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 #include "llvm/Transforms/Scalar.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/AssumptionTracker.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 #define DEBUG_TYPE "loop-unswitch"
58 STATISTIC(NumBranches, "Number of branches unswitched");
59 STATISTIC(NumSwitches, "Number of switches unswitched");
60 STATISTIC(NumSelects , "Number of selects unswitched");
61 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
62 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
63 STATISTIC(TotalInsts, "Total number of instructions analyzed");
65 // The specific value of 100 here was chosen based only on intuition and a
66 // few specific examples.
67 static cl::opt<unsigned>
68 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
69 cl::init(100), cl::Hidden);
73 class LUAnalysisCache {
75 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
78 typedef UnswitchedValsMap::iterator UnswitchedValsIt;
80 struct LoopProperties {
81 unsigned CanBeUnswitchedCount;
82 unsigned SizeEstimation;
83 UnswitchedValsMap UnswitchedVals;
86 // Here we use std::map instead of DenseMap, since we need to keep valid
87 // LoopProperties pointer for current loop for better performance.
88 typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
89 typedef LoopPropsMap::iterator LoopPropsMapIt;
91 LoopPropsMap LoopsProperties;
92 UnswitchedValsMap *CurLoopInstructions;
93 LoopProperties *CurrentLoopProperties;
95 // Max size of code we can produce on remained iterations.
101 CurLoopInstructions(nullptr), CurrentLoopProperties(nullptr),
105 // Analyze loop. Check its size, calculate is it possible to unswitch
106 // it. Returns true if we can unswitch this loop.
107 bool countLoop(const Loop *L, const TargetTransformInfo &TTI);
109 // Clean all data related to given loop.
110 void forgetLoop(const Loop *L);
112 // Mark case value as unswitched.
113 // Since SI instruction can be partly unswitched, in order to avoid
114 // extra unswitching in cloned loops keep track all unswitched values.
115 void setUnswitched(const SwitchInst *SI, const Value *V);
117 // Check was this case value unswitched before or not.
118 bool isUnswitched(const SwitchInst *SI, const Value *V);
120 // Clone all loop-unswitch related loop properties.
121 // Redistribute unswitching quotas.
122 // Note, that new loop data is stored inside the VMap.
123 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
124 const ValueToValueMapTy &VMap);
127 class LoopUnswitch : public LoopPass {
128 LoopInfo *LI; // Loop information
130 AssumptionTracker *AT;
132 // LoopProcessWorklist - Used to check if second loop needs processing
133 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
134 std::vector<Loop*> LoopProcessWorklist;
136 LUAnalysisCache BranchesInfo;
138 bool OptimizeForSize;
143 BasicBlock *loopHeader;
144 BasicBlock *loopPreheader;
146 // LoopBlocks contains all of the basic blocks of the loop, including the
147 // preheader of the loop, the body of the loop, and the exit blocks of the
148 // loop, in that order.
149 std::vector<BasicBlock*> LoopBlocks;
150 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
151 std::vector<BasicBlock*> NewBlocks;
154 static char ID; // Pass ID, replacement for typeid
155 explicit LoopUnswitch(bool Os = false) :
156 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
157 currentLoop(nullptr), DT(nullptr), loopHeader(nullptr),
158 loopPreheader(nullptr) {
159 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
162 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
163 bool processCurrentLoop();
165 /// This transformation requires natural loop information & requires that
166 /// loop preheaders be inserted into the CFG.
168 void getAnalysisUsage(AnalysisUsage &AU) const override {
169 AU.addRequired<AssumptionTracker>();
170 AU.addRequiredID(LoopSimplifyID);
171 AU.addPreservedID(LoopSimplifyID);
172 AU.addRequired<LoopInfo>();
173 AU.addPreserved<LoopInfo>();
174 AU.addRequiredID(LCSSAID);
175 AU.addPreservedID(LCSSAID);
176 AU.addPreserved<DominatorTreeWrapperPass>();
177 AU.addPreserved<ScalarEvolution>();
178 AU.addRequired<TargetTransformInfo>();
183 void releaseMemory() override {
184 BranchesInfo.forgetLoop(currentLoop);
187 void initLoopData() {
188 loopHeader = currentLoop->getHeader();
189 loopPreheader = currentLoop->getLoopPreheader();
192 /// Split all of the edges from inside the loop to their exit blocks.
