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/LLVMContext.h"
58 #include "llvm/Analysis/LoopPass.h"
59 #include "llvm/Analysis/ScalarEvolution.h"
60 #include "llvm/Analysis/Dominators.h"
61 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
62 #include "llvm/Transforms/Utils/Cloning.h"
63 #include "llvm/Transforms/Utils/Local.h"
64 #include "llvm/Support/Compiler.h"
65 #include "llvm/ADT/DepthFirstIterator.h"
66 #include "llvm/ADT/Statistic.h"
70 STATISTIC(NumIndexSplit, "Number of loop index split");
71 STATISTIC(NumIndexSplitRemoved, "Number of loops eliminated by loop index split");
72 STATISTIC(NumRestrictBounds, "Number of loop iteration space restricted");
76 class VISIBILITY_HIDDEN LoopIndexSplit : public LoopPass {
79 static char ID; // Pass ID, replacement for typeid
80 LoopIndexSplit() : LoopPass(&ID) {}
82 // Index split Loop L. Return true if loop is split.
83 bool runOnLoop(Loop *L, LPPassManager &LPM);
85 void getAnalysisUsage(AnalysisUsage &AU) const {
86 AU.addPreserved<ScalarEvolution>();
87 AU.addRequiredID(LCSSAID);
88 AU.addPreservedID(LCSSAID);
89 AU.addRequired<LoopInfo>();
90 AU.addPreserved<LoopInfo>();
91 AU.addRequiredID(LoopSimplifyID);
92 AU.addPreservedID(LoopSimplifyID);
93 AU.addRequired<DominatorTree>();
94 AU.addRequired<DominanceFrontier>();
95 AU.addPreserved<DominatorTree>();
96 AU.addPreserved<DominanceFrontier>();
100 /// processOneIterationLoop -- Eliminate loop if loop body is executed
101 /// only once. For example,
102 /// for (i = 0; i < N; ++i) {
108 bool processOneIterationLoop();
110 // -- Routines used by updateLoopIterationSpace();
112 /// updateLoopIterationSpace -- Update loop's iteration space if loop
113 /// body is executed for certain IV range only. For example,
115 /// for (i = 0; i < N; ++i) {
116 /// if ( i > A && i < B) {
120 /// is transformed to iterators from A to B, if A > 0 and B < N.
122 bool updateLoopIterationSpace();
124 /// restrictLoopBound - Op dominates loop body. Op compares an IV based value
125 /// with a loop invariant value. Update loop's lower and upper bound based on
126 /// the loop invariant value.
127 bool restrictLoopBound(ICmpInst &Op);
129 // --- Routines used by splitLoop(). --- /
133 /// removeBlocks - Remove basic block DeadBB and all blocks dominated by
134 /// DeadBB. This routine is used to remove split condition's dead branch,
135 /// dominated by DeadBB. LiveBB dominates split conidition's other branch.
136 void removeBlocks(BasicBlock *DeadBB, Loop *LP, BasicBlock *LiveBB);
138 /// moveExitCondition - Move exit condition EC into split condition block.
139 void moveExitCondition(BasicBlock *CondBB, BasicBlock *ActiveBB,
140 BasicBlock *ExitBB, ICmpInst *EC, ICmpInst *SC,
141 PHINode *IV, Instruction *IVAdd, Loop *LP,
144 /// updatePHINodes - CFG has been changed.
146 /// - ExitBB's single predecessor was Latch
147 /// - Latch's second successor was Header
149 /// - ExitBB's single predecessor was Header
150 /// - Latch's one and only successor was Header
152 /// Update ExitBB PHINodes' to reflect this change.
153 void updatePHINodes(BasicBlock *ExitBB, BasicBlock *Latch,
155 PHINode *IV, Instruction *IVIncrement, Loop *LP);
157 // --- Utility routines --- /
159 /// cleanBlock - A block is considered clean if all non terminal
160 /// instructions are either PHINodes or IV based values.
161 bool cleanBlock(BasicBlock *BB);
163 /// IVisLT - If Op is comparing IV based value with an loop invariant and
164 /// IV based value is less than the loop invariant then return the loop
165 /// invariant. Otherwise return NULL.
166 Value * IVisLT(ICmpInst &Op);
168 /// IVisLE - If Op is comparing IV based value with an loop invariant and
169 /// IV based value is less than or equal to the loop invariant then
170 /// return the loop invariant. Otherwise return NULL.
171 Value * IVisLE(ICmpInst &Op);
173 /// IVisGT - If Op is comparing IV based value with an loop invariant and
174 /// IV based value is greater than the loop invariant then return the loop
175 /// invariant. Otherwise return NULL.
176 Value * IVisGT(ICmpInst &Op);
178 /// IVisGE - If Op is comparing IV based value with an loop invariant and
179 /// IV based value is greater than or equal to the loop invariant then
180 /// return the loop invariant. Otherwise return NULL.
181 Value * IVisGE(ICmpInst &Op);
185 // Current Loop information.
190 DominanceFrontier *DF;
193 ICmpInst *ExitCondition;
194 ICmpInst *SplitCondition;
197 Instruction *IVIncrement;
198 SmallPtrSet<Value *, 4> IVBasedValues;
202 char LoopIndexSplit::ID = 0;
203 static RegisterPass<LoopIndexSplit>
204 X("loop-index-split", "Index Split Loops");
206 Pass *llvm::createLoopIndexSplitPass() {
207 return new LoopIndexSplit();
210 // Index split Loop L. Return true if loop is split.
