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 "LiveIntervalUnion.h"
18 #include "LiveRangeEdit.h"
19 #include "RegAllocBase.h"
21 #include "SpillPlacement.h"
23 #include "VirtRegMap.h"
24 #include "VirtRegRewriter.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"
48 static RegisterRegAlloc greedyRegAlloc("greedy", "greedy register allocator",
49 createGreedyRegisterAllocator);
52 class RAGreedy : public MachineFunctionPass, public RegAllocBase {
55 BitVector ReservedRegs;
60 MachineDominatorTree *DomTree;
61 MachineLoopInfo *Loops;
62 MachineLoopRanges *LoopRanges;
64 SpillPlacement *SpillPlacer;
67 std::auto_ptr<Spiller> SpillerInstance;
68 std::auto_ptr<SplitAnalysis> SA;
72 /// All basic blocks where the current register is live.
73 SmallVector<SpillPlacement::BlockConstraint, 8> SpillConstraints;
75 /// For every instruction in SA->UseSlots, store the previous non-copy
77 SmallVector<SlotIndex, 8> PrevSlot;
82 /// Return the pass name.
83 virtual const char* getPassName() const {
84 return "Greedy Register Allocator";
87 /// RAGreedy analysis usage.
88 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
90 virtual void releaseMemory();
92 virtual Spiller &spiller() { return *SpillerInstance; }
94 virtual float getPriority(LiveInterval *LI);
96 virtual unsigned selectOrSplit(LiveInterval&,
97 SmallVectorImpl<LiveInterval*>&);
99 /// Perform register allocation.
100 virtual bool runOnMachineFunction(MachineFunction &mf);
105 bool checkUncachedInterference(LiveInterval&, unsigned);
106 LiveInterval *getSingleInterference(LiveInterval&, unsigned);
107 bool reassignVReg(LiveInterval &InterferingVReg, unsigned OldPhysReg);
108 float calcInterferenceWeight(LiveInterval&, unsigned);
109 float calcInterferenceInfo(LiveInterval&, unsigned);
110 float calcGlobalSplitCost(const BitVector&);
111 void splitAroundRegion(LiveInterval&, unsigned, const BitVector&,
112 SmallVectorImpl<LiveInterval*>&);
113 void calcGapWeights(unsigned, SmallVectorImpl<float>&);
114 SlotIndex getPrevMappedIndex(const MachineInstr*);
115 void calcPrevSlots();
116 unsigned nextSplitPoint(unsigned);
118 unsigned tryReassignOrEvict(LiveInterval&, AllocationOrder&,
119 SmallVectorImpl<LiveInterval*>&);
120 unsigned tryRegionSplit(LiveInterval&, AllocationOrder&,
121 SmallVectorImpl<LiveInterval*>&);
122 unsigned tryLocalSplit(LiveInterval&, AllocationOrder&,
123 SmallVectorImpl<LiveInterval*>&);
124 unsigned trySplit(LiveInterval&, AllocationOrder&,
125 SmallVectorImpl<LiveInterval*>&);
126 unsigned trySpillInterferences(LiveInterval&, AllocationOrder&,
127 SmallVectorImpl<LiveInterval*>&);
129 } // end anonymous namespace
131 char RAGreedy::ID = 0;
133 FunctionPass* llvm::createGreedyRegisterAllocator() {
134 return new RAGreedy();
137 RAGreedy::RAGreedy(): MachineFunctionPass(ID) {
138 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
139 initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
140 initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
141 initializeStrongPHIEliminationPass(*PassRegistry::getPassRegistry());
142 initializeRegisterCoalescerAnalysisGroup(*PassRegistry::getPassRegistry());
143 initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
144 initializeLiveStacksPass(*PassRegistry::getPassRegistry());
145 initializeMachineDominatorTreePass(*PassRegistry::getPassRegistry());
146 initializeMachineLoopInfoPass(*PassRegistry::getPassRegistry());
147 initializeMachineLoopRangesPass(*PassRegistry::getPassRegistry());
148 initializeVirtRegMapPass(*PassRegistry::getPassRegistry());
149 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
150 initializeSpillPlacementPass(*PassRegistry::getPassRegistry());
153 void RAGreedy::getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.setPreservesCFG();
155 AU.addRequired<AliasAnalysis>();
156 AU.addPreserved<AliasAnalysis>();
157 AU.addRequired<LiveIntervals>();
158 AU.addRequired<SlotIndexes>();
159 AU.addPreserved<SlotIndexes>();
161 AU.addRequiredID(StrongPHIEliminationID);
162 AU.addRequiredTransitive<RegisterCoalescer>();
163 AU.addRequired<CalculateSpillWeights>();
164 AU.addRequired<LiveStacks>();
165 AU.addPreserved<LiveStacks>();
166 AU.addRequired<MachineDominatorTree>();
167 AU.addPreserved<MachineDominatorTree>();
168 AU.addRequired<MachineLoopInfo>();
169 AU.addPreserved<MachineLoopInfo>();
170 AU.addRequired<MachineLoopRanges>();
171 AU.addPreserved<MachineLoopRanges>();
172 AU.addRequired<VirtRegMap>();
173 AU.addPreserved<VirtRegMap>();
174 AU.addRequired<EdgeBundles>();
175 AU.addRequired<SpillPlacement>();
176 MachineFunctionPass::getAnalysisUsage(AU);
179 void RAGreedy::releaseMemory() {
180 SpillerInstance.reset(0);
181 RegAllocBase::releaseMemory();
184 float RAGreedy::getPriority(LiveInterval *LI) {
185 float Priority = LI->weight;
187 // Prioritize hinted registers so they are allocated first.
