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 = Context->getConstantInt(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 = Context->getConstantInt(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(Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
313 V1, V2, "lsp", InsertPt);
314 return SelectInst::Create(C, V1, V2, "lsp", InsertPt);
317 // Return max(V1, V2)
318 static Value *getMax(Value *V1, Value *V2, bool Sign, Instruction *InsertPt) {
320 Value *C = new ICmpInst(Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
321 V1, V2, "lsp", InsertPt);
322 return SelectInst::Create(C, V2, V1, "lsp", InsertPt);
325 /// processOneIterationLoop -- Eliminate loop if loop body is executed
326 /// only once. For example,
327 /// for (i = 0; i < N; ++i) {
333 bool LoopIndexSplit::processOneIterationLoop() {
334 SplitCondition = NULL;
335 BasicBlock *Latch = L->getLoopLatch();
336 BasicBlock *Header = L->getHeader();
337 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
338 if (!BR) return false;
339 if (!isa<BranchInst>(Latch->getTerminator())) return false;
340 if (BR->isUnconditional()) return false;
341 SplitCondition = dyn_cast<ICmpInst>(BR->getCondition());
342 if (!SplitCondition) return false;
343 if (SplitCondition == ExitCondition) return false;
344 if (SplitCondition->getPredicate() != ICmpInst::ICMP_EQ) return false;
345 if (BR->getOperand(1) != Latch) return false;
346 if (!IVBasedValues.count(SplitCondition->getOperand(0))
347 && !IVBasedValues.count(SplitCondition->getOperand(1)))
350 // If IV is used outside the loop then this loop traversal is required.
351 // FIXME: Calculate and use last IV value.
352 if (isUsedOutsideLoop(IVIncrement, L))
355 // If BR operands are not IV or not loop invariants then skip this loop.
356 Value *OPV = SplitCondition->getOperand(0);
357 Value *SplitValue = SplitCondition->getOperand(1);
358 if (!L->isLoopInvariant(SplitValue))
359 std::swap(OPV, SplitValue);
360 if (!L->isLoopInvariant(SplitValue))
362 Instruction *OPI = dyn_cast<Instruction>(OPV);
365 if (OPI->getParent() != Header || isUsedOutsideLoop(OPI, L))
367 Value *StartValue = IVStartValue;
368 Value *ExitValue = IVExitValue;;
371 // If BR operand is IV based then use this operand to calculate
372 // effective conditions for loop body.
373 BinaryOperator *BOPV = dyn_cast<BinaryOperator>(OPV);
376 if (BOPV->getOpcode() != Instruction::Add)
378 StartValue = BinaryOperator::CreateAdd(OPV, StartValue, "" , BR);
379 ExitValue = BinaryOperator::CreateAdd(OPV, ExitValue, "" , BR);
382 if (!cleanBlock(Header))
385 if (!cleanBlock(Latch))
388 // If the merge point for BR is not loop latch then skip this loop.
389 if (BR->getSuccessor(0) != Latch) {
390 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
391 assert (DF0 != DF->end() && "Unable to find dominance frontier");
392 if (!DF0->second.count(Latch))
396 if (BR->getSuccessor(1) != Latch) {
397 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
398 assert (DF1 != DF->end() && "Unable to find dominance frontier");
399 if (!DF1->second.count(Latch))
403 // Now, Current loop L contains compare instruction
404 // that compares induction variable, IndVar, against loop invariant. And
405 // entire (i.e. meaningful) loop body is dominated by this compare
406 // instruction. In such case eliminate
407 // loop structure surrounding this loop body. For example,
408 // for (int i = start; i < end; ++i) {
409 // if ( i == somevalue) {
413 // can be transformed into
414 // if (somevalue >= start && somevalue < end) {
419 // Replace index variable with split value in loop body. Loop body is executed
420 // only when index variable is equal to split value.
421 IndVar->replaceAllUsesWith(SplitValue);
423 // Replace split condition in header.
425 // SplitCondition : icmp eq i32 IndVar, SplitValue
427 // c1 = icmp uge i32 SplitValue, StartValue
428 // c2 = icmp ult i32 SplitValue, ExitValue
430 Instruction *C1 = new ICmpInst(ExitCondition->isSignedPredicate() ?
431 ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
432 SplitValue, StartValue, "lisplit", BR);
434 CmpInst::Predicate C2P = ExitCondition->getPredicate();
435 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
436 if (LatchBR->getOperand(0) != Header)
437 C2P = CmpInst::getInversePredicate(C2P);
438 Instruction *C2 = new ICmpInst(C2P, SplitValue, ExitValue, "lisplit", BR);
439 Instruction *NSplitCond = BinaryOperator::CreateAnd(C1, C2, "lisplit", BR);
441 SplitCondition->replaceAllUsesWith(NSplitCond);
442 SplitCondition->eraseFromParent();
444 // Remove Latch to Header edge.
