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 "VirtRegMap.h"
18 #include "llvm/CodeGen/CalcSpillWeights.h"
19 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineLoopInfo.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Target/TargetInstrInfo.h"
27 #include "llvm/Target/TargetMachine.h"
32 AllowSplit("spiller-splits-edges",
33 cl::desc("Allow critical edge splitting during spilling"));
35 //===----------------------------------------------------------------------===//
37 //===----------------------------------------------------------------------===//
39 SplitAnalysis::SplitAnalysis(const MachineFunction &mf,
40 const LiveIntervals &lis,
41 const MachineLoopInfo &mli)
45 tii_(*mf.getTarget().getInstrInfo()),
48 void SplitAnalysis::clear() {
55 bool SplitAnalysis::canAnalyzeBranch(const MachineBasicBlock *MBB) {
56 MachineBasicBlock *T, *F;
57 SmallVector<MachineOperand, 4> Cond;
58 return !tii_.AnalyzeBranch(const_cast<MachineBasicBlock&>(*MBB), T, F, Cond);
61 /// analyzeUses - Count instructions, basic blocks, and loops using curli.
62 void SplitAnalysis::analyzeUses() {
63 const MachineRegisterInfo &MRI = mf_.getRegInfo();
64 for (MachineRegisterInfo::reg_iterator I = MRI.reg_begin(curli_->reg);
65 MachineInstr *MI = I.skipInstruction();) {
66 if (MI->isDebugValue() || !usingInstrs_.insert(MI))
68 MachineBasicBlock *MBB = MI->getParent();
69 if (usingBlocks_[MBB]++)
71 if (MachineLoop *Loop = loops_.getLoopFor(MBB))
74 DEBUG(dbgs() << " counted "
75 << usingInstrs_.size() << " instrs, "
76 << usingBlocks_.size() << " blocks, "
77 << usingLoops_.size() << " loops.\n");
80 /// removeUse - Update statistics by noting that MI no longer uses curli.
81 void SplitAnalysis::removeUse(const MachineInstr *MI) {
82 if (!usingInstrs_.erase(MI))
85 // Decrement MBB count.
86 const MachineBasicBlock *MBB = MI->getParent();
87 BlockCountMap::iterator bi = usingBlocks_.find(MBB);
88 assert(bi != usingBlocks_.end() && "MBB missing");
89 assert(bi->second && "0 count in map");
92 // No more uses in MBB.
93 usingBlocks_.erase(bi);
95 // Decrement loop count.
96 MachineLoop *Loop = loops_.getLoopFor(MBB);
99 LoopCountMap::iterator li = usingLoops_.find(Loop);
100 assert(li != usingLoops_.end() && "Loop missing");
101 assert(li->second && "0 count in map");
104 // No more blocks in Loop.
105 usingLoops_.erase(li);
108 // Get three sets of basic blocks surrounding a loop: Blocks inside the loop,
109 // predecessor blocks, and exit blocks.
110 void SplitAnalysis::getLoopBlocks(const MachineLoop *Loop, LoopBlocks &Blocks) {
113 // Blocks in the loop.
114 Blocks.Loop.insert(Loop->block_begin(), Loop->block_end());
116 // Predecessor blocks.
117 const MachineBasicBlock *Header = Loop->getHeader();
118 for (MachineBasicBlock::const_pred_iterator I = Header->pred_begin(),
119 E = Header->pred_end(); I != E; ++I)
120 if (!Blocks.Loop.count(*I))
121 Blocks.Preds.insert(*I);
124 for (MachineLoop::block_iterator I = Loop->block_begin(),
125 E = Loop->block_end(); I != E; ++I) {
126 const MachineBasicBlock *MBB = *I;
127 for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(),
128 SE = MBB->succ_end(); SI != SE; ++SI)
129 if (!Blocks.Loop.count(*SI))
130 Blocks.Exits.insert(*SI);
134 /// analyzeLoopPeripheralUse - Return an enum describing how curli_ is used in
135 /// and around the Loop.
