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 std::swap(OPV, SplitValue);
354 if (!L->isLoopInvariant(SplitValue))
356 Instruction *OPI = dyn_cast<Instruction>(OPV);
359 if (OPI->getParent() != Header || isUsedOutsideLoop(OPI, L))
361 Value *StartValue = IVStartValue;
362 Value *ExitValue = IVExitValue;;
365 // If BR operand is IV based then use this operand to calculate
366 // effective conditions for loop body.
367 BinaryOperator *BOPV = dyn_cast<BinaryOperator>(OPV);
370 if (BOPV->getOpcode() != Instruction::Add)
372 StartValue = BinaryOperator::CreateAdd(OPV, StartValue, "" , BR);
373 ExitValue = BinaryOperator::CreateAdd(OPV, ExitValue, "" , BR);
376 if (!cleanBlock(Header))
379 if (!cleanBlock(Latch))
382 // If the merge point for BR is not loop latch then skip this loop.
383 if (BR->getSuccessor(0) != Latch) {
384 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
385 assert (DF0 != DF->end() && "Unable to find dominance frontier");
386 if (!DF0->second.count(Latch))
390 if (BR->getSuccessor(1) != Latch) {
391 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
392 assert (DF1 != DF->end() && "Unable to find dominance frontier");
393 if (!DF1->second.count(Latch))
397 // Now, Current loop L contains compare instruction
398 // that compares induction variable, IndVar, against loop invariant. And
399 // entire (i.e. meaningful) loop body is dominated by this compare
400 // instruction. In such case eliminate
401 // loop structure surrounding this loop body. For example,
402 // for (int i = start; i < end; ++i) {
403 // if ( i == somevalue) {
407 // can be transformed into
408 // if (somevalue >= start && somevalue < end) {
413 // Replace index variable with split value in loop body. Loop body is executed
414 // only when index variable is equal to split value.
415 IndVar->replaceAllUsesWith(SplitValue);
417 // Replace split condition in header.
419 // SplitCondition : icmp eq i32 IndVar, SplitValue
421 // c1 = icmp uge i32 SplitValue, StartValue
422 // c2 = icmp ult i32 SplitValue, ExitValue
424 Instruction *C1 = new ICmpInst(ExitCondition->isSignedPredicate() ?
425 ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
426 SplitValue, StartValue, "lisplit", BR);
428 CmpInst::Predicate C2P = ExitCondition->getPredicate();
429 BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
430 if (LatchBR->getOperand(0) != Header)
431 C2P = CmpInst::getInversePredicate(C2P);
432 Instruction *C2 = new ICmpInst(C2P, SplitValue, ExitValue, "lisplit", BR);
433 Instruction *NSplitCond = BinaryOperator::CreateAnd(C1, C2, "lisplit", BR);
435 SplitCondition->replaceAllUsesWith(NSplitCond);
436 SplitCondition->eraseFromParent();
438 // Remove Latch to Header edge.
439 BasicBlock *LatchSucc = NULL;
440 Header->removePredecessor(Latch);
441 for (succ_iterator SI = succ_begin(Latch), E = succ_end(Latch);
447 // Clean up latch block.
448 Value *LatchBRCond = LatchBR->getCondition();
449 LatchBR->setUnconditionalDest(LatchSucc);
450 RecursivelyDeleteTriviallyDeadInstructions(LatchBRCond);
452 LPM->deleteLoopFromQueue(L);
454 // Update Dominator Info.
455 // Only CFG change done is to remove Latch to Header edge. This
456 // does not change dominator tree because Latch did not dominate
459 DominanceFrontier::iterator HeaderDF = DF->find(Header);
460 if (HeaderDF != DF->end())
461 DF->removeFromFrontier(HeaderDF, Header);
463 DominanceFrontier::iterator LatchDF = DF->find(Latch);
464 if (LatchDF != DF->end())
465 DF->removeFromFrontier(LatchDF, Header);
468 ++NumIndexSplitRemoved;
472 /// restrictLoopBound - Op dominates loop body. Op compares an IV based value
473 /// with a loop invariant value. Update loop's lower and upper bound based on
474 /// the loop invariant value.
