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");
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 tryLocalSplit(LiveInterval&, AllocationOrder&,
270 SmallVectorImpl<LiveInterval*>&);
271 unsigned trySplit(LiveInterval&, AllocationOrder&,
272 SmallVectorImpl<LiveInterval*>&);
274 } // end anonymous namespace
276 char RAGreedy::ID = 0;
279 const char *const RAGreedy::StageName[] = {
289 // Hysteresis to use when comparing floats.
290 // This helps stabilize decisions based on float comparisons.
291 const float Hysteresis = 0.98f;
294 FunctionPass* llvm::createGreedyRegisterAllocator() {
295 return new RAGreedy();
298 RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
299 initializeLiveDebugVariablesPass(*PassRegistry::getPassRegistry());
300 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
301 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
302 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
303 initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
304 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
305 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
306 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
307 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
308 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
309 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
310 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
311 initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
314 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
315 AU.setPreservesCFG();
316 AU.addRequired<AliasAnalysis>();
317 AU.addPreserved<AliasAnalysis>();
318 AU.addRequired<LiveIntervals>();
319 AU.addRequired<SlotIndexes>();
320 AU.addPreserved<SlotIndexes>();
321 AU.addRequired<LiveDebugVariables>();
322 AU.addPreserved<LiveDebugVariables>();
324 AU.addRequiredID(StrongPHIEliminationID);
325 AU.addRequiredTransitive<RegisterCoalescer>();
326 AU.addRequired<CalculateSpillWeights>();
327 AU.addRequired<LiveStacks>();
328 AU.addPreserved<LiveStacks>();
329 AU.addRequired<MachineDominatorTree>();
330 AU.addPreserved<MachineDominatorTree>();
331 AU.addRequired<MachineLoopInfo>();
332 AU.addPreserved<MachineLoopInfo>();
333 AU.addRequired<VirtRegMap>();
334 AU.addPreserved<VirtRegMap>();
335 AU.addRequired<EdgeBundles>();
336 AU.addRequired<SpillPlacement>();
337 MachineFunctionPass::getAnalysisUsage(AU);
341 //===----------------------------------------------------------------------===//
342 // LiveRangeEdit delegate methods
343 //===----------------------------------------------------------------------===//
345 void RAGreedy::LRE_WillEraseInstruction(MachineInstr *MI) {
346 // LRE itself will remove from SlotIndexes and parent basic block.
347 VRM->RemoveMachineInstrFromMaps(MI);
350 bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
351 if (unsigned PhysReg = VRM->getPhys(VirtReg)) {
352 unassign(LIS->getInterval(VirtReg), PhysReg);
355 // Unassigned virtreg is probably in the priority queue.
356 // RegAllocBase will erase it after dequeueing.
360 void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
361 unsigned PhysReg = VRM->getPhys(VirtReg);
365 // Register is assigned, put it back on the queue for reassignment.
366 LiveInterval &LI = LIS->getInterval(VirtReg);
367 unassign(LI, PhysReg);
371 void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
372 // LRE may clone a virtual register because dead code elimination causes it to
373 // be split into connected components. The new components are much smaller
374 // than the original, so they should get a new chance at being assigned.
375 // same stage as the parent.
376 ExtraRegInfo[Old].Stage = RS_Assign;
377 ExtraRegInfo.grow(New);
378 ExtraRegInfo[New] = ExtraRegInfo[Old];
381 void RAGreedy::releaseMemory() {
382 SpillerInstance.reset(0);
383 ExtraRegInfo.clear();
385 RegAllocBase::releaseMemory();
388 void RAGreedy::enqueue(LiveInterval *LI) {
389 // Prioritize live ranges by size, assigning larger ranges first.
390 // The queue holds (size, reg) pairs.
391 const unsigned Size = LI->getSize();
392 const unsigned Reg = LI->reg;
393 assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
394 "Can only enqueue virtual registers");
397 ExtraRegInfo.grow(Reg);
398 if (ExtraRegInfo[Reg].Stage == RS_New)
399 ExtraRegInfo[Reg].Stage = RS_Assign;
401 if (ExtraRegInfo[Reg].Stage == RS_Split)
402 // Unsplit ranges that couldn't be allocated immediately are deferred until
403 // everything else has been allocated. Long ranges are allocated last so
404 // they are split against realistic interference.
405 Prio = (1u << 31) - Size;
407 // Everything else is allocated in long->short order. Long ranges that don't
408 // fit should be spilled ASAP so they don't create interference.
409 Prio = (1u << 31) + Size;
411 // Boost ranges that have a physical register hint.
412 if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
416 Queue.push(std::make_pair(Prio, Reg));
419 LiveInterval *RAGreedy::dequeue() {
422 LiveInterval *LI = &LIS->getInterval(Queue.top().second);
428 //===----------------------------------------------------------------------===//
430 //===----------------------------------------------------------------------===//
432 /// tryAssign - Try to assign VirtReg to an available register.
