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/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,
108 AssumptionCache *AC);
110 // Clean all data related to given loop.
111 void forgetLoop(const Loop *L);
113 // Mark case value as unswitched.
114 // Since SI instruction can be partly unswitched, in order to avoid
115 // extra unswitching in cloned loops keep track all unswitched values.
116 void setUnswitched(const SwitchInst *SI, const Value *V);
118 // Check was this case value unswitched before or not.
119 bool isUnswitched(const SwitchInst *SI, const Value *V);
121 // Clone all loop-unswitch related loop properties.
122 // Redistribute unswitching quotas.
123 // Note, that new loop data is stored inside the VMap.
124 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
125 const ValueToValueMapTy &VMap);
128 class LoopUnswitch : public LoopPass {
129 LoopInfo *LI; // Loop information
133 // LoopProcessWorklist - Used to check if second loop needs processing
134 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
135 std::vector<Loop*> LoopProcessWorklist;
137 LUAnalysisCache BranchesInfo;
139 bool OptimizeForSize;
144 BasicBlock *loopHeader;
145 BasicBlock *loopPreheader;
147 // LoopBlocks contains all of the basic blocks of the loop, including the
148 // preheader of the loop, the body of the loop, and the exit blocks of the
149 // loop, in that order.
150 std::vector<BasicBlock*> LoopBlocks;
151 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
152 std::vector<BasicBlock*> NewBlocks;
155 static char ID; // Pass ID, replacement for typeid
156 explicit LoopUnswitch(bool Os = false) :
157 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
158 currentLoop(nullptr), DT(nullptr), loopHeader(nullptr),
159 loopPreheader(nullptr) {
160 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
163 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
164 bool processCurrentLoop();
166 /// This transformation requires natural loop information & requires that
167 /// loop preheaders be inserted into the CFG.
169 void getAnalysisUsage(AnalysisUsage &AU) const override {
170 AU.addRequired<AssumptionCacheTracker>();
171 AU.addRequiredID(LoopSimplifyID);
172 AU.addPreservedID(LoopSimplifyID);
173 AU.addRequired<LoopInfoWrapperPass>();
174 AU.addPreserved<LoopInfoWrapperPass>();
175 AU.addRequiredID(LCSSAID);
176 AU.addPreservedID(LCSSAID);
177 AU.addPreserved<DominatorTreeWrapperPass>();
178 AU.addPreserved<ScalarEvolution>();
179 AU.addRequired<TargetTransformInfo>();
184 void releaseMemory() override {
185 BranchesInfo.forgetLoop(currentLoop);
188 void initLoopData() {
189 loopHeader = currentLoop->getHeader();
190 loopPreheader = currentLoop->getLoopPreheader();
193 /// Split all of the edges from inside the loop to their exit blocks.
194 /// Update the appropriate Phi nodes as we do so.
195 void SplitExitEdges(Loop *L, const SmallVectorImpl<BasicBlock *> &ExitBlocks);
197 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
198 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
199 BasicBlock *ExitBlock);
200 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
202 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
203 Constant *Val, bool isEqual);
205 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
206 BasicBlock *TrueDest,
207 BasicBlock *FalseDest,
208 Instruction *InsertPt);
210 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
211 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = nullptr,
212 BasicBlock **LoopExit = nullptr);
217 // Analyze loop. Check its size, calculate is it possible to unswitch
218 // it. Returns true if we can unswitch this loop.
219 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI,
220 AssumptionCache *AC) {
222 LoopPropsMapIt PropsIt;
224 std::tie(PropsIt, Inserted) =
225 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
227 LoopProperties &Props = PropsIt->second;
232 // Limit the number of instructions to avoid causing significant code
233 // expansion, and the number of basic blocks, to avoid loops with
234 // large numbers of branches which cause loop unswitching to go crazy.
235 // This is a very ad-hoc heuristic.
