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/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineLoopInfo.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetMachine.h"
33 AllowSplit("spiller-splits-edges",
34 cl::desc("Allow critical edge splitting during spilling"));
36 //===----------------------------------------------------------------------===//
38 //===----------------------------------------------------------------------===//
40 SplitAnalysis::SplitAnalysis(const MachineFunction &mf,
41 const LiveIntervals &lis,
42 const MachineLoopInfo &mli)
46 tii_(*mf.getTarget().getInstrInfo()),
49 void SplitAnalysis::clear() {
56 bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) {
57 MachineBasicBlock *T, *F;
58 SmallVector<MachineOperand, 4> Cond;
59 return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond);
62 /// analyzeUses - Count instructions, basic blocks, and loops using curli.
63 void SplitAnalysis::analyzeUses() {
64 const MachineRegisterInfo &MRI = mf_.getRegInfo();
65 for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg);
66 MachineInstr *MI = I.skipInstruction();) {
67 if (MI->isDebugValue() || !usingInstrs_.insert(MI))
69 MachineBasicBlock *MBB = MI->getParent();
70 if (usingBlocks_[MBB]++)
72 for (MachineLoop *Loop = loops_.getLoopFor(MBB); Loop;
73 Loop = Loop->getParentLoop())
76 DEBUG(dbgs() << " counted "
77 << usingInstrs_.size() << " instrs, "
78 << usingBlocks_.size() << " blocks, "
79 << usingLoops_.size() << " loops.\n");
82 // Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
83 // predecessor blocks, and exit blocks.
84 void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
87 // Blocks in the loop.
88 Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
90 // Predecessor blocks.
91 const MachineBasicBlock *Header = Loop->getHeader();
92 for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
93 E = Header->pred_end(); I != E; ++I)
94 if (!Blocks.Loop.count(*I))
95 Blocks.Preds.insert(*I);
98 for (MachineLoop::block_iterator I = Loop->block_begin(),
99 E = Loop->block_end(); I != E; ++I) {
100 const MachineBasicBlock *MBB = *I;
101 for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
102 SE = MBB->succ_end(); SI != SE; ++SI)
103 if (!Blocks.Loop.count(*SI))
104 Blocks.Exits.insert(*SI);
108 /// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
109 /// and around the Loop.
110 SplitAnalysis::LoopPeripheralUse SplitAnalysis::
111 analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
112 LoopPeripheralUse use = ContainedInLoop;
113 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
115 const MachineBasicBlock *MBB = I->first;
116 // Is this a peripheral block?
117 if (use < MultiPeripheral &&
118 (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
119 if (I->second > 1) use = MultiPeripheral;
120 else use = SinglePeripheral;
123 // Is it a loop block?
124 if (Blocks.Loop.count(MBB))
126 // It must be an unrelated block.
132 /// getCriticalExits - It may be necessary to partially break critical edges
133 /// leaving the loop if an exit block has phi uses of curli. Collect the exit
134 /// blocks that need special treatment into CriticalExits.
135 void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
136 BlockPtrSet &CriticalExits) {
137 CriticalExits.clear();
139 // A critical exit block contains a phi def of curli, and has a predecessor
140 // that is not in the loop nor a loop predecessor.
141 // For such an exit block, the edges carrying the new variable must be moved
142 // to a new pre-exit block.
143 for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
145 const MachineBasicBlock *Succ = *I;
146 SlotIndex SuccIdx = lis_.getMBBStartIdx(Succ);
147 VNInfo *SuccVNI = curli_->getVNInfoAt(SuccIdx);
148 // This exit may not have curli live in at all. No need to split.
151 // If this is not a PHI def, it is either using a value from before the
152 // loop, or a value defined inside the loop. Both are safe.
153 if (!SuccVNI->isPHIDef() || SuccVNI->def.getBaseIndex() != SuccIdx)
155 // This exit block does have a PHI. Does it also have a predecessor that is
156 // not a loop block or loop predecessor?
157 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
158 PE = Succ->pred_end(); PI != PE; ++PI) {
159 const MachineBasicBlock *Pred = *PI;
160 if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
162 // This is a critical exit block, and we need to split the exit edge.
163 CriticalExits.insert(Succ);
169 /// canSplitCriticalExits - Return true if it is possible to insert new exit
170 /// blocks before the blocks in CriticalExits.
