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/LiveVariables.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineInstr.h"
23 #include "llvm/CodeGen/MachineLoopInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/Passes.h"
26 #include "llvm/CodeGen/RegisterCoalescer.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/ADT/STLExtras.h"
38 STATISTIC(numJoins , "Number of interval joins performed");
39 STATISTIC(numCommutes , "Number of instruction commuting performed");
40 STATISTIC(numExtends , "Number of copies extended");
41 STATISTIC(numPeep , "Number of identity moves eliminated after coalescing");
42 STATISTIC(numAborts , "Number of times interval joining aborted");
44 char SimpleRegisterCoalescing::ID = 0;
47 EnableJoining("join-liveintervals",
48 cl::desc("Coalesce copies (default=true)"),
52 NewHeuristic("new-coalescer-heuristic",
53 cl::desc("Use new coalescer heuristic"),
57 CommuteDef("coalescer-commute-instrs",
58 cl::init(false), cl::Hidden);
61 CommuteLimit("commute-limit",
62 cl::init(-1), cl::Hidden);
64 RegisterPass<SimpleRegisterCoalescing>
65 X("simple-register-coalescing", "Simple Register Coalescing");
67 // Declare that we implement the RegisterCoalescer interface
68 RegisterAnalysisGroup<RegisterCoalescer, true/*The Default*/> V(X);
71 const PassInfo *llvm::SimpleRegisterCoalescingID = X.getPassInfo();
73 void SimpleRegisterCoalescing::getAnalysisUsage(AnalysisUsage &AU) const {
74 AU.addPreserved<LiveIntervals>();
75 AU.addPreserved<MachineLoopInfo>();
76 AU.addPreservedID(MachineDominatorsID);
77 AU.addPreservedID(PHIEliminationID);
78 AU.addPreservedID(TwoAddressInstructionPassID);
79 AU.addRequired<LiveVariables>();
80 AU.addRequired<LiveIntervals>();
81 AU.addRequired<MachineLoopInfo>();
82 MachineFunctionPass::getAnalysisUsage(AU);
85 /// AdjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA
86 /// being the source and IntB being the dest, thus this defines a value number
87 /// in IntB. If the source value number (in IntA) is defined by a copy from B,
88 /// see if we can merge these two pieces of B into a single value number,
89 /// eliminating a copy. For example:
93 /// B1 = A3 <- this copy
95 /// In this case, B0 can be extended to where the B1 copy lives, allowing the B1
96 /// value number to be replaced with B0 (which simplifies the B liveinterval).
98 /// This returns true if an interval was modified.
100 bool SimpleRegisterCoalescing::AdjustCopiesBackFrom(LiveInterval &IntA,
102 MachineInstr *CopyMI) {
103 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
105 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
106 // the example above.
107 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
108 VNInfo *BValNo = BLR->valno;
110 // Get the location that B is defined at. Two options: either this value has
111 // an unknown definition point or it is defined at CopyIdx. If unknown, we
113 if (!BValNo->copy) return false;
114 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
116 // AValNo is the value number in A that defines the copy, A3 in the example.
117 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyIdx-1);
118 VNInfo *AValNo = ALR->valno;
120 // If AValNo is defined as a copy from IntB, we can potentially process this.
121 // Get the instruction that defines this value number.
122 unsigned SrcReg = li_->getVNInfoSourceReg(AValNo);
123 if (!SrcReg) return false; // Not defined by a copy.
125 // If the value number is not defined by a copy instruction, ignore it.
127 // If the source register comes from an interval other than IntB, we can't
129 if (SrcReg != IntB.reg) return false;
131 // Get the LiveRange in IntB that this value number starts with.
132 LiveInterval::iterator ValLR = IntB.FindLiveRangeContaining(AValNo->def-1);
134 // Make sure that the end of the live range is inside the same block as
136 MachineInstr *ValLREndInst = li_->getInstructionFromIndex(ValLR->end-1);
138 ValLREndInst->getParent() != CopyMI->getParent()) return false;
140 // Okay, we now know that ValLR ends in the same block that the CopyMI
141 // live-range starts. If there are no intervening live ranges between them in
142 // IntB, we can merge them.
143 if (ValLR+1 != BLR) return false;
145 // If a live interval is a physical register, conservatively check if any
146 // of its sub-registers is overlapping the live interval of the virtual
147 // register. If so, do not coalesce.
148 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg) &&
149 *tri_->getSubRegisters(IntB.reg)) {
150 for (const unsigned* SR = tri_->getSubRegisters(IntB.reg); *SR; ++SR)
151 if (li_->hasInterval(*SR) && IntA.overlaps(li_->getInterval(*SR))) {
152 DOUT << "Interfere with sub-register ";
153 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
158 DOUT << "\nExtending: "; IntB.print(DOUT, tri_);
160 unsigned FillerStart = ValLR->end, FillerEnd = BLR->start;
161 // We are about to delete CopyMI, so need to remove it as the 'instruction
162 // that defines this value #'. Update the the valnum with the new defining
164 BValNo->def = FillerStart;
167 // Okay, we can merge them. We need to insert a new liverange:
168 // [ValLR.end, BLR.begin) of either value number, then we merge the
169 // two value numbers.
170 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
172 // If the IntB live range is assigned to a physical register, and if that
173 // physreg has aliases,
174 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
175 // Update the liveintervals of sub-registers.
176 for (const unsigned *AS = tri_->getSubRegisters(IntB.reg); *AS; ++AS) {
177 LiveInterval &AliasLI = li_->getInterval(*AS);
178 AliasLI.addRange(LiveRange(FillerStart, FillerEnd,
179 AliasLI.getNextValue(FillerStart, 0, li_->getVNInfoAllocator())));
183 // Okay, merge "B1" into the same value number as "B0".
184 if (BValNo != ValLR->valno)
185 IntB.MergeValueNumberInto(BValNo, ValLR->valno);
186 DOUT << " result = "; IntB.print(DOUT, tri_);
189 // If the source instruction was killing the source register before the
190 // merge, unset the isKill marker given the live range has been extended.
191 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true);
193 ValLREndInst->getOperand(UIdx).setIsKill(false);
199 /// HasOtherReachingDefs - Return true if there are definitions of IntB
200 /// other than BValNo val# that can reach uses of AValno val# of IntA.
201 bool SimpleRegisterCoalescing::HasOtherReachingDefs(LiveInterval &IntA,
205 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
207 if (AI->valno != AValNo) continue;
208 LiveInterval::Ranges::iterator BI =
209 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start);
210 if (BI != IntB.ranges.begin())
212 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) {
213 if (BI->valno == BValNo)
215 if (BI->start <= AI->start && BI->end > AI->start)
217 if (BI->start > AI->start && BI->start < AI->end)
224 /// RemoveCopyByCommutingDef - We found a non-trivially-coalescable copy with IntA
225 /// being the source and IntB being the dest, thus this defines a value number
226 /// in IntB. If the source value number (in IntA) is defined by a commutable
227 /// instruction and its other operand is coalesced to the copy dest register,
228 /// see if we can transform the copy into a noop by commuting the definition. For
231 /// A3 = op A2 B0<kill>
233 /// B1 = A3 <- this copy
235 /// = op A3 <- more uses
239 /// B2 = op B0 A2<kill>
241 /// B1 = B2 <- now an identify copy
243 /// = op B2 <- more uses
245 /// This returns true if an interval was modified.
247 bool SimpleRegisterCoalescing::RemoveCopyByCommutingDef(LiveInterval &IntA,
249 MachineInstr *CopyMI) {
250 if (!CommuteDef) return false;
252 unsigned CopyIdx = li_->getDefIndex(li_->getInstructionIndex(CopyMI));
254 // FIXME: For now, only eliminate the copy by commuting its def when the
255 // source register is a virtual register. We want to guard against cases
256 // where the copy is a back edge copy and commuting the def lengthen the
257 // live interval of the source register to the entire loop.
