1 //===-- RegAllocGreedy.cpp - greedy register allocator --------------------===//
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 defines the RAGreedy function pass for register allocation in
13 //===----------------------------------------------------------------------===//
15 #include "llvm/CodeGen/Passes.h"
16 #include "AllocationOrder.h"
17 #include "InterferenceCache.h"
18 #include "LiveDebugVariables.h"
19 #include "RegAllocBase.h"
20 #include "SpillPlacement.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/CalcSpillWeights.h"
26 #include "llvm/CodeGen/EdgeBundles.h"
27 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
28 #include "llvm/CodeGen/LiveRangeEdit.h"
29 #include "llvm/CodeGen/LiveRegMatrix.h"
30 #include "llvm/CodeGen/LiveStackAnalysis.h"
31 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
32 #include "llvm/CodeGen/MachineDominators.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/RegAllocRegistry.h"
37 #include "llvm/CodeGen/RegisterClassInfo.h"
38 #include "llvm/CodeGen/VirtRegMap.h"
39 #include "llvm/IR/LLVMContext.h"
40 #include "llvm/PassAnalysisSupport.h"
41 #include "llvm/Support/BranchProbability.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/Timer.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Target/TargetSubtargetInfo.h"
52 #define DEBUG_TYPE "regalloc"
54 STATISTIC(NumGlobalSplits, "Number of split global live ranges");
55 STATISTIC(NumLocalSplits, "Number of split local live ranges");
56 STATISTIC(NumEvicted, "Number of interferences evicted");
58 static cl::opt<SplitEditor::ComplementSpillMode>
59 SplitSpillMode("split-spill-mode", cl::Hidden,
60 cl::desc("Spill mode for splitting live ranges"),
61 cl::values(clEnumValN(SplitEditor::SM_Partition, "default", "Default"),
62 clEnumValN(SplitEditor::SM_Size, "size", "Optimize for size"),
63 clEnumValN(SplitEditor::SM_Speed, "speed", "Optimize for speed"),
65 cl::init(SplitEditor::SM_Partition));
67 static cl::opt<unsigned>
68 LastChanceRecoloringMaxDepth("lcr-max-depth", cl::Hidden,
69 cl::desc("Last chance recoloring max depth"),
72 static cl::opt<unsigned> LastChanceRecoloringMaxInterference(
73 "lcr-max-interf", cl::Hidden,
74 cl::desc("Last chance recoloring maximum number of considered"
75 " interference at a time"),
79 ExhaustiveSearch("exhaustive-register-search", cl::NotHidden,
80 cl::desc("Exhaustive Search for registers bypassing the depth "
81 "and interference cutoffs of last chance recoloring"));
83 static cl::opt<bool> EnableLocalReassignment(
84 "enable-local-reassign", cl::Hidden,
85 cl::desc("Local reassignment can yield better allocation decisions, but "
86 "may be compile time intensive"),
89 // FIXME: Find a good default for this flag and remove the flag.
90 static cl::opt<unsigned>
91 CSRFirstTimeCost("regalloc-csr-first-time-cost",
92 cl::desc("Cost for first time use of callee-saved register."),
93 cl::init(0), cl::Hidden);
95 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
96 createGreedyRegisterAllocator);
99 class RAGreedy : public MachineFunctionPass,
101 private LiveRangeEdit::Delegate {
102 // Convenient shortcuts.
103 typedef std::priority_queue<std::pair<unsigned, unsigned> > PQueue;
104 typedef SmallPtrSet<LiveInterval *, 4> SmallLISet;
105 typedef SmallSet<unsigned, 16> SmallVirtRegSet;
110 // Shortcuts to some useful interface.
111 const TargetInstrInfo *TII;
112 const TargetRegisterInfo *TRI;
113 RegisterClassInfo RCI;
116 SlotIndexes *Indexes;
117 MachineBlockFrequencyInfo *MBFI;
118 MachineDominatorTree *DomTree;
119 MachineLoopInfo *Loops;
120 EdgeBundles *Bundles;
121 SpillPlacement *SpillPlacer;
122 LiveDebugVariables *DebugVars;
125 std::unique_ptr<Spiller> SpillerInstance;
127 unsigned NextCascade;
129 // Live ranges pass through a number of stages as we try to allocate them.
130 // Some of the stages may also create new live ranges:
132 // - Region splitting.
133 // - Per-block splitting.
134 // - Local splitting.
137 // Ranges produced by one of the stages skip the previous stages when they are
138 // dequeued. This improves performance because we can skip interference checks
139 // that are unlikely to give any results. It also guarantees that the live
140 // range splitting algorithm terminates, something that is otherwise hard to
142 enum LiveRangeStage {
143 /// Newly created live range that has never been queued.
146 /// Only attempt assignment and eviction. Then requeue as RS_Split.
149 /// Attempt live range splitting if assignment is impossible.
152 /// Attempt more aggressive live range splitting that is guaranteed to make
153 /// progress. This is used for split products that may not be making
157 /// Live range will be spilled. No more splitting will be attempted.
160 /// There is nothing more we can do to this live range. Abort compilation
161 /// if it can't be assigned.
165 // Enum CutOffStage to keep a track whether the register allocation failed
166 // because of the cutoffs encountered in last chance recoloring.
167 // Note: This is used as bitmask. New value should be next power of 2.
169 // No cutoffs encountered
172 // lcr-max-depth cutoff encountered
175 // lcr-max-interf cutoff encountered
182 static const char *const StageName[];
185 // RegInfo - Keep additional information about each live range.
187 LiveRangeStage Stage;
189 // Cascade - Eviction loop prevention. See canEvictInterference().
192 RegInfo() : Stage(RS_New), Cascade(0) {}
195 IndexedMap<RegInfo, VirtReg2IndexFunctor> ExtraRegInfo;
197 LiveRangeStage getStage(const LiveInterval &VirtReg) const {
198 return ExtraRegInfo[VirtReg.reg].Stage;
201 void setStage(const LiveInterval &VirtReg, LiveRangeStage Stage) {
202 ExtraRegInfo.resize(MRI->getNumVirtRegs());
203 ExtraRegInfo[VirtReg.reg].Stage = Stage;
206 template<typename Iterator>
207 void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
208 ExtraRegInfo.resize(MRI->getNumVirtRegs());
209 for (;Begin != End; ++Begin) {
210 unsigned Reg = *Begin;
211 if (ExtraRegInfo[Reg].Stage == RS_New)
212 ExtraRegInfo[Reg].Stage = NewStage;
216 /// Cost of evicting interference.
217 struct EvictionCost {
218 unsigned BrokenHints; ///< Total number of broken hints.
219 float MaxWeight; ///< Maximum spill weight evicted.
221 EvictionCost(): BrokenHints(0), MaxWeight(0) {}
223 bool isMax() const { return BrokenHints == ~0u; }
225 void setMax() { BrokenHints = ~0u; }
227 void setBrokenHints(unsigned NHints) { BrokenHints = NHints; }
229 bool operator<(const EvictionCost &O) const {
230 return std::tie(BrokenHints, MaxWeight) <
231 std::tie(O.BrokenHints, O.MaxWeight);
236 std::unique_ptr<SplitAnalysis> SA;
237 std::unique_ptr<SplitEditor> SE;
239 /// Cached per-block interference maps
240 InterferenceCache IntfCache;
242 /// All basic blocks where the current register has uses.
243 SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints;
245 /// Global live range splitting candidate info.
246 struct GlobalSplitCandidate {
247 // Register intended for assignment, or 0.
250 // SplitKit interval index for this candidate.
253 // Interference for PhysReg.
254 InterferenceCache::Cursor Intf;
256 // Bundles where this candidate should be live.
257 BitVector LiveBundles;
258 SmallVector<unsigned, 8> ActiveBlocks;
260 void reset(InterferenceCache &Cache, unsigned Reg) {
263 Intf.setPhysReg(Cache, Reg);
265 ActiveBlocks.clear();
268 // Set B[i] = C for every live bundle where B[i] was NoCand.
269 unsigned getBundles(SmallVectorImpl<unsigned> &B, unsigned C) {
271 for (int i = LiveBundles.find_first(); i >= 0;
272 i = LiveBundles.find_next(i))
273 if (B[i] == NoCand) {
281 /// Candidate info for each PhysReg in AllocationOrder.
282 /// This vector never shrinks, but grows to the size of the largest register
284 SmallVector<GlobalSplitCandidate, 32> GlobalCand;
286 enum : unsigned { NoCand = ~0u };
288 /// Candidate map. Each edge bundle is assigned to a GlobalCand entry, or to
289 /// NoCand which indicates the stack interval.
290 SmallVector<unsigned, 32> BundleCand;
292 /// Callee-save register cost, calculated once per machine function.
293 BlockFrequency CSRCost;
295 /// Run or not the local reassignment heuristic. This information is
296 /// obtained from the TargetSubtargetInfo.
297 bool EnableLocalReassign;
299 /// Set of broken hints that may be reconciled later because of eviction.
300 SmallSetVector<LiveInterval *, 8> SetOfBrokenHints;
305 /// Return the pass name.
306 const char* getPassName() const override {
307 return "Greedy Register Allocator";
310 /// RAGreedy analysis usage.
311 void getAnalysisUsage(AnalysisUsage &AU) const override;
312 void releaseMemory() override;
313 Spiller &spiller() override { return *SpillerInstance; }
314 void enqueue(LiveInterval *LI) override;
315 LiveInterval *dequeue() override;
316 unsigned selectOrSplit(LiveInterval&, SmallVectorImpl<unsigned>&) override;
317 void aboutToRemoveInterval(LiveInterval &) override;
319 /// Perform register allocation.
320 bool runOnMachineFunction(MachineFunction &mf) override;
325 unsigned selectOrSplitImpl(LiveInterval &, SmallVectorImpl<unsigned> &,
326 SmallVirtRegSet &, unsigned = 0);
328 bool LRE_CanEraseVirtReg(unsigned) override;
329 void LRE_WillShrinkVirtReg(unsigned) override;
330 void LRE_DidCloneVirtReg(unsigned, unsigned) override;
331 void enqueue(PQueue &CurQueue, LiveInterval *LI);
332 LiveInterval *dequeue(PQueue &CurQueue);
334 BlockFrequency calcSpillCost();
335 bool addSplitConstraints(InterferenceCache::Cursor, BlockFrequency&);
336 void addThroughConstraints(InterferenceCache::Cursor, ArrayRef<unsigned>);
337 void growRegion(GlobalSplitCandidate &Cand);
338 BlockFrequency calcGlobalSplitCost(GlobalSplitCandidate&);
339 bool calcCompactRegion(GlobalSplitCandidate&);
340 void splitAroundRegion(LiveRangeEdit&, ArrayRef<unsigned>);
341 void calcGapWeights(unsigned, SmallVectorImpl<float>&);
342 unsigned canReassign(LiveInterval &VirtReg, unsigned PhysReg);
343 bool shouldEvict(LiveInterval &A, bool, LiveInterval &B, bool);
344 bool canEvictInterference(LiveInterval&, unsigned, bool, EvictionCost&);
345 void evictInterference(LiveInterval&, unsigned,
346 SmallVectorImpl<unsigned>&);
347 bool mayRecolorAllInterferences(unsigned PhysReg, LiveInterval &VirtReg,
348 SmallLISet &RecoloringCandidates,
349 const SmallVirtRegSet &FixedRegisters);
351 unsigned tryAssign(LiveInterval&, AllocationOrder&,
352 SmallVectorImpl<unsigned>&);
353 unsigned tryEvict(LiveInterval&, AllocationOrder&,
354 SmallVectorImpl<unsigned>&, unsigned = ~0u);
355 unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
356 SmallVectorImpl<unsigned>&);
357 /// Calculate cost of region splitting.
