1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
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 pass transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #define DEBUG_TYPE "loop-unswitch"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Analysis/CodeMetrics.h"
35 #include "llvm/Analysis/InstructionSimplify.h"
36 #include "llvm/Analysis/LoopInfo.h"
37 #include "llvm/Analysis/LoopPass.h"
38 #include "llvm/Analysis/ScalarEvolution.h"
39 #include "llvm/Analysis/TargetTransformInfo.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Dominators.h"
43 #include "llvm/IR/Function.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Cloning.h"
50 #include "llvm/Transforms/Utils/Local.h"
56 STATISTIC(NumBranches, "Number of branches unswitched");
57 STATISTIC(NumSwitches, "Number of switches unswitched");
58 STATISTIC(NumSelects , "Number of selects unswitched");
59 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
60 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
61 STATISTIC(TotalInsts, "Total number of instructions analyzed");
63 // The specific value of 100 here was chosen based only on intuition and a
64 // few specific examples.
65 static cl::opt<unsigned>
66 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
67 cl::init(100), cl::Hidden);
71 class LUAnalysisCache {
73 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
76 typedef UnswitchedValsMap::iterator UnswitchedValsIt;
78 struct LoopProperties {
79 unsigned CanBeUnswitchedCount;
80 unsigned SizeEstimation;
81 UnswitchedValsMap UnswitchedVals;
84 // Here we use std::map instead of DenseMap, since we need to keep valid
85 // LoopProperties pointer for current loop for better performance.
86 typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
87 typedef LoopPropsMap::iterator LoopPropsMapIt;
89 LoopPropsMap LoopsProperties;
90 UnswitchedValsMap *CurLoopInstructions;
91 LoopProperties *CurrentLoopProperties;
93 // Max size of code we can produce on remained iterations.
99 CurLoopInstructions(0), CurrentLoopProperties(0),
103 // Analyze loop. Check its size, calculate is it possible to unswitch
104 // it. Returns true if we can unswitch this loop.
105 bool countLoop(const Loop *L, const TargetTransformInfo &TTI);
107 // Clean all data related to given loop.
108 void forgetLoop(const Loop *L);
110 // Mark case value as unswitched.
111 // Since SI instruction can be partly unswitched, in order to avoid
112 // extra unswitching in cloned loops keep track all unswitched values.
113 void setUnswitched(const SwitchInst *SI, const Value *V);
115 // Check was this case value unswitched before or not.
116 bool isUnswitched(const SwitchInst *SI, const Value *V);
118 // Clone all loop-unswitch related loop properties.
119 // Redistribute unswitching quotas.
120 // Note, that new loop data is stored inside the VMap.
121 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
122 const ValueToValueMapTy &VMap);
125 class LoopUnswitch : public LoopPass {
126 LoopInfo *LI; // Loop information
129 // LoopProcessWorklist - Used to check if second loop needs processing
130 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
131 std::vector<Loop*> LoopProcessWorklist;
133 LUAnalysisCache BranchesInfo;
135 bool OptimizeForSize;
140 BasicBlock *loopHeader;
141 BasicBlock *loopPreheader;
143 // LoopBlocks contains all of the basic blocks of the loop, including the
144 // preheader of the loop, the body of the loop, and the exit blocks of the
145 // loop, in that order.
146 std::vector<BasicBlock*> LoopBlocks;
147 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
148 std::vector<BasicBlock*> NewBlocks;
151 static char ID; // Pass ID, replacement for typeid
152 explicit LoopUnswitch(bool Os = false) :
153 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
154 currentLoop(0), DT(0), loopHeader(0),
156 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
159 bool runOnLoop(Loop *L, LPPassManager &LPM);
160 bool processCurrentLoop();
162 /// This transformation requires natural loop information & requires that
163 /// loop preheaders be inserted into the CFG.
165 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
166 AU.addRequiredID(LoopSimplifyID);
167 AU.addPreservedID(LoopSimplifyID);
168 AU.addRequired<LoopInfo>();
169 AU.addPreserved<LoopInfo>();
170 AU.addRequiredID(LCSSAID);
171 AU.addPreservedID(LCSSAID);
172 AU.addPreserved<DominatorTreeWrapperPass>();
173 AU.addPreserved<ScalarEvolution>();
174 AU.addRequired<TargetTransformInfo>();
179 virtual void releaseMemory() {
180 BranchesInfo.forgetLoop(currentLoop);
183 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
185 void RemoveLoopFromWorklist(Loop *L) {
186 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
187 LoopProcessWorklist.end(), L);
188 if (I != LoopProcessWorklist.end())
189 LoopProcessWorklist.erase(I);
192 void initLoopData() {
193 loopHeader = currentLoop->getHeader();
194 loopPreheader = currentLoop->getLoopPreheader();
197 /// Split all of the edges from inside the loop to their exit blocks.
