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/ConstantFolding.h"
36 #include "llvm/Analysis/LoopInfo.h"
37 #include "llvm/Analysis/LoopPass.h"
38 #include "llvm/Analysis/Dominators.h"
39 #include "llvm/Transforms/Utils/Cloning.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/Support/Debug.h"
51 STATISTIC(NumBranches, "Number of branches unswitched");
52 STATISTIC(NumSwitches, "Number of switches unswitched");
53 STATISTIC(NumSelects , "Number of selects unswitched");
54 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
55 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
57 static cl::opt<unsigned>
58 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
59 cl::init(10), cl::Hidden);
62 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
63 LoopInfo *LI; // Loop information
66 // LoopProcessWorklist - Used to check if second loop needs processing
67 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
68 std::vector<Loop*> LoopProcessWorklist;
69 SmallPtrSet<Value *,8> UnswitchedVals;
75 DominanceFrontier *DF;
77 BasicBlock *loopHeader;
78 BasicBlock *loopPreheader;
80 // LoopBlocks contains all of the basic blocks of the loop, including the
81 // preheader of the loop, the body of the loop, and the exit blocks of the
82 // loop, in that order.
83 std::vector<BasicBlock*> LoopBlocks;
84 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
85 std::vector<BasicBlock*> NewBlocks;
88 static char ID; // Pass ID, replacement for typeid
89 explicit LoopUnswitch(bool Os = false) :
90 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
91 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
92 loopPreheader(NULL) {}
94 bool runOnLoop(Loop *L, LPPassManager &LPM);
95 bool processCurrentLoop();
97 /// This transformation requires natural loop information & requires that
98 /// loop preheaders be inserted into the CFG...
100 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
101 AU.addRequiredID(LoopSimplifyID);
102 AU.addPreservedID(LoopSimplifyID);
103 AU.addRequired<LoopInfo>();
104 AU.addPreserved<LoopInfo>();
105 AU.addRequiredID(LCSSAID);
106 AU.addPreservedID(LCSSAID);
107 AU.addPreserved<DominatorTree>();
108 AU.addPreserved<DominanceFrontier>();
113 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
115 void RemoveLoopFromWorklist(Loop *L) {
116 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
117 LoopProcessWorklist.end(), L);
118 if (I != LoopProcessWorklist.end())
119 LoopProcessWorklist.erase(I);
122 void initLoopData() {
123 loopHeader = currentLoop->getHeader();
124 loopPreheader = currentLoop->getLoopPreheader();
127 /// Split all of the edges from inside the loop to their exit blocks.
128 /// Update the appropriate Phi nodes as we do so.
129 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
131 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
132 unsigned getLoopUnswitchCost(Value *LIC);
133 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
134 BasicBlock *ExitBlock);
135 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
137 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
138 Constant *Val, bool isEqual);
140 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
141 BasicBlock *TrueDest,
142 BasicBlock *FalseDest,
143 Instruction *InsertPt);
145 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
146 void RemoveBlockIfDead(BasicBlock *BB,
147 std::vector<Instruction*> &Worklist, Loop *l);
148 void RemoveLoopFromHierarchy(Loop *L);
149 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
150 BasicBlock **LoopExit = 0);
154 char LoopUnswitch::ID = 0;
155 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
157 LoopPass *llvm::createLoopUnswitchPass(bool Os) {
158 return new LoopUnswitch(Os);
161 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
162 /// invariant in the loop, or has an invariant piece, return the invariant.
163 /// Otherwise, return null.
164 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
165 // Constants should be folded, not unswitched on!
166 if (isa<Constant>(Cond)) return false;
168 // TODO: Handle: br (VARIANT|INVARIANT).
169 // TODO: Hoist simple expressions out of loops.
170 if (L->isLoopInvariant(Cond)) return Cond;
172 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
173 if (BO->getOpcode() == Instruction::And ||
174 BO->getOpcode() == Instruction::Or) {
175 // If either the left or right side is invariant, we can unswitch on this,
176 // which will cause the branch to go away in one loop and the condition to
177 // simplify in the other one.
