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/ADT/STLExtras.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Compiler.h"
47 #include "llvm/Support/Debug.h"
52 STATISTIC(NumBranches, "Number of branches unswitched");
53 STATISTIC(NumSwitches, "Number of switches unswitched");
54 STATISTIC(NumSelects , "Number of selects unswitched");
55 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
56 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
58 static cl::opt<unsigned>
59 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
60 cl::init(10), cl::Hidden);
63 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
64 LoopInfo *LI; // Loop information
67 // LoopProcessWorklist - Used to check if second loop needs processing
68 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
69 std::vector<Loop*> LoopProcessWorklist;
70 SmallPtrSet<Value *,8> UnswitchedVals;
76 DominanceFrontier *DF;
78 BasicBlock *loopHeader;
79 BasicBlock *loopPreheader;
81 // LoopBlocks contains all of the basic blocks of the loop, including the
82 // preheader of the loop, the body of the loop, and the exit blocks of the
83 // loop, in that order.
84 std::vector<BasicBlock*> LoopBlocks;
85 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
86 std::vector<BasicBlock*> NewBlocks;
89 static char ID; // Pass ID, replacement for typeid
90 explicit LoopUnswitch(bool Os = false) :
91 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
92 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
93 loopPreheader(NULL) {}
95 bool runOnLoop(Loop *L, LPPassManager &LPM);
96 bool processCurrentLoop();
98 /// This transformation requires natural loop information & requires that
99 /// loop preheaders be inserted into the CFG...
101 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
102 AU.addRequiredID(LoopSimplifyID);
103 AU.addPreservedID(LoopSimplifyID);
104 AU.addRequired<LoopInfo>();
105 AU.addPreserved<LoopInfo>();
106 AU.addRequiredID(LCSSAID);
107 AU.addPreservedID(LCSSAID);
108 AU.addPreserved<DominatorTree>();
109 AU.addPreserved<DominanceFrontier>();
114 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
116 void RemoveLoopFromWorklist(Loop *L) {
117 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
118 LoopProcessWorklist.end(), L);
119 if (I != LoopProcessWorklist.end())
120 LoopProcessWorklist.erase(I);
123 void initLoopData() {
124 loopHeader = currentLoop->getHeader();
125 loopPreheader = currentLoop->getLoopPreheader();
128 /// Split all of the edges from inside the loop to their exit blocks.
129 /// Update the appropriate Phi nodes as we do so.
130 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
132 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
133 unsigned getLoopUnswitchCost(Value *LIC);
134 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
135 BasicBlock *ExitBlock);
136 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
138 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
139 Constant *Val, bool isEqual);
141 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
142 BasicBlock *TrueDest,
143 BasicBlock *FalseDest,
144 Instruction *InsertPt);
146 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
147 void RemoveBlockIfDead(BasicBlock *BB,
148 std::vector<Instruction*> &Worklist, Loop *l);
149 void RemoveLoopFromHierarchy(Loop *L);
150 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
151 BasicBlock **LoopExit = 0);
155 char LoopUnswitch::ID = 0;
156 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
158 Pass *llvm::createLoopUnswitchPass(bool Os) {
159 return new LoopUnswitch(Os);
162 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
163 /// invariant in the loop, or has an invariant piece, return the invariant.
164 /// Otherwise, return null.
165 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
166 // Constants should be folded, not unswitched on!
167 if (isa<Constant>(Cond)) return 0;
169 // TODO: Handle: br (VARIANT|INVARIANT).
170 // TODO: Hoist simple expressions out of loops.
171 if (L->isLoopInvariant(Cond)) return Cond;
173 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
174 if (BO->getOpcode() == Instruction::And ||
175 BO->getOpcode() == Instruction::Or) {
176 // If either the left or right side is invariant, we can unswitch on this,
177 // which will cause the branch to go away in one loop and the condition to
178 // simplify in the other one.
179 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
181 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
188 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
189 LI = &getAnalysis<LoopInfo>();
191 DF = getAnalysisToUpdate<DominanceFrontier>();
192 DT = getAnalysisToUpdate<DominatorTree>();
194 Function *F = currentLoop->getHeader()->getParent();
195 bool Changed = false;
197 assert(currentLoop->isLCSSAForm());
199 Changed |= processCurrentLoop();
203 // FIXME: Reconstruct dom info, because it is not preserved properly.