475 bool LoopIndexSplit::restrictLoopBound(ICmpInst &Op) {
476 bool Sign = Op.isSignedPredicate();
477 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
479 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
481 cast<BranchInst>(ExitCondition->getParent()->getTerminator());
482 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
483 BasicBlock *T = EBR->getSuccessor(0);
484 EBR->setSuccessor(0, EBR->getSuccessor(1));
485 EBR->setSuccessor(1, T);
488 // New upper and lower bounds.
491 if (Value *V = IVisLT(Op)) {
492 // Restrict upper bound.
493 if (IVisLE(*ExitCondition))
494 V = getMinusOne(V, Sign, PHTerm);
495 NUB = getMin(V, IVExitValue, Sign, PHTerm);
496 } else if (Value *V = IVisLE(Op)) {
497 // Restrict upper bound.
498 if (IVisLT(*ExitCondition))
499 V = getPlusOne(V, Sign, PHTerm);
500 NUB = getMin(V, IVExitValue, Sign, PHTerm);
501 } else if (Value *V = IVisGT(Op)) {
502 // Restrict lower bound.
503 V = getPlusOne(V, Sign, PHTerm);
504 NLB = getMax(V, IVStartValue, Sign, PHTerm);
505 } else if (Value *V = IVisGE(Op))
506 // Restrict lower bound.
507 NLB = getMax(V, IVStartValue, Sign, PHTerm);
513 unsigned i = IndVar->getBasicBlockIndex(L->getLoopPreheader());
514 IndVar->setIncomingValue(i, NLB);
518 unsigned i = (ExitCondition->getOperand(0) != IVExitValue);
519 ExitCondition->setOperand(i, NUB);
524 /// updateLoopIterationSpace -- Update loop's iteration space if loop
525 /// body is executed for certain IV range only. For example,
527 /// for (i = 0; i < N; ++i) {
528 /// if ( i > A && i < B) {
532 /// is transformed to iterators from A to B, if A > 0 and B < N.
534 bool LoopIndexSplit::updateLoopIterationSpace() {
535 SplitCondition = NULL;
536 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
537 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
539 BasicBlock *Latch = L->getLoopLatch();
540 BasicBlock *Header = L->getHeader();
541 BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
542 if (!BR) return false;
543 if (!isa<BranchInst>(Latch->getTerminator())) return false;
544 if (BR->isUnconditional()) return false;
545 BinaryOperator *AND = dyn_cast<BinaryOperator>(BR->getCondition());
546 if (!AND) return false;
547 if (AND->getOpcode() != Instruction::And) return false;
548 ICmpInst *Op0 = dyn_cast<ICmpInst>(AND->getOperand(0));
549 ICmpInst *Op1 = dyn_cast<ICmpInst>(AND->getOperand(1));
552 IVBasedValues.insert(AND);
553 IVBasedValues.insert(Op0);
554 IVBasedValues.insert(Op1);
555 if (!cleanBlock(Header)) return false;
556 BasicBlock *ExitingBlock = ExitCondition->getParent();
557 if (!cleanBlock(ExitingBlock)) return false;
559 // If the merge point for BR is not loop latch then skip this loop.
560 if (BR->getSuccessor(0) != Latch) {
561 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
562 assert (DF0 != DF->end() && "Unable to find dominance frontier");
563 if (!DF0->second.count(Latch))
567 if (BR->getSuccessor(1) != Latch) {
568 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
569 assert (DF1 != DF->end() && "Unable to find dominance frontier");
570 if (!DF1->second.count(Latch))
574 // Verify that loop exiting block has only two predecessor, where one pred
575 // is split condition block. The other predecessor will become exiting block's
576 // dominator after CFG is updated. TODO : Handle CFG's where exiting block has
577 // more then two predecessors. This requires extra work in updating dominator
579 BasicBlock *ExitingBBPred = NULL;
580 for (pred_iterator PI = pred_begin(ExitingBlock), PE = pred_end(ExitingBlock);
582 BasicBlock *BB = *PI;
591 if (!restrictLoopBound(*Op0))
594 if (!restrictLoopBound(*Op1))
598 if (BR->getSuccessor(0) == ExitingBlock)
599 BR->setUnconditionalDest(BR->getSuccessor(1));
601 BR->setUnconditionalDest(BR->getSuccessor(0));
603 AND->eraseFromParent();
604 if (Op0->use_empty())
605 Op0->eraseFromParent();
606 if (Op1->use_empty())
607 Op1->eraseFromParent();
609 // Update domiantor info. Now, ExitingBlock has only one predecessor,
610 // ExitingBBPred, and it is ExitingBlock's immediate domiantor.