193 /// Update the appropriate Phi nodes as we do so.
194 void SplitExitEdges(Loop *L, const SmallVectorImpl<BasicBlock *> &ExitBlocks);
196 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
197 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
198 BasicBlock *ExitBlock);
199 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
201 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
202 Constant *Val, bool isEqual);
204 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
205 BasicBlock *TrueDest,
206 BasicBlock *FalseDest,
207 Instruction *InsertPt);
209 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
210 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = nullptr,
211 BasicBlock **LoopExit = nullptr);
216 // Analyze loop. Check its size, calculate is it possible to unswitch
217 // it. Returns true if we can unswitch this loop.
218 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI) {
220 LoopPropsMapIt PropsIt;
222 std::tie(PropsIt, Inserted) =
223 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
225 LoopProperties &Props = PropsIt->second;
230 // Limit the number of instructions to avoid causing significant code
231 // expansion, and the number of basic blocks, to avoid loops with
232 // large numbers of branches which cause loop unswitching to go crazy.
233 // This is a very ad-hoc heuristic.
235 // FIXME: This is overly conservative because it does not take into
236 // consideration code simplification opportunities and code that can
237 // be shared by the resultant unswitched loops.
239 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
241 Metrics.analyzeBasicBlock(*I, TTI);
243 Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5);
244 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
245 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
247 if (Metrics.notDuplicatable) {
248 DEBUG(dbgs() << "NOT unswitching loop %"
249 << L->getHeader()->getName() << ", contents cannot be "
255 if (!Props.CanBeUnswitchedCount) {
256 DEBUG(dbgs() << "NOT unswitching loop %"
257 << L->getHeader()->getName() << ", cost too high: "
258 << L->getBlocks().size() << "\n");
262 // Be careful. This links are good only before new loop addition.
263 CurrentLoopProperties = &Props;
264 CurLoopInstructions = &Props.UnswitchedVals;
269 // Clean all data related to given loop.
270 void LUAnalysisCache::forgetLoop(const Loop *L) {
272 LoopPropsMapIt LIt = LoopsProperties.find(L);
274 if (LIt != LoopsProperties.end()) {
275 LoopProperties &Props = LIt->second;
276 MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation;
277 LoopsProperties.erase(LIt);
280 CurrentLoopProperties = nullptr;
281 CurLoopInstructions = nullptr;
284 // Mark case value as unswitched.
285 // Since SI instruction can be partly unswitched, in order to avoid
286 // extra unswitching in cloned loops keep track all unswitched values.
287 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
288 (*CurLoopInstructions)[SI].insert(V);
291 // Check was this case value unswitched before or not.
292 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
293 return (*CurLoopInstructions)[SI].count(V);
296 // Clone all loop-unswitch related loop properties.
297 // Redistribute unswitching quotas.
298 // Note, that new loop data is stored inside the VMap.
299 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
300 const ValueToValueMapTy &VMap) {
302 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
303 LoopProperties &OldLoopProps = *CurrentLoopProperties;
304 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
306 // Reallocate "can-be-unswitched quota"
308 --OldLoopProps.CanBeUnswitchedCount;
309 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
310 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
311 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
313 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
315 // Clone unswitched values info:
316 // for new loop switches we clone info about values that was
317 // already unswitched and has redundant successors.
318 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
319 const SwitchInst *OldInst = I->first;
320 Value *NewI = VMap.lookup(OldInst);
321 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
322 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
324 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
328 char LoopUnswitch::ID = 0;
329 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
331 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
332 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
333 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
334 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
335 INITIALIZE_PASS_DEPENDENCY(LCSSA)
336 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
339 Pass *llvm::createLoopUnswitchPass(bool Os) {
340 return new LoopUnswitch(Os);
343 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
344 /// invariant in the loop, or has an invariant piece, return the invariant.
