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 static cl::opt<bool> UseTerminalRule("terminal-rule",
62 cl::desc("Apply the terminal rule"),
63 cl::init(false), cl::Hidden);
65 /// Temporary flag to test critical edge unsplitting.
67 EnableJoinSplits("join-splitedges",
68 cl::desc("Coalesce copies on split edges (default=subtarget)"), cl::Hidden);
70 /// Temporary flag to test global copy optimization.
71 static cl::opt<cl::boolOrDefault>
72 EnableGlobalCopies("join-globalcopies",
73 cl::desc("Coalesce copies that span blocks (default=subtarget)"),
74 cl::init(cl::BOU_UNSET), cl::Hidden);
77 VerifyCoalescing("verify-coalescing",
78 cl::desc("Verify machine instrs before and after register coalescing"),
82 class RegisterCoalescer : public MachineFunctionPass,
83 private LiveRangeEdit::Delegate {
85 MachineRegisterInfo* MRI;
86 const TargetMachine* TM;
87 const TargetRegisterInfo* TRI;
88 const TargetInstrInfo* TII;
90 const MachineLoopInfo* Loops;
92 RegisterClassInfo RegClassInfo;
94 /// A LaneMask to remember on which subregister live ranges we need to call
95 /// shrinkToUses() later.
98 /// True if the main range of the currently coalesced intervals should be
99 /// checked for smaller live intervals.
100 bool ShrinkMainRange;
102 /// \brief True if the coalescer should aggressively coalesce global copies
103 /// in favor of keeping local copies.
104 bool JoinGlobalCopies;
106 /// \brief True if the coalescer should aggressively coalesce fall-thru
107 /// blocks exclusively containing copies.
110 /// Copy instructions yet to be coalesced.
111 SmallVector<MachineInstr*, 8> WorkList;
112 SmallVector<MachineInstr*, 8> LocalWorkList;
114 /// Set of instruction pointers that have been erased, and
115 /// that may be present in WorkList.
116 SmallPtrSet<MachineInstr*, 8> ErasedInstrs;
118 /// Dead instructions that are about to be deleted.
119 SmallVector<MachineInstr*, 8> DeadDefs;
121 /// Virtual registers to be considered for register class inflation.
122 SmallVector<unsigned, 8> InflateRegs;
124 /// Recursively eliminate dead defs in DeadDefs.
125 void eliminateDeadDefs();
127 /// LiveRangeEdit callback for eliminateDeadDefs().
128 void LRE_WillEraseInstruction(MachineInstr *MI) override;
130 /// Coalesce the LocalWorkList.
131 void coalesceLocals();
133 /// Join compatible live intervals
134 void joinAllIntervals();
136 /// Coalesce copies in the specified MBB, putting
137 /// copies that cannot yet be coalesced into WorkList.
138 void copyCoalesceInMBB(MachineBasicBlock *MBB);
140 /// Tries to coalesce all copies in CurrList. Returns true if any progress
142 bool copyCoalesceWorkList(MutableArrayRef<MachineInstr*> CurrList);
144 /// Attempt to join intervals corresponding to SrcReg/DstReg, which are the
145 /// src/dst of the copy instruction CopyMI. This returns true if the copy
146 /// was successfully coalesced away. If it is not currently possible to
147 /// coalesce this interval, but it may be possible if other things get
148 /// coalesced, then it returns true by reference in 'Again'.
149 bool joinCopy(MachineInstr *TheCopy, bool &Again);
151 /// Attempt to join these two intervals. On failure, this
152 /// returns false. The output "SrcInt" will not have been modified, so we
153 /// can use this information below to update aliases.
154 bool joinIntervals(CoalescerPair &CP);
156 /// Attempt joining two virtual registers. Return true on success.
157 bool joinVirtRegs(CoalescerPair &CP);
159 /// Attempt joining with a reserved physreg.
160 bool joinReservedPhysReg(CoalescerPair &CP);
162 /// Add the LiveRange @p ToMerge as a subregister liverange of @p LI.
163 /// Subranges in @p LI which only partially interfere with the desired
164 /// LaneMask are split as necessary. @p LaneMask are the lanes that
165 /// @p ToMerge will occupy in the coalescer register. @p LI has its subrange
166 /// lanemasks already adjusted to the coalesced register.
167 /// @returns false if live range conflicts couldn't get resolved.
168 bool mergeSubRangeInto(LiveInterval &LI, const LiveRange &ToMerge,
169 unsigned LaneMask, CoalescerPair &CP);
171 /// Join the liveranges of two subregisters. Joins @p RRange into
172 /// @p LRange, @p RRange may be invalid afterwards.
173 /// @returns false if live range conflicts couldn't get resolved.
174 bool joinSubRegRanges(LiveRange &LRange, LiveRange &RRange,
175 unsigned LaneMask, const CoalescerPair &CP);
177 /// We found a non-trivially-coalescable copy. If the source value number is
178 /// defined by a copy from the destination reg see if we can merge these two
179 /// destination reg valno# into a single value number, eliminating a copy.
180 /// This returns true if an interval was modified.
181 bool adjustCopiesBackFrom(const CoalescerPair &CP, MachineInstr *CopyMI);
183 /// Return true if there are definitions of IntB
184 /// other than BValNo val# that can reach uses of AValno val# of IntA.
185 bool hasOtherReachingDefs(LiveInterval &IntA, LiveInterval &IntB,
186 VNInfo *AValNo, VNInfo *BValNo);
188 /// We found a non-trivially-coalescable copy.
189 /// If the source value number is defined by a commutable instruction and
190 /// its other operand is coalesced to the copy dest register, see if we
191 /// can transform the copy into a noop by commuting the definition.
192 /// This returns true if an interval was modified.
193 bool removeCopyByCommutingDef(const CoalescerPair &CP,MachineInstr *CopyMI);
195 /// If the source of a copy is defined by a
196 /// trivial computation, replace the copy by rematerialize the definition.
197 bool reMaterializeTrivialDef(const CoalescerPair &CP, MachineInstr *CopyMI,
200 /// Return true if a copy involving a physreg should be joined.
201 bool canJoinPhys(const CoalescerPair &CP);
203 /// Replace all defs and uses of SrcReg to DstReg and update the subregister
204 /// number if it is not zero. If DstReg is a physical register and the
205 /// existing subregister number of the def / use being updated is not zero,
206 /// make sure to set it to the correct physical subregister.
207 void updateRegDefsUses(unsigned SrcReg, unsigned DstReg, unsigned SubIdx);
209 /// Handle copies of undef values.
210 /// Returns true if @p CopyMI was a copy of an undef value and eliminated.
211 bool eliminateUndefCopy(MachineInstr *CopyMI);
213 /// Check whether or not we should apply the terminal rule on the
214 /// destination (Dst) of \p Copy.
215 /// When the terminal rule applies, Copy is not profitable to
217 /// Dst is terminal if it has exactly one affinity (Dst, Src) and
218 /// at least one interference (Dst, Dst2). If Dst is terminal, the
219 /// terminal rule consists in checking that at least one of
220 /// interfering node, say Dst2, has an affinity of equal or greater
222 /// In that case, Dst2 and Dst will not be able to be both coalesced
223 /// with Src. Since Dst2 exposes more coalescing opportunities than
224 /// Dst, we can drop \p Copy.
225 bool applyTerminalRule(const MachineInstr &Copy) const;
227 /// Check whether or not \p LI is composed by multiple connected
228 /// components and if that is the case, fix that.
229 void splitNewRanges(LiveInterval *LI) {
230 ConnectedVNInfoEqClasses ConEQ(*LIS);
231 unsigned NumComps = ConEQ.Classify(LI);
234 SmallVector<LiveInterval*, 8> NewComps(1, LI);
235 for (unsigned i = 1; i != NumComps; ++i) {
236 unsigned VReg = MRI->createVirtualRegister(MRI->getRegClass(LI->reg));
237 NewComps.push_back(&LIS->createEmptyInterval(VReg));
240 ConEQ.Distribute(&NewComps[0], *MRI);
243 /// Wrapper method for \see LiveIntervals::shrinkToUses.
244 /// This method does the proper fixing of the live-ranges when the afore
245 /// mentioned method returns true.
246 void shrinkToUses(LiveInterval *LI,
247 SmallVectorImpl<MachineInstr * > *Dead = nullptr) {
248 if (LIS->shrinkToUses(LI, Dead))
249 // We may have created multiple connected components, split them.
254 static char ID; ///< Class identification, replacement for typeinfo
255 RegisterCoalescer() : MachineFunctionPass(ID) {
256 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
259 void getAnalysisUsage(AnalysisUsage &AU) const override;
261 void releaseMemory() override;
263 /// This is the pass entry point.
264 bool runOnMachineFunction(MachineFunction&) override;
266 /// Implement the dump method.
267 void print(raw_ostream &O, const Module* = nullptr) const override;
269 } // end anonymous namespace
271 char &llvm::RegisterCoalescerID = RegisterCoalescer::ID;
273 INITIALIZE_PASS_BEGIN(RegisterCoalescer, "simple-register-coalescing",
274 "Simple Register Coalescing", false, false)
275 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
276 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
277 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
278 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
279 INITIALIZE_PASS_END(RegisterCoalescer, "simple-register-coalescing",
280 "Simple Register Coalescing", false, false)
282 char RegisterCoalescer::ID = 0;
284 static bool isMoveInstr(const TargetRegisterInfo &tri, const MachineInstr *MI,
285 unsigned &Src, unsigned &Dst,
286 unsigned &SrcSub, unsigned &DstSub) {
288 Dst = MI->getOperand(0).getReg();
289 DstSub = MI->getOperand(0).getSubReg();
290 Src = MI->getOperand(1).getReg();
291 SrcSub = MI->getOperand(1).getSubReg();
292 } else if (MI->isSubregToReg()) {
293 Dst = MI->getOperand(0).getReg();
294 DstSub = tri.composeSubRegIndices(MI->getOperand(0).getSubReg(),
295 MI->getOperand(3).getImm());
296 Src = MI->getOperand(2).getReg();
297 SrcSub = MI->getOperand(2).getSubReg();
303 /// Return true if this block should be vacated by the coalescer to eliminate
304 /// branches. The important cases to handle in the coalescer are critical edges
305 /// split during phi elimination which contain only copies. Simple blocks that
306 /// contain non-branches should also be vacated, but this can be handled by an
307 /// earlier pass similar to early if-conversion.
308 static bool isSplitEdge(const MachineBasicBlock *MBB) {
309 if (MBB->pred_size() != 1 || MBB->succ_size() != 1)
312 for (const auto &MI : *MBB) {
313 if (!MI.isCopyLike() && !MI.isUnconditionalBranch())
319 bool CoalescerPair::setRegisters(const MachineInstr *MI) {
323 Flipped = CrossClass = false;
325 unsigned Src, Dst, SrcSub, DstSub;
326 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
328 Partial = SrcSub || DstSub;
330 // If one register is a physreg, it must be Dst.
331 if (TargetRegisterInfo::isPhysicalRegister(Src)) {
332 if (TargetRegisterInfo::isPhysicalRegister(Dst))
335 std::swap(SrcSub, DstSub);
339 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
341 if (TargetRegisterInfo::isPhysicalRegister(Dst)) {
342 // Eliminate DstSub on a physreg.
344 Dst = TRI.getSubReg(Dst, DstSub);
345 if (!Dst) return false;
349 // Eliminate SrcSub by picking a corresponding Dst superregister.
351 Dst = TRI.getMatchingSuperReg(Dst, SrcSub, MRI.getRegClass(Src));
352 if (!Dst) return false;
353 } else if (!MRI.getRegClass(Src)->contains(Dst)) {
357 // Both registers are virtual.
358 const TargetRegisterClass *SrcRC = MRI.getRegClass(Src);
359 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
361 // Both registers have subreg indices.
362 if (SrcSub && DstSub) {
363 // Copies between different sub-registers are never coalescable.
364 if (Src == Dst && SrcSub != DstSub)
367 NewRC = TRI.getCommonSuperRegClass(SrcRC, SrcSub, DstRC, DstSub,
372 // SrcReg will be merged with a sub-register of DstReg.
374 NewRC = TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSub);
376 // DstReg will be merged with a sub-register of SrcReg.
378 NewRC = TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSub);
380 // This is a straight copy without sub-registers.
381 NewRC = TRI.getCommonSubClass(DstRC, SrcRC);
384 // The combined constraint may be impossible to satisfy.
388 // Prefer SrcReg to be a sub-register of DstReg.
389 // FIXME: Coalescer should support subregs symmetrically.
390 if (DstIdx && !SrcIdx) {
392 std::swap(SrcIdx, DstIdx);
396 CrossClass = NewRC != DstRC || NewRC != SrcRC;
398 // Check our invariants
399 assert(TargetRegisterInfo::isVirtualRegister(Src) && "Src must be virtual");
400 assert(!(TargetRegisterInfo::isPhysicalRegister(Dst) && DstSub) &&
401 "Cannot have a physical SubIdx");
407 bool CoalescerPair::flip() {
408 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
410 std::swap(SrcReg, DstReg);
411 std::swap(SrcIdx, DstIdx);
416 bool CoalescerPair::isCoalescable(const MachineInstr *MI) const {
419 unsigned Src, Dst, SrcSub, DstSub;
420 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
423 // Find the virtual register that is SrcReg.
426 std::swap(SrcSub, DstSub);
427 } else if (Src != SrcReg) {
431 // Now check that Dst matches DstReg.
432 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
433 if (!TargetRegisterInfo::isPhysicalRegister(Dst))
435 assert(!DstIdx && !SrcIdx && "Inconsistent CoalescerPair state.");
436 // DstSub could be set for a physreg from INSERT_SUBREG.
438 Dst = TRI.getSubReg(Dst, DstSub);
441 return DstReg == Dst;
442 // This is a partial register copy. Check that the parts match.
443 return TRI.getSubReg(DstReg, SrcSub) == Dst;
445 // DstReg is virtual.
448 // Registers match, do the subregisters line up?
