1 //===- LoopIndexSplit.cpp - Loop Index Splitting Pass ---------------------===//
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 implements Loop Index Splitting Pass. This pass handles three
13 // [1] A loop may be eliminated if the body is executed exactly once.
16 // for (i = 0; i < N; ++i) {
27 // [2] A loop's iteration space may be shrunk if the loop body is executed
28 // for a proper sub-range of the loop's iteration space. For example,
30 // for (i = 0; i < N; ++i) {
31 // if (i > A && i < B) {
36 // is transformed to iterators from A to B, if A > 0 and B < N.
38 // [3] A loop may be split if the loop body is dominated by a branch.
41 // for (i = LB; i < UB; ++i) { if (i < SV) A; else B; }
43 // is transformed into
46 // for (i = LB; i < min(UB, AEV); ++i)
48 // for (i = max(LB, BSV); i < UB; ++i);
51 //===----------------------------------------------------------------------===//
53 #define DEBUG_TYPE "loop-index-split"
55 #include "llvm/Transforms/Scalar.h"
56 #include "llvm/IntrinsicInst.h"
57 #include "llvm/Analysis/LoopPass.h"
58 #include "llvm/Analysis/ScalarEvolution.h"
59 #include "llvm/Analysis/Dominators.h"
60 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
61 #include "llvm/Transforms/Utils/Cloning.h"
62 #include "llvm/Transforms/Utils/Local.h"
63 #include "llvm/Support/Compiler.h"
64 #include "llvm/ADT/DepthFirstIterator.h"
65 #include "llvm/ADT/Statistic.h"
69 STATISTIC(NumIndexSplit, "Number of loop index split");
70 STATISTIC(NumIndexSplitRemoved, "Number of loops eliminated by loop index split");
71 STATISTIC(NumRestrictBounds, "Number of loop iteration space restricted");
75 class VISIBILITY_HIDDEN LoopIndexSplit : public LoopPass {
78 static char ID; // Pass ID, replacement for typeid
79 LoopIndexSplit() : LoopPass(&ID) {}
81 // Index split Loop L. Return true if loop is split.
82 bool runOnLoop(Loop *L, LPPassManager &LPM);
84 void getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addPreserved<ScalarEvolution>();
86 AU.addRequiredID(LCSSAID);
87 AU.addPreservedID(LCSSAID);
88 AU.addRequired<LoopInfo>();
89 AU.addPreserved<LoopInfo>();
90 AU.addRequiredID(LoopSimplifyID);
91 AU.addPreservedID(LoopSimplifyID);
92 AU.addRequired<DominatorTree>();
93 AU.addRequired<DominanceFrontier>();
94 AU.addPreserved<DominatorTree>();
95 AU.addPreserved<DominanceFrontier>();
99 /// processOneIterationLoop -- Eliminate loop if loop body is executed
100 /// only once. For example,
101 /// for (i = 0; i < N; ++i) {
107 bool processOneIterationLoop();
109 // -- Routines used by updateLoopIterationSpace();
111 /// updateLoopIterationSpace -- Update loop's iteration space if loop
112 /// body is executed for certain IV range only. For example,
114 /// for (i = 0; i < N; ++i) {
115 /// if ( i > A && i < B) {
119 /// is transformed to iterators from A to B, if A > 0 and B < N.
121 bool updateLoopIterationSpace();
123 /// restrictLoopBound - Op dominates loop body. Op compares an IV based value
124 /// with a loop invariant value. Update loop's lower and upper bound based on
125 /// the loop invariant value.
126 bool restrictLoopBound(ICmpInst &Op);
128 // --- Routines used by splitLoop(). --- /
132 /// removeBlocks - Remove basic block DeadBB and all blocks dominated by
133 /// DeadBB. This routine is used to remove split condition's dead branch,
134 /// dominated by DeadBB. LiveBB dominates split conidition's other branch.
135 void removeBlocks(BasicBlock *DeadBB, Loop *LP, BasicBlock *LiveBB);
137 /// moveExitCondition - Move exit condition EC into split condition block.
138 void moveExitCondition(BasicBlock *CondBB, BasicBlock *ActiveBB,
139 BasicBlock *ExitBB, ICmpInst *EC, ICmpInst *SC,
140 PHINode *IV, Instruction *IVAdd, Loop *LP,
143 /// updatePHINodes - CFG has been changed.
145 /// - ExitBB's single predecessor was Latch
146 /// - Latch's second successor was Header
148 /// - ExitBB's single predecessor was Header
149 /// - Latch's one and only successor was Header
151 /// Update ExitBB PHINodes' to reflect this change.
152 void updatePHINodes(BasicBlock *ExitBB, BasicBlock *Latch,
154 PHINode *IV, Instruction *IVIncrement, Loop *LP);
156 // --- Utility routines --- /
158 /// cleanBlock - A block is considered clean if all non terminal
159 /// instructions are either PHINodes or IV based values.
160 bool cleanBlock(BasicBlock *BB);
162 /// IVisLT - If Op is comparing IV based value with an loop invariant and
163 /// IV based value is less than the loop invariant then return the loop
164 /// invariant. Otherwise return NULL.
165 Value * IVisLT(ICmpInst &Op);
167 /// IVisLE - If Op is comparing IV based value with an loop invariant and
168 /// IV based value is less than or equal to the loop invariant then
169 /// return the loop invariant. Otherwise return NULL.
170 Value * IVisLE(ICmpInst &Op);
172 /// IVisGT - If Op is comparing IV based value with an loop invariant and
173 /// IV based value is greater than the loop invariant then return the loop
174 /// invariant. Otherwise return NULL.
