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) {
341 bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
342 ValueMap::const_iterator i = valueMap_.find(ParentVNI);
343 return i != valueMap_.end() && i->second == 0;
346 // defValue - Introduce a li_ def for ParentVNI that could be later than
348 VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
349 assert(li_ && "call reset first");
350 assert(ParentVNI && "Mapping NULL value");
351 assert(Idx.isValid() && "Invalid SlotIndex");
352 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
354 // Create a new value.
355 VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
357 // Preserve the PHIDef bit.
358 if (ParentVNI->isPHIDef() && Idx == ParentVNI->def)
359 VNI->setIsPHIDef(true);
361 // Use insert for lookup, so we can add missing values with a second lookup.
362 std::pair<ValueMap::iterator,bool> InsP =
363 valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
365 // This is now a complex def. Mark with a NULL in valueMap.
367 InsP.first->second = 0;
373 // mapValue - Find the mapped value for ParentVNI at Idx.
374 // Potentially create phi-def values.
375 VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
377 assert(li_ && "call reset first");
378 assert(ParentVNI && "Mapping NULL value");
379 assert(Idx.isValid() && "Invalid SlotIndex");
380 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
382 // Use insert for lookup, so we can add missing values with a second lookup.
383 std::pair<ValueMap::iterator,bool> InsP =
384 valueMap_.insert(makeVV(ParentVNI, 0));
386 // This was an unknown value. Create a simple mapping.
388 if (simple) *simple = true;
389 return InsP.first->second = li_->createValueCopy(ParentVNI,
390 lis_.getVNInfoAllocator());
393 // This was a simple mapped value.
394 if (InsP.first->second) {
395 if (simple) *simple = true;
396 return InsP.first->second;
399 // This is a complex mapped value. There may be multiple defs, and we may need
400 // to create phi-defs.
401 if (simple) *simple = false;
402 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
403 assert(IdxMBB && "No MBB at Idx");
405 // Is there a def in the same MBB we can extend?
406 if (VNInfo *VNI = extendTo(IdxMBB, Idx))
409 // Now for the fun part. We know that ParentVNI potentially has multiple defs,
410 // and we may need to create even more phi-defs to preserve VNInfo SSA form.
411 // Perform a depth-first search for predecessor blocks where we know the
412 // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
414 // Track MBBs where we have created or learned the dominating value.
415 // This may change during the DFS as we create new phi-defs.
416 typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
417 MBBValueMap DomValue;
418 typedef SplitAnalysis::BlockPtrSet BlockPtrSet;
421 // Iterate over IdxMBB predecessors in a depth-first order.
422 // Skip begin() since that is always IdxMBB.
423 for (idf_ext_iterator<MachineBasicBlock*, BlockPtrSet>
424 IDFI = llvm::next(idf_ext_begin(IdxMBB, Visited)),
425 IDFE = idf_ext_end(IdxMBB, Visited); IDFI != IDFE;) {
426 MachineBasicBlock *MBB = *IDFI;
427 SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
429 // We are operating on the restricted CFG where ParentVNI is live.
430 if (parentli_.getVNInfoAt(End) != ParentVNI) {
435 // Do we have a dominating value in this block?
436 VNInfo *VNI = extendTo(MBB, End);
442 // Yes, VNI dominates MBB. Make sure we visit MBB again from other paths.
445 // Track the path back to IdxMBB, creating phi-defs
446 // as needed along the way.
447 for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
448 // Start from MBB's immediate successor. End at IdxMBB.
449 MachineBasicBlock *Succ = IDFI.getPath(PI-1);
450 std::pair<MBBValueMap::iterator, bool> InsP =
451 DomValue.insert(MBBValueMap::value_type(Succ, VNI));
453 // This is the first time we backtrack to Succ.
457 // We reached Succ again with the same VNI. Nothing is going to change.
458 VNInfo *OVNI = InsP.first->second;
462 // Succ already has a phi-def. No need to continue.
463 SlotIndex Start = lis_.getMBBStartIdx(Succ);
464 if (OVNI->def == Start)
467 // We have a collision between the old and new VNI at Succ. That means
468 // neither dominates and we need a new phi-def.
469 VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
470 VNI->setIsPHIDef(true);
471 InsP.first->second = VNI;
473 // Replace OVNI with VNI in the remaining path.
474 for (; PI > 1 ; --PI) {
475 MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
476 if (I == DomValue.end() || I->second != OVNI)
482 // No need to search the children, we found a dominating value.
486 // The search should at least find a dominating value for IdxMBB.
