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 #include "RegisterCoalescer.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
22 #include "llvm/CodeGen/LiveRangeEdit.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineInstr.h"
25 #include "llvm/CodeGen/MachineLoopInfo.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/Passes.h"
28 #include "llvm/CodeGen/RegisterClassInfo.h"
29 #include "llvm/CodeGen/VirtRegMap.h"
30 #include "llvm/IR/Value.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/Format.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Target/TargetInstrInfo.h"
38 #include "llvm/Target/TargetMachine.h"
39 #include "llvm/Target/TargetRegisterInfo.h"
40 #include "llvm/Target/TargetSubtargetInfo.h"
45 #define DEBUG_TYPE "regalloc"
47 STATISTIC(numJoins , "Number of interval joins performed");
48 STATISTIC(numCrossRCs , "Number of cross class joins performed");
49 STATISTIC(numCommutes , "Number of instruction commuting performed");
50 STATISTIC(numExtends , "Number of copies extended");
51 STATISTIC(NumReMats , "Number of instructions re-materialized");
52 STATISTIC(NumInflated , "Number of register classes inflated");
53 STATISTIC(NumLaneConflicts, "Number of dead lane conflicts tested");
54 STATISTIC(NumLaneResolves, "Number of dead lane conflicts resolved");
57 EnableJoining("join-liveintervals",
58 cl::desc("Coalesce copies (default=true)"),
61 /// Temporary flag to test critical edge unsplitting.
63 EnableJoinSplits("join-splitedges",
64 cl::desc("Coalesce copies on split edges (default=subtarget)"), cl::Hidden);
66 /// Temporary flag to test global copy optimization.
67 static cl::opt<cl::boolOrDefault>
68 EnableGlobalCopies("join-globalcopies",
69 cl::desc("Coalesce copies that span blocks (default=subtarget)"),
70 cl::init(cl::BOU_UNSET), cl::Hidden);
73 VerifyCoalescing("verify-coalescing",
74 cl::desc("Verify machine instrs before and after register coalescing"),
78 class RegisterCoalescer : public MachineFunctionPass,
79 private LiveRangeEdit::Delegate {
81 MachineRegisterInfo* MRI;
82 const TargetMachine* TM;
83 const TargetRegisterInfo* TRI;
84 const TargetInstrInfo* TII;
86 const MachineLoopInfo* Loops;
88 RegisterClassInfo RegClassInfo;
90 /// A LaneMask to remember on which subregister live ranges we need to call
91 /// shrinkToUses() later.
94 /// True if the main range of the currently coalesced intervals should be
95 /// checked for smaller live intervals.
98 /// \brief True if the coalescer should aggressively coalesce global copies
99 /// in favor of keeping local copies.
100 bool JoinGlobalCopies;
102 /// \brief True if the coalescer should aggressively coalesce fall-thru
103 /// blocks exclusively containing copies.
106 /// Copy instructions yet to be coalesced.
107 SmallVector<MachineInstr*, 8> WorkList;
108 SmallVector<MachineInstr*, 8> LocalWorkList;
110 /// Set of instruction pointers that have been erased, and
111 /// that may be present in WorkList.
112 SmallPtrSet<MachineInstr*, 8> ErasedInstrs;
114 /// Dead instructions that are about to be deleted.
115 SmallVector<MachineInstr*, 8> DeadDefs;
117 /// Virtual registers to be considered for register class inflation.
118 SmallVector<unsigned, 8> InflateRegs;
120 /// Recursively eliminate dead defs in DeadDefs.
121 void eliminateDeadDefs();
123 /// LiveRangeEdit callback for eliminateDeadDefs().
124 void LRE_WillEraseInstruction(MachineInstr *MI) override;
126 /// Coalesce the LocalWorkList.
127 void coalesceLocals();
129 /// Join compatible live intervals
130 void joinAllIntervals();
132 /// Coalesce copies in the specified MBB, putting
133 /// copies that cannot yet be coalesced into WorkList.
134 void copyCoalesceInMBB(MachineBasicBlock *MBB);
136 /// Tries to coalesce all copies in CurrList. Returns true if any progress
138 bool copyCoalesceWorkList(MutableArrayRef<MachineInstr*> CurrList);
140 /// Attempt to join intervals corresponding to SrcReg/DstReg, which are the
141 /// src/dst of the copy instruction CopyMI. This returns true if the copy
142 /// was successfully coalesced away. If it is not currently possible to
143 /// coalesce this interval, but it may be possible if other things get
144 /// coalesced, then it returns true by reference in 'Again'.
145 bool joinCopy(MachineInstr *TheCopy, bool &Again);
147 /// Attempt to join these two intervals. On failure, this
148 /// returns false. The output "SrcInt" will not have been modified, so we
149 /// can use this information below to update aliases.
150 bool joinIntervals(CoalescerPair &CP);
152 /// Attempt joining two virtual registers. Return true on success.
153 bool joinVirtRegs(CoalescerPair &CP);
155 /// Attempt joining with a reserved physreg.
156 bool joinReservedPhysReg(CoalescerPair &CP);
158 /// Add the LiveRange @p ToMerge as a subregister liverange of @p LI.
159 /// Subranges in @p LI which only partially interfere with the desired
160 /// LaneMask are split as necessary. @p LaneMask are the lanes that
161 /// @p ToMerge will occupy in the coalescer register. @p LI has its subrange
162 /// lanemasks already adjusted to the coalesced register.
163 /// @returns false if live range conflicts couldn't get resolved.
164 bool mergeSubRangeInto(LiveInterval &LI, const LiveRange &ToMerge,
165 unsigned LaneMask, CoalescerPair &CP);
167 /// Join the liveranges of two subregisters. Joins @p RRange into
168 /// @p LRange, @p RRange may be invalid afterwards.
169 /// @returns false if live range conflicts couldn't get resolved.
170 bool joinSubRegRanges(LiveRange &LRange, LiveRange &RRange,
171 unsigned LaneMask, const CoalescerPair &CP);
173 /// We found a non-trivially-coalescable copy. If the source value number is
174 /// defined by a copy from the destination reg see if we can merge these two
175 /// destination reg valno# into a single value number, eliminating a copy.
176 /// This returns true if an interval was modified.
177 bool adjustCopiesBackFrom(const CoalescerPair &CP, MachineInstr *CopyMI);
179 /// Return true if there are definitions of IntB
180 /// other than BValNo val# that can reach uses of AValno val# of IntA.
181 bool hasOtherReachingDefs(LiveInterval &IntA, LiveInterval &IntB,
182 VNInfo *AValNo, VNInfo *BValNo);
184 /// We found a non-trivially-coalescable copy.
185 /// If the source value number is defined by a commutable instruction and
186 /// its other operand is coalesced to the copy dest register, see if we
187 /// can transform the copy into a noop by commuting the definition.
188 /// This returns true if an interval was modified.
189 bool removeCopyByCommutingDef(const CoalescerPair &CP,MachineInstr *CopyMI);
191 /// If the source of a copy is defined by a
192 /// trivial computation, replace the copy by rematerialize the definition.
193 bool reMaterializeTrivialDef(CoalescerPair &CP, MachineInstr *CopyMI,
196 /// Return true if a copy involving a physreg should be joined.
197 bool canJoinPhys(const CoalescerPair &CP);
199 /// Replace all defs and uses of SrcReg to DstReg and update the subregister
200 /// number if it is not zero. If DstReg is a physical register and the
201 /// existing subregister number of the def / use being updated is not zero,
202 /// make sure to set it to the correct physical subregister.
203 void updateRegDefsUses(unsigned SrcReg, unsigned DstReg, unsigned SubIdx);
205 /// Handle copies of undef values.
206 /// Returns true if @p CopyMI was a copy of an undef value and eliminated.
207 bool eliminateUndefCopy(MachineInstr *CopyMI);
210 static char ID; ///< Class identification, replacement for typeinfo
211 RegisterCoalescer() : MachineFunctionPass(ID) {
212 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
215 void getAnalysisUsage(AnalysisUsage &AU) const override;
217 void releaseMemory() override;
219 /// This is the pass entry point.
220 bool runOnMachineFunction(MachineFunction&) override;
222 /// Implement the dump method.
223 void print(raw_ostream &O, const Module* = nullptr) const override;
225 } // end anonymous namespace
227 char &llvm::RegisterCoalescerID = RegisterCoalescer::ID;
229 INITIALIZE_PASS_BEGIN(RegisterCoalescer, "simple-register-coalescing",
230 "Simple Register Coalescing", false, false)
231 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
232 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
233 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
234 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
235 INITIALIZE_PASS_END(RegisterCoalescer, "simple-register-coalescing",
236 "Simple Register Coalescing", false, false)
238 char RegisterCoalescer::ID = 0;
240 static bool isMoveInstr(const TargetRegisterInfo &tri, const MachineInstr *MI,
241 unsigned &Src, unsigned &Dst,
242 unsigned &SrcSub, unsigned &DstSub) {
244 Dst = MI->getOperand(0).getReg();
245 DstSub = MI->getOperand(0).getSubReg();
246 Src = MI->getOperand(1).getReg();
247 SrcSub = MI->getOperand(1).getSubReg();
248 } else if (MI->isSubregToReg()) {
249 Dst = MI->getOperand(0).getReg();
250 DstSub = tri.composeSubRegIndices(MI->getOperand(0).getSubReg(),
251 MI->getOperand(3).getImm());
252 Src = MI->getOperand(2).getReg();
253 SrcSub = MI->getOperand(2).getSubReg();
259 /// Return true if this block should be vacated by the coalescer to eliminate
260 /// branches. The important cases to handle in the coalescer are critical edges
261 /// split during phi elimination which contain only copies. Simple blocks that
262 /// contain non-branches should also be vacated, but this can be handled by an
263 /// earlier pass similar to early if-conversion.
264 static bool isSplitEdge(const MachineBasicBlock *MBB) {
265 if (MBB->pred_size() != 1 || MBB->succ_size() != 1)
268 for (const auto &MI : *MBB) {
269 if (!MI.isCopyLike() && !MI.isUnconditionalBranch())
275 bool CoalescerPair::setRegisters(const MachineInstr *MI) {
279 Flipped = CrossClass = false;
281 unsigned Src, Dst, SrcSub, DstSub;
282 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
284 Partial = SrcSub || DstSub;
286 // If one register is a physreg, it must be Dst.
287 if (TargetRegisterInfo::isPhysicalRegister(Src)) {
288 if (TargetRegisterInfo::isPhysicalRegister(Dst))
291 std::swap(SrcSub, DstSub);
295 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
297 if (TargetRegisterInfo::isPhysicalRegister(Dst)) {
298 // Eliminate DstSub on a physreg.
300 Dst = TRI.getSubReg(Dst, DstSub);
301 if (!Dst) return false;
305 // Eliminate SrcSub by picking a corresponding Dst superregister.
307 Dst = TRI.getMatchingSuperReg(Dst, SrcSub, MRI.getRegClass(Src));
308 if (!Dst) return false;
309 } else if (!MRI.getRegClass(Src)->contains(Dst)) {
313 // Both registers are virtual.
314 const TargetRegisterClass *SrcRC = MRI.getRegClass(Src);
315 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
317 // Both registers have subreg indices.
318 if (SrcSub && DstSub) {
319 // Copies between different sub-registers are never coalescable.
320 if (Src == Dst && SrcSub != DstSub)
323 NewRC = TRI.getCommonSuperRegClass(SrcRC, SrcSub, DstRC, DstSub,
328 // SrcReg will be merged with a sub-register of DstReg.
330 NewRC = TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSub);
332 // DstReg will be merged with a sub-register of SrcReg.
334 NewRC = TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSub);
336 // This is a straight copy without sub-registers.
337 NewRC = TRI.getCommonSubClass(DstRC, SrcRC);
340 // The combined constraint may be impossible to satisfy.
344 // Prefer SrcReg to be a sub-register of DstReg.
345 // FIXME: Coalescer should support subregs symmetrically.
346 if (DstIdx && !SrcIdx) {
348 std::swap(SrcIdx, DstIdx);
352 CrossClass = NewRC != DstRC || NewRC != SrcRC;
354 // Check our invariants
355 assert(TargetRegisterInfo::isVirtualRegister(Src) && "Src must be virtual");
356 assert(!(TargetRegisterInfo::isPhysicalRegister(Dst) && DstSub) &&
357 "Cannot have a physical SubIdx");
363 bool CoalescerPair::flip() {
364 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
366 std::swap(SrcReg, DstReg);
367 std::swap(SrcIdx, DstIdx);
372 bool CoalescerPair::isCoalescable(const MachineInstr *MI) const {
375 unsigned Src, Dst, SrcSub, DstSub;
376 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
379 // Find the virtual register that is SrcReg.
382 std::swap(SrcSub, DstSub);
383 } else if (Src != SrcReg) {
387 // Now check that Dst matches DstReg.
388 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
389 if (!TargetRegisterInfo::isPhysicalRegister(Dst))
391 assert(!DstIdx && !SrcIdx && "Inconsistent CoalescerPair state.");
392 // DstSub could be set for a physreg from INSERT_SUBREG.
394 Dst = TRI.getSubReg(Dst, DstSub);
397 return DstReg == Dst;
398 // This is a partial register copy. Check that the parts match.
399 return TRI.getSubReg(DstReg, SrcSub) == Dst;
401 // DstReg is virtual.
404 // Registers match, do the subregisters line up?