172 SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
173 BlockPtrSet &CriticalExits) {
174 // If we don't allow critical edge splitting, require no critical exits.
176 return CriticalExits.empty();
178 for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
180 const MachineBasicBlock *Succ = *I;
181 // We want to insert a new pre-exit MBB before Succ, and change all the
182 // in-loop blocks to branch to the pre-exit instead of Succ.
183 // Check that all the in-loop predecessors can be changed.
184 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
185 PE = Succ->pred_end(); PI != PE; ++PI) {
186 const MachineBasicBlock *Pred = *PI;
187 // The external predecessors won't be altered.
188 if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
190 if (!canAnalyzeBranch(Pred))
194 // If Succ's layout predecessor falls through, that too must be analyzable.
195 // We need to insert the pre-exit block in the gap.
196 MachineFunction::const_iterator MFI = Succ;
197 if (MFI == mf_.begin())
199 if (!canAnalyzeBranch(--MFI))
202 // No problems found.
206 void SplitAnalysis::analyze(const LiveInterval *li) {
212 const MachineLoop *SplitAnalysis::getBestSplitLoop() {
213 assert(curli_ && "Call analyze() before getBestSplitLoop");
214 if (usingLoops_.empty())
219 BlockPtrSet CriticalExits;
221 // We split around loops where curli is used outside the periphery.
222 for (LoopCountMap::const_iterator I = usingLoops_.begin(),
223 E = usingLoops_.end(); I != E; ++I) {
224 const MachineLoop *Loop = I->first;
225 getLoopBlocks(Loop, Blocks);
227 switch(analyzeLoopPeripheralUse(Blocks)) {
230 case MultiPeripheral:
231 // FIXME: We could split a live range with multiple uses in a peripheral
232 // block and still make progress. However, it is possible that splitting
233 // another live range will insert copies into a peripheral block, and
234 // there is a small chance we can enter an infinity loop, inserting copies
236 // For safety, stick to splitting live ranges with uses outside the
238 DEBUG(dbgs() << " multiple peripheral uses in " << *Loop);
240 case ContainedInLoop:
241 DEBUG(dbgs() << " contained in " << *Loop);
243 case SinglePeripheral:
244 DEBUG(dbgs() << " single peripheral use in " << *Loop);
247 // Will it be possible to split around this loop?
248 getCriticalExits(Blocks, CriticalExits);
249 DEBUG(dbgs() << " " << CriticalExits.size() << " critical exits from "
251 if (!canSplitCriticalExits(Blocks, CriticalExits))
253 // This is a possible split.
257 DEBUG(dbgs() << " getBestSplitLoop found " << Loops.size()
258 << " candidate loops.\n");
263 // Pick the earliest loop.
264 // FIXME: Are there other heuristics to consider?
265 const MachineLoop *Best = 0;
267 for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
269 SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
270 if (!Best || Idx < BestIdx)
271 Best = *I, BestIdx = Idx;
273 DEBUG(dbgs() << " getBestSplitLoop found " << *Best);
277 /// getMultiUseBlocks - if curli has more than one use in a basic block, it
278 /// may be an advantage to split curli for the duration of the block.
279 bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
280 // If curli is local to one block, there is no point to splitting it.
281 if (usingBlocks_.size() <= 1)
283 // Add blocks with multiple uses.
284 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
291 // It doesn't pay to split a 2-instr block if it redefines curli.
292 VNInfo *VN1 = curli_->getVNInfoAt(lis_.getMBBStartIdx(I->first));
294 curli_->getVNInfoAt(lis_.getMBBEndIdx(I->first).getPrevIndex());
295 // live-in and live-out with a different value.
296 if (VN1 && VN2 && VN1 != VN2)
300 Blocks.insert(I->first);
302 return !Blocks.empty();
305 //===----------------------------------------------------------------------===//
307 //===----------------------------------------------------------------------===//
309 // Work around the fact that the std::pair constructors are broken for pointer
310 // pairs in some implementations. makeVV(x, 0) works.
