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 "LiveRegMatrix.h"
20 #include "RegAllocBase.h"
22 #include "SpillPlacement.h"
24 #include "VirtRegMap.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Function.h"
28 #include "llvm/PassAnalysisSupport.h"
29 #include "llvm/CodeGen/CalcSpillWeights.h"
30 #include "llvm/CodeGen/EdgeBundles.h"
31 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
32 #include "llvm/CodeGen/LiveRangeEdit.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 static cl::opt<SplitEditor::ComplementSpillMode>
56 SplitSpillMode("split-spill-mode", cl::Hidden,
57 cl::desc("Spill mode for splitting live ranges"),
58 cl::values(clEnumValN(SplitEditor::SM_Partition, "default", "Default"),
59 clEnumValN(SplitEditor::SM_Size, "size", "Optimize for size"),
60 clEnumValN(SplitEditor::SM_Speed, "speed", "Optimize for speed"),
62 cl::init(SplitEditor::SM_Partition));
64 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
65 createGreedyRegisterAllocator);
68 class RAGreedy : public MachineFunctionPass,
70 private LiveRangeEdit::Delegate {
77 MachineDominatorTree *DomTree;
78 MachineLoopInfo *Loops;
80 SpillPlacement *SpillPlacer;
81 LiveDebugVariables *DebugVars;
84 std::auto_ptr<Spiller> SpillerInstance;
85 std::priority_queue<std::pair<unsigned, unsigned> > Queue;
88 // Live ranges pass through a number of stages as we try to allocate them.
89 // Some of the stages may also create new live ranges:
91 // - Region splitting.
92 // - Per-block splitting.
96 // Ranges produced by one of the stages skip the previous stages when they are
97 // dequeued. This improves performance because we can skip interference checks
98 // that are unlikely to give any results. It also guarantees that the live
99 // range splitting algorithm terminates, something that is otherwise hard to
101 enum LiveRangeStage {
102 /// Newly created live range that has never been queued.
105 /// Only attempt assignment and eviction. Then requeue as RS_Split.
108 /// Attempt live range splitting if assignment is impossible.
111 /// Attempt more aggressive live range splitting that is guaranteed to make
112 /// progress. This is used for split products that may not be making
116 /// Live range will be spilled. No more splitting will be attempted.
119 /// There is nothing more we can do to this live range. Abort compilation
120 /// if it can't be assigned.
124 static const char *const StageName[];
126 // RegInfo - Keep additional information about each live range.
128 LiveRangeStage Stage;
130 // Cascade - Eviction loop prevention. See canEvictInterference().
133 RegInfo() : Stage(RS_New), Cascade(0) {}
136 IndexedMap<RegInfo, VirtReg2IndexFunctor> ExtraRegInfo;
138 LiveRangeStage getStage(const LiveInterval &VirtReg) const {
139 return ExtraRegInfo[VirtReg.reg].Stage;
142 void setStage(const LiveInterval &VirtReg, LiveRangeStage Stage) {
143 ExtraRegInfo.resize(MRI->getNumVirtRegs());
144 ExtraRegInfo[VirtReg.reg].Stage = Stage;
147 template<typename Iterator>
148 void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
149 ExtraRegInfo.resize(MRI->getNumVirtRegs());
150 for (;Begin != End; ++Begin) {
151 unsigned Reg = (*Begin)->reg;
152 if (ExtraRegInfo[Reg].Stage == RS_New)
153 ExtraRegInfo[Reg].Stage = NewStage;
157 /// Cost of evicting interference.
158 struct EvictionCost {
159 unsigned BrokenHints; ///< Total number of broken hints.
160 float MaxWeight; ///< Maximum spill weight evicted.
162 EvictionCost(unsigned B = 0) : BrokenHints(B), MaxWeight(0) {}
164 bool operator<(const EvictionCost &O) const {
165 if (BrokenHints != O.BrokenHints)
166 return BrokenHints < O.BrokenHints;
167 return MaxWeight < O.MaxWeight;
172 std::auto_ptr<SplitAnalysis> SA;
173 std::auto_ptr<SplitEditor> SE;
175 /// Cached per-block interference maps
176 InterferenceCache IntfCache;
178 /// All basic blocks where the current register has uses.
179 SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints;
181 /// Global live range splitting candidate info.
182 struct GlobalSplitCandidate {
183 // Register intended for assignment, or 0.
186 // SplitKit interval index for this candidate.
189 // Interference for PhysReg.
190 InterferenceCache::Cursor Intf;
192 // Bundles where this candidate should be live.
193 BitVector LiveBundles;
194 SmallVector<unsigned, 8> ActiveBlocks;
196 void reset(InterferenceCache &Cache, unsigned Reg) {
199 Intf.setPhysReg(Cache, Reg);
201 ActiveBlocks.clear();
204 // Set B[i] = C for every live bundle where B[i] was NoCand.
205 unsigned getBundles(SmallVectorImpl<unsigned> &B, unsigned C) {
207 for (int i = LiveBundles.find_first(); i >= 0;
208 i = LiveBundles.find_next(i))
209 if (B[i] == NoCand) {
217 /// Candidate info for for each PhysReg in AllocationOrder.
218 /// This vector never shrinks, but grows to the size of the largest register
220 SmallVector<GlobalSplitCandidate, 32> GlobalCand;
222 enum { NoCand = ~0u };
224 /// Candidate map. Each edge bundle is assigned to a GlobalCand entry, or to
225 /// NoCand which indicates the stack interval.
226 SmallVector<unsigned, 32> BundleCand;
231 /// Return the pass name.
232 virtual const char* getPassName() const {
233 return "Greedy Register Allocator";
236 /// RAGreedy analysis usage.
237 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
238 virtual void releaseMemory();
239 virtual Spiller &spiller() { return *SpillerInstance; }
240 virtual void enqueue(LiveInterval *LI);
241 virtual LiveInterval *dequeue();
242 virtual unsigned selectOrSplit(LiveInterval&,
243 SmallVectorImpl<LiveInterval*>&);
245 /// Perform register allocation.
246 virtual bool runOnMachineFunction(MachineFunction &mf);
251 bool LRE_CanEraseVirtReg(unsigned);
252 void LRE_WillShrinkVirtReg(unsigned);
253 void LRE_DidCloneVirtReg(unsigned, unsigned);
255 float calcSpillCost();
256 bool addSplitConstraints(InterferenceCache::Cursor, float&);
257 void addThroughConstraints(InterferenceCache::Cursor, ArrayRef<unsigned>);
258 void growRegion(GlobalSplitCandidate &Cand);
259 float calcGlobalSplitCost(GlobalSplitCandidate&);
260 bool calcCompactRegion(GlobalSplitCandidate&);
261 void splitAroundRegion(LiveRangeEdit&, ArrayRef<unsigned>);
262 void calcGapWeights(unsigned, SmallVectorImpl<float>&);
263 bool shouldEvict(LiveInterval &A, bool, LiveInterval &B, bool);
264 bool canEvictInterference(LiveInterval&, unsigned, bool, EvictionCost&);
265 void evictInterference(LiveInterval&, unsigned,
266 SmallVectorImpl<LiveInterval*>&);
268 unsigned tryAssign(LiveInterval&, AllocationOrder&,
269 SmallVectorImpl<LiveInterval*>&);
270 unsigned tryEvict(LiveInterval&, AllocationOrder&,
271 SmallVectorImpl<LiveInterval*>&, unsigned = ~0u);
272 unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
273 SmallVectorImpl<LiveInterval*>&);
274 unsigned tryBlockSplit(LiveInterval&, AllocationOrder&,
275 SmallVectorImpl<LiveInterval*>&);
276 unsigned tryInstructionSplit(LiveInterval&, AllocationOrder&,
277 SmallVectorImpl<LiveInterval*>&);
278 unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
279 SmallVectorImpl<LiveInterval*>&);
280 unsigned trySplit(LiveInterval&, AllocationOrder&,
281 SmallVectorImpl<LiveInterval*>&);
283 } // end anonymous namespace
285 char RAGreedy::ID = 0;
288 const char *const RAGreedy::StageName[] = {
298 // Hysteresis to use when comparing floats.