433 unsigned RAGreedy::tryAssign(LiveInterval &VirtReg,
434 AllocationOrder &Order,
435 SmallVectorImpl<LiveInterval*> &NewVRegs) {
438 while ((PhysReg = Order.next()))
439 if (!checkPhysRegInterference(VirtReg, PhysReg))
441 if (!PhysReg || Order.isHint(PhysReg))
444 // PhysReg is available, but there may be a better choice.
446 // If we missed a simple hint, try to cheaply evict interference from the
447 // preferred register.
448 if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg))
449 if (Order.isHint(Hint)) {
450 DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n');
451 EvictionCost MaxCost(1);
452 if (canEvictInterference(VirtReg, Hint, true, MaxCost)) {
453 evictInterference(VirtReg, Hint, NewVRegs);
458 // Try to evict interference from a cheaper alternative.
459 unsigned Cost = TRI->getCostPerUse(PhysReg);
461 // Most registers have 0 additional cost.
465 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " is available at cost " << Cost
467 unsigned CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost);
468 return CheapReg ? CheapReg : PhysReg;
472 //===----------------------------------------------------------------------===//
473 // Interference eviction
474 //===----------------------------------------------------------------------===//
476 /// shouldEvict - determine if A should evict the assigned live range B. The
477 /// eviction policy defined by this function together with the allocation order
478 /// defined by enqueue() decides which registers ultimately end up being split
481 /// Cascade numbers are used to prevent infinite loops if this function is a
484 /// @param A The live range to be assigned.
485 /// @param IsHint True when A is about to be assigned to its preferred
487 /// @param B The live range to be evicted.
488 /// @param BreaksHint True when B is already assigned to its preferred register.
489 bool RAGreedy::shouldEvict(LiveInterval &A, bool IsHint,
490 LiveInterval &B, bool BreaksHint) {
491 bool CanSplit = getStage(B) < RS_Spill;
493 // Be fairly aggressive about following hints as long as the evictee can be
495 if (CanSplit && IsHint && !BreaksHint)
498 return A.weight > B.weight;
501 /// canEvictInterference - Return true if all interferences between VirtReg and
502 /// PhysReg can be evicted. When OnlyCheap is set, don't do anything
504 /// @param VirtReg Live range that is about to be assigned.
505 /// @param PhysReg Desired register for assignment.
506 /// @prarm IsHint True when PhysReg is VirtReg's preferred register.
507 /// @param MaxCost Only look for cheaper candidates and update with new cost
508 /// when returning true.
509 /// @returns True when interference can be evicted cheaper than MaxCost.
510 bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
511 bool IsHint, EvictionCost &MaxCost) {
512 // Find VirtReg's cascade number. This will be unassigned if VirtReg was never
513 // involved in an eviction before. If a cascade number was assigned, deny
514 // evicting anything with the same or a newer cascade number. This prevents
515 // infinite eviction loops.
517 // This works out so a register without a cascade number is allowed to evict
518 // anything, and it can be evicted by anything.
519 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
521 Cascade = NextCascade;
524 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
525 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
526 // If there is 10 or more interferences, chances are one is heavier.
527 if (Q.collectInterferingVRegs(10) >= 10)
530 // Check if any interfering live range is heavier than MaxWeight.
531 for (unsigned i = Q.interferingVRegs().size(); i; --i) {
532 LiveInterval *Intf = Q.interferingVRegs()[i - 1];
533 if (TargetRegisterInfo::isPhysicalRegister(Intf->reg))
535 // Never evict spill products. They cannot split or spill.
536 if (getStage(*Intf) == RS_Done)
538 // Once a live range becomes small enough, it is urgent that we find a
539 // register for it. This is indicated by an infinite spill weight. These
540 // urgent live ranges get to evict almost anything.
541 bool Urgent = !VirtReg.isSpillable() && Intf->isSpillable();
542 // Only evict older cascades or live ranges without a cascade.
543 unsigned IntfCascade = ExtraRegInfo[Intf->reg].Cascade;
544 if (Cascade <= IntfCascade) {
547 // We permit breaking cascades for urgent evictions. It should be the
548 // last resort, though, so make it really expensive.
549 Cost.BrokenHints += 10;
551 // Would this break a satisfied hint?
552 bool BreaksHint = VRM->hasPreferredPhys(Intf->reg);
553 // Update eviction cost.
554 Cost.BrokenHints += BreaksHint;
555 Cost.MaxWeight = std::max(Cost.MaxWeight, Intf->weight);
556 // Abort if this would be too expensive.
557 if (!(Cost < MaxCost))
559 // Finally, apply the eviction policy for non-urgent evictions.
560 if (!Urgent && !shouldEvict(VirtReg, IsHint, *Intf, BreaksHint))
568 /// evictInterference - Evict any interferring registers that prevent VirtReg
569 /// from being assigned to Physreg. This assumes that canEvictInterference
571 void RAGreedy::evictInterference(LiveInterval &VirtReg, unsigned PhysReg,
572 SmallVectorImpl<LiveInterval*> &NewVRegs) {
573 // Make sure that VirtReg has a cascade number, and assign that cascade
574 // number to every evicted register. These live ranges than then only be
575 // evicted by a newer cascade, preventing infinite loops.
