1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/PostOrderIterator.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");
60 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
61 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;
75 static char ID; // Pass ID, replacement for typeid
76 explicit LoopUnswitch(bool Os = false) :
77 LoopPass((intptr_t)&ID), OptimizeForSize(Os), redoLoop(false) {}
79 bool runOnLoop(Loop *L, LPPassManager &LPM);
80 bool processLoop(Loop *L);
82 /// This transformation requires natural loop information & requires that
83 /// loop preheaders be inserted into the CFG...
85 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
86 AU.addRequiredID(LoopSimplifyID);
87 AU.addPreservedID(LoopSimplifyID);
88 AU.addRequired<LoopInfo>();
89 AU.addPreserved<LoopInfo>();
90 AU.addRequiredID(LCSSAID);
91 AU.addPreservedID(LCSSAID);
92 AU.addPreserved<DominatorTree>();
93 AU.addPreserved<DominanceFrontier>();
98 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
100 void RemoveLoopFromWorklist(Loop *L) {
101 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
102 LoopProcessWorklist.end(), L);
103 if (I != LoopProcessWorklist.end())
104 LoopProcessWorklist.erase(I);
107 /// Split all of the edges from inside the loop to their exit blocks. Update
108 /// the appropriate Phi nodes as we do so.
109 void SplitExitEdges(const SmallVector<BasicBlock *, 8> &ExitBlocks,
110 SmallVector<BasicBlock *, 8> &MiddleBlocks);
112 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L);
113 unsigned getLoopUnswitchCost(Loop *L, Value *LIC);
114 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
115 BasicBlock *ExitBlock);
116 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
118 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
119 Constant *Val, bool isEqual);
121 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
122 BasicBlock *TrueDest,
123 BasicBlock *FalseDest,
124 Instruction *InsertPt);
126 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
127 void RemoveBlockIfDead(BasicBlock *BB,
128 std::vector<Instruction*> &Worklist, Loop *l);
129 void RemoveLoopFromHierarchy(Loop *L);
131 char LoopUnswitch::ID = 0;
132 RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
135 LoopPass *llvm::createLoopUnswitchPass(bool Os) {
136 return new LoopUnswitch(Os);
139 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
140 /// invariant in the loop, or has an invariant piece, return the invariant.
141 /// Otherwise, return null.
142 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
143 // Constants should be folded, not unswitched on!
144 if (isa<Constant>(Cond)) return false;
146 // TODO: Handle: br (VARIANT|INVARIANT).
147 // TODO: Hoist simple expressions out of loops.
148 if (L->isLoopInvariant(Cond)) return Cond;
150 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
151 if (BO->getOpcode() == Instruction::And ||
152 BO->getOpcode() == Instruction::Or) {
153 // If either the left or right side is invariant, we can unswitch on this,
154 // which will cause the branch to go away in one loop and the condition to
155 // simplify in the other one.
156 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
158 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
165 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
166 LI = &getAnalysis<LoopInfo>();
168 bool Changed = false;
172 Changed |= processLoop(L);
178 /// processLoop - Do actual work and unswitch loop if possible and profitable.
179 bool LoopUnswitch::processLoop(Loop *L) {
180 assert(L->isLCSSAForm());
181 bool Changed = false;
183 // Loop over all of the basic blocks in the loop. If we find an interior
184 // block that is branching on a loop-invariant condition, we can unswitch this
186 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
188 TerminatorInst *TI = (*I)->getTerminator();
189 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
190 // If this isn't branching on an invariant condition, we can't unswitch
192 if (BI->isConditional()) {
193 // See if this, or some part of it, is loop invariant. If so, we can
194 // unswitch on it if we desire.
195 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed);
196 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
202 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
203 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
204 if (LoopCond && SI->getNumCases() > 1) {
205 // Find a value to unswitch on:
206 // FIXME: this should chose the most expensive case!
207 Constant *UnswitchVal = SI->getCaseValue(1);
208 // Do not process same value again and again.
209 if (!UnswitchedVals.insert(UnswitchVal))
212 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) {
219 // Scan the instructions to check for unswitchable values.
220 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
222 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
223 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
224 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
232 assert(L->isLCSSAForm());
237 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
238 /// 1. Exit the loop with no side effects.
239 /// 2. Branch to the latch block with no side-effects.
