1 //===- RegisterCoalescer.cpp - Generic Register Coalescing Interface -------==//
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 the generic RegisterCoalescer interface which
11 // is used as the common interface used by all clients and
12 // implementations of register coalescing.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "regalloc"
17 #include "RegisterCoalescer.h"
18 #include "LiveDebugVariables.h"
19 #include "RegisterClassInfo.h"
20 #include "VirtRegMap.h"
22 #include "llvm/Pass.h"
23 #include "llvm/Value.h"
24 #include "llvm/ADT/OwningPtr.h"
25 #include "llvm/ADT/STLExtras.h"
26 #include "llvm/ADT/SmallSet.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/AliasAnalysis.h"
29 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
30 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
31 #include "llvm/CodeGen/LiveRangeEdit.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineInstr.h"
34 #include "llvm/CodeGen/MachineInstr.h"
35 #include "llvm/CodeGen/MachineLoopInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/Passes.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetInstrInfo.h"
45 #include "llvm/Target/TargetMachine.h"
46 #include "llvm/Target/TargetOptions.h"
47 #include "llvm/Target/TargetRegisterInfo.h"
52 STATISTIC(numJoins , "Number of interval joins performed");
53 STATISTIC(numCrossRCs , "Number of cross class joins performed");
54 STATISTIC(numCommutes , "Number of instruction commuting performed");
55 STATISTIC(numExtends , "Number of copies extended");
56 STATISTIC(NumReMats , "Number of instructions re-materialized");
57 STATISTIC(numPeep , "Number of identity moves eliminated after coalescing");
58 STATISTIC(NumInflated , "Number of register classes inflated");
61 EnableJoining("join-liveintervals",
62 cl::desc("Coalesce copies (default=true)"),
66 VerifyCoalescing("verify-coalescing",
67 cl::desc("Verify machine instrs before and after register coalescing"),
71 class RegisterCoalescer : public MachineFunctionPass,
72 private LiveRangeEdit::Delegate {
74 MachineRegisterInfo* MRI;
75 const TargetMachine* TM;
76 const TargetRegisterInfo* TRI;
77 const TargetInstrInfo* TII;
79 LiveDebugVariables *LDV;
80 const MachineLoopInfo* Loops;
82 RegisterClassInfo RegClassInfo;
84 /// JoinedCopies - Keep track of copies eliminated due to coalescing.
86 SmallPtrSet<MachineInstr*, 32> JoinedCopies;
88 /// ReMatDefs - Keep track of definition instructions which have
90 SmallPtrSet<MachineInstr*, 8> ReMatDefs;
92 /// WorkList - Copy instructions yet to be coalesced.
93 SmallVector<MachineInstr*, 8> WorkList;
95 /// ErasedInstrs - Set of instruction pointers that have been erased, and
96 /// that may be present in WorkList.
97 SmallPtrSet<MachineInstr*, 8> ErasedInstrs;
99 /// Dead instructions that are about to be deleted.
100 SmallVector<MachineInstr*, 8> DeadDefs;
102 /// Recursively eliminate dead defs in DeadDefs.
103 void eliminateDeadDefs();
105 /// LiveRangeEdit callback.
106 void LRE_WillEraseInstruction(MachineInstr *MI);
108 /// joinAllIntervals - join compatible live intervals
109 void joinAllIntervals();
111 /// copyCoalesceInMBB - Coalesce copies in the specified MBB, putting
112 /// copies that cannot yet be coalesced into WorkList.
113 void copyCoalesceInMBB(MachineBasicBlock *MBB);
115 /// copyCoalesceWorkList - Try to coalesce all copies in WorkList after
116 /// position From. Return true if any progress was made.
117 bool copyCoalesceWorkList(unsigned From = 0);
119 /// joinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
120 /// which are the src/dst of the copy instruction CopyMI. This returns
121 /// true if the copy was successfully coalesced away. If it is not
122 /// currently possible to coalesce this interval, but it may be possible if
123 /// other things get coalesced, then it returns true by reference in
125 bool joinCopy(MachineInstr *TheCopy, bool &Again);
127 /// joinIntervals - Attempt to join these two intervals. On failure, this
128 /// returns false. The output "SrcInt" will not have been modified, so we
129 /// can use this information below to update aliases.
130 bool joinIntervals(CoalescerPair &CP);
132 /// Attempt joining with a reserved physreg.
133 bool joinReservedPhysReg(CoalescerPair &CP);
135 /// adjustCopiesBackFrom - We found a non-trivially-coalescable copy. If
136 /// the source value number is defined by a copy from the destination reg
137 /// see if we can merge these two destination reg valno# into a single
138 /// value number, eliminating a copy.
139 bool adjustCopiesBackFrom(const CoalescerPair &CP, MachineInstr *CopyMI);
141 /// hasOtherReachingDefs - Return true if there are definitions of IntB
142 /// other than BValNo val# that can reach uses of AValno val# of IntA.
143 bool hasOtherReachingDefs(LiveInterval &IntA, LiveInterval &IntB,
144 VNInfo *AValNo, VNInfo *BValNo);
146 /// removeCopyByCommutingDef - We found a non-trivially-coalescable copy.
147 /// If the source value number is defined by a commutable instruction and
148 /// its other operand is coalesced to the copy dest register, see if we
149 /// can transform the copy into a noop by commuting the definition.
150 bool removeCopyByCommutingDef(const CoalescerPair &CP,MachineInstr *CopyMI);
152 /// reMaterializeTrivialDef - If the source of a copy is defined by a
153 /// trivial computation, replace the copy by rematerialize the definition.
154 /// If PreserveSrcInt is true, make sure SrcInt is valid after the call.
155 bool reMaterializeTrivialDef(LiveInterval &SrcInt, bool PreserveSrcInt,
156 unsigned DstReg, MachineInstr *CopyMI);
158 /// canJoinPhys - Return true if a physreg copy should be joined.
159 bool canJoinPhys(CoalescerPair &CP);
161 /// updateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
162 /// update the subregister number if it is not zero. If DstReg is a
163 /// physical register and the existing subregister number of the def / use
164 /// being updated is not zero, make sure to set it to the correct physical
166 void updateRegDefsUses(unsigned SrcReg, unsigned DstReg, unsigned SubIdx);
168 /// removeDeadDef - If a def of a live interval is now determined dead,
169 /// remove the val# it defines. If the live interval becomes empty, remove
171 bool removeDeadDef(LiveInterval &li, MachineInstr *DefMI);
173 /// markAsJoined - Remember that CopyMI has already been joined.
174 void markAsJoined(MachineInstr *CopyMI);
176 /// eliminateUndefCopy - Handle copies of undef values.
177 bool eliminateUndefCopy(MachineInstr *CopyMI, const CoalescerPair &CP);
180 static char ID; // Class identification, replacement for typeinfo
181 RegisterCoalescer() : MachineFunctionPass(ID) {
182 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
185 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
187 virtual void releaseMemory();
189 /// runOnMachineFunction - pass entry point
190 virtual bool runOnMachineFunction(MachineFunction&);
192 /// print - Implement the dump method.