449 return TRI.composeSubRegIndices(SrcIdx, SrcSub) ==
450 TRI.composeSubRegIndices(DstIdx, DstSub);
454 void RegisterCoalescer::getAnalysisUsage(AnalysisUsage &AU) const {
455 AU.setPreservesCFG();
456 AU.addRequired<AliasAnalysis>();
457 AU.addRequired<LiveIntervals>();
458 AU.addPreserved<LiveIntervals>();
459 AU.addPreserved<SlotIndexes>();
460 AU.addRequired<MachineLoopInfo>();
461 AU.addPreserved<MachineLoopInfo>();
462 AU.addPreservedID(MachineDominatorsID);
463 MachineFunctionPass::getAnalysisUsage(AU);
466 void RegisterCoalescer::eliminateDeadDefs() {
467 SmallVector<unsigned, 8> NewRegs;
468 LiveRangeEdit(nullptr, NewRegs, *MF, *LIS,
469 nullptr, this).eliminateDeadDefs(DeadDefs);
472 void RegisterCoalescer::LRE_WillEraseInstruction(MachineInstr *MI) {
473 // MI may be in WorkList. Make sure we don't visit it.
474 ErasedInstrs.insert(MI);
477 bool RegisterCoalescer::adjustCopiesBackFrom(const CoalescerPair &CP,
478 MachineInstr *CopyMI) {
479 assert(!CP.isPartial() && "This doesn't work for partial copies.");
480 assert(!CP.isPhys() && "This doesn't work for physreg copies.");
483 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
485 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
486 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
488 // We have a non-trivially-coalescable copy with IntA being the source and
489 // IntB being the dest, thus this defines a value number in IntB. If the
490 // source value number (in IntA) is defined by a copy from B, see if we can
491 // merge these two pieces of B into a single value number, eliminating a copy.
496 // B1 = A3 <- this copy
498 // In this case, B0 can be extended to where the B1 copy lives, allowing the
499 // B1 value number to be replaced with B0 (which simplifies the B
502 // BValNo is a value number in B that is defined by a copy from A. 'B1' in
503 // the example above.
504 LiveInterval::iterator BS = IntB.FindSegmentContaining(CopyIdx);
505 if (BS == IntB.end()) return false;
506 VNInfo *BValNo = BS->valno;
508 // Get the location that B is defined at. Two options: either this value has
509 // an unknown definition point or it is defined at CopyIdx. If unknown, we
511 if (BValNo->def != CopyIdx) return false;
513 // AValNo is the value number in A that defines the copy, A3 in the example.
514 SlotIndex CopyUseIdx = CopyIdx.getRegSlot(true);
515 LiveInterval::iterator AS = IntA.FindSegmentContaining(CopyUseIdx);
516 // The live segment might not exist after fun with physreg coalescing.
517 if (AS == IntA.end()) return false;
518 VNInfo *AValNo = AS->valno;
520 // If AValNo is defined as a copy from IntB, we can potentially process this.
521 // Get the instruction that defines this value number.
522 MachineInstr *ACopyMI = LIS->getInstructionFromIndex(AValNo->def);
523 // Don't allow any partial copies, even if isCoalescable() allows them.
524 if (!CP.isCoalescable(ACopyMI) || !ACopyMI->isFullCopy())
527 // Get the Segment in IntB that this value number starts with.
528 LiveInterval::iterator ValS =
529 IntB.FindSegmentContaining(AValNo->def.getPrevSlot());
530 if (ValS == IntB.end())
533 // Make sure that the end of the live segment is inside the same block as
535 MachineInstr *ValSEndInst =
536 LIS->getInstructionFromIndex(ValS->end.getPrevSlot());
537 if (!ValSEndInst || ValSEndInst->getParent() != CopyMI->getParent())
540 // Okay, we now know that ValS ends in the same block that the CopyMI
541 // live-range starts. If there are no intervening live segments between them
542 // in IntB, we can merge them.
543 if (ValS+1 != BS) return false;
545 DEBUG(dbgs() << "Extending: " << PrintReg(IntB.reg, TRI));
547 SlotIndex FillerStart = ValS->end, FillerEnd = BS->start;
548 // We are about to delete CopyMI, so need to remove it as the 'instruction
549 // that defines this value #'. Update the valnum with the new defining
551 BValNo->def = FillerStart;
553 // Okay, we can merge them. We need to insert a new liverange:
554 // [ValS.end, BS.begin) of either value number, then we merge the
555 // two value numbers.
556 IntB.addSegment(LiveInterval::Segment(FillerStart, FillerEnd, BValNo));
558 // Okay, merge "B1" into the same value number as "B0".
559 if (BValNo != ValS->valno)
560 IntB.MergeValueNumberInto(BValNo, ValS->valno);
562 // Do the same for the subregister segments.
563 for (LiveInterval::SubRange &S : IntB.subranges()) {
564 VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
565 S.addSegment(LiveInterval::Segment(FillerStart, FillerEnd, SubBValNo));
566 VNInfo *SubValSNo = S.getVNInfoAt(AValNo->def.getPrevSlot());
567 if (SubBValNo != SubValSNo)
568 S.MergeValueNumberInto(SubBValNo, SubValSNo);
571 DEBUG(dbgs() << " result = " << IntB << '\n');
573 // If the source instruction was killing the source register before the
574 // merge, unset the isKill marker given the live range has been extended.
575 int UIdx = ValSEndInst->findRegisterUseOperandIdx(IntB.reg, true);
577 ValSEndInst->getOperand(UIdx).setIsKill(false);
580 // Rewrite the copy. If the copy instruction was killing the destination
581 // register before the merge, find the last use and trim the live range. That
582 // will also add the isKill marker.
583 CopyMI->substituteRegister(IntA.reg, IntB.reg, 0, *TRI);
584 if (AS->end == CopyIdx)
591 bool RegisterCoalescer::hasOtherReachingDefs(LiveInterval &IntA,
595 // If AValNo has PHI kills, conservatively assume that IntB defs can reach
597 if (LIS->hasPHIKill(IntA, AValNo))
600 for (LiveRange::Segment &ASeg : IntA.segments) {
601 if (ASeg.valno != AValNo) continue;
602 LiveInterval::iterator BI =
603 std::upper_bound(IntB.begin(), IntB.end(), ASeg.start);
604 if (BI != IntB.begin())
606 for (; BI != IntB.end() && ASeg.end >= BI->start; ++BI) {
607 if (BI->valno == BValNo)
609 if (BI->start <= ASeg.start && BI->end > ASeg.start)
611 if (BI->start > ASeg.start && BI->start < ASeg.end)
618 /// Copy segements with value number @p SrcValNo from liverange @p Src to live
619 /// range @Dst and use value number @p DstValNo there.
620 static void addSegmentsWithValNo(LiveRange &Dst, VNInfo *DstValNo,
621 const LiveRange &Src, const VNInfo *SrcValNo)
623 for (const LiveRange::Segment &S : Src.segments) {
624 if (S.valno != SrcValNo)
626 Dst.addSegment(LiveRange::Segment(S.start, S.end, DstValNo));
630 bool RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP,
631 MachineInstr *CopyMI) {
632 assert(!CP.isPhys());
635 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
637 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
639 // We found a non-trivially-coalescable copy with IntA being the source and
640 // IntB being the dest, thus this defines a value number in IntB. If the
641 // source value number (in IntA) is defined by a commutable instruction and
642 // its other operand is coalesced to the copy dest register, see if we can
643 // transform the copy into a noop by commuting the definition. For example,
645 // A3 = op A2 B0<kill>
647 // B1 = A3 <- this copy
649 // = op A3 <- more uses
653 // B2 = op B0 A2<kill>
655 // B1 = B2 <- now an identity copy
657 // = op B2 <- more uses
659 // BValNo is a value number in B that is defined by a copy from A. 'B1' in
660 // the example above.
661 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
662 VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx);
663 assert(BValNo != nullptr && BValNo->def == CopyIdx);
665 // AValNo is the value number in A that defines the copy, A3 in the example.
666 VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getRegSlot(true));
667 assert(AValNo && !AValNo->isUnused() && "COPY source not live");
668 if (AValNo->isPHIDef())
670 MachineInstr *DefMI = LIS->getInstructionFromIndex(AValNo->def);
673 if (!DefMI->isCommutable())
675 // If DefMI is a two-address instruction then commuting it will change the
676 // destination register.
677 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
678 assert(DefIdx != -1);
680 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
682 unsigned Op1, Op2, NewDstIdx;
683 if (!TII->findCommutedOpIndices(DefMI, Op1, Op2))
687 else if (Op2 == UseOpIdx)
692 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
693 unsigned NewReg = NewDstMO.getReg();
694 if (NewReg != IntB.reg || !IntB.Query(AValNo->def).isKill())
697 // Make sure there are no other definitions of IntB that would reach the
698 // uses which the new definition can reach.
699 if (hasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
702 // If some of the uses of IntA.reg is already coalesced away, return false.
703 // It's not possible to determine whether it's safe to perform the coalescing.
704 for (MachineOperand &MO : MRI->use_nodbg_operands(IntA.reg)) {
705 MachineInstr *UseMI = MO.getParent();
706 unsigned OpNo = &MO - &UseMI->getOperand(0);
707 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI);
708 LiveInterval::iterator US = IntA.FindSegmentContaining(UseIdx);
709 if (US == IntA.end() || US->valno != AValNo)
711 // If this use is tied to a def, we can't rewrite the register.
712 if (UseMI->isRegTiedToDefOperand(OpNo))
716 DEBUG(dbgs() << "\tremoveCopyByCommutingDef: " << AValNo->def << '\t'
719 // At this point we have decided that it is legal to do this
720 // transformation. Start by commuting the instruction.
721 MachineBasicBlock *MBB = DefMI->getParent();
722 MachineInstr *NewMI = TII->commuteInstruction(DefMI);
725 if (TargetRegisterInfo::isVirtualRegister(IntA.reg) &&
726 TargetRegisterInfo::isVirtualRegister(IntB.reg) &&
727 !MRI->constrainRegClass(IntB.reg, MRI->getRegClass(IntA.reg)))
729 if (NewMI != DefMI) {
730 LIS->ReplaceMachineInstrInMaps(DefMI, NewMI);
731 MachineBasicBlock::iterator Pos = DefMI;
732 MBB->insert(Pos, NewMI);
736 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
745 // Update uses of IntA of the specific Val# with IntB.
746 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(IntA.reg),
748 UI != UE; /* ++UI is below because of possible MI removal */) {
749 MachineOperand &UseMO = *UI;
753 MachineInstr *UseMI = UseMO.getParent();
754 if (UseMI->isDebugValue()) {
755 // FIXME These don't have an instruction index. Not clear we have enough
756 // info to decide whether to do this replacement or not. For now do it.
757 UseMO.setReg(NewReg);
760 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI).getRegSlot(true);
761 LiveInterval::iterator US = IntA.FindSegmentContaining(UseIdx);
762 assert(US != IntA.end() && "Use must be live");
763 if (US->valno != AValNo)
765 // Kill flags are no longer accurate. They are recomputed after RA.
766 UseMO.setIsKill(false);
767 if (TargetRegisterInfo::isPhysicalRegister(NewReg))
768 UseMO.substPhysReg(NewReg, *TRI);
770 UseMO.setReg(NewReg);
773 if (!UseMI->isCopy())
775 if (UseMI->getOperand(0).getReg() != IntB.reg ||
776 UseMI->getOperand(0).getSubReg())
779 // This copy will become a noop. If it's defining a new val#, merge it into
781 SlotIndex DefIdx = UseIdx.getRegSlot();
782 VNInfo *DVNI = IntB.getVNInfoAt(DefIdx);
785 DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI);
786 assert(DVNI->def == DefIdx);
787 BValNo = IntB.MergeValueNumberInto(DVNI, BValNo);
788 for (LiveInterval::SubRange &S : IntB.subranges()) {
789 VNInfo *SubDVNI = S.getVNInfoAt(DefIdx);
792 VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
793 assert(SubBValNo->def == CopyIdx);
794 S.MergeValueNumberInto(SubDVNI, SubBValNo);
797 ErasedInstrs.insert(UseMI);
798 LIS->RemoveMachineInstrFromMaps(UseMI);
799 UseMI->eraseFromParent();
802 // Extend BValNo by merging in IntA live segments of AValNo. Val# definition
804 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
805 if (IntB.hasSubRanges()) {
806 if (!IntA.hasSubRanges()) {
807 unsigned Mask = MRI->getMaxLaneMaskForVReg(IntA.reg);
808 IntA.createSubRangeFrom(Allocator, Mask, IntA);
810 SlotIndex AIdx = CopyIdx.getRegSlot(true);
811 for (LiveInterval::SubRange &SA : IntA.subranges()) {
812 VNInfo *ASubValNo = SA.getVNInfoAt(AIdx);
813 assert(ASubValNo != nullptr);
815 unsigned AMask = SA.LaneMask;
816 for (LiveInterval::SubRange &SB : IntB.subranges()) {
817 unsigned BMask = SB.LaneMask;
818 unsigned Common = BMask & AMask;
823 dbgs() << format("\t\tCopy+Merge %04X into %04X\n", BMask, Common));
824 unsigned BRest = BMask & ~AMask;
825 LiveInterval::SubRange *CommonRange;
828 DEBUG(dbgs() << format("\t\tReduce Lane to %04X\n", BRest));
829 // Duplicate SubRange for newly merged common stuff.
830 CommonRange = IntB.createSubRangeFrom(Allocator, Common, SB);
832 // We van reuse the L SubRange.
833 SB.LaneMask = Common;
836 LiveRange RangeCopy(SB, Allocator);
838 VNInfo *BSubValNo = CommonRange->getVNInfoAt(CopyIdx);
839 assert(BSubValNo->def == CopyIdx);
840 BSubValNo->def = ASubValNo->def;
841 addSegmentsWithValNo(*CommonRange, BSubValNo, SA, ASubValNo);
845 DEBUG(dbgs() << format("\t\tNew Lane %04X\n", AMask));
846 LiveRange *NewRange = IntB.createSubRange(Allocator, AMask);
847 VNInfo *BSubValNo = NewRange->getNextValue(CopyIdx, Allocator);
848 addSegmentsWithValNo(*NewRange, BSubValNo, SA, ASubValNo);
853 BValNo->def = AValNo->def;
854 addSegmentsWithValNo(IntB, BValNo, IntA, AValNo);
855 DEBUG(dbgs() << "\t\textended: " << IntB << '\n');
857 LIS->removeVRegDefAt(IntA, AValNo->def);
859 DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n');
864 /// Returns true if @p MI defines the full vreg @p Reg, as opposed to just
865 /// defining a subregister.
