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/LLVMContext.h"
36 #include "llvm/Analysis/ConstantFolding.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/LoopPass.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Transforms/Utils/Cloning.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/SmallPtrSet.h"
45 #include "llvm/ADT/STLExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Compiler.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");
60 static cl::opt<unsigned>
61 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
62 cl::init(10), cl::Hidden);
65 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
66 LoopInfo *LI; // Loop information
69 // LoopProcessWorklist - Used to check if second loop needs processing
70 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
71 std::vector<Loop*> LoopProcessWorklist;
72 SmallPtrSet<Value *,8> UnswitchedVals;
78 DominanceFrontier *DF;
80 BasicBlock *loopHeader;
81 BasicBlock *loopPreheader;
83 // LoopBlocks contains all of the basic blocks of the loop, including the
84 // preheader of the loop, the body of the loop, and the exit blocks of the
85 // loop, in that order.
86 std::vector<BasicBlock*> LoopBlocks;
87 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
88 std::vector<BasicBlock*> NewBlocks;
91 static char ID; // Pass ID, replacement for typeid
92 explicit LoopUnswitch(bool Os = false) :
93 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
94 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
95 loopPreheader(NULL) {}
97 bool runOnLoop(Loop *L, LPPassManager &LPM);
98 bool processCurrentLoop();
100 /// This transformation requires natural loop information & requires that
101 /// loop preheaders be inserted into the CFG...
103 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
104 AU.addRequiredID(LoopSimplifyID);
105 AU.addPreservedID(LoopSimplifyID);
106 AU.addRequired<LoopInfo>();
107 AU.addPreserved<LoopInfo>();
108 AU.addRequiredID(LCSSAID);
109 AU.addPreservedID(LCSSAID);
110 AU.addPreserved<DominatorTree>();
111 AU.addPreserved<DominanceFrontier>();
116 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
118 void RemoveLoopFromWorklist(Loop *L) {
119 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
120 LoopProcessWorklist.end(), L);
121 if (I != LoopProcessWorklist.end())
122 LoopProcessWorklist.erase(I);
125 void initLoopData() {
126 loopHeader = currentLoop->getHeader();
127 loopPreheader = currentLoop->getLoopPreheader();
130 /// Split all of the edges from inside the loop to their exit blocks.
131 /// Update the appropriate Phi nodes as we do so.
132 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
134 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
135 unsigned getLoopUnswitchCost(Value *LIC);
136 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
137 BasicBlock *ExitBlock);
138 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
140 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
141 Constant *Val, bool isEqual);
143 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
144 BasicBlock *TrueDest,
145 BasicBlock *FalseDest,
146 Instruction *InsertPt);
148 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
149 void RemoveBlockIfDead(BasicBlock *BB,
150 std::vector<Instruction*> &Worklist, Loop *l);
151 void RemoveLoopFromHierarchy(Loop *L);
152 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
153 BasicBlock **LoopExit = 0);
157 char LoopUnswitch::ID = 0;
158 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
160 Pass *llvm::createLoopUnswitchPass(bool Os) {
161 return new LoopUnswitch(Os);
164 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
165 /// invariant in the loop, or has an invariant piece, return the invariant.
166 /// Otherwise, return null.
167 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
168 // Constants should be folded, not unswitched on!
169 if (isa<Constant>(Cond)) return 0;
171 // TODO: Handle: br (VARIANT|INVARIANT).
173 // Hoist simple values out.
174 if (L->makeLoopInvariant(Cond, Changed))
177 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
178 if (BO->getOpcode() == Instruction::And ||
179 BO->getOpcode() == Instruction::Or) {
180 // If either the left or right side is invariant, we can unswitch on this,
181 // which will cause the branch to go away in one loop and the condition to
182 // simplify in the other one.
183 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
185 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
192 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
193 LI = &getAnalysis<LoopInfo>();
195 DF = getAnalysisIfAvailable<DominanceFrontier>();
196 DT = getAnalysisIfAvailable<DominatorTree>();
198 Function *F = currentLoop->getHeader()->getParent();
199 bool Changed = false;
201 assert(currentLoop->isLCSSAForm());
203 Changed |= processCurrentLoop();
207 // FIXME: Reconstruct dom info, because it is not preserved properly.