405 return TRI.composeSubRegIndices(SrcIdx, SrcSub) ==
406 TRI.composeSubRegIndices(DstIdx, DstSub);
410 void RegisterCoalescer::getAnalysisUsage(AnalysisUsage &AU) const {
411 AU.setPreservesCFG();
412 AU.addRequired<AliasAnalysis>();
413 AU.addRequired<LiveIntervals>();
414 AU.addPreserved<LiveIntervals>();
415 AU.addPreserved<SlotIndexes>();
416 AU.addRequired<MachineLoopInfo>();
417 AU.addPreserved<MachineLoopInfo>();
418 AU.addPreservedID(MachineDominatorsID);
419 MachineFunctionPass::getAnalysisUsage(AU);
422 void RegisterCoalescer::eliminateDeadDefs() {
423 SmallVector<unsigned, 8> NewRegs;
424 LiveRangeEdit(nullptr, NewRegs, *MF, *LIS,
425 nullptr, this).eliminateDeadDefs(DeadDefs);
428 void RegisterCoalescer::LRE_WillEraseInstruction(MachineInstr *MI) {
429 // MI may be in WorkList. Make sure we don't visit it.
430 ErasedInstrs.insert(MI);
433 bool RegisterCoalescer::adjustCopiesBackFrom(const CoalescerPair &CP,
434 MachineInstr *CopyMI) {
435 assert(!CP.isPartial() && "This doesn't work for partial copies.");
436 assert(!CP.isPhys() && "This doesn't work for physreg copies.");
439 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
441 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
442 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
444 // We have a non-trivially-coalescable copy with IntA being the source and
445 // IntB being the dest, thus this defines a value number in IntB. If the
446 // source value number (in IntA) is defined by a copy from B, see if we can
447 // merge these two pieces of B into a single value number, eliminating a copy.
452 // B1 = A3 <- this copy
454 // In this case, B0 can be extended to where the B1 copy lives, allowing the
455 // B1 value number to be replaced with B0 (which simplifies the B
458 // BValNo is a value number in B that is defined by a copy from A. 'B1' in
459 // the example above.
460 LiveInterval::iterator BS = IntB.FindSegmentContaining(CopyIdx);
461 if (BS == IntB.end()) return false;
462 VNInfo *BValNo = BS->valno;
464 // Get the location that B is defined at. Two options: either this value has
465 // an unknown definition point or it is defined at CopyIdx. If unknown, we
467 if (BValNo->def != CopyIdx) return false;
469 // AValNo is the value number in A that defines the copy, A3 in the example.
470 SlotIndex CopyUseIdx = CopyIdx.getRegSlot(true);
471 LiveInterval::iterator AS = IntA.FindSegmentContaining(CopyUseIdx);
472 // The live segment might not exist after fun with physreg coalescing.
473 if (AS == IntA.end()) return false;
474 VNInfo *AValNo = AS->valno;
476 // If AValNo is defined as a copy from IntB, we can potentially process this.
477 // Get the instruction that defines this value number.
478 MachineInstr *ACopyMI = LIS->getInstructionFromIndex(AValNo->def);
479 // Don't allow any partial copies, even if isCoalescable() allows them.
480 if (!CP.isCoalescable(ACopyMI) || !ACopyMI->isFullCopy())
483 // Get the Segment in IntB that this value number starts with.
484 LiveInterval::iterator ValS =
485 IntB.FindSegmentContaining(AValNo->def.getPrevSlot());
486 if (ValS == IntB.end())
489 // Make sure that the end of the live segment is inside the same block as
491 MachineInstr *ValSEndInst =
492 LIS->getInstructionFromIndex(ValS->end.getPrevSlot());
493 if (!ValSEndInst || ValSEndInst->getParent() != CopyMI->getParent())
496 // Okay, we now know that ValS ends in the same block that the CopyMI
497 // live-range starts. If there are no intervening live segments between them
498 // in IntB, we can merge them.
499 if (ValS+1 != BS) return false;
501 DEBUG(dbgs() << "Extending: " << PrintReg(IntB.reg, TRI));
503 SlotIndex FillerStart = ValS->end, FillerEnd = BS->start;
504 // We are about to delete CopyMI, so need to remove it as the 'instruction
505 // that defines this value #'. Update the valnum with the new defining
507 BValNo->def = FillerStart;
509 // Okay, we can merge them. We need to insert a new liverange:
510 // [ValS.end, BS.begin) of either value number, then we merge the
511 // two value numbers.
512 IntB.addSegment(LiveInterval::Segment(FillerStart, FillerEnd, BValNo));
514 // Okay, merge "B1" into the same value number as "B0".
515 if (BValNo != ValS->valno)
516 IntB.MergeValueNumberInto(BValNo, ValS->valno);
518 // Do the same for the subregister segments.
519 for (LiveInterval::SubRange &S : IntB.subranges()) {
520 VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
521 S.addSegment(LiveInterval::Segment(FillerStart, FillerEnd, SubBValNo));
522 VNInfo *SubValSNo = S.getVNInfoAt(AValNo->def.getPrevSlot());
523 if (SubBValNo != SubValSNo)
524 S.MergeValueNumberInto(SubBValNo, SubValSNo);
527 DEBUG(dbgs() << " result = " << IntB << '\n');
529 // If the source instruction was killing the source register before the
530 // merge, unset the isKill marker given the live range has been extended.
531 int UIdx = ValSEndInst->findRegisterUseOperandIdx(IntB.reg, true);
533 ValSEndInst->getOperand(UIdx).setIsKill(false);
536 // Rewrite the copy. If the copy instruction was killing the destination
537 // register before the merge, find the last use and trim the live range. That
538 // will also add the isKill marker.
539 CopyMI->substituteRegister(IntA.reg, IntB.reg, 0, *TRI);
540 if (AS->end == CopyIdx)
541 LIS->shrinkToUses(&IntA);
547 bool RegisterCoalescer::hasOtherReachingDefs(LiveInterval &IntA,
551 // If AValNo has PHI kills, conservatively assume that IntB defs can reach
553 if (LIS->hasPHIKill(IntA, AValNo))
556 for (LiveRange::Segment &ASeg : IntA.segments) {
557 if (ASeg.valno != AValNo) continue;
558 LiveInterval::iterator BI =
559 std::upper_bound(IntB.begin(), IntB.end(), ASeg.start);
560 if (BI != IntB.begin())
562 for (; BI != IntB.end() && ASeg.end >= BI->start; ++BI) {
563 if (BI->valno == BValNo)
565 if (BI->start <= ASeg.start && BI->end > ASeg.start)
567 if (BI->start > ASeg.start && BI->start < ASeg.end)
574 /// Copy segements with value number @p SrcValNo from liverange @p Src to live
575 /// range @Dst and use value number @p DstValNo there.
576 static void addSegmentsWithValNo(LiveRange &Dst, VNInfo *DstValNo,
577 const LiveRange &Src, const VNInfo *SrcValNo)
579 for (const LiveRange::Segment &S : Src.segments) {
580 if (S.valno != SrcValNo)
582 Dst.addSegment(LiveRange::Segment(S.start, S.end, DstValNo));
586 bool RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP,
587 MachineInstr *CopyMI) {
588 assert(!CP.isPhys());
591 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
593 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
595 // We found a non-trivially-coalescable copy with IntA being the source and
596 // IntB being the dest, thus this defines a value number in IntB. If the
597 // source value number (in IntA) is defined by a commutable instruction and
598 // its other operand is coalesced to the copy dest register, see if we can
599 // transform the copy into a noop by commuting the definition. For example,
601 // A3 = op A2 B0<kill>
603 // B1 = A3 <- this copy
605 // = op A3 <- more uses
609 // B2 = op B0 A2<kill>
611 // B1 = B2 <- now an identity copy
613 // = op B2 <- more uses
615 // BValNo is a value number in B that is defined by a copy from A. 'B1' in
616 // the example above.
617 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
618 VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx);
619 assert(BValNo != nullptr && BValNo->def == CopyIdx);
621 // AValNo is the value number in A that defines the copy, A3 in the example.
622 VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getRegSlot(true));
623 assert(AValNo && !AValNo->isUnused() && "COPY source not live");
624 if (AValNo->isPHIDef())
626 MachineInstr *DefMI = LIS->getInstructionFromIndex(AValNo->def);
629 if (!DefMI->isCommutable())
631 // If DefMI is a two-address instruction then commuting it will change the
632 // destination register.
633 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
634 assert(DefIdx != -1);
636 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
638 unsigned Op1, Op2, NewDstIdx;
639 if (!TII->findCommutedOpIndices(DefMI, Op1, Op2))
643 else if (Op2 == UseOpIdx)
648 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
649 unsigned NewReg = NewDstMO.getReg();
650 if (NewReg != IntB.reg || !IntB.Query(AValNo->def).isKill())
653 // Make sure there are no other definitions of IntB that would reach the
654 // uses which the new definition can reach.
655 if (hasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
658 // If some of the uses of IntA.reg is already coalesced away, return false.
659 // It's not possible to determine whether it's safe to perform the coalescing.
660 for (MachineOperand &MO : MRI->use_nodbg_operands(IntA.reg)) {
661 MachineInstr *UseMI = MO.getParent();
662 unsigned OpNo = &MO - &UseMI->getOperand(0);
663 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI);
664 LiveInterval::iterator US = IntA.FindSegmentContaining(UseIdx);
665 if (US == IntA.end() || US->valno != AValNo)
667 // If this use is tied to a def, we can't rewrite the register.
668 if (UseMI->isRegTiedToDefOperand(OpNo))
672 DEBUG(dbgs() << "\tremoveCopyByCommutingDef: " << AValNo->def << '\t'
675 // At this point we have decided that it is legal to do this
676 // transformation. Start by commuting the instruction.
677 MachineBasicBlock *MBB = DefMI->getParent();
678 MachineInstr *NewMI = TII->commuteInstruction(DefMI);
681 if (TargetRegisterInfo::isVirtualRegister(IntA.reg) &&
682 TargetRegisterInfo::isVirtualRegister(IntB.reg) &&
683 !MRI->constrainRegClass(IntB.reg, MRI->getRegClass(IntA.reg)))
685 if (NewMI != DefMI) {
686 LIS->ReplaceMachineInstrInMaps(DefMI, NewMI);
687 MachineBasicBlock::iterator Pos = DefMI;
688 MBB->insert(Pos, NewMI);
692 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
701 // Update uses of IntA of the specific Val# with IntB.
702 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(IntA.reg),
704 UI != UE; /* ++UI is below because of possible MI removal */) {
705 MachineOperand &UseMO = *UI;
709 MachineInstr *UseMI = UseMO.getParent();
710 if (UseMI->isDebugValue()) {
711 // FIXME These don't have an instruction index. Not clear we have enough
712 // info to decide whether to do this replacement or not. For now do it.
713 UseMO.setReg(NewReg);
716 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI).getRegSlot(true);
717 LiveInterval::iterator US = IntA.FindSegmentContaining(UseIdx);
718 assert(US != IntA.end() && "Use must be live");
719 if (US->valno != AValNo)
721 // Kill flags are no longer accurate. They are recomputed after RA.
722 UseMO.setIsKill(false);
723 if (TargetRegisterInfo::isPhysicalRegister(NewReg))
724 UseMO.substPhysReg(NewReg, *TRI);
726 UseMO.setReg(NewReg);
729 if (!UseMI->isCopy())
731 if (UseMI->getOperand(0).getReg() != IntB.reg ||
732 UseMI->getOperand(0).getSubReg())
735 // This copy will become a noop. If it's defining a new val#, merge it into
737 SlotIndex DefIdx = UseIdx.getRegSlot();
738 VNInfo *DVNI = IntB.getVNInfoAt(DefIdx);
741 DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI);
742 assert(DVNI->def == DefIdx);
743 BValNo = IntB.MergeValueNumberInto(DVNI, BValNo);
744 for (LiveInterval::SubRange &S : IntB.subranges()) {
745 VNInfo *SubDVNI = S.getVNInfoAt(DefIdx);
748 VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
749 assert(SubBValNo->def == CopyIdx);
750 S.MergeValueNumberInto(SubDVNI, SubBValNo);
753 ErasedInstrs.insert(UseMI);
754 LIS->RemoveMachineInstrFromMaps(UseMI);
755 UseMI->eraseFromParent();
758 // Extend BValNo by merging in IntA live segments of AValNo. Val# definition
760 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
761 if (IntB.hasSubRanges()) {
762 if (!IntA.hasSubRanges()) {
763 unsigned Mask = MRI->getMaxLaneMaskForVReg(IntA.reg);
764 IntA.createSubRangeFrom(Allocator, Mask, IntA);
766 SlotIndex AIdx = CopyIdx.getRegSlot(true);
767 for (LiveInterval::SubRange &SA : IntA.subranges()) {
768 VNInfo *ASubValNo = SA.getVNInfoAt(AIdx);
769 assert(ASubValNo != nullptr);
771 unsigned AMask = SA.LaneMask;
772 for (LiveInterval::SubRange &SB : IntB.subranges()) {
773 unsigned BMask = SB.LaneMask;
774 unsigned Common = BMask & AMask;
779 dbgs() << format("\t\tCopy+Merge %04X into %04X\n", BMask, Common));
780 unsigned BRest = BMask & ~AMask;
781 LiveInterval::SubRange *CommonRange;
784 DEBUG(dbgs() << format("\t\tReduce Lane to %04X\n", BRest));
785 // Duplicate SubRange for newly merged common stuff.
786 CommonRange = IntB.createSubRangeFrom(Allocator, Common, SB);
788 // We van reuse the L SubRange.