237 SmallPtrSet<const Value *, 32> EphValues;
238 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
240 // FIXME: This is overly conservative because it does not take into
241 // consideration code simplification opportunities and code that can
242 // be shared by the resultant unswitched loops.
244 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
246 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
248 Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5);
249 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
250 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
252 if (Metrics.notDuplicatable) {
253 DEBUG(dbgs() << "NOT unswitching loop %"
254 << L->getHeader()->getName() << ", contents cannot be "
260 if (!Props.CanBeUnswitchedCount) {
261 DEBUG(dbgs() << "NOT unswitching loop %"
262 << L->getHeader()->getName() << ", cost too high: "
263 << L->getBlocks().size() << "\n");
267 // Be careful. This links are good only before new loop addition.
268 CurrentLoopProperties = &Props;
269 CurLoopInstructions = &Props.UnswitchedVals;
274 // Clean all data related to given loop.
275 void LUAnalysisCache::forgetLoop(const Loop *L) {
277 LoopPropsMapIt LIt = LoopsProperties.find(L);
279 if (LIt != LoopsProperties.end()) {
280 LoopProperties &Props = LIt->second;
281 MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation;
282 LoopsProperties.erase(LIt);
285 CurrentLoopProperties = nullptr;
286 CurLoopInstructions = nullptr;
289 // Mark case value as unswitched.
290 // Since SI instruction can be partly unswitched, in order to avoid
291 // extra unswitching in cloned loops keep track all unswitched values.
292 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
293 (*CurLoopInstructions)[SI].insert(V);
296 // Check was this case value unswitched before or not.
297 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
298 return (*CurLoopInstructions)[SI].count(V);
301 // Clone all loop-unswitch related loop properties.
302 // Redistribute unswitching quotas.
303 // Note, that new loop data is stored inside the VMap.
304 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
305 const ValueToValueMapTy &VMap) {
307 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
308 LoopProperties &OldLoopProps = *CurrentLoopProperties;
309 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
311 // Reallocate "can-be-unswitched quota"
313 --OldLoopProps.CanBeUnswitchedCount;
314 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
315 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
316 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
318 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
320 // Clone unswitched values info:
321 // for new loop switches we clone info about values that was
322 // already unswitched and has redundant successors.
323 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
324 const SwitchInst *OldInst = I->first;
325 Value *NewI = VMap.lookup(OldInst);
326 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
327 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
329 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
333 char LoopUnswitch::ID = 0;
334 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
336 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
337 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
338 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
339 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
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 nullptr;
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 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
389 *L->getHeader()->getParent());
390 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
392 DominatorTreeWrapperPass *DTWP =
393 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
394 DT = DTWP ? &DTWP->getDomTree() : nullptr;
396 Function *F = currentLoop->getHeader()->getParent();
397 bool Changed = false;
399 assert(currentLoop->isLCSSAForm(*DT));
401 Changed |= processCurrentLoop();
405 // FIXME: Reconstruct dom info, because it is not preserved properly.
412 /// processCurrentLoop - Do actual work and unswitch loop if possible
414 bool LoopUnswitch::processCurrentLoop() {
415 bool Changed = false;
419 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
423 // Loops with indirectbr cannot be cloned.
424 if (!currentLoop->isSafeToClone())
427 // Without dedicated exits, splitting the exit edge may fail.
428 if (!currentLoop->hasDedicatedExits())
431 LLVMContext &Context = loopHeader->getContext();
433 // Probably we reach the quota of branches for this loop. If so
435 if (!BranchesInfo.countLoop(currentLoop, getAnalysis<TargetTransformInfo>(),
439 // Loop over all of the basic blocks in the loop. If we find an interior
440 // block that is branching on a loop-invariant condition, we can unswitch this
442 for (Loop::block_iterator I = currentLoop->block_begin(),
443 E = currentLoop->block_end(); I != E; ++I) {
444 TerminatorInst *TI = (*I)->getTerminator();
445 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
446 // If this isn't branching on an invariant condition, we can't unswitch
448 if (BI->isConditional()) {
449 // See if this, or some part of it, is loop invariant. If so, we can
450 // unswitch on it if we desire.