345 /// Otherwise, return null.
346 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
348 // We started analyze new instruction, increment scanned instructions counter.
351 // We can never unswitch on vector conditions.
352 if (Cond->getType()->isVectorTy())
355 // Constants should be folded, not unswitched on!
356 if (isa<Constant>(Cond)) return nullptr;
358 // TODO: Handle: br (VARIANT|INVARIANT).
360 // Hoist simple values out.
361 if (L->makeLoopInvariant(Cond, Changed))
364 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
365 if (BO->getOpcode() == Instruction::And ||
366 BO->getOpcode() == Instruction::Or) {
367 // If either the left or right side is invariant, we can unswitch on this,
368 // which will cause the branch to go away in one loop and the condition to
369 // simplify in the other one.
370 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
372 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
379 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
380 if (skipOptnoneFunction(L))
383 AT = &getAnalysis<AssumptionTracker>();
384 LI = &getAnalysis<LoopInfo>();
386 DominatorTreeWrapperPass *DTWP =
387 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
388 DT = DTWP ? &DTWP->getDomTree() : nullptr;
390 Function *F = currentLoop->getHeader()->getParent();
391 bool Changed = false;
393 assert(currentLoop->isLCSSAForm(*DT));
395 Changed |= processCurrentLoop();
399 // FIXME: Reconstruct dom info, because it is not preserved properly.
406 /// processCurrentLoop - Do actual work and unswitch loop if possible
408 bool LoopUnswitch::processCurrentLoop() {
409 bool Changed = false;
413 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
417 // Loops with indirectbr cannot be cloned.
418 if (!currentLoop->isSafeToClone())
421 // Without dedicated exits, splitting the exit edge may fail.
422 if (!currentLoop->hasDedicatedExits())
425 LLVMContext &Context = loopHeader->getContext();
427 // Probably we reach the quota of branches for this loop. If so
429 if (!BranchesInfo.countLoop(currentLoop, getAnalysis<TargetTransformInfo>()))
432 // Loop over all of the basic blocks in the loop. If we find an interior
433 // block that is branching on a loop-invariant condition, we can unswitch this
435 for (Loop::block_iterator I = currentLoop->block_begin(),
436 E = currentLoop->block_end(); I != E; ++I) {
437 TerminatorInst *TI = (*I)->getTerminator();
438 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
439 // If this isn't branching on an invariant condition, we can't unswitch
441 if (BI->isConditional()) {
442 // See if this, or some part of it, is loop invariant. If so, we can
443 // unswitch on it if we desire.
444 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
445 currentLoop, Changed);
446 if (LoopCond && UnswitchIfProfitable(LoopCond,
447 ConstantInt::getTrue(Context))) {
452 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
453 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
454 currentLoop, Changed);
455 unsigned NumCases = SI->getNumCases();
456 if (LoopCond && NumCases) {
457 // Find a value to unswitch on:
458 // FIXME: this should chose the most expensive case!
459 // FIXME: scan for a case with a non-critical edge?
460 Constant *UnswitchVal = nullptr;
462 // Do not process same value again and again.
463 // At this point we have some cases already unswitched and
464 // some not yet unswitched. Let's find the first not yet unswitched one.
465 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
467 Constant *UnswitchValCandidate = i.getCaseValue();
468 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
469 UnswitchVal = UnswitchValCandidate;
477 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
484 // Scan the instructions to check for unswitchable values.
485 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
487 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
488 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
489 currentLoop, Changed);
490 if (LoopCond && UnswitchIfProfitable(LoopCond,
491 ConstantInt::getTrue(Context))) {
500 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
501 /// loop with no side effects (including infinite loops).
