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 // defValue - Introduce a li_ def for ParentVNI that could be later than
356 VNInfo *LiveIntervalMap::defValue(const VNInfo *ParentVNI, SlotIndex Idx) {
357 assert(li_ && "call reset first");
358 assert(ParentVNI && "Mapping NULL value");
359 assert(Idx.isValid() && "Invalid SlotIndex");
360 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
362 // Is this a simple 1-1 mapping? Not likely.
363 if (Idx == ParentVNI->def)
364 return mapValue(ParentVNI, Idx);
366 // This is now a complex def. Mark with a NULL in valueMap.
367 valueMap_[ParentVNI] = 0;
369 // Should we insert a minimal snippet of VNI LiveRange, or can we count on
370 // callers to do that? We need it for lookups of complex values.
371 VNInfo *VNI = li_->getNextValue(Idx, 0, true, lis_.getVNInfoAllocator());
375 // mapValue - Find the mapped value for ParentVNI at Idx.
376 // Potentially create phi-def values.
377 VNInfo *LiveIntervalMap::mapValue(const VNInfo *ParentVNI, SlotIndex Idx) {
378 assert(li_ && "call reset first");
379 assert(ParentVNI && "Mapping NULL value");
380 assert(Idx.isValid() && "Invalid SlotIndex");
381 assert(parentli_.getVNInfoAt(Idx) == ParentVNI && "Bad ParentVNI");
383 // Use insert for lookup, so we can add missing values with a second lookup.
384 std::pair<ValueMap::iterator,bool> InsP =
385 valueMap_.insert(makeVV(ParentVNI, 0));
387 // This was an unknown value. Create a simple mapping.
389 return InsP.first->second = li_->createValueCopy(ParentVNI,
390 lis_.getVNInfoAllocator());
391 // This was a simple mapped value.
392 if (InsP.first->second)
393 return InsP.first->second;
395 // This is a complex mapped value. There may be multiple defs, and we may need
396 // to create phi-defs.
397 MachineBasicBlock *IdxMBB = lis_.getMBBFromIndex(Idx);
398 assert(IdxMBB && "No MBB at Idx");
400 // Is there a def in the same MBB we can extend?
401 if (VNInfo *VNI = extendTo(IdxMBB, Idx))
404 // Now for the fun part. We know that ParentVNI potentially has multiple defs,
405 // and we may need to create even more phi-defs to preserve VNInfo SSA form.
406 // Perform a depth-first search for predecessor blocks where we know the
407 // dominating VNInfo. Insert phi-def VNInfos along the path back to IdxMBB.
409 // Track MBBs where we have created or learned the dominating value.
410 // This may change during the DFS as we create new phi-defs.
411 typedef DenseMap<MachineBasicBlock*, VNInfo*> MBBValueMap;
412 MBBValueMap DomValue;
414 for (idf_iterator<MachineBasicBlock*>
415 IDFI = idf_begin(IdxMBB),
416 IDFE = idf_end(IdxMBB); IDFI != IDFE;) {
417 MachineBasicBlock *MBB = *IDFI;
418 SlotIndex End = lis_.getMBBEndIdx(MBB).getPrevSlot();
420 // We are operating on the restricted CFG where ParentVNI is live.
421 if (parentli_.getVNInfoAt(End) != ParentVNI) {
426 // Do we have a dominating value in this block?
427 VNInfo *VNI = extendTo(MBB, End);
433 // Yes, VNI dominates MBB. Track the path back to IdxMBB, creating phi-defs
434 // as needed along the way.
435 for (unsigned PI = IDFI.getPathLength()-1; PI != 0; --PI) {
436 // Start from MBB's immediate successor. End at IdxMBB.
437 MachineBasicBlock *Succ = IDFI.getPath(PI-1);
438 std::pair<MBBValueMap::iterator, bool> InsP =
439 DomValue.insert(MBBValueMap::value_type(Succ, VNI));
441 // This is the first time we backtrack to Succ.
445 // We reached Succ again with the same VNI. Nothing is going to change.
446 VNInfo *OVNI = InsP.first->second;
450 // Succ already has a phi-def. No need to continue.
