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/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineInstr.h"
22 #include "llvm/CodeGen/MachineLoopInfo.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/Passes.h"
25 #include "llvm/CodeGen/RegisterCoalescer.h"
26 #include "llvm/Target/TargetInstrInfo.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Target/TargetOptions.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/ADT/SmallSet.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/STLExtras.h"
39 STATISTIC(numJoins , "Number of interval joins performed");
40 STATISTIC(numCrossRCs , "Number of cross class joins performed");
41 STATISTIC(numCommutes , "Number of instruction commuting performed");
42 STATISTIC(numExtends , "Number of copies extended");
43 STATISTIC(NumReMats , "Number of instructions re-materialized");
44 STATISTIC(numPeep , "Number of identity moves eliminated after coalescing");
45 STATISTIC(numAborts , "Number of times interval joining aborted");
46 STATISTIC(numDeadValNo, "Number of valno def marked dead");
48 char SimpleRegisterCoalescing::ID = 0;
50 EnableJoining("join-liveintervals",
51 cl::desc("Coalesce copies (default=true)"),
55 NewHeuristic("new-coalescer-heuristic",
56 cl::desc("Use new coalescer heuristic"),
57 cl::init(false), cl::Hidden);
60 CrossClassJoin("join-cross-class-copies",
61 cl::desc("Coalesce cross register class copies"),
62 cl::init(false), cl::Hidden);
65 PhysJoinTweak("tweak-phys-join-heuristics",
66 cl::desc("Tweak heuristics for joining phys reg with vr"),
67 cl::init(false), cl::Hidden);
69 static RegisterPass<SimpleRegisterCoalescing>
70 X("simple-register-coalescing", "Simple Register Coalescing");
72 // Declare that we implement the RegisterCoalescer interface
73 static RegisterAnalysisGroup<RegisterCoalescer, true/*The Default*/> V(X);
75 const PassInfo *const llvm::SimpleRegisterCoalescingID = &X;
77 void SimpleRegisterCoalescing::getAnalysisUsage(AnalysisUsage &AU) const {
78 AU.addRequired<LiveIntervals>();
79 AU.addPreserved<LiveIntervals>();
80 AU.addRequired<MachineLoopInfo>();
81 AU.addPreserved<MachineLoopInfo>();
82 AU.addPreservedID(MachineDominatorsID);
84 AU.addPreservedID(StrongPHIEliminationID);
86 AU.addPreservedID(PHIEliminationID);
87 AU.addPreservedID(TwoAddressInstructionPassID);
88 MachineFunctionPass::getAnalysisUsage(AU);
91 /// AdjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA
92 /// being the source and IntB being the dest, thus this defines a value number
93 /// in IntB. If the source value number (in IntA) is defined by a copy from B,
94 /// see if we can merge these two pieces of B into a single value number,
95 /// eliminating a copy. For example:
99 /// B1 = A3 <- this copy
101 /// In this case, B0 can be extended to where the B1 copy lives, allowing the B1
102 /// value number to be replaced with B0 (which simplifies the B liveinterval).
104 /// This returns true if an interval was modified.
106 bool SimpleRegisterCoalescing::AdjustCopiesBackFrom(LiveInterval &IntA,
108 MachineInstr *CopyMI) {
109 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
111 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
112 // the example above.
113 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
114 assert(BLR != IntB.end() && "Live range not found!");
115 VNInfo *BValNo = BLR->valno;
117 // Get the location that B is defined at. Two options: either this value has
118 // an unknown definition point or it is defined at CopyIdx. If unknown, we
120 if (!BValNo->copy) return false;
121 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
123 // AValNo is the value number in A that defines the copy, A3 in the example.
124 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyIdx-1);
125 assert(ALR != IntA.end() && "Live range not found!");
126 VNInfo *AValNo = ALR->valno;
127 // If it's re-defined by an early clobber somewhere in the live range, then
128 // it's not safe to eliminate the copy. FIXME: This is a temporary workaround.
130 // 172 %ECX<def> = MOV32rr %reg1039<kill>
131 // 180 INLINEASM <es:subl $5,$1
132 // sbbl $3,$0>, 10, %EAX<def>, 14, %ECX<earlyclobber,def>, 9, %EAX<kill>,
133 // 36, <fi#0>, 1, %reg0, 0, 9, %ECX<kill>, 36, <fi#1>, 1, %reg0, 0
134 // 188 %EAX<def> = MOV32rr %EAX<kill>
135 // 196 %ECX<def> = MOV32rr %ECX<kill>
136 // 204 %ECX<def> = MOV32rr %ECX<kill>
137 // 212 %EAX<def> = MOV32rr %EAX<kill>
138 // 220 %EAX<def> = MOV32rr %EAX
139 // 228 %reg1039<def> = MOV32rr %ECX<kill>
140 // The early clobber operand ties ECX input to the ECX def.
142 // The live interval of ECX is represented as this:
143 // %reg20,inf = [46,47:1)[174,230:0) 0@174-(230) 1@46-(47)
144 // The coalescer has no idea there was a def in the middle of [174,230].
145 if (AValNo->hasRedefByEC())
148 // If AValNo is defined as a copy from IntB, we can potentially process this.
149 // Get the instruction that defines this value number.
150 unsigned SrcReg = li_->getVNInfoSourceReg(AValNo);
151 if (!SrcReg) return false; // Not defined by a copy.
153 // If the value number is not defined by a copy instruction, ignore it.
155 // If the source register comes from an interval other than IntB, we can't
157 if (SrcReg != IntB.reg) return false;
159 // Get the LiveRange in IntB that this value number starts with.
160 LiveInterval::iterator ValLR = IntB.FindLiveRangeContaining(AValNo->def-1);
161 assert(ValLR != IntB.end() && "Live range not found!");
163 // Make sure that the end of the live range is inside the same block as
165 MachineInstr *ValLREndInst = li_->getInstructionFromIndex(ValLR->end-1);
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))) {
181 DOUT << "Interfere with sub-register ";
182 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
187 DOUT << "\nExtending: "; IntB.print(DOUT, tri_);
189 unsigned FillerStart = ValLR->end, FillerEnd = BLR->start;
190 // We are about to delete CopyMI, so need to remove it as the 'instruction
191 // that defines this value #'. Update the the valnum with the new defining
193 BValNo->def = FillerStart;
196 // Okay, we can merge them. We need to insert a new liverange:
197 // [ValLR.end, BLR.begin) of either value number, then we merge the
198 // two value numbers.
199 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
201 // If the IntB live range is assigned to a physical register, and if that
202 // physreg has sub-registers, update their live intervals as well.
203 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
204 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
205 LiveInterval &SRLI = li_->getInterval(*SR);
206 SRLI.addRange(LiveRange(FillerStart, FillerEnd,
207 SRLI.getNextValue(FillerStart, 0, true,
208 li_->getVNInfoAllocator())));
212 // Okay, merge "B1" into the same value number as "B0".
213 if (BValNo != ValLR->valno) {
214 IntB.addKills(ValLR->valno, BValNo->kills);
215 IntB.MergeValueNumberInto(BValNo, ValLR->valno);
217 DOUT << " result = "; IntB.print(DOUT, tri_);
220 // If the source instruction was killing the source register before the
221 // merge, unset the isKill marker given the live range has been extended.
222 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true);
224 ValLREndInst->getOperand(UIdx).setIsKill(false);
225 IntB.removeKill(ValLR->valno, FillerStart);
232 /// HasOtherReachingDefs - Return true if there are definitions of IntB
233 /// other than BValNo val# that can reach uses of AValno val# of IntA.
234 bool SimpleRegisterCoalescing::HasOtherReachingDefs(LiveInterval &IntA,
238 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
240 if (AI->valno != AValNo) continue;
241 LiveInterval::Ranges::iterator BI =
242 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start);
243 if (BI != IntB.ranges.begin())
245 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) {
246 if (BI->valno == BValNo)
248 if (BI->start <= AI->start && BI->end > AI->start)
250 if (BI->start > AI->start && BI->start < AI->end)
257 /// RemoveCopyByCommutingDef - We found a non-trivially-coalescable copy with IntA
258 /// being the source and IntB being the dest, thus this defines a value number
259 /// in IntB. If the source value number (in IntA) is defined by a commutable
260 /// instruction and its other operand is coalesced to the copy dest register,
261 /// see if we can transform the copy into a noop by commuting the definition. For
264 /// A3 = op A2 B0<kill>
266 /// B1 = A3 <- this copy
268 /// = op A3 <- more uses
272 /// B2 = op B0 A2<kill>
274 /// B1 = B2 <- now an identify copy
276 /// = op B2 <- more uses
278 /// This returns true if an interval was modified.
280 bool SimpleRegisterCoalescing::RemoveCopyByCommutingDef(LiveInterval &IntA,
282 MachineInstr *CopyMI) {
283 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
285 // FIXME: For now, only eliminate the copy by commuting its def when the
286 // source register is a virtual register. We want to guard against cases
287 // where the copy is a back edge copy and commuting the def lengthen the
288 // live interval of the source register to the entire loop.
289 if (TargetRegisterInfo::isPhysicalRegister(IntA.reg))
292 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
293 // the example above.
294 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
295 assert(BLR != IntB.end() && "Live range not found!");
296 VNInfo *BValNo = BLR->valno;
298 // Get the location that B is defined at. Two options: either this value has
299 // an unknown definition point or it is defined at CopyIdx. If unknown, we
301 if (!BValNo->copy) return false;
302 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
304 // AValNo is the value number in A that defines the copy, A3 in the example.
305 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyIdx-1);
306 assert(ALR != IntA.end() && "Live range not found!");
307 VNInfo *AValNo = ALR->valno;
308 // If other defs can reach uses of this def, then it's not safe to perform
309 // the optimization. FIXME: Do isPHIDef and isDefAccurate both need to be
311 if (AValNo->isPHIDef() || !AValNo->isDefAccurate() ||
312 AValNo->isUnused() || AValNo->hasPHIKill())
314 MachineInstr *DefMI = li_->getInstructionFromIndex(AValNo->def);
315 const TargetInstrDesc &TID = DefMI->getDesc();
316 if (!TID.isCommutable())
318 // If DefMI is a two-address instruction then commuting it will change the
319 // destination register.
320 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
321 assert(DefIdx != -1);
323 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
325 unsigned Op1, Op2, NewDstIdx;
326 if (!tii_->findCommutedOpIndices(DefMI, Op1, Op2))
330 else if (Op2 == UseOpIdx)
335 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
336 unsigned NewReg = NewDstMO.getReg();
337 if (NewReg != IntB.reg || !NewDstMO.isKill())
340 // Make sure there are no other definitions of IntB that would reach the
341 // uses which the new definition can reach.
342 if (HasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
345 // If some of the uses of IntA.reg is already coalesced away, return false.
346 // It's not possible to determine whether it's safe to perform the coalescing.
347 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(IntA.reg),
348 UE = mri_->use_end(); UI != UE; ++UI) {
349 MachineInstr *UseMI = &*UI;
350 unsigned UseIdx = li_->getInstructionIndex(UseMI);
351 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
352 if (ULR == IntA.end())
354 if (ULR->valno == AValNo && JoinedCopies.count(UseMI))
358 // At this point we have decided that it is legal to do this
359 // transformation. Start by commuting the instruction.
360 MachineBasicBlock *MBB = DefMI->getParent();
361 MachineInstr *NewMI = tii_->commuteInstruction(DefMI);
364 if (NewMI != DefMI) {
365 li_->ReplaceMachineInstrInMaps(DefMI, NewMI);
366 MBB->insert(DefMI, NewMI);
369 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false);
370 NewMI->getOperand(OpIdx).setIsKill();
372 bool BHasPHIKill = BValNo->hasPHIKill();
373 SmallVector<VNInfo*, 4> BDeadValNos;
374 VNInfo::KillSet BKills;
375 std::map<unsigned, unsigned> BExtend;
377 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
386 // then do not add kills of A to the newly created B interval.
387 bool Extended = BLR->end > ALR->end && ALR->end != ALR->start;
389 BExtend[ALR->end] = BLR->end;
391 // Update uses of IntA of the specific Val# with IntB.
392 bool BHasSubRegs = false;
393 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg))
394 BHasSubRegs = *tri_->getSubRegisters(IntB.reg);
395 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(IntA.reg),
396 UE = mri_->use_end(); UI != UE;) {
397 MachineOperand &UseMO = UI.getOperand();
398 MachineInstr *UseMI = &*UI;
400 if (JoinedCopies.count(UseMI))
402 unsigned UseIdx = li_->getInstructionIndex(UseMI);
403 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
404 if (ULR == IntA.end() || ULR->valno != AValNo)
406 UseMO.setReg(NewReg);
409 if (UseMO.isKill()) {
411 UseMO.setIsKill(false);
413 BKills.push_back(VNInfo::KillInfo(false, li_->getUseIndex(UseIdx)+1));
415 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
416 if (!tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
418 if (DstReg == IntB.reg) {
419 // This copy will become a noop. If it's defining a new val#,
420 // remove that val# as well. However this live range is being
421 // extended to the end of the existing live range defined by the copy.
422 unsigned DefIdx = li_->getDefIndex(UseIdx);
423 const LiveRange *DLR = IntB.getLiveRangeContaining(DefIdx);
424 BHasPHIKill |= DLR->valno->hasPHIKill();
425 assert(DLR->valno->def == DefIdx);
426 BDeadValNos.push_back(DLR->valno);
427 BExtend[DLR->start] = DLR->end;
428 JoinedCopies.insert(UseMI);
429 // If this is a kill but it's going to be removed, the last use
430 // of the same val# is the new kill.
436 // We need to insert a new liverange: [ALR.start, LastUse). It may be we can
437 // simply extend BLR if CopyMI doesn't end the range.
