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/AssumptionCache.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/IR/Module.h"
46 #include "llvm/IR/MDBuilder.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Cloning.h"
52 #include "llvm/Transforms/Utils/Local.h"
58 #define DEBUG_TYPE "loop-unswitch"
60 STATISTIC(NumBranches, "Number of branches unswitched");
61 STATISTIC(NumSwitches, "Number of switches unswitched");
62 STATISTIC(NumSelects , "Number of selects unswitched");
63 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
64 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
65 STATISTIC(TotalInsts, "Total number of instructions analyzed");
67 // The specific value of 100 here was chosen based only on intuition and a
68 // few specific examples.
69 static cl::opt<unsigned>
70 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
71 cl::init(100), cl::Hidden);
75 class LUAnalysisCache {
77 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
80 typedef UnswitchedValsMap::iterator UnswitchedValsIt;
82 struct LoopProperties {
83 unsigned CanBeUnswitchedCount;
84 unsigned WasUnswitchedCount;
85 unsigned SizeEstimation;
86 UnswitchedValsMap UnswitchedVals;
89 // Here we use std::map instead of DenseMap, since we need to keep valid
90 // LoopProperties pointer for current loop for better performance.
91 typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
92 typedef LoopPropsMap::iterator LoopPropsMapIt;
94 LoopPropsMap LoopsProperties;
95 UnswitchedValsMap *CurLoopInstructions;
96 LoopProperties *CurrentLoopProperties;
98 // A loop unswitching with an estimated cost above this threshold
99 // is not performed. MaxSize is turned into unswitching quota for
100 // the current loop, and reduced correspondingly, though note that
101 // the quota is returned by releaseMemory() when the loop has been
102 // processed, so that MaxSize will return to its previous
103 // value. So in most cases MaxSize will equal the Threshold flag
104 // when a new loop is processed. An exception to that is that
105 // MaxSize will have a smaller value while processing nested loops
106 // that were introduced due to loop unswitching of an outer loop.
108 // FIXME: The way that MaxSize works is subtle and depends on the
109 // pass manager processing loops and calling releaseMemory() in a
110 // specific order. It would be good to find a more straightforward
111 // way of doing what MaxSize does.
116 : CurLoopInstructions(nullptr), CurrentLoopProperties(nullptr),
117 MaxSize(Threshold) {}
119 // Analyze loop. Check its size, calculate is it possible to unswitch
120 // it. Returns true if we can unswitch this loop.
121 bool countLoop(const Loop *L, const TargetTransformInfo &TTI,
122 AssumptionCache *AC);
124 // Clean all data related to given loop.
125 void forgetLoop(const Loop *L);
127 // Mark case value as unswitched.
128 // Since SI instruction can be partly unswitched, in order to avoid
129 // extra unswitching in cloned loops keep track all unswitched values.
130 void setUnswitched(const SwitchInst *SI, const Value *V);
132 // Check was this case value unswitched before or not.
133 bool isUnswitched(const SwitchInst *SI, const Value *V);
135 // Returns true if another unswitching could be done within the cost
137 bool CostAllowsUnswitching();
139 // Clone all loop-unswitch related loop properties.
140 // Redistribute unswitching quotas.
141 // Note, that new loop data is stored inside the VMap.
142 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
143 const ValueToValueMapTy &VMap);
146 class LoopUnswitch : public LoopPass {
147 LoopInfo *LI; // Loop information
151 // LoopProcessWorklist - Used to check if second loop needs processing
152 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
153 std::vector<Loop*> LoopProcessWorklist;
155 LUAnalysisCache BranchesInfo;
157 bool OptimizeForSize;
162 BasicBlock *loopHeader;
163 BasicBlock *loopPreheader;
165 // LoopBlocks contains all of the basic blocks of the loop, including the
166 // preheader of the loop, the body of the loop, and the exit blocks of the
167 // loop, in that order.
168 std::vector<BasicBlock*> LoopBlocks;
169 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
170 std::vector<BasicBlock*> NewBlocks;
173 static char ID; // Pass ID, replacement for typeid
174 explicit LoopUnswitch(bool Os = false) :
175 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
176 currentLoop(nullptr), DT(nullptr), loopHeader(nullptr),
177 loopPreheader(nullptr) {
178 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
181 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
182 bool processCurrentLoop();
184 /// This transformation requires natural loop information & requires that
185 /// loop preheaders be inserted into the CFG.
