1 //===- RegisterCoalescer.cpp - Generic Register Coalescing Interface -------==//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the generic RegisterCoalescer interface which
11 // is used as the common interface used by all clients and
12 // implementations of register coalescing.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "regalloc"
17 #include "RegisterCoalescer.h"
18 #include "LiveDebugVariables.h"
19 #include "VirtRegMap.h"
21 #include "llvm/Pass.h"
22 #include "llvm/Value.h"
23 #include "llvm/ADT/OwningPtr.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
29 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
30 #include "llvm/CodeGen/LiveRangeEdit.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineInstr.h"
33 #include "llvm/CodeGen/MachineInstr.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/CodeGen/RegisterClassInfo.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetInstrInfo.h"
45 #include "llvm/Target/TargetMachine.h"
46 #include "llvm/Target/TargetOptions.h"
47 #include "llvm/Target/TargetRegisterInfo.h"
52 STATISTIC(numJoins , "Number of interval joins performed");
53 STATISTIC(numCrossRCs , "Number of cross class joins performed");
54 STATISTIC(numCommutes , "Number of instruction commuting performed");
55 STATISTIC(numExtends , "Number of copies extended");
56 STATISTIC(NumReMats , "Number of instructions re-materialized");
57 STATISTIC(NumInflated , "Number of register classes inflated");
60 EnableJoining("join-liveintervals",
61 cl::desc("Coalesce copies (default=true)"),
65 VerifyCoalescing("verify-coalescing",
66 cl::desc("Verify machine instrs before and after register coalescing"),
70 class RegisterCoalescer : public MachineFunctionPass,
71 private LiveRangeEdit::Delegate {
73 MachineRegisterInfo* MRI;
74 const TargetMachine* TM;
75 const TargetRegisterInfo* TRI;
76 const TargetInstrInfo* TII;
78 LiveDebugVariables *LDV;
79 const MachineLoopInfo* Loops;
81 RegisterClassInfo RegClassInfo;
83 /// WorkList - Copy instructions yet to be coalesced.
84 SmallVector<MachineInstr*, 8> WorkList;
86 /// ErasedInstrs - Set of instruction pointers that have been erased, and
87 /// that may be present in WorkList.
88 SmallPtrSet<MachineInstr*, 8> ErasedInstrs;
90 /// Dead instructions that are about to be deleted.
91 SmallVector<MachineInstr*, 8> DeadDefs;
93 /// Virtual registers to be considered for register class inflation.
94 SmallVector<unsigned, 8> InflateRegs;
96 /// Recursively eliminate dead defs in DeadDefs.
97 void eliminateDeadDefs();
99 /// LiveRangeEdit callback.
100 void LRE_WillEraseInstruction(MachineInstr *MI);
102 /// joinAllIntervals - join compatible live intervals
103 void joinAllIntervals();
105 /// copyCoalesceInMBB - Coalesce copies in the specified MBB, putting
106 /// copies that cannot yet be coalesced into WorkList.
107 void copyCoalesceInMBB(MachineBasicBlock *MBB);
109 /// copyCoalesceWorkList - Try to coalesce all copies in WorkList after
110 /// position From. Return true if any progress was made.
111 bool copyCoalesceWorkList(unsigned From = 0);
113 /// joinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
114 /// which are the src/dst of the copy instruction CopyMI. This returns
115 /// true if the copy was successfully coalesced away. If it is not
116 /// currently possible to coalesce this interval, but it may be possible if
117 /// other things get coalesced, then it returns true by reference in
119 bool joinCopy(MachineInstr *TheCopy, bool &Again);
121 /// joinIntervals - Attempt to join these two intervals. On failure, this
122 /// returns false. The output "SrcInt" will not have been modified, so we
123 /// can use this information below to update aliases.
124 bool joinIntervals(CoalescerPair &CP);
126 /// Attempt joining with a reserved physreg.
127 bool joinReservedPhysReg(CoalescerPair &CP);
129 /// adjustCopiesBackFrom - We found a non-trivially-coalescable copy. If
130 /// the source value number is defined by a copy from the destination reg
131 /// see if we can merge these two destination reg valno# into a single
132 /// value number, eliminating a copy.
133 bool adjustCopiesBackFrom(const CoalescerPair &CP, MachineInstr *CopyMI);
135 /// hasOtherReachingDefs - Return true if there are definitions of IntB
136 /// other than BValNo val# that can reach uses of AValno val# of IntA.
137 bool hasOtherReachingDefs(LiveInterval &IntA, LiveInterval &IntB,
138 VNInfo *AValNo, VNInfo *BValNo);
140 /// removeCopyByCommutingDef - We found a non-trivially-coalescable copy.
141 /// If the source value number is defined by a commutable instruction and
142 /// its other operand is coalesced to the copy dest register, see if we
143 /// can transform the copy into a noop by commuting the definition.
144 bool removeCopyByCommutingDef(const CoalescerPair &CP,MachineInstr *CopyMI);
146 /// reMaterializeTrivialDef - If the source of a copy is defined by a
147 /// trivial computation, replace the copy by rematerialize the definition.
148 bool reMaterializeTrivialDef(LiveInterval &SrcInt, unsigned DstReg,
149 MachineInstr *CopyMI);
151 /// canJoinPhys - Return true if a physreg copy should be joined.
152 bool canJoinPhys(CoalescerPair &CP);
154 /// updateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
155 /// update the subregister number if it is not zero. If DstReg is a
156 /// physical register and the existing subregister number of the def / use
157 /// being updated is not zero, make sure to set it to the correct physical
159 void updateRegDefsUses(unsigned SrcReg, unsigned DstReg, unsigned SubIdx);
161 /// eliminateUndefCopy - Handle copies of undef values.
162 bool eliminateUndefCopy(MachineInstr *CopyMI, const CoalescerPair &CP);
165 static char ID; // Class identification, replacement for typeinfo
166 RegisterCoalescer() : MachineFunctionPass(ID) {
167 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
170 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
172 virtual void releaseMemory();
174 /// runOnMachineFunction - pass entry point
175 virtual bool runOnMachineFunction(MachineFunction&);
177 /// print - Implement the dump method.
178 virtual void print(raw_ostream &O, const Module* = 0) const;
180 } /// end anonymous namespace
182 char &llvm::RegisterCoalescerID = RegisterCoalescer::ID;
184 INITIALIZE_PASS_BEGIN(RegisterCoalescer, "simple-register-coalescing",
185 "Simple Register Coalescing", false, false)
186 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
187 INITIALIZE_PASS_DEPENDENCY(LiveDebugVariables)
188 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
189 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
190 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
191 INITIALIZE_PASS_END(RegisterCoalescer, "simple-register-coalescing",
192 "Simple Register Coalescing", false, false)
194 char RegisterCoalescer::ID = 0;
196 static unsigned compose(const TargetRegisterInfo &tri, unsigned a, unsigned b) {
199 return tri.composeSubRegIndices(a, b);
202 static bool isMoveInstr(const TargetRegisterInfo &tri, const MachineInstr *MI,
203 unsigned &Src, unsigned &Dst,
204 unsigned &SrcSub, unsigned &DstSub) {
206 Dst = MI->getOperand(0).getReg();
207 DstSub = MI->getOperand(0).getSubReg();
208 Src = MI->getOperand(1).getReg();
209 SrcSub = MI->getOperand(1).getSubReg();
210 } else if (MI->isSubregToReg()) {
211 Dst = MI->getOperand(0).getReg();
212 DstSub = compose(tri, MI->getOperand(0).getSubReg(),
213 MI->getOperand(3).getImm());
214 Src = MI->getOperand(2).getReg();
215 SrcSub = MI->getOperand(2).getSubReg();
221 bool CoalescerPair::setRegisters(const MachineInstr *MI) {
225 Flipped = CrossClass = false;
227 unsigned Src, Dst, SrcSub, DstSub;
228 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
230 Partial = SrcSub || DstSub;
232 // If one register is a physreg, it must be Dst.