193 virtual void print(raw_ostream &O, const Module* = 0) const;
195 } /// end anonymous namespace
197 char &llvm::RegisterCoalescerID = RegisterCoalescer::ID;
199 INITIALIZE_PASS_BEGIN(RegisterCoalescer, "simple-register-coalescing",
200 "Simple Register Coalescing", false, false)
201 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
202 INITIALIZE_PASS_DEPENDENCY(LiveDebugVariables)
203 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
204 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
205 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
206 INITIALIZE_PASS_END(RegisterCoalescer, "simple-register-coalescing",
207 "Simple Register Coalescing", false, false)
209 char RegisterCoalescer::ID = 0;
211 static unsigned compose(const TargetRegisterInfo &tri, unsigned a, unsigned b) {
214 return tri.composeSubRegIndices(a, b);
217 static bool isMoveInstr(const TargetRegisterInfo &tri, const MachineInstr *MI,
218 unsigned &Src, unsigned &Dst,
219 unsigned &SrcSub, unsigned &DstSub) {
221 Dst = MI->getOperand(0).getReg();
222 DstSub = MI->getOperand(0).getSubReg();
223 Src = MI->getOperand(1).getReg();
224 SrcSub = MI->getOperand(1).getSubReg();
225 } else if (MI->isSubregToReg()) {
226 Dst = MI->getOperand(0).getReg();
227 DstSub = compose(tri, MI->getOperand(0).getSubReg(),
228 MI->getOperand(3).getImm());
229 Src = MI->getOperand(2).getReg();
230 SrcSub = MI->getOperand(2).getSubReg();
236 bool CoalescerPair::setRegisters(const MachineInstr *MI) {
240 Flipped = CrossClass = false;
242 unsigned Src, Dst, SrcSub, DstSub;
243 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
245 Partial = SrcSub || DstSub;
247 // If one register is a physreg, it must be Dst.
248 if (TargetRegisterInfo::isPhysicalRegister(Src)) {
249 if (TargetRegisterInfo::isPhysicalRegister(Dst))
252 std::swap(SrcSub, DstSub);
256 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
258 if (TargetRegisterInfo::isPhysicalRegister(Dst)) {
259 // Eliminate DstSub on a physreg.
261 Dst = TRI.getSubReg(Dst, DstSub);
262 if (!Dst) return false;
266 // Eliminate SrcSub by picking a corresponding Dst superregister.
268 Dst = TRI.getMatchingSuperReg(Dst, SrcSub, MRI.getRegClass(Src));
269 if (!Dst) return false;
271 } else if (!MRI.getRegClass(Src)->contains(Dst)) {
275 // Both registers are virtual.
276 const TargetRegisterClass *SrcRC = MRI.getRegClass(Src);
277 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
279 // Both registers have subreg indices.
280 if (SrcSub && DstSub) {
281 // Copies between different sub-registers are never coalescable.
282 if (Src == Dst && SrcSub != DstSub)
285 NewRC = TRI.getCommonSuperRegClass(SrcRC, SrcSub, DstRC, DstSub,
290 // SrcReg will be merged with a sub-register of DstReg.
292 NewRC = TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSub);
294 // DstReg will be merged with a sub-register of SrcReg.
296 NewRC = TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSub);
298 // This is a straight copy without sub-registers.
299 NewRC = TRI.getCommonSubClass(DstRC, SrcRC);
302 // The combined constraint may be impossible to satisfy.
306 // Prefer SrcReg to be a sub-register of DstReg.
307 // FIXME: Coalescer should support subregs symmetrically.
308 if (DstIdx && !SrcIdx) {
310 std::swap(SrcIdx, DstIdx);
314 CrossClass = NewRC != DstRC || NewRC != SrcRC;
316 // Check our invariants
317 assert(TargetRegisterInfo::isVirtualRegister(Src) && "Src must be virtual");
318 assert(!(TargetRegisterInfo::isPhysicalRegister(Dst) && DstSub) &&
319 "Cannot have a physical SubIdx");
325 bool CoalescerPair::flip() {
326 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
328 std::swap(SrcReg, DstReg);
329 std::swap(SrcIdx, DstIdx);
334 bool CoalescerPair::isCoalescable(const MachineInstr *MI) const {
337 unsigned Src, Dst, SrcSub, DstSub;
338 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
341 // Find the virtual register that is SrcReg.
344 std::swap(SrcSub, DstSub);
345 } else if (Src != SrcReg) {
349 // Now check that Dst matches DstReg.
350 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
351 if (!TargetRegisterInfo::isPhysicalRegister(Dst))
353 assert(!DstIdx && !SrcIdx && "Inconsistent CoalescerPair state.");
354 // DstSub could be set for a physreg from INSERT_SUBREG.
356 Dst = TRI.getSubReg(Dst, DstSub);
359 return DstReg == Dst;
360 // This is a partial register copy. Check that the parts match.
361 return TRI.getSubReg(DstReg, SrcSub) == Dst;
363 // DstReg is virtual.
366 // Registers match, do the subregisters line up?
367 return compose(TRI, SrcIdx, SrcSub) == compose(TRI, DstIdx, DstSub);
371 void RegisterCoalescer::getAnalysisUsage(AnalysisUsage &AU) const {
372 AU.setPreservesCFG();
373 AU.addRequired<AliasAnalysis>();
374 AU.addRequired<LiveIntervals>();
375 AU.addPreserved<LiveIntervals>();
376 AU.addRequired<LiveDebugVariables>();
377 AU.addPreserved<LiveDebugVariables>();
378 AU.addPreserved<SlotIndexes>();
379 AU.addRequired<MachineLoopInfo>();
380 AU.addPreserved<MachineLoopInfo>();
381 AU.addPreservedID(MachineDominatorsID);
382 MachineFunctionPass::getAnalysisUsage(AU);
385 void RegisterCoalescer::markAsJoined(MachineInstr *CopyMI) {
386 /// Joined copies are not deleted immediately, but kept in JoinedCopies.
387 JoinedCopies.insert(CopyMI);
389 /// Mark all register operands of CopyMI as <undef> so they won't affect dead
390 /// code elimination.
391 for (MachineInstr::mop_iterator I = CopyMI->operands_begin(),
392 E = CopyMI->operands_end(); I != E; ++I)
397 void RegisterCoalescer::eliminateDeadDefs() {
398 SmallVector<LiveInterval*, 8> NewRegs;
399 LiveRangeEdit(0, NewRegs, *MF, *LIS, 0, this).eliminateDeadDefs(DeadDefs);
402 // Callback from eliminateDeadDefs().
403 void RegisterCoalescer::LRE_WillEraseInstruction(MachineInstr *MI) {
404 // MI may be in WorkList. Make sure we don't visit it.
405 ErasedInstrs.insert(MI);
408 /// adjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA
409 /// being the source and IntB being the dest, thus this defines a value number
410 /// in IntB. If the source value number (in IntA) is defined by a copy from B,
411 /// see if we can merge these two pieces of B into a single value number,
412 /// eliminating a copy. For example:
416 /// B1 = A3 <- this copy
418 /// In this case, B0 can be extended to where the B1 copy lives, allowing the B1
419 /// value number to be replaced with B0 (which simplifies the B liveinterval).
421 /// This returns true if an interval was modified.
423 bool RegisterCoalescer::adjustCopiesBackFrom(const CoalescerPair &CP,
424 MachineInstr *CopyMI) {
425 assert(!CP.isPartial() && "This doesn't work for partial copies.");
427 // Bail if there is no dst interval - can happen when merging physical subreg
429 if (!LIS->hasInterval(CP.getDstReg()))
433 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
435 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
436 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
438 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
439 // the example above.
440 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
441 if (BLR == IntB.end()) return false;
442 VNInfo *BValNo = BLR->valno;
444 // Get the location that B is defined at. Two options: either this value has
445 // an unknown definition point or it is defined at CopyIdx. If unknown, we
447 if (BValNo->def != CopyIdx) return false;
449 // AValNo is the value number in A that defines the copy, A3 in the example.
450 SlotIndex CopyUseIdx = CopyIdx.getRegSlot(true);
451 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyUseIdx);
452 // The live range might not exist after fun with physreg coalescing.
453 if (ALR == IntA.end()) return false;
454 VNInfo *AValNo = ALR->valno;
456 // If AValNo is defined as a copy from IntB, we can potentially process this.
457 // Get the instruction that defines this value number.
458 MachineInstr *ACopyMI = LIS->getInstructionFromIndex(AValNo->def);
459 if (!CP.isCoalescable(ACopyMI))
462 // Get the LiveRange in IntB that this value number starts with.