178 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
180 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
187 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
188 LI = &getAnalysis<LoopInfo>();
190 DF = getAnalysisToUpdate<DominanceFrontier>();
191 DT = getAnalysisToUpdate<DominatorTree>();
193 bool Changed = false;
195 assert(currentLoop->isLCSSAForm());
197 Changed |= processCurrentLoop();
203 /// processCurrentLoop - Do actual work and unswitch loop if possible
205 bool LoopUnswitch::processCurrentLoop() {
206 bool Changed = false;
208 // Loop over all of the basic blocks in the loop. If we find an interior
209 // block that is branching on a loop-invariant condition, we can unswitch this
211 for (Loop::block_iterator I = currentLoop->block_begin(),
212 E = currentLoop->block_end();
214 TerminatorInst *TI = (*I)->getTerminator();
215 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
216 // If this isn't branching on an invariant condition, we can't unswitch
218 if (BI->isConditional()) {
219 // See if this, or some part of it, is loop invariant. If so, we can
220 // unswitch on it if we desire.
221 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
222 currentLoop, Changed);
223 if (LoopCond && UnswitchIfProfitable(LoopCond,
224 ConstantInt::getTrue())) {
229 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
230 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
231 currentLoop, Changed);
232 if (LoopCond && SI->getNumCases() > 1) {
233 // Find a value to unswitch on:
234 // FIXME: this should chose the most expensive case!
235 Constant *UnswitchVal = SI->getCaseValue(1);
236 // Do not process same value again and again.
237 if (!UnswitchedVals.insert(UnswitchVal))
240 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
247 // Scan the instructions to check for unswitchable values.
248 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
250 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
251 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
252 currentLoop, Changed);
253 if (LoopCond && UnswitchIfProfitable(LoopCond,
254 ConstantInt::getTrue())) {
263 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
264 /// 1. Exit the loop with no side effects.
265 /// 2. Branch to the latch block with no side-effects.
267 /// If these conditions are true, we return true and set ExitBB to the block we
270 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
272 std::set<BasicBlock*> &Visited) {
273 if (!Visited.insert(BB).second) {
274 // Already visited and Ok, end of recursion.
276 } else if (!L->contains(BB)) {
277 // Otherwise, this is a loop exit, this is fine so long as this is the
279 if (ExitBB != 0) return false;
284 // Otherwise, this is an unvisited intra-loop node. Check all successors.
285 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
286 // Check to see if the successor is a trivial loop exit.
287 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
291 // Okay, everything after this looks good, check to make sure that this block
292 // doesn't include any side effects.
293 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
294 if (I->mayWriteToMemory())
300 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
301 /// leads to an exit from the specified loop, and has no side-effects in the
302 /// process. If so, return the block that is exited to, otherwise return null.
303 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
304 std::set<BasicBlock*> Visited;
305 Visited.insert(L->getHeader()); // Branches to header are ok.
306 BasicBlock *ExitBB = 0;
307 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
312 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
313 /// trivial: that is, that the condition controls whether or not the loop does
314 /// anything at all. If this is a trivial condition, unswitching produces no
315 /// code duplications (equivalently, it produces a simpler loop and a new empty
316 /// loop, which gets deleted).
318 /// If this is a trivial condition, return true, otherwise return false. When
319 /// returning true, this sets Cond and Val to the condition that controls the
320 /// trivial condition: when Cond dynamically equals Val, the loop is known to
321 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
324 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
325 BasicBlock **LoopExit) {
326 BasicBlock *Header = currentLoop->getHeader();
327 TerminatorInst *HeaderTerm = Header->getTerminator();
329 BasicBlock *LoopExitBB = 0;
330 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
331 // If the header block doesn't end with a conditional branch on Cond, we
333 if (!BI->isConditional() || BI->getCondition() != Cond)
336 // Check to see if a successor of the branch is guaranteed to go to the
337 // latch block or exit through a one exit block without having any
338 // side-effects. If so, determine the value of Cond that causes it to do
340 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
341 BI->getSuccessor(0)))) {
342 if (Val) *Val = ConstantInt::getTrue();
343 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
344 BI->getSuccessor(1)))) {
345 if (Val) *Val = ConstantInt::getFalse();
347 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
348 // If this isn't a switch on Cond, we can't handle it.
