1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #define DEBUG_TYPE "loop-unswitch"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Constants.h"
32 #include "llvm/DerivedTypes.h"
33 #include "llvm/Function.h"
34 #include "llvm/Instructions.h"
35 #include "llvm/Analysis/InlineCost.h"
36 #include "llvm/Analysis/InstructionSimplify.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/LoopPass.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Analysis/ScalarEvolution.h"
41 #include "llvm/Transforms/Utils/Cloning.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/STLExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
54 STATISTIC(NumBranches, "Number of branches unswitched");
55 STATISTIC(NumSwitches, "Number of switches unswitched");
56 STATISTIC(NumSelects , "Number of selects unswitched");
57 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
58 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
59 STATISTIC(TotalInsts, "Total number of instructions analyzed");
61 // The specific value of 50 here was chosen based only on intuition and a
62 // few specific examples.
63 static cl::opt<unsigned>
64 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
65 cl::init(100), cl::Hidden);
68 class LoopUnswitch : public LoopPass {
69 LoopInfo *LI; // Loop information
72 // LoopProcessWorklist - Used to check if second loop needs processing
73 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
74 std::vector<Loop*> LoopProcessWorklist;
76 struct LoopProperties {
77 unsigned CanBeUnswitchedCount;
78 unsigned SizeEstimation;
81 typedef DenseMap<const Loop*, LoopProperties> LoopPropsMap;
82 typedef LoopPropsMap::iterator LoopPropsMapIt;
83 LoopPropsMap LoopsProperties;
85 // Max size of code we can produce on remained iterations.
88 // FIXME: Consider custom class for this.
89 std::map<const SwitchInst*, SmallPtrSet<const Value *,8> > UnswitchedVals;
96 BasicBlock *loopHeader;
97 BasicBlock *loopPreheader;
99 // LoopBlocks contains all of the basic blocks of the loop, including the
100 // preheader of the loop, the body of the loop, and the exit blocks of the
101 // loop, in that order.
102 std::vector<BasicBlock*> LoopBlocks;
103 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
104 std::vector<BasicBlock*> NewBlocks;
107 static char ID; // Pass ID, replacement for typeid
108 explicit LoopUnswitch(bool Os = false) :
109 LoopPass(ID), MaxSize(Threshold), OptimizeForSize(Os), redoLoop(false),
110 currentLoop(NULL), DT(NULL), loopHeader(NULL),
111 loopPreheader(NULL) {
112 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
115 bool runOnLoop(Loop *L, LPPassManager &LPM);
116 bool processCurrentLoop();
118 /// This transformation requires natural loop information & requires that
119 /// loop preheaders be inserted into the CFG.
121 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
122 AU.addRequiredID(LoopSimplifyID);
123 AU.addPreservedID(LoopSimplifyID);
124 AU.addRequired<LoopInfo>();
125 AU.addPreserved<LoopInfo>();
126 AU.addRequiredID(LCSSAID);
127 AU.addPreservedID(LCSSAID);
128 AU.addPreserved<DominatorTree>();
129 AU.addPreserved<ScalarEvolution>();
134 virtual void releaseMemory() {
136 LoopPropsMapIt LIt = LoopsProperties.find(currentLoop);
138 if (LIt != LoopsProperties.end()) {
139 LoopProperties& Props = LIt->second;
140 MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation;
141 LoopsProperties.erase(LIt);
144 // We need to forget about all switches in the current loop.
145 // FIXME: Do it better than enumerating all blocks of code
146 // and see if it is a switch instruction.
147 for (Loop::block_iterator I = currentLoop->block_begin(),
148 E = currentLoop->block_end(); I != E; ++I) {
149 SwitchInst* SI = dyn_cast<SwitchInst>((*I)->getTerminator());
151 UnswitchedVals.erase(SI);
155 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
157 void RemoveLoopFromWorklist(Loop *L) {
158 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
159 LoopProcessWorklist.end(), L);
160 if (I != LoopProcessWorklist.end())
161 LoopProcessWorklist.erase(I);
164 /// For new loop switches we clone info about values that was
165 /// already unswitched and has redundant successors.
166 /// Note, that new loop data is stored inside the VMap.
