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){
421 Function *F = loopHeader->getParent();
423 // Do not unswitch if the function is optimized for size.
424 if (F->getNotes() & FN_NOTE_OptimizeForSize)
427 // Check to see if it would be profitable to unswitch current loop.
428 unsigned Cost = getLoopUnswitchCost(LoopCond);
430 // Do not do non-trivial unswitch while optimizing for size.
431 if (Cost && OptimizeForSize)
434 if (Cost > Threshold) {
435 // FIXME: this should estimate growth by the amount of code shared by the
436 // resultant unswitched loops.
438 DOUT << "NOT unswitching loop %"
439 << currentLoop->getHeader()->getName() << ", cost too high: "
440 << currentLoop->getBlocks().size() << "\n";
447 BasicBlock *ExitBlock;
448 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
449 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
451 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
454 // FIXME: Reconstruct dom info, because it is not preserved properly.
456 DT->runOnFunction(*F);
458 DF->runOnFunction(*F);
462 // RemapInstruction - Convert the instruction operands from referencing the
463 // current values into those specified by ValueMap.
465 static inline void RemapInstruction(Instruction *I,
466 DenseMap<const Value *, Value*> &ValueMap) {
467 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
468 Value *Op = I->getOperand(op);
469 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
470 if (It != ValueMap.end()) Op = It->second;
471 I->setOperand(op, Op);
475 /// CloneLoop - Recursively clone the specified loop and all of its children,
476 /// mapping the blocks with the specified map.
477 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
478 LoopInfo *LI, LPPassManager *LPM) {
479 Loop *New = new Loop();
481 LPM->insertLoop(New, PL);
483 // Add all of the blocks in L to the new loop.
484 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
486 if (LI->getLoopFor(*I) == L)
487 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
489 // Add all of the subloops to the new loop.
490 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
491 CloneLoop(*I, New, VM, LI, LPM);
496 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
497 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
498 /// code immediately before InsertPt.
499 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
500 BasicBlock *TrueDest,
501 BasicBlock *FalseDest,
502 Instruction *InsertPt) {
503 // Insert a conditional branch on LIC to the two preheaders. The original
504 // code is the true version and the new code is the false version.
505 Value *BranchVal = LIC;
506 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
507 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
508 else if (Val != ConstantInt::getTrue())
509 // We want to enter the new loop when the condition is true.
510 std::swap(TrueDest, FalseDest);
512 // Insert the new branch.
513 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
516 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
517 /// condition in it (a cond branch from its header block to its latch block,
518 /// where the path through the loop that doesn't execute its body has no
519 /// side-effects), unswitch it. This doesn't involve any code duplication, just
520 /// moving the conditional branch outside of the loop and updating loop info.
521 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
523 BasicBlock *ExitBlock) {
524 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
525 << loopHeader->getName() << " [" << L->getBlocks().size()
526 << " blocks] in Function " << L->getHeader()->getParent()->getName()
527 << " on cond: " << *Val << " == " << *Cond << "\n";
529 // First step, split the preheader, so that we know that there is a safe place
530 // to insert the conditional branch. We will change loopPreheader to have a
531 // conditional branch on Cond.
532 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
534 // Now that we have a place to insert the conditional branch, create a place
535 // to branch to: this is the exit block out of the loop that we should
538 // Split this block now, so that the loop maintains its exit block, and so
539 // that the jump from the preheader can execute the contents of the exit block
540 // without actually branching to it (the exit block should be dominated by the
541 // loop header, not the preheader).
542 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
543 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
545 // Okay, now we have a position to branch from and a position to branch to,
546 // insert the new conditional branch.
547 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
548 loopPreheader->getTerminator());
549 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
550 loopPreheader->getTerminator()->eraseFromParent();
552 // We need to reprocess this loop, it could be unswitched again.
555 // Now that we know that the loop is never entered when this condition is a
556 // particular value, rewrite the loop with this info. We know that this will
557 // at least eliminate the old branch.
558 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
562 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
563 /// blocks. Update the appropriate Phi nodes as we do so.
564 void LoopUnswitch::SplitExitEdges(Loop *L,
565 const SmallVector<BasicBlock *, 8> &ExitBlocks)
568 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
569 BasicBlock *ExitBlock = ExitBlocks[i];
570 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
572 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
573 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
574 BasicBlock* StartBlock = Preds[j];
575 BasicBlock* EndBlock;
576 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
577 EndBlock = NewExitBlock;
578 NewExitBlock = EndBlock->getSinglePredecessor();;
580 EndBlock = ExitBlock;
583 std::set<PHINode*> InsertedPHIs;
584 PHINode* OldLCSSA = 0;
585 for (BasicBlock::iterator I = EndBlock->begin();
586 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
587 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
588 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
589 OldLCSSA->getName() + ".us-lcssa",
590 NewExitBlock->getTerminator());
591 NewLCSSA->addIncoming(OldValue, StartBlock);
592 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
594 InsertedPHIs.insert(NewLCSSA);
597 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
598 for (BasicBlock::iterator I = NewExitBlock->begin();
599 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
601 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
602 OldLCSSA->getName() + ".us-lcssa",
604 OldLCSSA->replaceAllUsesWith(NewLCSSA);
605 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
613 /// UnswitchNontrivialCondition - We determined that the loop is profitable
614 /// to unswitch when LIC equal Val. Split it into loop versions and test the
615 /// condition outside of either loop. Return the loops created as Out1/Out2.