209 DT->runOnFunction(*F);
211 DF->runOnFunction(*F);
216 /// processCurrentLoop - Do actual work and unswitch loop if possible
218 bool LoopUnswitch::processCurrentLoop() {
219 bool Changed = false;
220 LLVMContext &Context = currentLoop->getHeader()->getContext();
222 // Loop over all of the basic blocks in the loop. If we find an interior
223 // block that is branching on a loop-invariant condition, we can unswitch this
225 for (Loop::block_iterator I = currentLoop->block_begin(),
226 E = currentLoop->block_end();
228 TerminatorInst *TI = (*I)->getTerminator();
229 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
230 // If this isn't branching on an invariant condition, we can't unswitch
232 if (BI->isConditional()) {
233 // See if this, or some part of it, is loop invariant. If so, we can
234 // unswitch on it if we desire.
235 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
236 currentLoop, Changed);
237 if (LoopCond && UnswitchIfProfitable(LoopCond,
238 ConstantInt::getTrue(Context))) {
243 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
244 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
245 currentLoop, Changed);
246 if (LoopCond && SI->getNumCases() > 1) {
247 // Find a value to unswitch on:
248 // FIXME: this should chose the most expensive case!
249 Constant *UnswitchVal = SI->getCaseValue(1);
250 // Do not process same value again and again.
251 if (!UnswitchedVals.insert(UnswitchVal))
254 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
261 // Scan the instructions to check for unswitchable values.
262 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
264 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
265 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
266 currentLoop, Changed);
267 if (LoopCond && UnswitchIfProfitable(LoopCond,
268 ConstantInt::getTrue(Context))) {
277 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
278 /// 1. Exit the loop with no side effects.
279 /// 2. Branch to the latch block with no side-effects.
281 /// If these conditions are true, we return true and set ExitBB to the block we
284 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
286 std::set<BasicBlock*> &Visited) {
287 if (!Visited.insert(BB).second) {
288 // Already visited and Ok, end of recursion.
290 } else if (!L->contains(BB)) {
291 // Otherwise, this is a loop exit, this is fine so long as this is the
293 if (ExitBB != 0) return false;
298 // Otherwise, this is an unvisited intra-loop node. Check all successors.
299 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
300 // Check to see if the successor is a trivial loop exit.
301 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
305 // Okay, everything after this looks good, check to make sure that this block
306 // doesn't include any side effects.
307 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
308 if (I->mayHaveSideEffects())
314 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
315 /// leads to an exit from the specified loop, and has no side-effects in the
316 /// process. If so, return the block that is exited to, otherwise return null.
317 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
318 std::set<BasicBlock*> Visited;
319 Visited.insert(L->getHeader()); // Branches to header are ok.
320 BasicBlock *ExitBB = 0;
321 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
326 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
327 /// trivial: that is, that the condition controls whether or not the loop does
328 /// anything at all. If this is a trivial condition, unswitching produces no
329 /// code duplications (equivalently, it produces a simpler loop and a new empty
330 /// loop, which gets deleted).
332 /// If this is a trivial condition, return true, otherwise return false. When
333 /// returning true, this sets Cond and Val to the condition that controls the
334 /// trivial condition: when Cond dynamically equals Val, the loop is known to
335 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
338 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
339 BasicBlock **LoopExit) {
340 BasicBlock *Header = currentLoop->getHeader();
341 TerminatorInst *HeaderTerm = Header->getTerminator();
342 LLVMContext &Context = Header->getContext();
344 BasicBlock *LoopExitBB = 0;
345 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
346 // If the header block doesn't end with a conditional branch on Cond, we
348 if (!BI->isConditional() || BI->getCondition() != Cond)
351 // Check to see if a successor of the branch is guaranteed to go to the
352 // latch block or exit through a one exit block without having any
353 // side-effects. If so, determine the value of Cond that causes it to do
355 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
356 BI->getSuccessor(0)))) {
357 if (Val) *Val = ConstantInt::getTrue(Context);
358 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
359 BI->getSuccessor(1)))) {
360 if (Val) *Val = ConstantInt::getFalse(Context);
362 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
363 // If this isn't a switch on Cond, we can't handle it.
