1 //===---------- SplitKit.cpp - Toolkit for splitting live ranges ----------===//
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 contains the SplitAnalysis class as well as mutator functions for
11 // live range splitting.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "splitter"
17 #include "LiveRangeEdit.h"
18 #include "VirtRegMap.h"
19 #include "llvm/CodeGen/CalcSpillWeights.h"
20 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
21 #include "llvm/CodeGen/MachineDominators.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineLoopInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetInstrInfo.h"
29 #include "llvm/Target/TargetMachine.h"
34 AllowSplit("spiller-splits-edges",
35 cl::desc("Allow critical edge splitting during spilling"));
37 //===----------------------------------------------------------------------===//
39 //===----------------------------------------------------------------------===//
41 SplitAnalysis::SplitAnalysis(const MachineFunction &mf,
42 const LiveIntervals &lis,
43 const MachineLoopInfo &mli)
47 tii_(*mf.getTarget().getInstrInfo()),
50 void SplitAnalysis::clear() {
57 bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) {
58 MachineBasicBlock *T, *F;
59 SmallVector<MachineOperand, 4> Cond;
60 return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond);
63 /// analyzeUses - Count instructions, basic blocks, and loops using curli.
64 void SplitAnalysis::analyzeUses() {
65 const MachineRegisterInfo &MRI = mf_.getRegInfo();
66 for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg);
67 MachineInstr *MI = I.skipInstruction();) {
68 if (MI->isDebugValue() || !usingInstrs_.insert(MI))
70 MachineBasicBlock *MBB = MI->getParent();
71 if (usingBlocks_[MBB]++)
73 for (MachineLoop *Loop = loops_.getLoopFor(MBB); Loop;
74 Loop = Loop->getParentLoop())
77 DEBUG(dbgs() << " counted "
78 << usingInstrs_.size() << " instrs, "
79 << usingBlocks_.size() << " blocks, "
80 << usingLoops_.size() << " loops.\n");
83 void SplitAnalysis::print(const BlockPtrSet &B, raw_ostream &OS) const {
84 for (BlockPtrSet::const_iterator I = B.begin(), E = B.end(); I != E; ++I) {
85 unsigned count = usingBlocks_.lookup(*I);
86 OS << " BB#" << (*I)->getNumber();
88 OS << '(' << count << ')';
92 // Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
93 // predecessor blocks, and exit blocks.
94 void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
97 // Blocks in the loop.
98 Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
100 // Predecessor blocks.
101 const MachineBasicBlock *Header = Loop->getHeader();
102 for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
103 E = Header->pred_end(); I != E; ++I)
104 if (!Blocks.Loop.count(*I))
105 Blocks.Preds.insert(*I);
108 for (MachineLoop::block_iterator I = Loop->block_begin(),
109 E = Loop->block_end(); I != E; ++I) {
110 const MachineBasicBlock *MBB = *I;
111 for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
112 SE = MBB->succ_end(); SI != SE; ++SI)
113 if (!Blocks.Loop.count(*SI))
114 Blocks.Exits.insert(*SI);
118 void SplitAnalysis::print(const LoopBlocks &B, raw_ostream &OS) const {
127 /// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
128 /// and around the Loop.
129 SplitAnalysis::LoopPeripheralUse SplitAnalysis::
130 analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
131 LoopPeripheralUse use = ContainedInLoop;
132 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
134 const MachineBasicBlock *MBB = I->first;
135 // Is this a peripheral block?
136 if (use < MultiPeripheral &&
137 (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
138 if (I->second > 1) use = MultiPeripheral;
139 else use = SinglePeripheral;
142 // Is it a loop block?
143 if (Blocks.Loop.count(MBB))
145 // It must be an unrelated block.
146 DEBUG(dbgs() << ", outside: BB#" << MBB->getNumber());
152 /// getCriticalExits - It may be necessary to partially break critical edges
153 /// leaving the loop if an exit block has predecessors from outside the loop
155 void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
156 BlockPtrSet &CriticalExits) {
157 CriticalExits.clear();
159 // A critical exit block has curli live-in, and has a predecessor that is not
160 // in the loop nor a loop predecessor. For such an exit block, the edges
161 // carrying the new variable must be moved to a new pre-exit block.
162 for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
164 const MachineBasicBlock *Exit = *I;
165 // A single-predecessor exit block is definitely not a critical edge.
166 if (Exit->pred_size() == 1)
168 // This exit may not have curli live in at all. No need to split.
