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() << "Interfere with sub-register ";
183 li_->getInterval(*SR).print(dbgs(), tri_);
190 dbgs() << "\nExtending: ";
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 (CopyMI->killsRegister(IntA.reg))
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 if (BI->start <= AI->start && BI->end > AI->start)
264 if (BI->start > AI->start && BI->start < AI->end)
272 TransferImplicitOps(MachineInstr *MI, MachineInstr *NewMI) {
273 for (unsigned i = MI->getDesc().getNumOperands(), e = MI->getNumOperands();
275 MachineOperand &MO = MI->getOperand(i);
276 if (MO.isReg() && MO.isImplicit())
277 NewMI->addOperand(MO);
281 /// RemoveCopyByCommutingDef - We found a non-trivially-coalescable copy with
282 /// IntA being the source and IntB being the dest, thus this defines a value
283 /// number in IntB. If the source value number (in IntA) is defined by a
284 /// commutable instruction and its other operand is coalesced to the copy dest
285 /// register, see if we can transform the copy into a noop by commuting the
286 /// definition. For example,
288 /// A3 = op A2 B0<kill>
290 /// B1 = A3 <- this copy
292 /// = op A3 <- more uses
296 /// B2 = op B0 A2<kill>
298 /// B1 = B2 <- now an identify copy
300 /// = op B2 <- more uses
302 /// This returns true if an interval was modified.
304 bool SimpleRegisterCoalescing::RemoveCopyByCommutingDef(LiveInterval &IntA,
306 MachineInstr *CopyMI) {
308 li_->getInstructionIndex(CopyMI).getDefIndex();
310 // FIXME: For now, only eliminate the copy by commuting its def when the
311 // source register is a virtual register. We want to guard against cases
312 // where the copy is a back edge copy and commuting the def lengthen the
313 // live interval of the source register to the entire loop.
314 if (TargetRegisterInfo::isPhysicalRegister(IntA.reg))
317 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
318 // the example above.
319 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
320 assert(BLR != IntB.end() && "Live range not found!");
321 VNInfo *BValNo = BLR->valno;
323 // Get the location that B is defined at. Two options: either this value has
324 // an unknown definition point or it is defined at CopyIdx. If unknown, we
326 if (!BValNo->getCopy()) return false;
327 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
329 // AValNo is the value number in A that defines the copy, A3 in the example.
330 LiveInterval::iterator ALR =
331 IntA.FindLiveRangeContaining(CopyIdx.getUseIndex()); //
333 assert(ALR != IntA.end() && "Live range not found!");
334 VNInfo *AValNo = ALR->valno;
335 // If other defs can reach uses of this def, then it's not safe to perform
336 // the optimization. FIXME: Do isPHIDef and isDefAccurate both need to be
338 if (AValNo->isPHIDef() || !AValNo->isDefAccurate() ||
339 AValNo->isUnused() || AValNo->hasPHIKill())
341 MachineInstr *DefMI = li_->getInstructionFromIndex(AValNo->def);
342 const TargetInstrDesc &TID = DefMI->getDesc();
343 if (!TID.isCommutable())
345 // If DefMI is a two-address instruction then commuting it will change the
346 // destination register.
347 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
348 assert(DefIdx != -1);
350 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
352 unsigned Op1, Op2, NewDstIdx;
353 if (!tii_->findCommutedOpIndices(DefMI, Op1, Op2))
357 else if (Op2 == UseOpIdx)
362 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
363 unsigned NewReg = NewDstMO.getReg();
364 if (NewReg != IntB.reg || !NewDstMO.isKill())
367 // Make sure there are no other definitions of IntB that would reach the
368 // uses which the new definition can reach.
369 if (HasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
372 // If some of the uses of IntA.reg is already coalesced away, return false.
373 // It's not possible to determine whether it's safe to perform the coalescing.
374 for (MachineRegisterInfo::use_nodbg_iterator UI =
375 mri_->use_nodbg_begin(IntA.reg),
376 UE = mri_->use_nodbg_end(); UI != UE; ++UI) {
377 MachineInstr *UseMI = &*UI;
378 SlotIndex UseIdx = li_->getInstructionIndex(UseMI);
379 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
380 if (ULR == IntA.end())
382 if (ULR->valno == AValNo && JoinedCopies.count(UseMI))
386 // At this point we have decided that it is legal to do this
387 // transformation. Start by commuting the instruction.
388 MachineBasicBlock *MBB = DefMI->getParent();
389 MachineInstr *NewMI = tii_->commuteInstruction(DefMI);
392 if (NewMI != DefMI) {
393 li_->ReplaceMachineInstrInMaps(DefMI, NewMI);
394 MBB->insert(DefMI, NewMI);
397 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false);
398 NewMI->getOperand(OpIdx).setIsKill();
400 bool BHasPHIKill = BValNo->hasPHIKill();
401 SmallVector<VNInfo*, 4> BDeadValNos;
402 VNInfo::KillSet BKills;
403 std::map<SlotIndex, SlotIndex> BExtend;
405 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
414 // then do not add kills of A to the newly created B interval.
415 bool Extended = BLR->end > ALR->end && ALR->end != ALR->start;
417 BExtend[ALR->end] = BLR->end;
419 // Update uses of IntA of the specific Val# with IntB.
420 bool BHasSubRegs = false;
421 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg))
422 BHasSubRegs = *tri_->getSubRegisters(IntB.reg);
423 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(IntA.reg),
424 UE = mri_->use_end(); UI != UE;) {
425 MachineOperand &UseMO = UI.getOperand();
426 MachineInstr *UseMI = &*UI;
428 if (JoinedCopies.count(UseMI))
430 if (UseMI->isDebugValue()) {
431 // FIXME These don't have an instruction index. Not clear we have enough
432 // info to decide whether to do this replacement or not. For now do it.
433 UseMO.setReg(NewReg);
436 SlotIndex UseIdx = li_->getInstructionIndex(UseMI).getUseIndex();
437 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
438 if (ULR == IntA.end() || ULR->valno != AValNo)
440 UseMO.setReg(NewReg);
443 if (UseMO.isKill()) {
445 UseMO.setIsKill(false);
447 BKills.push_back(UseIdx.getDefIndex());
449 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
450 if (!tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
452 if (DstReg == IntB.reg) {
453 // This copy will become a noop. If it's defining a new val#,
454 // remove that val# as well. However this live range is being
455 // extended to the end of the existing live range defined by the copy.
456 SlotIndex DefIdx = UseIdx.getDefIndex();
457 const LiveRange *DLR = IntB.getLiveRangeContaining(DefIdx);
458 BHasPHIKill |= DLR->valno->hasPHIKill();
459 assert(DLR->valno->def == DefIdx);
460 BDeadValNos.push_back(DLR->valno);
461 BExtend[DLR->start] = DLR->end;
462 JoinedCopies.insert(UseMI);
463 // If this is a kill but it's going to be removed, the last use
464 // of the same val# is the new kill.
470 // We need to insert a new liverange: [ALR.start, LastUse). It may be we can
471 // simply extend BLR if CopyMI doesn't end the range.
473 dbgs() << "\nExtending: ";
474 IntB.print(dbgs(), tri_);
477 // Remove val#'s defined by copies that will be coalesced away.
478 for (unsigned i = 0, e = BDeadValNos.size(); i != e; ++i) {
479 VNInfo *DeadVNI = BDeadValNos[i];
481 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
482 LiveInterval &SRLI = li_->getInterval(*SR);
483 const LiveRange *SRLR = SRLI.getLiveRangeContaining(DeadVNI->def);
484 SRLI.removeValNo(SRLR->valno);
487 IntB.removeValNo(BDeadValNos[i]);
490 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition
491 // is updated. Kills are also updated.
492 VNInfo *ValNo = BValNo;
493 ValNo->def = AValNo->def;
495 for (unsigned j = 0, ee = ValNo->kills.size(); j != ee; ++j) {
496 if (ValNo->kills[j] != BLR->end)
497 BKills.push_back(ValNo->kills[j]);
499 ValNo->kills.clear();
500 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
502 if (AI->valno != AValNo) continue;
503 SlotIndex End = AI->end;
504 std::map<SlotIndex, SlotIndex>::iterator
505 EI = BExtend.find(End);
506 if (EI != BExtend.end())
508 IntB.addRange(LiveRange(AI->start, End, ValNo));
510 // If the IntB live range is assigned to a physical register, and if that
511 // physreg has sub-registers, update their live intervals as well.
513 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
514 LiveInterval &SRLI = li_->getInterval(*SR);
515 SRLI.MergeInClobberRange(*li_, AI->start, End,
516 li_->getVNInfoAllocator());
520 IntB.addKills(ValNo, BKills);
521 ValNo->setHasPHIKill(BHasPHIKill);
524 dbgs() << " result = ";
525 IntB.print(dbgs(), tri_);
527 dbgs() << "\nShortening: ";
528 IntA.print(dbgs(), tri_);
531 IntA.removeValNo(AValNo);
534 dbgs() << " result = ";
535 IntA.print(dbgs(), tri_);
543 /// isSameOrFallThroughBB - Return true if MBB == SuccMBB or MBB simply
544 /// fallthoughs to SuccMBB.
545 static bool isSameOrFallThroughBB(MachineBasicBlock *MBB,
546 MachineBasicBlock *SuccMBB,
547 const TargetInstrInfo *tii_) {
550 MachineBasicBlock *TBB = 0, *FBB = 0;
551 SmallVector<MachineOperand, 4> Cond;
552 return !tii_->AnalyzeBranch(*MBB, TBB, FBB, Cond) && !TBB && !FBB &&
553 MBB->isSuccessor(SuccMBB);
556 /// removeRange - Wrapper for LiveInterval::removeRange. This removes a range
557 /// from a physical register live interval as well as from the live intervals
558 /// of its sub-registers.
559 static void removeRange(LiveInterval &li,
560 SlotIndex Start, SlotIndex End,
561 LiveIntervals *li_, const TargetRegisterInfo *tri_) {
562 li.removeRange(Start, End, true);
563 if (TargetRegisterInfo::isPhysicalRegister(li.reg)) {
564 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
565 if (!li_->hasInterval(*SR))
567 LiveInterval &sli = li_->getInterval(*SR);
568 SlotIndex RemoveStart = Start;
569 SlotIndex RemoveEnd = Start;
571 while (RemoveEnd != End) {
572 LiveInterval::iterator LR = sli.FindLiveRangeContaining(RemoveStart);
575 RemoveEnd = (LR->end < End) ? LR->end : End;
576 sli.removeRange(RemoveStart, RemoveEnd, true);
577 RemoveStart = RemoveEnd;
583 /// TrimLiveIntervalToLastUse - If there is a last use in the same basic block
584 /// as the copy instruction, trim the live interval to the last use and return
587 SimpleRegisterCoalescing::TrimLiveIntervalToLastUse(SlotIndex CopyIdx,
588 MachineBasicBlock *CopyMBB,
590 const LiveRange *LR) {
591 SlotIndex MBBStart = li_->getMBBStartIdx(CopyMBB);
592 SlotIndex LastUseIdx;
593 MachineOperand *LastUse =
594 lastRegisterUse(LR->start, CopyIdx.getPrevSlot(), li.reg, LastUseIdx);
596 MachineInstr *LastUseMI = LastUse->getParent();
597 if (!isSameOrFallThroughBB(LastUseMI->getParent(), CopyMBB, tii_)) {
604 // r1025<dead> = r1024<kill>
605 if (MBBStart < LR->end)
606 removeRange(li, MBBStart, LR->end, li_, tri_);
610 // There are uses before the copy, just shorten the live range to the end
612 LastUse->setIsKill();
613 removeRange(li, LastUseIdx.getDefIndex(), LR->end, li_, tri_);
614 LR->valno->addKill(LastUseIdx.getDefIndex());
615 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
616 if (tii_->isMoveInstr(*LastUseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
618 // Last use is itself an identity code.
619 int DeadIdx = LastUseMI->findRegisterDefOperandIdx(li.reg, false, tri_);
620 LastUseMI->getOperand(DeadIdx).setIsDead();
626 if (LR->start <= MBBStart && LR->end > MBBStart) {
627 if (LR->start == li_->getZeroIndex()) {
628 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
629 // Live-in to the function but dead. Remove it from entry live-in set.
630 mf_->begin()->removeLiveIn(li.reg);
632 // FIXME: Shorten intervals in BBs that reaches this BB.
638 /// ReMaterializeTrivialDef - If the source of a copy is defined by a trivial
639 /// computation, replace the copy by rematerialize the definition.