299 // This helps stabilize decisions based on float comparisons.
300 const float Hysteresis = 0.98f;
303 FunctionPass* llvm::createGreedyRegisterAllocator() {
304 return new RAGreedy();
307 RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
308 initializeLiveDebugVariablesPass(*PassRegistry::getPassRegistry());
309 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
310 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
311 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
312 initializeRegisterCoalescerPass(*PassRegistry::getPassRegistry());
313 initializeMachineSchedulerPass(*PassRegistry::getPassRegistry());
314 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
315 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
316 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
317 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
318 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
319 initializeLiveRegMatrixPass(*PassRegistry::getPassRegistry());
320 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
321 initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
324 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
325 AU.setPreservesCFG();
326 AU.addRequired<AliasAnalysis>();
327 AU.addPreserved<AliasAnalysis>();
328 AU.addRequired<LiveIntervals>();
329 AU.addPreserved<LiveIntervals>();
330 AU.addRequired<SlotIndexes>();
331 AU.addPreserved<SlotIndexes>();
332 AU.addRequired<LiveDebugVariables>();
333 AU.addPreserved<LiveDebugVariables>();
334 AU.addRequired<CalculateSpillWeights>();
335 AU.addRequired<LiveStacks>();
336 AU.addPreserved<LiveStacks>();
337 AU.addRequired<MachineDominatorTree>();
338 AU.addPreserved<MachineDominatorTree>();
339 AU.addRequired<MachineLoopInfo>();
340 AU.addPreserved<MachineLoopInfo>();
341 AU.addRequired<VirtRegMap>();
342 AU.addPreserved<VirtRegMap>();
343 AU.addRequired<LiveRegMatrix>();
344 AU.addPreserved<LiveRegMatrix>();
345 AU.addRequired<EdgeBundles>();
346 AU.addRequired<SpillPlacement>();
347 MachineFunctionPass::getAnalysisUsage(AU);
351 //===----------------------------------------------------------------------===//
352 // LiveRangeEdit delegate methods
353 //===----------------------------------------------------------------------===//
355 bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
356 if (VRM->hasPhys(VirtReg)) {
357 Matrix->unassign(LIS->getInterval(VirtReg));
360 // Unassigned virtreg is probably in the priority queue.
361 // RegAllocBase will erase it after dequeueing.
365 void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
366 if (!VRM->hasPhys(VirtReg))
369 // Register is assigned, put it back on the queue for reassignment.
370 LiveInterval &LI = LIS->getInterval(VirtReg);
371 Matrix->unassign(LI);
375 void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
376 // Cloning a register we haven't even heard about yet? Just ignore it.
377 if (!ExtraRegInfo.inBounds(Old))
380 // LRE may clone a virtual register because dead code elimination causes it to
381 // be split into connected components. The new components are much smaller
382 // than the original, so they should get a new chance at being assigned.
383 // same stage as the parent.
384 ExtraRegInfo[Old].Stage = RS_Assign;
385 ExtraRegInfo.grow(New);
386 ExtraRegInfo[New] = ExtraRegInfo[Old];
389 void RAGreedy::releaseMemory() {
390 SpillerInstance.reset(0);
391 ExtraRegInfo.clear();
393 RegAllocBase::releaseMemory();
396 void RAGreedy::enqueue(LiveInterval *LI) {
397 // Prioritize live ranges by size, assigning larger ranges first.
398 // The queue holds (size, reg) pairs.
399 const unsigned Size = LI->getSize();
400 const unsigned Reg = LI->reg;
401 assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
402 "Can only enqueue virtual registers");
405 ExtraRegInfo.grow(Reg);
406 if (ExtraRegInfo[Reg].Stage == RS_New)
407 ExtraRegInfo[Reg].Stage = RS_Assign;
409 if (ExtraRegInfo[Reg].Stage == RS_Split) {
410 // Unsplit ranges that couldn't be allocated immediately are deferred until
411 // everything else has been allocated.
414 // Everything is allocated in long->short order. Long ranges that don't fit
415 // should be spilled (or split) ASAP so they don't create interference.
416 Prio = (1u << 31) + Size;
418 // Boost ranges that have a physical register hint.
419 if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
423 Queue.push(std::make_pair(Prio, ~Reg));
426 LiveInterval *RAGreedy::dequeue() {
429 LiveInterval *LI = &LIS->getInterval(~Queue.top().second);
435 //===----------------------------------------------------------------------===//
437 //===----------------------------------------------------------------------===//
439 /// tryAssign - Try to assign VirtReg to an available register.
440 unsigned RAGreedy::tryAssign(LiveInterval &VirtReg,
441 AllocationOrder &Order,
442 SmallVectorImpl<LiveInterval*> &NewVRegs) {
445 while ((PhysReg = Order.next()))
446 if (!Matrix->checkInterference(VirtReg, PhysReg))
448 if (!PhysReg || Order.isHint(PhysReg))
451 // PhysReg is available, but there may be a better choice.
453 // If we missed a simple hint, try to cheaply evict interference from the
454 // preferred register.
455 if (unsigned Hint = MRI->getSimpleHint(VirtReg.reg))
456 if (Order.isHint(Hint)) {
457 DEBUG(dbgs() << "missed hint " << PrintReg(Hint, TRI) << '\n');
458 EvictionCost MaxCost(1);
459 if (canEvictInterference(VirtReg, Hint, true, MaxCost)) {
460 evictInterference(VirtReg, Hint, NewVRegs);
465 // Try to evict interference from a cheaper alternative.
466 unsigned Cost = TRI->getCostPerUse(PhysReg);
468 // Most registers have 0 additional cost.
472 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " is available at cost " << Cost
474 unsigned CheapReg = tryEvict(VirtReg, Order, NewVRegs, Cost);
475 return CheapReg ? CheapReg : PhysReg;
479 //===----------------------------------------------------------------------===//
480 // Interference eviction
481 //===----------------------------------------------------------------------===//
483 /// shouldEvict - determine if A should evict the assigned live range B. The
484 /// eviction policy defined by this function together with the allocation order
485 /// defined by enqueue() decides which registers ultimately end up being split
488 /// Cascade numbers are used to prevent infinite loops if this function is a
491 /// @param A The live range to be assigned.
492 /// @param IsHint True when A is about to be assigned to its preferred
494 /// @param B The live range to be evicted.
495 /// @param BreaksHint True when B is already assigned to its preferred register.
496 bool RAGreedy::shouldEvict(LiveInterval &A, bool IsHint,
497 LiveInterval &B, bool BreaksHint) {
498 bool CanSplit = getStage(B) < RS_Spill;
500 // Be fairly aggressive about following hints as long as the evictee can be
502 if (CanSplit && IsHint && !BreaksHint)
505 return A.weight > B.weight;
508 /// canEvictInterference - Return true if all interferences between VirtReg and
509 /// PhysReg can be evicted. When OnlyCheap is set, don't do anything
511 /// @param VirtReg Live range that is about to be assigned.