188 std::pair<unsigned, unsigned> Hint;
189 if (Hint.first || Hint.second) {
190 // The hint can be target specific, a virtual register, or a physreg.
193 // Prefer physreg hints above anything else.
194 if (Hint.first == 0 && TargetRegisterInfo::isPhysicalRegister(Hint.second))
201 //===----------------------------------------------------------------------===//
202 // Register Reassignment
203 //===----------------------------------------------------------------------===//
205 // Check interference without using the cache.
206 bool RAGreedy::checkUncachedInterference(LiveInterval &VirtReg,
208 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
209 LiveIntervalUnion::Query subQ(&VirtReg, &PhysReg2LiveUnion[*AliasI]);
210 if (subQ.checkInterference())
216 /// getSingleInterference - Return the single interfering virtual register
217 /// assigned to PhysReg. Return 0 if more than one virtual register is
219 LiveInterval *RAGreedy::getSingleInterference(LiveInterval &VirtReg,
221 // Check physreg and aliases.
222 LiveInterval *Interference = 0;
223 for (const unsigned *AliasI = TRI->getOverlaps(PhysReg); *AliasI; ++AliasI) {
224 LiveIntervalUnion::Query &Q = query(VirtReg, *AliasI);
225 if (Q.checkInterference()) {
228 Q.collectInterferingVRegs(1);
229 if (!Q.seenAllInterferences())
231 Interference = Q.interferingVRegs().front();
237 // Attempt to reassign this virtual register to a different physical register.
239 // FIXME: we are not yet caching these "second-level" interferences discovered
240 // in the sub-queries. These interferences can change with each call to
241 // selectOrSplit. However, we could implement a "may-interfere" cache that
242 // could be conservatively dirtied when we reassign or split.
244 // FIXME: This may result in a lot of alias queries. We could summarize alias
245 // live intervals in their parent register's live union, but it's messy.
246 bool RAGreedy::reassignVReg(LiveInterval &InterferingVReg,
247 unsigned WantedPhysReg) {
248 assert(TargetRegisterInfo::isVirtualRegister(InterferingVReg.reg) &&
249 "Can only reassign virtual registers");
250 assert(TRI->regsOverlap(WantedPhysReg, VRM->getPhys(InterferingVReg.reg)) &&
251 "inconsistent phys reg assigment");
253 AllocationOrder Order(InterferingVReg.reg, *VRM, ReservedRegs);
254 while (unsigned PhysReg = Order.next()) {
255 // Don't reassign to a WantedPhysReg alias.
256 if (TRI->regsOverlap(PhysReg, WantedPhysReg))
259 if (checkUncachedInterference(InterferingVReg, PhysReg))
262 // Reassign the interfering virtual reg to this physical reg.
263 unsigned OldAssign = VRM->getPhys(InterferingVReg.reg);
264 DEBUG(dbgs() << "reassigning: " << InterferingVReg << " from " <<
265 TRI->getName(OldAssign) << " to " << TRI->getName(PhysReg) << '\n');
266 unassign(InterferingVReg, OldAssign);
267 assign(InterferingVReg, PhysReg);
273 /// tryReassignOrEvict - Try to reassign a single interferences to a different
274 /// physreg, or evict a single interference with a lower spill weight.
275 /// @param VirtReg Currently unassigned virtual register.
276 /// @param Order Physregs to try.
277 /// @return Physreg to assign VirtReg, or 0.
278 unsigned RAGreedy::tryReassignOrEvict(LiveInterval &VirtReg,
279 AllocationOrder &Order,
280 SmallVectorImpl<LiveInterval*> &NewVRegs){
281 NamedRegionTimer T("Reassign", TimerGroupName, TimePassesIsEnabled);
283 // Keep track of the lightest single interference seen so far.
284 float BestWeight = VirtReg.weight;
285 LiveInterval *BestVirt = 0;
286 unsigned BestPhys = 0;
289 while (unsigned PhysReg = Order.next()) {
290 LiveInterval *InterferingVReg = getSingleInterference(VirtReg, PhysReg);
291 if (!InterferingVReg)
293 if (TargetRegisterInfo::isPhysicalRegister(InterferingVReg->reg))
295 if (reassignVReg(*InterferingVReg, PhysReg))
298 // Cannot reassign, is this an eviction candidate?
299 if (InterferingVReg->weight < BestWeight) {
300 BestVirt = InterferingVReg;
302 BestWeight = InterferingVReg->weight;
306 // Nothing reassigned, can we evict a lighter single interference?
308 DEBUG(dbgs() << "evicting lighter " << *BestVirt << '\n');
309 unassign(*BestVirt, VRM->getPhys(BestVirt->reg));
310 NewVRegs.push_back(BestVirt);
318 //===----------------------------------------------------------------------===//
320 //===----------------------------------------------------------------------===//
322 /// calcInterferenceInfo - Compute per-block outgoing and ingoing constraints
323 /// when considering interference from PhysReg. Also compute an optimistic local
324 /// cost of this interference pattern.
326 /// The final cost of a split is the local cost + global cost of preferences
327 /// broken by SpillPlacement.
329 float RAGreedy::calcInterferenceInfo(LiveInterval &VirtReg, unsigned PhysReg) {
330 // Reset interference dependent info.
331 SpillConstraints.resize(SA->LiveBlocks.size());
332 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
333 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
334 SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
335 BC.Number = BI.MBB->getNumber();
336 BC.Entry = (BI.Uses && BI.LiveIn) ?
337 SpillPlacement::PrefReg : SpillPlacement::DontCare;
338 BC.Exit = (BI.Uses && BI.LiveOut) ?
339 SpillPlacement::PrefReg : SpillPlacement::DontCare;
340 BI.OverlapEntry = BI.OverlapExit = false;
343 // Add interference info from each PhysReg alias.