167 void CloneUnswitchedVals(const ValueToValueMapTy& VMap,
168 const BasicBlock* SrcBB);
170 bool CountLoop(const Loop* L);
171 void CloneLoopProperties(const Loop* NewLoop, const Loop* OldLoop);
173 void initLoopData() {
174 loopHeader = currentLoop->getHeader();
175 loopPreheader = currentLoop->getLoopPreheader();
178 /// Split all of the edges from inside the loop to their exit blocks.
179 /// Update the appropriate Phi nodes as we do so.
180 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
182 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
183 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
184 BasicBlock *ExitBlock);
185 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
187 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
188 Constant *Val, bool isEqual);
190 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
191 BasicBlock *TrueDest,
192 BasicBlock *FalseDest,
193 Instruction *InsertPt);
195 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
196 void RemoveBlockIfDead(BasicBlock *BB,
197 std::vector<Instruction*> &Worklist, Loop *l);
198 void RemoveLoopFromHierarchy(Loop *L);
199 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
200 BasicBlock **LoopExit = 0);
204 char LoopUnswitch::ID = 0;
205 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
207 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
208 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
209 INITIALIZE_PASS_DEPENDENCY(LCSSA)
210 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
213 Pass *llvm::createLoopUnswitchPass(bool Os) {
214 return new LoopUnswitch(Os);
217 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
218 /// invariant in the loop, or has an invariant piece, return the invariant.
219 /// Otherwise, return null.
220 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
222 // We started analyze new instruction, increment scanned instructions counter.
225 // We can never unswitch on vector conditions.
226 if (Cond->getType()->isVectorTy())
229 // Constants should be folded, not unswitched on!
230 if (isa<Constant>(Cond)) return 0;
232 // TODO: Handle: br (VARIANT|INVARIANT).
234 // Hoist simple values out.
235 if (L->makeLoopInvariant(Cond, Changed))
238 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
239 if (BO->getOpcode() == Instruction::And ||
240 BO->getOpcode() == Instruction::Or) {
241 // If either the left or right side is invariant, we can unswitch on this,
242 // which will cause the branch to go away in one loop and the condition to
243 // simplify in the other one.
244 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
246 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
253 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
254 LI = &getAnalysis<LoopInfo>();
256 DT = getAnalysisIfAvailable<DominatorTree>();
258 Function *F = currentLoop->getHeader()->getParent();
259 bool Changed = false;
261 assert(currentLoop->isLCSSAForm(*DT));
263 Changed |= processCurrentLoop();
267 // FIXME: Reconstruct dom info, because it is not preserved properly.
269 DT->runOnFunction(*F);
274 /// processCurrentLoop - Do actual work and unswitch loop if possible
276 bool LoopUnswitch::processCurrentLoop() {
277 bool Changed = false;
281 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
285 LLVMContext &Context = loopHeader->getContext();
287 // Probably we reach the quota of branches for this loop. If so
289 if (!CountLoop(currentLoop))
292 // Loop over all of the basic blocks in the loop. If we find an interior
293 // block that is branching on a loop-invariant condition, we can unswitch this
295 for (Loop::block_iterator I = currentLoop->block_begin(),
296 E = currentLoop->block_end(); I != E; ++I) {
297 TerminatorInst *TI = (*I)->getTerminator();
298 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
299 // If this isn't branching on an invariant condition, we can't unswitch
301 if (BI->isConditional()) {
302 // See if this, or some part of it, is loop invariant. If so, we can
303 // unswitch on it if we desire.
304 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
305 currentLoop, Changed);
306 if (LoopCond && UnswitchIfProfitable(LoopCond,
307 ConstantInt::getTrue(Context))) {
312 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
313 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
314 currentLoop, Changed);
315 unsigned NumCases = SI->getNumCases();
316 if (LoopCond && NumCases > 1) {
317 // Find a value to unswitch on:
318 // FIXME: this should chose the most expensive case!
319 // FIXME: scan for a case with a non-critical edge?
320 Constant *UnswitchVal = NULL;
322 // Do not process same value again and again.
323 // At this point we have some cases already unswitched and
324 // some not yet unswitched. Let's find the first not yet unswitched one.
325 for (unsigned i = 1; i < NumCases; ++i) {
326 Constant* UnswitchValCandidate = SI->getCaseValue(i);
327 if (!UnswitchedVals[SI].count(UnswitchValCandidate)) {
328 UnswitchVal = UnswitchValCandidate;
336 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
343 // Scan the instructions to check for unswitchable values.
