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 #define DEBUG_TYPE "regalloc"
16 #include "AllocationOrder.h"
17 #include "InterferenceCache.h"
18 #include "LiveDebugVariables.h"
19 #include "LiveRangeEdit.h"
20 #include "RegAllocBase.h"
22 #include "SpillPlacement.h"
24 #include "VirtRegMap.h"
25 #include "RegisterCoalescer.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Function.h"
29 #include "llvm/PassAnalysisSupport.h"
30 #include "llvm/CodeGen/CalcSpillWeights.h"
31 #include "llvm/CodeGen/EdgeBundles.h"
32 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
33 #include "llvm/CodeGen/LiveStackAnalysis.h"
34 #include "llvm/CodeGen/MachineDominators.h"
35 #include "llvm/CodeGen/MachineFunctionPass.h"
36 #include "llvm/CodeGen/MachineLoopInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/Passes.h"
39 #include "llvm/CodeGen/RegAllocRegistry.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Support/Timer.h"
51 STATISTIC(NumGlobalSplits, "Number of split global live ranges");
52 STATISTIC(NumLocalSplits, "Number of split local live ranges");
53 STATISTIC(NumEvicted, "Number of interferences evicted");
55 cl::opt<bool> CompactRegions("compact-regions", cl::init(true));
57 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
58 createGreedyRegisterAllocator);
61 class RAGreedy : public MachineFunctionPass,
63 private LiveRangeEdit::Delegate {
71 MachineDominatorTree *DomTree;
72 MachineLoopInfo *Loops;
74 SpillPlacement *SpillPlacer;
75 LiveDebugVariables *DebugVars;
78 std::auto_ptr<Spiller> SpillerInstance;
79 std::priority_queue<std::pair<unsigned, unsigned> > Queue;
82 // Live ranges pass through a number of stages as we try to allocate them.
83 // Some of the stages may also create new live ranges:
85 // - Region splitting.
86 // - Per-block splitting.
90 // Ranges produced by one of the stages skip the previous stages when they are
91 // dequeued. This improves performance because we can skip interference checks
92 // that are unlikely to give any results. It also guarantees that the live
93 // range splitting algorithm terminates, something that is otherwise hard to
96 /// Newly created live range that has never been queued.
99 /// Only attempt assignment and eviction. Then requeue as RS_Split.
102 /// Attempt live range splitting if assignment is impossible.
105 /// Attempt more aggressive live range splitting that is guaranteed to make
106 /// progress. This is used for split products that may not be making
110 /// Live range will be spilled. No more splitting will be attempted.
113 /// There is nothing more we can do to this live range. Abort compilation
114 /// if it can't be assigned.
118 static const char *const StageName[];
120 // RegInfo - Keep additional information about each live range.
122 LiveRangeStage Stage;
124 // Cascade - Eviction loop prevention. See canEvictInterference().
127 RegInfo() : Stage(RS_New), Cascade(0) {}
130 IndexedMap<RegInfo, VirtReg2IndexFunctor> ExtraRegInfo;
132 LiveRangeStage getStage(const LiveInterval &VirtReg) const {
133 return ExtraRegInfo[VirtReg.reg].Stage;
136 void setStage(const LiveInterval &VirtReg, LiveRangeStage Stage) {
137 ExtraRegInfo.resize(MRI->getNumVirtRegs());
138 ExtraRegInfo[VirtReg.reg].Stage = Stage;
141 template<typename Iterator>
142 void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
143 ExtraRegInfo.resize(MRI->getNumVirtRegs());
144 for (;Begin != End; ++Begin) {
145 unsigned Reg = (*Begin)->reg;
146 if (ExtraRegInfo[Reg].Stage == RS_New)
147 ExtraRegInfo[Reg].Stage = NewStage;
151 /// Cost of evicting interference.
152 struct EvictionCost {
153 unsigned BrokenHints; ///< Total number of broken hints.
154 float MaxWeight; ///< Maximum spill weight evicted.
156 EvictionCost(unsigned B = 0) : BrokenHints(B), MaxWeight(0) {}
158 bool operator<(const EvictionCost &O) const {
159 if (BrokenHints != O.BrokenHints)
160 return BrokenHints < O.BrokenHints;
161 return MaxWeight < O.MaxWeight;
166 std::auto_ptr<SplitAnalysis> SA;
167 std::auto_ptr<SplitEditor> SE;
169 /// Cached per-block interference maps
170 InterferenceCache IntfCache;
172 /// All basic blocks where the current register has uses.
173 SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints;
175 /// Global live range splitting candidate info.
176 struct GlobalSplitCandidate {
177 // Register intended for assignment, or 0.
180 // SplitKit interval index for this candidate.
183 // Interference for PhysReg.
184 InterferenceCache::Cursor Intf;
186 // Bundles where this candidate should be live.
187 BitVector LiveBundles;
188 SmallVector<unsigned, 8> ActiveBlocks;
190 void reset(InterferenceCache &Cache, unsigned Reg) {
193 Intf.setPhysReg(Cache, Reg);
195 ActiveBlocks.clear();
198 // Set B[i] = C for every live bundle where B[i] was NoCand.
199 unsigned getBundles(SmallVectorImpl<unsigned> &B, unsigned C) {
201 for (int i = LiveBundles.find_first(); i >= 0;
202 i = LiveBundles.find_next(i))
203 if (B[i] == NoCand) {
211 /// Candidate info for for each PhysReg in AllocationOrder.
212 /// This vector never shrinks, but grows to the size of the largest register
214 SmallVector<GlobalSplitCandidate, 32> GlobalCand;
216 enum { NoCand = ~0u };
218 /// Candidate map. Each edge bundle is assigned to a GlobalCand entry, or to
219 /// NoCand which indicates the stack interval.
220 SmallVector<unsigned, 32> BundleCand;
225 /// Return the pass name.
226 virtual const char* getPassName() const {
227 return "Greedy Register Allocator";
230 /// RAGreedy analysis usage.
231 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
232 virtual void releaseMemory();
233 virtual Spiller &spiller() { return *SpillerInstance; }
234 virtual void enqueue(LiveInterval *LI);
235 virtual LiveInterval *dequeue();
236 virtual unsigned selectOrSplit(LiveInterval&,
237 SmallVectorImpl<LiveInterval*>&);
239 /// Perform register allocation.