241 /// If these conditions are true, we return true and set ExitBB to the block we
244 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
246 std::set<BasicBlock*> &Visited) {
247 if (!Visited.insert(BB).second) {
248 // Already visited and Ok, end of recursion.
250 } else if (!L->contains(BB)) {
251 // Otherwise, this is a loop exit, this is fine so long as this is the
253 if (ExitBB != 0) return false;
258 // Otherwise, this is an unvisited intra-loop node. Check all successors.
259 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
260 // Check to see if the successor is a trivial loop exit.
261 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
265 // Okay, everything after this looks good, check to make sure that this block
266 // doesn't include any side effects.
267 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
268 if (I->mayWriteToMemory())
274 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
275 /// leads to an exit from the specified loop, and has no side-effects in the
276 /// process. If so, return the block that is exited to, otherwise return null.
277 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
278 std::set<BasicBlock*> Visited;
279 Visited.insert(L->getHeader()); // Branches to header are ok.
280 BasicBlock *ExitBB = 0;
281 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
286 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
287 /// trivial: that is, that the condition controls whether or not the loop does
288 /// anything at all. If this is a trivial condition, unswitching produces no
289 /// code duplications (equivalently, it produces a simpler loop and a new empty
290 /// loop, which gets deleted).
292 /// If this is a trivial condition, return true, otherwise return false. When
293 /// returning true, this sets Cond and Val to the condition that controls the
294 /// trivial condition: when Cond dynamically equals Val, the loop is known to
295 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
298 static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0,
299 BasicBlock **LoopExit = 0) {
300 BasicBlock *Header = L->getHeader();
301 TerminatorInst *HeaderTerm = Header->getTerminator();
303 BasicBlock *LoopExitBB = 0;
304 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
305 // If the header block doesn't end with a conditional branch on Cond, we
307 if (!BI->isConditional() || BI->getCondition() != Cond)
310 // Check to see if a successor of the branch is guaranteed to go to the
311 // latch block or exit through a one exit block without having any
312 // side-effects. If so, determine the value of Cond that causes it to do
314 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) {
315 if (Val) *Val = ConstantInt::getTrue();
316 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) {
317 if (Val) *Val = ConstantInt::getFalse();
319 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
320 // If this isn't a switch on Cond, we can't handle it.
321 if (SI->getCondition() != Cond) return false;
323 // Check to see if a successor of the switch is guaranteed to go to the
324 // latch block or exit through a one exit block without having any
325 // side-effects. If so, determine the value of Cond that causes it to do
326 // this. Note that we can't trivially unswitch on the default case.
327 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
328 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) {
329 // Okay, we found a trivial case, remember the value that is trivial.
330 if (Val) *Val = SI->getCaseValue(i);
335 // If we didn't find a single unique LoopExit block, or if the loop exit block
336 // contains phi nodes, this isn't trivial.
337 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
338 return false; // Can't handle this.
340 if (LoopExit) *LoopExit = LoopExitBB;
342 // We already know that nothing uses any scalar values defined inside of this
343 // loop. As such, we just have to check to see if this loop will execute any
344 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
345 // part of the loop that the code *would* execute. We already checked the
346 // tail, check the header now.
347 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
348 if (I->mayWriteToMemory())
353 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
354 /// we choose to unswitch the specified loop on the specified value.
356 unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) {
357 // If the condition is trivial, always unswitch. There is no code growth for
359 if (IsTrivialUnswitchCondition(L, LIC))
362 // FIXME: This is really overly conservative. However, more liberal
363 // estimations have thus far resulted in excessive unswitching, which is bad
364 // both in compile time and in code size. This should be replaced once
365 // someone figures out how a good estimation.
366 return L->getBlocks().size();
369 // FIXME: this is brain dead. It should take into consideration code
371 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
374 // Do not include empty blocks in the cost calculation. This happen due to
375 // loop canonicalization and will be removed.
376 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
379 // Count basic blocks.
386 /// UnswitchIfProfitable - We have found that we can unswitch L when
387 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
388 /// unswitch the loop, reprocess the pieces, then return true.
389 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){
390 // Check to see if it would be profitable to unswitch this loop.
391 unsigned Cost = getLoopUnswitchCost(L, LoopCond);
393 // Do not do non-trivial unswitch while optimizing for size.