445 BasicBlock *LatchSucc = NULL;
446 Header->removePredecessor(Latch);
447 for (succ_iterator SI = succ_begin(Latch), E = succ_end(Latch);
453 // Clean up latch block.
454 Value *LatchBRCond = LatchBR->getCondition();
455 LatchBR->setUnconditionalDest(LatchSucc);
456 RecursivelyDeleteTriviallyDeadInstructions(LatchBRCond);
458 LPM->deleteLoopFromQueue(L);
460 // Update Dominator Info.
461 // Only CFG change done is to remove Latch to Header edge. This
462 // does not change dominator tree because Latch did not dominate
465 DominanceFrontier::iterator HeaderDF = DF->find(Header);
466 if (HeaderDF != DF->end())
467 DF->removeFromFrontier(HeaderDF, Header);
469 DominanceFrontier::iterator LatchDF = DF->find(Latch);
470 if (LatchDF != DF->end())
471 DF->removeFromFrontier(LatchDF, Header);
474 ++NumIndexSplitRemoved;
478 /// restrictLoopBound - Op dominates loop body. Op compares an IV based value
479 /// with a loop invariant value. Update loop's lower and upper bound based on
480 /// the loop invariant value.
481 bool LoopIndexSplit::restrictLoopBound(ICmpInst &Op) {
482 bool Sign = Op.isSignedPredicate();
483 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
485 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
487 cast<BranchInst>(ExitCondition->getParent()->getTerminator());
488 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
489 BasicBlock *T = EBR->getSuccessor(0);
490 EBR->setSuccessor(0, EBR->getSuccessor(1));
491 EBR->setSuccessor(1, T);
494 // New upper and lower bounds.
497 if (Value *V = IVisLT(Op)) {
498 // Restrict upper bound.
499 if (IVisLE(*ExitCondition))
500 V = getMinusOne(V, Sign, PHTerm, Context);
501 NUB = getMin(V, IVExitValue, Sign, PHTerm);
502 } else if (Value *V = IVisLE(Op)) {
503 // Restrict upper bound.
504 if (IVisLT(*ExitCondition))
505 V = getPlusOne(V, Sign, PHTerm, Context);
506 NUB = getMin(V, IVExitValue, Sign, PHTerm);
507 } else if (Value *V = IVisGT(Op)) {
508 // Restrict lower bound.
509 V = getPlusOne(V, Sign, PHTerm, Context);
510 NLB = getMax(V, IVStartValue, Sign, PHTerm);
511 } else if (Value *V = IVisGE(Op))
512 // Restrict lower bound.
513 NLB = getMax(V, IVStartValue, Sign, PHTerm);
519 unsigned i = IndVar->getBasicBlockIndex(L->getLoopPreheader());
520 IndVar->setIncomingValue(i, NLB);
524 unsigned i = (ExitCondition->getOperand(0) != IVExitValue);
525 ExitCondition->setOperand(i, NUB);
530 /// updateLoopIterationSpace -- Update loop's iteration space if loop
531 /// body is executed for certain IV range only. For example,
533 /// for (i = 0; i < N; ++i) {
534 /// if ( i > A && i < B) {
538 /// is transformed to iterators from A to B, if A > 0 and B < N.
540 bool LoopIndexSplit::updateLoopIterationSpace() {
541 SplitCondition = NULL;
542 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
543 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
545 BasicBlock *Latch = L->getLoopLatch();
546 BasicBlock *Header = L->getHeader();
547 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
548 if (!BR) return false;
549 if (!isa<BranchInst>(Latch->getTerminator())) return false;
550 if (BR->isUnconditional()) return false;
551 BinaryOperator *AND = dyn_cast<BinaryOperator>(BR->getCondition());
552 if (!AND) return false;
553 if (AND->getOpcode() != Instruction::And) return false;
554 ICmpInst *Op0 = dyn_cast<ICmpInst>(AND->getOperand(0));
555 ICmpInst *Op1 = dyn_cast<ICmpInst>(AND->getOperand(1));
558 IVBasedValues.insert(AND);
559 IVBasedValues.insert(Op0);
560 IVBasedValues.insert(Op1);
561 if (!cleanBlock(Header)) return false;
562 BasicBlock *ExitingBlock = ExitCondition->getParent();
563 if (!cleanBlock(ExitingBlock)) return false;
565 // If the merge point for BR is not loop latch then skip this loop.
