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 "LiveRangeEdit.h"
19 #include "RegAllocBase.h"
21 #include "SpillPlacement.h"
23 #include "VirtRegMap.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Function.h"
27 #include "llvm/PassAnalysisSupport.h"
28 #include "llvm/CodeGen/CalcSpillWeights.h"
29 #include "llvm/CodeGen/EdgeBundles.h"
30 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
31 #include "llvm/CodeGen/LiveStackAnalysis.h"
32 #include "llvm/CodeGen/MachineDominators.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineLoopRanges.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/CodeGen/RegAllocRegistry.h"
39 #include "llvm/CodeGen/RegisterCoalescer.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Support/Timer.h"
50 STATISTIC(NumGlobalSplits, "Number of split global live ranges");
51 STATISTIC(NumLocalSplits, "Number of split local live ranges");
52 STATISTIC(NumEvicted, "Number of interferences evicted");
54 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
55 createGreedyRegisterAllocator);
58 class RAGreedy : public MachineFunctionPass,
60 private LiveRangeEdit::Delegate {
64 BitVector ReservedRegs;
69 MachineDominatorTree *DomTree;
70 MachineLoopInfo *Loops;
71 MachineLoopRanges *LoopRanges;
73 SpillPlacement *SpillPlacer;
76 std::auto_ptr<Spiller> SpillerInstance;
77 std::priority_queue<std::pair<unsigned, unsigned> > Queue;
79 // Live ranges pass through a number of stages as we try to allocate them.
80 // Some of the stages may also create new live ranges:
82 // - Region splitting.
83 // - Per-block splitting.
87 // Ranges produced by one of the stages skip the previous stages when they are
88 // dequeued. This improves performance because we can skip interference checks
89 // that are unlikely to give any results. It also guarantees that the live
90 // range splitting algorithm terminates, something that is otherwise hard to
93 RS_New, ///< Never seen before.
94 RS_First, ///< First time in the queue.
95 RS_Second, ///< Second time in the queue.
96 RS_Region, ///< Produced by region splitting.
97 RS_Block, ///< Produced by per-block splitting.
98 RS_Local, ///< Produced by local splitting.
99 RS_Spill ///< Produced by spilling.
102 IndexedMap<unsigned char, VirtReg2IndexFunctor> LRStage;
104 LiveRangeStage getStage(const LiveInterval &VirtReg) const {
105 return LiveRangeStage(LRStage[VirtReg.reg]);
108 template<typename Iterator>
109 void setStage(Iterator Begin, Iterator End, LiveRangeStage NewStage) {
110 LRStage.resize(MRI->getNumVirtRegs());
111 for (;Begin != End; ++Begin) {
112 unsigned Reg = (*Begin)->reg;
113 if (LRStage[Reg] == RS_New)
114 LRStage[Reg] = NewStage;
119 std::auto_ptr<SplitAnalysis> SA;
120 std::auto_ptr<SplitEditor> SE;
122 /// Cached per-block interference maps
123 InterferenceCache IntfCache;
125 /// All basic blocks where the current register is live.
126 SmallVector<SpillPlacement::BlockConstraint, 8> SplitConstraints;
128 /// Global live range splitting candidate info.
129 struct GlobalSplitCandidate {
131 BitVector LiveBundles;
134 /// Candidate info for for each PhysReg in AllocationOrder.
135 /// This vector never shrinks, but grows to the size of the largest register
137 SmallVector<GlobalSplitCandidate, 32> GlobalCand;
139 /// For every instruction in SA->UseSlots, store the previous non-copy
141 SmallVector<SlotIndex, 8> PrevSlot;
146 /// Return the pass name.
147 virtual const char* getPassName() const {
148 return "Greedy Register Allocator";
151 /// RAGreedy analysis usage.
152 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
153 virtual void releaseMemory();
154 virtual Spiller &spiller() { return *SpillerInstance; }
155 virtual void enqueue(LiveInterval *LI);
156 virtual LiveInterval *dequeue();
157 virtual unsigned selectOrSplit(LiveInterval&,
158 SmallVectorImpl<LiveInterval*>&);
160 /// Perform register allocation.