187 void getAnalysisUsage(AnalysisUsage &AU) const override {
188 AU.addRequired<AssumptionCacheTracker>();
189 AU.addRequiredID(LoopSimplifyID);
190 AU.addPreservedID(LoopSimplifyID);
191 AU.addRequired<LoopInfoWrapperPass>();
192 AU.addPreserved<LoopInfoWrapperPass>();
193 AU.addRequiredID(LCSSAID);
194 AU.addPreservedID(LCSSAID);
195 AU.addPreserved<DominatorTreeWrapperPass>();
196 AU.addPreserved<ScalarEvolution>();
197 AU.addRequired<TargetTransformInfoWrapperPass>();
202 void releaseMemory() override {
203 BranchesInfo.forgetLoop(currentLoop);
206 void initLoopData() {
207 loopHeader = currentLoop->getHeader();
208 loopPreheader = currentLoop->getLoopPreheader();
211 /// Split all of the edges from inside the loop to their exit blocks.
212 /// Update the appropriate Phi nodes as we do so.
213 void SplitExitEdges(Loop *L, const SmallVectorImpl<BasicBlock *> &ExitBlocks);
215 bool TryTrivialLoopUnswitch(bool &Changed);
217 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,
218 TerminatorInst *TI = nullptr);
219 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
220 BasicBlock *ExitBlock, TerminatorInst *TI);
221 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L,
224 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
225 Constant *Val, bool isEqual);
227 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
228 BasicBlock *TrueDest,
229 BasicBlock *FalseDest,
230 Instruction *InsertPt,
233 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
237 // Analyze loop. Check its size, calculate is it possible to unswitch
238 // it. Returns true if we can unswitch this loop.
239 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI,
240 AssumptionCache *AC) {
242 LoopPropsMapIt PropsIt;
244 std::tie(PropsIt, Inserted) =
245 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
247 LoopProperties &Props = PropsIt->second;
252 // Limit the number of instructions to avoid causing significant code
253 // expansion, and the number of basic blocks, to avoid loops with
254 // large numbers of branches which cause loop unswitching to go crazy.
255 // This is a very ad-hoc heuristic.
257 SmallPtrSet<const Value *, 32> EphValues;
258 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
260 // FIXME: This is overly conservative because it does not take into
261 // consideration code simplification opportunities and code that can
262 // be shared by the resultant unswitched loops.
264 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
266 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
268 Props.SizeEstimation = Metrics.NumInsts;
269 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
270 Props.WasUnswitchedCount = 0;
271 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
273 if (Metrics.notDuplicatable) {
274 DEBUG(dbgs() << "NOT unswitching loop %"
275 << L->getHeader()->getName() << ", contents cannot be "
281 // Be careful. This links are good only before new loop addition.
282 CurrentLoopProperties = &Props;
283 CurLoopInstructions = &Props.UnswitchedVals;
288 // Clean all data related to given loop.
289 void LUAnalysisCache::forgetLoop(const Loop *L) {
291 LoopPropsMapIt LIt = LoopsProperties.find(L);
293 if (LIt != LoopsProperties.end()) {
294 LoopProperties &Props = LIt->second;
295 MaxSize += (Props.CanBeUnswitchedCount + Props.WasUnswitchedCount) *
296 Props.SizeEstimation;
297 LoopsProperties.erase(LIt);
300 CurrentLoopProperties = nullptr;
301 CurLoopInstructions = nullptr;
304 // Mark case value as unswitched.
305 // Since SI instruction can be partly unswitched, in order to avoid
306 // extra unswitching in cloned loops keep track all unswitched values.
307 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
308 (*CurLoopInstructions)[SI].insert(V);
311 // Check was this case value unswitched before or not.
312 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
313 return (*CurLoopInstructions)[SI].count(V);
316 bool LUAnalysisCache::CostAllowsUnswitching() {
317 return CurrentLoopProperties->CanBeUnswitchedCount > 0;
320 // Clone all loop-unswitch related loop properties.
321 // Redistribute unswitching quotas.
322 // Note, that new loop data is stored inside the VMap.