640 bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
643 MachineInstr *CopyMI) {
644 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI).getUseIndex();
645 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx);
646 assert(SrcLR != SrcInt.end() && "Live range not found!");
647 VNInfo *ValNo = SrcLR->valno;
648 // If other defs can reach uses of this def, then it's not safe to perform
649 // the optimization. FIXME: Do isPHIDef and isDefAccurate both need to be
651 if (ValNo->isPHIDef() || !ValNo->isDefAccurate() ||
652 ValNo->isUnused() || ValNo->hasPHIKill())
654 MachineInstr *DefMI = li_->getInstructionFromIndex(ValNo->def);
655 const TargetInstrDesc &TID = DefMI->getDesc();
656 if (!TID.isAsCheapAsAMove())
658 if (!tii_->isTriviallyReMaterializable(DefMI, AA))
660 bool SawStore = false;
661 if (!DefMI->isSafeToMove(tii_, AA, SawStore))
663 if (TID.getNumDefs() != 1)
665 if (!DefMI->isImplicitDef()) {
666 // Make sure the copy destination register class fits the instruction
667 // definition register class. The mismatch can happen as a result of earlier
668 // extract_subreg, insert_subreg, subreg_to_reg coalescing.
669 const TargetRegisterClass *RC = TID.OpInfo[0].getRegClass(tri_);
670 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
671 if (mri_->getRegClass(DstReg) != RC)
673 } else if (!RC->contains(DstReg))
677 // If destination register has a sub-register index on it, make sure it mtches
678 // the instruction register class.
680 const TargetInstrDesc &TID = DefMI->getDesc();
681 if (TID.getNumDefs() != 1)
683 const TargetRegisterClass *DstRC = mri_->getRegClass(DstReg);
684 const TargetRegisterClass *DstSubRC =
685 DstRC->getSubRegisterRegClass(DstSubIdx);
686 const TargetRegisterClass *DefRC = TID.OpInfo[0].getRegClass(tri_);
689 else if (DefRC != DstSubRC)
693 SlotIndex DefIdx = CopyIdx.getDefIndex();
694 const LiveRange *DLR= li_->getInterval(DstReg).getLiveRangeContaining(DefIdx);
695 DLR->valno->setCopy(0);
696 // Don't forget to update sub-register intervals.
697 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
698 for (const unsigned* SR = tri_->getSubRegisters(DstReg); *SR; ++SR) {
699 if (!li_->hasInterval(*SR))
701 const LiveRange *DLR =
702 li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
703 if (DLR && DLR->valno->getCopy() == CopyMI)
704 DLR->valno->setCopy(0);
708 // If copy kills the source register, find the last use and propagate
710 bool checkForDeadDef = false;
711 MachineBasicBlock *MBB = CopyMI->getParent();
712 if (CopyMI->killsRegister(SrcInt.reg))
713 if (!TrimLiveIntervalToLastUse(CopyIdx, MBB, SrcInt, SrcLR)) {
714 checkForDeadDef = true;
717 MachineBasicBlock::iterator MII =
718 llvm::next(MachineBasicBlock::iterator(CopyMI));
719 tii_->reMaterialize(*MBB, MII, DstReg, DstSubIdx, DefMI, tri_);
720 MachineInstr *NewMI = prior(MII);
722 if (checkForDeadDef) {
723 // PR4090 fix: Trim interval failed because there was no use of the
724 // source interval in this MBB. If the def is in this MBB too then we
725 // should mark it dead:
726 if (DefMI->getParent() == MBB) {
727 DefMI->addRegisterDead(SrcInt.reg, tri_);
728 SrcLR->end = SrcLR->start.getNextSlot();
732 // CopyMI may have implicit operands, transfer them over to the newly
733 // rematerialized instruction. And update implicit def interval valnos.
734 for (unsigned i = CopyMI->getDesc().getNumOperands(),
735 e = CopyMI->getNumOperands(); i != e; ++i) {
736 MachineOperand &MO = CopyMI->getOperand(i);
737 if (MO.isReg() && MO.isImplicit())
738 NewMI->addOperand(MO);
739 if (MO.isDef() && li_->hasInterval(MO.getReg())) {
740 unsigned Reg = MO.getReg();
741 const LiveRange *DLR =
742 li_->getInterval(Reg).getLiveRangeContaining(DefIdx);
743 if (DLR && DLR->valno->getCopy() == CopyMI)
744 DLR->valno->setCopy(0);
745 // Handle subregs as well
746 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
747 for (const unsigned* SR = tri_->getSubRegisters(Reg); *SR; ++SR) {
748 if (!li_->hasInterval(*SR))
750 const LiveRange *DLR =
751 li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
752 if (DLR && DLR->valno->getCopy() == CopyMI)
753 DLR->valno->setCopy(0);
759 TransferImplicitOps(CopyMI, NewMI);
760 li_->ReplaceMachineInstrInMaps(CopyMI, NewMI);
761 CopyMI->eraseFromParent();
762 ReMatCopies.insert(CopyMI);
763 ReMatDefs.insert(DefMI);
764 DEBUG(dbgs() << "Remat: " << *NewMI);
769 /// UpdateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
770 /// update the subregister number if it is not zero. If DstReg is a
771 /// physical register and the existing subregister number of the def / use
772 /// being updated is not zero, make sure to set it to the correct physical
775 SimpleRegisterCoalescing::UpdateRegDefsUses(unsigned SrcReg, unsigned DstReg,
777 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
778 if (DstIsPhys && SubIdx) {
779 // Figure out the real physical register we are updating with.
780 DstReg = tri_->getSubReg(DstReg, SubIdx);
784 // Copy the register use-list before traversing it. We may be adding operands
785 // and invalidating pointers.
786 SmallVector<std::pair<MachineInstr*, unsigned>, 32> reglist;
787 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(SrcReg),
788 E = mri_->reg_end(); I != E; ++I)
789 reglist.push_back(std::make_pair(&*I, I.getOperandNo()));
791 for (unsigned N=0; N != reglist.size(); ++N) {
792 MachineInstr *UseMI = reglist[N].first;
793 MachineOperand &O = UseMI->getOperand(reglist[N].second);
794 unsigned OldSubIdx = O.getSubReg();
796 unsigned UseDstReg = DstReg;
798 UseDstReg = tri_->getSubReg(DstReg, OldSubIdx);
800 unsigned CopySrcReg, CopyDstReg, CopySrcSubIdx, CopyDstSubIdx;
801 if (tii_->isMoveInstr(*UseMI, CopySrcReg, CopyDstReg,
802 CopySrcSubIdx, CopyDstSubIdx) &&
803 CopySrcReg != CopyDstReg &&
804 CopySrcReg == SrcReg && CopyDstReg != UseDstReg) {
805 // If the use is a copy and it won't be coalesced away, and its source
806 // is defined by a trivial computation, try to rematerialize it instead.
807 if (ReMaterializeTrivialDef(li_->getInterval(SrcReg), CopyDstReg,
808 CopyDstSubIdx, UseMI))
815 // Def and kill of subregister of a virtual register actually defs and
816 // kills the whole register. Add imp-defs and imp-kills as needed.
819 UseMI->addRegisterDead(DstReg, tri_, true);
821 UseMI->addRegisterDefined(DstReg, tri_);
822 } else if (!O.isUndef() &&
824 UseMI->isRegTiedToDefOperand(&O-&UseMI->getOperand(0))))
825 UseMI->addRegisterKilled(DstReg, tri_, true);
830 // Sub-register indexes goes from small to large. e.g.
831 // RAX: 1 -> AL, 2 -> AX, 3 -> EAX
832 // EAX: 1 -> AL, 2 -> AX
833 // So RAX's sub-register 2 is AX, RAX's sub-regsiter 3 is EAX, whose
834 // sub-register 2 is also AX.
835 if (SubIdx && OldSubIdx && SubIdx != OldSubIdx)
836 assert(OldSubIdx < SubIdx && "Conflicting sub-register index!");
839 // Remove would-be duplicated kill marker.
840 if (O.isKill() && UseMI->killsRegister(DstReg))
844 // After updating the operand, check if the machine instruction has
845 // become a copy. If so, update its val# information.
846 if (JoinedCopies.count(UseMI))
849 const TargetInstrDesc &TID = UseMI->getDesc();
850 unsigned CopySrcReg, CopyDstReg, CopySrcSubIdx, CopyDstSubIdx;
851 if (TID.getNumDefs() == 1 && TID.getNumOperands() > 2 &&
852 tii_->isMoveInstr(*UseMI, CopySrcReg, CopyDstReg,
853 CopySrcSubIdx, CopyDstSubIdx) &&
854 CopySrcReg != CopyDstReg &&
855 (TargetRegisterInfo::isVirtualRegister(CopyDstReg) ||
856 allocatableRegs_[CopyDstReg])) {
857 LiveInterval &LI = li_->getInterval(CopyDstReg);
859 li_->getInstructionIndex(UseMI).getDefIndex();
860 if (const LiveRange *DLR = LI.getLiveRangeContaining(DefIdx)) {
861 if (DLR->valno->def == DefIdx)
862 DLR->valno->setCopy(UseMI);
868 /// RemoveUnnecessaryKills - Remove kill markers that are no longer accurate
869 /// due to live range lengthening as the result of coalescing.
870 void SimpleRegisterCoalescing::RemoveUnnecessaryKills(unsigned Reg,
872 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(Reg),
873 UE = mri_->use_end(); UI != UE; ++UI) {
874 MachineOperand &UseMO = UI.getOperand();
877 MachineInstr *UseMI = UseMO.getParent();
879 li_->getInstructionIndex(UseMI).getUseIndex();
880 const LiveRange *LR = LI.getLiveRangeContaining(UseIdx);
882 (!LR->valno->isKill(UseIdx.getDefIndex()) &&
883 LR->valno->def != UseIdx.getDefIndex())) {
884 // Interesting problem. After coalescing reg1027's def and kill are both
885 // at the same point: %reg1027,0.000000e+00 = [56,814:0) 0@70-(814)
888 // 60 %reg1027<def> = t2MOVr %reg1027, 14, %reg0, %reg0
889 // 68 %reg1027<def> = t2LDRi12 %reg1027<kill>, 8, 14, %reg0
890 // 76 t2CMPzri %reg1038<kill,undef>, 0, 14, %reg0, %CPSR<imp-def>
891 // 84 %reg1027<def> = t2MOVr %reg1027, 14, %reg0, %reg0
892 // 96 t2Bcc mbb<bb5,0x2030910>, 1, %CPSR<kill>
894 // Do not remove the kill marker on t2LDRi12.
895 UseMO.setIsKill(false);
900 /// removeIntervalIfEmpty - Check if the live interval of a physical register
901 /// is empty, if so remove it and also remove the empty intervals of its
902 /// sub-registers. Return true if live interval is removed.
903 static bool removeIntervalIfEmpty(LiveInterval &li, LiveIntervals *li_,
904 const TargetRegisterInfo *tri_) {
906 if (TargetRegisterInfo::isPhysicalRegister(li.reg))
907 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
908 if (!li_->hasInterval(*SR))
910 LiveInterval &sli = li_->getInterval(*SR);
912 li_->removeInterval(*SR);
914 li_->removeInterval(li.reg);
920 /// ShortenDeadCopyLiveRange - Shorten a live range defined by a dead copy.
921 /// Return true if live interval is removed.
922 bool SimpleRegisterCoalescing::ShortenDeadCopyLiveRange(LiveInterval &li,
923 MachineInstr *CopyMI) {
924 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI);
925 LiveInterval::iterator MLR =
926 li.FindLiveRangeContaining(CopyIdx.getDefIndex());
928 return false; // Already removed by ShortenDeadCopySrcLiveRange.
929 SlotIndex RemoveStart = MLR->start;
930 SlotIndex RemoveEnd = MLR->end;
931 SlotIndex DefIdx = CopyIdx.getDefIndex();
932 // Remove the liverange that's defined by this.
933 if (RemoveStart == DefIdx && RemoveEnd == DefIdx.getStoreIndex()) {
934 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
935 return removeIntervalIfEmpty(li, li_, tri_);
940 /// RemoveDeadDef - If a def of a live interval is now determined dead, remove
941 /// the val# it defines. If the live interval becomes empty, remove it as well.