438 DOUT << "\nExtending: "; IntB.print(DOUT, tri_);
440 // Remove val#'s defined by copies that will be coalesced away.
441 for (unsigned i = 0, e = BDeadValNos.size(); i != e; ++i) {
442 VNInfo *DeadVNI = BDeadValNos[i];
444 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
445 LiveInterval &SRLI = li_->getInterval(*SR);
446 const LiveRange *SRLR = SRLI.getLiveRangeContaining(DeadVNI->def);
447 SRLI.removeValNo(SRLR->valno);
450 IntB.removeValNo(BDeadValNos[i]);
453 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition
454 // is updated. Kills are also updated.
455 VNInfo *ValNo = BValNo;
456 ValNo->def = AValNo->def;
458 for (unsigned j = 0, ee = ValNo->kills.size(); j != ee; ++j) {
459 unsigned Kill = ValNo->kills[j].killIdx;
460 if (Kill != BLR->end)
461 BKills.push_back(VNInfo::KillInfo(ValNo->kills[j].isPHIKill, Kill));
463 ValNo->kills.clear();
464 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
466 if (AI->valno != AValNo) continue;
467 unsigned End = AI->end;
468 std::map<unsigned, unsigned>::iterator EI = BExtend.find(End);
469 if (EI != BExtend.end())
471 IntB.addRange(LiveRange(AI->start, End, ValNo));
473 // If the IntB live range is assigned to a physical register, and if that
474 // physreg has sub-registers, update their live intervals as well.
476 for (const unsigned *SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR) {
477 LiveInterval &SRLI = li_->getInterval(*SR);
478 SRLI.MergeInClobberRange(AI->start, End, li_->getVNInfoAllocator());
482 IntB.addKills(ValNo, BKills);
483 ValNo->setHasPHIKill(BHasPHIKill);
485 DOUT << " result = "; IntB.print(DOUT, tri_);
488 DOUT << "\nShortening: "; IntA.print(DOUT, tri_);
489 IntA.removeValNo(AValNo);
490 DOUT << " result = "; IntA.print(DOUT, tri_);
497 /// isSameOrFallThroughBB - Return true if MBB == SuccMBB or MBB simply
498 /// fallthoughs to SuccMBB.
499 static bool isSameOrFallThroughBB(MachineBasicBlock *MBB,
500 MachineBasicBlock *SuccMBB,
501 const TargetInstrInfo *tii_) {
504 MachineBasicBlock *TBB = 0, *FBB = 0;
505 SmallVector<MachineOperand, 4> Cond;
506 return !tii_->AnalyzeBranch(*MBB, TBB, FBB, Cond) && !TBB && !FBB &&
507 MBB->isSuccessor(SuccMBB);
510 /// removeRange - Wrapper for LiveInterval::removeRange. This removes a range
511 /// from a physical register live interval as well as from the live intervals
512 /// of its sub-registers.
513 static void removeRange(LiveInterval &li, unsigned Start, unsigned End,
514 LiveIntervals *li_, const TargetRegisterInfo *tri_) {
515 li.removeRange(Start, End, true);
516 if (TargetRegisterInfo::isPhysicalRegister(li.reg)) {
517 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
518 if (!li_->hasInterval(*SR))
520 LiveInterval &sli = li_->getInterval(*SR);
521 unsigned RemoveEnd = Start;
522 while (RemoveEnd != End) {
523 LiveInterval::iterator LR = sli.FindLiveRangeContaining(Start);
526 RemoveEnd = (LR->end < End) ? LR->end : End;
527 sli.removeRange(Start, RemoveEnd, true);
534 /// TrimLiveIntervalToLastUse - If there is a last use in the same basic block
535 /// as the copy instruction, trim the live interval to the last use and return
538 SimpleRegisterCoalescing::TrimLiveIntervalToLastUse(unsigned CopyIdx,
539 MachineBasicBlock *CopyMBB,
541 const LiveRange *LR) {
542 unsigned MBBStart = li_->getMBBStartIdx(CopyMBB);
544 MachineOperand *LastUse = lastRegisterUse(LR->start, CopyIdx-1, li.reg,
547 MachineInstr *LastUseMI = LastUse->getParent();
548 if (!isSameOrFallThroughBB(LastUseMI->getParent(), CopyMBB, tii_)) {
555 // r1025<dead> = r1024<kill>
556 if (MBBStart < LR->end)
557 removeRange(li, MBBStart, LR->end, li_, tri_);
561 // There are uses before the copy, just shorten the live range to the end
563 LastUse->setIsKill();
564 removeRange(li, li_->getDefIndex(LastUseIdx), LR->end, li_, tri_);
565 li.addKill(LR->valno, LastUseIdx+1, false);
566 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
567 if (tii_->isMoveInstr(*LastUseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
569 // Last use is itself an identity code.
570 int DeadIdx = LastUseMI->findRegisterDefOperandIdx(li.reg, false, tri_);
571 LastUseMI->getOperand(DeadIdx).setIsDead();
577 if (LR->start <= MBBStart && LR->end > MBBStart) {
578 if (LR->start == 0) {
579 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
580 // Live-in to the function but dead. Remove it from entry live-in set.
581 mf_->begin()->removeLiveIn(li.reg);
583 // FIXME: Shorten intervals in BBs that reaches this BB.
589 /// ReMaterializeTrivialDef - If the source of a copy is defined by a trivial
590 /// computation, replace the copy by rematerialize the definition.
591 bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
593 MachineInstr *CopyMI) {
594 unsigned CopyIdx = li_->getUseIndex(li_->getInstructionIndex(CopyMI));
595 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx);
596 assert(SrcLR != SrcInt.end() && "Live range not found!");
597 VNInfo *ValNo = SrcLR->valno;
598 // If other defs can reach uses of this def, then it's not safe to perform
599 // the optimization. FIXME: Do isPHIDef and isDefAccurate both need to be
601 if (ValNo->isPHIDef() || !ValNo->isDefAccurate() ||
602 ValNo->isUnused() || ValNo->hasPHIKill())
604 MachineInstr *DefMI = li_->getInstructionFromIndex(ValNo->def);
605 const TargetInstrDesc &TID = DefMI->getDesc();
606 if (!TID.isAsCheapAsAMove())
608 if (!DefMI->getDesc().isRematerializable() ||
609 !tii_->isTriviallyReMaterializable(DefMI))
611 bool SawStore = false;
612 if (!DefMI->isSafeToMove(tii_, SawStore))
615 unsigned DefIdx = li_->getDefIndex(CopyIdx);
616 const LiveRange *DLR= li_->getInterval(DstReg).getLiveRangeContaining(DefIdx);
617 DLR->valno->copy = NULL;
618 // Don't forget to update sub-register intervals.
619 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
620 for (const unsigned* SR = tri_->getSubRegisters(DstReg); *SR; ++SR) {
621 if (!li_->hasInterval(*SR))
623 DLR = li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
624 if (DLR && DLR->valno->copy == CopyMI)
625 DLR->valno->copy = NULL;
629 // If copy kills the source register, find the last use and propagate
631 bool checkForDeadDef = false;
632 MachineBasicBlock *MBB = CopyMI->getParent();
633 if (CopyMI->killsRegister(SrcInt.reg))
634 if (!TrimLiveIntervalToLastUse(CopyIdx, MBB, SrcInt, SrcLR)) {
635 checkForDeadDef = true;
638 MachineBasicBlock::iterator MII = next(MachineBasicBlock::iterator(CopyMI));
639 tii_->reMaterialize(*MBB, MII, DstReg, DefMI);
640 MachineInstr *NewMI = prior(MII);
642 if (checkForDeadDef) {
643 // PR4090 fix: Trim interval failed because there was no use of the
644 // source interval in this MBB. If the def is in this MBB too then we
645 // should mark it dead:
646 if (DefMI->getParent() == MBB) {
647 DefMI->addRegisterDead(SrcInt.reg, tri_);
648 SrcLR->end = SrcLR->start + 1;
652 // CopyMI may have implicit operands, transfer them over to the newly
653 // rematerialized instruction. And update implicit def interval valnos.
654 for (unsigned i = CopyMI->getDesc().getNumOperands(),
655 e = CopyMI->getNumOperands(); i != e; ++i) {
656 MachineOperand &MO = CopyMI->getOperand(i);
657 if (MO.isReg() && MO.isImplicit())
658 NewMI->addOperand(MO);
659 if (MO.isDef() && li_->hasInterval(MO.getReg())) {
660 unsigned Reg = MO.getReg();
661 DLR = li_->getInterval(Reg).getLiveRangeContaining(DefIdx);
662 if (DLR && DLR->valno->copy == CopyMI)
663 DLR->valno->copy = NULL;
667 li_->ReplaceMachineInstrInMaps(CopyMI, NewMI);
668 CopyMI->eraseFromParent();
669 ReMatCopies.insert(CopyMI);
670 ReMatDefs.insert(DefMI);
675 /// isBackEdgeCopy - Returns true if CopyMI is a back edge copy.
677 bool SimpleRegisterCoalescing::isBackEdgeCopy(MachineInstr *CopyMI,
678 unsigned DstReg) const {
679 MachineBasicBlock *MBB = CopyMI->getParent();
680 const MachineLoop *L = loopInfo->getLoopFor(MBB);
683 if (MBB != L->getLoopLatch())
686 LiveInterval &LI = li_->getInterval(DstReg);
687 unsigned DefIdx = li_->getInstructionIndex(CopyMI);
688 LiveInterval::const_iterator DstLR =
689 LI.FindLiveRangeContaining(li_->getDefIndex(DefIdx));
690 if (DstLR == LI.end())
692 if (DstLR->valno->kills.size() == 1 && DstLR->valno->kills[0].isPHIKill)
697 /// UpdateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
698 /// update the subregister number if it is not zero. If DstReg is a
699 /// physical register and the existing subregister number of the def / use
700 /// being updated is not zero, make sure to set it to the correct physical
703 SimpleRegisterCoalescing::UpdateRegDefsUses(unsigned SrcReg, unsigned DstReg,
705 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
706 if (DstIsPhys && SubIdx) {
707 // Figure out the real physical register we are updating with.
708 DstReg = tri_->getSubReg(DstReg, SubIdx);
712 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(SrcReg),
713 E = mri_->reg_end(); I != E; ) {
714 MachineOperand &O = I.getOperand();
715 MachineInstr *UseMI = &*I;
717 unsigned OldSubIdx = O.getSubReg();
719 unsigned UseDstReg = DstReg;
721 UseDstReg = tri_->getSubReg(DstReg, OldSubIdx);
723 unsigned CopySrcReg, CopyDstReg, CopySrcSubIdx, CopyDstSubIdx;
724 if (tii_->isMoveInstr(*UseMI, CopySrcReg, CopyDstReg,
725 CopySrcSubIdx, CopyDstSubIdx) &&
726 CopySrcReg != CopyDstReg &&
727 CopySrcReg == SrcReg && CopyDstReg != UseDstReg) {
728 // If the use is a copy and it won't be coalesced away, and its source
729 // is defined by a trivial computation, try to rematerialize it instead.
730 if (ReMaterializeTrivialDef(li_->getInterval(SrcReg), CopyDstReg,UseMI))
739 // Sub-register indexes goes from small to large. e.g.
740 // RAX: 1 -> AL, 2 -> AX, 3 -> EAX
741 // EAX: 1 -> AL, 2 -> AX
742 // So RAX's sub-register 2 is AX, RAX's sub-regsiter 3 is EAX, whose
743 // sub-register 2 is also AX.
744 if (SubIdx && OldSubIdx && SubIdx != OldSubIdx)
745 assert(OldSubIdx < SubIdx && "Conflicting sub-register index!");
748 // Remove would-be duplicated kill marker.
749 if (O.isKill() && UseMI->killsRegister(DstReg))
753 // After updating the operand, check if the machine instruction has
754 // become a copy. If so, update its val# information.
755 if (JoinedCopies.count(UseMI))
758 const TargetInstrDesc &TID = UseMI->getDesc();
759 unsigned CopySrcReg, CopyDstReg, CopySrcSubIdx, CopyDstSubIdx;
760 if (TID.getNumDefs() == 1 && TID.getNumOperands() > 2 &&
761 tii_->isMoveInstr(*UseMI, CopySrcReg, CopyDstReg,
762 CopySrcSubIdx, CopyDstSubIdx) &&
763 CopySrcReg != CopyDstReg &&
764 (TargetRegisterInfo::isVirtualRegister(CopyDstReg) ||
765 allocatableRegs_[CopyDstReg])) {
766 LiveInterval &LI = li_->getInterval(CopyDstReg);
767 unsigned DefIdx = li_->getDefIndex(li_->getInstructionIndex(UseMI));
768 if (const LiveRange *DLR = LI.getLiveRangeContaining(DefIdx)) {
769 if (DLR->valno->def == DefIdx)
770 DLR->valno->copy = UseMI;
776 /// RemoveDeadImpDef - Remove implicit_def instructions which are "re-defining"
777 /// registers due to insert_subreg coalescing. e.g.
779 /// r1025 = implicit_def
780 /// r1025 = insert_subreg r1025, r1024
784 /// r1025 = implicit_def
785 /// r1025 = insert_subreg r1025, r1025
788 SimpleRegisterCoalescing::RemoveDeadImpDef(unsigned Reg, LiveInterval &LI) {
789 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(Reg),
790 E = mri_->reg_end(); I != E; ) {
791 MachineOperand &O = I.getOperand();
792 MachineInstr *DefMI = &*I;
796 if (DefMI->getOpcode() != TargetInstrInfo::IMPLICIT_DEF)
798 if (!LI.liveBeforeAndAt(li_->getInstructionIndex(DefMI)))
800 li_->RemoveMachineInstrFromMaps(DefMI);
801 DefMI->eraseFromParent();
805 /// RemoveUnnecessaryKills - Remove kill markers that are no longer accurate
806 /// due to live range lengthening as the result of coalescing.