169 if (!lis_.isLiveInToMBB(*curli_, Exit))
171 // Does this exit block have a predecessor that is not a loop block or loop
173 for (MachineBasicBlock::const_pred_iterator PI = Exit->pred_begin(),
174 PE = Exit->pred_end(); PI != PE; ++PI) {
175 const MachineBasicBlock *Pred = *PI;
176 if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
178 // This is a critical exit block, and we need to split the exit edge.
179 CriticalExits.insert(Exit);
185 void SplitAnalysis::getCriticalPreds(const SplitAnalysis::LoopBlocks &Blocks,
186 BlockPtrSet &CriticalPreds) {
187 CriticalPreds.clear();
189 // A critical predecessor block has curli live-out, and has a successor that
190 // has curli live-in and is not in the loop nor a loop exit block. For such a
191 // predecessor block, we must carry the value in both the 'inside' and
192 // 'outside' registers.
193 for (BlockPtrSet::iterator I = Blocks.Preds.begin(), E = Blocks.Preds.end();
195 const MachineBasicBlock *Pred = *I;
196 // Definitely not a critical edge.
197 if (Pred->succ_size() == 1)
199 // This block may not have curli live out at all if there is a PHI.
200 if (!lis_.isLiveOutOfMBB(*curli_, Pred))
202 // Does this block have a successor outside the loop?
203 for (MachineBasicBlock::const_pred_iterator SI = Pred->succ_begin(),
204 SE = Pred->succ_end(); SI != SE; ++SI) {
205 const MachineBasicBlock *Succ = *SI;
206 if (Blocks.Loop.count(Succ) || Blocks.Exits.count(Succ))
208 if (!lis_.isLiveInToMBB(*curli_, Succ))
210 // This is a critical predecessor block.
211 CriticalPreds.insert(Pred);
217 /// canSplitCriticalExits - Return true if it is possible to insert new exit
218 /// blocks before the blocks in CriticalExits.
220 SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
221 BlockPtrSet &CriticalExits) {
222 // If we don't allow critical edge splitting, require no critical exits.
224 return CriticalExits.empty();
226 for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
228 const MachineBasicBlock *Succ = *I;
229 // We want to insert a new pre-exit MBB before Succ, and change all the
230 // in-loop blocks to branch to the pre-exit instead of Succ.
231 // Check that all the in-loop predecessors can be changed.
232 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
233 PE = Succ->pred_end(); PI != PE; ++PI) {
234 const MachineBasicBlock *Pred = *PI;
235 // The external predecessors won't be altered.
236 if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
238 if (!canAnalyzeBranch(Pred))
242 // If Succ's layout predecessor falls through, that too must be analyzable.
243 // We need to insert the pre-exit block in the gap.
244 MachineFunction::const_iterator MFI = Succ;
245 if (MFI == mf_.begin())
247 if (!canAnalyzeBranch(--MFI))
250 // No problems found.
254 void SplitAnalysis::analyze(const LiveInterval *li) {
260 const MachineLoop *SplitAnalysis::getBestSplitLoop() {
261 assert(curli_ && "Call analyze() before getBestSplitLoop");
262 if (usingLoops_.empty())
267 BlockPtrSet CriticalExits;
269 // We split around loops where curli is used outside the periphery.
270 for (LoopCountMap::const_iterator I = usingLoops_.begin(),
271 E = usingLoops_.end(); I != E; ++I) {
272 const MachineLoop *Loop = I->first;
273 getLoopBlocks(Loop, Blocks);
274 DEBUG({ dbgs() << " "; print(Blocks, dbgs()); });
276 switch(analyzeLoopPeripheralUse(Blocks)) {
279 case MultiPeripheral:
280 // FIXME: We could split a live range with multiple uses in a peripheral
281 // block and still make progress. However, it is possible that splitting
282 // another live range will insert copies into a peripheral block, and
283 // there is a small chance we can enter an infinity loop, inserting copies
285 // For safety, stick to splitting live ranges with uses outside the
287 DEBUG(dbgs() << ": multiple peripheral uses\n");
289 case ContainedInLoop:
290 DEBUG(dbgs() << ": fully contained\n");
292 case SinglePeripheral:
293 DEBUG(dbgs() << ": single peripheral use\n");
296 // Will it be possible to split around this loop?
297 getCriticalExits(Blocks, CriticalExits);
298 DEBUG(dbgs() << ": " << CriticalExits.size() << " critical exits\n");
299 if (!canSplitCriticalExits(Blocks, CriticalExits))
301 // This is a possible split.
305 DEBUG(dbgs() << " getBestSplitLoop found " << Loops.size()
306 << " candidate loops.\n");
311 // Pick the earliest loop.
312 // FIXME: Are there other heuristics to consider?