451 SlotIndex Start = lis_.getMBBStartIdx(Succ);
452 if (OVNI->def == Start)
455 // We have a collision between the old and new VNI at Succ. That means
456 // neither dominates and we need a new phi-def.
457 VNI = li_->getNextValue(Start, 0, true, lis_.getVNInfoAllocator());
458 VNI->setIsPHIDef(true);
459 InsP.first->second = VNI;
461 // Replace OVNI with VNI in the remaining path.
462 for (; PI > 1 ; --PI) {
463 MBBValueMap::iterator I = DomValue.find(IDFI.getPath(PI-2));
464 if (I == DomValue.end() || I->second != OVNI)
470 // No need to search the children, we found a dominating value.
474 // The search should at least find a dominating value for IdxMBB.
475 assert(!DomValue.empty() && "Couldn't find a reaching definition");
477 // Since we went through the trouble of a full DFS visiting all reaching defs,
478 // the values in DomValue are now accurate. No more phi-defs are needed for
479 // these blocks, so we can color the live ranges.
480 // This makes the next mapValue call much faster.
482 for (MBBValueMap::iterator I = DomValue.begin(), E = DomValue.end(); I != E;
484 MachineBasicBlock *MBB = I->first;
485 VNInfo *VNI = I->second;
486 SlotIndex Start = lis_.getMBBStartIdx(MBB);
488 // Don't add full liveness to IdxMBB, stop at Idx.
490 li_->addRange(LiveRange(Start, Idx.getNextSlot(), VNI));
491 // The caller had better add some liveness to IdxVNI, or it leaks.
494 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
497 assert(IdxVNI && "Didn't find value for Idx");
501 // extendTo - Find the last li_ value defined in MBB at or before Idx. The
502 // parentli_ is assumed to be live at Idx. Extend the live range to Idx.
503 // Return the found VNInfo, or NULL.
504 VNInfo *LiveIntervalMap::extendTo(MachineBasicBlock *MBB, SlotIndex Idx) {
505 assert(li_ && "call reset first");
506 LiveInterval::iterator I = std::upper_bound(li_->begin(), li_->end(), Idx);
507 if (I == li_->begin())
510 if (I->start < lis_.getMBBStartIdx(MBB))
513 I->end = Idx.getNextSlot();
517 // addSimpleRange - Add a simple range from parentli_ to li_.
518 // ParentVNI must be live in the [Start;End) interval.
519 void LiveIntervalMap::addSimpleRange(SlotIndex Start, SlotIndex End,
520 const VNInfo *ParentVNI) {
521 assert(li_ && "call reset first");
522 VNInfo *VNI = mapValue(ParentVNI, Start);
523 // A simple mappoing is easy.
524 if (VNI->def == ParentVNI->def) {
525 li_->addRange(LiveRange(Start, End, VNI));
529 // ParentVNI is a complex value. We must map per MBB.
530 MachineFunction::iterator MBB = lis_.getMBBFromIndex(Start);
531 MachineFunction::iterator MBBE = lis_.getMBBFromIndex(End);
534 li_->addRange(LiveRange(Start, End, VNI));
539 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB), VNI));
541 // Run sequence of full blocks.
542 for (++MBB; MBB != MBBE; ++MBB) {
543 Start = lis_.getMBBStartIdx(MBB);
544 li_->addRange(LiveRange(Start, lis_.getMBBEndIdx(MBB),
545 mapValue(ParentVNI, Start)));
549 Start = lis_.getMBBStartIdx(MBB);
551 li_->addRange(LiveRange(Start, End, mapValue(ParentVNI, Start)));
554 /// addRange - Add live ranges to li_ where [Start;End) intersects parentli_.
555 /// All needed values whose def is not inside [Start;End) must be defined
556 /// beforehand so mapValue will work.
557 void LiveIntervalMap::addRange(SlotIndex Start, SlotIndex End) {
558 assert(li_ && "call reset first");
559 LiveInterval::const_iterator B = parentli_.begin(), E = parentli_.end();
560 LiveInterval::const_iterator I = std::lower_bound(B, E, Start);
562 // Check if --I begins before Start and overlaps.