463 LiveInterval::iterator ValLR =
464 IntB.FindLiveRangeContaining(AValNo->def.getPrevSlot());
465 if (ValLR == IntB.end())
468 // Make sure that the end of the live range is inside the same block as
470 MachineInstr *ValLREndInst =
471 LIS->getInstructionFromIndex(ValLR->end.getPrevSlot());
472 if (!ValLREndInst || ValLREndInst->getParent() != CopyMI->getParent())
475 // Okay, we now know that ValLR ends in the same block that the CopyMI
476 // live-range starts. If there are no intervening live ranges between them in
477 // IntB, we can merge them.
478 if (ValLR+1 != BLR) return false;
480 // If a live interval is a physical register, conservatively check if any
481 // of its aliases is overlapping the live interval of the virtual register.
482 // If so, do not coalesce.
483 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
484 for (const uint16_t *AS = TRI->getAliasSet(IntB.reg); *AS; ++AS)
485 if (LIS->hasInterval(*AS) && IntA.overlaps(LIS->getInterval(*AS))) {
487 dbgs() << "\t\tInterfere with alias ";
488 LIS->getInterval(*AS).print(dbgs(), TRI);
495 dbgs() << "Extending: ";
496 IntB.print(dbgs(), TRI);
499 SlotIndex FillerStart = ValLR->end, FillerEnd = BLR->start;
500 // We are about to delete CopyMI, so need to remove it as the 'instruction
501 // that defines this value #'. Update the valnum with the new defining
503 BValNo->def = FillerStart;
505 // Okay, we can merge them. We need to insert a new liverange:
506 // [ValLR.end, BLR.begin) of either value number, then we merge the
507 // two value numbers.
508 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
510 // If the IntB live range is assigned to a physical register, and if that
511 // physreg has sub-registers, update their live intervals as well.
512 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
513 for (const uint16_t *SR = TRI->getSubRegisters(IntB.reg); *SR; ++SR) {
514 if (!LIS->hasInterval(*SR))
516 LiveInterval &SRLI = LIS->getInterval(*SR);
517 SRLI.addRange(LiveRange(FillerStart, FillerEnd,
518 SRLI.getNextValue(FillerStart,
519 LIS->getVNInfoAllocator())));
523 // Okay, merge "B1" into the same value number as "B0".
524 if (BValNo != ValLR->valno) {
525 // If B1 is killed by a PHI, then the merged live range must also be killed
526 // by the same PHI, as B0 and B1 can not overlap.
527 bool HasPHIKill = BValNo->hasPHIKill();
528 IntB.MergeValueNumberInto(BValNo, ValLR->valno);
530 ValLR->valno->setHasPHIKill(true);
533 dbgs() << " result = ";
534 IntB.print(dbgs(), TRI);
538 // If the source instruction was killing the source register before the
539 // merge, unset the isKill marker given the live range has been extended.
540 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true);
542 ValLREndInst->getOperand(UIdx).setIsKill(false);
545 // Rewrite the copy. If the copy instruction was killing the destination
546 // register before the merge, find the last use and trim the live range. That
547 // will also add the isKill marker.
548 CopyMI->substituteRegister(IntA.reg, IntB.reg, 0, *TRI);
549 if (ALR->end == CopyIdx)
550 LIS->shrinkToUses(&IntA);
556 /// hasOtherReachingDefs - Return true if there are definitions of IntB
557 /// other than BValNo val# that can reach uses of AValno val# of IntA.
558 bool RegisterCoalescer::hasOtherReachingDefs(LiveInterval &IntA,
562 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
564 if (AI->valno != AValNo) continue;
565 LiveInterval::Ranges::iterator BI =
566 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start);
567 if (BI != IntB.ranges.begin())
569 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) {
570 if (BI->valno == BValNo)
572 if (BI->start <= AI->start && BI->end > AI->start)
574 if (BI->start > AI->start && BI->start < AI->end)
581 /// removeCopyByCommutingDef - We found a non-trivially-coalescable copy with
582 /// IntA being the source and IntB being the dest, thus this defines a value
583 /// number in IntB. If the source value number (in IntA) is defined by a
584 /// commutable instruction and its other operand is coalesced to the copy dest
585 /// register, see if we can transform the copy into a noop by commuting the
586 /// definition. For example,
588 /// A3 = op A2 B0<kill>
590 /// B1 = A3 <- this copy
592 /// = op A3 <- more uses
596 /// B2 = op B0 A2<kill>
598 /// B1 = B2 <- now an identify copy
600 /// = op B2 <- more uses
602 /// This returns true if an interval was modified.
604 bool RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP,
605 MachineInstr *CopyMI) {
606 // FIXME: For now, only eliminate the copy by commuting its def when the
607 // source register is a virtual register. We want to guard against cases
608 // where the copy is a back edge copy and commuting the def lengthen the
609 // live interval of the source register to the entire loop.
610 if (CP.isPhys() && CP.isFlipped())
613 // Bail if there is no dst interval.
614 if (!LIS->hasInterval(CP.getDstReg()))
617 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
620 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
622 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
624 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
625 // the example above.
626 VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx);
627 if (!BValNo || BValNo->def != CopyIdx)
630 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
632 // AValNo is the value number in A that defines the copy, A3 in the example.
633 VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getRegSlot(true));
634 assert(AValNo && "COPY source not live");
636 // If other defs can reach uses of this def, then it's not safe to perform
638 if (AValNo->isPHIDef() || AValNo->isUnused() || AValNo->hasPHIKill())
640 MachineInstr *DefMI = LIS->getInstructionFromIndex(AValNo->def);
643 if (!DefMI->isCommutable())
645 // If DefMI is a two-address instruction then commuting it will change the
646 // destination register.
647 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
648 assert(DefIdx != -1);
650 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
652 unsigned Op1, Op2, NewDstIdx;
653 if (!TII->findCommutedOpIndices(DefMI, Op1, Op2))
657 else if (Op2 == UseOpIdx)
662 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
663 unsigned NewReg = NewDstMO.getReg();
664 if (NewReg != IntB.reg || !NewDstMO.isKill())
667 // Make sure there are no other definitions of IntB that would reach the
668 // uses which the new definition can reach.
669 if (hasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
672 // Abort if the aliases of IntB.reg have values that are not simply the
673 // clobbers from the superreg.
674 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg))
675 for (const uint16_t *AS = TRI->getAliasSet(IntB.reg); *AS; ++AS)
676 if (LIS->hasInterval(*AS) &&
677 hasOtherReachingDefs(IntA, LIS->getInterval(*AS), AValNo, 0))
680 // If some of the uses of IntA.reg is already coalesced away, return false.
681 // It's not possible to determine whether it's safe to perform the coalescing.
682 for (MachineRegisterInfo::use_nodbg_iterator UI =
683 MRI->use_nodbg_begin(IntA.reg),
684 UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
685 MachineInstr *UseMI = &*UI;
686 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI);
687 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
688 if (ULR == IntA.end())
690 if (ULR->valno == AValNo && JoinedCopies.count(UseMI))
694 DEBUG(dbgs() << "\tremoveCopyByCommutingDef: " << AValNo->def << '\t'
697 // At this point we have decided that it is legal to do this
698 // transformation. Start by commuting the instruction.
699 MachineBasicBlock *MBB = DefMI->getParent();
700 MachineInstr *NewMI = TII->commuteInstruction(DefMI);
703 if (TargetRegisterInfo::isVirtualRegister(IntA.reg) &&
704 TargetRegisterInfo::isVirtualRegister(IntB.reg) &&
705 !MRI->constrainRegClass(IntB.reg, MRI->getRegClass(IntA.reg)))
707 if (NewMI != DefMI) {
708 LIS->ReplaceMachineInstrInMaps(DefMI, NewMI);
709 MachineBasicBlock::iterator Pos = DefMI;
710 MBB->insert(Pos, NewMI);
713 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false);
714 NewMI->getOperand(OpIdx).setIsKill();
716 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
725 // Update uses of IntA of the specific Val# with IntB.