240 virtual bool runOnMachineFunction(MachineFunction &mf);
245 void LRE_WillEraseInstruction(MachineInstr*);
246 bool LRE_CanEraseVirtReg(unsigned);
247 void LRE_WillShrinkVirtReg(unsigned);
248 void LRE_DidCloneVirtReg(unsigned, unsigned);
250 float calcSpillCost();
251 bool addSplitConstraints(InterferenceCache::Cursor, float&);
252 void addThroughConstraints(InterferenceCache::Cursor, ArrayRef<unsigned>);
253 void growRegion(GlobalSplitCandidate &Cand);
254 float calcGlobalSplitCost(GlobalSplitCandidate&);
255 bool calcCompactRegion(GlobalSplitCandidate&);
256 void splitAroundRegion(LiveRangeEdit&, ArrayRef<unsigned>);
257 void calcGapWeights(unsigned, SmallVectorImpl<float>&);
258 bool shouldEvict(LiveInterval &A, bool, LiveInterval &B, bool);
259 bool canEvictInterference(LiveInterval&, unsigned, bool, EvictionCost&);
260 void evictInterference(LiveInterval&, unsigned,
261 SmallVectorImpl<LiveInterval*>&);
263 unsigned tryAssign(LiveInterval&, AllocationOrder&,
264 SmallVectorImpl<LiveInterval*>&);
265 unsigned tryEvict(LiveInterval&, AllocationOrder&,
266 SmallVectorImpl<LiveInterval*>&, unsigned = ~0u);
267 unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
268 SmallVectorImpl<LiveInterval*>&);
269 unsigned tryBlockSplit(LiveInterval&, AllocationOrder&,
270 SmallVectorImpl<LiveInterval*>&);
271 unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
272 SmallVectorImpl<LiveInterval*>&);
273 unsigned trySplit(LiveInterval&, AllocationOrder&,
274 SmallVectorImpl<LiveInterval*>&);
276 } // end anonymous namespace
278 char RAGreedy::ID = 0;
281 const char *const RAGreedy::StageName[] = {
291 // Hysteresis to use when comparing floats.
292 // This helps stabilize decisions based on float comparisons.
293 const float Hysteresis = 0.98f;
296 FunctionPass* llvm::createGreedyRegisterAllocator() {
297 return new RAGreedy();
300 RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
301 initializeLiveDebugVariablesPass(*PassRegistry::getPassRegistry());
302 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
303 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
304 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
305 initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
306 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
307 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
308 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
309 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
310 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
311 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
312 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
313 initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
316 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
317 AU.setPreservesCFG();
318 AU.addRequired<AliasAnalysis>();
319 AU.addPreserved<AliasAnalysis>();
320 AU.addRequired<LiveIntervals>();
321 AU.addRequired<SlotIndexes>();
322 AU.addPreserved<SlotIndexes>();
323 AU.addRequired<LiveDebugVariables>();
324 AU.addPreserved<LiveDebugVariables>();
326 AU.addRequiredID(StrongPHIEliminationID);
327 AU.addRequiredTransitive<RegisterCoalescer>();
328 AU.addRequired<CalculateSpillWeights>();
329 AU.addRequired<LiveStacks>();
330 AU.addPreserved<LiveStacks>();
331 AU.addRequired<MachineDominatorTree>();
332 AU.addPreserved<MachineDominatorTree>();
333 AU.addRequired<MachineLoopInfo>();
334 AU.addPreserved<MachineLoopInfo>();
335 AU.addRequired<VirtRegMap>();
336 AU.addPreserved<VirtRegMap>();
337 AU.addRequired<EdgeBundles>();
338 AU.addRequired<SpillPlacement>();
339 MachineFunctionPass::getAnalysisUsage(AU);
343 //===----------------------------------------------------------------------===//
344 // LiveRangeEdit delegate methods
345 //===----------------------------------------------------------------------===//
347 void RAGreedy::LRE_WillEraseInstruction(MachineInstr *MI) {
348 // LRE itself will remove from SlotIndexes and parent basic block.
349 VRM->RemoveMachineInstrFromMaps(MI);
352 bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
353 if (unsigned PhysReg = VRM->getPhys(VirtReg)) {
354 unassign(LIS->getInterval(VirtReg), PhysReg);
357 // Unassigned virtreg is probably in the priority queue.
358 // RegAllocBase will erase it after dequeueing.
362 void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
363 unsigned PhysReg = VRM->getPhys(VirtReg);
367 // Register is assigned, put it back on the queue for reassignment.
368 LiveInterval &LI = LIS->getInterval(VirtReg);
369 unassign(LI, PhysReg);
373 void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
374 // LRE may clone a virtual register because dead code elimination causes it to
375 // be split into connected components. The new components are much smaller
376 // than the original, so they should get a new chance at being assigned.
377 // same stage as the parent.
378 ExtraRegInfo[Old].Stage = RS_Assign;
379 ExtraRegInfo.grow(New);
380 ExtraRegInfo[New] = ExtraRegInfo[Old];
383 void RAGreedy::releaseMemory() {
384 SpillerInstance.reset(0);
385 ExtraRegInfo.clear();
387 RegAllocBase::releaseMemory();
390 void RAGreedy::enqueue(LiveInterval *LI) {
391 // Prioritize live ranges by size, assigning larger ranges first.
392 // The queue holds (size, reg) pairs.
393 const unsigned Size = LI->getSize();
394 const unsigned Reg = LI->reg;
395 assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
396 "Can only enqueue virtual registers");
399 ExtraRegInfo.grow(Reg);
400 if (ExtraRegInfo[Reg].Stage == RS_New)
401 ExtraRegInfo[Reg].Stage = RS_Assign;
403 if (ExtraRegInfo[Reg].Stage == RS_Split) {
404 // Unsplit ranges that couldn't be allocated immediately are deferred until
405 // everything else has been allocated. Long ranges are allocated last so
406 // they are split against realistic interference.
410 Prio = (1u << 31) - Size;
412 // Everything else is allocated in long->short order. Long ranges that don't
413 // fit should be spilled ASAP so they don't create interference.