807 void SimpleRegisterCoalescing::RemoveUnnecessaryKills(unsigned Reg,
809 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(Reg),
810 UE = mri_->use_end(); UI != UE; ++UI) {
811 MachineOperand &UseMO = UI.getOperand();
812 if (UseMO.isKill()) {
813 MachineInstr *UseMI = UseMO.getParent();
814 unsigned UseIdx = li_->getUseIndex(li_->getInstructionIndex(UseMI));
815 const LiveRange *UI = LI.getLiveRangeContaining(UseIdx);
816 if (!UI || !LI.isKill(UI->valno, UseIdx+1))
817 UseMO.setIsKill(false);
822 /// removeIntervalIfEmpty - Check if the live interval of a physical register
823 /// is empty, if so remove it and also remove the empty intervals of its
824 /// sub-registers. Return true if live interval is removed.
825 static bool removeIntervalIfEmpty(LiveInterval &li, LiveIntervals *li_,
826 const TargetRegisterInfo *tri_) {
828 if (TargetRegisterInfo::isPhysicalRegister(li.reg))
829 for (const unsigned* SR = tri_->getSubRegisters(li.reg); *SR; ++SR) {
830 if (!li_->hasInterval(*SR))
832 LiveInterval &sli = li_->getInterval(*SR);
834 li_->removeInterval(*SR);
836 li_->removeInterval(li.reg);
842 /// ShortenDeadCopyLiveRange - Shorten a live range defined by a dead copy.
843 /// Return true if live interval is removed.
844 bool SimpleRegisterCoalescing::ShortenDeadCopyLiveRange(LiveInterval &li,
845 MachineInstr *CopyMI) {
846 unsigned CopyIdx = li_->getInstructionIndex(CopyMI);
847 LiveInterval::iterator MLR =
848 li.FindLiveRangeContaining(li_->getDefIndex(CopyIdx));
850 return false; // Already removed by ShortenDeadCopySrcLiveRange.
851 unsigned RemoveStart = MLR->start;
852 unsigned RemoveEnd = MLR->end;
853 // Remove the liverange that's defined by this.
854 if (RemoveEnd == li_->getDefIndex(CopyIdx)+1) {
855 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
856 return removeIntervalIfEmpty(li, li_, tri_);
861 /// RemoveDeadDef - If a def of a live interval is now determined dead, remove
862 /// the val# it defines. If the live interval becomes empty, remove it as well.
863 bool SimpleRegisterCoalescing::RemoveDeadDef(LiveInterval &li,
864 MachineInstr *DefMI) {
865 unsigned DefIdx = li_->getDefIndex(li_->getInstructionIndex(DefMI));
866 LiveInterval::iterator MLR = li.FindLiveRangeContaining(DefIdx);
867 if (DefIdx != MLR->valno->def)
869 li.removeValNo(MLR->valno);
870 return removeIntervalIfEmpty(li, li_, tri_);
873 /// PropagateDeadness - Propagate the dead marker to the instruction which
874 /// defines the val#.
875 static void PropagateDeadness(LiveInterval &li, MachineInstr *CopyMI,
876 unsigned &LRStart, LiveIntervals *li_,
877 const TargetRegisterInfo* tri_) {
878 MachineInstr *DefMI =
879 li_->getInstructionFromIndex(li_->getDefIndex(LRStart));
880 if (DefMI && DefMI != CopyMI) {
881 int DeadIdx = DefMI->findRegisterDefOperandIdx(li.reg, false, tri_);
883 DefMI->getOperand(DeadIdx).setIsDead();
884 // A dead def should have a single cycle interval.
890 /// ShortenDeadCopySrcLiveRange - Shorten a live range as it's artificially
891 /// extended by a dead copy. Mark the last use (if any) of the val# as kill as
892 /// ends the live range there. If there isn't another use, then this live range
893 /// is dead. Return true if live interval is removed.
895 SimpleRegisterCoalescing::ShortenDeadCopySrcLiveRange(LiveInterval &li,
896 MachineInstr *CopyMI) {
897 unsigned CopyIdx = li_->getInstructionIndex(CopyMI);
899 // FIXME: special case: function live in. It can be a general case if the
900 // first instruction index starts at > 0 value.
901 assert(TargetRegisterInfo::isPhysicalRegister(li.reg));
902 // Live-in to the function but dead. Remove it from entry live-in set.
903 if (mf_->begin()->isLiveIn(li.reg))
904 mf_->begin()->removeLiveIn(li.reg);
905 const LiveRange *LR = li.getLiveRangeContaining(CopyIdx);
906 removeRange(li, LR->start, LR->end, li_, tri_);
907 return removeIntervalIfEmpty(li, li_, tri_);
910 LiveInterval::iterator LR = li.FindLiveRangeContaining(CopyIdx-1);
912 // Livein but defined by a phi.
915 unsigned RemoveStart = LR->start;
916 unsigned RemoveEnd = li_->getDefIndex(CopyIdx)+1;
917 if (LR->end > RemoveEnd)
918 // More uses past this copy? Nothing to do.
921 // If there is a last use in the same bb, we can't remove the live range.
922 // Shorten the live interval and return.
923 MachineBasicBlock *CopyMBB = CopyMI->getParent();
924 if (TrimLiveIntervalToLastUse(CopyIdx, CopyMBB, li, LR))
927 MachineBasicBlock *StartMBB = li_->getMBBFromIndex(RemoveStart);
928 if (!isSameOrFallThroughBB(StartMBB, CopyMBB, tii_))
929 // If the live range starts in another mbb and the copy mbb is not a fall
930 // through mbb, then we can only cut the range from the beginning of the
932 RemoveStart = li_->getMBBStartIdx(CopyMBB) + 1;
934 if (LR->valno->def == RemoveStart) {
935 // If the def MI defines the val# and this copy is the only kill of the
936 // val#, then propagate the dead marker.
937 if (li.isOnlyLROfValNo(LR)) {
938 PropagateDeadness(li, CopyMI, RemoveStart, li_, tri_);
941 if (li.isKill(LR->valno, RemoveEnd))
942 li.removeKill(LR->valno, RemoveEnd);
945 removeRange(li, RemoveStart, RemoveEnd, li_, tri_);
946 return removeIntervalIfEmpty(li, li_, tri_);
949 /// CanCoalesceWithImpDef - Returns true if the specified copy instruction
950 /// from an implicit def to another register can be coalesced away.
951 bool SimpleRegisterCoalescing::CanCoalesceWithImpDef(MachineInstr *CopyMI,
953 LiveInterval &ImpLi) const{
954 if (!CopyMI->killsRegister(ImpLi.reg))
956 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
957 LiveInterval::iterator LR = li.FindLiveRangeContaining(CopyIdx);
960 if (LR->valno->hasPHIKill())
962 if (LR->valno->def != CopyIdx)
964 // Make sure all of val# uses are copies.
965 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(li.reg),
966 UE = mri_->use_end(); UI != UE;) {
967 MachineInstr *UseMI = &*UI;
969 if (JoinedCopies.count(UseMI))
971 unsigned UseIdx = li_->getUseIndex(li_->getInstructionIndex(UseMI));
972 LiveInterval::iterator ULR = li.FindLiveRangeContaining(UseIdx);
973 if (ULR == li.end() || ULR->valno != LR->valno)
975 // If the use is not a use, then it's not safe to coalesce the move.
976 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
977 if (!tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
978 if (UseMI->getOpcode() == TargetInstrInfo::INSERT_SUBREG &&
979 UseMI->getOperand(1).getReg() == li.reg)
988 /// TurnCopiesFromValNoToImpDefs - The specified value# is defined by an
989 /// implicit_def and it is being removed. Turn all copies from this value#
990 /// into implicit_defs.
991 void SimpleRegisterCoalescing::TurnCopiesFromValNoToImpDefs(LiveInterval &li,
993 SmallVector<MachineInstr*, 4> ImpDefs;
994 MachineOperand *LastUse = NULL;
995 unsigned LastUseIdx = li_->getUseIndex(VNI->def);
996 for (MachineRegisterInfo::reg_iterator RI = mri_->reg_begin(li.reg),
997 RE = mri_->reg_end(); RI != RE;) {
998 MachineOperand *MO = &RI.getOperand();
999 MachineInstr *MI = &*RI;
1002 if (MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
1003 ImpDefs.push_back(MI);
1006 if (JoinedCopies.count(MI))
1008 unsigned UseIdx = li_->getUseIndex(li_->getInstructionIndex(MI));
1009 LiveInterval::iterator ULR = li.FindLiveRangeContaining(UseIdx);
1010 if (ULR == li.end() || ULR->valno != VNI)
1012 // If the use is a copy, turn it into an identity copy.
1013 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
1014 if (tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
1016 // Change it to an implicit_def.
1017 MI->setDesc(tii_->get(TargetInstrInfo::IMPLICIT_DEF));
1018 for (int i = MI->getNumOperands() - 1, e = 0; i > e; --i)
1019 MI->RemoveOperand(i);
1020 // It's no longer a copy, update the valno it defines.
1021 unsigned DefIdx = li_->getDefIndex(UseIdx);
1022 LiveInterval &DstInt = li_->getInterval(DstReg);
1023 LiveInterval::iterator DLR = DstInt.FindLiveRangeContaining(DefIdx);
1024 assert(DLR != DstInt.end() && "Live range not found!");
1025 assert(DLR->valno->copy == MI);
1026 DLR->valno->copy = NULL;
1027 ReMatCopies.insert(MI);
1028 } else if (UseIdx > LastUseIdx) {
1029 LastUseIdx = UseIdx;
1034 LastUse->setIsKill();
1035 li.addKill(VNI, LastUseIdx+1, false);
1037 // Remove dead implicit_def's.
1038 while (!ImpDefs.empty()) {
1039 MachineInstr *ImpDef = ImpDefs.back();
1041 li_->RemoveMachineInstrFromMaps(ImpDef);
1042 ImpDef->eraseFromParent();
1047 /// isWinToJoinVRWithSrcPhysReg - Return true if it's worth while to join a
1048 /// a virtual destination register with physical source register.
1050 SimpleRegisterCoalescing::isWinToJoinVRWithSrcPhysReg(MachineInstr *CopyMI,
1051 MachineBasicBlock *CopyMBB,
1052 LiveInterval &DstInt,
1053 LiveInterval &SrcInt) {
1054 // If the virtual register live interval is long but it has low use desity,
1055 // do not join them, instead mark the physical register as its allocation
1057 const TargetRegisterClass *RC = mri_->getRegClass(DstInt.reg);
1058 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1059 unsigned Length = li_->getApproximateInstructionCount(DstInt);
1060 if (Length > Threshold &&
1061 (((float)std::distance(mri_->use_begin(DstInt.reg),
1062 mri_->use_end()) / Length) < (1.0 / Threshold)))
1065 // If the virtual register live interval extends into a loop, turn down
1067 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
1068 const MachineLoop *L = loopInfo->getLoopFor(CopyMBB);
1070 // Let's see if the virtual register live interval extends into the loop.
1071 LiveInterval::iterator DLR = DstInt.FindLiveRangeContaining(CopyIdx);
1072 assert(DLR != DstInt.end() && "Live range not found!");
1073 DLR = DstInt.FindLiveRangeContaining(DLR->end+1);
1074 if (DLR != DstInt.end()) {
1075 CopyMBB = li_->getMBBFromIndex(DLR->start);
1076 L = loopInfo->getLoopFor(CopyMBB);
1080 if (!L || Length <= Threshold)
1083 unsigned UseIdx = li_->getUseIndex(CopyIdx);
1084 LiveInterval::iterator SLR = SrcInt.FindLiveRangeContaining(UseIdx);
1085 MachineBasicBlock *SMBB = li_->getMBBFromIndex(SLR->start);
1086 if (loopInfo->getLoopFor(SMBB) != L) {
1087 if (!loopInfo->isLoopHeader(CopyMBB))
1089 // If vr's live interval extends pass the loop header, do not join.
1090 for (MachineBasicBlock::succ_iterator SI = CopyMBB->succ_begin(),
1091 SE = CopyMBB->succ_end(); SI != SE; ++SI) {
1092 MachineBasicBlock *SuccMBB = *SI;
1093 if (SuccMBB == CopyMBB)
1095 if (DstInt.overlaps(li_->getMBBStartIdx(SuccMBB),
1096 li_->getMBBEndIdx(SuccMBB)+1))
1103 /// isWinToJoinVRWithDstPhysReg - Return true if it's worth while to join a
1104 /// copy from a virtual source register to a physical destination register.
1106 SimpleRegisterCoalescing::isWinToJoinVRWithDstPhysReg(MachineInstr *CopyMI,
1107 MachineBasicBlock *CopyMBB,
1108 LiveInterval &DstInt,
1109 LiveInterval &SrcInt) {
1110 // If the virtual register live interval is long but it has low use desity,
1111 // do not join them, instead mark the physical register as its allocation
1113 const TargetRegisterClass *RC = mri_->getRegClass(SrcInt.reg);
1114 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1115 unsigned Length = li_->getApproximateInstructionCount(SrcInt);
1116 if (Length > Threshold &&
1117 (((float)std::distance(mri_->use_begin(SrcInt.reg),
1118 mri_->use_end()) / Length) < (1.0 / Threshold)))
1122 // Must be implicit_def.
1125 // If the virtual register live interval is defined or cross a loop, turn
1126 // down aggressiveness.