258 if (TargetRegisterInfo::isPhysicalRegister(IntA.reg))
261 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
262 // the example above.
263 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
264 VNInfo *BValNo = BLR->valno;
266 // Get the location that B is defined at. Two options: either this value has
267 // an unknown definition point or it is defined at CopyIdx. If unknown, we
269 if (!BValNo->copy) return false;
270 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
272 // AValNo is the value number in A that defines the copy, A3 in the example.
273 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyIdx-1);
274 VNInfo *AValNo = ALR->valno;
275 // If other defs can reach uses of this def, then it's not safe to perform
277 if (AValNo->def == ~0U || AValNo->def == ~1U || AValNo->hasPHIKill)
279 MachineInstr *DefMI = li_->getInstructionFromIndex(AValNo->def);
280 const TargetInstrDesc &TID = DefMI->getDesc();
282 if (!TID.isCommutable() ||
283 !tii_->CommuteChangesDestination(DefMI, NewDstIdx))
286 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
287 unsigned NewReg = NewDstMO.getReg();
288 if (NewReg != IntB.reg || !NewDstMO.isKill())
291 // Make sure there are no other definitions of IntB that would reach the
292 // uses which the new definition can reach.
293 if (HasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
296 if (CommuteLimit >= 0 && numCommutes >= (unsigned)CommuteLimit)
299 // At this point we have decided that it is legal to do this
300 // transformation. Start by commuting the instruction.
301 MachineBasicBlock *MBB = DefMI->getParent();
302 MachineInstr *NewMI = tii_->commuteInstruction(DefMI);
305 if (NewMI != DefMI) {
306 li_->ReplaceMachineInstrInMaps(DefMI, NewMI);
307 MBB->insert(DefMI, NewMI);
310 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg);
311 NewMI->getOperand(OpIdx).setIsKill();
313 // Update uses of IntA of the specific Val# with IntB.
314 bool BHasPHIKill = BValNo->hasPHIKill;
315 SmallVector<VNInfo*, 4> BDeadValNos;
316 SmallVector<unsigned, 4> BKills;
317 std::map<unsigned, unsigned> BExtend;
318 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(IntA.reg),
319 UE = mri_->use_end(); UI != UE;) {
320 MachineOperand &UseMO = UI.getOperand();
321 MachineInstr *UseMI = &*UI;
323 if (JoinedCopies.count(UseMI))
325 unsigned UseIdx = li_->getInstructionIndex(UseMI);
326 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
327 if (ULR->valno != AValNo)
329 UseMO.setReg(NewReg);
333 BKills.push_back(li_->getUseIndex(UseIdx)+1);
334 unsigned SrcReg, DstReg;
335 if (!tii_->isMoveInstr(*UseMI, SrcReg, DstReg))
337 if (DstReg == IntB.reg) {
338 // This copy will become a noop. If it's defining a new val#,
339 // remove that val# as well. However this live range is being
340 // extended to the end of the existing live range defined by the copy.
341 unsigned DefIdx = li_->getDefIndex(UseIdx);
342 LiveInterval::iterator DLR = IntB.FindLiveRangeContaining(DefIdx);
343 BHasPHIKill |= DLR->valno->hasPHIKill;
344 assert(DLR->valno->def == DefIdx);
345 BDeadValNos.push_back(DLR->valno);
346 BExtend[DLR->start] = DLR->end;
347 JoinedCopies.insert(UseMI);
348 // If this is a kill but it's going to be removed, the last use
349 // of the same val# is the new kill.
350 if (UseMO.isKill()) {
356 // We need to insert a new liverange: [ALR.start, LastUse). It may be we can
357 // simply extend BLR if CopyMI doesn't end the range.
358 DOUT << "\nExtending: "; IntB.print(DOUT, tri_);
360 IntB.removeValNo(BValNo);
361 for (unsigned i = 0, e = BDeadValNos.size(); i != e; ++i)
362 IntB.removeValNo(BDeadValNos[i]);
363 VNInfo *ValNo = IntB.getNextValue(ALR->start, 0, li_->getVNInfoAllocator());
364 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
366 if (AI->valno != AValNo) continue;
367 unsigned End = AI->end;
368 std::map<unsigned, unsigned>::iterator EI = BExtend.find(End);
369 if (EI != BExtend.end())
371 IntB.addRange(LiveRange(AI->start, End, ValNo));
373 IntB.addKills(ValNo, BKills);
374 ValNo->hasPHIKill = BHasPHIKill;
376 DOUT << " result = "; IntB.print(DOUT, tri_);
379 DOUT << "\nShortening: "; IntA.print(DOUT, tri_);
380 IntA.removeValNo(AValNo);
381 DOUT << " result = "; IntA.print(DOUT, tri_);
388 /// RemoveUnnecessaryKills - Remove kill markers that are no longer accurate
389 /// due to live range lengthening as the result of coalescing.
390 void SimpleRegisterCoalescing::RemoveUnnecessaryKills(unsigned Reg,
392 for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(Reg),
393 UE = mri_->use_end(); UI != UE; ++UI) {
394 MachineOperand &UseMO = UI.getOperand();
395 if (UseMO.isKill()) {
396 MachineInstr *UseMI = UseMO.getParent();
397 unsigned UseIdx = li_->getUseIndex(li_->getInstructionIndex(UseMI));
398 if (JoinedCopies.count(UseMI))
400 LiveInterval::const_iterator UI = LI.FindLiveRangeContaining(UseIdx);
401 assert(UI != LI.end());
402 if (!LI.isKill(UI->valno, UseIdx+1))
403 UseMO.setIsKill(false);
408 /// isBackEdgeCopy - Returns true if CopyMI is a back edge copy.
410 bool SimpleRegisterCoalescing::isBackEdgeCopy(MachineInstr *CopyMI,
412 MachineBasicBlock *MBB = CopyMI->getParent();
413 const MachineLoop *L = loopInfo->getLoopFor(MBB);
416 if (MBB != L->getLoopLatch())
419 LiveInterval &LI = li_->getInterval(DstReg);
420 unsigned DefIdx = li_->getInstructionIndex(CopyMI);
421 LiveInterval::const_iterator DstLR =
422 LI.FindLiveRangeContaining(li_->getDefIndex(DefIdx));
423 if (DstLR == LI.end())
425 unsigned KillIdx = li_->getInstructionIndex(&MBB->back()) + InstrSlots::NUM;
426 if (DstLR->valno->kills.size() == 1 &&
427 DstLR->valno->kills[0] == KillIdx && DstLR->valno->hasPHIKill)
432 /// UpdateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
433 /// update the subregister number if it is not zero. If DstReg is a
434 /// physical register and the existing subregister number of the def / use
435 /// being updated is not zero, make sure to set it to the correct physical
438 SimpleRegisterCoalescing::UpdateRegDefsUses(unsigned SrcReg, unsigned DstReg,
440 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
441 if (DstIsPhys && SubIdx) {
442 // Figure out the real physical register we are updating with.
443 DstReg = tri_->getSubReg(DstReg, SubIdx);
447 for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(SrcReg),
448 E = mri_->reg_end(); I != E; ) {
449 MachineOperand &O = I.getOperand();
452 unsigned UseSubIdx = O.getSubReg();
453 unsigned UseDstReg = DstReg;
455 UseDstReg = tri_->getSubReg(DstReg, UseSubIdx);
459 unsigned OldSubIdx = O.getSubReg();
460 // Sub-register indexes goes from small to large. e.g.