205 DT->runOnFunction(*F);
207 DF->runOnFunction(*F);
212 /// processCurrentLoop - Do actual work and unswitch loop if possible
214 bool LoopUnswitch::processCurrentLoop() {
215 bool Changed = false;
217 // Loop over all of the basic blocks in the loop. If we find an interior
218 // block that is branching on a loop-invariant condition, we can unswitch this
220 for (Loop::block_iterator I = currentLoop->block_begin(),
221 E = currentLoop->block_end();
223 TerminatorInst *TI = (*I)->getTerminator();
224 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
225 // If this isn't branching on an invariant condition, we can't unswitch
227 if (BI->isConditional()) {
228 // See if this, or some part of it, is loop invariant. If so, we can
229 // unswitch on it if we desire.
230 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
231 currentLoop, Changed);
232 if (LoopCond && UnswitchIfProfitable(LoopCond,
233 ConstantInt::getTrue())) {
238 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
239 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
240 currentLoop, Changed);
241 if (LoopCond && SI->getNumCases() > 1) {
242 // Find a value to unswitch on:
243 // FIXME: this should chose the most expensive case!
244 Constant *UnswitchVal = SI->getCaseValue(1);
245 // Do not process same value again and again.
246 if (!UnswitchedVals.insert(UnswitchVal))
249 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
256 // Scan the instructions to check for unswitchable values.
257 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
259 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
260 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
261 currentLoop, Changed);
262 if (LoopCond && UnswitchIfProfitable(LoopCond,
263 ConstantInt::getTrue())) {
272 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
273 /// 1. Exit the loop with no side effects.
274 /// 2. Branch to the latch block with no side-effects.
276 /// If these conditions are true, we return true and set ExitBB to the block we
279 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
281 std::set<BasicBlock*> &Visited) {
282 if (!Visited.insert(BB).second) {
283 // Already visited and Ok, end of recursion.
285 } else if (!L->contains(BB)) {
286 // Otherwise, this is a loop exit, this is fine so long as this is the
288 if (ExitBB != 0) return false;
293 // Otherwise, this is an unvisited intra-loop node. Check all successors.
294 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
295 // Check to see if the successor is a trivial loop exit.
296 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
300 // Okay, everything after this looks good, check to make sure that this block
301 // doesn't include any side effects.
302 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
303 if (I->mayWriteToMemory())
309 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
310 /// leads to an exit from the specified loop, and has no side-effects in the
311 /// process. If so, return the block that is exited to, otherwise return null.
312 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
313 std::set<BasicBlock*> Visited;
314 Visited.insert(L->getHeader()); // Branches to header are ok.
315 BasicBlock *ExitBB = 0;
316 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
321 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
322 /// trivial: that is, that the condition controls whether or not the loop does
323 /// anything at all. If this is a trivial condition, unswitching produces no
324 /// code duplications (equivalently, it produces a simpler loop and a new empty
325 /// loop, which gets deleted).
327 /// If this is a trivial condition, return true, otherwise return false. When
328 /// returning true, this sets Cond and Val to the condition that controls the
329 /// trivial condition: when Cond dynamically equals Val, the loop is known to
330 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
333 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
334 BasicBlock **LoopExit) {
335 BasicBlock *Header = currentLoop->getHeader();
336 TerminatorInst *HeaderTerm = Header->getTerminator();
338 BasicBlock *LoopExitBB = 0;
339 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
340 // If the header block doesn't end with a conditional branch on Cond, we
342 if (!BI->isConditional() || BI->getCondition() != Cond)
345 // Check to see if a successor of the branch is guaranteed to go to the
346 // latch block or exit through a one exit block without having any
347 // side-effects. If so, determine the value of Cond that causes it to do
349 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
350 BI->getSuccessor(0)))) {
351 if (Val) *Val = ConstantInt::getTrue();
352 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
353 BI->getSuccessor(1)))) {
354 if (Val) *Val = ConstantInt::getFalse();
356 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
357 // If this isn't a switch on Cond, we can't handle it.