233 if (TargetRegisterInfo::isPhysicalRegister(Src)) {
234 if (TargetRegisterInfo::isPhysicalRegister(Dst))
237 std::swap(SrcSub, DstSub);
241 const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
243 if (TargetRegisterInfo::isPhysicalRegister(Dst)) {
244 // Eliminate DstSub on a physreg.
246 Dst = TRI.getSubReg(Dst, DstSub);
247 if (!Dst) return false;
251 // Eliminate SrcSub by picking a corresponding Dst superregister.
253 Dst = TRI.getMatchingSuperReg(Dst, SrcSub, MRI.getRegClass(Src));
254 if (!Dst) return false;
256 } else if (!MRI.getRegClass(Src)->contains(Dst)) {
260 // Both registers are virtual.
261 const TargetRegisterClass *SrcRC = MRI.getRegClass(Src);
262 const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
264 // Both registers have subreg indices.
265 if (SrcSub && DstSub) {
266 // Copies between different sub-registers are never coalescable.
267 if (Src == Dst && SrcSub != DstSub)
270 NewRC = TRI.getCommonSuperRegClass(SrcRC, SrcSub, DstRC, DstSub,
275 // SrcReg will be merged with a sub-register of DstReg.
277 NewRC = TRI.getMatchingSuperRegClass(DstRC, SrcRC, DstSub);
279 // DstReg will be merged with a sub-register of SrcReg.
281 NewRC = TRI.getMatchingSuperRegClass(SrcRC, DstRC, SrcSub);
283 // This is a straight copy without sub-registers.
284 NewRC = TRI.getCommonSubClass(DstRC, SrcRC);
287 // The combined constraint may be impossible to satisfy.
291 // Prefer SrcReg to be a sub-register of DstReg.
292 // FIXME: Coalescer should support subregs symmetrically.
293 if (DstIdx && !SrcIdx) {
295 std::swap(SrcIdx, DstIdx);
299 CrossClass = NewRC != DstRC || NewRC != SrcRC;
301 // Check our invariants
302 assert(TargetRegisterInfo::isVirtualRegister(Src) && "Src must be virtual");
303 assert(!(TargetRegisterInfo::isPhysicalRegister(Dst) && DstSub) &&
304 "Cannot have a physical SubIdx");
310 bool CoalescerPair::flip() {
311 if (TargetRegisterInfo::isPhysicalRegister(DstReg))
313 std::swap(SrcReg, DstReg);
314 std::swap(SrcIdx, DstIdx);
319 bool CoalescerPair::isCoalescable(const MachineInstr *MI) const {
322 unsigned Src, Dst, SrcSub, DstSub;
323 if (!isMoveInstr(TRI, MI, Src, Dst, SrcSub, DstSub))
326 // Find the virtual register that is SrcReg.
329 std::swap(SrcSub, DstSub);
330 } else if (Src != SrcReg) {
334 // Now check that Dst matches DstReg.
335 if (TargetRegisterInfo::isPhysicalRegister(DstReg)) {
336 if (!TargetRegisterInfo::isPhysicalRegister(Dst))
338 assert(!DstIdx && !SrcIdx && "Inconsistent CoalescerPair state.");
339 // DstSub could be set for a physreg from INSERT_SUBREG.
341 Dst = TRI.getSubReg(Dst, DstSub);
344 return DstReg == Dst;
345 // This is a partial register copy. Check that the parts match.
346 return TRI.getSubReg(DstReg, SrcSub) == Dst;
348 // DstReg is virtual.
351 // Registers match, do the subregisters line up?
352 return compose(TRI, SrcIdx, SrcSub) == compose(TRI, DstIdx, DstSub);
356 void RegisterCoalescer::getAnalysisUsage(AnalysisUsage &AU) const {
357 AU.setPreservesCFG();
358 AU.addRequired<AliasAnalysis>();
359 AU.addRequired<LiveIntervals>();
360 AU.addPreserved<LiveIntervals>();
361 AU.addRequired<LiveDebugVariables>();
362 AU.addPreserved<LiveDebugVariables>();
363 AU.addPreserved<SlotIndexes>();
364 AU.addRequired<MachineLoopInfo>();
365 AU.addPreserved<MachineLoopInfo>();
366 AU.addPreservedID(MachineDominatorsID);
367 MachineFunctionPass::getAnalysisUsage(AU);
370 void RegisterCoalescer::eliminateDeadDefs() {
371 SmallVector<LiveInterval*, 8> NewRegs;
372 LiveRangeEdit(0, NewRegs, *MF, *LIS, 0, this).eliminateDeadDefs(DeadDefs);
375 // Callback from eliminateDeadDefs().
376 void RegisterCoalescer::LRE_WillEraseInstruction(MachineInstr *MI) {
377 // MI may be in WorkList. Make sure we don't visit it.
378 ErasedInstrs.insert(MI);
381 /// adjustCopiesBackFrom - We found a non-trivially-coalescable copy with IntA
382 /// being the source and IntB being the dest, thus this defines a value number
383 /// in IntB. If the source value number (in IntA) is defined by a copy from B,
384 /// see if we can merge these two pieces of B into a single value number,
385 /// eliminating a copy. For example:
389 /// B1 = A3 <- this copy
391 /// In this case, B0 can be extended to where the B1 copy lives, allowing the B1
392 /// value number to be replaced with B0 (which simplifies the B liveinterval).
394 /// This returns true if an interval was modified.
396 bool RegisterCoalescer::adjustCopiesBackFrom(const CoalescerPair &CP,
397 MachineInstr *CopyMI) {
398 assert(!CP.isPartial() && "This doesn't work for partial copies.");
399 assert(!CP.isPhys() && "This doesn't work for physreg copies.");
401 // Bail if there is no dst interval - can happen when merging physical subreg
403 if (!LIS->hasInterval(CP.getDstReg()))
407 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
409 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
410 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
412 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
413 // the example above.
414 LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
415 if (BLR == IntB.end()) return false;
416 VNInfo *BValNo = BLR->valno;
418 // Get the location that B is defined at. Two options: either this value has
419 // an unknown definition point or it is defined at CopyIdx. If unknown, we
421 if (BValNo->def != CopyIdx) return false;
423 // AValNo is the value number in A that defines the copy, A3 in the example.
424 SlotIndex CopyUseIdx = CopyIdx.getRegSlot(true);
425 LiveInterval::iterator ALR = IntA.FindLiveRangeContaining(CopyUseIdx);
426 // The live range might not exist after fun with physreg coalescing.
427 if (ALR == IntA.end()) return false;
428 VNInfo *AValNo = ALR->valno;
430 // If AValNo is defined as a copy from IntB, we can potentially process this.
431 // Get the instruction that defines this value number.
432 MachineInstr *ACopyMI = LIS->getInstructionFromIndex(AValNo->def);
433 if (!CP.isCoalescable(ACopyMI))
436 // Get the LiveRange in IntB that this value number starts with.
437 LiveInterval::iterator ValLR =
438 IntB.FindLiveRangeContaining(AValNo->def.getPrevSlot());
439 if (ValLR == IntB.end())
442 // Make sure that the end of the live range is inside the same block as
444 MachineInstr *ValLREndInst =
445 LIS->getInstructionFromIndex(ValLR->end.getPrevSlot());
446 if (!ValLREndInst || ValLREndInst->getParent() != CopyMI->getParent())
449 // Okay, we now know that ValLR ends in the same block that the CopyMI
450 // live-range starts. If there are no intervening live ranges between them in
451 // IntB, we can merge them.