866 static bool definesFullReg(const MachineInstr &MI, unsigned Reg) {
867 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) &&
868 "This code cannot handle physreg aliasing");
869 for (const MachineOperand &Op : MI.operands()) {
870 if (!Op.isReg() || !Op.isDef() || Op.getReg() != Reg)
872 // Return true if we define the full register or don't care about the value
873 // inside other subregisters.
874 if (Op.getSubReg() == 0 || Op.isUndef())
880 bool RegisterCoalescer::reMaterializeTrivialDef(const CoalescerPair &CP,
881 MachineInstr *CopyMI,
884 unsigned SrcReg = CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg();
885 unsigned SrcIdx = CP.isFlipped() ? CP.getDstIdx() : CP.getSrcIdx();
886 unsigned DstReg = CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg();
887 unsigned DstIdx = CP.isFlipped() ? CP.getSrcIdx() : CP.getDstIdx();
888 if (TargetRegisterInfo::isPhysicalRegister(SrcReg))
891 LiveInterval &SrcInt = LIS->getInterval(SrcReg);
892 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI);
893 VNInfo *ValNo = SrcInt.Query(CopyIdx).valueIn();
894 assert(ValNo && "CopyMI input register not live");
895 if (ValNo->isPHIDef() || ValNo->isUnused())
897 MachineInstr *DefMI = LIS->getInstructionFromIndex(ValNo->def);
900 if (DefMI->isCopyLike()) {
904 if (!TII->isAsCheapAsAMove(DefMI))
906 if (!TII->isTriviallyReMaterializable(DefMI, AA))
908 if (!definesFullReg(*DefMI, SrcReg))
910 bool SawStore = false;
911 if (!DefMI->isSafeToMove(AA, SawStore))
913 const MCInstrDesc &MCID = DefMI->getDesc();
914 if (MCID.getNumDefs() != 1)
916 // Only support subregister destinations when the def is read-undef.
917 MachineOperand &DstOperand = CopyMI->getOperand(0);
918 unsigned CopyDstReg = DstOperand.getReg();
919 if (DstOperand.getSubReg() && !DstOperand.isUndef())
922 // If both SrcIdx and DstIdx are set, correct rematerialization would widen
923 // the register substantially (beyond both source and dest size). This is bad
924 // for performance since it can cascade through a function, introducing many
925 // extra spills and fills (e.g. ARM can easily end up copying QQQQPR registers
926 // around after a few subreg copies).
927 if (SrcIdx && DstIdx)
930 const TargetRegisterClass *DefRC = TII->getRegClass(MCID, 0, TRI, *MF);
931 if (!DefMI->isImplicitDef()) {
932 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
933 unsigned NewDstReg = DstReg;
935 unsigned NewDstIdx = TRI->composeSubRegIndices(CP.getSrcIdx(),
936 DefMI->getOperand(0).getSubReg());
938 NewDstReg = TRI->getSubReg(DstReg, NewDstIdx);
940 // Finally, make sure that the physical subregister that will be
941 // constructed later is permitted for the instruction.
942 if (!DefRC->contains(NewDstReg))
945 // Theoretically, some stack frame reference could exist. Just make sure
946 // it hasn't actually happened.
947 assert(TargetRegisterInfo::isVirtualRegister(DstReg) &&
948 "Only expect to deal with virtual or physical registers");
952 MachineBasicBlock *MBB = CopyMI->getParent();
953 MachineBasicBlock::iterator MII =
954 std::next(MachineBasicBlock::iterator(CopyMI));
955 TII->reMaterialize(*MBB, MII, DstReg, SrcIdx, DefMI, *TRI);
956 MachineInstr *NewMI = std::prev(MII);
958 // In a situation like the following:
959 // %vreg0:subreg = instr ; DefMI, subreg = DstIdx
960 // %vreg1 = copy %vreg0:subreg ; CopyMI, SrcIdx = 0
961 // instead of widening %vreg1 to the register class of %vreg0 simply do:
963 const TargetRegisterClass *NewRC = CP.getNewRC();
965 MachineOperand &DefMO = NewMI->getOperand(0);
966 if (DefMO.getSubReg() == DstIdx) {
967 assert(SrcIdx == 0 && CP.isFlipped()
968 && "Shouldn't have SrcIdx+DstIdx at this point");
969 const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
970 const TargetRegisterClass *CommonRC =
971 TRI->getCommonSubClass(DefRC, DstRC);
972 if (CommonRC != nullptr) {
980 LIS->ReplaceMachineInstrInMaps(CopyMI, NewMI);
981 CopyMI->eraseFromParent();
982 ErasedInstrs.insert(CopyMI);
984 // NewMI may have dead implicit defs (E.g. EFLAGS for MOV<bits>r0 on X86).
985 // We need to remember these so we can add intervals once we insert
986 // NewMI into SlotIndexes.
987 SmallVector<unsigned, 4> NewMIImplDefs;
988 for (unsigned i = NewMI->getDesc().getNumOperands(),
989 e = NewMI->getNumOperands(); i != e; ++i) {
990 MachineOperand &MO = NewMI->getOperand(i);
991 if (MO.isReg() && MO.isDef()) {
992 assert(MO.isImplicit() && MO.isDead() &&
993 TargetRegisterInfo::isPhysicalRegister(MO.getReg()));
994 NewMIImplDefs.push_back(MO.getReg());
998 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
999 unsigned NewIdx = NewMI->getOperand(0).getSubReg();
1001 if (DefRC != nullptr) {
1003 NewRC = TRI->getMatchingSuperRegClass(NewRC, DefRC, NewIdx);
1005 NewRC = TRI->getCommonSubClass(NewRC, DefRC);
1006 assert(NewRC && "subreg chosen for remat incompatible with instruction");
1008 MRI->setRegClass(DstReg, NewRC);
1010 updateRegDefsUses(DstReg, DstReg, DstIdx);
1011 NewMI->getOperand(0).setSubReg(NewIdx);
1012 } else if (NewMI->getOperand(0).getReg() != CopyDstReg) {
1013 // The New instruction may be defining a sub-register of what's actually
1014 // been asked for. If so it must implicitly define the whole thing.
1015 assert(TargetRegisterInfo::isPhysicalRegister(DstReg) &&
1016 "Only expect virtual or physical registers in remat");
1017 NewMI->getOperand(0).setIsDead(true);
1018 NewMI->addOperand(MachineOperand::CreateReg(CopyDstReg,
1022 // Record small dead def live-ranges for all the subregisters
1023 // of the destination register.
1024 // Otherwise, variables that live through may miss some
1025 // interferences, thus creating invalid allocation.
1027 // vreg1 = somedef ; vreg1 GR8
1028 // vreg2 = remat ; vreg2 GR32
1029 // CL = COPY vreg2.sub_8bit
1030 // = somedef vreg1 ; vreg1 GR8
1032 // vreg1 = somedef ; vreg1 GR8
1033 // ECX<def, dead> = remat ; CL<imp-def>
1034 // = somedef vreg1 ; vreg1 GR8
1035 // vreg1 will see the inteferences with CL but not with CH since
1036 // no live-ranges would have been created for ECX.
1038 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
1039 for (MCRegUnitIterator Units(NewMI->getOperand(0).getReg(), TRI);
1040 Units.isValid(); ++Units)
1041 if (LiveRange *LR = LIS->getCachedRegUnit(*Units))
1042 LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
1045 if (NewMI->getOperand(0).getSubReg())
1046 NewMI->getOperand(0).setIsUndef();
1048 // CopyMI may have implicit operands, transfer them over to the newly
1049 // rematerialized instruction. And update implicit def interval valnos.
1050 for (unsigned i = CopyMI->getDesc().getNumOperands(),
1051 e = CopyMI->getNumOperands(); i != e; ++i) {
1052 MachineOperand &MO = CopyMI->getOperand(i);
1054 assert(MO.isImplicit() && "No explicit operands after implict operands.");
1055 // Discard VReg implicit defs.
1056 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
1057 NewMI->addOperand(MO);
1062 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
1063 for (unsigned i = 0, e = NewMIImplDefs.size(); i != e; ++i) {
1064 unsigned Reg = NewMIImplDefs[i];
1065 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units)
1066 if (LiveRange *LR = LIS->getCachedRegUnit(*Units))
1067 LR->createDeadDef(NewMIIdx.getRegSlot(), LIS->getVNInfoAllocator());
1070 DEBUG(dbgs() << "Remat: " << *NewMI);
1073 // The source interval can become smaller because we removed a use.
1074 shrinkToUses(&SrcInt, &DeadDefs);
1075 if (!DeadDefs.empty()) {
1076 // If the virtual SrcReg is completely eliminated, update all DBG_VALUEs
1077 // to describe DstReg instead.
1078 for (MachineOperand &UseMO : MRI->use_operands(SrcReg)) {
1079 MachineInstr *UseMI = UseMO.getParent();
1080 if (UseMI->isDebugValue()) {
1081 UseMO.setReg(DstReg);
1082 DEBUG(dbgs() << "\t\tupdated: " << *UseMI);
1085 eliminateDeadDefs();
1091 bool RegisterCoalescer::eliminateUndefCopy(MachineInstr *CopyMI) {
1092 // ProcessImpicitDefs may leave some copies of <undef> values, it only removes
1093 // local variables. When we have a copy like:
1095 // %vreg1 = COPY %vreg2<undef>
1097 // We delete the copy and remove the corresponding value number from %vreg1.
1098 // Any uses of that value number are marked as <undef>.
1100 // Note that we do not query CoalescerPair here but redo isMoveInstr as the
1101 // CoalescerPair may have a new register class with adjusted subreg indices
1103 unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
1104 isMoveInstr(*TRI, CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx);
1106 SlotIndex Idx = LIS->getInstructionIndex(CopyMI);
1107 const LiveInterval &SrcLI = LIS->getInterval(SrcReg);
1108 // CopyMI is undef iff SrcReg is not live before the instruction.
1109 if (SrcSubIdx != 0 && SrcLI.hasSubRanges()) {
1110 unsigned SrcMask = TRI->getSubRegIndexLaneMask(SrcSubIdx);
1111 for (const LiveInterval::SubRange &SR : SrcLI.subranges()) {
1112 if ((SR.LaneMask & SrcMask) == 0)
1117 } else if (SrcLI.liveAt(Idx))
1120 DEBUG(dbgs() << "\tEliminating copy of <undef> value\n");
1122 // Remove any DstReg segments starting at the instruction.
1123 LiveInterval &DstLI = LIS->getInterval(DstReg);
1124 SlotIndex RegIndex = Idx.getRegSlot();
1125 // Remove value or merge with previous one in case of a subregister def.
1126 if (VNInfo *PrevVNI = DstLI.getVNInfoAt(Idx)) {
1127 VNInfo *VNI = DstLI.getVNInfoAt(RegIndex);
1128 DstLI.MergeValueNumberInto(VNI, PrevVNI);
1130 // The affected subregister segments can be removed.
1131 unsigned DstMask = TRI->getSubRegIndexLaneMask(DstSubIdx);
1132 for (LiveInterval::SubRange &SR : DstLI.subranges()) {
1133 if ((SR.LaneMask & DstMask) == 0)
1136 VNInfo *SVNI = SR.getVNInfoAt(RegIndex);
1137 assert(SVNI != nullptr && SlotIndex::isSameInstr(SVNI->def, RegIndex));
1138 SR.removeValNo(SVNI);
1140 DstLI.removeEmptySubRanges();
1142 LIS->removeVRegDefAt(DstLI, RegIndex);
1144 // Mark uses as undef.
1145 for (MachineOperand &MO : MRI->reg_nodbg_operands(DstReg)) {
1146 if (MO.isDef() /*|| MO.isUndef()*/)
1148 const MachineInstr &MI = *MO.getParent();
1149 SlotIndex UseIdx = LIS->getInstructionIndex(&MI);
1150 unsigned UseMask = TRI->getSubRegIndexLaneMask(MO.getSubReg());
1152 if (UseMask != ~0u && DstLI.hasSubRanges()) {
1154 for (const LiveInterval::SubRange &SR : DstLI.subranges()) {
1155 if ((SR.LaneMask & UseMask) == 0)
1157 if (SR.liveAt(UseIdx)) {
1163 isLive = DstLI.liveAt(UseIdx);
1166 MO.setIsUndef(true);
1167 DEBUG(dbgs() << "\tnew undef: " << UseIdx << '\t' << MI);
1172 void RegisterCoalescer::updateRegDefsUses(unsigned SrcReg,
1175 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
1176 LiveInterval *DstInt = DstIsPhys ? nullptr : &LIS->getInterval(DstReg);
1178 SmallPtrSet<MachineInstr*, 8> Visited;
1179 for (MachineRegisterInfo::reg_instr_iterator
1180 I = MRI->reg_instr_begin(SrcReg), E = MRI->reg_instr_end();
1182 MachineInstr *UseMI = &*(I++);
1184 // Each instruction can only be rewritten once because sub-register
1185 // composition is not always idempotent. When SrcReg != DstReg, rewriting
1186 // the UseMI operands removes them from the SrcReg use-def chain, but when
1187 // SrcReg is DstReg we could encounter UseMI twice if it has multiple
1188 // operands mentioning the virtual register.
1189 if (SrcReg == DstReg && !Visited.insert(UseMI).second)
1192 SmallVector<unsigned,8> Ops;
1194 std::tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
1196 // If SrcReg wasn't read, it may still be the case that DstReg is live-in
1197 // because SrcReg is a sub-register.
1198 if (DstInt && !Reads && SubIdx)
1199 Reads = DstInt->liveAt(LIS->getInstructionIndex(UseMI));
1201 // Replace SrcReg with DstReg in all UseMI operands.
1202 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
1203 MachineOperand &MO = UseMI->getOperand(Ops[i]);
1205 // Adjust <undef> flags in case of sub-register joins. We don't want to
1206 // turn a full def into a read-modify-write sub-register def and vice
1208 if (SubIdx && MO.isDef())
1209 MO.setIsUndef(!Reads);
1211 // A subreg use of a partially undef (super) register may be a complete
1212 // undef use now and then has to be marked that way.