211 bool LoopIndexSplit::runOnLoop(Loop *IncomingLoop, LPPassManager &LPM_Ref) {
215 // FIXME - Nested loops make dominator info updates tricky.
216 if (!L->getSubLoops().empty())
219 DT = &getAnalysis<DominatorTree>();
220 LI = &getAnalysis<LoopInfo>();
221 DF = &getAnalysis<DominanceFrontier>();
223 // Initialize loop data.
224 IndVar = L->getCanonicalInductionVariable();
225 if (!IndVar) return false;
227 bool P1InLoop = L->contains(IndVar->getIncomingBlock(1));
228 IVStartValue = IndVar->getIncomingValue(!P1InLoop);
229 IVIncrement = dyn_cast<Instruction>(IndVar->getIncomingValue(P1InLoop));
230 if (!IVIncrement) return false;
232 IVBasedValues.clear();
233 IVBasedValues.insert(IndVar);
234 IVBasedValues.insert(IVIncrement);
235 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
237 for(BasicBlock::iterator BI = (*I)->begin(), BE = (*I)->end();
239 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BI))
240 if (BO != IVIncrement
241 && (BO->getOpcode() == Instruction::Add
242 || BO->getOpcode() == Instruction::Sub))
243 if (IVBasedValues.count(BO->getOperand(0))
244 && L->isLoopInvariant(BO->getOperand(1)))
245 IVBasedValues.insert(BO);
248 // Reject loop if loop exit condition is not suitable.
249 BasicBlock *ExitingBlock = L->getExitingBlock();
252 BranchInst *EBR = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
253 if (!EBR) return false;
254 ExitCondition = dyn_cast<ICmpInst>(EBR->getCondition());
255 if (!ExitCondition) return false;
256 if (ExitingBlock != L->getLoopLatch()) return false;
257 IVExitValue = ExitCondition->getOperand(1);
258 if (!L->isLoopInvariant(IVExitValue))
259 IVExitValue = ExitCondition->getOperand(0);
260 if (!L->isLoopInvariant(IVExitValue))
262 if (!IVBasedValues.count(
263 ExitCondition->getOperand(IVExitValue == ExitCondition->getOperand(0))))
266 // If start value is more then exit value where induction variable
267 // increments by 1 then we are potentially dealing with an infinite loop.
268 // Do not index split this loop.
269 if (ConstantInt *SV = dyn_cast<ConstantInt>(IVStartValue))
270 if (ConstantInt *EV = dyn_cast<ConstantInt>(IVExitValue))
271 if (SV->getSExtValue() > EV->getSExtValue())
274 if (processOneIterationLoop())
277 if (updateLoopIterationSpace())
286 // --- Helper routines ---
287 // isUsedOutsideLoop - Returns true iff V is used outside the loop L.
288 static bool isUsedOutsideLoop(Value *V, Loop *L) {
289 for(Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
290 if (!L->contains(cast<Instruction>(*UI)->getParent()))
296 static Value *getPlusOne(Value *V, bool Sign, Instruction *InsertPt,
297 LLVMContext &Context) {
298 Constant *One = ConstantInt::get(V->getType(), 1, Sign);
299 return BinaryOperator::CreateAdd(V, One, "lsp", InsertPt);
303 static Value *getMinusOne(Value *V, bool Sign, Instruction *InsertPt,
304 LLVMContext &Context) {
305 Constant *One = ConstantInt::get(V->getType(), 1, Sign);
306 return BinaryOperator::CreateSub(V, One, "lsp", InsertPt);
309 // Return min(V1, V1)
310 static Value *getMin(Value *V1, Value *V2, bool Sign, Instruction *InsertPt) {
312 Value *C = new ICmpInst(InsertPt,
313 Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
315 return SelectInst::Create(C, V1, V2, "lsp", InsertPt);
318 // Return max(V1, V2)
319 static Value *getMax(Value *V1, Value *V2, bool Sign, Instruction *InsertPt) {
321 Value *C = new ICmpInst(InsertPt,
322 Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
324 return SelectInst::Create(C, V2, V1, "lsp", InsertPt);
327 /// processOneIterationLoop -- Eliminate loop if loop body is executed
328 /// only once. For example,
329 /// for (i = 0; i < N; ++i) {
335 bool LoopIndexSplit::processOneIterationLoop() {
336 SplitCondition = NULL;
337 BasicBlock *Latch = L->getLoopLatch();
338 BasicBlock *Header = L->getHeader();
339 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
340 if (!BR) return false;
341 if (!isa<BranchInst>(Latch->getTerminator())) return false;
342 if (BR->isUnconditional()) return false;
343 SplitCondition = dyn_cast<ICmpInst>(BR->getCondition());
344 if (!SplitCondition) return false;
345 if (SplitCondition == ExitCondition) return false;
346 if (SplitCondition->getPredicate() != ICmpInst::ICMP_EQ) return false;
347 if (BR->getOperand(1) != Latch) return false;
348 if (!IVBasedValues.count(SplitCondition->getOperand(0))
349 && !IVBasedValues.count(SplitCondition->getOperand(1)))
352 // If IV is used outside the loop then this loop traversal is required.
353 // FIXME: Calculate and use last IV value.
354 if (isUsedOutsideLoop(IVIncrement, L))
357 // If BR operands are not IV or not loop invariants then skip this loop.