503 /// If true, we return true and set ExitBB to the block we
506 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
508 std::set<BasicBlock*> &Visited) {
509 if (!Visited.insert(BB).second) {
510 // Already visited. Without more analysis, this could indicate an infinite
514 if (!L->contains(BB)) {
515 // Otherwise, this is a loop exit, this is fine so long as this is the
517 if (ExitBB) return false;
522 // Otherwise, this is an unvisited intra-loop node. Check all successors.
523 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
524 // Check to see if the successor is a trivial loop exit.
525 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
529 // Okay, everything after this looks good, check to make sure that this block
530 // doesn't include any side effects.
531 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
532 if (I->mayHaveSideEffects())
538 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
539 /// leads to an exit from the specified loop, and has no side-effects in the
540 /// process. If so, return the block that is exited to, otherwise return null.
541 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
542 std::set<BasicBlock*> Visited;
543 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
544 BasicBlock *ExitBB = nullptr;
545 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
550 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
551 /// trivial: that is, that the condition controls whether or not the loop does
552 /// anything at all. If this is a trivial condition, unswitching produces no
553 /// code duplications (equivalently, it produces a simpler loop and a new empty
554 /// loop, which gets deleted).
556 /// If this is a trivial condition, return true, otherwise return false. When
557 /// returning true, this sets Cond and Val to the condition that controls the
558 /// trivial condition: when Cond dynamically equals Val, the loop is known to
559 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
562 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
563 BasicBlock **LoopExit) {
564 BasicBlock *Header = currentLoop->getHeader();
565 TerminatorInst *HeaderTerm = Header->getTerminator();
566 LLVMContext &Context = Header->getContext();
568 BasicBlock *LoopExitBB = nullptr;
569 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
570 // If the header block doesn't end with a conditional branch on Cond, we
572 if (!BI->isConditional() || BI->getCondition() != Cond)
575 // Check to see if a successor of the branch is guaranteed to
576 // exit through a unique exit block without having any
577 // side-effects. If so, determine the value of Cond that causes it to do
579 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
580 BI->getSuccessor(0)))) {
581 if (Val) *Val = ConstantInt::getTrue(Context);
582 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
583 BI->getSuccessor(1)))) {
584 if (Val) *Val = ConstantInt::getFalse(Context);
586 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
587 // If this isn't a switch on Cond, we can't handle it.
588 if (SI->getCondition() != Cond) return false;
590 // Check to see if a successor of the switch is guaranteed to go to the
591 // latch block or exit through a one exit block without having any
592 // side-effects. If so, determine the value of Cond that causes it to do
594 // Note that we can't trivially unswitch on the default case or
595 // on already unswitched cases.
596 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
598 BasicBlock *LoopExitCandidate;
599 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
600 i.getCaseSuccessor()))) {
601 // Okay, we found a trivial case, remember the value that is trivial.
602 ConstantInt *CaseVal = i.getCaseValue();
604 // Check that it was not unswitched before, since already unswitched
605 // trivial vals are looks trivial too.
606 if (BranchesInfo.isUnswitched(SI, CaseVal))
608 LoopExitBB = LoopExitCandidate;
609 if (Val) *Val = CaseVal;
615 // If we didn't find a single unique LoopExit block, or if the loop exit block
616 // contains phi nodes, this isn't trivial.
617 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
618 return false; // Can't handle this.
620 if (LoopExit) *LoopExit = LoopExitBB;
622 // We already know that nothing uses any scalar values defined inside of this
623 // loop. As such, we just have to check to see if this loop will execute any
624 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
625 // part of the loop that the code *would* execute. We already checked the
626 // tail, check the header now.
627 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
628 if (I->mayHaveSideEffects())
633 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
634 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
635 /// unswitch the loop, reprocess the pieces, then return true.
636 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
637 Function *F = loopHeader->getParent();
638 Constant *CondVal = nullptr;
639 BasicBlock *ExitBlock = nullptr;
641 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
642 // If the condition is trivial, always unswitch. There is no code growth
644 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
648 // Check to see if it would be profitable to unswitch current loop.