487 assert(!DomValue.empty() && "Couldn't find a reaching definition");
489 // Since we went through the trouble of a full DFS visiting all reaching defs,
490 // the values in DomValue are now accurate. No more phi-defs are needed for
491 // these blocks, so we can color the live ranges.
492 // This makes the next mapValue call much faster.
494 for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
496 MachineBasicBlock *MBB = I->first;
497 VNInfo *VNI = I->second;
498 SlotIndex Start = lis_.getMBBStartIdx(MBB);
500 // Don't add full liveness to IdxMBB, stop at Idx.
502 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
503 // The caller had better add some liveness to IdxVNI, or it leaks.
506 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
509 assert(IdxVNI && "Didn't find value for Idx");
513 // extendTo - Find the last li_ value defined in MBB at or before Idx. The
514 // parentli_ is assumed to be live at Idx. Extend the live range to Idx.
515 // Return the found VNInfo, or NULL.
516 VNInfo *LiveIntervalMap::extendTo(const MachineBasicBlock *MBB, SlotIndex Idx) {
517 assert(li_ && "call reset first");
518 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
519 if (I == li_->begin())
522 if (I->end <= lis_.getMBBStartIdx(MBB))
525 I->end = Idx.getNextSlot();
529 // addSimpleRange - Add a simple range from parentli_ to li_.
530 // ParentVNI must be live in the [Start;End) interval.
531 void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
532 const VNInfo *ParentVNI) {
533 assert(li_ && "call reset first");
535 VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
536 // A simple mapping is easy.
538 li_->addRange(LiveRange(Start, End, VNI));
542 // ParentVNI is a complex value. We must map per MBB.
543 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
544 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
547 li_->addRange(LiveRange(Start, End, VNI));
552 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
554 // Run sequence of full blocks.
555 for (++MBB; MBB != MBBE; ++MBB) {
556 Start = lis_.getMBBStartIdx(MBB);
557 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
558 mapValue(ParentVNI, Start)));
562 Start = lis_.getMBBStartIdx(MBB);
564 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
567 /// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
568 /// All needed values whose def is not inside [Start;End) must be defined
569 /// beforehand so mapValue will work.
570 void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
571 assert(li_ && "call reset first");
572 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
573 LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
575 // Check if --I begins before Start and overlaps.
579 addSimpleRange(Start, std::min(End, I->end), I->valno);
583 // The remaining ranges begin after Start.
584 for (;I != E && I->start < End; ++I)
585 addSimpleRange(I->start, std::min(End, I->end), I->valno);
588 VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
589 const VNInfo *ParentVNI,
590 MachineBasicBlock &MBB,
591 MachineBasicBlock::iterator I) {
592 const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
593 get(TargetOpcode::COPY);
594 MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
595 SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
596 VNInfo *VNI = defValue(ParentVNI, DefIdx);
598 li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
602 //===----------------------------------------------------------------------===//
604 //===----------------------------------------------------------------------===//
606 /// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
607 SplitEditor::SplitEditor(SplitAnalysis &sa,
610 MachineDominatorTree &mdt,
612 : sa_(sa), lis_(lis), vrm_(vrm),
613 mri_(vrm.getMachineFunction().getRegInfo()),
614 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
616 dupli_(lis_, mdt, edit.getParent()),
617 openli_(lis_, mdt, edit.getParent())
621 bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
622 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
623 if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
628 /// Create a new virtual register and live interval.
629 void SplitEditor::openIntv() {
630 assert(!openli_.getLI() && "Previous LI not closed before openIntv");
633 dupli_.reset(&edit_.create(mri_, lis_, vrm_));
635 openli_.reset(&edit_.create(mri_, lis_, vrm_));
638 /// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
639 /// not live before Idx, a COPY is not inserted.
640 void SplitEditor::enterIntvBefore(SlotIndex Idx) {
641 assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
642 DEBUG(dbgs() << " enterIntvBefore " << Idx);
643 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getUseIndex());
645 DEBUG(dbgs() << ": not live\n");
648 DEBUG(dbgs() << ": valno " << ParentVNI->id);
649 truncatedValues.insert(ParentVNI);
650 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
651 assert(MI && "enterIntvBefore called with invalid index");
652 VNInfo *VNI = openli_.defByCopyFrom(edit_.getReg(), ParentVNI,
653 *MI->getParent(), MI);
654 openli_.getLI()->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
655 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
658 /// enterIntvAtEnd - Enter openli at the end of MBB.