566 if (BR->getSuccessor(0) != Latch) {
567 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
568 assert (DF0 != DF->end() && "Unable to find dominance frontier");
569 if (!DF0->second.count(Latch))
573 if (BR->getSuccessor(1) != Latch) {
574 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
575 assert (DF1 != DF->end() && "Unable to find dominance frontier");
576 if (!DF1->second.count(Latch))
580 // Verify that loop exiting block has only two predecessor, where one pred
581 // is split condition block. The other predecessor will become exiting block's
582 // dominator after CFG is updated. TODO : Handle CFG's where exiting block has
583 // more then two predecessors. This requires extra work in updating dominator
585 BasicBlock *ExitingBBPred = NULL;
586 for (pred_iterator PI = pred_begin(ExitingBlock), PE = pred_end(ExitingBlock);
588 BasicBlock *BB = *PI;
597 if (!restrictLoopBound(*Op0))
600 if (!restrictLoopBound(*Op1))
604 if (BR->getSuccessor(0) == ExitingBlock)
605 BR->setUnconditionalDest(BR->getSuccessor(1));
607 BR->setUnconditionalDest(BR->getSuccessor(0));
609 AND->eraseFromParent();
610 if (Op0->use_empty())
611 Op0->eraseFromParent();
612 if (Op1->use_empty())
613 Op1->eraseFromParent();
615 // Update domiantor info. Now, ExitingBlock has only one predecessor,
616 // ExitingBBPred, and it is ExitingBlock's immediate domiantor.
617 DT->changeImmediateDominator(ExitingBlock, ExitingBBPred);
619 BasicBlock *ExitBlock = ExitingBlock->getTerminator()->getSuccessor(1);
620 if (L->contains(ExitBlock))
621 ExitBlock = ExitingBlock->getTerminator()->getSuccessor(0);
623 // If ExitingBlock is a member of the loop basic blocks' DF list then
624 // replace ExitingBlock with header and exit block in the DF list
625 DominanceFrontier::iterator ExitingBlockDF = DF->find(ExitingBlock);
626 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
629 if (BB == Header || BB == ExitingBlock)
631 DominanceFrontier::iterator BBDF = DF->find(BB);
632 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
633 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
634 while (DomSetI != DomSetE) {
635 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
637 BasicBlock *DFBB = *CurrentItr;
638 if (DFBB == ExitingBlock) {
639 BBDF->second.erase(DFBB);
640 for (DominanceFrontier::DomSetType::iterator
641 EBI = ExitingBlockDF->second.begin(),
642 EBE = ExitingBlockDF->second.end(); EBI != EBE; ++EBI)
643 BBDF->second.insert(*EBI);
651 /// removeBlocks - Remove basic block DeadBB and all blocks dominated by DeadBB.
652 /// This routine is used to remove split condition's dead branch, dominated by
653 /// DeadBB. LiveBB dominates split conidition's other branch.
654 void LoopIndexSplit::removeBlocks(BasicBlock *DeadBB, Loop *LP,
655 BasicBlock *LiveBB) {
657 // First update DeadBB's dominance frontier.
658 SmallVector<BasicBlock *, 8> FrontierBBs;
659 DominanceFrontier::iterator DeadBBDF = DF->find(DeadBB);
660 if (DeadBBDF != DF->end()) {
661 SmallVector<BasicBlock *, 8> PredBlocks;
663 DominanceFrontier::DomSetType DeadBBSet = DeadBBDF->second;
664 for (DominanceFrontier::DomSetType::iterator DeadBBSetI = DeadBBSet.begin(),
665 DeadBBSetE = DeadBBSet.end(); DeadBBSetI != DeadBBSetE; ++DeadBBSetI)
667 BasicBlock *FrontierBB = *DeadBBSetI;
668 FrontierBBs.push_back(FrontierBB);
670 // Rremove any PHI incoming edge from blocks dominated by DeadBB.