136 SplitAnalysis::LoopPeripheralUse SplitAnalysis::
137 analyzeLoopPeripheralUse(const SplitAnalysis::LoopBlocks &Blocks) {
138 LoopPeripheralUse use = ContainedInLoop;
139 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
141 const MachineBasicBlock *MBB = I->first;
142 // Is this a peripheral block?
143 if (use < MultiPeripheral &&
144 (Blocks.Preds.count(MBB) || Blocks.Exits.count(MBB))) {
145 if (I->second > 1) use = MultiPeripheral;
146 else use = SinglePeripheral;
149 // Is it a loop block?
150 if (Blocks.Loop.count(MBB))
152 // It must be an unrelated block.
158 /// getCriticalExits - It may be necessary to partially break critical edges
159 /// leaving the loop if an exit block has phi uses of curli. Collect the exit
160 /// blocks that need special treatment into CriticalExits.
161 void SplitAnalysis::getCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
162 BlockPtrSet &CriticalExits) {
163 CriticalExits.clear();
165 // A critical exit block contains a phi def of curli, and has a predecessor
166 // that is not in the loop nor a loop predecessor.
167 // For such an exit block, the edges carrying the new variable must be moved
168 // to a new pre-exit block.
169 for (BlockPtrSet::iterator I = Blocks.Exits.begin(), E = Blocks.Exits.end();
171 const MachineBasicBlock *Succ = *I;
172 SlotIndex SuccIdx = lis_.getMBBStartIdx(Succ);
173 VNInfo *SuccVNI = curli_->getVNInfoAt(SuccIdx);
174 // This exit may not have curli live in at all. No need to split.
177 // If this is not a PHI def, it is either using a value from before the
178 // loop, or a value defined inside the loop. Both are safe.
179 if (!SuccVNI->isPHIDef() || SuccVNI->def.getBaseIndex() != SuccIdx)
181 // This exit block does have a PHI. Does it also have a predecessor that is
182 // not a loop block or loop predecessor?
183 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
184 PE = Succ->pred_end(); PI != PE; ++PI) {
185 const MachineBasicBlock *Pred = *PI;
186 if (Blocks.Loop.count(Pred) || Blocks.Preds.count(Pred))
188 // This is a critical exit block, and we need to split the exit edge.
189 CriticalExits.insert(Succ);
195 /// canSplitCriticalExits - Return true if it is possible to insert new exit
196 /// blocks before the blocks in CriticalExits.
198 SplitAnalysis::canSplitCriticalExits(const SplitAnalysis::LoopBlocks &Blocks,
199 BlockPtrSet &CriticalExits) {
200 // If we don't allow critical edge splitting, require no critical exits.
202 return CriticalExits.empty();
204 for (BlockPtrSet::iterator I = CriticalExits.begin(), E = CriticalExits.end();
206 const MachineBasicBlock *Succ = *I;
207 // We want to insert a new pre-exit MBB before Succ, and change all the
208 // in-loop blocks to branch to the pre-exit instead of Succ.
209 // Check that all the in-loop predecessors can be changed.
210 for (MachineBasicBlock::const_pred_iterator PI = Succ->pred_begin(),
211 PE = Succ->pred_end(); PI != PE; ++PI) {
212 const MachineBasicBlock *Pred = *PI;
213 // The external predecessors won't be altered.
214 if (!Blocks.Loop.count(Pred) && !Blocks.Preds.count(Pred))
216 if (!canAnalyzeBranch(Pred))
220 // If Succ's layout predecessor falls through, that too must be analyzable.
221 // We need to insert the pre-exit block in the gap.
222 MachineFunction::const_iterator MFI = Succ;
223 if (MFI == mf_.begin())
225 if (!canAnalyzeBranch(--MFI))
228 // No problems found.