789 SB.LaneMask = Common;
792 LiveRange RangeCopy(SB, Allocator);
794 VNInfo *BSubValNo = CommonRange->getVNInfoAt(CopyIdx);
795 assert(BSubValNo->def == CopyIdx);
796 BSubValNo->def = ASubValNo->def;
797 addSegmentsWithValNo(*CommonRange, BSubValNo, SA, ASubValNo);
801 DEBUG(dbgs() << format("\t\tNew Lane %04X\n", AMask));
802 LiveRange *NewRange = IntB.createSubRange(Allocator, AMask);
803 VNInfo *BSubValNo = NewRange->getNextValue(CopyIdx, Allocator);
804 addSegmentsWithValNo(*NewRange, BSubValNo, SA, ASubValNo);
809 BValNo->def = AValNo->def;
810 addSegmentsWithValNo(IntB, BValNo, IntA, AValNo);
811 DEBUG(dbgs() << "\t\textended: " << IntB << '\n');
813 LIS->removeVRegDefAt(IntA, AValNo->def);
815 DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n');
820 /// Returns true if @p MI defines the full vreg @p Reg, as opposed to just
821 /// defining a subregister.
822 static bool definesFullReg(const MachineInstr &MI, unsigned Reg) {
823 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) &&
824 "This code cannot handle physreg aliasing");
825 for (const MachineOperand &Op : MI.operands()) {
826 if (!Op.isReg() || !Op.isDef() || Op.getReg() != Reg)
828 // Return true if we define the full register or don't care about the value
829 // inside other subregisters.
830 if (Op.getSubReg() == 0 || Op.isUndef())
836 bool RegisterCoalescer::reMaterializeTrivialDef(CoalescerPair &CP,
837 MachineInstr *CopyMI,
840 unsigned SrcReg = CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg();
841 unsigned SrcIdx = CP.isFlipped() ? CP.getDstIdx() : CP.getSrcIdx();
842 unsigned DstReg = CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg();
843 unsigned DstIdx = CP.isFlipped() ? CP.getSrcIdx() : CP.getDstIdx();
844 if (TargetRegisterInfo::isPhysicalRegister(SrcReg))
847 LiveInterval &SrcInt = LIS->getInterval(SrcReg);
848 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI);
849 VNInfo *ValNo = SrcInt.Query(CopyIdx).valueIn();
850 assert(ValNo && "CopyMI input register not live");
851 if (ValNo->isPHIDef() || ValNo->isUnused())
853 MachineInstr *DefMI = LIS->getInstructionFromIndex(ValNo->def);
856 if (DefMI->isCopyLike()) {
860 if (!TII->isAsCheapAsAMove(DefMI))
862 if (!TII->isTriviallyReMaterializable(DefMI, AA))
864 if (!definesFullReg(*DefMI, SrcReg))
866 bool SawStore = false;
867 if (!DefMI->isSafeToMove(TII, AA, SawStore))
869 const MCInstrDesc &MCID = DefMI->getDesc();
870 if (MCID.getNumDefs() != 1)
872 // Only support subregister destinations when the def is read-undef.
873 MachineOperand &DstOperand = CopyMI->getOperand(0);
874 unsigned CopyDstReg = DstOperand.getReg();
875 if (DstOperand.getSubReg() && !DstOperand.isUndef())
878 // If both SrcIdx and DstIdx are set, correct rematerialization would widen
879 // the register substantially (beyond both source and dest size). This is bad
880 // for performance since it can cascade through a function, introducing many
881 // extra spills and fills (e.g. ARM can easily end up copying QQQQPR registers
882 // around after a few subreg copies).
883 if (SrcIdx && DstIdx)
886 const TargetRegisterClass *DefRC = TII->getRegClass(MCID, 0, TRI, *MF);
887 if (!DefMI->isImplicitDef()) {
888 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
889 unsigned NewDstReg = DstReg;
891 unsigned NewDstIdx = TRI->composeSubRegIndices(CP.getSrcIdx(),
892 DefMI->getOperand(0).getSubReg());
894 NewDstReg = TRI->getSubReg(DstReg, NewDstIdx);
896 // Finally, make sure that the physical subregister that will be
897 // constructed later is permitted for the instruction.
898 if (!DefRC->contains(NewDstReg))
901 // Theoretically, some stack frame reference could exist. Just make sure
902 // it hasn't actually happened.
903 assert(TargetRegisterInfo::isVirtualRegister(DstReg) &&
904 "Only expect to deal with virtual or physical registers");
908 MachineBasicBlock *MBB = CopyMI->getParent();
909 MachineBasicBlock::iterator MII =
910 std::next(MachineBasicBlock::iterator(CopyMI));
911 TII->reMaterialize(*MBB, MII, DstReg, SrcIdx, DefMI, *TRI);
912 MachineInstr *NewMI = std::prev(MII);
914 LIS->ReplaceMachineInstrInMaps(CopyMI, NewMI);
915 CopyMI->eraseFromParent();
916 ErasedInstrs.insert(CopyMI);
918 // NewMI may have dead implicit defs (E.g. EFLAGS for MOV<bits>r0 on X86).
919 // We need to remember these so we can add intervals once we insert
920 // NewMI into SlotIndexes.
921 SmallVector<unsigned, 4> NewMIImplDefs;
922 for (unsigned i = NewMI->getDesc().getNumOperands(),
923 e = NewMI->getNumOperands(); i != e; ++i) {
924 MachineOperand &MO = NewMI->getOperand(i);
926 assert(MO.isDef() && MO.isImplicit() && MO.isDead() &&
927 TargetRegisterInfo::isPhysicalRegister(MO.getReg()));
928 NewMIImplDefs.push_back(MO.getReg());
932 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
933 const TargetRegisterClass *NewRC = CP.getNewRC();
934 unsigned NewIdx = NewMI->getOperand(0).getSubReg();
936 if (DefRC != nullptr) {
938 NewRC = TRI->getMatchingSuperRegClass(NewRC, DefRC, NewIdx);
940 NewRC = TRI->getCommonSubClass(NewRC, DefRC);
941 assert(NewRC && "subreg chosen for remat incompatible with instruction");
943 MRI->setRegClass(DstReg, NewRC);
945 updateRegDefsUses(DstReg, DstReg, DstIdx);
946 NewMI->getOperand(0).setSubReg(NewIdx);
947 } else if (NewMI->getOperand(0).getReg() != CopyDstReg) {
948 // The New instruction may be defining a sub-register of what's actually
949 // been asked for. If so it must implicitly define the whole thing.
950 assert(TargetRegisterInfo::isPhysicalRegister(DstReg) &&
951 "Only expect virtual or physical registers in remat");
952 NewMI->getOperand(0).setIsDead(true);
953 NewMI->addOperand(MachineOperand::CreateReg(CopyDstReg,
957 // Record small dead def live-ranges for all the subregisters
958 // of the destination register.
959 // Otherwise, variables that live through may miss some
960 // interferences, thus creating invalid allocation.
962 // vreg1 = somedef ; vreg1 GR8
963 // vreg2 = remat ; vreg2 GR32
964 // CL = COPY vreg2.sub_8bit
965 // = somedef vreg1 ; vreg1 GR8
967 // vreg1 = somedef ; vreg1 GR8
968 // ECX<def, dead> = remat ; CL<imp-def>
969 // = somedef vreg1 ; vreg1 GR8
970 // vreg1 will see the inteferences with CL but not with CH since
971 // no live-ranges would have been created for ECX.
973 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
974 for (MCRegUnitIterator Units(NewMI->getOperand(0).getReg(), TRI);
975 Units.isValid(); ++Units)
976 if (LiveRange *LR = LIS->getCachedRegUnit(*Units))
977 LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
980 if (NewMI->getOperand(0).getSubReg())
981 NewMI->getOperand(0).setIsUndef();
983 // CopyMI may have implicit operands, transfer them over to the newly
984 // rematerialized instruction. And update implicit def interval valnos.
985 for (unsigned i = CopyMI->getDesc().getNumOperands(),
986 e = CopyMI->getNumOperands(); i != e; ++i) {
987 MachineOperand &MO = CopyMI->getOperand(i);
989 assert(MO.isImplicit() && "No explicit operands after implict operands.");
990 // Discard VReg implicit defs.
991 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
992 NewMI->addOperand(MO);
997 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
998 for (unsigned i = 0, e = NewMIImplDefs.size(); i != e; ++i) {
999 unsigned Reg = NewMIImplDefs[i];
1000 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units)
1001 if (LiveRange *LR = LIS->getCachedRegUnit(*Units))
1002 LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
1005 DEBUG(dbgs() << "Remat: " << *NewMI);
1008 // The source interval can become smaller because we removed a use.
1009 LIS->shrinkToUses(&SrcInt, &DeadDefs);
1010 if (!DeadDefs.empty()) {
1011 // If the virtual SrcReg is completely eliminated, update all DBG_VALUEs
1012 // to describe DstReg instead.
1013 for (MachineOperand &UseMO : MRI->use_operands(SrcReg)) {
1014 MachineInstr *UseMI = UseMO.getParent();
1015 if (UseMI->isDebugValue()) {
1016 UseMO.setReg(DstReg);
1017 DEBUG(dbgs() << "\t\tupdated: " << *UseMI);
1020 eliminateDeadDefs();
1026 bool RegisterCoalescer::eliminateUndefCopy(MachineInstr *CopyMI) {
1027 // ProcessImpicitDefs may leave some copies of <undef> values, it only removes
1028 // local variables. When we have a copy like:
1030 // %vreg1 = COPY %vreg2<undef>
1032 // We delete the copy and remove the corresponding value number from %vreg1.
1033 // Any uses of that value number are marked as <undef>.
1035 // Note that we do not query CoalescerPair here but redo isMoveInstr as the
1036 // CoalescerPair may have a new register class with adjusted subreg indices
1038 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
1039 isMoveInstr(*TRI, CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx);
1041 SlotIndex Idx = LIS->getInstructionIndex(CopyMI);
1042 const LiveInterval &SrcLI = LIS->getInterval(SrcReg);
1043 // CopyMI is undef iff SrcReg is not live before the instruction.
1044 if (SrcSubIdx != 0 && SrcLI.hasSubRanges()) {
1045 unsigned SrcMask = TRI->getSubRegIndexLaneMask(SrcSubIdx);
1046 for (const LiveInterval::SubRange &SR : SrcLI.subranges()) {
1047 if ((SR.LaneMask & SrcMask) == 0)
1052 } else if (SrcLI.liveAt(Idx))
1055 DEBUG(dbgs() << "\tEliminating copy of <undef> value\n");
1057 // Remove any DstReg segments starting at the instruction.
1058 LiveInterval &DstLI = LIS->getInterval(DstReg);
1059 SlotIndex RegIndex = Idx.getRegSlot();
1060 // Remove value or merge with previous one in case of a subregister def.
1061 if (VNInfo *PrevVNI = DstLI.getVNInfoAt(Idx)) {
1062 VNInfo *VNI = DstLI.getVNInfoAt(RegIndex);
1063 DstLI.MergeValueNumberInto(VNI, PrevVNI);
1065 // The affected subregister segments can be removed.
1066 unsigned DstMask = TRI->getSubRegIndexLaneMask(DstSubIdx);
1067 for (LiveInterval::SubRange &SR : DstLI.subranges()) {
1068 if ((SR.LaneMask & DstMask) == 0)
1071 VNInfo *SVNI = SR.getVNInfoAt(RegIndex);
1072 assert(SVNI != nullptr && SlotIndex::isSameInstr(SVNI->def, RegIndex));
1073 SR.removeValNo(SVNI);
1075 DstLI.removeEmptySubRanges();
1077 LIS->removeVRegDefAt(DstLI, RegIndex);
1079 // Mark uses as undef.
1080 for (MachineOperand &MO : MRI->reg_nodbg_operands(DstReg)) {
1081 if (MO.isDef() /*|| MO.isUndef()*/)
1083 const MachineInstr &MI = *MO.getParent();
1084 SlotIndex UseIdx = LIS->getInstructionIndex(&MI);
1085 unsigned UseMask = TRI->getSubRegIndexLaneMask(MO.getSubReg());
1087 if (UseMask != ~0u && DstLI.hasSubRanges()) {
1089 for (const LiveInterval::SubRange &SR : DstLI.subranges()) {
1090 if ((SR.LaneMask & UseMask) == 0)
1092 if (SR.liveAt(UseIdx)) {
1098 isLive = DstLI.liveAt(UseIdx);
1101 MO.setIsUndef(true);
1102 DEBUG(dbgs() << "\tnew undef: " << UseIdx << '\t' << MI);
1107 void RegisterCoalescer::updateRegDefsUses(unsigned SrcReg,
1110 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1111 LiveInterval *DstInt = DstIsPhys ? nullptr : &LIS->getInterval(DstReg);
1113 SmallPtrSet<MachineInstr*, 8> Visited;
1114 for (MachineRegisterInfo::reg_instr_iterator
1115 I = MRI->reg_instr_begin(SrcReg), E = MRI->reg_instr_end();
1117 MachineInstr *UseMI = &*(I++);
1119 // Each instruction can only be rewritten once because sub-register
1120 // composition is not always idempotent. When SrcReg != DstReg, rewriting
1121 // the UseMI operands removes them from the SrcReg use-def chain, but when
1122 // SrcReg is DstReg we could encounter UseMI twice if it has multiple
1123 // operands mentioning the virtual register.
1124 if (SrcReg == DstReg && !Visited.insert(UseMI).second)
1127 SmallVector<unsigned,8> Ops;
1129 std::tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
1131 // If SrcReg wasn't read, it may still be the case that DstReg is live-in
1132 // because SrcReg is a sub-register.
1133 if (DstInt && !Reads && SubIdx)
1134 Reads = DstInt->liveAt(LIS->getInstructionIndex(UseMI));
1136 // Replace SrcReg with DstReg in all UseMI operands.