566 addSimpleRange(Start, std::min(End, I->end), I->valno);
570 // The remaining ranges begin after Start.
571 for (;I != E && I->start < End; ++I)
572 addSimpleRange(I->start, std::min(End, I->end), I->valno);
575 VNInfo *LiveIntervalMap::defByCopyFrom(unsigned Reg,
576 const VNInfo *ParentVNI,
577 MachineBasicBlock &MBB,
578 MachineBasicBlock::iterator I) {
579 const TargetInstrDesc &TID = MBB.getParent()->getTarget().getInstrInfo()->
580 get(TargetOpcode::COPY);
581 MachineInstr *MI = BuildMI(MBB, I, DebugLoc(), TID, li_->reg).addReg(Reg);
582 SlotIndex DefIdx = lis_.InsertMachineInstrInMaps(MI).getDefIndex();
583 VNInfo *VNI = defValue(ParentVNI, DefIdx);
585 li_->addRange(LiveRange(DefIdx, DefIdx.getNextSlot(), VNI));
589 //===----------------------------------------------------------------------===//
591 //===----------------------------------------------------------------------===//
593 /// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
594 SplitEditor::SplitEditor(SplitAnalysis &sa, LiveIntervals &lis, VirtRegMap &vrm,
595 SmallVectorImpl<LiveInterval*> &intervals)
596 : sa_(sa), lis_(lis), vrm_(vrm),
597 mri_(vrm.getMachineFunction().getRegInfo()),
598 tii_(*vrm.getMachineFunction().getTarget().getInstrInfo()),
599 curli_(sa_.getCurLI()),
600 dupli_(lis_, *curli_),
601 openli_(lis_, *curli_),
602 intervals_(intervals),
603 firstInterval(intervals_.size())
605 assert(curli_ && "SplitEditor created from empty SplitAnalysis");
607 // Make sure curli_ is assigned a stack slot, so all our intervals get the
608 // same slot as curli_.
609 if (vrm_.getStackSlot(curli_->reg) == VirtRegMap::NO_STACK_SLOT)
610 vrm_.assignVirt2StackSlot(curli_->reg);
614 LiveInterval *SplitEditor::createInterval() {
615 unsigned Reg = mri_.createVirtualRegister(mri_.getRegClass(curli_->reg));
616 LiveInterval &Intv = lis_.getOrCreateInterval(Reg);
618 vrm_.assignVirt2StackSlot(Reg, vrm_.getStackSlot(curli_->reg));
622 /// Create a new virtual register and live interval.
623 void SplitEditor::openIntv() {
624 assert(!openli_.getLI() && "Previous LI not closed before openIntv");
626 if (!dupli_.getLI()) {
627 // Create an interval for dupli that is a copy of curli.
628 dupli_.reset(createInterval());
629 dupli_.getLI()->Copy(*curli_, &mri_, lis_.getVNInfoAllocator());
632 openli_.reset(createInterval());
633 intervals_.push_back(openli_.getLI());
636 /// enterIntvBefore - Enter openli before the instruction at Idx. If curli is
637 /// not live before Idx, a COPY is not inserted.
638 void SplitEditor::enterIntvBefore(SlotIndex Idx) {
639 assert(openli_.getLI() && "openIntv not called before enterIntvBefore");
640 VNInfo *ParentVNI = curli_->getVNInfoAt(Idx.getUseIndex());
642 DEBUG(dbgs() << " enterIntvBefore " << Idx << ": not live\n");
645 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
646 assert(MI && "enterIntvBefore called with invalid index");
647 openli_.defByCopyFrom(curli_->reg, ParentVNI, *MI->getParent(), MI);
648 DEBUG(dbgs() << " enterIntvBefore " << Idx << ": " << *openli_.getLI()
652 /// enterIntvAtEnd - Enter openli at the end of MBB.
653 void SplitEditor::enterIntvAtEnd(MachineBasicBlock &MBB) {
654 assert(openli_.getLI() && "openIntv not called before enterIntvAtEnd");
655 SlotIndex End = lis_.getMBBEndIdx(&MBB);
656 VNInfo *ParentVNI = curli_->getVNInfoAt(End.getPrevSlot());
658 DEBUG(dbgs() << " enterIntvAtEnd " << End << ": not live\n");
661 VNInfo *VNI = openli_.defByCopyFrom(curli_->reg, ParentVNI,
662 MBB, MBB.getFirstTerminator());
663 // Make sure openli is live out of MBB.