344 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
346 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
347 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
348 currentLoop, Changed);
349 if (LoopCond && UnswitchIfProfitable(LoopCond,
350 ConstantInt::getTrue(Context))) {
359 /// For new loop switches we clone info about values that was
360 /// already unswitched and has redundant successors.
361 /// Not that new loop data is stored inside the VMap.
362 void LoopUnswitch::CloneUnswitchedVals(const ValueToValueMapTy& VMap,
363 const BasicBlock* SrcBB) {
365 const SwitchInst* SI = dyn_cast<SwitchInst>(SrcBB->getTerminator());
366 if (SI && UnswitchedVals.count(SI)) {
367 // Don't clone a totally simplified switch.
368 if (isa<Constant>(SI->getCondition()))
370 Value* I = VMap.lookup(SI);
371 assert(I && "All instructions that are in SrcBB must be in VMap.");
372 UnswitchedVals[cast<SwitchInst>(I)] = UnswitchedVals[SI];
376 bool LoopUnswitch::CountLoop(const Loop* L) {
377 std::pair<LoopPropsMapIt, bool> InsertRes =
378 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
380 LoopProperties& Props = InsertRes.first->second;
382 if (InsertRes.second) {
385 // Limit the number of instructions to avoid causing significant code
386 // expansion, and the number of basic blocks, to avoid loops with
387 // large numbers of branches which cause loop unswitching to go crazy.
388 // This is a very ad-hoc heuristic.
390 // FIXME: This is overly conservative because it does not take into
391 // consideration code simplification opportunities and code that can
392 // be shared by the resultant unswitched loops.
394 for (Loop::block_iterator I = L->block_begin(),
397 Metrics.analyzeBasicBlock(*I);
399 Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5);
400 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
401 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
404 if (!Props.CanBeUnswitchedCount) {
405 DEBUG(dbgs() << "NOT unswitching loop %"
406 << L->getHeader()->getName() << ", cost too high: "
407 << L->getBlocks().size() << "\n");
414 void LoopUnswitch::CloneLoopProperties(
415 const Loop* NewLoop, const Loop* OldLoop) {
417 LoopProperties& OldLoopProps = LoopsProperties[OldLoop];
418 LoopProperties& NewLoopProps = LoopsProperties[NewLoop];
420 --OldLoopProps.CanBeUnswitchedCount;
421 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
422 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
423 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
425 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
428 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
429 /// loop with no side effects (including infinite loops).
431 /// If true, we return true and set ExitBB to the block we
434 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
436 std::set<BasicBlock*> &Visited) {
437 if (!Visited.insert(BB).second) {
438 // Already visited. Without more analysis, this could indicate an infinite
441 } else if (!L->contains(BB)) {
442 // Otherwise, this is a loop exit, this is fine so long as this is the
444 if (ExitBB != 0) return false;
449 // Otherwise, this is an unvisited intra-loop node. Check all successors.
450 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
451 // Check to see if the successor is a trivial loop exit.
452 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
456 // Okay, everything after this looks good, check to make sure that this block
457 // doesn't include any side effects.
458 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
459 if (I->mayHaveSideEffects())
465 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
466 /// leads to an exit from the specified loop, and has no side-effects in the
467 /// process. If so, return the block that is exited to, otherwise return null.
468 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
469 std::set<BasicBlock*> Visited;
470 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
471 BasicBlock *ExitBB = 0;
472 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
477 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
478 /// trivial: that is, that the condition controls whether or not the loop does
479 /// anything at all. If this is a trivial condition, unswitching produces no
480 /// code duplications (equivalently, it produces a simpler loop and a new empty
481 /// loop, which gets deleted).