358 Value *OPV = SplitCondition->getOperand(0);
359 Value *SplitValue = SplitCondition->getOperand(1);
360 if (!L->isLoopInvariant(SplitValue))
361 std::swap(OPV, SplitValue);
362 if (!L->isLoopInvariant(SplitValue))
364 Instruction *OPI = dyn_cast<Instruction>(OPV);
367 if (OPI->getParent() != Header || isUsedOutsideLoop(OPI, L))
369 Value *StartValue = IVStartValue;
370 Value *ExitValue = IVExitValue;;
373 // If BR operand is IV based then use this operand to calculate
374 // effective conditions for loop body.
375 BinaryOperator *BOPV = dyn_cast<BinaryOperator>(OPV);
378 if (BOPV->getOpcode() != Instruction::Add)
380 StartValue = BinaryOperator::CreateAdd(OPV, StartValue, "" , BR);
381 ExitValue = BinaryOperator::CreateAdd(OPV, ExitValue, "" , BR);
384 if (!cleanBlock(Header))
387 if (!cleanBlock(Latch))
390 // If the merge point for BR is not loop latch then skip this loop.
391 if (BR->getSuccessor(0) != Latch) {
392 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
393 assert (DF0 != DF->end() && "Unable to find dominance frontier");
394 if (!DF0->second.count(Latch))
398 if (BR->getSuccessor(1) != Latch) {
399 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
400 assert (DF1 != DF->end() && "Unable to find dominance frontier");
401 if (!DF1->second.count(Latch))
405 // Now, Current loop L contains compare instruction
406 // that compares induction variable, IndVar, against loop invariant. And
407 // entire (i.e. meaningful) loop body is dominated by this compare
408 // instruction. In such case eliminate
409 // loop structure surrounding this loop body. For example,
410 // for (int i = start; i < end; ++i) {
411 // if ( i == somevalue) {
415 // can be transformed into
416 // if (somevalue >= start && somevalue < end) {
421 // Replace index variable with split value in loop body. Loop body is executed
422 // only when index variable is equal to split value.
423 IndVar->replaceAllUsesWith(SplitValue);
425 // Replace split condition in header.
427 // SplitCondition : icmp eq i32 IndVar, SplitValue
429 // c1 = icmp uge i32 SplitValue, StartValue
430 // c2 = icmp ult i32 SplitValue, ExitValue
432 Instruction *C1 = new ICmpInst(BR, ExitCondition->isSignedPredicate() ?
433 ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
434 SplitValue, StartValue, "lisplit");
436 CmpInst::Predicate C2P = ExitCondition->getPredicate();
437 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
438 if (LatchBR->getOperand(1) != Header)
439 C2P = CmpInst::getInversePredicate(C2P);
440 Instruction *C2 = new ICmpInst(BR, C2P, SplitValue, ExitValue, "lisplit");
441 Instruction *NSplitCond = BinaryOperator::CreateAnd(C1, C2, "lisplit", BR);
443 SplitCondition->replaceAllUsesWith(NSplitCond);
444 SplitCondition->eraseFromParent();
446 // Remove Latch to Header edge.
447 BasicBlock *LatchSucc = NULL;
448 Header->removePredecessor(Latch);
449 for (succ_iterator SI = succ_begin(Latch), E = succ_end(Latch);
455 // Clean up latch block.
456 Value *LatchBRCond = LatchBR->getCondition();
457 LatchBR->setUnconditionalDest(LatchSucc);
458 RecursivelyDeleteTriviallyDeadInstructions(LatchBRCond);
460 LPM->deleteLoopFromQueue(L);
462 // Update Dominator Info.
463 // Only CFG change done is to remove Latch to Header edge. This
464 // does not change dominator tree because Latch did not dominate
467 DominanceFrontier::iterator HeaderDF = DF->find(Header);
468 if (HeaderDF != DF->end())
469 DF->removeFromFrontier(HeaderDF, Header);
471 DominanceFrontier::iterator LatchDF = DF->find(Latch);
472 if (LatchDF != DF->end())
473 DF->removeFromFrontier(LatchDF, Header);
476 ++NumIndexSplitRemoved;
480 /// restrictLoopBound - Op dominates loop body. Op compares an IV based value
481 /// with a loop invariant value. Update loop's lower and upper bound based on
482 /// the loop invariant value.
483 bool LoopIndexSplit::restrictLoopBound(ICmpInst &Op) {
484 bool Sign = Op.isSignedPredicate();
485 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
487 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
489 cast<BranchInst>(ExitCondition->getParent()->getTerminator());
490 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
491 BasicBlock *T = EBR->getSuccessor(0);
492 EBR->setSuccessor(0, EBR->getSuccessor(1));
493 EBR->setSuccessor(1, T);
496 LLVMContext &Context = Op.getContext();
498 // New upper and lower bounds.
501 if (Value *V = IVisLT(Op)) {
502 // Restrict upper bound.
503 if (IVisLE(*ExitCondition))
504 V = getMinusOne(V, Sign, PHTerm, Context);
505 NUB = getMin(V, IVExitValue, Sign, PHTerm);
506 } else if (Value *V = IVisLE(Op)) {
507 // Restrict upper bound.
508 if (IVisLT(*ExitCondition))
509 V = getPlusOne(V, Sign, PHTerm, Context);
510 NUB = getMin(V, IVExitValue, Sign, PHTerm);
511 } else if (Value *V = IVisGT(Op)) {
512 // Restrict lower bound.
513 V = getPlusOne(V, Sign, PHTerm, Context);
514 NLB = getMax(V, IVStartValue, Sign, PHTerm);
515 } else if (Value *V = IVisGE(Op))
516 // Restrict lower bound.