358 if (SI->getCondition() != Cond) return false;
360 // Check to see if a successor of the switch is guaranteed to go to the
361 // latch block or exit through a one exit block without having any
362 // side-effects. If so, determine the value of Cond that causes it to do
363 // this. Note that we can't trivially unswitch on the default case.
364 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
365 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
366 SI->getSuccessor(i)))) {
367 // Okay, we found a trivial case, remember the value that is trivial.
368 if (Val) *Val = SI->getCaseValue(i);
373 // If we didn't find a single unique LoopExit block, or if the loop exit block
374 // contains phi nodes, this isn't trivial.
375 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
376 return false; // Can't handle this.
378 if (LoopExit) *LoopExit = LoopExitBB;
380 // We already know that nothing uses any scalar values defined inside of this
381 // loop. As such, we just have to check to see if this loop will execute any
382 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
383 // part of the loop that the code *would* execute. We already checked the
384 // tail, check the header now.
385 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
386 if (I->mayWriteToMemory())
391 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
392 /// we choose to unswitch current loop on the specified value.
394 unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
395 // If the condition is trivial, always unswitch. There is no code growth for
397 if (IsTrivialUnswitchCondition(LIC))
400 // FIXME: This is really overly conservative. However, more liberal
401 // estimations have thus far resulted in excessive unswitching, which is bad
402 // both in compile time and in code size. This should be replaced once
403 // someone figures out how a good estimation.
404 return currentLoop->getBlocks().size();
407 // FIXME: this is brain dead. It should take into consideration code
409 for (Loop::block_iterator I = currentLoop->block_begin(),
410 E = currentLoop->block_end();
413 // Do not include empty blocks in the cost calculation. This happen due to
414 // loop canonicalization and will be removed.
415 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
418 // Count basic blocks.
425 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
426 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
427 /// unswitch the loop, reprocess the pieces, then return true.
428 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
431 Function *F = loopHeader->getParent();
433 // Do not unswitch if the function is optimized for size.
434 if (!F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
437 // Check to see if it would be profitable to unswitch current loop.
438 unsigned Cost = getLoopUnswitchCost(LoopCond);
440 // Do not do non-trivial unswitch while optimizing for size.
441 if (Cost && OptimizeForSize)
444 if (Cost > Threshold) {
445 // FIXME: this should estimate growth by the amount of code shared by the
446 // resultant unswitched loops.
448 DOUT << "NOT unswitching loop %"
449 << currentLoop->getHeader()->getName() << ", cost too high: "
450 << currentLoop->getBlocks().size() << "\n";
455 BasicBlock *ExitBlock;
456 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
457 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
459 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
465 // RemapInstruction - Convert the instruction operands from referencing the
466 // current values into those specified by ValueMap.
468 static inline void RemapInstruction(Instruction *I,
469 DenseMap<const Value *, Value*> &ValueMap) {
470 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
471 Value *Op = I->getOperand(op);
472 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
473 if (It != ValueMap.end()) Op = It->second;
474 I->setOperand(op, Op);
478 /// CloneLoop - Recursively clone the specified loop and all of its children,
479 /// mapping the blocks with the specified map.
480 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
481 LoopInfo *LI, LPPassManager *LPM) {
482 Loop *New = new Loop();
484 LPM->insertLoop(New, PL);
486 // Add all of the blocks in L to the new loop.
487 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
489 if (LI->getLoopFor(*I) == L)
490 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
492 // Add all of the subloops to the new loop.
493 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
494 CloneLoop(*I, New, VM, LI, LPM);
499 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
500 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
501 /// code immediately before InsertPt.
502 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
503 BasicBlock *TrueDest,
504 BasicBlock *FalseDest,
505 Instruction *InsertPt) {
506 // Insert a conditional branch on LIC to the two preheaders. The original
507 // code is the true version and the new code is the false version.
508 Value *BranchVal = LIC;
509 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
510 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
511 else if (Val != ConstantInt::getTrue())
512 // We want to enter the new loop when the condition is true.
513 std::swap(TrueDest, FalseDest);
515 // Insert the new branch.