672 for(pred_iterator PI = pred_begin(FrontierBB), PE = pred_end(FrontierBB);
675 if (P == DeadBB || DT->dominates(DeadBB, P))
676 PredBlocks.push_back(P);
679 for(BasicBlock::iterator FBI = FrontierBB->begin(), FBE = FrontierBB->end();
681 if (PHINode *PN = dyn_cast<PHINode>(FBI)) {
682 for(SmallVector<BasicBlock *, 8>::iterator PI = PredBlocks.begin(),
683 PE = PredBlocks.end(); PI != PE; ++PI) {
685 PN->removeIncomingValue(P);
694 // Now remove DeadBB and all nodes dominated by DeadBB in df order.
695 SmallVector<BasicBlock *, 32> WorkList;
696 DomTreeNode *DN = DT->getNode(DeadBB);
697 for (df_iterator<DomTreeNode*> DI = df_begin(DN),
698 E = df_end(DN); DI != E; ++DI) {
699 BasicBlock *BB = DI->getBlock();
700 WorkList.push_back(BB);
701 BB->replaceAllUsesWith(UndefValue::get(Type::LabelTy));
704 while (!WorkList.empty()) {
705 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
706 LPM->deleteSimpleAnalysisValue(BB, LP);
707 for(BasicBlock::iterator BBI = BB->begin(), BBE = BB->end();
709 Instruction *I = BBI;
711 I->replaceAllUsesWith(UndefValue::get(I->getType()));
712 LPM->deleteSimpleAnalysisValue(I, LP);
713 I->eraseFromParent();
718 BB->eraseFromParent();
721 // Update Frontier BBs' dominator info.
722 while (!FrontierBBs.empty()) {
723 BasicBlock *FBB = FrontierBBs.back(); FrontierBBs.pop_back();
724 BasicBlock *NewDominator = FBB->getSinglePredecessor();
726 pred_iterator PI = pred_begin(FBB), PE = pred_end(FBB);
729 if (NewDominator != LiveBB) {
730 for(; PI != PE; ++PI) {
733 NewDominator = LiveBB;
736 NewDominator = DT->findNearestCommonDominator(NewDominator, P);
740 assert (NewDominator && "Unable to fix dominator info.");
741 DT->changeImmediateDominator(FBB, NewDominator);
742 DF->changeImmediateDominator(FBB, NewDominator, DT);
747 // moveExitCondition - Move exit condition EC into split condition block CondBB.
748 void LoopIndexSplit::moveExitCondition(BasicBlock *CondBB, BasicBlock *ActiveBB,
749 BasicBlock *ExitBB, ICmpInst *EC,
750 ICmpInst *SC, PHINode *IV,
751 Instruction *IVAdd, Loop *LP,
752 unsigned ExitValueNum) {
754 BasicBlock *ExitingBB = EC->getParent();
755 Instruction *CurrentBR = CondBB->getTerminator();
757 // Move exit condition into split condition block.
758 EC->moveBefore(CurrentBR);
759 EC->setOperand(ExitValueNum == 0 ? 1 : 0, IV);
761 // Move exiting block's branch into split condition block. Update its branch
763 BranchInst *ExitingBR = cast<BranchInst>(ExitingBB->getTerminator());
764 ExitingBR->moveBefore(CurrentBR);
765 BasicBlock *OrigDestBB = NULL;
766 if (ExitingBR->getSuccessor(0) == ExitBB) {
767 OrigDestBB = ExitingBR->getSuccessor(1);
768 ExitingBR->setSuccessor(1, ActiveBB);
771 OrigDestBB = ExitingBR->getSuccessor(0);
772 ExitingBR->setSuccessor(0, ActiveBB);
775 // Remove split condition and current split condition branch.
776 SC->eraseFromParent();
777 CurrentBR->eraseFromParent();
779 // Connect exiting block to original destination.
780 BranchInst::Create(OrigDestBB, ExitingBB);
783 updatePHINodes(ExitBB, ExitingBB, CondBB, IV, IVAdd, LP);
785 // Fix dominator info.
786 // ExitBB is now dominated by CondBB
787 DT->changeImmediateDominator(ExitBB, CondBB);
788 DF->changeImmediateDominator(ExitBB, CondBB, DT);
790 // Blocks outside the loop may have been in the dominance frontier of blocks
791 // inside the condition; this is now impossible because the blocks inside the
792 // condition no loger dominate the exit. Remove the relevant blocks from
793 // the dominance frontiers.