517 NLB = getMax(V, IVStartValue, Sign, PHTerm);
523 unsigned i = IndVar->getBasicBlockIndex(L->getLoopPreheader());
524 IndVar->setIncomingValue(i, NLB);
528 unsigned i = (ExitCondition->getOperand(0) != IVExitValue);
529 ExitCondition->setOperand(i, NUB);
534 /// updateLoopIterationSpace -- Update loop's iteration space if loop
535 /// body is executed for certain IV range only. For example,
537 /// for (i = 0; i < N; ++i) {
538 /// if ( i > A && i < B) {
542 /// is transformed to iterators from A to B, if A > 0 and B < N.
544 bool LoopIndexSplit::updateLoopIterationSpace() {
545 SplitCondition = NULL;
546 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
547 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
549 BasicBlock *Latch = L->getLoopLatch();
550 BasicBlock *Header = L->getHeader();
551 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
552 if (!BR) return false;
553 if (!isa<BranchInst>(Latch->getTerminator())) return false;
554 if (BR->isUnconditional()) return false;
555 BinaryOperator *AND = dyn_cast<BinaryOperator>(BR->getCondition());
556 if (!AND) return false;
557 if (AND->getOpcode() != Instruction::And) return false;
558 ICmpInst *Op0 = dyn_cast<ICmpInst>(AND->getOperand(0));
559 ICmpInst *Op1 = dyn_cast<ICmpInst>(AND->getOperand(1));
562 IVBasedValues.insert(AND);
563 IVBasedValues.insert(Op0);
564 IVBasedValues.insert(Op1);
565 if (!cleanBlock(Header)) return false;
566 BasicBlock *ExitingBlock = ExitCondition->getParent();
567 if (!cleanBlock(ExitingBlock)) return false;
569 // If the merge point for BR is not loop latch then skip this loop.
570 if (BR->getSuccessor(0) != Latch) {
571 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
572 assert (DF0 != DF->end() && "Unable to find dominance frontier");
573 if (!DF0->second.count(Latch))
577 if (BR->getSuccessor(1) != Latch) {
578 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
579 assert (DF1 != DF->end() && "Unable to find dominance frontier");
580 if (!DF1->second.count(Latch))
584 // Verify that loop exiting block has only two predecessor, where one pred
585 // is split condition block. The other predecessor will become exiting block's
586 // dominator after CFG is updated. TODO : Handle CFG's where exiting block has
587 // more then two predecessors. This requires extra work in updating dominator
589 BasicBlock *ExitingBBPred = NULL;
590 for (pred_iterator PI = pred_begin(ExitingBlock), PE = pred_end(ExitingBlock);
592 BasicBlock *BB = *PI;
601 if (!restrictLoopBound(*Op0))
604 if (!restrictLoopBound(*Op1))
608 if (BR->getSuccessor(0) == ExitingBlock)
609 BR->setUnconditionalDest(BR->getSuccessor(1));
611 BR->setUnconditionalDest(BR->getSuccessor(0));
613 AND->eraseFromParent();
614 if (Op0->use_empty())
615 Op0->eraseFromParent();
616 if (Op1->use_empty())
617 Op1->eraseFromParent();
619 // Update domiantor info. Now, ExitingBlock has only one predecessor,
620 // ExitingBBPred, and it is ExitingBlock's immediate domiantor.
621 DT->changeImmediateDominator(ExitingBlock, ExitingBBPred);
623 BasicBlock *ExitBlock = ExitingBlock->getTerminator()->getSuccessor(1);
624 if (L->contains(ExitBlock))
625 ExitBlock = ExitingBlock->getTerminator()->getSuccessor(0);
627 // If ExitingBlock is a member of the loop basic blocks' DF list then
628 // replace ExitingBlock with header and exit block in the DF list
629 DominanceFrontier::iterator ExitingBlockDF = DF->find(ExitingBlock);
630 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
633 if (BB == Header || BB == ExitingBlock)
635 DominanceFrontier::iterator BBDF = DF->find(BB);
636 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
637 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
638 while (DomSetI != DomSetE) {
639 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
641 BasicBlock *DFBB = *CurrentItr;
642 if (DFBB == ExitingBlock) {
643 BBDF->second.erase(DFBB);
644 for (DominanceFrontier::DomSetType::iterator
645 EBI = ExitingBlockDF->second.begin(),
646 EBE = ExitingBlockDF->second.end(); EBI != EBE; ++EBI)
647 BBDF->second.insert(*EBI);
655 /// removeBlocks - Remove basic block DeadBB and all blocks dominated by DeadBB.
656 /// This routine is used to remove split condition's dead branch, dominated by
657 /// DeadBB. LiveBB dominates split conidition's other branch.
658 void LoopIndexSplit::removeBlocks(BasicBlock *DeadBB, Loop *LP,
659 BasicBlock *LiveBB) {
661 // First update DeadBB's dominance frontier.
662 SmallVector<BasicBlock *, 8> FrontierBBs;
663 DominanceFrontier::iterator DeadBBDF = DF->find(DeadBB);
664 if (DeadBBDF != DF->end()) {
665 SmallVector<BasicBlock *, 8> PredBlocks;
667 DominanceFrontier::DomSetType DeadBBSet = DeadBBDF->second;
668 for (DominanceFrontier::DomSetType::iterator DeadBBSetI = DeadBBSet.begin(),
669 DeadBBSetE = DeadBBSet.end(); DeadBBSetI != DeadBBSetE; ++DeadBBSetI)
671 BasicBlock *FrontierBB = *DeadBBSetI;
672 FrontierBBs.push_back(FrontierBB);
674 // Rremove any PHI incoming edge from blocks dominated by DeadBB.