1213 if (SubIdx != 0 && MO.isUse() && MRI->shouldTrackSubRegLiveness(DstReg)) {
1214 if (!DstInt->hasSubRanges()) {
1215 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
1216 unsigned Mask = MRI->getMaxLaneMaskForVReg(DstInt->reg);
1217 DstInt->createSubRangeFrom(Allocator, Mask, *DstInt);
1219 unsigned Mask = TRI->getSubRegIndexLaneMask(SubIdx);
1220 bool IsUndef = true;
1221 SlotIndex MIIdx = UseMI->isDebugValue()
1222 ? LIS->getSlotIndexes()->getIndexBefore(UseMI)
1223 : LIS->getInstructionIndex(UseMI);
1224 SlotIndex UseIdx = MIIdx.getRegSlot(true);
1225 for (LiveInterval::SubRange &S : DstInt->subranges()) {
1226 if ((S.LaneMask & Mask) == 0)
1228 if (S.liveAt(UseIdx)) {
1234 MO.setIsUndef(true);
1235 // We found out some subregister use is actually reading an undefined
1236 // value. In some cases the whole vreg has become undefined at this
1237 // point so we have to potentially shrink the main range if the
1238 // use was ending a live segment there.
1239 LiveQueryResult Q = DstInt->Query(MIIdx);
1240 if (Q.valueOut() == nullptr)
1241 ShrinkMainRange = true;
1246 MO.substPhysReg(DstReg, *TRI);
1248 MO.substVirtReg(DstReg, SubIdx, *TRI);
1252 dbgs() << "\t\tupdated: ";
1253 if (!UseMI->isDebugValue())
1254 dbgs() << LIS->getInstructionIndex(UseMI) << "\t";
1260 bool RegisterCoalescer::canJoinPhys(const CoalescerPair &CP) {
1261 // Always join simple intervals that are defined by a single copy from a
1262 // reserved register. This doesn't increase register pressure, so it is
1263 // always beneficial.
1264 if (!MRI->isReserved(CP.getDstReg())) {
1265 DEBUG(dbgs() << "\tCan only merge into reserved registers.\n");
1269 LiveInterval &JoinVInt = LIS->getInterval(CP.getSrcReg());
1270 if (JoinVInt.containsOneValue())
1273 DEBUG(dbgs() << "\tCannot join complex intervals into reserved register.\n");
1277 bool RegisterCoalescer::joinCopy(MachineInstr *CopyMI, bool &Again) {
1280 DEBUG(dbgs() << LIS->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
1282 CoalescerPair CP(*TRI);
1283 if (!CP.setRegisters(CopyMI)) {
1284 DEBUG(dbgs() << "\tNot coalescable.\n");
1288 if (CP.getNewRC()) {
1289 auto SrcRC = MRI->getRegClass(CP.getSrcReg());
1290 auto DstRC = MRI->getRegClass(CP.getDstReg());
1291 unsigned SrcIdx = CP.getSrcIdx();
1292 unsigned DstIdx = CP.getDstIdx();
1293 if (CP.isFlipped()) {
1294 std::swap(SrcIdx, DstIdx);
1295 std::swap(SrcRC, DstRC);
1297 if (!TRI->shouldCoalesce(CopyMI, SrcRC, SrcIdx, DstRC, DstIdx,
1299 DEBUG(dbgs() << "\tSubtarget bailed on coalescing.\n");
1304 // Dead code elimination. This really should be handled by MachineDCE, but
1305 // sometimes dead copies slip through, and we can't generate invalid live
1307 if (!CP.isPhys() && CopyMI->allDefsAreDead()) {
1308 DEBUG(dbgs() << "\tCopy is dead.\n");
1309 DeadDefs.push_back(CopyMI);
1310 eliminateDeadDefs();
1314 // Eliminate undefs.
1315 if (!CP.isPhys() && eliminateUndefCopy(CopyMI)) {
1316 LIS->RemoveMachineInstrFromMaps(CopyMI);
1317 CopyMI->eraseFromParent();
1318 return false; // Not coalescable.
1321 // Coalesced copies are normally removed immediately, but transformations
1322 // like removeCopyByCommutingDef() can inadvertently create identity copies.
1323 // When that happens, just join the values and remove the copy.
1324 if (CP.getSrcReg() == CP.getDstReg()) {
1325 LiveInterval &LI = LIS->getInterval(CP.getSrcReg());
1326 DEBUG(dbgs() << "\tCopy already coalesced: " << LI << '\n');
1327 const SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI);
1328 LiveQueryResult LRQ = LI.Query(CopyIdx);
1329 if (VNInfo *DefVNI = LRQ.valueDefined()) {
1330 VNInfo *ReadVNI = LRQ.valueIn();
1331 assert(ReadVNI && "No value before copy and no <undef> flag.");
1332 assert(ReadVNI != DefVNI && "Cannot read and define the same value.");
1333 LI.MergeValueNumberInto(DefVNI, ReadVNI);
1335 // Process subregister liveranges.
1336 for (LiveInterval::SubRange &S : LI.subranges()) {
1337 LiveQueryResult SLRQ = S.Query(CopyIdx);
1338 if (VNInfo *SDefVNI = SLRQ.valueDefined()) {
1339 VNInfo *SReadVNI = SLRQ.valueIn();
1340 S.MergeValueNumberInto(SDefVNI, SReadVNI);
1343 DEBUG(dbgs() << "\tMerged values: " << LI << '\n');
1345 LIS->RemoveMachineInstrFromMaps(CopyMI);
1346 CopyMI->eraseFromParent();
1350 // Enforce policies.
1352 DEBUG(dbgs() << "\tConsidering merging " << PrintReg(CP.getSrcReg(), TRI)
1353 << " with " << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx())
1355 if (!canJoinPhys(CP)) {
1356 // Before giving up coalescing, if definition of source is defined by
1357 // trivial computation, try rematerializing it.
1359 if (reMaterializeTrivialDef(CP, CopyMI, IsDefCopy))
1362 Again = true; // May be possible to coalesce later.
1366 // When possible, let DstReg be the larger interval.
1367 if (!CP.isPartial() && LIS->getInterval(CP.getSrcReg()).size() >
1368 LIS->getInterval(CP.getDstReg()).size())
1372 dbgs() << "\tConsidering merging to "
1373 << TRI->getRegClassName(CP.getNewRC()) << " with ";
1374 if (CP.getDstIdx() && CP.getSrcIdx())
1375 dbgs() << PrintReg(CP.getDstReg()) << " in "
1376 << TRI->getSubRegIndexName(CP.getDstIdx()) << " and "
1377 << PrintReg(CP.getSrcReg()) << " in "
1378 << TRI->getSubRegIndexName(CP.getSrcIdx()) << '\n';
1380 dbgs() << PrintReg(CP.getSrcReg(), TRI) << " in "
1381 << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n';
1386 ShrinkMainRange = false;
1388 // Okay, attempt to join these two intervals. On failure, this returns false.
1389 // Otherwise, if one of the intervals being joined is a physreg, this method
1390 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1391 // been modified, so we can use this information below to update aliases.
1392 if (!joinIntervals(CP)) {
1393 // Coalescing failed.
1395 // If definition of source is defined by trivial computation, try
1396 // rematerializing it.
1398 if (reMaterializeTrivialDef(CP, CopyMI, IsDefCopy))
1401 // If we can eliminate the copy without merging the live segments, do so
1403 if (!CP.isPartial() && !CP.isPhys()) {
1404 if (adjustCopiesBackFrom(CP, CopyMI) ||
1405 removeCopyByCommutingDef(CP, CopyMI)) {
1406 LIS->RemoveMachineInstrFromMaps(CopyMI);
1407 CopyMI->eraseFromParent();
1408 DEBUG(dbgs() << "\tTrivial!\n");
1413 // Otherwise, we are unable to join the intervals.
1414 DEBUG(dbgs() << "\tInterference!\n");
1415 Again = true; // May be possible to coalesce later.
1419 // Coalescing to a virtual register that is of a sub-register class of the
1420 // other. Make sure the resulting register is set to the right register class.
1421 if (CP.isCrossClass()) {
1423 MRI->setRegClass(CP.getDstReg(), CP.getNewRC());
1426 // Removing sub-register copies can ease the register class constraints.
1427 // Make sure we attempt to inflate the register class of DstReg.
1428 if (!CP.isPhys() && RegClassInfo.isProperSubClass(CP.getNewRC()))
1429 InflateRegs.push_back(CP.getDstReg());
1431 // CopyMI has been erased by joinIntervals at this point. Remove it from
1432 // ErasedInstrs since copyCoalesceWorkList() won't add a successful join back
1433 // to the work list. This keeps ErasedInstrs from growing needlessly.
1434 ErasedInstrs.erase(CopyMI);
1436 // Rewrite all SrcReg operands to DstReg.
1437 // Also update DstReg operands to include DstIdx if it is set.
1439 updateRegDefsUses(CP.getDstReg(), CP.getDstReg(), CP.getDstIdx());
1440 updateRegDefsUses(CP.getSrcReg(), CP.getDstReg(), CP.getSrcIdx());
1442 // Shrink subregister ranges if necessary.
1443 if (ShrinkMask != 0) {
1444 LiveInterval &LI = LIS->getInterval(CP.getDstReg());
1445 for (LiveInterval::SubRange &S : LI.subranges()) {
1446 if ((S.LaneMask & ShrinkMask) == 0)
1448 DEBUG(dbgs() << "Shrink LaneUses (Lane "
1449 << format("%04X", S.LaneMask) << ")\n");
1450 LIS->shrinkToUses(S, LI.reg);
1453 if (ShrinkMainRange) {
1454 LiveInterval &LI = LIS->getInterval(CP.getDstReg());
1458 // SrcReg is guaranteed to be the register whose live interval that is
1460 LIS->removeInterval(CP.getSrcReg());
1462 // Update regalloc hint.
1463 TRI->updateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF);
1466 dbgs() << "\tSuccess: " << PrintReg(CP.getSrcReg(), TRI, CP.getSrcIdx())
1467 << " -> " << PrintReg(CP.getDstReg(), TRI, CP.getDstIdx()) << '\n';
1468 dbgs() << "\tResult = ";
1470 dbgs() << PrintReg(CP.getDstReg(), TRI);
1472 dbgs() << LIS->getInterval(CP.getDstReg());
1480 bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) {
1481 unsigned DstReg = CP.getDstReg();
1482 assert(CP.isPhys() && "Must be a physreg copy");
1483 assert(MRI->isReserved(DstReg) && "Not a reserved register");
1484 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
1485 DEBUG(dbgs() << "\t\tRHS = " << RHS << '\n');
1487 assert(RHS.containsOneValue() && "Invalid join with reserved register");
1489 // Optimization for reserved registers like ESP. We can only merge with a
1490 // reserved physreg if RHS has a single value that is a copy of DstReg.
1491 // The live range of the reserved register will look like a set of dead defs
1492 // - we don't properly track the live range of reserved registers.
1494 // Deny any overlapping intervals. This depends on all the reserved
1495 // register live ranges to look like dead defs.
1496 for (MCRegUnitIterator UI(DstReg, TRI); UI.isValid(); ++UI)
1497 if (RHS.overlaps(LIS->getRegUnit(*UI))) {
1498 DEBUG(dbgs() << "\t\tInterference: " << PrintRegUnit(*UI, TRI) << '\n');
1502 // Skip any value computations, we are not adding new values to the
1503 // reserved register. Also skip merging the live ranges, the reserved
1504 // register live range doesn't need to be accurate as long as all the
1507 // Delete the identity copy.
1508 MachineInstr *CopyMI;
1509 if (CP.isFlipped()) {
1510 CopyMI = MRI->getVRegDef(RHS.reg);
1512 if (!MRI->hasOneNonDBGUse(RHS.reg)) {
1513 DEBUG(dbgs() << "\t\tMultiple vreg uses!\n");
1517 MachineInstr *DestMI = MRI->getVRegDef(RHS.reg);
1518 CopyMI = &*MRI->use_instr_nodbg_begin(RHS.reg);
1519 const SlotIndex CopyRegIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
1520 const SlotIndex DestRegIdx = LIS->getInstructionIndex(DestMI).getRegSlot();
1522 // We checked above that there are no interfering defs of the physical
1523 // register. However, for this case, where we intent to move up the def of
1524 // the physical register, we also need to check for interfering uses.
1525 SlotIndexes *Indexes = LIS->getSlotIndexes();
1526 for (SlotIndex SI = Indexes->getNextNonNullIndex(DestRegIdx);
1527 SI != CopyRegIdx; SI = Indexes->getNextNonNullIndex(SI)) {
1528 MachineInstr *MI = LIS->getInstructionFromIndex(SI);
1529 if (MI->readsRegister(DstReg, TRI)) {
1530 DEBUG(dbgs() << "\t\tInterference (read): " << *MI);
1535 // We're going to remove the copy which defines a physical reserved
1536 // register, so remove its valno, etc.
1537 DEBUG(dbgs() << "\t\tRemoving phys reg def of " << DstReg << " at "
1538 << CopyRegIdx << "\n");
1540 LIS->removePhysRegDefAt(DstReg, CopyRegIdx);
1541 // Create a new dead def at the new def location.
1542 for (MCRegUnitIterator UI(DstReg, TRI); UI.isValid(); ++UI) {
1543 LiveRange &LR = LIS->getRegUnit(*UI);
1544 LR.createDeadDef(DestRegIdx, LIS->getVNInfoAllocator());
1548 LIS->RemoveMachineInstrFromMaps(CopyMI);
1549 CopyMI->eraseFromParent();
1551 // We don't track kills for reserved registers.
1552 MRI->clearKillFlags(CP.getSrcReg());
1557 //===----------------------------------------------------------------------===//
1558 // Interference checking and interval joining
1559 //===----------------------------------------------------------------------===//
1561 // In the easiest case, the two live ranges being joined are disjoint, and
1562 // there is no interference to consider. It is quite common, though, to have
1563 // overlapping live ranges, and we need to check if the interference can be
1566 // The live range of a single SSA value forms a sub-tree of the dominator tree.
1567 // This means that two SSA values overlap if and only if the def of one value
1568 // is contained in the live range of the other value. As a special case, the
1569 // overlapping values can be defined at the same index.
1571 // The interference from an overlapping def can be resolved in these cases:
1573 // 1. Coalescable copies. The value is defined by a copy that would become an
1574 // identity copy after joining SrcReg and DstReg. The copy instruction will
1575 // be removed, and the value will be merged with the source value.
1577 // There can be several copies back and forth, causing many values to be
1578 // merged into one. We compute a list of ultimate values in the joined live
1579 // range as well as a mappings from the old value numbers.
1581 // 2. IMPLICIT_DEF. This instruction is only inserted to ensure all PHI
1582 // predecessors have a live out value. It doesn't cause real interference,
1583 // and can be merged into the value it overlaps. Like a coalescable copy, it
1584 // can be erased after joining.