311 static inline std::pair<const VNInfo*, VNInfo*>
312 makeVV(const VNInfo *a, VNInfo *b) {
313 return std::make_pair(a, b);
316 void LiveIntervalMap::reset(LiveInterval *li) {
321 bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
322 ValueMap::const_iterator i = valueMap_.find(ParentVNI);
323 return i != valueMap_.end() && i->second == 0;
326 // defValue - Introduce a li_ def for ParentVNI that could be later than
328 VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
329 assert(li_ && "call reset first");
330 assert(ParentVNI && "Mapping NULL value");
331 assert(Idx.isValid() && "Invalid SlotIndex");
332 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
334 // Create a new value.
335 VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
337 // Use insert for lookup, so we can add missing values with a second lookup.
338 std::pair<ValueMap::iterator,bool> InsP =
339 valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
341 // This is now a complex def. Mark with a NULL in valueMap.
343 InsP.first->second = 0;
349 // mapValue - Find the mapped value for ParentVNI at Idx.
350 // Potentially create phi-def values.
351 VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
353 assert(li_ && "call reset first");
354 assert(ParentVNI && "Mapping NULL value");
355 assert(Idx.isValid() && "Invalid SlotIndex");
356 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
358 // Use insert for lookup, so we can add missing values with a second lookup.
359 std::pair<ValueMap::iterator,bool> InsP =
360 valueMap_.insert(makeVV(ParentVNI, 0));
362 // This was an unknown value. Create a simple mapping.
364 if (simple) *simple = true;
365 return InsP.first->second = li_->createValueCopy(ParentVNI,
366 lis_.getVNInfoAllocator());
369 // This was a simple mapped value.
370 if (InsP.first->second) {
371 if (simple) *simple = true;
372 return InsP.first->second;
375 // This is a complex mapped value. There may be multiple defs, and we may need
376 // to create phi-defs.
377 if (simple) *simple = false;
378 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
379 assert(IdxMBB && "No MBB at Idx");
381 // Is there a def in the same MBB we can extend?
382 if (VNInfo *VNI = extendTo(IdxMBB, Idx))
385 // Now for the fun part. We know that ParentVNI potentially has multiple defs,
386 // and we may need to create even more phi-defs to preserve VNInfo SSA form.
387 // Perform a depth-first search for predecessor blocks where we know the
388 // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
390 // Track MBBs where we have created or learned the dominating value.
391 // This may change during the DFS as we create new phi-defs.
392 typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
393 MBBValueMap DomValue;
394 typedef SplitAnalysis::BlockPtrSet BlockPtrSet;
397 // Iterate over IdxMBB predecessors in a depth-first order.
398 // Skip begin() since that is always IdxMBB.
399 for (idf_ext_iterator<MachineBasicBlock*, BlockPtrSet>
400 IDFI = llvm::next(idf_ext_begin(IdxMBB, Visited)),
401 IDFE = idf_ext_end(IdxMBB, Visited); IDFI != IDFE;) {
402 MachineBasicBlock *MBB = *IDFI;
403 SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
405 // We are operating on the restricted CFG where ParentVNI is live.
406 if (parentli_.getVNInfoAt(End) != ParentVNI) {
411 // Do we have a dominating value in this block?
412 VNInfo *VNI = extendTo(MBB, End);
418 // Yes, VNI dominates MBB. Make sure we visit MBB again from other paths.
421 // Track the path back to IdxMBB, creating phi-defs
422 // as needed along the way.
423 for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
424 // Start from MBB's immediate successor. End at IdxMBB.
425 MachineBasicBlock *Succ = IDFI.getPath(PI-1);
426 std::pair<MBBValueMap::iterator, bool> InsP =
427 DomValue.insert(MBBValueMap::value_type(Succ, VNI));
429 // This is the first time we backtrack to Succ.
433 // We reached Succ again with the same VNI. Nothing is going to change.
434 VNInfo *OVNI = InsP.first->second;
438 // Succ already has a phi-def. No need to continue.
439 SlotIndex Start = lis_.getMBBStartIdx(Succ);
440 if (OVNI->def == Start)
443 // We have a collision between the old and new VNI at Succ. That means
444 // neither dominates and we need a new phi-def.
445 VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
446 VNI->setIsPHIDef(true);
447 InsP.first->second = VNI;
449 // Replace OVNI with VNI in the remaining path.
450 for (; PI > 1 ; --PI) {
451 MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
452 if (I == DomValue.end() || I->second != OVNI)
458 // No need to search the children, we found a dominating value.