349 if (SI->getCondition() != Cond) return false;
351 // Check to see if a successor of the switch 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
354 // this. Note that we can't trivially unswitch on the default case.
355 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
356 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
357 SI->getSuccessor(i)))) {
358 // Okay, we found a trivial case, remember the value that is trivial.
359 if (Val) *Val = SI->getCaseValue(i);
364 // If we didn't find a single unique LoopExit block, or if the loop exit block
365 // contains phi nodes, this isn't trivial.
366 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
367 return false; // Can't handle this.
369 if (LoopExit) *LoopExit = LoopExitBB;
371 // We already know that nothing uses any scalar values defined inside of this
372 // loop. As such, we just have to check to see if this loop will execute any
373 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
374 // part of the loop that the code *would* execute. We already checked the
375 // tail, check the header now.
376 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
377 if (I->mayWriteToMemory())
382 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
383 /// we choose to unswitch current loop on the specified value.
385 unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
386 // If the condition is trivial, always unswitch. There is no code growth for
388 if (IsTrivialUnswitchCondition(LIC))
391 // FIXME: This is really overly conservative. However, more liberal
392 // estimations have thus far resulted in excessive unswitching, which is bad
393 // both in compile time and in code size. This should be replaced once
394 // someone figures out how a good estimation.
395 return currentLoop->getBlocks().size();
398 // FIXME: this is brain dead. It should take into consideration code
400 for (Loop::block_iterator I = currentLoop->block_begin(),
401 E = currentLoop->block_end();
404 // Do not include empty blocks in the cost calculation. This happen due to
405 // loop canonicalization and will be removed.
406 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
409 // Count basic blocks.
416 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
417 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
418 /// unswitch the loop, reprocess the pieces, then return true.
419 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
422 Function *F = loopHeader->getParent();
424 // Do not unswitch if the function is optimized for size.
425 if (F->getNotes() & FN_NOTE_OptimizeForSize)
428 // Check to see if it would be profitable to unswitch current loop.
429 unsigned Cost = getLoopUnswitchCost(LoopCond);
431 // Do not do non-trivial unswitch while optimizing for size.
432 if (Cost && OptimizeForSize)
435 if (Cost > Threshold) {
436 // FIXME: this should estimate growth by the amount of code shared by the
437 // resultant unswitched loops.
439 DOUT << "NOT unswitching loop %"
440 << currentLoop->getHeader()->getName() << ", cost too high: "
441 << currentLoop->getBlocks().size() << "\n";
446 BasicBlock *ExitBlock;
447 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
448 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
450 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
453 // FIXME: Reconstruct dom info, because it is not preserved properly.
455 DT->runOnFunction(*F);
457 DF->runOnFunction(*F);
461 // RemapInstruction - Convert the instruction operands from referencing the
462 // current values into those specified by ValueMap.
464 static inline void RemapInstruction(Instruction *I,
465 DenseMap<const Value *, Value*> &ValueMap) {
466 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
467 Value *Op = I->getOperand(op);
468 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
469 if (It != ValueMap.end()) Op = It->second;
470 I->setOperand(op, Op);
474 /// CloneLoop - Recursively clone the specified loop and all of its children,
475 /// mapping the blocks with the specified map.
476 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
477 LoopInfo *LI, LPPassManager *LPM) {
478 Loop *New = new Loop();
480 LPM->insertLoop(New, PL);
482 // Add all of the blocks in L to the new loop.
483 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
485 if (LI->getLoopFor(*I) == L)
486 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
488 // Add all of the subloops to the new loop.
489 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
490 CloneLoop(*I, New, VM, LI, LPM);
495 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
496 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
497 /// code immediately before InsertPt.
498 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
499 BasicBlock *TrueDest,
500 BasicBlock *FalseDest,
501 Instruction *InsertPt) {
502 // Insert a conditional branch on LIC to the two preheaders. The original
503 // code is the true version and the new code is the false version.
504 Value *BranchVal = LIC;
505 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
506 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
507 else if (Val != ConstantInt::getTrue())
508 // We want to enter the new loop when the condition is true.
509 std::swap(TrueDest, FalseDest);
511 // Insert the new branch.