576 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
578 Cascade = ExtraRegInfo[VirtReg.reg].Cascade = NextCascade++;
580 DEBUG(dbgs() << "evicting " << PrintReg(PhysReg, TRI)
581 << " interference: Cascade " << Cascade << '\n');
582 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
583 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
584 assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
585 for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
586 LiveInterval *Intf = Q.interferingVRegs()[i];
587 unassign(*Intf, VRM->getPhys(Intf->reg));
588 assert((ExtraRegInfo[Intf->reg].Cascade < Cascade ||
589 VirtReg.isSpillable() < Intf->isSpillable()) &&
590 "Cannot decrease cascade number, illegal eviction");
591 ExtraRegInfo[Intf->reg].Cascade = Cascade;
593 NewVRegs.push_back(Intf);
598 /// tryEvict - Try to evict all interferences for a physreg.
599 /// @param VirtReg Currently unassigned virtual register.
600 /// @param Order Physregs to try.
601 /// @return Physreg to assign VirtReg, or 0.
602 unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
603 AllocationOrder &Order,
604 SmallVectorImpl<LiveInterval*> &NewVRegs,
605 unsigned CostPerUseLimit) {
606 NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
608 // Keep track of the cheapest interference seen so far.
609 EvictionCost BestCost(~0u);
610 unsigned BestPhys = 0;
612 // When we are just looking for a reduced cost per use, don't break any
613 // hints, and only evict smaller spill weights.
614 if (CostPerUseLimit < ~0u) {
615 BestCost.BrokenHints = 0;
616 BestCost.MaxWeight = VirtReg.weight;
620 while (unsigned PhysReg = Order.next()) {
621 if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
623 // The first use of a callee-saved register in a function has cost 1.
624 // Don't start using a CSR when the CostPerUseLimit is low.
625 if (CostPerUseLimit == 1)
626 if (unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg))
627 if (!MRI->isPhysRegUsed(CSR)) {
628 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " would clobber CSR "
629 << PrintReg(CSR, TRI) << '\n');
633 if (!canEvictInterference(VirtReg, PhysReg, false, BestCost))
639 // Stop if the hint can be used.
640 if (Order.isHint(PhysReg))
647 evictInterference(VirtReg, BestPhys, NewVRegs);
652 //===----------------------------------------------------------------------===//
654 //===----------------------------------------------------------------------===//
656 /// addSplitConstraints - Fill out the SplitConstraints vector based on the
657 /// interference pattern in Physreg and its aliases. Add the constraints to
658 /// SpillPlacement and return the static cost of this split in Cost, assuming
659 /// that all preferences in SplitConstraints are met.
660 /// Return false if there are no bundles with positive bias.
661 bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf,
663 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
665 // Reset interference dependent info.
666 SplitConstraints.resize(UseBlocks.size());
667 float StaticCost = 0;
668 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
669 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
670 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
672 BC.Number = BI.MBB->getNumber();
673 Intf.moveToBlock(BC.Number);
674 BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
675 BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
677 if (!Intf.hasInterference())
680 // Number of spill code instructions to insert.
683 // Interference for the live-in value.
685 if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
686 BC.Entry = SpillPlacement::MustSpill, ++Ins;
687 else if (Intf.first() < BI.FirstUse)
688 BC.Entry = SpillPlacement::PrefSpill, ++Ins;
689 else if (Intf.first() < BI.LastUse)
693 // Interference for the live-out value.
695 if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
696 BC.Exit = SpillPlacement::MustSpill, ++Ins;
697 else if (Intf.last() > BI.LastUse)
698 BC.Exit = SpillPlacement::PrefSpill, ++Ins;
699 else if (Intf.last() > BI.FirstUse)
703 // Accumulate the total frequency of inserted spill code.
705 StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
709 // Add constraints for use-blocks. Note that these are the only constraints
710 // that may add a positive bias, it is downhill from here.
711 SpillPlacer->addConstraints(SplitConstraints);
712 return SpillPlacer->scanActiveBundles();
716 /// addThroughConstraints - Add constraints and links to SpillPlacer from the
717 /// live-through blocks in Blocks.
718 void RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf,
719 ArrayRef<unsigned> Blocks) {
720 const unsigned GroupSize = 8;
721 SpillPlacement::BlockConstraint BCS[GroupSize];
722 unsigned TBS[GroupSize];
723 unsigned B = 0, T = 0;
725 for (unsigned i = 0; i != Blocks.size(); ++i) {
726 unsigned Number = Blocks[i];
727 Intf.moveToBlock(Number);
729 if (!Intf.hasInterference()) {
730 assert(T < GroupSize && "Array overflow");
732 if (++T == GroupSize) {
733 SpillPlacer->addLinks(makeArrayRef(TBS, T));
739 assert(B < GroupSize && "Array overflow");
740 BCS[B].Number = Number;
742 // Interference for the live-in value.
743 if (Intf.first() <= Indexes->getMBBStartIdx(Number))
744 BCS[B].Entry = SpillPlacement::MustSpill;
746 BCS[B].Entry = SpillPlacement::PrefSpill;
748 // Interference for the live-out value.