794 for (Loop::block_iterator I = LP->block_begin(), E = LP->block_end();
796 if (*I == CondBB || !DT->dominates(CondBB, *I)) continue;
797 DominanceFrontier::iterator BBDF = DF->find(*I);
798 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
799 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
800 while (DomSetI != DomSetE) {
801 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
803 BasicBlock *DFBB = *CurrentItr;
804 if (!LP->contains(DFBB))
805 BBDF->second.erase(DFBB);
810 /// updatePHINodes - CFG has been changed.
812 /// - ExitBB's single predecessor was Latch
813 /// - Latch's second successor was Header
815 /// - ExitBB's single predecessor is Header
816 /// - Latch's one and only successor is Header
818 /// Update ExitBB PHINodes' to reflect this change.
819 void LoopIndexSplit::updatePHINodes(BasicBlock *ExitBB, BasicBlock *Latch,
821 PHINode *IV, Instruction *IVIncrement,
824 for (BasicBlock::iterator BI = ExitBB->begin(), BE = ExitBB->end();
826 PHINode *PN = dyn_cast<PHINode>(BI);
831 Value *V = PN->getIncomingValueForBlock(Latch);
832 if (PHINode *PHV = dyn_cast<PHINode>(V)) {
833 // PHV is in Latch. PHV has one use is in ExitBB PHINode. And one use
834 // in Header which is new incoming value for PN.
836 for (Value::use_iterator UI = PHV->use_begin(), E = PHV->use_end();
838 if (PHINode *U = dyn_cast<PHINode>(*UI))
839 if (LP->contains(U->getParent())) {
844 // Add incoming value from header only if PN has any use inside the loop.
846 PN->addIncoming(NewV, Header);
848 } else if (Instruction *PHI = dyn_cast<Instruction>(V)) {
849 // If this instruction is IVIncrement then IV is new incoming value
850 // from header otherwise this instruction must be incoming value from
851 // header because loop is in LCSSA form.
852 if (PHI == IVIncrement)
853 PN->addIncoming(IV, Header);
855 PN->addIncoming(V, Header);
857 // Otherwise this is an incoming value from header because loop is in
859 PN->addIncoming(V, Header);
861 // Remove incoming value from Latch.
862 PN->removeIncomingValue(Latch);
866 bool LoopIndexSplit::splitLoop() {
867 SplitCondition = NULL;
868 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
869 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
871 BasicBlock *Header = L->getHeader();
872 BasicBlock *Latch = L->getLoopLatch();
873 BranchInst *SBR = NULL; // Split Condition Branch
874 BranchInst *EBR = cast<BranchInst>(ExitCondition->getParent()->getTerminator());
875 // If Exiting block includes loop variant instructions then this
876 // loop may not be split safely.
877 BasicBlock *ExitingBlock = ExitCondition->getParent();
878 if (!cleanBlock(ExitingBlock)) return false;
880 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
882 BranchInst *BR = dyn_cast<BranchInst>((*I)->getTerminator());
883 if (!BR || BR->isUnconditional()) continue;
884 ICmpInst *CI = dyn_cast<ICmpInst>(BR->getCondition());
885 if (!CI || CI == ExitCondition
886 || CI->getPredicate() == ICmpInst::ICMP_NE
887 || CI->getPredicate() == ICmpInst::ICMP_EQ)
890 // Unable to handle triangle loops at the moment.
891 // In triangle loop, split condition is in header and one of the
892 // the split destination is loop latch. If split condition is EQ
893 // then such loops are already handle in processOneIterationLoop().
895 && (Latch == BR->getSuccessor(0) || Latch == BR->getSuccessor(1)))
898 // If the block does not dominate the latch then this is not a diamond.
899 // Such loop may not benefit from index split.
900 if (!DT->dominates((*I), Latch))
903 // If split condition branches heads do not have single predecessor,
904 // SplitCondBlock, then is not possible to remove inactive branch.
905 if (!BR->getSuccessor(0)->getSinglePredecessor()
906 || !BR->getSuccessor(1)->getSinglePredecessor())
909 // If the merge point for BR is not loop latch then skip this condition.
910 if (BR->getSuccessor(0) != Latch) {
911 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
912 assert (DF0 != DF->end() && "Unable to find dominance frontier");
913 if (!DF0->second.count(Latch))
917 if (BR->getSuccessor(1) != Latch) {
918 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
919 assert (DF1 != DF->end() && "Unable to find dominance frontier");
920 if (!DF1->second.count(Latch))
931 // If the predicate sign does not match then skip.