512 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
515 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
516 /// condition in it (a cond branch from its header block to its latch block,
517 /// where the path through the loop that doesn't execute its body has no
518 /// side-effects), unswitch it. This doesn't involve any code duplication, just
519 /// moving the conditional branch outside of the loop and updating loop info.
520 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
522 BasicBlock *ExitBlock) {
523 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
524 << loopHeader->getName() << " [" << L->getBlocks().size()
525 << " blocks] in Function " << L->getHeader()->getParent()->getName()
526 << " on cond: " << *Val << " == " << *Cond << "\n";
528 // First step, split the preheader, so that we know that there is a safe place
529 // to insert the conditional branch. We will change loopPreheader to have a
530 // conditional branch on Cond.
531 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
533 // Now that we have a place to insert the conditional branch, create a place
534 // to branch to: this is the exit block out of the loop that we should
537 // Split this block now, so that the loop maintains its exit block, and so
538 // that the jump from the preheader can execute the contents of the exit block
539 // without actually branching to it (the exit block should be dominated by the
540 // loop header, not the preheader).
541 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
542 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
544 // Okay, now we have a position to branch from and a position to branch to,
545 // insert the new conditional branch.
546 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
547 loopPreheader->getTerminator());
548 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
549 loopPreheader->getTerminator()->eraseFromParent();
551 // We need to reprocess this loop, it could be unswitched again.
554 // Now that we know that the loop is never entered when this condition is a
555 // particular value, rewrite the loop with this info. We know that this will
556 // at least eliminate the old branch.
557 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
561 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
562 /// blocks. Update the appropriate Phi nodes as we do so.
563 void LoopUnswitch::SplitExitEdges(Loop *L,
564 const SmallVector<BasicBlock *, 8> &ExitBlocks)
567 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
568 BasicBlock *ExitBlock = ExitBlocks[i];
569 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
571 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
572 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
573 BasicBlock* StartBlock = Preds[j];
574 BasicBlock* EndBlock;
575 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
576 EndBlock = NewExitBlock;
577 NewExitBlock = EndBlock->getSinglePredecessor();;
579 EndBlock = ExitBlock;
582 std::set<PHINode*> InsertedPHIs;
583 PHINode* OldLCSSA = 0;
584 for (BasicBlock::iterator I = EndBlock->begin();
585 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
586 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
587 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
588 OldLCSSA->getName() + ".us-lcssa",
589 NewExitBlock->getTerminator());
590 NewLCSSA->addIncoming(OldValue, StartBlock);
591 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
593 InsertedPHIs.insert(NewLCSSA);
596 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
597 for (BasicBlock::iterator I = NewExitBlock->begin();
598 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
600 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
601 OldLCSSA->getName() + ".us-lcssa",
603 OldLCSSA->replaceAllUsesWith(NewLCSSA);
604 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
612 /// UnswitchNontrivialCondition - We determined that the loop is profitable
613 /// to unswitch when LIC equal Val. Split it into loop versions and test the
614 /// condition outside of either loop. Return the loops created as Out1/Out2.
615 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
617 Function *F = loopHeader->getParent();
618 DOUT << "loop-unswitch: Unswitching loop %"
619 << loopHeader->getName() << " [" << L->getBlocks().size()
620 << " blocks] in Function " << F->getName()
621 << " when '" << *Val << "' == " << *LIC << "\n";
626 // First step, split the preheader and exit blocks, and add these blocks to
627 // the LoopBlocks list.
628 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
629 LoopBlocks.push_back(NewPreheader);
631 // We want the loop to come after the preheader, but before the exit blocks.
632 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
634 SmallVector<BasicBlock*, 8> ExitBlocks;
635 L->getUniqueExitBlocks(ExitBlocks);
637 // Split all of the edges from inside the loop to their exit blocks. Update
638 // the appropriate Phi nodes as we do so.
639 SplitExitEdges(L, ExitBlocks);
641 // The exit blocks may have been changed due to edge splitting, recompute.
643 L->getUniqueExitBlocks(ExitBlocks);
645 // Add exit blocks to the loop blocks.
646 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
648 // Next step, clone all of the basic blocks that make up the loop (including
649 // the loop preheader and exit blocks), keeping track of the mapping between
650 // the instructions and blocks.