161 virtual bool runOnMachineFunction(MachineFunction &mf);
166 void LRE_WillEraseInstruction(MachineInstr*);
167 bool LRE_CanEraseVirtReg(unsigned);
168 void LRE_WillShrinkVirtReg(unsigned);
169 void LRE_DidCloneVirtReg(unsigned, unsigned);
171 float calcSplitConstraints(unsigned);
172 float calcGlobalSplitCost(const BitVector&);
173 void splitAroundRegion(LiveInterval&, unsigned, const BitVector&,
174 SmallVectorImpl<LiveInterval*>&);
175 void calcGapWeights(unsigned, SmallVectorImpl<float>&);
176 SlotIndex getPrevMappedIndex(const MachineInstr*);
177 void calcPrevSlots();
178 unsigned nextSplitPoint(unsigned);
179 bool canEvictInterference(LiveInterval&, unsigned, float&);
181 unsigned tryEvict(LiveInterval&, AllocationOrder&,
182 SmallVectorImpl<LiveInterval*>&);
183 unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
184 SmallVectorImpl<LiveInterval*>&);
185 unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
186 SmallVectorImpl<LiveInterval*>&);
187 unsigned trySplit(LiveInterval&, AllocationOrder&,
188 SmallVectorImpl<LiveInterval*>&);
190 } // end anonymous namespace
192 char RAGreedy::ID = 0;
194 FunctionPass* llvm::createGreedyRegisterAllocator() {
195 return new RAGreedy();
198 RAGreedy::RAGreedy(): MachineFunctionPass(ID), LRStage(RS_New) {
199 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
200 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
201 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
202 initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
203 initializeRegisterCoalescerAnalysisGroup(*PassRegistry::getPassRegistry());
204 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
205 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
206 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
207 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
208 initializeMachineLoopRangesPass(*PassRegistry::getPassRegistry());
209 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
210 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
211 initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
214 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
215 AU.setPreservesCFG();
216 AU.addRequired<AliasAnalysis>();
217 AU.addPreserved<AliasAnalysis>();
218 AU.addRequired<LiveIntervals>();
219 AU.addRequired<SlotIndexes>();
220 AU.addPreserved<SlotIndexes>();
222 AU.addRequiredID(StrongPHIEliminationID);
223 AU.addRequiredTransitive<RegisterCoalescer>();
224 AU.addRequired<CalculateSpillWeights>();
225 AU.addRequired<LiveStacks>();
226 AU.addPreserved<LiveStacks>();
227 AU.addRequired<MachineDominatorTree>();
228 AU.addPreserved<MachineDominatorTree>();
229 AU.addRequired<MachineLoopInfo>();
230 AU.addPreserved<MachineLoopInfo>();
231 AU.addRequired<MachineLoopRanges>();
232 AU.addPreserved<MachineLoopRanges>();
233 AU.addRequired<VirtRegMap>();
234 AU.addPreserved<VirtRegMap>();
235 AU.addRequired<EdgeBundles>();
236 AU.addRequired<SpillPlacement>();
237 MachineFunctionPass::getAnalysisUsage(AU);
241 //===----------------------------------------------------------------------===//
242 // LiveRangeEdit delegate methods
243 //===----------------------------------------------------------------------===//
245 void RAGreedy::LRE_WillEraseInstruction(MachineInstr *MI) {
246 // LRE itself will remove from SlotIndexes and parent basic block.
247 VRM->RemoveMachineInstrFromMaps(MI);
250 bool RAGreedy::LRE_CanEraseVirtReg(unsigned VirtReg) {
251 if (unsigned PhysReg = VRM->getPhys(VirtReg)) {
252 unassign(LIS->getInterval(VirtReg), PhysReg);
255 // Unassigned virtreg is probably in the priority queue.
256 // RegAllocBase will erase it after dequeueing.
260 void RAGreedy::LRE_WillShrinkVirtReg(unsigned VirtReg) {
261 unsigned PhysReg = VRM->getPhys(VirtReg);
265 // Register is assigned, put it back on the queue for reassignment.
266 LiveInterval &LI = LIS->getInterval(VirtReg);
267 unassign(LI, PhysReg);
271 void RAGreedy::LRE_DidCloneVirtReg(unsigned New, unsigned Old) {
272 // LRE may clone a virtual register because dead code elimination causes it to
273 // be split into connected components. Ensure that the new register gets the
274 // same stage as the parent.
276 LRStage[New] = LRStage[Old];
279 void RAGreedy::releaseMemory() {
280 SpillerInstance.reset(0);
282 RegAllocBase::releaseMemory();
285 void RAGreedy::enqueue(LiveInterval *LI) {
286 // Prioritize live ranges by size, assigning larger ranges first.
287 // The queue holds (size, reg) pairs.
288 const unsigned Size = LI->getSize();
289 const unsigned Reg = LI->reg;
290 assert(TargetRegisterInfo::isVirtualRegister(Reg) &&
291 "Can only enqueue virtual registers");
295 if (LRStage[Reg] == RS_New)
296 LRStage[Reg] = RS_First;
298 if (LRStage[Reg] == RS_Second)
299 // Unsplit ranges that couldn't be allocated immediately are deferred until
300 // everything else has been allocated. Long ranges are allocated last so
301 // they are split against realistic interference.
302 Prio = (1u << 31) - Size;
304 // Everything else is allocated in long->short order. Long ranges that don't
305 // fit should be spilled ASAP so they don't create interference.