364 if (SI->getCondition() != Cond) return false;
366 // Check to see if a successor of the switch is guaranteed to go to the
367 // latch block or exit through a one exit block without having any
368 // side-effects. If so, determine the value of Cond that causes it to do
369 // this. Note that we can't trivially unswitch on the default case.
370 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
371 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
372 SI->getSuccessor(i)))) {
373 // Okay, we found a trivial case, remember the value that is trivial.
374 if (Val) *Val = SI->getCaseValue(i);
379 // If we didn't find a single unique LoopExit block, or if the loop exit block
380 // contains phi nodes, this isn't trivial.
381 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
382 return false; // Can't handle this.
384 if (LoopExit) *LoopExit = LoopExitBB;
386 // We already know that nothing uses any scalar values defined inside of this
387 // loop. As such, we just have to check to see if this loop will execute any
388 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
389 // part of the loop that the code *would* execute. We already checked the
390 // tail, check the header now.
391 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
392 if (I->mayHaveSideEffects())
397 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
398 /// we choose to unswitch current loop on the specified value.
400 unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
401 // If the condition is trivial, always unswitch. There is no code growth for
403 if (IsTrivialUnswitchCondition(LIC))
406 // FIXME: This is really overly conservative. However, more liberal
407 // estimations have thus far resulted in excessive unswitching, which is bad
408 // both in compile time and in code size. This should be replaced once
409 // someone figures out how a good estimation.
410 return currentLoop->getBlocks().size();
413 // FIXME: this is brain dead. It should take into consideration code
415 for (Loop::block_iterator I = currentLoop->block_begin(),
416 E = currentLoop->block_end();
419 // Do not include empty blocks in the cost calculation. This happen due to
420 // loop canonicalization and will be removed.
421 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
424 // Count basic blocks.
431 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
432 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
433 /// unswitch the loop, reprocess the pieces, then return true.
434 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
437 Function *F = loopHeader->getParent();
440 // Check to see if it would be profitable to unswitch current loop.
441 unsigned Cost = getLoopUnswitchCost(LoopCond);
443 // Do not do non-trivial unswitch while optimizing for size.
444 if (Cost && OptimizeForSize)
446 if (Cost && !F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
449 if (Cost > Threshold) {
450 // FIXME: this should estimate growth by the amount of code shared by the
451 // resultant unswitched loops.
453 DEBUG(errs() << "NOT unswitching loop %"
454 << currentLoop->getHeader()->getName() << ", cost too high: "
455 << currentLoop->getBlocks().size() << "\n");
460 BasicBlock *ExitBlock;
461 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
462 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
464 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
470 // RemapInstruction - Convert the instruction operands from referencing the
471 // current values into those specified by ValueMap.
473 static inline void RemapInstruction(Instruction *I,
474 DenseMap<const Value *, Value*> &ValueMap) {
475 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
476 Value *Op = I->getOperand(op);
477 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
478 if (It != ValueMap.end()) Op = It->second;
479 I->setOperand(op, Op);
483 /// CloneLoop - Recursively clone the specified loop and all of its children,
484 /// mapping the blocks with the specified map.
485 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
486 LoopInfo *LI, LPPassManager *LPM) {
487 Loop *New = new Loop();
489 LPM->insertLoop(New, PL);
491 // Add all of the blocks in L to the new loop.
492 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
494 if (LI->getLoopFor(*I) == L)
495 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
497 // Add all of the subloops to the new loop.
498 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
499 CloneLoop(*I, New, VM, LI, LPM);
504 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
505 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
506 /// code immediately before InsertPt.
507 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
508 BasicBlock *TrueDest,
509 BasicBlock *FalseDest,
510 Instruction *InsertPt) {
511 // Insert a conditional branch on LIC to the two preheaders. The original
512 // code is the true version and the new code is the false version.
513 Value *BranchVal = LIC;
514 if (!isa<ConstantInt>(Val) ||
515 Val->getType() != Type::getInt1Ty(LIC->getContext()))
516 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
517 else if (Val != ConstantInt::getTrue(Val->getContext()))
518 // We want to enter the new loop when the condition is true.
519 std::swap(TrueDest, FalseDest);
521 // Insert the new branch.
522 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
525 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
526 /// condition in it (a cond branch from its header block to its latch block,
527 /// where the path through the loop that doesn't execute its body has no
528 /// side-effects), unswitch it. This doesn't involve any code duplication, just
529 /// moving the conditional branch outside of the loop and updating loop info.