651 NewBlocks.reserve(LoopBlocks.size());
652 DenseMap<const Value*, Value*> ValueMap;
653 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
654 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
655 NewBlocks.push_back(New);
656 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
657 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
660 // Splice the newly inserted blocks into the function right before the
661 // original preheader.
662 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
663 NewBlocks[0], F->end());
665 // Now we create the new Loop object for the versioned loop.
666 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
667 Loop *ParentLoop = L->getParentLoop();
669 // Make sure to add the cloned preheader and exit blocks to the parent loop
671 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
674 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
675 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
676 // The new exit block should be in the same loop as the old one.
677 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
678 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
680 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
681 "Exit block should have been split to have one successor!");
682 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
684 // If the successor of the exit block had PHI nodes, add an entry for
687 for (BasicBlock::iterator I = ExitSucc->begin();
688 (PN = dyn_cast<PHINode>(I)); ++I) {
689 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
690 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
691 if (It != ValueMap.end()) V = It->second;
692 PN->addIncoming(V, NewExit);
696 // Rewrite the code to refer to itself.
697 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
698 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
699 E = NewBlocks[i]->end(); I != E; ++I)
700 RemapInstruction(I, ValueMap);
702 // Rewrite the original preheader to select between versions of the loop.
703 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
704 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
705 "Preheader splitting did not work correctly!");
707 // Emit the new branch that selects between the two versions of this loop.
708 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
709 LPM->deleteSimpleAnalysisValue(OldBR, L);
710 OldBR->eraseFromParent();
712 LoopProcessWorklist.push_back(NewLoop);
715 // Now we rewrite the original code to know that the condition is true and the
716 // new code to know that the condition is false.
717 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
719 // It's possible that simplifying one loop could cause the other to be
720 // deleted. If so, don't simplify it.
721 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
722 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
726 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
728 static void RemoveFromWorklist(Instruction *I,
729 std::vector<Instruction*> &Worklist) {
730 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
732 while (WI != Worklist.end()) {
733 unsigned Offset = WI-Worklist.begin();
735 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
739 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
740 /// program, replacing all uses with V and update the worklist.
741 static void ReplaceUsesOfWith(Instruction *I, Value *V,
742 std::vector<Instruction*> &Worklist,
743 Loop *L, LPPassManager *LPM) {
744 DOUT << "Replace with '" << *V << "': " << *I;
746 // Add uses to the worklist, which may be dead now.
747 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
748 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
749 Worklist.push_back(Use);
751 // Add users to the worklist which may be simplified now.
752 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
754 Worklist.push_back(cast<Instruction>(*UI));
755 LPM->deleteSimpleAnalysisValue(I, L);
756 RemoveFromWorklist(I, Worklist);
757 I->replaceAllUsesWith(V);
758 I->eraseFromParent();
762 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
763 /// information, and remove any dead successors it has.
765 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
766 std::vector<Instruction*> &Worklist,
768 if (pred_begin(BB) != pred_end(BB)) {
769 // This block isn't dead, since an edge to BB was just removed, see if there
770 // are any easy simplifications we can do now.
771 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
772 // If it has one pred, fold phi nodes in BB.
773 while (isa<PHINode>(BB->begin()))
774 ReplaceUsesOfWith(BB->begin(),
775 cast<PHINode>(BB->begin())->getIncomingValue(0),
778 // If this is the header of a loop and the only pred is the latch, we now
779 // have an unreachable loop.
780 if (Loop *L = LI->getLoopFor(BB))
781 if (loopHeader == BB && L->contains(Pred)) {
782 // Remove the branch from the latch to the header block, this makes
783 // the header dead, which will make the latch dead (because the header
784 // dominates the latch).
785 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
786 Pred->getTerminator()->eraseFromParent();
787 new UnreachableInst(Pred);
789 // The loop is now broken, remove it from LI.
790 RemoveLoopFromHierarchy(L);
792 // Reprocess the header, which now IS dead.
793 RemoveBlockIfDead(BB, Worklist, L);
797 // If pred ends in a uncond branch, add uncond branch to worklist so that
798 // the two blocks will get merged.
799 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
800 if (BI->isUnconditional())
801 Worklist.push_back(BI);
806 DOUT << "Nuking dead block: " << *BB;
808 // Remove the instructions in the basic block from the worklist.