313 const MachineLoop *Best = 0;
315 for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
317 SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
318 if (!Best || Idx < BestIdx)
319 Best = *I, BestIdx = Idx;
321 DEBUG(dbgs() << " getBestSplitLoop found " << *Best);
325 //===----------------------------------------------------------------------===//
327 //===----------------------------------------------------------------------===//
329 // Work around the fact that the std::pair constructors are broken for pointer
330 // pairs in some implementations. makeVV(x, 0) works.
331 static inline std::pair<const VNInfo*, VNInfo*>
332 makeVV(const VNInfo *a, VNInfo *b) {
333 return std::make_pair(a, b);
336 void LiveIntervalMap::reset(LiveInterval *li) {
339 liveOutCache_.clear();
342 bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
343 ValueMap::const_iterator i = valueMap_.find(ParentVNI);
344 return i != valueMap_.end() && i->second == 0;
347 // defValue - Introduce a li_ def for ParentVNI that could be later than
349 VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
350 assert(li_ && "call reset first");
351 assert(ParentVNI && "Mapping NULL value");
352 assert(Idx.isValid() && "Invalid SlotIndex");
353 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
355 // Create a new value.
356 VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
358 // Preserve the PHIDef bit.
359 if (ParentVNI->isPHIDef() && Idx == ParentVNI->def)
360 VNI->setIsPHIDef(true);
362 // Use insert for lookup, so we can add missing values with a second lookup.
363 std::pair<ValueMap::iterator,bool> InsP =
364 valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
366 // This is now a complex def. Mark with a NULL in valueMap.
368 InsP.first->second = 0;
374 // mapValue - Find the mapped value for ParentVNI at Idx.
375 // Potentially create phi-def values.
376 VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
378 assert(li_ && "call reset first");
379 assert(ParentVNI && "Mapping NULL value");
380 assert(Idx.isValid() && "Invalid SlotIndex");
381 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
383 // Use insert for lookup, so we can add missing values with a second lookup.
384 std::pair<ValueMap::iterator,bool> InsP =
385 valueMap_.insert(makeVV(ParentVNI, 0));
387 // This was an unknown value. Create a simple mapping.
389 if (simple) *simple = true;
390 return InsP.first->second = li_->createValueCopy(ParentVNI,
391 lis_.getVNInfoAllocator());
394 // This was a simple mapped value.
395 if (InsP.first->second) {
396 if (simple) *simple = true;
397 return InsP.first->second;
400 // This is a complex mapped value. There may be multiple defs, and we may need
401 // to create phi-defs.
402 if (simple) *simple = false;
403 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
404 assert(IdxMBB && "No MBB at Idx");
406 // Is there a def in the same MBB we can extend?
407 if (VNInfo *VNI = extendTo(IdxMBB, Idx))
410 // Now for the fun part. We know that ParentVNI potentially has multiple defs,
411 // and we may need to create even more phi-defs to preserve VNInfo SSA form.
412 // Perform a search for all predecessor blocks where we know the dominating
413 // VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
414 DEBUG(dbgs() << "\n Reaching defs for BB#" << IdxMBB->getNumber()
415 << " at " << Idx << " in " << *li_ << '\n');
417 // Blocks where li_ should be live-in.
418 SmallVector<MachineDomTreeNode*, 16> LiveIn;
419 LiveIn.push_back(mdt_[IdxMBB]);
421 // Using liveOutCache_ as a visited set, perform a BFS for all reaching defs.
422 for (unsigned i = 0; i != LiveIn.size(); ++i) {
423 MachineBasicBlock *MBB = LiveIn[i]->getBlock();
424 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
425 PE = MBB->pred_end(); PI != PE; ++PI) {
426 MachineBasicBlock *Pred = *PI;
427 // Is this a known live-out block?
428 std::pair<LiveOutMap::iterator,bool> LOIP =
429 liveOutCache_.insert(std::make_pair(Pred, LiveOutPair()));
430 // Yes, we have been here before.
432 if (VNInfo *VNI = LOIP.first->second.first) {
433 DEBUG(dbgs() << " known valno #" << VNI->id
434 << " at BB#" << Pred->getNumber() << '\n');
439 // Does Pred provide a live-out value?
440 SlotIndex Last = lis_.getMBBEndIdx(Pred).getPrevSlot();
441 if (VNInfo *VNI = extendTo(Pred, Last)) {
442 MachineBasicBlock *DefMBB = lis_.getMBBFromIndex(VNI->def);
443 DEBUG(dbgs() << " found valno #" << VNI->id
444 << " from BB#" << DefMBB->getNumber()
445 << " at BB#" << Pred->getNumber() << '\n');
446 LiveOutPair &LOP = LOIP.first->second;
448 LOP.second = mdt_[lis_.getMBBFromIndex(VNI->def)];
451 // No, we need a live-in value for Pred as well
453 LiveIn.push_back(mdt_[Pred]);
457 // We may need to add phi-def values to preserve the SSA form.