512 /// @param PhysReg Desired register for assignment.
513 /// @prarm IsHint True when PhysReg is VirtReg's preferred register.
514 /// @param MaxCost Only look for cheaper candidates and update with new cost
515 /// when returning true.
516 /// @returns True when interference can be evicted cheaper than MaxCost.
517 bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
518 bool IsHint, EvictionCost &MaxCost) {
519 // It is only possible to evict virtual register interference.
520 if (Matrix->checkInterference(VirtReg, PhysReg) > LiveRegMatrix::IK_VirtReg)
523 // Find VirtReg's cascade number. This will be unassigned if VirtReg was never
524 // involved in an eviction before. If a cascade number was assigned, deny
525 // evicting anything with the same or a newer cascade number. This prevents
526 // infinite eviction loops.
528 // This works out so a register without a cascade number is allowed to evict
529 // anything, and it can be evicted by anything.
530 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
532 Cascade = NextCascade;
535 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
536 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
537 // If there is 10 or more interferences, chances are one is heavier.
538 if (Q.collectInterferingVRegs(10) >= 10)
541 // Check if any interfering live range is heavier than MaxWeight.
542 for (unsigned i = Q.interferingVRegs().size(); i; --i) {
543 LiveInterval *Intf = Q.interferingVRegs()[i - 1];
544 assert(TargetRegisterInfo::isVirtualRegister(Intf->reg) &&
545 "Only expecting virtual register interference from query");
546 // Never evict spill products. They cannot split or spill.
547 if (getStage(*Intf) == RS_Done)
549 // Once a live range becomes small enough, it is urgent that we find a
550 // register for it. This is indicated by an infinite spill weight. These
551 // urgent live ranges get to evict almost anything.
553 // Also allow urgent evictions of unspillable ranges from a strictly
554 // larger allocation order.
555 bool Urgent = !VirtReg.isSpillable() &&
556 (Intf->isSpillable() ||
557 RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(VirtReg.reg)) <
558 RegClassInfo.getNumAllocatableRegs(MRI->getRegClass(Intf->reg)));
559 // Only evict older cascades or live ranges without a cascade.
560 unsigned IntfCascade = ExtraRegInfo[Intf->reg].Cascade;
561 if (Cascade <= IntfCascade) {
564 // We permit breaking cascades for urgent evictions. It should be the
565 // last resort, though, so make it really expensive.
566 Cost.BrokenHints += 10;
568 // Would this break a satisfied hint?
569 bool BreaksHint = VRM->hasPreferredPhys(Intf->reg);
570 // Update eviction cost.
571 Cost.BrokenHints += BreaksHint;
572 Cost.MaxWeight = std::max(Cost.MaxWeight, Intf->weight);
573 // Abort if this would be too expensive.
574 if (!(Cost < MaxCost))
576 // Finally, apply the eviction policy for non-urgent evictions.
577 if (!Urgent && !shouldEvict(VirtReg, IsHint, *Intf, BreaksHint))
585 /// evictInterference - Evict any interferring registers that prevent VirtReg
586 /// from being assigned to Physreg. This assumes that canEvictInterference
588 void RAGreedy::evictInterference(LiveInterval &VirtReg, unsigned PhysReg,
589 SmallVectorImpl<LiveInterval*> &NewVRegs) {
590 // Make sure that VirtReg has a cascade number, and assign that cascade
591 // number to every evicted register. These live ranges than then only be
592 // evicted by a newer cascade, preventing infinite loops.
593 unsigned Cascade = ExtraRegInfo[VirtReg.reg].Cascade;
595 Cascade = ExtraRegInfo[VirtReg.reg].Cascade = NextCascade++;
597 DEBUG(dbgs() << "evicting " << PrintReg(PhysReg, TRI)
598 << " interference: Cascade " << Cascade << '\n');
600 // Collect all interfering virtregs first.
601 SmallVector<LiveInterval*, 8> Intfs;
602 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
603 LiveIntervalUnion::Query &Q = Matrix->query(VirtReg, *Units);
604 assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
605 ArrayRef<LiveInterval*> IVR = Q.interferingVRegs();
606 Intfs.append(IVR.begin(), IVR.end());
609 // Evict them second. This will invalidate the queries.
610 for (unsigned i = 0, e = Intfs.size(); i != e; ++i) {
611 LiveInterval *Intf = Intfs[i];
612 // The same VirtReg may be present in multiple RegUnits. Skip duplicates.
613 if (!VRM->hasPhys(Intf->reg))
615 Matrix->unassign(*Intf);
616 assert((ExtraRegInfo[Intf->reg].Cascade < Cascade ||
617 VirtReg.isSpillable() < Intf->isSpillable()) &&
618 "Cannot decrease cascade number, illegal eviction");
619 ExtraRegInfo[Intf->reg].Cascade = Cascade;
621 NewVRegs.push_back(Intf);
625 /// tryEvict - Try to evict all interferences for a physreg.
626 /// @param VirtReg Currently unassigned virtual register.
627 /// @param Order Physregs to try.
628 /// @return Physreg to assign VirtReg, or 0.
629 unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
630 AllocationOrder &Order,
631 SmallVectorImpl<LiveInterval*> &NewVRegs,
632 unsigned CostPerUseLimit) {
633 NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
635 // Keep track of the cheapest interference seen so far.
636 EvictionCost BestCost(~0u);
637 unsigned BestPhys = 0;
639 // When we are just looking for a reduced cost per use, don't break any
640 // hints, and only evict smaller spill weights.
641 if (CostPerUseLimit < ~0u) {
642 BestCost.BrokenHints = 0;
643 BestCost.MaxWeight = VirtReg.weight;
647 while (unsigned PhysReg = Order.next()) {
648 if (TRI->getCostPerUse(PhysReg) >= CostPerUseLimit)
650 // The first use of a callee-saved register in a function has cost 1.
651 // Don't start using a CSR when the CostPerUseLimit is low.
652 if (CostPerUseLimit == 1)
653 if (unsigned CSR = RegClassInfo.getLastCalleeSavedAlias(PhysReg))
654 if (!MRI->isPhysRegUsed(CSR)) {
655 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << " would clobber CSR "
656 << PrintReg(CSR, TRI) << '\n');
660 if (!canEvictInterference(VirtReg, PhysReg, false, BestCost))
666 // Stop if the hint can be used.
667 if (Order.isHint(PhysReg))
674 evictInterference(VirtReg, BestPhys, NewVRegs);
679 //===----------------------------------------------------------------------===//
681 //===----------------------------------------------------------------------===//
683 /// addSplitConstraints - Fill out the SplitConstraints vector based on the
684 /// interference pattern in Physreg and its aliases. Add the constraints to
685 /// SpillPlacement and return the static cost of this split in Cost, assuming
686 /// that all preferences in SplitConstraints are met.
687 /// Return false if there are no bundles with positive bias.
688 bool RAGreedy::addSplitConstraints(InterferenceCache::Cursor Intf,
690 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
692 // Reset interference dependent info.