306 Prio = (1u << 31) + Size;
308 // Boost ranges that have a physical register hint.
309 if (TargetRegisterInfo::isPhysicalRegister(VRM->getRegAllocPref(Reg)))
313 Queue.push(std::make_pair(Prio, Reg));
316 LiveInterval *RAGreedy::dequeue() {
319 LiveInterval *LI = &LIS->getInterval(Queue.top().second);
324 //===----------------------------------------------------------------------===//
325 // Interference eviction
326 //===----------------------------------------------------------------------===//
328 /// canEvict - Return true if all interferences between VirtReg and PhysReg can
329 /// be evicted. Set maxWeight to the maximal spill weight of an interference.
330 bool RAGreedy::canEvictInterference(LiveInterval &VirtReg, unsigned PhysReg,
333 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
334 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
335 // If there is 10 or more interferences, chances are one is smaller.
336 if (Q.collectInterferingVRegs(10) >= 10)
339 // Check if any interfering live range is heavier than VirtReg.
340 for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
341 LiveInterval *Intf = Q.interferingVRegs()[i];
342 if (TargetRegisterInfo::isPhysicalRegister(Intf->reg))
344 if (Intf->weight >= VirtReg.weight)
346 Weight = std::max(Weight, Intf->weight);
353 /// tryEvict - Try to evict all interferences for a physreg.
354 /// @param VirtReg Currently unassigned virtual register.
355 /// @param Order Physregs to try.
356 /// @return Physreg to assign VirtReg, or 0.
357 unsigned RAGreedy::tryEvict(LiveInterval &VirtReg,
358 AllocationOrder &Order,
359 SmallVectorImpl<LiveInterval*> &NewVRegs){
360 NamedRegionTimer T("Evict", TimerGroupName, TimePassesIsEnabled);
362 // Keep track of the lightest single interference seen so far.
363 float BestWeight = 0;
364 unsigned BestPhys = 0;
367 while (unsigned PhysReg = Order.next()) {
369 if (!canEvictInterference(VirtReg, PhysReg, Weight))
372 // This is an eviction candidate.
373 DEBUG(dbgs() << "max " << PrintReg(PhysReg, TRI) << " interference = "
375 if (BestPhys && Weight >= BestWeight)
381 // Stop if the hint can be used.
382 if (Order.isHint(PhysReg))
389 DEBUG(dbgs() << "evicting " << PrintReg(BestPhys, TRI) << " interference\n");
390 for (const unsigned *AliasI = TRI->getOverlaps(BestPhys); *AliasI; ++AliasI) {
391 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
392 assert(Q.seenAllInterferences() && "Didn't check all interfererences.");
393 for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
394 LiveInterval *Intf = Q.interferingVRegs()[i];
395 unassign(*Intf, VRM->getPhys(Intf->reg));
397 NewVRegs.push_back(Intf);
404 //===----------------------------------------------------------------------===//
406 //===----------------------------------------------------------------------===//
408 /// calcSplitConstraints - Fill out the SplitConstraints vector based on the
409 /// interference pattern in Physreg and its aliases. Return the static cost of
410 /// this split, assuming that all preferences in SplitConstraints are met.
411 float RAGreedy::calcSplitConstraints(unsigned PhysReg) {
412 InterferenceCache::Cursor Intf(IntfCache, PhysReg);
414 // Reset interference dependent info.
415 SplitConstraints.resize(SA->LiveBlocks.size());
416 float StaticCost = 0;
417 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
418 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
419 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
421 BC.Number = BI.MBB->getNumber();
422 Intf.moveToBlock(BC.Number);
423 BC.Entry = (BI.Uses && BI.LiveIn) ?
424 SpillPlacement::PrefReg : SpillPlacement::DontCare;
425 BC.Exit = (BI.Uses && BI.LiveOut) ?
426 SpillPlacement::PrefReg : SpillPlacement::DontCare;
428 if (!Intf.hasInterference())
431 // Number of spill code instructions to insert.
434 // Interference for the live-in value.
436 if (Intf.first() <= Indexes->getMBBStartIdx(BC.Number))
437 BC.Entry = SpillPlacement::MustSpill, Ins += BI.Uses;
439 BC.Entry = SpillPlacement::PrefSpill;
440 else if (Intf.first() < BI.FirstUse)
441 BC.Entry = SpillPlacement::PrefSpill, ++Ins;
442 else if (Intf.first() < (BI.LiveThrough ? BI.LastUse : BI.Kill))
446 // Interference for the live-out value.