611 DT->changeImmediateDominator(ExitingBlock, ExitingBBPred);
613 BasicBlock *ExitBlock = ExitingBlock->getTerminator()->getSuccessor(1);
614 if (L->contains(ExitBlock))
615 ExitBlock = ExitingBlock->getTerminator()->getSuccessor(0);
617 // If ExitingBlock is a member of the loop basic blocks' DF list then
618 // replace ExitingBlock with header and exit block in the DF list
619 DominanceFrontier::iterator ExitingBlockDF = DF->find(ExitingBlock);
620 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
623 if (BB == Header || BB == ExitingBlock)
625 DominanceFrontier::iterator BBDF = DF->find(BB);
626 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
627 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
628 while (DomSetI != DomSetE) {
629 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
631 BasicBlock *DFBB = *CurrentItr;
632 if (DFBB == ExitingBlock) {
633 BBDF->second.erase(DFBB);
634 for (DominanceFrontier::DomSetType::iterator
635 EBI = ExitingBlockDF->second.begin(),
636 EBE = ExitingBlockDF->second.end(); EBI != EBE; ++EBI)
637 BBDF->second.insert(*EBI);
645 /// removeBlocks - Remove basic block DeadBB and all blocks dominated by DeadBB.
646 /// This routine is used to remove split condition's dead branch, dominated by
647 /// DeadBB. LiveBB dominates split conidition's other branch.
648 void LoopIndexSplit::removeBlocks(BasicBlock *DeadBB, Loop *LP,
649 BasicBlock *LiveBB) {
651 // First update DeadBB's dominance frontier.
652 SmallVector<BasicBlock *, 8> FrontierBBs;
653 DominanceFrontier::iterator DeadBBDF = DF->find(DeadBB);
654 if (DeadBBDF != DF->end()) {
655 SmallVector<BasicBlock *, 8> PredBlocks;
657 DominanceFrontier::DomSetType DeadBBSet = DeadBBDF->second;
658 for (DominanceFrontier::DomSetType::iterator DeadBBSetI = DeadBBSet.begin(),
659 DeadBBSetE = DeadBBSet.end(); DeadBBSetI != DeadBBSetE; ++DeadBBSetI)
661 BasicBlock *FrontierBB = *DeadBBSetI;
662 FrontierBBs.push_back(FrontierBB);
664 // Rremove any PHI incoming edge from blocks dominated by DeadBB.
666 for(pred_iterator PI = pred_begin(FrontierBB), PE = pred_end(FrontierBB);
669 if (P == DeadBB || DT->dominates(DeadBB, P))
670 PredBlocks.push_back(P);
673 for(BasicBlock::iterator FBI = FrontierBB->begin(), FBE = FrontierBB->end();
675 if (PHINode *PN = dyn_cast<PHINode>(FBI)) {
676 for(SmallVector<BasicBlock *, 8>::iterator PI = PredBlocks.begin(),
677 PE = PredBlocks.end(); PI != PE; ++PI) {
679 PN->removeIncomingValue(P);
688 // Now remove DeadBB and all nodes dominated by DeadBB in df order.
689 SmallVector<BasicBlock *, 32> WorkList;
690 DomTreeNode *DN = DT->getNode(DeadBB);
691 for (df_iterator<DomTreeNode*> DI = df_begin(DN),
692 E = df_end(DN); DI != E; ++DI) {
693 BasicBlock *BB = DI->getBlock();
694 WorkList.push_back(BB);
695 BB->replaceAllUsesWith(UndefValue::get(Type::LabelTy));
698 while (!WorkList.empty()) {
699 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
700 LPM->deleteSimpleAnalysisValue(BB, LP);
701 for(BasicBlock::iterator BBI = BB->begin(), BBE = BB->end();
703 Instruction *I = BBI;
705 I->replaceAllUsesWith(UndefValue::get(I->getType()));
706 LPM->deleteSimpleAnalysisValue(I, LP);
707 I->eraseFromParent();
712 BB->eraseFromParent();
715 // Update Frontier BBs' dominator info.