693 SplitConstraints.resize(UseBlocks.size());
694 float StaticCost = 0;
695 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
696 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
697 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
699 BC.Number = BI.MBB->getNumber();
700 Intf.moveToBlock(BC.Number);
701 BC.Entry = BI.LiveIn ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
702 BC.Exit = BI.LiveOut ? SpillPlacement::PrefReg : SpillPlacement::DontCare;
703 BC.ChangesValue = BI.FirstDef;
705 if (!Intf.hasInterference())
708 // Number of spill code instructions to insert.
711 // Interference for the live-in value.
713 if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
714 BC.Entry = SpillPlacement::MustSpill, ++Ins;
715 else if (Intf.first() < BI.FirstInstr)
716 BC.Entry = SpillPlacement::PrefSpill, ++Ins;
717 else if (Intf.first() < BI.LastInstr)
721 // Interference for the live-out value.
723 if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
724 BC.Exit = SpillPlacement::MustSpill, ++Ins;
725 else if (Intf.last() > BI.LastInstr)
726 BC.Exit = SpillPlacement::PrefSpill, ++Ins;
727 else if (Intf.last() > BI.FirstInstr)
731 // Accumulate the total frequency of inserted spill code.
733 StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
737 // Add constraints for use-blocks. Note that these are the only constraints
738 // that may add a positive bias, it is downhill from here.
739 SpillPlacer->addConstraints(SplitConstraints);
740 return SpillPlacer->scanActiveBundles();
744 /// addThroughConstraints - Add constraints and links to SpillPlacer from the
745 /// live-through blocks in Blocks.
746 void RAGreedy::addThroughConstraints(InterferenceCache::Cursor Intf,
747 ArrayRef<unsigned> Blocks) {
748 const unsigned GroupSize = 8;
749 SpillPlacement::BlockConstraint BCS[GroupSize];
750 unsigned TBS[GroupSize];
751 unsigned B = 0, T = 0;
753 for (unsigned i = 0; i != Blocks.size(); ++i) {
754 unsigned Number = Blocks[i];
755 Intf.moveToBlock(Number);
757 if (!Intf.hasInterference()) {
758 assert(T < GroupSize && "Array overflow");
760 if (++T == GroupSize) {
761 SpillPlacer->addLinks(makeArrayRef(TBS, T));
767 assert(B < GroupSize && "Array overflow");
768 BCS[B].Number = Number;
770 // Interference for the live-in value.
771 if (Intf.first() <= Indexes->getMBBStartIdx(Number))
772 BCS[B].Entry = SpillPlacement::MustSpill;
774 BCS[B].Entry = SpillPlacement::PrefSpill;
776 // Interference for the live-out value.
777 if (Intf.last() >= SA->getLastSplitPoint(Number))
778 BCS[B].Exit = SpillPlacement::MustSpill;
780 BCS[B].Exit = SpillPlacement::PrefSpill;
782 if (++B == GroupSize) {
783 ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
784 SpillPlacer->addConstraints(Array);
789 ArrayRef<SpillPlacement::BlockConstraint> Array(BCS, B);
790 SpillPlacer->addConstraints(Array);
791 SpillPlacer->addLinks(makeArrayRef(TBS, T));
794 void RAGreedy::growRegion(GlobalSplitCandidate &Cand) {
795 // Keep track of through blocks that have not been added to SpillPlacer.
796 BitVector Todo = SA->getThroughBlocks();
797 SmallVectorImpl<unsigned> &ActiveBlocks = Cand.ActiveBlocks;
798 unsigned AddedTo = 0;
800 unsigned Visited = 0;
804 ArrayRef<unsigned> NewBundles = SpillPlacer->getRecentPositive();
805 // Find new through blocks in the periphery of PrefRegBundles.
806 for (int i = 0, e = NewBundles.size(); i != e; ++i) {
807 unsigned Bundle = NewBundles[i];
808 // Look at all blocks connected to Bundle in the full graph.
809 ArrayRef<unsigned> Blocks = Bundles->getBlocks(Bundle);
810 for (ArrayRef<unsigned>::iterator I = Blocks.begin(), E = Blocks.end();
813 if (!Todo.test(Block))
816 // This is a new through block. Add it to SpillPlacer later.
817 ActiveBlocks.push_back(Block);
823 // Any new blocks to add?
824 if (ActiveBlocks.size() == AddedTo)
827 // Compute through constraints from the interference, or assume that all
828 // through blocks prefer spilling when forming compact regions.
829 ArrayRef<unsigned> NewBlocks = makeArrayRef(ActiveBlocks).slice(AddedTo);
831 addThroughConstraints(Cand.Intf, NewBlocks);
833 // Provide a strong negative bias on through blocks to prevent unwanted
834 // liveness on loop backedges.
835 SpillPlacer->addPrefSpill(NewBlocks, /* Strong= */ true);
836 AddedTo = ActiveBlocks.size();
838 // Perhaps iterating can enable more bundles?
839 SpillPlacer->iterate();
841 DEBUG(dbgs() << ", v=" << Visited);
844 /// calcCompactRegion - Compute the set of edge bundles that should be live
845 /// when splitting the current live range into compact regions. Compact
846 /// regions can be computed without looking at interference. They are the
847 /// regions formed by removing all the live-through blocks from the live range.
849 /// Returns false if the current live range is already compact, or if the
850 /// compact regions would form single block regions anyway.
851 bool RAGreedy::calcCompactRegion(GlobalSplitCandidate &Cand) {
852 // Without any through blocks, the live range is already compact.
853 if (!SA->getNumThroughBlocks())
856 // Compact regions don't correspond to any physreg.
857 Cand.reset(IntfCache, 0);
859 DEBUG(dbgs() << "Compact region bundles");
861 // Use the spill placer to determine the live bundles. GrowRegion pretends
862 // that all the through blocks have interference when PhysReg is unset.
863 SpillPlacer->prepare(Cand.LiveBundles);
865 // The static split cost will be zero since Cand.Intf reports no interference.
867 if (!addSplitConstraints(Cand.Intf, Cost)) {
868 DEBUG(dbgs() << ", none.\n");
873 SpillPlacer->finish();
875 if (!Cand.LiveBundles.any()) {
876 DEBUG(dbgs() << ", none.\n");
881 for (int i = Cand.LiveBundles.find_first(); i>=0;
882 i = Cand.LiveBundles.find_next(i))
883 dbgs() << " EB#" << i;
889 /// calcSpillCost - Compute how expensive it would be to split the live range in
890 /// SA around all use blocks instead of forming bundle regions.
891 float RAGreedy::calcSpillCost() {
893 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
894 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
895 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
896 unsigned Number = BI.MBB->getNumber();
897 // We normally only need one spill instruction - a load or a store.
898 Cost += SpillPlacer->getBlockFrequency(Number);
900 // Unless the value is redefined in the block.
901 if (BI.LiveIn && BI.LiveOut && BI.FirstDef)
902 Cost += SpillPlacer->getBlockFrequency(Number);
907 /// calcGlobalSplitCost - Return the global split cost of following the split
908 /// pattern in LiveBundles. This cost should be added to the local cost of the
909 /// interference pattern in SplitConstraints.
911 float RAGreedy::calcGlobalSplitCost(GlobalSplitCandidate &Cand) {
912 float GlobalCost = 0;
913 const BitVector &LiveBundles = Cand.LiveBundles;
914 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
915 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
916 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
917 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
918 bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)];
919 bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
923 Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
925 Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
927 GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
930 for (unsigned i = 0, e = Cand.ActiveBlocks.size(); i != e; ++i) {
931 unsigned Number = Cand.ActiveBlocks[i];
932 bool RegIn = LiveBundles[Bundles->getBundle(Number, 0)];
933 bool RegOut = LiveBundles[Bundles->getBundle(Number, 1)];
934 if (!RegIn && !RegOut)
936 if (RegIn && RegOut) {
937 // We need double spill code if this block has interference.