358 unsigned calculateRegionSplitCost(LiveInterval &VirtReg,
359 AllocationOrder &Order,
360 BlockFrequency &BestCost,
361 unsigned &NumCands, bool IgnoreCSR);
362 /// Perform region splitting.
363 unsigned doRegionSplit(LiveInterval &VirtReg, unsigned BestCand,
365 SmallVectorImpl<unsigned> &NewVRegs);
366 /// Check other options before using a callee-saved register for the first
368 unsigned tryAssignCSRFirstTime(LiveInterval &VirtReg, AllocationOrder &Order,
369 unsigned PhysReg, unsigned &CostPerUseLimit,
370 SmallVectorImpl<unsigned> &NewVRegs);
371 void initializeCSRCost();
372 unsigned tryBlockSplit(LiveInterval&, AllocationOrder&,
373 SmallVectorImpl<unsigned>&);
374 unsigned tryInstructionSplit(LiveInterval&, AllocationOrder&,
375 SmallVectorImpl<unsigned>&);
376 unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
377 SmallVectorImpl<unsigned>&);
378 unsigned trySplit(LiveInterval&, AllocationOrder&,
379 SmallVectorImpl<unsigned>&);
380 unsigned tryLastChanceRecoloring(LiveInterval &, AllocationOrder &,
381 SmallVectorImpl<unsigned> &,
382 SmallVirtRegSet &, unsigned);
383 bool tryRecoloringCandidates(PQueue &, SmallVectorImpl<unsigned> &,
384 SmallVirtRegSet &, unsigned);
385 void tryHintRecoloring(LiveInterval &);
386 void tryHintsRecoloring();
388 /// Model the information carried by one end of a copy.
390 /// The frequency of the copy.
392 /// The virtual register or physical register.
394 /// Its currently assigned register.
395 /// In case of a physical register Reg == PhysReg.
397 HintInfo(BlockFrequency Freq, unsigned Reg, unsigned PhysReg)
398 : Freq(Freq), Reg(Reg), PhysReg(PhysReg) {}
400 typedef SmallVector<HintInfo, 4> HintsInfo;
401 BlockFrequency getBrokenHintFreq(const HintsInfo &, unsigned);
402 void collectHintInfo(unsigned, HintsInfo &);
404 } // end anonymous namespace
406 char RAGreedy::ID = 0;
409 const char *const RAGreedy::StageName[] = {
419 // Hysteresis to use when comparing floats.
420 // This helps stabilize decisions based on float comparisons.
421 const float Hysteresis = (2007 / 2048.0f); // 0.97998046875
424 FunctionPass* llvm::createGreedyRegisterAllocator() {
425 return new RAGreedy();
428 RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
429 initializeLiveDebugVariablesPass(*PassRegistry::getPassRegistry());
430 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
431 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
432 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
433 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
434 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
435 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
436 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
437 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
438 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
439 initializeLiveRegMatrixPass(*PassRegistry::getPassRegistry());
440 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
441 initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
444 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
445 AU.setPreservesCFG();
446 AU.addRequired<MachineBlockFrequencyInfo>();
447 AU.addPreserved<MachineBlockFrequencyInfo>();
448 AU.addRequired<AliasAnalysis>();
449 AU.addPreserved<AliasAnalysis>();
450 AU.addRequired<LiveIntervals>();
451 AU.addPreserved<LiveIntervals>();
452 AU.addRequired<SlotIndexes>();
453 AU.addPreserved<SlotIndexes>();
454 AU.addRequired<LiveDebugVariables>();
455 AU.addPreserved<LiveDebugVariables>();
456 AU.addRequired<LiveStacks>();
457 AU.addPreserved<LiveStacks>();
458 AU.addRequired<MachineDominatorTree>();
459 AU.addPreserved<MachineDominatorTree>();
460 AU.addRequired<MachineLoopInfo>();
461 AU.addPreserved<MachineLoopInfo>();
462 AU.addRequired<VirtRegMap>();
463 AU.addPreserved<VirtRegMap>();
464 AU.addRequired<LiveRegMatrix>();
465 AU.addPreserved<LiveRegMatrix>();
466 AU.addRequired<EdgeBundles>();
467 AU.addRequired<SpillPlacement>();
468 MachineFunctionPass::getAnalysisUsage(AU);
472 //===----------------------------------------------------------------------===//
473 // LiveRangeEdit delegate methods
474 //===----------------------------------------------------------------------===//
476 bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
477 if (VRM->hasPhys(VirtReg)) {
478 LiveInterval &LI = LIS->getInterval(VirtReg);
479 Matrix->unassign(LI);
480 aboutToRemoveInterval(LI);
483 // Unassigned virtreg is probably in the priority queue.
484 // RegAllocBase will erase it after dequeueing.
488 void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
489 if (!VRM->hasPhys(VirtReg))
492 // Register is assigned, put it back on the queue for reassignment.
493 LiveInterval &LI = LIS->getInterval(VirtReg);
494 Matrix->unassign(LI);
498 void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
499 // Cloning a register we haven't even heard about yet? Just ignore it.
500 if (!ExtraRegInfo.inBounds(Old))
503 // LRE may clone a virtual register because dead code elimination causes it to
504 // be split into connected components. The new components are much smaller
505 // than the original, so they should get a new chance at being assigned.
506 // same stage as the parent.
507 ExtraRegInfo[Old].Stage = RS_Assign;
508 ExtraRegInfo.grow(New);
509 ExtraRegInfo[New] = ExtraRegInfo[Old];
512 void RAGreedy::releaseMemory() {
513 SpillerInstance.reset();
514 ExtraRegInfo.clear();
518 void RAGreedy::enqueue(LiveInterval *LI) { enqueue(Queue, LI); }
520 void RAGreedy::enqueue(PQueue &CurQueue, LiveInterval *LI) {
521 // Prioritize live ranges by size, assigning larger ranges first.
522 // The queue holds (size, reg) pairs.
523 const unsigned Size = LI->getSize();
524 const unsigned Reg = LI->reg;
525 assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
526 "Can only enqueue virtual registers");
529 ExtraRegInfo.grow(Reg);
530 if (ExtraRegInfo[Reg].Stage == RS_New)
531 ExtraRegInfo[Reg].Stage = RS_Assign;
533 if (ExtraRegInfo[Reg].Stage == RS_Split) {
534 // Unsplit ranges that couldn't be allocated immediately are deferred until
535 // everything else has been allocated.
538 // Giant live ranges fall back to the global assignment heuristic, which
539 // prevents excessive spilling in pathological cases.
540 bool ReverseLocal = TRI->reverseLocalAssignment();
541 bool ForceGlobal = !ReverseLocal &&
542 (Size / SlotIndex::InstrDist) > (2 * MRI->getRegClass(Reg)->getNumRegs());
544 if (ExtraRegInfo[Reg].Stage == RS_Assign && !ForceGlobal && !LI->empty() &&
545 LIS->intervalIsInOneMBB(*LI)) {
546 // Allocate original local ranges in linear instruction order. Since they
547 // are singly defined, this produces optimal coloring in the absence of
548 // global interference and other constraints.
550 Prio = LI->beginIndex().getInstrDistance(Indexes->getLastIndex());
552 // Allocating bottom up may allow many short LRGs to be assigned first
553 // to one of the cheap registers. This could be much faster for very
554 // large blocks on targets with many physical registers.
555 Prio = Indexes->getZeroIndex().getInstrDistance(LI->beginIndex());
559 // Allocate global and split ranges in long->short order. Long ranges that
560 // don't fit should be spilled (or split) ASAP so they don't create
561 // interference. Mark a bit to prioritize global above local ranges.
562 Prio = (1u << 29) + Size;
564 // Mark a higher bit to prioritize global and local above RS_Split.
567 // Boost ranges that have a physical register hint.
568 if (VRM->hasKnownPreference(Reg))
571 // The virtual register number is a tie breaker for same-sized ranges.
572 // Give lower vreg numbers higher priority to assign them first.
573 CurQueue.push(std::make_pair(Prio, ~Reg));
576 LiveInterval *RAGreedy::dequeue() { return dequeue(Queue); }
578 LiveInterval *RAGreedy::dequeue(PQueue &CurQueue) {
579 if (CurQueue.empty())
581 LiveInterval *LI = &LIS->getInterval(~CurQueue.top().second);
587 //===----------------------------------------------------------------------===//
589 //===----------------------------------------------------------------------===//
591 /// tryAssign - Try to assign VirtReg to an available register.
592 unsigned RAGreedy::tryAssign(LiveInterval &VirtReg,
593 AllocationOrder &Order,
594 SmallVectorImpl<unsigned> &NewVRegs) {
597 while ((PhysReg = Order.next()))
598 if (!Matrix->checkInterference(VirtReg, PhysReg))
600 if (!PhysReg || Order.isHint())
603 // PhysReg is available, but there may be a better choice.
605 // If we missed a simple hint, try to cheaply evict interference from the
606 // preferred register.
607 if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg))
608 if (Order.isHint(Hint)) {
609 DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n');
610 EvictionCost MaxCost;
611 MaxCost.setBrokenHints(1);
612 if (canEvictInterference(VirtReg, Hint, true, MaxCost)) {
613 evictInterference(VirtReg, Hint, NewVRegs);
618 // Try to evict interference from a cheaper alternative.
619 unsigned Cost = TRI->getCostPerUse(PhysReg);
621 // Most registers have 0 additional cost.
625 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " is available at cost " << Cost
627 unsigned CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost);
628 return CheapReg ? CheapReg : PhysReg;
632 //===----------------------------------------------------------------------===//
633 // Interference eviction
634 //===----------------------------------------------------------------------===//
636 unsigned RAGreedy::canReassign(LiveInterval &VirtReg, unsigned PrevReg) {
637 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
639 while ((PhysReg = Order.next())) {
640 if (PhysReg == PrevReg)
643 MCRegUnitIterator Units(PhysReg, TRI);
644 for (; Units.isValid(); ++Units) {
645 // Instantiate a "subquery", not to be confused with the Queries array.
646 LiveIntervalUnion::Query subQ(&VirtReg, &Matrix->getLiveUnions()[*Units]);
647 if (subQ.checkInterference())
650 // If no units have interference, break out with the current PhysReg.
651 if (!Units.isValid())
655 DEBUG(dbgs() << "can reassign: " << VirtReg << " from "
656 << PrintReg(PrevReg, TRI) << " to " << PrintReg(PhysReg, TRI)
661 /// shouldEvict - determine if A should evict the assigned live range B. The
662 /// eviction policy defined by this function together with the allocation order
663 /// defined by enqueue() decides which registers ultimately end up being split
666 /// Cascade numbers are used to prevent infinite loops if this function is a
669 /// @param A The live range to be assigned.