483 /// If this is a trivial condition, return true, otherwise return false. When
484 /// returning true, this sets Cond and Val to the condition that controls the
485 /// trivial condition: when Cond dynamically equals Val, the loop is known to
486 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
489 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
490 BasicBlock **LoopExit) {
491 BasicBlock *Header = currentLoop->getHeader();
492 TerminatorInst *HeaderTerm = Header->getTerminator();
493 LLVMContext &Context = Header->getContext();
495 BasicBlock *LoopExitBB = 0;
496 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
497 // If the header block doesn't end with a conditional branch on Cond, we
499 if (!BI->isConditional() || BI->getCondition() != Cond)
502 // Check to see if a successor of the branch is guaranteed to
503 // exit through a unique exit block without having any
504 // side-effects. If so, determine the value of Cond that causes it to do
506 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
507 BI->getSuccessor(0)))) {
508 if (Val) *Val = ConstantInt::getTrue(Context);
509 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
510 BI->getSuccessor(1)))) {
511 if (Val) *Val = ConstantInt::getFalse(Context);
513 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
514 // If this isn't a switch on Cond, we can't handle it.
515 if (SI->getCondition() != Cond) return false;
517 // Check to see if a successor of the switch is guaranteed to go to the
518 // latch block or exit through a one exit block without having any
519 // side-effects. If so, determine the value of Cond that causes it to do
521 // Note that we can't trivially unswitch on the default case or
522 // on already unswitched cases.
523 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) {
524 BasicBlock* LoopExitCandidate;
525 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
526 SI->getSuccessor(i)))) {
527 // Okay, we found a trivial case, remember the value that is trivial.
528 ConstantInt* CaseVal = SI->getCaseValue(i);
530 // Check that it was not unswitched before, since already unswitched
531 // trivial vals are looks trivial too.
532 if (UnswitchedVals[SI].count(CaseVal))
534 LoopExitBB = LoopExitCandidate;
535 if (Val) *Val = CaseVal;
541 // If we didn't find a single unique LoopExit block, or if the loop exit block
542 // contains phi nodes, this isn't trivial.
543 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
544 return false; // Can't handle this.
546 if (LoopExit) *LoopExit = LoopExitBB;
548 // We already know that nothing uses any scalar values defined inside of this
549 // loop. As such, we just have to check to see if this loop will execute any
550 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
551 // part of the loop that the code *would* execute. We already checked the
552 // tail, check the header now.
553 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
554 if (I->mayHaveSideEffects())
559 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
560 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
561 /// unswitch the loop, reprocess the pieces, then return true.
562 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
564 Function *F = loopHeader->getParent();
566 Constant *CondVal = 0;
567 BasicBlock *ExitBlock = 0;
568 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
569 // If the condition is trivial, always unswitch. There is no code growth
571 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
575 // Check to see if it would be profitable to unswitch current loop.
577 // Do not do non-trivial unswitch while optimizing for size.
578 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
581 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
585 /// CloneLoop - Recursively clone the specified loop and all of its children,
586 /// mapping the blocks with the specified map.
587 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
588 LoopInfo *LI, LPPassManager *LPM) {
589 Loop *New = new Loop();
590 LPM->insertLoop(New, PL);
592 // Add all of the blocks in L to the new loop.
593 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
595 if (LI->getLoopFor(*I) == L)
596 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
598 // Add all of the subloops to the new loop.
599 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
600 CloneLoop(*I, New, VM, LI, LPM);
605 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
606 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
607 /// code immediately before InsertPt.
608 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
609 BasicBlock *TrueDest,
610 BasicBlock *FalseDest,
611 Instruction *InsertPt) {
612 // Insert a conditional branch on LIC to the two preheaders. The original
613 // code is the true version and the new code is the false version.
614 Value *BranchVal = LIC;
615 if (!isa<ConstantInt>(Val) ||
616 Val->getType() != Type::getInt1Ty(LIC->getContext()))
617 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
618 else if (Val != ConstantInt::getTrue(Val->getContext()))
619 // We want to enter the new loop when the condition is true.
620 std::swap(TrueDest, FalseDest);
622 // Insert the new branch.
623 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
625 // If either edge is critical, split it. This helps preserve LoopSimplify
626 // form for enclosing loops.
627 SplitCriticalEdge(BI, 0, this);
628 SplitCriticalEdge(BI, 1, this);
631 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
632 /// condition in it (a cond branch from its header block to its latch block,
633 /// where the path through the loop that doesn't execute its body has no
634 /// side-effects), unswitch it. This doesn't involve any code duplication, just
635 /// moving the conditional branch outside of the loop and updating loop info.
