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);
63 DisablePhysicalJoin("disable-physical-join",
64 cl::desc("Avoid coalescing physical register copies"),
65 cl::init(false), cl::Hidden);
67 INITIALIZE_AG_PASS_BEGIN(SimpleRegisterCoalescing, RegisterCoalescer,
68 "simple-register-coalescing", "Simple Register Coalescing",
70 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
71 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
72 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
73 INITIALIZE_PASS_DEPENDENCY(StrongPHIElimination)
74 INITIALIZE_PASS_DEPENDENCY(PHIElimination)
75 INITIALIZE_PASS_DEPENDENCY(TwoAddressInstructionPass)
76 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
77 INITIALIZE_AG_PASS_END(SimpleRegisterCoalescing, RegisterCoalescer,
78 "simple-register-coalescing", "Simple Register Coalescing",
81 char &llvm::SimpleRegisterCoalescingID = SimpleRegisterCoalescing::ID;
83 void SimpleRegisterCoalescing::getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequired<AliasAnalysis>();
86 AU.addRequired<LiveIntervals>();
87 AU.addPreserved<LiveIntervals>();
88 AU.addPreserved<SlotIndexes>();
89 AU.addRequired<MachineLoopInfo>();
90 AU.addPreserved<MachineLoopInfo>();
91 AU.addPreservedID(MachineDominatorsID);
93 AU.addPreservedID(StrongPHIEliminationID);
95 AU.addPreservedID(PHIEliminationID);
96 AU.addPreservedID(TwoAddressInstructionPassID);
97 MachineFunctionPass::getAnalysisUsage(AU);
100 /// AdjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA
101 /// being the source and IntB being the dest, thus this defines a value number
102 /// in IntB. If the source value number (in IntA) is defined by a copy from B,
103 /// see if we can merge these two pieces of B into a single value number,
104 /// eliminating a copy. For example:
108 /// B1 = A3 <- this copy
110 /// In this case, B0 can be extended to where the B1 copy lives, allowing the B1
111 /// value number to be replaced with B0 (which simplifies the B liveinterval).
113 /// This returns true if an interval was modified.
115 bool SimpleRegisterCoalescing::AdjustCopiesBackFrom(const CoalescerPair &CP,
116 MachineInstr *CopyMI) {
117 // Bail if there is no dst interval - can happen when merging physical subreg
119 if (!li_->hasInterval(CP.getDstReg()))
123 li_->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
125 li_->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
126 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI).getDefIndex();
128 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
129 // the example above.
130 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
131 if (BLR == IntB.end()) return false;
132 VNInfo *BValNo = BLR->valno;
134 // Get the location that B is defined at. Two options: either this value has
135 // an unknown definition point or it is defined at CopyIdx. If unknown, we
137 if (!BValNo->isDefByCopy()) return false;
138 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
140 // AValNo is the value number in A that defines the copy, A3 in the example.
141 SlotIndex CopyUseIdx = CopyIdx.getUseIndex();
142 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyUseIdx);
143 // The live range might not exist after fun with physreg coalescing.
144 if (ALR == IntA.end()) return false;
145 VNInfo *AValNo = ALR->valno;
146 // If it's re-defined by an early clobber somewhere in the live range, then
147 // it's not safe to eliminate the copy. FIXME: This is a temporary workaround.
149 // 172 %ECX<def> = MOV32rr %reg1039<kill>
150 // 180 INLINEASM <es:subl $5,$1
151 // sbbl $3,$0>, 10, %EAX<def>, 14, %ECX<earlyclobber,def>, 9,
153 // 36, <fi#0>, 1, %reg0, 0, 9, %ECX<kill>, 36, <fi#1>, 1, %reg0, 0
154 // 188 %EAX<def> = MOV32rr %EAX<kill>
155 // 196 %ECX<def> = MOV32rr %ECX<kill>
156 // 204 %ECX<def> = MOV32rr %ECX<kill>
157 // 212 %EAX<def> = MOV32rr %EAX<kill>
158 // 220 %EAX<def> = MOV32rr %EAX
159 // 228 %reg1039<def> = MOV32rr %ECX<kill>
160 // The early clobber operand ties ECX input to the ECX def.
162 // The live interval of ECX is represented as this:
163 // %reg20,inf = [46,47:1)[174,230:0) 0@174-(230) 1@46-(47)
164 // The coalescer has no idea there was a def in the middle of [174,230].
165 if (AValNo->hasRedefByEC())
168 // If AValNo is defined as a copy from IntB, we can potentially process this.
169 // Get the instruction that defines this value number.
170 if (!CP.isCoalescable(AValNo->getCopy()))
173 // Get the LiveRange in IntB that this value number starts with.
174 LiveInterval::iterator ValLR =
175 IntB.FindLiveRangeContaining(AValNo->def.getPrevSlot());
176 if (ValLR == IntB.end())
179 // Make sure that the end of the live range is inside the same block as
181 MachineInstr *ValLREndInst =
182 li_->getInstructionFromIndex(ValLR->end.getPrevSlot());
183 if (!ValLREndInst || ValLREndInst->getParent() != CopyMI->getParent())
186 // Okay, we now know that ValLR ends in the same block that the CopyMI
187 // live-range starts. If there are no intervening live ranges between them in
188 // IntB, we can merge them.
189 if (ValLR+1 != BLR) return false;
191 // If a live interval is a physical register, conservatively check if any
192 // of its sub-registers is overlapping the live interval of the virtual
193 // register. If so, do not coalesce.
194 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg) &&
195 *tri_->getSubRegisters(IntB.reg)) {
196 for (const unsigned* SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR)
197 if (li_->hasInterval(*SR) && IntA.overlaps(li_->getInterval(*SR))) {
199 dbgs() << "\t\tInterfere with sub-register ";
200 li_->getInterval(*SR).print(dbgs(), tri_);
207 dbgs() << "Extending: ";
208 IntB.print(dbgs(), tri_);
211 SlotIndex FillerStart = ValLR->end, FillerEnd = BLR->start;
212 // We are about to delete CopyMI, so need to remove it as the 'instruction
213 // that defines this value #'. Update the valnum with the new defining
215 BValNo->def = FillerStart;
218 // Okay, we can merge them. We need to insert a new liverange:
219 // [ValLR.end, BLR.begin) of either value number, then we merge the
220 // two value numbers.
221 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
223 // If the IntB live range is assigned to a physical register, and if that
224 // physreg has sub-registers, update their live intervals as well.
225 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
226 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
227 if (!li_->hasInterval(*SR))
229 LiveInterval &SRLI = li_->getInterval(*SR);
230 SRLI.addRange(LiveRange(FillerStart, FillerEnd,
231 SRLI.getNextValue(FillerStart, 0,
232 li_->getVNInfoAllocator())));
236 // Okay, merge "B1" into the same value number as "B0".
237 if (BValNo != ValLR->valno) {
238 IntB.MergeValueNumberInto(BValNo, ValLR->valno);
241 dbgs() << " result = ";
242 IntB.print(dbgs(), tri_);
246 // If the source instruction was killing the source register before the
247 // merge, unset the isKill marker given the live range has been extended.
248 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true);
250 ValLREndInst->getOperand(UIdx).setIsKill(false);
253 // If the copy instruction was killing the destination register before the
254 // merge, find the last use and trim the live range. That will also add the
256 if (ALR->end == CopyIdx)
257 TrimLiveIntervalToLastUse(CopyUseIdx, CopyMI->getParent(), IntA, ALR);
263 /// HasOtherReachingDefs - Return true if there are definitions of IntB
264 /// other than BValNo val# that can reach uses of AValno val# of IntA.