344 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
345 if (!query(VirtReg, *AI).checkInterference())
347 LiveIntervalUnion::SegmentIter IntI =
348 PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
352 // Determine which blocks have interference live in or after the last split
354 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
355 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
356 SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
357 SlotIndex Start, Stop;
358 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
360 // Skip interference-free blocks.
361 if (IntI.start() >= Stop)
364 // Is the interference live-in?
366 IntI.advanceTo(Start);
369 if (IntI.start() <= Start)
370 BC.Entry = SpillPlacement::MustSpill;
373 // Is the interference overlapping the last split point?
375 if (IntI.stop() < BI.LastSplitPoint)
376 IntI.advanceTo(BI.LastSplitPoint.getPrevSlot());
379 if (IntI.start() < Stop)
380 BC.Exit = SpillPlacement::MustSpill;
384 // Rewind iterator and check other interferences.
385 IntI.find(VirtReg.beginIndex());
386 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
387 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
388 SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
389 SlotIndex Start, Stop;
390 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
392 // Skip interference-free blocks.
393 if (IntI.start() >= Stop)
396 // Handle transparent blocks with interference separately.
397 // Transparent blocks never incur any fixed cost.
398 if (BI.LiveThrough && !BI.Uses) {
399 IntI.advanceTo(Start);
402 if (IntI.start() >= Stop)
405 if (BC.Entry != SpillPlacement::MustSpill)
406 BC.Entry = SpillPlacement::PrefSpill;
407 if (BC.Exit != SpillPlacement::MustSpill)
408 BC.Exit = SpillPlacement::PrefSpill;
412 // Now we only have blocks with uses left.
413 // Check if the interference overlaps the uses.
414 assert(BI.Uses && "Non-transparent block without any uses");
416 // Check interference on entry.
417 if (BI.LiveIn && BC.Entry != SpillPlacement::MustSpill) {
418 IntI.advanceTo(Start);
421 // Not live in, but before the first use.
422 if (IntI.start() < BI.FirstUse)
423 BC.Entry = SpillPlacement::PrefSpill;
426 // Does interference overlap the uses in the entry segment
428 if (BI.LiveIn && !BI.OverlapEntry) {
429 IntI.advanceTo(BI.FirstUse);
432 // A live-through interval has no kill.
433 // Check [FirstUse;LastUse) instead.
434 if (IntI.start() < (BI.LiveThrough ? BI.LastUse : BI.Kill))
435 BI.OverlapEntry = true;
438 // Does interference overlap the uses in the exit segment [Def;LastUse)?
439 if (BI.LiveOut && !BI.LiveThrough && !BI.OverlapExit) {
440 IntI.advanceTo(BI.Def);
443 if (IntI.start() < BI.LastUse)
444 BI.OverlapExit = true;
447 // Check interference on exit.
448 if (BI.LiveOut && BC.Exit != SpillPlacement::MustSpill) {
449 // Check interference between LastUse and Stop.
450 if (BC.Exit != SpillPlacement::PrefSpill) {
451 IntI.advanceTo(BI.LastUse);
454 if (IntI.start() < Stop)
455 BC.Exit = SpillPlacement::PrefSpill;
461 // Accumulate a local cost of this interference pattern.
463 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
464 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
467 SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
468 unsigned Inserts = 0;
470 // Do we need spill code for the entry segment?
472 Inserts += BI.OverlapEntry || BC.Entry != SpillPlacement::PrefReg;
474 // For the exit segment?
476 Inserts += BI.OverlapExit || BC.Exit != SpillPlacement::PrefReg;
478 // The local cost of spill code in this block is the block frequency times
479 // the number of spill instructions inserted.
481 LocalCost += Inserts * SpillPlacer->getBlockFrequency(BI.MBB);
483 DEBUG(dbgs() << "Local cost of " << PrintReg(PhysReg, TRI) << " = "
484 << LocalCost << '\n');
488 /// calcGlobalSplitCost - Return the global split cost of following the split
489 /// pattern in LiveBundles. This cost should be added to the local cost of the
490 /// interference pattern in SpillConstraints.
492 float RAGreedy::calcGlobalSplitCost(const BitVector &LiveBundles) {
493 float GlobalCost = 0;
494 for (unsigned i = 0, e = SpillConstraints.size(); i != e; ++i) {
495 SpillPlacement::BlockConstraint &BC = SpillConstraints[i];
496 unsigned Inserts = 0;
497 // Broken entry preference?
498 Inserts += LiveBundles[Bundles->getBundle(BC.Number, 0)] !=
499 (BC.Entry == SpillPlacement::PrefReg);
500 // Broken exit preference?
501 Inserts += LiveBundles[Bundles->getBundle(BC.Number, 1)] !=
502 (BC.Exit == SpillPlacement::PrefReg);
505 Inserts * SpillPlacer->getBlockFrequency(SA->LiveBlocks[i].MBB);
507 DEBUG(dbgs() << "Global cost = " << GlobalCost << '\n');
511 /// splitAroundRegion - Split VirtReg around the region determined by
512 /// LiveBundles. Make an effort to avoid interference from PhysReg.
514 /// The 'register' interval is going to contain as many uses as possible while
515 /// avoiding interference. The 'stack' interval is the complement constructed by
516 /// SplitEditor. It will contain the rest.