516 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
519 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
520 /// condition in it (a cond branch from its header block to its latch block,
521 /// where the path through the loop that doesn't execute its body has no
522 /// side-effects), unswitch it. This doesn't involve any code duplication, just
523 /// moving the conditional branch outside of the loop and updating loop info.
524 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
526 BasicBlock *ExitBlock) {
527 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
528 << loopHeader->getName() << " [" << L->getBlocks().size()
529 << " blocks] in Function " << L->getHeader()->getParent()->getName()
530 << " on cond: " << *Val << " == " << *Cond << "\n";
532 // First step, split the preheader, so that we know that there is a safe place
533 // to insert the conditional branch. We will change loopPreheader to have a
534 // conditional branch on Cond.
535 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
537 // Now that we have a place to insert the conditional branch, create a place
538 // to branch to: this is the exit block out of the loop that we should
541 // Split this block now, so that the loop maintains its exit block, and so
542 // that the jump from the preheader can execute the contents of the exit block
543 // without actually branching to it (the exit block should be dominated by the
544 // loop header, not the preheader).
545 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
546 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
548 // Okay, now we have a position to branch from and a position to branch to,
549 // insert the new conditional branch.
550 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
551 loopPreheader->getTerminator());
552 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
553 loopPreheader->getTerminator()->eraseFromParent();
555 // We need to reprocess this loop, it could be unswitched again.
558 // Now that we know that the loop is never entered when this condition is a
559 // particular value, rewrite the loop with this info. We know that this will
560 // at least eliminate the old branch.
561 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
565 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
566 /// blocks. Update the appropriate Phi nodes as we do so.
567 void LoopUnswitch::SplitExitEdges(Loop *L,
568 const SmallVector<BasicBlock *, 8> &ExitBlocks)
571 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
572 BasicBlock *ExitBlock = ExitBlocks[i];
573 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
575 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
576 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
577 BasicBlock* StartBlock = Preds[j];
578 BasicBlock* EndBlock;
579 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
580 EndBlock = NewExitBlock;
581 NewExitBlock = EndBlock->getSinglePredecessor();;
583 EndBlock = ExitBlock;
586 std::set<PHINode*> InsertedPHIs;
587 PHINode* OldLCSSA = 0;
588 for (BasicBlock::iterator I = EndBlock->begin();
589 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
590 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
591 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
592 OldLCSSA->getName() + ".us-lcssa",
593 NewExitBlock->getTerminator());
594 NewLCSSA->addIncoming(OldValue, StartBlock);
595 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
597 InsertedPHIs.insert(NewLCSSA);
600 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
601 for (BasicBlock::iterator I = NewExitBlock->begin();
602 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
604 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
605 OldLCSSA->getName() + ".us-lcssa",
607 OldLCSSA->replaceAllUsesWith(NewLCSSA);
608 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
616 /// UnswitchNontrivialCondition - We determined that the loop is profitable
617 /// to unswitch when LIC equal Val. Split it into loop versions and test the
618 /// condition outside of either loop. Return the loops created as Out1/Out2.
619 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
621 Function *F = loopHeader->getParent();
622 DOUT << "loop-unswitch: Unswitching loop %"
623 << loopHeader->getName() << " [" << L->getBlocks().size()
624 << " blocks] in Function " << F->getName()
625 << " when '" << *Val << "' == " << *LIC << "\n";
630 // First step, split the preheader and exit blocks, and add these blocks to
631 // the LoopBlocks list.
632 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
633 LoopBlocks.push_back(NewPreheader);
635 // We want the loop to come after the preheader, but before the exit blocks.
636 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
638 SmallVector<BasicBlock*, 8> ExitBlocks;
639 L->getUniqueExitBlocks(ExitBlocks);
641 // Split all of the edges from inside the loop to their exit blocks. Update
642 // the appropriate Phi nodes as we do so.
643 SplitExitEdges(L, ExitBlocks);
645 // The exit blocks may have been changed due to edge splitting, recompute.
647 L->getUniqueExitBlocks(ExitBlocks);
649 // Add exit blocks to the loop blocks.
650 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
652 // Next step, clone all of the basic blocks that make up the loop (including
653 // the loop preheader and exit blocks), keeping track of the mapping between
654 // the instructions and blocks.