670 /// @param IsHint True when A is about to be assigned to its preferred
672 /// @param B The live range to be evicted.
673 /// @param BreaksHint True when B is already assigned to its preferred register.
674 bool RAGreedy::shouldEvict(LiveInterval &A, bool IsHint,
675 LiveInterval &B, bool BreaksHint) {
676 bool CanSplit = getStage(B) < RS_Spill;
678 // Be fairly aggressive about following hints as long as the evictee can be
680 if (CanSplit && IsHint && !BreaksHint)
683 if (A.weight > B.weight) {
684 DEBUG(dbgs() << "should evict: " << B << " w= " << B.weight << '\n');
690 /// canEvictInterference - Return true if all interferences between VirtReg and
691 /// PhysReg can be evicted.
693 /// @param VirtReg Live range that is about to be assigned.
694 /// @param PhysReg Desired register for assignment.
695 /// @param IsHint True when PhysReg is VirtReg's preferred register.
696 /// @param MaxCost Only look for cheaper candidates and update with new cost
697 /// when returning true.
698 /// @returns True when interference can be evicted cheaper than MaxCost.
699 bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
700 bool IsHint, EvictionCost &MaxCost) {
701 // It is only possible to evict virtual register interference.
702 if (Matrix->checkInterference(VirtReg, PhysReg) > LiveRegMatrix::IK_VirtReg)
705 bool IsLocal = LIS->intervalIsInOneMBB(VirtReg);
707 // Find VirtReg's cascade number. This will be unassigned if VirtReg was never
708 // involved in an eviction before. If a cascade number was assigned, deny
709 // evicting anything with the same or a newer cascade number. This prevents
710 // infinite eviction loops.
712 // This works out so a register without a cascade number is allowed to evict
713 // anything, and it can be evicted by anything.
714 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
716 Cascade = NextCascade;
719 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
720 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
721 // If there is 10 or more interferences, chances are one is heavier.
722 if (Q.collectInterferingVRegs(10) >= 10)
725 // Check if any interfering live range is heavier than MaxWeight.
726 for (unsigned i = Q.interferingVRegs().size(); i; --i) {
727 LiveInterval *Intf = Q.interferingVRegs()[i - 1];
728 assert(TargetRegisterInfo::isVirtualRegister(Intf->reg) &&
729 "Only expecting virtual register interference from query");
730 // Never evict spill products. They cannot split or spill.
731 if (getStage(*Intf) == RS_Done)
733 // Once a live range becomes small enough, it is urgent that we find a
734 // register for it. This is indicated by an infinite spill weight. These
735 // urgent live ranges get to evict almost anything.
737 // Also allow urgent evictions of unspillable ranges from a strictly
738 // larger allocation order.
739 bool Urgent = !VirtReg.isSpillable() &&
740 (Intf->isSpillable() ||
741 RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(VirtReg.reg)) <
742 RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(Intf->reg)));
743 // Only evict older cascades or live ranges without a cascade.
744 unsigned IntfCascade = ExtraRegInfo[Intf->reg].Cascade;
745 if (Cascade <= IntfCascade) {
748 // We permit breaking cascades for urgent evictions. It should be the
749 // last resort, though, so make it really expensive.
750 Cost.BrokenHints += 10;
752 // Would this break a satisfied hint?
753 bool BreaksHint = VRM->hasPreferredPhys(Intf->reg);
754 // Update eviction cost.
755 Cost.BrokenHints += BreaksHint;
756 Cost.MaxWeight = std::max(Cost.MaxWeight, Intf->weight);
757 // Abort if this would be too expensive.
758 if (!(Cost < MaxCost))
762 // Apply the eviction policy for non-urgent evictions.
763 if (!shouldEvict(VirtReg, IsHint, *Intf, BreaksHint))
765 // If !MaxCost.isMax(), then we're just looking for a cheap register.
766 // Evicting another local live range in this case could lead to suboptimal
768 if (!MaxCost.isMax() && IsLocal && LIS->intervalIsInOneMBB(*Intf) &&
769 (!EnableLocalReassign || !canReassign(*Intf, PhysReg))) {
778 /// evictInterference - Evict any interferring registers that prevent VirtReg
779 /// from being assigned to Physreg. This assumes that canEvictInterference
781 void RAGreedy::evictInterference(LiveInterval &VirtReg, unsigned PhysReg,
782 SmallVectorImpl<unsigned> &NewVRegs) {
783 // Make sure that VirtReg has a cascade number, and assign that cascade
784 // number to every evicted register. These live ranges than then only be
785 // evicted by a newer cascade, preventing infinite loops.
786 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
788 Cascade = ExtraRegInfo[VirtReg.reg].Cascade = NextCascade++;
790 DEBUG(dbgs() << "evicting " << PrintReg(PhysReg, TRI)
791 << " interference: Cascade " << Cascade << '\n');
793 // Collect all interfering virtregs first.
794 SmallVector<LiveInterval*, 8> Intfs;
795 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
796 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
797 assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
798 ArrayRef<LiveInterval*> IVR = Q.interferingVRegs();
799 Intfs.append(IVR.begin(), IVR.end());
802 // Evict them second. This will invalidate the queries.
803 for (unsigned i = 0, e = Intfs.size(); i != e; ++i) {
804 LiveInterval *Intf = Intfs[i];
805 // The same VirtReg may be present in multiple RegUnits. Skip duplicates.
806 if (!VRM->hasPhys(Intf->reg))
808 Matrix->unassign(*Intf);
809 assert((ExtraRegInfo[Intf->reg].Cascade < Cascade ||
810 VirtReg.isSpillable() < Intf->isSpillable()) &&
811 "Cannot decrease cascade number, illegal eviction");
812 ExtraRegInfo[Intf->reg].Cascade = Cascade;
814 NewVRegs.push_back(Intf->reg);
818 /// tryEvict - Try to evict all interferences for a physreg.
819 /// @param VirtReg Currently unassigned virtual register.
820 /// @param Order Physregs to try.
821 /// @return Physreg to assign VirtReg, or 0.
822 unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
823 AllocationOrder &Order,
824 SmallVectorImpl<unsigned> &NewVRegs,
825 unsigned CostPerUseLimit) {
826 NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
828 // Keep track of the cheapest interference seen so far.
829 EvictionCost BestCost;
831 unsigned BestPhys = 0;
832 unsigned OrderLimit = Order.getOrder().size();
834 // When we are just looking for a reduced cost per use, don't break any
835 // hints, and only evict smaller spill weights.
836 if (CostPerUseLimit < ~0u) {
837 BestCost.BrokenHints = 0;
838 BestCost.MaxWeight = VirtReg.weight;
840 // Check of any registers in RC are below CostPerUseLimit.
841 const TargetRegisterClass *RC = MRI->getRegClass(VirtReg.reg);
842 unsigned MinCost = RegClassInfo.getMinCost(RC);
843 if (MinCost >= CostPerUseLimit) {
844 DEBUG(dbgs() << TRI->getRegClassName(RC) << " minimum cost = " << MinCost
845 << ", no cheaper registers to be found.\n");
849 // It is normal for register classes to have a long tail of registers with
850 // the same cost. We don't need to look at them if they're too expensive.
851 if (TRI->getCostPerUse(Order.getOrder().back()) >= CostPerUseLimit) {
852 OrderLimit = RegClassInfo.getLastCostChange(RC);
853 DEBUG(dbgs() << "Only trying the first " << OrderLimit << " regs.\n");
858 while (unsigned PhysReg = Order.next(OrderLimit)) {
859 if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
861 // The first use of a callee-saved register in a function has cost 1.
862 // Don't start using a CSR when the CostPerUseLimit is low.
863 if (CostPerUseLimit == 1)
864 if (unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg))
865 if (!MRI->isPhysRegUsed(CSR)) {
866 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " would clobber CSR "
867 << PrintReg(CSR, TRI) << '\n');
871 if (!canEvictInterference(VirtReg, PhysReg, false, BestCost))
877 // Stop if the hint can be used.
885 evictInterference(VirtReg, BestPhys, NewVRegs);
890 //===----------------------------------------------------------------------===//
892 //===----------------------------------------------------------------------===//
894 /// addSplitConstraints - Fill out the SplitConstraints vector based on the
895 /// interference pattern in Physreg and its aliases. Add the constraints to
896 /// SpillPlacement and return the static cost of this split in Cost, assuming
897 /// that all preferences in SplitConstraints are met.
898 /// Return false if there are no bundles with positive bias.
899 bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf,
900 BlockFrequency &Cost) {
901 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
903 // Reset interference dependent info.
904 SplitConstraints.resize(UseBlocks.size());
905 BlockFrequency StaticCost = 0;
906 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
907 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
908 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
910 BC.Number = BI.MBB->getNumber();
911 Intf.moveToBlock(BC.Number);
912 BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
913 BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
914 BC.ChangesValue = BI.FirstDef.isValid();
916 if (!Intf.hasInterference())
919 // Number of spill code instructions to insert.
922 // Interference for the live-in value.
924 if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
925 BC.Entry = SpillPlacement::MustSpill, ++Ins;
926 else if (Intf.first() < BI.FirstInstr)
927 BC.Entry = SpillPlacement::PrefSpill, ++Ins;
928 else if (Intf.first() < BI.LastInstr)
932 // Interference for the live-out value.
934 if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
935 BC.Exit = SpillPlacement::MustSpill, ++Ins;
936 else if (Intf.last() > BI.LastInstr)
937 BC.Exit = SpillPlacement::PrefSpill, ++Ins;
938 else if (Intf.last() > BI.FirstInstr)
942 // Accumulate the total frequency of inserted spill code.
944 StaticCost += SpillPlacer->getBlockFrequency(BC.Number);
948 // Add constraints for use-blocks. Note that these are the only constraints
949 // that may add a positive bias, it is downhill from here.
950 SpillPlacer->addConstraints(SplitConstraints);
951 return SpillPlacer->scanActiveBundles();
955 /// addThroughConstraints - Add constraints and links to SpillPlacer from the
956 /// live-through blocks in Blocks.
957 void RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf,
958 ArrayRef<unsigned> Blocks) {
959 const unsigned GroupSize = 8;
960 SpillPlacement::BlockConstraint BCS[GroupSize];
961 unsigned TBS[GroupSize];
962 unsigned B = 0, T = 0;
964 for (unsigned i = 0; i != Blocks.size(); ++i) {
965 unsigned Number = Blocks[i];
966 Intf.moveToBlock(Number);
968 if (!Intf.hasInterference()) {
969 assert(T < GroupSize && "Array overflow");
971 if (++T == GroupSize) {
972 SpillPlacer->addLinks(makeArrayRef(TBS, T));
978 assert(B < GroupSize && "Array overflow");
979 BCS[B].Number = Number;
981 // Interference for the live-in value.
982 if (Intf.first() <= Indexes->getMBBStartIdx(Number))
983 BCS[B].Entry = SpillPlacement::MustSpill;
985 BCS[B].Entry = SpillPlacement::PrefSpill;
987 // Interference for the live-out value.