448 if (Intf.last() >= SA->getLastSplitPoint(BC.Number))
449 BC.Exit = SpillPlacement::MustSpill, Ins += BI.Uses;
451 BC.Exit = SpillPlacement::PrefSpill;
452 else if (Intf.last() > BI.LastUse)
453 BC.Exit = SpillPlacement::PrefSpill, ++Ins;
454 else if (Intf.last() > (BI.LiveThrough ? BI.FirstUse : BI.Def))
458 // Accumulate the total frequency of inserted spill code.
460 StaticCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
466 /// calcGlobalSplitCost - Return the global split cost of following the split
467 /// pattern in LiveBundles. This cost should be added to the local cost of the
468 /// interference pattern in SplitConstraints.
470 float RAGreedy::calcGlobalSplitCost(const BitVector &LiveBundles) {
471 float GlobalCost = 0;
472 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
473 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
474 SpillPlacement::BlockConstraint &BC = SplitConstraints[i];
475 bool RegIn = LiveBundles[Bundles->getBundle(BC.Number, 0)];
476 bool RegOut = LiveBundles[Bundles->getBundle(BC.Number, 1)];
480 Ins += RegIn != RegOut;
483 Ins += RegIn != (BC.Entry == SpillPlacement::PrefReg);
485 Ins += RegOut != (BC.Exit == SpillPlacement::PrefReg);
488 GlobalCost += Ins * SpillPlacer->getBlockFrequency(BC.Number);
493 /// splitAroundRegion - Split VirtReg around the region determined by
494 /// LiveBundles. Make an effort to avoid interference from PhysReg.
496 /// The 'register' interval is going to contain as many uses as possible while
497 /// avoiding interference. The 'stack' interval is the complement constructed by
498 /// SplitEditor. It will contain the rest.
500 void RAGreedy::splitAroundRegion(LiveInterval &VirtReg, unsigned PhysReg,
501 const BitVector &LiveBundles,
502 SmallVectorImpl<LiveInterval*> &NewVRegs) {
504 dbgs() << "Splitting around region for " << PrintReg(PhysReg, TRI)
506 for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
507 dbgs() << " EB#" << i;
511 InterferenceCache::Cursor Intf(IntfCache, PhysReg);
512 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
515 // Create the main cross-block interval.
518 // First add all defs that are live out of a block.
519 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
520 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
521 bool RegIn = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
522 bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
524 // Should the register be live out?
525 if (!BI.LiveOut || !RegOut)
528 SlotIndex Start, Stop;
529 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
530 Intf.moveToBlock(BI.MBB->getNumber());
531 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " -> EB#"
532 << Bundles->getBundle(BI.MBB->getNumber(), 1)
533 << " [" << Start << ';'
534 << SA->getLastSplitPoint(BI.MBB->getNumber()) << '-' << Stop
535 << ") intf [" << Intf.first() << ';' << Intf.last() << ')');
537 // The interference interval should either be invalid or overlap MBB.
538 assert((!Intf.hasInterference() || Intf.first() < Stop)
539 && "Bad interference");
540 assert((!Intf.hasInterference() || Intf.last() > Start)
541 && "Bad interference");
543 // Check interference leaving the block.
544 if (!Intf.hasInterference()) {
545 // Block is interference-free.
546 DEBUG(dbgs() << ", no interference");
548 assert(BI.LiveThrough && "No uses, but not live through block?");
549 // Block is live-through without interference.
550 DEBUG(dbgs() << ", no uses"
551 << (RegIn ? ", live-through.\n" : ", stack in.\n"));
553 SE->enterIntvAtEnd(*BI.MBB);
556 if (!BI.LiveThrough) {
557 DEBUG(dbgs() << ", not live-through.\n");
558 SE->useIntv(SE->enterIntvBefore(BI.Def), Stop);
562 // Block is live-through, but entry bundle is on the stack.
563 // Reload just before the first use.
564 DEBUG(dbgs() << ", not live-in, enter before first use.\n");
565 SE->useIntv(SE->enterIntvBefore(BI.FirstUse), Stop);
568 DEBUG(dbgs() << ", live-through.\n");
572 // Block has interference.
573 DEBUG(dbgs() << ", interference to " << Intf.last());
575 if (!BI.LiveThrough && Intf.last() <= BI.Def) {
576 // The interference doesn't reach the outgoing segment.
577 DEBUG(dbgs() << " doesn't affect def from " << BI.Def << '\n');
578 SE->useIntv(BI.Def, Stop);
584 // No uses in block, avoid interference by reloading as late as possible.
585 DEBUG(dbgs() << ", no uses.\n");
586 SlotIndex SegStart = SE->enterIntvAtEnd(*BI.MBB);
587 assert(SegStart >= Intf.last() && "Couldn't avoid interference");
591 SlotIndex LastSplitPoint = SA->getLastSplitPoint(BI.MBB->getNumber());
592 if (Intf.last().getBoundaryIndex() < BI.LastUse) {
593 // There are interference-free uses at the end of the block.