676 for(pred_iterator PI = pred_begin(FrontierBB), PE = pred_end(FrontierBB);
679 if (P == DeadBB || DT->dominates(DeadBB, P))
680 PredBlocks.push_back(P);
683 for(BasicBlock::iterator FBI = FrontierBB->begin(), FBE = FrontierBB->end();
685 if (PHINode *PN = dyn_cast<PHINode>(FBI)) {
686 for(SmallVector<BasicBlock *, 8>::iterator PI = PredBlocks.begin(),
687 PE = PredBlocks.end(); PI != PE; ++PI) {
689 PN->removeIncomingValue(P);
698 // Now remove DeadBB and all nodes dominated by DeadBB in df order.
699 SmallVector<BasicBlock *, 32> WorkList;
700 DomTreeNode *DN = DT->getNode(DeadBB);
701 for (df_iterator<DomTreeNode*> DI = df_begin(DN),
702 E = df_end(DN); DI != E; ++DI) {
703 BasicBlock *BB = DI->getBlock();
704 WorkList.push_back(BB);
705 BB->replaceAllUsesWith(UndefValue::get(
706 Type::getLabelTy(DeadBB->getContext())));
709 while (!WorkList.empty()) {
710 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
711 LPM->deleteSimpleAnalysisValue(BB, LP);
712 for(BasicBlock::iterator BBI = BB->begin(), BBE = BB->end();
714 Instruction *I = BBI;
716 I->replaceAllUsesWith(UndefValue::get(I->getType()));
717 LPM->deleteSimpleAnalysisValue(I, LP);
718 I->eraseFromParent();
723 BB->eraseFromParent();
726 // Update Frontier BBs' dominator info.
727 while (!FrontierBBs.empty()) {
728 BasicBlock *FBB = FrontierBBs.back(); FrontierBBs.pop_back();
729 BasicBlock *NewDominator = FBB->getSinglePredecessor();
731 pred_iterator PI = pred_begin(FBB), PE = pred_end(FBB);
734 if (NewDominator != LiveBB) {
735 for(; PI != PE; ++PI) {
738 NewDominator = LiveBB;
741 NewDominator = DT->findNearestCommonDominator(NewDominator, P);
745 assert (NewDominator && "Unable to fix dominator info.");
746 DT->changeImmediateDominator(FBB, NewDominator);
747 DF->changeImmediateDominator(FBB, NewDominator, DT);
752 // moveExitCondition - Move exit condition EC into split condition block CondBB.
753 void LoopIndexSplit::moveExitCondition(BasicBlock *CondBB, BasicBlock *ActiveBB,
754 BasicBlock *ExitBB, ICmpInst *EC,
755 ICmpInst *SC, PHINode *IV,
756 Instruction *IVAdd, Loop *LP,
757 unsigned ExitValueNum) {
759 BasicBlock *ExitingBB = EC->getParent();
760 Instruction *CurrentBR = CondBB->getTerminator();
762 // Move exit condition into split condition block.
763 EC->moveBefore(CurrentBR);
764 EC->setOperand(ExitValueNum == 0 ? 1 : 0, IV);
766 // Move exiting block's branch into split condition block. Update its branch
768 BranchInst *ExitingBR = cast<BranchInst>(ExitingBB->getTerminator());
769 ExitingBR->moveBefore(CurrentBR);
770 BasicBlock *OrigDestBB = NULL;
771 if (ExitingBR->getSuccessor(0) == ExitBB) {
772 OrigDestBB = ExitingBR->getSuccessor(1);
773 ExitingBR->setSuccessor(1, ActiveBB);
776 OrigDestBB = ExitingBR->getSuccessor(0);
777 ExitingBR->setSuccessor(0, ActiveBB);
780 // Remove split condition and current split condition branch.
781 SC->eraseFromParent();
782 CurrentBR->eraseFromParent();
784 // Connect exiting block to original destination.
785 BranchInst::Create(OrigDestBB, ExitingBB);
788 updatePHINodes(ExitBB, ExitingBB, CondBB, IV, IVAdd, LP);
790 // Fix dominator info.
791 // ExitBB is now dominated by CondBB
792 DT->changeImmediateDominator(ExitBB, CondBB);
793 DF->changeImmediateDominator(ExitBB, CondBB, DT);
795 // Blocks outside the loop may have been in the dominance frontier of blocks
796 // inside the condition; this is now impossible because the blocks inside the
797 // condition no loger dominate the exit. Remove the relevant blocks from
798 // the dominance frontiers.
799 for (Loop::block_iterator I = LP->block_begin(), E = LP->block_end();
801 if (*I == CondBB || !DT->dominates(CondBB, *I)) continue;
802 DominanceFrontier::iterator BBDF = DF->find(*I);
803 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
804 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
805 while (DomSetI != DomSetE) {
806 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
808 BasicBlock *DFBB = *CurrentItr;
809 if (!LP->contains(DFBB))
810 BBDF->second.erase(DFBB);
815 /// updatePHINodes - CFG has been changed.
817 /// - ExitBB's single predecessor was Latch
818 /// - Latch's second successor was Header
820 /// - ExitBB's single predecessor is Header
821 /// - Latch's one and only successor is Header
823 /// Update ExitBB PHINodes' to reflect this change.