394 if (Cost && OptimizeForSize)
397 if (Cost > Threshold) {
398 // FIXME: this should estimate growth by the amount of code shared by the
399 // resultant unswitched loops.
401 DOUT << "NOT unswitching loop %"
402 << L->getHeader()->getName() << ", cost too high: "
403 << L->getBlocks().size() << "\n";
407 // If this is a trivial condition to unswitch (which results in no code
408 // duplication), do it now.
410 BasicBlock *ExitBlock;
411 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) {
412 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock);
414 UnswitchNontrivialCondition(LoopCond, Val, L);
420 // RemapInstruction - Convert the instruction operands from referencing the
421 // current values into those specified by ValueMap.
423 static inline void RemapInstruction(Instruction *I,
424 DenseMap<const Value *, Value*> &ValueMap) {
425 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
426 Value *Op = I->getOperand(op);
427 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
428 if (It != ValueMap.end()) Op = It->second;
429 I->setOperand(op, Op);
433 // CloneDomInfo - NewBB is cloned from Orig basic block. Now clone Dominator
436 // If Orig block's immediate dominator is mapped in VM then use corresponding
437 // immediate dominator from the map. Otherwise Orig block's dominator is also
438 // NewBB's dominator.
440 // OrigPreheader is loop pre-header before this pass started
441 // updating CFG. NewPrehader is loops new pre-header. However, after CFG
442 // manipulation, loop L may not exist. So rely on input parameter NewPreheader.
443 void CloneDomInfo(BasicBlock *NewBB, BasicBlock *Orig,
444 BasicBlock *NewPreheader, BasicBlock *OrigPreheader,
445 BasicBlock *OrigHeader,
446 DominatorTree *DT, DominanceFrontier *DF,
447 DenseMap<const Value*, Value*> &VM) {
449 // If NewBB alreay has found its place in domiantor tree then no need to do
451 if (DT->getNode(NewBB))
454 // If Orig does not have any immediate domiantor then its clone, NewBB, does
455 // not need any immediate dominator.
456 DomTreeNode *OrigNode = DT->getNode(Orig);
459 DomTreeNode *OrigIDomNode = OrigNode->getIDom();
463 BasicBlock *OrigIDom = NULL;
465 // If Orig is original loop header then its immediate dominator is
467 if (Orig == OrigHeader)
468 OrigIDom = NewPreheader;
470 // If Orig is new pre-header then its immediate dominator is
471 // original pre-header.
472 else if (Orig == NewPreheader)
473 OrigIDom = OrigPreheader;
475 // Other as DT to find Orig's immediate dominator.
477 OrigIDom = OrigIDomNode->getBlock();
479 // Initially use Orig's immediate dominator as NewBB's immediate dominator.
480 BasicBlock *NewIDom = OrigIDom;
481 DenseMap<const Value*, Value*>::iterator I = VM.find(OrigIDom);
483 NewIDom = cast<BasicBlock>(I->second);
485 // If NewIDom does not have corresponding dominatore tree node then
487 if (!DT->getNode(NewIDom))
488 CloneDomInfo(NewIDom, OrigIDom, NewPreheader, OrigPreheader,
489 OrigHeader, DT, DF, VM);
492 DT->addNewBlock(NewBB, NewIDom);
494 // Copy cloned dominance frontiner set
495 DominanceFrontier::DomSetType NewDFSet;
497 DominanceFrontier::iterator DFI = DF->find(Orig);
498 if ( DFI != DF->end()) {
499 DominanceFrontier::DomSetType S = DFI->second;
500 for (DominanceFrontier::DomSetType::iterator I = S.begin(), E = S.end();
503 DenseMap<const Value*, Value*>::iterator IDM = VM.find(BB);
505 NewDFSet.insert(cast<BasicBlock>(IDM->second));
510 DF->addBasicBlock(NewBB, NewDFSet);
514 /// CloneLoop - Recursively clone the specified loop and all of its children,
515 /// mapping the blocks with the specified map.
516 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
517 LoopInfo *LI, LPPassManager *LPM) {
518 Loop *New = new Loop();
520 LPM->insertLoop(New, PL);
522 // Add all of the blocks in L to the new loop.
523 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
525 if (LI->getLoopFor(*I) == L)
526 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
528 // Add all of the subloops to the new loop.
529 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
530 CloneLoop(*I, New, VM, LI, LPM);
535 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
536 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
537 /// code immediately before InsertPt.