265 bool SimpleRegisterCoalescing::HasOtherReachingDefs(LiveInterval &IntA,
269 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
271 if (AI->valno != AValNo) continue;
272 LiveInterval::Ranges::iterator BI =
273 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start);
274 if (BI != IntB.ranges.begin())
276 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) {
277 if (BI->valno == BValNo)
279 if (BI->start <= AI->start && BI->end > AI->start)
281 if (BI->start > AI->start && BI->start < AI->end)
289 TransferImplicitOps(MachineInstr *MI, MachineInstr *NewMI) {
290 for (unsigned i = MI->getDesc().getNumOperands(), e = MI->getNumOperands();
292 MachineOperand &MO = MI->getOperand(i);
293 if (MO.isReg() && MO.isImplicit())
294 NewMI->addOperand(MO);
298 /// RemoveCopyByCommutingDef - We found a non-trivially-coalescable copy with
299 /// IntA being the source and IntB being the dest, thus this defines a value
300 /// number in IntB. If the source value number (in IntA) is defined by a
301 /// commutable instruction and its other operand is coalesced to the copy dest
302 /// register, see if we can transform the copy into a noop by commuting the
303 /// definition. For example,
305 /// A3 = op A2 B0<kill>
307 /// B1 = A3 <- this copy
309 /// = op A3 <- more uses
313 /// B2 = op B0 A2<kill>
315 /// B1 = B2 <- now an identify copy
317 /// = op B2 <- more uses
319 /// This returns true if an interval was modified.
321 bool SimpleRegisterCoalescing::RemoveCopyByCommutingDef(const CoalescerPair &CP,
322 MachineInstr *CopyMI) {
323 // FIXME: For now, only eliminate the copy by commuting its def when the
324 // source register is a virtual register. We want to guard against cases
325 // where the copy is a back edge copy and commuting the def lengthen the
326 // live interval of the source register to the entire loop.
327 if (CP.isPhys() && CP.isFlipped())
330 // Bail if there is no dst interval.
331 if (!li_->hasInterval(CP.getDstReg()))
334 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI).getDefIndex();
337 li_->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
339 li_->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
341 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
342 // the example above.
343 VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx);
344 if (!BValNo || !BValNo->isDefByCopy())
347 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
349 // AValNo is the value number in A that defines the copy, A3 in the example.
350 VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getUseIndex());
351 assert(AValNo && "COPY source not live");
353 // If other defs can reach uses of this def, then it's not safe to perform
355 if (AValNo->isPHIDef() || AValNo->isUnused() || AValNo->hasPHIKill())
357 MachineInstr *DefMI = li_->getInstructionFromIndex(AValNo->def);
360 const TargetInstrDesc &TID = DefMI->getDesc();
361 if (!TID.isCommutable())
363 // If DefMI is a two-address instruction then commuting it will change the
364 // destination register.
365 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
366 assert(DefIdx != -1);
368 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
370 unsigned Op1, Op2, NewDstIdx;
371 if (!tii_->findCommutedOpIndices(DefMI, Op1, Op2))
375 else if (Op2 == UseOpIdx)
380 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
381 unsigned NewReg = NewDstMO.getReg();
382 if (NewReg != IntB.reg || !NewDstMO.isKill())
385 // Make sure there are no other definitions of IntB that would reach the
386 // uses which the new definition can reach.
387 if (HasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
390 // Abort if the aliases of IntB.reg have values that are not simply the
391 // clobbers from the superreg.
392 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg))
393 for (const unsigned *AS = tri_->getAliasSet(IntB.reg); *AS; ++AS)
394 if (li_->hasInterval(*AS) &&
395 HasOtherReachingDefs(IntA, li_->getInterval(*AS), AValNo, 0))
398 // If some of the uses of IntA.reg is already coalesced away, return false.
399 // It's not possible to determine whether it's safe to perform the coalescing.
400 for (MachineRegisterInfo::use_nodbg_iterator UI =
401 mri_->use_nodbg_begin(IntA.reg),
402 UE = mri_->use_nodbg_end(); UI != UE; ++UI) {
403 MachineInstr *UseMI = &*UI;
404 SlotIndex UseIdx = li_->getInstructionIndex(UseMI);
405 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
406 if (ULR == IntA.end())
408 if (ULR->valno == AValNo && JoinedCopies.count(UseMI))
412 DEBUG(dbgs() << "\tRemoveCopyByCommutingDef: " << AValNo->def << '\t'
415 // At this point we have decided that it is legal to do this
416 // transformation. Start by commuting the instruction.
417 MachineBasicBlock *MBB = DefMI->getParent();
418 MachineInstr *NewMI = tii_->commuteInstruction(DefMI);
421 if (NewMI != DefMI) {
422 li_->ReplaceMachineInstrInMaps(DefMI, NewMI);
423 MBB->insert(DefMI, NewMI);
426 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false);
427 NewMI->getOperand(OpIdx).setIsKill();
429 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
438 // Update uses of IntA of the specific Val# with IntB.
439 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(IntA.reg),
440 UE = mri_->use_end(); UI != UE;) {
441 MachineOperand &UseMO = UI.getOperand();
442 MachineInstr *UseMI = &*UI;
444 if (JoinedCopies.count(UseMI))
446 if (UseMI->isDebugValue()) {
447 // FIXME These don't have an instruction index. Not clear we have enough
448 // info to decide whether to do this replacement or not. For now do it.
449 UseMO.setReg(NewReg);
452 SlotIndex UseIdx = li_->getInstructionIndex(UseMI).getUseIndex();
453 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
454 if (ULR == IntA.end() || ULR->valno != AValNo)
456 if (TargetRegisterInfo::isPhysicalRegister(NewReg))
457 UseMO.substPhysReg(NewReg, *tri_);
459 UseMO.setReg(NewReg);
462 if (!UseMI->isCopy())
464 if (UseMI->getOperand(0).getReg() != IntB.reg ||
465 UseMI->getOperand(0).getSubReg())
468 // This copy will become a noop. If it's defining a new val#, merge it into
470 SlotIndex DefIdx = UseIdx.getDefIndex();
471 VNInfo *DVNI = IntB.getVNInfoAt(DefIdx);
474 DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI);
475 assert(DVNI->def == DefIdx);
476 BValNo = IntB.MergeValueNumberInto(BValNo, DVNI);
477 JoinedCopies.insert(UseMI);
480 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition
482 VNInfo *ValNo = BValNo;
483 ValNo->def = AValNo->def;
485 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
487 if (AI->valno != AValNo) continue;
488 IntB.addRange(LiveRange(AI->start, AI->end, ValNo));
490 DEBUG(dbgs() << "\t\textended: " << IntB << '\n');
492 IntA.removeValNo(AValNo);
493 DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n');
498 /// isSameOrFallThroughBB - Return true if MBB == SuccMBB or MBB simply
499 /// fallthoughs to SuccMBB.
500 static bool isSameOrFallThroughBB(MachineBasicBlock *MBB,
501 MachineBasicBlock *SuccMBB,
502 const TargetInstrInfo *tii_) {
505 MachineBasicBlock *TBB = 0, *FBB = 0;
506 SmallVector<MachineOperand, 4> Cond;
507 return !tii_->AnalyzeBranch(*MBB, TBB, FBB, Cond) && !TBB && !FBB &&
508 MBB->isSuccessor(SuccMBB);
511 /// removeRange - Wrapper for LiveInterval::removeRange. This removes a range
512 /// from a physical register live interval as well as from the live intervals
513 /// of its sub-registers.