594 // Find the first use that can get the live-out register.
595 SmallVectorImpl<SlotIndex>::const_iterator UI =
596 std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
597 Intf.last().getBoundaryIndex());
598 assert(UI != SA->UseSlots.end() && "Couldn't find last use");
600 assert(Use <= BI.LastUse && "Couldn't find last use");
601 // Only attempt a split befroe the last split point.
602 if (Use.getBaseIndex() <= LastSplitPoint) {
603 DEBUG(dbgs() << ", free use at " << Use << ".\n");
604 SlotIndex SegStart = SE->enterIntvBefore(Use);
605 assert(SegStart >= Intf.last() && "Couldn't avoid interference");
606 assert(SegStart < LastSplitPoint && "Impossible split point");
607 SE->useIntv(SegStart, Stop);
612 // Interference is after the last use.
613 DEBUG(dbgs() << " after last use.\n");
614 SlotIndex SegStart = SE->enterIntvAtEnd(*BI.MBB);
615 assert(SegStart >= Intf.last() && "Couldn't avoid interference");
618 // Now all defs leading to live bundles are handled, do everything else.
619 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
620 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
621 bool RegIn = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
622 bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
624 // Is the register live-in?
625 if (!BI.LiveIn || !RegIn)
628 // We have an incoming register. Check for interference.
629 SlotIndex Start, Stop;
630 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
631 Intf.moveToBlock(BI.MBB->getNumber());
632 DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
633 << " -> BB#" << BI.MBB->getNumber() << " [" << Start << ';'
634 << SA->getLastSplitPoint(BI.MBB->getNumber()) << '-' << Stop
637 // Check interference entering the block.
638 if (!Intf.hasInterference()) {
639 // Block is interference-free.
640 DEBUG(dbgs() << ", no interference");
642 assert(BI.LiveThrough && "No uses, but not live through block?");
643 // Block is live-through without interference.
645 DEBUG(dbgs() << ", no uses, live-through.\n");
646 SE->useIntv(Start, Stop);
648 DEBUG(dbgs() << ", no uses, stack-out.\n");
649 SE->leaveIntvAtTop(*BI.MBB);
653 if (!BI.LiveThrough) {
654 DEBUG(dbgs() << ", killed in block.\n");
655 SE->useIntv(Start, SE->leaveIntvAfter(BI.Kill));
659 SlotIndex LastSplitPoint = SA->getLastSplitPoint(BI.MBB->getNumber());
660 // Block is live-through, but exit bundle is on the stack.
661 // Spill immediately after the last use.
662 if (BI.LastUse < LastSplitPoint) {
663 DEBUG(dbgs() << ", uses, stack-out.\n");
664 SE->useIntv(Start, SE->leaveIntvAfter(BI.LastUse));
667 // The last use is after the last split point, it is probably an
669 DEBUG(dbgs() << ", uses at " << BI.LastUse << " after split point "
670 << LastSplitPoint << ", stack-out.\n");
671 SlotIndex SegEnd = SE->leaveIntvBefore(LastSplitPoint);
672 SE->useIntv(Start, SegEnd);
673 // Run a double interval from the split to the last use.
674 // This makes it possible to spill the complement without affecting the
676 SE->overlapIntv(SegEnd, BI.LastUse);
679 // Register is live-through.
680 DEBUG(dbgs() << ", uses, live-through.\n");
681 SE->useIntv(Start, Stop);
685 // Block has interference.
686 DEBUG(dbgs() << ", interference from " << Intf.first());
688 if (!BI.LiveThrough && Intf.first() >= BI.Kill) {
689 // The interference doesn't reach the outgoing segment.
690 DEBUG(dbgs() << " doesn't affect kill at " << BI.Kill << '\n');
691 SE->useIntv(Start, BI.Kill);
696 // No uses in block, avoid interference by spilling as soon as possible.
697 DEBUG(dbgs() << ", no uses.\n");
698 SlotIndex SegEnd = SE->leaveIntvAtTop(*BI.MBB);
699 assert(SegEnd <= Intf.first() && "Couldn't avoid interference");
702 if (Intf.first().getBaseIndex() > BI.FirstUse) {
703 // There are interference-free uses at the beginning of the block.
704 // Find the last use that can get the register.
705 SmallVectorImpl<SlotIndex>::const_iterator UI =
706 std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
707 Intf.first().getBaseIndex());
708 assert(UI != SA->UseSlots.begin() && "Couldn't find first use");
709 SlotIndex Use = (--UI)->getBoundaryIndex();
710 DEBUG(dbgs() << ", free use at " << *UI << ".\n");
711 SlotIndex SegEnd = SE->leaveIntvAfter(Use);
712 assert(SegEnd <= Intf.first() && "Couldn't avoid interference");
713 SE->useIntv(Start, SegEnd);
717 // Interference is before the first use.
