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;
74 static char ID; // Pass ID, replacement for typeid
75 LoopUnswitch(bool Os = false) :
76 LoopPass((intptr_t)&ID), OptimizeForSize(Os) {}
78 bool runOnLoop(Loop *L, LPPassManager &LPM);
80 /// This transformation requires natural loop information & requires that
81 /// loop preheaders be inserted into the CFG...
83 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
84 AU.addRequiredID(LoopSimplifyID);
85 AU.addPreservedID(LoopSimplifyID);
86 AU.addPreserved<DominatorTree>();
87 AU.addPreserved<DominanceFrontier>();
88 AU.addRequired<LoopInfo>();
89 AU.addPreserved<LoopInfo>();
90 AU.addRequiredID(LCSSAID);
91 AU.addPreservedID(LCSSAID);
95 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
97 void RemoveLoopFromWorklist(Loop *L) {
98 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
99 LoopProcessWorklist.end(), L);
100 if (I != LoopProcessWorklist.end())
101 LoopProcessWorklist.erase(I);
104 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L);
105 unsigned getLoopUnswitchCost(Loop *L, Value *LIC);
106 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
107 BasicBlock *ExitBlock);
108 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
109 BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To);
110 BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt);
112 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
113 Constant *Val, bool isEqual);
115 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
116 BasicBlock *TrueDest,
117 BasicBlock *FalseDest,
118 Instruction *InsertPt);
120 void SimplifyCode(std::vector<Instruction*> &Worklist);
121 void RemoveBlockIfDead(BasicBlock *BB,
122 std::vector<Instruction*> &Worklist);
123 void RemoveLoopFromHierarchy(Loop *L);
125 char LoopUnswitch::ID = 0;
126 RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
129 LoopPass *llvm::createLoopUnswitchPass(bool Os) {
130 return new LoopUnswitch(Os);
133 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
134 /// invariant in the loop, or has an invariant piece, return the invariant.
135 /// Otherwise, return null.
136 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
137 // Constants should be folded, not unswitched on!
138 if (isa<Constant>(Cond)) return false;
140 // TODO: Handle: br (VARIANT|INVARIANT).
141 // TODO: Hoist simple expressions out of loops.
142 if (L->isLoopInvariant(Cond)) return Cond;
144 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
145 if (BO->getOpcode() == Instruction::And ||
146 BO->getOpcode() == Instruction::Or) {
147 // If either the left or right side is invariant, we can unswitch on this,
148 // which will cause the branch to go away in one loop and the condition to
149 // simplify in the other one.
150 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
152 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
159 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
160 assert(L->isLCSSAForm());
161 LI = &getAnalysis<LoopInfo>();
163 bool Changed = false;
165 // Loop over all of the basic blocks in the loop. If we find an interior
166 // block that is branching on a loop-invariant condition, we can unswitch this
168 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
170 TerminatorInst *TI = (*I)->getTerminator();
171 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
172 // If this isn't branching on an invariant condition, we can't unswitch
174 if (BI->isConditional()) {
175 // See if this, or some part of it, is loop invariant. If so, we can
176 // unswitch on it if we desire.
177 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed);
178 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
184 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
185 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
186 if (LoopCond && SI->getNumCases() > 1) {
187 // Find a value to unswitch on:
188 // FIXME: this should chose the most expensive case!
189 Constant *UnswitchVal = SI->getCaseValue(1);
190 // Do not process same value again and again.
191 if (!UnswitchedVals.insert(UnswitchVal))
194 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) {
201 // Scan the instructions to check for unswitchable values.
202 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
204 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
205 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
206 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
214 assert(L->isLCSSAForm());
219 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
220 /// 1. Exit the loop with no side effects.
221 /// 2. Branch to the latch block with no side-effects.
223 /// If these conditions are true, we return true and set ExitBB to the block we
226 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
228 std::set<BasicBlock*> &Visited) {
229 if (!Visited.insert(BB).second) {
230 // Already visited and Ok, end of recursion.
232 } else if (!L->contains(BB)) {
233 // Otherwise, this is a loop exit, this is fine so long as this is the
235 if (ExitBB != 0) return false;
240 // Otherwise, this is an unvisited intra-loop node. Check all successors.
241 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
242 // Check to see if the successor is a trivial loop exit.
243 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
247 // Okay, everything after this looks good, check to make sure that this block
248 // doesn't include any side effects.
249 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
250 if (I->mayWriteToMemory())
256 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
257 /// leads to an exit from the specified loop, and has no side-effects in the
258 /// process. If so, return the block that is exited to, otherwise return null.
259 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
260 std::set<BasicBlock*> Visited;
261 Visited.insert(L->getHeader()); // Branches to header are ok.
262 BasicBlock *ExitBB = 0;
263 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
268 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
269 /// trivial: that is, that the condition controls whether or not the loop does
270 /// anything at all. If this is a trivial condition, unswitching produces no
271 /// code duplications (equivalently, it produces a simpler loop and a new empty
272 /// loop, which gets deleted).