938 Cand.Intf.moveToBlock(Number);
939 if (Cand.Intf.hasInterference())
940 GlobalCost += 2*SpillPlacer->getBlockFrequency(Number);
943 // live-in / stack-out or stack-in live-out.
944 GlobalCost += SpillPlacer->getBlockFrequency(Number);
949 /// splitAroundRegion - Split the current live range around the regions
950 /// determined by BundleCand and GlobalCand.
952 /// Before calling this function, GlobalCand and BundleCand must be initialized
953 /// so each bundle is assigned to a valid candidate, or NoCand for the
954 /// stack-bound bundles. The shared SA/SE SplitAnalysis and SplitEditor
955 /// objects must be initialized for the current live range, and intervals
956 /// created for the used candidates.
958 /// @param LREdit The LiveRangeEdit object handling the current split.
959 /// @param UsedCands List of used GlobalCand entries. Every BundleCand value
960 /// must appear in this list.
961 void RAGreedy::splitAroundRegion(LiveRangeEdit &LREdit,
962 ArrayRef<unsigned> UsedCands) {
963 // These are the intervals created for new global ranges. We may create more
964 // intervals for local ranges.
965 const unsigned NumGlobalIntvs = LREdit.size();
966 DEBUG(dbgs() << "splitAroundRegion with " << NumGlobalIntvs << " globals.\n");
967 assert(NumGlobalIntvs && "No global intervals configured");
969 // Isolate even single instructions when dealing with a proper sub-class.
970 // That guarantees register class inflation for the stack interval because it
972 unsigned Reg = SA->getParent().reg;
973 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
975 // First handle all the blocks with uses.
976 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
977 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
978 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
979 unsigned Number = BI.MBB->getNumber();
980 unsigned IntvIn = 0, IntvOut = 0;
981 SlotIndex IntfIn, IntfOut;
983 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
984 if (CandIn != NoCand) {
985 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
986 IntvIn = Cand.IntvIdx;
987 Cand.Intf.moveToBlock(Number);
988 IntfIn = Cand.Intf.first();
992 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
993 if (CandOut != NoCand) {
994 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
995 IntvOut = Cand.IntvIdx;
996 Cand.Intf.moveToBlock(Number);
997 IntfOut = Cand.Intf.last();
1001 // Create separate intervals for isolated blocks with multiple uses.
1002 if (!IntvIn && !IntvOut) {
1003 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " isolated.\n");
1004 if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
1005 SE->splitSingleBlock(BI);
1009 if (IntvIn && IntvOut)
1010 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
1012 SE->splitRegInBlock(BI, IntvIn, IntfIn);
1014 SE->splitRegOutBlock(BI, IntvOut, IntfOut);
1017 // Handle live-through blocks. The relevant live-through blocks are stored in
1018 // the ActiveBlocks list with each candidate. We need to filter out
1020 BitVector Todo = SA->getThroughBlocks();
1021 for (unsigned c = 0; c != UsedCands.size(); ++c) {
1022 ArrayRef<unsigned> Blocks = GlobalCand[UsedCands[c]].ActiveBlocks;
1023 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1024 unsigned Number = Blocks[i];
1025 if (!Todo.test(Number))
1029 unsigned IntvIn = 0, IntvOut = 0;
1030 SlotIndex IntfIn, IntfOut;
1032 unsigned CandIn = BundleCand[Bundles->getBundle(Number, 0)];
1033 if (CandIn != NoCand) {
1034 GlobalSplitCandidate &Cand = GlobalCand[CandIn];
1035 IntvIn = Cand.IntvIdx;
1036 Cand.Intf.moveToBlock(Number);
1037 IntfIn = Cand.Intf.first();
1040 unsigned CandOut = BundleCand[Bundles->getBundle(Number, 1)];
1041 if (CandOut != NoCand) {
1042 GlobalSplitCandidate &Cand = GlobalCand[CandOut];
1043 IntvOut = Cand.IntvIdx;
1044 Cand.Intf.moveToBlock(Number);
1045 IntfOut = Cand.Intf.last();
1047 if (!IntvIn && !IntvOut)
1049 SE->splitLiveThroughBlock(Number, IntvIn, IntfIn, IntvOut, IntfOut);
1055 SmallVector<unsigned, 8> IntvMap;
1056 SE->finish(&IntvMap);
1057 DebugVars->splitRegister(Reg, LREdit.regs());
1059 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1060 unsigned OrigBlocks = SA->getNumLiveBlocks();
1062 // Sort out the new intervals created by splitting. We get four kinds:
1063 // - Remainder intervals should not be split again.
1064 // - Candidate intervals can be assigned to Cand.PhysReg.
1065 // - Block-local splits are candidates for local splitting.
1066 // - DCE leftovers should go back on the queue.
1067 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1068 LiveInterval &Reg = *LREdit.get(i);
1070 // Ignore old intervals from DCE.
1071 if (getStage(Reg) != RS_New)
1074 // Remainder interval. Don't try splitting again, spill if it doesn't
1076 if (IntvMap[i] == 0) {
1077 setStage(Reg, RS_Spill);
1081 // Global intervals. Allow repeated splitting as long as the number of live
1082 // blocks is strictly decreasing.
1083 if (IntvMap[i] < NumGlobalIntvs) {
1084 if (SA->countLiveBlocks(&Reg) >= OrigBlocks) {
1085 DEBUG(dbgs() << "Main interval covers the same " << OrigBlocks
1086 << " blocks as original.\n");
1087 // Don't allow repeated splitting as a safe guard against looping.
1088 setStage(Reg, RS_Split2);
1093 // Other intervals are treated as new. This includes local intervals created
1094 // for blocks with multiple uses, and anything created by DCE.
1098 MF->verify(this, "After splitting live range around region");
1101 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1102 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1103 unsigned NumCands = 0;
1104 unsigned BestCand = NoCand;
1106 SmallVector<unsigned, 8> UsedCands;
1108 // Check if we can split this live range around a compact region.
1109 bool HasCompact = calcCompactRegion(GlobalCand.front());
1111 // Yes, keep GlobalCand[0] as the compact region candidate.
1113 BestCost = HUGE_VALF;
1115 // No benefit from the compact region, our fallback will be per-block
1116 // splitting. Make sure we find a solution that is cheaper than spilling.
1117 BestCost = Hysteresis * calcSpillCost();
1118 DEBUG(dbgs() << "Cost of isolating all blocks = " << BestCost << '\n');
1122 while (unsigned PhysReg = Order.next()) {
1123 // Discard bad candidates before we run out of interference cache cursors.
1124 // This will only affect register classes with a lot of registers (>32).
