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 void SplitAnalysis::print(const BlockPtrSet &B, raw_ostream &OS) const {
83 for (BlockPtrSet::const_iterator I = B.begin(), E = B.end(); I != E; ++I) {
84 unsigned count = usingBlocks_.lookup(*I);
85 OS << " BB#" << (*I)->getNumber();
87 OS << '(' << count << ')';
91 // Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
92 // predecessor blocks, and exit blocks.
93 void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
96 // Blocks in the loop.
97 Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
99 // Predecessor blocks.
100 const MachineBasicBlock *Header = Loop->getHeader();
101 for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
102 E = Header->pred_end(); I != E; ++I)
103 if (!Blocks.Loop.count(*I))
104 Blocks.Preds.insert(*I);
107 for (MachineLoop::block_iterator I = Loop->block_begin(),
108 E = Loop->block_end(); I != E; ++I) {
109 const MachineBasicBlock *MBB = *I;
110 for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
111 SE = MBB->succ_end(); SI != SE; ++SI)
112 if (!Blocks.Loop.count(*SI))
113 Blocks.Exits.insert(*SI);
117 void SplitAnalysis::print(const LoopBlocks &B, raw_ostream &OS) const {
126 /// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
127 /// and around the Loop.
128 SplitAnalysis::LoopPeripheralUse SplitAnalysis::
129 analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
130 LoopPeripheralUse use = ContainedInLoop;
131 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
133 const MachineBasicBlock *MBB = I->first;
134 // Is this a peripheral block?
135 if (use < MultiPeripheral &&
136 (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
137 if (I->second > 1) use = MultiPeripheral;
138 else use = SinglePeripheral;
141 // Is it a loop block?
142 if (Blocks.Loop.count(MBB))
144 // It must be an unrelated block.
145 DEBUG(dbgs() << ", outside: BB#" << MBB->getNumber());
151 /// getCriticalExits - It may be necessary to partially break critical edges
152 /// leaving the loop if an exit block has predecessors from outside the loop
154 void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
155 BlockPtrSet &CriticalExits) {
156 CriticalExits.clear();
158 // A critical exit block has curli line-in, and has a predecessor that is not
159 // in the loop nor a loop predecessor. For such an exit block, the edges
160 // carrying the new variable must be moved to a new pre-exit block.
161 for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
163 const MachineBasicBlock *Exit = *I;
164 // A single-predecessor exit block is definitely not a critical edge.
165 if (Exit->pred_size() == 1)
167 // This exit may not have curli live in at all. No need to split.
168 if (!lis_.isLiveInToMBB(*curli_, Exit))
170 // Does this exit block have a predecessor that is not a loop block or loop
172 for (MachineBasicBlock::const_pred_iterator PI = Exit->pred_begin(),
173 PE = Exit->pred_end(); PI != PE; ++PI) {
174 const MachineBasicBlock *Pred = *PI;
175 if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
177 // This is a critical exit block, and we need to split the exit edge.
178 CriticalExits.insert(Exit);
184 /// canSplitCriticalExits - Return true if it is possible to insert new exit
185 /// blocks before the blocks in CriticalExits.
187 SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
188 BlockPtrSet &CriticalExits) {
189 // If we don't allow critical edge splitting, require no critical exits.
191 return CriticalExits.empty();
193 for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
195 const MachineBasicBlock *Succ = *I;
196 // We want to insert a new pre-exit MBB before Succ, and change all the
197 // in-loop blocks to branch to the pre-exit instead of Succ.
198 // Check that all the in-loop predecessors can be changed.
199 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
200 PE = Succ->pred_end(); PI != PE; ++PI) {
201 const MachineBasicBlock *Pred = *PI;
202 // The external predecessors won't be altered.
203 if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
205 if (!canAnalyzeBranch(Pred))
209 // If Succ's layout predecessor falls through, that too must be analyzable.
210 // We need to insert the pre-exit block in the gap.
211 MachineFunction::const_iterator MFI = Succ;
212 if (MFI == mf_.begin())
214 if (!canAnalyzeBranch(--MFI))
217 // No problems found.