636 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
638 BasicBlock *ExitBlock) {
639 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
640 << loopHeader->getName() << " [" << L->getBlocks().size()
641 << " blocks] in Function " << L->getHeader()->getParent()->getName()
642 << " on cond: " << *Val << " == " << *Cond << "\n");
644 // First step, split the preheader, so that we know that there is a safe place
645 // to insert the conditional branch. We will change loopPreheader to have a
646 // conditional branch on Cond.
647 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
649 // Now that we have a place to insert the conditional branch, create a place
650 // to branch to: this is the exit block out of the loop that we should
653 // Split this block now, so that the loop maintains its exit block, and so
654 // that the jump from the preheader can execute the contents of the exit block
655 // without actually branching to it (the exit block should be dominated by the
656 // loop header, not the preheader).
657 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
658 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
660 // Okay, now we have a position to branch from and a position to branch to,
661 // insert the new conditional branch.
662 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
663 loopPreheader->getTerminator());
664 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
665 loopPreheader->getTerminator()->eraseFromParent();
667 // We need to reprocess this loop, it could be unswitched again.
670 // Now that we know that the loop is never entered when this condition is a
671 // particular value, rewrite the loop with this info. We know that this will
672 // at least eliminate the old branch.
673 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
677 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
678 /// blocks. Update the appropriate Phi nodes as we do so.
679 void LoopUnswitch::SplitExitEdges(Loop *L,
680 const SmallVector<BasicBlock *, 8> &ExitBlocks){
682 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
683 BasicBlock *ExitBlock = ExitBlocks[i];
684 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
685 pred_end(ExitBlock));
687 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
688 // general, if we call it on all predecessors of all exits then it does.
689 if (!ExitBlock->isLandingPad()) {
690 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
692 SmallVector<BasicBlock*, 2> NewBBs;
693 SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
699 /// UnswitchNontrivialCondition - We determined that the loop is profitable
700 /// to unswitch when LIC equal Val. Split it into loop versions and test the
701 /// condition outside of either loop. Return the loops created as Out1/Out2.
702 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
704 Function *F = loopHeader->getParent();
705 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
706 << loopHeader->getName() << " [" << L->getBlocks().size()
707 << " blocks] in Function " << F->getName()
708 << " when '" << *Val << "' == " << *LIC << "\n");
710 if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
716 // First step, split the preheader and exit blocks, and add these blocks to
717 // the LoopBlocks list.
718 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
719 LoopBlocks.push_back(NewPreheader);
721 // We want the loop to come after the preheader, but before the exit blocks.
722 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
724 SmallVector<BasicBlock*, 8> ExitBlocks;
725 L->getUniqueExitBlocks(ExitBlocks);
727 // Split all of the edges from inside the loop to their exit blocks. Update
728 // the appropriate Phi nodes as we do so.
729 SplitExitEdges(L, ExitBlocks);
731 // The exit blocks may have been changed due to edge splitting, recompute.
733 L->getUniqueExitBlocks(ExitBlocks);
735 // Add exit blocks to the loop blocks.
736 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
738 // Next step, clone all of the basic blocks that make up the loop (including
739 // the loop preheader and exit blocks), keeping track of the mapping between
740 // the instructions and blocks.
741 NewBlocks.reserve(LoopBlocks.size());
742 ValueToValueMapTy VMap;
743 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
744 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
746 // Inherit simplified switches info for NewBB
747 // We needn't pass NewBB since its instructions are already contained
749 CloneUnswitchedVals(VMap, LoopBlocks[i]);
751 NewBlocks.push_back(NewBB);
752 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
753 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
756 // Splice the newly inserted blocks into the function right before the
757 // original preheader.
758 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
759 NewBlocks[0], F->end());
761 // Now we create the new Loop object for the versioned loop.
762 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
763 CloneLoopProperties(NewLoop, L);
764 Loop *ParentLoop = L->getParentLoop();
766 // Make sure to add the cloned preheader and exit blocks to the parent loop
768 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
771 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
772 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
773 // The new exit block should be in the same loop as the old one.
774 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
775 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
777 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
778 "Exit block should have been split to have one successor!");
779 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
781 // If the successor of the exit block had PHI nodes, add an entry for
784 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) {
785 PN = cast<PHINode>(I);
786 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
787 ValueToValueMapTy::iterator It = VMap.find(V);
788 if (It != VMap.end()) V = It->second;
789 PN->addIncoming(V, NewExit);
792 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
793 PN = PHINode::Create(LPad->getType(), 0, "",
794 ExitSucc->getFirstInsertionPt());
796 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
799 LandingPadInst *LPI = BB->getLandingPadInst();
800 LPI->replaceAllUsesWith(PN);
801 PN->addIncoming(LPI, BB);
806 // Rewrite the code to refer to itself.