942 bool SimpleRegisterCoalescing::RemoveDeadDef(LiveInterval &li,
943 MachineInstr *DefMI) {
944 SlotIndex DefIdx = li_->getInstructionIndex(DefMI).getDefIndex();
945 LiveInterval::iterator MLR = li.FindLiveRangeContaining(DefIdx);
946 if (DefIdx != MLR->valno->def)
948 li.removeValNo(MLR->valno);
949 return removeIntervalIfEmpty(li, li_, tri_);
952 /// PropagateDeadness - Propagate the dead marker to the instruction which
953 /// defines the val#.
954 static void PropagateDeadness(LiveInterval &li, MachineInstr *CopyMI,
955 SlotIndex &LRStart, LiveIntervals *li_,
956 const TargetRegisterInfo* tri_) {
957 MachineInstr *DefMI =
958 li_->getInstructionFromIndex(LRStart.getDefIndex());
959 if (DefMI && DefMI != CopyMI) {
960 int DeadIdx = DefMI->findRegisterDefOperandIdx(li.reg, false);
962 DefMI->getOperand(DeadIdx).setIsDead();
964 DefMI->addOperand(MachineOperand::CreateReg(li.reg,
965 /*def*/true, /*implicit*/true, /*kill*/false, /*dead*/true));
966 LRStart = LRStart.getNextSlot();
970 /// ShortenDeadCopySrcLiveRange - Shorten a live range as it's artificially
971 /// extended by a dead copy. Mark the last use (if any) of the val# as kill as
972 /// ends the live range there. If there isn't another use, then this live range
973 /// is dead. Return true if live interval is removed.
975 SimpleRegisterCoalescing::ShortenDeadCopySrcLiveRange(LiveInterval &li,
976 MachineInstr *CopyMI) {
977 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI);
978 if (CopyIdx == SlotIndex()) {
979 // FIXME: special case: function live in. It can be a general case if the
980 // first instruction index starts at > 0 value.
981 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
982 // Live-in to the function but dead. Remove it from entry live-in set.
983 if (mf_->begin()->isLiveIn(li.reg))
984 mf_->begin()->removeLiveIn(li.reg);
985 const LiveRange *LR = li.getLiveRangeContaining(CopyIdx);
986 removeRange(li, LR->start, LR->end, li_, tri_);
987 return removeIntervalIfEmpty(li, li_, tri_);
990 LiveInterval::iterator LR =
991 li.FindLiveRangeContaining(CopyIdx.getPrevIndex().getStoreIndex());
993 // Livein but defined by a phi.
996 SlotIndex RemoveStart = LR->start;
997 SlotIndex RemoveEnd = CopyIdx.getStoreIndex();
998 if (LR->end > RemoveEnd)
999 // More uses past this copy? Nothing to do.
1002 // If there is a last use in the same bb, we can't remove the live range.
1003 // Shorten the live interval and return.
1004 MachineBasicBlock *CopyMBB = CopyMI->getParent();
1005 if (TrimLiveIntervalToLastUse(CopyIdx, CopyMBB, li, LR))
1008 // There are other kills of the val#. Nothing to do.
1009 if (!li.isOnlyLROfValNo(LR))
1012 MachineBasicBlock *StartMBB = li_->getMBBFromIndex(RemoveStart);
1013 if (!isSameOrFallThroughBB(StartMBB, CopyMBB, tii_))
1014 // If the live range starts in another mbb and the copy mbb is not a fall
1015 // through mbb, then we can only cut the range from the beginning of the
1017 RemoveStart = li_->getMBBStartIdx(CopyMBB).getNextIndex().getBaseIndex();
1019 if (LR->valno->def == RemoveStart) {
1020 // If the def MI defines the val# and this copy is the only kill of the
1021 // val#, then propagate the dead marker.
1022 PropagateDeadness(li, CopyMI, RemoveStart, li_, tri_);
1025 if (LR->valno->isKill(RemoveEnd))
1026 LR->valno->removeKill(RemoveEnd);
1029 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
1030 return removeIntervalIfEmpty(li, li_, tri_);
1033 /// CanCoalesceWithImpDef - Returns true if the specified copy instruction
1034 /// from an implicit def to another register can be coalesced away.
1035 bool SimpleRegisterCoalescing::CanCoalesceWithImpDef(MachineInstr *CopyMI,
1037 LiveInterval &ImpLi) const{
1038 if (!CopyMI->killsRegister(ImpLi.reg))
1040 // Make sure this is the only use.
1041 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(ImpLi.reg),
1042 UE = mri_->use_end(); UI != UE;) {
1043 MachineInstr *UseMI = &*UI;
1045 if (CopyMI == UseMI || JoinedCopies.count(UseMI))
1053 /// isWinToJoinVRWithSrcPhysReg - Return true if it's worth while to join a
1054 /// a virtual destination register with physical source register.
1056 SimpleRegisterCoalescing::isWinToJoinVRWithSrcPhysReg(MachineInstr *CopyMI,
1057 MachineBasicBlock *CopyMBB,
1058 LiveInterval &DstInt,
1059 LiveInterval &SrcInt) {
1060 // If the virtual register live interval is long but it has low use desity,
1061 // do not join them, instead mark the physical register as its allocation
1063 const TargetRegisterClass *RC = mri_->getRegClass(DstInt.reg);
1064 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1065 unsigned Length = li_->getApproximateInstructionCount(DstInt);
1066 if (Length > Threshold &&
1067 (((float)std::distance(mri_->use_nodbg_begin(DstInt.reg),
1068 mri_->use_nodbg_end()) / Length) <
1072 // If the virtual register live interval extends into a loop, turn down
1075 li_->getInstructionIndex(CopyMI).getDefIndex();
1076 const MachineLoop *L = loopInfo->getLoopFor(CopyMBB);
1078 // Let's see if the virtual register live interval extends into the loop.
1079 LiveInterval::iterator DLR = DstInt.FindLiveRangeContaining(CopyIdx);
1080 assert(DLR != DstInt.end() && "Live range not found!");
1081 DLR = DstInt.FindLiveRangeContaining(DLR->end.getNextSlot());
1082 if (DLR != DstInt.end()) {
1083 CopyMBB = li_->getMBBFromIndex(DLR->start);
1084 L = loopInfo->getLoopFor(CopyMBB);
1088 if (!L || Length <= Threshold)
1091 SlotIndex UseIdx = CopyIdx.getUseIndex();
1092 LiveInterval::iterator SLR = SrcInt.FindLiveRangeContaining(UseIdx);
1093 MachineBasicBlock *SMBB = li_->getMBBFromIndex(SLR->start);
1094 if (loopInfo->getLoopFor(SMBB) != L) {
1095 if (!loopInfo->isLoopHeader(CopyMBB))
1097 // If vr's live interval extends pass the loop header, do not join.
1098 for (MachineBasicBlock::succ_iterator SI = CopyMBB->succ_begin(),
1099 SE = CopyMBB->succ_end(); SI != SE; ++SI) {
1100 MachineBasicBlock *SuccMBB = *SI;
1101 if (SuccMBB == CopyMBB)
1103 if (DstInt.overlaps(li_->getMBBStartIdx(SuccMBB),
1104 li_->getMBBEndIdx(SuccMBB)))
1111 /// isWinToJoinVRWithDstPhysReg - Return true if it's worth while to join a
1112 /// copy from a virtual source register to a physical destination register.
1114 SimpleRegisterCoalescing::isWinToJoinVRWithDstPhysReg(MachineInstr *CopyMI,
1115 MachineBasicBlock *CopyMBB,
1116 LiveInterval &DstInt,
1117 LiveInterval &SrcInt) {
1118 // If the virtual register live interval is long but it has low use density,
1119 // do not join them, instead mark the physical register as its allocation
1121 const TargetRegisterClass *RC = mri_->getRegClass(SrcInt.reg);
1122 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1123 unsigned Length = li_->getApproximateInstructionCount(SrcInt);
1124 if (Length > Threshold &&
1125 (((float)std::distance(mri_->use_nodbg_begin(SrcInt.reg),
1126 mri_->use_nodbg_end()) / Length) <
1131 // Must be implicit_def.
1134 // If the virtual register live interval is defined or cross a loop, turn
1135 // down aggressiveness.
1137 li_->getInstructionIndex(CopyMI).getDefIndex();
1138 SlotIndex UseIdx = CopyIdx.getUseIndex();
1139 LiveInterval::iterator SLR = SrcInt.FindLiveRangeContaining(UseIdx);
1140 assert(SLR != SrcInt.end() && "Live range not found!");
1141 SLR = SrcInt.FindLiveRangeContaining(SLR->start.getPrevSlot());
1142 if (SLR == SrcInt.end())
1144 MachineBasicBlock *SMBB = li_->getMBBFromIndex(SLR->start);
1145 const MachineLoop *L = loopInfo->getLoopFor(SMBB);
1147 if (!L || Length <= Threshold)
1150 if (loopInfo->getLoopFor(CopyMBB) != L) {
1151 if (SMBB != L->getLoopLatch())
1153 // If vr's live interval is extended from before the loop latch, do not
1155 for (MachineBasicBlock::pred_iterator PI = SMBB->pred_begin(),
1156 PE = SMBB->pred_end(); PI != PE; ++PI) {
1157 MachineBasicBlock *PredMBB = *PI;
1158 if (PredMBB == SMBB)
1160 if (SrcInt.overlaps(li_->getMBBStartIdx(PredMBB),
1161 li_->getMBBEndIdx(PredMBB)))
1168 /// isWinToJoinCrossClass - Return true if it's profitable to coalesce
1169 /// two virtual registers from different register classes.
1171 SimpleRegisterCoalescing::isWinToJoinCrossClass(unsigned SrcReg,
1173 const TargetRegisterClass *SrcRC,
1174 const TargetRegisterClass *DstRC,
1175 const TargetRegisterClass *NewRC) {
1176 unsigned NewRCCount = allocatableRCRegs_[NewRC].count();
1177 // This heuristics is good enough in practice, but it's obviously not *right*.
1178 // 4 is a magic number that works well enough for x86, ARM, etc. It filter
1179 // out all but the most restrictive register classes.
1180 if (NewRCCount > 4 ||
1181 // Early exit if the function is fairly small, coalesce aggressively if
1182 // that's the case. For really special register classes with 3 or
1183 // fewer registers, be a bit more careful.
1184 (li_->getFuncInstructionCount() / NewRCCount) < 8)
1186 LiveInterval &SrcInt = li_->getInterval(SrcReg);
1187 LiveInterval &DstInt = li_->getInterval(DstReg);
1188 unsigned SrcSize = li_->getApproximateInstructionCount(SrcInt);
1189 unsigned DstSize = li_->getApproximateInstructionCount(DstInt);
1190 if (SrcSize <= NewRCCount && DstSize <= NewRCCount)
1192 // Estimate *register use density*. If it doubles or more, abort.
1193 unsigned SrcUses = std::distance(mri_->use_nodbg_begin(SrcReg),
1194 mri_->use_nodbg_end());
1195 unsigned DstUses = std::distance(mri_->use_nodbg_begin(DstReg),
1196 mri_->use_nodbg_end());
1197 float NewDensity = ((float)(SrcUses + DstUses) / (SrcSize + DstSize)) /
1199 if (SrcRC != NewRC && SrcSize > NewRCCount) {
1200 unsigned SrcRCCount = allocatableRCRegs_[SrcRC].count();
1201 float Density = ((float)SrcUses / SrcSize) / SrcRCCount;
1202 if (NewDensity > Density * 2.0f)
1205 if (DstRC != NewRC && DstSize > NewRCCount) {
1206 unsigned DstRCCount = allocatableRCRegs_[DstRC].count();
1207 float Density = ((float)DstUses / DstSize) / DstRCCount;
1208 if (NewDensity > Density * 2.0f)
1214 /// HasIncompatibleSubRegDefUse - If we are trying to coalesce a virtual
1215 /// register with a physical register, check if any of the virtual register
1216 /// operand is a sub-register use or def. If so, make sure it won't result
1217 /// in an illegal extract_subreg or insert_subreg instruction. e.g.
1218 /// vr1024 = extract_subreg vr1025, 1
1220 /// vr1024 = mov8rr AH
1221 /// If vr1024 is coalesced with AH, the extract_subreg is now illegal since
1222 /// AH does not have a super-reg whose sub-register 1 is AH.