518 void RAGreedy::splitAroundRegion(LiveInterval &VirtReg, unsigned PhysReg,
519 const BitVector &LiveBundles,
520 SmallVectorImpl<LiveInterval*> &NewVRegs) {
522 dbgs() << "Splitting around region for " << PrintReg(PhysReg, TRI)
524 for (int i = LiveBundles.find_first(); i>=0; i = LiveBundles.find_next(i))
525 dbgs() << " EB#" << i;
529 // First compute interference ranges in the live blocks.
530 typedef std::pair<SlotIndex, SlotIndex> IndexPair;
531 SmallVector<IndexPair, 8> InterferenceRanges;
532 InterferenceRanges.resize(SA->LiveBlocks.size());
533 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
534 if (!query(VirtReg, *AI).checkInterference())
536 LiveIntervalUnion::SegmentIter IntI =
537 PhysReg2LiveUnion[*AI].find(VirtReg.beginIndex());
540 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
541 const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
542 IndexPair &IP = InterferenceRanges[i];
543 SlotIndex Start, Stop;
544 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
545 // Skip interference-free blocks.
546 if (IntI.start() >= Stop)
549 // First interference in block.
551 IntI.advanceTo(Start);
554 if (IntI.start() >= Stop)
556 if (!IP.first.isValid() || IntI.start() < IP.first)
557 IP.first = IntI.start();
560 // Last interference in block.
562 IntI.advanceTo(Stop);
563 if (!IntI.valid() || IntI.start() >= Stop)
565 if (IntI.stop() <= Start)
567 if (!IP.second.isValid() || IntI.stop() > IP.second)
568 IP.second = IntI.stop();
573 SmallVector<LiveInterval*, 4> SpillRegs;
574 LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
575 SplitEditor SE(*SA, *LIS, *VRM, *DomTree, LREdit);
577 // Create the main cross-block interval.
580 // First add all defs that are live out of a block.
581 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
582 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
583 bool RegIn = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
584 bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
586 // Should the register be live out?
587 if (!BI.LiveOut || !RegOut)
590 IndexPair &IP = InterferenceRanges[i];
591 SlotIndex Start, Stop;
592 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
594 DEBUG(dbgs() << "BB#" << BI.MBB->getNumber() << " -> EB#"
595 << Bundles->getBundle(BI.MBB->getNumber(), 1)
596 << " intf [" << IP.first << ';' << IP.second << ')');
598 // The interference interval should either be invalid or overlap MBB.
599 assert((!IP.first.isValid() || IP.first < Stop) && "Bad interference");
600 assert((!IP.second.isValid() || IP.second > Start) && "Bad interference");
602 // Check interference leaving the block.
603 if (!IP.second.isValid()) {
604 // Block is interference-free.
605 DEBUG(dbgs() << ", no interference");
607 assert(BI.LiveThrough && "No uses, but not live through block?");
608 // Block is live-through without interference.
609 DEBUG(dbgs() << ", no uses"
610 << (RegIn ? ", live-through.\n" : ", stack in.\n"));
612 SE.enterIntvAtEnd(*BI.MBB);
615 if (!BI.LiveThrough) {
616 DEBUG(dbgs() << ", not live-through.\n");
617 SE.useIntv(SE.enterIntvBefore(BI.Def), Stop);
621 // Block is live-through, but entry bundle is on the stack.
622 // Reload just before the first use.
623 DEBUG(dbgs() << ", not live-in, enter before first use.\n");
624 SE.useIntv(SE.enterIntvBefore(BI.FirstUse), Stop);
627 DEBUG(dbgs() << ", live-through.\n");
631 // Block has interference.
632 DEBUG(dbgs() << ", interference to " << IP.second);
634 if (!BI.LiveThrough && IP.second <= BI.Def) {
635 // The interference doesn't reach the outgoing segment.
636 DEBUG(dbgs() << " doesn't affect def from " << BI.Def << '\n');
637 SE.useIntv(BI.Def, Stop);
643 // No uses in block, avoid interference by reloading as late as possible.
644 DEBUG(dbgs() << ", no uses.\n");
645 SlotIndex SegStart = SE.enterIntvAtEnd(*BI.MBB);
646 assert(SegStart >= IP.second && "Couldn't avoid interference");
650 if (IP.second.getBoundaryIndex() < BI.LastUse) {
651 // There are interference-free uses at the end of the block.
652 // Find the first use that can get the live-out register.
653 SmallVectorImpl<SlotIndex>::const_iterator UI =
654 std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
655 IP.second.getBoundaryIndex());
656 assert(UI != SA->UseSlots.end() && "Couldn't find last use");
658 assert(Use <= BI.LastUse && "Couldn't find last use");
659 // Only attempt a split befroe the last split point.
660 if (Use.getBaseIndex() <= BI.LastSplitPoint) {
661 DEBUG(dbgs() << ", free use at " << Use << ".\n");
662 SlotIndex SegStart = SE.enterIntvBefore(Use);
663 assert(SegStart >= IP.second && "Couldn't avoid interference");
664 assert(SegStart < BI.LastSplitPoint && "Impossible split point");
665 SE.useIntv(SegStart, Stop);
670 // Interference is after the last use.
671 DEBUG(dbgs() << " after last use.\n");
672 SlotIndex SegStart = SE.enterIntvAtEnd(*BI.MBB);
673 assert(SegStart >= IP.second && "Couldn't avoid interference");
676 // Now all defs leading to live bundles are handled, do everything else.