718 DEBUG(dbgs() << " before first use.\n");
719 SlotIndex SegEnd = SE->leaveIntvAtTop(*BI.MBB);
720 assert(SegEnd <= Intf.first() && "Couldn't avoid interference");
725 // FIXME: Should we be more aggressive about splitting the stack region into
726 // per-block segments? The current approach allows the stack region to
727 // separate into connected components. Some components may be allocatable.
732 MF->verify(this, "After splitting live range around region");
735 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
736 SmallVectorImpl<LiveInterval*> &NewVRegs) {
737 BitVector LiveBundles, BestBundles;
739 unsigned BestReg = 0;
742 for (unsigned Cand = 0; unsigned PhysReg = Order.next(); ++Cand) {
743 if (GlobalCand.size() <= Cand)
744 GlobalCand.resize(Cand+1);
745 GlobalCand[Cand].PhysReg = PhysReg;
747 float Cost = calcSplitConstraints(PhysReg);
748 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << "\tstatic = " << Cost);
749 if (BestReg && Cost >= BestCost) {
750 DEBUG(dbgs() << " higher.\n");
754 SpillPlacer->placeSpills(SplitConstraints, LiveBundles);
755 // No live bundles, defer to splitSingleBlocks().
756 if (!LiveBundles.any()) {
757 DEBUG(dbgs() << " no bundles.\n");
761 Cost += calcGlobalSplitCost(LiveBundles);
763 dbgs() << ", total = " << Cost << " with bundles";
764 for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
765 dbgs() << " EB#" << i;
768 if (!BestReg || Cost < BestCost) {
770 BestCost = 0.98f * Cost; // Prevent rounding effects.
771 BestBundles.swap(LiveBundles);
778 splitAroundRegion(VirtReg, BestReg, BestBundles, NewVRegs);
779 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Region);
784 //===----------------------------------------------------------------------===//
786 //===----------------------------------------------------------------------===//
789 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted
790 /// in order to use PhysReg between two entries in SA->UseSlots.
792 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
794 void RAGreedy::calcGapWeights(unsigned PhysReg,
795 SmallVectorImpl<float> &GapWeight) {
796 assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
797 const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
798 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
799 const unsigned NumGaps = Uses.size()-1;
801 // Start and end points for the interference check.
802 SlotIndex StartIdx = BI.LiveIn ? BI.FirstUse.getBaseIndex() : BI.FirstUse;
803 SlotIndex StopIdx = BI.LiveOut ? BI.LastUse.getBoundaryIndex() : BI.LastUse;
805 GapWeight.assign(NumGaps, 0.0f);
807 // Add interference from each overlapping register.
808 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
809 if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
810 .checkInterference())
813 // We know that VirtReg is a continuous interval from FirstUse to LastUse,
814 // so we don't need InterferenceQuery.
816 // Interference that overlaps an instruction is counted in both gaps
817 // surrounding the instruction. The exception is interference before
818 // StartIdx and after StopIdx.
820 LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
821 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
822 // Skip the gaps before IntI.
823 while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
824 if (++Gap == NumGaps)
829 // Update the gaps covered by IntI.
830 const float weight = IntI.value()->weight;
831 for (; Gap != NumGaps; ++Gap) {
832 GapWeight[Gap] = std::max(GapWeight[Gap], weight);
833 if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
842 /// getPrevMappedIndex - Return the slot index of the last non-copy instruction
843 /// before MI that has a slot index. If MI is the first mapped instruction in
844 /// its block, return the block start index instead.
846 SlotIndex RAGreedy::getPrevMappedIndex(const MachineInstr *MI) {
847 assert(MI && "Missing MachineInstr");
848 const MachineBasicBlock *MBB = MI->getParent();
849 MachineBasicBlock::const_iterator B = MBB->begin(), I = MI;
851 if (!(--I)->isDebugValue() && !I->isCopy())
852 return Indexes->getInstructionIndex(I);
853 return Indexes->getMBBStartIdx(MBB);
856 /// calcPrevSlots - Fill in the PrevSlot array with the index of the previous
857 /// real non-copy instruction for each instruction in SA->UseSlots.
859 void RAGreedy::calcPrevSlots() {
860 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
862 PrevSlot.reserve(Uses.size());
863 for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
864 const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]);
865 PrevSlot.push_back(getPrevMappedIndex(MI).getDefIndex());
869 /// nextSplitPoint - Find the next index into SA->UseSlots > i such that it may
870 /// be beneficial to split before UseSlots[i].