175 Value * IVisGT(ICmpInst &Op);
177 /// IVisGE - If Op is comparing IV based value with an loop invariant and
178 /// IV based value is greater than or equal to the loop invariant then
179 /// return the loop invariant. Otherwise return NULL.
180 Value * IVisGE(ICmpInst &Op);
184 // Current Loop information.
189 DominanceFrontier *DF;
192 ICmpInst *ExitCondition;
193 ICmpInst *SplitCondition;
196 Instruction *IVIncrement;
197 SmallPtrSet<Value *, 4> IVBasedValues;
201 char LoopIndexSplit::ID = 0;
202 static RegisterPass<LoopIndexSplit>
203 X("loop-index-split", "Index Split Loops");
205 Pass *llvm::createLoopIndexSplitPass() {
206 return new LoopIndexSplit();
209 // Index split Loop L. Return true if loop is split.
210 bool LoopIndexSplit::runOnLoop(Loop *IncomingLoop, LPPassManager &LPM_Ref) {
214 // FIXME - Nested loops make dominator info updates tricky.
215 if (!L->getSubLoops().empty())
218 DT = &getAnalysis<DominatorTree>();
219 LI = &getAnalysis<LoopInfo>();
220 DF = &getAnalysis<DominanceFrontier>();
222 // Initialize loop data.
223 IndVar = L->getCanonicalInductionVariable();
224 if (!IndVar) return false;
226 bool P1InLoop = L->contains(IndVar->getIncomingBlock(1));
227 IVStartValue = IndVar->getIncomingValue(!P1InLoop);
228 IVIncrement = dyn_cast<Instruction>(IndVar->getIncomingValue(P1InLoop));
229 if (!IVIncrement) return false;
231 IVBasedValues.clear();
232 IVBasedValues.insert(IndVar);
233 IVBasedValues.insert(IVIncrement);
234 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
236 for(BasicBlock::iterator BI = (*I)->begin(), BE = (*I)->end();
238 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BI))
239 if (BO != IVIncrement
240 && (BO->getOpcode() == Instruction::Add
241 || BO->getOpcode() == Instruction::Sub))
242 if (IVBasedValues.count(BO->getOperand(0))
243 && L->isLoopInvariant(BO->getOperand(1)))
244 IVBasedValues.insert(BO);
247 // Reject loop if loop exit condition is not suitable.
248 BasicBlock *ExitingBlock = L->getExitingBlock();
251 BranchInst *EBR = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
252 if (!EBR) return false;
253 ExitCondition = dyn_cast<ICmpInst>(EBR->getCondition());
254 if (!ExitCondition) return false;
255 if (ExitingBlock != L->getLoopLatch()) return false;
256 IVExitValue = ExitCondition->getOperand(1);
257 if (!L->isLoopInvariant(IVExitValue))
258 IVExitValue = ExitCondition->getOperand(0);
259 if (!L->isLoopInvariant(IVExitValue))
262 // If start value is more then exit value where induction variable
263 // increments by 1 then we are potentially dealing with an infinite loop.
264 // Do not index split this loop.
265 if (ConstantInt *SV = dyn_cast<ConstantInt>(IVStartValue))
266 if (ConstantInt *EV = dyn_cast<ConstantInt>(IVExitValue))
267 if (SV->getSExtValue() > EV->getSExtValue())
270 if (processOneIterationLoop())
273 if (updateLoopIterationSpace())
282 // --- Helper routines ---
283 // isUsedOutsideLoop - Returns true iff V is used outside the loop L.
284 static bool isUsedOutsideLoop(Value *V, Loop *L) {
285 for(Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
286 if (!L->contains(cast<Instruction>(*UI)->getParent()))
292 static Value *getPlusOne(Value *V, bool Sign, Instruction *InsertPt) {
293 ConstantInt *One = ConstantInt::get(V->getType(), 1, Sign);
294 return BinaryOperator::CreateAdd(V, One, "lsp", InsertPt);
298 static Value *getMinusOne(Value *V, bool Sign, Instruction *InsertPt) {
299 ConstantInt *One = ConstantInt::get(V->getType(), 1, Sign);
300 return BinaryOperator::CreateSub(V, One, "lsp", InsertPt);
303 // Return min(V1, V1)
304 static Value *getMin(Value *V1, Value *V2, bool Sign, Instruction *InsertPt) {
306 Value *C = new ICmpInst(Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
307 V1, V2, "lsp", InsertPt);
308 return SelectInst::Create(C, V1, V2, "lsp", InsertPt);
311 // Return max(V1, V2)
312 static Value *getMax(Value *V1, Value *V2, bool Sign, Instruction *InsertPt) {
314 Value *C = new ICmpInst(Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
315 V1, V2, "lsp", InsertPt);
316 return SelectInst::Create(C, V2, V1, "lsp", InsertPt);
319 /// processOneIterationLoop -- Eliminate loop if loop body is executed
320 /// only once. For example,
321 /// for (i = 0; i < N; ++i) {
327 bool LoopIndexSplit::processOneIterationLoop() {
328 SplitCondition = NULL;
329 BasicBlock *Latch = L->getLoopLatch();
330 BasicBlock *Header = L->getHeader();
331 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
332 if (!BR) return false;
333 if (!isa<BranchInst>(Latch->getTerminator())) return false;
334 if (BR->isUnconditional()) return false;
335 SplitCondition = dyn_cast<ICmpInst>(BR->getCondition());
336 if (!SplitCondition) return false;
337 if (SplitCondition == ExitCondition) return false;
338 if (SplitCondition->getPredicate() != ICmpInst::ICMP_EQ) return false;
339 if (BR->getOperand(1) != Latch) return false;
340 if (!IVBasedValues.count(SplitCondition->getOperand(0))
341 && !IVBasedValues.count(SplitCondition->getOperand(1)))
344 // If IV is used outside the loop then this loop traversal is required.