807 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
808 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
809 E = NewBlocks[i]->end(); I != E; ++I)
810 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
812 // Rewrite the original preheader to select between versions of the loop.
813 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
814 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
815 "Preheader splitting did not work correctly!");
817 // Emit the new branch that selects between the two versions of this loop.
818 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
819 LPM->deleteSimpleAnalysisValue(OldBR, L);
820 OldBR->eraseFromParent();
822 LoopProcessWorklist.push_back(NewLoop);
825 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
826 // deletes the instruction (for example by simplifying a PHI that feeds into
827 // the condition that we're unswitching on), we don't rewrite the second
829 WeakVH LICHandle(LIC);
831 // Now we rewrite the original code to know that the condition is true and the
832 // new code to know that the condition is false.
833 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
835 // It's possible that simplifying one loop could cause the other to be
836 // changed to another value or a constant. If its a constant, don't simplify
838 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
839 LICHandle && !isa<Constant>(LICHandle))
840 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
843 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
845 static void RemoveFromWorklist(Instruction *I,
846 std::vector<Instruction*> &Worklist) {
847 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
849 while (WI != Worklist.end()) {
850 unsigned Offset = WI-Worklist.begin();
852 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
856 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
857 /// program, replacing all uses with V and update the worklist.
858 static void ReplaceUsesOfWith(Instruction *I, Value *V,
859 std::vector<Instruction*> &Worklist,
860 Loop *L, LPPassManager *LPM) {
861 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
863 // Add uses to the worklist, which may be dead now.
864 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
865 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
866 Worklist.push_back(Use);
868 // Add users to the worklist which may be simplified now.
869 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
871 Worklist.push_back(cast<Instruction>(*UI));
872 LPM->deleteSimpleAnalysisValue(I, L);
873 RemoveFromWorklist(I, Worklist);
874 I->replaceAllUsesWith(V);
875 I->eraseFromParent();
879 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
880 /// information, and remove any dead successors it has.
882 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
883 std::vector<Instruction*> &Worklist,
885 if (pred_begin(BB) != pred_end(BB)) {
886 // This block isn't dead, since an edge to BB was just removed, see if there
887 // are any easy simplifications we can do now.
888 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
889 // If it has one pred, fold phi nodes in BB.
890 while (isa<PHINode>(BB->begin()))
891 ReplaceUsesOfWith(BB->begin(),
892 cast<PHINode>(BB->begin())->getIncomingValue(0),
895 // If this is the header of a loop and the only pred is the latch, we now
896 // have an unreachable loop.
897 if (Loop *L = LI->getLoopFor(BB))
898 if (loopHeader == BB && L->contains(Pred)) {
899 // Remove the branch from the latch to the header block, this makes
900 // the header dead, which will make the latch dead (because the header
901 // dominates the latch).
902 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
903 Pred->getTerminator()->eraseFromParent();
904 new UnreachableInst(BB->getContext(), Pred);
906 // The loop is now broken, remove it from LI.
907 RemoveLoopFromHierarchy(L);
909 // Reprocess the header, which now IS dead.
910 RemoveBlockIfDead(BB, Worklist, L);
914 // If pred ends in a uncond branch, add uncond branch to worklist so that
915 // the two blocks will get merged.
916 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
917 if (BI->isUnconditional())
918 Worklist.push_back(BI);
923 DEBUG(dbgs() << "Nuking dead block: " << *BB);
925 // Remove the instructions in the basic block from the worklist.
926 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
927 RemoveFromWorklist(I, Worklist);
929 // Anything that uses the instructions in this basic block should have their
930 // uses replaced with undefs.
931 // If I is not void type then replaceAllUsesWith undef.
932 // This allows ValueHandlers and custom metadata to adjust itself.
933 if (!I->getType()->isVoidTy())
934 I->replaceAllUsesWith(UndefValue::get(I->getType()));
937 // If this is the edge to the header block for a loop, remove the loop and
938 // promote all subloops.