1586 // 3. Copy of external value. The overlapping def may be a copy of a value that
1587 // is already in the other register. This is like a coalescable copy, but
1588 // the live range of the source register must be trimmed after erasing the
1589 // copy instruction:
1592 // %dst = COPY %ext <-- Remove this COPY, trim the live range of %ext.
1594 // 4. Clobbering undefined lanes. Vector registers are sometimes built by
1595 // defining one lane at a time:
1597 // %dst:ssub0<def,read-undef> = FOO
1599 // %dst:ssub1<def> = COPY %src
1601 // The live range of %src overlaps the %dst value defined by FOO, but
1602 // merging %src into %dst:ssub1 is only going to clobber the ssub1 lane
1603 // which was undef anyway.
1605 // The value mapping is more complicated in this case. The final live range
1606 // will have different value numbers for both FOO and BAR, but there is no
1607 // simple mapping from old to new values. It may even be necessary to add
1610 // 5. Clobbering dead lanes. A def may clobber a lane of a vector register that
1611 // is live, but never read. This can happen because we don't compute
1612 // individual live ranges per lane.
1616 // %dst:ssub1<def> = COPY %src
1618 // This kind of interference is only resolved locally. If the clobbered
1619 // lane value escapes the block, the join is aborted.
1622 /// Track information about values in a single virtual register about to be
1623 /// joined. Objects of this class are always created in pairs - one for each
1624 /// side of the CoalescerPair (or one for each lane of a side of the coalescer
1627 /// Live range we work on.
1629 /// (Main) register we work on.
1632 /// Reg (and therefore the values in this liverange) will end up as
1633 /// subregister SubIdx in the coalesced register. Either CP.DstIdx or
1635 const unsigned SubIdx;
1636 /// The LaneMask that this liverange will occupy the coalesced register. May
1637 /// be smaller than the lanemask produced by SubIdx when merging subranges.
1638 const unsigned LaneMask;
1640 /// This is true when joining sub register ranges, false when joining main
1642 const bool SubRangeJoin;
1643 /// Whether the current LiveInterval tracks subregister liveness.
1644 const bool TrackSubRegLiveness;
1646 /// Values that will be present in the final live range.
1647 SmallVectorImpl<VNInfo*> &NewVNInfo;
1649 const CoalescerPair &CP;
1651 SlotIndexes *Indexes;
1652 const TargetRegisterInfo *TRI;
1654 /// Value number assignments. Maps value numbers in LI to entries in
1655 /// NewVNInfo. This is suitable for passing to LiveInterval::join().
1656 SmallVector<int, 8> Assignments;
1658 /// Conflict resolution for overlapping values.
1659 enum ConflictResolution {
1660 /// No overlap, simply keep this value.
1663 /// Merge this value into OtherVNI and erase the defining instruction.
1664 /// Used for IMPLICIT_DEF, coalescable copies, and copies from external
1668 /// Merge this value into OtherVNI but keep the defining instruction.
1669 /// This is for the special case where OtherVNI is defined by the same
1673 /// Keep this value, and have it replace OtherVNI where possible. This
1674 /// complicates value mapping since OtherVNI maps to two different values
1675 /// before and after this def.
1676 /// Used when clobbering undefined or dead lanes.
1679 /// Unresolved conflict. Visit later when all values have been mapped.
1682 /// Unresolvable conflict. Abort the join.
1686 /// Per-value info for LI. The lane bit masks are all relative to the final
1687 /// joined register, so they can be compared directly between SrcReg and
1690 ConflictResolution Resolution;
1692 /// Lanes written by this def, 0 for unanalyzed values.
1693 unsigned WriteLanes;
1695 /// Lanes with defined values in this register. Other lanes are undef and
1696 /// safe to clobber.
1697 unsigned ValidLanes;
1699 /// Value in LI being redefined by this def.
1702 /// Value in the other live range that overlaps this def, if any.
1705 /// Is this value an IMPLICIT_DEF that can be erased?
1707 /// IMPLICIT_DEF values should only exist at the end of a basic block that
1708 /// is a predecessor to a phi-value. These IMPLICIT_DEF instructions can be
1709 /// safely erased if they are overlapping a live value in the other live
1712 /// Weird control flow graphs and incomplete PHI handling in
1713 /// ProcessImplicitDefs can very rarely create IMPLICIT_DEF values with
1714 /// longer live ranges. Such IMPLICIT_DEF values should be treated like
1716 bool ErasableImplicitDef;
1718 /// True when the live range of this value will be pruned because of an
1719 /// overlapping CR_Replace value in the other live range.
1722 /// True once Pruned above has been computed.
1723 bool PrunedComputed;
1725 Val() : Resolution(CR_Keep), WriteLanes(0), ValidLanes(0),
1726 RedefVNI(nullptr), OtherVNI(nullptr), ErasableImplicitDef(false),
1727 Pruned(false), PrunedComputed(false) {}
1729 bool isAnalyzed() const { return WriteLanes != 0; }
1732 /// One entry per value number in LI.
1733 SmallVector<Val, 8> Vals;
1735 /// Compute the bitmask of lanes actually written by DefMI.
1736 /// Set Redef if there are any partial register definitions that depend on the
1737 /// previous value of the register.
1738 unsigned computeWriteLanes(const MachineInstr *DefMI, bool &Redef) const;
1740 /// Find the ultimate value that VNI was copied from.
1741 std::pair<const VNInfo*,unsigned> followCopyChain(const VNInfo *VNI) const;
1743 bool valuesIdentical(VNInfo *Val0, VNInfo *Val1, const JoinVals &Other) const;
1745 /// Analyze ValNo in this live range, and set all fields of Vals[ValNo].
1746 /// Return a conflict resolution when possible, but leave the hard cases as
1748 /// Recursively calls computeAssignment() on this and Other, guaranteeing that
1749 /// both OtherVNI and RedefVNI have been analyzed and mapped before returning.
1750 /// The recursion always goes upwards in the dominator tree, making loops
1752 ConflictResolution analyzeValue(unsigned ValNo, JoinVals &Other);
1754 /// Compute the value assignment for ValNo in RI.
1755 /// This may be called recursively by analyzeValue(), but never for a ValNo on
1757 void computeAssignment(unsigned ValNo, JoinVals &Other);
1759 /// Assuming ValNo is going to clobber some valid lanes in Other.LR, compute
1760 /// the extent of the tainted lanes in the block.
1762 /// Multiple values in Other.LR can be affected since partial redefinitions
1763 /// can preserve previously tainted lanes.
1765 /// 1 %dst = VLOAD <-- Define all lanes in %dst
1766 /// 2 %src = FOO <-- ValNo to be joined with %dst:ssub0
1767 /// 3 %dst:ssub1 = BAR <-- Partial redef doesn't clear taint in ssub0
1768 /// 4 %dst:ssub0 = COPY %src <-- Conflict resolved, ssub0 wasn't read
1770 /// For each ValNo in Other that is affected, add an (EndIndex, TaintedLanes)
1771 /// entry to TaintedVals.
1773 /// Returns false if the tainted lanes extend beyond the basic block.
1774 bool taintExtent(unsigned, unsigned, JoinVals&,
1775 SmallVectorImpl<std::pair<SlotIndex, unsigned> >&);
1777 /// Return true if MI uses any of the given Lanes from Reg.
1778 /// This does not include partial redefinitions of Reg.
1779 bool usesLanes(const MachineInstr *MI, unsigned, unsigned, unsigned) const;
1781 /// Determine if ValNo is a copy of a value number in LR or Other.LR that will
1784 /// %dst = COPY %src
1785 /// %src = COPY %dst <-- This value to be pruned.
1786 /// %dst = COPY %src <-- This value is a copy of a pruned value.
1787 bool isPrunedValue(unsigned ValNo, JoinVals &Other);
1790 JoinVals(LiveRange &LR, unsigned Reg, unsigned SubIdx, unsigned LaneMask,
1791 SmallVectorImpl<VNInfo*> &newVNInfo, const CoalescerPair &cp,
1792 LiveIntervals *lis, const TargetRegisterInfo *TRI, bool SubRangeJoin,
1793 bool TrackSubRegLiveness)
1794 : LR(LR), Reg(Reg), SubIdx(SubIdx), LaneMask(LaneMask),
1795 SubRangeJoin(SubRangeJoin), TrackSubRegLiveness(TrackSubRegLiveness),
1796 NewVNInfo(newVNInfo), CP(cp), LIS(lis), Indexes(LIS->getSlotIndexes()),
1797 TRI(TRI), Assignments(LR.getNumValNums(), -1), Vals(LR.getNumValNums())
1800 /// Analyze defs in LR and compute a value mapping in NewVNInfo.
1801 /// Returns false if any conflicts were impossible to resolve.
1802 bool mapValues(JoinVals &Other);
1804 /// Try to resolve conflicts that require all values to be mapped.
1805 /// Returns false if any conflicts were impossible to resolve.
1806 bool resolveConflicts(JoinVals &Other);
1808 /// Prune the live range of values in Other.LR where they would conflict with
1809 /// CR_Replace values in LR. Collect end points for restoring the live range
1811 void pruneValues(JoinVals &Other, SmallVectorImpl<SlotIndex> &EndPoints,
1814 /// Removes subranges starting at copies that get removed. This sometimes
1815 /// happens when undefined subranges are copied around. These ranges contain
1816 /// no usefull information and can be removed.
1817 void pruneSubRegValues(LiveInterval &LI, unsigned &ShrinkMask);
1819 /// Erase any machine instructions that have been coalesced away.
1820 /// Add erased instructions to ErasedInstrs.
1821 /// Add foreign virtual registers to ShrinkRegs if their live range ended at
1822 /// the erased instrs.
1823 void eraseInstrs(SmallPtrSetImpl<MachineInstr*> &ErasedInstrs,
1824 SmallVectorImpl<unsigned> &ShrinkRegs);
1826 /// Remove liverange defs at places where implicit defs will be removed.
1827 void removeImplicitDefs();
1829 /// Get the value assignments suitable for passing to LiveInterval::join.
1830 const int *getAssignments() const { return Assignments.data(); }
1832 } // end anonymous namespace
1834 unsigned JoinVals::computeWriteLanes(const MachineInstr *DefMI, bool &Redef)
1837 for (const MachineOperand &MO : DefMI->operands()) {
1838 if (!MO.isReg() || MO.getReg() != Reg || !MO.isDef())
1840 L |= TRI->getSubRegIndexLaneMask(
1841 TRI->composeSubRegIndices(SubIdx, MO.getSubReg()));
1848 std::pair<const VNInfo*, unsigned> JoinVals::followCopyChain(
1849 const VNInfo *VNI) const {
1850 unsigned Reg = this->Reg;
1852 while (!VNI->isPHIDef()) {
1853 SlotIndex Def = VNI->def;
1854 MachineInstr *MI = Indexes->getInstructionFromIndex(Def);
1855 assert(MI && "No defining instruction");
1856 if (!MI->isFullCopy())
1857 return std::make_pair(VNI, Reg);
1858 unsigned SrcReg = MI->getOperand(1).getReg();
1859 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1860 return std::make_pair(VNI, Reg);
1862 const LiveInterval &LI = LIS->getInterval(SrcReg);
1863 const VNInfo *ValueIn;
1864 // No subrange involved.
1865 if (!SubRangeJoin || !LI.hasSubRanges()) {
1866 LiveQueryResult LRQ = LI.Query(Def);
1867 ValueIn = LRQ.valueIn();
1869 // Query subranges. Pick the first matching one.
1871 for (const LiveInterval::SubRange &S : LI.subranges()) {
1872 // Transform lanemask to a mask in the joined live interval.
1873 unsigned SMask = TRI->composeSubRegIndexLaneMask(SubIdx, S.LaneMask);
1874 if ((SMask & LaneMask) == 0)
1876 LiveQueryResult LRQ = S.Query(Def);
1877 ValueIn = LRQ.valueIn();
1881 if (ValueIn == nullptr)
1886 return std::make_pair(VNI, Reg);
1889 bool JoinVals::valuesIdentical(VNInfo *Value0, VNInfo *Value1,
1890 const JoinVals &Other) const {
1891 const VNInfo *Orig0;
1893 std::tie(Orig0, Reg0) = followCopyChain(Value0);
1894 if (Orig0 == Value1)
1897 const VNInfo *Orig1;
1899 std::tie(Orig1, Reg1) = Other.followCopyChain(Value1);
1901 // The values are equal if they are defined at the same place and use the
1902 // same register. Note that we cannot compare VNInfos directly as some of
1903 // them might be from a copy created in mergeSubRangeInto() while the other
1904 // is from the original LiveInterval.
1905 return Orig0->def == Orig1->def && Reg0 == Reg1;
1908 JoinVals::ConflictResolution
1909 JoinVals::analyzeValue(unsigned ValNo, JoinVals &Other) {
1910 Val &V = Vals[ValNo];
1911 assert(!V.isAnalyzed() && "Value has already been analyzed!");
1912 VNInfo *VNI = LR.getValNumInfo(ValNo);
1913 if (VNI->isUnused()) {
1918 // Get the instruction defining this value, compute the lanes written.
1919 const MachineInstr *DefMI = nullptr;
1920 if (VNI->isPHIDef()) {
1921 // Conservatively assume that all lanes in a PHI are valid.
1922 unsigned Lanes = SubRangeJoin ? 1 : TRI->getSubRegIndexLaneMask(SubIdx);
1923 V.ValidLanes = V.WriteLanes = Lanes;
1925 DefMI = Indexes->getInstructionFromIndex(VNI->def);
1926 assert(DefMI != nullptr);
1928 // We don't care about the lanes when joining subregister ranges.
1929 V.WriteLanes = V.ValidLanes = 1;
1930 if (DefMI->isImplicitDef()) {
1932 V.ErasableImplicitDef = true;
1936 V.ValidLanes = V.WriteLanes = computeWriteLanes(DefMI, Redef);
1938 // If this is a read-modify-write instruction, there may be more valid
1939 // lanes than the ones written by this instruction.
1940 // This only covers partial redef operands. DefMI may have normal use
1941 // operands reading the register. They don't contribute valid lanes.