345 // FIXME: Calculate and use last IV value.
346 if (isUsedOutsideLoop(IVIncrement, L))
349 // If BR operands are not IV or not loop invariants then skip this loop.
350 Value *OPV = SplitCondition->getOperand(0);
351 Value *SplitValue = SplitCondition->getOperand(1);
352 if (!L->isLoopInvariant(SplitValue)) {
353 Value *T = SplitValue;
357 if (!L->isLoopInvariant(SplitValue))
359 Instruction *OPI = dyn_cast<Instruction>(OPV);
362 if (OPI->getParent() != Header || isUsedOutsideLoop(OPI, L))
364 Value *StartValue = IVStartValue;
365 Value *ExitValue = IVExitValue;;
368 // If BR operand is IV based then use this operand to calculate
369 // effective conditions for loop body.
370 BinaryOperator *BOPV = dyn_cast<BinaryOperator>(OPV);
373 if (BOPV->getOpcode() != Instruction::Add)
375 StartValue = BinaryOperator::CreateAdd(OPV, StartValue, "" , BR);
376 ExitValue = BinaryOperator::CreateAdd(OPV, ExitValue, "" , BR);
379 if (!cleanBlock(Header))
382 if (!cleanBlock(Latch))
385 // If the merge point for BR is not loop latch then skip this loop.
386 if (BR->getSuccessor(0) != Latch) {
387 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
388 assert (DF0 != DF->end() && "Unable to find dominance frontier");
389 if (!DF0->second.count(Latch))
393 if (BR->getSuccessor(1) != Latch) {
394 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
395 assert (DF1 != DF->end() && "Unable to find dominance frontier");
396 if (!DF1->second.count(Latch))
400 // Now, Current loop L contains compare instruction
401 // that compares induction variable, IndVar, against loop invariant. And
402 // entire (i.e. meaningful) loop body is dominated by this compare
403 // instruction. In such case eliminate
404 // loop structure surrounding this loop body. For example,
405 // for (int i = start; i < end; ++i) {
406 // if ( i == somevalue) {
410 // can be transformed into
411 // if (somevalue >= start && somevalue < end) {
416 // Replace index variable with split value in loop body. Loop body is executed
417 // only when index variable is equal to split value.
418 IndVar->replaceAllUsesWith(SplitValue);
420 // Replace split condition in header.
422 // SplitCondition : icmp eq i32 IndVar, SplitValue
424 // c1 = icmp uge i32 SplitValue, StartValue
425 // c2 = icmp ult i32 SplitValue, ExitValue
427 Instruction *C1 = new ICmpInst(ExitCondition->isSignedPredicate() ?
428 ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
429 SplitValue, StartValue, "lisplit", BR);
431 CmpInst::Predicate C2P = ExitCondition->getPredicate();
432 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
433 if (LatchBR->getOperand(0) != Header)
434 C2P = CmpInst::getInversePredicate(C2P);
435 Instruction *C2 = new ICmpInst(C2P, SplitValue, ExitValue, "lisplit", BR);
436 Instruction *NSplitCond = BinaryOperator::CreateAnd(C1, C2, "lisplit", BR);
438 SplitCondition->replaceAllUsesWith(NSplitCond);
439 SplitCondition->eraseFromParent();
441 // Remove Latch to Header edge.
442 BasicBlock *LatchSucc = NULL;
443 Header->removePredecessor(Latch);
444 for (succ_iterator SI = succ_begin(Latch), E = succ_end(Latch);
450 // Clean up latch block.
451 Value *LatchBRCond = LatchBR->getCondition();
452 LatchBR->setUnconditionalDest(LatchSucc);
453 RecursivelyDeleteTriviallyDeadInstructions(LatchBRCond);
455 LPM->deleteLoopFromQueue(L);
457 // Update Dominator Info.
458 // Only CFG change done is to remove Latch to Header edge. This
459 // does not change dominator tree because Latch did not dominate
462 DominanceFrontier::iterator HeaderDF = DF->find(Header);
463 if (HeaderDF != DF->end())
464 DF->removeFromFrontier(HeaderDF, Header);
466 DominanceFrontier::iterator LatchDF = DF->find(Latch);
467 if (LatchDF != DF->end())
468 DF->removeFromFrontier(LatchDF, Header);
471 ++NumIndexSplitRemoved;
475 /// restrictLoopBound - Op dominates loop body. Op compares an IV based value
476 /// with a loop invariant value. Update loop's lower and upper bound based on
477 /// the loop invariant value.
478 bool LoopIndexSplit::restrictLoopBound(ICmpInst &Op) {
479 bool Sign = Op.isSignedPredicate();
480 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
482 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
484 cast<BranchInst>(ExitCondition->getParent()->getTerminator());
485 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
486 BasicBlock *T = EBR->getSuccessor(0);
487 EBR->setSuccessor(0, EBR->getSuccessor(1));
488 EBR->setSuccessor(1, T);
491 // New upper and lower bounds.
494 if (Value *V = IVisLT(Op)) {
495 // Restrict upper bound.
496 if (IVisLE(*ExitCondition))
497 V = getMinusOne(V, Sign, PHTerm);
498 NUB = getMin(V, IVExitValue, Sign, PHTerm);
499 } else if (Value *V = IVisLE(Op)) {
500 // Restrict upper bound.