1137 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1138 MachineOperand &MO = UseMI->getOperand(Ops[i]);
1140 // Adjust <undef> flags in case of sub-register joins. We don't want to
1141 // turn a full def into a read-modify-write sub-register def and vice
1143 if (SubIdx && MO.isDef())
1144 MO.setIsUndef(!Reads);
1146 // A subreg use of a partially undef (super) register may be a complete
1147 // undef use now and then has to be marked that way.
1148 if (SubIdx != 0 && MO.isUse() && MRI->tracksSubRegLiveness()) {
1149 if (!DstInt->hasSubRanges()) {
1150 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
1151 unsigned Mask = MRI->getMaxLaneMaskForVReg(DstInt->reg);
1152 DstInt->createSubRangeFrom(Allocator, Mask, *DstInt);
1154 unsigned Mask = TRI->getSubRegIndexLaneMask(SubIdx);
1155 bool IsUndef = true;
1156 SlotIndex MIIdx = UseMI->isDebugValue()
1157 ? LIS->getSlotIndexes()->getIndexBefore(UseMI)
1158 : LIS->getInstructionIndex(UseMI);
1159 SlotIndex UseIdx = MIIdx.getRegSlot(true);
1160 for (LiveInterval::SubRange &S : DstInt->subranges()) {
1161 if ((S.LaneMask & Mask) == 0)
1163 if (S.liveAt(UseIdx)) {
1169 MO.setIsUndef(true);
1170 // We found out some subregister use is actually reading an undefined
1171 // value. In some cases the whole vreg has become undefined at this
1172 // point so we have to potentially shrink the main range if the
1173 // use was ending a live segment there.
1174 LiveQueryResult Q = DstInt->Query(MIIdx);
1175 if (Q.valueOut() == nullptr)
1176 ShrinkMainRange = true;
1181 MO.substPhysReg(DstReg, *TRI);
1183 MO.substVirtReg(DstReg, SubIdx, *TRI);
1187 dbgs() << "\t\tupdated: ";
1188 if (!UseMI->isDebugValue())
1189 dbgs() << LIS->getInstructionIndex(UseMI) << "\t";
1195 bool RegisterCoalescer::canJoinPhys(const CoalescerPair &CP) {
1196 // Always join simple intervals that are defined by a single copy from a
1197 // reserved register. This doesn't increase register pressure, so it is
1198 // always beneficial.
1199 if (!MRI->isReserved(CP.getDstReg())) {
1200 DEBUG(dbgs() << "\tCan only merge into reserved registers.\n");
1204 LiveInterval &JoinVInt = LIS->getInterval(CP.getSrcReg());
1205 if (JoinVInt.containsOneValue())
1208 DEBUG(dbgs() << "\tCannot join complex intervals into reserved register.\n");
1212 bool RegisterCoalescer::joinCopy(MachineInstr *CopyMI, bool &Again) {
1215 DEBUG(dbgs() << LIS->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
1217 CoalescerPair CP(*TRI);
1218 if (!CP.setRegisters(CopyMI)) {
1219 DEBUG(dbgs() << "\tNot coalescable.\n");
1223 if (CP.getNewRC()) {
1224 auto SrcRC = MRI->getRegClass(CP.getSrcReg());
1225 auto DstRC = MRI->getRegClass(CP.getDstReg());
1226 unsigned SrcIdx = CP.getSrcIdx();
1227 unsigned DstIdx = CP.getDstIdx();
1228 if (CP.isFlipped()) {
1229 std::swap(SrcIdx, DstIdx);
1230 std::swap(SrcRC, DstRC);
1232 if (!TRI->shouldCoalesce(CopyMI, SrcRC, SrcIdx, DstRC, DstIdx,
1234 DEBUG(dbgs() << "\tSubtarget bailed on coalescing.\n");
1239 // Dead code elimination. This really should be handled by MachineDCE, but
1240 // sometimes dead copies slip through, and we can't generate invalid live
1242 if (!CP.isPhys() && CopyMI->allDefsAreDead()) {
1243 DEBUG(dbgs() << "\tCopy is dead.\n");
1244 DeadDefs.push_back(CopyMI);
1245 eliminateDeadDefs();
1249 // Eliminate undefs.
1250 if (!CP.isPhys() && eliminateUndefCopy(CopyMI)) {
1251 LIS->RemoveMachineInstrFromMaps(CopyMI);
1252 CopyMI->eraseFromParent();
1253 return false; // Not coalescable.
1256 // Coalesced copies are normally removed immediately, but transformations
1257 // like removeCopyByCommutingDef() can inadvertently create identity copies.
1258 // When that happens, just join the values and remove the copy.
1259 if (CP.getSrcReg() == CP.getDstReg()) {
1260 LiveInterval &LI = LIS->getInterval(CP.getSrcReg());
1261 DEBUG(dbgs() << "\tCopy already coalesced: " << LI << '\n');
1262 const SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI);
1263 LiveQueryResult LRQ = LI.Query(CopyIdx);
1264 if (VNInfo *DefVNI = LRQ.valueDefined()) {
1265 VNInfo *ReadVNI = LRQ.valueIn();
1266 assert(ReadVNI && "No value before copy and no <undef> flag.");
1267 assert(ReadVNI != DefVNI && "Cannot read and define the same value.");
1268 LI.MergeValueNumberInto(DefVNI, ReadVNI);
1270 // Process subregister liveranges.
1271 for (LiveInterval::SubRange &S : LI.subranges()) {
1272 LiveQueryResult SLRQ = S.Query(CopyIdx);
1273 if (VNInfo *SDefVNI = SLRQ.valueDefined()) {
1274 VNInfo *SReadVNI = SLRQ.valueIn();
1275 S.MergeValueNumberInto(SDefVNI, SReadVNI);
1278 DEBUG(dbgs() << "\tMerged values: " << LI << '\n');
1280 LIS->RemoveMachineInstrFromMaps(CopyMI);
1281 CopyMI->eraseFromParent();
1285 // Enforce policies.
1287 DEBUG(dbgs() << "\tConsidering merging " << PrintReg(CP.getSrcReg(), TRI)
1288 << " with " << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx())
1290 if (!canJoinPhys(CP)) {
1291 // Before giving up coalescing, if definition of source is defined by
1292 // trivial computation, try rematerializing it.
1294 if (reMaterializeTrivialDef(CP, CopyMI, IsDefCopy))
1297 Again = true; // May be possible to coalesce later.
1301 // When possible, let DstReg be the larger interval.
1302 if (!CP.isPartial() && LIS->getInterval(CP.getSrcReg()).size() >
1303 LIS->getInterval(CP.getDstReg()).size())
1307 dbgs() << "\tConsidering merging to "
1308 << TRI->getRegClassName(CP.getNewRC()) << " with ";
1309 if (CP.getDstIdx() && CP.getSrcIdx())
1310 dbgs() << PrintReg(CP.getDstReg()) << " in "
1311 << TRI->getSubRegIndexName(CP.getDstIdx()) << " and "
1312 << PrintReg(CP.getSrcReg()) << " in "
1313 << TRI->getSubRegIndexName(CP.getSrcIdx()) << '\n';
1315 dbgs() << PrintReg(CP.getSrcReg(), TRI) << " in "
1316 << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n';
1321 ShrinkMainRange = false;
1323 // Okay, attempt to join these two intervals. On failure, this returns false.
1324 // Otherwise, if one of the intervals being joined is a physreg, this method
1325 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1326 // been modified, so we can use this information below to update aliases.
1327 if (!joinIntervals(CP)) {
1328 // Coalescing failed.
1330 // If definition of source is defined by trivial computation, try
1331 // rematerializing it.
1333 if (reMaterializeTrivialDef(CP, CopyMI, IsDefCopy))
1336 // If we can eliminate the copy without merging the live segments, do so
1338 if (!CP.isPartial() && !CP.isPhys()) {
1339 if (adjustCopiesBackFrom(CP, CopyMI) ||
1340 removeCopyByCommutingDef(CP, CopyMI)) {
1341 LIS->RemoveMachineInstrFromMaps(CopyMI);
1342 CopyMI->eraseFromParent();
1343 DEBUG(dbgs() << "\tTrivial!\n");
1348 // Otherwise, we are unable to join the intervals.
1349 DEBUG(dbgs() << "\tInterference!\n");
1350 Again = true; // May be possible to coalesce later.
1354 // Coalescing to a virtual register that is of a sub-register class of the
1355 // other. Make sure the resulting register is set to the right register class.
1356 if (CP.isCrossClass()) {
1358 MRI->setRegClass(CP.getDstReg(), CP.getNewRC());
1361 // Removing sub-register copies can ease the register class constraints.
1362 // Make sure we attempt to inflate the register class of DstReg.
1363 if (!CP.isPhys() && RegClassInfo.isProperSubClass(CP.getNewRC()))
1364 InflateRegs.push_back(CP.getDstReg());
1366 // CopyMI has been erased by joinIntervals at this point. Remove it from
1367 // ErasedInstrs since copyCoalesceWorkList() won't add a successful join back
1368 // to the work list. This keeps ErasedInstrs from growing needlessly.
1369 ErasedInstrs.erase(CopyMI);
1371 // Rewrite all SrcReg operands to DstReg.
1372 // Also update DstReg operands to include DstIdx if it is set.
1374 updateRegDefsUses(CP.getDstReg(), CP.getDstReg(), CP.getDstIdx());
1375 updateRegDefsUses(CP.getSrcReg(), CP.getDstReg(), CP.getSrcIdx());
1377 // Shrink subregister ranges if necessary.
1378 if (ShrinkMask != 0) {
1379 LiveInterval &LI = LIS->getInterval(CP.getDstReg());
1380 for (LiveInterval::SubRange &S : LI.subranges()) {
1381 if ((S.LaneMask & ShrinkMask) == 0)
1383 DEBUG(dbgs() << "Shrink LaneUses (Lane "
1384 << format("%04X", S.LaneMask) << ")\n");
1385 LIS->shrinkToUses(S, LI.reg);
1388 if (ShrinkMainRange) {
1389 LiveInterval &LI = LIS->getInterval(CP.getDstReg());
1390 LIS->shrinkToUses(&LI);
1393 // SrcReg is guaranteed to be the register whose live interval that is
1395 LIS->removeInterval(CP.getSrcReg());
1397 // Update regalloc hint.
1398 TRI->updateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF);
1401 dbgs() << "\tSuccess: " << PrintReg(CP.getSrcReg(), TRI, CP.getSrcIdx())
1402 << " -> " << PrintReg(CP.getDstReg(), TRI, CP.getDstIdx()) << '\n';
1403 dbgs() << "\tResult = ";
1405 dbgs() << PrintReg(CP.getDstReg(), TRI);
1407 dbgs() << LIS->getInterval(CP.getDstReg());
1415 bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) {
1416 unsigned DstReg = CP.getDstReg();
1417 assert(CP.isPhys() && "Must be a physreg copy");
1418 assert(MRI->isReserved(DstReg) && "Not a reserved register");
1419 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
1420 DEBUG(dbgs() << "\t\tRHS = " << RHS << '\n');
1422 assert(RHS.containsOneValue() && "Invalid join with reserved register");
1424 // Optimization for reserved registers like ESP. We can only merge with a
1425 // reserved physreg if RHS has a single value that is a copy of DstReg.
1426 // The live range of the reserved register will look like a set of dead defs
1427 // - we don't properly track the live range of reserved registers.
1429 // Deny any overlapping intervals. This depends on all the reserved
1430 // register live ranges to look like dead defs.
1431 for (MCRegUnitIterator UI(DstReg, TRI); UI.isValid(); ++UI)
1432 if (RHS.overlaps(LIS->getRegUnit(*UI))) {
1433 DEBUG(dbgs() << "\t\tInterference: " << PrintRegUnit(*UI, TRI) << '\n');
1437 // Skip any value computations, we are not adding new values to the
1438 // reserved register. Also skip merging the live ranges, the reserved
1439 // register live range doesn't need to be accurate as long as all the
1442 // Delete the identity copy.
1443 MachineInstr *CopyMI;
1444 if (CP.isFlipped()) {
1445 CopyMI = MRI->getVRegDef(RHS.reg);
1447 if (!MRI->hasOneNonDBGUse(RHS.reg)) {
1448 DEBUG(dbgs() << "\t\tMultiple vreg uses!\n");
1452 MachineInstr *DestMI = MRI->getVRegDef(RHS.reg);
1453 CopyMI = &*MRI->use_instr_nodbg_begin(RHS.reg);
1454 const SlotIndex CopyRegIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
1455 const SlotIndex DestRegIdx = LIS->getInstructionIndex(DestMI).getRegSlot();
1457 // We checked above that there are no interfering defs of the physical
1458 // register. However, for this case, where we intent to move up the def of
1459 // the physical register, we also need to check for interfering uses.
1460 SlotIndexes *Indexes = LIS->getSlotIndexes();
1461 for (SlotIndex SI = Indexes->getNextNonNullIndex(DestRegIdx);
1462 SI != CopyRegIdx; SI = Indexes->getNextNonNullIndex(SI)) {
1463 MachineInstr *MI = LIS->getInstructionFromIndex(SI);
1464 if (MI->readsRegister(DstReg, TRI)) {
1465 DEBUG(dbgs() << "\t\tInterference (read): " << *MI);
1470 // We're going to remove the copy which defines a physical reserved
1471 // register, so remove its valno, etc.
1472 DEBUG(dbgs() << "\t\tRemoving phys reg def of " << DstReg << " at "
1473 << CopyRegIdx << "\n");
1475 LIS->removePhysRegDefAt(DstReg, CopyRegIdx);
1476 // Create a new dead def at the new def location.