1943 // This adds ssub1 to the set of valid lanes in %src:
1945 // %src:ssub1<def> = FOO
1947 // This leaves only ssub1 valid, making any other lanes undef:
1949 // %src:ssub1<def,read-undef> = FOO %src:ssub2
1951 // The <read-undef> flag on the def operand means that old lane values are
1954 V.RedefVNI = LR.Query(VNI->def).valueIn();
1955 assert((TrackSubRegLiveness || V.RedefVNI) &&
1956 "Instruction is reading nonexistent value");
1957 if (V.RedefVNI != nullptr) {
1958 computeAssignment(V.RedefVNI->id, Other);
1959 V.ValidLanes |= Vals[V.RedefVNI->id].ValidLanes;
1963 // An IMPLICIT_DEF writes undef values.
1964 if (DefMI->isImplicitDef()) {
1965 // We normally expect IMPLICIT_DEF values to be live only until the end
1966 // of their block. If the value is really live longer and gets pruned in
1967 // another block, this flag is cleared again.
1968 V.ErasableImplicitDef = true;
1969 V.ValidLanes &= ~V.WriteLanes;
1974 // Find the value in Other that overlaps VNI->def, if any.
1975 LiveQueryResult OtherLRQ = Other.LR.Query(VNI->def);
1977 // It is possible that both values are defined by the same instruction, or
1978 // the values are PHIs defined in the same block. When that happens, the two
1979 // values should be merged into one, but not into any preceding value.
1980 // The first value defined or visited gets CR_Keep, the other gets CR_Merge.
1981 if (VNInfo *OtherVNI = OtherLRQ.valueDefined()) {
1982 assert(SlotIndex::isSameInstr(VNI->def, OtherVNI->def) && "Broken LRQ");
1984 // One value stays, the other is merged. Keep the earlier one, or the first
1986 if (OtherVNI->def < VNI->def)
1987 Other.computeAssignment(OtherVNI->id, *this);
1988 else if (VNI->def < OtherVNI->def && OtherLRQ.valueIn()) {
1989 // This is an early-clobber def overlapping a live-in value in the other
1990 // register. Not mergeable.
1991 V.OtherVNI = OtherLRQ.valueIn();
1992 return CR_Impossible;
1994 V.OtherVNI = OtherVNI;
1995 Val &OtherV = Other.Vals[OtherVNI->id];
1996 // Keep this value, check for conflicts when analyzing OtherVNI.
1997 if (!OtherV.isAnalyzed())
1999 // Both sides have been analyzed now.
2000 // Allow overlapping PHI values. Any real interference would show up in a
2001 // predecessor, the PHI itself can't introduce any conflicts.
2002 if (VNI->isPHIDef())
2004 if (V.ValidLanes & OtherV.ValidLanes)
2005 // Overlapping lanes can't be resolved.
2006 return CR_Impossible;
2011 // No simultaneous def. Is Other live at the def?
2012 V.OtherVNI = OtherLRQ.valueIn();
2014 // No overlap, no conflict.
2017 assert(!SlotIndex::isSameInstr(VNI->def, V.OtherVNI->def) && "Broken LRQ");
2019 // We have overlapping values, or possibly a kill of Other.
2020 // Recursively compute assignments up the dominator tree.
2021 Other.computeAssignment(V.OtherVNI->id, *this);
2022 Val &OtherV = Other.Vals[V.OtherVNI->id];
2024 // Check if OtherV is an IMPLICIT_DEF that extends beyond its basic block.
2025 // This shouldn't normally happen, but ProcessImplicitDefs can leave such
2026 // IMPLICIT_DEF instructions behind, and there is nothing wrong with it
2029 // WHen it happens, treat that IMPLICIT_DEF as a normal value, and don't try
2030 // to erase the IMPLICIT_DEF instruction.
2031 if (OtherV.ErasableImplicitDef && DefMI &&
2032 DefMI->getParent() != Indexes->getMBBFromIndex(V.OtherVNI->def)) {
2033 DEBUG(dbgs() << "IMPLICIT_DEF defined at " << V.OtherVNI->def
2034 << " extends into BB#" << DefMI->getParent()->getNumber()
2035 << ", keeping it.\n");
2036 OtherV.ErasableImplicitDef = false;
2039 // Allow overlapping PHI values. Any real interference would show up in a
2040 // predecessor, the PHI itself can't introduce any conflicts.
2041 if (VNI->isPHIDef())
2044 // Check for simple erasable conflicts.
2045 if (DefMI->isImplicitDef()) {
2046 // We need the def for the subregister if there is nothing else live at the
2047 // subrange at this point.
2048 if (TrackSubRegLiveness
2049 && (V.WriteLanes & (OtherV.ValidLanes | OtherV.WriteLanes)) == 0)
2054 // Include the non-conflict where DefMI is a coalescable copy that kills
2055 // OtherVNI. We still want the copy erased and value numbers merged.
2056 if (CP.isCoalescable(DefMI)) {
2057 // Some of the lanes copied from OtherVNI may be undef, making them undef
2059 V.ValidLanes &= ~V.WriteLanes | OtherV.ValidLanes;
2063 // This may not be a real conflict if DefMI simply kills Other and defines
2065 if (OtherLRQ.isKill() && OtherLRQ.endPoint() <= VNI->def)
2068 // Handle the case where VNI and OtherVNI can be proven to be identical:
2070 // %other = COPY %ext
2071 // %this = COPY %ext <-- Erase this copy
2073 if (DefMI->isFullCopy() && !CP.isPartial()
2074 && valuesIdentical(VNI, V.OtherVNI, Other))
2077 // If the lanes written by this instruction were all undef in OtherVNI, it is
2078 // still safe to join the live ranges. This can't be done with a simple value
2079 // mapping, though - OtherVNI will map to multiple values:
2081 // 1 %dst:ssub0 = FOO <-- OtherVNI
2082 // 2 %src = BAR <-- VNI
2083 // 3 %dst:ssub1 = COPY %src<kill> <-- Eliminate this copy.
2085 // 5 QUUX %src<kill>
2087 // Here OtherVNI will map to itself in [1;2), but to VNI in [2;5). CR_Replace
2088 // handles this complex value mapping.
2089 if ((V.WriteLanes & OtherV.ValidLanes) == 0)
2092 // If the other live range is killed by DefMI and the live ranges are still
2093 // overlapping, it must be because we're looking at an early clobber def:
2095 // %dst<def,early-clobber> = ASM %src<kill>
2097 // In this case, it is illegal to merge the two live ranges since the early
2098 // clobber def would clobber %src before it was read.
2099 if (OtherLRQ.isKill()) {
2100 // This case where the def doesn't overlap the kill is handled above.
2101 assert(VNI->def.isEarlyClobber() &&
2102 "Only early clobber defs can overlap a kill");
2103 return CR_Impossible;
2106 // VNI is clobbering live lanes in OtherVNI, but there is still the
2107 // possibility that no instructions actually read the clobbered lanes.
2108 // If we're clobbering all the lanes in OtherVNI, at least one must be read.
2109 // Otherwise Other.RI wouldn't be live here.
2110 if ((TRI->getSubRegIndexLaneMask(Other.SubIdx) & ~V.WriteLanes) == 0)
2111 return CR_Impossible;
2113 // We need to verify that no instructions are reading the clobbered lanes. To
2114 // save compile time, we'll only check that locally. Don't allow the tainted
2115 // value to escape the basic block.
2116 MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
2117 if (OtherLRQ.endPoint() >= Indexes->getMBBEndIdx(MBB))
2118 return CR_Impossible;
2120 // There are still some things that could go wrong besides clobbered lanes
2121 // being read, for example OtherVNI may be only partially redefined in MBB,
2122 // and some clobbered lanes could escape the block. Save this analysis for
2123 // resolveConflicts() when all values have been mapped. We need to know
2124 // RedefVNI and WriteLanes for any later defs in MBB, and we can't compute
2125 // that now - the recursive analyzeValue() calls must go upwards in the
2127 return CR_Unresolved;
2130 void JoinVals::computeAssignment(unsigned ValNo, JoinVals &Other) {
2131 Val &V = Vals[ValNo];
2132 if (V.isAnalyzed()) {
2133 // Recursion should always move up the dominator tree, so ValNo is not
2134 // supposed to reappear before it has been assigned.
2135 assert(Assignments[ValNo] != -1 && "Bad recursion?");
2138 switch ((V.Resolution = analyzeValue(ValNo, Other))) {
2141 // Merge this ValNo into OtherVNI.
2142 assert(V.OtherVNI && "OtherVNI not assigned, can't merge.");
2143 assert(Other.Vals[V.OtherVNI->id].isAnalyzed() && "Missing recursion");
2144 Assignments[ValNo] = Other.Assignments[V.OtherVNI->id];
2145 DEBUG(dbgs() << "\t\tmerge " << PrintReg(Reg) << ':' << ValNo << '@'
2146 << LR.getValNumInfo(ValNo)->def << " into "
2147 << PrintReg(Other.Reg) << ':' << V.OtherVNI->id << '@'
2148 << V.OtherVNI->def << " --> @"
2149 << NewVNInfo[Assignments[ValNo]]->def << '\n');
2152 case CR_Unresolved: {
2153 // The other value is going to be pruned if this join is successful.
2154 assert(V.OtherVNI && "OtherVNI not assigned, can't prune");
2155 Val &OtherV = Other.Vals[V.OtherVNI->id];
2156 // We cannot erase an IMPLICIT_DEF if we don't have valid values for all
2158 if ((OtherV.WriteLanes & ~V.ValidLanes) != 0 && TrackSubRegLiveness)
2159 OtherV.ErasableImplicitDef = false;
2160 OtherV.Pruned = true;
2164 // This value number needs to go in the final joined live range.
2165 Assignments[ValNo] = NewVNInfo.size();
2166 NewVNInfo.push_back(LR.getValNumInfo(ValNo));
2171 bool JoinVals::mapValues(JoinVals &Other) {
2172 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2173 computeAssignment(i, Other);
2174 if (Vals[i].Resolution == CR_Impossible) {
2175 DEBUG(dbgs() << "\t\tinterference at " << PrintReg(Reg) << ':' << i
2176 << '@' << LR.getValNumInfo(i)->def << '\n');
2184 taintExtent(unsigned ValNo, unsigned TaintedLanes, JoinVals &Other,
2185 SmallVectorImpl<std::pair<SlotIndex, unsigned> > &TaintExtent) {
2186 VNInfo *VNI = LR.getValNumInfo(ValNo);
2187 MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
2188 SlotIndex MBBEnd = Indexes->getMBBEndIdx(MBB);
2190 // Scan Other.LR from VNI.def to MBBEnd.
2191 LiveInterval::iterator OtherI = Other.LR.find(VNI->def);
2192 assert(OtherI != Other.LR.end() && "No conflict?");
2194 // OtherI is pointing to a tainted value. Abort the join if the tainted
2195 // lanes escape the block.
2196 SlotIndex End = OtherI->end;
2197 if (End >= MBBEnd) {
2198 DEBUG(dbgs() << "\t\ttaints global " << PrintReg(Other.Reg) << ':'
2199 << OtherI->valno->id << '@' << OtherI->start << '\n');
2202 DEBUG(dbgs() << "\t\ttaints local " << PrintReg(Other.Reg) << ':'
2203 << OtherI->valno->id << '@' << OtherI->start
2204 << " to " << End << '\n');
2205 // A dead def is not a problem.
2208 TaintExtent.push_back(std::make_pair(End, TaintedLanes));
2210 // Check for another def in the MBB.
2211 if (++OtherI == Other.LR.end() || OtherI->start >= MBBEnd)
2214 // Lanes written by the new def are no longer tainted.
2215 const Val &OV = Other.Vals[OtherI->valno->id];
2216 TaintedLanes &= ~OV.WriteLanes;
2219 } while (TaintedLanes);
2223 bool JoinVals::usesLanes(const MachineInstr *MI, unsigned Reg, unsigned SubIdx,
2224 unsigned Lanes) const {
2225 if (MI->isDebugValue())
2227 for (const MachineOperand &MO : MI->operands()) {
2228 if (!MO.isReg() || MO.isDef() || MO.getReg() != Reg)
2232 if (Lanes & TRI->getSubRegIndexLaneMask(
2233 TRI->composeSubRegIndices(SubIdx, MO.getSubReg())))
2239 bool JoinVals::resolveConflicts(JoinVals &Other) {
2240 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2242 assert (V.Resolution != CR_Impossible && "Unresolvable conflict");
2243 if (V.Resolution != CR_Unresolved)
2245 DEBUG(dbgs() << "\t\tconflict at " << PrintReg(Reg) << ':' << i
2246 << '@' << LR.getValNumInfo(i)->def << '\n');
2251 assert(V.OtherVNI && "Inconsistent conflict resolution.");
2252 VNInfo *VNI = LR.getValNumInfo(i);
2253 const Val &OtherV = Other.Vals[V.OtherVNI->id];
2255 // VNI is known to clobber some lanes in OtherVNI. If we go ahead with the
2256 // join, those lanes will be tainted with a wrong value. Get the extent of
2257 // the tainted lanes.
2258 unsigned TaintedLanes = V.WriteLanes & OtherV.ValidLanes;
2259 SmallVector<std::pair<SlotIndex, unsigned>, 8> TaintExtent;
2260 if (!taintExtent(i, TaintedLanes, Other, TaintExtent))
2261 // Tainted lanes would extend beyond the basic block.
2264 assert(!TaintExtent.empty() && "There should be at least one conflict.");
2266 // Now look at the instructions from VNI->def to TaintExtent (inclusive).
2267 MachineBasicBlock *MBB = Indexes->getMBBFromIndex(VNI->def);
2268 MachineBasicBlock::iterator MI = MBB->begin();
2269 if (!VNI->isPHIDef()) {
2270 MI = Indexes->getInstructionFromIndex(VNI->def);
2271 // No need to check the instruction defining VNI for reads.
2274 assert(!SlotIndex::isSameInstr(VNI->def, TaintExtent.front().first) &&
2275 "Interference ends on VNI->def. Should have been handled earlier");
2276 MachineInstr *LastMI =
2277 Indexes->getInstructionFromIndex(TaintExtent.front().first);
2278 assert(LastMI && "Range must end at a proper instruction");
2279 unsigned TaintNum = 0;
2281 assert(MI != MBB->end() && "Bad LastMI");
2282 if (usesLanes(MI, Other.Reg, Other.SubIdx, TaintedLanes)) {
2283 DEBUG(dbgs() << "\t\ttainted lanes used by: " << *MI);
2286 // LastMI is the last instruction to use the current value.