501 if (IVisLT(*ExitCondition))
502 V = getPlusOne(V, Sign, PHTerm);
503 NUB = getMin(V, IVExitValue, Sign, PHTerm);
504 } else if (Value *V = IVisGT(Op)) {
505 // Restrict lower bound.
506 V = getPlusOne(V, Sign, PHTerm);
507 NLB = getMax(V, IVStartValue, Sign, PHTerm);
508 } else if (Value *V = IVisGE(Op))
509 // Restrict lower bound.
510 NLB = getMax(V, IVStartValue, Sign, PHTerm);
516 unsigned i = IndVar->getBasicBlockIndex(L->getLoopPreheader());
517 IndVar->setIncomingValue(i, NLB);
521 unsigned i = (ExitCondition->getOperand(0) != IVExitValue);
522 ExitCondition->setOperand(i, NUB);
527 /// updateLoopIterationSpace -- Update loop's iteration space if loop
528 /// body is executed for certain IV range only. For example,
530 /// for (i = 0; i < N; ++i) {
531 /// if ( i > A && i < B) {
535 /// is transformed to iterators from A to B, if A > 0 and B < N.
537 bool LoopIndexSplit::updateLoopIterationSpace() {
538 SplitCondition = NULL;
539 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
540 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
542 BasicBlock *Latch = L->getLoopLatch();
543 BasicBlock *Header = L->getHeader();
544 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
545 if (!BR) return false;
546 if (!isa<BranchInst>(Latch->getTerminator())) return false;
547 if (BR->isUnconditional()) return false;
548 BinaryOperator *AND = dyn_cast<BinaryOperator>(BR->getCondition());
549 if (!AND) return false;
550 if (AND->getOpcode() != Instruction::And) return false;
551 ICmpInst *Op0 = dyn_cast<ICmpInst>(AND->getOperand(0));
552 ICmpInst *Op1 = dyn_cast<ICmpInst>(AND->getOperand(1));
555 IVBasedValues.insert(AND);
556 IVBasedValues.insert(Op0);
557 IVBasedValues.insert(Op1);
558 if (!cleanBlock(Header)) return false;
559 BasicBlock *ExitingBlock = ExitCondition->getParent();
560 if (!cleanBlock(ExitingBlock)) return false;
562 // If the merge point for BR is not loop latch then skip this loop.
563 if (BR->getSuccessor(0) != Latch) {
564 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
565 assert (DF0 != DF->end() && "Unable to find dominance frontier");
566 if (!DF0->second.count(Latch))
570 if (BR->getSuccessor(1) != Latch) {
571 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
572 assert (DF1 != DF->end() && "Unable to find dominance frontier");
573 if (!DF1->second.count(Latch))
577 // Verify that loop exiting block has only two predecessor, where one pred
578 // is split condition block. The other predecessor will become exiting block's
579 // dominator after CFG is updated. TODO : Handle CFG's where exiting block has
580 // more then two predecessors. This requires extra work in updating dominator
582 BasicBlock *ExitingBBPred = NULL;
583 for (pred_iterator PI = pred_begin(ExitingBlock), PE = pred_end(ExitingBlock);
585 BasicBlock *BB = *PI;
594 if (!restrictLoopBound(*Op0))
597 if (!restrictLoopBound(*Op1))
601 if (BR->getSuccessor(0) == ExitingBlock)
602 BR->setUnconditionalDest(BR->getSuccessor(1));
604 BR->setUnconditionalDest(BR->getSuccessor(0));
606 AND->eraseFromParent();
607 if (Op0->use_empty())
608 Op0->eraseFromParent();
609 if (Op1->use_empty())
610 Op1->eraseFromParent();
612 // Update domiantor info. Now, ExitingBlock has only one predecessor,
613 // ExitingBBPred, and it is ExitingBlock's immediate domiantor.
614 DT->changeImmediateDominator(ExitingBlock, ExitingBBPred);
616 BasicBlock *ExitBlock = ExitingBlock->getTerminator()->getSuccessor(1);
617 if (L->contains(ExitBlock))
618 ExitBlock = ExitingBlock->getTerminator()->getSuccessor(0);
620 // If ExitingBlock is a member of the loop basic blocks' DF list then
621 // replace ExitingBlock with header and exit block in the DF list
622 DominanceFrontier::iterator ExitingBlockDF = DF->find(ExitingBlock);
623 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
626 if (BB == Header || BB == ExitingBlock)
628 DominanceFrontier::iterator BBDF = DF->find(BB);
629 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
630 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
631 while (DomSetI != DomSetE) {
632 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
634 BasicBlock *DFBB = *CurrentItr;
635 if (DFBB == ExitingBlock) {
636 BBDF->second.erase(DFBB);
637 for (DominanceFrontier::DomSetType::iterator
638 EBI = ExitingBlockDF->second.begin(),
639 EBE = ExitingBlockDF->second.end(); EBI != EBE; ++EBI)
640 BBDF->second.insert(*EBI);
648 /// removeBlocks - Remove basic block DeadBB and all blocks dominated by DeadBB.
649 /// This routine is used to remove split condition's dead branch, dominated by
650 /// DeadBB. LiveBB dominates split conidition's other branch.
651 void LoopIndexSplit::removeBlocks(BasicBlock *DeadBB, Loop *LP,
652 BasicBlock *LiveBB) {
654 // First update DeadBB's dominance frontier.