1224 SimpleRegisterCoalescing::HasIncompatibleSubRegDefUse(MachineInstr *CopyMI,
1227 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(VirtReg),
1228 E = mri_->reg_end(); I != E; ++I) {
1229 MachineOperand &O = I.getOperand();
1232 MachineInstr *MI = &*I;
1233 if (MI == CopyMI || JoinedCopies.count(MI))
1235 unsigned SubIdx = O.getSubReg();
1236 if (SubIdx && !tri_->getSubReg(PhysReg, SubIdx))
1238 if (MI->isExtractSubreg()) {
1239 SubIdx = MI->getOperand(2).getImm();
1240 if (O.isUse() && !tri_->getSubReg(PhysReg, SubIdx))
1243 unsigned SrcReg = MI->getOperand(1).getReg();
1244 const TargetRegisterClass *RC =
1245 TargetRegisterInfo::isPhysicalRegister(SrcReg)
1246 ? tri_->getPhysicalRegisterRegClass(SrcReg)
1247 : mri_->getRegClass(SrcReg);
1248 if (!tri_->getMatchingSuperReg(PhysReg, SubIdx, RC))
1252 if (MI->isInsertSubreg() || MI->isSubregToReg()) {
1253 SubIdx = MI->getOperand(3).getImm();
1254 if (VirtReg == MI->getOperand(0).getReg()) {
1255 if (!tri_->getSubReg(PhysReg, SubIdx))
1258 unsigned DstReg = MI->getOperand(0).getReg();
1259 const TargetRegisterClass *RC =
1260 TargetRegisterInfo::isPhysicalRegister(DstReg)
1261 ? tri_->getPhysicalRegisterRegClass(DstReg)
1262 : mri_->getRegClass(DstReg);
1263 if (!tri_->getMatchingSuperReg(PhysReg, SubIdx, RC))
1272 /// CanJoinExtractSubRegToPhysReg - Return true if it's possible to coalesce
1273 /// an extract_subreg where dst is a physical register, e.g.
1274 /// cl = EXTRACT_SUBREG reg1024, 1
1276 SimpleRegisterCoalescing::CanJoinExtractSubRegToPhysReg(unsigned DstReg,
1277 unsigned SrcReg, unsigned SubIdx,
1278 unsigned &RealDstReg) {
1279 const TargetRegisterClass *RC = mri_->getRegClass(SrcReg);
1280 RealDstReg = tri_->getMatchingSuperReg(DstReg, SubIdx, RC);
1281 assert(RealDstReg && "Invalid extract_subreg instruction!");
1283 LiveInterval &RHS = li_->getInterval(SrcReg);
1284 // For this type of EXTRACT_SUBREG, conservatively
1285 // check if the live interval of the source register interfere with the
1286 // actual super physical register we are trying to coalesce with.
1287 if (li_->hasInterval(RealDstReg) &&
1288 RHS.overlaps(li_->getInterval(RealDstReg))) {
1290 dbgs() << "Interfere with register ";
1291 li_->getInterval(RealDstReg).print(dbgs(), tri_);
1293 return false; // Not coalescable
1295 for (const unsigned* SR = tri_->getSubRegisters(RealDstReg); *SR; ++SR)
1296 // Do not check DstReg or its sub-register. JoinIntervals() will take care
1298 if (*SR != DstReg &&
1299 !tri_->isSubRegister(DstReg, *SR) &&
1300 li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
1302 dbgs() << "Interfere with sub-register ";
1303 li_->getInterval(*SR).print(dbgs(), tri_);
1305 return false; // Not coalescable
1310 /// CanJoinInsertSubRegToPhysReg - Return true if it's possible to coalesce
1311 /// an insert_subreg where src is a physical register, e.g.
1312 /// reg1024 = INSERT_SUBREG reg1024, c1, 0
1314 SimpleRegisterCoalescing::CanJoinInsertSubRegToPhysReg(unsigned DstReg,
1315 unsigned SrcReg, unsigned SubIdx,
1316 unsigned &RealSrcReg) {
1317 const TargetRegisterClass *RC = mri_->getRegClass(DstReg);
1318 RealSrcReg = tri_->getMatchingSuperReg(SrcReg, SubIdx, RC);
1319 assert(RealSrcReg && "Invalid extract_subreg instruction!");
1321 LiveInterval &LHS = li_->getInterval(DstReg);
1322 if (li_->hasInterval(RealSrcReg) &&
1323 LHS.overlaps(li_->getInterval(RealSrcReg))) {
1325 dbgs() << "Interfere with register ";
1326 li_->getInterval(RealSrcReg).print(dbgs(), tri_);
1328 return false; // Not coalescable
1330 for (const unsigned* SR = tri_->getSubRegisters(RealSrcReg); *SR; ++SR)
1331 // Do not check SrcReg or its sub-register. JoinIntervals() will take care
1333 if (*SR != SrcReg &&
1334 !tri_->isSubRegister(SrcReg, *SR) &&
1335 li_->hasInterval(*SR) && LHS.overlaps(li_->getInterval(*SR))) {
1337 dbgs() << "Interfere with sub-register ";
1338 li_->getInterval(*SR).print(dbgs(), tri_);
1340 return false; // Not coalescable
1345 /// getRegAllocPreference - Return register allocation preference register.
1347 static unsigned getRegAllocPreference(unsigned Reg, MachineFunction &MF,
1348 MachineRegisterInfo *MRI,
1349 const TargetRegisterInfo *TRI) {
1350 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1352 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
1353 return TRI->ResolveRegAllocHint(Hint.first, Hint.second, MF);
1356 /// JoinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
1357 /// which are the src/dst of the copy instruction CopyMI. This returns true
1358 /// if the copy was successfully coalesced away. If it is not currently
1359 /// possible to coalesce this interval, but it may be possible if other
1360 /// things get coalesced, then it returns true by reference in 'Again'.
1361 bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) {
1362 MachineInstr *CopyMI = TheCopy.MI;
1365 if (JoinedCopies.count(CopyMI) || ReMatCopies.count(CopyMI))
1366 return false; // Already done.
1368 DEBUG(dbgs() << li_->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
1370 unsigned SrcReg, DstReg, SrcSubIdx = 0, DstSubIdx = 0;
1371 bool isExtSubReg = CopyMI->isExtractSubreg();
1372 bool isInsSubReg = CopyMI->isInsertSubreg();
1373 bool isSubRegToReg = CopyMI->isSubregToReg();
1374 unsigned SubIdx = 0;
1376 DstReg = CopyMI->getOperand(0).getReg();
1377 DstSubIdx = CopyMI->getOperand(0).getSubReg();
1378 SrcReg = CopyMI->getOperand(1).getReg();
1379 SrcSubIdx = CopyMI->getOperand(2).getImm();
1380 } else if (isInsSubReg || isSubRegToReg) {
1381 DstReg = CopyMI->getOperand(0).getReg();
1382 DstSubIdx = CopyMI->getOperand(3).getImm();
1383 SrcReg = CopyMI->getOperand(2).getReg();
1384 SrcSubIdx = CopyMI->getOperand(2).getSubReg();
1385 if (SrcSubIdx && SrcSubIdx != DstSubIdx) {
1386 // r1025 = INSERT_SUBREG r1025, r1024<2>, 2 Then r1024 has already been
1387 // coalesced to a larger register so the subreg indices cancel out.
1388 DEBUG(dbgs() << "\tSource of insert_subreg or subreg_to_reg is already "
1389 "coalesced to another register.\n");
1390 return false; // Not coalescable.
1392 } else if (tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
1393 if (SrcSubIdx && DstSubIdx && SrcSubIdx != DstSubIdx) {
1394 // e.g. %reg16404:1<def> = MOV8rr %reg16412:2<kill>
1396 return false; // Not coalescable.
1399 llvm_unreachable("Unrecognized copy instruction!");
1402 // If they are already joined we continue.
1403 if (SrcReg == DstReg) {
1404 DEBUG(dbgs() << "\tCopy already coalesced.\n");
1405 return false; // Not coalescable.
1408 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
1409 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1411 // If they are both physical registers, we cannot join them.
1412 if (SrcIsPhys && DstIsPhys) {
1413 DEBUG(dbgs() << "\tCan not coalesce physregs.\n");
1414 return false; // Not coalescable.
1417 // We only join virtual registers with allocatable physical registers.
1418 if (SrcIsPhys && !allocatableRegs_[SrcReg]) {
1419 DEBUG(dbgs() << "\tSrc reg is unallocatable physreg.\n");
1420 return false; // Not coalescable.
1422 if (DstIsPhys && !allocatableRegs_[DstReg]) {
1423 DEBUG(dbgs() << "\tDst reg is unallocatable physreg.\n");
1424 return false; // Not coalescable.
1427 // Check that a physical source register is compatible with dst regclass
1429 unsigned SrcSubReg = SrcSubIdx ?
1430 tri_->getSubReg(SrcReg, SrcSubIdx) : SrcReg;
1431 const TargetRegisterClass *DstRC = mri_->getRegClass(DstReg);
1432 const TargetRegisterClass *DstSubRC = DstRC;
1434 DstSubRC = DstRC->getSubRegisterRegClass(DstSubIdx);
1435 assert(DstSubRC && "Illegal subregister index");
1436 if (!DstSubRC->contains(SrcSubReg)) {
1437 DEBUG(dbgs() << "\tIncompatible destination regclass: "
1438 << tri_->getName(SrcSubReg) << " not in "
1439 << DstSubRC->getName() << ".\n");
1440 return false; // Not coalescable.
1444 // Check that a physical dst register is compatible with source regclass
1446 unsigned DstSubReg = DstSubIdx ?
1447 tri_->getSubReg(DstReg, DstSubIdx) : DstReg;
1448 const TargetRegisterClass *SrcRC = mri_->getRegClass(SrcReg);
1449 const TargetRegisterClass *SrcSubRC = SrcRC;
1451 SrcSubRC = SrcRC->getSubRegisterRegClass(SrcSubIdx);
1452 assert(SrcSubRC && "Illegal subregister index");
1453 if (!SrcSubRC->contains(DstSubReg)) {
1454 DEBUG(dbgs() << "\tIncompatible source regclass: "
1455 << tri_->getName(DstSubReg) << " not in "
1456 << SrcSubRC->getName() << ".\n");
1458 return false; // Not coalescable.
1462 // Should be non-null only when coalescing to a sub-register class.
1463 bool CrossRC = false;
1464 const TargetRegisterClass *SrcRC= SrcIsPhys ? 0 : mri_->getRegClass(SrcReg);
1465 const TargetRegisterClass *DstRC= DstIsPhys ? 0 : mri_->getRegClass(DstReg);
1466 const TargetRegisterClass *NewRC = NULL;
1467 unsigned RealDstReg = 0;
1468 unsigned RealSrcReg = 0;
1469 if (isExtSubReg || isInsSubReg || isSubRegToReg) {
1470 SubIdx = CopyMI->getOperand(isExtSubReg ? 2 : 3).getImm();
1471 if (SrcIsPhys && isExtSubReg) {
1472 // r1024 = EXTRACT_SUBREG EAX, 0 then r1024 is really going to be
1473 // coalesced with AX.
1474 unsigned DstSubIdx = CopyMI->getOperand(0).getSubReg();
1476 // r1024<2> = EXTRACT_SUBREG EAX, 2. Then r1024 has already been
1477 // coalesced to a larger register so the subreg indices cancel out.
1478 if (DstSubIdx != SubIdx) {
1479 DEBUG(dbgs() << "\t Sub-register indices mismatch.\n");
1480 return false; // Not coalescable.
1483 SrcReg = tri_->getSubReg(SrcReg, SubIdx);
1485 } else if (DstIsPhys && (isInsSubReg || isSubRegToReg)) {
1486 // EAX = INSERT_SUBREG EAX, r1024, 0
1487 unsigned SrcSubIdx = CopyMI->getOperand(2).getSubReg();
1489 // EAX = INSERT_SUBREG EAX, r1024<2>, 2 Then r1024 has already been
1490 // coalesced to a larger register so the subreg indices cancel out.
1491 if (SrcSubIdx != SubIdx) {
1492 DEBUG(dbgs() << "\t Sub-register indices mismatch.\n");
1493 return false; // Not coalescable.
1496 DstReg = tri_->getSubReg(DstReg, SubIdx);
1498 } else if ((DstIsPhys && isExtSubReg) ||
1499 (SrcIsPhys && (isInsSubReg || isSubRegToReg))) {
1500 if (!isSubRegToReg && CopyMI->getOperand(1).getSubReg()) {
1501 DEBUG(dbgs() << "\tSrc of extract_subreg already coalesced with reg"
1502 << " of a super-class.\n");
1503 return false; // Not coalescable.
1506 // FIXME: The following checks are somewhat conservative. Perhaps a better
1507 // way to implement this is to treat this as coalescing a vr with the
1508 // super physical register.