1125 if (NumCands == IntfCache.getMaxCursors()) {
1126 unsigned WorstCount = ~0u;
1128 for (unsigned i = 0; i != NumCands; ++i) {
1129 if (i == BestCand || !GlobalCand[i].PhysReg)
1131 unsigned Count = GlobalCand[i].LiveBundles.count();
1132 if (Count < WorstCount)
1133 Worst = i, WorstCount = Count;
1136 GlobalCand[Worst] = GlobalCand[NumCands];
1137 if (BestCand == NumCands)
1141 if (GlobalCand.size() <= NumCands)
1142 GlobalCand.resize(NumCands+1);
1143 GlobalSplitCandidate &Cand = GlobalCand[NumCands];
1144 Cand.reset(IntfCache, PhysReg);
1146 SpillPlacer->prepare(Cand.LiveBundles);
1148 if (!addSplitConstraints(Cand.Intf, Cost)) {
1149 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tno positive bundles\n");
1152 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
1153 if (Cost >= BestCost) {
1155 if (BestCand == NoCand)
1156 dbgs() << " worse than no bundles\n";
1158 dbgs() << " worse than "
1159 << PrintReg(GlobalCand[BestCand].PhysReg, TRI) << '\n';
1165 SpillPlacer->finish();
1167 // No live bundles, defer to splitSingleBlocks().
1168 if (!Cand.LiveBundles.any()) {
1169 DEBUG(dbgs() << " no bundles.\n");
1173 Cost += calcGlobalSplitCost(Cand);
1175 dbgs() << ", total = " << Cost << " with bundles";
1176 for (int i = Cand.LiveBundles.find_first(); i>=0;
1177 i = Cand.LiveBundles.find_next(i))
1178 dbgs() << " EB#" << i;
1181 if (Cost < BestCost) {
1182 BestCand = NumCands;
1183 BestCost = Hysteresis * Cost; // Prevent rounding effects.
1188 // No solutions found, fall back to single block splitting.
1189 if (!HasCompact && BestCand == NoCand)
1192 // Prepare split editor.
1193 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1194 SE->reset(LREdit, SplitSpillMode);
1196 // Assign all edge bundles to the preferred candidate, or NoCand.
1197 BundleCand.assign(Bundles->getNumBundles(), NoCand);
1199 // Assign bundles for the best candidate region.
1200 if (BestCand != NoCand) {
1201 GlobalSplitCandidate &Cand = GlobalCand[BestCand];
1202 if (unsigned B = Cand.getBundles(BundleCand, BestCand)) {
1203 UsedCands.push_back(BestCand);
1204 Cand.IntvIdx = SE->openIntv();
1205 DEBUG(dbgs() << "Split for " << PrintReg(Cand.PhysReg, TRI) << " in "
1206 << B << " bundles, intv " << Cand.IntvIdx << ".\n");
1211 // Assign bundles for the compact region.
1213 GlobalSplitCandidate &Cand = GlobalCand.front();
1214 assert(!Cand.PhysReg && "Compact region has no physreg");
1215 if (unsigned B = Cand.getBundles(BundleCand, 0)) {
1216 UsedCands.push_back(0);
1217 Cand.IntvIdx = SE->openIntv();
1218 DEBUG(dbgs() << "Split for compact region in " << B << " bundles, intv "
1219 << Cand.IntvIdx << ".\n");
1224 splitAroundRegion(LREdit, UsedCands);
1229 //===----------------------------------------------------------------------===//
1230 // Per-Block Splitting
1231 //===----------------------------------------------------------------------===//
1233 /// tryBlockSplit - Split a global live range around every block with uses. This
1234 /// creates a lot of local live ranges, that will be split by tryLocalSplit if
1235 /// they don't allocate.
1236 unsigned RAGreedy::tryBlockSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1237 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1238 assert(&SA->getParent() == &VirtReg && "Live range wasn't analyzed");
1239 unsigned Reg = VirtReg.reg;
1240 bool SingleInstrs = RegClassInfo.isProperSubClass(MRI->getRegClass(Reg));
1241 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1242 SE->reset(LREdit, SplitSpillMode);
1243 ArrayRef<SplitAnalysis::BlockInfo> UseBlocks = SA->getUseBlocks();
1244 for (unsigned i = 0; i != UseBlocks.size(); ++i) {
1245 const SplitAnalysis::BlockInfo &BI = UseBlocks[i];
1246 if (SA->shouldSplitSingleBlock(BI, SingleInstrs))
1247 SE->splitSingleBlock(BI);
1249 // No blocks were split.
1253 // We did split for some blocks.
1254 SmallVector<unsigned, 8> IntvMap;
1255 SE->finish(&IntvMap);
1257 // Tell LiveDebugVariables about the new ranges.
1258 DebugVars->splitRegister(Reg, LREdit.regs());
1260 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1262 // Sort out the new intervals created by splitting. The remainder interval
1263 // goes straight to spilling, the new local ranges get to stay RS_New.
1264 for (unsigned i = 0, e = LREdit.size(); i != e; ++i) {
1265 LiveInterval &LI = *LREdit.get(i);
1266 if (getStage(LI) == RS_New && IntvMap[i] == 0)
1267 setStage(LI, RS_Spill);
1271 MF->verify(this, "After splitting live range around basic blocks");
1276 //===----------------------------------------------------------------------===//
1277 // Per-Instruction Splitting
1278 //===----------------------------------------------------------------------===//
1280 /// tryInstructionSplit - Split a live range around individual instructions.
1281 /// This is normally not worthwhile since the spiller is doing essentially the
1282 /// same thing. However, when the live range is in a constrained register
1283 /// class, it may help to insert copies such that parts of the live range can
1284 /// be moved to a larger register class.
1286 /// This is similar to spilling to a larger register class.
1288 RAGreedy::tryInstructionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1289 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1290 // There is no point to this if there are no larger sub-classes.
1291 if (!RegClassInfo.isProperSubClass(MRI->getRegClass(VirtReg.reg)))
1294 // Always enable split spill mode, since we're effectively spilling to a
1296 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1297 SE->reset(LREdit, SplitEditor::SM_Size);
1299 ArrayRef<SlotIndex> Uses = SA->getUseSlots();
1300 if (Uses.size() <= 1)
1303 DEBUG(dbgs() << "Split around " << Uses.size() << " individual instrs.\n");
1305 // Split around every non-copy instruction.
1306 for (unsigned i = 0; i != Uses.size(); ++i) {
1307 if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]))
1308 if (MI->isFullCopy()) {
1309 DEBUG(dbgs() << " skip:\t" << Uses[i] << '\t' << *MI);
1313 SlotIndex SegStart = SE->enterIntvBefore(Uses[i]);
1314 SlotIndex SegStop = SE->leaveIntvAfter(Uses[i]);
1315 SE->useIntv(SegStart, SegStop);
1318 if (LREdit.empty()) {
1319 DEBUG(dbgs() << "All uses were copies.\n");
1323 SmallVector<unsigned, 8> IntvMap;
1324 SE->finish(&IntvMap);
1325 DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
1326 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1328 // Assign all new registers to RS_Spill. This was the last chance.
1329 setStage(LREdit.begin(), LREdit.end(), RS_Spill);
1334 //===----------------------------------------------------------------------===//
1336 //===----------------------------------------------------------------------===//
1339 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted
1340 /// in order to use PhysReg between two entries in SA->UseSlots.
1342 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
1344 void RAGreedy::calcGapWeights(unsigned PhysReg,
1345 SmallVectorImpl<float> &GapWeight) {
1346 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1347 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1348 ArrayRef<SlotIndex> Uses = SA->getUseSlots();
1349 const unsigned NumGaps = Uses.size()-1;
1351 // Start and end points for the interference check.
1352 SlotIndex StartIdx =
1353 BI.LiveIn ? BI.FirstInstr.getBaseIndex() : BI.FirstInstr;
1355 BI.LiveOut ? BI.LastInstr.getBoundaryIndex() : BI.LastInstr;
1357 GapWeight.assign(NumGaps, 0.0f);
1359 // Add interference from each overlapping register.