198 /// Update the appropriate Phi nodes as we do so.
199 void SplitExitEdges(Loop *L, const SmallVectorImpl<BasicBlock *> &ExitBlocks);
201 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
202 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
203 BasicBlock *ExitBlock);
204 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
206 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
207 Constant *Val, bool isEqual);
209 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
210 BasicBlock *TrueDest,
211 BasicBlock *FalseDest,
212 Instruction *InsertPt);
214 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
215 void RemoveLoopFromHierarchy(Loop *L);
216 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
217 BasicBlock **LoopExit = 0);
222 // Analyze loop. Check its size, calculate is it possible to unswitch
223 // it. Returns true if we can unswitch this loop.
224 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI) {
226 LoopPropsMapIt PropsIt;
228 llvm::tie(PropsIt, Inserted) =
229 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
231 LoopProperties &Props = PropsIt->second;
236 // Limit the number of instructions to avoid causing significant code
237 // expansion, and the number of basic blocks, to avoid loops with
238 // large numbers of branches which cause loop unswitching to go crazy.
239 // This is a very ad-hoc heuristic.
241 // FIXME: This is overly conservative because it does not take into
242 // consideration code simplification opportunities and code that can
243 // be shared by the resultant unswitched loops.
245 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
247 Metrics.analyzeBasicBlock(*I, TTI);
249 Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5);
250 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
251 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
253 if (Metrics.notDuplicatable) {
254 DEBUG(dbgs() << "NOT unswitching loop %"
255 << L->getHeader()->getName() << ", contents cannot be "
261 if (!Props.CanBeUnswitchedCount) {
262 DEBUG(dbgs() << "NOT unswitching loop %"
263 << L->getHeader()->getName() << ", cost too high: "
264 << L->getBlocks().size() << "\n");
268 // Be careful. This links are good only before new loop addition.
269 CurrentLoopProperties = &Props;
270 CurLoopInstructions = &Props.UnswitchedVals;
275 // Clean all data related to given loop.
276 void LUAnalysisCache::forgetLoop(const Loop *L) {
278 LoopPropsMapIt LIt = LoopsProperties.find(L);
280 if (LIt != LoopsProperties.end()) {
281 LoopProperties &Props = LIt->second;
282 MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation;
283 LoopsProperties.erase(LIt);
286 CurrentLoopProperties = 0;
287 CurLoopInstructions = 0;
290 // Mark case value as unswitched.
291 // Since SI instruction can be partly unswitched, in order to avoid
292 // extra unswitching in cloned loops keep track all unswitched values.
293 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
294 (*CurLoopInstructions)[SI].insert(V);
297 // Check was this case value unswitched before or not.
298 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
299 return (*CurLoopInstructions)[SI].count(V);
302 // Clone all loop-unswitch related loop properties.
303 // Redistribute unswitching quotas.
304 // Note, that new loop data is stored inside the VMap.
305 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
306 const ValueToValueMapTy &VMap) {
308 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
309 LoopProperties &OldLoopProps = *CurrentLoopProperties;
310 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
312 // Reallocate "can-be-unswitched quota"
314 --OldLoopProps.CanBeUnswitchedCount;
315 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
316 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
317 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
319 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
321 // Clone unswitched values info:
322 // for new loop switches we clone info about values that was
323 // already unswitched and has redundant successors.
324 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
325 const SwitchInst *OldInst = I->first;
326 Value *NewI = VMap.lookup(OldInst);
327 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
328 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
330 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
334 char LoopUnswitch::ID = 0;
335 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
337 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
338 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
339 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
340 INITIALIZE_PASS_DEPENDENCY(LCSSA)
341 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
344 Pass *llvm::createLoopUnswitchPass(bool Os) {
345 return new LoopUnswitch(Os);
348 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
349 /// invariant in the loop, or has an invariant piece, return the invariant.
350 /// Otherwise, return null.
351 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
353 // We started analyze new instruction, increment scanned instructions counter.
356 // We can never unswitch on vector conditions.