677 for (unsigned i = 0, e = SA->LiveBlocks.size(); i != e; ++i) {
678 SplitAnalysis::BlockInfo &BI = SA->LiveBlocks[i];
679 bool RegIn = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 0)];
680 bool RegOut = LiveBundles[Bundles->getBundle(BI.MBB->getNumber(), 1)];
682 // Is the register live-in?
683 if (!BI.LiveIn || !RegIn)
686 // We have an incoming register. Check for interference.
687 IndexPair &IP = InterferenceRanges[i];
688 SlotIndex Start, Stop;
689 tie(Start, Stop) = Indexes->getMBBRange(BI.MBB);
691 DEBUG(dbgs() << "EB#" << Bundles->getBundle(BI.MBB->getNumber(), 0)
692 << " -> BB#" << BI.MBB->getNumber());
694 // Check interference entering the block.
695 if (!IP.first.isValid()) {
696 // Block is interference-free.
697 DEBUG(dbgs() << ", no interference");
699 assert(BI.LiveThrough && "No uses, but not live through block?");
700 // Block is live-through without interference.
702 DEBUG(dbgs() << ", no uses, live-through.\n");
703 SE.useIntv(Start, Stop);
705 DEBUG(dbgs() << ", no uses, stack-out.\n");
706 SE.leaveIntvAtTop(*BI.MBB);
710 if (!BI.LiveThrough) {
711 DEBUG(dbgs() << ", killed in block.\n");
712 SE.useIntv(Start, SE.leaveIntvAfter(BI.Kill));
716 // Block is live-through, but exit bundle is on the stack.
717 // Spill immediately after the last use.
718 if (BI.LastUse < BI.LastSplitPoint) {
719 DEBUG(dbgs() << ", uses, stack-out.\n");
720 SE.useIntv(Start, SE.leaveIntvAfter(BI.LastUse));
723 // The last use is after the last split point, it is probably an
725 DEBUG(dbgs() << ", uses at " << BI.LastUse << " after split point "
726 << BI.LastSplitPoint << ", stack-out.\n");
727 SlotIndex SegEnd = SE.leaveIntvBefore(BI.LastSplitPoint);
728 SE.useIntv(Start, SegEnd);
729 // Run a double interval from the split to the last use.
730 // This makes it possible to spill the complement without affecting the
732 SE.overlapIntv(SegEnd, BI.LastUse);
735 // Register is live-through.
736 DEBUG(dbgs() << ", uses, live-through.\n");
737 SE.useIntv(Start, Stop);
741 // Block has interference.
742 DEBUG(dbgs() << ", interference from " << IP.first);
744 if (!BI.LiveThrough && IP.first >= BI.Kill) {
745 // The interference doesn't reach the outgoing segment.
746 DEBUG(dbgs() << " doesn't affect kill at " << BI.Kill << '\n');
747 SE.useIntv(Start, BI.Kill);
752 // No uses in block, avoid interference by spilling as soon as possible.
753 DEBUG(dbgs() << ", no uses.\n");
754 SlotIndex SegEnd = SE.leaveIntvAtTop(*BI.MBB);
755 assert(SegEnd <= IP.first && "Couldn't avoid interference");
758 if (IP.first.getBaseIndex() > BI.FirstUse) {
759 // There are interference-free uses at the beginning of the block.
760 // Find the last use that can get the register.
761 SmallVectorImpl<SlotIndex>::const_iterator UI =
762 std::lower_bound(SA->UseSlots.begin(), SA->UseSlots.end(),
763 IP.first.getBaseIndex());
764 assert(UI != SA->UseSlots.begin() && "Couldn't find first use");
765 SlotIndex Use = (--UI)->getBoundaryIndex();
766 DEBUG(dbgs() << ", free use at " << *UI << ".\n");
767 SlotIndex SegEnd = SE.leaveIntvAfter(Use);
768 assert(SegEnd <= IP.first && "Couldn't avoid interference");
769 SE.useIntv(Start, SegEnd);
773 // Interference is before the first use.
774 DEBUG(dbgs() << " before first use.\n");
775 SlotIndex SegEnd = SE.leaveIntvAtTop(*BI.MBB);
776 assert(SegEnd <= IP.first && "Couldn't avoid interference");
781 // FIXME: Should we be more aggressive about splitting the stack region into
782 // per-block segments? The current approach allows the stack region to
783 // separate into connected components. Some components may be allocatable.
787 MF->verify(this, "After splitting live range around region");
790 // Make sure that at least one of the new intervals can allocate to PhysReg.
791 // That was the whole point of splitting the live range.
793 for (LiveRangeEdit::iterator I = LREdit.begin(), E = LREdit.end(); I != E;
795 if (!checkUncachedInterference(**I, PhysReg)) {
799 assert(found && "No allocatable intervals after pointless splitting");
804 unsigned RAGreedy::tryRegionSplit(LiveInterval &VirtReg, AllocationOrder &Order,
805 SmallVectorImpl<LiveInterval*> &NewVRegs) {
806 BitVector LiveBundles, BestBundles;
808 unsigned BestReg = 0;
810 while (unsigned PhysReg = Order.next()) {
811 float Cost = calcInterferenceInfo(VirtReg, PhysReg);
812 if (BestReg && Cost >= BestCost)
815 SpillPlacer->placeSpills(SpillConstraints, LiveBundles);
816 // No live bundles, defer to splitSingleBlocks().
817 if (!LiveBundles.any())
820 Cost += calcGlobalSplitCost(LiveBundles);
821 if (!BestReg || Cost < BestCost) {
824 BestBundles.swap(LiveBundles);
831 splitAroundRegion(VirtReg, BestReg, BestBundles, NewVRegs);
836 //===----------------------------------------------------------------------===//
838 //===----------------------------------------------------------------------===//
841 /// calcGapWeights - Compute the maximum spill weight that needs to be evicted
842 /// in order to use PhysReg between two entries in SA->UseSlots.