232 void SplitAnalysis::analyze(const LiveInterval *li) {
238 const MachineLoop *SplitAnalysis::getBestSplitLoop() {
239 assert(curli_ && "Call analyze() before getBestSplitLoop");
240 if (usingLoops_.empty())
243 LoopPtrSet Loops, SecondLoops;
245 BlockPtrSet CriticalExits;
247 // Find first-class and second class candidate loops.
248 // We prefer to split around loops where curli is used outside the periphery.
249 for (LoopCountMap::const_iterator I = usingLoops_.begin(),
250 E = usingLoops_.end(); I != E; ++I) {
251 const MachineLoop *Loop = I->first;
252 getLoopBlocks(Loop, Blocks);
254 // FIXME: We need an SSA updater to properly handle multiple exit blocks.
255 if (Blocks.Exits.size() > 1) {
256 DEBUG(dbgs() << " multiple exits from " << *Loop);
261 switch(analyzeLoopPeripheralUse(Blocks)) {
265 case MultiPeripheral:
268 case ContainedInLoop:
269 DEBUG(dbgs() << " contained in " << *Loop);
271 case SinglePeripheral:
272 DEBUG(dbgs() << " single peripheral use in " << *Loop);
275 // Will it be possible to split around this loop?
276 getCriticalExits(Blocks, CriticalExits);
277 DEBUG(dbgs() << " " << CriticalExits.size() << " critical exits from "
279 if (!canSplitCriticalExits(Blocks, CriticalExits))
281 // This is a possible split.
286 DEBUG(dbgs() << " getBestSplitLoop found " << Loops.size() << " + "
287 << SecondLoops.size() << " candidate loops.\n");
289 // If there are no first class loops available, look at second class loops.
296 // Pick the earliest loop.
297 // FIXME: Are there other heuristics to consider?
298 const MachineLoop *Best = 0;
300 for (LoopPtrSet::const_iterator I = Loops.begin(), E = Loops.end(); I != E;
302 SlotIndex Idx = lis_.getMBBStartIdx((*I)->getHeader());
303 if (!Best || Idx < BestIdx)
304 Best = *I, BestIdx = Idx;
306 DEBUG(dbgs() << " getBestSplitLoop found " << *Best);
310 /// getMultiUseBlocks - if curli has more than one use in a basic block, it
311 /// may be an advantage to split curli for the duration of the block.
312 bool SplitAnalysis::getMultiUseBlocks(BlockPtrSet &Blocks) {
313 // If curli is local to one block, there is no point to splitting it.
314 if (usingBlocks_.size() <= 1)
316 // Add blocks with multiple uses.
317 for (BlockCountMap::iterator I = usingBlocks_.begin(), E = usingBlocks_.end();
324 // It doesn't pay to split a 2-instr block if it redefines curli.
325 VNInfo *VN1 = curli_->getVNInfoAt(lis_.getMBBStartIdx(I->first));
327 curli_->getVNInfoAt(lis_.getMBBEndIdx(I->first).getPrevIndex());
328 // live-in and live-out with a different value.
329 if (VN1 && VN2 && VN1 != VN2)
333 Blocks.insert(I->first);
335 return !Blocks.empty();
338 //===----------------------------------------------------------------------===//
340 //===----------------------------------------------------------------------===//
342 // Work around the fact that the std::pair constructors are broken for pointer
343 // pairs in some implementations. makeVV(x, 0) works.
344 static inline std::pair<const VNInfo*, VNInfo*>
345 makeVV(const VNInfo *a, VNInfo *b) {
346 return std::make_pair(a, b);
349 void LiveIntervalMap::reset(LiveInterval *li) {
354 bool LiveIntervalMap::isComplexMapped(const VNInfo *ParentVNI) const {
355 ValueMap::const_iterator i = valueMap_.find(ParentVNI);
356 return i != valueMap_.end() && i->second == 0;
359 // defValue - Introduce a li_ def for ParentVNI that could be later than
361 VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
362 assert(li_ && "call reset first");
363 assert(ParentVNI && "Mapping NULL value");
364 assert(Idx.isValid() && "Invalid SlotIndex");
365 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
367 // Create a new value.