1477 for (MCRegUnitIterator UI(DstReg, TRI); UI.isValid(); ++UI) {
1478 LiveRange &LR = LIS->getRegUnit(*UI);
1479 LR.createDeadDef(DestRegIdx, LIS->getVNInfoAllocator());
1483 LIS->RemoveMachineInstrFromMaps(CopyMI);
1484 CopyMI->eraseFromParent();
1486 // We don't track kills for reserved registers.
1487 MRI->clearKillFlags(CP.getSrcReg());
1492 //===----------------------------------------------------------------------===//
1493 // Interference checking and interval joining
1494 //===----------------------------------------------------------------------===//
1496 // In the easiest case, the two live ranges being joined are disjoint, and
1497 // there is no interference to consider. It is quite common, though, to have
1498 // overlapping live ranges, and we need to check if the interference can be
1501 // The live range of a single SSA value forms a sub-tree of the dominator tree.
1502 // This means that two SSA values overlap if and only if the def of one value
1503 // is contained in the live range of the other value. As a special case, the
1504 // overlapping values can be defined at the same index.
1506 // The interference from an overlapping def can be resolved in these cases:
1508 // 1. Coalescable copies. The value is defined by a copy that would become an
1509 // identity copy after joining SrcReg and DstReg. The copy instruction will
1510 // be removed, and the value will be merged with the source value.
1512 // There can be several copies back and forth, causing many values to be
1513 // merged into one. We compute a list of ultimate values in the joined live
1514 // range as well as a mappings from the old value numbers.
1516 // 2. IMPLICIT_DEF. This instruction is only inserted to ensure all PHI
1517 // predecessors have a live out value. It doesn't cause real interference,
1518 // and can be merged into the value it overlaps. Like a coalescable copy, it
1519 // can be erased after joining.
1521 // 3. Copy of external value. The overlapping def may be a copy of a value that
1522 // is already in the other register. This is like a coalescable copy, but
1523 // the live range of the source register must be trimmed after erasing the
1524 // copy instruction:
1527 // %dst = COPY %ext <-- Remove this COPY, trim the live range of %ext.
1529 // 4. Clobbering undefined lanes. Vector registers are sometimes built by
1530 // defining one lane at a time:
1532 // %dst:ssub0<def,read-undef> = FOO
1534 // %dst:ssub1<def> = COPY %src
1536 // The live range of %src overlaps the %dst value defined by FOO, but
1537 // merging %src into %dst:ssub1 is only going to clobber the ssub1 lane
1538 // which was undef anyway.
1540 // The value mapping is more complicated in this case. The final live range
1541 // will have different value numbers for both FOO and BAR, but there is no
1542 // simple mapping from old to new values. It may even be necessary to add
1545 // 5. Clobbering dead lanes. A def may clobber a lane of a vector register that
1546 // is live, but never read. This can happen because we don't compute
1547 // individual live ranges per lane.
1551 // %dst:ssub1<def> = COPY %src
1553 // This kind of interference is only resolved locally. If the clobbered
1554 // lane value escapes the block, the join is aborted.
1557 /// Track information about values in a single virtual register about to be
1558 /// joined. Objects of this class are always created in pairs - one for each
1559 /// side of the CoalescerPair (or one for each lane of a side of the coalescer
1562 /// Live range we work on.
1564 /// (Main) register we work on.
1567 /// Reg (and therefore the values in this liverange) will end up as
1568 /// subregister SubIdx in the coalesced register. Either CP.DstIdx or
1570 const unsigned SubIdx;
1571 /// The LaneMask that this liverange will occupy the coalesced register. May
1572 /// be smaller than the lanemask produced by SubIdx when merging subranges.
1573 const unsigned LaneMask;
1575 /// This is true when joining sub register ranges, false when joining main
1577 const bool SubRangeJoin;
1578 /// Whether the current LiveInterval tracks subregister liveness.
1579 const bool TrackSubRegLiveness;
1581 /// Values that will be present in the final live range.
1582 SmallVectorImpl<VNInfo*> &NewVNInfo;
1584 const CoalescerPair &CP;
1586 SlotIndexes *Indexes;
1587 const TargetRegisterInfo *TRI;
1589 /// Value number assignments. Maps value numbers in LI to entries in
1590 /// NewVNInfo. This is suitable for passing to LiveInterval::join().
1591 SmallVector<int, 8> Assignments;
1593 /// Conflict resolution for overlapping values.
1594 enum ConflictResolution {
1595 /// No overlap, simply keep this value.
1598 /// Merge this value into OtherVNI and erase the defining instruction.
1599 /// Used for IMPLICIT_DEF, coalescable copies, and copies from external
1603 /// Merge this value into OtherVNI but keep the defining instruction.
1604 /// This is for the special case where OtherVNI is defined by the same
1608 /// Keep this value, and have it replace OtherVNI where possible. This
1609 /// complicates value mapping since OtherVNI maps to two different values
1610 /// before and after this def.
1611 /// Used when clobbering undefined or dead lanes.
1614 /// Unresolved conflict. Visit later when all values have been mapped.
1617 /// Unresolvable conflict. Abort the join.
1621 /// Per-value info for LI. The lane bit masks are all relative to the final
1622 /// joined register, so they can be compared directly between SrcReg and
1625 ConflictResolution Resolution;
1627 /// Lanes written by this def, 0 for unanalyzed values.
1628 unsigned WriteLanes;
1630 /// Lanes with defined values in this register. Other lanes are undef and
1631 /// safe to clobber.
1632 unsigned ValidLanes;
1634 /// Value in LI being redefined by this def.
1637 /// Value in the other live range that overlaps this def, if any.
1640 /// Is this value an IMPLICIT_DEF that can be erased?
1642 /// IMPLICIT_DEF values should only exist at the end of a basic block that
1643 /// is a predecessor to a phi-value. These IMPLICIT_DEF instructions can be
1644 /// safely erased if they are overlapping a live value in the other live
1647 /// Weird control flow graphs and incomplete PHI handling in
1648 /// ProcessImplicitDefs can very rarely create IMPLICIT_DEF values with
1649 /// longer live ranges. Such IMPLICIT_DEF values should be treated like
1651 bool ErasableImplicitDef;
1653 /// True when the live range of this value will be pruned because of an
1654 /// overlapping CR_Replace value in the other live range.
1657 /// True once Pruned above has been computed.
1658 bool PrunedComputed;
1660 Val() : Resolution(CR_Keep), WriteLanes(0), ValidLanes(0),
1661 RedefVNI(nullptr), OtherVNI(nullptr), ErasableImplicitDef(false),
1662 Pruned(false), PrunedComputed(false) {}
1664 bool isAnalyzed() const { return WriteLanes != 0; }
1667 /// One entry per value number in LI.
1668 SmallVector<Val, 8> Vals;
1670 /// Compute the bitmask of lanes actually written by DefMI.
1671 /// Set Redef if there are any partial register definitions that depend on the
1672 /// previous value of the register.
1673 unsigned computeWriteLanes(const MachineInstr *DefMI, bool &Redef) const;
1675 /// Find the ultimate value that VNI was copied from.
1676 std::pair<const VNInfo*,unsigned> followCopyChain(const VNInfo *VNI) const;
1678 bool valuesIdentical(VNInfo *Val0, VNInfo *Val1, const JoinVals &Other) const;
1680 /// Analyze ValNo in this live range, and set all fields of Vals[ValNo].
1681 /// Return a conflict resolution when possible, but leave the hard cases as
1683 /// Recursively calls computeAssignment() on this and Other, guaranteeing that
1684 /// both OtherVNI and RedefVNI have been analyzed and mapped before returning.
1685 /// The recursion always goes upwards in the dominator tree, making loops
1687 ConflictResolution analyzeValue(unsigned ValNo, JoinVals &Other);
1689 /// Compute the value assignment for ValNo in RI.
1690 /// This may be called recursively by analyzeValue(), but never for a ValNo on
1692 void computeAssignment(unsigned ValNo, JoinVals &Other);
1694 /// Assuming ValNo is going to clobber some valid lanes in Other.LR, compute
1695 /// the extent of the tainted lanes in the block.
1697 /// Multiple values in Other.LR can be affected since partial redefinitions
1698 /// can preserve previously tainted lanes.
1700 /// 1 %dst = VLOAD <-- Define all lanes in %dst
1701 /// 2 %src = FOO <-- ValNo to be joined with %dst:ssub0
1702 /// 3 %dst:ssub1 = BAR <-- Partial redef doesn't clear taint in ssub0
1703 /// 4 %dst:ssub0 = COPY %src <-- Conflict resolved, ssub0 wasn't read
1705 /// For each ValNo in Other that is affected, add an (EndIndex, TaintedLanes)
1706 /// entry to TaintedVals.
1708 /// Returns false if the tainted lanes extend beyond the basic block.
1709 bool taintExtent(unsigned, unsigned, JoinVals&,
1710 SmallVectorImpl<std::pair<SlotIndex, unsigned> >&);
1712 /// Return true if MI uses any of the given Lanes from Reg.
1713 /// This does not include partial redefinitions of Reg.
1714 bool usesLanes(const MachineInstr *MI, unsigned, unsigned, unsigned) const;
1716 /// Determine if ValNo is a copy of a value number in LR or Other.LR that will
1719 /// %dst = COPY %src
1720 /// %src = COPY %dst <-- This value to be pruned.
1721 /// %dst = COPY %src <-- This value is a copy of a pruned value.
1722 bool isPrunedValue(unsigned ValNo, JoinVals &Other);
1725 JoinVals(LiveRange &LR, unsigned Reg, unsigned SubIdx, unsigned LaneMask,
1726 SmallVectorImpl<VNInfo*> &newVNInfo, const CoalescerPair &cp,
1727 LiveIntervals *lis, const TargetRegisterInfo *TRI, bool SubRangeJoin,
1728 bool TrackSubRegLiveness)
1729 : LR(LR), Reg(Reg), SubIdx(SubIdx), LaneMask(LaneMask),
1730 SubRangeJoin(SubRangeJoin), TrackSubRegLiveness(TrackSubRegLiveness),
1731 NewVNInfo(newVNInfo), CP(cp), LIS(lis), Indexes(LIS->getSlotIndexes()),
1732 TRI(TRI), Assignments(LR.getNumValNums(), -1), Vals(LR.getNumValNums())
1735 /// Analyze defs in LR and compute a value mapping in NewVNInfo.
1736 /// Returns false if any conflicts were impossible to resolve.
1737 bool mapValues(JoinVals &Other);
1739 /// Try to resolve conflicts that require all values to be mapped.
1740 /// Returns false if any conflicts were impossible to resolve.
1741 bool resolveConflicts(JoinVals &Other);
1743 /// Prune the live range of values in Other.LR where they would conflict with
1744 /// CR_Replace values in LR. Collect end points for restoring the live range
1746 void pruneValues(JoinVals &Other, SmallVectorImpl<SlotIndex> &EndPoints,
1749 /// Removes subranges starting at copies that get removed. This sometimes
1750 /// happens when undefined subranges are copied around. These ranges contain
1751 /// no usefull information and can be removed.
1752 void pruneSubRegValues(LiveInterval &LI, unsigned &ShrinkMask);
1754 /// Erase any machine instructions that have been coalesced away.
1755 /// Add erased instructions to ErasedInstrs.
1756 /// Add foreign virtual registers to ShrinkRegs if their live range ended at
1757 /// the erased instrs.
1758 void eraseInstrs(SmallPtrSetImpl<MachineInstr*> &ErasedInstrs,
1759 SmallVectorImpl<unsigned> &ShrinkRegs);
1761 /// Get the value assignments suitable for passing to LiveInterval::join.
1762 const int *getAssignments() const { return Assignments.data(); }
1764 } // end anonymous namespace
1766 unsigned JoinVals::computeWriteLanes(const MachineInstr *DefMI, bool &Redef)
1769 for (ConstMIOperands MO(DefMI); MO.isValid(); ++MO) {
1770 if (!MO->isReg() || MO->getReg() != Reg || !MO->isDef())
1772 L |= TRI->getSubRegIndexLaneMask(
1773 TRI->composeSubRegIndices(SubIdx, MO->getSubReg()));
1780 std::pair<const VNInfo*, unsigned> JoinVals::followCopyChain(
1781 const VNInfo *VNI) const {
1782 unsigned Reg = this->Reg;
1784 while (!VNI->isPHIDef()) {
1785 SlotIndex Def = VNI->def;
1786 MachineInstr *MI = Indexes->getInstructionFromIndex(Def);
1787 assert(MI && "No defining instruction");
1788 if (!MI->isFullCopy())
1789 return std::make_pair(VNI, Reg);
1790 unsigned SrcReg = MI->getOperand(1).getReg();
1791 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1792 return std::make_pair(VNI, Reg);
1794 const LiveInterval &LI = LIS->getInterval(SrcReg);
1795 const VNInfo *ValueIn;
1796 // No subrange involved.
1797 if (!SubRangeJoin || !LI.hasSubRanges()) {
1798 LiveQueryResult LRQ = LI.Query(Def);
1799 ValueIn = LRQ.valueIn();
1801 // Query subranges. Pick the first matching one.
1803 for (const LiveInterval::SubRange &S : LI.subranges()) {
1804 // Transform lanemask to a mask in the joined live interval.
1805 unsigned SMask = TRI->composeSubRegIndexLaneMask(SubIdx, S.LaneMask);
1806 if ((SMask & LaneMask) == 0)
1808 LiveQueryResult LRQ = S.Query(Def);
1809 ValueIn = LRQ.valueIn();
1813 if (ValueIn == nullptr)
1818 return std::make_pair(VNI, Reg);
1821 bool JoinVals::valuesIdentical(VNInfo *Value0, VNInfo *Value1,
1822 const JoinVals &Other) const {
1823 const VNInfo *Orig0;
1825 std::tie(Orig0, Reg0) = followCopyChain(Value0);
1826 if (Orig0 == Value1)
1829 const VNInfo *Orig1;
1831 std::tie(Orig1, Reg1) = Other.followCopyChain(Value1);
1833 // The values are equal if they are defined at the same place and use the
1834 // same register. Note that we cannot compare VNInfos directly as some of
1835 // them might be from a copy created in mergeSubRangeInto() while the other
1836 // is from the original LiveInterval.