1127 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
1128 unsigned UseIdx = li_->getUseIndex(CopyIdx);
1129 LiveInterval::iterator SLR = SrcInt.FindLiveRangeContaining(UseIdx);
1130 assert(SLR != SrcInt.end() && "Live range not found!");
1131 SLR = SrcInt.FindLiveRangeContaining(SLR->start-1);
1132 if (SLR == SrcInt.end())
1134 MachineBasicBlock *SMBB = li_->getMBBFromIndex(SLR->start);
1135 const MachineLoop *L = loopInfo->getLoopFor(SMBB);
1137 if (!L || Length <= Threshold)
1140 if (loopInfo->getLoopFor(CopyMBB) != L) {
1141 if (SMBB != L->getLoopLatch())
1143 // If vr's live interval is extended from before the loop latch, do not
1145 for (MachineBasicBlock::pred_iterator PI = SMBB->pred_begin(),
1146 PE = SMBB->pred_end(); PI != PE; ++PI) {
1147 MachineBasicBlock *PredMBB = *PI;
1148 if (PredMBB == SMBB)
1150 if (SrcInt.overlaps(li_->getMBBStartIdx(PredMBB),
1151 li_->getMBBEndIdx(PredMBB)+1))
1158 /// isWinToJoinCrossClass - Return true if it's profitable to coalesce
1159 /// two virtual registers from different register classes.
1161 SimpleRegisterCoalescing::isWinToJoinCrossClass(unsigned LargeReg,
1163 unsigned Threshold) {
1164 // Then make sure the intervals are *short*.
1165 LiveInterval &LargeInt = li_->getInterval(LargeReg);
1166 LiveInterval &SmallInt = li_->getInterval(SmallReg);
1167 unsigned LargeSize = li_->getApproximateInstructionCount(LargeInt);
1168 unsigned SmallSize = li_->getApproximateInstructionCount(SmallInt);
1169 if (SmallSize > Threshold || LargeSize > Threshold)
1170 if ((float)std::distance(mri_->use_begin(SmallReg),
1171 mri_->use_end()) / SmallSize <
1172 (float)std::distance(mri_->use_begin(LargeReg),
1173 mri_->use_end()) / LargeSize)
1178 /// HasIncompatibleSubRegDefUse - If we are trying to coalesce a virtual
1179 /// register with a physical register, check if any of the virtual register
1180 /// operand is a sub-register use or def. If so, make sure it won't result
1181 /// in an illegal extract_subreg or insert_subreg instruction. e.g.
1182 /// vr1024 = extract_subreg vr1025, 1
1184 /// vr1024 = mov8rr AH
1185 /// If vr1024 is coalesced with AH, the extract_subreg is now illegal since
1186 /// AH does not have a super-reg whose sub-register 1 is AH.
1188 SimpleRegisterCoalescing::HasIncompatibleSubRegDefUse(MachineInstr *CopyMI,
1191 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(VirtReg),
1192 E = mri_->reg_end(); I != E; ++I) {
1193 MachineOperand &O = I.getOperand();
1194 MachineInstr *MI = &*I;
1195 if (MI == CopyMI || JoinedCopies.count(MI))
1197 unsigned SubIdx = O.getSubReg();
1198 if (SubIdx && !tri_->getSubReg(PhysReg, SubIdx))
1200 if (MI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) {
1201 SubIdx = MI->getOperand(2).getImm();
1202 if (O.isUse() && !tri_->getSubReg(PhysReg, SubIdx))
1205 unsigned SrcReg = MI->getOperand(1).getReg();
1206 const TargetRegisterClass *RC =
1207 TargetRegisterInfo::isPhysicalRegister(SrcReg)
1208 ? tri_->getPhysicalRegisterRegClass(SrcReg)
1209 : mri_->getRegClass(SrcReg);
1210 if (!tri_->getMatchingSuperReg(PhysReg, SubIdx, RC))
1214 if (MI->getOpcode() == TargetInstrInfo::INSERT_SUBREG ||
1215 MI->getOpcode() == TargetInstrInfo::SUBREG_TO_REG) {
1216 SubIdx = MI->getOperand(3).getImm();
1217 if (VirtReg == MI->getOperand(0).getReg()) {
1218 if (!tri_->getSubReg(PhysReg, SubIdx))
1221 unsigned DstReg = MI->getOperand(0).getReg();
1222 const TargetRegisterClass *RC =
1223 TargetRegisterInfo::isPhysicalRegister(DstReg)
1224 ? tri_->getPhysicalRegisterRegClass(DstReg)
1225 : mri_->getRegClass(DstReg);
1226 if (!tri_->getMatchingSuperReg(PhysReg, SubIdx, RC))
1235 /// CanJoinExtractSubRegToPhysReg - Return true if it's possible to coalesce
1236 /// an extract_subreg where dst is a physical register, e.g.
1237 /// cl = EXTRACT_SUBREG reg1024, 1
1239 SimpleRegisterCoalescing::CanJoinExtractSubRegToPhysReg(unsigned DstReg,
1240 unsigned SrcReg, unsigned SubIdx,
1241 unsigned &RealDstReg) {
1242 const TargetRegisterClass *RC = mri_->getRegClass(SrcReg);
1243 RealDstReg = tri_->getMatchingSuperReg(DstReg, SubIdx, RC);
1244 assert(RealDstReg && "Invalid extract_subreg instruction!");
1246 // For this type of EXTRACT_SUBREG, conservatively
1247 // check if the live interval of the source register interfere with the
1248 // actual super physical register we are trying to coalesce with.
1249 LiveInterval &RHS = li_->getInterval(SrcReg);
1250 if (li_->hasInterval(RealDstReg) &&
1251 RHS.overlaps(li_->getInterval(RealDstReg))) {
1252 DOUT << "Interfere with register ";
1253 DEBUG(li_->getInterval(RealDstReg).print(DOUT, tri_));
1254 return false; // Not coalescable
1256 for (const unsigned* SR = tri_->getSubRegisters(RealDstReg); *SR; ++SR)
1257 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
1258 DOUT << "Interfere with sub-register ";
1259 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
1260 return false; // Not coalescable
1265 /// CanJoinInsertSubRegToPhysReg - Return true if it's possible to coalesce
1266 /// an insert_subreg where src is a physical register, e.g.
1267 /// reg1024 = INSERT_SUBREG reg1024, c1, 0
1269 SimpleRegisterCoalescing::CanJoinInsertSubRegToPhysReg(unsigned DstReg,
1270 unsigned SrcReg, unsigned SubIdx,
1271 unsigned &RealSrcReg) {
1272 const TargetRegisterClass *RC = mri_->getRegClass(DstReg);
1273 RealSrcReg = tri_->getMatchingSuperReg(SrcReg, SubIdx, RC);
1274 assert(RealSrcReg && "Invalid extract_subreg instruction!");
1276 LiveInterval &RHS = li_->getInterval(DstReg);
1277 if (li_->hasInterval(RealSrcReg) &&
1278 RHS.overlaps(li_->getInterval(RealSrcReg))) {
1279 DOUT << "Interfere with register ";
1280 DEBUG(li_->getInterval(RealSrcReg).print(DOUT, tri_));
1281 return false; // Not coalescable
1283 for (const unsigned* SR = tri_->getSubRegisters(RealSrcReg); *SR; ++SR)
1284 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
1285 DOUT << "Interfere with sub-register ";
1286 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
1287 return false; // Not coalescable
1292 /// getRegAllocPreference - Return register allocation preference register.
1294 static unsigned getRegAllocPreference(unsigned Reg, MachineFunction &MF,
1295 MachineRegisterInfo *MRI,
1296 const TargetRegisterInfo *TRI) {
1297 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1299 std::pair<unsigned, unsigned> Hint = MRI->getRegAllocationHint(Reg);
1300 return TRI->ResolveRegAllocHint(Hint.first, Hint.second, MF);
1303 /// JoinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
1304 /// which are the src/dst of the copy instruction CopyMI. This returns true
1305 /// if the copy was successfully coalesced away. If it is not currently
1306 /// possible to coalesce this interval, but it may be possible if other
1307 /// things get coalesced, then it returns true by reference in 'Again'.
1308 bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) {
1309 MachineInstr *CopyMI = TheCopy.MI;
1312 if (JoinedCopies.count(CopyMI) || ReMatCopies.count(CopyMI))
1313 return false; // Already done.
1315 DOUT << li_->getInstructionIndex(CopyMI) << '\t' << *CopyMI;
1317 unsigned SrcReg, DstReg, SrcSubIdx = 0, DstSubIdx = 0;
1318 bool isExtSubReg = CopyMI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG;
1319 bool isInsSubReg = CopyMI->getOpcode() == TargetInstrInfo::INSERT_SUBREG;
1320 bool isSubRegToReg = CopyMI->getOpcode() == TargetInstrInfo::SUBREG_TO_REG;
1321 unsigned SubIdx = 0;
1323 DstReg = CopyMI->getOperand(0).getReg();
1324 DstSubIdx = CopyMI->getOperand(0).getSubReg();
1325 SrcReg = CopyMI->getOperand(1).getReg();
1326 SrcSubIdx = CopyMI->getOperand(2).getImm();
1327 } else if (isInsSubReg || isSubRegToReg) {
1328 if (CopyMI->getOperand(2).getSubReg()) {
1329 DOUT << "\tSource of insert_subreg is already coalesced "
1330 << "to another register.\n";
1331 return false; // Not coalescable.
1333 DstReg = CopyMI->getOperand(0).getReg();
1334 DstSubIdx = CopyMI->getOperand(3).getImm();
1335 SrcReg = CopyMI->getOperand(2).getReg();
1336 } else if (!tii_->isMoveInstr(*CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)){
1337 LLVM_UNREACHABLE("Unrecognized copy instruction!");
1340 // If they are already joined we continue.
1341 if (SrcReg == DstReg) {
1342 DOUT << "\tCopy already coalesced.\n";
1343 return false; // Not coalescable.
1346 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
1347 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1349 // If they are both physical registers, we cannot join them.
1350 if (SrcIsPhys && DstIsPhys) {
1351 DOUT << "\tCan not coalesce physregs.\n";
1352 return false; // Not coalescable.
1355 // We only join virtual registers with allocatable physical registers.
1356 if (SrcIsPhys && !allocatableRegs_[SrcReg]) {
1357 DOUT << "\tSrc reg is unallocatable physreg.\n";
1358 return false; // Not coalescable.
1360 if (DstIsPhys && !allocatableRegs_[DstReg]) {
1361 DOUT << "\tDst reg is unallocatable physreg.\n";
1362 return false; // Not coalescable.
1365 // Check that a physical source register is compatible with dst regclass
1367 unsigned SrcSubReg = SrcSubIdx ?
1368 tri_->getSubReg(SrcReg, SrcSubIdx) : SrcReg;
1369 const TargetRegisterClass *DstRC = mri_->getRegClass(DstReg);
1370 const TargetRegisterClass *DstSubRC = DstRC;
1372 DstSubRC = DstRC->getSubRegisterRegClass(DstSubIdx);
1373 assert(DstSubRC && "Illegal subregister index");
1374 if (!DstSubRC->contains(SrcSubReg)) {
1375 DOUT << "\tIncompatible destination regclass: "
1376 << tri_->getName(SrcSubReg) << " not in " << DstSubRC->getName()
1378 return false; // Not coalescable.
1382 // Check that a physical dst register is compatible with source regclass
1384 unsigned DstSubReg = DstSubIdx ?
1385 tri_->getSubReg(DstReg, DstSubIdx) : DstReg;
1386 const TargetRegisterClass *SrcRC = mri_->getRegClass(SrcReg);
1387 const TargetRegisterClass *SrcSubRC = SrcRC;
1389 SrcSubRC = SrcRC->getSubRegisterRegClass(SrcSubIdx);
1390 assert(SrcSubRC && "Illegal subregister index");
1391 if (!SrcSubRC->contains(DstReg)) {
1392 DOUT << "\tIncompatible source regclass: "
1393 << tri_->getName(DstSubReg) << " not in " << SrcSubRC->getName()
1395 return false; // Not coalescable.
1399 // Should be non-null only when coalescing to a sub-register class.
1400 bool CrossRC = false;
1401 const TargetRegisterClass *NewRC = NULL;
1402 MachineBasicBlock *CopyMBB = CopyMI->getParent();
1403 unsigned RealDstReg = 0;
1404 unsigned RealSrcReg = 0;
1405 if (isExtSubReg || isInsSubReg || isSubRegToReg) {
1406 SubIdx = CopyMI->getOperand(isExtSubReg ? 2 : 3).getImm();
1407 if (SrcIsPhys && isExtSubReg) {
1408 // r1024 = EXTRACT_SUBREG EAX, 0 then r1024 is really going to be
1409 // coalesced with AX.
1410 unsigned DstSubIdx = CopyMI->getOperand(0).getSubReg();
1412 // r1024<2> = EXTRACT_SUBREG EAX, 2. Then r1024 has already been
1413 // coalesced to a larger register so the subreg indices cancel out.
1414 if (DstSubIdx != SubIdx) {
1415 DOUT << "\t Sub-register indices mismatch.\n";
1416 return false; // Not coalescable.
1419 SrcReg = tri_->getSubReg(SrcReg, SubIdx);
1421 } else if (DstIsPhys && (isInsSubReg || isSubRegToReg)) {
1422 // EAX = INSERT_SUBREG EAX, r1024, 0
1423 unsigned SrcSubIdx = CopyMI->getOperand(2).getSubReg();
1425 // EAX = INSERT_SUBREG EAX, r1024<2>, 2 Then r1024 has already been
1426 // coalesced to a larger register so the subreg indices cancel out.
1427 if (SrcSubIdx != SubIdx) {
1428 DOUT << "\t Sub-register indices mismatch.\n";
1429 return false; // Not coalescable.