844 /// GapWeight[i] represents the gap between UseSlots[i] and UseSlots[i+1].
846 void RAGreedy::calcGapWeights(unsigned PhysReg,
847 SmallVectorImpl<float> &GapWeight) {
848 assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
849 const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
850 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
851 const unsigned NumGaps = Uses.size()-1;
853 // Start and end points for the interference check.
854 SlotIndex StartIdx = BI.LiveIn ? BI.FirstUse.getBaseIndex() : BI.FirstUse;
855 SlotIndex StopIdx = BI.LiveOut ? BI.LastUse.getBoundaryIndex() : BI.LastUse;
857 GapWeight.assign(NumGaps, 0.0f);
859 // Add interference from each overlapping register.
860 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
861 if (!query(const_cast<LiveInterval&>(SA->getParent()), *AI)
862 .checkInterference())
865 // We know that VirtReg is a continuous interval from FirstUse to LastUse,
866 // so we don't need InterferenceQuery.
868 // Interference that overlaps an instruction is counted in both gaps
869 // surrounding the instruction. The exception is interference before
870 // StartIdx and after StopIdx.
872 LiveIntervalUnion::SegmentIter IntI = PhysReg2LiveUnion[*AI].find(StartIdx);
873 for (unsigned Gap = 0; IntI.valid() && IntI.start() < StopIdx; ++IntI) {
874 // Skip the gaps before IntI.
875 while (Uses[Gap+1].getBoundaryIndex() < IntI.start())
876 if (++Gap == NumGaps)
881 // Update the gaps covered by IntI.
882 const float weight = IntI.value()->weight;
883 for (; Gap != NumGaps; ++Gap) {
884 GapWeight[Gap] = std::max(GapWeight[Gap], weight);
885 if (Uses[Gap+1].getBaseIndex() >= IntI.stop())
894 /// getPrevMappedIndex - Return the slot index of the last non-copy instruction
895 /// before MI that has a slot index. If MI is the first mapped instruction in
896 /// its block, return the block start index instead.
898 SlotIndex RAGreedy::getPrevMappedIndex(const MachineInstr *MI) {
899 assert(MI && "Missing MachineInstr");
900 const MachineBasicBlock *MBB = MI->getParent();
901 MachineBasicBlock::const_iterator B = MBB->begin(), I = MI;
903 if (!(--I)->isDebugValue() && !I->isCopy())
904 return Indexes->getInstructionIndex(I);
905 return Indexes->getMBBStartIdx(MBB);
908 /// calcPrevSlots - Fill in the PrevSlot array with the index of the previous
909 /// real non-copy instruction for each instruction in SA->UseSlots.
911 void RAGreedy::calcPrevSlots() {
912 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
914 PrevSlot.reserve(Uses.size());
915 for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
916 const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]);
917 PrevSlot.push_back(getPrevMappedIndex(MI).getDefIndex());
921 /// nextSplitPoint - Find the next index into SA->UseSlots > i such that it may
922 /// be beneficial to split before UseSlots[i].
924 /// 0 is always a valid split point
925 unsigned RAGreedy::nextSplitPoint(unsigned i) {
926 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
927 const unsigned Size = Uses.size();
928 assert(i != Size && "No split points after the end");
929 // Allow split before i when Uses[i] is not adjacent to the previous use.
930 while (++i != Size && PrevSlot[i].getBaseIndex() <= Uses[i-1].getBaseIndex())
935 /// tryLocalSplit - Try to split VirtReg into smaller intervals inside its only
938 unsigned RAGreedy::tryLocalSplit(LiveInterval &VirtReg, AllocationOrder &Order,
939 SmallVectorImpl<LiveInterval*> &NewVRegs) {
940 assert(SA->LiveBlocks.size() == 1 && "Not a local interval");
941 const SplitAnalysis::BlockInfo &BI = SA->LiveBlocks.front();
943 // Note that it is possible to have an interval that is live-in or live-out
944 // while only covering a single block - A phi-def can use undef values from
945 // predecessors, and the block could be a single-block loop.
946 // We don't bother doing anything clever about such a case, we simply assume
947 // that the interval is continuous from FirstUse to LastUse. We should make
948 // sure that we don't do anything illegal to such an interval, though.
950 const SmallVectorImpl<SlotIndex> &Uses = SA->UseSlots;
951 if (Uses.size() <= 2)
953 const unsigned NumGaps = Uses.size()-1;
956 dbgs() << "tryLocalSplit: ";
957 for (unsigned i = 0, e = Uses.size(); i != e; ++i)
958 dbgs() << ' ' << SA->UseSlots[i];
962 // For every use, find the previous mapped non-copy instruction.
963 // We use this to detect valid split points, and to estimate new interval
967 unsigned BestBefore = NumGaps;
968 unsigned BestAfter = 0;
971 const float blockFreq = SpillPlacer->getBlockFrequency(BI.MBB);
972 SmallVector<float, 8> GapWeight;
975 while (unsigned PhysReg = Order.next()) {
976 // Keep track of the largest spill weight that would need to be evicted in
977 // order to make use of PhysReg between UseSlots[i] and UseSlots[i+1].
978 calcGapWeights(PhysReg, GapWeight);
980 // Try to find the best sequence of gaps to close.
981 // The new spill weight must be larger than any gap interference.
983 // We will split before Uses[SplitBefore] and after Uses[SplitAfter].