1510 if (!CanJoinExtractSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealDstReg))
1511 return false; // Not coalescable
1513 if (!CanJoinInsertSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealSrcReg))
1514 return false; // Not coalescable
1518 unsigned OldSubIdx = isExtSubReg ? CopyMI->getOperand(0).getSubReg()
1519 : CopyMI->getOperand(2).getSubReg();
1521 if (OldSubIdx == SubIdx && !differingRegisterClasses(SrcReg, DstReg))
1522 // r1024<2> = EXTRACT_SUBREG r1025, 2. Then r1024 has already been
1523 // coalesced to a larger register so the subreg indices cancel out.
1524 // Also check if the other larger register is of the same register
1525 // class as the would be resulting register.
1528 DEBUG(dbgs() << "\t Sub-register indices mismatch.\n");
1529 return false; // Not coalescable.
1533 if (!DstIsPhys && !SrcIsPhys) {
1534 if (isInsSubReg || isSubRegToReg) {
1535 NewRC = tri_->getMatchingSuperRegClass(DstRC, SrcRC, SubIdx);
1536 } else // extract_subreg {
1537 NewRC = tri_->getMatchingSuperRegClass(SrcRC, DstRC, SubIdx);
1540 DEBUG(dbgs() << "\t Conflicting sub-register indices.\n");
1541 return false; // Not coalescable
1544 if (!isWinToJoinCrossClass(SrcReg, DstReg, SrcRC, DstRC, NewRC)) {
1545 DEBUG(dbgs() << "\tAvoid coalescing to constrainted register class: "
1546 << SrcRC->getName() << "/"
1547 << DstRC->getName() << " -> "
1548 << NewRC->getName() << ".\n");
1549 Again = true; // May be possible to coalesce later.
1554 } else if (differingRegisterClasses(SrcReg, DstReg)) {
1555 if (DisableCrossClassJoin)
1559 // FIXME: What if the result of a EXTRACT_SUBREG is then coalesced
1560 // with another? If it's the resulting destination register, then
1561 // the subidx must be propagated to uses (but only those defined
1562 // by the EXTRACT_SUBREG). If it's being coalesced into another
1563 // register, it should be safe because register is assumed to have
1564 // the register class of the super-register.
1566 // Process moves where one of the registers have a sub-register index.
1567 MachineOperand *DstMO = CopyMI->findRegisterDefOperand(DstReg);
1568 MachineOperand *SrcMO = CopyMI->findRegisterUseOperand(SrcReg);
1569 SubIdx = DstMO->getSubReg();
1571 if (SrcMO->getSubReg())
1572 // FIXME: can we handle this?
1574 // This is not an insert_subreg but it looks like one.
1575 // e.g. %reg1024:4 = MOV32rr %EAX
1578 if (!CanJoinInsertSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealSrcReg))
1579 return false; // Not coalescable
1583 SubIdx = SrcMO->getSubReg();
1585 // This is not a extract_subreg but it looks like one.
1586 // e.g. %cl = MOV16rr %reg1024:1
1589 if (!CanJoinExtractSubRegToPhysReg(DstReg, SrcReg, SubIdx,RealDstReg))
1590 return false; // Not coalescable
1596 // Now determine the register class of the joined register.
1597 if (!SrcIsPhys && !DstIsPhys) {
1600 SubIdx ? tri_->getMatchingSuperRegClass(SrcRC, DstRC, SubIdx) : SrcRC;
1601 } else if (isInsSubReg) {
1603 SubIdx ? tri_->getMatchingSuperRegClass(DstRC, SrcRC, SubIdx) : DstRC;
1605 NewRC = getCommonSubClass(SrcRC, DstRC);
1609 DEBUG(dbgs() << "\tDisjoint regclasses: "
1610 << SrcRC->getName() << ", "
1611 << DstRC->getName() << ".\n");
1612 return false; // Not coalescable.
1615 // If we are joining two virtual registers and the resulting register
1616 // class is more restrictive (fewer register, smaller size). Check if it's
1617 // worth doing the merge.
1618 if (!isWinToJoinCrossClass(SrcReg, DstReg, SrcRC, DstRC, NewRC)) {
1619 DEBUG(dbgs() << "\tAvoid coalescing to constrainted register class: "
1620 << SrcRC->getName() << "/"
1621 << DstRC->getName() << " -> "
1622 << NewRC->getName() << ".\n");
1623 // Allow the coalescer to try again in case either side gets coalesced to
1624 // a physical register that's compatible with the other side. e.g.
1625 // r1024 = MOV32to32_ r1025
1626 // But later r1024 is assigned EAX then r1025 may be coalesced with EAX.
1627 Again = true; // May be possible to coalesce later.
1633 // Will it create illegal extract_subreg / insert_subreg?
1634 if (SrcIsPhys && HasIncompatibleSubRegDefUse(CopyMI, DstReg, SrcReg))
1636 if (DstIsPhys && HasIncompatibleSubRegDefUse(CopyMI, SrcReg, DstReg))
1639 LiveInterval &SrcInt = li_->getInterval(SrcReg);
1640 LiveInterval &DstInt = li_->getInterval(DstReg);
1641 assert(SrcInt.reg == SrcReg && DstInt.reg == DstReg &&
1642 "Register mapping is horribly broken!");
1645 dbgs() << "\t\tInspecting "; SrcInt.print(dbgs(), tri_);
1646 dbgs() << " and "; DstInt.print(dbgs(), tri_);
1650 // Save a copy of the virtual register live interval. We'll manually
1651 // merge this into the "real" physical register live interval this is
1653 OwningPtr<LiveInterval> SavedLI;
1655 SavedLI.reset(li_->dupInterval(&SrcInt));
1656 else if (RealSrcReg)
1657 SavedLI.reset(li_->dupInterval(&DstInt));
1659 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg) {
1660 // Check if it is necessary to propagate "isDead" property.
1661 MachineOperand *mopd = CopyMI->findRegisterDefOperand(DstReg, false);
1662 bool isDead = mopd->isDead();
1664 // We need to be careful about coalescing a source physical register with a
1665 // virtual register. Once the coalescing is done, it cannot be broken and
1666 // these are not spillable! If the destination interval uses are far away,
1667 // think twice about coalescing them!
1668 if (!isDead && (SrcIsPhys || DstIsPhys)) {
1669 // If the virtual register live interval is long but it has low use
1670 // density, do not join them, instead mark the physical register as its
1671 // allocation preference.
1672 LiveInterval &JoinVInt = SrcIsPhys ? DstInt : SrcInt;
1673 LiveInterval &JoinPInt = SrcIsPhys ? SrcInt : DstInt;
1674 unsigned JoinVReg = SrcIsPhys ? DstReg : SrcReg;
1675 unsigned JoinPReg = SrcIsPhys ? SrcReg : DstReg;
1677 // Don't join with physregs that have a ridiculous number of live
1678 // ranges. The data structure performance is really bad when that
1680 if (JoinPInt.ranges.size() > 1000) {
1681 mri_->setRegAllocationHint(JoinVInt.reg, 0, JoinPReg);
1684 << "\tPhysical register live interval too complicated, abort!\n");
1688 const TargetRegisterClass *RC = mri_->getRegClass(JoinVReg);
1689 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1690 unsigned Length = li_->getApproximateInstructionCount(JoinVInt);
1691 float Ratio = 1.0 / Threshold;
1692 if (Length > Threshold &&
1693 (((float)std::distance(mri_->use_nodbg_begin(JoinVReg),
1694 mri_->use_nodbg_end()) / Length) < Ratio)) {
1695 // Before giving up coalescing, if definition of source is defined by
1696 // trivial computation, try rematerializing it.
1697 if (ReMaterializeTrivialDef(SrcInt, DstReg, DstSubIdx, CopyMI))
1700 mri_->setRegAllocationHint(JoinVInt.reg, 0, JoinPReg);
1702 DEBUG(dbgs() << "\tMay tie down a physical register, abort!\n");
1703 Again = true; // May be possible to coalesce later.
1709 // Okay, attempt to join these two intervals. On failure, this returns false.
1710 // Otherwise, if one of the intervals being joined is a physreg, this method
1711 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1712 // been modified, so we can use this information below to update aliases.
1713 bool Swapped = false;
1714 // If SrcInt is implicitly defined, it's safe to coalesce.
1715 if (SrcInt.empty()) {
1716 if (!CanCoalesceWithImpDef(CopyMI, DstInt, SrcInt)) {
1717 // Only coalesce an empty interval (defined by implicit_def) with
1718 // another interval which has a valno defined by the CopyMI and the CopyMI
1719 // is a kill of the implicit def.
1720 DEBUG(dbgs() << "Not profitable!\n");
1723 } else if (!JoinIntervals(DstInt, SrcInt, Swapped)) {
1724 // Coalescing failed.
1726 // If definition of source is defined by trivial computation, try
1727 // rematerializing it.
1728 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg &&
1729 ReMaterializeTrivialDef(SrcInt, DstReg, DstSubIdx, CopyMI))
1732 // If we can eliminate the copy without merging the live ranges, do so now.
1733 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg &&
1734 (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI) ||
1735 RemoveCopyByCommutingDef(SrcInt, DstInt, CopyMI))) {
1736 JoinedCopies.insert(CopyMI);
1737 DEBUG(dbgs() << "Trivial!\n");
1741 // Otherwise, we are unable to join the intervals.
1742 DEBUG(dbgs() << "Interference!\n");
1743 Again = true; // May be possible to coalesce later.
1747 LiveInterval *ResSrcInt = &SrcInt;
1748 LiveInterval *ResDstInt = &DstInt;
1750 std::swap(SrcReg, DstReg);
1751 std::swap(ResSrcInt, ResDstInt);
1753 assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1754 "LiveInterval::join didn't work right!");
1756 // If we're about to merge live ranges into a physical register live interval,
1757 // we have to update any aliased register's live ranges to indicate that they
1758 // have clobbered values for this range.
1759 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
1760 // If this is a extract_subreg where dst is a physical register, e.g.
1761 // cl = EXTRACT_SUBREG reg1024, 1
1762 // then create and update the actual physical register allocated to RHS.
1763 if (RealDstReg || RealSrcReg) {
1764 LiveInterval &RealInt =
1765 li_->getOrCreateInterval(RealDstReg ? RealDstReg : RealSrcReg);
1766 for (LiveInterval::const_vni_iterator I = SavedLI->vni_begin(),
1767 E = SavedLI->vni_end(); I != E; ++I) {
1768 const VNInfo *ValNo = *I;
1769 VNInfo *NewValNo = RealInt.getNextValue(ValNo->def, ValNo->getCopy(),
1770 false, // updated at *
1771 li_->getVNInfoAllocator());
1772 NewValNo->setFlags(ValNo->getFlags()); // * updated here.
1773 RealInt.addKills(NewValNo, ValNo->kills);
1774 RealInt.MergeValueInAsValue(*SavedLI, ValNo, NewValNo);
1776 RealInt.weight += SavedLI->weight;
1777 DstReg = RealDstReg ? RealDstReg : RealSrcReg;
1780 // Update the liveintervals of sub-registers.
1781 for (const unsigned *AS = tri_->getSubRegisters(DstReg); *AS; ++AS)
1782 li_->getOrCreateInterval(*AS).MergeInClobberRanges(*li_, *ResSrcInt,
1783 li_->getVNInfoAllocator());
1786 // If this is a EXTRACT_SUBREG, make sure the result of coalescing is the
1787 // larger super-register.
1788 if ((isExtSubReg || isInsSubReg || isSubRegToReg) &&
1789 !SrcIsPhys && !DstIsPhys) {
1790 if ((isExtSubReg && !Swapped) ||
1791 ((isInsSubReg || isSubRegToReg) && Swapped)) {
1792 ResSrcInt->Copy(*ResDstInt, mri_, li_->getVNInfoAllocator());
1793 std::swap(SrcReg, DstReg);
1794 std::swap(ResSrcInt, ResDstInt);
1798 // Coalescing to a virtual register that is of a sub-register class of the
1799 // other. Make sure the resulting register is set to the right register class.
1803 // This may happen even if it's cross-rc coalescing. e.g.
1804 // %reg1026<def> = SUBREG_TO_REG 0, %reg1037<kill>, 4
1805 // reg1026 -> GR64, reg1037 -> GR32_ABCD. The resulting register will have to
1806 // be allocate a register from GR64_ABCD.
1808 mri_->setRegClass(DstReg, NewRC);
1810 // Remember to delete the copy instruction.