462 // The search should at least find a dominating value for IdxMBB.
463 assert(!DomValue.empty() && "Couldn't find a reaching definition");
465 // Since we went through the trouble of a full DFS visiting all reaching defs,
466 // the values in DomValue are now accurate. No more phi-defs are needed for
467 // these blocks, so we can color the live ranges.
468 // This makes the next mapValue call much faster.
470 for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
472 MachineBasicBlock *MBB = I->first;
473 VNInfo *VNI = I->second;
474 SlotIndex Start = lis_.getMBBStartIdx(MBB);
476 // Don't add full liveness to IdxMBB, stop at Idx.
478 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
479 // The caller had better add some liveness to IdxVNI, or it leaks.
482 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
485 assert(IdxVNI && "Didn't find value for Idx");
489 // extendTo - Find the last li_ value defined in MBB at or before Idx. The
490 // parentli_ is assumed to be live at Idx. Extend the live range to Idx.
491 // Return the found VNInfo, or NULL.
492 VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
493 assert(li_ && "call reset first");
494 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
495 if (I == li_->begin())
498 if (I->end <= lis_.getMBBStartIdx(MBB))
501 I->end = Idx.getNextSlot();
505 // addSimpleRange - Add a simple range from parentli_ to li_.
506 // ParentVNI must be live in the [Start;End) interval.
507 void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
508 const VNInfo *ParentVNI) {
509 assert(li_ && "call reset first");
511 VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
512 // A simple mapping is easy.
514 li_->addRange(LiveRange(Start, End, VNI));
518 // ParentVNI is a complex value. We must map per MBB.
519 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
520 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
523 li_->addRange(LiveRange(Start, End, VNI));
528 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
530 // Run sequence of full blocks.
531 for (++MBB; MBB != MBBE; ++MBB) {
532 Start = lis_.getMBBStartIdx(MBB);
533 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
534 mapValue(ParentVNI, Start)));
538 Start = lis_.getMBBStartIdx(MBB);
540 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
543 /// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
544 /// All needed values whose def is not inside [Start;End) must be defined
545 /// beforehand so mapValue will work.
546 void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
547 assert(li_ && "call reset first");
548 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
549 LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
551 // Check if --I begins before Start and overlaps.
555 addSimpleRange(Start, std::min(End, I->end), I->valno);
559 // The remaining ranges begin after Start.
560 for (;I != E && I->start < End; ++I)
561 addSimpleRange(I->start, std::min(End, I->end), I->valno);
564 VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
565 const VNInfo *ParentVNI,
566 MachineBasicBlock &MBB,
567 MachineBasicBlock::iterator I) {
568 const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
569 get(TargetOpcode::COPY);
570 MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
571 SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
572 VNInfo *VNI = defValue(ParentVNI, DefIdx);
574 li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
578 //===----------------------------------------------------------------------===//
580 //===----------------------------------------------------------------------===//
582 /// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
583 SplitEditor::SplitEditor(SplitAnalysis &sa, LiveIntervals &lis, VirtRegMap &vrm,
585 : sa_(sa), lis_(lis), vrm_(vrm),
586 mri_(vrm.getMachineFunction().getRegInfo()),
587 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
589 curli_(sa_.getCurLI()),
590 dupli_(lis_, *curli_),
591 openli_(lis_, *curli_)
593 assert(curli_ && "SplitEditor created from empty SplitAnalysis");
596 bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
597 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
598 if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
603 /// Create a new virtual register and live interval.
604 void SplitEditor::openIntv() {
605 assert(!openli_.getLI() && "Previous LI not closed before openIntv");
608 dupli_.reset(&edit_.create(mri_, lis_, vrm_));
610 openli_.reset(&edit_.create(mri_, lis_, vrm_));
613 /// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
614 /// not live before Idx, a COPY is not inserted.
615 void SplitEditor::enterIntvBefore(SlotIndex Idx) {
616 assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
617 DEBUG(dbgs() << " enterIntvBefore " << Idx);
618 VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getUseIndex());
620 DEBUG(dbgs() << ": not live\n");
623 DEBUG(dbgs() << ": valno " << ParentVNI->id);
624 truncatedValues.insert(ParentVNI);
625 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
626 assert(MI && "enterIntvBefore called with invalid index");
627 VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
628 *MI->getParent(), MI);
629 openli_.getLI()->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
630 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
633 /// enterIntvAtEnd - Enter openli at the end of MBB.