514 static void removeRange(LiveInterval &li,
515 SlotIndex Start, SlotIndex End,
516 LiveIntervals *li_, const TargetRegisterInfo *tri_) {
517 li.removeRange(Start, End, true);
518 if (TargetRegisterInfo::isPhysicalRegister(li.reg)) {
519 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
520 if (!li_->hasInterval(*SR))
522 LiveInterval &sli = li_->getInterval(*SR);
523 SlotIndex RemoveStart = Start;
524 SlotIndex RemoveEnd = Start;
526 while (RemoveEnd != End) {
527 LiveInterval::iterator LR = sli.FindLiveRangeContaining(RemoveStart);
530 RemoveEnd = (LR->end < End) ? LR->end : End;
531 sli.removeRange(RemoveStart, RemoveEnd, true);
532 RemoveStart = RemoveEnd;
538 /// TrimLiveIntervalToLastUse - If there is a last use in the same basic block
539 /// as the copy instruction, trim the live interval to the last use and return
542 SimpleRegisterCoalescing::TrimLiveIntervalToLastUse(SlotIndex CopyIdx,
543 MachineBasicBlock *CopyMBB,
545 const LiveRange *LR) {
546 SlotIndex MBBStart = li_->getMBBStartIdx(CopyMBB);
547 SlotIndex LastUseIdx;
548 MachineOperand *LastUse =
549 lastRegisterUse(LR->start, CopyIdx.getPrevSlot(), li.reg, LastUseIdx);
551 MachineInstr *LastUseMI = LastUse->getParent();
552 if (!isSameOrFallThroughBB(LastUseMI->getParent(), CopyMBB, tii_)) {
559 // r1025<dead> = r1024<kill>
560 if (MBBStart < LR->end)
561 removeRange(li, MBBStart, LR->end, li_, tri_);
565 // There are uses before the copy, just shorten the live range to the end
567 LastUse->setIsKill();
568 removeRange(li, LastUseIdx.getDefIndex(), LR->end, li_, tri_);
569 if (LastUseMI->isCopy()) {
570 MachineOperand &DefMO = LastUseMI->getOperand(0);
571 if (DefMO.getReg() == li.reg && !DefMO.getSubReg())
578 if (LR->start <= MBBStart && LR->end > MBBStart) {
579 if (LR->start == li_->getZeroIndex()) {
580 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
581 // Live-in to the function but dead. Remove it from entry live-in set.
582 mf_->begin()->removeLiveIn(li.reg);
584 // FIXME: Shorten intervals in BBs that reaches this BB.
590 /// ReMaterializeTrivialDef - If the source of a copy is defined by a trivial
591 /// computation, replace the copy by rematerialize the definition.
592 bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
595 MachineInstr *CopyMI) {
596 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI).getUseIndex();
597 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx);
598 assert(SrcLR != SrcInt.end() && "Live range not found!");
599 VNInfo *ValNo = SrcLR->valno;
600 // If other defs can reach uses of this def, then it's not safe to perform
602 if (ValNo->isPHIDef() || ValNo->isUnused() || ValNo->hasPHIKill())
604 MachineInstr *DefMI = li_->getInstructionFromIndex(ValNo->def);
607 assert(DefMI && "Defining instruction disappeared");
608 const TargetInstrDesc &TID = DefMI->getDesc();
609 if (!TID.isAsCheapAsAMove())
611 if (!tii_->isTriviallyReMaterializable(DefMI, AA))
613 bool SawStore = false;
614 if (!DefMI->isSafeToMove(tii_, AA, SawStore))
616 if (TID.getNumDefs() != 1)
618 if (!DefMI->isImplicitDef()) {
619 // Make sure the copy destination register class fits the instruction
620 // definition register class. The mismatch can happen as a result of earlier
621 // extract_subreg, insert_subreg, subreg_to_reg coalescing.
622 const TargetRegisterClass *RC = TID.OpInfo[0].getRegClass(tri_);
623 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
624 if (mri_->getRegClass(DstReg) != RC)
626 } else if (!RC->contains(DstReg))
630 // If destination register has a sub-register index on it, make sure it mtches
631 // the instruction register class.
633 const TargetInstrDesc &TID = DefMI->getDesc();
634 if (TID.getNumDefs() != 1)
636 const TargetRegisterClass *DstRC = mri_->getRegClass(DstReg);
637 const TargetRegisterClass *DstSubRC =
638 DstRC->getSubRegisterRegClass(DstSubIdx);
639 const TargetRegisterClass *DefRC = TID.OpInfo[0].getRegClass(tri_);
642 else if (DefRC != DstSubRC)
646 RemoveCopyFlag(DstReg, CopyMI);
648 // If copy kills the source register, find the last use and propagate
650 bool checkForDeadDef = false;
651 MachineBasicBlock *MBB = CopyMI->getParent();
652 if (SrcLR->end == CopyIdx.getDefIndex())
653 if (!TrimLiveIntervalToLastUse(CopyIdx, MBB, SrcInt, SrcLR)) {
654 checkForDeadDef = true;
657 MachineBasicBlock::iterator MII =
658 llvm::next(MachineBasicBlock::iterator(CopyMI));
659 tii_->reMaterialize(*MBB, MII, DstReg, DstSubIdx, DefMI, *tri_);
660 MachineInstr *NewMI = prior(MII);
662 if (checkForDeadDef) {
663 // PR4090 fix: Trim interval failed because there was no use of the
664 // source interval in this MBB. If the def is in this MBB too then we
665 // should mark it dead:
666 if (DefMI->getParent() == MBB) {
667 DefMI->addRegisterDead(SrcInt.reg, tri_);
668 SrcLR->end = SrcLR->start.getNextSlot();
672 // CopyMI may have implicit operands, transfer them over to the newly
673 // rematerialized instruction. And update implicit def interval valnos.
674 for (unsigned i = CopyMI->getDesc().getNumOperands(),
675 e = CopyMI->getNumOperands(); i != e; ++i) {
676 MachineOperand &MO = CopyMI->getOperand(i);
677 if (MO.isReg() && MO.isImplicit())
678 NewMI->addOperand(MO);
680 RemoveCopyFlag(MO.getReg(), CopyMI);
683 TransferImplicitOps(CopyMI, NewMI);
684 li_->ReplaceMachineInstrInMaps(CopyMI, NewMI);
685 CopyMI->eraseFromParent();
686 ReMatCopies.insert(CopyMI);
687 ReMatDefs.insert(DefMI);
688 DEBUG(dbgs() << "Remat: " << *NewMI);
693 /// UpdateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
694 /// update the subregister number if it is not zero. If DstReg is a
695 /// physical register and the existing subregister number of the def / use
696 /// being updated is not zero, make sure to set it to the correct physical
699 SimpleRegisterCoalescing::UpdateRegDefsUses(const CoalescerPair &CP) {
700 bool DstIsPhys = CP.isPhys();
701 unsigned SrcReg = CP.getSrcReg();
702 unsigned DstReg = CP.getDstReg();
703 unsigned SubIdx = CP.getSubIdx();
705 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(SrcReg);
706 MachineInstr *UseMI = I.skipInstruction();) {
707 // A PhysReg copy that won't be coalesced can perhaps be rematerialized
710 if (UseMI->isCopy() &&
711 !UseMI->getOperand(1).getSubReg() &&
712 !UseMI->getOperand(0).getSubReg() &&
713 UseMI->getOperand(1).getReg() == SrcReg &&
714 UseMI->getOperand(0).getReg() != SrcReg &&
715 UseMI->getOperand(0).getReg() != DstReg &&
716 !JoinedCopies.count(UseMI) &&
717 ReMaterializeTrivialDef(li_->getInterval(SrcReg),
718 UseMI->getOperand(0).getReg(), 0, UseMI))
722 SmallVector<unsigned,8> Ops;
724 tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
725 bool Kills = false, Deads = false;
727 // Replace SrcReg with DstReg in all UseMI operands.