357 if (Cond->getType()->isVectorTy())
360 // Constants should be folded, not unswitched on!
361 if (isa<Constant>(Cond)) return 0;
363 // TODO: Handle: br (VARIANT|INVARIANT).
365 // Hoist simple values out.
366 if (L->makeLoopInvariant(Cond, Changed))
369 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
370 if (BO->getOpcode() == Instruction::And ||
371 BO->getOpcode() == Instruction::Or) {
372 // If either the left or right side is invariant, we can unswitch on this,
373 // which will cause the branch to go away in one loop and the condition to
374 // simplify in the other one.
375 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
377 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
384 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
385 if (skipOptnoneFunction(L))
388 LI = &getAnalysis<LoopInfo>();
390 DominatorTreeWrapperPass *DTWP =
391 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
392 DT = DTWP ? &DTWP->getDomTree() : 0;
394 Function *F = currentLoop->getHeader()->getParent();
395 bool Changed = false;
397 assert(currentLoop->isLCSSAForm(*DT));
399 Changed |= processCurrentLoop();
403 // FIXME: Reconstruct dom info, because it is not preserved properly.
410 /// processCurrentLoop - Do actual work and unswitch loop if possible
412 bool LoopUnswitch::processCurrentLoop() {
413 bool Changed = false;
417 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
421 // Loops with indirectbr cannot be cloned.
422 if (!currentLoop->isSafeToClone())
425 // Without dedicated exits, splitting the exit edge may fail.
426 if (!currentLoop->hasDedicatedExits())
429 LLVMContext &Context = loopHeader->getContext();
431 // Probably we reach the quota of branches for this loop. If so
433 if (!BranchesInfo.countLoop(currentLoop, getAnalysis<TargetTransformInfo>()))
436 // Loop over all of the basic blocks in the loop. If we find an interior
437 // block that is branching on a loop-invariant condition, we can unswitch this
439 for (Loop::block_iterator I = currentLoop->block_begin(),
440 E = currentLoop->block_end(); I != E; ++I) {
441 TerminatorInst *TI = (*I)->getTerminator();
442 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
443 // If this isn't branching on an invariant condition, we can't unswitch
445 if (BI->isConditional()) {
446 // See if this, or some part of it, is loop invariant. If so, we can
447 // unswitch on it if we desire.
448 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
449 currentLoop, Changed);
450 if (LoopCond && UnswitchIfProfitable(LoopCond,
451 ConstantInt::getTrue(Context))) {
456 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
457 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
458 currentLoop, Changed);
459 unsigned NumCases = SI->getNumCases();
460 if (LoopCond && NumCases) {
461 // Find a value to unswitch on:
462 // FIXME: this should chose the most expensive case!
463 // FIXME: scan for a case with a non-critical edge?
464 Constant *UnswitchVal = 0;
466 // Do not process same value again and again.
467 // At this point we have some cases already unswitched and
468 // some not yet unswitched. Let's find the first not yet unswitched one.
469 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
471 Constant *UnswitchValCandidate = i.getCaseValue();
472 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
473 UnswitchVal = UnswitchValCandidate;
481 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
488 // Scan the instructions to check for unswitchable values.
489 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
491 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
492 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
493 currentLoop, Changed);
494 if (LoopCond && UnswitchIfProfitable(LoopCond,
495 ConstantInt::getTrue(Context))) {
504 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
505 /// loop with no side effects (including infinite loops).
507 /// If true, we return true and set ExitBB to the block we
510 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
512 std::set<BasicBlock*> &Visited) {
513 if (!Visited.insert(BB).second) {
514 // Already visited. Without more analysis, this could indicate an infinite
518 if (!L->contains(BB)) {
519 // Otherwise, this is a loop exit, this is fine so long as this is the
521 if (ExitBB != 0) return false;
526 // Otherwise, this is an unvisited intra-loop node. Check all successors.
527 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
528 // Check to see if the successor is a trivial loop exit.
529 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
533 // Okay, everything after this looks good, check to make sure that this block
534 // doesn't include any side effects.
535 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
536 if (I->mayHaveSideEffects())
542 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
543 /// leads to an exit from the specified loop, and has no side-effects in the
544 /// process. If so, return the block that is exited to, otherwise return null.