655 SmallVector<BasicBlock *, 8> FrontierBBs;
656 DominanceFrontier::iterator DeadBBDF = DF->find(DeadBB);
657 if (DeadBBDF != DF->end()) {
658 SmallVector<BasicBlock *, 8> PredBlocks;
660 DominanceFrontier::DomSetType DeadBBSet = DeadBBDF->second;
661 for (DominanceFrontier::DomSetType::iterator DeadBBSetI = DeadBBSet.begin(),
662 DeadBBSetE = DeadBBSet.end(); DeadBBSetI != DeadBBSetE; ++DeadBBSetI)
664 BasicBlock *FrontierBB = *DeadBBSetI;
665 FrontierBBs.push_back(FrontierBB);
667 // Rremove any PHI incoming edge from blocks dominated by DeadBB.
669 for(pred_iterator PI = pred_begin(FrontierBB), PE = pred_end(FrontierBB);
672 if (P == DeadBB || DT->dominates(DeadBB, P))
673 PredBlocks.push_back(P);
676 for(BasicBlock::iterator FBI = FrontierBB->begin(), FBE = FrontierBB->end();
678 if (PHINode *PN = dyn_cast<PHINode>(FBI)) {
679 for(SmallVector<BasicBlock *, 8>::iterator PI = PredBlocks.begin(),
680 PE = PredBlocks.end(); PI != PE; ++PI) {
682 PN->removeIncomingValue(P);
691 // Now remove DeadBB and all nodes dominated by DeadBB in df order.
692 SmallVector<BasicBlock *, 32> WorkList;
693 DomTreeNode *DN = DT->getNode(DeadBB);
694 for (df_iterator<DomTreeNode*> DI = df_begin(DN),
695 E = df_end(DN); DI != E; ++DI) {
696 BasicBlock *BB = DI->getBlock();
697 WorkList.push_back(BB);
698 BB->replaceAllUsesWith(UndefValue::get(Type::LabelTy));
701 while (!WorkList.empty()) {
702 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
703 LPM->deleteSimpleAnalysisValue(BB, LP);
704 for(BasicBlock::iterator BBI = BB->begin(), BBE = BB->end();
706 Instruction *I = BBI;
708 I->replaceAllUsesWith(UndefValue::get(I->getType()));
709 LPM->deleteSimpleAnalysisValue(I, LP);
710 I->eraseFromParent();
715 BB->eraseFromParent();
718 // Update Frontier BBs' dominator info.
719 while (!FrontierBBs.empty()) {
720 BasicBlock *FBB = FrontierBBs.back(); FrontierBBs.pop_back();
721 BasicBlock *NewDominator = FBB->getSinglePredecessor();
723 pred_iterator PI = pred_begin(FBB), PE = pred_end(FBB);
726 if (NewDominator != LiveBB) {
727 for(; PI != PE; ++PI) {
730 NewDominator = LiveBB;
733 NewDominator = DT->findNearestCommonDominator(NewDominator, P);
737 assert (NewDominator && "Unable to fix dominator info.");
738 DT->changeImmediateDominator(FBB, NewDominator);
739 DF->changeImmediateDominator(FBB, NewDominator, DT);
744 // moveExitCondition - Move exit condition EC into split condition block CondBB.
745 void LoopIndexSplit::moveExitCondition(BasicBlock *CondBB, BasicBlock *ActiveBB,
746 BasicBlock *ExitBB, ICmpInst *EC,
747 ICmpInst *SC, PHINode *IV,
748 Instruction *IVAdd, Loop *LP,
749 unsigned ExitValueNum) {
751 BasicBlock *ExitingBB = EC->getParent();
752 Instruction *CurrentBR = CondBB->getTerminator();
754 // Move exit condition into split condition block.
755 EC->moveBefore(CurrentBR);
756 EC->setOperand(ExitValueNum == 0 ? 1 : 0, IV);
758 // Move exiting block's branch into split condition block. Update its branch
760 BranchInst *ExitingBR = cast<BranchInst>(ExitingBB->getTerminator());
761 ExitingBR->moveBefore(CurrentBR);
762 BasicBlock *OrigDestBB = NULL;
763 if (ExitingBR->getSuccessor(0) == ExitBB) {
764 OrigDestBB = ExitingBR->getSuccessor(1);
765 ExitingBR->setSuccessor(1, ActiveBB);
768 OrigDestBB = ExitingBR->getSuccessor(0);
769 ExitingBR->setSuccessor(0, ActiveBB);
772 // Remove split condition and current split condition branch.
773 SC->eraseFromParent();
774 CurrentBR->eraseFromParent();
776 // Connect exiting block to original destination.
777 BranchInst::Create(OrigDestBB, ExitingBB);
780 updatePHINodes(ExitBB, ExitingBB, CondBB, IV, IVAdd, LP);
782 // Fix dominator info.
783 // ExitBB is now dominated by CondBB
784 DT->changeImmediateDominator(ExitBB, CondBB);
785 DF->changeImmediateDominator(ExitBB, CondBB, DT);
787 // Basicblocks dominated by ActiveBB may have ExitingBB or
788 // a basic block outside the loop in their DF list. If so,
789 // replace it with CondBB.
790 DomTreeNode *Node = DT->getNode(ActiveBB);
791 for (df_iterator<DomTreeNode *> DI = df_begin(Node), DE = df_end(Node);
793 BasicBlock *BB = DI->getBlock();
794 DominanceFrontier::iterator BBDF = DF->find(BB);
795 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
796 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
797 while (DomSetI != DomSetE) {
798 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
800 BasicBlock *DFBB = *CurrentItr;
801 if (DFBB == ExitingBB || !L->contains(DFBB)) {
802 BBDF->second.erase(DFBB);
803 BBDF->second.insert(CondBB);
809 /// updatePHINodes - CFG has been changed.