726 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(IntA.reg),
727 UE = MRI->use_end(); UI != UE;) {
728 MachineOperand &UseMO = UI.getOperand();
729 MachineInstr *UseMI = &*UI;
731 if (JoinedCopies.count(UseMI))
733 if (UseMI->isDebugValue()) {
734 // FIXME These don't have an instruction index. Not clear we have enough
735 // info to decide whether to do this replacement or not. For now do it.
736 UseMO.setReg(NewReg);
739 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI).getRegSlot(true);
740 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
741 if (ULR == IntA.end() || ULR->valno != AValNo)
743 if (TargetRegisterInfo::isPhysicalRegister(NewReg))
744 UseMO.substPhysReg(NewReg, *TRI);
746 UseMO.setReg(NewReg);
749 if (!UseMI->isCopy())
751 if (UseMI->getOperand(0).getReg() != IntB.reg ||
752 UseMI->getOperand(0).getSubReg())
755 // This copy will become a noop. If it's defining a new val#, merge it into
757 SlotIndex DefIdx = UseIdx.getRegSlot();
758 VNInfo *DVNI = IntB.getVNInfoAt(DefIdx);
761 DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI);
762 assert(DVNI->def == DefIdx);
763 BValNo = IntB.MergeValueNumberInto(BValNo, DVNI);
767 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition
769 VNInfo *ValNo = BValNo;
770 ValNo->def = AValNo->def;
771 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
773 if (AI->valno != AValNo) continue;
774 IntB.addRange(LiveRange(AI->start, AI->end, ValNo));
776 DEBUG(dbgs() << "\t\textended: " << IntB << '\n');
778 IntA.removeValNo(AValNo);
779 DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n');
784 /// reMaterializeTrivialDef - If the source of a copy is defined by a trivial
785 /// computation, replace the copy by rematerialize the definition.
786 bool RegisterCoalescer::reMaterializeTrivialDef(LiveInterval &SrcInt,
789 MachineInstr *CopyMI) {
790 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot(true);
791 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx);
792 assert(SrcLR != SrcInt.end() && "Live range not found!");
793 VNInfo *ValNo = SrcLR->valno;
794 if (ValNo->isPHIDef() || ValNo->isUnused())
796 MachineInstr *DefMI = LIS->getInstructionFromIndex(ValNo->def);
799 assert(DefMI && "Defining instruction disappeared");
800 if (!DefMI->isAsCheapAsAMove())
802 if (!TII->isTriviallyReMaterializable(DefMI, AA))
804 bool SawStore = false;
805 if (!DefMI->isSafeToMove(TII, AA, SawStore))
807 const MCInstrDesc &MCID = DefMI->getDesc();
808 if (MCID.getNumDefs() != 1)
810 if (!DefMI->isImplicitDef()) {
811 // Make sure the copy destination register class fits the instruction
812 // definition register class. The mismatch can happen as a result of earlier
813 // extract_subreg, insert_subreg, subreg_to_reg coalescing.
814 const TargetRegisterClass *RC = TII->getRegClass(MCID, 0, TRI, *MF);
815 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
816 if (MRI->getRegClass(DstReg) != RC)
818 } else if (!RC->contains(DstReg))
822 MachineBasicBlock *MBB = CopyMI->getParent();
823 MachineBasicBlock::iterator MII =
824 llvm::next(MachineBasicBlock::iterator(CopyMI));
825 TII->reMaterialize(*MBB, MII, DstReg, 0, DefMI, *TRI);
826 MachineInstr *NewMI = prior(MII);
828 // NewMI may have dead implicit defs (E.g. EFLAGS for MOV<bits>r0 on X86).
829 // We need to remember these so we can add intervals once we insert
830 // NewMI into SlotIndexes.
831 SmallVector<unsigned, 4> NewMIImplDefs;
832 for (unsigned i = NewMI->getDesc().getNumOperands(),
833 e = NewMI->getNumOperands(); i != e; ++i) {
834 MachineOperand &MO = NewMI->getOperand(i);
836 assert(MO.isDef() && MO.isImplicit() && MO.isDead() &&
837 TargetRegisterInfo::isPhysicalRegister(MO.getReg()));
838 NewMIImplDefs.push_back(MO.getReg());
842 // CopyMI may have implicit operands, transfer them over to the newly
843 // rematerialized instruction. And update implicit def interval valnos.
844 for (unsigned i = CopyMI->getDesc().getNumOperands(),
845 e = CopyMI->getNumOperands(); i != e; ++i) {
846 MachineOperand &MO = CopyMI->getOperand(i);
848 assert(MO.isImplicit() && "No explicit operands after implict operands.");
849 // Discard VReg implicit defs.
850 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
851 NewMI->addOperand(MO);
856 LIS->ReplaceMachineInstrInMaps(CopyMI, NewMI);
858 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
859 for (unsigned i = 0, e = NewMIImplDefs.size(); i != e; ++i) {
860 unsigned reg = NewMIImplDefs[i];
861 LiveInterval &li = LIS->getInterval(reg);
862 VNInfo *DeadDefVN = li.getNextValue(NewMIIdx.getRegSlot(),
863 LIS->getVNInfoAllocator());
864 LiveRange lr(NewMIIdx.getRegSlot(), NewMIIdx.getDeadSlot(), DeadDefVN);
868 CopyMI->eraseFromParent();
869 ErasedInstrs.insert(CopyMI);
870 ReMatDefs.insert(DefMI);
871 DEBUG(dbgs() << "Remat: " << *NewMI);
874 // The source interval can become smaller because we removed a use.
876 LIS->shrinkToUses(&SrcInt);
881 /// eliminateUndefCopy - ProcessImpicitDefs may leave some copies of <undef>
882 /// values, it only removes local variables. When we have a copy like:
884 /// %vreg1 = COPY %vreg2<undef>
886 /// We delete the copy and remove the corresponding value number from %vreg1.
887 /// Any uses of that value number are marked as <undef>.
888 bool RegisterCoalescer::eliminateUndefCopy(MachineInstr *CopyMI,
889 const CoalescerPair &CP) {
890 SlotIndex Idx = LIS->getInstructionIndex(CopyMI);
891 LiveInterval *SrcInt = &LIS->getInterval(CP.getSrcReg());
892 if (SrcInt->liveAt(Idx))
894 LiveInterval *DstInt = &LIS->getInterval(CP.getDstReg());
895 if (DstInt->liveAt(Idx))
898 // No intervals are live-in to CopyMI - it is undef.
903 VNInfo *DeadVNI = DstInt->getVNInfoAt(Idx.getRegSlot());
904 assert(DeadVNI && "No value defined in DstInt");
905 DstInt->removeValNo(DeadVNI);
907 // Find new undef uses.
908 for (MachineRegisterInfo::reg_nodbg_iterator
909 I = MRI->reg_nodbg_begin(DstInt->reg), E = MRI->reg_nodbg_end();
911 MachineOperand &MO = I.getOperand();
912 if (MO.isDef() || MO.isUndef())
914 MachineInstr *MI = MO.getParent();
915 SlotIndex Idx = LIS->getInstructionIndex(MI);
916 if (DstInt->liveAt(Idx))
919 DEBUG(dbgs() << "\tnew undef: " << Idx << '\t' << *MI);
924 /// updateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
925 /// update the subregister number if it is not zero. If DstReg is a
926 /// physical register and the existing subregister number of the def / use
927 /// being updated is not zero, make sure to set it to the correct physical
929 void RegisterCoalescer::updateRegDefsUses(unsigned SrcReg,
932 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
933 LiveInterval &DstInt = LIS->getInterval(DstReg);
935 // Update LiveDebugVariables.