368 VNInfo *VNI = li_->getNextValue(Idx, 0, lis_.getVNInfoAllocator());
370 // Use insert for lookup, so we can add missing values with a second lookup.
371 std::pair<ValueMap::iterator,bool> InsP =
372 valueMap_.insert(makeVV(ParentVNI, Idx == ParentVNI->def ? VNI : 0));
374 // This is now a complex def. Mark with a NULL in valueMap.
376 InsP.first->second = 0;
382 // mapValue - Find the mapped value for ParentVNI at Idx.
383 // Potentially create phi-def values.
384 VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx,
386 assert(li_ && "call reset first");
387 assert(ParentVNI && "Mapping NULL value");
388 assert(Idx.isValid() && "Invalid SlotIndex");
389 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
391 // Use insert for lookup, so we can add missing values with a second lookup.
392 std::pair<ValueMap::iterator,bool> InsP =
393 valueMap_.insert(makeVV(ParentVNI, 0));
395 // This was an unknown value. Create a simple mapping.
397 if (simple) *simple = true;
398 return InsP.first->second = li_->createValueCopy(ParentVNI,
399 lis_.getVNInfoAllocator());
402 // This was a simple mapped value.
403 if (InsP.first->second) {
404 if (simple) *simple = true;
405 return InsP.first->second;
408 // This is a complex mapped value. There may be multiple defs, and we may need
409 // to create phi-defs.
410 if (simple) *simple = false;
411 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
412 assert(IdxMBB && "No MBB at Idx");
414 // Is there a def in the same MBB we can extend?
415 if (VNInfo *VNI = extendTo(IdxMBB, Idx))
418 // Now for the fun part. We know that ParentVNI potentially has multiple defs,
419 // and we may need to create even more phi-defs to preserve VNInfo SSA form.
420 // Perform a depth-first search for predecessor blocks where we know the
421 // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
423 // Track MBBs where we have created or learned the dominating value.
424 // This may change during the DFS as we create new phi-defs.
425 typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
426 MBBValueMap DomValue;
428 for (idf_iterator<MachineBasicBlock*>
429 IDFI = idf_begin(IdxMBB),
430 IDFE = idf_end(IdxMBB); IDFI != IDFE;) {
431 MachineBasicBlock *MBB = *IDFI;
432 SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
434 // We are operating on the restricted CFG where ParentVNI is live.
435 if (parentli_.getVNInfoAt(End) != ParentVNI) {
440 // Do we have a dominating value in this block?
441 VNInfo *VNI = extendTo(MBB, End);
447 // Yes, VNI dominates MBB. Track the path back to IdxMBB, creating phi-defs
448 // as needed along the way.
449 for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
450 // Start from MBB's immediate successor. End at IdxMBB.
451 MachineBasicBlock *Succ = IDFI.getPath(PI-1);
452 std::pair<MBBValueMap::iterator, bool> InsP =
453 DomValue.insert(MBBValueMap::value_type(Succ, VNI));
455 // This is the first time we backtrack to Succ.
459 // We reached Succ again with the same VNI. Nothing is going to change.
460 VNInfo *OVNI = InsP.first->second;
464 // Succ already has a phi-def. No need to continue.
465 SlotIndex Start = lis_.getMBBStartIdx(Succ);
466 if (OVNI->def == Start)
469 // We have a collision between the old and new VNI at Succ. That means
470 // neither dominates and we need a new phi-def.
471 VNI = li_->getNextValue(Start, 0, lis_.getVNInfoAllocator());
472 VNI->setIsPHIDef(true);
473 InsP.first->second = VNI;
475 // Replace OVNI with VNI in the remaining path.
476 for (; PI > 1 ; --PI) {
477 MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
478 if (I == DomValue.end() || I->second != OVNI)
484 // No need to search the children, we found a dominating value.
488 // The search should at least find a dominating value for IdxMBB.