728 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
729 MachineOperand &MO = UseMI->getOperand(Ops[i]);
730 Kills |= MO.isKill();
731 Deads |= MO.isDead();
734 MO.substPhysReg(DstReg, *tri_);
736 MO.substVirtReg(DstReg, SubIdx, *tri_);
739 // This instruction is a copy that will be removed.
740 if (JoinedCopies.count(UseMI))
744 // If UseMI was a simple SrcReg def, make sure we didn't turn it into a
745 // read-modify-write of DstReg.
747 UseMI->addRegisterDead(DstReg, tri_);
748 else if (!Reads && Writes)
749 UseMI->addRegisterDefined(DstReg, tri_);
751 // Kill flags apply to the whole physical register.
752 if (DstIsPhys && Kills)
753 UseMI->addRegisterKilled(DstReg, tri_);
757 dbgs() << "\t\tupdated: ";
758 if (!UseMI->isDebugValue())
759 dbgs() << li_->getInstructionIndex(UseMI) << "\t";
765 /// removeIntervalIfEmpty - Check if the live interval of a physical register
766 /// is empty, if so remove it and also remove the empty intervals of its
767 /// sub-registers. Return true if live interval is removed.
768 static bool removeIntervalIfEmpty(LiveInterval &li, LiveIntervals *li_,
769 const TargetRegisterInfo *tri_) {
771 if (TargetRegisterInfo::isPhysicalRegister(li.reg))
772 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
773 if (!li_->hasInterval(*SR))
775 LiveInterval &sli = li_->getInterval(*SR);
777 li_->removeInterval(*SR);
779 li_->removeInterval(li.reg);
785 /// ShortenDeadCopyLiveRange - Shorten a live range defined by a dead copy.
786 /// Return true if live interval is removed.
787 bool SimpleRegisterCoalescing::ShortenDeadCopyLiveRange(LiveInterval &li,
788 MachineInstr *CopyMI) {
789 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI);
790 LiveInterval::iterator MLR =
791 li.FindLiveRangeContaining(CopyIdx.getDefIndex());
793 return false; // Already removed by ShortenDeadCopySrcLiveRange.
794 SlotIndex RemoveStart = MLR->start;
795 SlotIndex RemoveEnd = MLR->end;
796 SlotIndex DefIdx = CopyIdx.getDefIndex();
797 // Remove the liverange that's defined by this.
798 if (RemoveStart == DefIdx && RemoveEnd == DefIdx.getStoreIndex()) {
799 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
800 return removeIntervalIfEmpty(li, li_, tri_);
805 /// RemoveDeadDef - If a def of a live interval is now determined dead, remove
806 /// the val# it defines. If the live interval becomes empty, remove it as well.
807 bool SimpleRegisterCoalescing::RemoveDeadDef(LiveInterval &li,
808 MachineInstr *DefMI) {
809 SlotIndex DefIdx = li_->getInstructionIndex(DefMI).getDefIndex();
810 LiveInterval::iterator MLR = li.FindLiveRangeContaining(DefIdx);
811 if (DefIdx != MLR->valno->def)
813 li.removeValNo(MLR->valno);
814 return removeIntervalIfEmpty(li, li_, tri_);
817 void SimpleRegisterCoalescing::RemoveCopyFlag(unsigned DstReg,
818 const MachineInstr *CopyMI) {
819 SlotIndex DefIdx = li_->getInstructionIndex(CopyMI).getDefIndex();
820 if (li_->hasInterval(DstReg)) {
821 LiveInterval &LI = li_->getInterval(DstReg);
822 if (const LiveRange *LR = LI.getLiveRangeContaining(DefIdx))
823 if (LR->valno->def == DefIdx)
824 LR->valno->setCopy(0);
826 if (!TargetRegisterInfo::isPhysicalRegister(DstReg))
828 for (const unsigned* AS = tri_->getAliasSet(DstReg); *AS; ++AS) {
829 if (!li_->hasInterval(*AS))
831 LiveInterval &LI = li_->getInterval(*AS);
832 if (const LiveRange *LR = LI.getLiveRangeContaining(DefIdx))
833 if (LR->valno->def == DefIdx)
834 LR->valno->setCopy(0);
838 /// PropagateDeadness - Propagate the dead marker to the instruction which
839 /// defines the val#.
840 static void PropagateDeadness(LiveInterval &li, MachineInstr *CopyMI,
841 SlotIndex &LRStart, LiveIntervals *li_,
842 const TargetRegisterInfo* tri_) {
843 MachineInstr *DefMI =
844 li_->getInstructionFromIndex(LRStart.getDefIndex());
845 if (DefMI && DefMI != CopyMI) {
846 int DeadIdx = DefMI->findRegisterDefOperandIdx(li.reg);
848 DefMI->getOperand(DeadIdx).setIsDead();
850 DefMI->addOperand(MachineOperand::CreateReg(li.reg,
851 /*def*/true, /*implicit*/true, /*kill*/false, /*dead*/true));
852 LRStart = LRStart.getNextSlot();
856 /// ShortenDeadCopySrcLiveRange - Shorten a live range as it's artificially
857 /// extended by a dead copy. Mark the last use (if any) of the val# as kill as
858 /// ends the live range there. If there isn't another use, then this live range
859 /// is dead. Return true if live interval is removed.
861 SimpleRegisterCoalescing::ShortenDeadCopySrcLiveRange(LiveInterval &li,
862 MachineInstr *CopyMI) {
863 SlotIndex CopyIdx = li_->getInstructionIndex(CopyMI);
864 if (CopyIdx == SlotIndex()) {
865 // FIXME: special case: function live in. It can be a general case if the
866 // first instruction index starts at > 0 value.
867 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
868 // Live-in to the function but dead. Remove it from entry live-in set.
869 if (mf_->begin()->isLiveIn(li.reg))
870 mf_->begin()->removeLiveIn(li.reg);
871 if (const LiveRange *LR = li.getLiveRangeContaining(CopyIdx))
872 removeRange(li, LR->start, LR->end, li_, tri_);
873 return removeIntervalIfEmpty(li, li_, tri_);
876 LiveInterval::iterator LR =
877 li.FindLiveRangeContaining(CopyIdx.getPrevIndex().getStoreIndex());
879 // Livein but defined by a phi.
882 SlotIndex RemoveStart = LR->start;
883 SlotIndex RemoveEnd = CopyIdx.getStoreIndex();
884 if (LR->end > RemoveEnd)
885 // More uses past this copy? Nothing to do.
888 // If there is a last use in the same bb, we can't remove the live range.
889 // Shorten the live interval and return.
890 MachineBasicBlock *CopyMBB = CopyMI->getParent();
891 if (TrimLiveIntervalToLastUse(CopyIdx, CopyMBB, li, LR))
894 // There are other kills of the val#. Nothing to do.
895 if (!li.isOnlyLROfValNo(LR))
898 MachineBasicBlock *StartMBB = li_->getMBBFromIndex(RemoveStart);
899 if (!isSameOrFallThroughBB(StartMBB, CopyMBB, tii_))
900 // If the live range starts in another mbb and the copy mbb is not a fall
901 // through mbb, then we can only cut the range from the beginning of the
903 RemoveStart = li_->getMBBStartIdx(CopyMBB).getNextIndex().getBaseIndex();
905 if (LR->valno->def == RemoveStart) {
906 // If the def MI defines the val# and this copy is the only kill of the
907 // val#, then propagate the dead marker.
908 PropagateDeadness(li, CopyMI, RemoveStart, li_, tri_);
912 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
913 return removeIntervalIfEmpty(li, li_, tri_);
917 /// isWinToJoinCrossClass - Return true if it's profitable to coalesce
918 /// two virtual registers from different register classes.