323 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
324 const ValueToValueMapTy &VMap) {
326 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
327 LoopProperties &OldLoopProps = *CurrentLoopProperties;
328 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
330 // Reallocate "can-be-unswitched quota"
332 --OldLoopProps.CanBeUnswitchedCount;
333 ++OldLoopProps.WasUnswitchedCount;
334 NewLoopProps.WasUnswitchedCount = 0;
335 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
336 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
337 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
339 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
341 // Clone unswitched values info:
342 // for new loop switches we clone info about values that was
343 // already unswitched and has redundant successors.
344 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
345 const SwitchInst *OldInst = I->first;
346 Value *NewI = VMap.lookup(OldInst);
347 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
348 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
350 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
354 char LoopUnswitch::ID = 0;
355 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
357 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
358 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
359 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
360 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
361 INITIALIZE_PASS_DEPENDENCY(LCSSA)
362 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
365 Pass *llvm::createLoopUnswitchPass(bool Os) {
366 return new LoopUnswitch(Os);
369 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
370 /// invariant in the loop, or has an invariant piece, return the invariant.
371 /// Otherwise, return null.
372 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
374 // We started analyze new instruction, increment scanned instructions counter.
377 // We can never unswitch on vector conditions.
378 if (Cond->getType()->isVectorTy())
381 // Constants should be folded, not unswitched on!
382 if (isa<Constant>(Cond)) return nullptr;
384 // TODO: Handle: br (VARIANT|INVARIANT).
386 // Hoist simple values out.
387 if (L->makeLoopInvariant(Cond, Changed))
390 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
391 if (BO->getOpcode() == Instruction::And ||
392 BO->getOpcode() == Instruction::Or) {
393 // If either the left or right side is invariant, we can unswitch on this,
394 // which will cause the branch to go away in one loop and the condition to
395 // simplify in the other one.
396 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
398 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
405 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
406 if (skipOptnoneFunction(L))
409 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
410 *L->getHeader()->getParent());
411 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
413 DominatorTreeWrapperPass *DTWP =
414 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
415 DT = DTWP ? &DTWP->getDomTree() : nullptr;
417 Function *F = currentLoop->getHeader()->getParent();
418 bool Changed = false;
420 assert(currentLoop->isLCSSAForm(*DT));
422 Changed |= processCurrentLoop();
426 // FIXME: Reconstruct dom info, because it is not preserved properly.
433 /// processCurrentLoop - Do actual work and unswitch loop if possible
435 bool LoopUnswitch::processCurrentLoop() {
436 bool Changed = false;
440 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
444 // Loops with indirectbr cannot be cloned.
445 if (!currentLoop->isSafeToClone())
448 // Without dedicated exits, splitting the exit edge may fail.
449 if (!currentLoop->hasDedicatedExits())
452 LLVMContext &Context = loopHeader->getContext();
454 // Probably we reach the quota of branches for this loop. If so
456 if (!BranchesInfo.countLoop(
457 currentLoop, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
458 *currentLoop->getHeader()->getParent()),
462 // Trivial unswitch condition can only occur at loop header basic block because
463 // that condition needs to control whether or not the loop does anything at all.
464 // Try trivial unswitch first before loop over other basic blocks in the loop.
465 if (TryTrivialLoopUnswitch(Changed)) {
469 // Loop over all of the basic blocks in the loop. If we find an interior
470 // block that is branching on a loop-invariant condition, we can unswitch this
472 for (Loop::block_iterator I = currentLoop->block_begin(),
473 E = currentLoop->block_end(); I != E; ++I) {
474 TerminatorInst *TI = (*I)->getTerminator();
475 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
476 // If this isn't branching on an invariant condition, we can't unswitch
478 if (BI->isConditional()) {
479 // See if this, or some part of it, is loop invariant. If so, we can
480 // unswitch on it if we desire.
481 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
482 currentLoop, Changed);
484 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context), TI)) {
489 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
490 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
491 currentLoop, Changed);
492 unsigned NumCases = SI->getNumCases();
493 if (LoopCond && NumCases) {
494 // Find a value to unswitch on:
495 // FIXME: this should chose the most expensive case!
496 // FIXME: scan for a case with a non-critical edge?
497 Constant *UnswitchVal = nullptr;
499 // Do not process same value again and again.
500 // At this point we have some cases already unswitched and
501 // some not yet unswitched. Let's find the first not yet unswitched one.