650 // Do not do non-trivial unswitch while optimizing for size.
651 if (OptimizeForSize ||
652 F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
653 Attribute::OptimizeForSize))
656 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
660 /// CloneLoop - Recursively clone the specified loop and all of its children,
661 /// mapping the blocks with the specified map.
662 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
663 LoopInfo *LI, LPPassManager *LPM) {
664 Loop *New = new Loop();
665 LPM->insertLoop(New, PL);
667 // Add all of the blocks in L to the new loop.
668 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
670 if (LI->getLoopFor(*I) == L)
671 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
673 // Add all of the subloops to the new loop.
674 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
675 CloneLoop(*I, New, VM, LI, LPM);
680 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
681 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
682 /// code immediately before InsertPt.
683 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
684 BasicBlock *TrueDest,
685 BasicBlock *FalseDest,
686 Instruction *InsertPt) {
687 // Insert a conditional branch on LIC to the two preheaders. The original
688 // code is the true version and the new code is the false version.
689 Value *BranchVal = LIC;
690 if (!isa<ConstantInt>(Val) ||
691 Val->getType() != Type::getInt1Ty(LIC->getContext()))
692 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
693 else if (Val != ConstantInt::getTrue(Val->getContext()))
694 // We want to enter the new loop when the condition is true.
695 std::swap(TrueDest, FalseDest);
697 // Insert the new branch.
698 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
700 // If either edge is critical, split it. This helps preserve LoopSimplify
701 // form for enclosing loops.
702 SplitCriticalEdge(BI, 0, this, false, false, true);
703 SplitCriticalEdge(BI, 1, this, false, false, true);
706 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
707 /// condition in it (a cond branch from its header block to its latch block,
708 /// where the path through the loop that doesn't execute its body has no
709 /// side-effects), unswitch it. This doesn't involve any code duplication, just
710 /// moving the conditional branch outside of the loop and updating loop info.
711 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
713 BasicBlock *ExitBlock) {
714 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
715 << loopHeader->getName() << " [" << L->getBlocks().size()
716 << " blocks] in Function " << L->getHeader()->getParent()->getName()
717 << " on cond: " << *Val << " == " << *Cond << "\n");
719 // First step, split the preheader, so that we know that there is a safe place
720 // to insert the conditional branch. We will change loopPreheader to have a
721 // conditional branch on Cond.
722 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
724 // Now that we have a place to insert the conditional branch, create a place
725 // to branch to: this is the exit block out of the loop that we should
728 // Split this block now, so that the loop maintains its exit block, and so
729 // that the jump from the preheader can execute the contents of the exit block
730 // without actually branching to it (the exit block should be dominated by the
731 // loop header, not the preheader).
732 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
733 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
735 // Okay, now we have a position to branch from and a position to branch to,
736 // insert the new conditional branch.
737 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
738 loopPreheader->getTerminator());
739 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
740 loopPreheader->getTerminator()->eraseFromParent();
742 // We need to reprocess this loop, it could be unswitched again.
745 // Now that we know that the loop is never entered when this condition is a
746 // particular value, rewrite the loop with this info. We know that this will
747 // at least eliminate the old branch.
748 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
752 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
753 /// blocks. Update the appropriate Phi nodes as we do so.
754 void LoopUnswitch::SplitExitEdges(Loop *L,
755 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
757 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
758 BasicBlock *ExitBlock = ExitBlocks[i];
759 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
760 pred_end(ExitBlock));
762 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
763 // general, if we call it on all predecessors of all exits then it does.
764 if (!ExitBlock->isLandingPad()) {
765 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
767 SmallVector<BasicBlock*, 2> NewBBs;
768 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
774 /// UnswitchNontrivialCondition - We determined that the loop is profitable
775 /// to unswitch when LIC equal Val. Split it into loop versions and test the
776 /// condition outside of either loop. Return the loops created as Out1/Out2.