749 if (Intf.last() >= SA->getLastSplitPoint(Number))
750 BCS[B].Exit = SpillPlacement::MustSpill;
752 BCS[B].Exit = SpillPlacement::PrefSpill;
754 if (++B == GroupSize) {
755 ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
756 SpillPlacer->addConstraints(Array);
761 ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
762 SpillPlacer->addConstraints(Array);
763 SpillPlacer->addLinks(makeArrayRef(TBS, T));
766 void RAGreedy::growRegion(GlobalSplitCandidate &Cand) {
767 // Keep track of through blocks that have not been added to SpillPlacer.
768 BitVector Todo = SA->getThroughBlocks();
769 SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks;
770 unsigned AddedTo = 0;
772 unsigned Visited = 0;
776 ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive();
777 // Find new through blocks in the periphery of PrefRegBundles.
778 for (int i = 0, e = NewBundles.size(); i != e; ++i) {
779 unsigned Bundle = NewBundles[i];
780 // Look at all blocks connected to Bundle in the full graph.
781 ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle);
782 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
785 if (!Todo.test(Block))
788 // This is a new through block. Add it to SpillPlacer later.
789 ActiveBlocks.push_back(Block);
795 // Any new blocks to add?
796 if (ActiveBlocks.size() == AddedTo)
799 // Compute through constraints from the interference, or assume that all
800 // through blocks prefer spilling when forming compact regions.
801 ArrayRef<unsigned> NewBlocks = makeArrayRef(ActiveBlocks).slice(AddedTo);
803 addThroughConstraints(Cand.Intf, NewBlocks);
805 SpillPlacer->addPrefSpill(NewBlocks);
806 AddedTo = ActiveBlocks.size();
808 // Perhaps iterating can enable more bundles?
809 SpillPlacer->iterate();
811 DEBUG(dbgs() << ", v=" << Visited);
814 /// calcCompactRegion - Compute the set of edge bundles that should be live
815 /// when splitting the current live range into compact regions. Compact
816 /// regions can be computed without looking at interference. They are the
817 /// regions formed by removing all the live-through blocks from the live range.
819 /// Returns false if the current live range is already compact, or if the
820 /// compact regions would form single block regions anyway.
821 bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) {
822 // Without any through blocks, the live range is already compact.
823 if (!SA->getNumThroughBlocks())
826 // Compact regions don't correspond to any physreg.
827 Cand.reset(IntfCache, 0);
829 DEBUG(dbgs() << "Compact region bundles");
831 // Use the spill placer to determine the live bundles. GrowRegion pretends
832 // that all the through blocks have interference when PhysReg is unset.
833 SpillPlacer->prepare(Cand.LiveBundles);
835 // The static split cost will be zero since Cand.Intf reports no interference.
837 if (!addSplitConstraints(Cand.Intf, Cost)) {
838 DEBUG(dbgs() << ", none.\n");
843 SpillPlacer->finish();
845 if (!Cand.LiveBundles.any()) {
846 DEBUG(dbgs() << ", none.\n");
851 for (int i = Cand.LiveBundles.find_first(); i>=0;
852 i = Cand.LiveBundles.find_next(i))
853 dbgs() << " EB#" << i;
859 /// calcSpillCost - Compute how expensive it would be to split the live range in
860 /// SA around all use blocks instead of forming bundle regions.
861 float RAGreedy::calcSpillCost() {
863 const LiveInterval &LI = SA->getParent();
864 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
865 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
866 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
867 unsigned Number = BI.MBB->getNumber();
868 // We normally only need one spill instruction - a load or a store.
869 Cost += SpillPlacer->getBlockFrequency(Number);
871 // Unless the value is redefined in the block.
872 if (BI.LiveIn && BI.LiveOut) {
873 SlotIndex Start, Stop;
874 tie(Start, Stop) = Indexes->getMBBRange(Number);
875 LiveInterval::const_iterator I = LI.find(Start);
876 assert(I != LI.end() && "Expected live-in value");
877 // Is there a different live-out value? If so, we need an extra spill
880 Cost += SpillPlacer->getBlockFrequency(Number);
886 /// calcGlobalSplitCost - Return the global split cost of following the split
887 /// pattern in LiveBundles. This cost should be added to the local cost of the
888 /// interference pattern in SplitConstraints.
890 float RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) {
891 float GlobalCost = 0;
892 const BitVector &LiveBundles = Cand.LiveBundles;
893 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
894 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
895 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
896 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
897 bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)];
898 bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
902 Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
904 Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
906 GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
909 for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
910 unsigned Number = Cand.ActiveBlocks[i];
911 bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)];
912 bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)];
913 if (!RegIn && !RegOut)
915 if (RegIn && RegOut) {
916 // We need double spill code if this block has interference.
917 Cand.Intf.moveToBlock(Number);
918 if (Cand.Intf.hasInterference())
919 GlobalCost += 2*SpillPlacer->getBlockFrequency(Number);
922 // live-in / stack-out or stack-in live-out.
923 GlobalCost += SpillPlacer->getBlockFrequency(Number);
928 /// splitAroundRegion - Split the current live range around the regions
929 /// determined by BundleCand and GlobalCand.