984 unsigned SplitBefore = 0, SplitAfter = nextSplitPoint(1) - 1;
986 // MaxGap should always be max(GapWeight[SplitBefore..SplitAfter-1]).
987 // It is the spill weight that needs to be evicted.
988 float MaxGap = GapWeight[0];
989 for (unsigned i = 1; i != SplitAfter; ++i)
990 MaxGap = std::max(MaxGap, GapWeight[i]);
993 // Live before/after split?
994 const bool LiveBefore = SplitBefore != 0 || BI.LiveIn;
995 const bool LiveAfter = SplitAfter != NumGaps || BI.LiveOut;
997 DEBUG(dbgs() << PrintReg(PhysReg, TRI) << ' '
998 << Uses[SplitBefore] << '-' << Uses[SplitAfter]
1001 // Stop before the interval gets so big we wouldn't be making progress.
1002 if (!LiveBefore && !LiveAfter) {
1003 DEBUG(dbgs() << " all\n");
1006 // Should the interval be extended or shrunk?
1008 if (MaxGap < HUGE_VALF) {
1009 // Estimate the new spill weight.
1011 // Each instruction reads and writes the register, except the first
1012 // instr doesn't read when !FirstLive, and the last instr doesn't write
1015 // We will be inserting copies before and after, so the total number of
1016 // reads and writes is 2 * EstUses.
1018 const unsigned EstUses = 2*(SplitAfter - SplitBefore) +
1019 2*(LiveBefore + LiveAfter);
1021 // Try to guess the size of the new interval. This should be trivial,
1022 // but the slot index of an inserted copy can be a lot smaller than the
1023 // instruction it is inserted before if there are many dead indexes
1026 // We measure the distance from the instruction before SplitBefore to
1027 // get a conservative estimate.
1029 // The final distance can still be different if inserting copies
1030 // triggers a slot index renumbering.
1032 const float EstWeight = normalizeSpillWeight(blockFreq * EstUses,
1033 PrevSlot[SplitBefore].distance(Uses[SplitAfter]));
1034 // Would this split be possible to allocate?
1035 // Never allocate all gaps, we wouldn't be making progress.
1036 float Diff = EstWeight - MaxGap;
1037 DEBUG(dbgs() << " w=" << EstWeight << " d=" << Diff);
1040 if (Diff > BestDiff) {
1041 DEBUG(dbgs() << " (best)");
1043 BestBefore = SplitBefore;
1044 BestAfter = SplitAfter;
1051 SplitBefore = nextSplitPoint(SplitBefore);
1052 if (SplitBefore < SplitAfter) {
1053 DEBUG(dbgs() << " shrink\n");
1054 // Recompute the max when necessary.
1055 if (GapWeight[SplitBefore - 1] >= MaxGap) {
1056 MaxGap = GapWeight[SplitBefore];
1057 for (unsigned i = SplitBefore + 1; i != SplitAfter; ++i)
1058 MaxGap = std::max(MaxGap, GapWeight[i]);
1065 // Try to extend the interval.
1066 if (SplitAfter >= NumGaps) {
1067 DEBUG(dbgs() << " end\n");
1071 DEBUG(dbgs() << " extend\n");
1072 for (unsigned e = nextSplitPoint(SplitAfter + 1) - 1;
1073 SplitAfter != e; ++SplitAfter)
1074 MaxGap = std::max(MaxGap, GapWeight[SplitAfter]);
1079 // Didn't find any candidates?
1080 if (BestBefore == NumGaps)
1083 DEBUG(dbgs() << "Best local split range: " << Uses[BestBefore]
1084 << '-' << Uses[BestAfter] << ", " << BestDiff
1085 << ", " << (BestAfter - BestBefore + 1) << " instrs\n");
1087 SmallVector<LiveInterval*, 4> SpillRegs;
1088 LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
1089 SplitEditor SE(*SA, *LIS, *VRM, *DomTree, LREdit);
1092 SlotIndex SegStart = SE.enterIntvBefore(Uses[BestBefore]);
1093 SlotIndex SegStop = SE.leaveIntvAfter(Uses[BestAfter]);
1094 SE.useIntv(SegStart, SegStop);
1101 //===----------------------------------------------------------------------===//
1102 // Live Range Splitting
1103 //===----------------------------------------------------------------------===//
1105 /// trySplit - Try to split VirtReg or one of its interferences, making it
1107 /// @return Physreg when VirtReg may be assigned and/or new NewVRegs.
1108 unsigned RAGreedy::trySplit(LiveInterval &VirtReg, AllocationOrder &Order,
1109 SmallVectorImpl<LiveInterval*>&NewVRegs) {
1110 NamedRegionTimer T("Splitter", TimerGroupName, TimePassesIsEnabled);
1111 SA->analyze(&VirtReg);
1113 // Local intervals are handled separately.
1114 if (LIS->intervalIsInOneMBB(VirtReg))
1115 return tryLocalSplit(VirtReg, Order, NewVRegs);
1117 // First try to split around a region spanning multiple blocks.
1118 unsigned PhysReg = tryRegionSplit(VirtReg, Order, NewVRegs);
1119 if (PhysReg || !NewVRegs.empty())
1122 // Then isolate blocks with multiple uses.
1123 SplitAnalysis::BlockPtrSet Blocks;
1124 if (SA->getMultiUseBlocks(Blocks)) {
1125 SmallVector<LiveInterval*, 4> SpillRegs;
1126 LiveRangeEdit LREdit(VirtReg, NewVRegs, SpillRegs);
1127 SplitEditor(*SA, *LIS, *VRM, *DomTree, LREdit).splitSingleBlocks(Blocks);
1129 MF->verify(this, "After splitting live range around basic blocks");
1132 // Don't assign any physregs.