451 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
452 currentLoop, Changed);
453 if (LoopCond && UnswitchIfProfitable(LoopCond,
454 ConstantInt::getTrue(Context))) {
459 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
460 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
461 currentLoop, Changed);
462 unsigned NumCases = SI->getNumCases();
463 if (LoopCond && NumCases) {
464 // Find a value to unswitch on:
465 // FIXME: this should chose the most expensive case!
466 // FIXME: scan for a case with a non-critical edge?
467 Constant *UnswitchVal = nullptr;
469 // Do not process same value again and again.
470 // At this point we have some cases already unswitched and
471 // some not yet unswitched. Let's find the first not yet unswitched one.
472 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
474 Constant *UnswitchValCandidate = i.getCaseValue();
475 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
476 UnswitchVal = UnswitchValCandidate;
484 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
491 // Scan the instructions to check for unswitchable values.
492 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
494 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
495 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
496 currentLoop, Changed);
497 if (LoopCond && UnswitchIfProfitable(LoopCond,
498 ConstantInt::getTrue(Context))) {
507 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
508 /// loop with no side effects (including infinite loops).
510 /// If true, we return true and set ExitBB to the block we
513 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
515 std::set<BasicBlock*> &Visited) {
516 if (!Visited.insert(BB).second) {
517 // Already visited. Without more analysis, this could indicate an infinite
521 if (!L->contains(BB)) {
522 // Otherwise, this is a loop exit, this is fine so long as this is the
524 if (ExitBB) return false;
529 // Otherwise, this is an unvisited intra-loop node. Check all successors.
530 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
531 // Check to see if the successor is a trivial loop exit.
532 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
536 // Okay, everything after this looks good, check to make sure that this block
537 // doesn't include any side effects.
538 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
539 if (I->mayHaveSideEffects())
545 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
546 /// leads to an exit from the specified loop, and has no side-effects in the
547 /// process. If so, return the block that is exited to, otherwise return null.
548 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
549 std::set<BasicBlock*> Visited;
550 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
551 BasicBlock *ExitBB = nullptr;
552 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
557 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
558 /// trivial: that is, that the condition controls whether or not the loop does
559 /// anything at all. If this is a trivial condition, unswitching produces no
560 /// code duplications (equivalently, it produces a simpler loop and a new empty
561 /// loop, which gets deleted).
563 /// If this is a trivial condition, return true, otherwise return false. When
564 /// returning true, this sets Cond and Val to the condition that controls the
565 /// trivial condition: when Cond dynamically equals Val, the loop is known to
566 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
569 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
570 BasicBlock **LoopExit) {
571 BasicBlock *Header = currentLoop->getHeader();
572 TerminatorInst *HeaderTerm = Header->getTerminator();
573 LLVMContext &Context = Header->getContext();
575 BasicBlock *LoopExitBB = nullptr;
576 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
577 // If the header block doesn't end with a conditional branch on Cond, we
579 if (!BI->isConditional() || BI->getCondition() != Cond)
582 // Check to see if a successor of the branch is guaranteed to
583 // exit through a unique exit block without having any
584 // side-effects. If so, determine the value of Cond that causes it to do
586 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
587 BI->getSuccessor(0)))) {
588 if (Val) *Val = ConstantInt::getTrue(Context);
589 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
590 BI->getSuccessor(1)))) {
591 if (Val) *Val = ConstantInt::getFalse(Context);
593 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
594 // If this isn't a switch on Cond, we can't handle it.
595 if (SI->getCondition() != Cond) return false;
597 // Check to see if a successor of the switch is guaranteed to go to the
598 // latch block or exit through a one exit block without having any
599 // side-effects. If so, determine the value of Cond that causes it to do
601 // Note that we can't trivially unswitch on the default case or
602 // on already unswitched cases.