414 Prio = (1u << 31) + Size;
416 // Boost ranges that have a physical register hint.
417 if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
421 Queue.push(std::make_pair(Prio, Reg));
424 LiveInterval *RAGreedy::dequeue() {
427 LiveInterval *LI = &LIS->getInterval(Queue.top().second);
433 //===----------------------------------------------------------------------===//
435 //===----------------------------------------------------------------------===//
437 /// tryAssign - Try to assign VirtReg to an available register.
438 unsigned RAGreedy::tryAssign(LiveInterval &VirtReg,
439 AllocationOrder &Order,
440 SmallVectorImpl<LiveInterval*> &NewVRegs) {
443 while ((PhysReg = Order.next()))
444 if (!checkPhysRegInterference(VirtReg, PhysReg))
446 if (!PhysReg || Order.isHint(PhysReg))
449 // PhysReg is available, but there may be a better choice.
451 // If we missed a simple hint, try to cheaply evict interference from the
452 // preferred register.
453 if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg))
454 if (Order.isHint(Hint)) {
455 DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n');
456 EvictionCost MaxCost(1);
457 if (canEvictInterference(VirtReg, Hint, true, MaxCost)) {
458 evictInterference(VirtReg, Hint, NewVRegs);
463 // Try to evict interference from a cheaper alternative.
464 unsigned Cost = TRI->getCostPerUse(PhysReg);
466 // Most registers have 0 additional cost.
470 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " is available at cost " << Cost
472 unsigned CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost);
473 return CheapReg ? CheapReg : PhysReg;
477 //===----------------------------------------------------------------------===//
478 // Interference eviction
479 //===----------------------------------------------------------------------===//
481 /// shouldEvict - determine if A should evict the assigned live range B. The
482 /// eviction policy defined by this function together with the allocation order
483 /// defined by enqueue() decides which registers ultimately end up being split
486 /// Cascade numbers are used to prevent infinite loops if this function is a
489 /// @param A The live range to be assigned.
490 /// @param IsHint True when A is about to be assigned to its preferred
492 /// @param B The live range to be evicted.
493 /// @param BreaksHint True when B is already assigned to its preferred register.
494 bool RAGreedy::shouldEvict(LiveInterval &A, bool IsHint,
495 LiveInterval &B, bool BreaksHint) {
496 bool CanSplit = getStage(B) < RS_Spill;
498 // Be fairly aggressive about following hints as long as the evictee can be
500 if (CanSplit && IsHint && !BreaksHint)
503 return A.weight > B.weight;
506 /// canEvictInterference - Return true if all interferences between VirtReg and
507 /// PhysReg can be evicted. When OnlyCheap is set, don't do anything
509 /// @param VirtReg Live range that is about to be assigned.
510 /// @param PhysReg Desired register for assignment.
511 /// @prarm IsHint True when PhysReg is VirtReg's preferred register.
512 /// @param MaxCost Only look for cheaper candidates and update with new cost
513 /// when returning true.
514 /// @returns True when interference can be evicted cheaper than MaxCost.
515 bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
516 bool IsHint, EvictionCost &MaxCost) {
517 // Find VirtReg's cascade number. This will be unassigned if VirtReg was never
518 // involved in an eviction before. If a cascade number was assigned, deny
519 // evicting anything with the same or a newer cascade number. This prevents
520 // infinite eviction loops.
522 // This works out so a register without a cascade number is allowed to evict
523 // anything, and it can be evicted by anything.
524 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
526 Cascade = NextCascade;
529 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
530 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
531 // If there is 10 or more interferences, chances are one is heavier.
532 if (Q.collectInterferingVRegs(10) >= 10)
535 // Check if any interfering live range is heavier than MaxWeight.
536 for (unsigned i = Q.interferingVRegs().size(); i; --i) {
537 LiveInterval *Intf = Q.interferingVRegs()[i - 1];
538 if (TargetRegisterInfo::isPhysicalRegister(Intf->reg))
540 // Never evict spill products. They cannot split or spill.
541 if (getStage(*Intf) == RS_Done)
543 // Once a live range becomes small enough, it is urgent that we find a
544 // register for it. This is indicated by an infinite spill weight. These
545 // urgent live ranges get to evict almost anything.
546 bool Urgent = !VirtReg.isSpillable() && Intf->isSpillable();
547 // Only evict older cascades or live ranges without a cascade.
548 unsigned IntfCascade = ExtraRegInfo[Intf->reg].Cascade;
549 if (Cascade <= IntfCascade) {
552 // We permit breaking cascades for urgent evictions. It should be the
553 // last resort, though, so make it really expensive.
554 Cost.BrokenHints += 10;
556 // Would this break a satisfied hint?
557 bool BreaksHint = VRM->hasPreferredPhys(Intf->reg);
558 // Update eviction cost.
559 Cost.BrokenHints += BreaksHint;
560 Cost.MaxWeight = std::max(Cost.MaxWeight, Intf->weight);
561 // Abort if this would be too expensive.
562 if (!(Cost < MaxCost))
564 // Finally, apply the eviction policy for non-urgent evictions.
565 if (!Urgent && !shouldEvict(VirtReg, IsHint, *Intf, BreaksHint))
573 /// evictInterference - Evict any interferring registers that prevent VirtReg
574 /// from being assigned to Physreg. This assumes that canEvictInterference
576 void RAGreedy::evictInterference(LiveInterval &VirtReg, unsigned PhysReg,
577 SmallVectorImpl<LiveInterval*> &NewVRegs) {
578 // Make sure that VirtReg has a cascade number, and assign that cascade
579 // number to every evicted register. These live ranges than then only be
580 // evicted by a newer cascade, preventing infinite loops.
581 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
583 Cascade = ExtraRegInfo[VirtReg.reg].Cascade = NextCascade++;
585 DEBUG(dbgs() << "evicting " << PrintReg(PhysReg, TRI)
586 << " interference: Cascade " << Cascade << '\n');
587 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
588 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
589 assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
590 for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
591 LiveInterval *Intf = Q.interferingVRegs()[i];
592 unassign(*Intf, VRM->getPhys(Intf->reg));
593 assert((ExtraRegInfo[Intf->reg].Cascade < Cascade ||
594 VirtReg.isSpillable() < Intf->isSpillable()) &&
595 "Cannot decrease cascade number, illegal eviction");
596 ExtraRegInfo[Intf->reg].Cascade = Cascade;
598 NewVRegs.push_back(Intf);
603 /// tryEvict - Try to evict all interferences for a physreg.