1432 DstReg = tri_->getSubReg(DstReg, SubIdx);
1434 } else if ((DstIsPhys && isExtSubReg) ||
1435 (SrcIsPhys && (isInsSubReg || isSubRegToReg))) {
1436 if (!isSubRegToReg && CopyMI->getOperand(1).getSubReg()) {
1437 DOUT << "\tSrc of extract_subreg already coalesced with reg"
1438 << " of a super-class.\n";
1439 return false; // Not coalescable.
1443 if (!CanJoinExtractSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealDstReg))
1444 return false; // Not coalescable
1446 if (!CanJoinInsertSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealSrcReg))
1447 return false; // Not coalescable
1451 unsigned OldSubIdx = isExtSubReg ? CopyMI->getOperand(0).getSubReg()
1452 : CopyMI->getOperand(2).getSubReg();
1454 if (OldSubIdx == SubIdx && !differingRegisterClasses(SrcReg, DstReg))
1455 // r1024<2> = EXTRACT_SUBREG r1025, 2. Then r1024 has already been
1456 // coalesced to a larger register so the subreg indices cancel out.
1457 // Also check if the other larger register is of the same register
1458 // class as the would be resulting register.
1461 DOUT << "\t Sub-register indices mismatch.\n";
1462 return false; // Not coalescable.
1466 unsigned LargeReg = isExtSubReg ? SrcReg : DstReg;
1467 unsigned SmallReg = isExtSubReg ? DstReg : SrcReg;
1468 unsigned Limit= allocatableRCRegs_[mri_->getRegClass(SmallReg)].count();
1469 if (!isWinToJoinCrossClass(LargeReg, SmallReg, Limit)) {
1470 Again = true; // May be possible to coalesce later.
1475 } else if (differingRegisterClasses(SrcReg, DstReg)) {
1476 if (!CrossClassJoin)
1480 // FIXME: What if the result of a EXTRACT_SUBREG is then coalesced
1481 // with another? If it's the resulting destination register, then
1482 // the subidx must be propagated to uses (but only those defined
1483 // by the EXTRACT_SUBREG). If it's being coalesced into another
1484 // register, it should be safe because register is assumed to have
1485 // the register class of the super-register.
1487 // Process moves where one of the registers have a sub-register index.
1488 MachineOperand *DstMO = CopyMI->findRegisterDefOperand(DstReg);
1489 MachineOperand *SrcMO = CopyMI->findRegisterUseOperand(SrcReg);
1490 SubIdx = DstMO->getSubReg();
1492 if (SrcMO->getSubReg())
1493 // FIXME: can we handle this?
1495 // This is not an insert_subreg but it looks like one.
1496 // e.g. %reg1024:4 = MOV32rr %EAX
1499 if (!CanJoinInsertSubRegToPhysReg(DstReg, SrcReg, SubIdx, RealSrcReg))
1500 return false; // Not coalescable
1504 SubIdx = SrcMO->getSubReg();
1506 // This is not a extract_subreg but it looks like one.
1507 // e.g. %cl = MOV16rr %reg1024:1
1510 if (!CanJoinExtractSubRegToPhysReg(DstReg, SrcReg, SubIdx,RealDstReg))
1511 return false; // Not coalescable
1517 const TargetRegisterClass *SrcRC= SrcIsPhys ? 0 : mri_->getRegClass(SrcReg);
1518 const TargetRegisterClass *DstRC= DstIsPhys ? 0 : mri_->getRegClass(DstReg);
1519 unsigned LargeReg = SrcReg;
1520 unsigned SmallReg = DstReg;
1523 // Now determine the register class of the joined register.
1525 if (SubIdx && DstRC && DstRC->isASubClass()) {
1526 // This is a move to a sub-register class. However, the source is a
1527 // sub-register of a larger register class. We don't know what should
1528 // the register class be. FIXME.
1532 Limit = allocatableRCRegs_[DstRC].count();
1533 } else if (!SrcIsPhys && !DstIsPhys) {
1534 NewRC = getCommonSubClass(SrcRC, DstRC);
1536 DOUT << "\tDisjoint regclasses: "
1537 << SrcRC->getName() << ", "
1538 << DstRC->getName() << ".\n";
1539 return false; // Not coalescable.
1541 if (DstRC->getSize() > SrcRC->getSize())
1542 std::swap(LargeReg, SmallReg);
1545 // If we are joining two virtual registers and the resulting register
1546 // class is more restrictive (fewer register, smaller size). Check if it's
1547 // worth doing the merge.
1548 if (!SrcIsPhys && !DstIsPhys &&
1549 (isExtSubReg || DstRC->isASubClass()) &&
1550 !isWinToJoinCrossClass(LargeReg, SmallReg,
1551 allocatableRCRegs_[NewRC].count())) {
1552 DOUT << "\tSrc/Dest are different register classes.\n";
1553 // Allow the coalescer to try again in case either side gets coalesced to
1554 // a physical register that's compatible with the other side. e.g.
1555 // r1024 = MOV32to32_ r1025
1556 // But later r1024 is assigned EAX then r1025 may be coalesced with EAX.
1557 Again = true; // May be possible to coalesce later.
1562 // Will it create illegal extract_subreg / insert_subreg?
1563 if (SrcIsPhys && HasIncompatibleSubRegDefUse(CopyMI, DstReg, SrcReg))
1565 if (DstIsPhys && HasIncompatibleSubRegDefUse(CopyMI, SrcReg, DstReg))
1568 LiveInterval &SrcInt = li_->getInterval(SrcReg);
1569 LiveInterval &DstInt = li_->getInterval(DstReg);
1570 assert(SrcInt.reg == SrcReg && DstInt.reg == DstReg &&
1571 "Register mapping is horribly broken!");
1573 DOUT << "\t\tInspecting "; SrcInt.print(DOUT, tri_);
1574 DOUT << " and "; DstInt.print(DOUT, tri_);
1577 // Save a copy of the virtual register live interval. We'll manually
1578 // merge this into the "real" physical register live interval this is
1580 LiveInterval *SavedLI = 0;
1582 SavedLI = li_->dupInterval(&SrcInt);
1583 else if (RealSrcReg)
1584 SavedLI = li_->dupInterval(&DstInt);
1586 // Check if it is necessary to propagate "isDead" property.
1587 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg) {
1588 MachineOperand *mopd = CopyMI->findRegisterDefOperand(DstReg, false);
1589 bool isDead = mopd->isDead();
1591 // We need to be careful about coalescing a source physical register with a
1592 // virtual register. Once the coalescing is done, it cannot be broken and
1593 // these are not spillable! If the destination interval uses are far away,
1594 // think twice about coalescing them!
1595 if (!isDead && (SrcIsPhys || DstIsPhys)) {
1596 // If the copy is in a loop, take care not to coalesce aggressively if the
1597 // src is coming in from outside the loop (or the dst is out of the loop).
1598 // If it's not in a loop, then determine whether to join them base purely
1599 // by the length of the interval.
1600 if (PhysJoinTweak) {
1602 if (!isWinToJoinVRWithSrcPhysReg(CopyMI, CopyMBB, DstInt, SrcInt)) {
1603 mri_->setRegAllocationHint(DstInt.reg, 0, SrcReg);
1605 DOUT << "\tMay tie down a physical register, abort!\n";
1606 Again = true; // May be possible to coalesce later.
1610 if (!isWinToJoinVRWithDstPhysReg(CopyMI, CopyMBB, DstInt, SrcInt)) {
1611 mri_->setRegAllocationHint(SrcInt.reg, 0, DstReg);
1613 DOUT << "\tMay tie down a physical register, abort!\n";
1614 Again = true; // May be possible to coalesce later.
1619 // If the virtual register live interval is long but it has low use desity,
1620 // do not join them, instead mark the physical register as its allocation
1622 LiveInterval &JoinVInt = SrcIsPhys ? DstInt : SrcInt;
1623 unsigned JoinVReg = SrcIsPhys ? DstReg : SrcReg;
1624 unsigned JoinPReg = SrcIsPhys ? SrcReg : DstReg;
1625 const TargetRegisterClass *RC = mri_->getRegClass(JoinVReg);
1626 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
1627 if (TheCopy.isBackEdge)
1628 Threshold *= 2; // Favors back edge copies.
1630 unsigned Length = li_->getApproximateInstructionCount(JoinVInt);
1631 float Ratio = 1.0 / Threshold;
1632 if (Length > Threshold &&
1633 (((float)std::distance(mri_->use_begin(JoinVReg),
1634 mri_->use_end()) / Length) < Ratio)) {
1635 mri_->setRegAllocationHint(JoinVInt.reg, 0, JoinPReg);
1637 DOUT << "\tMay tie down a physical register, abort!\n";
1638 Again = true; // May be possible to coalesce later.
1645 // Okay, attempt to join these two intervals. On failure, this returns false.
1646 // Otherwise, if one of the intervals being joined is a physreg, this method
1647 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1648 // been modified, so we can use this information below to update aliases.
1649 bool Swapped = false;
1650 // If SrcInt is implicitly defined, it's safe to coalesce.
1651 bool isEmpty = SrcInt.empty();
1652 if (isEmpty && !CanCoalesceWithImpDef(CopyMI, DstInt, SrcInt)) {
1653 // Only coalesce an empty interval (defined by implicit_def) with
1654 // another interval which has a valno defined by the CopyMI and the CopyMI
1655 // is a kill of the implicit def.
1656 DOUT << "Not profitable!\n";
1660 if (!isEmpty && !JoinIntervals(DstInt, SrcInt, Swapped)) {
1661 // Coalescing failed.
1663 // If definition of source is defined by trivial computation, try
1664 // rematerializing it.
1665 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg &&
1666 ReMaterializeTrivialDef(SrcInt, DstInt.reg, CopyMI))
1669 // If we can eliminate the copy without merging the live ranges, do so now.
1670 if (!isExtSubReg && !isInsSubReg && !isSubRegToReg &&
1671 (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI) ||
1672 RemoveCopyByCommutingDef(SrcInt, DstInt, CopyMI))) {
1673 JoinedCopies.insert(CopyMI);
1677 // Otherwise, we are unable to join the intervals.
1678 DOUT << "Interference!\n";
1679 Again = true; // May be possible to coalesce later.
1683 LiveInterval *ResSrcInt = &SrcInt;
1684 LiveInterval *ResDstInt = &DstInt;
1686 std::swap(SrcReg, DstReg);
1687 std::swap(ResSrcInt, ResDstInt);
1689 assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1690 "LiveInterval::join didn't work right!");
1692 // If we're about to merge live ranges into a physical register live interval,
1693 // we have to update any aliased register's live ranges to indicate that they
1694 // have clobbered values for this range.
1695 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
1696 // If this is a extract_subreg where dst is a physical register, e.g.
1697 // cl = EXTRACT_SUBREG reg1024, 1
1698 // then create and update the actual physical register allocated to RHS.
1699 if (RealDstReg || RealSrcReg) {
1700 LiveInterval &RealInt =
1701 li_->getOrCreateInterval(RealDstReg ? RealDstReg : RealSrcReg);
1702 for (LiveInterval::const_vni_iterator I = SavedLI->vni_begin(),
1703 E = SavedLI->vni_end(); I != E; ++I) {
1704 const VNInfo *ValNo = *I;
1705 VNInfo *NewValNo = RealInt.getNextValue(ValNo->def, ValNo->copy,
1706 false, // updated at *
1707 li_->getVNInfoAllocator());
1708 NewValNo->setFlags(ValNo->getFlags()); // * updated here.
1709 RealInt.addKills(NewValNo, ValNo->kills);
1710 RealInt.MergeValueInAsValue(*SavedLI, ValNo, NewValNo);
1712 RealInt.weight += SavedLI->weight;
1713 DstReg = RealDstReg ? RealDstReg : RealSrcReg;
1716 // Update the liveintervals of sub-registers.
1717 for (const unsigned *AS = tri_->getSubRegisters(DstReg); *AS; ++AS)
1718 li_->getOrCreateInterval(*AS).MergeInClobberRanges(*ResSrcInt,
1719 li_->getVNInfoAllocator());
1722 // If this is a EXTRACT_SUBREG, make sure the result of coalescing is the
1723 // larger super-register.
1724 if ((isExtSubReg || isInsSubReg || isSubRegToReg) &&
1725 !SrcIsPhys && !DstIsPhys) {
1726 if ((isExtSubReg && !Swapped) ||
1727 ((isInsSubReg || isSubRegToReg) && Swapped)) {
1728 ResSrcInt->Copy(*ResDstInt, mri_, li_->getVNInfoAllocator());
1729 std::swap(SrcReg, DstReg);
1730 std::swap(ResSrcInt, ResDstInt);
1734 // Coalescing to a virtual register that is of a sub-register class of the
1735 // other. Make sure the resulting register is set to the right register class.
1739 mri_->setRegClass(DstReg, NewRC);
1743 // Add all copies that define val# in the source interval into the queue.
1744 for (LiveInterval::const_vni_iterator i = ResSrcInt->vni_begin(),
1745 e = ResSrcInt->vni_end(); i != e; ++i) {
1746 const VNInfo *vni = *i;
1747 // FIXME: Do isPHIDef and isDefAccurate both need to be tested?
1748 if (!vni->def || vni->isUnused() || vni->isPHIDef() || !vni->isDefAccurate())
1750 MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
1751 unsigned NewSrcReg, NewDstReg, NewSrcSubIdx, NewDstSubIdx;
1753 JoinedCopies.count(CopyMI) == 0 &&
1754 tii_->isMoveInstr(*CopyMI, NewSrcReg, NewDstReg,
1755 NewSrcSubIdx, NewDstSubIdx)) {
1756 unsigned LoopDepth = loopInfo->getLoopDepth(CopyMBB);
1757 JoinQueue->push(CopyRec(CopyMI, LoopDepth,
1758 isBackEdgeCopy(CopyMI, DstReg)));
1763 // Remember to delete the copy instruction.
