
2747 // %DO<def> = INSERT_SUBREG %D0<undef>, %S0<kill>, 1
2748 // or else the scavenger may complain. LowerSubregs will
2749 // delete them later.
2752 if (MI->allDefsAreDead()) {
2753 LiveInterval &li = li_->getInterval(DstReg);
2754 if (!ShortenDeadCopySrcLiveRange(li, MI))
2755 ShortenDeadCopyLiveRange(li, MI);
2759 mii = llvm::next(mii);
2761 li_->RemoveMachineInstrFromMaps(MI);
2762 mii = mbbi->erase(mii);
2768 // Now check if this is a remat'ed def instruction which is now dead.
2769 if (ReMatDefs.count(MI)) {
2771 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2772 const MachineOperand &MO = MI->getOperand(i);
2775 unsigned Reg = MO.getReg();
2778 if (TargetRegisterInfo::isVirtualRegister(Reg))
2779 DeadDefs.push_back(Reg);
2782 if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
2783 !mri_->use_nodbg_empty(Reg)) {
2789 while (!DeadDefs.empty()) {
2790 unsigned DeadDef = DeadDefs.back();
2791 DeadDefs.pop_back();
2792 RemoveDeadDef(li_->getInterval(DeadDef), MI);
2794 li_->RemoveMachineInstrFromMaps(mii);
2795 mii = mbbi->erase(mii);
2801 // If the move will be an identity move delete it
2802 bool isMove= tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx);
2803 if (isMove && SrcReg == DstReg && SrcSubIdx == DstSubIdx) {
2804 if (li_->hasInterval(SrcReg)) {
2805 LiveInterval &RegInt = li_->getInterval(SrcReg);
2806 // If def of this move instruction is dead, remove its live range
2807 // from the dstination register's live interval.
2808 if (MI->registerDefIsDead(DstReg)) {
2809 if (!ShortenDeadCopySrcLiveRange(RegInt, MI))
2810 ShortenDeadCopyLiveRange(RegInt, MI);
2813 li_->RemoveMachineInstrFromMaps(MI);
2814 mii = mbbi->erase(mii);
2821 // Check for now unnecessary kill flags.
2822 if (li_->isNotInMIMap(MI)) continue;
2823 SlotIndex UseIdx = li_->getInstructionIndex(MI).getUseIndex();
2824 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2825 MachineOperand &MO = MI->getOperand(i);
2826 if (!MO.isReg() || !MO.isKill()) continue;
2827 unsigned reg = MO.getReg();
2828 if (!reg || !li_->hasInterval(reg)) continue;
2829 LiveInterval &LI = li_->getInterval(reg);
2830 const LiveRange *LR = LI.getLiveRangeContaining(UseIdx);
2832 (!LR->valno->isKill(UseIdx.getDefIndex()) &&
2833 LR->valno->def != UseIdx.getDefIndex()))
2834 MO.setIsKill(false);
2843 /// print - Implement the dump method.
2844 void SimpleRegisterCoalescing::print(raw_ostream &O, const Module* m) const {
2848 RegisterCoalescer* llvm::createSimpleRegisterCoalescer() {
2849 return new SimpleRegisterCoalescing();
2852 // Make sure that anything that uses RegisterCoalescer pulls in this file...
2853 DEFINING_FILE_FOR(SimpleRegisterCoalescing)