502 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
504 Constant *UnswitchValCandidate = i.getCaseValue();
505 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
506 UnswitchVal = UnswitchValCandidate;
514 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
521 // Scan the instructions to check for unswitchable values.
522 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
524 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
525 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
526 currentLoop, Changed);
527 if (LoopCond && UnswitchIfProfitable(LoopCond,
528 ConstantInt::getTrue(Context))) {
537 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
538 /// loop with no side effects (including infinite loops).
540 /// If true, we return true and set ExitBB to the block we
543 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
545 std::set<BasicBlock*> &Visited) {
546 if (!Visited.insert(BB).second) {
547 // Already visited. Without more analysis, this could indicate an infinite
551 if (!L->contains(BB)) {
552 // Otherwise, this is a loop exit, this is fine so long as this is the
554 if (ExitBB) return false;
559 // Otherwise, this is an unvisited intra-loop node. Check all successors.
560 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
561 // Check to see if the successor is a trivial loop exit.
562 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
566 // Okay, everything after this looks good, check to make sure that this block
567 // doesn't include any side effects.
568 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
569 if (I->mayHaveSideEffects())
575 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
576 /// leads to an exit from the specified loop, and has no side-effects in the
577 /// process. If so, return the block that is exited to, otherwise return null.
578 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
579 std::set<BasicBlock*> Visited;
580 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
581 BasicBlock *ExitBB = nullptr;
582 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
587 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
588 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
589 /// unswitch the loop, reprocess the pieces, then return true.
590 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,
591 TerminatorInst *TI) {
592 Function *F = loopHeader->getParent();
594 // Check to see if it would be profitable to unswitch current loop.
595 if (!BranchesInfo.CostAllowsUnswitching()) {
596 DEBUG(dbgs() << "NOT unswitching loop %"
597 << currentLoop->getHeader()->getName()
598 << " at non-trivial condition '" << *Val
599 << "' == " << *LoopCond << "\n"
600 << ". Cost too high.\n");
604 // Do not do non-trivial unswitch while optimizing for size.
605 if (OptimizeForSize || F->hasFnAttribute(Attribute::OptimizeForSize))
608 UnswitchNontrivialCondition(LoopCond, Val, currentLoop, TI);
612 /// CloneLoop - Recursively clone the specified loop and all of its children,
613 /// mapping the blocks with the specified map.
614 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
615 LoopInfo *LI, LPPassManager *LPM) {
616 Loop *New = new Loop();
617 LPM->insertLoop(New, PL);
619 // Add all of the blocks in L to the new loop.
620 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
622 if (LI->getLoopFor(*I) == L)
623 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
625 // Add all of the subloops to the new loop.
626 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
627 CloneLoop(*I, New, VM, LI, LPM);
632 static void copyMetadata(Instruction *DstInst, const Instruction *SrcInst,
634 if (!SrcInst || !SrcInst->hasMetadata())
637 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
638 SrcInst->getAllMetadata(MDs);
639 for (auto &MD : MDs) {
643 case LLVMContext::MD_prof:
644 if (Swapped && MD.second->getNumOperands() == 3 &&
645 isa<MDString>(MD.second->getOperand(0))) {
646 MDString *MDName = cast<MDString>(MD.second->getOperand(0));
647 if (MDName->getString() == "branch_weights") {
648 auto *ValT = cast_or_null<ConstantAsMetadata>(
649 MD.second->getOperand(1))->getValue();
650 auto *ValF = cast_or_null<ConstantAsMetadata>(
651 MD.second->getOperand(2))->getValue();
652 assert(ValT && ValF && "Invalid Operands of branch_weights");
654 MDBuilder(DstInst->getParent()->getContext())
655 .createBranchWeights(cast<ConstantInt>(ValF)->getZExtValue(),
656 cast<ConstantInt>(ValT)->getZExtValue());
661 case LLVMContext::MD_dbg:
662 DstInst->setMetadata(MD.first, MD.second);
667 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
668 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
669 /// code immediately before InsertPt.
670 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
671 BasicBlock *TrueDest,
672 BasicBlock *FalseDest,
673 Instruction *InsertPt,
674 TerminatorInst *TI) {
675 // Insert a conditional branch on LIC to the two preheaders. The original
676 // code is the true version and the new code is the false version.