716 while (!FrontierBBs.empty()) {
717 BasicBlock *FBB = FrontierBBs.back(); FrontierBBs.pop_back();
718 BasicBlock *NewDominator = FBB->getSinglePredecessor();
720 pred_iterator PI = pred_begin(FBB), PE = pred_end(FBB);
723 if (NewDominator != LiveBB) {
724 for(; PI != PE; ++PI) {
727 NewDominator = LiveBB;
730 NewDominator = DT->findNearestCommonDominator(NewDominator, P);
734 assert (NewDominator && "Unable to fix dominator info.");
735 DT->changeImmediateDominator(FBB, NewDominator);
736 DF->changeImmediateDominator(FBB, NewDominator, DT);
741 // moveExitCondition - Move exit condition EC into split condition block CondBB.
742 void LoopIndexSplit::moveExitCondition(BasicBlock *CondBB, BasicBlock *ActiveBB,
743 BasicBlock *ExitBB, ICmpInst *EC,
744 ICmpInst *SC, PHINode *IV,
745 Instruction *IVAdd, Loop *LP,
746 unsigned ExitValueNum) {
748 BasicBlock *ExitingBB = EC->getParent();
749 Instruction *CurrentBR = CondBB->getTerminator();
751 // Move exit condition into split condition block.
752 EC->moveBefore(CurrentBR);
753 EC->setOperand(ExitValueNum == 0 ? 1 : 0, IV);
755 // Move exiting block's branch into split condition block. Update its branch
757 BranchInst *ExitingBR = cast<BranchInst>(ExitingBB->getTerminator());
758 ExitingBR->moveBefore(CurrentBR);
759 BasicBlock *OrigDestBB = NULL;
760 if (ExitingBR->getSuccessor(0) == ExitBB) {
761 OrigDestBB = ExitingBR->getSuccessor(1);
762 ExitingBR->setSuccessor(1, ActiveBB);
765 OrigDestBB = ExitingBR->getSuccessor(0);
766 ExitingBR->setSuccessor(0, ActiveBB);
769 // Remove split condition and current split condition branch.
770 SC->eraseFromParent();
771 CurrentBR->eraseFromParent();
773 // Connect exiting block to original destination.
774 BranchInst::Create(OrigDestBB, ExitingBB);
777 updatePHINodes(ExitBB, ExitingBB, CondBB, IV, IVAdd, LP);
779 // Fix dominator info.
780 // ExitBB is now dominated by CondBB
781 DT->changeImmediateDominator(ExitBB, CondBB);
782 DF->changeImmediateDominator(ExitBB, CondBB, DT);
784 // Blocks outside the loop may have been in the dominance frontier of blocks
785 // inside the condition; this is now impossible because the blocks inside the
786 // condition no loger dominate the exit. Remove the relevant blocks from
787 // the dominance frontiers.
788 for (Loop::block_iterator I = LP->block_begin(), E = LP->block_end();
790 if (*I == CondBB || !DT->dominates(CondBB, *I)) continue;
791 DominanceFrontier::iterator BBDF = DF->find(*I);
792 DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
793 DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
794 while (DomSetI != DomSetE) {
795 DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
797 BasicBlock *DFBB = *CurrentItr;
798 if (!LP->contains(DFBB))
799 BBDF->second.erase(DFBB);
804 /// updatePHINodes - CFG has been changed.
806 /// - ExitBB's single predecessor was Latch
807 /// - Latch's second successor was Header
809 /// - ExitBB's single predecessor is Header
810 /// - Latch's one and only successor is Header
812 /// Update ExitBB PHINodes' to reflect this change.
813 void LoopIndexSplit::updatePHINodes(BasicBlock *ExitBB, BasicBlock *Latch,
815 PHINode *IV, Instruction *IVIncrement,
818 for (BasicBlock::iterator BI = ExitBB->begin(), BE = ExitBB->end();
820 PHINode *PN = dyn_cast<PHINode>(BI);
825 Value *V = PN->getIncomingValueForBlock(Latch);
826 if (PHINode *PHV = dyn_cast<PHINode>(V)) {
827 // PHV is in Latch. PHV has one use is in ExitBB PHINode. And one use
828 // in Header which is new incoming value for PN.