452 if (ValLR+1 != BLR) return false;
454 DEBUG(dbgs() << "Extending: " << PrintReg(IntB.reg, TRI));
456 SlotIndex FillerStart = ValLR->end, FillerEnd = BLR->start;
457 // We are about to delete CopyMI, so need to remove it as the 'instruction
458 // that defines this value #'. Update the valnum with the new defining
460 BValNo->def = FillerStart;
462 // Okay, we can merge them. We need to insert a new liverange:
463 // [ValLR.end, BLR.begin) of either value number, then we merge the
464 // two value numbers.
465 IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
467 // If the IntB live range is assigned to a physical register, and if that
468 // physreg has sub-registers, update their live intervals as well.
469 if (TargetRegisterInfo::isPhysicalRegister(IntB.reg)) {
470 for (MCSubRegIterator SR(IntB.reg, TRI); SR.isValid(); ++SR) {
471 if (!LIS->hasInterval(*SR))
473 LiveInterval &SRLI = LIS->getInterval(*SR);
474 SRLI.addRange(LiveRange(FillerStart, FillerEnd,
475 SRLI.getNextValue(FillerStart,
476 LIS->getVNInfoAllocator())));
480 // Okay, merge "B1" into the same value number as "B0".
481 if (BValNo != ValLR->valno) {
482 // If B1 is killed by a PHI, then the merged live range must also be killed
483 // by the same PHI, as B0 and B1 can not overlap.
484 bool HasPHIKill = BValNo->hasPHIKill();
485 IntB.MergeValueNumberInto(BValNo, ValLR->valno);
487 ValLR->valno->setHasPHIKill(true);
489 DEBUG(dbgs() << " result = " << IntB << '\n');
491 // If the source instruction was killing the source register before the
492 // merge, unset the isKill marker given the live range has been extended.
493 int UIdx = ValLREndInst->findRegisterUseOperandIdx(IntB.reg, true);
495 ValLREndInst->getOperand(UIdx).setIsKill(false);
498 // Rewrite the copy. If the copy instruction was killing the destination
499 // register before the merge, find the last use and trim the live range. That
500 // will also add the isKill marker.
501 CopyMI->substituteRegister(IntA.reg, IntB.reg, 0, *TRI);
502 if (ALR->end == CopyIdx)
503 LIS->shrinkToUses(&IntA);
509 /// hasOtherReachingDefs - Return true if there are definitions of IntB
510 /// other than BValNo val# that can reach uses of AValno val# of IntA.
511 bool RegisterCoalescer::hasOtherReachingDefs(LiveInterval &IntA,
515 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
517 if (AI->valno != AValNo) continue;
518 LiveInterval::Ranges::iterator BI =
519 std::upper_bound(IntB.ranges.begin(), IntB.ranges.end(), AI->start);
520 if (BI != IntB.ranges.begin())
522 for (; BI != IntB.ranges.end() && AI->end >= BI->start; ++BI) {
523 if (BI->valno == BValNo)
525 if (BI->start <= AI->start && BI->end > AI->start)
527 if (BI->start > AI->start && BI->start < AI->end)
534 /// removeCopyByCommutingDef - We found a non-trivially-coalescable copy with
535 /// IntA being the source and IntB being the dest, thus this defines a value
536 /// number in IntB. If the source value number (in IntA) is defined by a
537 /// commutable instruction and its other operand is coalesced to the copy dest
538 /// register, see if we can transform the copy into a noop by commuting the
539 /// definition. For example,
541 /// A3 = op A2 B0<kill>
543 /// B1 = A3 <- this copy
545 /// = op A3 <- more uses
549 /// B2 = op B0 A2<kill>
551 /// B1 = B2 <- now an identify copy
553 /// = op B2 <- more uses
555 /// This returns true if an interval was modified.
557 bool RegisterCoalescer::removeCopyByCommutingDef(const CoalescerPair &CP,
558 MachineInstr *CopyMI) {
559 assert (!CP.isPhys());
561 // Bail if there is no dst interval.
562 if (!LIS->hasInterval(CP.getDstReg()))
565 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot();
568 LIS->getInterval(CP.isFlipped() ? CP.getDstReg() : CP.getSrcReg());
570 LIS->getInterval(CP.isFlipped() ? CP.getSrcReg() : CP.getDstReg());
572 // BValNo is a value number in B that is defined by a copy from A. 'B3' in
573 // the example above.
574 VNInfo *BValNo = IntB.getVNInfoAt(CopyIdx);
575 if (!BValNo || BValNo->def != CopyIdx)
578 assert(BValNo->def == CopyIdx && "Copy doesn't define the value?");
580 // AValNo is the value number in A that defines the copy, A3 in the example.
581 VNInfo *AValNo = IntA.getVNInfoAt(CopyIdx.getRegSlot(true));
582 assert(AValNo && "COPY source not live");
584 // If other defs can reach uses of this def, then it's not safe to perform
586 if (AValNo->isPHIDef() || AValNo->isUnused() || AValNo->hasPHIKill())
588 MachineInstr *DefMI = LIS->getInstructionFromIndex(AValNo->def);
591 if (!DefMI->isCommutable())
593 // If DefMI is a two-address instruction then commuting it will change the
594 // destination register.
595 int DefIdx = DefMI->findRegisterDefOperandIdx(IntA.reg);
596 assert(DefIdx != -1);
598 if (!DefMI->isRegTiedToUseOperand(DefIdx, &UseOpIdx))
600 unsigned Op1, Op2, NewDstIdx;
601 if (!TII->findCommutedOpIndices(DefMI, Op1, Op2))
605 else if (Op2 == UseOpIdx)
610 MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
611 unsigned NewReg = NewDstMO.getReg();
612 if (NewReg != IntB.reg || !NewDstMO.isKill())
615 // Make sure there are no other definitions of IntB that would reach the
616 // uses which the new definition can reach.
617 if (hasOtherReachingDefs(IntA, IntB, AValNo, BValNo))
620 // If some of the uses of IntA.reg is already coalesced away, return false.
621 // It's not possible to determine whether it's safe to perform the coalescing.
622 for (MachineRegisterInfo::use_nodbg_iterator UI =
623 MRI->use_nodbg_begin(IntA.reg),
624 UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
625 MachineInstr *UseMI = &*UI;
626 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI);
627 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
628 if (ULR == IntA.end() || ULR->valno != AValNo)
630 // If this use is tied to a def, we can't rewrite the register.
631 if (UseMI->isRegTiedToDefOperand(UI.getOperandNo()))
635 DEBUG(dbgs() << "\tremoveCopyByCommutingDef: " << AValNo->def << '\t'
638 // At this point we have decided that it is legal to do this
639 // transformation. Start by commuting the instruction.
640 MachineBasicBlock *MBB = DefMI->getParent();
641 MachineInstr *NewMI = TII->commuteInstruction(DefMI);
644 if (TargetRegisterInfo::isVirtualRegister(IntA.reg) &&
645 TargetRegisterInfo::isVirtualRegister(IntB.reg) &&
646 !MRI->constrainRegClass(IntB.reg, MRI->getRegClass(IntA.reg)))
648 if (NewMI != DefMI) {
649 LIS->ReplaceMachineInstrInMaps(DefMI, NewMI);
650 MachineBasicBlock::iterator Pos = DefMI;
651 MBB->insert(Pos, NewMI);
654 unsigned OpIdx = NewMI->findRegisterUseOperandIdx(IntA.reg, false);
655 NewMI->getOperand(OpIdx).setIsKill();
657 // If ALR and BLR overlaps and end of BLR extends beyond end of ALR, e.g.