2287 if (&*MI == LastMI) {
2288 if (++TaintNum == TaintExtent.size())
2290 LastMI = Indexes->getInstructionFromIndex(TaintExtent[TaintNum].first);
2291 assert(LastMI && "Range must end at a proper instruction");
2292 TaintedLanes = TaintExtent[TaintNum].second;
2297 // The tainted lanes are unused.
2298 V.Resolution = CR_Replace;
2304 bool JoinVals::isPrunedValue(unsigned ValNo, JoinVals &Other) {
2305 Val &V = Vals[ValNo];
2306 if (V.Pruned || V.PrunedComputed)
2309 if (V.Resolution != CR_Erase && V.Resolution != CR_Merge)
2312 // Follow copies up the dominator tree and check if any intermediate value
2314 V.PrunedComputed = true;
2315 V.Pruned = Other.isPrunedValue(V.OtherVNI->id, *this);
2319 void JoinVals::pruneValues(JoinVals &Other,
2320 SmallVectorImpl<SlotIndex> &EndPoints,
2321 bool changeInstrs) {
2322 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2323 SlotIndex Def = LR.getValNumInfo(i)->def;
2324 switch (Vals[i].Resolution) {
2328 // This value takes precedence over the value in Other.LR.
2329 LIS->pruneValue(Other.LR, Def, &EndPoints);
2330 // Check if we're replacing an IMPLICIT_DEF value. The IMPLICIT_DEF
2331 // instructions are only inserted to provide a live-out value for PHI
2332 // predecessors, so the instruction should simply go away once its value
2333 // has been replaced.
2334 Val &OtherV = Other.Vals[Vals[i].OtherVNI->id];
2335 bool EraseImpDef = OtherV.ErasableImplicitDef &&
2336 OtherV.Resolution == CR_Keep;
2337 if (!Def.isBlock()) {
2339 // Remove <def,read-undef> flags. This def is now a partial redef.
2340 // Also remove <def,dead> flags since the joined live range will
2341 // continue past this instruction.
2342 for (MachineOperand &MO :
2343 Indexes->getInstructionFromIndex(Def)->operands()) {
2344 if (MO.isReg() && MO.isDef() && MO.getReg() == Reg) {
2345 MO.setIsUndef(EraseImpDef);
2346 MO.setIsDead(false);
2350 // This value will reach instructions below, but we need to make sure
2351 // the live range also reaches the instruction at Def.
2353 EndPoints.push_back(Def);
2355 DEBUG(dbgs() << "\t\tpruned " << PrintReg(Other.Reg) << " at " << Def
2356 << ": " << Other.LR << '\n');
2361 if (isPrunedValue(i, Other)) {
2362 // This value is ultimately a copy of a pruned value in LR or Other.LR.
2363 // We can no longer trust the value mapping computed by
2364 // computeAssignment(), the value that was originally copied could have
2366 LIS->pruneValue(LR, Def, &EndPoints);
2367 DEBUG(dbgs() << "\t\tpruned all of " << PrintReg(Reg) << " at "
2368 << Def << ": " << LR << '\n');
2373 llvm_unreachable("Unresolved conflicts");
2378 void JoinVals::pruneSubRegValues(LiveInterval &LI, unsigned &ShrinkMask)
2380 // Look for values being erased.
2381 bool DidPrune = false;
2382 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2383 if (Vals[i].Resolution != CR_Erase)
2386 // Check subranges at the point where the copy will be removed.
2387 SlotIndex Def = LR.getValNumInfo(i)->def;
2388 for (LiveInterval::SubRange &S : LI.subranges()) {
2389 LiveQueryResult Q = S.Query(Def);
2391 // If a subrange starts at the copy then an undefined value has been
2392 // copied and we must remove that subrange value as well.
2393 VNInfo *ValueOut = Q.valueOutOrDead();
2394 if (ValueOut != nullptr && Q.valueIn() == nullptr) {
2395 DEBUG(dbgs() << "\t\tPrune sublane " << format("%04X", S.LaneMask)
2396 << " at " << Def << "\n");
2397 LIS->pruneValue(S, Def, nullptr);
2399 // Mark value number as unused.
2400 ValueOut->markUnused();
2403 // If a subrange ends at the copy, then a value was copied but only
2404 // partially used later. Shrink the subregister range apropriately.
2405 if (Q.valueIn() != nullptr && Q.valueOut() == nullptr) {
2406 DEBUG(dbgs() << "\t\tDead uses at sublane "
2407 << format("%04X", S.LaneMask) << " at " << Def << "\n");
2408 ShrinkMask |= S.LaneMask;
2413 LI.removeEmptySubRanges();
2416 void JoinVals::removeImplicitDefs() {
2417 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2419 if (V.Resolution != CR_Keep || !V.ErasableImplicitDef || !V.Pruned)
2422 VNInfo *VNI = LR.getValNumInfo(i);
2424 LR.removeValNo(VNI);
2428 void JoinVals::eraseInstrs(SmallPtrSetImpl<MachineInstr*> &ErasedInstrs,
2429 SmallVectorImpl<unsigned> &ShrinkRegs) {
2430 for (unsigned i = 0, e = LR.getNumValNums(); i != e; ++i) {
2431 // Get the def location before markUnused() below invalidates it.
2432 SlotIndex Def = LR.getValNumInfo(i)->def;
2433 switch (Vals[i].Resolution) {
2435 // If an IMPLICIT_DEF value is pruned, it doesn't serve a purpose any
2436 // longer. The IMPLICIT_DEF instructions are only inserted by
2437 // PHIElimination to guarantee that all PHI predecessors have a value.
2438 if (!Vals[i].ErasableImplicitDef || !Vals[i].Pruned)
2440 // Remove value number i from LR.
2441 VNInfo *VNI = LR.getValNumInfo(i);
2442 LR.removeValNo(VNI);
2443 // Note that this VNInfo is reused and still referenced in NewVNInfo,
2444 // make it appear like an unused value number.
2446 DEBUG(dbgs() << "\t\tremoved " << i << '@' << Def << ": " << LR << '\n');
2451 MachineInstr *MI = Indexes->getInstructionFromIndex(Def);
2452 assert(MI && "No instruction to erase");
2454 unsigned Reg = MI->getOperand(1).getReg();
2455 if (TargetRegisterInfo::isVirtualRegister(Reg) &&
2456 Reg != CP.getSrcReg() && Reg != CP.getDstReg())
2457 ShrinkRegs.push_back(Reg);
2459 ErasedInstrs.insert(MI);
2460 DEBUG(dbgs() << "\t\terased:\t" << Def << '\t' << *MI);
2461 LIS->RemoveMachineInstrFromMaps(MI);
2462 MI->eraseFromParent();
2471 bool RegisterCoalescer::joinSubRegRanges(LiveRange &LRange, LiveRange &RRange,
2473 const CoalescerPair &CP) {
2474 SmallVector<VNInfo*, 16> NewVNInfo;
2475 JoinVals RHSVals(RRange, CP.getSrcReg(), CP.getSrcIdx(), LaneMask,
2476 NewVNInfo, CP, LIS, TRI, true, true);
2477 JoinVals LHSVals(LRange, CP.getDstReg(), CP.getDstIdx(), LaneMask,
2478 NewVNInfo, CP, LIS, TRI, true, true);
2480 // Compute NewVNInfo and resolve conflicts (see also joinVirtRegs())
2481 // We should be able to resolve all conflicts here as we could successfully do
2482 // it on the mainrange already. There is however a problem when multiple
2483 // ranges get mapped to the "overflow" lane mask bit which creates unexpected
2485 if (!LHSVals.mapValues(RHSVals) || !RHSVals.mapValues(LHSVals)) {
2486 DEBUG(dbgs() << "*** Couldn't join subrange!\n");
2489 if (!LHSVals.resolveConflicts(RHSVals) ||
2490 !RHSVals.resolveConflicts(LHSVals)) {
2491 DEBUG(dbgs() << "*** Couldn't join subrange!\n");
2495 // The merging algorithm in LiveInterval::join() can't handle conflicting
2496 // value mappings, so we need to remove any live ranges that overlap a
2497 // CR_Replace resolution. Collect a set of end points that can be used to
2498 // restore the live range after joining.
2499 SmallVector<SlotIndex, 8> EndPoints;
2500 LHSVals.pruneValues(RHSVals, EndPoints, false);
2501 RHSVals.pruneValues(LHSVals, EndPoints, false);
2503 LHSVals.removeImplicitDefs();
2504 RHSVals.removeImplicitDefs();
2509 // Join RRange into LHS.
2510 LRange.join(RRange, LHSVals.getAssignments(), RHSVals.getAssignments(),
2513 DEBUG(dbgs() << "\t\tjoined lanes: " << LRange << "\n");
2514 if (EndPoints.empty())
2517 // Recompute the parts of the live range we had to remove because of
2518 // CR_Replace conflicts.
2519 DEBUG(dbgs() << "\t\trestoring liveness to " << EndPoints.size()
2520 << " points: " << LRange << '\n');
2521 LIS->extendToIndices(LRange, EndPoints);
2525 bool RegisterCoalescer::mergeSubRangeInto(LiveInterval &LI,
2526 const LiveRange &ToMerge,
2527 unsigned LaneMask, CoalescerPair &CP) {
2528 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
2529 for (LiveInterval::SubRange &R : LI.subranges()) {
2530 unsigned RMask = R.LaneMask;
2531 // LaneMask of subregisters common to subrange R and ToMerge.
2532 unsigned Common = RMask & LaneMask;
2533 // There is nothing to do without common subregs.
2537 DEBUG(dbgs() << format("\t\tCopy+Merge %04X into %04X\n", RMask, Common));
2538 // LaneMask of subregisters contained in the R range but not in ToMerge,
2539 // they have to split into their own subrange.
2540 unsigned LRest = RMask & ~LaneMask;
2541 LiveInterval::SubRange *CommonRange;
2544 DEBUG(dbgs() << format("\t\tReduce Lane to %04X\n", LRest));
2545 // Duplicate SubRange for newly merged common stuff.
2546 CommonRange = LI.createSubRangeFrom(Allocator, Common, R);
2548 // Reuse the existing range.
2549 R.LaneMask = Common;
2552 LiveRange RangeCopy(ToMerge, Allocator);
2553 if (!joinSubRegRanges(*CommonRange, RangeCopy, Common, CP))
2558 if (LaneMask != 0) {
2559 DEBUG(dbgs() << format("\t\tNew Lane %04X\n", LaneMask));
2560 LI.createSubRangeFrom(Allocator, LaneMask, ToMerge);
2565 bool RegisterCoalescer::joinVirtRegs(CoalescerPair &CP) {
2566 SmallVector<VNInfo*, 16> NewVNInfo;
2567 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
2568 LiveInterval &LHS = LIS->getInterval(CP.getDstReg());
2569 bool TrackSubRegLiveness = MRI->shouldTrackSubRegLiveness(*CP.getNewRC());
2570 JoinVals RHSVals(RHS, CP.getSrcReg(), CP.getSrcIdx(), 0, NewVNInfo, CP, LIS,
2571 TRI, false, TrackSubRegLiveness);
2572 JoinVals LHSVals(LHS, CP.getDstReg(), CP.getDstIdx(), 0, NewVNInfo, CP, LIS,
2573 TRI, false, TrackSubRegLiveness);
2575 DEBUG(dbgs() << "\t\tRHS = " << RHS
2576 << "\n\t\tLHS = " << LHS
2579 // First compute NewVNInfo and the simple value mappings.
2580 // Detect impossible conflicts early.
2581 if (!LHSVals.mapValues(RHSVals) || !RHSVals.mapValues(LHSVals))
2584 // Some conflicts can only be resolved after all values have been mapped.
2585 if (!LHSVals.resolveConflicts(RHSVals) || !RHSVals.resolveConflicts(LHSVals))
2588 // All clear, the live ranges can be merged.
2589 if (RHS.hasSubRanges() || LHS.hasSubRanges()) {
2590 BumpPtrAllocator &Allocator = LIS->getVNInfoAllocator();
2592 // Transform lanemasks from the LHS to masks in the coalesced register and
2593 // create initial subranges if necessary.
2594 unsigned DstIdx = CP.getDstIdx();
2595 if (!LHS.hasSubRanges()) {
2596 unsigned Mask = DstIdx == 0 ? CP.getNewRC()->getLaneMask()
2597 : TRI->getSubRegIndexLaneMask(DstIdx);
2598 // LHS must support subregs or we wouldn't be in this codepath.
2600 LHS.createSubRangeFrom(Allocator, Mask, LHS);
2601 } else if (DstIdx != 0) {
2602 // Transform LHS lanemasks to new register class if necessary.
2603 for (LiveInterval::SubRange &R : LHS.subranges()) {
2604 unsigned Mask = TRI->composeSubRegIndexLaneMask(DstIdx, R.LaneMask);
2608 DEBUG(dbgs() << "\t\tLHST = " << PrintReg(CP.getDstReg())
2609 << ' ' << LHS << '\n');
2611 // Determine lanemasks of RHS in the coalesced register and merge subranges.
2612 unsigned SrcIdx = CP.getSrcIdx();
2614 if (!RHS.hasSubRanges()) {
2615 unsigned Mask = SrcIdx == 0 ? CP.getNewRC()->getLaneMask()
2616 : TRI->getSubRegIndexLaneMask(SrcIdx);
2617 if (!mergeSubRangeInto(LHS, RHS, Mask, CP))
2620 // Pair up subranges and merge.
2621 for (LiveInterval::SubRange &R : RHS.subranges()) {
2622 unsigned Mask = TRI->composeSubRegIndexLaneMask(SrcIdx, R.LaneMask);
2623 if (!mergeSubRangeInto(LHS, R, Mask, CP)) {
2630 // This shouldn't have happened :-(
2631 // However we are aware of at least one existing problem where we
2632 // can't merge subranges when multiple ranges end up in the
2633 // "overflow bit" 32. As a workaround we drop all subregister ranges
2634 // which means we loose some precision but are back to a well defined
2636 assert((CP.getNewRC()->getLaneMask() & 0x80000000u)
2637 && "SubRange merge should only fail when merging into bit 32.");
2638 DEBUG(dbgs() << "\tSubrange join aborted!\n");
2639 LHS.clearSubRanges();
2640 RHS.clearSubRanges();
2642 DEBUG(dbgs() << "\tJoined SubRanges " << LHS << "\n");
2644 LHSVals.pruneSubRegValues(LHS, ShrinkMask);
2645 RHSVals.pruneSubRegValues(LHS, ShrinkMask);
2649 // The merging algorithm in LiveInterval::join() can't handle conflicting
2650 // value mappings, so we need to remove any live ranges that overlap a
2651 // CR_Replace resolution. Collect a set of end points that can be used to
2652 // restore the live range after joining.