604 /// @param VirtReg Currently unassigned virtual register.
605 /// @param Order Physregs to try.
606 /// @return Physreg to assign VirtReg, or 0.
607 unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
608 AllocationOrder &Order,
609 SmallVectorImpl<LiveInterval*> &NewVRegs,
610 unsigned CostPerUseLimit) {
611 NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
613 // Keep track of the cheapest interference seen so far.
614 EvictionCost BestCost(~0u);
615 unsigned BestPhys = 0;
617 // When we are just looking for a reduced cost per use, don't break any
618 // hints, and only evict smaller spill weights.
619 if (CostPerUseLimit < ~0u) {
620 BestCost.BrokenHints = 0;
621 BestCost.MaxWeight = VirtReg.weight;
625 while (unsigned PhysReg = Order.next()) {
626 if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
628 // The first use of a callee-saved register in a function has cost 1.
629 // Don't start using a CSR when the CostPerUseLimit is low.
630 if (CostPerUseLimit == 1)
631 if (unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg))
632 if (!MRI->isPhysRegUsed(CSR)) {
633 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " would clobber CSR "
634 << PrintReg(CSR, TRI) << '\n');
638 if (!canEvictInterference(VirtReg, PhysReg, false, BestCost))
644 // Stop if the hint can be used.
645 if (Order.isHint(PhysReg))
652 evictInterference(VirtReg, BestPhys, NewVRegs);
657 //===----------------------------------------------------------------------===//
659 //===----------------------------------------------------------------------===//
661 /// addSplitConstraints - Fill out the SplitConstraints vector based on the
662 /// interference pattern in Physreg and its aliases. Add the constraints to
663 /// SpillPlacement and return the static cost of this split in Cost, assuming
664 /// that all preferences in SplitConstraints are met.
665 /// Return false if there are no bundles with positive bias.
666 bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf,
668 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
670 // Reset interference dependent info.
671 SplitConstraints.resize(UseBlocks.size());
672 float StaticCost = 0;
673 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
674 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
675 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
677 BC.Number = BI.MBB->getNumber();
678 Intf.moveToBlock(BC.Number);
679 BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
680 BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
681 BC.ChangesValue = BI.FirstDef;
683 if (!Intf.hasInterference())
686 // Number of spill code instructions to insert.
689 // Interference for the live-in value.
691 if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
692 BC.Entry = SpillPlacement::MustSpill, ++Ins;
693 else if (Intf.first() < BI.FirstInstr)
694 BC.Entry = SpillPlacement::PrefSpill, ++Ins;
695 else if (Intf.first() < BI.LastInstr)
699 // Interference for the live-out value.
701 if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
702 BC.Exit = SpillPlacement::MustSpill, ++Ins;
703 else if (Intf.last() > BI.LastInstr)
704 BC.Exit = SpillPlacement::PrefSpill, ++Ins;
705 else if (Intf.last() > BI.FirstInstr)
709 // Accumulate the total frequency of inserted spill code.
711 StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
715 // Add constraints for use-blocks. Note that these are the only constraints
716 // that may add a positive bias, it is downhill from here.
717 SpillPlacer->addConstraints(SplitConstraints);
718 return SpillPlacer->scanActiveBundles();
722 /// addThroughConstraints - Add constraints and links to SpillPlacer from the
723 /// live-through blocks in Blocks.
724 void RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf,
725 ArrayRef<unsigned> Blocks) {
726 const unsigned GroupSize = 8;
727 SpillPlacement::BlockConstraint BCS[GroupSize];
728 unsigned TBS[GroupSize];
729 unsigned B = 0, T = 0;
731 for (unsigned i = 0; i != Blocks.size(); ++i) {
732 unsigned Number = Blocks[i];
733 Intf.moveToBlock(Number);
735 if (!Intf.hasInterference()) {
736 assert(T < GroupSize && "Array overflow");
738 if (++T == GroupSize) {
739 SpillPlacer->addLinks(makeArrayRef(TBS, T));
745 assert(B < GroupSize && "Array overflow");
746 BCS[B].Number = Number;
748 // Interference for the live-in value.
749 if (Intf.first() <= Indexes->getMBBStartIdx(Number))
750 BCS[B].Entry = SpillPlacement::MustSpill;
752 BCS[B].Entry = SpillPlacement::PrefSpill;
754 // Interference for the live-out value.
755 if (Intf.last() >= SA->getLastSplitPoint(Number))
756 BCS[B].Exit = SpillPlacement::MustSpill;
758 BCS[B].Exit = SpillPlacement::PrefSpill;
760 if (++B == GroupSize) {
761 ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
762 SpillPlacer->addConstraints(Array);
767 ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
768 SpillPlacer->addConstraints(Array);
769 SpillPlacer->addLinks(makeArrayRef(TBS, T));
772 void RAGreedy::growRegion(GlobalSplitCandidate &Cand) {
773 // Keep track of through blocks that have not been added to SpillPlacer.
774 BitVector Todo = SA->getThroughBlocks();
775 SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks;
776 unsigned AddedTo = 0;
778 unsigned Visited = 0;
782 ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive();
783 // Find new through blocks in the periphery of PrefRegBundles.
784 for (int i = 0, e = NewBundles.size(); i != e; ++i) {
785 unsigned Bundle = NewBundles[i];
786 // Look at all blocks connected to Bundle in the full graph.
787 ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle);
788 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
791 if (!Todo.test(Block))
794 // This is a new through block. Add it to SpillPlacer later.
795 ActiveBlocks.push_back(Block);
801 // Any new blocks to add?
802 if (ActiveBlocks.size() == AddedTo)
805 // Compute through constraints from the interference, or assume that all
806 // through blocks prefer spilling when forming compact regions.
807 ArrayRef<unsigned> NewBlocks = makeArrayRef(ActiveBlocks).slice(AddedTo);
809 addThroughConstraints(Cand.Intf, NewBlocks);
811 // Provide a strong negative bias on through blocks to prevent unwanted
812 // liveness on loop backedges.