664 openli_.getLI()->addRange(LiveRange(VNI->def, End, VNI));
665 DEBUG(dbgs() << " enterIntvAtEnd: " << *openli_.getLI() << '\n');
668 /// useIntv - indicate that all instructions in MBB should use openli.
669 void SplitEditor::useIntv(const MachineBasicBlock &MBB) {
670 useIntv(lis_.getMBBStartIdx(&MBB), lis_.getMBBEndIdx(&MBB));
673 void SplitEditor::useIntv(SlotIndex Start, SlotIndex End) {
674 assert(openli_.getLI() && "openIntv not called before useIntv");
675 openli_.addRange(Start, End);
676 DEBUG(dbgs() << " use [" << Start << ';' << End << "): "
677 << *openli_.getLI() << '\n');
680 /// leaveIntvAfter - Leave openli after the instruction at Idx.
681 void SplitEditor::leaveIntvAfter(SlotIndex Idx) {
682 assert(openli_.getLI() && "openIntv not called before leaveIntvAfter");
684 // The interval must be live beyond the instruction at Idx.
685 SlotIndex EndIdx = Idx.getNextIndex().getBaseIndex();
686 VNInfo *ParentVNI = curli_->getVNInfoAt(EndIdx);
688 DEBUG(dbgs() << " leaveIntvAfter " << Idx << ": not live\n");
692 MachineInstr *MI = lis_.getInstructionFromIndex(Idx);
693 assert(MI && "leaveIntvAfter called with invalid index");
695 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
696 *MI->getParent(), MI);
698 // Finally we must make sure that openli is properly extended from Idx to the
700 openli_.mapValue(ParentVNI, VNI->def.getUseIndex());
702 DEBUG(dbgs() << " leaveIntvAfter " << Idx << ": " << *openli_.getLI()
706 /// leaveIntvAtTop - Leave the interval at the top of MBB.
707 /// Currently, only one value can leave the interval.
708 void SplitEditor::leaveIntvAtTop(MachineBasicBlock &MBB) {
709 assert(openli_.getLI() && "openIntv not called before leaveIntvAtTop");
711 SlotIndex Start = lis_.getMBBStartIdx(&MBB);
712 VNInfo *ParentVNI = curli_->getVNInfoAt(Start);
714 // Is curli even live-in to MBB?
716 DEBUG(dbgs() << " leaveIntvAtTop at " << Start << ": not live\n");
720 // We are going to insert a back copy, so we must have a dupli_.
721 VNInfo *VNI = dupli_.defByCopyFrom(openli_.getLI()->reg, ParentVNI,
724 // Finally we must make sure that openli is properly extended from Start to
726 openli_.mapValue(ParentVNI, VNI->def.getUseIndex());
728 DEBUG(dbgs() << " leaveIntvAtTop at " << Start << ": " << *openli_.getLI()
732 /// closeIntv - Indicate that we are done editing the currently open
733 /// LiveInterval, and ranges can be trimmed.
734 void SplitEditor::closeIntv() {
735 assert(openli_.getLI() && "openIntv not called before closeIntv");
737 DEBUG(dbgs() << " closeIntv cleaning up\n");
738 DEBUG(dbgs() << " open " << *openli_.getLI() << '\n');
740 for (LiveInterval::iterator I = openli_.getLI()->begin(),
741 E = openli_.getLI()->end(); I != E; ++I) {
742 dupli_.getLI()->removeRange(I->start, I->end);
744 // FIXME: A block branching to the entry block may also branch elsewhere
745 // curli is live. We need both openli and curli to be live in that case.
746 DEBUG(dbgs() << " dup2 " << *dupli_.getLI() << '\n');
750 /// rewrite - after all the new live ranges have been created, rewrite
751 /// instructions using curli to use the new intervals.