666 // Update uses of IntA of the specific Val# with IntB.
667 for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(IntA.reg),
668 UE = MRI->use_end(); UI != UE;) {
669 MachineOperand &UseMO = UI.getOperand();
670 MachineInstr *UseMI = &*UI;
672 if (UseMI->isDebugValue()) {
673 // FIXME These don't have an instruction index. Not clear we have enough
674 // info to decide whether to do this replacement or not. For now do it.
675 UseMO.setReg(NewReg);
678 SlotIndex UseIdx = LIS->getInstructionIndex(UseMI).getRegSlot(true);
679 LiveInterval::iterator ULR = IntA.FindLiveRangeContaining(UseIdx);
680 if (ULR == IntA.end() || ULR->valno != AValNo)
682 if (TargetRegisterInfo::isPhysicalRegister(NewReg))
683 UseMO.substPhysReg(NewReg, *TRI);
685 UseMO.setReg(NewReg);
688 if (!UseMI->isCopy())
690 if (UseMI->getOperand(0).getReg() != IntB.reg ||
691 UseMI->getOperand(0).getSubReg())
694 // This copy will become a noop. If it's defining a new val#, merge it into
696 SlotIndex DefIdx = UseIdx.getRegSlot();
697 VNInfo *DVNI = IntB.getVNInfoAt(DefIdx);
700 DEBUG(dbgs() << "\t\tnoop: " << DefIdx << '\t' << *UseMI);
701 assert(DVNI->def == DefIdx);
702 BValNo = IntB.MergeValueNumberInto(BValNo, DVNI);
703 ErasedInstrs.insert(UseMI);
704 LIS->RemoveMachineInstrFromMaps(UseMI);
705 UseMI->eraseFromParent();
708 // Extend BValNo by merging in IntA live ranges of AValNo. Val# definition
710 VNInfo *ValNo = BValNo;
711 ValNo->def = AValNo->def;
712 for (LiveInterval::iterator AI = IntA.begin(), AE = IntA.end();
714 if (AI->valno != AValNo) continue;
715 IntB.addRange(LiveRange(AI->start, AI->end, ValNo));
717 DEBUG(dbgs() << "\t\textended: " << IntB << '\n');
719 IntA.removeValNo(AValNo);
720 DEBUG(dbgs() << "\t\ttrimmed: " << IntA << '\n');
725 /// reMaterializeTrivialDef - If the source of a copy is defined by a trivial
726 /// computation, replace the copy by rematerialize the definition.
727 bool RegisterCoalescer::reMaterializeTrivialDef(LiveInterval &SrcInt,
729 MachineInstr *CopyMI) {
730 SlotIndex CopyIdx = LIS->getInstructionIndex(CopyMI).getRegSlot(true);
731 LiveInterval::iterator SrcLR = SrcInt.FindLiveRangeContaining(CopyIdx);
732 assert(SrcLR != SrcInt.end() && "Live range not found!");
733 VNInfo *ValNo = SrcLR->valno;
734 if (ValNo->isPHIDef() || ValNo->isUnused())
736 MachineInstr *DefMI = LIS->getInstructionFromIndex(ValNo->def);
739 assert(DefMI && "Defining instruction disappeared");
740 if (!DefMI->isAsCheapAsAMove())
742 if (!TII->isTriviallyReMaterializable(DefMI, AA))
744 bool SawStore = false;
745 if (!DefMI->isSafeToMove(TII, AA, SawStore))
747 const MCInstrDesc &MCID = DefMI->getDesc();
748 if (MCID.getNumDefs() != 1)
750 if (!DefMI->isImplicitDef()) {
751 // Make sure the copy destination register class fits the instruction
752 // definition register class. The mismatch can happen as a result of earlier
753 // extract_subreg, insert_subreg, subreg_to_reg coalescing.
754 const TargetRegisterClass *RC = TII->getRegClass(MCID, 0, TRI, *MF);
755 if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
756 if (MRI->getRegClass(DstReg) != RC)
758 } else if (!RC->contains(DstReg))
762 MachineBasicBlock *MBB = CopyMI->getParent();
763 MachineBasicBlock::iterator MII =
764 llvm::next(MachineBasicBlock::iterator(CopyMI));
765 TII->reMaterialize(*MBB, MII, DstReg, 0, DefMI, *TRI);
766 MachineInstr *NewMI = prior(MII);
768 // NewMI may have dead implicit defs (E.g. EFLAGS for MOV<bits>r0 on X86).
769 // We need to remember these so we can add intervals once we insert
770 // NewMI into SlotIndexes.
771 SmallVector<unsigned, 4> NewMIImplDefs;
772 for (unsigned i = NewMI->getDesc().getNumOperands(),
773 e = NewMI->getNumOperands(); i != e; ++i) {
774 MachineOperand &MO = NewMI->getOperand(i);
776 assert(MO.isDef() && MO.isImplicit() && MO.isDead() &&
777 TargetRegisterInfo::isPhysicalRegister(MO.getReg()));
778 NewMIImplDefs.push_back(MO.getReg());
782 // CopyMI may have implicit operands, transfer them over to the newly
783 // rematerialized instruction. And update implicit def interval valnos.
784 for (unsigned i = CopyMI->getDesc().getNumOperands(),
785 e = CopyMI->getNumOperands(); i != e; ++i) {
786 MachineOperand &MO = CopyMI->getOperand(i);
788 assert(MO.isImplicit() && "No explicit operands after implict operands.");
789 // Discard VReg implicit defs.
790 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
791 NewMI->addOperand(MO);
796 LIS->ReplaceMachineInstrInMaps(CopyMI, NewMI);
798 SlotIndex NewMIIdx = LIS->getInstructionIndex(NewMI);
799 for (unsigned i = 0, e = NewMIImplDefs.size(); i != e; ++i) {
800 unsigned reg = NewMIImplDefs[i];
801 LiveInterval &li = LIS->getInterval(reg);
802 VNInfo *DeadDefVN = li.getNextValue(NewMIIdx.getRegSlot(),
803 LIS->getVNInfoAllocator());
804 LiveRange lr(NewMIIdx.getRegSlot(), NewMIIdx.getDeadSlot(), DeadDefVN);
808 CopyMI->eraseFromParent();
809 ErasedInstrs.insert(CopyMI);
810 DEBUG(dbgs() << "Remat: " << *NewMI);
813 // The source interval can become smaller because we removed a use.
814 LIS->shrinkToUses(&SrcInt, &DeadDefs);
815 if (!DeadDefs.empty())
821 /// eliminateUndefCopy - ProcessImpicitDefs may leave some copies of <undef>
822 /// values, it only removes local variables. When we have a copy like:
824 /// %vreg1 = COPY %vreg2<undef>
826 /// We delete the copy and remove the corresponding value number from %vreg1.
827 /// Any uses of that value number are marked as <undef>.
828 bool RegisterCoalescer::eliminateUndefCopy(MachineInstr *CopyMI,
829 const CoalescerPair &CP) {
830 SlotIndex Idx = LIS->getInstructionIndex(CopyMI);
831 LiveInterval *SrcInt = &LIS->getInterval(CP.getSrcReg());
832 if (SrcInt->liveAt(Idx))
834 LiveInterval *DstInt = &LIS->getInterval(CP.getDstReg());
835 if (DstInt->liveAt(Idx))
838 // No intervals are live-in to CopyMI - it is undef.
843 VNInfo *DeadVNI = DstInt->getVNInfoAt(Idx.getRegSlot());
844 assert(DeadVNI && "No value defined in DstInt");
845 DstInt->removeValNo(DeadVNI);
847 // Find new undef uses.