813 SpillPlacer->addPrefSpill(NewBlocks, /* Strong= */ true);
814 AddedTo = ActiveBlocks.size();
816 // Perhaps iterating can enable more bundles?
817 SpillPlacer->iterate();
819 DEBUG(dbgs() << ", v=" << Visited);
822 /// calcCompactRegion - Compute the set of edge bundles that should be live
823 /// when splitting the current live range into compact regions. Compact
824 /// regions can be computed without looking at interference. They are the
825 /// regions formed by removing all the live-through blocks from the live range.
827 /// Returns false if the current live range is already compact, or if the
828 /// compact regions would form single block regions anyway.
829 bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) {
830 // Without any through blocks, the live range is already compact.
831 if (!SA->getNumThroughBlocks())
834 // Compact regions don't correspond to any physreg.
835 Cand.reset(IntfCache, 0);
837 DEBUG(dbgs() << "Compact region bundles");
839 // Use the spill placer to determine the live bundles. GrowRegion pretends
840 // that all the through blocks have interference when PhysReg is unset.
841 SpillPlacer->prepare(Cand.LiveBundles);
843 // The static split cost will be zero since Cand.Intf reports no interference.
845 if (!addSplitConstraints(Cand.Intf, Cost)) {
846 DEBUG(dbgs() << ", none.\n");
851 SpillPlacer->finish();
853 if (!Cand.LiveBundles.any()) {
854 DEBUG(dbgs() << ", none.\n");
859 for (int i = Cand.LiveBundles.find_first(); i>=0;
860 i = Cand.LiveBundles.find_next(i))
861 dbgs() << " EB#" << i;
867 /// calcSpillCost - Compute how expensive it would be to split the live range in
868 /// SA around all use blocks instead of forming bundle regions.
869 float RAGreedy::calcSpillCost() {
871 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
872 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
873 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
874 unsigned Number = BI.MBB->getNumber();
875 // We normally only need one spill instruction - a load or a store.
876 Cost += SpillPlacer->getBlockFrequency(Number);
878 // Unless the value is redefined in the block.
879 if (BI.LiveIn && BI.LiveOut && BI.FirstDef)
880 Cost += SpillPlacer->getBlockFrequency(Number);
885 /// calcGlobalSplitCost - Return the global split cost of following the split
886 /// pattern in LiveBundles. This cost should be added to the local cost of the
887 /// interference pattern in SplitConstraints.
889 float RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) {
890 float GlobalCost = 0;
891 const BitVector &LiveBundles = Cand.LiveBundles;
892 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
893 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
894 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
895 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
896 bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)];
897 bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
901 Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
903 Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
905 GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
908 for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
909 unsigned Number = Cand.ActiveBlocks[i];
910 bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)];
911 bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)];
912 if (!RegIn && !RegOut)
914 if (RegIn && RegOut) {
915 // We need double spill code if this block has interference.
916 Cand.Intf.moveToBlock(Number);
917 if (Cand.Intf.hasInterference())
918 GlobalCost += 2*SpillPlacer->getBlockFrequency(Number);
921 // live-in / stack-out or stack-in live-out.
922 GlobalCost += SpillPlacer->getBlockFrequency(Number);
927 /// splitAroundRegion - Split the current live range around the regions
928 /// determined by BundleCand and GlobalCand.
930 /// Before calling this function, GlobalCand and BundleCand must be initialized
931 /// so each bundle is assigned to a valid candidate, or NoCand for the
932 /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor
933 /// objects must be initialized for the current live range, and intervals
934 /// created for the used candidates.
936 /// @param LREdit The LiveRangeEdit object handling the current split.
937 /// @param UsedCands List of used GlobalCand entries. Every BundleCand value
938 /// must appear in this list.
939 void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit,
940 ArrayRef<unsigned> UsedCands) {
941 // These are the intervals created for new global ranges. We may create more
942 // intervals for local ranges.
943 const unsigned NumGlobalIntvs = LREdit.size();
944 DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs << " globals.\n");
945 assert(NumGlobalIntvs && "No global intervals configured");
947 // Isolate even single instructions when dealing with a proper sub-class.
948 // That guarantees register class inflation for the stack interval because it
950 unsigned Reg = SA->getParent().reg;
951 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
953 // First handle all the blocks with uses.
954 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
955 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
956 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
957 unsigned Number = BI.MBB->getNumber();
958 unsigned IntvIn = 0, IntvOut = 0;
959 SlotIndex IntfIn, IntfOut;
961 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
962 if (CandIn != NoCand) {
963 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
964 IntvIn = Cand.IntvIdx;
965 Cand.Intf.moveToBlock(Number);
966 IntfIn = Cand.Intf.first();
970 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
971 if (CandOut != NoCand) {
972 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
973 IntvOut = Cand.IntvIdx;
974 Cand.Intf.moveToBlock(Number);
975 IntfOut = Cand.Intf.last();
979 // Create separate intervals for isolated blocks with multiple uses.
980 if (!IntvIn && !IntvOut) {
981 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
982 if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
983 SE->splitSingleBlock(BI);
987 if (IntvIn && IntvOut)
988 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
990 SE->splitRegInBlock(BI, IntvIn, IntfIn);
992 SE->splitRegOutBlock(BI, IntvOut, IntfOut);
995 // Handle live-through blocks. The relevant live-through blocks are stored in
996 // the ActiveBlocks list with each candidate. We need to filter out
998 BitVector Todo = SA->getThroughBlocks();
999 for (unsigned c = 0; c != UsedCands.size(); ++c) {
1000 ArrayRef<unsigned> Blocks = GlobalCand[UsedCands[c]].ActiveBlocks;
1001 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1002 unsigned Number = Blocks[i];
1003 if (!Todo.test(Number))
1007 unsigned IntvIn = 0, IntvOut = 0;
1008 SlotIndex IntfIn, IntfOut;
1010 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
1011 if (CandIn != NoCand) {
1012 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
1013 IntvIn = Cand.IntvIdx;
1014 Cand.Intf.moveToBlock(Number);
1015 IntfIn = Cand.Intf.first();
1018 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
1019 if (CandOut != NoCand) {
1020 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
1021 IntvOut = Cand.IntvIdx;
1022 Cand.Intf.moveToBlock(Number);
1023 IntfOut = Cand.Intf.last();
1025 if (!IntvIn && !IntvOut)
1027 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
1033 SmallVector<unsigned, 8> IntvMap;
1034 SE->finish(&IntvMap);
1035 DebugVars->splitRegister(Reg, LREdit.regs());
1037 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1038 unsigned OrigBlocks = SA->getNumLiveBlocks();
1040 // Sort out the new intervals created by splitting. We get four kinds:
1041 // - Remainder intervals should not be split again.