988 if (Intf.last() >= SA->getLastSplitPoint(Number))
989 BCS[B].Exit = SpillPlacement::MustSpill;
991 BCS[B].Exit = SpillPlacement::PrefSpill;
993 if (++B == GroupSize) {
994 SpillPlacer->addConstraints(makeArrayRef(BCS, B));
999 SpillPlacer->addConstraints(makeArrayRef(BCS, B));
1000 SpillPlacer->addLinks(makeArrayRef(TBS, T));
1003 void RAGreedy::growRegion(GlobalSplitCandidate &Cand) {
1004 // Keep track of through blocks that have not been added to SpillPlacer.
1005 BitVector Todo = SA->getThroughBlocks();
1006 SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks;
1007 unsigned AddedTo = 0;
1009 unsigned Visited = 0;
1013 ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive();
1014 // Find new through blocks in the periphery of PrefRegBundles.
1015 for (int i = 0, e = NewBundles.size(); i != e; ++i) {
1016 unsigned Bundle = NewBundles[i];
1017 // Look at all blocks connected to Bundle in the full graph.
1018 ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle);
1019 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
1021 unsigned Block = *I;
1022 if (!Todo.test(Block))
1025 // This is a new through block. Add it to SpillPlacer later.
1026 ActiveBlocks.push_back(Block);
1032 // Any new blocks to add?
1033 if (ActiveBlocks.size() == AddedTo)
1036 // Compute through constraints from the interference, or assume that all
1037 // through blocks prefer spilling when forming compact regions.
1038 auto NewBlocks = makeArrayRef(ActiveBlocks).slice(AddedTo);
1040 addThroughConstraints(Cand.Intf, NewBlocks);
1042 // Provide a strong negative bias on through blocks to prevent unwanted
1043 // liveness on loop backedges.
1044 SpillPlacer->addPrefSpill(NewBlocks, /* Strong= */ true);
1045 AddedTo = ActiveBlocks.size();
1047 // Perhaps iterating can enable more bundles?
1048 SpillPlacer->iterate();
1050 DEBUG(dbgs() << ", v=" << Visited);
1053 /// calcCompactRegion - Compute the set of edge bundles that should be live
1054 /// when splitting the current live range into compact regions. Compact
1055 /// regions can be computed without looking at interference. They are the
1056 /// regions formed by removing all the live-through blocks from the live range.
1058 /// Returns false if the current live range is already compact, or if the
1059 /// compact regions would form single block regions anyway.
1060 bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) {
1061 // Without any through blocks, the live range is already compact.
1062 if (!SA->getNumThroughBlocks())
1065 // Compact regions don't correspond to any physreg.
1066 Cand.reset(IntfCache, 0);
1068 DEBUG(dbgs() << "Compact region bundles");
1070 // Use the spill placer to determine the live bundles. GrowRegion pretends
1071 // that all the through blocks have interference when PhysReg is unset.
1072 SpillPlacer->prepare(Cand.LiveBundles);
1074 // The static split cost will be zero since Cand.Intf reports no interference.
1075 BlockFrequency Cost;
1076 if (!addSplitConstraints(Cand.Intf, Cost)) {
1077 DEBUG(dbgs() << ", none.\n");
1082 SpillPlacer->finish();
1084 if (!Cand.LiveBundles.any()) {
1085 DEBUG(dbgs() << ", none.\n");
1090 for (int i = Cand.LiveBundles.find_first(); i>=0;
1091 i = Cand.LiveBundles.find_next(i))
1092 dbgs() << " EB#" << i;
1098 /// calcSpillCost - Compute how expensive it would be to split the live range in
1099 /// SA around all use blocks instead of forming bundle regions.
1100 BlockFrequency RAGreedy::calcSpillCost() {
1101 BlockFrequency Cost = 0;
1102 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
1103 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
1104 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
1105 unsigned Number = BI.MBB->getNumber();
1106 // We normally only need one spill instruction - a load or a store.
1107 Cost += SpillPlacer->getBlockFrequency(Number);
1109 // Unless the value is redefined in the block.
1110 if (BI.LiveIn && BI.LiveOut && BI.FirstDef)
1111 Cost += SpillPlacer->getBlockFrequency(Number);
1116 /// calcGlobalSplitCost - Return the global split cost of following the split
1117 /// pattern in LiveBundles. This cost should be added to the local cost of the
1118 /// interference pattern in SplitConstraints.
1120 BlockFrequency RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) {
1121 BlockFrequency GlobalCost = 0;
1122 const BitVector &LiveBundles = Cand.LiveBundles;
1123 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
1124 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
1125 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
1126 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
1127 bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)];
1128 bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
1132 Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
1134 Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
1136 GlobalCost += SpillPlacer->getBlockFrequency(BC.Number);
1139 for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
1140 unsigned Number = Cand.ActiveBlocks[i];
1141 bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)];
1142 bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)];
1143 if (!RegIn && !RegOut)
1145 if (RegIn && RegOut) {
1146 // We need double spill code if this block has interference.
1147 Cand.Intf.moveToBlock(Number);
1148 if (Cand.Intf.hasInterference()) {
1149 GlobalCost += SpillPlacer->getBlockFrequency(Number);
1150 GlobalCost += SpillPlacer->getBlockFrequency(Number);
1154 // live-in / stack-out or stack-in live-out.
1155 GlobalCost += SpillPlacer->getBlockFrequency(Number);
1160 /// splitAroundRegion - Split the current live range around the regions
1161 /// determined by BundleCand and GlobalCand.
1163 /// Before calling this function, GlobalCand and BundleCand must be initialized
1164 /// so each bundle is assigned to a valid candidate, or NoCand for the
1165 /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor
1166 /// objects must be initialized for the current live range, and intervals
1167 /// created for the used candidates.
1169 /// @param LREdit The LiveRangeEdit object handling the current split.
1170 /// @param UsedCands List of used GlobalCand entries. Every BundleCand value
1171 /// must appear in this list.
1172 void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit,
1173 ArrayRef<unsigned> UsedCands) {
1174 // These are the intervals created for new global ranges. We may create more
1175 // intervals for local ranges.
1176 const unsigned NumGlobalIntvs = LREdit.size();
1177 DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs << " globals.\n");
1178 assert(NumGlobalIntvs && "No global intervals configured");
1180 // Isolate even single instructions when dealing with a proper sub-class.
1181 // That guarantees register class inflation for the stack interval because it
1183 unsigned Reg = SA->getParent().reg;
1184 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
1186 // First handle all the blocks with uses.
1187 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
1188 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
1189 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
1190 unsigned Number = BI.MBB->getNumber();
1191 unsigned IntvIn = 0, IntvOut = 0;
1192 SlotIndex IntfIn, IntfOut;
1194 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
1195 if (CandIn != NoCand) {
1196 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
1197 IntvIn = Cand.IntvIdx;
1198 Cand.Intf.moveToBlock(Number);
1199 IntfIn = Cand.Intf.first();
1203 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
1204 if (CandOut != NoCand) {
1205 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
1206 IntvOut = Cand.IntvIdx;
1207 Cand.Intf.moveToBlock(Number);
1208 IntfOut = Cand.Intf.last();
1212 // Create separate intervals for isolated blocks with multiple uses.
1213 if (!IntvIn && !IntvOut) {
1214 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
1215 if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
1216 SE->splitSingleBlock(BI);
1220 if (IntvIn && IntvOut)
1221 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
1223 SE->splitRegInBlock(BI, IntvIn, IntfIn);
1225 SE->splitRegOutBlock(BI, IntvOut, IntfOut);
1228 // Handle live-through blocks. The relevant live-through blocks are stored in
1229 // the ActiveBlocks list with each candidate. We need to filter out
1231 BitVector Todo = SA->getThroughBlocks();
1232 for (unsigned c = 0; c != UsedCands.size(); ++c) {
1233 ArrayRef<unsigned> Blocks = GlobalCand[UsedCands[c]].ActiveBlocks;
1234 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1235 unsigned Number = Blocks[i];
1236 if (!Todo.test(Number))
1240 unsigned IntvIn = 0, IntvOut = 0;
1241 SlotIndex IntfIn, IntfOut;
1243 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
1244 if (CandIn != NoCand) {
1245 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
1246 IntvIn = Cand.IntvIdx;
1247 Cand.Intf.moveToBlock(Number);
1248 IntfIn = Cand.Intf.first();
1251 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
1252 if (CandOut != NoCand) {
1253 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
1254 IntvOut = Cand.IntvIdx;
1255 Cand.Intf.moveToBlock(Number);
1256 IntfOut = Cand.Intf.last();
1258 if (!IntvIn && !IntvOut)
1260 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
1266 SmallVector<unsigned, 8> IntvMap;
1267 SE->finish(&IntvMap);
1268 DebugVars->splitRegister(Reg, LREdit.regs(), *LIS);
1270 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1271 unsigned OrigBlocks = SA->getNumLiveBlocks();
1273 // Sort out the new intervals created by splitting. We get four kinds:
1274 // - Remainder intervals should not be split again.
1275 // - Candidate intervals can be assigned to Cand.PhysReg.
1276 // - Block-local splits are candidates for local splitting.
1277 // - DCE leftovers should go back on the queue.
1278 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1279 LiveInterval &Reg = LIS->getInterval(LREdit.get(i));
1281 // Ignore old intervals from DCE.
1282 if (getStage(Reg) != RS_New)
1285 // Remainder interval. Don't try splitting again, spill if it doesn't
1287 if (IntvMap[i] == 0) {
1288 setStage(Reg, RS_Spill);
1292 // Global intervals. Allow repeated splitting as long as the number of live
1293 // blocks is strictly decreasing.
1294 if (IntvMap[i] < NumGlobalIntvs) {
1295 if (SA->countLiveBlocks(&Reg) >= OrigBlocks) {
1296 DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks
1297 << " blocks as original.\n");
1298 // Don't allow repeated splitting as a safe guard against looping.
1299 setStage(Reg, RS_Split2);
1304 // Other intervals are treated as new. This includes local intervals created
1305 // for blocks with multiple uses, and anything created by DCE.
1309 MF->verify(this, "After splitting live range around region");
1312 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1313 SmallVectorImpl<unsigned> &NewVRegs) {
1314 unsigned NumCands = 0;
1315 BlockFrequency BestCost;
1317 // Check if we can split this live range around a compact region.
1318 bool HasCompact = calcCompactRegion(GlobalCand.front());
1320 // Yes, keep GlobalCand[0] as the compact region candidate.
1322 BestCost = BlockFrequency::getMaxFrequency();
1324 // No benefit from the compact region, our fallback will be per-block
1325 // splitting. Make sure we find a solution that is cheaper than spilling.
1326 BestCost = calcSpillCost();
1327 DEBUG(dbgs() << "Cost of isolating all blocks = ";
1328 MBFI->printBlockFreq(dbgs(), BestCost) << '\n');
1332 calculateRegionSplitCost(VirtReg, Order, BestCost, NumCands,
1333 false/*IgnoreCSR*/);
1335 // No solutions found, fall back to single block splitting.
1336 if (!HasCompact && BestCand == NoCand)
1339 return doRegionSplit(VirtReg, BestCand, HasCompact, NewVRegs);
1342 unsigned RAGreedy::calculateRegionSplitCost(LiveInterval &VirtReg,
1343 AllocationOrder &Order,
1344 BlockFrequency &BestCost,
1347 unsigned BestCand = NoCand;
1349 while (unsigned PhysReg = Order.next()) {
1350 if (unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg))
1351 if (IgnoreCSR && !MRI->isPhysRegUsed(CSR))
1354 // Discard bad candidates before we run out of interference cache cursors.