932 if (ExitCondition->isSignedPredicate() != SplitCondition->isSignedPredicate())
935 unsigned EVOpNum = (ExitCondition->getOperand(1) == IVExitValue);
936 unsigned SVOpNum = IVBasedValues.count(SplitCondition->getOperand(0));
937 Value *SplitValue = SplitCondition->getOperand(SVOpNum);
938 if (!L->isLoopInvariant(SplitValue))
940 if (!IVBasedValues.count(SplitCondition->getOperand(!SVOpNum)))
943 // Normalize loop conditions so that it is easier to calculate new loop
945 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
946 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
947 BasicBlock *T = EBR->getSuccessor(0);
948 EBR->setSuccessor(0, EBR->getSuccessor(1));
949 EBR->setSuccessor(1, T);
952 if (IVisGT(*SplitCondition) || IVisGE(*SplitCondition)) {
953 SplitCondition->setPredicate(SplitCondition->getInversePredicate());
954 BasicBlock *T = SBR->getSuccessor(0);
955 SBR->setSuccessor(0, SBR->getSuccessor(1));
956 SBR->setSuccessor(1, T);
959 //[*] Calculate new loop bounds.
960 Value *AEV = SplitValue;
961 Value *BSV = SplitValue;
962 bool Sign = SplitCondition->isSignedPredicate();
963 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
965 if (IVisLT(*ExitCondition)) {
966 if (IVisLT(*SplitCondition)) {
969 else if (IVisLE(*SplitCondition)) {
970 AEV = getPlusOne(SplitValue, Sign, PHTerm, Context);
971 BSV = getPlusOne(SplitValue, Sign, PHTerm, Context);
973 assert (0 && "Unexpected split condition!");
976 else if (IVisLE(*ExitCondition)) {
977 if (IVisLT(*SplitCondition)) {
978 AEV = getMinusOne(SplitValue, Sign, PHTerm, Context);
980 else if (IVisLE(*SplitCondition)) {
981 BSV = getPlusOne(SplitValue, Sign, PHTerm, Context);
983 assert (0 && "Unexpected split condition!");
986 assert (0 && "Unexpected exit condition!");
988 AEV = getMin(AEV, IVExitValue, Sign, PHTerm);
989 BSV = getMax(BSV, IVStartValue, Sign, PHTerm);
992 DenseMap<const Value *, Value *> ValueMap;
993 Loop *BLoop = CloneLoop(L, LPM, LI, ValueMap, this);
996 // [*] ALoop's exiting edge enters BLoop's header.
997 // ALoop's original exit block becomes BLoop's exit block.
998 PHINode *B_IndVar = cast<PHINode>(ValueMap[IndVar]);
999 BasicBlock *A_ExitingBlock = ExitCondition->getParent();
1000 BranchInst *A_ExitInsn =
1001 dyn_cast<BranchInst>(A_ExitingBlock->getTerminator());
1002 assert (A_ExitInsn && "Unable to find suitable loop exit branch");
1003 BasicBlock *B_ExitBlock = A_ExitInsn->getSuccessor(1);
1004 BasicBlock *B_Header = BLoop->getHeader();
1005 if (ALoop->contains(B_ExitBlock)) {
1006 B_ExitBlock = A_ExitInsn->getSuccessor(0);
1007 A_ExitInsn->setSuccessor(0, B_Header);
1009 A_ExitInsn->setSuccessor(1, B_Header);
1011 // [*] Update ALoop's exit value using new exit value.
1012 ExitCondition->setOperand(EVOpNum, AEV);
1014 // [*] Update BLoop's header phi nodes. Remove incoming PHINode's from
1015 // original loop's preheader. Add incoming PHINode values from
1016 // ALoop's exiting block. Update BLoop header's domiantor info.
1018 // Collect inverse map of Header PHINodes.
1019 DenseMap<Value *, Value *> InverseMap;
1020 for (BasicBlock::iterator BI = ALoop->getHeader()->begin(),
1021 BE = ALoop->getHeader()->end(); BI != BE; ++BI) {
1022 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1023 PHINode *PNClone = cast<PHINode>(ValueMap[PN]);
1024 InverseMap[PNClone] = PN;
1029 BasicBlock *A_Preheader = ALoop->getLoopPreheader();
1030 for (BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1032 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1033 // Remove incoming value from original preheader.