659 void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
660 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
661 SlotIndex End = lis_.getMBBEndIdx(&MBB);
662 DEBUG(dbgs() << " enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
663 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(End.getPrevSlot());
665 DEBUG(dbgs() << ": not live\n");
668 DEBUG(dbgs() << ": valno " << ParentVNI->id);
669 truncatedValues.insert(ParentVNI);
670 VNInfo *VNI = openli_.defByCopyFrom(edit_.getReg(), ParentVNI,
671 MBB, MBB.getFirstTerminator());
672 // Make sure openli is live out of MBB.
673 openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
674 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
677 /// useIntv - indicate that all instructions in MBB should use openli.
678 void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
679 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
682 void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
683 assert(openli_.getLI() && "openIntv not called before useIntv");
684 openli_.addRange(Start, End);
685 DEBUG(dbgs() << " use [" << Start << ';' << End << "): "
686 << *openli_.getLI() << '\n');
689 /// leaveIntvAfter - Leave openli after the instruction at Idx.
690 void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
691 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
692 DEBUG(dbgs() << " leaveIntvAfter " << Idx);
694 // The interval must be live beyond the instruction at Idx.
695 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getBoundaryIndex());
697 DEBUG(dbgs() << ": not live\n");
700 DEBUG(dbgs() << ": valno " << ParentVNI->id);
702 MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
703 MachineBasicBlock *MBB = MII->getParent();
704 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI, *MBB,
707 // Finally we must make sure that openli is properly extended from Idx to the
709 openli_.addSimpleRange(Idx.getBoundaryIndex(), VNI->def, ParentVNI);
710 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
713 /// leaveIntvAtTop - Leave the interval at the top of MBB.
714 /// Currently, only one value can leave the interval.
715 void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
716 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
717 SlotIndex Start = lis_.getMBBStartIdx(&MBB);
718 DEBUG(dbgs() << " leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
720 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Start);
722 DEBUG(dbgs() << ": not live\n");
726 // We are going to insert a back copy, so we must have a dupli_.
727 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
730 // Finally we must make sure that openli is properly extended from Start to
732 openli_.addSimpleRange(Start, VNI->def, ParentVNI);
733 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
736 /// closeIntv - Indicate that we are done editing the currently open
737 /// LiveInterval, and ranges can be trimmed.
738 void SplitEditor::closeIntv() {
739 assert(openli_.getLI() && "openIntv not called before closeIntv");
741 DEBUG(dbgs() << " closeIntv cleaning up\n");
742 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n');
746 /// rewrite - Rewrite all uses of reg to use the new registers.
747 void SplitEditor::rewrite(unsigned reg) {
748 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
749 RE = mri_.reg_end(); RI != RE;) {
750 MachineOperand &MO = RI.getOperand();
751 MachineInstr *MI = MO.getParent();
753 if (MI->isDebugValue()) {
754 DEBUG(dbgs() << "Zapping " << *MI);
755 // FIXME: We can do much better with debug values.
759 SlotIndex Idx = lis_.getInstructionIndex(MI);
760 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
761 LiveInterval *LI = 0;
762 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
764 LiveInterval *testli = *I;
765 if (testli->liveAt(Idx)) {
770 DEBUG(dbgs() << " rewr BB#" << MI->getParent()->getNumber() << '\t'<< Idx);
771 assert(LI && "No register was live at use");
773 DEBUG(dbgs() << '\t' << *MI);
778 SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
779 // Build vector of iterator pairs from the intervals.
780 typedef std::pair<LiveInterval::const_iterator,
781 LiveInterval::const_iterator> IIPair;
782 SmallVector<IIPair, 8> Iters;
783 for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
785 if (*LI == dupli_.getLI())
787 LiveInterval::const_iterator I = (*LI)->find(Start);
788 LiveInterval::const_iterator E = (*LI)->end();
790 Iters.push_back(std::make_pair(I, E));
793 SlotIndex sidx = Start;
794 // Break [Start;End) into segments that don't overlap any intervals.
796 SlotIndex next = sidx, eidx = End;
797 // Find overlapping intervals.
798 for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
799 LiveInterval::const_iterator I = Iters[i].first;
800 // Interval I is overlapping [sidx;eidx). Trim sidx.
801 if (I->start <= sidx) {
803 // Move to the next run, remove iters when all are consumed.
804 I = ++Iters[i].first;
805 if (I == Iters[i].second) {
806 Iters.erase(Iters.begin() + i);
811 // Trim eidx too if needed.
812 if (I->start >= eidx)
817 // Now, [sidx;eidx) doesn't overlap anything in intervals_.