931 /// Before calling this function, GlobalCand and BundleCand must be initialized
932 /// so each bundle is assigned to a valid candidate, or NoCand for the
933 /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor
934 /// objects must be initialized for the current live range, and intervals
935 /// created for the used candidates.
937 /// @param LREdit The LiveRangeEdit object handling the current split.
938 /// @param UsedCands List of used GlobalCand entries. Every BundleCand value
939 /// must appear in this list.
940 void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit,
941 ArrayRef<unsigned> UsedCands) {
942 // These are the intervals created for new global ranges. We may create more
943 // intervals for local ranges.
944 const unsigned NumGlobalIntvs = LREdit.size();
945 DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs << " globals.\n");
946 assert(NumGlobalIntvs && "No global intervals configured");
948 // First handle all the blocks with uses.
949 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
950 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
951 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
952 unsigned Number = BI.MBB->getNumber();
953 unsigned IntvIn = 0, IntvOut = 0;
954 SlotIndex IntfIn, IntfOut;
956 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
957 if (CandIn != NoCand) {
958 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
959 IntvIn = Cand.IntvIdx;
960 Cand.Intf.moveToBlock(Number);
961 IntfIn = Cand.Intf.first();
965 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
966 if (CandOut != NoCand) {
967 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
968 IntvOut = Cand.IntvIdx;
969 Cand.Intf.moveToBlock(Number);
970 IntfOut = Cand.Intf.last();
974 // Create separate intervals for isolated blocks with multiple uses.
975 if (!IntvIn && !IntvOut) {
976 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
977 if (!BI.isOneInstr())
978 SE->splitSingleBlock(BI);
982 if (IntvIn && IntvOut)
983 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
985 SE->splitRegInBlock(BI, IntvIn, IntfIn);
987 SE->splitRegOutBlock(BI, IntvOut, IntfOut);
990 // Handle live-through blocks. The relevant live-through blocks are stored in
991 // the ActiveBlocks list with each candidate. We need to filter out
993 BitVector Todo = SA->getThroughBlocks();
994 for (unsigned c = 0; c != UsedCands.size(); ++c) {
995 ArrayRef<unsigned> Blocks = GlobalCand[UsedCands[c]].ActiveBlocks;
996 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
997 unsigned Number = Blocks[i];
998 if (!Todo.test(Number))
1002 unsigned IntvIn = 0, IntvOut = 0;
1003 SlotIndex IntfIn, IntfOut;
1005 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
1006 if (CandIn != NoCand) {
1007 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
1008 IntvIn = Cand.IntvIdx;
1009 Cand.Intf.moveToBlock(Number);
1010 IntfIn = Cand.Intf.first();
1013 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
1014 if (CandOut != NoCand) {
1015 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
1016 IntvOut = Cand.IntvIdx;
1017 Cand.Intf.moveToBlock(Number);
1018 IntfOut = Cand.Intf.last();
1020 if (!IntvIn && !IntvOut)
1022 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
1028 SmallVector<unsigned, 8> IntvMap;
1029 SE->finish(&IntvMap);
1030 DebugVars->splitRegister(SA->getParent().reg, LREdit.regs());
1032 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1033 unsigned OrigBlocks = SA->getNumLiveBlocks();
1035 // Sort out the new intervals created by splitting. We get four kinds:
1036 // - Remainder intervals should not be split again.
1037 // - Candidate intervals can be assigned to Cand.PhysReg.
1038 // - Block-local splits are candidates for local splitting.
1039 // - DCE leftovers should go back on the queue.
1040 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1041 LiveInterval &Reg = *LREdit.get(i);
1043 // Ignore old intervals from DCE.
1044 if (getStage(Reg) != RS_New)
1047 // Remainder interval. Don't try splitting again, spill if it doesn't
1049 if (IntvMap[i] == 0) {
1050 setStage(Reg, RS_Spill);
1054 // Global intervals. Allow repeated splitting as long as the number of live
1055 // blocks is strictly decreasing.
1056 if (IntvMap[i] < NumGlobalIntvs) {
1057 if (SA->countLiveBlocks(&Reg) >= OrigBlocks) {
1058 DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks
1059 << " blocks as original.\n");
1060 // Don't allow repeated splitting as a safe guard against looping.
1061 setStage(Reg, RS_Split2);
1066 // Other intervals are treated as new. This includes local intervals created
1067 // for blocks with multiple uses, and anything created by DCE.
1071 MF->verify(this, "After splitting live range around region");
1074 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1075 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1076 unsigned NumCands = 0;
1077 unsigned BestCand = NoCand;
1079 SmallVector<unsigned, 8> UsedCands;
1081 // Check if we can split this live range around a compact region.
1082 bool HasCompact = CompactRegions && calcCompactRegion(GlobalCand.front());
1084 // Yes, keep GlobalCand[0] as the compact region candidate.
1086 BestCost = HUGE_VALF;
1088 // No benefit from the compact region, our fallback will be per-block
1089 // splitting. Make sure we find a solution that is cheaper than spilling.
1090 BestCost = Hysteresis * calcSpillCost();
1091 DEBUG(dbgs() << "Cost of isolating all blocks = " << BestCost << '\n');
1095 while (unsigned PhysReg = Order.next()) {
1096 // Discard bad candidates before we run out of interference cache cursors.