752 bool SplitEditor::rewrite() {
753 assert(!openli_.getLI() && "Previous LI not closed before rewrite");
754 const LiveInterval *curli = sa_.getCurLI();
755 for (MachineRegisterInfo::reg_iterator RI = mri_.reg_begin(curli->reg),
756 RE = mri_.reg_end(); RI != RE;) {
757 MachineOperand &MO = RI.getOperand();
758 MachineInstr *MI = MO.getParent();
760 if (MI->isDebugValue()) {
761 DEBUG(dbgs() << "Zapping " << *MI);
762 // FIXME: We can do much better with debug values.
766 SlotIndex Idx = lis_.getInstructionIndex(MI);
767 Idx = MO.isUse() ? Idx.getUseIndex() : Idx.getDefIndex();
768 LiveInterval *LI = dupli_.getLI();
769 for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
770 LiveInterval *testli = intervals_[i];
771 if (testli->liveAt(Idx)) {
779 DEBUG(dbgs() << " rewrite " << Idx << '\t' << *MI);
783 // dupli_ goes in last, after rewriting.
784 if (dupli_.getLI()) {
785 if (dupli_.getLI()->empty()) {
786 DEBUG(dbgs() << " dupli became empty?\n");
787 lis_.removeInterval(dupli_.getLI()->reg);
790 dupli_.getLI()->RenumberValues(lis_);
791 intervals_.push_back(dupli_.getLI());
795 // Calculate spill weight and allocation hints for new intervals.
796 VirtRegAuxInfo vrai(vrm_.getMachineFunction(), lis_, sa_.loops_);
797 for (unsigned i = firstInterval, e = intervals_.size(); i != e; ++i) {
798 LiveInterval &li = *intervals_[i];
799 vrai.CalculateRegClass(li.reg);
800 vrai.CalculateWeightAndHint(li);
801 DEBUG(dbgs() << " new interval " << mri_.getRegClass(li.reg)->getName()
802 << ":" << li << '\n');
804 return dupli_.getLI();
808 //===----------------------------------------------------------------------===//
810 //===----------------------------------------------------------------------===//
812 bool SplitEditor::splitAroundLoop(const MachineLoop *Loop) {
813 SplitAnalysis::LoopBlocks Blocks;
814 sa_.getLoopBlocks(Loop, Blocks);
816 // Break critical edges as needed.
817 SplitAnalysis::BlockPtrSet CriticalExits;
818 sa_.getCriticalExits(Blocks, CriticalExits);
819 assert(CriticalExits.empty() && "Cannot break critical exits yet");
821 // Create new live interval for the loop.
824 // Insert copies in the predecessors.
825 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Preds.begin(),
826 E = Blocks.Preds.end(); I != E; ++I) {
827 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
831 // Switch all loop blocks.
832 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Loop.begin(),
833 E = Blocks.Loop.end(); I != E; ++I)
836 // Insert back copies in the exit blocks.
837 for (SplitAnalysis::BlockPtrSet::iterator I = Blocks.Exits.begin(),
838 E = Blocks.Exits.end(); I != E; ++I) {
839 MachineBasicBlock &MBB = const_cast<MachineBasicBlock&>(**I);
849 //===----------------------------------------------------------------------===//
850 // Single Block Splitting
851 //===----------------------------------------------------------------------===//
853 /// splitSingleBlocks - Split curli into a separate live interval inside each
854 /// basic block in Blocks. Return true if curli has been completely replaced,
855 /// false if curli is still intact, and needs to be spilled or split further.
856 bool SplitEditor::splitSingleBlocks(const SplitAnalysis::BlockPtrSet &Blocks) {
857 DEBUG(dbgs() << " splitSingleBlocks for " << Blocks.size() << " blocks.\n");
858 // Determine the first and last instruction using curli in each block.