489 assert(!DomValue.empty() && "Couldn't find a reaching definition");
491 // Since we went through the trouble of a full DFS visiting all reaching defs,
492 // the values in DomValue are now accurate. No more phi-defs are needed for
493 // these blocks, so we can color the live ranges.
494 // This makes the next mapValue call much faster.
496 for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
498 MachineBasicBlock *MBB = I->first;
499 VNInfo *VNI = I->second;
500 SlotIndex Start = lis_.getMBBStartIdx(MBB);
502 // Don't add full liveness to IdxMBB, stop at Idx.
504 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
505 // The caller had better add some liveness to IdxVNI, or it leaks.
508 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
511 assert(IdxVNI && "Didn't find value for Idx");
515 // extendTo - Find the last li_ value defined in MBB at or before Idx. The
516 // parentli_ is assumed to be live at Idx. Extend the live range to Idx.
517 // Return the found VNInfo, or NULL.
518 VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
519 assert(li_ && "call reset first");
520 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
521 if (I == li_->begin())
524 if (I->start < lis_.getMBBStartIdx(MBB))
527 I->end = Idx.getNextSlot();
531 // addSimpleRange - Add a simple range from parentli_ to li_.
532 // ParentVNI must be live in the [Start;End) interval.
533 void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
534 const VNInfo *ParentVNI) {
535 assert(li_ && "call reset first");
537 VNInfo *VNI = mapValue(ParentVNI, Start, &simple);
538 // A simple mapping is easy.
540 li_->addRange(LiveRange(Start, End, VNI));
544 // ParentVNI is a complex value. We must map per MBB.
545 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
546 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End);
549 li_->addRange(LiveRange(Start, End, VNI));
554 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
556 // Run sequence of full blocks.
557 for (++MBB; MBB != MBBE; ++MBB) {
558 Start = lis_.getMBBStartIdx(MBB);
559 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
560 mapValue(ParentVNI, Start)));
564 Start = lis_.getMBBStartIdx(MBB);
566 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
569 /// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
570 /// All needed values whose def is not inside [Start;End) must be defined
571 /// beforehand so mapValue will work.
572 void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
573 assert(li_ && "call reset first");
574 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
575 LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
577 // Check if --I begins before Start and overlaps.
581 addSimpleRange(Start, std::min(End, I->end), I->valno);
585 // The remaining ranges begin after Start.
586 for (;I != E && I->start < End; ++I)
587 addSimpleRange(I->start, std::min(End, I->end), I->valno);
590 VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
591 const VNInfo *ParentVNI,
592 MachineBasicBlock &MBB,
593 MachineBasicBlock::iterator I) {
594 const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
595 get(TargetOpcode::COPY);
596 MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
597 SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
598 VNInfo *VNI = defValue(ParentVNI, DefIdx);
600 li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
604 //===----------------------------------------------------------------------===//
606 //===----------------------------------------------------------------------===//
608 /// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
609 SplitEditor::SplitEditor(SplitAnalysis &sa, LiveIntervals &lis, VirtRegMap &vrm,
610 SmallVectorImpl<LiveInterval*> &intervals)
611 : sa_(sa), lis_(lis), vrm_(vrm),
612 mri_(vrm.getMachineFunction().getRegInfo()),
613 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
614 curli_(sa_.getCurLI()),
615 dupli_(lis_, *curli_),
616 openli_(lis_, *curli_),
617 intervals_(intervals),
618 firstInterval(intervals_.size())
620 assert(curli_ && "SplitEditor created from empty SplitAnalysis");
622 // Make sure curli_ is assigned a stack slot, so all our intervals get the
623 // same slot as curli_.
624 if (vrm_.getStackSlot(curli_->reg) == VirtRegMap::NO_STACK_SLOT)
625 vrm_.assignVirt2StackSlot(curli_->reg);
629 LiveInterval *SplitEditor::createInterval() {
630 unsigned Reg = mri_.createVirtualRegister(mri_.getRegClass(curli_->reg));
631 LiveInterval &Intv = lis_.getOrCreateInterval(Reg);
633 vrm_.assignVirt2StackSlot(Reg, vrm_.getStackSlot(curli_->reg));
637 bool SplitEditor::intervalsLiveAt(SlotIndex Idx) const {
638 for (int i = firstInterval, e = intervals_.size(); i != e; ++i)
639 if (intervals_[i]->liveAt(Idx))
644 /// Create a new virtual register and live interval.