1097 // This will only affect register classes with a lot of registers (>32).
1098 if (NumCands == IntfCache.getMaxCursors()) {
1099 unsigned WorstCount = ~0u;
1101 for (unsigned i = 0; i != NumCands; ++i) {
1102 if (i == BestCand || !GlobalCand[i].PhysReg)
1104 unsigned Count = GlobalCand[i].LiveBundles.count();
1105 if (Count < WorstCount)
1106 Worst = i, WorstCount = Count;
1109 GlobalCand[Worst] = GlobalCand[NumCands];
1112 if (GlobalCand.size() <= NumCands)
1113 GlobalCand.resize(NumCands+1);
1114 GlobalSplitCandidate &Cand = GlobalCand[NumCands];
1115 Cand.reset(IntfCache, PhysReg);
1117 SpillPlacer->prepare(Cand.LiveBundles);
1119 if (!addSplitConstraints(Cand.Intf, Cost)) {
1120 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n");
1123 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
1124 if (Cost >= BestCost) {
1126 if (BestCand == NoCand)
1127 dbgs() << " worse than no bundles\n";
1129 dbgs() << " worse than "
1130 << PrintReg(GlobalCand[BestCand].PhysReg, TRI) << '\n';
1136 SpillPlacer->finish();
1138 // No live bundles, defer to splitSingleBlocks().
1139 if (!Cand.LiveBundles.any()) {
1140 DEBUG(dbgs() << " no bundles.\n");
1144 Cost += calcGlobalSplitCost(Cand);
1146 dbgs() << ", total = " << Cost << " with bundles";
1147 for (int i = Cand.LiveBundles.find_first(); i>=0;
1148 i = Cand.LiveBundles.find_next(i))
1149 dbgs() << " EB#" << i;
1152 if (Cost < BestCost) {
1153 BestCand = NumCands;
1154 BestCost = Hysteresis * Cost; // Prevent rounding effects.
1159 // No solutions found, fall back to single block splitting.
1160 if (!HasCompact && BestCand == NoCand)
1163 // Prepare split editor.
1164 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1167 // Assign all edge bundles to the preferred candidate, or NoCand.
1168 BundleCand.assign(Bundles->getNumBundles(), NoCand);
1170 // Assign bundles for the best candidate region.
1171 if (BestCand != NoCand) {
1172 GlobalSplitCandidate &Cand = GlobalCand[BestCand];
1173 if (unsigned B = Cand.getBundles(BundleCand, BestCand)) {
1174 UsedCands.push_back(BestCand);
1175 Cand.IntvIdx = SE->openIntv();
1176 DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in "
1177 << B << " bundles, intv " << Cand.IntvIdx << ".\n");
1181 // Assign bundles for the compact region.
1183 GlobalSplitCandidate &Cand = GlobalCand.front();
1184 assert(!Cand.PhysReg && "Compact region has no physreg");
1185 if (unsigned B = Cand.getBundles(BundleCand, 0)) {
1186 UsedCands.push_back(0);
1187 Cand.IntvIdx = SE->openIntv();
1188 DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv "
1189 << Cand.IntvIdx << ".\n");
1193 splitAroundRegion(LREdit, UsedCands);
1198 //===----------------------------------------------------------------------===//
1200 //===----------------------------------------------------------------------===//
1203 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted
1204 /// in order to use PhysReg between two entries in SA->UseSlots.
1206 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
1208 void RAGreedy::calcGapWeights(unsigned PhysReg,
1209 SmallVectorImpl<float> &GapWeight) {
1210 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1211 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1212 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
1213 const unsigned NumGaps = Uses.size()-1;
1215 // Start and end points for the interference check.
1216 SlotIndex StartIdx = BI.LiveIn ? BI.FirstUse.getBaseIndex() : BI.FirstUse;
1217 SlotIndex StopIdx = BI.LiveOut ? BI.LastUse.getBoundaryIndex() : BI.LastUse;
1219 GapWeight.assign(NumGaps, 0.0f);
1221 // Add interference from each overlapping register.
1222 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
1223 if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
1224 .checkInterference())
1227 // We know that VirtReg is a continuous interval from FirstUse to LastUse,
1228 // so we don't need InterferenceQuery.
1230 // Interference that overlaps an instruction is counted in both gaps
1231 // surrounding the instruction. The exception is interference before
1232 // StartIdx and after StopIdx.
1234 LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
1235 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
1236 // Skip the gaps before IntI.
1237 while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
1238 if (++Gap == NumGaps)
1243 // Update the gaps covered by IntI.
1244 const float weight = IntI.value()->weight;
1245 for (; Gap != NumGaps; ++Gap) {
1246 GapWeight[Gap] = std::max(GapWeight[Gap], weight);
1247 if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
1256 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
1259 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1260 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1261 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1262 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1264 // Note that it is possible to have an interval that is live-in or live-out
1265 // while only covering a single block - A phi-def can use undef values from
1266 // predecessors, and the block could be a single-block loop.