936 LDV->renameRegister(SrcReg, DstReg, SubIdx);
938 for (MachineRegisterInfo::reg_iterator I = MRI->reg_begin(SrcReg);
939 MachineInstr *UseMI = I.skipInstruction();) {
940 bool AlreadyJoined = JoinedCopies.count(UseMI);
942 // A PhysReg copy that won't be coalesced can perhaps be rematerialized
945 if (UseMI->isFullCopy() &&
946 UseMI->getOperand(1).getReg() == SrcReg &&
947 UseMI->getOperand(0).getReg() != SrcReg &&
948 UseMI->getOperand(0).getReg() != DstReg &&
950 reMaterializeTrivialDef(LIS->getInterval(SrcReg), false,
951 UseMI->getOperand(0).getReg(), UseMI))
955 SmallVector<unsigned,8> Ops;
957 tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
959 // If SrcReg wasn't read, it may still be the case that DstReg is live-in
960 // because SrcReg is a sub-register.
961 if (!Reads && SubIdx && !AlreadyJoined)
962 Reads = DstInt.liveAt(LIS->getInstructionIndex(UseMI));
964 // Replace SrcReg with DstReg in all UseMI operands.
965 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
966 MachineOperand &MO = UseMI->getOperand(Ops[i]);
968 // Adjust <undef> flags in case of sub-register joins. We don't want to
969 // turn a full def into a read-modify-write sub-register def and vice
971 if (SubIdx && !AlreadyJoined && MO.isDef())
972 MO.setIsUndef(!Reads);
975 MO.substPhysReg(DstReg, *TRI);
977 MO.substVirtReg(DstReg, SubIdx, *TRI);
980 // This instruction is a copy that will be removed.
985 dbgs() << "\t\tupdated: ";
986 if (!UseMI->isDebugValue())
987 dbgs() << LIS->getInstructionIndex(UseMI) << "\t";
993 /// removeIntervalIfEmpty - Check if the live interval of a physical register
994 /// is empty, if so remove it and also remove the empty intervals of its
995 /// sub-registers. Return true if live interval is removed.
996 static bool removeIntervalIfEmpty(LiveInterval &li, LiveIntervals *LIS,
997 const TargetRegisterInfo *TRI) {
999 if (TargetRegisterInfo::isPhysicalRegister(li.reg))
1000 for (const uint16_t* SR = TRI->getSubRegisters(li.reg); *SR; ++SR) {
1001 if (!LIS->hasInterval(*SR))
1003 LiveInterval &sli = LIS->getInterval(*SR);
1005 LIS->removeInterval(*SR);
1007 LIS->removeInterval(li.reg);
1013 /// removeDeadDef - If a def of a live interval is now determined dead, remove
1014 /// the val# it defines. If the live interval becomes empty, remove it as well.
1015 bool RegisterCoalescer::removeDeadDef(LiveInterval &li, MachineInstr *DefMI) {
1016 SlotIndex DefIdx = LIS->getInstructionIndex(DefMI).getRegSlot();
1017 LiveInterval::iterator MLR = li.FindLiveRangeContaining(DefIdx);
1018 if (DefIdx != MLR->valno->def)
1020 li.removeValNo(MLR->valno);
1021 return removeIntervalIfEmpty(li, LIS, TRI);
1024 /// canJoinPhys - Return true if a copy involving a physreg should be joined.
1025 bool RegisterCoalescer::canJoinPhys(CoalescerPair &CP) {
1026 /// Always join simple intervals that are defined by a single copy from a
1027 /// reserved register. This doesn't increase register pressure, so it is
1028 /// always beneficial.
1029 if (!RegClassInfo.isReserved(CP.getDstReg())) {
1030 DEBUG(dbgs() << "\tCan only merge into reserved registers.\n");
1034 LiveInterval &JoinVInt = LIS->getInterval(CP.getSrcReg());
1035 if (CP.isFlipped() && JoinVInt.containsOneValue())
1038 DEBUG(dbgs() << "\tCannot join defs into reserved register.\n");
1042 /// joinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
1043 /// which are the src/dst of the copy instruction CopyMI. This returns true
1044 /// if the copy was successfully coalesced away. If it is not currently
1045 /// possible to coalesce this interval, but it may be possible if other
1046 /// things get coalesced, then it returns true by reference in 'Again'.
1047 bool RegisterCoalescer::joinCopy(MachineInstr *CopyMI, bool &Again) {
1050 if (JoinedCopies.count(CopyMI))
1051 return false; // Already done.
1053 DEBUG(dbgs() << LIS->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
1055 CoalescerPair CP(*TII, *TRI);
1056 if (!CP.setRegisters(CopyMI)) {
1057 DEBUG(dbgs() << "\tNot coalescable.\n");
1061 // Dead code elimination. This really should be handled by MachineDCE, but
1062 // sometimes dead copies slip through, and we can't generate invalid live
1064 if (!CP.isPhys() && CopyMI->allDefsAreDead()) {
1065 DEBUG(dbgs() << "\tCopy is dead.\n");
1066 DeadDefs.push_back(CopyMI);
1067 eliminateDeadDefs();
1071 // If they are already joined we continue.
1072 if (CP.getSrcReg() == CP.getDstReg()) {
1073 markAsJoined(CopyMI);
1074 DEBUG(dbgs() << "\tCopy already coalesced.\n");
1075 return false; // Not coalescable.
1078 // Eliminate undefs.
1079 if (!CP.isPhys() && eliminateUndefCopy(CopyMI, CP)) {
1080 markAsJoined(CopyMI);
1081 DEBUG(dbgs() << "\tEliminated copy of <undef> value.\n");
1082 return false; // Not coalescable.
1085 // Enforce policies.
1087 DEBUG(dbgs() << "\tConsidering merging " << PrintReg(CP.getSrcReg(), TRI)
1088 << " with " << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx())
1090 if (!canJoinPhys(CP)) {
1091 // Before giving up coalescing, if definition of source is defined by
1092 // trivial computation, try rematerializing it.
1093 if (!CP.isFlipped() &&
1094 reMaterializeTrivialDef(LIS->getInterval(CP.getSrcReg()), true,
1095 CP.getDstReg(), CopyMI))
1101 dbgs() << "\tConsidering merging to " << CP.getNewRC()->getName()
1103 if (CP.getDstIdx() && CP.getSrcIdx())
1104 dbgs() << PrintReg(CP.getDstReg()) << " in "
1105 << TRI->getSubRegIndexName(CP.getDstIdx()) << " and "
1106 << PrintReg(CP.getSrcReg()) << " in "
1107 << TRI->getSubRegIndexName(CP.getSrcIdx()) << '\n';
1109 dbgs() << PrintReg(CP.getSrcReg(), TRI) << " in "
1110 << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n';
1113 // When possible, let DstReg be the larger interval.
1114 if (!CP.isPartial() && LIS->getInterval(CP.getSrcReg()).ranges.size() >
1115 LIS->getInterval(CP.getDstReg()).ranges.size())
1119 // Okay, attempt to join these two intervals. On failure, this returns false.
1120 // Otherwise, if one of the intervals being joined is a physreg, this method
1121 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1122 // been modified, so we can use this information below to update aliases.
1123 if (!joinIntervals(CP)) {
1124 // Coalescing failed.
1126 // If definition of source is defined by trivial computation, try
1127 // rematerializing it.
1128 if (!CP.isFlipped() &&
1129 reMaterializeTrivialDef(LIS->getInterval(CP.getSrcReg()), true,
1130 CP.getDstReg(), CopyMI))
1133 // If we can eliminate the copy without merging the live ranges, do so now.
1134 if (!CP.isPartial()) {
1135 if (adjustCopiesBackFrom(CP, CopyMI) ||
1136 removeCopyByCommutingDef(CP, CopyMI)) {
1137 markAsJoined(CopyMI);
1138 DEBUG(dbgs() << "\tTrivial!\n");
1143 // Otherwise, we are unable to join the intervals.