655 NewBlocks.reserve(LoopBlocks.size());
656 DenseMap<const Value*, Value*> ValueMap;
657 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
658 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
659 NewBlocks.push_back(New);
660 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
661 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
664 // Splice the newly inserted blocks into the function right before the
665 // original preheader.
666 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
667 NewBlocks[0], F->end());
669 // Now we create the new Loop object for the versioned loop.
670 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
671 Loop *ParentLoop = L->getParentLoop();
673 // Make sure to add the cloned preheader and exit blocks to the parent loop
675 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
678 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
679 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
680 // The new exit block should be in the same loop as the old one.
681 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
682 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
684 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
685 "Exit block should have been split to have one successor!");
686 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
688 // If the successor of the exit block had PHI nodes, add an entry for
691 for (BasicBlock::iterator I = ExitSucc->begin();
692 (PN = dyn_cast<PHINode>(I)); ++I) {
693 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
694 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
695 if (It != ValueMap.end()) V = It->second;
696 PN->addIncoming(V, NewExit);
700 // Rewrite the code to refer to itself.
701 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
702 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
703 E = NewBlocks[i]->end(); I != E; ++I)
704 RemapInstruction(I, ValueMap);
706 // Rewrite the original preheader to select between versions of the loop.
707 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
708 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
709 "Preheader splitting did not work correctly!");
711 // Emit the new branch that selects between the two versions of this loop.
712 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
713 LPM->deleteSimpleAnalysisValue(OldBR, L);
714 OldBR->eraseFromParent();
716 LoopProcessWorklist.push_back(NewLoop);
719 // Now we rewrite the original code to know that the condition is true and the
720 // new code to know that the condition is false.
721 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
723 // It's possible that simplifying one loop could cause the other to be
724 // deleted. If so, don't simplify it.
725 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
726 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
730 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
732 static void RemoveFromWorklist(Instruction *I,
733 std::vector<Instruction*> &Worklist) {
734 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
736 while (WI != Worklist.end()) {
737 unsigned Offset = WI-Worklist.begin();
739 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
743 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
744 /// program, replacing all uses with V and update the worklist.
745 static void ReplaceUsesOfWith(Instruction *I, Value *V,
746 std::vector<Instruction*> &Worklist,
747 Loop *L, LPPassManager *LPM) {
748 DOUT << "Replace with '" << *V << "': " << *I;
750 // Add uses to the worklist, which may be dead now.
751 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
752 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
753 Worklist.push_back(Use);
755 // Add users to the worklist which may be simplified now.
756 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
758 Worklist.push_back(cast<Instruction>(*UI));
759 LPM->deleteSimpleAnalysisValue(I, L);
760 RemoveFromWorklist(I, Worklist);
761 I->replaceAllUsesWith(V);
762 I->eraseFromParent();
766 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
767 /// information, and remove any dead successors it has.
769 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
770 std::vector<Instruction*> &Worklist,
772 if (pred_begin(BB) != pred_end(BB)) {
773 // This block isn't dead, since an edge to BB was just removed, see if there
774 // are any easy simplifications we can do now.
775 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
776 // If it has one pred, fold phi nodes in BB.
777 while (isa<PHINode>(BB->begin()))
778 ReplaceUsesOfWith(BB->begin(),
779 cast<PHINode>(BB->begin())->getIncomingValue(0),
782 // If this is the header of a loop and the only pred is the latch, we now
783 // have an unreachable loop.
784 if (Loop *L = LI->getLoopFor(BB))
785 if (loopHeader == BB && L->contains(Pred)) {
786 // Remove the branch from the latch to the header block, this makes
787 // the header dead, which will make the latch dead (because the header
788 // dominates the latch).
789 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
790 Pred->getTerminator()->eraseFromParent();
791 new UnreachableInst(Pred);
793 // The loop is now broken, remove it from LI.
794 RemoveLoopFromHierarchy(L);
796 // Reprocess the header, which now IS dead.
797 RemoveBlockIfDead(BB, Worklist, L);
801 // If pred ends in a uncond branch, add uncond branch to worklist so that
802 // the two blocks will get merged.
803 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
804 if (BI->isUnconditional())
805 Worklist.push_back(BI);
810 DOUT << "Nuking dead block: " << *BB;
812 // Remove the instructions in the basic block from the worklist.