830 for (Value::use_iterator UI = PHV->use_begin(), E = PHV->use_end();
832 if (PHINode *U = dyn_cast<PHINode>(*UI))
833 if (LP->contains(U->getParent())) {
838 // Add incoming value from header only if PN has any use inside the loop.
840 PN->addIncoming(NewV, Header);
842 } else if (Instruction *PHI = dyn_cast<Instruction>(V)) {
843 // If this instruction is IVIncrement then IV is new incoming value
844 // from header otherwise this instruction must be incoming value from
845 // header because loop is in LCSSA form.
846 if (PHI == IVIncrement)
847 PN->addIncoming(IV, Header);
849 PN->addIncoming(V, Header);
851 // Otherwise this is an incoming value from header because loop is in
853 PN->addIncoming(V, Header);
855 // Remove incoming value from Latch.
856 PN->removeIncomingValue(Latch);
860 bool LoopIndexSplit::splitLoop() {
861 SplitCondition = NULL;
862 if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
863 || ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
865 BasicBlock *Header = L->getHeader();
866 BasicBlock *Latch = L->getLoopLatch();
867 BranchInst *SBR = NULL; // Split Condition Branch
868 BranchInst *EBR = cast<BranchInst>(ExitCondition->getParent()->getTerminator());
869 // If Exiting block includes loop variant instructions then this
870 // loop may not be split safely.
871 BasicBlock *ExitingBlock = ExitCondition->getParent();
872 if (!cleanBlock(ExitingBlock)) return false;
874 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
876 BranchInst *BR = dyn_cast<BranchInst>((*I)->getTerminator());
877 if (!BR || BR->isUnconditional()) continue;
878 ICmpInst *CI = dyn_cast<ICmpInst>(BR->getCondition());
879 if (!CI || CI == ExitCondition
880 || CI->getPredicate() == ICmpInst::ICMP_NE
881 || CI->getPredicate() == ICmpInst::ICMP_EQ)
884 // Unable to handle triangle loops at the moment.
885 // In triangle loop, split condition is in header and one of the
886 // the split destination is loop latch. If split condition is EQ
887 // then such loops are already handle in processOneIterationLoop().
889 && (Latch == BR->getSuccessor(0) || Latch == BR->getSuccessor(1)))
892 // If the block does not dominate the latch then this is not a diamond.
893 // Such loop may not benefit from index split.
894 if (!DT->dominates((*I), Latch))
897 // If split condition branches heads do not have single predecessor,
898 // SplitCondBlock, then is not possible to remove inactive branch.
899 if (!BR->getSuccessor(0)->getSinglePredecessor()
900 || !BR->getSuccessor(1)->getSinglePredecessor())
903 // If the merge point for BR is not loop latch then skip this condition.
904 if (BR->getSuccessor(0) != Latch) {
905 DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
906 assert (DF0 != DF->end() && "Unable to find dominance frontier");
907 if (!DF0->second.count(Latch))
911 if (BR->getSuccessor(1) != Latch) {
912 DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
913 assert (DF1 != DF->end() && "Unable to find dominance frontier");
914 if (!DF1->second.count(Latch))
925 // If the predicate sign does not match then skip.
926 if (ExitCondition->isSignedPredicate() != SplitCondition->isSignedPredicate())
929 unsigned EVOpNum = (ExitCondition->getOperand(1) == IVExitValue);
930 unsigned SVOpNum = IVBasedValues.count(SplitCondition->getOperand(0));
931 Value *SplitValue = SplitCondition->getOperand(SVOpNum);
932 if (!L->isLoopInvariant(SplitValue))
934 if (!IVBasedValues.count(SplitCondition->getOperand(!SVOpNum)))
937 // Normalize loop conditions so that it is easier to calculate new loop
939 if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
940 ExitCondition->setPredicate(ExitCondition->getInversePredicate());
941 BasicBlock *T = EBR->getSuccessor(0);
942 EBR->setSuccessor(0, EBR->getSuccessor(1));
943 EBR->setSuccessor(1, T);
946 if (IVisGT(*SplitCondition) || IVisGE(*SplitCondition)) {
947 SplitCondition->setPredicate(SplitCondition->getInversePredicate());
948 BasicBlock *T = SBR->getSuccessor(0);
949 SBR->setSuccessor(0, SBR->getSuccessor(1));
950 SBR->setSuccessor(1, T);
953 //[*] Calculate new loop bounds.