538 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
539 BasicBlock *TrueDest,
540 BasicBlock *FalseDest,
541 Instruction *InsertPt) {
542 // Insert a conditional branch on LIC to the two preheaders. The original
543 // code is the true version and the new code is the false version.
544 Value *BranchVal = LIC;
545 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
546 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
547 else if (Val != ConstantInt::getTrue())
548 // We want to enter the new loop when the condition is true.
549 std::swap(TrueDest, FalseDest);
551 // Insert the new branch.
552 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt);
557 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
558 /// condition in it (a cond branch from its header block to its latch block,
559 /// where the path through the loop that doesn't execute its body has no
560 /// side-effects), unswitch it. This doesn't involve any code duplication, just
561 /// moving the conditional branch outside of the loop and updating loop info.
562 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
564 BasicBlock *ExitBlock) {
565 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
566 << L->getHeader()->getName() << " [" << L->getBlocks().size()
567 << " blocks] in Function " << L->getHeader()->getParent()->getName()
568 << " on cond: " << *Val << " == " << *Cond << "\n";
570 // First step, split the preheader, so that we know that there is a safe place
571 // to insert the conditional branch. We will change 'OrigPH' to have a
572 // conditional branch on Cond.
573 BasicBlock *OrigPH = L->getLoopPreheader();
574 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader(), this);
576 // Now that we have a place to insert the conditional branch, create a place
577 // to branch to: this is the exit block out of the loop that we should
580 // Split this block now, so that the loop maintains its exit block, and so
581 // that the jump from the preheader can execute the contents of the exit block
582 // without actually branching to it (the exit block should be dominated by the
583 // loop header, not the preheader).
584 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
585 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
587 // Okay, now we have a position to branch from and a position to branch to,
588 // insert the new conditional branch.
589 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
590 OrigPH->getTerminator());
591 LPM->deleteSimpleAnalysisValue(OrigPH->getTerminator(), L);
592 OrigPH->getTerminator()->eraseFromParent();
594 // We need to reprocess this loop, it could be unswitched again.
597 // Now that we know that the loop is never entered when this condition is a
598 // particular value, rewrite the loop with this info. We know that this will
599 // at least eliminate the old branch.
600 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
605 /// Split all of the edges from inside the loop to their exit blocks. Update
606 /// the appropriate Phi nodes as we do so.
607 void LoopUnswitch::SplitExitEdges(const SmallVector<BasicBlock *, 8> &ExitBlocks,
608 SmallVector<BasicBlock *, 8> &MiddleBlocks) {
609 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
610 BasicBlock *ExitBlock = ExitBlocks[i];
611 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
613 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
614 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock, this);
615 MiddleBlocks.push_back(MiddleBlock);
616 BasicBlock* StartBlock = Preds[j];
617 BasicBlock* EndBlock;
618 if (MiddleBlock->getSinglePredecessor() == ExitBlock) {
619 EndBlock = MiddleBlock;
620 MiddleBlock = EndBlock->getSinglePredecessor();;
622 EndBlock = ExitBlock;
625 std::set<PHINode*> InsertedPHIs;
626 PHINode* OldLCSSA = 0;
627 for (BasicBlock::iterator I = EndBlock->begin();
628 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
629 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock);
630 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(),
631 OldLCSSA->getName() + ".us-lcssa",
632 MiddleBlock->getTerminator());
633 NewLCSSA->addIncoming(OldValue, StartBlock);
634 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock),
636 InsertedPHIs.insert(NewLCSSA);
639 BasicBlock::iterator InsertPt = EndBlock->begin();
640 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt;
641 for (BasicBlock::iterator I = MiddleBlock->begin();
642 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
644 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(),
645 OldLCSSA->getName() + ".us-lcssa",
647 OldLCSSA->replaceAllUsesWith(NewLCSSA);
648 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock);
654 /// UnswitchNontrivialCondition - We determined that the loop is profitable
655 /// to unswitch when LIC equal Val. Split it into loop versions and test the
656 /// condition outside of either loop. Return the loops created as Out1/Out2.