1837 return Orig0->def == Orig1->def && Reg0 == Reg1;
1840 JoinVals::ConflictResolution
1841 JoinVals::analyzeValue(unsigned ValNo, JoinVals &Other) {
1842 Val &V = Vals[ValNo];
1843 assert(!V.isAnalyzed() && "Value has already been analyzed!");
1844 VNInfo *VNI = LR.getValNumInfo(ValNo);
1845 if (VNI->isUnused()) {
1850 // Get the instruction defining this value, compute the lanes written.
1851 const MachineInstr *DefMI = nullptr;
1852 if (VNI->isPHIDef()) {
1853 // Conservatively assume that all lanes in a PHI are valid.
1854 unsigned Lanes = SubRangeJoin ? 1 : TRI->getSubRegIndexLaneMask(SubIdx);
1855 V.ValidLanes = V.WriteLanes = Lanes;
1857 DefMI = Indexes->getInstructionFromIndex(VNI->def);
1858 assert(DefMI != nullptr);
1860 // We don't care about the lanes when joining subregister ranges.
1861 V.ValidLanes = V.WriteLanes = 1;
1864 V.ValidLanes = V.WriteLanes = computeWriteLanes(DefMI, Redef);
1866 // If this is a read-modify-write instruction, there may be more valid
1867 // lanes than the ones written by this instruction.
1868 // This only covers partial redef operands. DefMI may have normal use
1869 // operands reading the register. They don't contribute valid lanes.
1871 // This adds ssub1 to the set of valid lanes in %src:
1873 // %src:ssub1<def> = FOO
1875 // This leaves only ssub1 valid, making any other lanes undef:
1877 // %src:ssub1<def,read-undef> = FOO %src:ssub2
1879 // The <read-undef> flag on the def operand means that old lane values are
1882 V.RedefVNI = LR.Query(VNI->def).valueIn();
1883 assert((TrackSubRegLiveness || V.RedefVNI) &&
1884 "Instruction is reading nonexistent value");
1885 if (V.RedefVNI != nullptr) {
1886 computeAssignment(V.RedefVNI->id, Other);
1887 V.ValidLanes |= Vals[V.RedefVNI->id].ValidLanes;
1891 // An IMPLICIT_DEF writes undef values.
1892 if (DefMI->isImplicitDef()) {
1893 // We normally expect IMPLICIT_DEF values to be live only until the end
1894 // of their block. If the value is really live longer and gets pruned in
1895 // another block, this flag is cleared again.
1896 V.ErasableImplicitDef = true;
1897 V.ValidLanes &= ~V.WriteLanes;
1902 // Find the value in Other that overlaps VNI->def, if any.
1903 LiveQueryResult OtherLRQ = Other.LR.Query(VNI->def);
1905 // It is possible that both values are defined by the same instruction, or
1906 // the values are PHIs defined in the same block. When that happens, the two
1907 // values should be merged into one, but not into any preceding value.
1908 // The first value defined or visited gets CR_Keep, the other gets CR_Merge.
1909 if (VNInfo *OtherVNI = OtherLRQ.valueDefined()) {
1910 assert(SlotIndex::isSameInstr(VNI->def, OtherVNI->def) && "Broken LRQ");
1912 // One value stays, the other is merged. Keep the earlier one, or the first
1914 if (OtherVNI->def < VNI->def)
1915 Other.computeAssignment(OtherVNI->id, *this);
1916 else if (VNI->def < OtherVNI->def && OtherLRQ.valueIn()) {
1917 // This is an early-clobber def overlapping a live-in value in the other
1918 // register. Not mergeable.
1919 V.OtherVNI = OtherLRQ.valueIn();
1920 return CR_Impossible;
1922 V.OtherVNI = OtherVNI;
1923 Val &OtherV = Other.Vals[OtherVNI->id];
1924 // Keep this value, check for conflicts when analyzing OtherVNI.
1925 if (!OtherV.isAnalyzed())
1927 // Both sides have been analyzed now.
1928 // Allow overlapping PHI values. Any real interference would show up in a
1929 // predecessor, the PHI itself can't introduce any conflicts.
1930 if (VNI->isPHIDef())
1932 if (V.ValidLanes & OtherV.ValidLanes)
1933 // Overlapping lanes can't be resolved.
1934 return CR_Impossible;
1939 // No simultaneous def. Is Other live at the def?
1940 V.OtherVNI = OtherLRQ.valueIn();
1942 // No overlap, no conflict.
1945 assert(!SlotIndex::isSameInstr(VNI->def, V.OtherVNI->def) && "Broken LRQ");
1947 // We have overlapping values, or possibly a kill of Other.
1948 // Recursively compute assignments up the dominator tree.
1949 Other.computeAssignment(V.OtherVNI->id, *this);
1950 Val &OtherV = Other.Vals[V.OtherVNI->id];
1952 // Check if OtherV is an IMPLICIT_DEF that extends beyond its basic block.
1953 // This shouldn't normally happen, but ProcessImplicitDefs can leave such
1954 // IMPLICIT_DEF instructions behind, and there is nothing wrong with it
1957 // WHen it happens, treat that IMPLICIT_DEF as a normal value, and don't try
1958 // to erase the IMPLICIT_DEF instruction.
1959 if (OtherV.ErasableImplicitDef && DefMI &&
1960 DefMI->getParent() != Indexes->getMBBFromIndex(V.OtherVNI->def)) {
1961 DEBUG(dbgs() << "IMPLICIT_DEF defined at " << V.OtherVNI->def
1962 << " extends into BB#" << DefMI->getParent()->getNumber()
1963 << ", keeping it.\n");
1964 OtherV.ErasableImplicitDef = false;
1967 // Allow overlapping PHI values. Any real interference would show up in a
1968 // predecessor, the PHI itself can't introduce any conflicts.
1969 if (VNI->isPHIDef())
1972 // Check for simple erasable conflicts.
1973 if (DefMI->isImplicitDef()) {
1974 // We need the def for the subregister if there is nothing else live at the
1975 // subrange at this point.
1976 if (TrackSubRegLiveness
1977 && (V.WriteLanes & (OtherV.ValidLanes | OtherV.WriteLanes)) == 0)
1982 // Include the non-conflict where DefMI is a coalescable copy that kills
1983 // OtherVNI. We still want the copy erased and value numbers merged.
1984 if (CP.isCoalescable(DefMI)) {
1985 // Some of the lanes copied from OtherVNI may be undef, making them undef
1987 V.ValidLanes &= ~V.WriteLanes | OtherV.ValidLanes;
1991 // This may not be a real conflict if DefMI simply kills Other and defines
1993 if (OtherLRQ.isKill() && OtherLRQ.endPoint() <= VNI->def)
1996 // Handle the case where VNI and OtherVNI can be proven to be identical:
1998 // %other = COPY %ext
1999 // %this = COPY %ext <-- Erase this copy
2001 if (DefMI->isFullCopy() && !CP.isPartial()
2002 && valuesIdentical(VNI, V.OtherVNI, Other))
2005 // If the lanes written by this instruction were all undef in OtherVNI, it is
2006 // still safe to join the live ranges. This can't be done with a simple value
2007 // mapping, though - OtherVNI will map to multiple values:
2009 // 1 %dst:ssub0 = FOO <-- OtherVNI
2010 // 2 %src = BAR <-- VNI
2011 // 3 %dst:ssub1 = COPY %src<kill> <-- Eliminate this copy.
2013 // 5 QUUX %src<kill>
2015 // Here OtherVNI will map to itself in [1;2), but to VNI in [2;5). CR_Replace
2016 // handles this complex value mapping.
2017 if ((V.WriteLanes & OtherV.ValidLanes) == 0)
2020 // If the other live range is killed by DefMI and the live ranges are still
2021 // overlapping, it must be because we're looking at an early clobber def:
2023 // %dst<def,early-clobber> = ASM %src<kill>
2025 // In this case, it is illegal to merge the two live ranges since the early
2026 // clobber def would clobber %src before it was read.
2027 if (OtherLRQ.isKill()) {
2028 // This case where the def doesn't overlap the kill is handled above.
2029 assert(VNI->def.isEarlyClobber() &&
2030 "Only early clobber defs can overlap a kill");
2031 return CR_Impossible;
2034 // VNI is clobbering live lanes in OtherVNI, but there is still the
2035 // possibility that no instructions actually read the clobbered lanes.
2036 // If we're clobbering all the lanes in OtherVNI, at least one must be read.
2037 // Otherwise Other.RI wouldn't be live here.
2038 if ((TRI->getSubRegIndexLaneMask(Other.SubIdx) & ~V.WriteLanes) == 0)
2039 return CR_Impossible;
2041 // We need to verify that no instructions are reading the clobbered lanes. To
2042 // save compile time, we'll only check that locally. Don't allow the tainted
2043 // value to escape the basic block.
2044 MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
2045 if (OtherLRQ.endPoint() >= Indexes->getMBBEndIdx(MBB))
2046 return CR_Impossible;
2048 // There are still some things that could go wrong besides clobbered lanes
2049 // being read, for example OtherVNI may be only partially redefined in MBB,
2050 // and some clobbered lanes could escape the block. Save this analysis for
2051 // resolveConflicts() when all values have been mapped. We need to know
2052 // RedefVNI and WriteLanes for any later defs in MBB, and we can't compute
2053 // that now - the recursive analyzeValue() calls must go upwards in the
2055 return CR_Unresolved;
2058 void JoinVals::computeAssignment(unsigned ValNo, JoinVals &Other) {
2059 Val &V = Vals[ValNo];
2060 if (V.isAnalyzed()) {
2061 // Recursion should always move up the dominator tree, so ValNo is not
2062 // supposed to reappear before it has been assigned.
2063 assert(Assignments[ValNo] != -1 && "Bad recursion?");
2066 switch ((V.Resolution = analyzeValue(ValNo, Other))) {
2069 // Merge this ValNo into OtherVNI.
2070 assert(V.OtherVNI && "OtherVNI not assigned, can't merge.");
2071 assert(Other.Vals[V.OtherVNI->id].isAnalyzed() && "Missing recursion");
2072 Assignments[ValNo] = Other.Assignments[V.OtherVNI->id];
2073 DEBUG(dbgs() << "\t\tmerge " << PrintReg(Reg) << ':' << ValNo << '@'
2074 << LR.getValNumInfo(ValNo)->def << " into "
2075 << PrintReg(Other.Reg) << ':' << V.OtherVNI->id << '@'
2076 << V.OtherVNI->def << " --> @"
2077 << NewVNInfo[Assignments[ValNo]]->def << '\n');
2080 case CR_Unresolved: {
2081 // The other value is going to be pruned if this join is successful.
2082 assert(V.OtherVNI && "OtherVNI not assigned, can't prune");
2083 Val &OtherV = Other.Vals[V.OtherVNI->id];
2084 // We cannot erase an IMPLICIT_DEF if we don't have valid values for all
2086 if ((OtherV.WriteLanes & ~V.ValidLanes) != 0 && TrackSubRegLiveness)
2087 OtherV.ErasableImplicitDef = false;
2088 OtherV.Pruned = true;
2092 // This value number needs to go in the final joined live range.
2093 Assignments[ValNo] = NewVNInfo.size();
2094 NewVNInfo.push_back(LR.getValNumInfo(ValNo));
2099 bool JoinVals::mapValues(JoinVals &Other) {
2100 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2101 computeAssignment(i, Other);
2102 if (Vals[i].Resolution == CR_Impossible) {
2103 DEBUG(dbgs() << "\t\tinterference at " << PrintReg(Reg) << ':' << i
2104 << '@' << LR.getValNumInfo(i)->def << '\n');
2112 taintExtent(unsigned ValNo, unsigned TaintedLanes, JoinVals &Other,
2113 SmallVectorImpl<std::pair<SlotIndex, unsigned> > &TaintExtent) {
2114 VNInfo *VNI = LR.getValNumInfo(ValNo);
2115 MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
2116 SlotIndex MBBEnd = Indexes->getMBBEndIdx(MBB);
2118 // Scan Other.LR from VNI.def to MBBEnd.
2119 LiveInterval::iterator OtherI = Other.LR.find(VNI->def);
2120 assert(OtherI != Other.LR.end() && "No conflict?");
2122 // OtherI is pointing to a tainted value. Abort the join if the tainted
2123 // lanes escape the block.
2124 SlotIndex End = OtherI->end;
2125 if (End >= MBBEnd) {
2126 DEBUG(dbgs() << "\t\ttaints global " << PrintReg(Other.Reg) << ':'
2127 << OtherI->valno->id << '@' << OtherI->start << '\n');
2130 DEBUG(dbgs() << "\t\ttaints local " << PrintReg(Other.Reg) << ':'
2131 << OtherI->valno->id << '@' << OtherI->start
2132 << " to " << End << '\n');
2133 // A dead def is not a problem.
2136 TaintExtent.push_back(std::make_pair(End, TaintedLanes));
2138 // Check for another def in the MBB.
2139 if (++OtherI == Other.LR.end() || OtherI->start >= MBBEnd)
2142 // Lanes written by the new def are no longer tainted.