1360 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
1361 if (!Matrix->query(const_cast<LiveInterval&>(SA->getParent()), *Units)
1362 .checkInterference())
1365 // We know that VirtReg is a continuous interval from FirstInstr to
1366 // LastInstr, so we don't need InterferenceQuery.
1368 // Interference that overlaps an instruction is counted in both gaps
1369 // surrounding the instruction. The exception is interference before
1370 // StartIdx and after StopIdx.
1372 LiveIntervalUnion::SegmentIter IntI =
1373 Matrix->getLiveUnions()[*Units] .find(StartIdx);
1374 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
1375 // Skip the gaps before IntI.
1376 while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
1377 if (++Gap == NumGaps)
1382 // Update the gaps covered by IntI.
1383 const float weight = IntI.value()->weight;
1384 for (; Gap != NumGaps; ++Gap) {
1385 GapWeight[Gap] = std::max(GapWeight[Gap], weight);
1386 if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
1394 // Add fixed interference.
1395 for (MCRegUnitIterator Units(PhysReg, TRI); Units.isValid(); ++Units) {
1396 const LiveInterval &LI = LIS->getRegUnit(*Units);
1397 LiveInterval::const_iterator I = LI.find(StartIdx);
1398 LiveInterval::const_iterator E = LI.end();
1400 // Same loop as above. Mark any overlapped gaps as HUGE_VALF.
1401 for (unsigned Gap = 0; I != E && I->start < StopIdx; ++I) {
1402 while (Uses[Gap+1].getBoundaryIndex() < I->start)
1403 if (++Gap == NumGaps)
1408 for (; Gap != NumGaps; ++Gap) {
1409 GapWeight[Gap] = HUGE_VALF;
1410 if (Uses[Gap+1].getBaseIndex() >= I->end)
1419 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
1422 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
1423 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1424 assert(SA->getUseBlocks().size() == 1 && "Not a local interval");
1425 const SplitAnalysis::BlockInfo &BI = SA->getUseBlocks().front();
1427 // Note that it is possible to have an interval that is live-in or live-out
1428 // while only covering a single block - A phi-def can use undef values from
1429 // predecessors, and the block could be a single-block loop.
1430 // We don't bother doing anything clever about such a case, we simply assume
1431 // that the interval is continuous from FirstInstr to LastInstr. We should
1432 // make sure that we don't do anything illegal to such an interval, though.
1434 ArrayRef<SlotIndex> Uses = SA->getUseSlots();
1435 if (Uses.size() <= 2)
1437 const unsigned NumGaps = Uses.size()-1;
1440 dbgs() << "tryLocalSplit: ";
1441 for (unsigned i = 0, e = Uses.size(); i != e; ++i)
1442 dbgs() << ' ' << Uses[i];
1446 // If VirtReg is live across any register mask operands, compute a list of
1447 // gaps with register masks.
1448 SmallVector<unsigned, 8> RegMaskGaps;
1449 if (Matrix->checkRegMaskInterference(VirtReg)) {
1450 // Get regmask slots for the whole block.
1451 ArrayRef<SlotIndex> RMS = LIS->getRegMaskSlotsInBlock(BI.MBB->getNumber());
1452 DEBUG(dbgs() << RMS.size() << " regmasks in block:");
1453 // Constrain to VirtReg's live range.
1454 unsigned ri = std::lower_bound(RMS.begin(), RMS.end(),
1455 Uses.front().getRegSlot()) - RMS.begin();
1456 unsigned re = RMS.size();
1457 for (unsigned i = 0; i != NumGaps && ri != re; ++i) {
1458 // Look for Uses[i] <= RMS <= Uses[i+1].
1459 assert(!SlotIndex::isEarlierInstr(RMS[ri], Uses[i]));
1460 if (SlotIndex::isEarlierInstr(Uses[i+1], RMS[ri]))
1462 // Skip a regmask on the same instruction as the last use. It doesn't
1463 // overlap the live range.
1464 if (SlotIndex::isSameInstr(Uses[i+1], RMS[ri]) && i+1 == NumGaps)
1466 DEBUG(dbgs() << ' ' << RMS[ri] << ':' << Uses[i] << '-' << Uses[i+1]);
1467 RegMaskGaps.push_back(i);
1468 // Advance ri to the next gap. A regmask on one of the uses counts in
1470 while (ri != re && SlotIndex::isEarlierInstr(RMS[ri], Uses[i+1]))
1473 DEBUG(dbgs() << '\n');
1476 // Since we allow local split results to be split again, there is a risk of
1477 // creating infinite loops. It is tempting to require that the new live
1478 // ranges have less instructions than the original. That would guarantee
1479 // convergence, but it is too strict. A live range with 3 instructions can be
1480 // split 2+3 (including the COPY), and we want to allow that.
1482 // Instead we use these rules:
1484 // 1. Allow any split for ranges with getStage() < RS_Split2. (Except for the
1485 // noop split, of course).
1486 // 2. Require progress be made for ranges with getStage() == RS_Split2. All
1487 // the new ranges must have fewer instructions than before the split.
1488 // 3. New ranges with the same number of instructions are marked RS_Split2,
1489 // smaller ranges are marked RS_New.
1491 // These rules allow a 3 -> 2+3 split once, which we need. They also prevent
1492 // excessive splitting and infinite loops.
1494 bool ProgressRequired = getStage(VirtReg) >= RS_Split2;
1496 // Best split candidate.
1497 unsigned BestBefore = NumGaps;
1498 unsigned BestAfter = 0;
1501 const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
1502 SmallVector<float, 8> GapWeight;
1505 while (unsigned PhysReg = Order.next()) {
1506 // Keep track of the largest spill weight that would need to be evicted in
1507 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
1508 calcGapWeights(PhysReg, GapWeight);
1510 // Remove any gaps with regmask clobbers.
1511 if (Matrix->checkRegMaskInterference(VirtReg, PhysReg))
1512 for (unsigned i = 0, e = RegMaskGaps.size(); i != e; ++i)
1513 GapWeight[RegMaskGaps[i]] = HUGE_VALF;
1515 // Try to find the best sequence of gaps to close.
1516 // The new spill weight must be larger than any gap interference.
1518 // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
1519 unsigned SplitBefore = 0, SplitAfter = 1;
1521 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
1522 // It is the spill weight that needs to be evicted.
1523 float MaxGap = GapWeight[0];
1526 // Live before/after split?
1527 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
1528 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
1530 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
1531 << Uses[SplitBefore] << '-' << Uses[SplitAfter]
1532 << " i=" << MaxGap);
1534 // Stop before the interval gets so big we wouldn't be making progress.
1535 if (!LiveBefore && !LiveAfter) {
1536 DEBUG(dbgs() << " all\n");
1539 // Should the interval be extended or shrunk?
1542 // How many gaps would the new range have?
1543 unsigned NewGaps = LiveBefore + SplitAfter - SplitBefore + LiveAfter;
1545 // Legally, without causing looping?
1546 bool Legal = !ProgressRequired || NewGaps < NumGaps;
1548 if (Legal && MaxGap < HUGE_VALF) {
1549 // Estimate the new spill weight. Each instruction reads or writes the
1550 // register. Conservatively assume there are no read-modify-write
1553 // Try to guess the size of the new interval.