954 Value *AEV = SplitValue;
955 Value *BSV = SplitValue;
956 bool Sign = SplitCondition->isSignedPredicate();
957 Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
959 if (IVisLT(*ExitCondition)) {
960 if (IVisLT(*SplitCondition)) {
963 else if (IVisLE(*SplitCondition)) {
964 AEV = getPlusOne(SplitValue, Sign, PHTerm);
965 BSV = getPlusOne(SplitValue, Sign, PHTerm);
967 assert (0 && "Unexpected split condition!");
970 else if (IVisLE(*ExitCondition)) {
971 if (IVisLT(*SplitCondition)) {
972 AEV = getMinusOne(SplitValue, Sign, PHTerm);
974 else if (IVisLE(*SplitCondition)) {
975 BSV = getPlusOne(SplitValue, Sign, PHTerm);
977 assert (0 && "Unexpected split condition!");
980 assert (0 && "Unexpected exit condition!");
982 AEV = getMin(AEV, IVExitValue, Sign, PHTerm);
983 BSV = getMax(BSV, IVStartValue, Sign, PHTerm);
986 DenseMap<const Value *, Value *> ValueMap;
987 Loop *BLoop = CloneLoop(L, LPM, LI, ValueMap, this);
990 // [*] ALoop's exiting edge enters BLoop's header.
991 // ALoop's original exit block becomes BLoop's exit block.
992 PHINode *B_IndVar = cast<PHINode>(ValueMap[IndVar]);
993 BasicBlock *A_ExitingBlock = ExitCondition->getParent();
994 BranchInst *A_ExitInsn =
995 dyn_cast<BranchInst>(A_ExitingBlock->getTerminator());
996 assert (A_ExitInsn && "Unable to find suitable loop exit branch");
997 BasicBlock *B_ExitBlock = A_ExitInsn->getSuccessor(1);
998 BasicBlock *B_Header = BLoop->getHeader();
999 if (ALoop->contains(B_ExitBlock)) {
1000 B_ExitBlock = A_ExitInsn->getSuccessor(0);
1001 A_ExitInsn->setSuccessor(0, B_Header);
1003 A_ExitInsn->setSuccessor(1, B_Header);
1005 // [*] Update ALoop's exit value using new exit value.
1006 ExitCondition->setOperand(EVOpNum, AEV);
1008 // [*] Update BLoop's header phi nodes. Remove incoming PHINode's from
1009 // original loop's preheader. Add incoming PHINode values from
1010 // ALoop's exiting block. Update BLoop header's domiantor info.
1012 // Collect inverse map of Header PHINodes.
1013 DenseMap<Value *, Value *> InverseMap;
1014 for (BasicBlock::iterator BI = ALoop->getHeader()->begin(),
1015 BE = ALoop->getHeader()->end(); BI != BE; ++BI) {
1016 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1017 PHINode *PNClone = cast<PHINode>(ValueMap[PN]);
1018 InverseMap[PNClone] = PN;
1023 BasicBlock *A_Preheader = ALoop->getLoopPreheader();
1024 for (BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1026 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1027 // Remove incoming value from original preheader.
1028 PN->removeIncomingValue(A_Preheader);
1030 // Add incoming value from A_ExitingBlock.
1032 PN->addIncoming(BSV, A_ExitingBlock);
1034 PHINode *OrigPN = cast<PHINode>(InverseMap[PN]);
1036 // If loop header is also loop exiting block then
1037 // OrigPN is incoming value for B loop header.
1038 if (A_ExitingBlock == ALoop->getHeader())
1041 V2 = OrigPN->getIncomingValueForBlock(A_ExitingBlock);
1042 PN->addIncoming(V2, A_ExitingBlock);
1048 DT->changeImmediateDominator(B_Header, A_ExitingBlock);
1049 DF->changeImmediateDominator(B_Header, A_ExitingBlock, DT);
1051 // [*] Update BLoop's exit block. Its new predecessor is BLoop's exit
1052 // block. Remove incoming PHINode values from ALoop's exiting block.