1811 JoinedCopies.insert(CopyMI);
1813 // Some live range has been lengthened due to colaescing, eliminate the
1814 // unnecessary kills.
1815 RemoveUnnecessaryKills(SrcReg, *ResDstInt);
1816 if (TargetRegisterInfo::isVirtualRegister(DstReg))
1817 RemoveUnnecessaryKills(DstReg, *ResDstInt);
1819 UpdateRegDefsUses(SrcReg, DstReg, SubIdx);
1821 // If we have extended the live range of a physical register, make sure we
1822 // update live-in lists as well.
1823 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
1824 const LiveInterval &VRegInterval = li_->getInterval(SrcReg);
1825 SmallVector<MachineBasicBlock*, 16> BlockSeq;
1826 for (LiveInterval::const_iterator I = VRegInterval.begin(),
1827 E = VRegInterval.end(); I != E; ++I ) {
1828 li_->findLiveInMBBs(I->start, I->end, BlockSeq);
1829 for (unsigned idx = 0, size = BlockSeq.size(); idx != size; ++idx) {
1830 MachineBasicBlock &block = *BlockSeq[idx];
1831 if (!block.isLiveIn(DstReg))
1832 block.addLiveIn(DstReg);
1838 // SrcReg is guarateed to be the register whose live interval that is
1840 li_->removeInterval(SrcReg);
1842 // Update regalloc hint.
1843 tri_->UpdateRegAllocHint(SrcReg, DstReg, *mf_);
1845 // Manually deleted the live interval copy.
1851 // If resulting interval has a preference that no longer fits because of subreg
1852 // coalescing, just clear the preference.
1853 unsigned Preference = getRegAllocPreference(ResDstInt->reg, *mf_, mri_, tri_);
1854 if (Preference && (isExtSubReg || isInsSubReg || isSubRegToReg) &&
1855 TargetRegisterInfo::isVirtualRegister(ResDstInt->reg)) {
1856 const TargetRegisterClass *RC = mri_->getRegClass(ResDstInt->reg);
1857 if (!RC->contains(Preference))
1858 mri_->setRegAllocationHint(ResDstInt->reg, 0, 0);
1862 dbgs() << "\n\t\tJoined. Result = ";
1863 ResDstInt->print(dbgs(), tri_);
1871 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
1872 /// compute what the resultant value numbers for each value in the input two
1873 /// ranges will be. This is complicated by copies between the two which can
1874 /// and will commonly cause multiple value numbers to be merged into one.
1876 /// VN is the value number that we're trying to resolve. InstDefiningValue
1877 /// keeps track of the new InstDefiningValue assignment for the result
1878 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
1879 /// whether a value in this or other is a copy from the opposite set.
1880 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
1881 /// already been assigned.
1883 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
1884 /// contains the value number the copy is from.
1886 static unsigned ComputeUltimateVN(VNInfo *VNI,
1887 SmallVector<VNInfo*, 16> &NewVNInfo,
1888 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
1889 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
1890 SmallVector<int, 16> &ThisValNoAssignments,
1891 SmallVector<int, 16> &OtherValNoAssignments) {
1892 unsigned VN = VNI->id;
1894 // If the VN has already been computed, just return it.
1895 if (ThisValNoAssignments[VN] >= 0)
1896 return ThisValNoAssignments[VN];
1897 assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers");
1899 // If this val is not a copy from the other val, then it must be a new value
1900 // number in the destination.
1901 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
1902 if (I == ThisFromOther.end()) {
1903 NewVNInfo.push_back(VNI);
1904 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
1906 VNInfo *OtherValNo = I->second;
1908 // Otherwise, this *is* a copy from the RHS. If the other side has already
1909 // been computed, return it.
1910 if (OtherValNoAssignments[OtherValNo->id] >= 0)
1911 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
1913 // Mark this value number as currently being computed, then ask what the
1914 // ultimate value # of the other value is.
1915 ThisValNoAssignments[VN] = -2;
1916 unsigned UltimateVN =
1917 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
1918 OtherValNoAssignments, ThisValNoAssignments);
1919 return ThisValNoAssignments[VN] = UltimateVN;
1922 static bool InVector(VNInfo *Val, const SmallVector<VNInfo*, 8> &V) {
1923 return std::find(V.begin(), V.end(), Val) != V.end();
1926 static bool isValNoDefMove(const MachineInstr *MI, unsigned DR, unsigned SR,
1927 const TargetInstrInfo *TII,
1928 const TargetRegisterInfo *TRI) {
1929 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
1930 if (TII->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
1932 else if (MI->isExtractSubreg()) {
1933 DstReg = MI->getOperand(0).getReg();
1934 SrcReg = MI->getOperand(1).getReg();
1935 } else if (MI->isSubregToReg() ||
1936 MI->isInsertSubreg()) {
1937 DstReg = MI->getOperand(0).getReg();
1938 SrcReg = MI->getOperand(2).getReg();
1941 return (SrcReg == SR || TRI->isSuperRegister(SR, SrcReg)) &&
1942 (DstReg == DR || TRI->isSuperRegister(DR, DstReg));
1945 /// RangeIsDefinedByCopyFromReg - Return true if the specified live range of
1946 /// the specified live interval is defined by a copy from the specified
1948 bool SimpleRegisterCoalescing::RangeIsDefinedByCopyFromReg(LiveInterval &li,
1951 unsigned SrcReg = li_->getVNInfoSourceReg(LR->valno);
1954 // FIXME: Do isPHIDef and isDefAccurate both need to be tested?
1955 if ((LR->valno->isPHIDef() || !LR->valno->isDefAccurate()) &&
1956 TargetRegisterInfo::isPhysicalRegister(li.reg) &&
1957 *tri_->getSuperRegisters(li.reg)) {
1958 // It's a sub-register live interval, we may not have precise information.
1960 MachineInstr *DefMI = li_->getInstructionFromIndex(LR->start);
1961 if (DefMI && isValNoDefMove(DefMI, li.reg, Reg, tii_, tri_)) {
1962 // Cache computed info.
1963 LR->valno->def = LR->start;
1964 LR->valno->setCopy(DefMI);
1972 /// ValueLiveAt - Return true if the LiveRange pointed to by the given
1973 /// iterator, or any subsequent range with the same value number,
1974 /// is live at the given point.
1975 bool SimpleRegisterCoalescing::ValueLiveAt(LiveInterval::iterator LRItr,
1976 LiveInterval::iterator LREnd,
1977 SlotIndex defPoint) const {
1978 for (const VNInfo *valno = LRItr->valno;
1979 (LRItr != LREnd) && (LRItr->valno == valno); ++LRItr) {
1980 if (LRItr->contains(defPoint))
1988 /// SimpleJoin - Attempt to joint the specified interval into this one. The
1989 /// caller of this method must guarantee that the RHS only contains a single
1990 /// value number and that the RHS is not defined by a copy from this
1991 /// interval. This returns false if the intervals are not joinable, or it
1992 /// joins them and returns true.
1993 bool SimpleRegisterCoalescing::SimpleJoin(LiveInterval &LHS, LiveInterval &RHS){
1994 assert(RHS.containsOneValue());
1996 // Some number (potentially more than one) value numbers in the current
1997 // interval may be defined as copies from the RHS. Scan the overlapping
1998 // portions of the LHS and RHS, keeping track of this and looking for
1999 // overlapping live ranges that are NOT defined as copies. If these exist, we
2002 LiveInterval::iterator LHSIt = LHS.begin(), LHSEnd = LHS.end();
2003 LiveInterval::iterator RHSIt = RHS.begin(), RHSEnd = RHS.end();
2005 if (LHSIt->start < RHSIt->start) {
2006 LHSIt = std::upper_bound(LHSIt, LHSEnd, RHSIt->start);
2007 if (LHSIt != LHS.begin()) --LHSIt;
2008 } else if (RHSIt->start < LHSIt->start) {
2009 RHSIt = std::upper_bound(RHSIt, RHSEnd, LHSIt->start);
2010 if (RHSIt != RHS.begin()) --RHSIt;
2013 SmallVector<VNInfo*, 8> EliminatedLHSVals;
2016 // Determine if these live intervals overlap.
2017 bool Overlaps = false;
2018 if (LHSIt->start <= RHSIt->start)
2019 Overlaps = LHSIt->end > RHSIt->start;
2021 Overlaps = RHSIt->end > LHSIt->start;
2023 // If the live intervals overlap, there are two interesting cases: if the
2024 // LHS interval is defined by a copy from the RHS, it's ok and we record
2025 // that the LHS value # is the same as the RHS. If it's not, then we cannot
2026 // coalesce these live ranges and we bail out.
2028 // If we haven't already recorded that this value # is safe, check it.
2029 if (!InVector(LHSIt->valno, EliminatedLHSVals)) {
2030 // If it's re-defined by an early clobber somewhere in the live range,
2031 // then conservatively abort coalescing.
2032 if (LHSIt->valno->hasRedefByEC())
2034 // Copy from the RHS?
2035 if (!RangeIsDefinedByCopyFromReg(LHS, LHSIt, RHS.reg))
2036 return false; // Nope, bail out.
2038 if (ValueLiveAt(LHSIt, LHS.end(), RHSIt->valno->def))
2039 // Here is an interesting situation:
2041 // vr1025 = copy vr1024
2046 // Even though vr1025 is copied from vr1024, it's not safe to
2047 // coalesce them since the live range of vr1025 intersects the
2048 // def of vr1024. This happens because vr1025 is assigned the
2049 // value of the previous iteration of vr1024.
2051 EliminatedLHSVals.push_back(LHSIt->valno);
2054 // We know this entire LHS live range is okay, so skip it now.
2055 if (++LHSIt == LHSEnd) break;
2059 if (LHSIt->end < RHSIt->end) {
2060 if (++LHSIt == LHSEnd) break;
2062 // One interesting case to check here. It's possible that we have
2063 // something like "X3 = Y" which defines a new value number in the LHS,
2064 // and is the last use of this liverange of the RHS. In this case, we
2065 // want to notice this copy (so that it gets coalesced away) even though
2066 // the live ranges don't actually overlap.
2067 if (LHSIt->start == RHSIt->end) {
2068 if (InVector(LHSIt->valno, EliminatedLHSVals)) {
2069 // We already know that this value number is going to be merged in
2070 // if coalescing succeeds. Just skip the liverange.
2071 if (++LHSIt == LHSEnd) break;
2073 // If it's re-defined by an early clobber somewhere in the live range,
2074 // then conservatively abort coalescing.
2075 if (LHSIt->valno->hasRedefByEC())
2077 // Otherwise, if this is a copy from the RHS, mark it as being merged
2079 if (RangeIsDefinedByCopyFromReg(LHS, LHSIt, RHS.reg)) {
2080 if (ValueLiveAt(LHSIt, LHS.end(), RHSIt->valno->def))
2081 // Here is an interesting situation:
2083 // vr1025 = copy vr1024
2088 // Even though vr1025 is copied from vr1024, it's not safe to
2089 // coalesced them since live range of vr1025 intersects the
2090 // def of vr1024. This happens because vr1025 is assigned the
2091 // value of the previous iteration of vr1024.
2093 EliminatedLHSVals.push_back(LHSIt->valno);
2095 // We know this entire LHS live range is okay, so skip it now.
2096 if (++LHSIt == LHSEnd) break;
2101 if (++RHSIt == RHSEnd) break;
2105 // If we got here, we know that the coalescing will be successful and that
2106 // the value numbers in EliminatedLHSVals will all be merged together. Since
2107 // the most common case is that EliminatedLHSVals has a single number, we
2108 // optimize for it: if there is more than one value, we merge them all into
2109 // the lowest numbered one, then handle the interval as if we were merging
2110 // with one value number.
2111 VNInfo *LHSValNo = NULL;
2112 if (EliminatedLHSVals.size() > 1) {
2113 // Loop through all the equal value numbers merging them into the smallest
2115 VNInfo *Smallest = EliminatedLHSVals[0];
2116 for (unsigned i = 1, e = EliminatedLHSVals.size(); i != e; ++i) {
2117 if (EliminatedLHSVals[i]->id < Smallest->id) {
2118 // Merge the current notion of the smallest into the smaller one.
2119 LHS.MergeValueNumberInto(Smallest, EliminatedLHSVals[i]);
2120 Smallest = EliminatedLHSVals[i];
2122 // Merge into the smallest.