939 if (Loop *BBLoop = LI->getLoopFor(BB)) {
940 if (BBLoop->getLoopLatch() == BB) {
941 RemoveLoopFromHierarchy(BBLoop);
942 if (currentLoop == BBLoop) {
949 // Remove the block from the loop info, which removes it from any loops it
954 // Remove phi node entries in successors for this block.
955 TerminatorInst *TI = BB->getTerminator();
956 SmallVector<BasicBlock*, 4> Succs;
957 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
958 Succs.push_back(TI->getSuccessor(i));
959 TI->getSuccessor(i)->removePredecessor(BB);
962 // Unique the successors, remove anything with multiple uses.
963 array_pod_sort(Succs.begin(), Succs.end());
964 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
966 // Remove the basic block, including all of the instructions contained in it.
967 LPM->deleteSimpleAnalysisValue(BB, L);
968 BB->eraseFromParent();
969 // Remove successor blocks here that are not dead, so that we know we only
970 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
971 // then getting removed before we revisit them, which is badness.
973 for (unsigned i = 0; i != Succs.size(); ++i)
974 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
975 // One exception is loop headers. If this block was the preheader for a
976 // loop, then we DO want to visit the loop so the loop gets deleted.
977 // We know that if the successor is a loop header, that this loop had to
978 // be the preheader: the case where this was the latch block was handled
979 // above and headers can only have two predecessors.
980 if (!LI->isLoopHeader(Succs[i])) {
981 Succs.erase(Succs.begin()+i);
986 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
987 RemoveBlockIfDead(Succs[i], Worklist, L);
990 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
991 /// become unwrapped, either because the backedge was deleted, or because the
992 /// edge into the header was removed. If the edge into the header from the
993 /// latch block was removed, the loop is unwrapped but subloops are still alive,
994 /// so they just reparent loops. If the loops are actually dead, they will be
996 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
997 LPM->deleteLoopFromQueue(L);
998 RemoveLoopFromWorklist(L);
1001 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
1002 // the value specified by Val in the specified loop, or we know it does NOT have
1003 // that value. Rewrite any uses of LIC or of properties correlated to it.
1004 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
1007 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
1009 // FIXME: Support correlated properties, like:
1016 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
1017 // selects, switches.
1018 std::vector<Instruction*> Worklist;
1019 LLVMContext &Context = Val->getContext();
1022 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
1023 // in the loop with the appropriate one directly.
1024 if (IsEqual || (isa<ConstantInt>(Val) &&
1025 Val->getType()->isIntegerTy(1))) {
1030 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
1031 !cast<ConstantInt>(Val)->getZExtValue());
1033 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
1035 Instruction *U = dyn_cast<Instruction>(*UI);
1036 if (!U || !L->contains(U))
1038 Worklist.push_back(U);
1041 for (std::vector<Instruction*>::iterator UI = Worklist.begin();
1042 UI != Worklist.end(); ++UI)
1043 (*UI)->replaceUsesOfWith(LIC, Replacement);
1045 SimplifyCode(Worklist, L);
1049 // Otherwise, we don't know the precise value of LIC, but we do know that it
1050 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1051 // can. This case occurs when we unswitch switch statements.
1052 for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
1054 Instruction *U = dyn_cast<Instruction>(*UI);
1055 if (!U || !L->contains(U))
1058 Worklist.push_back(U);
1060 // TODO: We could do other simplifications, for example, turning
1061 // 'icmp eq LIC, Val' -> false.
1063 // If we know that LIC is not Val, use this info to simplify code.
1064 SwitchInst *SI = dyn_cast<SwitchInst>(U);
1065 if (SI == 0 || !isa<ConstantInt>(Val)) continue;
1067 unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1068 if (DeadCase == 0) continue; // Default case is live for multiple values.
1070 // Found a dead case value. Don't remove PHI nodes in the
1071 // successor if they become single-entry, those PHI nodes may
1072 // be in the Users list.
1074 BasicBlock *Switch = SI->getParent();
1075 BasicBlock *SISucc = SI->getSuccessor(DeadCase);
1076 BasicBlock *Latch = L->getLoopLatch();
1078 UnswitchedVals[SI].insert(Val);
1080 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1081 // If the DeadCase successor dominates the loop latch, then the
1082 // transformation isn't safe since it will delete the sole predecessor edge
1084 if (Latch && DT->dominates(SISucc, Latch))
1087 // FIXME: This is a hack. We need to keep the successor around
1088 // and hooked up so as to preserve the loop structure, because
1089 // trying to update it is complicated. So instead we preserve the
1090 // loop structure and put the block on a dead code path.