809 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
810 RemoveFromWorklist(I, Worklist);
812 // Anything that uses the instructions in this basic block should have their
813 // uses replaced with undefs.
815 I->replaceAllUsesWith(UndefValue::get(I->getType()));
818 // If this is the edge to the header block for a loop, remove the loop and
819 // promote all subloops.
820 if (Loop *BBLoop = LI->getLoopFor(BB)) {
821 if (BBLoop->getLoopLatch() == BB)
822 RemoveLoopFromHierarchy(BBLoop);
825 // Remove the block from the loop info, which removes it from any loops it
830 // Remove phi node entries in successors for this block.
831 TerminatorInst *TI = BB->getTerminator();
832 std::vector<BasicBlock*> Succs;
833 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
834 Succs.push_back(TI->getSuccessor(i));
835 TI->getSuccessor(i)->removePredecessor(BB);
838 // Unique the successors, remove anything with multiple uses.
839 std::sort(Succs.begin(), Succs.end());
840 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
842 // Remove the basic block, including all of the instructions contained in it.
843 LPM->deleteSimpleAnalysisValue(BB, L);
844 BB->eraseFromParent();
845 // Remove successor blocks here that are not dead, so that we know we only
846 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
847 // then getting removed before we revisit them, which is badness.
849 for (unsigned i = 0; i != Succs.size(); ++i)
850 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
851 // One exception is loop headers. If this block was the preheader for a
852 // loop, then we DO want to visit the loop so the loop gets deleted.
853 // We know that if the successor is a loop header, that this loop had to
854 // be the preheader: the case where this was the latch block was handled
855 // above and headers can only have two predecessors.
856 if (!LI->isLoopHeader(Succs[i])) {
857 Succs.erase(Succs.begin()+i);
862 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
863 RemoveBlockIfDead(Succs[i], Worklist, L);
866 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
867 /// become unwrapped, either because the backedge was deleted, or because the
868 /// edge into the header was removed. If the edge into the header from the
869 /// latch block was removed, the loop is unwrapped but subloops are still alive,
870 /// so they just reparent loops. If the loops are actually dead, they will be
872 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
873 LPM->deleteLoopFromQueue(L);
874 RemoveLoopFromWorklist(L);
877 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
878 // the value specified by Val in the specified loop, or we know it does NOT have
879 // that value. Rewrite any uses of LIC or of properties correlated to it.
880 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
883 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
885 // FIXME: Support correlated properties, like:
892 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
893 // selects, switches.
894 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
895 std::vector<Instruction*> Worklist;
897 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
898 // in the loop with the appropriate one directly.
899 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
904 Replacement = ConstantInt::get(Type::Int1Ty,
905 !cast<ConstantInt>(Val)->getZExtValue());
907 for (unsigned i = 0, e = Users.size(); i != e; ++i)
908 if (Instruction *U = cast<Instruction>(Users[i])) {
909 if (!L->contains(U->getParent()))
911 U->replaceUsesOfWith(LIC, Replacement);
912 Worklist.push_back(U);
915 // Otherwise, we don't know the precise value of LIC, but we do know that it
916 // is certainly NOT "Val". As such, simplify any uses in the loop that we
917 // can. This case occurs when we unswitch switch statements.
918 for (unsigned i = 0, e = Users.size(); i != e; ++i)
919 if (Instruction *U = cast<Instruction>(Users[i])) {
920 if (!L->contains(U->getParent()))
923 Worklist.push_back(U);
925 // If we know that LIC is not Val, use this info to simplify code.
926 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
927 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
928 if (SI->getCaseValue(i) == Val) {
929 // Found a dead case value. Don't remove PHI nodes in the
930 // successor if they become single-entry, those PHI nodes may
931 // be in the Users list.
933 // FIXME: This is a hack. We need to keep the successor around
934 // and hooked up so as to preserve the loop structure, because
935 // trying to update it is complicated. So instead we preserve the
936 // loop structure and put the block on an dead code path.