1764 JoinedCopies.insert(CopyMI);
1766 // Some live range has been lengthened due to colaescing, eliminate the
1767 // unnecessary kills.
1768 RemoveUnnecessaryKills(SrcReg, *ResDstInt);
1769 if (TargetRegisterInfo::isVirtualRegister(DstReg))
1770 RemoveUnnecessaryKills(DstReg, *ResDstInt);
1775 // r1024 = implicit_def
1778 RemoveDeadImpDef(DstReg, *ResDstInt);
1779 UpdateRegDefsUses(SrcReg, DstReg, SubIdx);
1781 // SrcReg is guarateed to be the register whose live interval that is
1783 li_->removeInterval(SrcReg);
1785 // Update regalloc hint.
1786 tri_->UpdateRegAllocHint(SrcReg, DstReg, *mf_);
1788 // Manually deleted the live interval copy.
1795 // Now the copy is being coalesced away, the val# previously defined
1796 // by the copy is being defined by an IMPLICIT_DEF which defines a zero
1797 // length interval. Remove the val#.
1798 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
1799 const LiveRange *LR = ResDstInt->getLiveRangeContaining(CopyIdx);
1800 VNInfo *ImpVal = LR->valno;
1801 assert(ImpVal->def == CopyIdx);
1802 unsigned NextDef = LR->end;
1803 TurnCopiesFromValNoToImpDefs(*ResDstInt, ImpVal);
1804 ResDstInt->removeValNo(ImpVal);
1805 LR = ResDstInt->FindLiveRangeContaining(NextDef);
1806 if (LR != ResDstInt->end() && LR->valno->def == NextDef) {
1807 // Special case: vr1024 = implicit_def
1808 // vr1024 = insert_subreg vr1024, vr1025, c
1809 // The insert_subreg becomes a "copy" that defines a val# which can itself
1810 // be coalesced away.
1811 MachineInstr *DefMI = li_->getInstructionFromIndex(NextDef);
1812 if (DefMI->getOpcode() == TargetInstrInfo::INSERT_SUBREG)
1813 LR->valno->copy = DefMI;
1817 // If resulting interval has a preference that no longer fits because of subreg
1818 // coalescing, just clear the preference.
1819 unsigned Preference = getRegAllocPreference(ResDstInt->reg, *mf_, mri_, tri_);
1820 if (Preference && (isExtSubReg || isInsSubReg || isSubRegToReg) &&
1821 TargetRegisterInfo::isVirtualRegister(ResDstInt->reg)) {
1822 const TargetRegisterClass *RC = mri_->getRegClass(ResDstInt->reg);
1823 if (!RC->contains(Preference))
1824 mri_->setRegAllocationHint(ResDstInt->reg, 0, 0);
1827 DOUT << "\n\t\tJoined. Result = "; ResDstInt->print(DOUT, tri_);
1834 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
1835 /// compute what the resultant value numbers for each value in the input two
1836 /// ranges will be. This is complicated by copies between the two which can
1837 /// and will commonly cause multiple value numbers to be merged into one.
1839 /// VN is the value number that we're trying to resolve. InstDefiningValue
1840 /// keeps track of the new InstDefiningValue assignment for the result
1841 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
1842 /// whether a value in this or other is a copy from the opposite set.
1843 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
1844 /// already been assigned.
1846 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
1847 /// contains the value number the copy is from.
1849 static unsigned ComputeUltimateVN(VNInfo *VNI,
1850 SmallVector<VNInfo*, 16> &NewVNInfo,
1851 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
1852 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
1853 SmallVector<int, 16> &ThisValNoAssignments,
1854 SmallVector<int, 16> &OtherValNoAssignments) {
1855 unsigned VN = VNI->id;
1857 // If the VN has already been computed, just return it.
1858 if (ThisValNoAssignments[VN] >= 0)
1859 return ThisValNoAssignments[VN];
1860 // assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?");
1862 // If this val is not a copy from the other val, then it must be a new value
1863 // number in the destination.
1864 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
1865 if (I == ThisFromOther.end()) {
1866 NewVNInfo.push_back(VNI);
1867 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
1869 VNInfo *OtherValNo = I->second;
1871 // Otherwise, this *is* a copy from the RHS. If the other side has already
1872 // been computed, return it.
1873 if (OtherValNoAssignments[OtherValNo->id] >= 0)
1874 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
1876 // Mark this value number as currently being computed, then ask what the
1877 // ultimate value # of the other value is.
1878 ThisValNoAssignments[VN] = -2;
1879 unsigned UltimateVN =
1880 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
1881 OtherValNoAssignments, ThisValNoAssignments);
1882 return ThisValNoAssignments[VN] = UltimateVN;
1885 static bool InVector(VNInfo *Val, const SmallVector<VNInfo*, 8> &V) {
1886 return std::find(V.begin(), V.end(), Val) != V.end();
1889 /// RangeIsDefinedByCopyFromReg - Return true if the specified live range of
1890 /// the specified live interval is defined by a copy from the specified
1892 bool SimpleRegisterCoalescing::RangeIsDefinedByCopyFromReg(LiveInterval &li,
1895 unsigned SrcReg = li_->getVNInfoSourceReg(LR->valno);
1898 // FIXME: Do isPHIDef and isDefAccurate both need to be tested?
1899 if ((LR->valno->isPHIDef() || !LR->valno->isDefAccurate()) &&
1900 TargetRegisterInfo::isPhysicalRegister(li.reg) &&
1901 *tri_->getSuperRegisters(li.reg)) {
1902 // It's a sub-register live interval, we may not have precise information.
1904 MachineInstr *DefMI = li_->getInstructionFromIndex(LR->start);
1905 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
1907 tii_->isMoveInstr(*DefMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
1908 DstReg == li.reg && SrcReg == Reg) {
1909 // Cache computed info.
1910 LR->valno->def = LR->start;
1911 LR->valno->copy = DefMI;
1918 /// SimpleJoin - Attempt to joint the specified interval into this one. The
1919 /// caller of this method must guarantee that the RHS only contains a single
1920 /// value number and that the RHS is not defined by a copy from this
1921 /// interval. This returns false if the intervals are not joinable, or it
1922 /// joins them and returns true.
1923 bool SimpleRegisterCoalescing::SimpleJoin(LiveInterval &LHS, LiveInterval &RHS){
1924 assert(RHS.containsOneValue());
1926 // Some number (potentially more than one) value numbers in the current
1927 // interval may be defined as copies from the RHS. Scan the overlapping
1928 // portions of the LHS and RHS, keeping track of this and looking for
1929 // overlapping live ranges that are NOT defined as copies. If these exist, we
1932 LiveInterval::iterator LHSIt = LHS.begin(), LHSEnd = LHS.end();
1933 LiveInterval::iterator RHSIt = RHS.begin(), RHSEnd = RHS.end();
1935 if (LHSIt->start < RHSIt->start) {
1936 LHSIt = std::upper_bound(LHSIt, LHSEnd, RHSIt->start);
1937 if (LHSIt != LHS.begin()) --LHSIt;
1938 } else if (RHSIt->start < LHSIt->start) {
1939 RHSIt = std::upper_bound(RHSIt, RHSEnd, LHSIt->start);
1940 if (RHSIt != RHS.begin()) --RHSIt;
1943 SmallVector<VNInfo*, 8> EliminatedLHSVals;
1946 // Determine if these live intervals overlap.
1947 bool Overlaps = false;
1948 if (LHSIt->start <= RHSIt->start)
1949 Overlaps = LHSIt->end > RHSIt->start;
1951 Overlaps = RHSIt->end > LHSIt->start;
1953 // If the live intervals overlap, there are two interesting cases: if the
1954 // LHS interval is defined by a copy from the RHS, it's ok and we record
1955 // that the LHS value # is the same as the RHS. If it's not, then we cannot
1956 // coalesce these live ranges and we bail out.
1958 // If we haven't already recorded that this value # is safe, check it.
1959 if (!InVector(LHSIt->valno, EliminatedLHSVals)) {
1960 // Copy from the RHS?
1961 if (!RangeIsDefinedByCopyFromReg(LHS, LHSIt, RHS.reg))
1962 return false; // Nope, bail out.
1964 if (LHSIt->contains(RHSIt->valno->def))
1965 // Here is an interesting situation:
1967 // vr1025 = copy vr1024
1972 // Even though vr1025 is copied from vr1024, it's not safe to
1973 // coalesce them since the live range of vr1025 intersects the
1974 // def of vr1024. This happens because vr1025 is assigned the
1975 // value of the previous iteration of vr1024.
1977 EliminatedLHSVals.push_back(LHSIt->valno);
1980 // We know this entire LHS live range is okay, so skip it now.
1981 if (++LHSIt == LHSEnd) break;
1985 if (LHSIt->end < RHSIt->end) {
1986 if (++LHSIt == LHSEnd) break;
1988 // One interesting case to check here. It's possible that we have
1989 // something like "X3 = Y" which defines a new value number in the LHS,
1990 // and is the last use of this liverange of the RHS. In this case, we
1991 // want to notice this copy (so that it gets coalesced away) even though
1992 // the live ranges don't actually overlap.
1993 if (LHSIt->start == RHSIt->end) {
1994 if (InVector(LHSIt->valno, EliminatedLHSVals)) {
1995 // We already know that this value number is going to be merged in
1996 // if coalescing succeeds. Just skip the liverange.
1997 if (++LHSIt == LHSEnd) break;
1999 // Otherwise, if this is a copy from the RHS, mark it as being merged
2001 if (RangeIsDefinedByCopyFromReg(LHS, LHSIt, RHS.reg)) {
2002 if (LHSIt->contains(RHSIt->valno->def))
2003 // Here is an interesting situation:
2005 // vr1025 = copy vr1024
2010 // Even though vr1025 is copied from vr1024, it's not safe to
2011 // coalesced them since live range of vr1025 intersects the
2012 // def of vr1024. This happens because vr1025 is assigned the
2013 // value of the previous iteration of vr1024.
2015 EliminatedLHSVals.push_back(LHSIt->valno);
2017 // We know this entire LHS live range is okay, so skip it now.
2018 if (++LHSIt == LHSEnd) break;
2023 if (++RHSIt == RHSEnd) break;
2027 // If we got here, we know that the coalescing will be successful and that
2028 // the value numbers in EliminatedLHSVals will all be merged together. Since
2029 // the most common case is that EliminatedLHSVals has a single number, we
2030 // optimize for it: if there is more than one value, we merge them all into
2031 // the lowest numbered one, then handle the interval as if we were merging
2032 // with one value number.
2033 VNInfo *LHSValNo = NULL;
2034 if (EliminatedLHSVals.size() > 1) {
2035 // Loop through all the equal value numbers merging them into the smallest
2037 VNInfo *Smallest = EliminatedLHSVals[0];
2038 for (unsigned i = 1, e = EliminatedLHSVals.size(); i != e; ++i) {
2039 if (EliminatedLHSVals[i]->id < Smallest->id) {
2040 // Merge the current notion of the smallest into the smaller one.
2041 LHS.MergeValueNumberInto(Smallest, EliminatedLHSVals[i]);
2042 Smallest = EliminatedLHSVals[i];
2044 // Merge into the smallest.
2045 LHS.MergeValueNumberInto(EliminatedLHSVals[i], Smallest);
2048 LHSValNo = Smallest;
2049 } else if (EliminatedLHSVals.empty()) {
2050 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
2051 *tri_->getSuperRegisters(LHS.reg))
2052 // Imprecise sub-register information. Can't handle it.
2054 LLVM_UNREACHABLE("No copies from the RHS?");
2056 LHSValNo = EliminatedLHSVals[0];
2059 // Okay, now that there is a single LHS value number that we're merging the
2060 // RHS into, update the value number info for the LHS to indicate that the
2061 // value number is defined where the RHS value number was.
2062 const VNInfo *VNI = RHS.getValNumInfo(0);
2063 LHSValNo->def = VNI->def;
2064 LHSValNo->copy = VNI->copy;
2066 // Okay, the final step is to loop over the RHS live intervals, adding them to
2068 if (VNI->hasPHIKill())
2069 LHSValNo->setHasPHIKill(true);
2070 LHS.addKills(LHSValNo, VNI->kills);
2071 LHS.MergeRangesInAsValue(RHS, LHSValNo);
2072 LHS.weight += RHS.weight;
2074 // Update regalloc hint if both are virtual registers.
2075 if (TargetRegisterInfo::isVirtualRegister(LHS.reg) &&
2076 TargetRegisterInfo::isVirtualRegister(RHS.reg)) {
2077 std::pair<unsigned, unsigned> RHSPref = mri_->getRegAllocationHint(RHS.reg);
2078 std::pair<unsigned, unsigned> LHSPref = mri_->getRegAllocationHint(LHS.reg);
2079 if (RHSPref != LHSPref)
2080 mri_->setRegAllocationHint(LHS.reg, RHSPref.first, RHSPref.second);
2083 // Update the liveintervals of sub-registers.
2084 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg))
2085 for (const unsigned *AS = tri_->getSubRegisters(LHS.reg); *AS; ++AS)
2086 li_->getOrCreateInterval(*AS).MergeInClobberRanges(LHS,
2087 li_->getVNInfoAllocator());
2092 /// JoinIntervals - Attempt to join these two intervals. On failure, this
2093 /// returns false. Otherwise, if one of the intervals being joined is a
2094 /// physreg, this method always canonicalizes LHS to be it. The output
2095 /// "RHS" will not have been modified, so we can use this information
2096 /// below to update aliases.