545 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
546 std::set<BasicBlock*> Visited;
547 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
548 BasicBlock *ExitBB = 0;
549 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
554 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
555 /// trivial: that is, that the condition controls whether or not the loop does
556 /// anything at all. If this is a trivial condition, unswitching produces no
557 /// code duplications (equivalently, it produces a simpler loop and a new empty
558 /// loop, which gets deleted).
560 /// If this is a trivial condition, return true, otherwise return false. When
561 /// returning true, this sets Cond and Val to the condition that controls the
562 /// trivial condition: when Cond dynamically equals Val, the loop is known to
563 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
566 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
567 BasicBlock **LoopExit) {
568 BasicBlock *Header = currentLoop->getHeader();
569 TerminatorInst *HeaderTerm = Header->getTerminator();
570 LLVMContext &Context = Header->getContext();
572 BasicBlock *LoopExitBB = 0;
573 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
574 // If the header block doesn't end with a conditional branch on Cond, we
576 if (!BI->isConditional() || BI->getCondition() != Cond)
579 // Check to see if a successor of the branch is guaranteed to
580 // exit through a unique exit block without having any
581 // side-effects. If so, determine the value of Cond that causes it to do
583 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
584 BI->getSuccessor(0)))) {
585 if (Val) *Val = ConstantInt::getTrue(Context);
586 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
587 BI->getSuccessor(1)))) {
588 if (Val) *Val = ConstantInt::getFalse(Context);
590 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
591 // If this isn't a switch on Cond, we can't handle it.
592 if (SI->getCondition() != Cond) return false;
594 // Check to see if a successor of the switch is guaranteed to go to the
595 // latch block or exit through a one exit block without having any
596 // side-effects. If so, determine the value of Cond that causes it to do
598 // Note that we can't trivially unswitch on the default case or
599 // on already unswitched cases.
600 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
602 BasicBlock *LoopExitCandidate;
603 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
604 i.getCaseSuccessor()))) {
605 // Okay, we found a trivial case, remember the value that is trivial.
606 ConstantInt *CaseVal = i.getCaseValue();
608 // Check that it was not unswitched before, since already unswitched
609 // trivial vals are looks trivial too.
610 if (BranchesInfo.isUnswitched(SI, CaseVal))
612 LoopExitBB = LoopExitCandidate;
613 if (Val) *Val = CaseVal;
619 // If we didn't find a single unique LoopExit block, or if the loop exit block
620 // contains phi nodes, this isn't trivial.
621 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
622 return false; // Can't handle this.
624 if (LoopExit) *LoopExit = LoopExitBB;
626 // We already know that nothing uses any scalar values defined inside of this
627 // loop. As such, we just have to check to see if this loop will execute any
628 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
629 // part of the loop that the code *would* execute. We already checked the
630 // tail, check the header now.
631 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
632 if (I->mayHaveSideEffects())
637 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
638 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
639 /// unswitch the loop, reprocess the pieces, then return true.
640 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
641 Function *F = loopHeader->getParent();
642 Constant *CondVal = 0;
643 BasicBlock *ExitBlock = 0;
645 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
646 // If the condition is trivial, always unswitch. There is no code growth
648 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
652 // Check to see if it would be profitable to unswitch current loop.
654 // Do not do non-trivial unswitch while optimizing for size.
655 if (OptimizeForSize ||
656 F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
657 Attribute::OptimizeForSize))
660 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
664 /// CloneLoop - Recursively clone the specified loop and all of its children,
665 /// mapping the blocks with the specified map.
666 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
667 LoopInfo *LI, LPPassManager *LPM) {
668 Loop *New = new Loop();
669 LPM->insertLoop(New, PL);
671 // Add all of the blocks in L to the new loop.
672 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
674 if (LI->getLoopFor(*I) == L)
675 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
677 // Add all of the subloops to the new loop.
678 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
679 CloneLoop(*I, New, VM, LI, LPM);
684 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
685 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
686 /// code immediately before InsertPt.
687 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
688 BasicBlock *TrueDest,
689 BasicBlock *FalseDest,
690 Instruction *InsertPt) {
691 // Insert a conditional branch on LIC to the two preheaders. The original
692 // code is the true version and the new code is the false version.
693 Value *BranchVal = LIC;
694 if (!isa<ConstantInt>(Val) ||
695 Val->getType() != Type::getInt1Ty(LIC->getContext()))
696 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
697 else if (Val != ConstantInt::getTrue(Val->getContext()))
698 // We want to enter the new loop when the condition is true.
699 std::swap(TrueDest, FalseDest);
701 // Insert the new branch.