938 BasicBlock *SISucc = SI->getSuccessor(i);
939 BasicBlock* Old = SI->getParent();
940 BasicBlock* Split = SplitBlock(Old, SI, this);
942 Instruction* OldTerm = Old->getTerminator();
943 BranchInst::Create(Split, SISucc,
944 ConstantInt::getTrue(), OldTerm);
946 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
947 Old->getTerminator()->eraseFromParent();
950 for (BasicBlock::iterator II = SISucc->begin();
951 (PN = dyn_cast<PHINode>(II)); ++II) {
952 Value *InVal = PN->removeIncomingValue(Split, false);
953 PN->addIncoming(InVal, Old);
962 // TODO: We could do other simplifications, for example, turning
963 // LIC == Val -> false.
967 SimplifyCode(Worklist, L);
970 /// SimplifyCode - Okay, now that we have simplified some instructions in the
971 /// loop, walk over it and constant prop, dce, and fold control flow where
972 /// possible. Note that this is effectively a very simple loop-structure-aware
973 /// optimizer. During processing of this loop, L could very well be deleted, so
974 /// it must not be used.
976 /// FIXME: When the loop optimizer is more mature, separate this out to a new
979 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
980 while (!Worklist.empty()) {
981 Instruction *I = Worklist.back();
984 // Simple constant folding.
985 if (Constant *C = ConstantFoldInstruction(I)) {
986 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
991 if (isInstructionTriviallyDead(I)) {
992 DOUT << "Remove dead instruction '" << *I;
994 // Add uses to the worklist, which may be dead now.
995 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
996 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
997 Worklist.push_back(Use);
998 LPM->deleteSimpleAnalysisValue(I, L);
999 RemoveFromWorklist(I, Worklist);
1000 I->eraseFromParent();
1005 // Special case hacks that appear commonly in unswitched code.
1006 switch (I->getOpcode()) {
1007 case Instruction::Select:
1008 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1009 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1014 case Instruction::And:
1015 if (isa<ConstantInt>(I->getOperand(0)) &&
1016 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1017 cast<BinaryOperator>(I)->swapOperands();
1018 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1019 if (CB->getType() == Type::Int1Ty) {
1020 if (CB->isOne()) // X & 1 -> X
1021 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1023 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1027 case Instruction::Or:
1028 if (isa<ConstantInt>(I->getOperand(0)) &&
1029 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1030 cast<BinaryOperator>(I)->swapOperands();
1031 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1032 if (CB->getType() == Type::Int1Ty) {
1033 if (CB->isOne()) // X | 1 -> 1
1034 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1036 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1040 case Instruction::Br: {
1041 BranchInst *BI = cast<BranchInst>(I);
1042 if (BI->isUnconditional()) {
1043 // If BI's parent is the only pred of the successor, fold the two blocks
1045 BasicBlock *Pred = BI->getParent();
1046 BasicBlock *Succ = BI->getSuccessor(0);
1047 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1048 if (!SinglePred) continue; // Nothing to do.
1049 assert(SinglePred == Pred && "CFG broken");
1051 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1052 << Succ->getName() << "\n";
1054 // Resolve any single entry PHI nodes in Succ.
1055 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1056 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1058 // Move all of the successor contents from Succ to Pred.
1059 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1061 LPM->deleteSimpleAnalysisValue(BI, L);
1062 BI->eraseFromParent();
1063 RemoveFromWorklist(BI, Worklist);
1065 // If Succ has any successors with PHI nodes, update them to have
1066 // entries coming from Pred instead of Succ.
1067 Succ->replaceAllUsesWith(Pred);
1069 // Remove Succ from the loop tree.
1070 LI->removeBlock(Succ);
1071 LPM->deleteSimpleAnalysisValue(Succ, L);
1072 Succ->eraseFromParent();
1074 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1075 // Conditional branch. Turn it into an unconditional branch, then
1076 // remove dead blocks.
1077 break; // FIXME: Enable.
1079 DOUT << "Folded branch: " << *BI;
1080 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1081 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1082 DeadSucc->removePredecessor(BI->getParent(), true);
1083 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1084 LPM->deleteSimpleAnalysisValue(BI, L);
1085 BI->eraseFromParent();
1086 RemoveFromWorklist(BI, Worklist);
1089 RemoveBlockIfDead(DeadSucc, Worklist, L);