777 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
779 Function *F = loopHeader->getParent();
780 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
781 << loopHeader->getName() << " [" << L->getBlocks().size()
782 << " blocks] in Function " << F->getName()
783 << " when '" << *Val << "' == " << *LIC << "\n");
785 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
791 // First step, split the preheader and exit blocks, and add these blocks to
792 // the LoopBlocks list.
793 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
794 LoopBlocks.push_back(NewPreheader);
796 // We want the loop to come after the preheader, but before the exit blocks.
797 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
799 SmallVector<BasicBlock*, 8> ExitBlocks;
800 L->getUniqueExitBlocks(ExitBlocks);
802 // Split all of the edges from inside the loop to their exit blocks. Update
803 // the appropriate Phi nodes as we do so.
804 SplitExitEdges(L, ExitBlocks);
806 // The exit blocks may have been changed due to edge splitting, recompute.
808 L->getUniqueExitBlocks(ExitBlocks);
810 // Add exit blocks to the loop blocks.
811 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
813 // Next step, clone all of the basic blocks that make up the loop (including
814 // the loop preheader and exit blocks), keeping track of the mapping between
815 // the instructions and blocks.
816 NewBlocks.reserve(LoopBlocks.size());
817 ValueToValueMapTy VMap;
818 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
819 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
821 NewBlocks.push_back(NewBB);
822 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
823 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
826 // Splice the newly inserted blocks into the function right before the
827 // original preheader.
828 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
829 NewBlocks[0], F->end());
831 // FIXME: We could register any cloned assumptions instead of clearing the
832 // whole function's cache.
833 AT->forgetCachedAssumptions(F);
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 (User *U : I->users())
943 Worklist.push_back(cast<Instruction>(U));
944 LPM->deleteSimpleAnalysisValue(I, L);
945 RemoveFromWorklist(I, Worklist);
946 I->replaceAllUsesWith(V);
947 I->eraseFromParent();
951 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
952 // the value specified by Val in the specified loop, or we know it does NOT have
953 // that value. Rewrite any uses of LIC or of properties correlated to it.
954 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
957 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
959 // FIXME: Support correlated properties, like:
966 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
967 // selects, switches.
968 std::vector<Instruction*> Worklist;
969 LLVMContext &Context = Val->getContext();
971 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
972 // in the loop with the appropriate one directly.
973 if (IsEqual || (isa<ConstantInt>(Val) &&
974 Val->getType()->isIntegerTy(1))) {
979 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
980 !cast<ConstantInt>(Val)->getZExtValue());
982 for (User *U : LIC->users()) {
983 Instruction *UI = dyn_cast<Instruction>(U);
984 if (!UI || !L->contains(UI))
986 Worklist.push_back(UI);
989 for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
990 UE = Worklist.end(); UI != UE; ++UI)
991 (*UI)->replaceUsesOfWith(LIC, Replacement);
993 SimplifyCode(Worklist, L);
997 // Otherwise, we don't know the precise value of LIC, but we do know that it
998 // is certainly NOT "Val". As such, simplify any uses in the loop that we
999 // can. This case occurs when we unswitch switch statements.
1000 for (User *U : LIC->users()) {
1001 Instruction *UI = dyn_cast<Instruction>(U);
1002 if (!UI || !L->contains(UI))
1005 Worklist.push_back(UI);
1007 // TODO: We could do other simplifications, for example, turning
1008 // 'icmp eq LIC, Val' -> false.
1010 // If we know that LIC is not Val, use this info to simplify code.
1011 SwitchInst *SI = dyn_cast<SwitchInst>(UI);
1012 if (!SI || !isa<ConstantInt>(Val)) continue;
1014 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1015 // Default case is live for multiple values.
1016 if (DeadCase == SI->case_default()) continue;
1018 // Found a dead case value. Don't remove PHI nodes in the
1019 // successor if they become single-entry, those PHI nodes may
1020 // be in the Users list.