920 SimpleRegisterCoalescing::isWinToJoinCrossClass(unsigned SrcReg,
922 const TargetRegisterClass *SrcRC,
923 const TargetRegisterClass *DstRC,
924 const TargetRegisterClass *NewRC) {
925 unsigned NewRCCount = allocatableRCRegs_[NewRC].count();
926 // This heuristics is good enough in practice, but it's obviously not *right*.
927 // 4 is a magic number that works well enough for x86, ARM, etc. It filter
928 // out all but the most restrictive register classes.
929 if (NewRCCount > 4 ||
930 // Early exit if the function is fairly small, coalesce aggressively if
931 // that's the case. For really special register classes with 3 or
932 // fewer registers, be a bit more careful.
933 (li_->getFuncInstructionCount() / NewRCCount) < 8)
935 LiveInterval &SrcInt = li_->getInterval(SrcReg);
936 LiveInterval &DstInt = li_->getInterval(DstReg);
937 unsigned SrcSize = li_->getApproximateInstructionCount(SrcInt);
938 unsigned DstSize = li_->getApproximateInstructionCount(DstInt);
939 if (SrcSize <= NewRCCount && DstSize <= NewRCCount)
941 // Estimate *register use density*. If it doubles or more, abort.
942 unsigned SrcUses = std::distance(mri_->use_nodbg_begin(SrcReg),
943 mri_->use_nodbg_end());
944 unsigned DstUses = std::distance(mri_->use_nodbg_begin(DstReg),
945 mri_->use_nodbg_end());
946 unsigned NewUses = SrcUses + DstUses;
947 unsigned NewSize = SrcSize + DstSize;
948 if (SrcRC != NewRC && SrcSize > NewRCCount) {
949 unsigned SrcRCCount = allocatableRCRegs_[SrcRC].count();
950 if (NewUses*SrcSize*SrcRCCount > 2*SrcUses*NewSize*NewRCCount)
953 if (DstRC != NewRC && DstSize > NewRCCount) {
954 unsigned DstRCCount = allocatableRCRegs_[DstRC].count();
955 if (NewUses*DstSize*DstRCCount > 2*DstUses*NewSize*NewRCCount)
962 /// JoinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
963 /// which are the src/dst of the copy instruction CopyMI. This returns true
964 /// if the copy was successfully coalesced away. If it is not currently
965 /// possible to coalesce this interval, but it may be possible if other
966 /// things get coalesced, then it returns true by reference in 'Again'.
967 bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) {
968 MachineInstr *CopyMI = TheCopy.MI;
971 if (JoinedCopies.count(CopyMI) || ReMatCopies.count(CopyMI))
972 return false; // Already done.
974 DEBUG(dbgs() << li_->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
976 CoalescerPair CP(*tii_, *tri_);
977 if (!CP.setRegisters(CopyMI)) {
978 DEBUG(dbgs() << "\tNot coalescable.\n");
982 // If they are already joined we continue.
983 if (CP.getSrcReg() == CP.getDstReg()) {
984 DEBUG(dbgs() << "\tCopy already coalesced.\n");
985 return false; // Not coalescable.
988 if (DisablePhysicalJoin && CP.isPhys()) {
989 DEBUG(dbgs() << "\tPhysical joins disabled.\n");
993 DEBUG(dbgs() << "\tConsidering merging %reg" << CP.getSrcReg());
997 DEBUG(dbgs() <<" with physreg %" << tri_->getName(CP.getDstReg()) << "\n");
998 // Only coalesce to allocatable physreg.
999 if (!li_->isAllocatable(CP.getDstReg())) {
1000 DEBUG(dbgs() << "\tRegister is an unallocatable physreg.\n");
1001 return false; // Not coalescable.
1005 dbgs() << " with reg%" << CP.getDstReg();
1007 dbgs() << ":" << tri_->getSubRegIndexName(CP.getSubIdx());
1008 dbgs() << " to " << CP.getNewRC()->getName() << "\n";
1011 // Avoid constraining virtual register regclass too much.
1012 if (CP.isCrossClass()) {
1013 if (DisableCrossClassJoin) {
1014 DEBUG(dbgs() << "\tCross-class joins disabled.\n");
1017 if (!isWinToJoinCrossClass(CP.getSrcReg(), CP.getDstReg(),
1018 mri_->getRegClass(CP.getSrcReg()),
1019 mri_->getRegClass(CP.getDstReg()),
1021 DEBUG(dbgs() << "\tAvoid coalescing to constrained register class: "
1022 << CP.getNewRC()->getName() << ".\n");
1023 Again = true; // May be possible to coalesce later.
1028 // When possible, let DstReg be the larger interval.
1029 if (!CP.getSubIdx() && li_->getInterval(CP.getSrcReg()).ranges.size() >
1030 li_->getInterval(CP.getDstReg()).ranges.size())
1034 // We need to be careful about coalescing a source physical register with a
1035 // virtual register. Once the coalescing is done, it cannot be broken and
1036 // these are not spillable! If the destination interval uses are far away,
1037 // think twice about coalescing them!
1038 // FIXME: Why are we skipping this test for partial copies?
1039 // CodeGen/X86/phys_subreg_coalesce-3.ll needs it.
1040 if (!CP.isPartial() && CP.isPhys()) {
1041 LiveInterval &JoinVInt = li_->getInterval(CP.getSrcReg());
1043 // Don't join with physregs that have a ridiculous number of live
1044 // ranges. The data structure performance is really bad when that
1046 if (li_->hasInterval(CP.getDstReg()) &&
1047 li_->getInterval(CP.getDstReg()).ranges.size() > 1000) {
1050 << "\tPhysical register live interval too complicated, abort!\n");
1054 const TargetRegisterClass *RC = mri_->getRegClass(CP.getSrcReg());
1055 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1056 unsigned Length = li_->getApproximateInstructionCount(JoinVInt);
1057 if (Length > Threshold &&
1058 std::distance(mri_->use_nodbg_begin(CP.getSrcReg()),
1059 mri_->use_nodbg_end()) * Threshold < Length) {
1060 // Before giving up coalescing, if definition of source is defined by
1061 // trivial computation, try rematerializing it.
1062 if (!CP.isFlipped() &&
1063 ReMaterializeTrivialDef(JoinVInt, CP.getDstReg(), 0, CopyMI))
1067 DEBUG(dbgs() << "\tMay tie down a physical register, abort!\n");
1068 Again = true; // May be possible to coalesce later.
1073 // Okay, attempt to join these two intervals. On failure, this returns false.
1074 // Otherwise, if one of the intervals being joined is a physreg, this method
1075 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1076 // been modified, so we can use this information below to update aliases.
1077 if (!JoinIntervals(CP)) {
1078 // Coalescing failed.
1080 // If definition of source is defined by trivial computation, try
1081 // rematerializing it.
1082 if (!CP.isFlipped() &&
1083 ReMaterializeTrivialDef(li_->getInterval(CP.getSrcReg()),
1084 CP.getDstReg(), 0, CopyMI))
1087 // If we can eliminate the copy without merging the live ranges, do so now.
1088 if (!CP.isPartial()) {
1089 if (AdjustCopiesBackFrom(CP, CopyMI) ||
1090 RemoveCopyByCommutingDef(CP, CopyMI)) {
1091 JoinedCopies.insert(CopyMI);
1092 DEBUG(dbgs() << "\tTrivial!\n");
1097 // Otherwise, we are unable to join the intervals.
1098 DEBUG(dbgs() << "\tInterference!\n");
1099 Again = true; // May be possible to coalesce later.
1103 // Coalescing to a virtual register that is of a sub-register class of the
1104 // other. Make sure the resulting register is set to the right register class.
1105 if (CP.isCrossClass()) {
1107 mri_->setRegClass(CP.getDstReg(), CP.getNewRC());
1110 // Remember to delete the copy instruction.