1144 DEBUG(dbgs() << "\tInterference!\n");
1145 Again = true; // May be possible to coalesce later.
1149 // Coalescing to a virtual register that is of a sub-register class of the
1150 // other. Make sure the resulting register is set to the right register class.
1151 if (CP.isCrossClass()) {
1153 MRI->setRegClass(CP.getDstReg(), CP.getNewRC());
1156 // Remember to delete the copy instruction.
1157 markAsJoined(CopyMI);
1159 // Rewrite all SrcReg operands to DstReg.
1160 // Also update DstReg operands to include DstIdx if it is set.
1162 updateRegDefsUses(CP.getDstReg(), CP.getDstReg(), CP.getDstIdx());
1163 updateRegDefsUses(CP.getSrcReg(), CP.getDstReg(), CP.getSrcIdx());
1165 // SrcReg is guaranteed to be the register whose live interval that is
1167 LIS->removeInterval(CP.getSrcReg());
1169 // Update regalloc hint.
1170 TRI->UpdateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF);
1173 LiveInterval &DstInt = LIS->getInterval(CP.getDstReg());
1174 dbgs() << "\tJoined. Result = ";
1175 DstInt.print(dbgs(), TRI);
1183 /// Attempt joining with a reserved physreg.
1184 bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) {
1185 assert(CP.isPhys() && "Must be a physreg copy");
1186 assert(RegClassInfo.isReserved(CP.getDstReg()) && "Not a reserved register");
1187 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
1188 DEBUG({ dbgs() << "\t\tRHS = "; RHS.print(dbgs(), TRI); dbgs() << "\n"; });
1190 assert(CP.isFlipped() && RHS.containsOneValue() &&
1191 "Invalid join with reserved register");
1193 // Optimization for reserved registers like ESP. We can only merge with a
1194 // reserved physreg if RHS has a single value that is a copy of CP.DstReg().
1195 // The live range of the reserved register will look like a set of dead defs
1196 // - we don't properly track the live range of reserved registers.
1198 // Deny any overlapping intervals. This depends on all the reserved
1199 // register live ranges to look like dead defs.
1200 for (const uint16_t *AS = TRI->getOverlaps(CP.getDstReg()); *AS; ++AS) {
1201 if (!LIS->hasInterval(*AS)) {
1202 // Make sure at least DstReg itself exists before attempting a join.
1203 if (*AS == CP.getDstReg())
1204 LIS->getOrCreateInterval(CP.getDstReg());
1207 if (RHS.overlaps(LIS->getInterval(*AS))) {
1208 DEBUG(dbgs() << "\t\tInterference: " << PrintReg(*AS, TRI) << '\n');
1212 // Skip any value computations, we are not adding new values to the
1213 // reserved register. Also skip merging the live ranges, the reserved
1214 // register live range doesn't need to be accurate as long as all the
1219 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
1220 /// compute what the resultant value numbers for each value in the input two
1221 /// ranges will be. This is complicated by copies between the two which can
1222 /// and will commonly cause multiple value numbers to be merged into one.
1224 /// VN is the value number that we're trying to resolve. InstDefiningValue
1225 /// keeps track of the new InstDefiningValue assignment for the result
1226 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
1227 /// whether a value in this or other is a copy from the opposite set.
1228 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
1229 /// already been assigned.
1231 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
1232 /// contains the value number the copy is from.
1234 static unsigned ComputeUltimateVN(VNInfo *VNI,
1235 SmallVector<VNInfo*, 16> &NewVNInfo,
1236 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
1237 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
1238 SmallVector<int, 16> &ThisValNoAssignments,
1239 SmallVector<int, 16> &OtherValNoAssignments) {
1240 unsigned VN = VNI->id;
1242 // If the VN has already been computed, just return it.
1243 if (ThisValNoAssignments[VN] >= 0)
1244 return ThisValNoAssignments[VN];
1245 assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers");
1247 // If this val is not a copy from the other val, then it must be a new value
1248 // number in the destination.
1249 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
1250 if (I == ThisFromOther.end()) {
1251 NewVNInfo.push_back(VNI);
1252 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
1254 VNInfo *OtherValNo = I->second;
1256 // Otherwise, this *is* a copy from the RHS. If the other side has already
1257 // been computed, return it.
1258 if (OtherValNoAssignments[OtherValNo->id] >= 0)
1259 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
1261 // Mark this value number as currently being computed, then ask what the
1262 // ultimate value # of the other value is.
1263 ThisValNoAssignments[VN] = -2;
1264 unsigned UltimateVN =
1265 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
1266 OtherValNoAssignments, ThisValNoAssignments);
1267 return ThisValNoAssignments[VN] = UltimateVN;
1271 // Find out if we have something like
1274 // if so, we can pretend this is actually
1277 // which allows us to coalesce A and B.
1278 // VNI is the definition of B. LR is the life range of A that includes
1279 // the slot just before B. If we return true, we add "B = X" to DupCopies.
1280 // This implies that A dominates B.
1281 static bool RegistersDefinedFromSameValue(LiveIntervals &li,
1282 const TargetRegisterInfo &tri,
1286 SmallVector<MachineInstr*, 8> &DupCopies) {
1287 // FIXME: This is very conservative. For example, we don't handle
1288 // physical registers.
1290 MachineInstr *MI = li.getInstructionFromIndex(VNI->def);
1292 if (!MI || !MI->isFullCopy() || CP.isPartial() || CP.isPhys())
1295 unsigned Dst = MI->getOperand(0).getReg();
1296 unsigned Src = MI->getOperand(1).getReg();
1298 if (!TargetRegisterInfo::isVirtualRegister(Src) ||
1299 !TargetRegisterInfo::isVirtualRegister(Dst))
1302 unsigned A = CP.getDstReg();
1303 unsigned B = CP.getSrcReg();
1309 VNInfo *Other = LR->valno;
1310 const MachineInstr *OtherMI = li.getInstructionFromIndex(Other->def);
1312 if (!OtherMI || !OtherMI->isFullCopy())
1315 unsigned OtherDst = OtherMI->getOperand(0).getReg();
1316 unsigned OtherSrc = OtherMI->getOperand(1).getReg();
1318 if (!TargetRegisterInfo::isVirtualRegister(OtherSrc) ||
1319 !TargetRegisterInfo::isVirtualRegister(OtherDst))
1322 assert(OtherDst == B);
1324 if (Src != OtherSrc)
1327 // If the copies use two different value numbers of X, we cannot merge
1329 LiveInterval &SrcInt = li.getInterval(Src);
1330 // getVNInfoBefore returns NULL for undef copies. In this case, the
1331 // optimization is still safe.
1332 if (SrcInt.getVNInfoBefore(Other->def) != SrcInt.getVNInfoBefore(VNI->def))
1335 DupCopies.push_back(MI);
1340 /// joinIntervals - Attempt to join these two intervals. On failure, this
1342 bool RegisterCoalescer::joinIntervals(CoalescerPair &CP) {
1343 // Handle physreg joins separately.
1345 return joinReservedPhysReg(CP);
1347 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
1348 DEBUG({ dbgs() << "\t\tRHS = "; RHS.print(dbgs(), TRI); dbgs() << "\n"; });
1350 // Compute the final value assignment, assuming that the live ranges can be
1352 SmallVector<int, 16> LHSValNoAssignments;
1353 SmallVector<int, 16> RHSValNoAssignments;
1354 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
1355 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
1356 SmallVector<VNInfo*, 16> NewVNInfo;
1358 SmallVector<MachineInstr*, 8> DupCopies;
1360 LiveInterval &LHS = LIS->getOrCreateInterval(CP.getDstReg());
1361 DEBUG({ dbgs() << "\t\tLHS = "; LHS.print(dbgs(), TRI); dbgs() << "\n"; });
1363 // Loop over the value numbers of the LHS, seeing if any are defined from
1365 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1368 if (VNI->isUnused() || VNI->isPHIDef())
1370 MachineInstr *MI = LIS->getInstructionFromIndex(VNI->def);
1371 assert(MI && "Missing def");
1372 if (!MI->isCopyLike()) // Src not defined by a copy?