702 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
704 // If either edge is critical, split it. This helps preserve LoopSimplify
705 // form for enclosing loops.
706 SplitCriticalEdge(BI, 0, this, false, false, true);
707 SplitCriticalEdge(BI, 1, this, false, false, true);
710 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
711 /// condition in it (a cond branch from its header block to its latch block,
712 /// where the path through the loop that doesn't execute its body has no
713 /// side-effects), unswitch it. This doesn't involve any code duplication, just
714 /// moving the conditional branch outside of the loop and updating loop info.
715 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
717 BasicBlock *ExitBlock) {
718 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
719 << loopHeader->getName() << " [" << L->getBlocks().size()
720 << " blocks] in Function " << L->getHeader()->getParent()->getName()
721 << " on cond: " << *Val << " == " << *Cond << "\n");
723 // First step, split the preheader, so that we know that there is a safe place
724 // to insert the conditional branch. We will change loopPreheader to have a
725 // conditional branch on Cond.
726 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
728 // Now that we have a place to insert the conditional branch, create a place
729 // to branch to: this is the exit block out of the loop that we should
732 // Split this block now, so that the loop maintains its exit block, and so
733 // that the jump from the preheader can execute the contents of the exit block
734 // without actually branching to it (the exit block should be dominated by the
735 // loop header, not the preheader).
736 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
737 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
739 // Okay, now we have a position to branch from and a position to branch to,
740 // insert the new conditional branch.
741 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
742 loopPreheader->getTerminator());
743 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
744 loopPreheader->getTerminator()->eraseFromParent();
746 // We need to reprocess this loop, it could be unswitched again.
749 // Now that we know that the loop is never entered when this condition is a
750 // particular value, rewrite the loop with this info. We know that this will
751 // at least eliminate the old branch.
752 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
756 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
757 /// blocks. Update the appropriate Phi nodes as we do so.
758 void LoopUnswitch::SplitExitEdges(Loop *L,
759 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
761 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
762 BasicBlock *ExitBlock = ExitBlocks[i];
763 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
764 pred_end(ExitBlock));
766 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
767 // general, if we call it on all predecessors of all exits then it does.
768 if (!ExitBlock->isLandingPad()) {
769 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
771 SmallVector<BasicBlock*, 2> NewBBs;
772 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
778 /// UnswitchNontrivialCondition - We determined that the loop is profitable
779 /// to unswitch when LIC equal Val. Split it into loop versions and test the
780 /// condition outside of either loop. Return the loops created as Out1/Out2.
781 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
783 Function *F = loopHeader->getParent();
784 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
785 << loopHeader->getName() << " [" << L->getBlocks().size()
786 << " blocks] in Function " << F->getName()
787 << " when '" << *Val << "' == " << *LIC << "\n");
789 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
795 // First step, split the preheader and exit blocks, and add these blocks to
796 // the LoopBlocks list.
797 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
798 LoopBlocks.push_back(NewPreheader);
800 // We want the loop to come after the preheader, but before the exit blocks.
801 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
803 SmallVector<BasicBlock*, 8> ExitBlocks;
804 L->getUniqueExitBlocks(ExitBlocks);
806 // Split all of the edges from inside the loop to their exit blocks. Update
807 // the appropriate Phi nodes as we do so.
808 SplitExitEdges(L, ExitBlocks);
810 // The exit blocks may have been changed due to edge splitting, recompute.
812 L->getUniqueExitBlocks(ExitBlocks);
814 // Add exit blocks to the loop blocks.
815 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
817 // Next step, clone all of the basic blocks that make up the loop (including
818 // the loop preheader and exit blocks), keeping track of the mapping between
819 // the instructions and blocks.
820 NewBlocks.reserve(LoopBlocks.size());
821 ValueToValueMapTy VMap;
822 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
823 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
825 NewBlocks.push_back(NewBB);
826 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
827 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
830 // Splice the newly inserted blocks into the function right before the
831 // original preheader.
832 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
833 NewBlocks[0], F->end());
835 // Now we create the new Loop object for the versioned loop.
836 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
838 // Recalculate unswitching quota, inherit simplified switches info for NewBB,
839 // Probably clone more loop-unswitch related loop properties.
840 BranchesInfo.cloneData(NewLoop, L, VMap);
842 Loop *ParentLoop = L->getParentLoop();
844 // Make sure to add the cloned preheader and exit blocks to the parent loop
846 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
849 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
850 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
851 // The new exit block should be in the same loop as the old one.