221 void SplitAnalysis::analyze(const LiveInterval *li) {
227 const MachineLoop *SplitAnalysis::getBestSplitLoop() {
228 assert(curli_ && "Call analyze() before getBestSplitLoop");
229 if (usingLoops_.empty())
234 BlockPtrSet CriticalExits;
236 // We split around loops where curli is used outside the periphery.
237 for (LoopCountMap::const_iterator I = usingLoops_.begin(),
238 E = usingLoops_.end(); I != E; ++I) {
239 const MachineLoop *Loop = I->first;
240 getLoopBlocks(Loop, Blocks);
241 DEBUG({ dbgs() << " "; print(Blocks, dbgs()); });
243 switch(analyzeLoopPeripheralUse(Blocks)) {
246 case MultiPeripheral:
247 // FIXME: We could split a live range with multiple uses in a peripheral
248 // block and still make progress. However, it is possible that splitting
249 // another live range will insert copies into a peripheral block, and
250 // there is a small chance we can enter an infinity loop, inserting copies
252 // For safety, stick to splitting live ranges with uses outside the
254 DEBUG(dbgs() << ": multiple peripheral uses\n");
256 case ContainedInLoop:
257 DEBUG(dbgs() << ": fully contained\n");
259 case SinglePeripheral:
260 DEBUG(dbgs() << ": single peripheral use\n");
263 // Will it be possible to split around this loop?
264 getCriticalExits(Blocks, CriticalExits);
265 DEBUG(dbgs() << ": " << CriticalExits.size() << " critical exits\n");
266 if (!canSplitCriticalExits(Blocks, CriticalExits))
268 // This is a possible split.
272 DEBUG(dbgs() << " getBestSplitLoop found " << Loops.size()
273 << " candidate loops.\n");
278 // Pick the earliest loop.
279 // FIXME: Are there other heuristics to consider?
280 const MachineLoop *Best = 0;
282 for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
284 SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
285 if (!Best || Idx < BestIdx)
286 Best = *I, BestIdx = Idx;
288 DEBUG(dbgs() << " getBestSplitLoop found " << *Best);
292 //===----------------------------------------------------------------------===//
294 //===----------------------------------------------------------------------===//
296 // Work around the fact that the std::pair constructors are broken for pointer
297 // pairs in some implementations. makeVV(x, 0) works.
298 static inline std::pair<const VNInfo*, VNInfo*>
299 makeVV(const VNInfo *a, VNInfo *b) {
300 return std::make_pair(a, b);
303 void LiveIntervalMap::reset(LiveInterval *li) {
308 bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
309 ValueMap::const_iterator i = valueMap_.find(ParentVNI);
310 return i != valueMap_.end() && i->second == 0;
313 // defValue - Introduce a li_ def for ParentVNI that could be later than
315 VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
316 assert(li_ && "call reset first");
317 assert(ParentVNI && "Mapping NULL value");
318 assert(Idx.isValid() && "Invalid SlotIndex");
319 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
321 // Create a new value.
322 VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
324 // Use insert for lookup, so we can add missing values with a second lookup.
325 std::pair<ValueMap::iterator,bool> InsP =
326 valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
328 // This is now a complex def. Mark with a NULL in valueMap.
330 InsP.first->second = 0;
336 // mapValue - Find the mapped value for ParentVNI at Idx.
337 // Potentially create phi-def values.
338 VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
340 assert(li_ && "call reset first");
341 assert(ParentVNI && "Mapping NULL value");
342 assert(Idx.isValid() && "Invalid SlotIndex");
343 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
345 // Use insert for lookup, so we can add missing values with a second lookup.
346 std::pair<ValueMap::iterator,bool> InsP =
347 valueMap_.insert(makeVV(ParentVNI, 0));
349 // This was an unknown value. Create a simple mapping.
351 if (simple) *simple = true;
352 return InsP.first->second = li_->createValueCopy(ParentVNI,
353 lis_.getVNInfoAllocator());
356 // This was a simple mapped value.
357 if (InsP.first->second) {
358 if (simple) *simple = true;
359 return InsP.first->second;
362 // This is a complex mapped value. There may be multiple defs, and we may need
363 // to create phi-defs.
364 if (simple) *simple = false;
365 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
366 assert(IdxMBB && "No MBB at Idx");
368 // Is there a def in the same MBB we can extend?