1042 // - Candidate intervals can be assigned to Cand.PhysReg.
1043 // - Block-local splits are candidates for local splitting.
1044 // - DCE leftovers should go back on the queue.
1045 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1046 LiveInterval &Reg = *LREdit.get(i);
1048 // Ignore old intervals from DCE.
1049 if (getStage(Reg) != RS_New)
1052 // Remainder interval. Don't try splitting again, spill if it doesn't
1054 if (IntvMap[i] == 0) {
1055 setStage(Reg, RS_Spill);
1059 // Global intervals. Allow repeated splitting as long as the number of live
1060 // blocks is strictly decreasing.
1061 if (IntvMap[i] < NumGlobalIntvs) {
1062 if (SA->countLiveBlocks(&Reg) >= OrigBlocks) {
1063 DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks
1064 << " blocks as original.\n");
1065 // Don't allow repeated splitting as a safe guard against looping.
1066 setStage(Reg, RS_Split2);
1071 // Other intervals are treated as new. This includes local intervals created
1072 // for blocks with multiple uses, and anything created by DCE.
1076 MF->verify(this, "After splitting live range around region");
1079 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1080 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1081 unsigned NumCands = 0;
1082 unsigned BestCand = NoCand;
1084 SmallVector<unsigned, 8> UsedCands;
1086 // Check if we can split this live range around a compact region.
1087 bool HasCompact = CompactRegions && calcCompactRegion(GlobalCand.front());
1089 // Yes, keep GlobalCand[0] as the compact region candidate.
1091 BestCost = HUGE_VALF;
1093 // No benefit from the compact region, our fallback will be per-block
1094 // splitting. Make sure we find a solution that is cheaper than spilling.
1095 BestCost = Hysteresis * calcSpillCost();
1096 DEBUG(dbgs() << "Cost of isolating all blocks = " << BestCost << '\n');
1100 while (unsigned PhysReg = Order.next()) {
1101 // Discard bad candidates before we run out of interference cache cursors.
1102 // This will only affect register classes with a lot of registers (>32).
1103 if (NumCands == IntfCache.getMaxCursors()) {
1104 unsigned WorstCount = ~0u;
1106 for (unsigned i = 0; i != NumCands; ++i) {
1107 if (i == BestCand || !GlobalCand[i].PhysReg)
1109 unsigned Count = GlobalCand[i].LiveBundles.count();
1110 if (Count < WorstCount)
1111 Worst = i, WorstCount = Count;
1114 GlobalCand[Worst] = GlobalCand[NumCands];
1117 if (GlobalCand.size() <= NumCands)
1118 GlobalCand.resize(NumCands+1);
1119 GlobalSplitCandidate &Cand = GlobalCand[NumCands];
1120 Cand.reset(IntfCache, PhysReg);
1122 SpillPlacer->prepare(Cand.LiveBundles);
1124 if (!addSplitConstraints(Cand.Intf, Cost)) {
1125 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n");
1128 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
1129 if (Cost >= BestCost) {
1131 if (BestCand == NoCand)
1132 dbgs() << " worse than no bundles\n";
1134 dbgs() << " worse than "
1135 << PrintReg(GlobalCand[BestCand].PhysReg, TRI) << '\n';
1141 SpillPlacer->finish();
1143 // No live bundles, defer to splitSingleBlocks().
1144 if (!Cand.LiveBundles.any()) {
1145 DEBUG(dbgs() << " no bundles.\n");
1149 Cost += calcGlobalSplitCost(Cand);
1151 dbgs() << ", total = " << Cost << " with bundles";
1152 for (int i = Cand.LiveBundles.find_first(); i>=0;
1153 i = Cand.LiveBundles.find_next(i))
1154 dbgs() << " EB#" << i;
1157 if (Cost < BestCost) {
1158 BestCand = NumCands;
1159 BestCost = Hysteresis * Cost; // Prevent rounding effects.
1164 // No solutions found, fall back to single block splitting.
1165 if (!HasCompact && BestCand == NoCand)
1168 // Prepare split editor.
1169 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1172 // Assign all edge bundles to the preferred candidate, or NoCand.
1173 BundleCand.assign(Bundles->getNumBundles(), NoCand);
1175 // Assign bundles for the best candidate region.
1176 if (BestCand != NoCand) {
1177 GlobalSplitCandidate &Cand = GlobalCand[BestCand];
1178 if (unsigned B = Cand.getBundles(BundleCand, BestCand)) {
1179 UsedCands.push_back(BestCand);
1180 Cand.IntvIdx = SE->openIntv();
1181 DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in "
1182 << B << " bundles, intv " << Cand.IntvIdx << ".\n");
1187 // Assign bundles for the compact region.
1189 GlobalSplitCandidate &Cand = GlobalCand.front();
1190 assert(!Cand.PhysReg && "Compact region has no physreg");
1191 if (unsigned B = Cand.getBundles(BundleCand, 0)) {
1192 UsedCands.push_back(0);
1193 Cand.IntvIdx = SE->openIntv();
1194 DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv "
1195 << Cand.IntvIdx << ".\n");
1200 splitAroundRegion(LREdit, UsedCands);
1205 //===----------------------------------------------------------------------===//
1206 // Per-Block Splitting
1207 //===----------------------------------------------------------------------===//
1209 /// tryBlockSplit - Split a global live range around every block with uses. This
1210 /// creates a lot of local live ranges, that will be split by tryLocalSplit if
1211 /// they don't allocate.
1212 unsigned RAGreedy::tryBlockSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1213 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1214 assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed");
1215 unsigned Reg = VirtReg.reg;
1216 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
1217 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1219 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
1220 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
1221 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
1222 if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
1223 SE->splitSingleBlock(BI);
1225 // No blocks were split.