848 for (MachineRegisterInfo::reg_nodbg_iterator
849 I = MRI->reg_nodbg_begin(DstInt->reg), E = MRI->reg_nodbg_end();
851 MachineOperand &MO = I.getOperand();
852 if (MO.isDef() || MO.isUndef())
854 MachineInstr *MI = MO.getParent();
855 SlotIndex Idx = LIS->getInstructionIndex(MI);
856 if (DstInt->liveAt(Idx))
859 DEBUG(dbgs() << "\tnew undef: " << Idx << '\t' << *MI);
864 /// updateRegDefsUses - Replace all defs and uses of SrcReg to DstReg and
865 /// update the subregister number if it is not zero. If DstReg is a
866 /// physical register and the existing subregister number of the def / use
867 /// being updated is not zero, make sure to set it to the correct physical
869 void RegisterCoalescer::updateRegDefsUses(unsigned SrcReg,
872 bool DstIsPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
873 LiveInterval *DstInt = DstIsPhys ? 0 : &LIS->getInterval(DstReg);
875 // Update LiveDebugVariables.
876 LDV->renameRegister(SrcReg, DstReg, SubIdx);
878 for (MachineRegisterInfo::reg_iterator I = MRI->reg_begin(SrcReg);
879 MachineInstr *UseMI = I.skipInstruction();) {
880 SmallVector<unsigned,8> Ops;
882 tie(Reads, Writes) = UseMI->readsWritesVirtualRegister(SrcReg, &Ops);
884 // If SrcReg wasn't read, it may still be the case that DstReg is live-in
885 // because SrcReg is a sub-register.
886 if (DstInt && !Reads && SubIdx)
887 Reads = DstInt->liveAt(LIS->getInstructionIndex(UseMI));
889 // Replace SrcReg with DstReg in all UseMI operands.
890 for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
891 MachineOperand &MO = UseMI->getOperand(Ops[i]);
893 // Adjust <undef> flags in case of sub-register joins. We don't want to
894 // turn a full def into a read-modify-write sub-register def and vice
896 if (SubIdx && MO.isDef())
897 MO.setIsUndef(!Reads);
900 MO.substPhysReg(DstReg, *TRI);
902 MO.substVirtReg(DstReg, SubIdx, *TRI);
906 dbgs() << "\t\tupdated: ";
907 if (!UseMI->isDebugValue())
908 dbgs() << LIS->getInstructionIndex(UseMI) << "\t";
914 /// canJoinPhys - Return true if a copy involving a physreg should be joined.
915 bool RegisterCoalescer::canJoinPhys(CoalescerPair &CP) {
916 /// Always join simple intervals that are defined by a single copy from a
917 /// reserved register. This doesn't increase register pressure, so it is
918 /// always beneficial.
919 if (!RegClassInfo.isReserved(CP.getDstReg())) {
920 DEBUG(dbgs() << "\tCan only merge into reserved registers.\n");
924 LiveInterval &JoinVInt = LIS->getInterval(CP.getSrcReg());
925 if (CP.isFlipped() && JoinVInt.containsOneValue())
928 DEBUG(dbgs() << "\tCannot join defs into reserved register.\n");
932 /// joinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
933 /// which are the src/dst of the copy instruction CopyMI. This returns true
934 /// if the copy was successfully coalesced away. If it is not currently
935 /// possible to coalesce this interval, but it may be possible if other
936 /// things get coalesced, then it returns true by reference in 'Again'.
937 bool RegisterCoalescer::joinCopy(MachineInstr *CopyMI, bool &Again) {
940 DEBUG(dbgs() << LIS->getInstructionIndex(CopyMI) << '\t' << *CopyMI);
942 CoalescerPair CP(*TRI);
943 if (!CP.setRegisters(CopyMI)) {
944 DEBUG(dbgs() << "\tNot coalescable.\n");
948 // Dead code elimination. This really should be handled by MachineDCE, but
949 // sometimes dead copies slip through, and we can't generate invalid live
951 if (!CP.isPhys() && CopyMI->allDefsAreDead()) {
952 DEBUG(dbgs() << "\tCopy is dead.\n");
953 DeadDefs.push_back(CopyMI);
959 if (!CP.isPhys() && eliminateUndefCopy(CopyMI, CP)) {
960 DEBUG(dbgs() << "\tEliminated copy of <undef> value.\n");
961 LIS->RemoveMachineInstrFromMaps(CopyMI);
962 CopyMI->eraseFromParent();
963 return false; // Not coalescable.
966 // Coalesced copies are normally removed immediately, but transformations
967 // like removeCopyByCommutingDef() can inadvertently create identity copies.
968 // When that happens, just join the values and remove the copy.
969 if (CP.getSrcReg() == CP.getDstReg()) {
970 LiveInterval &LI = LIS->getInterval(CP.getSrcReg());
971 DEBUG(dbgs() << "\tCopy already coalesced: " << LI << '\n');
972 LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(CopyMI));
973 if (VNInfo *DefVNI = LRQ.valueDefined()) {
974 VNInfo *ReadVNI = LRQ.valueIn();
975 assert(ReadVNI && "No value before copy and no <undef> flag.");
976 assert(ReadVNI != DefVNI && "Cannot read and define the same value.");
977 LI.MergeValueNumberInto(DefVNI, ReadVNI);
978 DEBUG(dbgs() << "\tMerged values: " << LI << '\n');
980 LIS->RemoveMachineInstrFromMaps(CopyMI);
981 CopyMI->eraseFromParent();
987 DEBUG(dbgs() << "\tConsidering merging " << PrintReg(CP.getSrcReg(), TRI)
988 << " with " << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx())
990 if (!canJoinPhys(CP)) {
991 // Before giving up coalescing, if definition of source is defined by
992 // trivial computation, try rematerializing it.
993 if (!CP.isFlipped() &&
994 reMaterializeTrivialDef(LIS->getInterval(CP.getSrcReg()),
995 CP.getDstReg(), CopyMI))
1001 dbgs() << "\tConsidering merging to " << CP.getNewRC()->getName()
1003 if (CP.getDstIdx() && CP.getSrcIdx())
1004 dbgs() << PrintReg(CP.getDstReg()) << " in "
1005 << TRI->getSubRegIndexName(CP.getDstIdx()) << " and "
1006 << PrintReg(CP.getSrcReg()) << " in "
1007 << TRI->getSubRegIndexName(CP.getSrcIdx()) << '\n';
1009 dbgs() << PrintReg(CP.getSrcReg(), TRI) << " in "
1010 << PrintReg(CP.getDstReg(), TRI, CP.getSrcIdx()) << '\n';
1013 // When possible, let DstReg be the larger interval.
1014 if (!CP.isPartial() && LIS->getInterval(CP.getSrcReg()).ranges.size() >
1015 LIS->getInterval(CP.getDstReg()).ranges.size())
1019 // Okay, attempt to join these two intervals. On failure, this returns false.
1020 // Otherwise, if one of the intervals being joined is a physreg, this method
1021 // always canonicalizes DstInt to be it. The output "SrcInt" will not have
1022 // been modified, so we can use this information below to update aliases.
1023 if (!joinIntervals(CP)) {
1024 // Coalescing failed.
1026 // If definition of source is defined by trivial computation, try
1027 // rematerializing it.
1028 if (!CP.isFlipped() &&
1029 reMaterializeTrivialDef(LIS->getInterval(CP.getSrcReg()),
1030 CP.getDstReg(), CopyMI))
1033 // If we can eliminate the copy without merging the live ranges, do so now.