813 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
814 RemoveFromWorklist(I, Worklist);
816 // Anything that uses the instructions in this basic block should have their
817 // uses replaced with undefs.
819 I->replaceAllUsesWith(UndefValue::get(I->getType()));
822 // If this is the edge to the header block for a loop, remove the loop and
823 // promote all subloops.
824 if (Loop *BBLoop = LI->getLoopFor(BB)) {
825 if (BBLoop->getLoopLatch() == BB)
826 RemoveLoopFromHierarchy(BBLoop);
829 // Remove the block from the loop info, which removes it from any loops it
834 // Remove phi node entries in successors for this block.
835 TerminatorInst *TI = BB->getTerminator();
836 SmallVector<BasicBlock*, 4> Succs;
837 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
838 Succs.push_back(TI->getSuccessor(i));
839 TI->getSuccessor(i)->removePredecessor(BB);
842 // Unique the successors, remove anything with multiple uses.
843 array_pod_sort(Succs.begin(), Succs.end());
844 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
846 // Remove the basic block, including all of the instructions contained in it.
847 LPM->deleteSimpleAnalysisValue(BB, L);
848 BB->eraseFromParent();
849 // Remove successor blocks here that are not dead, so that we know we only
850 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
851 // then getting removed before we revisit them, which is badness.
853 for (unsigned i = 0; i != Succs.size(); ++i)
854 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
855 // One exception is loop headers. If this block was the preheader for a
856 // loop, then we DO want to visit the loop so the loop gets deleted.
857 // We know that if the successor is a loop header, that this loop had to
858 // be the preheader: the case where this was the latch block was handled
859 // above and headers can only have two predecessors.
860 if (!LI->isLoopHeader(Succs[i])) {
861 Succs.erase(Succs.begin()+i);
866 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
867 RemoveBlockIfDead(Succs[i], Worklist, L);
870 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
871 /// become unwrapped, either because the backedge was deleted, or because the
872 /// edge into the header was removed. If the edge into the header from the
873 /// latch block was removed, the loop is unwrapped but subloops are still alive,
874 /// so they just reparent loops. If the loops are actually dead, they will be
876 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
877 LPM->deleteLoopFromQueue(L);
878 RemoveLoopFromWorklist(L);
881 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
882 // the value specified by Val in the specified loop, or we know it does NOT have
883 // that value. Rewrite any uses of LIC or of properties correlated to it.
884 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
887 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
889 // FIXME: Support correlated properties, like:
896 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
897 // selects, switches.
898 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
899 std::vector<Instruction*> Worklist;
901 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
902 // in the loop with the appropriate one directly.
903 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
908 Replacement = ConstantInt::get(Type::Int1Ty,
909 !cast<ConstantInt>(Val)->getZExtValue());
911 for (unsigned i = 0, e = Users.size(); i != e; ++i)
912 if (Instruction *U = cast<Instruction>(Users[i])) {
913 if (!L->contains(U->getParent()))
915 U->replaceUsesOfWith(LIC, Replacement);
916 Worklist.push_back(U);
919 // Otherwise, we don't know the precise value of LIC, but we do know that it
920 // is certainly NOT "Val". As such, simplify any uses in the loop that we
921 // can. This case occurs when we unswitch switch statements.
922 for (unsigned i = 0, e = Users.size(); i != e; ++i)
923 if (Instruction *U = cast<Instruction>(Users[i])) {
924 if (!L->contains(U->getParent()))
927 Worklist.push_back(U);
929 // If we know that LIC is not Val, use this info to simplify code.
930 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
931 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
932 if (SI->getCaseValue(i) == Val) {
933 // Found a dead case value. Don't remove PHI nodes in the
934 // successor if they become single-entry, those PHI nodes may
935 // be in the Users list.
937 // FIXME: This is a hack. We need to keep the successor around
938 // and hooked up so as to preserve the loop structure, because
939 // trying to update it is complicated. So instead we preserve the
940 // loop structure and put the block on an dead code path.