530 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
532 BasicBlock *ExitBlock) {
533 DEBUG(errs() << "loop-unswitch: Trivial-Unswitch loop %"
534 << loopHeader->getName() << " [" << L->getBlocks().size()
535 << " blocks] in Function " << L->getHeader()->getParent()->getName()
536 << " on cond: " << *Val << " == " << *Cond << "\n");
538 // First step, split the preheader, so that we know that there is a safe place
539 // to insert the conditional branch. We will change loopPreheader to have a
540 // conditional branch on Cond.
541 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
543 // Now that we have a place to insert the conditional branch, create a place
544 // to branch to: this is the exit block out of the loop that we should
547 // Split this block now, so that the loop maintains its exit block, and so
548 // that the jump from the preheader can execute the contents of the exit block
549 // without actually branching to it (the exit block should be dominated by the
550 // loop header, not the preheader).
551 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
552 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
554 // Okay, now we have a position to branch from and a position to branch to,
555 // insert the new conditional branch.
556 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
557 loopPreheader->getTerminator());
558 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
559 loopPreheader->getTerminator()->eraseFromParent();
561 // We need to reprocess this loop, it could be unswitched again.
564 // Now that we know that the loop is never entered when this condition is a
565 // particular value, rewrite the loop with this info. We know that this will
566 // at least eliminate the old branch.
567 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
571 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
572 /// blocks. Update the appropriate Phi nodes as we do so.
573 void LoopUnswitch::SplitExitEdges(Loop *L,
574 const SmallVector<BasicBlock *, 8> &ExitBlocks)
577 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
578 BasicBlock *ExitBlock = ExitBlocks[i];
579 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
581 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
582 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
583 BasicBlock* StartBlock = Preds[j];
584 BasicBlock* EndBlock;
585 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
586 EndBlock = NewExitBlock;
587 NewExitBlock = EndBlock->getSinglePredecessor();
589 EndBlock = ExitBlock;
592 std::set<PHINode*> InsertedPHIs;
593 PHINode* OldLCSSA = 0;
594 for (BasicBlock::iterator I = EndBlock->begin();
595 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
596 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
597 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
598 OldLCSSA->getName() + ".us-lcssa",
599 NewExitBlock->getTerminator());
600 NewLCSSA->addIncoming(OldValue, StartBlock);
601 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
603 InsertedPHIs.insert(NewLCSSA);
606 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
607 for (BasicBlock::iterator I = NewExitBlock->begin();
608 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
610 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
611 OldLCSSA->getName() + ".us-lcssa",
613 OldLCSSA->replaceAllUsesWith(NewLCSSA);
614 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
622 /// UnswitchNontrivialCondition - We determined that the loop is profitable
623 /// to unswitch when LIC equal Val. Split it into loop versions and test the
624 /// condition outside of either loop. Return the loops created as Out1/Out2.
625 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
627 Function *F = loopHeader->getParent();
628 DEBUG(errs() << "loop-unswitch: Unswitching loop %"
629 << loopHeader->getName() << " [" << L->getBlocks().size()
630 << " blocks] in Function " << F->getName()
631 << " when '" << *Val << "' == " << *LIC << "\n");
636 // First step, split the preheader and exit blocks, and add these blocks to
637 // the LoopBlocks list.
638 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
639 LoopBlocks.push_back(NewPreheader);
641 // We want the loop to come after the preheader, but before the exit blocks.
642 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
644 SmallVector<BasicBlock*, 8> ExitBlocks;
645 L->getUniqueExitBlocks(ExitBlocks);
647 // Split all of the edges from inside the loop to their exit blocks. Update
648 // the appropriate Phi nodes as we do so.
649 SplitExitEdges(L, ExitBlocks);
651 // The exit blocks may have been changed due to edge splitting, recompute.
653 L->getUniqueExitBlocks(ExitBlocks);
655 // Add exit blocks to the loop blocks.
656 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
658 // Next step, clone all of the basic blocks that make up the loop (including
659 // the loop preheader and exit blocks), keeping track of the mapping between
660 // the instructions and blocks.
661 NewBlocks.reserve(LoopBlocks.size());
662 DenseMap<const Value*, Value*> ValueMap;
663 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
664 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
665 NewBlocks.push_back(New);
666 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
667 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
670 // Splice the newly inserted blocks into the function right before the
671 // original preheader.