1022 BasicBlock *Switch = SI->getParent();
1023 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1024 BasicBlock *Latch = L->getLoopLatch();
1026 BranchesInfo.setUnswitched(SI, Val);
1028 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1029 // If the DeadCase successor dominates the loop latch, then the
1030 // transformation isn't safe since it will delete the sole predecessor edge
1032 if (Latch && DT->dominates(SISucc, Latch))
1035 // FIXME: This is a hack. We need to keep the successor around
1036 // and hooked up so as to preserve the loop structure, because
1037 // trying to update it is complicated. So instead we preserve the
1038 // loop structure and put the block on a dead code path.
1039 SplitEdge(Switch, SISucc, this);
1040 // Compute the successors instead of relying on the return value
1041 // of SplitEdge, since it may have split the switch successor
1043 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1044 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1045 // Create an "unreachable" destination.
1046 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1047 Switch->getParent(),
1049 new UnreachableInst(Context, Abort);
1050 // Force the new case destination to branch to the "unreachable"
1051 // block while maintaining a (dead) CFG edge to the old block.
1052 NewSISucc->getTerminator()->eraseFromParent();
1053 BranchInst::Create(Abort, OldSISucc,
1054 ConstantInt::getTrue(Context), NewSISucc);
1055 // Release the PHI operands for this edge.
1056 for (BasicBlock::iterator II = NewSISucc->begin();
1057 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1058 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1059 UndefValue::get(PN->getType()));
1060 // Tell the domtree about the new block. We don't fully update the
1061 // domtree here -- instead we force it to do a full recomputation
1062 // after the pass is complete -- but we do need to inform it of
1065 DT->addNewBlock(Abort, NewSISucc);
1068 SimplifyCode(Worklist, L);
1071 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1072 /// loop, walk over it and constant prop, dce, and fold control flow where
1073 /// possible. Note that this is effectively a very simple loop-structure-aware
1074 /// optimizer. During processing of this loop, L could very well be deleted, so
1075 /// it must not be used.
1077 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1080 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1081 while (!Worklist.empty()) {
1082 Instruction *I = Worklist.back();
1083 Worklist.pop_back();
1086 if (isInstructionTriviallyDead(I)) {
1087 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1089 // Add uses to the worklist, which may be dead now.
1090 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1091 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1092 Worklist.push_back(Use);
1093 LPM->deleteSimpleAnalysisValue(I, L);
1094 RemoveFromWorklist(I, Worklist);
1095 I->eraseFromParent();
1100 // See if instruction simplification can hack this up. This is common for
1101 // things like "select false, X, Y" after unswitching made the condition be
1102 // 'false'. TODO: update the domtree properly so we can pass it here.
1103 if (Value *V = SimplifyInstruction(I))
1104 if (LI->replacementPreservesLCSSAForm(I, V)) {
1105 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1109 // Special case hacks that appear commonly in unswitched code.
1110 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1111 if (BI->isUnconditional()) {
1112 // If BI's parent is the only pred of the successor, fold the two blocks
1114 BasicBlock *Pred = BI->getParent();
1115 BasicBlock *Succ = BI->getSuccessor(0);
1116 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1117 if (!SinglePred) continue; // Nothing to do.
1118 assert(SinglePred == Pred && "CFG broken");
1120 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1121 << Succ->getName() << "\n");
1123 // Resolve any single entry PHI nodes in Succ.
1124 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1125 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1127 // If Succ has any successors with PHI nodes, update them to have
1128 // entries coming from Pred instead of Succ.
1129 Succ->replaceAllUsesWith(Pred);
1131 // Move all of the successor contents from Succ to Pred.
1132 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1134 LPM->deleteSimpleAnalysisValue(BI, L);
1135 BI->eraseFromParent();
1136 RemoveFromWorklist(BI, Worklist);
1138 // Remove Succ from the loop tree.
1139 LI->removeBlock(Succ);
1140 LPM->deleteSimpleAnalysisValue(Succ, L);
1141 Succ->eraseFromParent();