677 Value *BranchVal = LIC;
678 bool Swapped = false;
679 if (!isa<ConstantInt>(Val) ||
680 Val->getType() != Type::getInt1Ty(LIC->getContext()))
681 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
682 else if (Val != ConstantInt::getTrue(Val->getContext())) {
683 // We want to enter the new loop when the condition is true.
684 std::swap(TrueDest, FalseDest);
688 // Insert the new branch.
689 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
690 copyMetadata(BI, TI, Swapped);
692 // If either edge is critical, split it. This helps preserve LoopSimplify
693 // form for enclosing loops.
694 auto Options = CriticalEdgeSplittingOptions(DT, LI).setPreserveLCSSA();
695 SplitCriticalEdge(BI, 0, Options);
696 SplitCriticalEdge(BI, 1, Options);
699 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
700 /// condition in it (a cond branch from its header block to its latch block,
701 /// where the path through the loop that doesn't execute its body has no
702 /// side-effects), unswitch it. This doesn't involve any code duplication, just
703 /// moving the conditional branch outside of the loop and updating loop info.
704 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
705 BasicBlock *ExitBlock,
706 TerminatorInst *TI) {
707 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
708 << loopHeader->getName() << " [" << L->getBlocks().size()
709 << " blocks] in Function "
710 << L->getHeader()->getParent()->getName() << " on cond: " << *Val
711 << " == " << *Cond << "\n");
713 // First step, split the preheader, so that we know that there is a safe place
714 // to insert the conditional branch. We will change loopPreheader to have a
715 // conditional branch on Cond.
716 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, DT, LI);
718 // Now that we have a place to insert the conditional branch, create a place
719 // to branch to: this is the exit block out of the loop that we should
722 // Split this block now, so that the loop maintains its exit block, and so
723 // that the jump from the preheader can execute the contents of the exit block
724 // without actually branching to it (the exit block should be dominated by the
725 // loop header, not the preheader).
726 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
727 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), DT, LI);
729 // Okay, now we have a position to branch from and a position to branch to,
730 // insert the new conditional branch.
731 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
732 loopPreheader->getTerminator(), TI);
733 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
734 loopPreheader->getTerminator()->eraseFromParent();
736 // We need to reprocess this loop, it could be unswitched again.
739 // Now that we know that the loop is never entered when this condition is a
740 // particular value, rewrite the loop with this info. We know that this will
741 // at least eliminate the old branch.
742 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
746 /// TryTrivialLoopUnswitch - Check if loop header block's terminator is a trivial
747 /// unswitch condition: that is, the condition controls whether or not the loop
748 /// does anything at all (which means it can only occur in loop header). If it is
749 /// a trivial condition, unswitching produces no code duplications (equivalently,
750 /// it produces a simpler loop and a new empty loop, which gets deleted). Therefore
751 /// always unswitch trivial condition.
753 bool LoopUnswitch::TryTrivialLoopUnswitch(bool &Changed) {
754 BasicBlock *Header = currentLoop->getHeader();
755 TerminatorInst *HeaderTerm = Header->getTerminator();
756 LLVMContext &Context = Header->getContext();
758 // Check if this loop will execute any side-effecting instructions (e.g.
759 // stores, calls, volatile loads) in the part of the loop that the code
760 // *would* execute. Check the header first.
761 for (BasicBlock::iterator I :*Header)
762 if (I->mayHaveSideEffects())
765 // CondVal is the condition that controls the trivial condition.
766 // LoopExitBB is the BasicBlock that loop exits when meets trivial condition.
767 Constant *CondVal = nullptr;
768 BasicBlock *LoopExitBB = nullptr;
770 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
771 // If this isn't branching on an invariant condition, we can't unswitch it.
772 if (!BI->isConditional())
775 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
776 currentLoop, Changed);
778 // Unswitch only if the trivial condition itself is an LIV (not
779 // partial LIV which could occur in and/or)
780 if (!LoopCond || LoopCond != BI->getCondition())
783 // Check to see if a successor of the branch is guaranteed to
784 // exit through a unique exit block without having any
785 // side-effects. If so, determine the value of Cond that causes
787 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
788 BI->getSuccessor(0)))) {
789 CondVal = ConstantInt::getTrue(Context);
790 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
791 BI->getSuccessor(1)))) {
792 CondVal = ConstantInt::getFalse(Context);
795 // If we didn't find a single unique LoopExit block, or if the loop exit block
796 // contains phi nodes, this isn't trivial.