672 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
673 NewBlocks[0], F->end());
675 // Now we create the new Loop object for the versioned loop.
676 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
677 Loop *ParentLoop = L->getParentLoop();
679 // Make sure to add the cloned preheader and exit blocks to the parent loop
681 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
684 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
685 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
686 // The new exit block should be in the same loop as the old one.
687 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
688 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
690 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
691 "Exit block should have been split to have one successor!");
692 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
694 // If the successor of the exit block had PHI nodes, add an entry for
697 for (BasicBlock::iterator I = ExitSucc->begin();
698 (PN = dyn_cast<PHINode>(I)); ++I) {
699 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
700 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
701 if (It != ValueMap.end()) V = It->second;
702 PN->addIncoming(V, NewExit);
706 // Rewrite the code to refer to itself.
707 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
708 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
709 E = NewBlocks[i]->end(); I != E; ++I)
710 RemapInstruction(I, ValueMap);
712 // Rewrite the original preheader to select between versions of the loop.
713 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
714 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
715 "Preheader splitting did not work correctly!");
717 // Emit the new branch that selects between the two versions of this loop.
718 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
719 LPM->deleteSimpleAnalysisValue(OldBR, L);
720 OldBR->eraseFromParent();
722 LoopProcessWorklist.push_back(NewLoop);
725 // Now we rewrite the original code to know that the condition is true and the
726 // new code to know that the condition is false.
727 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
729 // It's possible that simplifying one loop could cause the other to be
730 // deleted. If so, don't simplify it.
731 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
732 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
736 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
738 static void RemoveFromWorklist(Instruction *I,
739 std::vector<Instruction*> &Worklist) {
740 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
742 while (WI != Worklist.end()) {
743 unsigned Offset = WI-Worklist.begin();
745 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
749 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
750 /// program, replacing all uses with V and update the worklist.
751 static void ReplaceUsesOfWith(Instruction *I, Value *V,
752 std::vector<Instruction*> &Worklist,
753 Loop *L, LPPassManager *LPM) {
754 DEBUG(errs() << "Replace with '" << *V << "': " << *I);
756 // Add uses to the worklist, which may be dead now.
757 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
758 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
759 Worklist.push_back(Use);
761 // Add users to the worklist which may be simplified now.
762 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
764 Worklist.push_back(cast<Instruction>(*UI));
765 LPM->deleteSimpleAnalysisValue(I, L);
766 RemoveFromWorklist(I, Worklist);
767 I->replaceAllUsesWith(V);
768 I->eraseFromParent();
772 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
773 /// information, and remove any dead successors it has.
775 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
776 std::vector<Instruction*> &Worklist,
778 if (pred_begin(BB) != pred_end(BB)) {
779 // This block isn't dead, since an edge to BB was just removed, see if there
780 // are any easy simplifications we can do now.
781 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
782 // If it has one pred, fold phi nodes in BB.
783 while (isa<PHINode>(BB->begin()))
784 ReplaceUsesOfWith(BB->begin(),
785 cast<PHINode>(BB->begin())->getIncomingValue(0),
788 // If this is the header of a loop and the only pred is the latch, we now
789 // have an unreachable loop.
790 if (Loop *L = LI->getLoopFor(BB))
791 if (loopHeader == BB && L->contains(Pred)) {
792 // Remove the branch from the latch to the header block, this makes
793 // the header dead, which will make the latch dead (because the header
794 // dominates the latch).
795 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
796 Pred->getTerminator()->eraseFromParent();
797 new UnreachableInst(BB->getContext(), Pred);
799 // The loop is now broken, remove it from LI.
800 RemoveLoopFromHierarchy(L);
802 // Reprocess the header, which now IS dead.
803 RemoveBlockIfDead(BB, Worklist, L);
807 // If pred ends in a uncond branch, add uncond branch to worklist so that
808 // the two blocks will get merged.
809 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
810 if (BI->isUnconditional())
811 Worklist.push_back(BI);
816 DEBUG(errs() << "Nuking dead block: " << *BB);
818 // Remove the instructions in the basic block from the worklist.
819 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
820 RemoveFromWorklist(I, Worklist);
822 // Anything that uses the instructions in this basic block should have their
823 // uses replaced with undefs.