2098 SimpleRegisterCoalescing::JoinIntervals(LiveInterval &LHS, LiveInterval &RHS,
2100 // Compute the final value assignment, assuming that the live ranges can be
2102 SmallVector<int, 16> LHSValNoAssignments;
2103 SmallVector<int, 16> RHSValNoAssignments;
2104 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
2105 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
2106 SmallVector<VNInfo*, 16> NewVNInfo;
2108 // If a live interval is a physical register, conservatively check if any
2109 // of its sub-registers is overlapping the live interval of the virtual
2110 // register. If so, do not coalesce.
2111 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
2112 *tri_->getSubRegisters(LHS.reg)) {
2113 // If it's coalescing a virtual register to a physical register, estimate
2114 // its live interval length. This is the *cost* of scanning an entire live
2115 // interval. If the cost is low, we'll do an exhaustive check instead.
2117 // If this is something like this:
2125 // That is, the live interval of v1024 crosses a bb. Then we can't rely on
2126 // less conservative check. It's possible a sub-register is defined before
2127 // v1024 (or live in) and live out of BB1.
2128 if (RHS.containsOneValue() &&
2129 li_->intervalIsInOneMBB(RHS) &&
2130 li_->getApproximateInstructionCount(RHS) <= 10) {
2131 // Perform a more exhaustive check for some common cases.
2132 if (li_->conflictsWithPhysRegRef(RHS, LHS.reg, true, JoinedCopies))
2135 for (const unsigned* SR = tri_->getSubRegisters(LHS.reg); *SR; ++SR)
2136 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
2137 DOUT << "Interfere with sub-register ";
2138 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
2142 } else if (TargetRegisterInfo::isPhysicalRegister(RHS.reg) &&
2143 *tri_->getSubRegisters(RHS.reg)) {
2144 if (LHS.containsOneValue() &&
2145 li_->getApproximateInstructionCount(LHS) <= 10) {
2146 // Perform a more exhaustive check for some common cases.
2147 if (li_->conflictsWithPhysRegRef(LHS, RHS.reg, false, JoinedCopies))
2150 for (const unsigned* SR = tri_->getSubRegisters(RHS.reg); *SR; ++SR)
2151 if (li_->hasInterval(*SR) && LHS.overlaps(li_->getInterval(*SR))) {
2152 DOUT << "Interfere with sub-register ";
2153 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
2159 // Compute ultimate value numbers for the LHS and RHS values.
2160 if (RHS.containsOneValue()) {
2161 // Copies from a liveinterval with a single value are simple to handle and
2162 // very common, handle the special case here. This is important, because
2163 // often RHS is small and LHS is large (e.g. a physreg).
2165 // Find out if the RHS is defined as a copy from some value in the LHS.
2166 int RHSVal0DefinedFromLHS = -1;
2168 VNInfo *RHSValNoInfo = NULL;
2169 VNInfo *RHSValNoInfo0 = RHS.getValNumInfo(0);
2170 unsigned RHSSrcReg = li_->getVNInfoSourceReg(RHSValNoInfo0);
2171 if (RHSSrcReg == 0 || RHSSrcReg != LHS.reg) {
2172 // If RHS is not defined as a copy from the LHS, we can use simpler and
2173 // faster checks to see if the live ranges are coalescable. This joiner
2174 // can't swap the LHS/RHS intervals though.
2175 if (!TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
2176 return SimpleJoin(LHS, RHS);
2178 RHSValNoInfo = RHSValNoInfo0;
2181 // It was defined as a copy from the LHS, find out what value # it is.
2182 RHSValNoInfo = LHS.getLiveRangeContaining(RHSValNoInfo0->def-1)->valno;
2183 RHSValID = RHSValNoInfo->id;
2184 RHSVal0DefinedFromLHS = RHSValID;
2187 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
2188 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
2189 NewVNInfo.resize(LHS.getNumValNums(), NULL);
2191 // Okay, *all* of the values in LHS that are defined as a copy from RHS
2192 // should now get updated.
2193 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2196 unsigned VN = VNI->id;
2197 if (unsigned LHSSrcReg = li_->getVNInfoSourceReg(VNI)) {
2198 if (LHSSrcReg != RHS.reg) {
2199 // If this is not a copy from the RHS, its value number will be
2200 // unmodified by the coalescing.
2201 NewVNInfo[VN] = VNI;
2202 LHSValNoAssignments[VN] = VN;
2203 } else if (RHSValID == -1) {
2204 // Otherwise, it is a copy from the RHS, and we don't already have a
2205 // value# for it. Keep the current value number, but remember it.
2206 LHSValNoAssignments[VN] = RHSValID = VN;
2207 NewVNInfo[VN] = RHSValNoInfo;
2208 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
2210 // Otherwise, use the specified value #.
2211 LHSValNoAssignments[VN] = RHSValID;
2212 if (VN == (unsigned)RHSValID) { // Else this val# is dead.
2213 NewVNInfo[VN] = RHSValNoInfo;
2214 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
2218 NewVNInfo[VN] = VNI;
2219 LHSValNoAssignments[VN] = VN;
2223 assert(RHSValID != -1 && "Didn't find value #?");
2224 RHSValNoAssignments[0] = RHSValID;
2225 if (RHSVal0DefinedFromLHS != -1) {
2226 // This path doesn't go through ComputeUltimateVN so just set
2228 RHSValsDefinedFromLHS[RHSValNoInfo0] = (VNInfo*)1;
2231 // Loop over the value numbers of the LHS, seeing if any are defined from
2233 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2236 if (VNI->isUnused() || VNI->copy == 0) // Src not defined by a copy?
2239 // DstReg is known to be a register in the LHS interval. If the src is
2240 // from the RHS interval, we can use its value #.
2241 if (li_->getVNInfoSourceReg(VNI) != RHS.reg)
2244 // Figure out the value # from the RHS.
2245 LHSValsDefinedFromRHS[VNI]=RHS.getLiveRangeContaining(VNI->def-1)->valno;
2248 // Loop over the value numbers of the RHS, seeing if any are defined from
2250 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
2253 if (VNI->isUnused() || VNI->copy == 0) // Src not defined by a copy?
2256 // DstReg is known to be a register in the RHS interval. If the src is
2257 // from the LHS interval, we can use its value #.
2258 if (li_->getVNInfoSourceReg(VNI) != LHS.reg)
2261 // Figure out the value # from the LHS.
2262 RHSValsDefinedFromLHS[VNI]=LHS.getLiveRangeContaining(VNI->def-1)->valno;
2265 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
2266 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
2267 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
2269 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
2272 unsigned VN = VNI->id;
2273 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused())
2275 ComputeUltimateVN(VNI, NewVNInfo,
2276 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
2277 LHSValNoAssignments, RHSValNoAssignments);
2279 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
2282 unsigned VN = VNI->id;
2283 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused())
2285 // If this value number isn't a copy from the LHS, it's a new number.
2286 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
2287 NewVNInfo.push_back(VNI);
2288 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
2292 ComputeUltimateVN(VNI, NewVNInfo,
2293 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
2294 RHSValNoAssignments, LHSValNoAssignments);
2298 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
2299 // interval lists to see if these intervals are coalescable.
2300 LiveInterval::const_iterator I = LHS.begin();
2301 LiveInterval::const_iterator IE = LHS.end();
2302 LiveInterval::const_iterator J = RHS.begin();
2303 LiveInterval::const_iterator JE = RHS.end();
2305 // Skip ahead until the first place of potential sharing.
2306 if (I->start < J->start) {
2307 I = std::upper_bound(I, IE, J->start);
2308 if (I != LHS.begin()) --I;
2309 } else if (J->start < I->start) {
2310 J = std::upper_bound(J, JE, I->start);
2311 if (J != RHS.begin()) --J;
2315 // Determine if these two live ranges overlap.
2317 if (I->start < J->start) {
2318 Overlaps = I->end > J->start;
2320 Overlaps = J->end > I->start;
2323 // If so, check value # info to determine if they are really different.
2325 // If the live range overlap will map to the same value number in the
2326 // result liverange, we can still coalesce them. If not, we can't.
2327 if (LHSValNoAssignments[I->valno->id] !=
2328 RHSValNoAssignments[J->valno->id])
2332 if (I->end < J->end) {
2341 // Update kill info. Some live ranges are extended due to copy coalescing.
2342 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
2343 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
2344 VNInfo *VNI = I->first;
2345 unsigned LHSValID = LHSValNoAssignments[VNI->id];
2346 LiveInterval::removeKill(NewVNInfo[LHSValID], VNI->def);
2347 if (VNI->hasPHIKill())
2348 NewVNInfo[LHSValID]->setHasPHIKill(true);
2349 RHS.addKills(NewVNInfo[LHSValID], VNI->kills);
2352 // Update kill info. Some live ranges are extended due to copy coalescing.
2353 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
2354 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
2355 VNInfo *VNI = I->first;
2356 unsigned RHSValID = RHSValNoAssignments[VNI->id];
2357 LiveInterval::removeKill(NewVNInfo[RHSValID], VNI->def);
2358 if (VNI->hasPHIKill())
2359 NewVNInfo[RHSValID]->setHasPHIKill(true);
2360 LHS.addKills(NewVNInfo[RHSValID], VNI->kills);
2363 // If we get here, we know that we can coalesce the live ranges. Ask the
2364 // intervals to coalesce themselves now.
2365 if ((RHS.ranges.size() > LHS.ranges.size() &&
2366 TargetRegisterInfo::isVirtualRegister(LHS.reg)) ||
2367 TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
2368 RHS.join(LHS, &RHSValNoAssignments[0], &LHSValNoAssignments[0], NewVNInfo,
2372 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo,
2380 // DepthMBBCompare - Comparison predicate that sort first based on the loop
2381 // depth of the basic block (the unsigned), and then on the MBB number.
2382 struct DepthMBBCompare {
2383 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
2384 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
2385 if (LHS.first > RHS.first) return true; // Deeper loops first
2386 return LHS.first == RHS.first &&
2387 LHS.second->getNumber() < RHS.second->getNumber();
2392 /// getRepIntervalSize - Returns the size of the interval that represents the
2393 /// specified register.
2395 unsigned JoinPriorityQueue<SF>::getRepIntervalSize(unsigned Reg) {
2396 return Rc->getRepIntervalSize(Reg);
2399 /// CopyRecSort::operator - Join priority queue sorting function.
2401 bool CopyRecSort::operator()(CopyRec left, CopyRec right) const {
2402 // Inner loops first.
2403 if (left.LoopDepth > right.LoopDepth)
2405 else if (left.LoopDepth == right.LoopDepth)
2406 if (left.isBackEdge && !right.isBackEdge)
2411 void SimpleRegisterCoalescing::CopyCoalesceInMBB(MachineBasicBlock *MBB,
2412 std::vector<CopyRec> &TryAgain) {
2413 DOUT << ((Value*)MBB->getBasicBlock())->getName() << ":\n";
2415 std::vector<CopyRec> VirtCopies;
2416 std::vector<CopyRec> PhysCopies;
2417 std::vector<CopyRec> ImpDefCopies;
2418 unsigned LoopDepth = loopInfo->getLoopDepth(MBB);
2419 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2421 MachineInstr *Inst = MII++;
2423 // If this isn't a copy nor a extract_subreg, we can't join intervals.
2424 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2425 if (Inst->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) {
2426 DstReg = Inst->getOperand(0).getReg();
2427 SrcReg = Inst->getOperand(1).getReg();
2428 } else if (Inst->getOpcode() == TargetInstrInfo::INSERT_SUBREG ||
2429 Inst->getOpcode() == TargetInstrInfo::SUBREG_TO_REG) {
2430 DstReg = Inst->getOperand(0).getReg();
2431 SrcReg = Inst->getOperand(2).getReg();
2432 } else if (!tii_->isMoveInstr(*Inst, SrcReg, DstReg, SrcSubIdx, DstSubIdx))
2435 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
2436 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
2438 JoinQueue->push(CopyRec(Inst, LoopDepth, isBackEdgeCopy(Inst, DstReg)));
2440 if (li_->hasInterval(SrcReg) && li_->getInterval(SrcReg).empty())
2441 ImpDefCopies.push_back(CopyRec(Inst, 0, false));
2442 else if (SrcIsPhys || DstIsPhys)
2443 PhysCopies.push_back(CopyRec(Inst, 0, false));
2445 VirtCopies.push_back(CopyRec(Inst, 0, false));
2452 // Try coalescing implicit copies first, followed by copies to / from
2453 // physical registers, then finally copies from virtual registers to
2454 // virtual registers.
2455 for (unsigned i = 0, e = ImpDefCopies.size(); i != e; ++i) {
2456 CopyRec &TheCopy = ImpDefCopies[i];
2458 if (!JoinCopy(TheCopy, Again))
2460 TryAgain.push_back(TheCopy);
2462 for (unsigned i = 0, e = PhysCopies.size(); i != e; ++i) {
2463 CopyRec &TheCopy = PhysCopies[i];
2465 if (!JoinCopy(TheCopy, Again))
2467 TryAgain.push_back(TheCopy);
2469 for (unsigned i = 0, e = VirtCopies.size(); i != e; ++i) {
2470 CopyRec &TheCopy = VirtCopies[i];
2472 if (!JoinCopy(TheCopy, Again))
2474 TryAgain.push_back(TheCopy);
2478 void SimpleRegisterCoalescing::joinIntervals() {
2479 DOUT << "********** JOINING INTERVALS ***********\n";
2482 JoinQueue = new JoinPriorityQueue<CopyRecSort>(this);
2484 std::vector<CopyRec> TryAgainList;
2485 if (loopInfo->empty()) {
2486 // If there are no loops in the function, join intervals in function order.
2487 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();
2489 CopyCoalesceInMBB(I, TryAgainList);
2491 // Otherwise, join intervals in inner loops before other intervals.