2143 const Val &OV = Other.Vals[OtherI->valno->id];
2144 TaintedLanes &= ~OV.WriteLanes;
2147 } while (TaintedLanes);
2151 bool JoinVals::usesLanes(const MachineInstr *MI, unsigned Reg, unsigned SubIdx,
2152 unsigned Lanes) const {
2153 if (MI->isDebugValue())
2155 for (ConstMIOperands MO(MI); MO.isValid(); ++MO) {
2156 if (!MO->isReg() || MO->isDef() || MO->getReg() != Reg)
2158 if (!MO->readsReg())
2160 if (Lanes & TRI->getSubRegIndexLaneMask(
2161 TRI->composeSubRegIndices(SubIdx, MO->getSubReg())))
2167 bool JoinVals::resolveConflicts(JoinVals &Other) {
2168 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2170 assert (V.Resolution != CR_Impossible && "Unresolvable conflict");
2171 if (V.Resolution != CR_Unresolved)
2173 DEBUG(dbgs() << "\t\tconflict at " << PrintReg(Reg) << ':' << i
2174 << '@' << LR.getValNumInfo(i)->def << '\n');
2179 assert(V.OtherVNI && "Inconsistent conflict resolution.");
2180 VNInfo *VNI = LR.getValNumInfo(i);
2181 const Val &OtherV = Other.Vals[V.OtherVNI->id];
2183 // VNI is known to clobber some lanes in OtherVNI. If we go ahead with the
2184 // join, those lanes will be tainted with a wrong value. Get the extent of
2185 // the tainted lanes.
2186 unsigned TaintedLanes = V.WriteLanes & OtherV.ValidLanes;
2187 SmallVector<std::pair<SlotIndex, unsigned>, 8> TaintExtent;
2188 if (!taintExtent(i, TaintedLanes, Other, TaintExtent))
2189 // Tainted lanes would extend beyond the basic block.
2192 assert(!TaintExtent.empty() && "There should be at least one conflict.");
2194 // Now look at the instructions from VNI->def to TaintExtent (inclusive).
2195 MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
2196 MachineBasicBlock::iterator MI = MBB->begin();
2197 if (!VNI->isPHIDef()) {
2198 MI = Indexes->getInstructionFromIndex(VNI->def);
2199 // No need to check the instruction defining VNI for reads.
2202 assert(!SlotIndex::isSameInstr(VNI->def, TaintExtent.front().first) &&
2203 "Interference ends on VNI->def. Should have been handled earlier");
2204 MachineInstr *LastMI =
2205 Indexes->getInstructionFromIndex(TaintExtent.front().first);
2206 assert(LastMI && "Range must end at a proper instruction");
2207 unsigned TaintNum = 0;
2209 assert(MI != MBB->end() && "Bad LastMI");
2210 if (usesLanes(MI, Other.Reg, Other.SubIdx, TaintedLanes)) {
2211 DEBUG(dbgs() << "\t\ttainted lanes used by: " << *MI);
2214 // LastMI is the last instruction to use the current value.
2215 if (&*MI == LastMI) {
2216 if (++TaintNum == TaintExtent.size())
2218 LastMI = Indexes->getInstructionFromIndex(TaintExtent[TaintNum].first);
2219 assert(LastMI && "Range must end at a proper instruction");
2220 TaintedLanes = TaintExtent[TaintNum].second;
2225 // The tainted lanes are unused.
2226 V.Resolution = CR_Replace;
2232 bool JoinVals::isPrunedValue(unsigned ValNo, JoinVals &Other) {
2233 Val &V = Vals[ValNo];
2234 if (V.Pruned || V.PrunedComputed)
2237 if (V.Resolution != CR_Erase && V.Resolution != CR_Merge)
2240 // Follow copies up the dominator tree and check if any intermediate value
2242 V.PrunedComputed = true;
2243 V.Pruned = Other.isPrunedValue(V.OtherVNI->id, *this);
2247 void JoinVals::pruneValues(JoinVals &Other,
2248 SmallVectorImpl<SlotIndex> &EndPoints,
2249 bool changeInstrs) {
2250 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2251 SlotIndex Def = LR.getValNumInfo(i)->def;
2252 switch (Vals[i].Resolution) {
2256 // This value takes precedence over the value in Other.LR.
2257 LIS->pruneValue(Other.LR, Def, &EndPoints);
2258 // Check if we're replacing an IMPLICIT_DEF value. The IMPLICIT_DEF
2259 // instructions are only inserted to provide a live-out value for PHI
2260 // predecessors, so the instruction should simply go away once its value
2261 // has been replaced.
2262 Val &OtherV = Other.Vals[Vals[i].OtherVNI->id];
2263 bool EraseImpDef = OtherV.ErasableImplicitDef &&
2264 OtherV.Resolution == CR_Keep;
2265 if (!Def.isBlock()) {
2267 // Remove <def,read-undef> flags. This def is now a partial redef.
2268 // Also remove <def,dead> flags since the joined live range will
2269 // continue past this instruction.
2270 for (MIOperands MO(Indexes->getInstructionFromIndex(Def));
2271 MO.isValid(); ++MO) {
2272 if (MO->isReg() && MO->isDef() && MO->getReg() == Reg) {
2273 MO->setIsUndef(EraseImpDef);
2274 MO->setIsDead(false);
2278 // This value will reach instructions below, but we need to make sure
2279 // the live range also reaches the instruction at Def.
2281 EndPoints.push_back(Def);
2283 DEBUG(dbgs() << "\t\tpruned " << PrintReg(Other.Reg) << " at " << Def
2284 << ": " << Other.LR << '\n');
2289 if (isPrunedValue(i, Other)) {
2290 // This value is ultimately a copy of a pruned value in LR or Other.LR.
2291 // We can no longer trust the value mapping computed by
2292 // computeAssignment(), the value that was originally copied could have
2294 LIS->pruneValue(LR, Def, &EndPoints);
2295 DEBUG(dbgs() << "\t\tpruned all of " << PrintReg(Reg) << " at "
2296 << Def << ": " << LR << '\n');
2301 llvm_unreachable("Unresolved conflicts");
2306 void JoinVals::pruneSubRegValues(LiveInterval &LI, unsigned &ShrinkMask)
2308 // Look for values being erased.
2309 bool DidPrune = false;
2310 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2311 if (Vals[i].Resolution != CR_Erase)
2314 // Check subranges at the point where the copy will be removed.
2315 SlotIndex Def = LR.getValNumInfo(i)->def;
2316 for (LiveInterval::SubRange &S : LI.subranges()) {
2317 LiveQueryResult Q = S.Query(Def);
2319 // If a subrange starts at the copy then an undefined value has been
2320 // copied and we must remove that subrange value as well.
2321 VNInfo *ValueOut = Q.valueOutOrDead();
2322 if (ValueOut != nullptr && Q.valueIn() == nullptr) {
2323 DEBUG(dbgs() << "\t\tPrune sublane " << format("%04X", S.LaneMask)
2324 << " at " << Def << "\n");
2325 LIS->pruneValue(S, Def, nullptr);
2327 // Mark value number as unused.
2328 ValueOut->markUnused();
2331 // If a subrange ends at the copy, then a value was copied but only
2332 // partially used later. Shrink the subregister range apropriately.
2333 if (Q.valueIn() != nullptr && Q.valueOut() == nullptr) {
2334 DEBUG(dbgs() << "\t\tDead uses at sublane "
2335 << format("%04X", S.LaneMask) << " at " << Def << "\n");
2336 ShrinkMask |= S.LaneMask;
2341 LI.removeEmptySubRanges();
2344 void JoinVals::eraseInstrs(SmallPtrSetImpl<MachineInstr*> &ErasedInstrs,
2345 SmallVectorImpl<unsigned> &ShrinkRegs) {
2346 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2347 // Get the def location before markUnused() below invalidates it.
2348 SlotIndex Def = LR.getValNumInfo(i)->def;
2349 switch (Vals[i].Resolution) {
2351 // If an IMPLICIT_DEF value is pruned, it doesn't serve a purpose any
2352 // longer. The IMPLICIT_DEF instructions are only inserted by
2353 // PHIElimination to guarantee that all PHI predecessors have a value.
2354 if (!Vals[i].ErasableImplicitDef || !Vals[i].Pruned)
2356 // Remove value number i from LR.
2357 VNInfo *VNI = LR.getValNumInfo(i);
2358 LR.removeValNo(VNI);
2359 // Note that this VNInfo is reused and still referenced in NewVNInfo,
2360 // make it appear like an unused value number.
2362 DEBUG(dbgs() << "\t\tremoved " << i << '@' << Def << ": " << LR << '\n');
2367 MachineInstr *MI = Indexes->getInstructionFromIndex(Def);
2368 assert(MI && "No instruction to erase");
2370 unsigned Reg = MI->getOperand(1).getReg();
2371 if (TargetRegisterInfo::isVirtualRegister(Reg) &&
2372 Reg != CP.getSrcReg() && Reg != CP.getDstReg())
2373 ShrinkRegs.push_back(Reg);
2375 ErasedInstrs.insert(MI);
2376 DEBUG(dbgs() << "\t\terased:\t" << Def << '\t' << *MI);
2377 LIS->RemoveMachineInstrFromMaps(MI);
2378 MI->eraseFromParent();
2387 bool RegisterCoalescer::joinSubRegRanges(LiveRange &LRange, LiveRange &RRange,
2389 const CoalescerPair &CP) {
2390 SmallVector<VNInfo*, 16> NewVNInfo;
2391 JoinVals RHSVals(RRange, CP.getSrcReg(), CP.getSrcIdx(), LaneMask,
2392 NewVNInfo, CP, LIS, TRI, true, true);
2393 JoinVals LHSVals(LRange, CP.getDstReg(), CP.getDstIdx(), LaneMask,
2394 NewVNInfo, CP, LIS, TRI, true, true);
2396 // Compute NewVNInfo and resolve conflicts (see also joinVirtRegs())
2397 // We should be able to resolve all conflicts here as we could successfully do
2398 // it on the mainrange already. There is however a problem when multiple
2399 // ranges get mapped to the "overflow" lane mask bit which creates unexpected
2401 if (!LHSVals.mapValues(RHSVals) || !RHSVals.mapValues(LHSVals)) {
2402 DEBUG(dbgs() << "*** Couldn't join subrange!\n");
2405 if (!LHSVals.resolveConflicts(RHSVals) ||
2406 !RHSVals.resolveConflicts(LHSVals)) {
2407 DEBUG(dbgs() << "*** Couldn't join subrange!\n");
2411 // The merging algorithm in LiveInterval::join() can't handle conflicting
2412 // value mappings, so we need to remove any live ranges that overlap a
2413 // CR_Replace resolution. Collect a set of end points that can be used to
2414 // restore the live range after joining.
2415 SmallVector<SlotIndex, 8> EndPoints;
2416 LHSVals.pruneValues(RHSVals, EndPoints, false);
2417 RHSVals.pruneValues(LHSVals, EndPoints, false);
2422 // Join RRange into LHS.
2423 LRange.join(RRange, LHSVals.getAssignments(), RHSVals.getAssignments(),
2426 DEBUG(dbgs() << "\t\tjoined lanes: " << LRange << "\n");
2427 if (EndPoints.empty())
2430 // Recompute the parts of the live range we had to remove because of
2431 // CR_Replace conflicts.
2432 DEBUG(dbgs() << "\t\trestoring liveness to " << EndPoints.size()
2433 << " points: " << LRange << '\n');
2434 LIS->extendToIndices(LRange, EndPoints);
2438 bool RegisterCoalescer::mergeSubRangeInto(LiveInterval &LI,
2439 const LiveRange &ToMerge,
2440 unsigned LaneMask, CoalescerPair &CP) {
2441 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
2442 for (LiveInterval::SubRange &R : LI.subranges()) {
2443 unsigned RMask = R.LaneMask;
2444 // LaneMask of subregisters common to subrange R and ToMerge.
2445 unsigned Common = RMask & LaneMask;
2446 // There is nothing to do without common subregs.
2450 DEBUG(dbgs() << format("\t\tCopy+Merge %04X into %04X\n", RMask, Common));
2451 // LaneMask of subregisters contained in the R range but not in ToMerge,
2452 // they have to split into their own subrange.
2453 unsigned LRest = RMask & ~LaneMask;
2454 LiveInterval::SubRange *CommonRange;
2457 DEBUG(dbgs() << format("\t\tReduce Lane to %04X\n", LRest));
2458 // Duplicate SubRange for newly merged common stuff.
2459 CommonRange = LI.createSubRangeFrom(Allocator, Common, R);
2461 // Reuse the existing range.
2462 R.LaneMask = Common;
2465 LiveRange RangeCopy(ToMerge, Allocator);
2466 if (!joinSubRegRanges(*CommonRange, RangeCopy, Common, CP))
2471 if (LaneMask != 0) {
2472 DEBUG(dbgs() << format("\t\tNew Lane %04X\n", LaneMask));
2473 LI.createSubRangeFrom(Allocator, LaneMask, ToMerge);
2478 bool RegisterCoalescer::joinVirtRegs(CoalescerPair &CP) {
2479 SmallVector<VNInfo*, 16> NewVNInfo;
2480 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
2481 LiveInterval &LHS = LIS->getInterval(CP.getDstReg());
2482 bool TrackSubRegLiveness = MRI->tracksSubRegLiveness();
2483 JoinVals RHSVals(RHS, CP.getSrcReg(), CP.getSrcIdx(), 0, NewVNInfo, CP, LIS,
2484 TRI, false, TrackSubRegLiveness);
2485 JoinVals LHSVals(LHS, CP.getDstReg(), CP.getDstIdx(), 0, NewVNInfo, CP, LIS,
2486 TRI, false, TrackSubRegLiveness);
2488 DEBUG(dbgs() << "\t\tRHS = " << RHS
2489 << "\n\t\tLHS = " << LHS
2492 // First compute NewVNInfo and the simple value mappings.