369 if (VNInfo *VNI = extendTo(IdxMBB, Idx))
372 // Now for the fun part. We know that ParentVNI potentially has multiple defs,
373 // and we may need to create even more phi-defs to preserve VNInfo SSA form.
374 // Perform a depth-first search for predecessor blocks where we know the
375 // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
377 // Track MBBs where we have created or learned the dominating value.
378 // This may change during the DFS as we create new phi-defs.
379 typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
380 MBBValueMap DomValue;
381 typedef SplitAnalysis::BlockPtrSet BlockPtrSet;
384 // Iterate over IdxMBB predecessors in a depth-first order.
385 // Skip begin() since that is always IdxMBB.
386 for (idf_ext_iterator<MachineBasicBlock*, BlockPtrSet>
387 IDFI = llvm::next(idf_ext_begin(IdxMBB, Visited)),
388 IDFE = idf_ext_end(IdxMBB, Visited); IDFI != IDFE;) {
389 MachineBasicBlock *MBB = *IDFI;
390 SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
392 // We are operating on the restricted CFG where ParentVNI is live.
393 if (parentli_.getVNInfoAt(End) != ParentVNI) {
398 // Do we have a dominating value in this block?
399 VNInfo *VNI = extendTo(MBB, End);
405 // Yes, VNI dominates MBB. Make sure we visit MBB again from other paths.
408 // Track the path back to IdxMBB, creating phi-defs
409 // as needed along the way.
410 for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
411 // Start from MBB's immediate successor. End at IdxMBB.
412 MachineBasicBlock *Succ = IDFI.getPath(PI-1);
413 std::pair<MBBValueMap::iterator, bool> InsP =
414 DomValue.insert(MBBValueMap::value_type(Succ, VNI));
416 // This is the first time we backtrack to Succ.
420 // We reached Succ again with the same VNI. Nothing is going to change.
421 VNInfo *OVNI = InsP.first->second;
425 // Succ already has a phi-def. No need to continue.
426 SlotIndex Start = lis_.getMBBStartIdx(Succ);
427 if (OVNI->def == Start)
430 // We have a collision between the old and new VNI at Succ. That means
431 // neither dominates and we need a new phi-def.
432 VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
433 VNI->setIsPHIDef(true);
434 InsP.first->second = VNI;
436 // Replace OVNI with VNI in the remaining path.
437 for (; PI > 1 ; --PI) {
438 MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
439 if (I == DomValue.end() || I->second != OVNI)
445 // No need to search the children, we found a dominating value.
449 // The search should at least find a dominating value for IdxMBB.
450 assert(!DomValue.empty() && "Couldn't find a reaching definition");
452 // Since we went through the trouble of a full DFS visiting all reaching defs,
453 // the values in DomValue are now accurate. No more phi-defs are needed for
454 // these blocks, so we can color the live ranges.
455 // This makes the next mapValue call much faster.
457 for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
459 MachineBasicBlock *MBB = I->first;
460 VNInfo *VNI = I->second;
461 SlotIndex Start = lis_.getMBBStartIdx(MBB);
463 // Don't add full liveness to IdxMBB, stop at Idx.
465 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
466 // The caller had better add some liveness to IdxVNI, or it leaks.
469 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
472 assert(IdxVNI && "Didn't find value for Idx");
476 // extendTo - Find the last li_ value defined in MBB at or before Idx. The
477 // parentli_ is assumed to be live at Idx. Extend the live range to Idx.
478 // Return the found VNInfo, or NULL.
479 VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
480 assert(li_ && "call reset first");
481 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
482 if (I == li_->begin())
485 if (I->end <= lis_.getMBBStartIdx(MBB))
488 I->end = Idx.getNextSlot();
492 // addSimpleRange - Add a simple range from parentli_ to li_.
493 // ParentVNI must be live in the [Start;End) interval.
494 void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
495 const VNInfo *ParentVNI) {
496 assert(li_ && "call reset first");
498 VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
499 // A simple mapping is easy.
501 li_->addRange(LiveRange(Start, End, VNI));
505 // ParentVNI is a complex value. We must map per MBB.