859 typedef std::pair<SlotIndex,SlotIndex> IndexPair;
860 typedef DenseMap<const MachineBasicBlock*,IndexPair> IndexPairMap;
861 IndexPairMap MBBRange;
862 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
863 E = sa_.usingInstrs_.end(); I != E; ++I) {
864 const MachineBasicBlock *MBB = (*I)->getParent();
865 if (!Blocks.count(MBB))
867 SlotIndex Idx = lis_.getInstructionIndex(*I);
868 DEBUG(dbgs() << " BB#" << MBB->getNumber() << '\t' << Idx << '\t' << **I);
869 IndexPair &IP = MBBRange[MBB];
870 if (!IP.first.isValid() || Idx < IP.first)
872 if (!IP.second.isValid() || Idx > IP.second)
876 // Create a new interval for each block.
877 for (SplitAnalysis::BlockPtrSet::const_iterator I = Blocks.begin(),
878 E = Blocks.end(); I != E; ++I) {
879 IndexPair &IP = MBBRange[*I];
880 DEBUG(dbgs() << " splitting for BB#" << (*I)->getNumber() << ": ["
881 << IP.first << ';' << IP.second << ")\n");
882 assert(IP.first.isValid() && IP.second.isValid());
885 enterIntvBefore(IP.first);
886 useIntv(IP.first.getBaseIndex(), IP.second.getBoundaryIndex());
887 leaveIntvAfter(IP.second);
894 //===----------------------------------------------------------------------===//
895 // Sub Block Splitting
896 //===----------------------------------------------------------------------===//
898 /// getBlockForInsideSplit - If curli is contained inside a single basic block,
899 /// and it wou pay to subdivide the interval inside that block, return it.
900 /// Otherwise return NULL. The returned block can be passed to
901 /// SplitEditor::splitInsideBlock.
902 const MachineBasicBlock *SplitAnalysis::getBlockForInsideSplit() {
903 // The interval must be exclusive to one block.
904 if (usingBlocks_.size() != 1)
906 // Don't to this for less than 4 instructions. We want to be sure that
907 // splitting actually reduces the instruction count per interval.
908 if (usingInstrs_.size() < 4)
910 return usingBlocks_.begin()->first;
913 /// splitInsideBlock - Split curli into multiple intervals inside MBB. Return
914 /// true if curli has been completely replaced, false if curli is still
915 /// intact, and needs to be spilled or split further.
916 bool SplitEditor::splitInsideBlock(const MachineBasicBlock *MBB) {
917 SmallVector<SlotIndex, 32> Uses;
918 Uses.reserve(sa_.usingInstrs_.size());
919 for (SplitAnalysis::InstrPtrSet::const_iterator I = sa_.usingInstrs_.begin(),
920 E = sa_.usingInstrs_.end(); I != E; ++I)
921 if ((*I)->getParent() == MBB)
922 Uses.push_back(lis_.getInstructionIndex(*I));
923 DEBUG(dbgs() << " splitInsideBlock BB#" << MBB->getNumber() << " for "
924 << Uses.size() << " instructions.\n");
925 assert(Uses.size() >= 3 && "Need at least 3 instructions");
926 array_pod_sort(Uses.begin(), Uses.end());
928 // Simple algorithm: Find the largest gap between uses as determined by slot
929 // indices. Create new intervals for instructions before the gap and after the
931 unsigned bestPos = 0;
933 DEBUG(dbgs() << " dist (" << Uses[0]);
934 for (unsigned i = 1, e = Uses.size(); i != e; ++i) {
935 int g = Uses[i-1].distance(Uses[i]);
936 DEBUG(dbgs() << ") -" << g << "- (" << Uses[i]);
938 bestPos = i, bestGap = g;
940 DEBUG(dbgs() << "), best: -" << bestGap << "-\n");
942 // bestPos points to the first use after the best gap.
943 assert(bestPos > 0 && "Invalid gap");
945 // FIXME: Don't create intervals for low densities.
947 // First interval before the gap. Don't create single-instr intervals.
950 enterIntvBefore(Uses.front());
951 useIntv(Uses.front().getBaseIndex(), Uses[bestPos-1].getBoundaryIndex());
952 leaveIntvAfter(Uses[bestPos-1]);
956 // Second interval after the gap.
957 if (bestPos < Uses.size()-1) {
959 enterIntvBefore(Uses[bestPos]);
960 useIntv(Uses[bestPos].getBaseIndex(), Uses.back().getBoundaryIndex());
961 leaveIntvAfter(Uses.back());