1053 // Add new incoming values from BLoop's incoming exiting value.
1054 // Update BLoop exit block's dominator info..
1055 BasicBlock *B_ExitingBlock = cast<BasicBlock>(ValueMap[A_ExitingBlock]);
1056 for (BasicBlock::iterator BI = B_ExitBlock->begin(), BE = B_ExitBlock->end();
1058 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1059 PN->addIncoming(ValueMap[PN->getIncomingValueForBlock(A_ExitingBlock)],
1061 PN->removeIncomingValue(A_ExitingBlock);
1066 DT->changeImmediateDominator(B_ExitBlock, B_ExitingBlock);
1067 DF->changeImmediateDominator(B_ExitBlock, B_ExitingBlock, DT);
1069 //[*] Split ALoop's exit edge. This creates a new block which
1070 // serves two purposes. First one is to hold PHINode defnitions
1071 // to ensure that ALoop's LCSSA form. Second use it to act
1072 // as a preheader for BLoop.
1073 BasicBlock *A_ExitBlock = SplitEdge(A_ExitingBlock, B_Header, this);
1075 //[*] Preserve ALoop's LCSSA form. Create new forwarding PHINodes
1076 // in A_ExitBlock to redefine outgoing PHI definitions from ALoop.
1077 for(BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
1079 if (PHINode *PN = dyn_cast<PHINode>(BI)) {
1080 Value *V1 = PN->getIncomingValueForBlock(A_ExitBlock);
1081 PHINode *newPHI = PHINode::Create(PN->getType(), PN->getName());
1082 newPHI->addIncoming(V1, A_ExitingBlock);
1083 A_ExitBlock->getInstList().push_front(newPHI);
1084 PN->removeIncomingValue(A_ExitBlock);
1085 PN->addIncoming(newPHI, A_ExitBlock);
1090 //[*] Eliminate split condition's inactive branch from ALoop.
1091 BasicBlock *A_SplitCondBlock = SplitCondition->getParent();
1092 BranchInst *A_BR = cast<BranchInst>(A_SplitCondBlock->getTerminator());
1093 BasicBlock *A_InactiveBranch = NULL;
1094 BasicBlock *A_ActiveBranch = NULL;
1095 A_ActiveBranch = A_BR->getSuccessor(0);
1096 A_InactiveBranch = A_BR->getSuccessor(1);
1097 A_BR->setUnconditionalDest(A_ActiveBranch);
1098 removeBlocks(A_InactiveBranch, L, A_ActiveBranch);
1100 //[*] Eliminate split condition's inactive branch in from BLoop.
1101 BasicBlock *B_SplitCondBlock = cast<BasicBlock>(ValueMap[A_SplitCondBlock]);
1102 BranchInst *B_BR = cast<BranchInst>(B_SplitCondBlock->getTerminator());
1103 BasicBlock *B_InactiveBranch = NULL;
1104 BasicBlock *B_ActiveBranch = NULL;
1105 B_ActiveBranch = B_BR->getSuccessor(1);
1106 B_InactiveBranch = B_BR->getSuccessor(0);
1107 B_BR->setUnconditionalDest(B_ActiveBranch);
1108 removeBlocks(B_InactiveBranch, BLoop, B_ActiveBranch);
1110 BasicBlock *A_Header = ALoop->getHeader();
1111 if (A_ExitingBlock == A_Header)
1114 //[*] Move exit condition into split condition block to avoid
1115 // executing dead loop iteration.
1116 ICmpInst *B_ExitCondition = cast<ICmpInst>(ValueMap[ExitCondition]);
1117 Instruction *B_IndVarIncrement = cast<Instruction>(ValueMap[IVIncrement]);
1118 ICmpInst *B_SplitCondition = cast<ICmpInst>(ValueMap[SplitCondition]);
1120 moveExitCondition(A_SplitCondBlock, A_ActiveBranch, A_ExitBlock, ExitCondition,
1121 cast<ICmpInst>(SplitCondition), IndVar, IVIncrement,
1124 moveExitCondition(B_SplitCondBlock, B_ActiveBranch,
1125 B_ExitBlock, B_ExitCondition,
1126 B_SplitCondition, B_IndVar, B_IndVarIncrement,
1133 /// cleanBlock - A block is considered clean if all non terminal instructions
1134 /// are either, PHINodes, IV based.