645 void SplitEditor::openIntv() {
646 assert(!openli_.getLI() && "Previous LI not closed before openIntv");
649 dupli_.reset(createInterval());
651 openli_.reset(createInterval());
652 intervals_.push_back(openli_.getLI());
655 /// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
656 /// not live before Idx, a COPY is not inserted.
657 void SplitEditor::enterIntvBefore(SlotIndex Idx) {
658 assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
659 VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getUseIndex());
661 DEBUG(dbgs() << " enterIntvBefore " << Idx << ": not live\n");
664 truncatedValues.insert(ParentVNI);
665 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
666 assert(MI && "enterIntvBefore called with invalid index");
667 openli_.defByCopyFrom(curli_->reg, ParentVNI, *MI->getParent(), MI);
668 DEBUG(dbgs() << " enterIntvBefore " << Idx << ": " << *openli_.getLI()
672 /// enterIntvAtEnd - Enter openli at the end of MBB.
673 void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
674 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
675 SlotIndex End = lis_.getMBBEndIdx(&MBB);
676 VNInfo *ParentVNI = curli_->getVNInfoAt(End.getPrevSlot());
678 DEBUG(dbgs() << " enterIntvAtEnd " << End << ": not live\n");
681 truncatedValues.insert(ParentVNI);
682 VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
683 MBB, MBB.getFirstTerminator());
684 // Make sure openli is live out of MBB.
685 openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
686 DEBUG(dbgs() << " enterIntvAtEnd: " << *openli_.getLI() << '\n');
689 /// useIntv - indicate that all instructions in MBB should use openli.
690 void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
691 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
694 void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
695 assert(openli_.getLI() && "openIntv not called before useIntv");
696 openli_.addRange(Start, End);
697 DEBUG(dbgs() << " use [" << Start << ';' << End << "): "
698 << *openli_.getLI() << '\n');
701 /// leaveIntvAfter - Leave openli after the instruction at Idx.
702 void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
703 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
705 // The interval must be live beyond the instruction at Idx.
706 SlotIndex EndIdx = Idx.getNextIndex().getBaseIndex();
707 VNInfo *ParentVNI = curli_->getVNInfoAt(EndIdx);
709 DEBUG(dbgs() << " leaveIntvAfter " << Idx << ": not live\n");
713 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
714 assert(MI && "leaveIntvAfter called with invalid index");
716 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
717 *MI->getParent(), MI);
719 // Finally we must make sure that openli is properly extended from Idx to the
721 openli_.mapValue(ParentVNI, VNI->def.getUseIndex());
723 DEBUG(dbgs() << " leaveIntvAfter " << Idx << ": " << *openli_.getLI()
727 /// leaveIntvAtTop - Leave the interval at the top of MBB.
728 /// Currently, only one value can leave the interval.
729 void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
730 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
732 SlotIndex Start = lis_.getMBBStartIdx(&MBB);
733 VNInfo *ParentVNI = curli_->getVNInfoAt(Start);
735 // Is curli even live-in to MBB?
737 DEBUG(dbgs() << " leaveIntvAtTop at " << Start << ": not live\n");
741 // We are going to insert a back copy, so we must have a dupli_.
742 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
745 // Finally we must make sure that openli is properly extended from Start to
747 openli_.mapValue(ParentVNI, VNI->def.getUseIndex());
749 DEBUG(dbgs() << " leaveIntvAtTop at " << Start << ": " << *openli_.getLI()
753 /// closeIntv - Indicate that we are done editing the currently open
754 /// LiveInterval, and ranges can be trimmed.