824 void LoopIndexSplit::updatePHINodes(BasicBlock *ExitBB, BasicBlock *Latch,
826 PHINode *IV, Instruction *IVIncrement,
829 for (BasicBlock::iterator BI = ExitBB->begin(), BE = ExitBB->end();
831 PHINode *PN = dyn_cast<PHINode>(BI);
836 Value *V = PN->getIncomingValueForBlock(Latch);
837 if (PHINode *PHV = dyn_cast<PHINode>(V)) {
838 // PHV is in Latch. PHV has one use is in ExitBB PHINode. And one use
839 // in Header which is new incoming value for PN.
841 for (Value::use_iterator UI = PHV->use_begin(), E = PHV->use_end();
843 if (PHINode *U = dyn_cast<PHINode>(*UI))
844 if (LP->contains(U->getParent())) {
849 // Add incoming value from header only if PN has any use inside the loop.
851 PN->addIncoming(NewV, Header);
853 } else if (Instruction *PHI = dyn_cast<Instruction>(V)) {
854 // If this instruction is IVIncrement then IV is new incoming value
855 // from header otherwise this instruction must be incoming value from
856 // header because loop is in LCSSA form.
857 if (PHI == IVIncrement)
858 PN->addIncoming(IV, Header);
860 PN->addIncoming(V, Header);
862 // Otherwise this is an incoming value from header because loop is in
864 PN->addIncoming(V, Header);
866 // Remove incoming value from Latch.
867 PN->removeIncomingValue(Latch);
871 bool LoopIndexSplit::splitLoop() {
872 SplitCondition = NULL;
873 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
874 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
876 BasicBlock *Header = L->getHeader();
877 BasicBlock *Latch = L->getLoopLatch();
878 BranchInst *SBR = NULL; // Split Condition Branch
879 BranchInst *EBR = cast<BranchInst>(ExitCondition->getParent()->getTerminator());
880 // If Exiting block includes loop variant instructions then this
881 // loop may not be split safely.
882 BasicBlock *ExitingBlock = ExitCondition->getParent();
883 if (!cleanBlock(ExitingBlock)) return false;
885 LLVMContext &Context = Header->getContext();
887 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
889 BranchInst *BR = dyn_cast<BranchInst>((*I)->getTerminator());
890 if (!BR || BR->isUnconditional()) continue;
891 ICmpInst *CI = dyn_cast<ICmpInst>(BR->getCondition());
892 if (!CI || CI == ExitCondition
893 || CI->getPredicate() == ICmpInst::ICMP_NE
894 || CI->getPredicate() == ICmpInst::ICMP_EQ)
897 // Unable to handle triangle loops at the moment.
898 // In triangle loop, split condition is in header and one of the
899 // the split destination is loop latch. If split condition is EQ
900 // then such loops are already handle in processOneIterationLoop().
902 && (Latch == BR->getSuccessor(0) || Latch == BR->getSuccessor(1)))
905 // If the block does not dominate the latch then this is not a diamond.
906 // Such loop may not benefit from index split.
907 if (!DT->dominates((*I), Latch))
910 // If split condition branches heads do not have single predecessor,
911 // SplitCondBlock, then is not possible to remove inactive branch.
912 if (!BR->getSuccessor(0)->getSinglePredecessor()
913 || !BR->getSuccessor(1)->getSinglePredecessor())
916 // If the merge point for BR is not loop latch then skip this condition.
917 if (BR->getSuccessor(0) != Latch) {
918 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
919 assert (DF0 != DF->end() && "Unable to find dominance frontier");
920 if (!DF0->second.count(Latch))
924 if (BR->getSuccessor(1) != Latch) {
925 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
926 assert (DF1 != DF->end() && "Unable to find dominance frontier");
927 if (!DF1->second.count(Latch))
938 // If the predicate sign does not match then skip.
939 if (ExitCondition->isSignedPredicate() != SplitCondition->isSignedPredicate())
942 unsigned EVOpNum = (ExitCondition->getOperand(1) == IVExitValue);
943 unsigned SVOpNum = IVBasedValues.count(SplitCondition->getOperand(0));
944 Value *SplitValue = SplitCondition->getOperand(SVOpNum);
945 if (!L->isLoopInvariant(SplitValue))
947 if (!IVBasedValues.count(SplitCondition->getOperand(!SVOpNum)))
950 // Normalize loop conditions so that it is easier to calculate new loop
952 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
953 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
954 BasicBlock *T = EBR->getSuccessor(0);
955 EBR->setSuccessor(0, EBR->getSuccessor(1));
956 EBR->setSuccessor(1, T);
959 if (IVisGT(*SplitCondition) || IVisGE(*SplitCondition)) {
960 SplitCondition->setPredicate(SplitCondition->getInversePredicate());
961 BasicBlock *T = SBR->getSuccessor(0);
962 SBR->setSuccessor(0, SBR->getSuccessor(1));
963 SBR->setSuccessor(1, T);
966 //[*] Calculate new loop bounds.
967 Value *AEV = SplitValue;
968 Value *BSV = SplitValue;
969 bool Sign = SplitCondition->isSignedPredicate();
970 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
972 if (IVisLT(*ExitCondition)) {
973 if (IVisLT(*SplitCondition)) {
976 else if (IVisLE(*SplitCondition)) {
977 AEV = getPlusOne(SplitValue, Sign, PHTerm, Context);
978 BSV = getPlusOne(SplitValue, Sign, PHTerm, Context);
980 assert (0 && "Unexpected split condition!");
983 else if (IVisLE(*ExitCondition)) {
984 if (IVisLT(*SplitCondition)) {
985 AEV = getMinusOne(SplitValue, Sign, PHTerm, Context);
987 else if (IVisLE(*SplitCondition)) {
988 BSV = getPlusOne(SplitValue, Sign, PHTerm, Context);
990 assert (0 && "Unexpected split condition!");
993 assert (0 && "Unexpected exit condition!");
995 AEV = getMin(AEV, IVExitValue, Sign, PHTerm);
996 BSV = getMax(BSV, IVStartValue, Sign, PHTerm);
999 DenseMap<const Value *, Value *> ValueMap;
1000 Loop *BLoop = CloneLoop(L, LPM, LI, ValueMap, this);
1003 // [*] ALoop's exiting edge enters BLoop's header.