852 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
853 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
855 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
856 "Exit block should have been split to have one successor!");
857 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
859 // If the successor of the exit block had PHI nodes, add an entry for
861 for (BasicBlock::iterator I = ExitSucc->begin();
862 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
863 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
864 ValueToValueMapTy::iterator It = VMap.find(V);
865 if (It != VMap.end()) V = It->second;
866 PN->addIncoming(V, NewExit);
869 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
870 PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
871 ExitSucc->getFirstInsertionPt());
873 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
876 LandingPadInst *LPI = BB->getLandingPadInst();
877 LPI->replaceAllUsesWith(PN);
878 PN->addIncoming(LPI, BB);
883 // Rewrite the code to refer to itself.
884 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
885 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
886 E = NewBlocks[i]->end(); I != E; ++I)
887 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
889 // Rewrite the original preheader to select between versions of the loop.
890 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
891 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
892 "Preheader splitting did not work correctly!");
894 // Emit the new branch that selects between the two versions of this loop.
895 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
896 LPM->deleteSimpleAnalysisValue(OldBR, L);
897 OldBR->eraseFromParent();
899 LoopProcessWorklist.push_back(NewLoop);
902 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
903 // deletes the instruction (for example by simplifying a PHI that feeds into
904 // the condition that we're unswitching on), we don't rewrite the second
906 WeakVH LICHandle(LIC);
908 // Now we rewrite the original code to know that the condition is true and the
909 // new code to know that the condition is false.
910 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
912 // It's possible that simplifying one loop could cause the other to be
913 // changed to another value or a constant. If its a constant, don't simplify
915 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
916 LICHandle && !isa<Constant>(LICHandle))
917 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
920 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
922 static void RemoveFromWorklist(Instruction *I,
923 std::vector<Instruction*> &Worklist) {
925 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
929 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
930 /// program, replacing all uses with V and update the worklist.
931 static void ReplaceUsesOfWith(Instruction *I, Value *V,
932 std::vector<Instruction*> &Worklist,
933 Loop *L, LPPassManager *LPM) {
934 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
936 // Add uses to the worklist, which may be dead now.
937 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
938 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
939 Worklist.push_back(Use);
941 // Add users to the worklist which may be simplified now.
942 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
944 Worklist.push_back(cast<Instruction>(*UI));
945 LPM->deleteSimpleAnalysisValue(I, L);
946 RemoveFromWorklist(I, Worklist);
947 I->replaceAllUsesWith(V);
948 I->eraseFromParent();
952 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
953 /// become unwrapped, either because the backedge was deleted, or because the
954 /// edge into the header was removed. If the edge into the header from the
955 /// latch block was removed, the loop is unwrapped but subloops are still alive,
956 /// so they just reparent loops. If the loops are actually dead, they will be
958 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
959 LPM->deleteLoopFromQueue(L);
960 RemoveLoopFromWorklist(L);
963 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
964 // the value specified by Val in the specified loop, or we know it does NOT have
965 // that value. Rewrite any uses of LIC or of properties correlated to it.
966 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
969 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
971 // FIXME: Support correlated properties, like:
978 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
979 // selects, switches.
980 std::vector<Instruction*> Worklist;
981 LLVMContext &Context = Val->getContext();
983 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
984 // in the loop with the appropriate one directly.
985 if (IsEqual || (isa<ConstantInt>(Val) &&
986 Val->getType()->isIntegerTy(1))) {
991 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
992 !cast<ConstantInt>(Val)->getZExtValue());
994 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
996 Instruction *U = dyn_cast<Instruction>(*UI);
997 if (!U || !L->contains(U))
999 Worklist.push_back(U);
1002 for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
1003 UE = Worklist.end(); UI != UE; ++UI)
1004 (*UI)->replaceUsesOfWith(LIC, Replacement);
1006 SimplifyCode(Worklist, L);
1010 // Otherwise, we don't know the precise value of LIC, but we do know that it
1011 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1012 // can. This case occurs when we unswitch switch statements.
1013 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
1015 Instruction *U = dyn_cast<Instruction>(*UI);
1016 if (!U || !L->contains(U))
1019 Worklist.push_back(U);
1021 // TODO: We could do other simplifications, for example, turning
1022 // 'icmp eq LIC, Val' -> false.
1024 // If we know that LIC is not Val, use this info to simplify code.