1375 // Figure out the value # from the RHS.
1376 LiveRange *lr = RHS.getLiveRangeContaining(VNI->def.getPrevSlot());
1377 // The copy could be to an aliased physreg.
1380 // DstReg is known to be a register in the LHS interval. If the src is
1381 // from the RHS interval, we can use its value #.
1382 if (!CP.isCoalescable(MI) &&
1383 !RegistersDefinedFromSameValue(*LIS, *TRI, CP, VNI, lr, DupCopies))
1386 LHSValsDefinedFromRHS[VNI] = lr->valno;
1389 // Loop over the value numbers of the RHS, seeing if any are defined from
1391 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1394 if (VNI->isUnused() || VNI->isPHIDef())
1396 MachineInstr *MI = LIS->getInstructionFromIndex(VNI->def);
1397 assert(MI && "Missing def");
1398 if (!MI->isCopyLike()) // Src not defined by a copy?
1401 // Figure out the value # from the LHS.
1402 LiveRange *lr = LHS.getLiveRangeContaining(VNI->def.getPrevSlot());
1403 // The copy could be to an aliased physreg.
1406 // DstReg is known to be a register in the RHS interval. If the src is
1407 // from the LHS interval, we can use its value #.
1408 if (!CP.isCoalescable(MI) &&
1409 !RegistersDefinedFromSameValue(*LIS, *TRI, CP, VNI, lr, DupCopies))
1412 RHSValsDefinedFromLHS[VNI] = lr->valno;
1415 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
1416 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
1417 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
1419 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1422 unsigned VN = VNI->id;
1423 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused())
1425 ComputeUltimateVN(VNI, NewVNInfo,
1426 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
1427 LHSValNoAssignments, RHSValNoAssignments);
1429 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1432 unsigned VN = VNI->id;
1433 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused())
1435 // If this value number isn't a copy from the LHS, it's a new number.
1436 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
1437 NewVNInfo.push_back(VNI);
1438 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
1442 ComputeUltimateVN(VNI, NewVNInfo,
1443 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
1444 RHSValNoAssignments, LHSValNoAssignments);
1447 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
1448 // interval lists to see if these intervals are coalescable.
1449 LiveInterval::const_iterator I = LHS.begin();
1450 LiveInterval::const_iterator IE = LHS.end();
1451 LiveInterval::const_iterator J = RHS.begin();
1452 LiveInterval::const_iterator JE = RHS.end();
1454 // Skip ahead until the first place of potential sharing.
1455 if (I != IE && J != JE) {
1456 if (I->start < J->start) {
1457 I = std::upper_bound(I, IE, J->start);
1458 if (I != LHS.begin()) --I;
1459 } else if (J->start < I->start) {
1460 J = std::upper_bound(J, JE, I->start);
1461 if (J != RHS.begin()) --J;
1465 while (I != IE && J != JE) {
1466 // Determine if these two live ranges overlap.
1468 if (I->start < J->start) {
1469 Overlaps = I->end > J->start;
1471 Overlaps = J->end > I->start;
1474 // If so, check value # info to determine if they are really different.
1476 // If the live range overlap will map to the same value number in the
1477 // result liverange, we can still coalesce them. If not, we can't.
1478 if (LHSValNoAssignments[I->valno->id] !=
1479 RHSValNoAssignments[J->valno->id])
1483 if (I->end < J->end)
1489 // Update kill info. Some live ranges are extended due to copy coalescing.
1490 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
1491 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
1492 VNInfo *VNI = I->first;
1493 unsigned LHSValID = LHSValNoAssignments[VNI->id];
1494 if (VNI->hasPHIKill())
1495 NewVNInfo[LHSValID]->setHasPHIKill(true);
1498 // Update kill info. Some live ranges are extended due to copy coalescing.
1499 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
1500 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
1501 VNInfo *VNI = I->first;
1502 unsigned RHSValID = RHSValNoAssignments[VNI->id];
1503 if (VNI->hasPHIKill())
1504 NewVNInfo[RHSValID]->setHasPHIKill(true);
1507 if (LHSValNoAssignments.empty())
1508 LHSValNoAssignments.push_back(-1);
1509 if (RHSValNoAssignments.empty())
1510 RHSValNoAssignments.push_back(-1);
1512 SmallVector<unsigned, 8> SourceRegisters;
1513 for (SmallVector<MachineInstr*, 8>::iterator I = DupCopies.begin(),
1514 E = DupCopies.end(); I != E; ++I) {
1515 MachineInstr *MI = *I;
1517 // We have pretended that the assignment to B in
1520 // was actually a copy from A. Now that we decided to coalesce A and B,
1521 // transform the code into
1524 // and mark the X as coalesced to keep the illusion.
1525 unsigned Src = MI->getOperand(1).getReg();
1526 SourceRegisters.push_back(Src);
1527 MI->getOperand(0).substVirtReg(Src, 0, *TRI);
1532 // If B = X was the last use of X in a liverange, we have to shrink it now
1533 // that B = X is gone.
1534 for (SmallVector<unsigned, 8>::iterator I = SourceRegisters.begin(),
1535 E = SourceRegisters.end(); I != E; ++I) {
1536 LIS->shrinkToUses(&LIS->getInterval(*I));
1539 // If we get here, we know that we can coalesce the live ranges. Ask the
1540 // intervals to coalesce themselves now.
1541 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo,
1547 // DepthMBBCompare - Comparison predicate that sort first based on the loop
1548 // depth of the basic block (the unsigned), and then on the MBB number.
1549 struct DepthMBBCompare {
1550 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
1551 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
1552 // Deeper loops first
1553 if (LHS.first != RHS.first)
1554 return LHS.first > RHS.first;
1556 // Prefer blocks that are more connected in the CFG. This takes care of
1557 // the most difficult copies first while intervals are short.
1558 unsigned cl = LHS.second->pred_size() + LHS.second->succ_size();
1559 unsigned cr = RHS.second->pred_size() + RHS.second->succ_size();
1563 // As a last resort, sort by block number.
1564 return LHS.second->getNumber() < RHS.second->getNumber();
1569 // Try joining WorkList copies starting from index From.
1570 // Null out any successful joins.
1571 bool RegisterCoalescer::copyCoalesceWorkList(unsigned From) {
1572 assert(From <= WorkList.size() && "Out of range");
1573 bool Progress = false;
1574 for (unsigned i = From, e = WorkList.size(); i != e; ++i) {
1577 // Skip instruction pointers that have already been erased, for example by
1578 // dead code elimination.
1579 if (ErasedInstrs.erase(WorkList[i])) {
1584 bool Success = joinCopy(WorkList[i], Again);
1585 Progress |= Success;
1586 if (Success || !Again)
1593 RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) {
1594 DEBUG(dbgs() << MBB->getName() << ":\n");
1596 // Collect all copy-like instructions in MBB. Don't start coalescing anything
1597 // yet, it might invalidate the iterator.
1598 const unsigned PrevSize = WorkList.size();
1599 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
1601 if (MII->isCopyLike())
1602 WorkList.push_back(MII);
1604 // Try coalescing the collected copies immediately, and remove the nulls.
1605 // This prevents the WorkList from getting too large since most copies are
1606 // joinable on the first attempt.