461 // RAX: 0 -> AL, 1 -> AH, 2 -> AX, 3 -> EAX
462 // EAX: 0 -> AL, 1 -> AH, 2 -> AX
463 // So RAX's sub-register 2 is AX, RAX's sub-regsiter 3 is EAX, whose
464 // sub-register 2 is also AX.
465 if (SubIdx && OldSubIdx && SubIdx != OldSubIdx)
466 assert(OldSubIdx < SubIdx && "Conflicting sub-register index!");
474 /// JoinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
475 /// which are the src/dst of the copy instruction CopyMI. This returns true
476 /// if the copy was successfully coalesced away. If it is not currently
477 /// possible to coalesce this interval, but it may be possible if other
478 /// things get coalesced, then it returns true by reference in 'Again'.
479 bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) {
480 MachineInstr *CopyMI = TheCopy.MI;
483 if (JoinedCopies.count(CopyMI))
484 return false; // Already done.
486 DOUT << li_->getInstructionIndex(CopyMI) << '\t' << *CopyMI;
490 bool isExtSubReg = CopyMI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG;
493 DstReg = CopyMI->getOperand(0).getReg();
494 SrcReg = CopyMI->getOperand(1).getReg();
495 } else if (!tii_->isMoveInstr(*CopyMI, SrcReg, DstReg)) {
496 assert(0 && "Unrecognized copy instruction!");
500 // If they are already joined we continue.
501 if (SrcReg == DstReg) {
502 DOUT << "\tCopy already coalesced.\n";
503 return false; // Not coalescable.
506 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
507 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
509 // If they are both physical registers, we cannot join them.
510 if (SrcIsPhys && DstIsPhys) {
511 DOUT << "\tCan not coalesce physregs.\n";
512 return false; // Not coalescable.
515 // We only join virtual registers with allocatable physical registers.
516 if (SrcIsPhys && !allocatableRegs_[SrcReg]) {
517 DOUT << "\tSrc reg is unallocatable physreg.\n";
518 return false; // Not coalescable.
520 if (DstIsPhys && !allocatableRegs_[DstReg]) {
521 DOUT << "\tDst reg is unallocatable physreg.\n";
522 return false; // Not coalescable.
525 unsigned RealDstReg = 0;
527 SubIdx = CopyMI->getOperand(2).getImm();
529 // r1024 = EXTRACT_SUBREG EAX, 0 then r1024 is really going to be
530 // coalesced with AX.
531 SrcReg = tri_->getSubReg(SrcReg, SubIdx);
533 } else if (DstIsPhys) {
534 // If this is a extract_subreg where dst is a physical register, e.g.
535 // cl = EXTRACT_SUBREG reg1024, 1
536 // then create and update the actual physical register allocated to RHS.
537 const TargetRegisterClass *RC = mri_->getRegClass(SrcReg);
538 for (const unsigned *SRs = tri_->getSuperRegisters(DstReg);
539 unsigned SR = *SRs; ++SRs) {
540 if (DstReg == tri_->getSubReg(SR, SubIdx) &&
546 assert(RealDstReg && "Invalid extra_subreg instruction!");
548 // For this type of EXTRACT_SUBREG, conservatively
549 // check if the live interval of the source register interfere with the
550 // actual super physical register we are trying to coalesce with.
551 LiveInterval &RHS = li_->getInterval(SrcReg);
552 if (li_->hasInterval(RealDstReg) &&
553 RHS.overlaps(li_->getInterval(RealDstReg))) {
554 DOUT << "Interfere with register ";
555 DEBUG(li_->getInterval(RealDstReg).print(DOUT, tri_));
556 return false; // Not coalescable
558 for (const unsigned* SR = tri_->getSubRegisters(RealDstReg); *SR; ++SR)
559 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
560 DOUT << "Interfere with sub-register ";
561 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
562 return false; // Not coalescable
566 unsigned SrcSize= li_->getInterval(SrcReg).getSize() / InstrSlots::NUM;
567 unsigned DstSize= li_->getInterval(DstReg).getSize() / InstrSlots::NUM;
568 const TargetRegisterClass *RC = mri_->getRegClass(DstReg);
569 unsigned Threshold = allocatableRCRegs_[RC].count();
570 // Be conservative. If both sides are virtual registers, do not coalesce
571 // if this will cause a high use density interval to target a smaller set
573 if (DstSize > Threshold || SrcSize > Threshold) {
574 LiveVariables::VarInfo &svi = lv_->getVarInfo(SrcReg);
575 LiveVariables::VarInfo &dvi = lv_->getVarInfo(DstReg);
576 if ((float)dvi.NumUses / DstSize < (float)svi.NumUses / SrcSize) {
577 Again = true; // May be possible to coalesce later.
582 } else if (differingRegisterClasses(SrcReg, DstReg)) {
583 // FIXME: What if the resul of a EXTRACT_SUBREG is then coalesced
584 // with another? If it's the resulting destination register, then
585 // the subidx must be propagated to uses (but only those defined
586 // by the EXTRACT_SUBREG). If it's being coalesced into another
587 // register, it should be safe because register is assumed to have
588 // the register class of the super-register.
590 // If they are not of the same register class, we cannot join them.
591 DOUT << "\tSrc/Dest are different register classes.\n";
592 // Allow the coalescer to try again in case either side gets coalesced to
593 // a physical register that's compatible with the other side. e.g.
594 // r1024 = MOV32to32_ r1025
595 // but later r1024 is assigned EAX then r1025 may be coalesced with EAX.
596 Again = true; // May be possible to coalesce later.
600 LiveInterval &SrcInt = li_->getInterval(SrcReg);
601 LiveInterval &DstInt = li_->getInterval(DstReg);
602 assert(SrcInt.reg == SrcReg && DstInt.reg == DstReg &&
603 "Register mapping is horribly broken!");
605 DOUT << "\t\tInspecting "; SrcInt.print(DOUT, tri_);
606 DOUT << " and "; DstInt.print(DOUT, tri_);
609 // Check if it is necessary to propagate "isDead" property before intervals
611 MachineOperand *mopd = CopyMI->findRegisterDefOperand(DstReg);
612 bool isDead = mopd->isDead();
613 bool isShorten = false;
614 unsigned SrcStart = 0, RemoveStart = 0;
615 unsigned SrcEnd = 0, RemoveEnd = 0;
617 unsigned CopyIdx = li_->getInstructionIndex(CopyMI);
618 LiveInterval::iterator SrcLR =
619 SrcInt.FindLiveRangeContaining(li_->getUseIndex(CopyIdx));
620 RemoveStart = SrcStart = SrcLR->start;
621 RemoveEnd = SrcEnd = SrcLR->end;
622 // The instruction which defines the src is only truly dead if there are
623 // no intermediate uses and there isn't a use beyond the copy.
624 // FIXME: find the last use, mark is kill and shorten the live range.
625 if (SrcEnd > li_->getDefIndex(CopyIdx)) {
629 MachineOperand *LastUse =
630 lastRegisterUse(SrcStart, CopyIdx, SrcReg, LastUseIdx);
632 // Shorten the liveinterval to the end of last use.
633 LastUse->setIsKill();
636 RemoveStart = li_->getDefIndex(LastUseIdx);
639 MachineInstr *SrcMI = li_->getInstructionFromIndex(SrcStart);
641 MachineOperand *mops = findDefOperand(SrcMI, SrcReg);
643 // A dead def should have a single cycle interval.
650 // We need to be careful about coalescing a source physical register with a
651 // virtual register. Once the coalescing is done, it cannot be broken and
652 // these are not spillable! If the destination interval uses are far away,
653 // think twice about coalescing them!