506 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
507 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End.getPrevSlot());
510 li_->addRange(LiveRange(Start, End, VNI));
515 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
517 // Run sequence of full blocks.
518 for (++MBB; MBB != MBBE; ++MBB) {
519 Start = lis_.getMBBStartIdx(MBB);
520 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
521 mapValue(ParentVNI, Start)));
525 Start = lis_.getMBBStartIdx(MBB);
527 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
530 /// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
531 /// All needed values whose def is not inside [Start;End) must be defined
532 /// beforehand so mapValue will work.
533 void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
534 assert(li_ && "call reset first");
535 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
536 LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
538 // Check if --I begins before Start and overlaps.
542 addSimpleRange(Start, std::min(End, I->end), I->valno);
546 // The remaining ranges begin after Start.
547 for (;I != E && I->start < End; ++I)
548 addSimpleRange(I->start, std::min(End, I->end), I->valno);
551 VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
552 const VNInfo *ParentVNI,
553 MachineBasicBlock &MBB,
554 MachineBasicBlock::iterator I) {
555 const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
556 get(TargetOpcode::COPY);
557 MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
558 SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
559 VNInfo *VNI = defValue(ParentVNI, DefIdx);
561 li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
565 //===----------------------------------------------------------------------===//
567 //===----------------------------------------------------------------------===//
569 /// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
570 SplitEditor::SplitEditor(SplitAnalysis &sa, LiveIntervals &lis, VirtRegMap &vrm,
572 : sa_(sa), lis_(lis), vrm_(vrm),
573 mri_(vrm.getMachineFunction().getRegInfo()),
574 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
576 dupli_(lis_, edit.getParent()),
577 openli_(lis_, edit.getParent())
581 bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
582 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I)
583 if (*I != dupli_.getLI() && (*I)->liveAt(Idx))
588 /// Create a new virtual register and live interval.
589 void SplitEditor::openIntv() {
590 assert(!openli_.getLI() && "Previous LI not closed before openIntv");
593 dupli_.reset(&edit_.create(mri_, lis_, vrm_));
595 openli_.reset(&edit_.create(mri_, lis_, vrm_));
598 /// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
599 /// not live before Idx, a COPY is not inserted.
600 void SplitEditor::enterIntvBefore(SlotIndex Idx) {
601 assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
602 DEBUG(dbgs() << " enterIntvBefore " << Idx);
603 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getUseIndex());
605 DEBUG(dbgs() << ": not live\n");
608 DEBUG(dbgs() << ": valno " << ParentVNI->id);
609 truncatedValues.insert(ParentVNI);
610 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
611 assert(MI && "enterIntvBefore called with invalid index");
612 VNInfo *VNI = openli_.defByCopyFrom(edit_.getReg(), ParentVNI,
613 *MI->getParent(), MI);
614 openli_.getLI()->addRange(LiveRange(VNI->def, Idx.getDefIndex(), VNI));
615 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
618 /// enterIntvAtEnd - Enter openli at the end of MBB.
619 void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
620 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
621 SlotIndex End = lis_.getMBBEndIdx(&MBB);
622 DEBUG(dbgs() << " enterIntvAtEnd BB#" << MBB.getNumber() << ", " << End);
623 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(End.getPrevSlot());
625 DEBUG(dbgs() << ": not live\n");
628 DEBUG(dbgs() << ": valno " << ParentVNI->id);
629 truncatedValues.insert(ParentVNI);
630 VNInfo *VNI = openli_.defByCopyFrom(edit_.getReg(), ParentVNI,
631 MBB, MBB.getFirstTerminator());
632 // Make sure openli is live out of MBB.
633 openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
634 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
637 /// useIntv - indicate that all instructions in MBB should use openli.
638 void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
639 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
642 void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
643 assert(openli_.getLI() && "openIntv not called before useIntv");
644 openli_.addRange(Start, End);
645 DEBUG(dbgs() << " use [" << Start << ';' << End << "): "
646 << *openli_.getLI() << '\n');
649 /// leaveIntvAfter - Leave openli after the instruction at Idx.
650 void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
651 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
652 DEBUG(dbgs() << " leaveIntvAfter " << Idx);
654 // The interval must be live beyond the instruction at Idx.