825 I->replaceAllUsesWith(UndefValue::get(I->getType()));
828 // If this is the edge to the header block for a loop, remove the loop and
829 // promote all subloops.
830 if (Loop *BBLoop = LI->getLoopFor(BB)) {
831 if (BBLoop->getLoopLatch() == BB)
832 RemoveLoopFromHierarchy(BBLoop);
835 // Remove the block from the loop info, which removes it from any loops it
840 // Remove phi node entries in successors for this block.
841 TerminatorInst *TI = BB->getTerminator();
842 SmallVector<BasicBlock*, 4> Succs;
843 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
844 Succs.push_back(TI->getSuccessor(i));
845 TI->getSuccessor(i)->removePredecessor(BB);
848 // Unique the successors, remove anything with multiple uses.
849 array_pod_sort(Succs.begin(), Succs.end());
850 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
852 // Remove the basic block, including all of the instructions contained in it.
853 LPM->deleteSimpleAnalysisValue(BB, L);
854 BB->eraseFromParent();
855 // Remove successor blocks here that are not dead, so that we know we only
856 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
857 // then getting removed before we revisit them, which is badness.
859 for (unsigned i = 0; i != Succs.size(); ++i)
860 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
861 // One exception is loop headers. If this block was the preheader for a
862 // loop, then we DO want to visit the loop so the loop gets deleted.
863 // We know that if the successor is a loop header, that this loop had to
864 // be the preheader: the case where this was the latch block was handled
865 // above and headers can only have two predecessors.
866 if (!LI->isLoopHeader(Succs[i])) {
867 Succs.erase(Succs.begin()+i);
872 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
873 RemoveBlockIfDead(Succs[i], Worklist, L);
876 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
877 /// become unwrapped, either because the backedge was deleted, or because the
878 /// edge into the header was removed. If the edge into the header from the
879 /// latch block was removed, the loop is unwrapped but subloops are still alive,
880 /// so they just reparent loops. If the loops are actually dead, they will be
882 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
883 LPM->deleteLoopFromQueue(L);
884 RemoveLoopFromWorklist(L);
887 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
888 // the value specified by Val in the specified loop, or we know it does NOT have
889 // that value. Rewrite any uses of LIC or of properties correlated to it.
890 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
893 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
895 // FIXME: Support correlated properties, like:
902 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
903 // selects, switches.
904 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
905 std::vector<Instruction*> Worklist;
906 LLVMContext &Context = Val->getContext();
909 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
910 // in the loop with the appropriate one directly.
911 if (IsEqual || (isa<ConstantInt>(Val) &&
912 Val->getType() == Type::getInt1Ty(Val->getContext()))) {
917 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
918 !cast<ConstantInt>(Val)->getZExtValue());
920 for (unsigned i = 0, e = Users.size(); i != e; ++i)
921 if (Instruction *U = cast<Instruction>(Users[i])) {
922 if (!L->contains(U->getParent()))
924 U->replaceUsesOfWith(LIC, Replacement);
925 Worklist.push_back(U);
928 // Otherwise, we don't know the precise value of LIC, but we do know that it
929 // is certainly NOT "Val". As such, simplify any uses in the loop that we
930 // can. This case occurs when we unswitch switch statements.
931 for (unsigned i = 0, e = Users.size(); i != e; ++i)
932 if (Instruction *U = cast<Instruction>(Users[i])) {
933 if (!L->contains(U->getParent()))
936 Worklist.push_back(U);
938 // If we know that LIC is not Val, use this info to simplify code.
939 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
940 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
941 if (SI->getCaseValue(i) == Val) {
942 // Found a dead case value. Don't remove PHI nodes in the
943 // successor if they become single-entry, those PHI nodes may
944 // be in the Users list.
946 // FIXME: This is a hack. We need to keep the successor around
947 // and hooked up so as to preserve the loop structure, because
948 // trying to update it is complicated. So instead we preserve the
949 // loop structure and put the block on an dead code path.