872 /// 0 is always a valid split point
873 unsigned RAGreedy::nextSplitPoint(unsigned i) {
874 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
875 const unsigned Size = Uses.size();
876 assert(i != Size && "No split points after the end");
877 // Allow split before i when Uses[i] is not adjacent to the previous use.
878 while (++i != Size && PrevSlot[i].getBaseIndex() <= Uses[i-1].getBaseIndex())
883 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
886 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
887 SmallVectorImpl<LiveInterval*> &NewVRegs) {
888 assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
889 const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
891 // Note that it is possible to have an interval that is live-in or live-out
892 // while only covering a single block - A phi-def can use undef values from
893 // predecessors, and the block could be a single-block loop.
894 // We don't bother doing anything clever about such a case, we simply assume
895 // that the interval is continuous from FirstUse to LastUse. We should make
896 // sure that we don't do anything illegal to such an interval, though.
898 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
899 if (Uses.size() <= 2)
901 const unsigned NumGaps = Uses.size()-1;
904 dbgs() << "tryLocalSplit: ";
905 for (unsigned i = 0, e = Uses.size(); i != e; ++i)
906 dbgs() << ' ' << SA->UseSlots[i];
910 // For every use, find the previous mapped non-copy instruction.
911 // We use this to detect valid split points, and to estimate new interval
915 unsigned BestBefore = NumGaps;
916 unsigned BestAfter = 0;
919 const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB->getNumber());
920 SmallVector<float, 8> GapWeight;
923 while (unsigned PhysReg = Order.next()) {
924 // Keep track of the largest spill weight that would need to be evicted in
925 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
926 calcGapWeights(PhysReg, GapWeight);
928 // Try to find the best sequence of gaps to close.
929 // The new spill weight must be larger than any gap interference.
931 // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
932 unsigned SplitBefore = 0, SplitAfter = nextSplitPoint(1) - 1;
934 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
935 // It is the spill weight that needs to be evicted.
936 float MaxGap = GapWeight[0];
937 for (unsigned i = 1; i != SplitAfter; ++i)
938 MaxGap = std::max(MaxGap, GapWeight[i]);
941 // Live before/after split?
942 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
943 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
945 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
946 << Uses[SplitBefore] << '-' << Uses[SplitAfter]
949 // Stop before the interval gets so big we wouldn't be making progress.
950 if (!LiveBefore && !LiveAfter) {
951 DEBUG(dbgs() << " all\n");
954 // Should the interval be extended or shrunk?
956 if (MaxGap < HUGE_VALF) {
957 // Estimate the new spill weight.
959 // Each instruction reads and writes the register, except the first
960 // instr doesn't read when !FirstLive, and the last instr doesn't write
963 // We will be inserting copies before and after, so the total number of
964 // reads and writes is 2 * EstUses.
966 const unsigned EstUses = 2*(SplitAfter - SplitBefore) +
967 2*(LiveBefore + LiveAfter);
969 // Try to guess the size of the new interval. This should be trivial,
970 // but the slot index of an inserted copy can be a lot smaller than the
971 // instruction it is inserted before if there are many dead indexes
974 // We measure the distance from the instruction before SplitBefore to
975 // get a conservative estimate.
977 // The final distance can still be different if inserting copies
978 // triggers a slot index renumbering.
980 const float EstWeight = normalizeSpillWeight(blockFreq * EstUses,
981 PrevSlot[SplitBefore].distance(Uses[SplitAfter]));
982 // Would this split be possible to allocate?
983 // Never allocate all gaps, we wouldn't be making progress.
984 float Diff = EstWeight - MaxGap;
985 DEBUG(dbgs() << " w=" << EstWeight << " d=" << Diff);
988 if (Diff > BestDiff) {
989 DEBUG(dbgs() << " (best)");
991 BestBefore = SplitBefore;
992 BestAfter = SplitAfter;
999 SplitBefore = nextSplitPoint(SplitBefore);
1000 if (SplitBefore < SplitAfter) {
1001 DEBUG(dbgs() << " shrink\n");
1002 // Recompute the max when necessary.
1003 if (GapWeight[SplitBefore - 1] >= MaxGap) {
1004 MaxGap = GapWeight[SplitBefore];
1005 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
1006 MaxGap = std::max(MaxGap, GapWeight[i]);
1013 // Try to extend the interval.
1014 if (SplitAfter >= NumGaps) {
1015 DEBUG(dbgs() << " end\n");
1019 DEBUG(dbgs() << " extend\n");
1020 for (unsigned e = nextSplitPoint(SplitAfter + 1) - 1;
1021 SplitAfter != e; ++SplitAfter)
1022 MaxGap = std::max(MaxGap, GapWeight[SplitAfter]);
1027 // Didn't find any candidates?