1137 //===----------------------------------------------------------------------===//
1139 //===----------------------------------------------------------------------===//
1141 /// calcInterferenceWeight - Calculate the combined spill weight of
1142 /// interferences when assigning VirtReg to PhysReg.
1143 float RAGreedy::calcInterferenceWeight(LiveInterval &VirtReg, unsigned PhysReg){
1145 for (const unsigned *AI = TRI->getOverlaps(PhysReg); *AI; ++AI) {
1146 LiveIntervalUnion::Query &Q = query(VirtReg, *AI);
1147 Q.collectInterferingVRegs();
1148 if (Q.seenUnspillableVReg())
1150 for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i)
1151 Sum += Q.interferingVRegs()[i]->weight;
1156 /// trySpillInterferences - Try to spill interfering registers instead of the
1157 /// current one. Only do it if the accumulated spill weight is smaller than the
1158 /// current spill weight.
1159 unsigned RAGreedy::trySpillInterferences(LiveInterval &VirtReg,
1160 AllocationOrder &Order,
1161 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1162 NamedRegionTimer T("Spill Interference", TimerGroupName, TimePassesIsEnabled);
1163 unsigned BestPhys = 0;
1164 float BestWeight = 0;
1167 while (unsigned PhysReg = Order.next()) {
1168 float Weight = calcInterferenceWeight(VirtReg, PhysReg);
1169 if (Weight == HUGE_VALF || Weight >= VirtReg.weight)
1171 if (!BestPhys || Weight < BestWeight)
1172 BestPhys = PhysReg, BestWeight = Weight;
1175 // No candidates found.
1179 // Collect all interfering registers.
1180 SmallVector<LiveInterval*, 8> Spills;
1181 for (const unsigned *AI = TRI->getOverlaps(BestPhys); *AI; ++AI) {
1182 LiveIntervalUnion::Query &Q = query(VirtReg, *AI);
1183 Spills.append(Q.interferingVRegs().begin(), Q.interferingVRegs().end());
1184 for (unsigned i = 0, e = Q.interferingVRegs().size(); i != e; ++i) {
1185 LiveInterval *VReg = Q.interferingVRegs()[i];
1186 unassign(*VReg, *AI);
1191 DEBUG(dbgs() << "spilling " << Spills.size() << " interferences with weight "
1192 << BestWeight << '\n');
1193 for (unsigned i = 0, e = Spills.size(); i != e; ++i)
1194 spiller().spill(Spills[i], NewVRegs, Spills);
1199 //===----------------------------------------------------------------------===//
1201 //===----------------------------------------------------------------------===//
1203 unsigned RAGreedy::selectOrSplit(LiveInterval &VirtReg,
1204 SmallVectorImpl<LiveInterval*> &NewVRegs) {
1205 // First try assigning a free register.
1206 AllocationOrder Order(VirtReg.reg, *VRM, ReservedRegs);
1207 while (unsigned PhysReg = Order.next()) {
1208 if (!checkPhysRegInterference(VirtReg, PhysReg))
1212 // Try to reassign interferences.
1213 if (unsigned PhysReg = tryReassignOrEvict(VirtReg, Order, NewVRegs))
1216 assert(NewVRegs.empty() && "Cannot append to existing NewVRegs");
1218 // Try splitting VirtReg or interferences.
1219 unsigned PhysReg = trySplit(VirtReg, Order, NewVRegs);
1220 if (PhysReg || !NewVRegs.empty())
1223 // Try to spill another interfering reg with less spill weight.
1224 PhysReg = trySpillInterferences(VirtReg, Order, NewVRegs);
1228 // Finally spill VirtReg itself.
1229 NamedRegionTimer T("Spiller", TimerGroupName, TimePassesIsEnabled);
1230 SmallVector<LiveInterval*, 1> pendingSpills;
1231 spiller().spill(&VirtReg, NewVRegs, pendingSpills);
1233 // The live virtual register requesting allocation was spilled, so tell
1234 // the caller not to allocate anything during this round.
1238 bool RAGreedy::runOnMachineFunction(MachineFunction &mf) {
1239 DEBUG(dbgs() << "********** GREEDY REGISTER ALLOCATION **********\n"
1240 << "********** Function: "
1241 << ((Value*)mf.getFunction())->getName() << '\n');
1245 MF->verify(this, "Before greedy register allocator");
1247 RegAllocBase::init(getAnalysis<VirtRegMap>(), getAnalysis<LiveIntervals>());
1248 Indexes = &getAnalysis<SlotIndexes>();
1249 DomTree = &getAnalysis<MachineDominatorTree>();
1250 ReservedRegs = TRI->getReservedRegs(*MF);
1251 SpillerInstance.reset(createInlineSpiller(*this, *MF, *VRM));
1252 Loops = &getAnalysis<MachineLoopInfo>();
1253 LoopRanges = &getAnalysis<MachineLoopRanges>();
1254 Bundles = &getAnalysis<EdgeBundles>();
1255 SpillPlacer = &getAnalysis<SpillPlacement>();
1257 SA.reset(new SplitAnalysis(*MF, *LIS, *Loops));
1261 LIS->addKillFlags();
1265 NamedRegionTimer T("Rewriter", TimerGroupName, TimePassesIsEnabled);
1266 std::auto_ptr<VirtRegRewriter> rewriter(createVirtRegRewriter());
1267 rewriter->runOnMachineFunction(*MF, *VRM, LIS);
1270 // The pass output is in VirtRegMap. Release all the transient data.