654 if (!mopd->isDead() && (SrcIsPhys || DstIsPhys) && !isExtSubReg) {
655 LiveInterval &JoinVInt = SrcIsPhys ? DstInt : SrcInt;
656 unsigned JoinVReg = SrcIsPhys ? DstReg : SrcReg;
657 unsigned JoinPReg = SrcIsPhys ? SrcReg : DstReg;
658 const TargetRegisterClass *RC = mri_->getRegClass(JoinVReg);
659 unsigned Threshold = allocatableRCRegs_[RC].count() * 2;
660 if (TheCopy.isBackEdge)
661 Threshold *= 2; // Favors back edge copies.
663 // If the virtual register live interval is long but it has low use desity,
664 // do not join them, instead mark the physical register as its allocation
666 unsigned Length = JoinVInt.getSize() / InstrSlots::NUM;
667 LiveVariables::VarInfo &vi = lv_->getVarInfo(JoinVReg);
668 if (Length > Threshold &&
669 (((float)vi.NumUses / Length) < (1.0 / Threshold))) {
670 JoinVInt.preference = JoinPReg;
672 DOUT << "\tMay tie down a physical register, abort!\n";
673 Again = true; // May be possible to coalesce later.
678 // Okay, attempt to join these two intervals. On failure, this returns false.
679 // Otherwise, if one of the intervals being joined is a physreg, this method
680 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
681 // been modified, so we can use this information below to update aliases.
682 bool Swapped = false;
683 if (JoinIntervals(DstInt, SrcInt, Swapped)) {
685 // Result of the copy is dead. Propagate this property.
687 assert(TargetRegisterInfo::isPhysicalRegister(SrcReg) &&
688 "Live-in must be a physical register!");
689 // Live-in to the function but dead. Remove it from entry live-in set.
690 // JoinIntervals may end up swapping the two intervals.
691 mf_->begin()->removeLiveIn(SrcReg);
693 MachineInstr *SrcMI = li_->getInstructionFromIndex(SrcStart);
695 MachineOperand *mops = findDefOperand(SrcMI, SrcReg);
702 if (isShorten || isDead) {
703 // Shorten the destination live interval.
705 SrcInt.removeRange(RemoveStart, RemoveEnd, true);
708 // Coalescing failed.
710 // If we can eliminate the copy without merging the live ranges, do so now.
712 (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI) ||
713 RemoveCopyByCommutingDef(SrcInt, DstInt, CopyMI))) {
714 JoinedCopies.insert(CopyMI);
718 // Otherwise, we are unable to join the intervals.
719 DOUT << "Interference!\n";
720 Again = true; // May be possible to coalesce later.
724 LiveInterval *ResSrcInt = &SrcInt;
725 LiveInterval *ResDstInt = &DstInt;
727 std::swap(SrcReg, DstReg);
728 std::swap(ResSrcInt, ResDstInt);
730 assert(TargetRegisterInfo::isVirtualRegister(SrcReg) &&
731 "LiveInterval::join didn't work right!");
733 // If we're about to merge live ranges into a physical register live range,
734 // we have to update any aliased register's live ranges to indicate that they
735 // have clobbered values for this range.
736 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
737 // Unset unnecessary kills.
738 if (!ResDstInt->containsOneValue()) {
739 for (LiveInterval::Ranges::const_iterator I = ResSrcInt->begin(),
740 E = ResSrcInt->end(); I != E; ++I)
741 unsetRegisterKills(I->start, I->end, DstReg);
744 // If this is a extract_subreg where dst is a physical register, e.g.
745 // cl = EXTRACT_SUBREG reg1024, 1
746 // then create and update the actual physical register allocated to RHS.
748 LiveInterval &RealDstInt = li_->getOrCreateInterval(RealDstReg);
749 SmallSet<const VNInfo*, 4> CopiedValNos;
750 for (LiveInterval::Ranges::const_iterator I = ResSrcInt->ranges.begin(),
751 E = ResSrcInt->ranges.end(); I != E; ++I) {
752 LiveInterval::const_iterator DstLR =
753 ResDstInt->FindLiveRangeContaining(I->start);
754 assert(DstLR != ResDstInt->end() && "Invalid joined interval!");
755 const VNInfo *DstValNo = DstLR->valno;
756 if (CopiedValNos.insert(DstValNo)) {
757 VNInfo *ValNo = RealDstInt.getNextValue(DstValNo->def, DstValNo->copy,
758 li_->getVNInfoAllocator());
759 ValNo->hasPHIKill = DstValNo->hasPHIKill;
760 RealDstInt.addKills(ValNo, DstValNo->kills);
761 RealDstInt.MergeValueInAsValue(*ResDstInt, DstValNo, ValNo);
767 // Update the liveintervals of sub-registers.
768 for (const unsigned *AS = tri_->getSubRegisters(DstReg); *AS; ++AS)
769 li_->getOrCreateInterval(*AS).MergeInClobberRanges(*ResSrcInt,
770 li_->getVNInfoAllocator());
772 // Merge use info if the destination is a virtual register.
773 LiveVariables::VarInfo& dVI = lv_->getVarInfo(DstReg);
774 LiveVariables::VarInfo& sVI = lv_->getVarInfo(SrcReg);
775 dVI.NumUses += sVI.NumUses;
778 // If this is a EXTRACT_SUBREG, make sure the result of coalescing is the
779 // larger super-register.
780 if (isExtSubReg && !SrcIsPhys && !DstIsPhys) {
782 ResSrcInt->Copy(*ResDstInt, li_->getVNInfoAllocator());
783 std::swap(SrcReg, DstReg);
784 std::swap(ResSrcInt, ResDstInt);
789 // Add all copies that define val# in the source interval into the queue.
790 for (LiveInterval::const_vni_iterator i = ResSrcInt->vni_begin(),
791 e = ResSrcInt->vni_end(); i != e; ++i) {
792 const VNInfo *vni = *i;
793 if (!vni->def || vni->def == ~1U || vni->def == ~0U)
795 MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
796 unsigned NewSrcReg, NewDstReg;
798 JoinedCopies.count(CopyMI) == 0 &&
799 tii_->isMoveInstr(*CopyMI, NewSrcReg, NewDstReg)) {
800 unsigned LoopDepth = loopInfo->getLoopDepth(CopyMI->getParent());
801 JoinQueue->push(CopyRec(CopyMI, LoopDepth,
802 isBackEdgeCopy(CopyMI, DstReg)));
807 DOUT << "\n\t\tJoined. Result = "; ResDstInt->print(DOUT, tri_);
810 // Remember to delete the copy instruction.
811 JoinedCopies.insert(CopyMI);
813 // Some live range has been lengthened due to colaescing, eliminate the
814 // unnecessary kills.
815 RemoveUnnecessaryKills(SrcReg, *ResDstInt);
816 if (TargetRegisterInfo::isVirtualRegister(DstReg))
817 RemoveUnnecessaryKills(DstReg, *ResDstInt);
819 // SrcReg is guarateed to be the register whose live interval that is
821 li_->removeInterval(SrcReg);
822 UpdateRegDefsUses(SrcReg, DstReg, SubIdx);
828 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
829 /// compute what the resultant value numbers for each value in the input two
830 /// ranges will be. This is complicated by copies between the two which can
831 /// and will commonly cause multiple value numbers to be merged into one.
833 /// VN is the value number that we're trying to resolve. InstDefiningValue
834 /// keeps track of the new InstDefiningValue assignment for the result
835 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
836 /// whether a value in this or other is a copy from the opposite set.
837 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
838 /// already been assigned.
840 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
841 /// contains the value number the copy is from.