634 void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
635 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
636 SlotIndex End = lis_.getMBBEndIdx(&MBB);
637 DEBUG(dbgs() << " enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
638 VNInfo *ParentVNI = curli_->getVNInfoAt(End.getPrevSlot());
640 DEBUG(dbgs() << ": not live\n");
643 DEBUG(dbgs() << ": valno " << ParentVNI->id);
644 truncatedValues.insert(ParentVNI);
645 VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
646 MBB, MBB.getFirstTerminator());
647 // Make sure openli is live out of MBB.
648 openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
649 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
652 /// useIntv - indicate that all instructions in MBB should use openli.
653 void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
654 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
657 void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
658 assert(openli_.getLI() && "openIntv not called before useIntv");
659 openli_.addRange(Start, End);
660 DEBUG(dbgs() << " use [" << Start << ';' << End << "): "
661 << *openli_.getLI() << '\n');
664 /// leaveIntvAfter - Leave openli after the instruction at Idx.
665 void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
666 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
667 DEBUG(dbgs() << " leaveIntvAfter " << Idx);
669 // The interval must be live beyond the instruction at Idx.
670 VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getBoundaryIndex());
672 DEBUG(dbgs() << ": not live\n");
675 DEBUG(dbgs() << ": valno " << ParentVNI->id);
677 MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
678 MachineBasicBlock *MBB = MII->getParent();
679 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI, *MBB,
682 // Finally we must make sure that openli is properly extended from Idx to the
684 openli_.addSimpleRange(Idx.getBoundaryIndex(), VNI->def, ParentVNI);
685 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
688 /// leaveIntvAtTop - Leave the interval at the top of MBB.
689 /// Currently, only one value can leave the interval.
690 void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
691 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
692 SlotIndex Start = lis_.getMBBStartIdx(&MBB);
693 DEBUG(dbgs() << " leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
695 VNInfo *ParentVNI = curli_->getVNInfoAt(Start);
697 DEBUG(dbgs() << ": not live\n");
701 // We are going to insert a back copy, so we must have a dupli_.
702 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
705 // Finally we must make sure that openli is properly extended from Start to
707 openli_.addSimpleRange(Start, VNI->def, ParentVNI);
708 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
711 /// closeIntv - Indicate that we are done editing the currently open
712 /// LiveInterval, and ranges can be trimmed.
713 void SplitEditor::closeIntv() {
714 assert(openli_.getLI() && "openIntv not called before closeIntv");
716 DEBUG(dbgs() << " closeIntv cleaning up\n");
717 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n');
721 /// rewrite - Rewrite all uses of reg to use the new registers.
722 void SplitEditor::rewrite(unsigned reg) {
723 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
724 RE = mri_.reg_end(); RI != RE;) {
725 MachineOperand &MO = RI.getOperand();
726 MachineInstr *MI = MO.getParent();
728 if (MI->isDebugValue()) {
729 DEBUG(dbgs() << "Zapping " << *MI);
730 // FIXME: We can do much better with debug values.
734 SlotIndex Idx = lis_.getInstructionIndex(MI);
735 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
736 LiveInterval *LI = 0;
737 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
739 LiveInterval *testli = *I;
740 if (testli->liveAt(Idx)) {
745 assert(LI && "No register was live at use");
747 DEBUG(dbgs() << " rewrite BB#" << MI->getParent()->getNumber() << '\t'
748 << Idx << '\t' << *MI);
753 SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
754 // Build vector of iterator pairs from the intervals.
755 typedef std::pair<LiveInterval::const_iterator,
756 LiveInterval::const_iterator> IIPair;
757 SmallVector<IIPair, 8> Iters;
758 for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
760 LiveInterval::const_iterator I = (*LI)->find(Start);
761 LiveInterval::const_iterator E = (*LI)->end();
763 Iters.push_back(std::make_pair(I, E));
766 SlotIndex sidx = Start;
767 // Break [Start;End) into segments that don't overlap any intervals.
769 SlotIndex next = sidx, eidx = End;
770 // Find overlapping intervals.