1355 // This will only affect register classes with a lot of registers (>32).
1356 if (NumCands == IntfCache.getMaxCursors()) {
1357 unsigned WorstCount = ~0u;
1359 for (unsigned i = 0; i != NumCands; ++i) {
1360 if (i == BestCand || !GlobalCand[i].PhysReg)
1362 unsigned Count = GlobalCand[i].LiveBundles.count();
1363 if (Count < WorstCount)
1364 Worst = i, WorstCount = Count;
1367 GlobalCand[Worst] = GlobalCand[NumCands];
1368 if (BestCand == NumCands)
1372 if (GlobalCand.size() <= NumCands)
1373 GlobalCand.resize(NumCands+1);
1374 GlobalSplitCandidate &Cand = GlobalCand[NumCands];
1375 Cand.reset(IntfCache, PhysReg);
1377 SpillPlacer->prepare(Cand.LiveBundles);
1378 BlockFrequency Cost;
1379 if (!addSplitConstraints(Cand.Intf, Cost)) {
1380 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n");
1383 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = ";
1384 MBFI->printBlockFreq(dbgs(), Cost));
1385 if (Cost >= BestCost) {
1387 if (BestCand == NoCand)
1388 dbgs() << " worse than no bundles\n";
1390 dbgs() << " worse than "
1391 << PrintReg(GlobalCand[BestCand].PhysReg, TRI) << '\n';
1397 SpillPlacer->finish();
1399 // No live bundles, defer to splitSingleBlocks().
1400 if (!Cand.LiveBundles.any()) {
1401 DEBUG(dbgs() << " no bundles.\n");
1405 Cost += calcGlobalSplitCost(Cand);
1407 dbgs() << ", total = "; MBFI->printBlockFreq(dbgs(), Cost)
1409 for (int i = Cand.LiveBundles.find_first(); i>=0;
1410 i = Cand.LiveBundles.find_next(i))
1411 dbgs() << " EB#" << i;
1414 if (Cost < BestCost) {
1415 BestCand = NumCands;
1423 unsigned RAGreedy::doRegionSplit(LiveInterval &VirtReg, unsigned BestCand,
1425 SmallVectorImpl<unsigned> &NewVRegs) {
1426 SmallVector<unsigned, 8> UsedCands;
1427 // Prepare split editor.
1428 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1429 SE->reset(LREdit, SplitSpillMode);
1431 // Assign all edge bundles to the preferred candidate, or NoCand.
1432 BundleCand.assign(Bundles->getNumBundles(), NoCand);
1434 // Assign bundles for the best candidate region.
1435 if (BestCand != NoCand) {
1436 GlobalSplitCandidate &Cand = GlobalCand[BestCand];
1437 if (unsigned B = Cand.getBundles(BundleCand, BestCand)) {
1438 UsedCands.push_back(BestCand);
1439 Cand.IntvIdx = SE->openIntv();
1440 DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in "
1441 << B << " bundles, intv " << Cand.IntvIdx << ".\n");
1446 // Assign bundles for the compact region.
1448 GlobalSplitCandidate &Cand = GlobalCand.front();
1449 assert(!Cand.PhysReg && "Compact region has no physreg");
1450 if (unsigned B = Cand.getBundles(BundleCand, 0)) {
1451 UsedCands.push_back(0);
1452 Cand.IntvIdx = SE->openIntv();
1453 DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv "
1454 << Cand.IntvIdx << ".\n");
1459 splitAroundRegion(LREdit, UsedCands);
1464 //===----------------------------------------------------------------------===//
1465 // Per-Block Splitting
1466 //===----------------------------------------------------------------------===//
1468 /// tryBlockSplit - Split a global live range around every block with uses. This
1469 /// creates a lot of local live ranges, that will be split by tryLocalSplit if
1470 /// they don't allocate.
1471 unsigned RAGreedy::tryBlockSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1472 SmallVectorImpl<unsigned> &NewVRegs) {
1473 assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed");
1474 unsigned Reg = VirtReg.reg;
1475 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
1476 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1477 SE->reset(LREdit, SplitSpillMode);
1478 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
1479 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
1480 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
1481 if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
1482 SE->splitSingleBlock(BI);
1484 // No blocks were split.
1488 // We did split for some blocks.
1489 SmallVector<unsigned, 8> IntvMap;
1490 SE->finish(&IntvMap);
1492 // Tell LiveDebugVariables about the new ranges.
1493 DebugVars->splitRegister(Reg, LREdit.regs(), *LIS);
1495 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1497 // Sort out the new intervals created by splitting. The remainder interval
1498 // goes straight to spilling, the new local ranges get to stay RS_New.
1499 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1500 LiveInterval &LI = LIS->getInterval(LREdit.get(i));
1501 if (getStage(LI) == RS_New && IntvMap[i] == 0)
1502 setStage(LI, RS_Spill);
1506 MF->verify(this, "After splitting live range around basic blocks");
1511 //===----------------------------------------------------------------------===//
1512 // Per-Instruction Splitting
1513 //===----------------------------------------------------------------------===//
1515 /// Get the number of allocatable registers that match the constraints of \p Reg
1516 /// on \p MI and that are also in \p SuperRC.
1517 static unsigned getNumAllocatableRegsForConstraints(
1518 const MachineInstr *MI, unsigned Reg, const TargetRegisterClass *SuperRC,
1519 const TargetInstrInfo *TII, const TargetRegisterInfo *TRI,
1520 const RegisterClassInfo &RCI) {
1521 assert(SuperRC && "Invalid register class");
1523 const TargetRegisterClass *ConstrainedRC =
1524 MI->getRegClassConstraintEffectForVReg(Reg, SuperRC, TII, TRI,
1525 /* ExploreBundle */ true);
1528 return RCI.getNumAllocatableRegs(ConstrainedRC);
1531 /// tryInstructionSplit - Split a live range around individual instructions.
1532 /// This is normally not worthwhile since the spiller is doing essentially the
1533 /// same thing. However, when the live range is in a constrained register
1534 /// class, it may help to insert copies such that parts of the live range can
1535 /// be moved to a larger register class.
1537 /// This is similar to spilling to a larger register class.
1539 RAGreedy::tryInstructionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1540 SmallVectorImpl<unsigned> &NewVRegs) {
1541 const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg);
1542 // There is no point to this if there are no larger sub-classes.
1543 if (!RegClassInfo.isProperSubClass(CurRC))
1546 // Always enable split spill mode, since we're effectively spilling to a
1548 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1549 SE->reset(LREdit, SplitEditor::SM_Size);
1551 ArrayRef<SlotIndex> Uses = SA->getUseSlots();
1552 if (Uses.size() <= 1)
1555 DEBUG(dbgs() << "Split around " << Uses.size() << " individual instrs.\n");
1557 const TargetRegisterClass *SuperRC =
1558 TRI->getLargestLegalSuperClass(CurRC, *MF);
1559 unsigned SuperRCNumAllocatableRegs = RCI.getNumAllocatableRegs(SuperRC);
1560 // Split around every non-copy instruction if this split will relax
1561 // the constraints on the virtual register.
1562 // Otherwise, splitting just inserts uncoalescable copies that do not help
1564 for (unsigned i = 0; i != Uses.size(); ++i) {
1565 if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]))
1566 if (MI->isFullCopy() ||
1567 SuperRCNumAllocatableRegs ==
1568 getNumAllocatableRegsForConstraints(MI, VirtReg.reg, SuperRC, TII,
1570 DEBUG(dbgs() << " skip:\t" << Uses[i] << '\t' << *MI);
1574 SlotIndex SegStart = SE->enterIntvBefore(Uses[i]);
1575 SlotIndex SegStop = SE->leaveIntvAfter(Uses[i]);
1576 SE->useIntv(SegStart, SegStop);
1579 if (LREdit.empty()) {
1580 DEBUG(dbgs() << "All uses were copies.\n");
1584 SmallVector<unsigned, 8> IntvMap;
1585 SE->finish(&IntvMap);
1586 DebugVars->splitRegister(VirtReg.reg, LREdit.regs(), *LIS);
1587 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1589 // Assign all new registers to RS_Spill. This was the last chance.
1590 setStage(LREdit.begin(), LREdit.end(), RS_Spill);
1595 //===----------------------------------------------------------------------===//
1597 //===----------------------------------------------------------------------===//
1600 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted
1601 /// in order to use PhysReg between two entries in SA->UseSlots.
1603 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
1605 void RAGreedy::calcGapWeights(unsigned PhysReg,
1606 SmallVectorImpl<float> &GapWeight) {
1607 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1608 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1609 ArrayRef<SlotIndex> Uses = SA->getUseSlots();
1610 const unsigned NumGaps = Uses.size()-1;
1612 // Start and end points for the interference check.
1613 SlotIndex StartIdx =
1614 BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr;
1616 BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr;
1618 GapWeight.assign(NumGaps, 0.0f);
1620 // Add interference from each overlapping register.
1621 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
1622 if (!Matrix->query(const_cast<LiveInterval&>(SA->getParent()), *Units)
1623 .checkInterference())
1626 // We know that VirtReg is a continuous interval from FirstInstr to
1627 // LastInstr, so we don't need InterferenceQuery.
1629 // Interference that overlaps an instruction is counted in both gaps
1630 // surrounding the instruction. The exception is interference before
1631 // StartIdx and after StopIdx.
1633 LiveIntervalUnion::SegmentIter IntI =
1634 Matrix->getLiveUnions()[*Units] .find(StartIdx);
1635 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
1636 // Skip the gaps before IntI.
1637 while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
1638 if (++Gap == NumGaps)
1643 // Update the gaps covered by IntI.
1644 const float weight = IntI.value()->weight;
1645 for (; Gap != NumGaps; ++Gap) {
1646 GapWeight[Gap] = std::max(GapWeight[Gap], weight);
1647 if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
1655 // Add fixed interference.
1656 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
1657 const LiveRange &LR = LIS->getRegUnit(*Units);
1658 LiveRange::const_iterator I = LR.find(StartIdx);
1659 LiveRange::const_iterator E = LR.end();
1661 // Same loop as above. Mark any overlapped gaps as HUGE_VALF.
1662 for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) {
1663 while (Uses[Gap+1].getBoundaryIndex() < I->start)
1664 if (++Gap == NumGaps)
1669 for (; Gap != NumGaps; ++Gap) {
1670 GapWeight[Gap] = llvm::huge_valf;
1671 if (Uses[Gap+1].getBaseIndex() >= I->end)
1680 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
1683 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1684 SmallVectorImpl<unsigned> &NewVRegs) {
1685 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1686 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1688 // Note that it is possible to have an interval that is live-in or live-out
1689 // while only covering a single block - A phi-def can use undef values from
1690 // predecessors, and the block could be a single-block loop.
1691 // We don't bother doing anything clever about such a case, we simply assume
1692 // that the interval is continuous from FirstInstr to LastInstr. We should
1693 // make sure that we don't do anything illegal to such an interval, though.