843 static unsigned ComputeUltimateVN(VNInfo *VNI,
844 SmallVector<VNInfo*, 16> &NewVNInfo,
845 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
846 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
847 SmallVector<int, 16> &ThisValNoAssignments,
848 SmallVector<int, 16> &OtherValNoAssignments) {
849 unsigned VN = VNI->id;
851 // If the VN has already been computed, just return it.
852 if (ThisValNoAssignments[VN] >= 0)
853 return ThisValNoAssignments[VN];
854 // assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?");
856 // If this val is not a copy from the other val, then it must be a new value
857 // number in the destination.
858 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
859 if (I == ThisFromOther.end()) {
860 NewVNInfo.push_back(VNI);
861 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
863 VNInfo *OtherValNo = I->second;
865 // Otherwise, this *is* a copy from the RHS. If the other side has already
866 // been computed, return it.
867 if (OtherValNoAssignments[OtherValNo->id] >= 0)
868 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
870 // Mark this value number as currently being computed, then ask what the
871 // ultimate value # of the other value is.
872 ThisValNoAssignments[VN] = -2;
873 unsigned UltimateVN =
874 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
875 OtherValNoAssignments, ThisValNoAssignments);
876 return ThisValNoAssignments[VN] = UltimateVN;
879 static bool InVector(VNInfo *Val, const SmallVector<VNInfo*, 8> &V) {
880 return std::find(V.begin(), V.end(), Val) != V.end();
883 /// SimpleJoin - Attempt to joint the specified interval into this one. The
884 /// caller of this method must guarantee that the RHS only contains a single
885 /// value number and that the RHS is not defined by a copy from this
886 /// interval. This returns false if the intervals are not joinable, or it
887 /// joins them and returns true.
888 bool SimpleRegisterCoalescing::SimpleJoin(LiveInterval &LHS, LiveInterval &RHS){
889 assert(RHS.containsOneValue());
891 // Some number (potentially more than one) value numbers in the current
892 // interval may be defined as copies from the RHS. Scan the overlapping
893 // portions of the LHS and RHS, keeping track of this and looking for
894 // overlapping live ranges that are NOT defined as copies. If these exist, we
897 LiveInterval::iterator LHSIt = LHS.begin(), LHSEnd = LHS.end();
898 LiveInterval::iterator RHSIt = RHS.begin(), RHSEnd = RHS.end();
900 if (LHSIt->start < RHSIt->start) {
901 LHSIt = std::upper_bound(LHSIt, LHSEnd, RHSIt->start);
902 if (LHSIt != LHS.begin()) --LHSIt;
903 } else if (RHSIt->start < LHSIt->start) {
904 RHSIt = std::upper_bound(RHSIt, RHSEnd, LHSIt->start);
905 if (RHSIt != RHS.begin()) --RHSIt;
908 SmallVector<VNInfo*, 8> EliminatedLHSVals;
911 // Determine if these live intervals overlap.
912 bool Overlaps = false;
913 if (LHSIt->start <= RHSIt->start)
914 Overlaps = LHSIt->end > RHSIt->start;
916 Overlaps = RHSIt->end > LHSIt->start;
918 // If the live intervals overlap, there are two interesting cases: if the
919 // LHS interval is defined by a copy from the RHS, it's ok and we record
920 // that the LHS value # is the same as the RHS. If it's not, then we cannot
921 // coalesce these live ranges and we bail out.
923 // If we haven't already recorded that this value # is safe, check it.
924 if (!InVector(LHSIt->valno, EliminatedLHSVals)) {
925 // Copy from the RHS?
926 unsigned SrcReg = li_->getVNInfoSourceReg(LHSIt->valno);
927 if (SrcReg != RHS.reg)
928 return false; // Nope, bail out.
930 EliminatedLHSVals.push_back(LHSIt->valno);
933 // We know this entire LHS live range is okay, so skip it now.
934 if (++LHSIt == LHSEnd) break;
938 if (LHSIt->end < RHSIt->end) {
939 if (++LHSIt == LHSEnd) break;
941 // One interesting case to check here. It's possible that we have
942 // something like "X3 = Y" which defines a new value number in the LHS,
943 // and is the last use of this liverange of the RHS. In this case, we
944 // want to notice this copy (so that it gets coalesced away) even though
945 // the live ranges don't actually overlap.
946 if (LHSIt->start == RHSIt->end) {
947 if (InVector(LHSIt->valno, EliminatedLHSVals)) {
948 // We already know that this value number is going to be merged in
949 // if coalescing succeeds. Just skip the liverange.
950 if (++LHSIt == LHSEnd) break;
952 // Otherwise, if this is a copy from the RHS, mark it as being merged
954 if (li_->getVNInfoSourceReg(LHSIt->valno) == RHS.reg) {
955 EliminatedLHSVals.push_back(LHSIt->valno);
957 // We know this entire LHS live range is okay, so skip it now.
958 if (++LHSIt == LHSEnd) break;
963 if (++RHSIt == RHSEnd) break;
967 // If we got here, we know that the coalescing will be successful and that
968 // the value numbers in EliminatedLHSVals will all be merged together. Since
969 // the most common case is that EliminatedLHSVals has a single number, we
970 // optimize for it: if there is more than one value, we merge them all into
971 // the lowest numbered one, then handle the interval as if we were merging
972 // with one value number.
974 if (EliminatedLHSVals.size() > 1) {
975 // Loop through all the equal value numbers merging them into the smallest
977 VNInfo *Smallest = EliminatedLHSVals[0];
978 for (unsigned i = 1, e = EliminatedLHSVals.size(); i != e; ++i) {
979 if (EliminatedLHSVals[i]->id < Smallest->id) {
980 // Merge the current notion of the smallest into the smaller one.
981 LHS.MergeValueNumberInto(Smallest, EliminatedLHSVals[i]);
982 Smallest = EliminatedLHSVals[i];
984 // Merge into the smallest.
985 LHS.MergeValueNumberInto(EliminatedLHSVals[i], Smallest);
990 assert(!EliminatedLHSVals.empty() && "No copies from the RHS?");
991 LHSValNo = EliminatedLHSVals[0];
994 // Okay, now that there is a single LHS value number that we're merging the
995 // RHS into, update the value number info for the LHS to indicate that the
996 // value number is defined where the RHS value number was.
997 const VNInfo *VNI = RHS.getValNumInfo(0);
998 LHSValNo->def = VNI->def;
999 LHSValNo->copy = VNI->copy;
1001 // Okay, the final step is to loop over the RHS live intervals, adding them to
1003 LHSValNo->hasPHIKill |= VNI->hasPHIKill;
1004 LHS.addKills(LHSValNo, VNI->kills);
1005 LHS.MergeRangesInAsValue(RHS, LHSValNo);
1006 LHS.weight += RHS.weight;
1007 if (RHS.preference && !LHS.preference)
1008 LHS.preference = RHS.preference;
1013 /// JoinIntervals - Attempt to join these two intervals. On failure, this
1014 /// returns false. Otherwise, if one of the intervals being joined is a
1015 /// physreg, this method always canonicalizes LHS to be it. The output
1016 /// "RHS" will not have been modified, so we can use this information
1017 /// below to update aliases.
1018 bool SimpleRegisterCoalescing::JoinIntervals(LiveInterval &LHS,
1019 LiveInterval &RHS, bool &Swapped) {
1020 // Compute the final value assignment, assuming that the live ranges can be
1022 SmallVector<int, 16> LHSValNoAssignments;
1023 SmallVector<int, 16> RHSValNoAssignments;
1024 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
1025 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
1026 SmallVector<VNInfo*, 16> NewVNInfo;
1028 // If a live interval is a physical register, conservatively check if any
1029 // of its sub-registers is overlapping the live interval of the virtual
1030 // register. If so, do not coalesce.