655 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Idx.getBoundaryIndex());
657 DEBUG(dbgs() << ": not live\n");
660 DEBUG(dbgs() << ": valno " << ParentVNI->id);
662 MachineBasicBlock::iterator MII = lis_.getInstructionFromIndex(Idx);
663 MachineBasicBlock *MBB = MII->getParent();
664 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI, *MBB,
667 // Finally we must make sure that openli is properly extended from Idx to the
669 openli_.addSimpleRange(Idx.getBoundaryIndex(), VNI->def, ParentVNI);
670 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
673 /// leaveIntvAtTop - Leave the interval at the top of MBB.
674 /// Currently, only one value can leave the interval.
675 void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
676 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
677 SlotIndex Start = lis_.getMBBStartIdx(&MBB);
678 DEBUG(dbgs() << " leaveIntvAtTop BB#" << MBB.getNumber() << ", " << Start);
680 VNInfo *ParentVNI = edit_.getParent().getVNInfoAt(Start);
682 DEBUG(dbgs() << ": not live\n");
686 // We are going to insert a back copy, so we must have a dupli_.
687 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
690 // Finally we must make sure that openli is properly extended from Start to
692 openli_.addSimpleRange(Start, VNI->def, ParentVNI);
693 DEBUG(dbgs() << ": " << *openli_.getLI() << '\n');
696 /// closeIntv - Indicate that we are done editing the currently open
697 /// LiveInterval, and ranges can be trimmed.
698 void SplitEditor::closeIntv() {
699 assert(openli_.getLI() && "openIntv not called before closeIntv");
701 DEBUG(dbgs() << " closeIntv cleaning up\n");
702 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n');
706 /// rewrite - Rewrite all uses of reg to use the new registers.
707 void SplitEditor::rewrite(unsigned reg) {
708 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(reg),
709 RE = mri_.reg_end(); RI != RE;) {
710 MachineOperand &MO = RI.getOperand();
711 MachineInstr *MI = MO.getParent();
713 if (MI->isDebugValue()) {
714 DEBUG(dbgs() << "Zapping " << *MI);
715 // FIXME: We can do much better with debug values.
719 SlotIndex Idx = lis_.getInstructionIndex(MI);
720 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
721 LiveInterval *LI = 0;
722 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E;
724 LiveInterval *testli = *I;
725 if (testli->liveAt(Idx)) {
730 DEBUG(dbgs() << " rewr BB#" << MI->getParent()->getNumber() << '\t'<< Idx);
731 assert(LI && "No register was live at use");
733 DEBUG(dbgs() << '\t' << *MI);
738 SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
739 // Build vector of iterator pairs from the intervals.
740 typedef std::pair<LiveInterval::const_iterator,
741 LiveInterval::const_iterator> IIPair;
742 SmallVector<IIPair, 8> Iters;
743 for (LiveRangeEdit::iterator LI = edit_.begin(), LE = edit_.end(); LI != LE;
745 if (*LI == dupli_.getLI())
747 LiveInterval::const_iterator I = (*LI)->find(Start);
748 LiveInterval::const_iterator E = (*LI)->end();
750 Iters.push_back(std::make_pair(I, E));
753 SlotIndex sidx = Start;
754 // Break [Start;End) into segments that don't overlap any intervals.
756 SlotIndex next = sidx, eidx = End;
757 // Find overlapping intervals.
758 for (unsigned i = 0; i != Iters.size() && sidx < eidx; ++i) {
759 LiveInterval::const_iterator I = Iters[i].first;
760 // Interval I is overlapping [sidx;eidx). Trim sidx.
761 if (I->start <= sidx) {
763 // Move to the next run, remove iters when all are consumed.
764 I = ++Iters[i].first;
765 if (I == Iters[i].second) {
766 Iters.erase(Iters.begin() + i);
771 // Trim eidx too if needed.
772 if (I->start >= eidx)
777 // Now, [sidx;eidx) doesn't overlap anything in intervals_.
779 dupli_.addSimpleRange(sidx, eidx, VNI);
780 // If the interval end was truncated, we can try again from next.
787 void SplitEditor::computeRemainder() {
788 // First we need to fill in the live ranges in dupli.