1695 ArrayRef<SlotIndex> Uses = SA->getUseSlots();
1696 if (Uses.size() <= 2)
1698 const unsigned NumGaps = Uses.size()-1;
1701 dbgs() << "tryLocalSplit: ";
1702 for (unsigned i = 0, e = Uses.size(); i != e; ++i)
1703 dbgs() << ' ' << Uses[i];
1707 // If VirtReg is live across any register mask operands, compute a list of
1708 // gaps with register masks.
1709 SmallVector<unsigned, 8> RegMaskGaps;
1710 if (Matrix->checkRegMaskInterference(VirtReg)) {
1711 // Get regmask slots for the whole block.
1712 ArrayRef<SlotIndex> RMS = LIS->getRegMaskSlotsInBlock(BI.MBB->getNumber());
1713 DEBUG(dbgs() << RMS.size() << " regmasks in block:");
1714 // Constrain to VirtReg's live range.
1715 unsigned ri = std::lower_bound(RMS.begin(), RMS.end(),
1716 Uses.front().getRegSlot()) - RMS.begin();
1717 unsigned re = RMS.size();
1718 for (unsigned i = 0; i != NumGaps && ri != re; ++i) {
1719 // Look for Uses[i] <= RMS <= Uses[i+1].
1720 assert(!SlotIndex::isEarlierInstr(RMS[ri], Uses[i]));
1721 if (SlotIndex::isEarlierInstr(Uses[i+1], RMS[ri]))
1723 // Skip a regmask on the same instruction as the last use. It doesn't
1724 // overlap the live range.
1725 if (SlotIndex::isSameInstr(Uses[i+1], RMS[ri]) && i+1 == NumGaps)
1727 DEBUG(dbgs() << ' ' << RMS[ri] << ':' << Uses[i] << '-' << Uses[i+1]);
1728 RegMaskGaps.push_back(i);
1729 // Advance ri to the next gap. A regmask on one of the uses counts in
1731 while (ri != re && SlotIndex::isEarlierInstr(RMS[ri], Uses[i+1]))
1734 DEBUG(dbgs() << '\n');
1737 // Since we allow local split results to be split again, there is a risk of
1738 // creating infinite loops. It is tempting to require that the new live
1739 // ranges have less instructions than the original. That would guarantee
1740 // convergence, but it is too strict. A live range with 3 instructions can be
1741 // split 2+3 (including the COPY), and we want to allow that.
1743 // Instead we use these rules:
1745 // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the
1746 // noop split, of course).
1747 // 2. Require progress be made for ranges with getStage() == RS_Split2. All
1748 // the new ranges must have fewer instructions than before the split.
1749 // 3. New ranges with the same number of instructions are marked RS_Split2,
1750 // smaller ranges are marked RS_New.
1752 // These rules allow a 3 -> 2+3 split once, which we need. They also prevent
1753 // excessive splitting and infinite loops.
1755 bool ProgressRequired = getStage(VirtReg) >= RS_Split2;
1757 // Best split candidate.
1758 unsigned BestBefore = NumGaps;
1759 unsigned BestAfter = 0;
1762 const float blockFreq =
1763 SpillPlacer->getBlockFrequency(BI.MBB->getNumber()).getFrequency() *
1764 (1.0f / MBFI->getEntryFreq());
1765 SmallVector<float, 8> GapWeight;
1768 while (unsigned PhysReg = Order.next()) {
1769 // Keep track of the largest spill weight that would need to be evicted in
1770 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
1771 calcGapWeights(PhysReg, GapWeight);
1773 // Remove any gaps with regmask clobbers.
1774 if (Matrix->checkRegMaskInterference(VirtReg, PhysReg))
1775 for (unsigned i = 0, e = RegMaskGaps.size(); i != e; ++i)
1776 GapWeight[RegMaskGaps[i]] = llvm::huge_valf;
1778 // Try to find the best sequence of gaps to close.
1779 // The new spill weight must be larger than any gap interference.
1781 // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
1782 unsigned SplitBefore = 0, SplitAfter = 1;
1784 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
1785 // It is the spill weight that needs to be evicted.
1786 float MaxGap = GapWeight[0];
1789 // Live before/after split?
1790 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
1791 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
1793 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
1794 << Uses[SplitBefore] << '-' << Uses[SplitAfter]
1795 << " i=" << MaxGap);
1797 // Stop before the interval gets so big we wouldn't be making progress.
1798 if (!LiveBefore && !LiveAfter) {
1799 DEBUG(dbgs() << " all\n");
1802 // Should the interval be extended or shrunk?
1805 // How many gaps would the new range have?
1806 unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter;
1808 // Legally, without causing looping?
1809 bool Legal = !ProgressRequired || NewGaps < NumGaps;
1811 if (Legal && MaxGap < llvm::huge_valf) {
1812 // Estimate the new spill weight. Each instruction reads or writes the
1813 // register. Conservatively assume there are no read-modify-write
1816 // Try to guess the size of the new interval.
1817 const float EstWeight = normalizeSpillWeight(
1818 blockFreq * (NewGaps + 1),
1819 Uses[SplitBefore].distance(Uses[SplitAfter]) +
1820 (LiveBefore + LiveAfter) * SlotIndex::InstrDist,
1822 // Would this split be possible to allocate?
1823 // Never allocate all gaps, we wouldn't be making progress.
1824 DEBUG(dbgs() << " w=" << EstWeight);
1825 if (EstWeight * Hysteresis >= MaxGap) {
1827 float Diff = EstWeight - MaxGap;
1828 if (Diff > BestDiff) {
1829 DEBUG(dbgs() << " (best)");
1830 BestDiff = Hysteresis * Diff;
1831 BestBefore = SplitBefore;
1832 BestAfter = SplitAfter;
1839 if (++SplitBefore < SplitAfter) {
1840 DEBUG(dbgs() << " shrink\n");
1841 // Recompute the max when necessary.
1842 if (GapWeight[SplitBefore - 1] >= MaxGap) {
1843 MaxGap = GapWeight[SplitBefore];
1844 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
1845 MaxGap = std::max(MaxGap, GapWeight[i]);
1852 // Try to extend the interval.
1853 if (SplitAfter >= NumGaps) {
1854 DEBUG(dbgs() << " end\n");
1858 DEBUG(dbgs() << " extend\n");
1859 MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]);
1863 // Didn't find any candidates?
1864 if (BestBefore == NumGaps)
1867 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
1868 << '-' << Uses[BestAfter] << ", " << BestDiff
1869 << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
1871 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1875 SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
1876 SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]);
1877 SE->useIntv(SegStart, SegStop);
1878 SmallVector<unsigned, 8> IntvMap;
1879 SE->finish(&IntvMap);
1880 DebugVars->splitRegister(VirtReg.reg, LREdit.regs(), *LIS);
1882 // If the new range has the same number of instructions as before, mark it as
1883 // RS_Split2 so the next split will be forced to make progress. Otherwise,
1884 // leave the new intervals as RS_New so they can compete.
1885 bool LiveBefore = BestBefore != 0 || BI.LiveIn;
1886 bool LiveAfter = BestAfter != NumGaps || BI.LiveOut;
1887 unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter;
1888 if (NewGaps >= NumGaps) {
1889 DEBUG(dbgs() << "Tagging non-progress ranges: ");
1890 assert(!ProgressRequired && "Didn't make progress when it was required.");
1891 for (unsigned i = 0, e = IntvMap.size(); i != e; ++i)
1892 if (IntvMap[i] == 1) {
1893 setStage(LIS->getInterval(LREdit.get(i)), RS_Split2);
1894 DEBUG(dbgs() << PrintReg(LREdit.get(i)));
1896 DEBUG(dbgs() << '\n');
1903 //===----------------------------------------------------------------------===//
1904 // Live Range Splitting
1905 //===----------------------------------------------------------------------===//
1907 /// trySplit - Try to split VirtReg or one of its interferences, making it
1909 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
1910 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
1911 SmallVectorImpl<unsigned>&NewVRegs) {
1912 // Ranges must be Split2 or less.
1913 if (getStage(VirtReg) >= RS_Spill)
1916 // Local intervals are handled separately.
1917 if (LIS->intervalIsInOneMBB(VirtReg)) {
1918 NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
1919 SA->analyze(&VirtReg);
1920 unsigned PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs);
1921 if (PhysReg || !NewVRegs.empty())
1923 return tryInstructionSplit(VirtReg, Order, NewVRegs);
1926 NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
1928 SA->analyze(&VirtReg);
1930 // FIXME: SplitAnalysis may repair broken live ranges coming from the
1931 // coalescer. That may cause the range to become allocatable which means that
1932 // tryRegionSplit won't be making progress. This check should be replaced with
1933 // an assertion when the coalescer is fixed.
1934 if (SA->didRepairRange()) {
1935 // VirtReg has changed, so all cached queries are invalid.
1936 Matrix->invalidateVirtRegs();
1937 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1941 // First try to split around a region spanning multiple blocks. RS_Split2
1942 // ranges already made dubious progress with region splitting, so they go
1943 // straight to single block splitting.
1944 if (getStage(VirtReg) < RS_Split2) {
1945 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
1946 if (PhysReg || !NewVRegs.empty())
1950 // Then isolate blocks.
1951 return tryBlockSplit(VirtReg, Order, NewVRegs);
1954 //===----------------------------------------------------------------------===//
1955 // Last Chance Recoloring
1956 //===----------------------------------------------------------------------===//
1958 /// mayRecolorAllInterferences - Check if the virtual registers that
1959 /// interfere with \p VirtReg on \p PhysReg (or one of its aliases) may be
1960 /// recolored to free \p PhysReg.
1961 /// When true is returned, \p RecoloringCandidates has been augmented with all
1962 /// the live intervals that need to be recolored in order to free \p PhysReg
1964 /// \p FixedRegisters contains all the virtual registers that cannot be
1967 RAGreedy::mayRecolorAllInterferences(unsigned PhysReg, LiveInterval &VirtReg,
1968 SmallLISet &RecoloringCandidates,
1969 const SmallVirtRegSet &FixedRegisters) {
1970 const TargetRegisterClass *CurRC = MRI->getRegClass(VirtReg.reg);
1972 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
1973 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
1974 // If there is LastChanceRecoloringMaxInterference or more interferences,
1975 // chances are one would not be recolorable.
1976 if (Q.collectInterferingVRegs(LastChanceRecoloringMaxInterference) >=
1977 LastChanceRecoloringMaxInterference && !ExhaustiveSearch) {
1978 DEBUG(dbgs() << "Early abort: too many interferences.\n");
1979 CutOffInfo |= CO_Interf;
1982 for (unsigned i = Q.interferingVRegs().size(); i; --i) {
1983 LiveInterval *Intf = Q.interferingVRegs()[i - 1];
1984 // If Intf is done and sit on the same register class as VirtReg,
1985 // it would not be recolorable as it is in the same state as VirtReg.
1986 if ((getStage(*Intf) == RS_Done &&
1987 MRI->getRegClass(Intf->reg) == CurRC) ||
1988 FixedRegisters.count(Intf->reg)) {
1989 DEBUG(dbgs() << "Early abort: the inteference is not recolorable.\n");
1992 RecoloringCandidates.insert(Intf);
1998 /// tryLastChanceRecoloring - Try to assign a color to \p VirtReg by recoloring
1999 /// its interferences.