2492 // Unfortunately we can't just iterate over loop hierarchy here because
2493 // there may be more MBB's than BB's. Collect MBB's for sorting.
2495 // Join intervals in the function prolog first. We want to join physical
2496 // registers with virtual registers before the intervals got too long.
2497 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
2498 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();I != E;++I){
2499 MachineBasicBlock *MBB = I;
2500 MBBs.push_back(std::make_pair(loopInfo->getLoopDepth(MBB), I));
2503 // Sort by loop depth.
2504 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
2506 // Finally, join intervals in loop nest order.
2507 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
2508 CopyCoalesceInMBB(MBBs[i].second, TryAgainList);
2511 // Joining intervals can allow other intervals to be joined. Iteratively join
2512 // until we make no progress.
2514 SmallVector<CopyRec, 16> TryAgain;
2515 bool ProgressMade = true;
2516 while (ProgressMade) {
2517 ProgressMade = false;
2518 while (!JoinQueue->empty()) {
2519 CopyRec R = JoinQueue->pop();
2521 bool Success = JoinCopy(R, Again);
2523 ProgressMade = true;
2525 TryAgain.push_back(R);
2529 while (!TryAgain.empty()) {
2530 JoinQueue->push(TryAgain.back());
2531 TryAgain.pop_back();
2536 bool ProgressMade = true;
2537 while (ProgressMade) {
2538 ProgressMade = false;
2540 for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
2541 CopyRec &TheCopy = TryAgainList[i];
2544 bool Success = JoinCopy(TheCopy, Again);
2545 if (Success || !Again) {
2546 TheCopy.MI = 0; // Mark this one as done.
2547 ProgressMade = true;
2558 /// Return true if the two specified registers belong to different register
2559 /// classes. The registers may be either phys or virt regs.
2561 SimpleRegisterCoalescing::differingRegisterClasses(unsigned RegA,
2562 unsigned RegB) const {
2563 // Get the register classes for the first reg.
2564 if (TargetRegisterInfo::isPhysicalRegister(RegA)) {
2565 assert(TargetRegisterInfo::isVirtualRegister(RegB) &&
2566 "Shouldn't consider two physregs!");
2567 return !mri_->getRegClass(RegB)->contains(RegA);
2570 // Compare against the regclass for the second reg.
2571 const TargetRegisterClass *RegClassA = mri_->getRegClass(RegA);
2572 if (TargetRegisterInfo::isVirtualRegister(RegB)) {
2573 const TargetRegisterClass *RegClassB = mri_->getRegClass(RegB);
2574 return RegClassA != RegClassB;
2576 return !RegClassA->contains(RegB);
2579 /// lastRegisterUse - Returns the last use of the specific register between
2580 /// cycles Start and End or NULL if there are no uses.
2582 SimpleRegisterCoalescing::lastRegisterUse(unsigned Start, unsigned End,
2583 unsigned Reg, unsigned &UseIdx) const{
2585 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2586 MachineOperand *LastUse = NULL;
2587 for (MachineRegisterInfo::use_iterator I = mri_->use_begin(Reg),
2588 E = mri_->use_end(); I != E; ++I) {
2589 MachineOperand &Use = I.getOperand();
2590 MachineInstr *UseMI = Use.getParent();
2591 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2592 if (tii_->isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
2594 // Ignore identity copies.
2596 unsigned Idx = li_->getInstructionIndex(UseMI);
2597 if (Idx >= Start && Idx < End && Idx >= UseIdx) {
2599 UseIdx = li_->getUseIndex(Idx);
2605 int e = (End-1) / InstrSlots::NUM * InstrSlots::NUM;
2608 // Skip deleted instructions
2609 MachineInstr *MI = li_->getInstructionFromIndex(e);
2610 while ((e - InstrSlots::NUM) >= s && !MI) {
2611 e -= InstrSlots::NUM;
2612 MI = li_->getInstructionFromIndex(e);
2614 if (e < s || MI == NULL)
2617 // Ignore identity copies.
2618 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2619 if (!(tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) &&
2621 for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
2622 MachineOperand &Use = MI->getOperand(i);
2623 if (Use.isReg() && Use.isUse() && Use.getReg() &&
2624 tri_->regsOverlap(Use.getReg(), Reg)) {
2625 UseIdx = li_->getUseIndex(e);
2630 e -= InstrSlots::NUM;
2637 void SimpleRegisterCoalescing::printRegName(unsigned reg) const {
2638 if (TargetRegisterInfo::isPhysicalRegister(reg))
2639 cerr << tri_->getName(reg);
2641 cerr << "%reg" << reg;
2644 void SimpleRegisterCoalescing::releaseMemory() {
2645 JoinedCopies.clear();
2646 ReMatCopies.clear();
2650 static bool isZeroLengthInterval(LiveInterval *li) {
2651 for (LiveInterval::Ranges::const_iterator
2652 i = li->ranges.begin(), e = li->ranges.end(); i != e; ++i)
2653 if (i->end - i->start > LiveInterval::InstrSlots::NUM)
2658 /// TurnCopyIntoImpDef - If source of the specified copy is an implicit def,
2659 /// turn the copy into an implicit def.
2661 SimpleRegisterCoalescing::TurnCopyIntoImpDef(MachineBasicBlock::iterator &I,
2662 MachineBasicBlock *MBB,
2663 unsigned DstReg, unsigned SrcReg) {
2664 MachineInstr *CopyMI = &*I;
2665 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
2666 if (!li_->hasInterval(SrcReg))
2668 LiveInterval &SrcInt = li_->getInterval(SrcReg);
2669 if (!SrcInt.empty())
2671 if (!li_->hasInterval(DstReg))
2673 LiveInterval &DstInt = li_->getInterval(DstReg);
2674 const LiveRange *DstLR = DstInt.getLiveRangeContaining(CopyIdx);
2675 // If the valno extends beyond this basic block, then it's not safe to delete
2676 // the val# or else livein information won't be correct.
2677 MachineBasicBlock *EndMBB = li_->getMBBFromIndex(DstLR->end);
2680 DstInt.removeValNo(DstLR->valno);
2681 CopyMI->setDesc(tii_->get(TargetInstrInfo::IMPLICIT_DEF));
2682 for (int i = CopyMI->getNumOperands() - 1, e = 0; i > e; --i)
2683 CopyMI->RemoveOperand(i);
2684 CopyMI->getOperand(0).setIsUndef();
2685 bool NoUse = mri_->use_empty(SrcReg);
2687 for (MachineRegisterInfo::reg_iterator RI = mri_->reg_begin(SrcReg),
2688 RE = mri_->reg_end(); RI != RE; ) {
2689 assert(RI.getOperand().isDef());
2690 MachineInstr *DefMI = &*RI;
2692 // The implicit_def source has no other uses, delete it.
2693 assert(DefMI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF);
2694 li_->RemoveMachineInstrFromMaps(DefMI);
2695 DefMI->eraseFromParent();
2699 // Mark uses of implicit_def isUndef.
2700 for (MachineRegisterInfo::use_iterator RI = mri_->use_begin(DstReg),
2701 RE = mri_->use_end(); RI != RE; ++RI) {
2702 assert((*RI).getParent() == MBB);
2703 RI.getOperand().setIsUndef();
2711 bool SimpleRegisterCoalescing::runOnMachineFunction(MachineFunction &fn) {
2713 mri_ = &fn.getRegInfo();
2714 tm_ = &fn.getTarget();
2715 tri_ = tm_->getRegisterInfo();
2716 tii_ = tm_->getInstrInfo();
2717 li_ = &getAnalysis<LiveIntervals>();
2718 loopInfo = &getAnalysis<MachineLoopInfo>();
2720 DOUT << "********** SIMPLE REGISTER COALESCING **********\n"
2721 << "********** Function: "
2722 << ((Value*)mf_->getFunction())->getName() << '\n';
2724 allocatableRegs_ = tri_->getAllocatableSet(fn);
2725 for (TargetRegisterInfo::regclass_iterator I = tri_->regclass_begin(),
2726 E = tri_->regclass_end(); I != E; ++I)
2727 allocatableRCRegs_.insert(std::make_pair(*I,
2728 tri_->getAllocatableSet(fn, *I)));
2730 // Join (coalesce) intervals if requested.
2731 if (EnableJoining) {
2734 DOUT << "********** INTERVALS POST JOINING **********\n";
2735 for (LiveIntervals::iterator I = li_->begin(), E = li_->end(); I != E; ++I){
2736 I->second->print(DOUT, tri_);
2742 // Perform a final pass over the instructions and compute spill weights
2743 // and remove identity moves.
2744 SmallVector<unsigned, 4> DeadDefs;
2745 for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end();
2746 mbbi != mbbe; ++mbbi) {
2747 MachineBasicBlock* mbb = mbbi;
2748 unsigned loopDepth = loopInfo->getLoopDepth(mbb);
2750 for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
2752 MachineInstr *MI = mii;
2753 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
2754 if (JoinedCopies.count(MI)) {
2755 // Delete all coalesced copies.
2756 if (!tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
2757 assert((MI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG ||
2758 MI->getOpcode() == TargetInstrInfo::INSERT_SUBREG ||
2759 MI->getOpcode() == TargetInstrInfo::SUBREG_TO_REG) &&
2760 "Unrecognized copy instruction");
2761 DstReg = MI->getOperand(0).getReg();
2763 if (MI->registerDefIsDead(DstReg)) {
2764 LiveInterval &li = li_->getInterval(DstReg);
2765 if (!ShortenDeadCopySrcLiveRange(li, MI))
2766 ShortenDeadCopyLiveRange(li, MI);
2768 li_->RemoveMachineInstrFromMaps(MI);
2769 mii = mbbi->erase(mii);
2774 // Now check if this is a remat'ed def instruction which is now dead.
2775 if (ReMatDefs.count(MI)) {
2777 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2778 const MachineOperand &MO = MI->getOperand(i);
2781 unsigned Reg = MO.getReg();
2784 if (TargetRegisterInfo::isVirtualRegister(Reg))
2785 DeadDefs.push_back(Reg);
2788 if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
2789 !mri_->use_empty(Reg)) {
2795 while (!DeadDefs.empty()) {
2796 unsigned DeadDef = DeadDefs.back();
2797 DeadDefs.pop_back();
2798 RemoveDeadDef(li_->getInterval(DeadDef), MI);
2800 li_->RemoveMachineInstrFromMaps(mii);
2801 mii = mbbi->erase(mii);
2807 // If the move will be an identity move delete it
2808 bool isMove= tii_->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx);
2809 if (isMove && SrcReg == DstReg) {
2810 if (li_->hasInterval(SrcReg)) {
2811 LiveInterval &RegInt = li_->getInterval(SrcReg);
2812 // If def of this move instruction is dead, remove its live range
2813 // from the dstination register's live interval.
2814 if (MI->registerDefIsDead(DstReg)) {
2815 if (!ShortenDeadCopySrcLiveRange(RegInt, MI))
2816 ShortenDeadCopyLiveRange(RegInt, MI);
2819 li_->RemoveMachineInstrFromMaps(MI);
2820 mii = mbbi->erase(mii);
2822 } else if (!isMove || !TurnCopyIntoImpDef(mii, mbb, DstReg, SrcReg)) {
2823 SmallSet<unsigned, 4> UniqueUses;
2824 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
2825 const MachineOperand &mop = MI->getOperand(i);
2826 if (mop.isReg() && mop.getReg() &&
2827 TargetRegisterInfo::isVirtualRegister(mop.getReg())) {
2828 unsigned reg = mop.getReg();
2829 // Multiple uses of reg by the same instruction. It should not
2830 // contribute to spill weight again.
2831 if (UniqueUses.count(reg) != 0)
2833 LiveInterval &RegInt = li_->getInterval(reg);
2835 li_->getSpillWeight(mop.isDef(), mop.isUse(), loopDepth);
2836 UniqueUses.insert(reg);
2844 for (LiveIntervals::iterator I = li_->begin(), E = li_->end(); I != E; ++I) {
2845 LiveInterval &LI = *I->second;
2846 if (TargetRegisterInfo::isVirtualRegister(LI.reg)) {
2847 // If the live interval length is essentially zero, i.e. in every live
2848 // range the use follows def immediately, it doesn't make sense to spill
2849 // it and hope it will be easier to allocate for this li.
2850 if (isZeroLengthInterval(&LI))
2851 LI.weight = HUGE_VALF;
2853 bool isLoad = false;
2854 SmallVector<LiveInterval*, 4> SpillIs;
2855 if (li_->isReMaterializable(LI, SpillIs, isLoad)) {
2856 // If all of the definitions of the interval are re-materializable,
2857 // it is a preferred candidate for spilling. If non of the defs are
2858 // loads, then it's potentially very cheap to re-materialize.
2859 // FIXME: this gets much more complicated once we support non-trivial
2860 // re-materialization.
2868 // Slightly prefer live interval that has been assigned a preferred reg.
2869 std::pair<unsigned, unsigned> Hint = mri_->getRegAllocationHint(LI.reg);
2870 if (Hint.first || Hint.second)
2873 // Divide the weight of the interval by its size. This encourages
2874 // spilling of intervals that are large and have few uses, and
2875 // discourages spilling of small intervals with many uses.
2876 LI.weight /= li_->getApproximateInstructionCount(LI) * InstrSlots::NUM;
2884 /// print - Implement the dump method.
2885 void SimpleRegisterCoalescing::print(std::ostream &O, const Module* m) const {
2889 RegisterCoalescer* llvm::createSimpleRegisterCoalescer() {
2890 return new SimpleRegisterCoalescing();
2893 // Make sure that anything that uses RegisterCoalescer pulls in this file...
2894 DEFINING_FILE_FOR(SimpleRegisterCoalescing)