458 // This is essentially the same iterative algorithm that SSAUpdater uses,
459 // except we already have a dominator tree, so we don't have to recompute it.
464 DEBUG(dbgs() << " Iterating over " << LiveIn.size() << " blocks.\n");
465 // Propagate live-out values down the dominator tree, inserting phi-defs when
466 // necessary. Since LiveIn was created by a BFS, going backwards makes it more
467 // likely for us to visit immediate dominators before their children.
468 for (unsigned i = LiveIn.size(); i; --i) {
469 MachineDomTreeNode *Node = LiveIn[i-1];
470 MachineBasicBlock *MBB = Node->getBlock();
471 MachineDomTreeNode *IDom = Node->getIDom();
472 LiveOutPair IDomValue;
473 // We need a live-in value to a block with no immediate dominator?
474 // This is probably an unreachable block that has survived somehow.
475 bool needPHI = !IDom;
477 // Get the IDom live-out value.
479 LiveOutMap::iterator I = liveOutCache_.find(IDom->getBlock());
480 if (I != liveOutCache_.end())
481 IDomValue = I->second;
483 // If IDom is outside our set of live-out blocks, there must be new
484 // defs, and we need a phi-def here.
488 // IDom dominates all of our predecessors, but it may not be the immediate
489 // dominator. Check if any of them have live-out values that are properly
490 // dominated by IDom. If so, we need a phi-def here.
492 for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(),
493 PE = MBB->pred_end(); PI != PE; ++PI) {
494 LiveOutPair Value = liveOutCache_[*PI];
495 if (!Value.first || Value.first == IDomValue.first)
497 // This predecessor is carrying something other than IDomValue.
498 // It could be because IDomValue hasn't propagated yet, or it could be
499 // because MBB is in the dominance frontier of that value.
500 if (mdt_.dominates(IDom, Value.second)) {
507 // Create a phi-def if required.
510 SlotIndex Start = lis_.getMBBStartIdx(MBB);
511 VNInfo *VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
512 VNI->setIsPHIDef(true);
513 DEBUG(dbgs() << " - BB#" << MBB->getNumber()
514 << " phi-def #" << VNI->id << " at " << Start << '\n');
515 // We no longer need li_ to be live-in.
516 LiveIn.erase(LiveIn.begin()+(i-1));
517 // Blocks in LiveIn are either IdxMBB, or have a value live-through.
520 // Check if we need to update live-out info.
521 LiveOutMap::iterator I = liveOutCache_.find(MBB);
522 if (I == liveOutCache_.end() || I->second.second == Node) {
523 // We already have a live-out defined in MBB, so this must be IdxMBB.
524 assert(MBB == IdxMBB && "Adding phi-def to known live-out");
525 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
527 // This phi-def is also live-out, so color the whole block.
528 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
529 I->second = LiveOutPair(VNI, Node);
531 } else if (IDomValue.first) {
532 // No phi-def here. Remember incoming value for IdxMBB.
534 IdxVNI = IDomValue.first;
535 // Propagate IDomValue if needed:
536 // MBB is live-out and doesn't define its own value.
537 LiveOutMap::iterator I = liveOutCache_.find(MBB);
538 if (I != liveOutCache_.end() && I->second.second != Node &&
539 I->second.first != IDomValue.first) {
541 I->second = IDomValue;
542 DEBUG(dbgs() << " - BB#" << MBB->getNumber()
543 << " idom valno #" << IDomValue.first->id
544 << " from BB#" << IDom->getBlock()->getNumber() << '\n');
548 DEBUG(dbgs() << " - made " << Changes << " changes.\n");
551 assert(IdxVNI && "Didn't find value for Idx");
554 // Check the liveOutCache_ invariants.
555 for (LiveOutMap::iterator I = liveOutCache_.begin(), E = liveOutCache_.end();
557 assert(I->first && "Null MBB entry in cache");
558 assert(I->second.first && "Null VNInfo in cache");
559 assert(I->second.second && "Null DomTreeNode in cache");
560 if (I->second.second->getBlock() == I->first)
562 for (MachineBasicBlock::pred_iterator PI = I->first->pred_begin(),
563 PE = I->first->pred_end(); PI != PE; ++PI)
564 assert(liveOutCache_.lookup(*PI) == I->second && "Bad invariant");
568 // Since we went through the trouble of a full BFS visiting all reaching defs,
569 // the values in LiveIn are now accurate. No more phi-defs are needed
570 // for these blocks, so we can color the live ranges.