755 void SplitEditor::closeIntv() {
756 assert(openli_.getLI() && "openIntv not called before closeIntv");
758 DEBUG(dbgs() << " closeIntv cleaning up\n");
759 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n');
764 SplitEditor::addTruncSimpleRange(SlotIndex Start, SlotIndex End, VNInfo *VNI) {
765 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 (int i = firstInterval, e = intervals_.size(); i != e && sidx < eidx;
773 LiveInterval::const_iterator I = intervals_[i]->find(sidx);
774 LiveInterval::const_iterator E = intervals_[i]->end();
777 // Interval I is overlapping [sidx;eidx). Trim sidx.
778 if (I->start <= sidx) {
783 // Trim eidx too if needed.
784 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 /// rewrite - after all the new live ranges have been created, rewrite
801 /// instructions using curli to use the new intervals.
802 bool SplitEditor::rewrite() {
803 assert(!openli_.getLI() && "Previous LI not closed before rewrite");
805 // First we need to fill in the live ranges in dupli.
806 // If values were redefined, we need a full recoloring with SSA update.
807 // If values were truncated, we only need to truncate the ranges.
808 // If values were partially rematted, we should shrink to uses.
809 // If values were fully rematted, they should be omitted.
810 // FIXME: If a single value is redefined, just move the def and truncate.
812 // Values that are fully contained in the split intervals.
813 SmallPtrSet<const VNInfo*, 8> deadValues;
815 // Map all curli values that should have live defs in dupli.
816 for (LiveInterval::const_vni_iterator I = curli_->vni_begin(),
817 E = curli_->vni_end(); I != E; ++I) {
818 const VNInfo *VNI = *I;
819 // Original def is contained in the split intervals.
820 if (intervalsLiveAt(VNI->def)) {
821 // Did this value escape?
822 if (dupli_.isMapped(VNI))
823 truncatedValues.insert(VNI);
825 deadValues.insert(VNI);
828 // Add minimal live range at the definition.
829 VNInfo *DVNI = dupli_.defValue(VNI, VNI->def);
830 dupli_.getLI()->addRange(LiveRange(VNI->def, VNI->def.getNextSlot(), DVNI));
833 // Add all ranges to dupli.
834 for (LiveInterval::const_iterator I = curli_->begin(), E = curli_->end();
836 const LiveRange &LR = *I;
837 if (truncatedValues.count(LR.valno)) {
838 // recolor after removing intervals_.
839 addTruncSimpleRange(LR.start, LR.end, LR.valno);
840 } else if (!deadValues.count(LR.valno)) {
841 // recolor without truncation.
842 dupli_.addSimpleRange(LR.start, LR.end, LR.valno);
847 const LiveInterval *curli = sa_.getCurLI();
848 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(curli->reg),
849 RE = mri_.reg_end(); RI != RE;) {
850 MachineOperand &MO = RI.getOperand();
851 MachineInstr *MI = MO.getParent();
853 if (MI->isDebugValue()) {
854 DEBUG(dbgs() << "Zapping " << *MI);
855 // FIXME: We can do much better with debug values.
859 SlotIndex Idx = lis_.getInstructionIndex(MI);
860 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
861 LiveInterval *LI = dupli_.getLI();
862 for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
863 LiveInterval *testli = intervals_[i];
864 if (testli->liveAt(Idx)) {
872 DEBUG(dbgs() << " rewrite " << Idx << '\t' << *MI);
876 // dupli_ goes in last, after rewriting.
877 if (dupli_.getLI()) {
878 if (dupli_.getLI()->empty()) {
879 DEBUG(dbgs() << " dupli became empty?\n");
880 lis_.removeInterval(dupli_.getLI()->reg);
883 dupli_.getLI()->RenumberValues(lis_);
884 intervals_.push_back(dupli_.getLI());
888 // Calculate spill weight and allocation hints for new intervals.