1554 const float EstWeight = normalizeSpillWeight(blockFreq * (NewGaps + 1),
1555 Uses[SplitBefore].distance(Uses[SplitAfter]) +
1556 (LiveBefore + LiveAfter)*SlotIndex::InstrDist);
1557 // Would this split be possible to allocate?
1558 // Never allocate all gaps, we wouldn't be making progress.
1559 DEBUG(dbgs() << " w=" << EstWeight);
1560 if (EstWeight * Hysteresis >= MaxGap) {
1562 float Diff = EstWeight - MaxGap;
1563 if (Diff > BestDiff) {
1564 DEBUG(dbgs() << " (best)");
1565 BestDiff = Hysteresis * Diff;
1566 BestBefore = SplitBefore;
1567 BestAfter = SplitAfter;
1574 if (++SplitBefore < SplitAfter) {
1575 DEBUG(dbgs() << " shrink\n");
1576 // Recompute the max when necessary.
1577 if (GapWeight[SplitBefore - 1] >= MaxGap) {
1578 MaxGap = GapWeight[SplitBefore];
1579 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
1580 MaxGap = std::max(MaxGap, GapWeight[i]);
1587 // Try to extend the interval.
1588 if (SplitAfter >= NumGaps) {
1589 DEBUG(dbgs() << " end\n");
1593 DEBUG(dbgs() << " extend\n");
1594 MaxGap = std::max(MaxGap, GapWeight[SplitAfter++]);
1598 // Didn't find any candidates?
1599 if (BestBefore == NumGaps)
1602 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
1603 << '-' << Uses[BestAfter] << ", " << BestDiff
1604 << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
1606 LiveRangeEdit LREdit(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1610 SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
1611 SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]);
1612 SE->useIntv(SegStart, SegStop);
1613 SmallVector<unsigned, 8> IntvMap;
1614 SE->finish(&IntvMap);
1615 DebugVars->splitRegister(VirtReg.reg, LREdit.regs());
1617 // If the new range has the same number of instructions as before, mark it as
1618 // RS_Split2 so the next split will be forced to make progress. Otherwise,
1619 // leave the new intervals as RS_New so they can compete.
1620 bool LiveBefore = BestBefore != 0 || BI.LiveIn;
1621 bool LiveAfter = BestAfter != NumGaps || BI.LiveOut;
1622 unsigned NewGaps = LiveBefore + BestAfter - BestBefore + LiveAfter;
1623 if (NewGaps >= NumGaps) {
1624 DEBUG(dbgs() << "Tagging non-progress ranges: ");
1625 assert(!ProgressRequired && "Didn't make progress when it was required.");
1626 for (unsigned i = 0, e = IntvMap.size(); i != e; ++i)
1627 if (IntvMap[i] == 1) {
1628 setStage(*LREdit.get(i), RS_Split2);
1629 DEBUG(dbgs() << PrintReg(LREdit.get(i)->reg));
1631 DEBUG(dbgs() << '\n');
1638 //===----------------------------------------------------------------------===//
1639 // Live Range Splitting
1640 //===----------------------------------------------------------------------===//
1642 /// trySplit - Try to split VirtReg or one of its interferences, making it
1644 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
1645 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
1646 SmallVectorImpl<LiveInterval*>&NewVRegs) {
1647 // Ranges must be Split2 or less.
1648 if (getStage(VirtReg) >= RS_Spill)
1651 // Local intervals are handled separately.
1652 if (LIS->intervalIsInOneMBB(VirtReg)) {
1653 NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
1654 SA->analyze(&VirtReg);
1655 unsigned PhysReg = tryLocalSplit(VirtReg, Order, NewVRegs);
1656 if (PhysReg || !NewVRegs.empty())
1658 return tryInstructionSplit(VirtReg, Order, NewVRegs);
1661 NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
1663 SA->analyze(&VirtReg);
1665 // FIXME: SplitAnalysis may repair broken live ranges coming from the
1666 // coalescer. That may cause the range to become allocatable which means that
1667 // tryRegionSplit won't be making progress. This check should be replaced with
1668 // an assertion when the coalescer is fixed.
1669 if (SA->didRepairRange()) {
1670 // VirtReg has changed, so all cached queries are invalid.
1671 Matrix->invalidateVirtRegs();
1672 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1676 // First try to split around a region spanning multiple blocks. RS_Split2
1677 // ranges already made dubious progress with region splitting, so they go
1678 // straight to single block splitting.
1679 if (getStage(VirtReg) < RS_Split2) {
1680 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
1681 if (PhysReg || !NewVRegs.empty())
1685 // Then isolate blocks.
1686 return tryBlockSplit(VirtReg, Order, NewVRegs);
1690 //===----------------------------------------------------------------------===//
1692 //===----------------------------------------------------------------------===//
1694 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
1695 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1696 // First try assigning a free register.
1697 AllocationOrder Order(VirtReg.reg, *VRM, RegClassInfo);
1698 if (unsigned PhysReg = tryAssign(VirtReg, Order, NewVRegs))
1701 LiveRangeStage Stage = getStage(VirtReg);
1702 DEBUG(dbgs() << StageName[Stage]
1703 << " Cascade " << ExtraRegInfo[VirtReg.reg].Cascade << '\n');
1705 // Try to evict a less worthy live range, but only for ranges from the primary
1706 // queue. The RS_Split ranges already failed to do this, and they should not
1707 // get a second chance until they have been split.
1708 if (Stage != RS_Split)
1709 if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
1712 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
1714 // The first time we see a live range, don't try to split or spill.
1715 // Wait until the second time, when all smaller ranges have been allocated.
1716 // This gives a better picture of the interference to split around.
1717 if (Stage < RS_Split) {
1718 setStage(VirtReg, RS_Split);
1719 DEBUG(dbgs() << "wait for second round\n");
1720 NewVRegs.push_back(&VirtReg);
1724 // If we couldn't allocate a register from spilling, there is probably some
1725 // invalid inline assembly. The base class wil report it.
1726 if (Stage >= RS_Done || !VirtReg.isSpillable())
1729 // Try splitting VirtReg or interferences.
1730 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
1731 if (PhysReg || !NewVRegs.empty())
1734 // Finally spill VirtReg itself.
1735 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
1736 LiveRangeEdit LRE(&VirtReg, NewVRegs, *MF, *LIS, VRM, this);
1737 spiller().spill(LRE);
1738 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Done);
1741 MF->verify(this, "After spilling");
1743 // The live virtual register requesting allocation was spilled, so tell
1744 // the caller not to allocate anything during this round.
1748 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
1749 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
1750 << "********** Function: "
1751 << ((Value*)mf.getFunction())->getName() << '\n');
1755 MF->verify(this, "Before greedy register allocator");
1757 RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
1758 Matrix = &getAnalysis<LiveRegMatrix>();
1759 Indexes = &getAnalysis<SlotIndexes>();
1760 DomTree = &getAnalysis<MachineDominatorTree>();
1761 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
1762 Loops = &getAnalysis<MachineLoopInfo>();
1763 Bundles = &getAnalysis<EdgeBundles>();
1764 SpillPlacer = &getAnalysis<SpillPlacement>();
1765 DebugVars = &getAnalysis<LiveDebugVariables>();
1767 SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
1768 SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
1769 ExtraRegInfo.clear();
1770 ExtraRegInfo.resize(MRI->getNumVirtRegs());
1772 IntfCache.init(MF, Matrix->getLiveUnions(), Indexes, LIS, TRI);
1773 GlobalCand.resize(32); // This will grow as needed.