657 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
659 Function *F = L->getHeader()->getParent();
660 DOUT << "loop-unswitch: Unswitching loop %"
661 << L->getHeader()->getName() << " [" << L->getBlocks().size()
662 << " blocks] in Function " << F->getName()
663 << " when '" << *Val << "' == " << *LIC << "\n";
665 // LoopBlocks contains all of the basic blocks of the loop, including the
666 // preheader of the loop, the body of the loop, and the exit blocks of the
667 // loop, in that order.
668 std::vector<BasicBlock*> LoopBlocks;
670 // First step, split the preheader and exit blocks, and add these blocks to
671 // the LoopBlocks list.
672 BasicBlock *OrigHeader = L->getHeader();
673 BasicBlock *OrigPreheader = L->getLoopPreheader();
674 BasicBlock *NewPreheader = SplitEdge(OrigPreheader, L->getHeader(), this);
675 LoopBlocks.push_back(NewPreheader);
677 // We want the loop to come after the preheader, but before the exit blocks.
678 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
680 SmallVector<BasicBlock*, 8> ExitBlocks;
681 L->getUniqueExitBlocks(ExitBlocks);
683 // Split all of the edges from inside the loop to their exit blocks. Update
684 // the appropriate Phi nodes as we do so.
685 SmallVector<BasicBlock *,8> MiddleBlocks;
686 SplitExitEdges(ExitBlocks, MiddleBlocks);
688 // The exit blocks may have been changed due to edge splitting, recompute.
690 L->getUniqueExitBlocks(ExitBlocks);
692 // Add exit blocks to the loop blocks.
693 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
695 DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>();
696 DominatorTree *DT = getAnalysisToUpdate<DominatorTree>();
698 // Next step, clone all of the basic blocks that make up the loop (including
699 // the loop preheader and exit blocks), keeping track of the mapping between
700 // the instructions and blocks.
701 std::vector<BasicBlock*> NewBlocks;
702 NewBlocks.reserve(LoopBlocks.size());
703 DenseMap<const Value*, Value*> ValueMap;
704 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
705 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
706 NewBlocks.push_back(New);
707 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
708 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
711 // OutSiders are basic block that are dominated by original header and
712 // at the same time they are not part of loop.
713 SmallPtrSet<BasicBlock *, 8> OutSiders;
715 DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader);
716 for(std::vector<DomTreeNode*>::iterator DI = OrigHeaderNode->begin(),
717 DE = OrigHeaderNode->end(); DI != DE; ++DI) {
718 BasicBlock *B = (*DI)->getBlock();
720 DenseMap<const Value*, Value*>::iterator VI = ValueMap.find(B);
721 if (VI == ValueMap.end())
726 // Splice the newly inserted blocks into the function right before the
727 // original preheader.
728 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
729 NewBlocks[0], F->end());
731 // Now we create the new Loop object for the versioned loop.
732 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
733 Loop *ParentLoop = L->getParentLoop();
735 // Make sure to add the cloned preheader and exit blocks to the parent loop
737 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
740 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
741 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
742 // The new exit block should be in the same loop as the old one.
743 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
744 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
746 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
747 "Exit block should have been split to have one successor!");
748 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
750 // If the successor of the exit block had PHI nodes, add an entry for
753 for (BasicBlock::iterator I = ExitSucc->begin();
754 (PN = dyn_cast<PHINode>(I)); ++I) {
755 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
756 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
757 if (It != ValueMap.end()) V = It->second;
758 PN->addIncoming(V, NewExit);
762 // Rewrite the code to refer to itself.
763 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
764 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
765 E = NewBlocks[i]->end(); I != E; ++I)
766 RemapInstruction(I, ValueMap);
768 // Rewrite the original preheader to select between versions of the loop.
769 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator());
770 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
771 "Preheader splitting did not work correctly!");
773 // Emit the new branch that selects between the two versions of this loop.
774 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
775 LPM->deleteSimpleAnalysisValue(OldBR, L);
776 OldBR->eraseFromParent();
778 // Update dominator info
781 // Clone dominator info for all cloned basic block.
782 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
783 BasicBlock *LBB = LoopBlocks[i];
784 BasicBlock *NBB = NewBlocks[i];
785 CloneDomInfo(NBB, LBB, NewPreheader, OrigPreheader,
786 OrigHeader, DT, DF, ValueMap);
788 // Remove any OutSiders from LBB and NBB's dominance frontier.