951 BasicBlock *SISucc = SI->getSuccessor(i);
952 BasicBlock* Old = SI->getParent();
953 BasicBlock* Split = SplitBlock(Old, SI, this);
955 Instruction* OldTerm = Old->getTerminator();
956 BranchInst::Create(Split, SISucc,
957 ConstantInt::getTrue(Context), OldTerm);
959 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
960 Old->getTerminator()->eraseFromParent();
963 for (BasicBlock::iterator II = SISucc->begin();
964 (PN = dyn_cast<PHINode>(II)); ++II) {
965 Value *InVal = PN->removeIncomingValue(Split, false);
966 PN->addIncoming(InVal, Old);
975 // TODO: We could do other simplifications, for example, turning
976 // LIC == Val -> false.
980 SimplifyCode(Worklist, L);
983 /// SimplifyCode - Okay, now that we have simplified some instructions in the
984 /// loop, walk over it and constant prop, dce, and fold control flow where
985 /// possible. Note that this is effectively a very simple loop-structure-aware
986 /// optimizer. During processing of this loop, L could very well be deleted, so
987 /// it must not be used.
989 /// FIXME: When the loop optimizer is more mature, separate this out to a new
992 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
993 while (!Worklist.empty()) {
994 Instruction *I = Worklist.back();
997 // Simple constant folding.
998 if (Constant *C = ConstantFoldInstruction(I, I->getContext())) {
999 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
1004 if (isInstructionTriviallyDead(I)) {
1005 DEBUG(errs() << "Remove dead instruction '" << *I);
1007 // Add uses to the worklist, which may be dead now.
1008 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1009 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1010 Worklist.push_back(Use);
1011 LPM->deleteSimpleAnalysisValue(I, L);
1012 RemoveFromWorklist(I, Worklist);
1013 I->eraseFromParent();
1018 // Special case hacks that appear commonly in unswitched code.
1019 switch (I->getOpcode()) {
1020 case Instruction::Select:
1021 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1022 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1027 case Instruction::And:
1028 if (isa<ConstantInt>(I->getOperand(0)) &&
1030 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
1031 cast<BinaryOperator>(I)->swapOperands();
1032 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1033 if (CB->getType() == Type::getInt1Ty(I->getContext())) {
1034 if (CB->isOne()) // X & 1 -> X
1035 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1037 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1041 case Instruction::Or:
1042 if (isa<ConstantInt>(I->getOperand(0)) &&
1044 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
1045 cast<BinaryOperator>(I)->swapOperands();
1046 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1047 if (CB->getType() == Type::getInt1Ty(I->getContext())) {
1048 if (CB->isOne()) // X | 1 -> 1
1049 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1051 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1055 case Instruction::Br: {
1056 BranchInst *BI = cast<BranchInst>(I);
1057 if (BI->isUnconditional()) {
1058 // If BI's parent is the only pred of the successor, fold the two blocks
1060 BasicBlock *Pred = BI->getParent();
1061 BasicBlock *Succ = BI->getSuccessor(0);
1062 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1063 if (!SinglePred) continue; // Nothing to do.
1064 assert(SinglePred == Pred && "CFG broken");
1066 DEBUG(errs() << "Merging blocks: " << Pred->getName() << " <- "
1067 << Succ->getName() << "\n");
1069 // Resolve any single entry PHI nodes in Succ.
1070 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1071 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1073 // Move all of the successor contents from Succ to Pred.
1074 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1076 LPM->deleteSimpleAnalysisValue(BI, L);
1077 BI->eraseFromParent();
1078 RemoveFromWorklist(BI, Worklist);
1080 // If Succ has any successors with PHI nodes, update them to have
1081 // entries coming from Pred instead of Succ.
1082 Succ->replaceAllUsesWith(Pred);
1084 // Remove Succ from the loop tree.
1085 LI->removeBlock(Succ);
1086 LPM->deleteSimpleAnalysisValue(Succ, L);
1087 Succ->eraseFromParent();
1089 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1090 // Conditional branch. Turn it into an unconditional branch, then
1091 // remove dead blocks.
1092 break; // FIXME: Enable.
1094 DEBUG(errs() << "Folded branch: " << *BI);
1095 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1096 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1097 DeadSucc->removePredecessor(BI->getParent(), true);
1098 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1099 LPM->deleteSimpleAnalysisValue(BI, L);
1100 BI->eraseFromParent();
1101 RemoveFromWorklist(BI, Worklist);
1104 RemoveBlockIfDead(DeadSucc, Worklist, L);