1034 if (!CP.isPartial() && !CP.isPhys()) {
1035 if (adjustCopiesBackFrom(CP, CopyMI) ||
1036 removeCopyByCommutingDef(CP, CopyMI)) {
1037 LIS->RemoveMachineInstrFromMaps(CopyMI);
1038 CopyMI->eraseFromParent();
1039 DEBUG(dbgs() << "\tTrivial!\n");
1044 // Otherwise, we are unable to join the intervals.
1045 DEBUG(dbgs() << "\tInterference!\n");
1046 Again = true; // May be possible to coalesce later.
1050 // Coalescing to a virtual register that is of a sub-register class of the
1051 // other. Make sure the resulting register is set to the right register class.
1052 if (CP.isCrossClass()) {
1054 MRI->setRegClass(CP.getDstReg(), CP.getNewRC());
1057 // Removing sub-register copies can ease the register class constraints.
1058 // Make sure we attempt to inflate the register class of DstReg.
1059 if (!CP.isPhys() && RegClassInfo.isProperSubClass(CP.getNewRC()))
1060 InflateRegs.push_back(CP.getDstReg());
1062 // CopyMI has been erased by joinIntervals at this point. Remove it from
1063 // ErasedInstrs since copyCoalesceWorkList() won't add a successful join back
1064 // to the work list. This keeps ErasedInstrs from growing needlessly.
1065 ErasedInstrs.erase(CopyMI);
1067 // Rewrite all SrcReg operands to DstReg.
1068 // Also update DstReg operands to include DstIdx if it is set.
1070 updateRegDefsUses(CP.getDstReg(), CP.getDstReg(), CP.getDstIdx());
1071 updateRegDefsUses(CP.getSrcReg(), CP.getDstReg(), CP.getSrcIdx());
1073 // SrcReg is guaranteed to be the register whose live interval that is
1075 LIS->removeInterval(CP.getSrcReg());
1077 // Update regalloc hint.
1078 TRI->UpdateRegAllocHint(CP.getSrcReg(), CP.getDstReg(), *MF);
1080 DEBUG(dbgs() << "\tJoined. Result = " << PrintReg(CP.getDstReg(), TRI)
1081 << ' ' << LIS->getInterval(CP.getDstReg()) << '\n');
1087 /// Attempt joining with a reserved physreg.
1088 bool RegisterCoalescer::joinReservedPhysReg(CoalescerPair &CP) {
1089 assert(CP.isPhys() && "Must be a physreg copy");
1090 assert(RegClassInfo.isReserved(CP.getDstReg()) && "Not a reserved register");
1091 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
1092 DEBUG(dbgs() << "\t\tRHS = " << PrintReg(CP.getSrcReg()) << ' ' << RHS
1095 assert(CP.isFlipped() && RHS.containsOneValue() &&
1096 "Invalid join with reserved register");
1098 // Optimization for reserved registers like ESP. We can only merge with a
1099 // reserved physreg if RHS has a single value that is a copy of CP.DstReg().
1100 // The live range of the reserved register will look like a set of dead defs
1101 // - we don't properly track the live range of reserved registers.
1103 // Deny any overlapping intervals. This depends on all the reserved
1104 // register live ranges to look like dead defs.
1105 if (LIS->trackingRegUnits()) {
1106 for (MCRegUnitIterator UI(CP.getDstReg(), TRI); UI.isValid(); ++UI)
1107 if (RHS.overlaps(LIS->getRegUnit(*UI))) {
1108 DEBUG(dbgs() << "\t\tInterference: " << PrintRegUnit(*UI, TRI) << '\n');
1112 for (MCRegAliasIterator AS(CP.getDstReg(), TRI, true); AS.isValid(); ++AS) {
1113 if (!LIS->hasInterval(*AS)) {
1114 // Make sure at least DstReg itself exists before attempting a join.
1115 if (*AS == CP.getDstReg())
1116 LIS->getOrCreateInterval(CP.getDstReg());
1119 if (RHS.overlaps(LIS->getInterval(*AS))) {
1120 DEBUG(dbgs() << "\t\tInterference: " << PrintReg(*AS, TRI) << '\n');
1125 // Skip any value computations, we are not adding new values to the
1126 // reserved register. Also skip merging the live ranges, the reserved
1127 // register live range doesn't need to be accurate as long as all the
1130 // We don't track kills for reserved registers.
1131 MRI->clearKillFlags(CP.getSrcReg());
1136 /// ComputeUltimateVN - Assuming we are going to join two live intervals,
1137 /// compute what the resultant value numbers for each value in the input two
1138 /// ranges will be. This is complicated by copies between the two which can
1139 /// and will commonly cause multiple value numbers to be merged into one.
1141 /// VN is the value number that we're trying to resolve. InstDefiningValue
1142 /// keeps track of the new InstDefiningValue assignment for the result
1143 /// LiveInterval. ThisFromOther/OtherFromThis are sets that keep track of
1144 /// whether a value in this or other is a copy from the opposite set.
1145 /// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
1146 /// already been assigned.
1148 /// ThisFromOther[x] - If x is defined as a copy from the other interval, this
1149 /// contains the value number the copy is from.
1151 static unsigned ComputeUltimateVN(VNInfo *VNI,
1152 SmallVector<VNInfo*, 16> &NewVNInfo,
1153 DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
1154 DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
1155 SmallVector<int, 16> &ThisValNoAssignments,
1156 SmallVector<int, 16> &OtherValNoAssignments) {
1157 unsigned VN = VNI->id;
1159 // If the VN has already been computed, just return it.
1160 if (ThisValNoAssignments[VN] >= 0)
1161 return ThisValNoAssignments[VN];
1162 assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers");
1164 // If this val is not a copy from the other val, then it must be a new value
1165 // number in the destination.
1166 DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
1167 if (I == ThisFromOther.end()) {
1168 NewVNInfo.push_back(VNI);
1169 return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
1171 VNInfo *OtherValNo = I->second;
1173 // Otherwise, this *is* a copy from the RHS. If the other side has already
1174 // been computed, return it.
1175 if (OtherValNoAssignments[OtherValNo->id] >= 0)
1176 return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
1178 // Mark this value number as currently being computed, then ask what the
1179 // ultimate value # of the other value is.
1180 ThisValNoAssignments[VN] = -2;
1181 unsigned UltimateVN =
1182 ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
1183 OtherValNoAssignments, ThisValNoAssignments);
1184 return ThisValNoAssignments[VN] = UltimateVN;
1188 // Find out if we have something like
1191 // if so, we can pretend this is actually
1194 // which allows us to coalesce A and B.
1195 // VNI is the definition of B. LR is the life range of A that includes
1196 // the slot just before B. If we return true, we add "B = X" to DupCopies.
1197 // This implies that A dominates B.
1198 static bool RegistersDefinedFromSameValue(LiveIntervals &li,
1199 const TargetRegisterInfo &tri,
1203 SmallVector<MachineInstr*, 8> &DupCopies) {
1204 // FIXME: This is very conservative. For example, we don't handle
1205 // physical registers.