571 // This makes the next mapValue call much faster.
572 for (unsigned i = 0, e = LiveIn.size(); i != e; ++i) {
573 MachineBasicBlock *MBB = LiveIn[i]->getBlock();
574 SlotIndex Start = lis_.getMBBStartIdx(MBB);
576 li_->addRange(LiveRange(Start, Idx.getNextSlot(), IdxVNI));
579 // Anything in LiveIn other than IdxMBB is live-through.
580 VNInfo *VNI = liveOutCache_.lookup(MBB).first;
581 assert(VNI && "Missing block value");
582 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
588 // extendTo - Find the last li_ value defined in MBB at or before Idx. The
589 // parentli_ is assumed to be live at Idx. Extend the live range to Idx.
590 // Return the found VNInfo, or NULL.
591 VNInfo *LiveIntervalMap::extendTo(const MachineBasicBlock *MBB, SlotIndex Idx) {
592 assert(li_ && "call reset first");
593 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
594 if (I == li_->begin())
597 if (I->end <= lis_.getMBBStartIdx(MBB))
600 I->end = Idx.getNextSlot();
604 // addSimpleRange - Add a simple range from parentli_ to li_.
605 // ParentVNI must be live in the [Start;End) interval.
606 void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
607 const VNInfo *ParentVNI) {
608 assert(li_ && "call reset first");
610 VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
611 // A simple mapping is easy.
613 li_->addRange(LiveRange(Start, End, VNI));
617 // ParentVNI is a complex value. We must map per MBB.
618 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
619 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
622 li_->addRange(LiveRange(Start, End, VNI));
627 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
629 // Run sequence of full blocks.
630 for (++MBB; MBB != MBBE; ++MBB) {
631 Start = lis_.getMBBStartIdx(MBB);
632 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
633 mapValue(ParentVNI, Start)));
637 Start = lis_.getMBBStartIdx(MBB);
639 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
642 /// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
643 /// All needed values whose def is not inside [Start;End) must be defined
644 /// beforehand so mapValue will work.
645 void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
646 assert(li_ && "call reset first");
647 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
648 LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
650 // Check if --I begins before Start and overlaps.
654 addSimpleRange(Start, std::min(End, I->end), I->valno);
658 // The remaining ranges begin after Start.
659 for (;I != E && I->start < End; ++I)
660 addSimpleRange(I->start, std::min(End, I->end), I->valno);
663 VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
664 const VNInfo *ParentVNI,
665 MachineBasicBlock &MBB,
666 MachineBasicBlock::iterator I) {
667 const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
668 get(TargetOpcode::COPY);
669 MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
670 SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
671 VNInfo *VNI = defValue(ParentVNI, DefIdx);
673 li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
677 //===----------------------------------------------------------------------===//
679 //===----------------------------------------------------------------------===//
681 /// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
682 SplitEditor::SplitEditor(SplitAnalysis &sa,
685 MachineDominatorTree &mdt,
687 : sa_(sa), lis_(lis), vrm_(vrm),
688 mri_(vrm.getMachineFunction().getRegInfo()),
689 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
691 dupli_(lis_, mdt, edit.getParent()),
692 openli_(lis_, mdt, edit.getParent())
696 bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
697 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
698 if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
703 /// Create a new virtual register and live interval.
704 void SplitEditor::openIntv() {
705 assert(!openli_.getLI() && "Previous LI not closed before openIntv");
708 dupli_.reset(&edit_.create(mri_, lis_, vrm_));
710 openli_.reset(&edit_.create(mri_, lis_, vrm_));
713 /// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
714 /// not live before Idx, a COPY is not inserted.
715 void SplitEditor::enterIntvBefore(SlotIndex Idx) {
716 assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
717 DEBUG(dbgs() << " enterIntvBefore " << Idx);
718 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getUseIndex());
720 DEBUG(dbgs() << ": not live\n");
723 DEBUG(dbgs() << ": valno " << ParentVNI->id);
724 truncatedValues.insert(ParentVNI);
725 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
726 assert(MI && "enterIntvBefore called with invalid index");
727 VNInfo *VNI = openli_.defByCopyFrom(edit_.getReg(), ParentVNI,
728 *MI->getParent(), MI);
729 openli_.getLI()->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
730 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
733 /// enterIntvAtEnd - Enter openli at the end of MBB.