1111 JoinedCopies.insert(CopyMI);
1113 UpdateRegDefsUses(CP);
1115 // If we have extended the live range of a physical register, make sure we
1116 // update live-in lists as well.
1118 SmallVector<MachineBasicBlock*, 16> BlockSeq;
1119 // JoinIntervals invalidates the VNInfos in SrcInt, but we only need the
1120 // ranges for this, and they are preserved.
1121 LiveInterval &SrcInt = li_->getInterval(CP.getSrcReg());
1122 for (LiveInterval::const_iterator I = SrcInt.begin(), E = SrcInt.end();
1124 li_->findLiveInMBBs(I->start, I->end, BlockSeq);
1125 for (unsigned idx = 0, size = BlockSeq.size(); idx != size; ++idx) {
1126 MachineBasicBlock &block = *BlockSeq[idx];
1127 if (!block.isLiveIn(CP.getDstReg()))
1128 block.addLiveIn(CP.getDstReg());
1134 // SrcReg is guarateed to be the register whose live interval that is
1136 li_->removeInterval(CP.getSrcReg());
1138 // Update regalloc hint.
1139 tri_->UpdateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *mf_);
1142 LiveInterval &DstInt = li_->getInterval(CP.getDstReg());
1143 dbgs() << "\tJoined. Result = ";
1144 DstInt.print(dbgs(), tri_);
1152 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
1153 /// compute what the resultant value numbers for each value in the input two
1154 /// ranges will be. This is complicated by copies between the two which can
1155 /// and will commonly cause multiple value numbers to be merged into one.
1157 /// VN is the value number that we're trying to resolve. InstDefiningValue
1158 /// keeps track of the new InstDefiningValue assignment for the result
1159 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
1160 /// whether a value in this or other is a copy from the opposite set.
1161 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
1162 /// already been assigned.
1164 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
1165 /// contains the value number the copy is from.
1167 static unsigned ComputeUltimateVN(VNInfo *VNI,
1168 SmallVector<VNInfo*, 16> &NewVNInfo,
1169 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
1170 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
1171 SmallVector<int, 16> &ThisValNoAssignments,
1172 SmallVector<int, 16> &OtherValNoAssignments) {
1173 unsigned VN = VNI->id;
1175 // If the VN has already been computed, just return it.
1176 if (ThisValNoAssignments[VN] >= 0)
1177 return ThisValNoAssignments[VN];
1178 assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers");
1180 // If this val is not a copy from the other val, then it must be a new value
1181 // number in the destination.
1182 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
1183 if (I == ThisFromOther.end()) {
1184 NewVNInfo.push_back(VNI);
1185 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
1187 VNInfo *OtherValNo = I->second;
1189 // Otherwise, this *is* a copy from the RHS. If the other side has already
1190 // been computed, return it.
1191 if (OtherValNoAssignments[OtherValNo->id] >= 0)
1192 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
1194 // Mark this value number as currently being computed, then ask what the
1195 // ultimate value # of the other value is.
1196 ThisValNoAssignments[VN] = -2;
1197 unsigned UltimateVN =
1198 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
1199 OtherValNoAssignments, ThisValNoAssignments);
1200 return ThisValNoAssignments[VN] = UltimateVN;
1203 /// JoinIntervals - Attempt to join these two intervals. On failure, this
1205 bool SimpleRegisterCoalescing::JoinIntervals(CoalescerPair &CP) {
1206 LiveInterval &RHS = li_->getInterval(CP.getSrcReg());
1207 DEBUG({ dbgs() << "\t\tRHS = "; RHS.print(dbgs(), tri_); dbgs() << "\n"; });
1209 // If a live interval is a physical register, check for interference with any
1210 // aliases. The interference check implemented here is a bit more conservative
1211 // than the full interfeence check below. We allow overlapping live ranges
1212 // only when one is a copy of the other.
1214 for (const unsigned *AS = tri_->getAliasSet(CP.getDstReg()); *AS; ++AS){
1215 if (!li_->hasInterval(*AS))
1217 const LiveInterval &LHS = li_->getInterval(*AS);
1218 LiveInterval::const_iterator LI = LHS.begin();
1219 for (LiveInterval::const_iterator RI = RHS.begin(), RE = RHS.end();
1221 LI = std::lower_bound(LI, LHS.end(), RI->start);
1222 // Does LHS have an overlapping live range starting before RI?
1223 if ((LI != LHS.begin() && LI[-1].end > RI->start) &&
1224 (RI->start != RI->valno->def ||
1225 !CP.isCoalescable(li_->getInstructionFromIndex(RI->start)))) {
1227 dbgs() << "\t\tInterference from alias: ";
1228 LHS.print(dbgs(), tri_);
1229 dbgs() << "\n\t\tOverlap at " << RI->start << " and no copy.\n";
1234 // Check that LHS ranges beginning in this range are copies.
1235 for (; LI != LHS.end() && LI->start < RI->end; ++LI) {
1236 if (LI->start != LI->valno->def ||
1237 !CP.isCoalescable(li_->getInstructionFromIndex(LI->start))) {
1239 dbgs() << "\t\tInterference from alias: ";
1240 LHS.print(dbgs(), tri_);
1241 dbgs() << "\n\t\tDef at " << LI->start << " is not a copy.\n";
1250 // Compute the final value assignment, assuming that the live ranges can be
1252 SmallVector<int, 16> LHSValNoAssignments;
1253 SmallVector<int, 16> RHSValNoAssignments;
1254 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
1255 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
1256 SmallVector<VNInfo*, 16> NewVNInfo;
1258 LiveInterval &LHS = li_->getOrCreateInterval(CP.getDstReg());
1259 DEBUG({ dbgs() << "\t\tLHS = "; LHS.print(dbgs(), tri_); dbgs() << "\n"; });
1261 // Loop over the value numbers of the LHS, seeing if any are defined from
1263 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1266 if (VNI->isUnused() || !VNI->isDefByCopy()) // Src not defined by a copy?
1269 // Never join with a register that has EarlyClobber redefs.
1270 if (VNI->hasRedefByEC())
1273 // DstReg is known to be a register in the LHS interval. If the src is
1274 // from the RHS interval, we can use its value #.
1275 if (!CP.isCoalescable(VNI->getCopy()))
1278 // Figure out the value # from the RHS.
1279 LiveRange *lr = RHS.getLiveRangeContaining(VNI->def.getPrevSlot());
1280 // The copy could be to an aliased physreg.
1282 LHSValsDefinedFromRHS[VNI] = lr->valno;
1285 // Loop over the value numbers of the RHS, seeing if any are defined from
1287 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1290 if (VNI->isUnused() || !VNI->isDefByCopy()) // Src not defined by a copy?
1293 // Never join with a register that has EarlyClobber redefs.
1294 if (VNI->hasRedefByEC())
1297 // DstReg is known to be a register in the RHS interval. If the src is
1298 // from the LHS interval, we can use its value #.
1299 if (!CP.isCoalescable(VNI->getCopy()))
1302 // Figure out the value # from the LHS.
1303 LiveRange *lr = LHS.getLiveRangeContaining(VNI->def.getPrevSlot());
1304 // The copy could be to an aliased physreg.
1306 RHSValsDefinedFromLHS[VNI] = lr->valno;
1309 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
1310 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
1311 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
1313 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1316 unsigned VN = VNI->id;
1317 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused())
1319 ComputeUltimateVN(VNI, NewVNInfo,
1320 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
1321 LHSValNoAssignments, RHSValNoAssignments);
1323 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1326 unsigned VN = VNI->id;
1327 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused())
1329 // If this value number isn't a copy from the LHS, it's a new number.
1330 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
1331 NewVNInfo.push_back(VNI);
1332 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
1336 ComputeUltimateVN(VNI, NewVNInfo,
1337 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
1338 RHSValNoAssignments, LHSValNoAssignments);
1341 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
1342 // interval lists to see if these intervals are coalescable.