1031 if (TargetRegisterInfo::isPhysicalRegister(LHS.reg) &&
1032 *tri_->getSubRegisters(LHS.reg)) {
1033 for (const unsigned* SR = tri_->getSubRegisters(LHS.reg); *SR; ++SR)
1034 if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
1035 DOUT << "Interfere with sub-register ";
1036 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
1039 } else if (TargetRegisterInfo::isPhysicalRegister(RHS.reg) &&
1040 *tri_->getSubRegisters(RHS.reg)) {
1041 for (const unsigned* SR = tri_->getSubRegisters(RHS.reg); *SR; ++SR)
1042 if (li_->hasInterval(*SR) && LHS.overlaps(li_->getInterval(*SR))) {
1043 DOUT << "Interfere with sub-register ";
1044 DEBUG(li_->getInterval(*SR).print(DOUT, tri_));
1049 // Compute ultimate value numbers for the LHS and RHS values.
1050 if (RHS.containsOneValue()) {
1051 // Copies from a liveinterval with a single value are simple to handle and
1052 // very common, handle the special case here. This is important, because
1053 // often RHS is small and LHS is large (e.g. a physreg).
1055 // Find out if the RHS is defined as a copy from some value in the LHS.
1056 int RHSVal0DefinedFromLHS = -1;
1058 VNInfo *RHSValNoInfo = NULL;
1059 VNInfo *RHSValNoInfo0 = RHS.getValNumInfo(0);
1060 unsigned RHSSrcReg = li_->getVNInfoSourceReg(RHSValNoInfo0);
1061 if ((RHSSrcReg == 0 || RHSSrcReg != LHS.reg)) {
1062 // If RHS is not defined as a copy from the LHS, we can use simpler and
1063 // faster checks to see if the live ranges are coalescable. This joiner
1064 // can't swap the LHS/RHS intervals though.
1065 if (!TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
1066 return SimpleJoin(LHS, RHS);
1068 RHSValNoInfo = RHSValNoInfo0;
1071 // It was defined as a copy from the LHS, find out what value # it is.
1072 RHSValNoInfo = LHS.getLiveRangeContaining(RHSValNoInfo0->def-1)->valno;
1073 RHSValID = RHSValNoInfo->id;
1074 RHSVal0DefinedFromLHS = RHSValID;
1077 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
1078 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
1079 NewVNInfo.resize(LHS.getNumValNums(), NULL);
1081 // Okay, *all* of the values in LHS that are defined as a copy from RHS
1082 // should now get updated.
1083 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1086 unsigned VN = VNI->id;
1087 if (unsigned LHSSrcReg = li_->getVNInfoSourceReg(VNI)) {
1088 if (LHSSrcReg != RHS.reg) {
1089 // If this is not a copy from the RHS, its value number will be
1090 // unmodified by the coalescing.
1091 NewVNInfo[VN] = VNI;
1092 LHSValNoAssignments[VN] = VN;
1093 } else if (RHSValID == -1) {
1094 // Otherwise, it is a copy from the RHS, and we don't already have a
1095 // value# for it. Keep the current value number, but remember it.
1096 LHSValNoAssignments[VN] = RHSValID = VN;
1097 NewVNInfo[VN] = RHSValNoInfo;
1098 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
1100 // Otherwise, use the specified value #.
1101 LHSValNoAssignments[VN] = RHSValID;
1102 if (VN == (unsigned)RHSValID) { // Else this val# is dead.
1103 NewVNInfo[VN] = RHSValNoInfo;
1104 LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
1108 NewVNInfo[VN] = VNI;
1109 LHSValNoAssignments[VN] = VN;
1113 assert(RHSValID != -1 && "Didn't find value #?");
1114 RHSValNoAssignments[0] = RHSValID;
1115 if (RHSVal0DefinedFromLHS != -1) {
1116 // This path doesn't go through ComputeUltimateVN so just set
1118 RHSValsDefinedFromLHS[RHSValNoInfo0] = (VNInfo*)1;
1121 // Loop over the value numbers of the LHS, seeing if any are defined from
1123 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1126 if (VNI->def == ~1U || VNI->copy == 0) // Src not defined by a copy?
1129 // DstReg is known to be a register in the LHS interval. If the src is
1130 // from the RHS interval, we can use its value #.
1131 if (li_->getVNInfoSourceReg(VNI) != RHS.reg)
1134 // Figure out the value # from the RHS.
1135 LHSValsDefinedFromRHS[VNI]=RHS.getLiveRangeContaining(VNI->def-1)->valno;
1138 // Loop over the value numbers of the RHS, seeing if any are defined from
1140 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1143 if (VNI->def == ~1U || VNI->copy == 0) // Src not defined by a copy?
1146 // DstReg is known to be a register in the RHS interval. If the src is
1147 // from the LHS interval, we can use its value #.
1148 if (li_->getVNInfoSourceReg(VNI) != LHS.reg)
1151 // Figure out the value # from the LHS.
1152 RHSValsDefinedFromLHS[VNI]=LHS.getLiveRangeContaining(VNI->def-1)->valno;
1155 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
1156 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
1157 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
1159 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1162 unsigned VN = VNI->id;
1163 if (LHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
1165 ComputeUltimateVN(VNI, NewVNInfo,
1166 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
1167 LHSValNoAssignments, RHSValNoAssignments);
1169 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1172 unsigned VN = VNI->id;
1173 if (RHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
1175 // If this value number isn't a copy from the LHS, it's a new number.
1176 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
1177 NewVNInfo.push_back(VNI);
1178 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
1182 ComputeUltimateVN(VNI, NewVNInfo,
1183 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
1184 RHSValNoAssignments, LHSValNoAssignments);
1188 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
1189 // interval lists to see if these intervals are coalescable.
1190 LiveInterval::const_iterator I = LHS.begin();
1191 LiveInterval::const_iterator IE = LHS.end();
1192 LiveInterval::const_iterator J = RHS.begin();
1193 LiveInterval::const_iterator JE = RHS.end();
1195 // Skip ahead until the first place of potential sharing.
1196 if (I->start < J->start) {
1197 I = std::upper_bound(I, IE, J->start);
1198 if (I != LHS.begin()) --I;
1199 } else if (J->start < I->start) {
1200 J = std::upper_bound(J, JE, I->start);
1201 if (J != RHS.begin()) --J;
1205 // Determine if these two live ranges overlap.
1207 if (I->start < J->start) {
1208 Overlaps = I->end > J->start;
1210 Overlaps = J->end > I->start;
1213 // If so, check value # info to determine if they are really different.
1215 // If the live range overlap will map to the same value number in the
1216 // result liverange, we can still coalesce them. If not, we can't.
1217 if (LHSValNoAssignments[I->valno->id] !=
1218 RHSValNoAssignments[J->valno->id])
1222 if (I->end < J->end) {
1231 // Update kill info. Some live ranges are extended due to copy coalescing.
1232 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
1233 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
1234 VNInfo *VNI = I->first;
1235 unsigned LHSValID = LHSValNoAssignments[VNI->id];
1236 LiveInterval::removeKill(NewVNInfo[LHSValID], VNI->def);
1237 NewVNInfo[LHSValID]->hasPHIKill |= VNI->hasPHIKill;
1238 RHS.addKills(NewVNInfo[LHSValID], VNI->kills);
1241 // Update kill info. Some live ranges are extended due to copy coalescing.