734 void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
735 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
736 SlotIndex End = lis_.getMBBEndIdx(&MBB);
737 DEBUG(dbgs() << " enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
738 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(End.getPrevSlot());
740 DEBUG(dbgs() << ": not live\n");
743 DEBUG(dbgs() << ": valno " << ParentVNI->id);
744 truncatedValues.insert(ParentVNI);
745 VNInfo *VNI = openli_.defByCopyFrom(edit_.getReg(), ParentVNI,
746 MBB, MBB.getFirstTerminator());
747 // Make sure openli is live out of MBB.
748 openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
749 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
752 /// useIntv - indicate that all instructions in MBB should use openli.
753 void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
754 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
757 void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
758 assert(openli_.getLI() && "openIntv not called before useIntv");
759 openli_.addRange(Start, End);
760 DEBUG(dbgs() << " use [" << Start << ';' << End << "): "
761 << *openli_.getLI() << '\n');
764 /// leaveIntvAfter - Leave openli after the instruction at Idx.
765 void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
766 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
767 DEBUG(dbgs() << " leaveIntvAfter " << Idx);
769 // The interval must be live beyond the instruction at Idx.
770 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getBoundaryIndex());
772 DEBUG(dbgs() << ": not live\n");
775 DEBUG(dbgs() << ": valno " << ParentVNI->id);
777 MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
778 MachineBasicBlock *MBB = MII->getParent();
779 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI, *MBB,
782 // Finally we must make sure that openli is properly extended from Idx to the
784 openli_.addSimpleRange(Idx.getBoundaryIndex(), VNI->def, ParentVNI);
785 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
788 /// leaveIntvAtTop - Leave the interval at the top of MBB.
789 /// Currently, only one value can leave the interval.
790 void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
791 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
792 SlotIndex Start = lis_.getMBBStartIdx(&MBB);
793 DEBUG(dbgs() << " leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
795 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Start);
797 DEBUG(dbgs() << ": not live\n");
801 // We are going to insert a back copy, so we must have a dupli_.
802 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
805 // Finally we must make sure that openli is properly extended from Start to
807 openli_.addSimpleRange(Start, VNI->def, ParentVNI);
808 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
811 /// closeIntv - Indicate that we are done editing the currently open
812 /// LiveInterval, and ranges can be trimmed.
813 void SplitEditor::closeIntv() {
814 assert(openli_.getLI() && "openIntv not called before closeIntv");
816 DEBUG(dbgs() << " closeIntv cleaning up\n");
817 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n');
821 /// rewrite - Rewrite all uses of reg to use the new registers.
822 void SplitEditor::rewrite(unsigned reg) {
823 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
824 RE = mri_.reg_end(); RI != RE;) {
825 MachineOperand &MO = RI.getOperand();
826 MachineInstr *MI = MO.getParent();
828 if (MI->isDebugValue()) {
829 DEBUG(dbgs() << "Zapping " << *MI);
830 // FIXME: We can do much better with debug values.
834 SlotIndex Idx = lis_.getInstructionIndex(MI);
835 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
836 LiveInterval *LI = 0;
837 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
839 LiveInterval *testli = *I;
840 if (testli->liveAt(Idx)) {
845 DEBUG(dbgs() << " rewr BB#" << MI->getParent()->getNumber() << '\t'<< Idx);
846 assert(LI && "No register was live at use");
848 DEBUG(dbgs() << '\t' << *MI);
853 SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
854 // Build vector of iterator pairs from the intervals.
855 typedef std::pair<LiveInterval::const_iterator,
856 LiveInterval::const_iterator> IIPair;
857 SmallVector<IIPair, 8> Iters;
858 for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
860 if (*LI == dupli_.getLI())
862 LiveInterval::const_iterator I = (*LI)->find(Start);
863 LiveInterval::const_iterator E = (*LI)->end();
865 Iters.push_back(std::make_pair(I, E));
868 SlotIndex sidx = Start;
869 // Break [Start;End) into segments that don't overlap any intervals.
871 SlotIndex next = sidx, eidx = End;
872 // Find overlapping intervals.
873 for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
874 LiveInterval::const_iterator I = Iters[i].first;
875 // Interval I is overlapping [sidx;eidx). Trim sidx.
876 if (I->start <= sidx) {
878 // Move to the next run, remove iters when all are consumed.
879 I = ++Iters[i].first;
880 if (I == Iters[i].second) {
881 Iters.erase(Iters.begin() + i);
886 // Trim eidx too if needed.
887 if (I->start >= eidx)
892 // Now, [sidx;eidx) doesn't overlap anything in intervals_.