1267 // We don't bother doing anything clever about such a case, we simply assume
1268 // that the interval is continuous from FirstUse to LastUse. We should make
1269 // sure that we don't do anything illegal to such an interval, though.
1271 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
1272 if (Uses.size() <= 2)
1274 const unsigned NumGaps = Uses.size()-1;
1277 dbgs() << "tryLocalSplit: ";
1278 for (unsigned i = 0, e = Uses.size(); i != e; ++i)
1279 dbgs() << ' ' << SA->UseSlots[i];
1283 // Since we allow local split results to be split again, there is a risk of
1284 // creating infinite loops. It is tempting to require that the new live
1285 // ranges have less instructions than the original. That would guarantee
1286 // convergence, but it is too strict. A live range with 3 instructions can be
1287 // split 2+3 (including the COPY), and we want to allow that.
1289 // Instead we use these rules:
1291 // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the
1292 // noop split, of course).
1293 // 2. Require progress be made for ranges with getStage() == RS_Split2. All
1294 // the new ranges must have fewer instructions than before the split.
1295 // 3. New ranges with the same number of instructions are marked RS_Split2,
1296 // smaller ranges are marked RS_New.
1298 // These rules allow a 3 -> 2+3 split once, which we need. They also prevent
1299 // excessive splitting and infinite loops.
1301 bool ProgressRequired = getStage(VirtReg) >= RS_Split2;
1303 // Best split candidate.
1304 unsigned BestBefore = NumGaps;
1305 unsigned BestAfter = 0;
1308 const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
1309 SmallVector<float, 8> GapWeight;
1312 while (unsigned PhysReg = Order.next()) {
1313 // Keep track of the largest spill weight that would need to be evicted in
1314 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
1315 calcGapWeights(PhysReg, GapWeight);
1317 // Try to find the best sequence of gaps to close.
1318 // The new spill weight must be larger than any gap interference.
1320 // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
1321 unsigned SplitBefore = 0, SplitAfter = 1;
1323 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
1324 // It is the spill weight that needs to be evicted.
1325 float MaxGap = GapWeight[0];
1328 // Live before/after split?
1329 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
1330 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
1332 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
1333 << Uses[SplitBefore] << '-' << Uses[SplitAfter]
1334 << " i=" << MaxGap);
1336 // Stop before the interval gets so big we wouldn't be making progress.
1337 if (!LiveBefore && !LiveAfter) {
1338 DEBUG(dbgs() << " all\n");
1341 // Should the interval be extended or shrunk?
1344 // How many gaps would the new range have?
1345 unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter;
1347 // Legally, without causing looping?
1348 bool Legal = !ProgressRequired || NewGaps < NumGaps;
1350 if (Legal && MaxGap < HUGE_VALF) {
1351 // Estimate the new spill weight. Each instruction reads or writes the
1352 // register. Conservatively assume there are no read-modify-write
1355 // Try to guess the size of the new interval.
1356 const float EstWeight = normalizeSpillWeight(blockFreq * (NewGaps + 1),
1357 Uses[SplitBefore].distance(Uses[SplitAfter]) +
1358 (LiveBefore + LiveAfter)*SlotIndex::InstrDist);
1359 // Would this split be possible to allocate?
1360 // Never allocate all gaps, we wouldn't be making progress.
1361 DEBUG(dbgs() << " w=" << EstWeight);
1362 if (EstWeight * Hysteresis >= MaxGap) {
1364 float Diff = EstWeight - MaxGap;
1365 if (Diff > BestDiff) {
1366 DEBUG(dbgs() << " (best)");
1367 BestDiff = Hysteresis * Diff;
1368 BestBefore = SplitBefore;
1369 BestAfter = SplitAfter;
1376 if (++SplitBefore < SplitAfter) {
1377 DEBUG(dbgs() << " shrink\n");
1378 // Recompute the max when necessary.
1379 if (GapWeight[SplitBefore - 1] >= MaxGap) {
1380 MaxGap = GapWeight[SplitBefore];
1381 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
1382 MaxGap = std::max(MaxGap, GapWeight[i]);
1389 // Try to extend the interval.
1390 if (SplitAfter >= NumGaps) {
1391 DEBUG(dbgs() << " end\n");
1395 DEBUG(dbgs() << " extend\n");
1396 MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]);
1400 // Didn't find any candidates?
1401 if (BestBefore == NumGaps)
1404 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
1405 << '-' << Uses[BestAfter] << ", " << BestDiff
1406 << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
1408 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1412 SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
1413 SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]);
1414 SE->useIntv(SegStart, SegStop);
1415 SmallVector<unsigned, 8> IntvMap;
1416 SE->finish(&IntvMap);
1417 DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
1419 // If the new range has the same number of instructions as before, mark it as
1420 // RS_Split2 so the next split will be forced to make progress. Otherwise,
1421 // leave the new intervals as RS_New so they can compete.