1229 // We did split for some blocks.
1230 SmallVector<unsigned, 8> IntvMap;
1231 SE->finish(&IntvMap);
1233 // Tell LiveDebugVariables about the new ranges.
1234 DebugVars->splitRegister(Reg, LREdit.regs());
1236 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1238 // Sort out the new intervals created by splitting. The remainder interval
1239 // goes straight to spilling, the new local ranges get to stay RS_New.
1240 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1241 LiveInterval &LI = *LREdit.get(i);
1242 if (getStage(LI) == RS_New && IntvMap[i] == 0)
1243 setStage(LI, RS_Spill);
1247 MF->verify(this, "After splitting live range around basic blocks");
1251 //===----------------------------------------------------------------------===//
1253 //===----------------------------------------------------------------------===//
1256 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted
1257 /// in order to use PhysReg between two entries in SA->UseSlots.
1259 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
1261 void RAGreedy::calcGapWeights(unsigned PhysReg,
1262 SmallVectorImpl<float> &GapWeight) {
1263 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1264 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1265 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
1266 const unsigned NumGaps = Uses.size()-1;
1268 // Start and end points for the interference check.
1269 SlotIndex StartIdx =
1270 BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr;
1272 BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr;
1274 GapWeight.assign(NumGaps, 0.0f);
1276 // Add interference from each overlapping register.
1277 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
1278 if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
1279 .checkInterference())
1282 // We know that VirtReg is a continuous interval from FirstInstr to
1283 // LastInstr, so we don't need InterferenceQuery.
1285 // Interference that overlaps an instruction is counted in both gaps
1286 // surrounding the instruction. The exception is interference before
1287 // StartIdx and after StopIdx.
1289 LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
1290 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
1291 // Skip the gaps before IntI.
1292 while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
1293 if (++Gap == NumGaps)
1298 // Update the gaps covered by IntI.
1299 const float weight = IntI.value()->weight;
1300 for (; Gap != NumGaps; ++Gap) {
1301 GapWeight[Gap] = std::max(GapWeight[Gap], weight);
1302 if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
1311 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
1314 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1315 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1316 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1317 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1319 // Note that it is possible to have an interval that is live-in or live-out
1320 // while only covering a single block - A phi-def can use undef values from
1321 // predecessors, and the block could be a single-block loop.
1322 // We don't bother doing anything clever about such a case, we simply assume
1323 // that the interval is continuous from FirstInstr to LastInstr. We should
1324 // make sure that we don't do anything illegal to such an interval, though.
1326 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
1327 if (Uses.size() <= 2)
1329 const unsigned NumGaps = Uses.size()-1;
1332 dbgs() << "tryLocalSplit: ";
1333 for (unsigned i = 0, e = Uses.size(); i != e; ++i)
1334 dbgs() << ' ' << SA->UseSlots[i];
1338 // Since we allow local split results to be split again, there is a risk of
1339 // creating infinite loops. It is tempting to require that the new live
1340 // ranges have less instructions than the original. That would guarantee
1341 // convergence, but it is too strict. A live range with 3 instructions can be
1342 // split 2+3 (including the COPY), and we want to allow that.
1344 // Instead we use these rules:
1346 // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the
1347 // noop split, of course).
1348 // 2. Require progress be made for ranges with getStage() == RS_Split2. All
1349 // the new ranges must have fewer instructions than before the split.
1350 // 3. New ranges with the same number of instructions are marked RS_Split2,
1351 // smaller ranges are marked RS_New.
1353 // These rules allow a 3 -> 2+3 split once, which we need. They also prevent
1354 // excessive splitting and infinite loops.
1356 bool ProgressRequired = getStage(VirtReg) >= RS_Split2;
1358 // Best split candidate.
1359 unsigned BestBefore = NumGaps;
1360 unsigned BestAfter = 0;
1363 const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
1364 SmallVector<float, 8> GapWeight;
1367 while (unsigned PhysReg = Order.next()) {
1368 // Keep track of the largest spill weight that would need to be evicted in
1369 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
1370 calcGapWeights(PhysReg, GapWeight);
1372 // Try to find the best sequence of gaps to close.
1373 // The new spill weight must be larger than any gap interference.
1375 // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
1376 unsigned SplitBefore = 0, SplitAfter = 1;
1378 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
1379 // It is the spill weight that needs to be evicted.
1380 float MaxGap = GapWeight[0];
1383 // Live before/after split?
1384 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
1385 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
1387 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
1388 << Uses[SplitBefore] << '-' << Uses[SplitAfter]
1389 << " i=" << MaxGap);
1391 // Stop before the interval gets so big we wouldn't be making progress.
1392 if (!LiveBefore && !LiveAfter) {
1393 DEBUG(dbgs() << " all\n");
1396 // Should the interval be extended or shrunk?
1399 // How many gaps would the new range have?
1400 unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter;
1402 // Legally, without causing looping?
1403 bool Legal = !ProgressRequired || NewGaps < NumGaps;
1405 if (Legal && MaxGap < HUGE_VALF) {
1406 // Estimate the new spill weight. Each instruction reads or writes the
1407 // register. Conservatively assume there are no read-modify-write
1410 // Try to guess the size of the new interval.
1411 const float EstWeight = normalizeSpillWeight(blockFreq * (NewGaps + 1),
1412 Uses[SplitBefore].distance(Uses[SplitAfter]) +
1413 (LiveBefore + LiveAfter)*SlotIndex::InstrDist);
1414 // Would this split be possible to allocate?
1415 // Never allocate all gaps, we wouldn't be making progress.
1416 DEBUG(dbgs() << " w=" << EstWeight);
1417 if (EstWeight * Hysteresis >= MaxGap) {
1419 float Diff = EstWeight - MaxGap;
1420 if (Diff > BestDiff) {
1421 DEBUG(dbgs() << " (best)");
1422 BestDiff = Hysteresis * Diff;
1423 BestBefore = SplitBefore;
1424 BestAfter = SplitAfter;
1431 if (++SplitBefore < SplitAfter) {
1432 DEBUG(dbgs() << " shrink\n");
1433 // Recompute the max when necessary.