894 dupli_.addSimpleRange(sidx, eidx, VNI);
895 // If the interval end was truncated, we can try again from next.
902 void SplitEditor::computeRemainder() {
903 // First we need to fill in the live ranges in dupli.
904 // If values were redefined, we need a full recoloring with SSA update.
905 // If values were truncated, we only need to truncate the ranges.
906 // If values were partially rematted, we should shrink to uses.
907 // If values were fully rematted, they should be omitted.
908 // FIXME: If a single value is redefined, just move the def and truncate.
909 LiveInterval &parent = edit_.getParent();
911 // Values that are fully contained in the split intervals.
912 SmallPtrSet<const VNInfo*, 8> deadValues;
913 // Map all curli values that should have live defs in dupli.
914 for (LiveInterval::const_vni_iterator I = parent.vni_begin(),
915 E = parent.vni_end(); I != E; ++I) {
916 const VNInfo *VNI = *I;
917 // Original def is contained in the split intervals.
918 if (intervalsLiveAt(VNI->def)) {
919 // Did this value escape?
920 if (dupli_.isMapped(VNI))
921 truncatedValues.insert(VNI);
923 deadValues.insert(VNI);
926 // Add minimal live range at the definition.
927 VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
928 dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
931 // Add all ranges to dupli.
932 for (LiveInterval::const_iterator I = parent.begin(), E = parent.end();
934 const LiveRange &LR = *I;
935 if (truncatedValues.count(LR.valno)) {
936 // recolor after removing intervals_.
937 addTruncSimpleRange(LR.start, LR.end, LR.valno);
938 } else if (!deadValues.count(LR.valno)) {
939 // recolor without truncation.
940 dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
944 // Extend dupli_ to be live out of any critical loop predecessors.
945 // This means we have multiple registers live out of those blocks.
946 // The alternative would be to split the critical edges.
947 if (criticalPreds_.empty())
949 for (SplitAnalysis::BlockPtrSet::iterator I = criticalPreds_.begin(),
950 E = criticalPreds_.end(); I != E; ++I)
951 dupli_.extendTo(*I, lis_.getMBBEndIdx(*I).getPrevSlot());
952 criticalPreds_.clear();
955 void SplitEditor::finish() {
956 assert(!openli_.getLI() && "Previous LI not closed before rewrite");
957 assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
959 // Complete dupli liveness.
962 // Get rid of unused values and set phi-kill flags.
963 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
964 (*I)->RenumberValues(lis_);
966 // Rewrite instructions.
967 rewrite(edit_.getReg());
969 // Now check if any registers were separated into multiple components.
970 ConnectedVNInfoEqClasses ConEQ(lis_);
971 for (unsigned i = 0, e = edit_.size(); i != e; ++i) {
972 // Don't use iterators, they are invalidated by create() below.
973 LiveInterval *li = edit_.get(i);
974 unsigned NumComp = ConEQ.Classify(li);
977 DEBUG(dbgs() << " " << NumComp << " components: " << *li << '\n');
978 SmallVector<LiveInterval*, 8> dups;
980 for (unsigned i = 1; i != NumComp; ++i)
981 dups.push_back(&edit_.create(mri_, lis_, vrm_));
982 ConEQ.Distribute(&dups[0]);
983 // Rewrite uses to the new regs.
987 // Calculate spill weight and allocation hints for new intervals.
988 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
989 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
990 LiveInterval &li = **I;
991 vrai.CalculateRegClass(li.reg);
992 vrai.CalculateWeightAndHint(li);
993 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
994 << ":" << li << '\n');
999 //===----------------------------------------------------------------------===//
1001 //===----------------------------------------------------------------------===//
1003 void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
1004 SplitAnalysis::LoopBlocks Blocks;
1005 sa_.getLoopBlocks(Loop, Blocks);
1008 dbgs() << " splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n';
1011 // Break critical edges as needed.
1012 SplitAnalysis::BlockPtrSet CriticalExits;
1013 sa_.getCriticalExits(Blocks, CriticalExits);
1014 assert(CriticalExits.empty() && "Cannot break critical exits yet");
1016 // Get critical predecessors so computeRemainder can deal with them.
1017 sa_.getCriticalPreds(Blocks, criticalPreds_);
1019 // Create new live interval for the loop.
1022 // Insert copies in the predecessors.
1023 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
1024 E = Blocks.Preds.end(); I != E; ++I) {
1025 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
1026 enterIntvAtEnd(MBB);
1029 // Switch all loop blocks.
1030 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
1031 E = Blocks.Loop.end(); I != E; ++I)
1034 // Insert back copies in the exit blocks.