1422 bool LiveBefore = BestBefore != 0 || BI.LiveIn;
1423 bool LiveAfter = BestAfter != NumGaps || BI.LiveOut;
1424 unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter;
1425 if (NewGaps >= NumGaps) {
1426 DEBUG(dbgs() << "Tagging non-progress ranges: ");
1427 assert(!ProgressRequired && "Didn't make progress when it was required.");
1428 for (unsigned i = 0, e = IntvMap.size(); i != e; ++i)
1429 if (IntvMap[i] == 1) {
1430 setStage(*LREdit.get(i), RS_Split2);
1431 DEBUG(dbgs() << PrintReg(LREdit.get(i)->reg));
1433 DEBUG(dbgs() << '\n');
1440 //===----------------------------------------------------------------------===//
1441 // Live Range Splitting
1442 //===----------------------------------------------------------------------===//
1444 /// trySplit - Try to split VirtReg or one of its interferences, making it
1446 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
1447 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
1448 SmallVectorImpl<LiveInterval*>&NewVRegs) {
1449 // Local intervals are handled separately.
1450 if (LIS->intervalIsInOneMBB(VirtReg)) {
1451 NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
1452 SA->analyze(&VirtReg);
1453 return tryLocalSplit(VirtReg, Order, NewVRegs);
1456 NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
1458 // Ranges must be Split2 or less.
1459 if (getStage(VirtReg) >= RS_Spill)
1462 SA->analyze(&VirtReg);
1464 // FIXME: SplitAnalysis may repair broken live ranges coming from the
1465 // coalescer. That may cause the range to become allocatable which means that
1466 // tryRegionSplit won't be making progress. This check should be replaced with
1467 // an assertion when the coalescer is fixed.
1468 if (SA->didRepairRange()) {
1469 // VirtReg has changed, so all cached queries are invalid.
1470 invalidateVirtRegs();
1471 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1475 // First try to split around a region spanning multiple blocks. RS_Split2
1476 // ranges already made dubious progress with region splitting, so they go
1477 // straight to single block splitting.
1478 if (getStage(VirtReg) < RS_Split2) {
1479 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
1480 if (PhysReg || !NewVRegs.empty())
1484 // Then isolate blocks with multiple uses.
1485 SplitAnalysis::BlockPtrSet Blocks;
1486 if (SA->getMultiUseBlocks(Blocks)) {
1487 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1489 SE->splitSingleBlocks(Blocks);
1490 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Spill);
1492 MF->verify(this, "After splitting live range around basic blocks");
1495 // Don't assign any physregs.
1500 //===----------------------------------------------------------------------===//
1502 //===----------------------------------------------------------------------===//
1504 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
1505 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1506 // First try assigning a free register.
1507 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
1508 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1511 LiveRangeStage Stage = getStage(VirtReg);
1512 DEBUG(dbgs() << StageName[Stage]
1513 << " Cascade " << ExtraRegInfo[VirtReg.reg].Cascade << '\n');
1515 // Try to evict a less worthy live range, but only for ranges from the primary
1516 // queue. The RS_Split ranges already failed to do this, and they should not
1517 // get a second chance until they have been split.
1518 if (Stage != RS_Split)
1519 if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
1522 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
1524 // The first time we see a live range, don't try to split or spill.
1525 // Wait until the second time, when all smaller ranges have been allocated.
1526 // This gives a better picture of the interference to split around.
1527 if (Stage < RS_Split) {
1528 setStage(VirtReg, RS_Split);
1529 DEBUG(dbgs() << "wait for second round\n");
1530 NewVRegs.push_back(&VirtReg);
1534 // If we couldn't allocate a register from spilling, there is probably some
1535 // invalid inline assembly. The base class wil report it.
1536 if (Stage >= RS_Done || !VirtReg.isSpillable())
1539 // Try splitting VirtReg or interferences.
1540 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
1541 if (PhysReg || !NewVRegs.empty())
1544 // Finally spill VirtReg itself.
1545 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
1546 LiveRangeEdit LRE(VirtReg, NewVRegs, this);
1547 spiller().spill(LRE);
1548 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done);
1551 MF->verify(this, "After spilling");
1553 // The live virtual register requesting allocation was spilled, so tell
1554 // the caller not to allocate anything during this round.
1558 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
1559 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
1560 << "********** Function: "
1561 << ((Value*)mf.getFunction())->getName() << '\n');
1565 MF->verify(this, "Before greedy register allocator");
1567 RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
1568 Indexes = &getAnalysis<SlotIndexes>();
1569 DomTree = &getAnalysis<MachineDominatorTree>();
1570 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
1571 Loops = &getAnalysis<MachineLoopInfo>();
1572 Bundles = &getAnalysis<EdgeBundles>();
1573 SpillPlacer = &getAnalysis<SpillPlacement>();
1574 DebugVars = &getAnalysis<LiveDebugVariables>();
1576 SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
1577 SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
1578 ExtraRegInfo.clear();
1579 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1581 IntfCache.init(MF, &PhysReg2LiveUnion[0], Indexes, TRI);
1582 GlobalCand.resize(32); // This will grow as needed.
1586 LIS->addKillFlags();
1590 NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
1591 VRM->rewrite(Indexes);
1594 // Write out new DBG_VALUE instructions.
1595 DebugVars->emitDebugValues(VRM);
1597 // The pass output is in VirtRegMap. Release all the transient data.