603 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
605 BasicBlock *LoopExitCandidate;
606 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
607 i.getCaseSuccessor()))) {
608 // Okay, we found a trivial case, remember the value that is trivial.
609 ConstantInt *CaseVal = i.getCaseValue();
611 // Check that it was not unswitched before, since already unswitched
612 // trivial vals are looks trivial too.
613 if (BranchesInfo.isUnswitched(SI, CaseVal))
615 LoopExitBB = LoopExitCandidate;
616 if (Val) *Val = CaseVal;
622 // If we didn't find a single unique LoopExit block, or if the loop exit block
623 // contains phi nodes, this isn't trivial.
624 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
625 return false; // Can't handle this.
627 if (LoopExit) *LoopExit = LoopExitBB;
629 // We already know that nothing uses any scalar values defined inside of this
630 // loop. As such, we just have to check to see if this loop will execute any
631 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
632 // part of the loop that the code *would* execute. We already checked the
633 // tail, check the header now.
634 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
635 if (I->mayHaveSideEffects())
640 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
641 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
642 /// unswitch the loop, reprocess the pieces, then return true.
643 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
644 Function *F = loopHeader->getParent();
645 Constant *CondVal = nullptr;
646 BasicBlock *ExitBlock = nullptr;
648 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
649 // If the condition is trivial, always unswitch. There is no code growth
651 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
655 // Check to see if it would be profitable to unswitch current loop.
657 // Do not do non-trivial unswitch while optimizing for size.
658 if (OptimizeForSize ||
659 F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
660 Attribute::OptimizeForSize))
663 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
667 /// CloneLoop - Recursively clone the specified loop and all of its children,
668 /// mapping the blocks with the specified map.
669 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
670 LoopInfo *LI, LPPassManager *LPM) {
671 Loop *New = new Loop();
672 LPM->insertLoop(New, PL);
674 // Add all of the blocks in L to the new loop.
675 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
677 if (LI->getLoopFor(*I) == L)
678 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
680 // Add all of the subloops to the new loop.
681 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
682 CloneLoop(*I, New, VM, LI, LPM);
687 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
688 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
689 /// code immediately before InsertPt.
690 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
691 BasicBlock *TrueDest,
692 BasicBlock *FalseDest,
693 Instruction *InsertPt) {
694 // Insert a conditional branch on LIC to the two preheaders. The original
695 // code is the true version and the new code is the false version.
696 Value *BranchVal = LIC;
697 if (!isa<ConstantInt>(Val) ||
698 Val->getType() != Type::getInt1Ty(LIC->getContext()))
699 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
700 else if (Val != ConstantInt::getTrue(Val->getContext()))
701 // We want to enter the new loop when the condition is true.
702 std::swap(TrueDest, FalseDest);
704 // Insert the new branch.
705 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
707 // If either edge is critical, split it. This helps preserve LoopSimplify
708 // form for enclosing loops.
709 SplitCriticalEdge(BI, 0, this, false, false, true);
710 SplitCriticalEdge(BI, 1, this, false, false, true);
713 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
714 /// condition in it (a cond branch from its header block to its latch block,
715 /// where the path through the loop that doesn't execute its body has no
716 /// side-effects), unswitch it. This doesn't involve any code duplication, just
717 /// moving the conditional branch outside of the loop and updating loop info.
718 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
720 BasicBlock *ExitBlock) {
721 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
722 << loopHeader->getName() << " [" << L->getBlocks().size()
723 << " blocks] in Function " << L->getHeader()->getParent()->getName()
724 << " on cond: " << *Val << " == " << *Cond << "\n");
726 // First step, split the preheader, so that we know that there is a safe place
727 // to insert the conditional branch. We will change loopPreheader to have a
728 // conditional branch on Cond.
729 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
731 // Now that we have a place to insert the conditional branch, create a place
732 // to branch to: this is the exit block out of the loop that we should
735 // Split this block now, so that the loop maintains its exit block, and so
736 // that the jump from the preheader can execute the contents of the exit block
737 // without actually branching to it (the exit block should be dominated by the
738 // loop header, not the preheader).