1034 PN->removeIncomingValue(A_Preheader);
1036 // Add incoming value from A_ExitingBlock.
1038 PN->addIncoming(BSV, A_ExitingBlock);
1040 PHINode *OrigPN = cast<PHINode>(InverseMap[PN]);
1042 // If loop header is also loop exiting block then
1043 // OrigPN is incoming value for B loop header.
1044 if (A_ExitingBlock == ALoop->getHeader())
1047 V2 = OrigPN->getIncomingValueForBlock(A_ExitingBlock);
1048 PN->addIncoming(V2, A_ExitingBlock);
1054 DT->changeImmediateDominator(B_Header, A_ExitingBlock);
1055 DF->changeImmediateDominator(B_Header, A_ExitingBlock, DT);
1057 // [*] Update BLoop's exit block. Its new predecessor is BLoop's exit
1058 // block. Remove incoming PHINode values from ALoop's exiting block.
1059 // Add new incoming values from BLoop's incoming exiting value.
1060 // Update BLoop exit block's dominator info..
1061 BasicBlock *B_ExitingBlock = cast<BasicBlock>(ValueMap[A_ExitingBlock]);
1062 for (BasicBlock::iterator BI = B_ExitBlock->begin(), BE = B_ExitBlock->end();
1064 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1065 PN->addIncoming(ValueMap[PN->getIncomingValueForBlock(A_ExitingBlock)],
1067 PN->removeIncomingValue(A_ExitingBlock);
1072 DT->changeImmediateDominator(B_ExitBlock, B_ExitingBlock);
1073 DF->changeImmediateDominator(B_ExitBlock, B_ExitingBlock, DT);
1075 //[*] Split ALoop's exit edge. This creates a new block which
1076 // serves two purposes. First one is to hold PHINode defnitions
1077 // to ensure that ALoop's LCSSA form. Second use it to act
1078 // as a preheader for BLoop.
1079 BasicBlock *A_ExitBlock = SplitEdge(A_ExitingBlock, B_Header, this);
1081 //[*] Preserve ALoop's LCSSA form. Create new forwarding PHINodes
1082 // in A_ExitBlock to redefine outgoing PHI definitions from ALoop.
1083 for(BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1085 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1086 Value *V1 = PN->getIncomingValueForBlock(A_ExitBlock);
1087 PHINode *newPHI = PHINode::Create(PN->getType(), PN->getName());
1088 newPHI->addIncoming(V1, A_ExitingBlock);
1089 A_ExitBlock->getInstList().push_front(newPHI);
1090 PN->removeIncomingValue(A_ExitBlock);
1091 PN->addIncoming(newPHI, A_ExitBlock);
1096 //[*] Eliminate split condition's inactive branch from ALoop.
1097 BasicBlock *A_SplitCondBlock = SplitCondition->getParent();
1098 BranchInst *A_BR = cast<BranchInst>(A_SplitCondBlock->getTerminator());
1099 BasicBlock *A_InactiveBranch = NULL;
1100 BasicBlock *A_ActiveBranch = NULL;
1101 A_ActiveBranch = A_BR->getSuccessor(0);
1102 A_InactiveBranch = A_BR->getSuccessor(1);
1103 A_BR->setUnconditionalDest(A_ActiveBranch);
1104 removeBlocks(A_InactiveBranch, L, A_ActiveBranch);
1106 //[*] Eliminate split condition's inactive branch in from BLoop.
1107 BasicBlock *B_SplitCondBlock = cast<BasicBlock>(ValueMap[A_SplitCondBlock]);
1108 BranchInst *B_BR = cast<BranchInst>(B_SplitCondBlock->getTerminator());
1109 BasicBlock *B_InactiveBranch = NULL;
1110 BasicBlock *B_ActiveBranch = NULL;
1111 B_ActiveBranch = B_BR->getSuccessor(1);
1112 B_InactiveBranch = B_BR->getSuccessor(0);
1113 B_BR->setUnconditionalDest(B_ActiveBranch);
1114 removeBlocks(B_InactiveBranch, BLoop, B_ActiveBranch);
1116 BasicBlock *A_Header = ALoop->getHeader();
1117 if (A_ExitingBlock == A_Header)
1120 //[*] Move exit condition into split condition block to avoid
1121 // executing dead loop iteration.