811 /// - ExitBB's single predecessor was Latch
812 /// - Latch's second successor was Header
814 /// - ExitBB's single predecessor is Header
815 /// - Latch's one and only successor is Header
817 /// Update ExitBB PHINodes' to reflect this change.
818 void LoopIndexSplit::updatePHINodes(BasicBlock *ExitBB, BasicBlock *Latch,
820 PHINode *IV, Instruction *IVIncrement,
823 for (BasicBlock::iterator BI = ExitBB->begin(), BE = ExitBB->end();
825 PHINode *PN = dyn_cast<PHINode>(BI);
830 Value *V = PN->getIncomingValueForBlock(Latch);
831 if (PHINode *PHV = dyn_cast<PHINode>(V)) {
832 // PHV is in Latch. PHV has one use is in ExitBB PHINode. And one use
833 // in Header which is new incoming value for PN.
835 for (Value::use_iterator UI = PHV->use_begin(), E = PHV->use_end();
837 if (PHINode *U = dyn_cast<PHINode>(*UI))
838 if (LP->contains(U->getParent())) {
843 // Add incoming value from header only if PN has any use inside the loop.
845 PN->addIncoming(NewV, Header);
847 } else if (Instruction *PHI = dyn_cast<Instruction>(V)) {
848 // If this instruction is IVIncrement then IV is new incoming value
849 // from header otherwise this instruction must be incoming value from
850 // header because loop is in LCSSA form.
851 if (PHI == IVIncrement)
852 PN->addIncoming(IV, Header);
854 PN->addIncoming(V, Header);
856 // Otherwise this is an incoming value from header because loop is in
858 PN->addIncoming(V, Header);
860 // Remove incoming value from Latch.
861 PN->removeIncomingValue(Latch);
865 bool LoopIndexSplit::splitLoop() {
866 SplitCondition = NULL;
867 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
868 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
870 BasicBlock *Header = L->getHeader();
871 BasicBlock *Latch = L->getLoopLatch();
872 BranchInst *SBR = NULL; // Split Condition Branch
873 BranchInst *EBR = cast<BranchInst>(ExitCondition->getParent()->getTerminator());
874 // If Exiting block includes loop variant instructions then this
875 // loop may not be split safely.
876 BasicBlock *ExitingBlock = ExitCondition->getParent();
877 if (!cleanBlock(ExitingBlock)) return false;
879 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
881 BranchInst *BR = dyn_cast<BranchInst>((*I)->getTerminator());
882 if (!BR || BR->isUnconditional()) continue;
883 ICmpInst *CI = dyn_cast<ICmpInst>(BR->getCondition());
884 if (!CI || CI == ExitCondition
885 || CI->getPredicate() == ICmpInst::ICMP_NE
886 || CI->getPredicate() == ICmpInst::ICMP_EQ)
889 // Unable to handle triangle loops at the moment.
890 // In triangle loop, split condition is in header and one of the
891 // the split destination is loop latch. If split condition is EQ
892 // then such loops are already handle in processOneIterationLoop().
894 && (Latch == BR->getSuccessor(0) || Latch == BR->getSuccessor(1)))
897 // If the block does not dominate the latch then this is not a diamond.
898 // Such loop may not benefit from index split.
899 if (!DT->dominates((*I), Latch))
902 // If split condition branches heads do not have single predecessor,
903 // SplitCondBlock, then is not possible to remove inactive branch.
904 if (!BR->getSuccessor(0)->getSinglePredecessor()
905 || !BR->getSuccessor(1)->getSinglePredecessor())
908 // If the merge point for BR is not loop latch then skip this condition.
909 if (BR->getSuccessor(0) != Latch) {
910 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
911 assert (DF0 != DF->end() && "Unable to find dominance frontier");
912 if (!DF0->second.count(Latch))
916 if (BR->getSuccessor(1) != Latch) {
917 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
918 assert (DF1 != DF->end() && "Unable to find dominance frontier");
919 if (!DF1->second.count(Latch))
930 // If the predicate sign does not match then skip.
931 if (ExitCondition->isSignedPredicate() != SplitCondition->isSignedPredicate())
934 unsigned EVOpNum = (ExitCondition->getOperand(1) == IVExitValue);
935 unsigned SVOpNum = IVBasedValues.count(SplitCondition->getOperand(0));
936 Value *SplitValue = SplitCondition->getOperand(SVOpNum);
937 if (!L->isLoopInvariant(SplitValue))
939 if (!IVBasedValues.count(SplitCondition->getOperand(!SVOpNum)))
942 // Normalize loop conditions so that it is easier to calculate new loop
944 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
945 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
946 BasicBlock *T = EBR->getSuccessor(0);
947 EBR->setSuccessor(0, EBR->getSuccessor(1));
948 EBR->setSuccessor(1, T);
951 if (IVisGT(*SplitCondition) || IVisGE(*SplitCondition)) {
952 SplitCondition->setPredicate(SplitCondition->getInversePredicate());
953 BasicBlock *T = SBR->getSuccessor(0);
954 SBR->setSuccessor(0, SBR->getSuccessor(1));
955 SBR->setSuccessor(1, T);
958 //[*] Calculate new loop bounds.