2000 /// Last chance recoloring chooses a color for \p VirtReg and recolors every
2001 /// virtual register that was using it. The recoloring process may recursively
2002 /// use the last chance recoloring. Therefore, when a virtual register has been
2003 /// assigned a color by this mechanism, it is marked as Fixed, i.e., it cannot
2004 /// be last-chance-recolored again during this recoloring "session".
2007 /// vA can use {R1, R2 }
2008 /// vB can use { R2, R3}
2009 /// vC can use {R1 }
2010 /// Where vA, vB, and vC cannot be split anymore (they are reloads for
2011 /// instance) and they all interfere.
2013 /// vA is assigned R1
2014 /// vB is assigned R2
2015 /// vC tries to evict vA but vA is already done.
2016 /// Regular register allocation fails.
2018 /// Last chance recoloring kicks in:
2019 /// vC does as if vA was evicted => vC uses R1.
2020 /// vC is marked as fixed.
2021 /// vA needs to find a color.
2022 /// None are available.
2023 /// vA cannot evict vC: vC is a fixed virtual register now.
2024 /// vA does as if vB was evicted => vA uses R2.
2025 /// vB needs to find a color.
2026 /// R3 is available.
2027 /// Recoloring => vC = R1, vA = R2, vB = R3
2029 /// \p Order defines the preferred allocation order for \p VirtReg.
2030 /// \p NewRegs will contain any new virtual register that have been created
2031 /// (split, spill) during the process and that must be assigned.
2032 /// \p FixedRegisters contains all the virtual registers that cannot be
2034 /// \p Depth gives the current depth of the last chance recoloring.
2035 /// \return a physical register that can be used for VirtReg or ~0u if none
2037 unsigned RAGreedy::tryLastChanceRecoloring(LiveInterval &VirtReg,
2038 AllocationOrder &Order,
2039 SmallVectorImpl<unsigned> &NewVRegs,
2040 SmallVirtRegSet &FixedRegisters,
2042 DEBUG(dbgs() << "Try last chance recoloring for " << VirtReg << '\n');
2043 // Ranges must be Done.
2044 assert((getStage(VirtReg) >= RS_Done || !VirtReg.isSpillable()) &&
2045 "Last chance recoloring should really be last chance");
2046 // Set the max depth to LastChanceRecoloringMaxDepth.
2047 // We may want to reconsider that if we end up with a too large search space
2048 // for target with hundreds of registers.
2049 // Indeed, in that case we may want to cut the search space earlier.
2050 if (Depth >= LastChanceRecoloringMaxDepth && !ExhaustiveSearch) {
2051 DEBUG(dbgs() << "Abort because max depth has been reached.\n");
2052 CutOffInfo |= CO_Depth;
2056 // Set of Live intervals that will need to be recolored.
2057 SmallLISet RecoloringCandidates;
2058 // Record the original mapping virtual register to physical register in case
2059 // the recoloring fails.
2060 DenseMap<unsigned, unsigned> VirtRegToPhysReg;
2061 // Mark VirtReg as fixed, i.e., it will not be recolored pass this point in
2062 // this recoloring "session".
2063 FixedRegisters.insert(VirtReg.reg);
2066 while (unsigned PhysReg = Order.next()) {
2067 DEBUG(dbgs() << "Try to assign: " << VirtReg << " to "
2068 << PrintReg(PhysReg, TRI) << '\n');
2069 RecoloringCandidates.clear();
2070 VirtRegToPhysReg.clear();
2072 // It is only possible to recolor virtual register interference.
2073 if (Matrix->checkInterference(VirtReg, PhysReg) >
2074 LiveRegMatrix::IK_VirtReg) {
2075 DEBUG(dbgs() << "Some inteferences are not with virtual registers.\n");
2080 // Early give up on this PhysReg if it is obvious we cannot recolor all
2081 // the interferences.
2082 if (!mayRecolorAllInterferences(PhysReg, VirtReg, RecoloringCandidates,
2084 DEBUG(dbgs() << "Some inteferences cannot be recolored.\n");
2088 // RecoloringCandidates contains all the virtual registers that interfer
2089 // with VirtReg on PhysReg (or one of its aliases).
2090 // Enqueue them for recoloring and perform the actual recoloring.
2091 PQueue RecoloringQueue;
2092 for (SmallLISet::iterator It = RecoloringCandidates.begin(),
2093 EndIt = RecoloringCandidates.end();
2094 It != EndIt; ++It) {
2095 unsigned ItVirtReg = (*It)->reg;
2096 enqueue(RecoloringQueue, *It);
2097 assert(VRM->hasPhys(ItVirtReg) &&
2098 "Interferences are supposed to be with allocated vairables");
2100 // Record the current allocation.
2101 VirtRegToPhysReg[ItVirtReg] = VRM->getPhys(ItVirtReg);
2102 // unset the related struct.
2103 Matrix->unassign(**It);
2106 // Do as if VirtReg was assigned to PhysReg so that the underlying
2107 // recoloring has the right information about the interferes and
2108 // available colors.
2109 Matrix->assign(VirtReg, PhysReg);
2111 // Save the current recoloring state.
2112 // If we cannot recolor all the interferences, we will have to start again
2113 // at this point for the next physical register.
2114 SmallVirtRegSet SaveFixedRegisters(FixedRegisters);
2115 if (tryRecoloringCandidates(RecoloringQueue, NewVRegs, FixedRegisters,
2117 // Do not mess up with the global assignment process.
2118 // I.e., VirtReg must be unassigned.
2119 Matrix->unassign(VirtReg);
2123 DEBUG(dbgs() << "Fail to assign: " << VirtReg << " to "
2124 << PrintReg(PhysReg, TRI) << '\n');
2126 // The recoloring attempt failed, undo the changes.
2127 FixedRegisters = SaveFixedRegisters;
2128 Matrix->unassign(VirtReg);
2130 for (SmallLISet::iterator It = RecoloringCandidates.begin(),
2131 EndIt = RecoloringCandidates.end();
2132 It != EndIt; ++It) {
2133 unsigned ItVirtReg = (*It)->reg;
2134 if (VRM->hasPhys(ItVirtReg))
2135 Matrix->unassign(**It);
2136 Matrix->assign(**It, VirtRegToPhysReg[ItVirtReg]);
2140 // Last chance recoloring did not worked either, give up.
2144 /// tryRecoloringCandidates - Try to assign a new color to every register
2145 /// in \RecoloringQueue.
2146 /// \p NewRegs will contain any new virtual register created during the
2147 /// recoloring process.
2148 /// \p FixedRegisters[in/out] contains all the registers that have been
2150 /// \return true if all virtual registers in RecoloringQueue were successfully
2151 /// recolored, false otherwise.
2152 bool RAGreedy::tryRecoloringCandidates(PQueue &RecoloringQueue,
2153 SmallVectorImpl<unsigned> &NewVRegs,
2154 SmallVirtRegSet &FixedRegisters,
2156 while (!RecoloringQueue.empty()) {
2157 LiveInterval *LI = dequeue(RecoloringQueue);
2158 DEBUG(dbgs() << "Try to recolor: " << *LI << '\n');
2160 PhysReg = selectOrSplitImpl(*LI, NewVRegs, FixedRegisters, Depth + 1);
2161 if (PhysReg == ~0u || !PhysReg)
2163 DEBUG(dbgs() << "Recoloring of " << *LI
2164 << " succeeded with: " << PrintReg(PhysReg, TRI) << '\n');
2165 Matrix->assign(*LI, PhysReg);
2166 FixedRegisters.insert(LI->reg);
2171 //===----------------------------------------------------------------------===//
2173 //===----------------------------------------------------------------------===//
2175 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
2176 SmallVectorImpl<unsigned> &NewVRegs) {
2177 CutOffInfo = CO_None;
2178 LLVMContext &Ctx = MF->getFunction()->getContext();
2179 SmallVirtRegSet FixedRegisters;
2180 unsigned Reg = selectOrSplitImpl(VirtReg, NewVRegs, FixedRegisters);
2181 if (Reg == ~0U && (CutOffInfo != CO_None)) {
2182 uint8_t CutOffEncountered = CutOffInfo & (CO_Depth | CO_Interf);
2183 if (CutOffEncountered == CO_Depth)
2184 Ctx.emitError("register allocation failed: maximum depth for recoloring "
2185 "reached. Use -fexhaustive-register-search to skip "
2187 else if (CutOffEncountered == CO_Interf)
2188 Ctx.emitError("register allocation failed: maximum interference for "
2189 "recoloring reached. Use -fexhaustive-register-search "
2191 else if (CutOffEncountered == (CO_Depth | CO_Interf))
2192 Ctx.emitError("register allocation failed: maximum interference and "
2193 "depth for recoloring reached. Use "
2194 "-fexhaustive-register-search to skip cutoffs");
2199 /// Using a CSR for the first time has a cost because it causes push|pop
2200 /// to be added to prologue|epilogue. Splitting a cold section of the live
2201 /// range can have lower cost than using the CSR for the first time;
2202 /// Spilling a live range in the cold path can have lower cost than using
2203 /// the CSR for the first time. Returns the physical register if we decide
2204 /// to use the CSR; otherwise return 0.
2205 unsigned RAGreedy::tryAssignCSRFirstTime(LiveInterval &VirtReg,
2206 AllocationOrder &Order,
2208 unsigned &CostPerUseLimit,
2209 SmallVectorImpl<unsigned> &NewVRegs) {
2210 if (getStage(VirtReg) == RS_Spill && VirtReg.isSpillable()) {
2211 // We choose spill over using the CSR for the first time if the spill cost
2212 // is lower than CSRCost.
2213 SA->analyze(&VirtReg);
2214 if (calcSpillCost() >= CSRCost)
2217 // We are going to spill, set CostPerUseLimit to 1 to make sure that
2218 // we will not use a callee-saved register in tryEvict.
2219 CostPerUseLimit = 1;
2222 if (getStage(VirtReg) < RS_Split) {
2223 // We choose pre-splitting over using the CSR for the first time if
2224 // the cost of splitting is lower than CSRCost.
2225 SA->analyze(&VirtReg);
2226 unsigned NumCands = 0;
2227 BlockFrequency BestCost = CSRCost; // Don't modify CSRCost.
2228 unsigned BestCand = calculateRegionSplitCost(VirtReg, Order, BestCost,
2229 NumCands, true /*IgnoreCSR*/);
2230 if (BestCand == NoCand)
2231 // Use the CSR if we can't find a region split below CSRCost.
2234 // Perform the actual pre-splitting.
2235 doRegionSplit(VirtReg, BestCand, false/*HasCompact*/, NewVRegs);
2241 void RAGreedy::aboutToRemoveInterval(LiveInterval &LI) {
2242 // Do not keep invalid information around.
2243 SetOfBrokenHints.remove(&LI);
2246 void RAGreedy::initializeCSRCost() {
2247 // We use the larger one out of the command-line option and the value report
2249 CSRCost = BlockFrequency(
2250 std::max((unsigned)CSRFirstTimeCost, TRI->getCSRFirstUseCost()));
2251 if (!CSRCost.getFrequency())
2254 // Raw cost is relative to Entry == 2^14; scale it appropriately.