2123 LHS.MergeValueNumberInto(EliminatedLHSVals[i], Smallest);
2126 LHSValNo = Smallest;
2127 } else if (EliminatedLHSVals.empty()) {
2128 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
2129 *tri_->getSuperRegisters(LHS.reg))
2130 // Imprecise sub-register information. Can't handle it.
2132 llvm_unreachable("No copies from the RHS?");
2134 LHSValNo = EliminatedLHSVals[0];
2137 // Okay, now that there is a single LHS value number that we're merging the
2138 // RHS into, update the value number info for the LHS to indicate that the
2139 // value number is defined where the RHS value number was.
2140 const VNInfo *VNI = RHS.getValNumInfo(0);
2141 LHSValNo->def = VNI->def;
2142 LHSValNo->setCopy(VNI->getCopy());
2144 // Okay, the final step is to loop over the RHS live intervals, adding them to
2146 if (VNI->hasPHIKill())
2147 LHSValNo->setHasPHIKill(true);
2148 LHS.addKills(LHSValNo, VNI->kills);
2149 LHS.MergeRangesInAsValue(RHS, LHSValNo);
2151 LHS.ComputeJoinedWeight(RHS);
2153 // Update regalloc hint if both are virtual registers.
2154 if (TargetRegisterInfo::isVirtualRegister(LHS.reg) &&
2155 TargetRegisterInfo::isVirtualRegister(RHS.reg)) {
2156 std::pair<unsigned, unsigned> RHSPref = mri_->getRegAllocationHint(RHS.reg);
2157 std::pair<unsigned, unsigned> LHSPref = mri_->getRegAllocationHint(LHS.reg);
2158 if (RHSPref != LHSPref)
2159 mri_->setRegAllocationHint(LHS.reg, RHSPref.first, RHSPref.second);
2162 // Update the liveintervals of sub-registers.
2163 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg))
2164 for (const unsigned *AS = tri_->getSubRegisters(LHS.reg); *AS; ++AS)
2165 li_->getOrCreateInterval(*AS).MergeInClobberRanges(*li_, LHS,
2166 li_->getVNInfoAllocator());
2171 /// JoinIntervals - Attempt to join these two intervals. On failure, this
2172 /// returns false. Otherwise, if one of the intervals being joined is a
2173 /// physreg, this method always canonicalizes LHS to be it. The output
2174 /// "RHS" will not have been modified, so we can use this information
2175 /// below to update aliases.
2177 SimpleRegisterCoalescing::JoinIntervals(LiveInterval &LHS, LiveInterval &RHS,
2179 // Compute the final value assignment, assuming that the live ranges can be
2181 SmallVector<int, 16> LHSValNoAssignments;
2182 SmallVector<int, 16> RHSValNoAssignments;
2183 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
2184 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
2185 SmallVector<VNInfo*, 16> NewVNInfo;
2187 // If a live interval is a physical register, conservatively check if any
2188 // of its sub-registers is overlapping the live interval of the virtual
2189 // register. If so, do not coalesce.
2190 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
2191 *tri_->getSubRegisters(LHS.reg)) {
2192 // If it's coalescing a virtual register to a physical register, estimate
2193 // its live interval length. This is the *cost* of scanning an entire live
2194 // interval. If the cost is low, we'll do an exhaustive check instead.
2196 // If this is something like this:
2204 // That is, the live interval of v1024 crosses a bb. Then we can't rely on
2205 // less conservative check. It's possible a sub-register is defined before
2206 // v1024 (or live in) and live out of BB1.
2207 if (RHS.containsOneValue() &&
2208 li_->intervalIsInOneMBB(RHS) &&
2209 li_->getApproximateInstructionCount(RHS) <= 10) {
2210 // Perform a more exhaustive check for some common cases.
2211 if (li_->conflictsWithSubPhysRegRef(RHS, LHS.reg, true, JoinedCopies))
2214 for (const unsigned* SR = tri_->getSubRegisters(LHS.reg); *SR; ++SR)
2215 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
2217 dbgs() << "Interfere with sub-register ";
2218 li_->getInterval(*SR).print(dbgs(), tri_);
2223 } else if (TargetRegisterInfo::isPhysicalRegister(RHS.reg) &&
2224 *tri_->getSubRegisters(RHS.reg)) {
2225 if (LHS.containsOneValue() &&
2226 li_->getApproximateInstructionCount(LHS) <= 10) {
2227 // Perform a more exhaustive check for some common cases.
2228 if (li_->conflictsWithSubPhysRegRef(LHS, RHS.reg, false, JoinedCopies))
2231 for (const unsigned* SR = tri_->getSubRegisters(RHS.reg); *SR; ++SR)
2232 if (li_->hasInterval(*SR) && LHS.overlaps(li_->getInterval(*SR))) {
2234 dbgs() << "Interfere with sub-register ";
2235 li_->getInterval(*SR).print(dbgs(), tri_);
2242 // Compute ultimate value numbers for the LHS and RHS values.
2243 if (RHS.containsOneValue()) {
2244 // Copies from a liveinterval with a single value are simple to handle and
2245 // very common, handle the special case here. This is important, because
2246 // often RHS is small and LHS is large (e.g. a physreg).
2248 // Find out if the RHS is defined as a copy from some value in the LHS.
2249 int RHSVal0DefinedFromLHS = -1;
2251 VNInfo *RHSValNoInfo = NULL;
2252 VNInfo *RHSValNoInfo0 = RHS.getValNumInfo(0);
2253 unsigned RHSSrcReg = li_->getVNInfoSourceReg(RHSValNoInfo0);
2254 if (RHSSrcReg == 0 || RHSSrcReg != LHS.reg) {
2255 // If RHS is not defined as a copy from the LHS, we can use simpler and
2256 // faster checks to see if the live ranges are coalescable. This joiner
2257 // can't swap the LHS/RHS intervals though.
2258 if (!TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
2259 return SimpleJoin(LHS, RHS);
2261 RHSValNoInfo = RHSValNoInfo0;
2264 // It was defined as a copy from the LHS, find out what value # it is.
2266 LHS.getLiveRangeContaining(RHSValNoInfo0->def.getPrevSlot())->valno;
2267 RHSValID = RHSValNoInfo->id;
2268 RHSVal0DefinedFromLHS = RHSValID;
2271 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
2272 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
2273 NewVNInfo.resize(LHS.getNumValNums(), NULL);
2275 // Okay, *all* of the values in LHS that are defined as a copy from RHS
2276 // should now get updated.
2277 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2280 unsigned VN = VNI->id;
2281 if (unsigned LHSSrcReg = li_->getVNInfoSourceReg(VNI)) {
2282 if (LHSSrcReg != RHS.reg) {
2283 // If this is not a copy from the RHS, its value number will be
2284 // unmodified by the coalescing.
2285 NewVNInfo[VN] = VNI;
2286 LHSValNoAssignments[VN] = VN;
2287 } else if (RHSValID == -1) {
2288 // Otherwise, it is a copy from the RHS, and we don't already have a
2289 // value# for it. Keep the current value number, but remember it.
2290 LHSValNoAssignments[VN] = RHSValID = VN;
2291 NewVNInfo[VN] = RHSValNoInfo;
2292 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
2294 // Otherwise, use the specified value #.
2295 LHSValNoAssignments[VN] = RHSValID;
2296 if (VN == (unsigned)RHSValID) { // Else this val# is dead.
2297 NewVNInfo[VN] = RHSValNoInfo;
2298 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
2302 NewVNInfo[VN] = VNI;
2303 LHSValNoAssignments[VN] = VN;
2307 assert(RHSValID != -1 && "Didn't find value #?");
2308 RHSValNoAssignments[0] = RHSValID;
2309 if (RHSVal0DefinedFromLHS != -1) {
2310 // This path doesn't go through ComputeUltimateVN so just set
2312 RHSValsDefinedFromLHS[RHSValNoInfo0] = (VNInfo*)1;
2315 // Loop over the value numbers of the LHS, seeing if any are defined from
2317 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2320 if (VNI->isUnused() || VNI->getCopy() == 0) // Src not defined by a copy?
2323 // DstReg is known to be a register in the LHS interval. If the src is
2324 // from the RHS interval, we can use its value #.
2325 if (li_->getVNInfoSourceReg(VNI) != RHS.reg)
2328 // Figure out the value # from the RHS.
2329 LiveRange *lr = RHS.getLiveRangeContaining(VNI->def.getPrevSlot());
2330 assert(lr && "Cannot find live range");
2331 LHSValsDefinedFromRHS[VNI] = lr->valno;
2334 // Loop over the value numbers of the RHS, seeing if any are defined from
2336 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
2339 if (VNI->isUnused() || VNI->getCopy() == 0) // Src not defined by a copy?
2342 // DstReg is known to be a register in the RHS interval. If the src is
2343 // from the LHS interval, we can use its value #.
2344 if (li_->getVNInfoSourceReg(VNI) != LHS.reg)
2347 // Figure out the value # from the LHS.
2348 LiveRange *lr = LHS.getLiveRangeContaining(VNI->def.getPrevSlot());
2349 assert(lr && "Cannot find live range");
2350 RHSValsDefinedFromLHS[VNI] = lr->valno;
2353 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
2354 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
2355 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
2357 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2360 unsigned VN = VNI->id;
2361 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused())
2363 ComputeUltimateVN(VNI, NewVNInfo,
2364 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
2365 LHSValNoAssignments, RHSValNoAssignments);
2367 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
2370 unsigned VN = VNI->id;
2371 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused())
2373 // If this value number isn't a copy from the LHS, it's a new number.
2374 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
2375 NewVNInfo.push_back(VNI);
2376 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
2380 ComputeUltimateVN(VNI, NewVNInfo,
2381 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
2382 RHSValNoAssignments, LHSValNoAssignments);
2386 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
2387 // interval lists to see if these intervals are coalescable.
2388 LiveInterval::const_iterator I = LHS.begin();
2389 LiveInterval::const_iterator IE = LHS.end();
2390 LiveInterval::const_iterator J = RHS.begin();
2391 LiveInterval::const_iterator JE = RHS.end();
2393 // Skip ahead until the first place of potential sharing.
2394 if (I->start < J->start) {
2395 I = std::upper_bound(I, IE, J->start);
2396 if (I != LHS.begin()) --I;
2397 } else if (J->start < I->start) {
2398 J = std::upper_bound(J, JE, I->start);
2399 if (J != RHS.begin()) --J;
2403 // Determine if these two live ranges overlap.
2405 if (I->start < J->start) {
2406 Overlaps = I->end > J->start;
2408 Overlaps = J->end > I->start;
2411 // If so, check value # info to determine if they are really different.
2413 // If the live range overlap will map to the same value number in the
2414 // result liverange, we can still coalesce them. If not, we can't.
2415 if (LHSValNoAssignments[I->valno->id] !=
2416 RHSValNoAssignments[J->valno->id])
2418 // If it's re-defined by an early clobber somewhere in the live range,
2419 // then conservatively abort coalescing.
2420 if (NewVNInfo[LHSValNoAssignments[I->valno->id]]->hasRedefByEC())
2424 if (I->end < J->end) {
2433 // Update kill info. Some live ranges are extended due to copy coalescing.
2434 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
2435 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
2436 VNInfo *VNI = I->first;
2437 unsigned LHSValID = LHSValNoAssignments[VNI->id];
2438 NewVNInfo[LHSValID]->removeKill(VNI->def);
2439 if (VNI->hasPHIKill())
2440 NewVNInfo[LHSValID]->setHasPHIKill(true);
2441 RHS.addKills(NewVNInfo[LHSValID], VNI->kills);
2444 // Update kill info. Some live ranges are extended due to copy coalescing.
2445 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
2446 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
2447 VNInfo *VNI = I->first;
2448 unsigned RHSValID = RHSValNoAssignments[VNI->id];
2449 NewVNInfo[RHSValID]->removeKill(VNI->def);
2450 if (VNI->hasPHIKill())
2451 NewVNInfo[RHSValID]->setHasPHIKill(true);
2452 LHS.addKills(NewVNInfo[RHSValID], VNI->kills);
2455 // If we get here, we know that we can coalesce the live ranges. Ask the
2456 // intervals to coalesce themselves now.
2457 if ((RHS.ranges.size() > LHS.ranges.size() &&
2458 TargetRegisterInfo::isVirtualRegister(LHS.reg)) ||
2459 TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
2460 RHS.join(LHS, &RHSValNoAssignments[0], &LHSValNoAssignments[0], NewVNInfo,
2464 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo,
2472 // DepthMBBCompare - Comparison predicate that sort first based on the loop
2473 // depth of the basic block (the unsigned), and then on the MBB number.