1242 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
1243 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
1244 VNInfo *VNI = I->first;
1245 unsigned RHSValID = RHSValNoAssignments[VNI->id];
1246 LiveInterval::removeKill(NewVNInfo[RHSValID], VNI->def);
1247 NewVNInfo[RHSValID]->hasPHIKill |= VNI->hasPHIKill;
1248 LHS.addKills(NewVNInfo[RHSValID], VNI->kills);
1251 // If we get here, we know that we can coalesce the live ranges. Ask the
1252 // intervals to coalesce themselves now.
1253 if ((RHS.ranges.size() > LHS.ranges.size() &&
1254 TargetRegisterInfo::isVirtualRegister(LHS.reg)) ||
1255 TargetRegisterInfo::isPhysicalRegister(RHS.reg)) {
1256 RHS.join(LHS, &RHSValNoAssignments[0], &LHSValNoAssignments[0], NewVNInfo);
1259 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo);
1266 // DepthMBBCompare - Comparison predicate that sort first based on the loop
1267 // depth of the basic block (the unsigned), and then on the MBB number.
1268 struct DepthMBBCompare {
1269 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
1270 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
1271 if (LHS.first > RHS.first) return true; // Deeper loops first
1272 return LHS.first == RHS.first &&
1273 LHS.second->getNumber() < RHS.second->getNumber();
1278 /// getRepIntervalSize - Returns the size of the interval that represents the
1279 /// specified register.
1281 unsigned JoinPriorityQueue<SF>::getRepIntervalSize(unsigned Reg) {
1282 return Rc->getRepIntervalSize(Reg);
1285 /// CopyRecSort::operator - Join priority queue sorting function.
1287 bool CopyRecSort::operator()(CopyRec left, CopyRec right) const {
1288 // Inner loops first.
1289 if (left.LoopDepth > right.LoopDepth)
1291 else if (left.LoopDepth == right.LoopDepth)
1292 if (left.isBackEdge && !right.isBackEdge)
1297 void SimpleRegisterCoalescing::CopyCoalesceInMBB(MachineBasicBlock *MBB,
1298 std::vector<CopyRec> &TryAgain) {
1299 DOUT << ((Value*)MBB->getBasicBlock())->getName() << ":\n";
1301 std::vector<CopyRec> VirtCopies;
1302 std::vector<CopyRec> PhysCopies;
1303 unsigned LoopDepth = loopInfo->getLoopDepth(MBB);
1304 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
1306 MachineInstr *Inst = MII++;
1308 // If this isn't a copy nor a extract_subreg, we can't join intervals.
1309 unsigned SrcReg, DstReg;
1310 if (Inst->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) {
1311 DstReg = Inst->getOperand(0).getReg();
1312 SrcReg = Inst->getOperand(1).getReg();
1313 } else if (!tii_->isMoveInstr(*Inst, SrcReg, DstReg))
1316 bool SrcIsPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
1317 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1319 JoinQueue->push(CopyRec(Inst, LoopDepth, isBackEdgeCopy(Inst, DstReg)));
1321 if (SrcIsPhys || DstIsPhys)
1322 PhysCopies.push_back(CopyRec(Inst, 0, false));
1324 VirtCopies.push_back(CopyRec(Inst, 0, false));
1331 // Try coalescing physical register + virtual register first.
1332 for (unsigned i = 0, e = PhysCopies.size(); i != e; ++i) {
1333 CopyRec &TheCopy = PhysCopies[i];
1335 if (!JoinCopy(TheCopy, Again))
1337 TryAgain.push_back(TheCopy);
1339 for (unsigned i = 0, e = VirtCopies.size(); i != e; ++i) {
1340 CopyRec &TheCopy = VirtCopies[i];
1342 if (!JoinCopy(TheCopy, Again))
1344 TryAgain.push_back(TheCopy);
1348 void SimpleRegisterCoalescing::joinIntervals() {
1349 DOUT << "********** JOINING INTERVALS ***********\n";
1352 JoinQueue = new JoinPriorityQueue<CopyRecSort>(this);
1354 std::vector<CopyRec> TryAgainList;
1355 if (loopInfo->begin() == loopInfo->end()) {
1356 // If there are no loops in the function, join intervals in function order.
1357 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();
1359 CopyCoalesceInMBB(I, TryAgainList);
1361 // Otherwise, join intervals in inner loops before other intervals.
1362 // Unfortunately we can't just iterate over loop hierarchy here because
1363 // there may be more MBB's than BB's. Collect MBB's for sorting.
1365 // Join intervals in the function prolog first. We want to join physical
1366 // registers with virtual registers before the intervals got too long.
1367 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
1368 for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();I != E;++I){
1369 MachineBasicBlock *MBB = I;
1370 MBBs.push_back(std::make_pair(loopInfo->getLoopDepth(MBB), I));
1373 // Sort by loop depth.
1374 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
1376 // Finally, join intervals in loop nest order.
1377 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
1378 CopyCoalesceInMBB(MBBs[i].second, TryAgainList);
1381 // Joining intervals can allow other intervals to be joined. Iteratively join
1382 // until we make no progress.
1384 SmallVector<CopyRec, 16> TryAgain;
1385 bool ProgressMade = true;
1386 while (ProgressMade) {
1387 ProgressMade = false;
1388 while (!JoinQueue->empty()) {
1389 CopyRec R = JoinQueue->pop();
1391 bool Success = JoinCopy(R, Again);
1393 ProgressMade = true;
1395 TryAgain.push_back(R);
1399 while (!TryAgain.empty()) {
1400 JoinQueue->push(TryAgain.back());
1401 TryAgain.pop_back();
1406 bool ProgressMade = true;
1407 while (ProgressMade) {
1408 ProgressMade = false;
1410 for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
1411 CopyRec &TheCopy = TryAgainList[i];
1414 bool Success = JoinCopy(TheCopy, Again);
1415 if (Success || !Again) {
1416 TheCopy.MI = 0; // Mark this one as done.
1417 ProgressMade = true;
1428 /// Return true if the two specified registers belong to different register
1429 /// classes. The registers may be either phys or virt regs.
1430 bool SimpleRegisterCoalescing::differingRegisterClasses(unsigned RegA,
1431 unsigned RegB) const {
1433 // Get the register classes for the first reg.
1434 if (TargetRegisterInfo::isPhysicalRegister(RegA)) {
1435 assert(TargetRegisterInfo::isVirtualRegister(RegB) &&
1436 "Shouldn't consider two physregs!");
1437 return !mri_->getRegClass(RegB)->contains(RegA);
1440 // Compare against the regclass for the second reg.
1441 const TargetRegisterClass *RegClass = mri_->getRegClass(RegA);
1442 if (TargetRegisterInfo::isVirtualRegister(RegB))
1443 return RegClass != mri_->getRegClass(RegB);
1445 return !RegClass->contains(RegB);
1448 /// lastRegisterUse - Returns the last use of the specific register between
1449 /// cycles Start and End or NULL if there are no uses.
1451 SimpleRegisterCoalescing::lastRegisterUse(unsigned Start, unsigned End,
1452 unsigned Reg, unsigned &UseIdx) const{
1454 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1455 MachineOperand *LastUse = NULL;
1456 for (MachineRegisterInfo::use_iterator I = mri_->use_begin(Reg),
1457 E = mri_->use_end(); I != E; ++I) {
1458 MachineOperand &Use = I.getOperand();
1459 MachineInstr *UseMI = Use.getParent();
1460 unsigned Idx = li_->getInstructionIndex(UseMI);
1461 if (Idx >= Start && Idx < End && Idx >= UseIdx) {
1469 int e = (End-1) / InstrSlots::NUM * InstrSlots::NUM;
1472 // Skip deleted instructions
1473 MachineInstr *MI = li_->getInstructionFromIndex(e);
1474 while ((e - InstrSlots::NUM) >= s && !MI) {
1475 e -= InstrSlots::NUM;
1476 MI = li_->getInstructionFromIndex(e);
1478 if (e < s || MI == NULL)
1481 for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
1482 MachineOperand &Use = MI->getOperand(i);
1483 if (Use.isRegister() && Use.isUse() && Use.getReg() &&
1484 tri_->regsOverlap(Use.getReg(), Reg)) {
1490 e -= InstrSlots::NUM;
1497 /// findDefOperand - Returns the MachineOperand that is a def of the specific
1498 /// register. It returns NULL if the def is not found.