889 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
890 for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
891 LiveInterval &li = *intervals_[i];
892 vrai.CalculateRegClass(li.reg);
893 vrai.CalculateWeightAndHint(li);
894 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
895 << ":" << li << '\n');
897 return dupli_.getLI();
901 //===----------------------------------------------------------------------===//
903 //===----------------------------------------------------------------------===//
905 bool SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
906 SplitAnalysis::LoopBlocks Blocks;
907 sa_.getLoopBlocks(Loop, Blocks);
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. Return true if curli has been completely replaced,
948 /// false if curli is still intact, and needs to be spilled or split further.
949 bool SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
950 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
951 // Determine the first and last instruction using curli in each block.
952 typedef std::pair<SlotIndex,SlotIndex> IndexPair;
953 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
954 IndexPairMap MBBRange;
955 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
956 E = sa_.usingInstrs_.end(); I != E; ++I) {
957 const MachineBasicBlock *MBB = (*I)->getParent();
958 if (!Blocks.count(MBB))
960 SlotIndex Idx = lis_.getInstructionIndex(*I);
961 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
962 IndexPair &IP = MBBRange[MBB];
963 if (!IP.first.isValid() || Idx < IP.first)
965 if (!IP.second.isValid() || Idx > IP.second)
969 // Create a new interval for each block.
970 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
971 E = Blocks.end(); I != E; ++I) {
972 IndexPair &IP = MBBRange[*I];
973 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": ["
974 << IP.first << ';' << IP.second << ")\n");
975 assert(IP.first.isValid() && IP.second.isValid());
978 enterIntvBefore(IP.first);
979 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
980 leaveIntvAfter(IP.second);
987 //===----------------------------------------------------------------------===//
988 // Sub Block Splitting
989 //===----------------------------------------------------------------------===//
991 /// getBlockForInsideSplit - If curli is contained inside a single basic block,
992 /// and it wou pay to subdivide the interval inside that block, return it.
993 /// Otherwise return NULL. The returned block can be passed to
994 /// SplitEditor::splitInsideBlock.
995 const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
996 // The interval must be exclusive to one block.
997 if (usingBlocks_.size() != 1)
999 // Don't to this for less than 4 instructions. We want to be sure that
1000 // splitting actually reduces the instruction count per interval.
1001 if (usingInstrs_.size() < 4)
1003 return usingBlocks_.begin()->first;
1006 /// splitInsideBlock - Split curli into multiple intervals inside MBB. Return
1007 /// true if curli has been completely replaced, false if curli is still
1008 /// intact, and needs to be spilled or split further.
1009 bool SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
1010 SmallVector<SlotIndex, 32> Uses;
1011 Uses.reserve(sa_.usingInstrs_.size());
1012 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
1013 E = sa_.usingInstrs_.end(); I != E; ++I)
1014 if ((*I)->getParent() == MBB)
1015 Uses.push_back(lis_.getInstructionIndex(*I));
1016 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for "
1017 << Uses.size() << " instructions.\n");
1018 assert(Uses.size() >= 3 && "Need at least 3 instructions");
1019 array_pod_sort(Uses.begin(), Uses.end());
1021 // Simple algorithm: Find the largest gap between uses as determined by slot
1022 // indices. Create new intervals for instructions before the gap and after the
1024 unsigned bestPos = 0;
1026 DEBUG(dbgs() << " dist (" << Uses[0]);
1027 for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
1028 int g = Uses[i-1].distance(Uses[i]);
1029 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
1031 bestPos = i, bestGap = g;
1033 DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
1035 // bestPos points to the first use after the best gap.
1036 assert(bestPos > 0 && "Invalid gap");
1038 // FIXME: Don't create intervals for low densities.
1040 // First interval before the gap. Don't create single-instr intervals.
1043 enterIntvBefore(Uses.front());
1044 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
1045 leaveIntvAfter(Uses[bestPos-1]);
1049 // Second interval after the gap.
1050 if (bestPos < Uses.size()-1) {
1052 enterIntvBefore(Uses[bestPos]);
1053 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
1054 leaveIntvAfter(Uses.back());