1004 // ALoop's original exit block becomes BLoop's exit block.
1005 PHINode *B_IndVar = cast<PHINode>(ValueMap[IndVar]);
1006 BasicBlock *A_ExitingBlock = ExitCondition->getParent();
1007 BranchInst *A_ExitInsn =
1008 dyn_cast<BranchInst>(A_ExitingBlock->getTerminator());
1009 assert (A_ExitInsn && "Unable to find suitable loop exit branch");
1010 BasicBlock *B_ExitBlock = A_ExitInsn->getSuccessor(1);
1011 BasicBlock *B_Header = BLoop->getHeader();
1012 if (ALoop->contains(B_ExitBlock)) {
1013 B_ExitBlock = A_ExitInsn->getSuccessor(0);
1014 A_ExitInsn->setSuccessor(0, B_Header);
1016 A_ExitInsn->setSuccessor(1, B_Header);
1018 // [*] Update ALoop's exit value using new exit value.
1019 ExitCondition->setOperand(EVOpNum, AEV);
1021 // [*] Update BLoop's header phi nodes. Remove incoming PHINode's from
1022 // original loop's preheader. Add incoming PHINode values from
1023 // ALoop's exiting block. Update BLoop header's domiantor info.
1025 // Collect inverse map of Header PHINodes.
1026 DenseMap<Value *, Value *> InverseMap;
1027 for (BasicBlock::iterator BI = ALoop->getHeader()->begin(),
1028 BE = ALoop->getHeader()->end(); BI != BE; ++BI) {
1029 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1030 PHINode *PNClone = cast<PHINode>(ValueMap[PN]);
1031 InverseMap[PNClone] = PN;
1036 BasicBlock *A_Preheader = ALoop->getLoopPreheader();
1037 for (BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1039 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1040 // Remove incoming value from original preheader.
1041 PN->removeIncomingValue(A_Preheader);
1043 // Add incoming value from A_ExitingBlock.
1045 PN->addIncoming(BSV, A_ExitingBlock);
1047 PHINode *OrigPN = cast<PHINode>(InverseMap[PN]);
1049 // If loop header is also loop exiting block then
1050 // OrigPN is incoming value for B loop header.
1051 if (A_ExitingBlock == ALoop->getHeader())
1054 V2 = OrigPN->getIncomingValueForBlock(A_ExitingBlock);
1055 PN->addIncoming(V2, A_ExitingBlock);
1061 DT->changeImmediateDominator(B_Header, A_ExitingBlock);
1062 DF->changeImmediateDominator(B_Header, A_ExitingBlock, DT);
1064 // [*] Update BLoop's exit block. Its new predecessor is BLoop's exit
1065 // block. Remove incoming PHINode values from ALoop's exiting block.
1066 // Add new incoming values from BLoop's incoming exiting value.
1067 // Update BLoop exit block's dominator info..
1068 BasicBlock *B_ExitingBlock = cast<BasicBlock>(ValueMap[A_ExitingBlock]);
1069 for (BasicBlock::iterator BI = B_ExitBlock->begin(), BE = B_ExitBlock->end();
1071 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1072 PN->addIncoming(ValueMap[PN->getIncomingValueForBlock(A_ExitingBlock)],
1074 PN->removeIncomingValue(A_ExitingBlock);
1079 DT->changeImmediateDominator(B_ExitBlock, B_ExitingBlock);
1080 DF->changeImmediateDominator(B_ExitBlock, B_ExitingBlock, DT);
1082 //[*] Split ALoop's exit edge. This creates a new block which
1083 // serves two purposes. First one is to hold PHINode defnitions
1084 // to ensure that ALoop's LCSSA form. Second use it to act
1085 // as a preheader for BLoop.
1086 BasicBlock *A_ExitBlock = SplitEdge(A_ExitingBlock, B_Header, this);
1088 //[*] Preserve ALoop's LCSSA form. Create new forwarding PHINodes
1089 // in A_ExitBlock to redefine outgoing PHI definitions from ALoop.
1090 for(BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1092 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1093 Value *V1 = PN->getIncomingValueForBlock(A_ExitBlock);
1094 PHINode *newPHI = PHINode::Create(PN->getType(), PN->getName());
1095 newPHI->addIncoming(V1, A_ExitingBlock);
1096 A_ExitBlock->getInstList().push_front(newPHI);
1097 PN->removeIncomingValue(A_ExitBlock);
1098 PN->addIncoming(newPHI, A_ExitBlock);
1103 //[*] Eliminate split condition's inactive branch from ALoop.
1104 BasicBlock *A_SplitCondBlock = SplitCondition->getParent();
1105 BranchInst *A_BR = cast<BranchInst>(A_SplitCondBlock->getTerminator());
1106 BasicBlock *A_InactiveBranch = NULL;
1107 BasicBlock *A_ActiveBranch = NULL;
1108 A_ActiveBranch = A_BR->getSuccessor(0);
1109 A_InactiveBranch = A_BR->getSuccessor(1);
1110 A_BR->setUnconditionalDest(A_ActiveBranch);
1111 removeBlocks(A_InactiveBranch, L, A_ActiveBranch);
1113 //[*] Eliminate split condition's inactive branch in from BLoop.