789 DominanceFrontier::iterator LBBI = DF->find(LBB);
790 if (LBBI != DF->end()) {
791 DominanceFrontier::DomSetType &LBSet = LBBI->second;
792 for (DominanceFrontier::DomSetType::iterator LI = LBSet.begin(),
793 LE = LBSet.end(); LI != LE; /* NULL */) {
794 BasicBlock *B = *LI++;
795 if (OutSiders.count(B))
796 DF->removeFromFrontier(LBBI, B);
800 // Remove any OutSiders from LBB and NBB's dominance frontier.
801 DominanceFrontier::iterator NBBI = DF->find(NBB);
802 if (NBBI != DF->end()) {
803 DominanceFrontier::DomSetType NBSet = NBBI->second;
804 for (DominanceFrontier::DomSetType::iterator NI = NBSet.begin(),
805 NE = NBSet.end(); NI != NE; /* NULL */) {
806 BasicBlock *B = *NI++;
807 if (OutSiders.count(B))
808 DF->removeFromFrontier(NBBI, B);
813 // MiddleBlocks are dominated by original pre header. SplitEdge updated
814 // MiddleBlocks' dominance frontier appropriately.
815 for (unsigned i = 0, e = MiddleBlocks.size(); i != e; ++i) {
816 BasicBlock *MBB = MiddleBlocks[i];
817 if (!MBB->getSinglePredecessor())
818 DT->changeImmediateDominator(MBB, OrigPreheader);
821 // All Outsiders are now dominated by original pre header.
822 for (SmallPtrSet<BasicBlock *, 8>::iterator OI = OutSiders.begin(),
823 OE = OutSiders.end(); OI != OE; ++OI) {
824 BasicBlock *OB = *OI;
825 DT->changeImmediateDominator(OB, OrigPreheader);
828 // New loop headers are dominated by original preheader
829 DT->changeImmediateDominator(NewBlocks[0], OrigPreheader);
830 DT->changeImmediateDominator(LoopBlocks[0], OrigPreheader);
833 LoopProcessWorklist.push_back(NewLoop);
836 // Now we rewrite the original code to know that the condition is true and the
837 // new code to know that the condition is false.
838 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
840 // It's possible that simplifying one loop could cause the other to be
841 // deleted. If so, don't simplify it.
842 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
843 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
846 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
848 static void RemoveFromWorklist(Instruction *I,
849 std::vector<Instruction*> &Worklist) {
850 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
852 while (WI != Worklist.end()) {
853 unsigned Offset = WI-Worklist.begin();
855 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
859 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
860 /// program, replacing all uses with V and update the worklist.
861 static void ReplaceUsesOfWith(Instruction *I, Value *V,
862 std::vector<Instruction*> &Worklist,
863 Loop *L, LPPassManager *LPM) {
864 DOUT << "Replace with '" << *V << "': " << *I;
866 // Add uses to the worklist, which may be dead now.
867 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
868 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
869 Worklist.push_back(Use);
871 // Add users to the worklist which may be simplified now.
872 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
874 Worklist.push_back(cast<Instruction>(*UI));
875 LPM->deleteSimpleAnalysisValue(I, L);
876 RemoveFromWorklist(I, Worklist);
877 I->replaceAllUsesWith(V);
878 I->eraseFromParent();
882 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
883 /// information, and remove any dead successors it has.
885 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
886 std::vector<Instruction*> &Worklist,
888 if (pred_begin(BB) != pred_end(BB)) {
889 // This block isn't dead, since an edge to BB was just removed, see if there
890 // are any easy simplifications we can do now.
891 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
892 // If it has one pred, fold phi nodes in BB.
893 while (isa<PHINode>(BB->begin()))
894 ReplaceUsesOfWith(BB->begin(),
895 cast<PHINode>(BB->begin())->getIncomingValue(0),
898 // If this is the header of a loop and the only pred is the latch, we now
899 // have an unreachable loop.
900 if (Loop *L = LI->getLoopFor(BB))
901 if (L->getHeader() == BB && L->contains(Pred)) {
902 // Remove the branch from the latch to the header block, this makes
903 // the header dead, which will make the latch dead (because the header
904 // dominates the latch).
905 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
906 Pred->getTerminator()->eraseFromParent();
907 new UnreachableInst(Pred);
909 // The loop is now broken, remove it from LI.
910 RemoveLoopFromHierarchy(L);
912 // Reprocess the header, which now IS dead.