739 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
740 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), DT, LI);
742 // Okay, now we have a position to branch from and a position to branch to,
743 // insert the new conditional branch.
744 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
745 loopPreheader->getTerminator());
746 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
747 loopPreheader->getTerminator()->eraseFromParent();
749 // We need to reprocess this loop, it could be unswitched again.
752 // Now that we know that the loop is never entered when this condition is a
753 // particular value, rewrite the loop with this info. We know that this will
754 // at least eliminate the old branch.
755 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
759 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
760 /// blocks. Update the appropriate Phi nodes as we do so.
761 void LoopUnswitch::SplitExitEdges(Loop *L,
762 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
764 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
765 BasicBlock *ExitBlock = ExitBlocks[i];
766 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
767 pred_end(ExitBlock));
769 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
770 // general, if we call it on all predecessors of all exits then it does.
771 if (!ExitBlock->isLandingPad()) {
772 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
774 SmallVector<BasicBlock*, 2> NewBBs;
775 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
781 /// UnswitchNontrivialCondition - We determined that the loop is profitable
782 /// to unswitch when LIC equal Val. Split it into loop versions and test the
783 /// condition outside of either loop. Return the loops created as Out1/Out2.
784 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
786 Function *F = loopHeader->getParent();
787 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
788 << loopHeader->getName() << " [" << L->getBlocks().size()
789 << " blocks] in Function " << F->getName()
790 << " when '" << *Val << "' == " << *LIC << "\n");
792 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
798 // First step, split the preheader and exit blocks, and add these blocks to
799 // the LoopBlocks list.
800 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
801 LoopBlocks.push_back(NewPreheader);
803 // We want the loop to come after the preheader, but before the exit blocks.
804 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
806 SmallVector<BasicBlock*, 8> ExitBlocks;
807 L->getUniqueExitBlocks(ExitBlocks);
809 // Split all of the edges from inside the loop to their exit blocks. Update
810 // the appropriate Phi nodes as we do so.
811 SplitExitEdges(L, ExitBlocks);
813 // The exit blocks may have been changed due to edge splitting, recompute.
815 L->getUniqueExitBlocks(ExitBlocks);
817 // Add exit blocks to the loop blocks.
818 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
820 // Next step, clone all of the basic blocks that make up the loop (including
821 // the loop preheader and exit blocks), keeping track of the mapping between
822 // the instructions and blocks.
823 NewBlocks.reserve(LoopBlocks.size());
824 ValueToValueMapTy VMap;
825 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
826 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
828 NewBlocks.push_back(NewBB);
829 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
830 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
833 // Splice the newly inserted blocks into the function right before the
834 // original preheader.
835 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
836 NewBlocks[0], F->end());
838 // FIXME: We could register any cloned assumptions instead of clearing the
839 // whole function's cache.
842 // Now we create the new Loop object for the versioned loop.
843 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
845 // Recalculate unswitching quota, inherit simplified switches info for NewBB,
846 // Probably clone more loop-unswitch related loop properties.
847 BranchesInfo.cloneData(NewLoop, L, VMap);
849 Loop *ParentLoop = L->getParentLoop();
851 // Make sure to add the cloned preheader and exit blocks to the parent loop
853 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
856 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
857 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
858 // The new exit block should be in the same loop as the old one.
859 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
860 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
862 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
863 "Exit block should have been split to have one successor!");
864 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
866 // If the successor of the exit block had PHI nodes, add an entry for
868 for (BasicBlock::iterator I = ExitSucc->begin();
869 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
870 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
871 ValueToValueMapTy::iterator It = VMap.find(V);
872 if (It != VMap.end()) V = It->second;
873 PN->addIncoming(V, NewExit);
876 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
877 PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
878 ExitSucc->getFirstInsertionPt());
880 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
883 LandingPadInst *LPI = BB->getLandingPadInst();
884 LPI->replaceAllUsesWith(PN);
885 PN->addIncoming(LPI, BB);
890 // Rewrite the code to refer to itself.