1025 SwitchInst *SI = dyn_cast<SwitchInst>(U);
1026 if (SI == 0 || !isa<ConstantInt>(Val)) continue;
1028 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1029 // Default case is live for multiple values.
1030 if (DeadCase == SI->case_default()) continue;
1032 // Found a dead case value. Don't remove PHI nodes in the
1033 // successor if they become single-entry, those PHI nodes may
1034 // be in the Users list.
1036 BasicBlock *Switch = SI->getParent();
1037 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1038 BasicBlock *Latch = L->getLoopLatch();
1040 BranchesInfo.setUnswitched(SI, Val);
1042 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1043 // If the DeadCase successor dominates the loop latch, then the
1044 // transformation isn't safe since it will delete the sole predecessor edge
1046 if (Latch && DT->dominates(SISucc, Latch))
1049 // FIXME: This is a hack. We need to keep the successor around
1050 // and hooked up so as to preserve the loop structure, because
1051 // trying to update it is complicated. So instead we preserve the
1052 // loop structure and put the block on a dead code path.
1053 SplitEdge(Switch, SISucc, this);
1054 // Compute the successors instead of relying on the return value
1055 // of SplitEdge, since it may have split the switch successor
1057 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1058 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1059 // Create an "unreachable" destination.
1060 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1061 Switch->getParent(),
1063 new UnreachableInst(Context, Abort);
1064 // Force the new case destination to branch to the "unreachable"
1065 // block while maintaining a (dead) CFG edge to the old block.
1066 NewSISucc->getTerminator()->eraseFromParent();
1067 BranchInst::Create(Abort, OldSISucc,
1068 ConstantInt::getTrue(Context), NewSISucc);
1069 // Release the PHI operands for this edge.
1070 for (BasicBlock::iterator II = NewSISucc->begin();
1071 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1072 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1073 UndefValue::get(PN->getType()));
1074 // Tell the domtree about the new block. We don't fully update the
1075 // domtree here -- instead we force it to do a full recomputation
1076 // after the pass is complete -- but we do need to inform it of
1079 DT->addNewBlock(Abort, NewSISucc);
1082 SimplifyCode(Worklist, L);
1085 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1086 /// loop, walk over it and constant prop, dce, and fold control flow where
1087 /// possible. Note that this is effectively a very simple loop-structure-aware
1088 /// optimizer. During processing of this loop, L could very well be deleted, so
1089 /// it must not be used.
1091 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1094 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1095 while (!Worklist.empty()) {
1096 Instruction *I = Worklist.back();
1097 Worklist.pop_back();
1100 if (isInstructionTriviallyDead(I)) {
1101 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1103 // Add uses to the worklist, which may be dead now.
1104 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1105 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1106 Worklist.push_back(Use);
1107 LPM->deleteSimpleAnalysisValue(I, L);
1108 RemoveFromWorklist(I, Worklist);
1109 I->eraseFromParent();
1114 // See if instruction simplification can hack this up. This is common for
1115 // things like "select false, X, Y" after unswitching made the condition be
1116 // 'false'. TODO: update the domtree properly so we can pass it here.
1117 if (Value *V = SimplifyInstruction(I))
1118 if (LI->replacementPreservesLCSSAForm(I, V)) {
1119 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1123 // Special case hacks that appear commonly in unswitched code.
1124 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1125 if (BI->isUnconditional()) {
1126 // If BI's parent is the only pred of the successor, fold the two blocks
1128 BasicBlock *Pred = BI->getParent();
1129 BasicBlock *Succ = BI->getSuccessor(0);
1130 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1131 if (!SinglePred) continue; // Nothing to do.
1132 assert(SinglePred == Pred && "CFG broken");
1134 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1135 << Succ->getName() << "\n");
1137 // Resolve any single entry PHI nodes in Succ.
1138 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1139 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1141 // If Succ has any successors with PHI nodes, update them to have
1142 // entries coming from Pred instead of Succ.
1143 Succ->replaceAllUsesWith(Pred);
1145 // Move all of the successor contents from Succ to Pred.
1146 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1148 LPM->deleteSimpleAnalysisValue(BI, L);
1149 BI->eraseFromParent();
1150 RemoveFromWorklist(BI, Worklist);
1152 // Remove Succ from the loop tree.
1153 LI->removeBlock(Succ);
1154 LPM->deleteSimpleAnalysisValue(Succ, L);
1155 Succ->eraseFromParent();