797 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
798 return false; // Can't handle this.
800 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB, HeaderTerm);
803 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
804 // If this isn't switching on an invariant condition, we can't unswitch it.
805 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
806 currentLoop, Changed);
808 // Unswitch only if the trivial condition itself is an LIV (not
809 // partial LIV which could occur in and/or)
810 if (!LoopCond || LoopCond != SI->getCondition())
813 // Check to see if a successor of the switch is guaranteed to go to the
814 // latch block or exit through a one exit block without having any
815 // side-effects. If so, determine the value of Cond that causes it to do
817 // Note that we can't trivially unswitch on the default case or
818 // on already unswitched cases.
819 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
821 BasicBlock *LoopExitCandidate;
822 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
823 i.getCaseSuccessor()))) {
824 // Okay, we found a trivial case, remember the value that is trivial.
825 ConstantInt *CaseVal = i.getCaseValue();
827 // Check that it was not unswitched before, since already unswitched
828 // trivial vals are looks trivial too.
829 if (BranchesInfo.isUnswitched(SI, CaseVal))
831 LoopExitBB = LoopExitCandidate;
837 // If we didn't find a single unique LoopExit block, or if the loop exit block
838 // contains phi nodes, this isn't trivial.
839 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
840 return false; // Can't handle this.
842 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB, nullptr);
849 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
850 /// blocks. Update the appropriate Phi nodes as we do so.
851 void LoopUnswitch::SplitExitEdges(Loop *L,
852 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
854 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
855 BasicBlock *ExitBlock = ExitBlocks[i];
856 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
857 pred_end(ExitBlock));
859 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
860 // general, if we call it on all predecessors of all exits then it does.
861 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", DT, LI,
862 /*PreserveLCSSA*/ true);
866 /// UnswitchNontrivialCondition - We determined that the loop is profitable
867 /// to unswitch when LIC equal Val. Split it into loop versions and test the
868 /// condition outside of either loop. Return the loops created as Out1/Out2.
869 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
870 Loop *L, TerminatorInst *TI) {
871 Function *F = loopHeader->getParent();
872 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
873 << loopHeader->getName() << " [" << L->getBlocks().size()
874 << " blocks] in Function " << F->getName()
875 << " when '" << *Val << "' == " << *LIC << "\n");
877 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
883 // First step, split the preheader and exit blocks, and add these blocks to
884 // the LoopBlocks list.
885 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, DT, LI);
886 LoopBlocks.push_back(NewPreheader);
888 // We want the loop to come after the preheader, but before the exit blocks.
889 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
891 SmallVector<BasicBlock*, 8> ExitBlocks;
892 L->getUniqueExitBlocks(ExitBlocks);
894 // Split all of the edges from inside the loop to their exit blocks. Update
895 // the appropriate Phi nodes as we do so.
896 SplitExitEdges(L, ExitBlocks);
898 // The exit blocks may have been changed due to edge splitting, recompute.
900 L->getUniqueExitBlocks(ExitBlocks);
902 // Add exit blocks to the loop blocks.
903 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
905 // Next step, clone all of the basic blocks that make up the loop (including
906 // the loop preheader and exit blocks), keeping track of the mapping between
907 // the instructions and blocks.
908 NewBlocks.reserve(LoopBlocks.size());
909 ValueToValueMapTy VMap;
910 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
911 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
913 NewBlocks.push_back(NewBB);
914 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
915 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
918 // Splice the newly inserted blocks into the function right before the
919 // original preheader.
920 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
921 NewBlocks[0], F->end());
923 // FIXME: We could register any cloned assumptions instead of clearing the
924 // whole function's cache.
927 // Now we create the new Loop object for the versioned loop.
928 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
930 // Recalculate unswitching quota, inherit simplified switches info for NewBB,
931 // Probably clone more loop-unswitch related loop properties.
932 BranchesInfo.cloneData(NewLoop, L, VMap);
934 Loop *ParentLoop = L->getParentLoop();
936 // Make sure to add the cloned preheader and exit blocks to the parent loop
938 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
941 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
942 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
943 // The new exit block should be in the same loop as the old one.