1028 if (BestBefore == NumGaps)
1031 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
1032 << '-' << Uses[BestAfter] << ", " << BestDiff
1033 << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
1035 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1039 SlotIndex SegStart = SE->enterIntvBefore(Uses[BestBefore]);
1040 SlotIndex SegStop = SE->leaveIntvAfter(Uses[BestAfter]);
1041 SE->useIntv(SegStart, SegStop);
1044 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Local);
1050 //===----------------------------------------------------------------------===//
1051 // Live Range Splitting
1052 //===----------------------------------------------------------------------===//
1054 /// trySplit - Try to split VirtReg or one of its interferences, making it
1056 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
1057 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
1058 SmallVectorImpl<LiveInterval*>&NewVRegs) {
1059 // Local intervals are handled separately.
1060 if (LIS->intervalIsInOneMBB(VirtReg)) {
1061 NamedRegionTimer T("Local Splitting", TimerGroupName, TimePassesIsEnabled);
1062 SA->analyze(&VirtReg);
1063 return tryLocalSplit(VirtReg, Order, NewVRegs);
1066 NamedRegionTimer T("Global Splitting", TimerGroupName, TimePassesIsEnabled);
1068 // Don't iterate global splitting.
1069 // Move straight to spilling if this range was produced by a global split.
1070 LiveRangeStage Stage = getStage(VirtReg);
1071 if (Stage >= RS_Block)
1074 SA->analyze(&VirtReg);
1076 // First try to split around a region spanning multiple blocks.
1077 if (Stage < RS_Region) {
1078 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
1079 if (PhysReg || !NewVRegs.empty())
1083 // Then isolate blocks with multiple uses.
1084 if (Stage < RS_Block) {
1085 SplitAnalysis::BlockPtrSet Blocks;
1086 if (SA->getMultiUseBlocks(Blocks)) {
1087 LiveRangeEdit LREdit(VirtReg, NewVRegs, this);
1089 SE->splitSingleBlocks(Blocks);
1090 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Block);
1092 MF->verify(this, "After splitting live range around basic blocks");
1096 // Don't assign any physregs.
1101 //===----------------------------------------------------------------------===//
1103 //===----------------------------------------------------------------------===//
1105 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
1106 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1107 // First try assigning a free register.
1108 AllocationOrder Order(VirtReg.reg, *VRM, ReservedRegs);
1109 while (unsigned PhysReg = Order.next()) {
1110 if (!checkPhysRegInterference(VirtReg, PhysReg))
1114 if (unsigned PhysReg = tryEvict(VirtReg, Order, NewVRegs))
1117 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
1119 // The first time we see a live range, don't try to split or spill.
1120 // Wait until the second time, when all smaller ranges have been allocated.
1121 // This gives a better picture of the interference to split around.
1122 LiveRangeStage Stage = getStage(VirtReg);
1123 if (Stage == RS_First) {
1124 LRStage[VirtReg.reg] = RS_Second;
1125 DEBUG(dbgs() << "wait for second round\n");
1126 NewVRegs.push_back(&VirtReg);
1130 assert(Stage < RS_Spill && "Cannot allocate after spilling");
1132 // Try splitting VirtReg or interferences.
1133 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
1134 if (PhysReg || !NewVRegs.empty())
1137 // Finally spill VirtReg itself.
1138 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
1139 LiveRangeEdit LRE(VirtReg, NewVRegs, this);
1140 spiller().spill(LRE);
1141 setStage(NewVRegs.begin(), NewVRegs.end(), RS_Spill);
1144 MF->verify(this, "After spilling");
1146 // The live virtual register requesting allocation was spilled, so tell
1147 // the caller not to allocate anything during this round.
1151 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
1152 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
1153 << "********** Function: "
1154 << ((Value*)mf.getFunction())->getName() << '\n');
1158 MF->verify(this, "Before greedy register allocator");
1160 RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
1161 Indexes = &getAnalysis<SlotIndexes>();
1162 DomTree = &getAnalysis<MachineDominatorTree>();
1163 ReservedRegs = TRI->getReservedRegs(*MF);
1164 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
1165 Loops = &getAnalysis<MachineLoopInfo>();
1166 LoopRanges = &getAnalysis<MachineLoopRanges>();
1167 Bundles = &getAnalysis<EdgeBundles>();
1168 SpillPlacer = &getAnalysis<SpillPlacement>();
1170 SA.reset(new SplitAnalysis(*VRM, *LIS, *Loops));
1171 SE.reset(new SplitEditor(*SA, *LIS, *VRM, *DomTree));
1173 LRStage.resize(MRI->getNumVirtRegs());
1174 IntfCache.init(MF, &PhysReg2LiveUnion[0], Indexes, TRI);
1178 LIS->addKillFlags();
1182 NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
1183 VRM->rewrite(Indexes);
1186 // The pass output is in VirtRegMap. Release all the transient data.