1207 MachineInstr *MI = li.getInstructionFromIndex(VNI->def);
1209 if (!MI || !MI->isFullCopy() || CP.isPartial() || CP.isPhys())
1212 unsigned Dst = MI->getOperand(0).getReg();
1213 unsigned Src = MI->getOperand(1).getReg();
1215 if (!TargetRegisterInfo::isVirtualRegister(Src) ||
1216 !TargetRegisterInfo::isVirtualRegister(Dst))
1219 unsigned A = CP.getDstReg();
1220 unsigned B = CP.getSrcReg();
1226 const MachineInstr *OtherMI = li.getInstructionFromIndex(OtherVNI->def);
1228 if (!OtherMI || !OtherMI->isFullCopy())
1231 unsigned OtherDst = OtherMI->getOperand(0).getReg();
1232 unsigned OtherSrc = OtherMI->getOperand(1).getReg();
1234 if (!TargetRegisterInfo::isVirtualRegister(OtherSrc) ||
1235 !TargetRegisterInfo::isVirtualRegister(OtherDst))
1238 assert(OtherDst == B);
1240 if (Src != OtherSrc)
1243 // If the copies use two different value numbers of X, we cannot merge
1245 LiveInterval &SrcInt = li.getInterval(Src);
1246 // getVNInfoBefore returns NULL for undef copies. In this case, the
1247 // optimization is still safe.
1248 if (SrcInt.getVNInfoBefore(OtherVNI->def) != SrcInt.getVNInfoBefore(VNI->def))
1251 DupCopies.push_back(MI);
1256 /// joinIntervals - Attempt to join these two intervals. On failure, this
1258 bool RegisterCoalescer::joinIntervals(CoalescerPair &CP) {
1259 // Handle physreg joins separately.
1261 return joinReservedPhysReg(CP);
1263 LiveInterval &RHS = LIS->getInterval(CP.getSrcReg());
1264 DEBUG(dbgs() << "\t\tRHS = " << PrintReg(CP.getSrcReg()) << ' ' << RHS
1267 // Compute the final value assignment, assuming that the live ranges can be
1269 SmallVector<int, 16> LHSValNoAssignments;
1270 SmallVector<int, 16> RHSValNoAssignments;
1271 DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
1272 DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
1273 SmallVector<VNInfo*, 16> NewVNInfo;
1275 SmallVector<MachineInstr*, 8> DupCopies;
1276 SmallVector<MachineInstr*, 8> DeadCopies;
1278 LiveInterval &LHS = LIS->getOrCreateInterval(CP.getDstReg());
1279 DEBUG(dbgs() << "\t\tLHS = " << PrintReg(CP.getDstReg(), TRI) << ' ' << LHS
1282 // Loop over the value numbers of the LHS, seeing if any are defined from
1284 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1287 if (VNI->isUnused() || VNI->isPHIDef())
1289 MachineInstr *MI = LIS->getInstructionFromIndex(VNI->def);
1290 assert(MI && "Missing def");
1291 if (!MI->isCopyLike()) // Src not defined by a copy?
1294 // Figure out the value # from the RHS.
1295 VNInfo *OtherVNI = RHS.getVNInfoBefore(VNI->def);
1296 // The copy could be to an aliased physreg.
1300 // DstReg is known to be a register in the LHS interval. If the src is
1301 // from the RHS interval, we can use its value #.
1302 if (CP.isCoalescable(MI))
1303 DeadCopies.push_back(MI);
1304 else if (!RegistersDefinedFromSameValue(*LIS, *TRI, CP, VNI, OtherVNI,
1308 LHSValsDefinedFromRHS[VNI] = OtherVNI;
1311 // Loop over the value numbers of the RHS, seeing if any are defined from
1313 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1316 if (VNI->isUnused() || VNI->isPHIDef())
1318 MachineInstr *MI = LIS->getInstructionFromIndex(VNI->def);
1319 assert(MI && "Missing def");
1320 if (!MI->isCopyLike()) // Src not defined by a copy?
1323 // Figure out the value # from the LHS.
1324 VNInfo *OtherVNI = LHS.getVNInfoBefore(VNI->def);
1325 // The copy could be to an aliased physreg.
1329 // DstReg is known to be a register in the RHS interval. If the src is
1330 // from the LHS interval, we can use its value #.
1331 if (CP.isCoalescable(MI))
1332 DeadCopies.push_back(MI);
1333 else if (!RegistersDefinedFromSameValue(*LIS, *TRI, CP, VNI, OtherVNI,
1337 RHSValsDefinedFromLHS[VNI] = OtherVNI;
1340 LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
1341 RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
1342 NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
1344 for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
1347 unsigned VN = VNI->id;
1348 if (LHSValNoAssignments[VN] >= 0 || VNI->isUnused())
1350 ComputeUltimateVN(VNI, NewVNInfo,
1351 LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
1352 LHSValNoAssignments, RHSValNoAssignments);
1354 for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
1357 unsigned VN = VNI->id;
1358 if (RHSValNoAssignments[VN] >= 0 || VNI->isUnused())
1360 // If this value number isn't a copy from the LHS, it's a new number.
1361 if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
1362 NewVNInfo.push_back(VNI);
1363 RHSValNoAssignments[VN] = NewVNInfo.size()-1;
1367 ComputeUltimateVN(VNI, NewVNInfo,
1368 RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
1369 RHSValNoAssignments, LHSValNoAssignments);
1372 // Armed with the mappings of LHS/RHS values to ultimate values, walk the
1373 // interval lists to see if these intervals are coalescable.
1374 LiveInterval::const_iterator I = LHS.begin();
1375 LiveInterval::const_iterator IE = LHS.end();
1376 LiveInterval::const_iterator J = RHS.begin();
1377 LiveInterval::const_iterator JE = RHS.end();
1379 // Collect interval end points that will no longer be kills.
1380 SmallVector<MachineInstr*, 8> LHSOldKills;
1381 SmallVector<MachineInstr*, 8> RHSOldKills;
1383 // Skip ahead until the first place of potential sharing.
1384 if (I != IE && J != JE) {
1385 if (I->start < J->start) {
1386 I = std::upper_bound(I, IE, J->start);
1387 if (I != LHS.begin()) --I;
1388 } else if (J->start < I->start) {
1389 J = std::upper_bound(J, JE, I->start);
1390 if (J != RHS.begin()) --J;
1394 while (I != IE && J != JE) {
1395 // Determine if these two live ranges overlap.
1396 // If so, check value # info to determine if they are really different.
1397 if (I->end > J->start && J->end > I->start) {
1398 // If the live range overlap will map to the same value number in the
1399 // result liverange, we can still coalesce them. If not, we can't.
1400 if (LHSValNoAssignments[I->valno->id] !=
1401 RHSValNoAssignments[J->valno->id])
1404 // Extended live ranges should no longer be killed.
1405 if (!I->end.isBlock() && I->end < J->end)
1406 if (MachineInstr *MI = LIS->getInstructionFromIndex(I->end))
1407 LHSOldKills.push_back(MI);
1408 if (!J->end.isBlock() && J->end < I->end)
1409 if (MachineInstr *MI = LIS->getInstructionFromIndex(J->end))
1410 RHSOldKills.push_back(MI);
1413 if (I->end < J->end)
1419 // Update kill info. Some live ranges are extended due to copy coalescing.
1420 for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
1421 E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
1422 VNInfo *VNI = I->first;
1423 unsigned LHSValID = LHSValNoAssignments[VNI->id];
1424 if (VNI->hasPHIKill())
1425 NewVNInfo[LHSValID]->setHasPHIKill(true);
1428 // Update kill info. Some live ranges are extended due to copy coalescing.
1429 for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
1430 E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
1431 VNInfo *VNI = I->first;
1432 unsigned RHSValID = RHSValNoAssignments[VNI->id];
1433 if (VNI->hasPHIKill())
1434 NewVNInfo[RHSValID]->setHasPHIKill(true);
1437 // Clear kill flags where live ranges are extended.
1438 while (!LHSOldKills.empty())
1439 LHSOldKills.pop_back_val()->clearRegisterKills(LHS.reg, TRI);
1440 while (!RHSOldKills.empty())
1441 RHSOldKills.pop_back_val()->clearRegisterKills(RHS.reg, TRI);
1443 if (LHSValNoAssignments.empty())
1444 LHSValNoAssignments.push_back(-1);
1445 if (RHSValNoAssignments.empty())
1446 RHSValNoAssignments.push_back(-1);
1448 // Now erase all the redundant copies.