616 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
618 Function *F = loopHeader->getParent();
619 DOUT << "loop-unswitch: Unswitching loop %"
620 << loopHeader->getName() << " [" << L->getBlocks().size()
621 << " blocks] in Function " << F->getName()
622 << " when '" << *Val << "' == " << *LIC << "\n";
627 // First step, split the preheader and exit blocks, and add these blocks to
628 // the LoopBlocks list.
629 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
630 LoopBlocks.push_back(NewPreheader);
632 // We want the loop to come after the preheader, but before the exit blocks.
633 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
635 SmallVector<BasicBlock*, 8> ExitBlocks;
636 L->getUniqueExitBlocks(ExitBlocks);
638 // Split all of the edges from inside the loop to their exit blocks. Update
639 // the appropriate Phi nodes as we do so.
640 SplitExitEdges(L, ExitBlocks);
642 // The exit blocks may have been changed due to edge splitting, recompute.
644 L->getUniqueExitBlocks(ExitBlocks);
646 // Add exit blocks to the loop blocks.
647 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
649 // Next step, clone all of the basic blocks that make up the loop (including
650 // the loop preheader and exit blocks), keeping track of the mapping between
651 // the instructions and blocks.
652 NewBlocks.reserve(LoopBlocks.size());
653 DenseMap<const Value*, Value*> ValueMap;
654 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
655 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
656 NewBlocks.push_back(New);
657 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
658 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
661 // Splice the newly inserted blocks into the function right before the
662 // original preheader.
663 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
664 NewBlocks[0], F->end());
666 // Now we create the new Loop object for the versioned loop.
667 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
668 Loop *ParentLoop = L->getParentLoop();
670 // Make sure to add the cloned preheader and exit blocks to the parent loop
672 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
675 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
676 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
677 // The new exit block should be in the same loop as the old one.
678 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
679 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
681 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
682 "Exit block should have been split to have one successor!");
683 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
685 // If the successor of the exit block had PHI nodes, add an entry for
688 for (BasicBlock::iterator I = ExitSucc->begin();
689 (PN = dyn_cast<PHINode>(I)); ++I) {
690 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
691 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
692 if (It != ValueMap.end()) V = It->second;
693 PN->addIncoming(V, NewExit);
697 // Rewrite the code to refer to itself.
698 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
699 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
700 E = NewBlocks[i]->end(); I != E; ++I)
701 RemapInstruction(I, ValueMap);
703 // Rewrite the original preheader to select between versions of the loop.
704 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
705 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
706 "Preheader splitting did not work correctly!");
708 // Emit the new branch that selects between the two versions of this loop.
709 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
710 LPM->deleteSimpleAnalysisValue(OldBR, L);
711 OldBR->eraseFromParent();
713 LoopProcessWorklist.push_back(NewLoop);
716 // Now we rewrite the original code to know that the condition is true and the
717 // new code to know that the condition is false.
718 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
720 // It's possible that simplifying one loop could cause the other to be
721 // deleted. If so, don't simplify it.
722 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
723 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
727 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
729 static void RemoveFromWorklist(Instruction *I,
730 std::vector<Instruction*> &Worklist) {
731 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
733 while (WI != Worklist.end()) {
734 unsigned Offset = WI-Worklist.begin();
736 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
740 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
741 /// program, replacing all uses with V and update the worklist.
742 static void ReplaceUsesOfWith(Instruction *I, Value *V,
743 std::vector<Instruction*> &Worklist,
744 Loop *L, LPPassManager *LPM) {
745 DOUT << "Replace with '" << *V << "': " << *I;
747 // Add uses to the worklist, which may be dead now.
748 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
749 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
750 Worklist.push_back(Use);
752 // Add users to the worklist which may be simplified now.
753 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
755 Worklist.push_back(cast<Instruction>(*UI));
756 LPM->deleteSimpleAnalysisValue(I, L);
757 RemoveFromWorklist(I, Worklist);
758 I->replaceAllUsesWith(V);
759 I->eraseFromParent();
763 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
764 /// information, and remove any dead successors it has.
766 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
767 std::vector<Instruction*> &Worklist,
769 if (pred_begin(BB) != pred_end(BB)) {
770 // This block isn't dead, since an edge to BB was just removed, see if there
771 // are any easy simplifications we can do now.