789 // If values were redefined, we need a full recoloring with SSA update.
790 // If values were truncated, we only need to truncate the ranges.
791 // If values were partially rematted, we should shrink to uses.
792 // If values were fully rematted, they should be omitted.
793 // FIXME: If a single value is redefined, just move the def and truncate.
794 LiveInterval &parent = edit_.getParent();
796 // Values that are fully contained in the split intervals.
797 SmallPtrSet<const VNInfo*, 8> deadValues;
798 // Map all curli values that should have live defs in dupli.
799 for (LiveInterval::const_vni_iterator I = parent.vni_begin(),
800 E = parent.vni_end(); I != E; ++I) {
801 const VNInfo *VNI = *I;
802 // Original def is contained in the split intervals.
803 if (intervalsLiveAt(VNI->def)) {
804 // Did this value escape?
805 if (dupli_.isMapped(VNI))
806 truncatedValues.insert(VNI);
808 deadValues.insert(VNI);
811 // Add minimal live range at the definition.
812 VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
813 dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
816 // Add all ranges to dupli.
817 for (LiveInterval::const_iterator I = parent.begin(), E = parent.end();
819 const LiveRange &LR = *I;
820 if (truncatedValues.count(LR.valno)) {
821 // recolor after removing intervals_.
822 addTruncSimpleRange(LR.start, LR.end, LR.valno);
823 } else if (!deadValues.count(LR.valno)) {
824 // recolor without truncation.
825 dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
830 void SplitEditor::finish() {
831 assert(!openli_.getLI() && "Previous LI not closed before rewrite");
832 assert(dupli_.getLI() && "No dupli for rewrite. Noop spilt?");
834 // Complete dupli liveness.
837 // Get rid of unused values and set phi-kill flags.
838 dupli_.getLI()->RenumberValues(lis_);
840 // Now check if dupli was separated into multiple connected components.
841 ConnectedVNInfoEqClasses ConEQ(lis_);
842 if (unsigned NumComp = ConEQ.Classify(dupli_.getLI())) {
843 DEBUG(dbgs() << " Remainder has " << NumComp << " connected components: "
844 << *dupli_.getLI() << '\n');
845 // Did the remainder break up? Create intervals for all the components.
847 SmallVector<LiveInterval*, 8> dups;
848 dups.push_back(dupli_.getLI());
849 for (unsigned i = 1; i != NumComp; ++i)
850 dups.push_back(&edit_.create(mri_, lis_, vrm_));
851 ConEQ.Distribute(&dups[0]);
852 // Rewrite uses to the new regs.
853 rewrite(dupli_.getLI()->reg);
857 // Rewrite instructions.
858 rewrite(edit_.getReg());
860 // Calculate spill weight and allocation hints for new intervals.
861 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
862 for (LiveRangeEdit::iterator I = edit_.begin(), E = edit_.end(); I != E; ++I){
863 LiveInterval &li = **I;
864 vrai.CalculateRegClass(li.reg);
865 vrai.CalculateWeightAndHint(li);
866 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
867 << ":" << li << '\n');
872 //===----------------------------------------------------------------------===//
874 //===----------------------------------------------------------------------===//
876 void SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
877 SplitAnalysis::LoopBlocks Blocks;
878 sa_.getLoopBlocks(Loop, Blocks);
881 dbgs() << " splitAround"; sa_.print(Blocks, dbgs()); dbgs() << '\n';
884 // Break critical edges as needed.
885 SplitAnalysis::BlockPtrSet CriticalExits;
886 sa_.getCriticalExits(Blocks, CriticalExits);
887 assert(CriticalExits.empty() && "Cannot break critical exits yet");
889 // Create new live interval for the loop.
892 // Insert copies in the predecessors.
893 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
894 E = Blocks.Preds.end(); I != E; ++I) {
895 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
899 // Switch all loop blocks.
900 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
901 E = Blocks.Loop.end(); I != E; ++I)
904 // Insert back copies in the exit blocks.
905 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
906 E = Blocks.Exits.end(); I != E; ++I) {
907 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
917 //===----------------------------------------------------------------------===//
918 // Single Block Splitting
919 //===----------------------------------------------------------------------===//
921 /// getMultiUseBlocks - if curli has more than one use in a basic block, it
922 /// may be an advantage to split curli for the duration of the block.