819 dupli_.addSimpleRange(sidx, eidx, VNI);
820 // If the interval end was truncated, we can try again from next.
827 void SplitEditor::computeRemainder() {
828 // First we need to fill in the live ranges in dupli.
829 // If values were redefined, we need a full recoloring with SSA update.
830 // If values were truncated, we only need to truncate the ranges.
831 // If values were partially rematted, we should shrink to uses.
832 // If values were fully rematted, they should be omitted.
833 // FIXME: If a single value is redefined, just move the def and truncate.
834 LiveInterval &parent = edit_.getParent();
836 // Values that are fully contained in the split intervals.
837 SmallPtrSet<const VNInfo*, 8> deadValues;
838 // Map all curli values that should have live defs in dupli.
839 for (LiveInterval::const_vni_iterator I = parent.vni_begin(),
840 E = parent.vni_end(); I != E; ++I) {
841 const VNInfo *VNI = *I;
842 // Original def is contained in the split intervals.
843 if (intervalsLiveAt(VNI->def)) {
844 // Did this value escape?
845 if (dupli_.isMapped(VNI))
846 truncatedValues.insert(VNI);
848 deadValues.insert(VNI);
851 // Add minimal live range at the definition.
852 VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
853 dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
856 // Add all ranges to dupli.
857 for (LiveInterval::const_iterator I = parent.begin(), E = parent.end();
859 const LiveRange &LR = *I;
860 if (truncatedValues.count(LR.valno)) {
861 // recolor after removing intervals_.
862 addTruncSimpleRange(LR.start, LR.end, LR.valno);
863 } else if (!deadValues.count(LR.valno)) {
864 // recolor without truncation.
865 dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
869 // Extend dupli_ to be live out of any critical loop predecessors.
870 // This means we have multiple registers live out of those blocks.
871 // The alternative would be to split the critical edges.
872 if (criticalPreds_.empty())
874 for (SplitAnalysis::BlockPtrSet::iterator I = criticalPreds_.begin(),
875 E = criticalPreds_.end(); I != E; ++I)
876 dupli_.extendTo(*I, lis_.getMBBEndIdx(*I).getPrevSlot());
877 criticalPreds_.clear();
880 void SplitEditor::finish() {
881 assert(!openli_.getLI() && "Previous LI not closed before rewrite");
882 assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
884 // Complete dupli liveness.
887 // Get rid of unused values and set phi-kill flags.
888 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
889 (*I)->RenumberValues(lis_);
891 // Rewrite instructions.
892 rewrite(edit_.getReg());
894 // Now check if any registers were separated into multiple components.
895 ConnectedVNInfoEqClasses ConEQ(lis_);
896 for (unsigned i = 0, e = edit_.size(); i != e; ++i) {
897 // Don't use iterators, they are invalidated by create() below.
898 LiveInterval *li = edit_.get(i);
899 unsigned NumComp = ConEQ.Classify(li);
902 DEBUG(dbgs() << " " << NumComp << " components: " << *li << '\n');
903 SmallVector<LiveInterval*, 8> dups;
905 for (unsigned i = 1; i != NumComp; ++i)
906 dups.push_back(&edit_.create(mri_, lis_, vrm_));
907 ConEQ.Distribute(&dups[0]);
908 // Rewrite uses to the new regs.
912 // Calculate spill weight and allocation hints for new intervals.
913 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
914 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
915 LiveInterval &li = **I;
916 vrai.CalculateRegClass(li.reg);
917 vrai.CalculateWeightAndHint(li);
918 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
919 << ":" << li << '\n');
924 //===----------------------------------------------------------------------===//
926 //===----------------------------------------------------------------------===//
928 void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
929 SplitAnalysis::LoopBlocks Blocks;
930 sa_.getLoopBlocks(Loop, Blocks);
933 dbgs() << " splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n';
936 // Break critical edges as needed.
937 SplitAnalysis::BlockPtrSet CriticalExits;
938 sa_.getCriticalExits(Blocks, CriticalExits);
939 assert(CriticalExits.empty() && "Cannot break critical exits yet");
941 // Get critical predecessors so computeRemainder can deal with them.
942 sa_.getCriticalPreds(Blocks, criticalPreds_);
944 // Create new live interval for the loop.
947 // Insert copies in the predecessors.
948 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
949 E = Blocks.Preds.end(); I != E; ++I) {
950 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
954 // Switch all loop blocks.
955 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
956 E = Blocks.Loop.end(); I != E; ++I)
959 // Insert back copies in the exit blocks.