1607 if (copyCoalesceWorkList(PrevSize))
1608 WorkList.erase(std::remove(WorkList.begin() + PrevSize, WorkList.end(),
1609 (MachineInstr*)0), WorkList.end());
1612 void RegisterCoalescer::joinAllIntervals() {
1613 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
1614 assert(WorkList.empty() && "Old data still around.");
1616 if (Loops->empty()) {
1617 // If there are no loops in the function, join intervals in function order.
1618 for (MachineFunction::iterator I = MF->begin(), E = MF->end();
1620 copyCoalesceInMBB(I);
1622 // Otherwise, join intervals in inner loops before other intervals.
1623 // Unfortunately we can't just iterate over loop hierarchy here because
1624 // there may be more MBB's than BB's. Collect MBB's for sorting.
1626 // Join intervals in the function prolog first. We want to join physical
1627 // registers with virtual registers before the intervals got too long.
1628 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
1629 for (MachineFunction::iterator I = MF->begin(), E = MF->end();I != E;++I){
1630 MachineBasicBlock *MBB = I;
1631 MBBs.push_back(std::make_pair(Loops->getLoopDepth(MBB), I));
1634 // Sort by loop depth.
1635 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
1637 // Finally, join intervals in loop nest order.
1638 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
1639 copyCoalesceInMBB(MBBs[i].second);
1642 // Joining intervals can allow other intervals to be joined. Iteratively join
1643 // until we make no progress.
1644 while (copyCoalesceWorkList())
1648 void RegisterCoalescer::releaseMemory() {
1649 JoinedCopies.clear();
1650 ErasedInstrs.clear();
1656 bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) {
1658 MRI = &fn.getRegInfo();
1659 TM = &fn.getTarget();
1660 TRI = TM->getRegisterInfo();
1661 TII = TM->getInstrInfo();
1662 LIS = &getAnalysis<LiveIntervals>();
1663 LDV = &getAnalysis<LiveDebugVariables>();
1664 AA = &getAnalysis<AliasAnalysis>();
1665 Loops = &getAnalysis<MachineLoopInfo>();
1667 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
1668 << "********** Function: "
1669 << ((Value*)MF->getFunction())->getName() << '\n');
1671 if (VerifyCoalescing)
1672 MF->verify(this, "Before register coalescing");
1674 RegClassInfo.runOnMachineFunction(fn);
1676 // Join (coalesce) intervals if requested.
1677 if (EnableJoining) {
1680 dbgs() << "********** INTERVALS POST JOINING **********\n";
1681 for (LiveIntervals::iterator I = LIS->begin(), E = LIS->end();
1683 I->second->print(dbgs(), TRI);
1689 // Perform a final pass over the instructions and compute spill weights
1690 // and remove identity moves.
1691 SmallVector<unsigned, 4> DeadDefs, InflateRegs;
1692 for (MachineFunction::iterator mbbi = MF->begin(), mbbe = MF->end();
1693 mbbi != mbbe; ++mbbi) {
1694 MachineBasicBlock* mbb = mbbi;
1695 for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
1697 MachineInstr *MI = mii;
1698 if (JoinedCopies.count(MI)) {
1699 // Delete all coalesced copies.
1700 bool DoDelete = true;
1701 assert(MI->isCopyLike() && "Unrecognized copy instruction");
1702 unsigned SrcReg = MI->getOperand(MI->isSubregToReg() ? 2 : 1).getReg();
1703 unsigned DstReg = MI->getOperand(0).getReg();
1705 // Collect candidates for register class inflation.
1706 if (TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1707 RegClassInfo.isProperSubClass(MRI->getRegClass(SrcReg)))
1708 InflateRegs.push_back(SrcReg);
1709 if (TargetRegisterInfo::isVirtualRegister(DstReg) &&
1710 RegClassInfo.isProperSubClass(MRI->getRegClass(DstReg)))
1711 InflateRegs.push_back(DstReg);
1713 if (TargetRegisterInfo::isPhysicalRegister(SrcReg) &&
1714 MI->getNumOperands() > 2)
1715 // Do not delete extract_subreg, insert_subreg of physical
1716 // registers unless the definition is dead. e.g.
1717 // %DO<def> = INSERT_SUBREG %D0<undef>, %S0<kill>, 1
1718 // or else the scavenger may complain. LowerSubregs will
1719 // delete them later.
1722 if (MI->allDefsAreDead()) {
1723 if (TargetRegisterInfo::isVirtualRegister(SrcReg) &&
1724 LIS->hasInterval(SrcReg))
1725 LIS->shrinkToUses(&LIS->getInterval(SrcReg));
1729 // We need the instruction to adjust liveness, so make it a KILL.
1730 if (MI->isSubregToReg()) {
1731 MI->RemoveOperand(3);
1732 MI->RemoveOperand(1);
1734 MI->setDesc(TII->get(TargetOpcode::KILL));
1735 mii = llvm::next(mii);
1737 LIS->RemoveMachineInstrFromMaps(MI);
1738 mii = mbbi->erase(mii);
1744 // Now check if this is a remat'ed def instruction which is now dead.
1745 if (ReMatDefs.count(MI)) {
1747 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1748 const MachineOperand &MO = MI->getOperand(i);
1751 unsigned Reg = MO.getReg();
1754 DeadDefs.push_back(Reg);
1755 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1756 // Remat may also enable register class inflation.
1757 if (RegClassInfo.isProperSubClass(MRI->getRegClass(Reg)))
1758 InflateRegs.push_back(Reg);
1762 if (TargetRegisterInfo::isPhysicalRegister(Reg) ||
1763 !MRI->use_nodbg_empty(Reg)) {
1769 while (!DeadDefs.empty()) {
1770 unsigned DeadDef = DeadDefs.back();
1771 DeadDefs.pop_back();
1772 removeDeadDef(LIS->getInterval(DeadDef), MI);
1774 LIS->RemoveMachineInstrFromMaps(mii);
1775 mii = mbbi->erase(mii);
1783 // Check for now unnecessary kill flags.
1784 if (LIS->isNotInMIMap(MI)) continue;
1785 SlotIndex DefIdx = LIS->getInstructionIndex(MI).getRegSlot();
1786 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1787 MachineOperand &MO = MI->getOperand(i);
1788 if (!MO.isReg() || !MO.isKill()) continue;
1789 unsigned reg = MO.getReg();
1790 if (!reg || !LIS->hasInterval(reg)) continue;
1791 if (!LIS->getInterval(reg).killedAt(DefIdx)) {
1792 MO.setIsKill(false);
1795 // When leaving a kill flag on a physreg, check if any subregs should
1797 if (!TargetRegisterInfo::isPhysicalRegister(reg))
1799 for (const uint16_t *SR = TRI->getSubRegisters(reg);
1800 unsigned S = *SR; ++SR)
1801 if (LIS->hasInterval(S) && LIS->getInterval(S).liveAt(DefIdx))
1802 MI->addRegisterDefined(S, TRI);
1807 // After deleting a lot of copies, register classes may be less constrained.
1808 // Removing sub-register opreands may alow GR32_ABCD -> GR32 and DPR_VFP2 ->
1810 array_pod_sort(InflateRegs.begin(), InflateRegs.end());
1811 InflateRegs.erase(std::unique(InflateRegs.begin(), InflateRegs.end()),
1813 DEBUG(dbgs() << "Trying to inflate " << InflateRegs.size() << " regs.\n");
1814 for (unsigned i = 0, e = InflateRegs.size(); i != e; ++i) {
1815 unsigned Reg = InflateRegs[i];
1816 if (MRI->reg_nodbg_empty(Reg))
1818 if (MRI->recomputeRegClass(Reg, *TM)) {
1819 DEBUG(dbgs() << PrintReg(Reg) << " inflated to "
1820 << MRI->getRegClass(Reg)->getName() << '\n');
1827 if (VerifyCoalescing)
1828 MF->verify(this, "After register coalescing");
1832 /// print - Implement the dump method.
1833 void RegisterCoalescer::print(raw_ostream &O, const Module* m) const {