2653 SmallVector<SlotIndex, 8> EndPoints;
2654 LHSVals.pruneValues(RHSVals, EndPoints, true);
2655 RHSVals.pruneValues(LHSVals, EndPoints, true);
2657 // Erase COPY and IMPLICIT_DEF instructions. This may cause some external
2658 // registers to require trimming.
2659 SmallVector<unsigned, 8> ShrinkRegs;
2660 LHSVals.eraseInstrs(ErasedInstrs, ShrinkRegs);
2661 RHSVals.eraseInstrs(ErasedInstrs, ShrinkRegs);
2662 while (!ShrinkRegs.empty())
2663 shrinkToUses(&LIS->getInterval(ShrinkRegs.pop_back_val()));
2665 // Join RHS into LHS.
2666 LHS.join(RHS, LHSVals.getAssignments(), RHSVals.getAssignments(), NewVNInfo);
2668 // Kill flags are going to be wrong if the live ranges were overlapping.
2669 // Eventually, we should simply clear all kill flags when computing live
2670 // ranges. They are reinserted after register allocation.
2671 MRI->clearKillFlags(LHS.reg);
2672 MRI->clearKillFlags(RHS.reg);
2674 if (!EndPoints.empty()) {
2675 // Recompute the parts of the live range we had to remove because of
2676 // CR_Replace conflicts.
2677 DEBUG(dbgs() << "\t\trestoring liveness to " << EndPoints.size()
2678 << " points: " << LHS << '\n');
2679 LIS->extendToIndices((LiveRange&)LHS, EndPoints);
2685 bool RegisterCoalescer::joinIntervals(CoalescerPair &CP) {
2686 return CP.isPhys() ? joinReservedPhysReg(CP) : joinVirtRegs(CP);
2690 /// Information concerning MBB coalescing priority.
2691 struct MBBPriorityInfo {
2692 MachineBasicBlock *MBB;
2696 MBBPriorityInfo(MachineBasicBlock *mbb, unsigned depth, bool issplit)
2697 : MBB(mbb), Depth(depth), IsSplit(issplit) {}
2701 /// C-style comparator that sorts first based on the loop depth of the basic
2702 /// block (the unsigned), and then on the MBB number.
2704 /// EnableGlobalCopies assumes that the primary sort key is loop depth.
2705 static int compareMBBPriority(const MBBPriorityInfo *LHS,
2706 const MBBPriorityInfo *RHS) {
2707 // Deeper loops first
2708 if (LHS->Depth != RHS->Depth)
2709 return LHS->Depth > RHS->Depth ? -1 : 1;
2711 // Try to unsplit critical edges next.
2712 if (LHS->IsSplit != RHS->IsSplit)
2713 return LHS->IsSplit ? -1 : 1;
2715 // Prefer blocks that are more connected in the CFG. This takes care of
2716 // the most difficult copies first while intervals are short.
2717 unsigned cl = LHS->MBB->pred_size() + LHS->MBB->succ_size();
2718 unsigned cr = RHS->MBB->pred_size() + RHS->MBB->succ_size();
2720 return cl > cr ? -1 : 1;
2722 // As a last resort, sort by block number.
2723 return LHS->MBB->getNumber() < RHS->MBB->getNumber() ? -1 : 1;
2726 /// \returns true if the given copy uses or defines a local live range.
2727 static bool isLocalCopy(MachineInstr *Copy, const LiveIntervals *LIS) {
2728 if (!Copy->isCopy())
2731 if (Copy->getOperand(1).isUndef())
2734 unsigned SrcReg = Copy->getOperand(1).getReg();
2735 unsigned DstReg = Copy->getOperand(0).getReg();
2736 if (TargetRegisterInfo::isPhysicalRegister(SrcReg)
2737 || TargetRegisterInfo::isPhysicalRegister(DstReg))
2740 return LIS->intervalIsInOneMBB(LIS->getInterval(SrcReg))
2741 || LIS->intervalIsInOneMBB(LIS->getInterval(DstReg));
2744 bool RegisterCoalescer::
2745 copyCoalesceWorkList(MutableArrayRef<MachineInstr*> CurrList) {
2746 bool Progress = false;
2747 for (unsigned i = 0, e = CurrList.size(); i != e; ++i) {
2750 // Skip instruction pointers that have already been erased, for example by
2751 // dead code elimination.
2752 if (ErasedInstrs.erase(CurrList[i])) {
2753 CurrList[i] = nullptr;
2757 bool Success = joinCopy(CurrList[i], Again);
2758 Progress |= Success;
2759 if (Success || !Again)
2760 CurrList[i] = nullptr;
2765 /// Check if DstReg is a terminal node.
2766 /// I.e., it does not have any affinity other than \p Copy.
2767 static bool isTerminalReg(unsigned DstReg, const MachineInstr &Copy,
2768 const MachineRegisterInfo *MRI) {
2769 assert(Copy.isCopyLike());
2770 // Check if the destination of this copy as any other affinity.
2771 for (const MachineInstr &MI : MRI->reg_nodbg_instructions(DstReg))
2772 if (&MI != &Copy && MI.isCopyLike())
2777 bool RegisterCoalescer::applyTerminalRule(const MachineInstr &Copy) const {
2778 assert(Copy.isCopyLike());
2779 if (!UseTerminalRule)
2781 unsigned DstReg, DstSubReg, SrcReg, SrcSubReg;
2782 isMoveInstr(*TRI, &Copy, SrcReg, DstReg, SrcSubReg, DstSubReg);
2783 // Check if the destination of this copy has any other affinity.
2784 if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
2785 // If SrcReg is a physical register, the copy won't be coalesced.
2786 // Ignoring it may have other side effect (like missing
2787 // rematerialization). So keep it.
2788 TargetRegisterInfo::isPhysicalRegister(SrcReg) ||
2789 !isTerminalReg(DstReg, Copy, MRI))
2792 // DstReg is a terminal node. Check if it inteferes with any other
2793 // copy involving SrcReg.
2794 const MachineBasicBlock *OrigBB = Copy.getParent();
2795 const LiveInterval &DstLI = LIS->getInterval(DstReg);
2796 for (const MachineInstr &MI : MRI->reg_nodbg_instructions(SrcReg)) {
2797 // Technically we should check if the weight of the new copy is
2798 // interesting compared to the other one and update the weight
2799 // of the copies accordingly. However, this would only work if
2800 // we would gather all the copies first then coalesce, whereas
2801 // right now we interleave both actions.
2802 // For now, just consider the copies that are in the same block.
2803 if (&MI == &Copy || !MI.isCopyLike() || MI.getParent() != OrigBB)
2805 unsigned OtherReg, OtherSubReg, OtherSrcReg, OtherSrcSubReg;
2806 isMoveInstr(*TRI, &Copy, OtherSrcReg, OtherReg, OtherSrcSubReg,
2808 if (OtherReg == SrcReg)
2809 OtherReg = OtherSrcReg;
2810 // Check if OtherReg is a non-terminal.
2811 if (TargetRegisterInfo::isPhysicalRegister(OtherReg) ||
2812 isTerminalReg(OtherReg, MI, MRI))
2814 // Check that OtherReg interfere with DstReg.
2815 if (LIS->getInterval(OtherReg).overlaps(DstLI)) {
2816 DEBUG(dbgs() << "Apply terminal rule for: " << PrintReg(DstReg) << '\n');
2824 RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) {
2825 DEBUG(dbgs() << MBB->getName() << ":\n");
2827 // Collect all copy-like instructions in MBB. Don't start coalescing anything
2828 // yet, it might invalidate the iterator.
2829 const unsigned PrevSize = WorkList.size();
2830 if (JoinGlobalCopies) {
2831 SmallVector<MachineInstr*, 2> LocalTerminals;
2832 SmallVector<MachineInstr*, 2> GlobalTerminals;
2833 // Coalesce copies bottom-up to coalesce local defs before local uses. They
2834 // are not inherently easier to resolve, but slightly preferable until we
2835 // have local live range splitting. In particular this is required by
2836 // cmp+jmp macro fusion.
2837 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2839 if (!MII->isCopyLike())
2841 bool ApplyTerminalRule = applyTerminalRule(*MII);
2842 if (isLocalCopy(&(*MII), LIS)) {
2843 if (ApplyTerminalRule)
2844 LocalTerminals.push_back(&(*MII));
2846 LocalWorkList.push_back(&(*MII));
2848 if (ApplyTerminalRule)
2849 GlobalTerminals.push_back(&(*MII));
2851 WorkList.push_back(&(*MII));
2854 // Append the copies evicted by the terminal rule at the end of the list.
2855 LocalWorkList.append(LocalTerminals.begin(), LocalTerminals.end());
2856 WorkList.append(GlobalTerminals.begin(), GlobalTerminals.end());
2859 SmallVector<MachineInstr*, 2> Terminals;
2860 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
2862 if (MII->isCopyLike()) {
2863 if (applyTerminalRule(*MII))
2864 Terminals.push_back(&(*MII));
2866 WorkList.push_back(MII);
2868 // Append the copies evicted by the terminal rule at the end of the list.
2869 WorkList.append(Terminals.begin(), Terminals.end());
2871 // Try coalescing the collected copies immediately, and remove the nulls.
2872 // This prevents the WorkList from getting too large since most copies are
2873 // joinable on the first attempt.
2874 MutableArrayRef<MachineInstr*>
2875 CurrList(WorkList.begin() + PrevSize, WorkList.end());
2876 if (copyCoalesceWorkList(CurrList))
2877 WorkList.erase(std::remove(WorkList.begin() + PrevSize, WorkList.end(),
2878 (MachineInstr*)nullptr), WorkList.end());
2881 void RegisterCoalescer::coalesceLocals() {
2882 copyCoalesceWorkList(LocalWorkList);
2883 for (unsigned j = 0, je = LocalWorkList.size(); j != je; ++j) {
2884 if (LocalWorkList[j])
2885 WorkList.push_back(LocalWorkList[j]);
2887 LocalWorkList.clear();
2890 void RegisterCoalescer::joinAllIntervals() {
2891 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
2892 assert(WorkList.empty() && LocalWorkList.empty() && "Old data still around.");
2894 std::vector<MBBPriorityInfo> MBBs;
2895 MBBs.reserve(MF->size());
2896 for (MachineFunction::iterator I = MF->begin(), E = MF->end();I != E;++I){
2897 MachineBasicBlock *MBB = I;
2898 MBBs.push_back(MBBPriorityInfo(MBB, Loops->getLoopDepth(MBB),
2899 JoinSplitEdges && isSplitEdge(MBB)));
2901 array_pod_sort(MBBs.begin(), MBBs.end(), compareMBBPriority);
2903 // Coalesce intervals in MBB priority order.
2904 unsigned CurrDepth = UINT_MAX;
2905 for (unsigned i = 0, e = MBBs.size(); i != e; ++i) {
2906 // Try coalescing the collected local copies for deeper loops.
2907 if (JoinGlobalCopies && MBBs[i].Depth < CurrDepth) {
2909 CurrDepth = MBBs[i].Depth;
2911 copyCoalesceInMBB(MBBs[i].MBB);
2915 // Joining intervals can allow other intervals to be joined. Iteratively join
2916 // until we make no progress.
2917 while (copyCoalesceWorkList(WorkList))
2921 void RegisterCoalescer::releaseMemory() {
2922 ErasedInstrs.clear();
2925 InflateRegs.clear();
2928 bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) {
2930 MRI = &fn.getRegInfo();
2931 TM = &fn.getTarget();
2932 const TargetSubtargetInfo &STI = fn.getSubtarget();
2933 TRI = STI.getRegisterInfo();
2934 TII = STI.getInstrInfo();
2935 LIS = &getAnalysis<LiveIntervals>();
2936 AA = &getAnalysis<AliasAnalysis>();
2937 Loops = &getAnalysis<MachineLoopInfo>();
2938 if (EnableGlobalCopies == cl::BOU_UNSET)
2939 JoinGlobalCopies = STI.enableJoinGlobalCopies();
2941 JoinGlobalCopies = (EnableGlobalCopies == cl::BOU_TRUE);
2943 // The MachineScheduler does not currently require JoinSplitEdges. This will
2944 // either be enabled unconditionally or replaced by a more general live range
2945 // splitting optimization.
2946 JoinSplitEdges = EnableJoinSplits;
2948 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
2949 << "********** Function: " << MF->getName() << '\n');
2951 if (VerifyCoalescing)
2952 MF->verify(this, "Before register coalescing");
2954 RegClassInfo.runOnMachineFunction(fn);
2956 // Join (coalesce) intervals if requested.
2960 // After deleting a lot of copies, register classes may be less constrained.
2961 // Removing sub-register operands may allow GR32_ABCD -> GR32 and DPR_VFP2 ->
2963 array_pod_sort(InflateRegs.begin(), InflateRegs.end());
2964 InflateRegs.erase(std::unique(InflateRegs.begin(), InflateRegs.end()),
2966 DEBUG(dbgs() << "Trying to inflate " << InflateRegs.size() << " regs.\n");
2967 for (unsigned i = 0, e = InflateRegs.size(); i != e; ++i) {
2968 unsigned Reg = InflateRegs[i];
2969 if (MRI->reg_nodbg_empty(Reg))
2971 if (MRI->recomputeRegClass(Reg)) {
2972 DEBUG(dbgs() << PrintReg(Reg) << " inflated to "
2973 << TRI->getRegClassName(MRI->getRegClass(Reg)) << '\n');
2974 LiveInterval &LI = LIS->getInterval(Reg);
2975 unsigned MaxMask = MRI->getMaxLaneMaskForVReg(Reg);
2977 // If the inflated register class does not support subregisters anymore
2978 // remove the subranges.
2979 LI.clearSubRanges();
2982 // If subranges are still supported, then the same subregs should still
2984 for (LiveInterval::SubRange &S : LI.subranges()) {
2985 assert ((S.LaneMask & ~MaxMask) == 0);
2994 if (VerifyCoalescing)
2995 MF->verify(this, "After register coalescing");
2999 void RegisterCoalescer::print(raw_ostream &O, const Module* m) const {