1499 MachineOperand *SimpleRegisterCoalescing::findDefOperand(MachineInstr *MI,
1500 unsigned Reg) const {
1501 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1502 MachineOperand &MO = MI->getOperand(i);
1503 if (MO.isRegister() && MO.isDef() &&
1504 tri_->regsOverlap(MO.getReg(), Reg))
1510 /// unsetRegisterKills - Unset IsKill property of all uses of specific register
1511 /// between cycles Start and End.
1512 void SimpleRegisterCoalescing::unsetRegisterKills(unsigned Start, unsigned End,
1514 int e = (End-1) / InstrSlots::NUM * InstrSlots::NUM;
1517 // Skip deleted instructions
1518 MachineInstr *MI = li_->getInstructionFromIndex(e);
1519 while ((e - InstrSlots::NUM) >= s && !MI) {
1520 e -= InstrSlots::NUM;
1521 MI = li_->getInstructionFromIndex(e);
1523 if (e < s || MI == NULL)
1526 for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
1527 MachineOperand &MO = MI->getOperand(i);
1528 if (MO.isRegister() && MO.isKill() && MO.getReg() &&
1529 tri_->regsOverlap(MO.getReg(), Reg)) {
1530 MO.setIsKill(false);
1534 e -= InstrSlots::NUM;
1538 void SimpleRegisterCoalescing::printRegName(unsigned reg) const {
1539 if (TargetRegisterInfo::isPhysicalRegister(reg))
1540 cerr << tri_->getName(reg);
1542 cerr << "%reg" << reg;
1545 void SimpleRegisterCoalescing::releaseMemory() {
1546 JoinedCopies.clear();
1549 static bool isZeroLengthInterval(LiveInterval *li) {
1550 for (LiveInterval::Ranges::const_iterator
1551 i = li->ranges.begin(), e = li->ranges.end(); i != e; ++i)
1552 if (i->end - i->start > LiveIntervals::InstrSlots::NUM)
1557 bool SimpleRegisterCoalescing::runOnMachineFunction(MachineFunction &fn) {
1559 mri_ = &fn.getRegInfo();
1560 tm_ = &fn.getTarget();
1561 tri_ = tm_->getRegisterInfo();
1562 tii_ = tm_->getInstrInfo();
1563 li_ = &getAnalysis<LiveIntervals>();
1564 lv_ = &getAnalysis<LiveVariables>();
1565 loopInfo = &getAnalysis<MachineLoopInfo>();
1567 DOUT << "********** SIMPLE REGISTER COALESCING **********\n"
1568 << "********** Function: "
1569 << ((Value*)mf_->getFunction())->getName() << '\n';
1571 allocatableRegs_ = tri_->getAllocatableSet(fn);
1572 for (TargetRegisterInfo::regclass_iterator I = tri_->regclass_begin(),
1573 E = tri_->regclass_end(); I != E; ++I)
1574 allocatableRCRegs_.insert(std::make_pair(*I,
1575 tri_->getAllocatableSet(fn, *I)));
1577 // Join (coalesce) intervals if requested.
1578 if (EnableJoining) {
1580 DOUT << "********** INTERVALS POST JOINING **********\n";
1581 for (LiveIntervals::iterator I = li_->begin(), E = li_->end(); I != E; ++I){
1582 I->second.print(DOUT, tri_);
1586 // Delete all coalesced copies.
1587 for (SmallPtrSet<MachineInstr*,32>::iterator I = JoinedCopies.begin(),
1588 E = JoinedCopies.end(); I != E; ++I) {
1589 li_->RemoveMachineInstrFromMaps(*I);
1590 (*I)->eraseFromParent();
1595 // Perform a final pass over the instructions and compute spill weights
1596 // and remove identity moves.
1597 for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end();
1598 mbbi != mbbe; ++mbbi) {
1599 MachineBasicBlock* mbb = mbbi;
1600 unsigned loopDepth = loopInfo->getLoopDepth(mbb);
1602 for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
1604 // if the move will be an identity move delete it
1605 unsigned srcReg, dstReg;
1606 if (tii_->isMoveInstr(*mii, srcReg, dstReg) && srcReg == dstReg) {
1607 // remove from def list
1608 LiveInterval &RegInt = li_->getOrCreateInterval(srcReg);
1609 MachineOperand *MO = mii->findRegisterDefOperand(dstReg);
1610 // If def of this move instruction is dead, remove its live range from
1611 // the dstination register's live interval.
1613 unsigned MoveIdx = li_->getDefIndex(li_->getInstructionIndex(mii));
1614 LiveInterval::iterator MLR = RegInt.FindLiveRangeContaining(MoveIdx);
1615 RegInt.removeRange(MLR->start, MoveIdx+1, true);
1617 li_->removeInterval(srcReg);
1619 li_->RemoveMachineInstrFromMaps(mii);
1620 mii = mbbi->erase(mii);
1623 SmallSet<unsigned, 4> UniqueUses;
1624 for (unsigned i = 0, e = mii->getNumOperands(); i != e; ++i) {
1625 const MachineOperand &mop = mii->getOperand(i);
1626 if (mop.isRegister() && mop.getReg() &&
1627 TargetRegisterInfo::isVirtualRegister(mop.getReg())) {
1628 unsigned reg = mop.getReg();
1629 // Multiple uses of reg by the same instruction. It should not
1630 // contribute to spill weight again.
1631 if (UniqueUses.count(reg) != 0)
1633 LiveInterval &RegInt = li_->getInterval(reg);
1635 li_->getSpillWeight(mop.isDef(), mop.isUse(), loopDepth);
1636 UniqueUses.insert(reg);
1644 for (LiveIntervals::iterator I = li_->begin(), E = li_->end(); I != E; ++I) {
1645 LiveInterval &LI = I->second;
1646 if (TargetRegisterInfo::isVirtualRegister(LI.reg)) {
1647 // If the live interval length is essentially zero, i.e. in every live
1648 // range the use follows def immediately, it doesn't make sense to spill
1649 // it and hope it will be easier to allocate for this li.
1650 if (isZeroLengthInterval(&LI))
1651 LI.weight = HUGE_VALF;
1653 bool isLoad = false;
1654 if (li_->isReMaterializable(LI, isLoad)) {
1655 // If all of the definitions of the interval are re-materializable,
1656 // it is a preferred candidate for spilling. If non of the defs are
1657 // loads, then it's potentially very cheap to re-materialize.
1658 // FIXME: this gets much more complicated once we support non-trivial
1659 // re-materialization.
1667 // Slightly prefer live interval that has been assigned a preferred reg.
1671 // Divide the weight of the interval by its size. This encourages
1672 // spilling of intervals that are large and have few uses, and
1673 // discourages spilling of small intervals with many uses.
1674 LI.weight /= LI.getSize();
1682 /// print - Implement the dump method.
1683 void SimpleRegisterCoalescing::print(std::ostream &O, const Module* m) const {
1687 RegisterCoalescer* llvm::createSimpleRegisterCoalescer() {
1688 return new SimpleRegisterCoalescing();
1691 // Make sure that anything that uses RegisterCoalescer pulls in this file...
1692 DEFINING_FILE_FOR(SimpleRegisterCoalescing)