1122 ICmpInst *B_ExitCondition = cast<ICmpInst>(ValueMap[ExitCondition]);
1123 Instruction *B_IndVarIncrement = cast<Instruction>(ValueMap[IVIncrement]);
1124 ICmpInst *B_SplitCondition = cast<ICmpInst>(ValueMap[SplitCondition]);
1126 moveExitCondition(A_SplitCondBlock, A_ActiveBranch, A_ExitBlock, ExitCondition,
1127 cast<ICmpInst>(SplitCondition), IndVar, IVIncrement,
1130 moveExitCondition(B_SplitCondBlock, B_ActiveBranch,
1131 B_ExitBlock, B_ExitCondition,
1132 B_SplitCondition, B_IndVar, B_IndVarIncrement,
1139 /// cleanBlock - A block is considered clean if all non terminal instructions
1140 /// are either, PHINodes, IV based.
1141 bool LoopIndexSplit::cleanBlock(BasicBlock *BB) {
1142 Instruction *Terminator = BB->getTerminator();
1143 for(BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1145 Instruction *I = BI;
1147 if (isa<PHINode>(I) || I == Terminator || I == ExitCondition
1148 || I == SplitCondition || IVBasedValues.count(I)
1149 || isa<DbgInfoIntrinsic>(I))
1152 if (I->mayHaveSideEffects())
1155 // I is used only inside this block then it is OK.
1156 bool usedOutsideBB = false;
1157 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1159 Instruction *U = cast<Instruction>(UI);
1160 if (U->getParent() != BB)
1161 usedOutsideBB = true;
1166 // Otherwise we have a instruction that may not allow loop spliting.
1172 /// IVisLT - If Op is comparing IV based value with an loop invariant and
1173 /// IV based value is less than the loop invariant then return the loop
1174 /// invariant. Otherwise return NULL.
1175 Value * LoopIndexSplit::IVisLT(ICmpInst &Op) {
1176 ICmpInst::Predicate P = Op.getPredicate();
1177 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1178 && IVBasedValues.count(Op.getOperand(0))
1179 && L->isLoopInvariant(Op.getOperand(1)))
1180 return Op.getOperand(1);
1182 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1183 && IVBasedValues.count(Op.getOperand(1))
1184 && L->isLoopInvariant(Op.getOperand(0)))
1185 return Op.getOperand(0);
1190 /// IVisLE - If Op is comparing IV based value with an loop invariant and
1191 /// IV based value is less than or equal to the loop invariant then
1192 /// return the loop invariant. Otherwise return NULL.
1193 Value * LoopIndexSplit::IVisLE(ICmpInst &Op) {
1194 ICmpInst::Predicate P = Op.getPredicate();
1195 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1196 && IVBasedValues.count(Op.getOperand(0))
1197 && L->isLoopInvariant(Op.getOperand(1)))
1198 return Op.getOperand(1);
1200 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1201 && IVBasedValues.count(Op.getOperand(1))
1202 && L->isLoopInvariant(Op.getOperand(0)))
1203 return Op.getOperand(0);
1208 /// IVisGT - If Op is comparing IV based value with an loop invariant and
1209 /// IV based value is greater than the loop invariant then return the loop
1210 /// invariant. Otherwise return NULL.
1211 Value * LoopIndexSplit::IVisGT(ICmpInst &Op) {
1212 ICmpInst::Predicate P = Op.getPredicate();
1213 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1214 && IVBasedValues.count(Op.getOperand(0))
1215 && L->isLoopInvariant(Op.getOperand(1)))
1216 return Op.getOperand(1);
1218 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1219 && IVBasedValues.count(Op.getOperand(1))
1220 && L->isLoopInvariant(Op.getOperand(0)))
1221 return Op.getOperand(0);
1226 /// IVisGE - If Op is comparing IV based value with an loop invariant and
1227 /// IV based value is greater than or equal to the loop invariant then
1228 /// return the loop invariant. Otherwise return NULL.
1229 Value * LoopIndexSplit::IVisGE(ICmpInst &Op) {
1230 ICmpInst::Predicate P = Op.getPredicate();
1231 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1232 && IVBasedValues.count(Op.getOperand(0))
1233 && L->isLoopInvariant(Op.getOperand(1)))
1234 return Op.getOperand(1);
1236 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1237 && IVBasedValues.count(Op.getOperand(1))
1238 && L->isLoopInvariant(Op.getOperand(0)))
1239 return Op.getOperand(0);