1343 LiveInterval::const_iterator I = LHS.begin();
1344 LiveInterval::const_iterator IE = LHS.end();
1345 LiveInterval::const_iterator J = RHS.begin();
1346 LiveInterval::const_iterator JE = RHS.end();
1348 // Skip ahead until the first place of potential sharing.
1349 if (I != IE && J != JE) {
1350 if (I->start < J->start) {
1351 I = std::upper_bound(I, IE, J->start);
1352 if (I != LHS.begin()) --I;
1353 } else if (J->start < I->start) {
1354 J = std::upper_bound(J, JE, I->start);
1355 if (J != RHS.begin()) --J;
1359 while (I != IE && J != JE) {
1360 // Determine if these two live ranges overlap.
1362 if (I->start < J->start) {
1363 Overlaps = I->end > J->start;
1365 Overlaps = J->end > I->start;
1368 // If so, check value # info to determine if they are really different.
1370 // If the live range overlap will map to the same value number in the
1371 // result liverange, we can still coalesce them. If not, we can't.
1372 if (LHSValNoAssignments[I->valno->id] !=
1373 RHSValNoAssignments[J->valno->id])
1375 // If it's re-defined by an early clobber somewhere in the live range,
1376 // then conservatively abort coalescing.
1377 if (NewVNInfo[LHSValNoAssignments[I->valno->id]]->hasRedefByEC())
1381 if (I->end < J->end)
1387 // Update kill info. Some live ranges are extended due to copy coalescing.
1388 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
1389 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
1390 VNInfo *VNI = I->first;
1391 unsigned LHSValID = LHSValNoAssignments[VNI->id];
1392 if (VNI->hasPHIKill())
1393 NewVNInfo[LHSValID]->setHasPHIKill(true);
1396 // Update kill info. Some live ranges are extended due to copy coalescing.
1397 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
1398 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
1399 VNInfo *VNI = I->first;
1400 unsigned RHSValID = RHSValNoAssignments[VNI->id];
1401 if (VNI->hasPHIKill())
1402 NewVNInfo[RHSValID]->setHasPHIKill(true);
1405 if (LHSValNoAssignments.empty())
1406 LHSValNoAssignments.push_back(-1);
1407 if (RHSValNoAssignments.empty())
1408 RHSValNoAssignments.push_back(-1);
1410 // If we get here, we know that we can coalesce the live ranges. Ask the
1411 // intervals to coalesce themselves now.
1412 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo,
1418 // DepthMBBCompare - Comparison predicate that sort first based on the loop
1419 // depth of the basic block (the unsigned), and then on the MBB number.
1420 struct DepthMBBCompare {
1421 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
1422 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
1423 // Deeper loops first
1424 if (LHS.first != RHS.first)
1425 return LHS.first > RHS.first;
1427 // Prefer blocks that are more connected in the CFG. This takes care of
1428 // the most difficult copies first while intervals are short.
1429 unsigned cl = LHS.second->pred_size() + LHS.second->succ_size();
1430 unsigned cr = RHS.second->pred_size() + RHS.second->succ_size();
1434 // As a last resort, sort by block number.
1435 return LHS.second->getNumber() < RHS.second->getNumber();
1440 void SimpleRegisterCoalescing::CopyCoalesceInMBB(MachineBasicBlock *MBB,
1441 std::vector<CopyRec> &TryAgain) {
1442 DEBUG(dbgs() << MBB->getName() << ":\n");
1444 std::vector<CopyRec> VirtCopies;
1445 std::vector<CopyRec> PhysCopies;
1446 std::vector<CopyRec> ImpDefCopies;
1447 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
1449 MachineInstr *Inst = MII++;
1451 // If this isn't a copy nor a extract_subreg, we can't join intervals.
1452 unsigned SrcReg, DstReg;
1453 if (Inst->isCopy()) {
1454 DstReg = Inst->getOperand(0).getReg();
1455 SrcReg = Inst->getOperand(1).getReg();
1456 } else if (Inst->isSubregToReg()) {
1457 DstReg = Inst->getOperand(0).getReg();
1458 SrcReg = Inst->getOperand(2).getReg();
1462 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
1463 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1464 if (li_->hasInterval(SrcReg) && li_->getInterval(SrcReg).empty())
1465 ImpDefCopies.push_back(CopyRec(Inst, 0));
1466 else if (SrcIsPhys || DstIsPhys)
1467 PhysCopies.push_back(CopyRec(Inst, 0));
1469 VirtCopies.push_back(CopyRec(Inst, 0));
1472 // Try coalescing implicit copies and insert_subreg <undef> first,
1473 // followed by copies to / from physical registers, then finally copies
1474 // from virtual registers to virtual registers.
1475 for (unsigned i = 0, e = ImpDefCopies.size(); i != e; ++i) {
1476 CopyRec &TheCopy = ImpDefCopies[i];
1478 if (!JoinCopy(TheCopy, Again))
1480 TryAgain.push_back(TheCopy);
1482 for (unsigned i = 0, e = PhysCopies.size(); i != e; ++i) {
1483 CopyRec &TheCopy = PhysCopies[i];
1485 if (!JoinCopy(TheCopy, Again))
1487 TryAgain.push_back(TheCopy);
1489 for (unsigned i = 0, e = VirtCopies.size(); i != e; ++i) {
1490 CopyRec &TheCopy = VirtCopies[i];
1492 if (!JoinCopy(TheCopy, Again))
1494 TryAgain.push_back(TheCopy);
1498 void SimpleRegisterCoalescing::joinIntervals() {
1499 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
1501 std::vector<CopyRec> TryAgainList;
1502 if (loopInfo->empty()) {
1503 // If there are no loops in the function, join intervals in function order.
1504 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();
1506 CopyCoalesceInMBB(I, TryAgainList);
1508 // Otherwise, join intervals in inner loops before other intervals.
1509 // Unfortunately we can't just iterate over loop hierarchy here because
1510 // there may be more MBB's than BB's. Collect MBB's for sorting.
1512 // Join intervals in the function prolog first. We want to join physical
1513 // registers with virtual registers before the intervals got too long.
1514 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
1515 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();I != E;++I){
1516 MachineBasicBlock *MBB = I;
1517 MBBs.push_back(std::make_pair(loopInfo->getLoopDepth(MBB), I));
1520 // Sort by loop depth.
1521 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
1523 // Finally, join intervals in loop nest order.
1524 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
1525 CopyCoalesceInMBB(MBBs[i].second, TryAgainList);
1528 // Joining intervals can allow other intervals to be joined. Iteratively join
1529 // until we make no progress.
1530 bool ProgressMade = true;
1531 while (ProgressMade) {
1532 ProgressMade = false;
1534 for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
1535 CopyRec &TheCopy = TryAgainList[i];
1540 bool Success = JoinCopy(TheCopy, Again);
1541 if (Success || !Again) {
1542 TheCopy.MI = 0; // Mark this one as done.
1543 ProgressMade = true;
1549 /// Return true if the two specified registers belong to different register
1550 /// classes. The registers may be either phys or virt regs.
1552 SimpleRegisterCoalescing::differingRegisterClasses(unsigned RegA,
1553 unsigned RegB) const {
1554 // Get the register classes for the first reg.
1555 if (TargetRegisterInfo::isPhysicalRegister(RegA)) {
1556 assert(TargetRegisterInfo::isVirtualRegister(RegB) &&
1557 "Shouldn't consider two physregs!");
1558 return !mri_->getRegClass(RegB)->contains(RegA);
1561 // Compare against the regclass for the second reg.