1114 BasicBlock *B_SplitCondBlock = cast<BasicBlock>(ValueMap[A_SplitCondBlock]);
1115 BranchInst *B_BR = cast<BranchInst>(B_SplitCondBlock->getTerminator());
1116 BasicBlock *B_InactiveBranch = NULL;
1117 BasicBlock *B_ActiveBranch = NULL;
1118 B_ActiveBranch = B_BR->getSuccessor(1);
1119 B_InactiveBranch = B_BR->getSuccessor(0);
1120 B_BR->setUnconditionalDest(B_ActiveBranch);
1121 removeBlocks(B_InactiveBranch, BLoop, B_ActiveBranch);
1123 BasicBlock *A_Header = ALoop->getHeader();
1124 if (A_ExitingBlock == A_Header)
1127 //[*] Move exit condition into split condition block to avoid
1128 // executing dead loop iteration.
1129 ICmpInst *B_ExitCondition = cast<ICmpInst>(ValueMap[ExitCondition]);
1130 Instruction *B_IndVarIncrement = cast<Instruction>(ValueMap[IVIncrement]);
1131 ICmpInst *B_SplitCondition = cast<ICmpInst>(ValueMap[SplitCondition]);
1133 moveExitCondition(A_SplitCondBlock, A_ActiveBranch, A_ExitBlock, ExitCondition,
1134 cast<ICmpInst>(SplitCondition), IndVar, IVIncrement,
1137 moveExitCondition(B_SplitCondBlock, B_ActiveBranch,
1138 B_ExitBlock, B_ExitCondition,
1139 B_SplitCondition, B_IndVar, B_IndVarIncrement,
1146 /// cleanBlock - A block is considered clean if all non terminal instructions
1147 /// are either, PHINodes, IV based.
1148 bool LoopIndexSplit::cleanBlock(BasicBlock *BB) {
1149 Instruction *Terminator = BB->getTerminator();
1150 for(BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1152 Instruction *I = BI;
1154 if (isa<PHINode>(I) || I == Terminator || I == ExitCondition
1155 || I == SplitCondition || IVBasedValues.count(I)
1156 || isa<DbgInfoIntrinsic>(I))
1159 if (I->mayHaveSideEffects())
1162 // I is used only inside this block then it is OK.
1163 bool usedOutsideBB = false;
1164 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1166 Instruction *U = cast<Instruction>(UI);
1167 if (U->getParent() != BB)
1168 usedOutsideBB = true;
1173 // Otherwise we have a instruction that may not allow loop spliting.
1179 /// IVisLT - If Op is comparing IV based value with an loop invariant and
1180 /// IV based value is less than the loop invariant then return the loop
1181 /// invariant. Otherwise return NULL.
1182 Value * LoopIndexSplit::IVisLT(ICmpInst &Op) {
1183 ICmpInst::Predicate P = Op.getPredicate();
1184 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1185 && IVBasedValues.count(Op.getOperand(0))
1186 && L->isLoopInvariant(Op.getOperand(1)))
1187 return Op.getOperand(1);
1189 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1190 && IVBasedValues.count(Op.getOperand(1))
1191 && L->isLoopInvariant(Op.getOperand(0)))
1192 return Op.getOperand(0);
1197 /// IVisLE - If Op is comparing IV based value with an loop invariant and
1198 /// IV based value is less than or equal to the loop invariant then
1199 /// return the loop invariant. Otherwise return NULL.
1200 Value * LoopIndexSplit::IVisLE(ICmpInst &Op) {
1201 ICmpInst::Predicate P = Op.getPredicate();
1202 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1203 && IVBasedValues.count(Op.getOperand(0))
1204 && L->isLoopInvariant(Op.getOperand(1)))
1205 return Op.getOperand(1);
1207 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1208 && IVBasedValues.count(Op.getOperand(1))
1209 && L->isLoopInvariant(Op.getOperand(0)))
1210 return Op.getOperand(0);
1215 /// IVisGT - If Op is comparing IV based value with an loop invariant and
1216 /// IV based value is greater than the loop invariant then return the loop
1217 /// invariant. Otherwise return NULL.
1218 Value * LoopIndexSplit::IVisGT(ICmpInst &Op) {
1219 ICmpInst::Predicate P = Op.getPredicate();
1220 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1221 && IVBasedValues.count(Op.getOperand(0))
1222 && L->isLoopInvariant(Op.getOperand(1)))
1223 return Op.getOperand(1);
1225 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1226 && IVBasedValues.count(Op.getOperand(1))
1227 && L->isLoopInvariant(Op.getOperand(0)))
1228 return Op.getOperand(0);
1233 /// IVisGE - If Op is comparing IV based value with an loop invariant and
1234 /// IV based value is greater than or equal to the loop invariant then
1235 /// return the loop invariant. Otherwise return NULL.
1236 Value * LoopIndexSplit::IVisGE(ICmpInst &Op) {
1237 ICmpInst::Predicate P = Op.getPredicate();
1238 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1239 && IVBasedValues.count(Op.getOperand(0))
1240 && L->isLoopInvariant(Op.getOperand(1)))
1241 return Op.getOperand(1);
1243 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1244 && IVBasedValues.count(Op.getOperand(1))
1245 && L->isLoopInvariant(Op.getOperand(0)))
1246 return Op.getOperand(0);