1434 if (GapWeight[SplitBefore - 1] >= MaxGap) {
1435 MaxGap = GapWeight[SplitBefore];
1436 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
1437 MaxGap = std::max(MaxGap, GapWeight[i]);
1444 // Try to extend the interval.
1445 if (SplitAfter >= NumGaps) {
1446 DEBUG(dbgs() << " end\n");
1450 DEBUG(dbgs() << " extend\n");
1451 MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]);
1455 // Didn't find any candidates?
1456 if (BestBefore == NumGaps)
1459 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
1460 << '-' << Uses[BestAfter] << ", " << BestDiff
1461 << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
1463 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1467 SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
1468 SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]);
1469 SE->useIntv(SegStart, SegStop);
1470 SmallVector<unsigned, 8> IntvMap;
1471 SE->finish(&IntvMap);
1472 DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
1474 // If the new range has the same number of instructions as before, mark it as
1475 // RS_Split2 so the next split will be forced to make progress. Otherwise,
1476 // leave the new intervals as RS_New so they can compete.
1477 bool LiveBefore = BestBefore != 0 || BI.LiveIn;
1478 bool LiveAfter = BestAfter != NumGaps || BI.LiveOut;
1479 unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter;
1480 if (NewGaps >= NumGaps) {
1481 DEBUG(dbgs() << "Tagging non-progress ranges: ");
1482 assert(!ProgressRequired && "Didn't make progress when it was required.");
1483 for (unsigned i = 0, e = IntvMap.size(); i != e; ++i)
1484 if (IntvMap[i] == 1) {
1485 setStage(*LREdit.get(i), RS_Split2);
1486 DEBUG(dbgs() << PrintReg(LREdit.get(i)->reg));
1488 DEBUG(dbgs() << '\n');
1495 //===----------------------------------------------------------------------===//
1496 // Live Range Splitting
1497 //===----------------------------------------------------------------------===//
1499 /// trySplit - Try to split VirtReg or one of its interferences, making it
1501 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
1502 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
1503 SmallVectorImpl<LiveInterval*>&NewVRegs) {
1504 // Ranges must be Split2 or less.
1505 if (getStage(VirtReg) >= RS_Spill)
1508 // Local intervals are handled separately.
1509 if (LIS->intervalIsInOneMBB(VirtReg)) {
1510 NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
1511 SA->analyze(&VirtReg);
1512 return tryLocalSplit(VirtReg, Order, NewVRegs);
1515 NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
1517 SA->analyze(&VirtReg);
1519 // FIXME: SplitAnalysis may repair broken live ranges coming from the
1520 // coalescer. That may cause the range to become allocatable which means that
1521 // tryRegionSplit won't be making progress. This check should be replaced with
1522 // an assertion when the coalescer is fixed.
1523 if (SA->didRepairRange()) {
1524 // VirtReg has changed, so all cached queries are invalid.
1525 invalidateVirtRegs();
1526 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1530 // First try to split around a region spanning multiple blocks. RS_Split2
1531 // ranges already made dubious progress with region splitting, so they go
1532 // straight to single block splitting.
1533 if (getStage(VirtReg) < RS_Split2) {
1534 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
1535 if (PhysReg || !NewVRegs.empty())
1539 // Then isolate blocks.
1540 return tryBlockSplit(VirtReg, Order, NewVRegs);
1544 //===----------------------------------------------------------------------===//
1546 //===----------------------------------------------------------------------===//
1548 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
1549 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1550 // First try assigning a free register.
1551 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
1552 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1555 LiveRangeStage Stage = getStage(VirtReg);
1556 DEBUG(dbgs() << StageName[Stage]
1557 << " Cascade " << ExtraRegInfo[VirtReg.reg].Cascade << '\n');
1559 // Try to evict a less worthy live range, but only for ranges from the primary
1560 // queue. The RS_Split ranges already failed to do this, and they should not
1561 // get a second chance until they have been split.
1562 if (Stage != RS_Split)
1563 if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
1566 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
1568 // The first time we see a live range, don't try to split or spill.
1569 // Wait until the second time, when all smaller ranges have been allocated.
1570 // This gives a better picture of the interference to split around.
1571 if (Stage < RS_Split) {
1572 setStage(VirtReg, RS_Split);
1573 DEBUG(dbgs() << "wait for second round\n");
1574 NewVRegs.push_back(&VirtReg);
1578 // If we couldn't allocate a register from spilling, there is probably some
1579 // invalid inline assembly. The base class wil report it.
1580 if (Stage >= RS_Done || !VirtReg.isSpillable())
1583 // Try splitting VirtReg or interferences.
1584 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
1585 if (PhysReg || !NewVRegs.empty())
1588 // Finally spill VirtReg itself.
1589 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
1590 LiveRangeEdit LRE(VirtReg, NewVRegs, this);
1591 spiller().spill(LRE);
1592 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done);
1595 MF->verify(this, "After spilling");
1597 // The live virtual register requesting allocation was spilled, so tell
1598 // the caller not to allocate anything during this round.
1602 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
1603 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
1604 << "********** Function: "
1605 << ((Value*)mf.getFunction())->getName() << '\n');
1609 MF->verify(this, "Before greedy register allocator");
1611 RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
1612 Indexes = &getAnalysis<SlotIndexes>();
1613 DomTree = &getAnalysis<MachineDominatorTree>();
1614 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
1615 Loops = &getAnalysis<MachineLoopInfo>();
1616 Bundles = &getAnalysis<EdgeBundles>();
1617 SpillPlacer = &getAnalysis<SpillPlacement>();
1618 DebugVars = &getAnalysis<LiveDebugVariables>();
1620 SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
1621 SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
1622 ExtraRegInfo.clear();
1623 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1625 IntfCache.init(MF, &PhysReg2LiveUnion[0], Indexes, TRI);
1626 GlobalCand.resize(32); // This will grow as needed.
1630 LIS->addKillFlags();
1634 NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
1635 VRM->rewrite(Indexes);
1638 // Write out new DBG_VALUE instructions.
1640 NamedRegionTimer T("Emit Debug Info", TimerGroupName, TimePassesIsEnabled);
1641 DebugVars->emitDebugValues(VRM);
1644 // The pass output is in VirtRegMap. Release all the transient data.