771 for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
772 LiveInterval::const_iterator I = Iters[i].first;
773 // Interval I is overlapping [sidx;eidx). Trim sidx.
774 if (I->start <= sidx) {
776 // Move to the next run, remove iters when all are consumed.
777 I = ++Iters[i].first;
778 if (I == Iters[i].second) {
779 Iters.erase(Iters.begin() + i);
784 // Trim eidx too if needed.
785 if (I->start >= eidx)
790 // Now, [sidx;eidx) doesn't overlap anything in intervals_.
792 dupli_.addSimpleRange(sidx, eidx, VNI);
793 // If the interval end was truncated, we can try again from next.
800 void SplitEditor::computeRemainder() {
801 // First we need to fill in the live ranges in dupli.
802 // If values were redefined, we need a full recoloring with SSA update.
803 // If values were truncated, we only need to truncate the ranges.
804 // If values were partially rematted, we should shrink to uses.
805 // If values were fully rematted, they should be omitted.
806 // FIXME: If a single value is redefined, just move the def and truncate.
808 // Values that are fully contained in the split intervals.
809 SmallPtrSet<const VNInfo*, 8> deadValues;
811 // Map all curli values that should have live defs in dupli.
812 for (LiveInterval::const_vni_iterator I = curli_->vni_begin(),
813 E = curli_->vni_end(); I != E; ++I) {
814 const VNInfo *VNI = *I;
815 // Original def is contained in the split intervals.
816 if (intervalsLiveAt(VNI->def)) {
817 // Did this value escape?
818 if (dupli_.isMapped(VNI))
819 truncatedValues.insert(VNI);
821 deadValues.insert(VNI);
824 // Add minimal live range at the definition.
825 VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
826 dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
829 // Add all ranges to dupli.
830 for (LiveInterval::const_iterator I = curli_->begin(), E = curli_->end();
832 const LiveRange &LR = *I;
833 if (truncatedValues.count(LR.valno)) {
834 // recolor after removing intervals_.
835 addTruncSimpleRange(LR.start, LR.end, LR.valno);
836 } else if (!deadValues.count(LR.valno)) {
837 // recolor without truncation.
838 dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
843 void SplitEditor::finish() {
844 assert(!openli_.getLI() && "Previous LI not closed before rewrite");
845 assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
847 // Complete dupli liveness.
850 // Get rid of unused values and set phi-kill flags.
851 dupli_.getLI()->RenumberValues(lis_);
853 // Now check if dupli was separated into multiple connected components.
854 ConnectedVNInfoEqClasses ConEQ(lis_);
855 if (unsigned NumComp = ConEQ.Classify(dupli_.getLI())) {
856 DEBUG(dbgs() << " Remainder has " << NumComp << " connected components: "
857 << *dupli_.getLI() << '\n');
858 // Did the remainder break up? Create intervals for all the components.
860 SmallVector<LiveInterval*, 8> dups;
861 dups.push_back(dupli_.getLI());
862 for (unsigned i = 1; i != NumComp; ++i)
863 dups.push_back(&edit_.create(mri_, lis_, vrm_));
864 ConEQ.Distribute(&dups[0]);
865 // Rewrite uses to the new regs.
866 rewrite(dupli_.getLI()->reg);
870 // Rewrite instructions.
871 rewrite(curli_->reg);
873 // Calculate spill weight and allocation hints for new intervals.
874 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
875 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
876 LiveInterval &li = **I;
877 vrai.CalculateRegClass(li.reg);
878 vrai.CalculateWeightAndHint(li);
879 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
880 << ":" << li << '\n');
885 //===----------------------------------------------------------------------===//
887 //===----------------------------------------------------------------------===//
889 void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
890 SplitAnalysis::LoopBlocks Blocks;
891 sa_.getLoopBlocks(Loop, Blocks);
894 dbgs() << " splitAroundLoop";
895 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
896 E = Blocks.Loop.end(); I != E; ++I)
897 dbgs() << " BB#" << (*I)->getNumber();
898 dbgs() << ", preds:";
899 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
900 E = Blocks.Preds.end(); I != E; ++I)
901 dbgs() << " BB#" << (*I)->getNumber();
902 dbgs() << ", exits:";
903 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
904 E = Blocks.Exits.end(); I != E; ++I)
905 dbgs() << " BB#" << (*I)->getNumber();
909 // Break critical edges as needed.