923 bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
924 // If curli is local to one block, there is no point to splitting it.
925 if (usingBlocks_.size() <= 1)
927 // Add blocks with multiple uses.
928 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
935 // When there are only two uses and curli is both live in and live out,
936 // we don't really win anything by isolating the block since we would be
937 // inserting two copies.
938 // The remaing register would still have two uses in the block. (Unless it
939 // separates into disconnected components).
940 if (lis_.isLiveInToMBB(*curli_, I->first) &&
941 lis_.isLiveOutOfMBB(*curli_, I->first))
945 Blocks.insert(I->first);
947 return !Blocks.empty();
950 /// splitSingleBlocks - Split curli into a separate live interval inside each
951 /// basic block in Blocks.
952 void SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
953 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
954 // Determine the first and last instruction using curli in each block.
955 typedef std::pair<SlotIndex,SlotIndex> IndexPair;
956 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
957 IndexPairMap MBBRange;
958 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
959 E = sa_.usingInstrs_.end(); I != E; ++I) {
960 const MachineBasicBlock *MBB = (*I)->getParent();
961 if (!Blocks.count(MBB))
963 SlotIndex Idx = lis_.getInstructionIndex(*I);
964 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
965 IndexPair &IP = MBBRange[MBB];
966 if (!IP.first.isValid() || Idx < IP.first)
968 if (!IP.second.isValid() || Idx > IP.second)
972 // Create a new interval for each block.
973 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
974 E = Blocks.end(); I != E; ++I) {
975 IndexPair &IP = MBBRange[*I];
976 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": ["
977 << IP.first << ';' << IP.second << ")\n");
978 assert(IP.first.isValid() && IP.second.isValid());
981 enterIntvBefore(IP.first);
982 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
983 leaveIntvAfter(IP.second);
990 //===----------------------------------------------------------------------===//
991 // Sub Block Splitting
992 //===----------------------------------------------------------------------===//
994 /// getBlockForInsideSplit - If curli is contained inside a single basic block,
995 /// and it wou pay to subdivide the interval inside that block, return it.
996 /// Otherwise return NULL. The returned block can be passed to
997 /// SplitEditor::splitInsideBlock.
998 const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
999 // The interval must be exclusive to one block.
1000 if (usingBlocks_.size() != 1)
1002 // Don't to this for less than 4 instructions. We want to be sure that
1003 // splitting actually reduces the instruction count per interval.
1004 if (usingInstrs_.size() < 4)
1006 return usingBlocks_.begin()->first;
1009 /// splitInsideBlock - Split curli into multiple intervals inside MBB.
1010 void SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1011 SmallVector<SlotIndex, 32> Uses;
1012 Uses.reserve(sa_.usingInstrs_.size());
1013 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1014 E = sa_.usingInstrs_.end(); I != E; ++I)
1015 if ((*I)->getParent() == MBB)
1016 Uses.push_back(lis_.getInstructionIndex(*I));
1017 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for "
1018 << Uses.size() << " instructions.\n");
1019 assert(Uses.size() >= 3 && "Need at least 3 instructions");
1020 array_pod_sort(Uses.begin(), Uses.end());
1022 // Simple algorithm: Find the largest gap between uses as determined by slot
1023 // indices. Create new intervals for instructions before the gap and after the
1025 unsigned bestPos = 0;
1027 DEBUG(dbgs() << " dist (" << Uses[0]);
1028 for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1029 int g = Uses[i-1].distance(Uses[i]);
1030 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1032 bestPos = i, bestGap = g;
1034 DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1036 // bestPos points to the first use after the best gap.
1037 assert(bestPos > 0 && "Invalid gap");
1039 // FIXME: Don't create intervals for low densities.
1041 // First interval before the gap. Don't create single-instr intervals.
1044 enterIntvBefore(Uses.front());
1045 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1046 leaveIntvAfter(Uses[bestPos-1]);
1050 // Second interval after the gap.
1051 if (bestPos < Uses.size()-1) {
1053 enterIntvBefore(Uses[bestPos]);
1054 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1055 leaveIntvAfter(Uses.back());