944 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
945 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
947 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
948 "Exit block should have been split to have one successor!");
949 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
951 // If the successor of the exit block had PHI nodes, add an entry for
953 for (BasicBlock::iterator I = ExitSucc->begin();
954 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
955 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
956 ValueToValueMapTy::iterator It = VMap.find(V);
957 if (It != VMap.end()) V = It->second;
958 PN->addIncoming(V, NewExit);
961 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
962 PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
963 ExitSucc->getFirstInsertionPt());
965 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
968 LandingPadInst *LPI = BB->getLandingPadInst();
969 LPI->replaceAllUsesWith(PN);
970 PN->addIncoming(LPI, BB);
975 // Rewrite the code to refer to itself.
976 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
977 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
978 E = NewBlocks[i]->end(); I != E; ++I)
979 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
981 // Rewrite the original preheader to select between versions of the loop.
982 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
983 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
984 "Preheader splitting did not work correctly!");
986 // Emit the new branch that selects between the two versions of this loop.
987 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR,
989 LPM->deleteSimpleAnalysisValue(OldBR, L);
990 OldBR->eraseFromParent();
992 LoopProcessWorklist.push_back(NewLoop);
995 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
996 // deletes the instruction (for example by simplifying a PHI that feeds into
997 // the condition that we're unswitching on), we don't rewrite the second
999 WeakVH LICHandle(LIC);
1001 // Now we rewrite the original code to know that the condition is true and the
1002 // new code to know that the condition is false.
1003 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
1005 // It's possible that simplifying one loop could cause the other to be
1006 // changed to another value or a constant. If its a constant, don't simplify
1008 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
1009 LICHandle && !isa<Constant>(LICHandle))
1010 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
1013 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
1015 static void RemoveFromWorklist(Instruction *I,
1016 std::vector<Instruction*> &Worklist) {
1018 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
1022 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
1023 /// program, replacing all uses with V and update the worklist.
1024 static void ReplaceUsesOfWith(Instruction *I, Value *V,
1025 std::vector<Instruction*> &Worklist,
1026 Loop *L, LPPassManager *LPM) {
1027 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
1029 // Add uses to the worklist, which may be dead now.
1030 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1031 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1032 Worklist.push_back(Use);
1034 // Add users to the worklist which may be simplified now.
1035 for (User *U : I->users())
1036 Worklist.push_back(cast<Instruction>(U));
1037 LPM->deleteSimpleAnalysisValue(I, L);
1038 RemoveFromWorklist(I, Worklist);
1039 I->replaceAllUsesWith(V);
1040 I->eraseFromParent();
1044 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
1045 // the value specified by Val in the specified loop, or we know it does NOT have
1046 // that value. Rewrite any uses of LIC or of properties correlated to it.
1047 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
1050 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
1052 // FIXME: Support correlated properties, like:
1059 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
1060 // selects, switches.
1061 std::vector<Instruction*> Worklist;
1062 LLVMContext &Context = Val->getContext();
1064 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
1065 // in the loop with the appropriate one directly.
1066 if (IsEqual || (isa<ConstantInt>(Val) &&
1067 Val->getType()->isIntegerTy(1))) {
1072 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
1073 !cast<ConstantInt>(Val)->getZExtValue());
1075 for (User *U : LIC->users()) {
1076 Instruction *UI = dyn_cast<Instruction>(U);
1077 if (!UI || !L->contains(UI))
1079 Worklist.push_back(UI);
1082 for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
1083 UE = Worklist.end(); UI != UE; ++UI)
1084 (*UI)->replaceUsesOfWith(LIC, Replacement);
1086 SimplifyCode(Worklist, L);
1090 // Otherwise, we don't know the precise value of LIC, but we do know that it
1091 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1092 // can. This case occurs when we unswitch switch statements.
1093 for (User *U : LIC->users()) {
1094 Instruction *UI = dyn_cast<Instruction>(U);
1095 if (!UI || !L->contains(UI))
1098 Worklist.push_back(UI);
1100 // TODO: We could do other simplifications, for example, turning
1101 // 'icmp eq LIC, Val' -> false.
1103 // If we know that LIC is not Val, use this info to simplify code.
1104 SwitchInst *SI = dyn_cast<SwitchInst>(UI);
1105 if (!SI || !isa<ConstantInt>(Val)) continue;
1107 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1108 // Default case is live for multiple values.