1562 const TargetRegisterClass *RegClassA = mri_->getRegClass(RegA);
1563 if (TargetRegisterInfo::isVirtualRegister(RegB)) {
1564 const TargetRegisterClass *RegClassB = mri_->getRegClass(RegB);
1565 return RegClassA != RegClassB;
1567 return !RegClassA->contains(RegB);
1570 /// lastRegisterUse - Returns the last (non-debug) use of the specific register
1571 /// between cycles Start and End or NULL if there are no uses.
1573 SimpleRegisterCoalescing::lastRegisterUse(SlotIndex Start,
1576 SlotIndex &UseIdx) const{
1577 UseIdx = SlotIndex();
1578 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1579 MachineOperand *LastUse = NULL;
1580 for (MachineRegisterInfo::use_nodbg_iterator I = mri_->use_nodbg_begin(Reg),
1581 E = mri_->use_nodbg_end(); I != E; ++I) {
1582 MachineOperand &Use = I.getOperand();
1583 MachineInstr *UseMI = Use.getParent();
1584 if (UseMI->isIdentityCopy())
1586 SlotIndex Idx = li_->getInstructionIndex(UseMI);
1587 // FIXME: Should this be Idx != UseIdx? SlotIndex() will return something
1588 // that compares higher than any other interval.
1589 if (Idx >= Start && Idx < End && Idx >= UseIdx) {
1591 UseIdx = Idx.getUseIndex();
1597 SlotIndex s = Start;
1598 SlotIndex e = End.getPrevSlot().getBaseIndex();
1600 // Skip deleted instructions
1601 MachineInstr *MI = li_->getInstructionFromIndex(e);
1602 while (e != SlotIndex() && e.getPrevIndex() >= s && !MI) {
1603 e = e.getPrevIndex();
1604 MI = li_->getInstructionFromIndex(e);
1606 if (e < s || MI == NULL)
1609 // Ignore identity copies.
1610 if (!MI->isIdentityCopy())
1611 for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
1612 MachineOperand &Use = MI->getOperand(i);
1613 if (Use.isReg() && Use.isUse() && Use.getReg() &&
1614 tri_->regsOverlap(Use.getReg(), Reg)) {
1615 UseIdx = e.getUseIndex();
1620 e = e.getPrevIndex();
1626 void SimpleRegisterCoalescing::releaseMemory() {
1627 JoinedCopies.clear();
1628 ReMatCopies.clear();
1632 bool SimpleRegisterCoalescing::runOnMachineFunction(MachineFunction &fn) {
1634 mri_ = &fn.getRegInfo();
1635 tm_ = &fn.getTarget();
1636 tri_ = tm_->getRegisterInfo();
1637 tii_ = tm_->getInstrInfo();
1638 li_ = &getAnalysis<LiveIntervals>();
1639 AA = &getAnalysis<AliasAnalysis>();
1640 loopInfo = &getAnalysis<MachineLoopInfo>();
1642 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
1643 << "********** Function: "
1644 << ((Value*)mf_->getFunction())->getName() << '\n');
1646 for (TargetRegisterInfo::regclass_iterator I = tri_->regclass_begin(),
1647 E = tri_->regclass_end(); I != E; ++I)
1648 allocatableRCRegs_.insert(std::make_pair(*I,
1649 tri_->getAllocatableSet(fn, *I)));
1651 // Join (coalesce) intervals if requested.
1652 if (EnableJoining) {
1655 dbgs() << "********** INTERVALS POST JOINING **********\n";
1656 for (LiveIntervals::iterator I = li_->begin(), E = li_->end();
1658 I->second->print(dbgs(), tri_);
1664 // Perform a final pass over the instructions and compute spill weights
1665 // and remove identity moves.
1666 SmallVector<unsigned, 4> DeadDefs;
1667 for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end();
1668 mbbi != mbbe; ++mbbi) {
1669 MachineBasicBlock* mbb = mbbi;
1670 for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
1672 MachineInstr *MI = mii;
1673 if (JoinedCopies.count(MI)) {
1674 // Delete all coalesced copies.
1675 bool DoDelete = true;
1676 assert(MI->isCopyLike() && "Unrecognized copy instruction");
1677 unsigned SrcReg = MI->getOperand(MI->isSubregToReg() ? 2 : 1).getReg();
1678 if (TargetRegisterInfo::isPhysicalRegister(SrcReg) &&
1679 MI->getNumOperands() > 2)
1680 // Do not delete extract_subreg, insert_subreg of physical
1681 // registers unless the definition is dead. e.g.
1682 // %DO<def> = INSERT_SUBREG %D0<undef>, %S0<kill>, 1
1683 // or else the scavenger may complain. LowerSubregs will
1684 // delete them later.
1687 if (MI->allDefsAreDead()) {
1688 LiveInterval &li = li_->getInterval(SrcReg);
1689 if (!ShortenDeadCopySrcLiveRange(li, MI))
1690 ShortenDeadCopyLiveRange(li, MI);
1694 // We need the instruction to adjust liveness, so make it a KILL.
1695 if (MI->isSubregToReg()) {
1696 MI->RemoveOperand(3);
1697 MI->RemoveOperand(1);
1699 MI->setDesc(tii_->get(TargetOpcode::KILL));
1700 mii = llvm::next(mii);
1702 li_->RemoveMachineInstrFromMaps(MI);
1703 mii = mbbi->erase(mii);
1709 // Now check if this is a remat'ed def instruction which is now dead.
1710 if (ReMatDefs.count(MI)) {
1712 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1713 const MachineOperand &MO = MI->getOperand(i);
1716 unsigned Reg = MO.getReg();
1719 if (TargetRegisterInfo::isVirtualRegister(Reg))
1720 DeadDefs.push_back(Reg);
1723 if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
1724 !mri_->use_nodbg_empty(Reg)) {
1730 while (!DeadDefs.empty()) {
1731 unsigned DeadDef = DeadDefs.back();
1732 DeadDefs.pop_back();
1733 RemoveDeadDef(li_->getInterval(DeadDef), MI);
1735 li_->RemoveMachineInstrFromMaps(mii);
1736 mii = mbbi->erase(mii);
1742 // If the move will be an identity move delete it
1743 if (MI->isIdentityCopy()) {
1744 unsigned SrcReg = MI->getOperand(1).getReg();
1745 if (li_->hasInterval(SrcReg)) {
1746 LiveInterval &RegInt = li_->getInterval(SrcReg);
1747 // If def of this move instruction is dead, remove its live range
1748 // from the destination register's live interval.
1749 if (MI->allDefsAreDead()) {
1750 if (!ShortenDeadCopySrcLiveRange(RegInt, MI))
1751 ShortenDeadCopyLiveRange(RegInt, MI);
1754 li_->RemoveMachineInstrFromMaps(MI);
1755 mii = mbbi->erase(mii);
1762 // Check for now unnecessary kill flags.
1763 if (li_->isNotInMIMap(MI)) continue;
1764 SlotIndex DefIdx = li_->getInstructionIndex(MI).getDefIndex();
1765 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1766 MachineOperand &MO = MI->getOperand(i);
1767 if (!MO.isReg() || !MO.isKill()) continue;
1768 unsigned reg = MO.getReg();
1769 if (!reg || !li_->hasInterval(reg)) continue;
1770 if (!li_->getInterval(reg).killedAt(DefIdx))
1771 MO.setIsKill(false);
1780 /// print - Implement the dump method.
1781 void SimpleRegisterCoalescing::print(raw_ostream &O, const Module* m) const {
1785 RegisterCoalescer* llvm::createSimpleRegisterCoalescer() {
1786 return new SimpleRegisterCoalescing();
1789 // Make sure that anything that uses RegisterCoalescer pulls in this file...
1790 DEFINING_FILE_FOR(SimpleRegisterCoalescing)