2493 // Detect impossible conflicts early.
2494 if (!LHSVals.mapValues(RHSVals) || !RHSVals.mapValues(LHSVals))
2497 // Some conflicts can only be resolved after all values have been mapped.
2498 if (!LHSVals.resolveConflicts(RHSVals) || !RHSVals.resolveConflicts(LHSVals))
2501 // All clear, the live ranges can be merged.
2502 if (RHS.hasSubRanges() || LHS.hasSubRanges()) {
2503 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
2505 // Transform lanemasks from the LHS to masks in the coalesced register and
2506 // create initial subranges if necessary.
2507 unsigned DstIdx = CP.getDstIdx();
2508 if (!LHS.hasSubRanges()) {
2509 unsigned Mask = DstIdx == 0 ? CP.getNewRC()->getLaneMask()
2510 : TRI->getSubRegIndexLaneMask(DstIdx);
2511 // LHS must support subregs or we wouldn't be in this codepath.
2513 LHS.createSubRangeFrom(Allocator, Mask, LHS);
2514 } else if (DstIdx != 0) {
2515 // Transform LHS lanemasks to new register class if necessary.
2516 for (LiveInterval::SubRange &R : LHS.subranges()) {
2517 unsigned Mask = TRI->composeSubRegIndexLaneMask(DstIdx, R.LaneMask);
2521 DEBUG(dbgs() << "\t\tLHST = " << PrintReg(CP.getDstReg())
2522 << ' ' << LHS << '\n');
2524 // Determine lanemasks of RHS in the coalesced register and merge subranges.
2525 unsigned SrcIdx = CP.getSrcIdx();
2527 if (!RHS.hasSubRanges()) {
2528 unsigned Mask = SrcIdx == 0 ? CP.getNewRC()->getLaneMask()
2529 : TRI->getSubRegIndexLaneMask(SrcIdx);
2530 if (!mergeSubRangeInto(LHS, RHS, Mask, CP))
2533 // Pair up subranges and merge.
2534 for (LiveInterval::SubRange &R : RHS.subranges()) {
2535 unsigned Mask = TRI->composeSubRegIndexLaneMask(SrcIdx, R.LaneMask);
2536 if (!mergeSubRangeInto(LHS, R, Mask, CP)) {
2543 // This shouldn't have happened :-(
2544 // However we are aware of at least one existing problem where we
2545 // can't merge subranges when multiple ranges end up in the
2546 // "overflow bit" 32. As a workaround we drop all subregister ranges
2547 // which means we loose some precision but are back to a well defined
2549 assert((CP.getNewRC()->getLaneMask() & 0x80000000u)
2550 && "SubRange merge should only fail when merging into bit 32.");
2551 DEBUG(dbgs() << "\tSubrange join aborted!\n");
2552 LHS.clearSubRanges();
2553 RHS.clearSubRanges();
2555 DEBUG(dbgs() << "\tJoined SubRanges " << LHS << "\n");
2557 LHSVals.pruneSubRegValues(LHS, ShrinkMask);
2558 RHSVals.pruneSubRegValues(LHS, ShrinkMask);
2562 // The merging algorithm in LiveInterval::join() can't handle conflicting
2563 // value mappings, so we need to remove any live ranges that overlap a
2564 // CR_Replace resolution. Collect a set of end points that can be used to
2565 // restore the live range after joining.
2566 SmallVector<SlotIndex, 8> EndPoints;
2567 LHSVals.pruneValues(RHSVals, EndPoints, true);
2568 RHSVals.pruneValues(LHSVals, EndPoints, true);
2570 // Erase COPY and IMPLICIT_DEF instructions. This may cause some external
2571 // registers to require trimming.
2572 SmallVector<unsigned, 8> ShrinkRegs;
2573 LHSVals.eraseInstrs(ErasedInstrs, ShrinkRegs);
2574 RHSVals.eraseInstrs(ErasedInstrs, ShrinkRegs);
2575 while (!ShrinkRegs.empty())
2576 LIS->shrinkToUses(&LIS->getInterval(ShrinkRegs.pop_back_val()));
2578 // Join RHS into LHS.
2579 LHS.join(RHS, LHSVals.getAssignments(), RHSVals.getAssignments(), NewVNInfo);
2581 // Kill flags are going to be wrong if the live ranges were overlapping.
2582 // Eventually, we should simply clear all kill flags when computing live
2583 // ranges. They are reinserted after register allocation.
2584 MRI->clearKillFlags(LHS.reg);
2585 MRI->clearKillFlags(RHS.reg);
2587 if (!EndPoints.empty()) {
2588 // Recompute the parts of the live range we had to remove because of
2589 // CR_Replace conflicts.
2590 DEBUG(dbgs() << "\t\trestoring liveness to " << EndPoints.size()
2591 << " points: " << LHS << '\n');
2592 LIS->extendToIndices((LiveRange&)LHS, EndPoints);
2598 bool RegisterCoalescer::joinIntervals(CoalescerPair &CP) {
2599 return CP.isPhys() ? joinReservedPhysReg(CP) : joinVirtRegs(CP);
2603 /// Information concerning MBB coalescing priority.
2604 struct MBBPriorityInfo {
2605 MachineBasicBlock *MBB;
2609 MBBPriorityInfo(MachineBasicBlock *mbb, unsigned depth, bool issplit)
2610 : MBB(mbb), Depth(depth), IsSplit(issplit) {}
2614 /// C-style comparator that sorts first based on the loop depth of the basic
2615 /// block (the unsigned), and then on the MBB number.
2617 /// EnableGlobalCopies assumes that the primary sort key is loop depth.
2618 static int compareMBBPriority(const MBBPriorityInfo *LHS,
2619 const MBBPriorityInfo *RHS) {
2620 // Deeper loops first
2621 if (LHS->Depth != RHS->Depth)
2622 return LHS->Depth > RHS->Depth ? -1 : 1;
2624 // Try to unsplit critical edges next.
2625 if (LHS->IsSplit != RHS->IsSplit)
2626 return LHS->IsSplit ? -1 : 1;
2628 // Prefer blocks that are more connected in the CFG. This takes care of
2629 // the most difficult copies first while intervals are short.
2630 unsigned cl = LHS->MBB->pred_size() + LHS->MBB->succ_size();
2631 unsigned cr = RHS->MBB->pred_size() + RHS->MBB->succ_size();
2633 return cl > cr ? -1 : 1;
2635 // As a last resort, sort by block number.
2636 return LHS->MBB->getNumber() < RHS->MBB->getNumber() ? -1 : 1;
2639 /// \returns true if the given copy uses or defines a local live range.
2640 static bool isLocalCopy(MachineInstr *Copy, const LiveIntervals *LIS) {
2641 if (!Copy->isCopy())
2644 if (Copy->getOperand(1).isUndef())
2647 unsigned SrcReg = Copy->getOperand(1).getReg();
2648 unsigned DstReg = Copy->getOperand(0).getReg();
2649 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)
2650 || TargetRegisterInfo::isPhysicalRegister(DstReg))
2653 return LIS->intervalIsInOneMBB(LIS->getInterval(SrcReg))
2654 || LIS->intervalIsInOneMBB(LIS->getInterval(DstReg));
2657 bool RegisterCoalescer::
2658 copyCoalesceWorkList(MutableArrayRef<MachineInstr*> CurrList) {
2659 bool Progress = false;
2660 for (unsigned i = 0, e = CurrList.size(); i != e; ++i) {
2663 // Skip instruction pointers that have already been erased, for example by
2664 // dead code elimination.
2665 if (ErasedInstrs.erase(CurrList[i])) {
2666 CurrList[i] = nullptr;
2670 bool Success = joinCopy(CurrList[i], Again);
2671 Progress |= Success;
2672 if (Success || !Again)
2673 CurrList[i] = nullptr;
2679 RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) {
2680 DEBUG(dbgs() << MBB->getName() << ":\n");
2682 // Collect all copy-like instructions in MBB. Don't start coalescing anything
2683 // yet, it might invalidate the iterator.
2684 const unsigned PrevSize = WorkList.size();
2685 if (JoinGlobalCopies) {
2686 // Coalesce copies bottom-up to coalesce local defs before local uses. They
2687 // are not inherently easier to resolve, but slightly preferable until we
2688 // have local live range splitting. In particular this is required by
2689 // cmp+jmp macro fusion.
2690 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2692 if (!MII->isCopyLike())
2694 if (isLocalCopy(&(*MII), LIS))
2695 LocalWorkList.push_back(&(*MII));
2697 WorkList.push_back(&(*MII));
2701 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2703 if (MII->isCopyLike())
2704 WorkList.push_back(MII);
2706 // Try coalescing the collected copies immediately, and remove the nulls.
2707 // This prevents the WorkList from getting too large since most copies are
2708 // joinable on the first attempt.
2709 MutableArrayRef<MachineInstr*>
2710 CurrList(WorkList.begin() + PrevSize, WorkList.end());
2711 if (copyCoalesceWorkList(CurrList))
2712 WorkList.erase(std::remove(WorkList.begin() + PrevSize, WorkList.end(),
2713 (MachineInstr*)nullptr), WorkList.end());
2716 void RegisterCoalescer::coalesceLocals() {
2717 copyCoalesceWorkList(LocalWorkList);
2718 for (unsigned j = 0, je = LocalWorkList.size(); j != je; ++j) {
2719 if (LocalWorkList[j])
2720 WorkList.push_back(LocalWorkList[j]);
2722 LocalWorkList.clear();
2725 void RegisterCoalescer::joinAllIntervals() {
2726 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
2727 assert(WorkList.empty() && LocalWorkList.empty() && "Old data still around.");
2729 std::vector<MBBPriorityInfo> MBBs;
2730 MBBs.reserve(MF->size());
2731 for (MachineFunction::iterator I = MF->begin(), E = MF->end();I != E;++I){
2732 MachineBasicBlock *MBB = I;
2733 MBBs.push_back(MBBPriorityInfo(MBB, Loops->getLoopDepth(MBB),
2734 JoinSplitEdges && isSplitEdge(MBB)));
2736 array_pod_sort(MBBs.begin(), MBBs.end(), compareMBBPriority);
2738 // Coalesce intervals in MBB priority order.
2739 unsigned CurrDepth = UINT_MAX;
2740 for (unsigned i = 0, e = MBBs.size(); i != e; ++i) {
2741 // Try coalescing the collected local copies for deeper loops.
2742 if (JoinGlobalCopies && MBBs[i].Depth < CurrDepth) {
2744 CurrDepth = MBBs[i].Depth;
2746 copyCoalesceInMBB(MBBs[i].MBB);
2750 // Joining intervals can allow other intervals to be joined. Iteratively join
2751 // until we make no progress.
2752 while (copyCoalesceWorkList(WorkList))
2756 void RegisterCoalescer::releaseMemory() {
2757 ErasedInstrs.clear();
2760 InflateRegs.clear();
2763 bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) {
2765 MRI = &fn.getRegInfo();
2766 TM = &fn.getTarget();
2767 const TargetSubtargetInfo &STI = fn.getSubtarget();
2768 TRI = STI.getRegisterInfo();
2769 TII = STI.getInstrInfo();
2770 LIS = &getAnalysis<LiveIntervals>();
2771 AA = &getAnalysis<AliasAnalysis>();
2772 Loops = &getAnalysis<MachineLoopInfo>();
2773 if (EnableGlobalCopies == cl::BOU_UNSET)
2774 JoinGlobalCopies = STI.useMachineScheduler();
2776 JoinGlobalCopies = (EnableGlobalCopies == cl::BOU_TRUE);
2778 // The MachineScheduler does not currently require JoinSplitEdges. This will
2779 // either be enabled unconditionally or replaced by a more general live range
2780 // splitting optimization.
2781 JoinSplitEdges = EnableJoinSplits;
2783 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
2784 << "********** Function: " << MF->getName() << '\n');
2786 if (VerifyCoalescing)
2787 MF->verify(this, "Before register coalescing");
2789 RegClassInfo.runOnMachineFunction(fn);
2791 // Join (coalesce) intervals if requested.
2795 // After deleting a lot of copies, register classes may be less constrained.
2796 // Removing sub-register operands may allow GR32_ABCD -> GR32 and DPR_VFP2 ->
2798 array_pod_sort(InflateRegs.begin(), InflateRegs.end());
2799 InflateRegs.erase(std::unique(InflateRegs.begin(), InflateRegs.end()),
2801 DEBUG(dbgs() << "Trying to inflate " << InflateRegs.size() << " regs.\n");
2802 for (unsigned i = 0, e = InflateRegs.size(); i != e; ++i) {
2803 unsigned Reg = InflateRegs[i];
2804 if (MRI->reg_nodbg_empty(Reg))
2806 if (MRI->recomputeRegClass(Reg)) {
2807 DEBUG(dbgs() << PrintReg(Reg) << " inflated to "
2808 << TRI->getRegClassName(MRI->getRegClass(Reg)) << '\n');
2809 LiveInterval &LI = LIS->getInterval(Reg);
2810 unsigned MaxMask = MRI->getMaxLaneMaskForVReg(Reg);
2812 // If the inflated register class does not support subregisters anymore
2813 // remove the subranges.
2814 LI.clearSubRanges();
2817 // If subranges are still supported, then the same subregs should still
2819 for (LiveInterval::SubRange &S : LI.subranges()) {
2820 assert ((S.LaneMask & ~MaxMask) == 0);
2829 if (VerifyCoalescing)
2830 MF->verify(this, "After register coalescing");
2834 void RegisterCoalescer::print(raw_ostream &O, const Module* m) const {