2474 struct DepthMBBCompare {
2475 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
2476 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
2477 // Deeper loops first
2478 if (LHS.first != RHS.first)
2479 return LHS.first > RHS.first;
2481 // Prefer blocks that are more connected in the CFG. This takes care of
2482 // the most difficult copies first while intervals are short.
2483 unsigned cl = LHS.second->pred_size() + LHS.second->succ_size();
2484 unsigned cr = RHS.second->pred_size() + RHS.second->succ_size();
2488 // As a last resort, sort by block number.
2489 return LHS.second->getNumber() < RHS.second->getNumber();
2494 void SimpleRegisterCoalescing::CopyCoalesceInMBB(MachineBasicBlock *MBB,
2495 std::vector<CopyRec> &TryAgain) {
2496 DEBUG(dbgs() << MBB->getName() << ":\n");
2498 std::vector<CopyRec> VirtCopies;
2499 std::vector<CopyRec> PhysCopies;
2500 std::vector<CopyRec> ImpDefCopies;
2501 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2503 MachineInstr *Inst = MII++;
2505 // If this isn't a copy nor a extract_subreg, we can't join intervals.
2506 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2507 bool isInsUndef = false;
2508 if (Inst->isExtractSubreg()) {
2509 DstReg = Inst->getOperand(0).getReg();
2510 SrcReg = Inst->getOperand(1).getReg();
2511 } else if (Inst->isInsertSubreg()) {
2512 DstReg = Inst->getOperand(0).getReg();
2513 SrcReg = Inst->getOperand(2).getReg();
2514 if (Inst->getOperand(1).isUndef())
2516 } else if (Inst->isInsertSubreg() || Inst->isSubregToReg()) {
2517 DstReg = Inst->getOperand(0).getReg();
2518 SrcReg = Inst->getOperand(2).getReg();
2519 } else if (!tii_->isMoveInstr(*Inst, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
2522 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
2523 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
2525 (li_->hasInterval(SrcReg) && li_->getInterval(SrcReg).empty()))
2526 ImpDefCopies.push_back(CopyRec(Inst, 0));
2527 else if (SrcIsPhys || DstIsPhys)
2528 PhysCopies.push_back(CopyRec(Inst, 0));
2530 VirtCopies.push_back(CopyRec(Inst, 0));
2533 // Try coalescing implicit copies and insert_subreg <undef> first,
2534 // followed by copies to / from physical registers, then finally copies
2535 // from virtual registers to virtual registers.
2536 for (unsigned i = 0, e = ImpDefCopies.size(); i != e; ++i) {
2537 CopyRec &TheCopy = ImpDefCopies[i];
2539 if (!JoinCopy(TheCopy, Again))
2541 TryAgain.push_back(TheCopy);
2543 for (unsigned i = 0, e = PhysCopies.size(); i != e; ++i) {
2544 CopyRec &TheCopy = PhysCopies[i];
2546 if (!JoinCopy(TheCopy, Again))
2548 TryAgain.push_back(TheCopy);
2550 for (unsigned i = 0, e = VirtCopies.size(); i != e; ++i) {
2551 CopyRec &TheCopy = VirtCopies[i];
2553 if (!JoinCopy(TheCopy, Again))
2555 TryAgain.push_back(TheCopy);
2559 void SimpleRegisterCoalescing::joinIntervals() {
2560 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
2562 std::vector<CopyRec> TryAgainList;
2563 if (loopInfo->empty()) {
2564 // If there are no loops in the function, join intervals in function order.
2565 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();
2567 CopyCoalesceInMBB(I, TryAgainList);
2569 // Otherwise, join intervals in inner loops before other intervals.
2570 // Unfortunately we can't just iterate over loop hierarchy here because
2571 // there may be more MBB's than BB's. Collect MBB's for sorting.
2573 // Join intervals in the function prolog first. We want to join physical
2574 // registers with virtual registers before the intervals got too long.
2575 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
2576 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();I != E;++I){
2577 MachineBasicBlock *MBB = I;
2578 MBBs.push_back(std::make_pair(loopInfo->getLoopDepth(MBB), I));
2581 // Sort by loop depth.
2582 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
2584 // Finally, join intervals in loop nest order.
2585 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
2586 CopyCoalesceInMBB(MBBs[i].second, TryAgainList);
2589 // Joining intervals can allow other intervals to be joined. Iteratively join
2590 // until we make no progress.
2591 bool ProgressMade = true;
2592 while (ProgressMade) {
2593 ProgressMade = false;
2595 for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
2596 CopyRec &TheCopy = TryAgainList[i];
2601 bool Success = JoinCopy(TheCopy, Again);
2602 if (Success || !Again) {
2603 TheCopy.MI = 0; // Mark this one as done.
2604 ProgressMade = true;
2610 /// Return true if the two specified registers belong to different register
2611 /// classes. The registers may be either phys or virt regs.
2613 SimpleRegisterCoalescing::differingRegisterClasses(unsigned RegA,
2614 unsigned RegB) const {
2615 // Get the register classes for the first reg.
2616 if (TargetRegisterInfo::isPhysicalRegister(RegA)) {
2617 assert(TargetRegisterInfo::isVirtualRegister(RegB) &&
2618 "Shouldn't consider two physregs!");
2619 return !mri_->getRegClass(RegB)->contains(RegA);
2622 // Compare against the regclass for the second reg.
2623 const TargetRegisterClass *RegClassA = mri_->getRegClass(RegA);
2624 if (TargetRegisterInfo::isVirtualRegister(RegB)) {
2625 const TargetRegisterClass *RegClassB = mri_->getRegClass(RegB);
2626 return RegClassA != RegClassB;
2628 return !RegClassA->contains(RegB);
2631 /// lastRegisterUse - Returns the last (non-debug) use of the specific register
2632 /// between cycles Start and End or NULL if there are no uses.
2634 SimpleRegisterCoalescing::lastRegisterUse(SlotIndex Start,
2637 SlotIndex &UseIdx) const{
2638 UseIdx = SlotIndex();
2639 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2640 MachineOperand *LastUse = NULL;
2641 for (MachineRegisterInfo::use_nodbg_iterator I = mri_->use_nodbg_begin(Reg),
2642 E = mri_->use_nodbg_end(); I != E; ++I) {
2643 MachineOperand &Use = I.getOperand();
2644 MachineInstr *UseMI = Use.getParent();
2645 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2646 if (tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
2648 // Ignore identity copies.
2650 SlotIndex Idx = li_->getInstructionIndex(UseMI);
2651 // FIXME: Should this be Idx != UseIdx? SlotIndex() will return something
2652 // that compares higher than any other interval.
2653 if (Idx >= Start && Idx < End && Idx >= UseIdx) {
2655 UseIdx = Idx.getUseIndex();
2661 SlotIndex s = Start;
2662 SlotIndex e = End.getPrevSlot().getBaseIndex();
2664 // Skip deleted instructions
2665 MachineInstr *MI = li_->getInstructionFromIndex(e);
2666 while (e != SlotIndex() && e.getPrevIndex() >= s && !MI) {
2667 e = e.getPrevIndex();
2668 MI = li_->getInstructionFromIndex(e);
2670 if (e < s || MI == NULL)
2673 // Ignore identity copies.
2674 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2675 if (!(tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
2677 for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
2678 MachineOperand &Use = MI->getOperand(i);
2679 if (Use.isReg() && Use.isUse() && Use.getReg() &&
2680 tri_->regsOverlap(Use.getReg(), Reg)) {
2681 UseIdx = e.getUseIndex();
2686 e = e.getPrevIndex();
2692 void SimpleRegisterCoalescing::printRegName(unsigned reg) const {
2693 if (TargetRegisterInfo::isPhysicalRegister(reg))
2694 dbgs() << tri_->getName(reg);
2696 dbgs() << "%reg" << reg;
2699 void SimpleRegisterCoalescing::releaseMemory() {
2700 JoinedCopies.clear();
2701 ReMatCopies.clear();
2705 bool SimpleRegisterCoalescing::runOnMachineFunction(MachineFunction &fn) {
2707 mri_ = &fn.getRegInfo();
2708 tm_ = &fn.getTarget();
2709 tri_ = tm_->getRegisterInfo();
2710 tii_ = tm_->getInstrInfo();
2711 li_ = &getAnalysis<LiveIntervals>();
2712 AA = &getAnalysis<AliasAnalysis>();
2713 loopInfo = &getAnalysis<MachineLoopInfo>();
2715 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
2716 << "********** Function: "
2717 << ((Value*)mf_->getFunction())->getName() << '\n');
2719 allocatableRegs_ = tri_->getAllocatableSet(fn);
2720 for (TargetRegisterInfo::regclass_iterator I = tri_->regclass_begin(),
2721 E = tri_->regclass_end(); I != E; ++I)
2722 allocatableRCRegs_.insert(std::make_pair(*I,
2723 tri_->getAllocatableSet(fn, *I)));
2725 // Join (coalesce) intervals if requested.
2726 if (EnableJoining) {
2729 dbgs() << "********** INTERVALS POST JOINING **********\n";
2730 for (LiveIntervals::iterator I = li_->begin(), E = li_->end();
2732 I->second->print(dbgs(), tri_);
2738 // Perform a final pass over the instructions and compute spill weights
2739 // and remove identity moves.
2740 SmallVector<unsigned, 4> DeadDefs;
2741 for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end();
2742 mbbi != mbbe; ++mbbi) {
2743 MachineBasicBlock* mbb = mbbi;
2744 for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
2746 MachineInstr *MI = mii;
2747 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2748 if (JoinedCopies.count(MI)) {
2749 // Delete all coalesced copies.
2750 bool DoDelete = true;
2751 if (!tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
2752 assert((MI->isExtractSubreg() || MI->isInsertSubreg() ||
2753 MI->isSubregToReg()) && "Unrecognized copy instruction");
2754 DstReg = MI->getOperand(0).getReg();
2755 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
2756 // Do not delete extract_subreg, insert_subreg of physical
2757 // registers unless the definition is dead. e.g.
2758 // %DO<def> = INSERT_SUBREG %D0<undef>, %S0<kill>, 1
2759 // or else the scavenger may complain. LowerSubregs will
2760 // delete them later.
2763 if (MI->allDefsAreDead()) {
2764 LiveInterval &li = li_->getInterval(DstReg);
2765 if (!ShortenDeadCopySrcLiveRange(li, MI))
2766 ShortenDeadCopyLiveRange(li, MI);
2770 mii = llvm::next(mii);
2772 li_->RemoveMachineInstrFromMaps(MI);
2773 mii = mbbi->erase(mii);
2779 // Now check if this is a remat'ed def instruction which is now dead.
2780 if (ReMatDefs.count(MI)) {
2782 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2783 const MachineOperand &MO = MI->getOperand(i);
2786 unsigned Reg = MO.getReg();
2789 if (TargetRegisterInfo::isVirtualRegister(Reg))
2790 DeadDefs.push_back(Reg);
2793 if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
2794 !mri_->use_nodbg_empty(Reg)) {
2800 while (!DeadDefs.empty()) {
2801 unsigned DeadDef = DeadDefs.back();
2802 DeadDefs.pop_back();
2803 RemoveDeadDef(li_->getInterval(DeadDef), MI);
2805 li_->RemoveMachineInstrFromMaps(mii);
2806 mii = mbbi->erase(mii);
2812 // If the move will be an identity move delete it
2813 bool isMove= tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx);
2814 if (isMove && SrcReg == DstReg) {
2815 if (li_->hasInterval(SrcReg)) {
2816 LiveInterval &RegInt = li_->getInterval(SrcReg);
2817 // If def of this move instruction is dead, remove its live range
2818 // from the dstination register's live interval.
2819 if (MI->registerDefIsDead(DstReg)) {
2820 if (!ShortenDeadCopySrcLiveRange(RegInt, MI))
2821 ShortenDeadCopyLiveRange(RegInt, MI);
2824 li_->RemoveMachineInstrFromMaps(MI);
2825 mii = mbbi->erase(mii);
2837 /// print - Implement the dump method.
2838 void SimpleRegisterCoalescing::print(raw_ostream &O, const Module* m) const {
2842 RegisterCoalescer* llvm::createSimpleRegisterCoalescer() {
2843 return new SimpleRegisterCoalescing();
2846 // Make sure that anything that uses RegisterCoalescer pulls in this file...
2847 DEFINING_FILE_FOR(SimpleRegisterCoalescing)