1091 SplitEdge(Switch, SISucc, this);
1092 // Compute the successors instead of relying on the return value
1093 // of SplitEdge, since it may have split the switch successor
1095 BasicBlock *NewSISucc = SI->getSuccessor(DeadCase);
1096 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1097 // Create an "unreachable" destination.
1098 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1099 Switch->getParent(),
1101 new UnreachableInst(Context, Abort);
1102 // Force the new case destination to branch to the "unreachable"
1103 // block while maintaining a (dead) CFG edge to the old block.
1104 NewSISucc->getTerminator()->eraseFromParent();
1105 BranchInst::Create(Abort, OldSISucc,
1106 ConstantInt::getTrue(Context), NewSISucc);
1107 // Release the PHI operands for this edge.
1108 for (BasicBlock::iterator II = NewSISucc->begin();
1109 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1110 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1111 UndefValue::get(PN->getType()));
1112 // Tell the domtree about the new block. We don't fully update the
1113 // domtree here -- instead we force it to do a full recomputation
1114 // after the pass is complete -- but we do need to inform it of
1117 DT->addNewBlock(Abort, NewSISucc);
1120 SimplifyCode(Worklist, L);
1123 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1124 /// loop, walk over it and constant prop, dce, and fold control flow where
1125 /// possible. Note that this is effectively a very simple loop-structure-aware
1126 /// optimizer. During processing of this loop, L could very well be deleted, so
1127 /// it must not be used.
1129 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1132 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1133 while (!Worklist.empty()) {
1134 Instruction *I = Worklist.back();
1135 Worklist.pop_back();
1138 if (isInstructionTriviallyDead(I)) {
1139 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1141 // Add uses to the worklist, which may be dead now.
1142 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1143 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1144 Worklist.push_back(Use);
1145 LPM->deleteSimpleAnalysisValue(I, L);
1146 RemoveFromWorklist(I, Worklist);
1147 I->eraseFromParent();
1152 // See if instruction simplification can hack this up. This is common for
1153 // things like "select false, X, Y" after unswitching made the condition be
1155 if (Value *V = SimplifyInstruction(I, 0, 0, DT))
1156 if (LI->replacementPreservesLCSSAForm(I, V)) {
1157 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1161 // Special case hacks that appear commonly in unswitched code.
1162 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1163 if (BI->isUnconditional()) {
1164 // If BI's parent is the only pred of the successor, fold the two blocks
1166 BasicBlock *Pred = BI->getParent();
1167 BasicBlock *Succ = BI->getSuccessor(0);
1168 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1169 if (!SinglePred) continue; // Nothing to do.
1170 assert(SinglePred == Pred && "CFG broken");
1172 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1173 << Succ->getName() << "\n");
1175 // Resolve any single entry PHI nodes in Succ.
1176 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1177 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1179 // If Succ has any successors with PHI nodes, update them to have
1180 // entries coming from Pred instead of Succ.
1181 Succ->replaceAllUsesWith(Pred);
1183 // Move all of the successor contents from Succ to Pred.
1184 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1186 LPM->deleteSimpleAnalysisValue(BI, L);
1187 BI->eraseFromParent();
1188 RemoveFromWorklist(BI, Worklist);
1190 // Remove Succ from the loop tree.
1191 LI->removeBlock(Succ);
1192 LPM->deleteSimpleAnalysisValue(Succ, L);
1193 Succ->eraseFromParent();
1198 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1199 // Conditional branch. Turn it into an unconditional branch, then
1200 // remove dead blocks.
1201 continue; // FIXME: Enable.
1203 DEBUG(dbgs() << "Folded branch: " << *BI);
1204 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1205 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1206 DeadSucc->removePredecessor(BI->getParent(), true);
1207 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1208 LPM->deleteSimpleAnalysisValue(BI, L);
1209 BI->eraseFromParent();
1210 RemoveFromWorklist(BI, Worklist);
1213 RemoveBlockIfDead(DeadSucc, Worklist, L);