960 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
961 E = Blocks.Exits.end(); I != E; ++I) {
962 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
972 //===----------------------------------------------------------------------===//
973 // Single Block Splitting
974 //===----------------------------------------------------------------------===//
976 /// getMultiUseBlocks - if curli has more than one use in a basic block, it
977 /// may be an advantage to split curli for the duration of the block.
978 bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
979 // If curli is local to one block, there is no point to splitting it.
980 if (usingBlocks_.size() <= 1)
982 // Add blocks with multiple uses.
983 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
990 // When there are only two uses and curli is both live in and live out,
991 // we don't really win anything by isolating the block since we would be
992 // inserting two copies.
993 // The remaing register would still have two uses in the block. (Unless it
994 // separates into disconnected components).
995 if (lis_.isLiveInToMBB(*curli_, I->first) &&
996 lis_.isLiveOutOfMBB(*curli_, I->first))
1000 Blocks.insert(I->first);
1002 return !Blocks.empty();
1005 /// splitSingleBlocks - Split curli into a separate live interval inside each
1006 /// basic block in Blocks.
1007 void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
1008 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
1009 // Determine the first and last instruction using curli in each block.
1010 typedef std::pair<SlotIndex,SlotIndex> IndexPair;
1011 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
1012 IndexPairMap MBBRange;
1013 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1014 E = sa_.usingInstrs_.end(); I != E; ++I) {
1015 const MachineBasicBlock *MBB = (*I)->getParent();
1016 if (!Blocks.count(MBB))
1018 SlotIndex Idx = lis_.getInstructionIndex(*I);
1019 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
1020 IndexPair &IP = MBBRange[MBB];
1021 if (!IP.first.isValid() || Idx < IP.first)
1023 if (!IP.second.isValid() || Idx > IP.second)
1027 // Create a new interval for each block.
1028 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
1029 E = Blocks.end(); I != E; ++I) {
1030 IndexPair &IP = MBBRange[*I];
1031 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": ["
1032 << IP.first << ';' << IP.second << ")\n");
1033 assert(IP.first.isValid() && IP.second.isValid());
1036 enterIntvBefore(IP.first);
1037 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
1038 leaveIntvAfter(IP.second);
1045 //===----------------------------------------------------------------------===//
1046 // Sub Block Splitting
1047 //===----------------------------------------------------------------------===//
1049 /// getBlockForInsideSplit - If curli is contained inside a single basic block,
1050 /// and it wou pay to subdivide the interval inside that block, return it.
1051 /// Otherwise return NULL. The returned block can be passed to
1052 /// SplitEditor::splitInsideBlock.
1053 const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
1054 // The interval must be exclusive to one block.
1055 if (usingBlocks_.size() != 1)
1057 // Don't to this for less than 4 instructions. We want to be sure that
1058 // splitting actually reduces the instruction count per interval.
1059 if (usingInstrs_.size() < 4)
1061 return usingBlocks_.begin()->first;
1064 /// splitInsideBlock - Split curli into multiple intervals inside MBB.
1065 void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1066 SmallVector<SlotIndex, 32> Uses;
1067 Uses.reserve(sa_.usingInstrs_.size());
1068 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1069 E = sa_.usingInstrs_.end(); I != E; ++I)
1070 if ((*I)->getParent() == MBB)
1071 Uses.push_back(lis_.getInstructionIndex(*I));
1072 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for "
1073 << Uses.size() << " instructions.\n");
1074 assert(Uses.size() >= 3 && "Need at least 3 instructions");
1075 array_pod_sort(Uses.begin(), Uses.end());
1077 // Simple algorithm: Find the largest gap between uses as determined by slot
1078 // indices. Create new intervals for instructions before the gap and after the
1080 unsigned bestPos = 0;
1082 DEBUG(dbgs() << " dist (" << Uses[0]);
1083 for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1084 int g = Uses[i-1].distance(Uses[i]);
1085 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1087 bestPos = i, bestGap = g;
1089 DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1091 // bestPos points to the first use after the best gap.
1092 assert(bestPos > 0 && "Invalid gap");
1094 // FIXME: Don't create intervals for low densities.
1096 // First interval before the gap. Don't create single-instr intervals.
1099 enterIntvBefore(Uses.front());
1100 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1101 leaveIntvAfter(Uses[bestPos-1]);
1105 // Second interval after the gap.
1106 if (bestPos < Uses.size()-1) {
1108 enterIntvBefore(Uses[bestPos]);
1109 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1110 leaveIntvAfter(Uses.back());