959 Value *AEV = SplitValue;
960 Value *BSV = SplitValue;
961 bool Sign = SplitCondition->isSignedPredicate();
962 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
964 if (IVisLT(*ExitCondition)) {
965 if (IVisLT(*SplitCondition)) {
968 else if (IVisLE(*SplitCondition)) {
969 AEV = getPlusOne(SplitValue, Sign, PHTerm);
970 BSV = getPlusOne(SplitValue, Sign, PHTerm);
972 assert (0 && "Unexpected split condition!");
975 else if (IVisLE(*ExitCondition)) {
976 if (IVisLT(*SplitCondition)) {
977 AEV = getMinusOne(SplitValue, Sign, PHTerm);
979 else if (IVisLE(*SplitCondition)) {
980 BSV = getPlusOne(SplitValue, Sign, PHTerm);
982 assert (0 && "Unexpected split condition!");
985 assert (0 && "Unexpected exit condition!");
987 AEV = getMin(AEV, IVExitValue, Sign, PHTerm);
988 BSV = getMax(BSV, IVStartValue, Sign, PHTerm);
991 DenseMap<const Value *, Value *> ValueMap;
992 Loop *BLoop = CloneLoop(L, LPM, LI, ValueMap, this);
995 // [*] ALoop's exiting edge enters BLoop's header.
996 // ALoop's original exit block becomes BLoop's exit block.
997 PHINode *B_IndVar = cast<PHINode>(ValueMap[IndVar]);
998 BasicBlock *A_ExitingBlock = ExitCondition->getParent();
999 BranchInst *A_ExitInsn =
1000 dyn_cast<BranchInst>(A_ExitingBlock->getTerminator());
1001 assert (A_ExitInsn && "Unable to find suitable loop exit branch");
1002 BasicBlock *B_ExitBlock = A_ExitInsn->getSuccessor(1);
1003 BasicBlock *B_Header = BLoop->getHeader();
1004 if (ALoop->contains(B_ExitBlock)) {
1005 B_ExitBlock = A_ExitInsn->getSuccessor(0);
1006 A_ExitInsn->setSuccessor(0, B_Header);
1008 A_ExitInsn->setSuccessor(1, B_Header);
1010 // [*] Update ALoop's exit value using new exit value.
1011 ExitCondition->setOperand(EVOpNum, AEV);
1013 // [*] Update BLoop's header phi nodes. Remove incoming PHINode's from
1014 // original loop's preheader. Add incoming PHINode values from
1015 // ALoop's exiting block. Update BLoop header's domiantor info.
1017 // Collect inverse map of Header PHINodes.
1018 DenseMap<Value *, Value *> InverseMap;
1019 for (BasicBlock::iterator BI = ALoop->getHeader()->begin(),
1020 BE = ALoop->getHeader()->end(); BI != BE; ++BI) {
1021 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1022 PHINode *PNClone = cast<PHINode>(ValueMap[PN]);
1023 InverseMap[PNClone] = PN;
1028 BasicBlock *A_Preheader = ALoop->getLoopPreheader();
1029 for (BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1031 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1032 // Remove incoming value from original preheader.
1033 PN->removeIncomingValue(A_Preheader);
1035 // Add incoming value from A_ExitingBlock.
1037 PN->addIncoming(BSV, A_ExitingBlock);
1039 PHINode *OrigPN = cast<PHINode>(InverseMap[PN]);
1041 // If loop header is also loop exiting block then
1042 // OrigPN is incoming value for B loop header.
1043 if (A_ExitingBlock == ALoop->getHeader())
1046 V2 = OrigPN->getIncomingValueForBlock(A_ExitingBlock);
1047 PN->addIncoming(V2, A_ExitingBlock);
1053 DT->changeImmediateDominator(B_Header, A_ExitingBlock);
1054 DF->changeImmediateDominator(B_Header, A_ExitingBlock, DT);
1056 // [*] Update BLoop's exit block. Its new predecessor is BLoop's exit
1057 // block. Remove incoming PHINode values from ALoop's exiting block.
1058 // Add new incoming values from BLoop's incoming exiting value.
1059 // Update BLoop exit block's dominator info..
1060 BasicBlock *B_ExitingBlock = cast<BasicBlock>(ValueMap[A_ExitingBlock]);
1061 for (BasicBlock::iterator BI = B_ExitBlock->begin(), BE = B_ExitBlock->end();
1063 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1064 PN->addIncoming(ValueMap[PN->getIncomingValueForBlock(A_ExitingBlock)],
1066 PN->removeIncomingValue(A_ExitingBlock);
1071 DT->changeImmediateDominator(B_ExitBlock, B_ExitingBlock);
1072 DF->changeImmediateDominator(B_ExitBlock, B_ExitingBlock, DT);
1074 //[*] Split ALoop's exit edge. This creates a new block which
1075 // serves two purposes. First one is to hold PHINode defnitions
1076 // to ensure that ALoop's LCSSA form. Second use it to act
1077 // as a preheader for BLoop.
1078 BasicBlock *A_ExitBlock = SplitEdge(A_ExitingBlock, B_Header, this);
1080 //[*] Preserve ALoop's LCSSA form. Create new forwarding PHINodes
1081 // in A_ExitBlock to redefine outgoing PHI definitions from ALoop.
1082 for(BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1084 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1085 Value *V1 = PN->getIncomingValueForBlock(A_ExitBlock);
1086 PHINode *newPHI = PHINode::Create(PN->getType(), PN->getName());
1087 newPHI->addIncoming(V1, A_ExitingBlock);
1088 A_ExitBlock->getInstList().push_front(newPHI);
1089 PN->removeIncomingValue(A_ExitBlock);
1090 PN->addIncoming(newPHI, A_ExitBlock);
1095 //[*] Eliminate split condition's inactive branch from ALoop.
1096 BasicBlock *A_SplitCondBlock = SplitCondition->getParent();
1097 BranchInst *A_BR = cast<BranchInst>(A_SplitCondBlock->getTerminator());
1098 BasicBlock *A_InactiveBranch = NULL;
1099 BasicBlock *A_ActiveBranch = NULL;
1100 A_ActiveBranch = A_BR->getSuccessor(0);
1101 A_InactiveBranch = A_BR->getSuccessor(1);
1102 A_BR->setUnconditionalDest(A_ActiveBranch);
1103 removeBlocks(A_InactiveBranch, L, A_ActiveBranch);
1105 //[*] Eliminate split condition's inactive branch in from BLoop.