910 SplitAnalysis::BlockPtrSet CriticalExits;
911 sa_.getCriticalExits(Blocks, CriticalExits);
912 assert(CriticalExits.empty() && "Cannot break critical exits yet");
914 // Create new live interval for the loop.
917 // Insert copies in the predecessors.
918 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
919 E = Blocks.Preds.end(); I != E; ++I) {
920 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
924 // Switch all loop blocks.
925 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
926 E = Blocks.Loop.end(); I != E; ++I)
929 // Insert back copies in the exit blocks.
930 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
931 E = Blocks.Exits.end(); I != E; ++I) {
932 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
942 //===----------------------------------------------------------------------===//
943 // Single Block Splitting
944 //===----------------------------------------------------------------------===//
946 /// splitSingleBlocks - Split curli into a separate live interval inside each
947 /// basic block in Blocks.
948 void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
949 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
950 // Determine the first and last instruction using curli in each block.
951 typedef std::pair<SlotIndex,SlotIndex> IndexPair;
952 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
953 IndexPairMap MBBRange;
954 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
955 E = sa_.usingInstrs_.end(); I != E; ++I) {
956 const MachineBasicBlock *MBB = (*I)->getParent();
957 if (!Blocks.count(MBB))
959 SlotIndex Idx = lis_.getInstructionIndex(*I);
960 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
961 IndexPair &IP = MBBRange[MBB];
962 if (!IP.first.isValid() || Idx < IP.first)
964 if (!IP.second.isValid() || Idx > IP.second)
968 // Create a new interval for each block.
969 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
970 E = Blocks.end(); I != E; ++I) {
971 IndexPair &IP = MBBRange[*I];
972 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": ["
973 << IP.first << ';' << IP.second << ")\n");
974 assert(IP.first.isValid() && IP.second.isValid());
977 enterIntvBefore(IP.first);
978 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
979 leaveIntvAfter(IP.second);
986 //===----------------------------------------------------------------------===//
987 // Sub Block Splitting
988 //===----------------------------------------------------------------------===//
990 /// getBlockForInsideSplit - If curli is contained inside a single basic block,
991 /// and it wou pay to subdivide the interval inside that block, return it.
992 /// Otherwise return NULL. The returned block can be passed to
993 /// SplitEditor::splitInsideBlock.
994 const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
995 // The interval must be exclusive to one block.
996 if (usingBlocks_.size() != 1)
998 // Don't to this for less than 4 instructions. We want to be sure that
999 // splitting actually reduces the instruction count per interval.
1000 if (usingInstrs_.size() < 4)
1002 return usingBlocks_.begin()->first;
1005 /// splitInsideBlock - Split curli into multiple intervals inside MBB.
1006 void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1007 SmallVector<SlotIndex, 32> Uses;
1008 Uses.reserve(sa_.usingInstrs_.size());
1009 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1010 E = sa_.usingInstrs_.end(); I != E; ++I)
1011 if ((*I)->getParent() == MBB)
1012 Uses.push_back(lis_.getInstructionIndex(*I));
1013 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for "
1014 << Uses.size() << " instructions.\n");
1015 assert(Uses.size() >= 3 && "Need at least 3 instructions");
1016 array_pod_sort(Uses.begin(), Uses.end());
1018 // Simple algorithm: Find the largest gap between uses as determined by slot
1019 // indices. Create new intervals for instructions before the gap and after the
1021 unsigned bestPos = 0;
1023 DEBUG(dbgs() << " dist (" << Uses[0]);
1024 for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1025 int g = Uses[i-1].distance(Uses[i]);
1026 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1028 bestPos = i, bestGap = g;
1030 DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1032 // bestPos points to the first use after the best gap.
1033 assert(bestPos > 0 && "Invalid gap");
1035 // FIXME: Don't create intervals for low densities.
1037 // First interval before the gap. Don't create single-instr intervals.
1040 enterIntvBefore(Uses.front());
1041 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1042 leaveIntvAfter(Uses[bestPos-1]);
1046 // Second interval after the gap.
1047 if (bestPos < Uses.size()-1) {
1049 enterIntvBefore(Uses[bestPos]);
1050 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1051 leaveIntvAfter(Uses.back());