913 RemoveBlockIfDead(BB, Worklist, L);
917 // If pred ends in a uncond branch, add uncond branch to worklist so that
918 // the two blocks will get merged.
919 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
920 if (BI->isUnconditional())
921 Worklist.push_back(BI);
926 DOUT << "Nuking dead block: " << *BB;
928 // Remove the instructions in the basic block from the worklist.
929 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
930 RemoveFromWorklist(I, Worklist);
932 // Anything that uses the instructions in this basic block should have their
933 // uses replaced with undefs.
935 I->replaceAllUsesWith(UndefValue::get(I->getType()));
938 // If this is the edge to the header block for a loop, remove the loop and
939 // promote all subloops.
940 if (Loop *BBLoop = LI->getLoopFor(BB)) {
941 if (BBLoop->getLoopLatch() == BB)
942 RemoveLoopFromHierarchy(BBLoop);
945 // Remove the block from the loop info, which removes it from any loops it
950 // Remove phi node entries in successors for this block.
951 TerminatorInst *TI = BB->getTerminator();
952 std::vector<BasicBlock*> Succs;
953 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
954 Succs.push_back(TI->getSuccessor(i));
955 TI->getSuccessor(i)->removePredecessor(BB);
958 // Unique the successors, remove anything with multiple uses.
959 std::sort(Succs.begin(), Succs.end());
960 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
962 // Remove the basic block, including all of the instructions contained in it.
963 LPM->deleteSimpleAnalysisValue(BB, L);
964 BB->eraseFromParent();
965 // Remove successor blocks here that are not dead, so that we know we only
966 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
967 // then getting removed before we revisit them, which is badness.
969 for (unsigned i = 0; i != Succs.size(); ++i)
970 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
971 // One exception is loop headers. If this block was the preheader for a
972 // loop, then we DO want to visit the loop so the loop gets deleted.
973 // We know that if the successor is a loop header, that this loop had to
974 // be the preheader: the case where this was the latch block was handled
975 // above and headers can only have two predecessors.
976 if (!LI->isLoopHeader(Succs[i])) {
977 Succs.erase(Succs.begin()+i);
982 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
983 RemoveBlockIfDead(Succs[i], Worklist, L);
986 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
987 /// become unwrapped, either because the backedge was deleted, or because the
988 /// edge into the header was removed. If the edge into the header from the
989 /// latch block was removed, the loop is unwrapped but subloops are still alive,
990 /// so they just reparent loops. If the loops are actually dead, they will be
992 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
993 LPM->deleteLoopFromQueue(L);
994 RemoveLoopFromWorklist(L);
999 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
1000 // the value specified by Val in the specified loop, or we know it does NOT have
1001 // that value. Rewrite any uses of LIC or of properties correlated to it.
1002 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
1005 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
1007 // FIXME: Support correlated properties, like:
1014 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
1015 // selects, switches.
1016 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
1017 std::vector<Instruction*> Worklist;
1019 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
1020 // in the loop with the appropriate one directly.
1021 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
1026 Replacement = ConstantInt::get(Type::Int1Ty,
1027 !cast<ConstantInt>(Val)->getZExtValue());
1029 for (unsigned i = 0, e = Users.size(); i != e; ++i)
1030 if (Instruction *U = cast<Instruction>(Users[i])) {
1031 if (!L->contains(U->getParent()))
1033 U->replaceUsesOfWith(LIC, Replacement);
1034 Worklist.push_back(U);
1037 // Otherwise, we don't know the precise value of LIC, but we do know that it
1038 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1039 // can. This case occurs when we unswitch switch statements.
1040 for (unsigned i = 0, e = Users.size(); i != e; ++i)
1041 if (Instruction *U = cast<Instruction>(Users[i])) {
1042 if (!L->contains(U->getParent()))
1045 Worklist.push_back(U);
1047 // If we know that LIC is not Val, use this info to simplify code.
1048 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
1049 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
1050 if (SI->getCaseValue(i) == Val) {
1051 // Found a dead case value. Don't remove PHI nodes in the
1052 // successor if they become single-entry, those PHI nodes may
1053 // be in the Users list.
1055 // FIXME: This is a hack. We need to keep the successor around
1056 // and hooked up so as to preserve the loop structure, because
1057 // trying to update it is complicated. So instead we preserve the
1058 // loop structure and put the block on an dead code path.