1106 BasicBlock *B_SplitCondBlock = cast<BasicBlock>(ValueMap[A_SplitCondBlock]);
1107 BranchInst *B_BR = cast<BranchInst>(B_SplitCondBlock->getTerminator());
1108 BasicBlock *B_InactiveBranch = NULL;
1109 BasicBlock *B_ActiveBranch = NULL;
1110 B_ActiveBranch = B_BR->getSuccessor(1);
1111 B_InactiveBranch = B_BR->getSuccessor(0);
1112 B_BR->setUnconditionalDest(B_ActiveBranch);
1113 removeBlocks(B_InactiveBranch, BLoop, B_ActiveBranch);
1115 BasicBlock *A_Header = ALoop->getHeader();
1116 if (A_ExitingBlock == A_Header)
1119 //[*] Move exit condition into split condition block to avoid
1120 // executing dead loop iteration.
1121 ICmpInst *B_ExitCondition = cast<ICmpInst>(ValueMap[ExitCondition]);
1122 Instruction *B_IndVarIncrement = cast<Instruction>(ValueMap[IVIncrement]);
1123 ICmpInst *B_SplitCondition = cast<ICmpInst>(ValueMap[SplitCondition]);
1125 moveExitCondition(A_SplitCondBlock, A_ActiveBranch, A_ExitBlock, ExitCondition,
1126 cast<ICmpInst>(SplitCondition), IndVar, IVIncrement,
1129 moveExitCondition(B_SplitCondBlock, B_ActiveBranch,
1130 B_ExitBlock, B_ExitCondition,
1131 B_SplitCondition, B_IndVar, B_IndVarIncrement,
1138 /// cleanBlock - A block is considered clean if all non terminal instructions
1139 /// are either, PHINodes, IV based.
1140 bool LoopIndexSplit::cleanBlock(BasicBlock *BB) {
1141 Instruction *Terminator = BB->getTerminator();
1142 for(BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1144 Instruction *I = BI;
1146 if (isa<PHINode>(I) || I == Terminator || I == ExitCondition
1147 || I == SplitCondition || IVBasedValues.count(I)
1148 || isa<DbgInfoIntrinsic>(I))
1151 if (I->mayWriteToMemory())
1154 // I is used only inside this block then it is OK.
1155 bool usedOutsideBB = false;
1156 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1158 Instruction *U = cast<Instruction>(UI);
1159 if (U->getParent() != BB)
1160 usedOutsideBB = true;
1165 // Otherwise we have a instruction that may not allow loop spliting.
1171 /// IVisLT - If Op is comparing IV based value with an loop invariant and
1172 /// IV based value is less than the loop invariant then return the loop
1173 /// invariant. Otherwise return NULL.
1174 Value * LoopIndexSplit::IVisLT(ICmpInst &Op) {
1175 ICmpInst::Predicate P = Op.getPredicate();
1176 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1177 && IVBasedValues.count(Op.getOperand(0))
1178 && L->isLoopInvariant(Op.getOperand(1)))
1179 return Op.getOperand(1);
1181 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1182 && IVBasedValues.count(Op.getOperand(1))
1183 && L->isLoopInvariant(Op.getOperand(0)))
1184 return Op.getOperand(0);
1189 /// IVisLE - If Op is comparing IV based value with an loop invariant and
1190 /// IV based value is less than or equal to the loop invariant then
1191 /// return the loop invariant. Otherwise return NULL.
1192 Value * LoopIndexSplit::IVisLE(ICmpInst &Op) {
1193 ICmpInst::Predicate P = Op.getPredicate();
1194 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1195 && IVBasedValues.count(Op.getOperand(0))
1196 && L->isLoopInvariant(Op.getOperand(1)))
1197 return Op.getOperand(1);
1199 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1200 && IVBasedValues.count(Op.getOperand(1))
1201 && L->isLoopInvariant(Op.getOperand(0)))
1202 return Op.getOperand(0);
1207 /// IVisGT - If Op is comparing IV based value with an loop invariant and
1208 /// IV based value is greater than the loop invariant then return the loop
1209 /// invariant. Otherwise return NULL.
1210 Value * LoopIndexSplit::IVisGT(ICmpInst &Op) {
1211 ICmpInst::Predicate P = Op.getPredicate();
1212 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1213 && IVBasedValues.count(Op.getOperand(0))
1214 && L->isLoopInvariant(Op.getOperand(1)))
1215 return Op.getOperand(1);
1217 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1218 && IVBasedValues.count(Op.getOperand(1))
1219 && L->isLoopInvariant(Op.getOperand(0)))
1220 return Op.getOperand(0);
1225 /// IVisGE - If Op is comparing IV based value with an loop invariant and
1226 /// IV based value is greater than or equal to the loop invariant then
1227 /// return the loop invariant. Otherwise return NULL.
1228 Value * LoopIndexSplit::IVisGE(ICmpInst &Op) {
1229 ICmpInst::Predicate P = Op.getPredicate();
1230 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1231 && IVBasedValues.count(Op.getOperand(0))
1232 && L->isLoopInvariant(Op.getOperand(1)))
1233 return Op.getOperand(1);
1235 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1236 && IVBasedValues.count(Op.getOperand(1))
1237 && L->isLoopInvariant(Op.getOperand(0)))
1238 return Op.getOperand(0);