1449 for (unsigned i = 0, e = DeadCopies.size(); i != e; ++i) {
1450 MachineInstr *MI = DeadCopies[i];
1451 if (!ErasedInstrs.insert(MI))
1453 DEBUG(dbgs() << "\t\terased:\t" << LIS->getInstructionIndex(MI)
1455 LIS->RemoveMachineInstrFromMaps(MI);
1456 MI->eraseFromParent();
1459 SmallVector<unsigned, 8> SourceRegisters;
1460 for (SmallVector<MachineInstr*, 8>::iterator I = DupCopies.begin(),
1461 E = DupCopies.end(); I != E; ++I) {
1462 MachineInstr *MI = *I;
1463 if (!ErasedInstrs.insert(MI))
1466 // We have pretended that the assignment to B in
1469 // was actually a copy from A. Now that we decided to coalesce A and B,
1470 // transform the code into
1472 unsigned Src = MI->getOperand(1).getReg();
1473 SourceRegisters.push_back(Src);
1474 LIS->RemoveMachineInstrFromMaps(MI);
1475 MI->eraseFromParent();
1478 // If B = X was the last use of X in a liverange, we have to shrink it now
1479 // that B = X is gone.
1480 for (SmallVector<unsigned, 8>::iterator I = SourceRegisters.begin(),
1481 E = SourceRegisters.end(); I != E; ++I) {
1482 LIS->shrinkToUses(&LIS->getInterval(*I));
1485 // If we get here, we know that we can coalesce the live ranges. Ask the
1486 // intervals to coalesce themselves now.
1487 LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo,
1493 // DepthMBBCompare - Comparison predicate that sort first based on the loop
1494 // depth of the basic block (the unsigned), and then on the MBB number.
1495 struct DepthMBBCompare {
1496 typedef std::pair<unsigned, MachineBasicBlock*> DepthMBBPair;
1497 bool operator()(const DepthMBBPair &LHS, const DepthMBBPair &RHS) const {
1498 // Deeper loops first
1499 if (LHS.first != RHS.first)
1500 return LHS.first > RHS.first;
1502 // Prefer blocks that are more connected in the CFG. This takes care of
1503 // the most difficult copies first while intervals are short.
1504 unsigned cl = LHS.second->pred_size() + LHS.second->succ_size();
1505 unsigned cr = RHS.second->pred_size() + RHS.second->succ_size();
1509 // As a last resort, sort by block number.
1510 return LHS.second->getNumber() < RHS.second->getNumber();
1515 // Try joining WorkList copies starting from index From.
1516 // Null out any successful joins.
1517 bool RegisterCoalescer::copyCoalesceWorkList(unsigned From) {
1518 assert(From <= WorkList.size() && "Out of range");
1519 bool Progress = false;
1520 for (unsigned i = From, e = WorkList.size(); i != e; ++i) {
1523 // Skip instruction pointers that have already been erased, for example by
1524 // dead code elimination.
1525 if (ErasedInstrs.erase(WorkList[i])) {
1530 bool Success = joinCopy(WorkList[i], Again);
1531 Progress |= Success;
1532 if (Success || !Again)
1539 RegisterCoalescer::copyCoalesceInMBB(MachineBasicBlock *MBB) {
1540 DEBUG(dbgs() << MBB->getName() << ":\n");
1542 // Collect all copy-like instructions in MBB. Don't start coalescing anything
1543 // yet, it might invalidate the iterator.
1544 const unsigned PrevSize = WorkList.size();
1545 for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
1547 if (MII->isCopyLike())
1548 WorkList.push_back(MII);
1550 // Try coalescing the collected copies immediately, and remove the nulls.
1551 // This prevents the WorkList from getting too large since most copies are
1552 // joinable on the first attempt.
1553 if (copyCoalesceWorkList(PrevSize))
1554 WorkList.erase(std::remove(WorkList.begin() + PrevSize, WorkList.end(),
1555 (MachineInstr*)0), WorkList.end());
1558 void RegisterCoalescer::joinAllIntervals() {
1559 DEBUG(dbgs() << "********** JOINING INTERVALS ***********\n");
1560 assert(WorkList.empty() && "Old data still around.");
1562 if (Loops->empty()) {
1563 // If there are no loops in the function, join intervals in function order.
1564 for (MachineFunction::iterator I = MF->begin(), E = MF->end();
1566 copyCoalesceInMBB(I);
1568 // Otherwise, join intervals in inner loops before other intervals.
1569 // Unfortunately we can't just iterate over loop hierarchy here because
1570 // there may be more MBB's than BB's. Collect MBB's for sorting.
1572 // Join intervals in the function prolog first. We want to join physical
1573 // registers with virtual registers before the intervals got too long.
1574 std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
1575 for (MachineFunction::iterator I = MF->begin(), E = MF->end();I != E;++I){
1576 MachineBasicBlock *MBB = I;
1577 MBBs.push_back(std::make_pair(Loops->getLoopDepth(MBB), I));
1580 // Sort by loop depth.
1581 std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
1583 // Finally, join intervals in loop nest order.
1584 for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
1585 copyCoalesceInMBB(MBBs[i].second);
1588 // Joining intervals can allow other intervals to be joined. Iteratively join
1589 // until we make no progress.
1590 while (copyCoalesceWorkList())
1594 void RegisterCoalescer::releaseMemory() {
1595 ErasedInstrs.clear();
1598 InflateRegs.clear();
1601 bool RegisterCoalescer::runOnMachineFunction(MachineFunction &fn) {
1603 MRI = &fn.getRegInfo();
1604 TM = &fn.getTarget();
1605 TRI = TM->getRegisterInfo();
1606 TII = TM->getInstrInfo();
1607 LIS = &getAnalysis<LiveIntervals>();
1608 LDV = &getAnalysis<LiveDebugVariables>();
1609 AA = &getAnalysis<AliasAnalysis>();
1610 Loops = &getAnalysis<MachineLoopInfo>();
1612 DEBUG(dbgs() << "********** SIMPLE REGISTER COALESCING **********\n"
1613 << "********** Function: "
1614 << ((Value*)MF->getFunction())->getName() << '\n');
1616 if (VerifyCoalescing)
1617 MF->verify(this, "Before register coalescing");
1619 RegClassInfo.runOnMachineFunction(fn);
1621 // Join (coalesce) intervals if requested.
1625 // After deleting a lot of copies, register classes may be less constrained.
1626 // Removing sub-register operands may allow GR32_ABCD -> GR32 and DPR_VFP2 ->
1628 array_pod_sort(InflateRegs.begin(), InflateRegs.end());
1629 InflateRegs.erase(std::unique(InflateRegs.begin(), InflateRegs.end()),
1631 DEBUG(dbgs() << "Trying to inflate " << InflateRegs.size() << " regs.\n");
1632 for (unsigned i = 0, e = InflateRegs.size(); i != e; ++i) {
1633 unsigned Reg = InflateRegs[i];
1634 if (MRI->reg_nodbg_empty(Reg))
1636 if (MRI->recomputeRegClass(Reg, *TM)) {
1637 DEBUG(dbgs() << PrintReg(Reg) << " inflated to "
1638 << MRI->getRegClass(Reg)->getName() << '\n');
1645 if (VerifyCoalescing)
1646 MF->verify(this, "After register coalescing");
1650 /// print - Implement the dump method.
1651 void RegisterCoalescer::print(raw_ostream &O, const Module* m) const {