2255 uint64_t ActualEntry = MBFI->getEntryFreq();
2260 uint64_t FixedEntry = 1 << 14;
2261 if (ActualEntry < FixedEntry)
2262 CSRCost *= BranchProbability(ActualEntry, FixedEntry);
2263 else if (ActualEntry <= UINT32_MAX)
2264 // Invert the fraction and divide.
2265 CSRCost /= BranchProbability(FixedEntry, ActualEntry);
2267 // Can't use BranchProbability in general, since it takes 32-bit numbers.
2268 CSRCost = CSRCost.getFrequency() * (ActualEntry / FixedEntry);
2271 /// \brief Collect the hint info for \p Reg.
2272 /// The results are stored into \p Out.
2273 /// \p Out is not cleared before being populated.
2274 void RAGreedy::collectHintInfo(unsigned Reg, HintsInfo &Out) {
2275 for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) {
2276 if (!Instr.isFullCopy())
2278 // Look for the other end of the copy.
2279 unsigned OtherReg = Instr.getOperand(0).getReg();
2280 if (OtherReg == Reg) {
2281 OtherReg = Instr.getOperand(1).getReg();
2282 if (OtherReg == Reg)
2285 // Get the current assignment.
2286 unsigned OtherPhysReg = TargetRegisterInfo::isPhysicalRegister(OtherReg)
2288 : VRM->getPhys(OtherReg);
2289 // Push the collected information.
2290 Out.push_back(HintInfo(MBFI->getBlockFreq(Instr.getParent()), OtherReg,
2295 /// \brief Using the given \p List, compute the cost of the broken hints if
2296 /// \p PhysReg was used.
2297 /// \return The cost of \p List for \p PhysReg.
2298 BlockFrequency RAGreedy::getBrokenHintFreq(const HintsInfo &List,
2300 BlockFrequency Cost = 0;
2301 for (const HintInfo &Info : List) {
2302 if (Info.PhysReg != PhysReg)
2308 /// \brief Using the register assigned to \p VirtReg, try to recolor
2309 /// all the live ranges that are copy-related with \p VirtReg.
2310 /// The recoloring is then propagated to all the live-ranges that have
2311 /// been recolored and so on, until no more copies can be coalesced or
2312 /// it is not profitable.
2313 /// For a given live range, profitability is determined by the sum of the
2314 /// frequencies of the non-identity copies it would introduce with the old
2315 /// and new register.
2316 void RAGreedy::tryHintRecoloring(LiveInterval &VirtReg) {
2317 // We have a broken hint, check if it is possible to fix it by
2318 // reusing PhysReg for the copy-related live-ranges. Indeed, we evicted
2319 // some register and PhysReg may be available for the other live-ranges.
2320 SmallSet<unsigned, 4> Visited;
2321 SmallVector<unsigned, 2> RecoloringCandidates;
2323 unsigned Reg = VirtReg.reg;
2324 unsigned PhysReg = VRM->getPhys(Reg);
2325 // Start the recoloring algorithm from the input live-interval, then
2326 // it will propagate to the ones that are copy-related with it.
2327 Visited.insert(Reg);
2328 RecoloringCandidates.push_back(Reg);
2330 DEBUG(dbgs() << "Trying to reconcile hints for: " << PrintReg(Reg, TRI) << '('
2331 << PrintReg(PhysReg, TRI) << ")\n");
2334 Reg = RecoloringCandidates.pop_back_val();
2336 // We cannot recolor physcal register.
2337 if (TargetRegisterInfo::isPhysicalRegister(Reg))
2340 assert(VRM->hasPhys(Reg) && "We have unallocated variable!!");
2342 // Get the live interval mapped with this virtual register to be able
2343 // to check for the interference with the new color.
2344 LiveInterval &LI = LIS->getInterval(Reg);
2345 unsigned CurrPhys = VRM->getPhys(Reg);
2346 // Check that the new color matches the register class constraints and
2347 // that it is free for this live range.
2348 if (CurrPhys != PhysReg && (!MRI->getRegClass(Reg)->contains(PhysReg) ||
2349 Matrix->checkInterference(LI, PhysReg)))
2352 DEBUG(dbgs() << PrintReg(Reg, TRI) << '(' << PrintReg(CurrPhys, TRI)
2353 << ") is recolorable.\n");
2355 // Gather the hint info.
2357 collectHintInfo(Reg, Info);
2358 // Check if recoloring the live-range will increase the cost of the
2359 // non-identity copies.
2360 if (CurrPhys != PhysReg) {
2361 DEBUG(dbgs() << "Checking profitability:\n");
2362 BlockFrequency OldCopiesCost = getBrokenHintFreq(Info, CurrPhys);
2363 BlockFrequency NewCopiesCost = getBrokenHintFreq(Info, PhysReg);
2364 DEBUG(dbgs() << "Old Cost: " << OldCopiesCost.getFrequency()
2365 << "\nNew Cost: " << NewCopiesCost.getFrequency() << '\n');
2366 if (OldCopiesCost < NewCopiesCost) {
2367 DEBUG(dbgs() << "=> Not profitable.\n");
2370 // At this point, the cost is either cheaper or equal. If it is
2371 // equal, we consider this is profitable because it may expose
2372 // more recoloring opportunities.
2373 DEBUG(dbgs() << "=> Profitable.\n");
2374 // Recolor the live-range.
2375 Matrix->unassign(LI);
2376 Matrix->assign(LI, PhysReg);
2378 // Push all copy-related live-ranges to keep reconciling the broken
2380 for (const HintInfo &HI : Info) {
2381 if (Visited.insert(HI.Reg).second)
2382 RecoloringCandidates.push_back(HI.Reg);
2384 } while (!RecoloringCandidates.empty());
2387 /// \brief Try to recolor broken hints.
2388 /// Broken hints may be repaired by recoloring when an evicted variable
2389 /// freed up a register for a larger live-range.
2390 /// Consider the following example:
2398 /// Let us assume b gets split:
2408 /// Because of how the allocation work, b, c, and d may be assigned different
2409 /// colors. Now, if a gets evicted later:
2419 /// e = ld SpillSlot
2421 /// This is likely that we can assign the same register for b, c, and d,
2422 /// getting rid of 2 copies.
2423 void RAGreedy::tryHintsRecoloring() {
2424 for (LiveInterval *LI : SetOfBrokenHints) {
2425 assert(TargetRegisterInfo::isVirtualRegister(LI->reg) &&
2426 "Recoloring is possible only for virtual registers");
2427 // Some dead defs may be around (e.g., because of debug uses).
2429 if (!VRM->hasPhys(LI->reg))
2431 tryHintRecoloring(*LI);
2435 unsigned RAGreedy::selectOrSplitImpl(LiveInterval &VirtReg,
2436 SmallVectorImpl<unsigned> &NewVRegs,
2437 SmallVirtRegSet &FixedRegisters,
2439 unsigned CostPerUseLimit = ~0u;
2440 // First try assigning a free register.
2441 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
2442 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs)) {
2443 // We check other options if we are using a CSR for the first time.
2444 bool CSRFirstUse = false;
2445 if (unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg))
2446 if (!MRI->isPhysRegUsed(CSR))
2449 // When NewVRegs is not empty, we may have made decisions such as evicting
2450 // a virtual register, go with the earlier decisions and use the physical
2452 if (CSRCost.getFrequency() && CSRFirstUse && NewVRegs.empty()) {
2453 unsigned CSRReg = tryAssignCSRFirstTime(VirtReg, Order, PhysReg,
2454 CostPerUseLimit, NewVRegs);
2455 if (CSRReg || !NewVRegs.empty())
2456 // Return now if we decide to use a CSR or create new vregs due to
2463 LiveRangeStage Stage = getStage(VirtReg);
2464 DEBUG(dbgs() << StageName[Stage]
2465 << " Cascade " << ExtraRegInfo[VirtReg.reg].Cascade << '\n');
2467 // Try to evict a less worthy live range, but only for ranges from the primary
2468 // queue. The RS_Split ranges already failed to do this, and they should not
2469 // get a second chance until they have been split.
2470 if (Stage != RS_Split)
2471 if (unsigned PhysReg =
2472 tryEvict(VirtReg, Order, NewVRegs, CostPerUseLimit)) {
2473 unsigned Hint = MRI->getSimpleHint(VirtReg.reg);
2474 // If VirtReg has a hint and that hint is broken record this
2475 // virtual register as a recoloring candidate for broken hint.
2476 // Indeed, since we evicted a variable in its neighborhood it is
2477 // likely we can at least partially recolor some of the
2478 // copy-related live-ranges.
2479 if (Hint && Hint != PhysReg)
2480 SetOfBrokenHints.insert(&VirtReg);
2484 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
2486 // The first time we see a live range, don't try to split or spill.
2487 // Wait until the second time, when all smaller ranges have been allocated.
2488 // This gives a better picture of the interference to split around.
2489 if (Stage < RS_Split) {
2490 setStage(VirtReg, RS_Split);
2491 DEBUG(dbgs() << "wait for second round\n");
2492 NewVRegs.push_back(VirtReg.reg);
2496 // If we couldn't allocate a register from spilling, there is probably some
2497 // invalid inline assembly. The base class wil report it.
2498 if (Stage >= RS_Done || !VirtReg.isSpillable())
2499 return tryLastChanceRecoloring(VirtReg, Order, NewVRegs, FixedRegisters,
2502 // Try splitting VirtReg or interferences.
2503 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
2504 if (PhysReg || !NewVRegs.empty())
2507 // Finally spill VirtReg itself.
2508 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
2509 LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
2510 spiller().spill(LRE);
2511 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done);
2514 MF->verify(this, "After spilling");
2516 // The live virtual register requesting allocation was spilled, so tell
2517 // the caller not to allocate anything during this round.
2521 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
2522 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
2523 << "********** Function: " << mf.getName() << '\n');
2526 TRI = MF->getSubtarget().getRegisterInfo();
2527 TII = MF->getSubtarget().getInstrInfo();
2528 RCI.runOnMachineFunction(mf);
2530 EnableLocalReassign = EnableLocalReassignment ||
2531 MF->getSubtarget().enableRALocalReassignment(
2532 MF->getTarget().getOptLevel());
2535 MF->verify(this, "Before greedy register allocator");
2537 RegAllocBase::init(getAnalysis<VirtRegMap>(),
2538 getAnalysis<LiveIntervals>(),
2539 getAnalysis<LiveRegMatrix>());
2540 Indexes = &getAnalysis<SlotIndexes>();
2541 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
2542 DomTree = &getAnalysis<MachineDominatorTree>();
2543 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
2544 Loops = &getAnalysis<MachineLoopInfo>();
2545 Bundles = &getAnalysis<EdgeBundles>();
2546 SpillPlacer = &getAnalysis<SpillPlacement>();
2547 DebugVars = &getAnalysis<LiveDebugVariables>();
2549 initializeCSRCost();
2551 calculateSpillWeightsAndHints(*LIS, mf, *Loops, *MBFI);
2555 SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
2556 SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree, *MBFI));
2557 ExtraRegInfo.clear();
2558 ExtraRegInfo.resize(MRI->getNumVirtRegs());
2560 IntfCache.init(MF, Matrix->getLiveUnions(), Indexes, LIS, TRI);
2561 GlobalCand.resize(32); // This will grow as needed.
2562 SetOfBrokenHints.clear();
2565 tryHintsRecoloring();