1135 bool LoopIndexSplit::cleanBlock(BasicBlock *BB) {
1136 Instruction *Terminator = BB->getTerminator();
1137 for(BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1139 Instruction *I = BI;
1141 if (isa<PHINode>(I) || I == Terminator || I == ExitCondition
1142 || I == SplitCondition || IVBasedValues.count(I)
1143 || isa<DbgInfoIntrinsic>(I))
1146 if (I->mayHaveSideEffects())
1149 // I is used only inside this block then it is OK.
1150 bool usedOutsideBB = false;
1151 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1153 Instruction *U = cast<Instruction>(UI);
1154 if (U->getParent() != BB)
1155 usedOutsideBB = true;
1160 // Otherwise we have a instruction that may not allow loop spliting.
1166 /// IVisLT - If Op is comparing IV based value with an loop invariant and
1167 /// IV based value is less than the loop invariant then return the loop
1168 /// invariant. Otherwise return NULL.
1169 Value * LoopIndexSplit::IVisLT(ICmpInst &Op) {
1170 ICmpInst::Predicate P = Op.getPredicate();
1171 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1172 && IVBasedValues.count(Op.getOperand(0))
1173 && L->isLoopInvariant(Op.getOperand(1)))
1174 return Op.getOperand(1);
1176 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1177 && IVBasedValues.count(Op.getOperand(1))
1178 && L->isLoopInvariant(Op.getOperand(0)))
1179 return Op.getOperand(0);
1184 /// IVisLE - If Op is comparing IV based value with an loop invariant and
1185 /// IV based value is less than or equal to the loop invariant then
1186 /// return the loop invariant. Otherwise return NULL.
1187 Value * LoopIndexSplit::IVisLE(ICmpInst &Op) {
1188 ICmpInst::Predicate P = Op.getPredicate();
1189 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1190 && IVBasedValues.count(Op.getOperand(0))
1191 && L->isLoopInvariant(Op.getOperand(1)))
1192 return Op.getOperand(1);
1194 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1195 && IVBasedValues.count(Op.getOperand(1))
1196 && L->isLoopInvariant(Op.getOperand(0)))
1197 return Op.getOperand(0);
1202 /// IVisGT - If Op is comparing IV based value with an loop invariant and
1203 /// IV based value is greater than the loop invariant then return the loop
1204 /// invariant. Otherwise return NULL.
1205 Value * LoopIndexSplit::IVisGT(ICmpInst &Op) {
1206 ICmpInst::Predicate P = Op.getPredicate();
1207 if ((P == ICmpInst::ICMP_SGT || P == ICmpInst::ICMP_UGT)
1208 && IVBasedValues.count(Op.getOperand(0))
1209 && L->isLoopInvariant(Op.getOperand(1)))
1210 return Op.getOperand(1);
1212 if ((P == ICmpInst::ICMP_SLT || P == ICmpInst::ICMP_ULT)
1213 && IVBasedValues.count(Op.getOperand(1))
1214 && L->isLoopInvariant(Op.getOperand(0)))
1215 return Op.getOperand(0);
1220 /// IVisGE - If Op is comparing IV based value with an loop invariant and
1221 /// IV based value is greater than or equal to the loop invariant then
1222 /// return the loop invariant. Otherwise return NULL.
1223 Value * LoopIndexSplit::IVisGE(ICmpInst &Op) {
1224 ICmpInst::Predicate P = Op.getPredicate();
1225 if ((P == ICmpInst::ICMP_SGE || P == ICmpInst::ICMP_UGE)
1226 && IVBasedValues.count(Op.getOperand(0))
1227 && L->isLoopInvariant(Op.getOperand(1)))
1228 return Op.getOperand(1);
1230 if ((P == ICmpInst::ICMP_SLE || P == ICmpInst::ICMP_ULE)
1231 && IVBasedValues.count(Op.getOperand(1))
1232 && L->isLoopInvariant(Op.getOperand(0)))
1233 return Op.getOperand(0);