891 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
892 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
893 E = NewBlocks[i]->end(); I != E; ++I)
894 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
896 // Rewrite the original preheader to select between versions of the loop.
897 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
898 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
899 "Preheader splitting did not work correctly!");
901 // Emit the new branch that selects between the two versions of this loop.
902 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
903 LPM->deleteSimpleAnalysisValue(OldBR, L);
904 OldBR->eraseFromParent();
906 LoopProcessWorklist.push_back(NewLoop);
909 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
910 // deletes the instruction (for example by simplifying a PHI that feeds into
911 // the condition that we're unswitching on), we don't rewrite the second
913 WeakVH LICHandle(LIC);
915 // Now we rewrite the original code to know that the condition is true and the
916 // new code to know that the condition is false.
917 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
919 // It's possible that simplifying one loop could cause the other to be
920 // changed to another value or a constant. If its a constant, don't simplify
922 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
923 LICHandle && !isa<Constant>(LICHandle))
924 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
927 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
929 static void RemoveFromWorklist(Instruction *I,
930 std::vector<Instruction*> &Worklist) {
932 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
936 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
937 /// program, replacing all uses with V and update the worklist.
938 static void ReplaceUsesOfWith(Instruction *I, Value *V,
939 std::vector<Instruction*> &Worklist,
940 Loop *L, LPPassManager *LPM) {
941 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
943 // Add uses to the worklist, which may be dead now.
944 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
945 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
946 Worklist.push_back(Use);
948 // Add users to the worklist which may be simplified now.
949 for (User *U : I->users())
950 Worklist.push_back(cast<Instruction>(U));
951 LPM->deleteSimpleAnalysisValue(I, L);
952 RemoveFromWorklist(I, Worklist);
953 I->replaceAllUsesWith(V);
954 I->eraseFromParent();
958 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
959 // the value specified by Val in the specified loop, or we know it does NOT have
960 // that value. Rewrite any uses of LIC or of properties correlated to it.
961 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
964 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
966 // FIXME: Support correlated properties, like:
973 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
974 // selects, switches.
975 std::vector<Instruction*> Worklist;
976 LLVMContext &Context = Val->getContext();
978 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
979 // in the loop with the appropriate one directly.
980 if (IsEqual || (isa<ConstantInt>(Val) &&
981 Val->getType()->isIntegerTy(1))) {
986 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
987 !cast<ConstantInt>(Val)->getZExtValue());
989 for (User *U : LIC->users()) {
990 Instruction *UI = dyn_cast<Instruction>(U);
991 if (!UI || !L->contains(UI))
993 Worklist.push_back(UI);
996 for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
997 UE = Worklist.end(); UI != UE; ++UI)
998 (*UI)->replaceUsesOfWith(LIC, Replacement);
1000 SimplifyCode(Worklist, L);
1004 // Otherwise, we don't know the precise value of LIC, but we do know that it
1005 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1006 // can. This case occurs when we unswitch switch statements.
1007 for (User *U : LIC->users()) {
1008 Instruction *UI = dyn_cast<Instruction>(U);
1009 if (!UI || !L->contains(UI))
1012 Worklist.push_back(UI);
1014 // TODO: We could do other simplifications, for example, turning
1015 // 'icmp eq LIC, Val' -> false.
1017 // If we know that LIC is not Val, use this info to simplify code.
1018 SwitchInst *SI = dyn_cast<SwitchInst>(UI);
1019 if (!SI || !isa<ConstantInt>(Val)) continue;
1021 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1022 // Default case is live for multiple values.
1023 if (DeadCase == SI->case_default()) continue;
1025 // Found a dead case value. Don't remove PHI nodes in the
1026 // successor if they become single-entry, those PHI nodes may
1027 // be in the Users list.