772 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
773 // If it has one pred, fold phi nodes in BB.
774 while (isa<PHINode>(BB->begin()))
775 ReplaceUsesOfWith(BB->begin(),
776 cast<PHINode>(BB->begin())->getIncomingValue(0),
779 // If this is the header of a loop and the only pred is the latch, we now
780 // have an unreachable loop.
781 if (Loop *L = LI->getLoopFor(BB))
782 if (loopHeader == BB && L->contains(Pred)) {
783 // Remove the branch from the latch to the header block, this makes
784 // the header dead, which will make the latch dead (because the header
785 // dominates the latch).
786 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
787 Pred->getTerminator()->eraseFromParent();
788 new UnreachableInst(Pred);
790 // The loop is now broken, remove it from LI.
791 RemoveLoopFromHierarchy(L);
793 // Reprocess the header, which now IS dead.
794 RemoveBlockIfDead(BB, Worklist, L);
798 // If pred ends in a uncond branch, add uncond branch to worklist so that
799 // the two blocks will get merged.
800 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
801 if (BI->isUnconditional())
802 Worklist.push_back(BI);
807 DOUT << "Nuking dead block: " << *BB;
809 // Remove the instructions in the basic block from the worklist.
810 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
811 RemoveFromWorklist(I, Worklist);
813 // Anything that uses the instructions in this basic block should have their
814 // uses replaced with undefs.
816 I->replaceAllUsesWith(UndefValue::get(I->getType()));
819 // If this is the edge to the header block for a loop, remove the loop and
820 // promote all subloops.
821 if (Loop *BBLoop = LI->getLoopFor(BB)) {
822 if (BBLoop->getLoopLatch() == BB)
823 RemoveLoopFromHierarchy(BBLoop);
826 // Remove the block from the loop info, which removes it from any loops it
831 // Remove phi node entries in successors for this block.
832 TerminatorInst *TI = BB->getTerminator();
833 std::vector<BasicBlock*> Succs;
834 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
835 Succs.push_back(TI->getSuccessor(i));
836 TI->getSuccessor(i)->removePredecessor(BB);
839 // Unique the successors, remove anything with multiple uses.
840 std::sort(Succs.begin(), Succs.end());
841 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
843 // Remove the basic block, including all of the instructions contained in it.
844 LPM->deleteSimpleAnalysisValue(BB, L);
845 BB->eraseFromParent();
846 // Remove successor blocks here that are not dead, so that we know we only
847 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
848 // then getting removed before we revisit them, which is badness.
850 for (unsigned i = 0; i != Succs.size(); ++i)
851 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
852 // One exception is loop headers. If this block was the preheader for a
853 // loop, then we DO want to visit the loop so the loop gets deleted.
854 // We know that if the successor is a loop header, that this loop had to
855 // be the preheader: the case where this was the latch block was handled
856 // above and headers can only have two predecessors.
857 if (!LI->isLoopHeader(Succs[i])) {
858 Succs.erase(Succs.begin()+i);
863 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
864 RemoveBlockIfDead(Succs[i], Worklist, L);
867 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
868 /// become unwrapped, either because the backedge was deleted, or because the
869 /// edge into the header was removed. If the edge into the header from the
870 /// latch block was removed, the loop is unwrapped but subloops are still alive,
871 /// so they just reparent loops. If the loops are actually dead, they will be
873 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
874 LPM->deleteLoopFromQueue(L);
875 RemoveLoopFromWorklist(L);
878 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
879 // the value specified by Val in the specified loop, or we know it does NOT have
880 // that value. Rewrite any uses of LIC or of properties correlated to it.
881 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
884 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
886 // FIXME: Support correlated properties, like:
893 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
894 // selects, switches.
895 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
896 std::vector<Instruction*> Worklist;
898 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
899 // in the loop with the appropriate one directly.
900 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
905 Replacement = ConstantInt::get(Type::Int1Ty,
906 !cast<ConstantInt>(Val)->getZExtValue());
908 for (unsigned i = 0, e = Users.size(); i != e; ++i)
909 if (Instruction *U = cast<Instruction>(Users[i])) {
910 if (!L->contains(U->getParent()))
912 U->replaceUsesOfWith(LIC, Replacement);
913 Worklist.push_back(U);
916 // Otherwise, we don't know the precise value of LIC, but we do know that it
917 // is certainly NOT "Val". As such, simplify any uses in the loop that we
918 // can. This case occurs when we unswitch switch statements.
919 for (unsigned i = 0, e = Users.size(); i != e; ++i)
920 if (Instruction *U = cast<Instruction>(Users[i])) {
921 if (!L->contains(U->getParent()))
924 Worklist.push_back(U);
926 // If we know that LIC is not Val, use this info to simplify code.