1109 if (DeadCase == SI->case_default()) continue;
1111 // Found a dead case value. Don't remove PHI nodes in the
1112 // successor if they become single-entry, those PHI nodes may
1113 // be in the Users list.
1115 BasicBlock *Switch = SI->getParent();
1116 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1117 BasicBlock *Latch = L->getLoopLatch();
1119 BranchesInfo.setUnswitched(SI, Val);
1121 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1122 // If the DeadCase successor dominates the loop latch, then the
1123 // transformation isn't safe since it will delete the sole predecessor edge
1125 if (Latch && DT->dominates(SISucc, Latch))
1128 // FIXME: This is a hack. We need to keep the successor around
1129 // and hooked up so as to preserve the loop structure, because
1130 // trying to update it is complicated. So instead we preserve the
1131 // loop structure and put the block on a dead code path.
1132 SplitEdge(Switch, SISucc, DT, LI);
1133 // Compute the successors instead of relying on the return value
1134 // of SplitEdge, since it may have split the switch successor
1136 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1137 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1138 // Create an "unreachable" destination.
1139 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1140 Switch->getParent(),
1142 new UnreachableInst(Context, Abort);
1143 // Force the new case destination to branch to the "unreachable"
1144 // block while maintaining a (dead) CFG edge to the old block.
1145 NewSISucc->getTerminator()->eraseFromParent();
1146 BranchInst::Create(Abort, OldSISucc,
1147 ConstantInt::getTrue(Context), NewSISucc);
1148 // Release the PHI operands for this edge.
1149 for (BasicBlock::iterator II = NewSISucc->begin();
1150 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1151 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1152 UndefValue::get(PN->getType()));
1153 // Tell the domtree about the new block. We don't fully update the
1154 // domtree here -- instead we force it to do a full recomputation
1155 // after the pass is complete -- but we do need to inform it of
1158 DT->addNewBlock(Abort, NewSISucc);
1161 SimplifyCode(Worklist, L);
1164 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1165 /// loop, walk over it and constant prop, dce, and fold control flow where
1166 /// possible. Note that this is effectively a very simple loop-structure-aware
1167 /// optimizer. During processing of this loop, L could very well be deleted, so
1168 /// it must not be used.
1170 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1173 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1174 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1175 while (!Worklist.empty()) {
1176 Instruction *I = Worklist.back();
1177 Worklist.pop_back();
1180 if (isInstructionTriviallyDead(I)) {
1181 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1183 // Add uses to the worklist, which may be dead now.
1184 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1185 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1186 Worklist.push_back(Use);
1187 LPM->deleteSimpleAnalysisValue(I, L);
1188 RemoveFromWorklist(I, Worklist);
1189 I->eraseFromParent();
1194 // See if instruction simplification can hack this up. This is common for
1195 // things like "select false, X, Y" after unswitching made the condition be
1196 // 'false'. TODO: update the domtree properly so we can pass it here.
1197 if (Value *V = SimplifyInstruction(I, DL))
1198 if (LI->replacementPreservesLCSSAForm(I, V)) {
1199 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1203 // Special case hacks that appear commonly in unswitched code.
1204 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1205 if (BI->isUnconditional()) {
1206 // If BI's parent is the only pred of the successor, fold the two blocks
1208 BasicBlock *Pred = BI->getParent();
1209 BasicBlock *Succ = BI->getSuccessor(0);
1210 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1211 if (!SinglePred) continue; // Nothing to do.
1212 assert(SinglePred == Pred && "CFG broken");
1214 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1215 << Succ->getName() << "\n");
1217 // Resolve any single entry PHI nodes in Succ.
1218 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1219 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1221 // If Succ has any successors with PHI nodes, update them to have
1222 // entries coming from Pred instead of Succ.
1223 Succ->replaceAllUsesWith(Pred);
1225 // Move all of the successor contents from Succ to Pred.
1226 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1228 LPM->deleteSimpleAnalysisValue(BI, L);
1229 BI->eraseFromParent();
1230 RemoveFromWorklist(BI, Worklist);
1232 // Remove Succ from the loop tree.
1233 LI->removeBlock(Succ);
1234 LPM->deleteSimpleAnalysisValue(Succ, L);
1235 Succ->eraseFromParent();