1060 BasicBlock* Old = SI->getParent();
1061 BasicBlock* Split = SplitBlock(Old, SI, this);
1063 Instruction* OldTerm = Old->getTerminator();
1064 new BranchInst(Split, SI->getSuccessor(i),
1065 ConstantInt::getTrue(), OldTerm);
1067 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
1068 Old->getTerminator()->eraseFromParent();
1071 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin();
1072 (PN = dyn_cast<PHINode>(II)); ++II) {
1073 Value *InVal = PN->removeIncomingValue(Split, false);
1074 PN->addIncoming(InVal, Old);
1083 // TODO: We could do other simplifications, for example, turning
1084 // LIC == Val -> false.
1088 SimplifyCode(Worklist, L);
1091 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1092 /// loop, walk over it and constant prop, dce, and fold control flow where
1093 /// possible. Note that this is effectively a very simple loop-structure-aware
1094 /// optimizer. During processing of this loop, L could very well be deleted, so
1095 /// it must not be used.
1097 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1100 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1101 while (!Worklist.empty()) {
1102 Instruction *I = Worklist.back();
1103 Worklist.pop_back();
1105 // Simple constant folding.
1106 if (Constant *C = ConstantFoldInstruction(I)) {
1107 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
1112 if (isInstructionTriviallyDead(I)) {
1113 DOUT << "Remove dead instruction '" << *I;
1115 // Add uses to the worklist, which may be dead now.
1116 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1117 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1118 Worklist.push_back(Use);
1119 LPM->deleteSimpleAnalysisValue(I, L);
1120 RemoveFromWorklist(I, Worklist);
1121 I->eraseFromParent();
1126 // Special case hacks that appear commonly in unswitched code.
1127 switch (I->getOpcode()) {
1128 case Instruction::Select:
1129 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1130 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1135 case Instruction::And:
1136 if (isa<ConstantInt>(I->getOperand(0)) &&
1137 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1138 cast<BinaryOperator>(I)->swapOperands();
1139 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1140 if (CB->getType() == Type::Int1Ty) {
1141 if (CB->isOne()) // X & 1 -> X
1142 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1144 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1148 case Instruction::Or:
1149 if (isa<ConstantInt>(I->getOperand(0)) &&
1150 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1151 cast<BinaryOperator>(I)->swapOperands();
1152 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1153 if (CB->getType() == Type::Int1Ty) {
1154 if (CB->isOne()) // X | 1 -> 1
1155 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1157 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1161 case Instruction::Br: {
1162 BranchInst *BI = cast<BranchInst>(I);
1163 if (BI->isUnconditional()) {
1164 // If BI's parent is the only pred of the successor, fold the two blocks
1166 BasicBlock *Pred = BI->getParent();
1167 BasicBlock *Succ = BI->getSuccessor(0);
1168 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1169 if (!SinglePred) continue; // Nothing to do.
1170 assert(SinglePred == Pred && "CFG broken");
1172 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1173 << Succ->getName() << "\n";
1175 // Resolve any single entry PHI nodes in Succ.
1176 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1177 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1179 // Move all of the successor contents from Succ to Pred.
1180 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1182 LPM->deleteSimpleAnalysisValue(BI, L);
1183 BI->eraseFromParent();
1184 RemoveFromWorklist(BI, Worklist);
1186 // If Succ has any successors with PHI nodes, update them to have
1187 // entries coming from Pred instead of Succ.
1188 Succ->replaceAllUsesWith(Pred);
1190 // Remove Succ from the loop tree.
1191 LI->removeBlock(Succ);
1192 LPM->deleteSimpleAnalysisValue(Succ, L);
1193 Succ->eraseFromParent();
1195 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1196 // Conditional branch. Turn it into an unconditional branch, then
1197 // remove dead blocks.
1198 break; // FIXME: Enable.
1200 DOUT << "Folded branch: " << *BI;
1201 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1202 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1203 DeadSucc->removePredecessor(BI->getParent(), true);
1204 Worklist.push_back(new BranchInst(LiveSucc, BI));
1205 LPM->deleteSimpleAnalysisValue(BI, L);
1206 BI->eraseFromParent();
1207 RemoveFromWorklist(BI, Worklist);
1210 RemoveBlockIfDead(DeadSucc, Worklist, L);