1035 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
1036 E = Blocks.Exits.end(); I != E; ++I) {
1037 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
1038 leaveIntvAtTop(MBB);
1047 //===----------------------------------------------------------------------===//
1048 // Single Block Splitting
1049 //===----------------------------------------------------------------------===//
1051 /// getMultiUseBlocks - if curli has more than one use in a basic block, it
1052 /// may be an advantage to split curli for the duration of the block.
1053 bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
1054 // If curli is local to one block, there is no point to splitting it.
1055 if (usingBlocks_.size() <= 1)
1057 // Add blocks with multiple uses.
1058 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
1060 switch (I->second) {
1065 // When there are only two uses and curli is both live in and live out,
1066 // we don't really win anything by isolating the block since we would be
1067 // inserting two copies.
1068 // The remaing register would still have two uses in the block. (Unless it
1069 // separates into disconnected components).
1070 if (lis_.isLiveInToMBB(*curli_, I->first) &&
1071 lis_.isLiveOutOfMBB(*curli_, I->first))
1075 Blocks.insert(I->first);
1077 return !Blocks.empty();
1080 /// splitSingleBlocks - Split curli into a separate live interval inside each
1081 /// basic block in Blocks.
1082 void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
1083 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
1084 // Determine the first and last instruction using curli in each block.
1085 typedef std::pair<SlotIndex,SlotIndex> IndexPair;
1086 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
1087 IndexPairMap MBBRange;
1088 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1089 E = sa_.usingInstrs_.end(); I != E; ++I) {
1090 const MachineBasicBlock *MBB = (*I)->getParent();
1091 if (!Blocks.count(MBB))
1093 SlotIndex Idx = lis_.getInstructionIndex(*I);
1094 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
1095 IndexPair &IP = MBBRange[MBB];
1096 if (!IP.first.isValid() || Idx < IP.first)
1098 if (!IP.second.isValid() || Idx > IP.second)
1102 // Create a new interval for each block.
1103 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
1104 E = Blocks.end(); I != E; ++I) {
1105 IndexPair &IP = MBBRange[*I];
1106 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": ["
1107 << IP.first << ';' << IP.second << ")\n");
1108 assert(IP.first.isValid() && IP.second.isValid());
1111 enterIntvBefore(IP.first);
1112 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
1113 leaveIntvAfter(IP.second);
1120 //===----------------------------------------------------------------------===//
1121 // Sub Block Splitting
1122 //===----------------------------------------------------------------------===//
1124 /// getBlockForInsideSplit - If curli is contained inside a single basic block,
1125 /// and it wou pay to subdivide the interval inside that block, return it.
1126 /// Otherwise return NULL. The returned block can be passed to
1127 /// SplitEditor::splitInsideBlock.
1128 const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
1129 // The interval must be exclusive to one block.
1130 if (usingBlocks_.size() != 1)
1132 // Don't to this for less than 4 instructions. We want to be sure that
1133 // splitting actually reduces the instruction count per interval.
1134 if (usingInstrs_.size() < 4)
1136 return usingBlocks_.begin()->first;
1139 /// splitInsideBlock - Split curli into multiple intervals inside MBB.
1140 void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1141 SmallVector<SlotIndex, 32> Uses;
1142 Uses.reserve(sa_.usingInstrs_.size());
1143 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1144 E = sa_.usingInstrs_.end(); I != E; ++I)
1145 if ((*I)->getParent() == MBB)
1146 Uses.push_back(lis_.getInstructionIndex(*I));
1147 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for "
1148 << Uses.size() << " instructions.\n");
1149 assert(Uses.size() >= 3 && "Need at least 3 instructions");
1150 array_pod_sort(Uses.begin(), Uses.end());
1152 // Simple algorithm: Find the largest gap between uses as determined by slot
1153 // indices. Create new intervals for instructions before the gap and after the
1155 unsigned bestPos = 0;
1157 DEBUG(dbgs() << " dist (" << Uses[0]);
1158 for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1159 int g = Uses[i-1].distance(Uses[i]);
1160 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1162 bestPos = i, bestGap = g;
1164 DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1166 // bestPos points to the first use after the best gap.
1167 assert(bestPos > 0 && "Invalid gap");
1169 // FIXME: Don't create intervals for low densities.
1171 // First interval before the gap. Don't create single-instr intervals.
1174 enterIntvBefore(Uses.front());
1175 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1176 leaveIntvAfter(Uses[bestPos-1]);
1180 // Second interval after the gap.
1181 if (bestPos < Uses.size()-1) {
1183 enterIntvBefore(Uses[bestPos]);
1184 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1185 leaveIntvAfter(Uses.back());