942 BasicBlock *SISucc = SI->getSuccessor(i);
943 BasicBlock* Old = SI->getParent();
944 BasicBlock* Split = SplitBlock(Old, SI, this);
946 Instruction* OldTerm = Old->getTerminator();
947 BranchInst::Create(Split, SISucc,
948 ConstantInt::getTrue(), OldTerm);
950 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
951 Old->getTerminator()->eraseFromParent();
954 for (BasicBlock::iterator II = SISucc->begin();
955 (PN = dyn_cast<PHINode>(II)); ++II) {
956 Value *InVal = PN->removeIncomingValue(Split, false);
957 PN->addIncoming(InVal, Old);
966 // TODO: We could do other simplifications, for example, turning
967 // LIC == Val -> false.
971 SimplifyCode(Worklist, L);
974 /// SimplifyCode - Okay, now that we have simplified some instructions in the
975 /// loop, walk over it and constant prop, dce, and fold control flow where
976 /// possible. Note that this is effectively a very simple loop-structure-aware
977 /// optimizer. During processing of this loop, L could very well be deleted, so
978 /// it must not be used.
980 /// FIXME: When the loop optimizer is more mature, separate this out to a new
983 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
984 while (!Worklist.empty()) {
985 Instruction *I = Worklist.back();
988 // Simple constant folding.
989 if (Constant *C = ConstantFoldInstruction(I)) {
990 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
995 if (isInstructionTriviallyDead(I)) {
996 DOUT << "Remove dead instruction '" << *I;
998 // Add uses to the worklist, which may be dead now.
999 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1000 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1001 Worklist.push_back(Use);
1002 LPM->deleteSimpleAnalysisValue(I, L);
1003 RemoveFromWorklist(I, Worklist);
1004 I->eraseFromParent();
1009 // Special case hacks that appear commonly in unswitched code.
1010 switch (I->getOpcode()) {
1011 case Instruction::Select:
1012 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1013 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1018 case Instruction::And:
1019 if (isa<ConstantInt>(I->getOperand(0)) &&
1020 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1021 cast<BinaryOperator>(I)->swapOperands();
1022 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1023 if (CB->getType() == Type::Int1Ty) {
1024 if (CB->isOne()) // X & 1 -> X
1025 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1027 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1031 case Instruction::Or:
1032 if (isa<ConstantInt>(I->getOperand(0)) &&
1033 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1034 cast<BinaryOperator>(I)->swapOperands();
1035 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1036 if (CB->getType() == Type::Int1Ty) {
1037 if (CB->isOne()) // X | 1 -> 1
1038 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1040 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1044 case Instruction::Br: {
1045 BranchInst *BI = cast<BranchInst>(I);
1046 if (BI->isUnconditional()) {
1047 // If BI's parent is the only pred of the successor, fold the two blocks
1049 BasicBlock *Pred = BI->getParent();
1050 BasicBlock *Succ = BI->getSuccessor(0);
1051 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1052 if (!SinglePred) continue; // Nothing to do.
1053 assert(SinglePred == Pred && "CFG broken");
1055 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1056 << Succ->getName() << "\n";
1058 // Resolve any single entry PHI nodes in Succ.
1059 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1060 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1062 // Move all of the successor contents from Succ to Pred.
1063 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1065 LPM->deleteSimpleAnalysisValue(BI, L);
1066 BI->eraseFromParent();
1067 RemoveFromWorklist(BI, Worklist);
1069 // If Succ has any successors with PHI nodes, update them to have
1070 // entries coming from Pred instead of Succ.
1071 Succ->replaceAllUsesWith(Pred);
1073 // Remove Succ from the loop tree.
1074 LI->removeBlock(Succ);
1075 LPM->deleteSimpleAnalysisValue(Succ, L);
1076 Succ->eraseFromParent();
1078 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1079 // Conditional branch. Turn it into an unconditional branch, then
1080 // remove dead blocks.
1081 break; // FIXME: Enable.
1083 DOUT << "Folded branch: " << *BI;
1084 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1085 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1086 DeadSucc->removePredecessor(BI->getParent(), true);
1087 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1088 LPM->deleteSimpleAnalysisValue(BI, L);
1089 BI->eraseFromParent();
1090 RemoveFromWorklist(BI, Worklist);
1093 RemoveBlockIfDead(DeadSucc, Worklist, L);