1029 BasicBlock *Switch = SI->getParent();
1030 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1031 BasicBlock *Latch = L->getLoopLatch();
1033 BranchesInfo.setUnswitched(SI, Val);
1035 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1036 // If the DeadCase successor dominates the loop latch, then the
1037 // transformation isn't safe since it will delete the sole predecessor edge
1039 if (Latch && DT->dominates(SISucc, Latch))
1042 // FIXME: This is a hack. We need to keep the successor around
1043 // and hooked up so as to preserve the loop structure, because
1044 // trying to update it is complicated. So instead we preserve the
1045 // loop structure and put the block on a dead code path.
1046 SplitEdge(Switch, SISucc, this);
1047 // Compute the successors instead of relying on the return value
1048 // of SplitEdge, since it may have split the switch successor
1050 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1051 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1052 // Create an "unreachable" destination.
1053 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1054 Switch->getParent(),
1056 new UnreachableInst(Context, Abort);
1057 // Force the new case destination to branch to the "unreachable"
1058 // block while maintaining a (dead) CFG edge to the old block.
1059 NewSISucc->getTerminator()->eraseFromParent();
1060 BranchInst::Create(Abort, OldSISucc,
1061 ConstantInt::getTrue(Context), NewSISucc);
1062 // Release the PHI operands for this edge.
1063 for (BasicBlock::iterator II = NewSISucc->begin();
1064 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1065 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1066 UndefValue::get(PN->getType()));
1067 // Tell the domtree about the new block. We don't fully update the
1068 // domtree here -- instead we force it to do a full recomputation
1069 // after the pass is complete -- but we do need to inform it of
1072 DT->addNewBlock(Abort, NewSISucc);
1075 SimplifyCode(Worklist, L);
1078 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1079 /// loop, walk over it and constant prop, dce, and fold control flow where
1080 /// possible. Note that this is effectively a very simple loop-structure-aware
1081 /// optimizer. During processing of this loop, L could very well be deleted, so
1082 /// it must not be used.
1084 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1087 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1088 while (!Worklist.empty()) {
1089 Instruction *I = Worklist.back();
1090 Worklist.pop_back();
1093 if (isInstructionTriviallyDead(I)) {
1094 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1096 // Add uses to the worklist, which may be dead now.
1097 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1098 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1099 Worklist.push_back(Use);
1100 LPM->deleteSimpleAnalysisValue(I, L);
1101 RemoveFromWorklist(I, Worklist);
1102 I->eraseFromParent();
1107 // See if instruction simplification can hack this up. This is common for
1108 // things like "select false, X, Y" after unswitching made the condition be
1109 // 'false'. TODO: update the domtree properly so we can pass it here.
1110 if (Value *V = SimplifyInstruction(I))
1111 if (LI->replacementPreservesLCSSAForm(I, V)) {
1112 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1116 // Special case hacks that appear commonly in unswitched code.
1117 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1118 if (BI->isUnconditional()) {
1119 // If BI's parent is the only pred of the successor, fold the two blocks
1121 BasicBlock *Pred = BI->getParent();
1122 BasicBlock *Succ = BI->getSuccessor(0);
1123 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1124 if (!SinglePred) continue; // Nothing to do.
1125 assert(SinglePred == Pred && "CFG broken");
1127 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1128 << Succ->getName() << "\n");
1130 // Resolve any single entry PHI nodes in Succ.
1131 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1132 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1134 // If Succ has any successors with PHI nodes, update them to have
1135 // entries coming from Pred instead of Succ.
1136 Succ->replaceAllUsesWith(Pred);
1138 // Move all of the successor contents from Succ to Pred.
1139 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1141 LPM->deleteSimpleAnalysisValue(BI, L);
1142 BI->eraseFromParent();
1143 RemoveFromWorklist(BI, Worklist);
1145 // Remove Succ from the loop tree.
1146 LI->removeBlock(Succ);
1147 LPM->deleteSimpleAnalysisValue(Succ, L);
1148 Succ->eraseFromParent();