927 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
928 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
929 if (SI->getCaseValue(i) == Val) {
930 // Found a dead case value. Don't remove PHI nodes in the
931 // successor if they become single-entry, those PHI nodes may
932 // be in the Users list.
934 // FIXME: This is a hack. We need to keep the successor around
935 // and hooked up so as to preserve the loop structure, because
936 // trying to update it is complicated. So instead we preserve the
937 // loop structure and put the block on an dead code path.
939 BasicBlock *SISucc = SI->getSuccessor(i);
940 BasicBlock* Old = SI->getParent();
941 BasicBlock* Split = SplitBlock(Old, SI, this);
943 Instruction* OldTerm = Old->getTerminator();
944 BranchInst::Create(Split, SISucc,
945 ConstantInt::getTrue(), OldTerm);
947 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
948 Old->getTerminator()->eraseFromParent();
951 for (BasicBlock::iterator II = SISucc->begin();
952 (PN = dyn_cast<PHINode>(II)); ++II) {
953 Value *InVal = PN->removeIncomingValue(Split, false);
954 PN->addIncoming(InVal, Old);
963 // TODO: We could do other simplifications, for example, turning
964 // LIC == Val -> false.
968 SimplifyCode(Worklist, L);
971 /// SimplifyCode - Okay, now that we have simplified some instructions in the
972 /// loop, walk over it and constant prop, dce, and fold control flow where
973 /// possible. Note that this is effectively a very simple loop-structure-aware
974 /// optimizer. During processing of this loop, L could very well be deleted, so
975 /// it must not be used.
977 /// FIXME: When the loop optimizer is more mature, separate this out to a new
980 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
981 while (!Worklist.empty()) {
982 Instruction *I = Worklist.back();
985 // Simple constant folding.
986 if (Constant *C = ConstantFoldInstruction(I)) {
987 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
992 if (isInstructionTriviallyDead(I)) {
993 DOUT << "Remove dead instruction '" << *I;
995 // Add uses to the worklist, which may be dead now.
996 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
997 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
998 Worklist.push_back(Use);
999 LPM->deleteSimpleAnalysisValue(I, L);
1000 RemoveFromWorklist(I, Worklist);
1001 I->eraseFromParent();
1006 // Special case hacks that appear commonly in unswitched code.
1007 switch (I->getOpcode()) {
1008 case Instruction::Select:
1009 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1010 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1015 case Instruction::And:
1016 if (isa<ConstantInt>(I->getOperand(0)) &&
1017 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1018 cast<BinaryOperator>(I)->swapOperands();
1019 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1020 if (CB->getType() == Type::Int1Ty) {
1021 if (CB->isOne()) // X & 1 -> X
1022 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1024 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1028 case Instruction::Or:
1029 if (isa<ConstantInt>(I->getOperand(0)) &&
1030 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1031 cast<BinaryOperator>(I)->swapOperands();
1032 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1033 if (CB->getType() == Type::Int1Ty) {
1034 if (CB->isOne()) // X | 1 -> 1
1035 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1037 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1041 case Instruction::Br: {
1042 BranchInst *BI = cast<BranchInst>(I);
1043 if (BI->isUnconditional()) {
1044 // If BI's parent is the only pred of the successor, fold the two blocks
1046 BasicBlock *Pred = BI->getParent();
1047 BasicBlock *Succ = BI->getSuccessor(0);
1048 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1049 if (!SinglePred) continue; // Nothing to do.
1050 assert(SinglePred == Pred && "CFG broken");
1052 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1053 << Succ->getName() << "\n";
1055 // Resolve any single entry PHI nodes in Succ.
1056 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1057 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1059 // Move all of the successor contents from Succ to Pred.
1060 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1062 LPM->deleteSimpleAnalysisValue(BI, L);
1063 BI->eraseFromParent();
1064 RemoveFromWorklist(BI, Worklist);
1066 // If Succ has any successors with PHI nodes, update them to have
1067 // entries coming from Pred instead of Succ.
1068 Succ->replaceAllUsesWith(Pred);
1070 // Remove Succ from the loop tree.
1071 LI->removeBlock(Succ);
1072 LPM->deleteSimpleAnalysisValue(Succ, L);
1073 Succ->eraseFromParent();
1075 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1076 // Conditional branch. Turn it into an unconditional branch, then
1077 // remove dead blocks.
1078 break; // FIXME: Enable.
1080 DOUT << "Folded branch: " << *BI;
1081 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1082 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1083 DeadSucc->removePredecessor(BI->getParent(), true);
1084 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1085 LPM->deleteSimpleAnalysisValue(BI, L);
1086 BI->eraseFromParent();
1087 RemoveFromWorklist(BI, Worklist);
1090 RemoveBlockIfDead(DeadSucc, Worklist, L);