274 /// If this is a trivial condition, return true, otherwise return false. When
275 /// returning true, this sets Cond and Val to the condition that controls the
276 /// trivial condition: when Cond dynamically equals Val, the loop is known to
277 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
280 static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0,
281 BasicBlock **LoopExit = 0) {
282 BasicBlock *Header = L->getHeader();
283 TerminatorInst *HeaderTerm = Header->getTerminator();
285 BasicBlock *LoopExitBB = 0;
286 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
287 // If the header block doesn't end with a conditional branch on Cond, we
289 if (!BI->isConditional() || BI->getCondition() != Cond)
292 // Check to see if a successor of the branch is guaranteed to go to the
293 // latch block or exit through a one exit block without having any
294 // side-effects. If so, determine the value of Cond that causes it to do
296 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) {
297 if (Val) *Val = ConstantInt::getTrue();
298 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) {
299 if (Val) *Val = ConstantInt::getFalse();
301 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
302 // If this isn't a switch on Cond, we can't handle it.
303 if (SI->getCondition() != Cond) return false;
305 // Check to see if a successor of the switch is guaranteed to go to the
306 // latch block or exit through a one exit block without having any
307 // side-effects. If so, determine the value of Cond that causes it to do
308 // this. Note that we can't trivially unswitch on the default case.
309 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
310 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) {
311 // Okay, we found a trivial case, remember the value that is trivial.
312 if (Val) *Val = SI->getCaseValue(i);
317 // If we didn't find a single unique LoopExit block, or if the loop exit block
318 // contains phi nodes, this isn't trivial.
319 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
320 return false; // Can't handle this.
322 if (LoopExit) *LoopExit = LoopExitBB;
324 // We already know that nothing uses any scalar values defined inside of this
325 // loop. As such, we just have to check to see if this loop will execute any
326 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
327 // part of the loop that the code *would* execute. We already checked the
328 // tail, check the header now.
329 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
330 if (I->mayWriteToMemory())
335 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
336 /// we choose to unswitch the specified loop on the specified value.
338 unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) {
339 // If the condition is trivial, always unswitch. There is no code growth for
341 if (IsTrivialUnswitchCondition(L, LIC))
344 // FIXME: This is really overly conservative. However, more liberal
345 // estimations have thus far resulted in excessive unswitching, which is bad
346 // both in compile time and in code size. This should be replaced once
347 // someone figures out how a good estimation.
348 return L->getBlocks().size();
351 // FIXME: this is brain dead. It should take into consideration code
353 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
356 // Do not include empty blocks in the cost calculation. This happen due to
357 // loop canonicalization and will be removed.
358 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
361 // Count basic blocks.
368 /// UnswitchIfProfitable - We have found that we can unswitch L when
369 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
370 /// unswitch the loop, reprocess the pieces, then return true.
371 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){
372 // Check to see if it would be profitable to unswitch this loop.
373 unsigned Cost = getLoopUnswitchCost(L, LoopCond);
375 // Do not do non-trivial unswitch while optimizing for size.
376 if (Cost && OptimizeForSize)
379 if (Cost > Threshold) {
380 // FIXME: this should estimate growth by the amount of code shared by the
381 // resultant unswitched loops.
383 DOUT << "NOT unswitching loop %"
384 << L->getHeader()->getName() << ", cost too high: "
385 << L->getBlocks().size() << "\n";
389 // If this is a trivial condition to unswitch (which results in no code
390 // duplication), do it now.
392 BasicBlock *ExitBlock;
393 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) {
394 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock);
396 UnswitchNontrivialCondition(LoopCond, Val, L);
402 /// SplitBlock - Split the specified block at the specified instruction - every
403 /// thing before SplitPt stays in Old and everything starting with SplitPt moves
404 /// to a new block. The two blocks are joined by an unconditional branch and
405 /// the loop info is updated.
407 BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) {
408 BasicBlock::iterator SplitIt = SplitPt;
409 while (isa<PHINode>(SplitIt))
411 BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split");
413 // The new block lives in whichever loop the old one did.
414 if (Loop *L = LI->getLoopFor(Old))
415 L->addBasicBlockToLoop(New, *LI);
417 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>())
418 DT->addNewBlock(New, Old);
420 if (DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>())
427 BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) {
428 TerminatorInst *LatchTerm = BB->getTerminator();
429 unsigned SuccNum = 0;
430 for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) {
431 assert(i != e && "Didn't find edge?");
432 if (LatchTerm->getSuccessor(i) == Succ) {
438 // If this is a critical edge, let SplitCriticalEdge do it.
439 if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this))
440 return LatchTerm->getSuccessor(SuccNum);
442 // If the edge isn't critical, then BB has a single successor or Succ has a
443 // single pred. Split the block.
444 BasicBlock::iterator SplitPoint;
445 if (BasicBlock *SP = Succ->getSinglePredecessor()) {
446 // If the successor only has a single pred, split the top of the successor
448 assert(SP == BB && "CFG broken");
449 return SplitBlock(Succ, Succ->begin());
451 // Otherwise, if BB has a single successor, split it at the bottom of the
453 assert(BB->getTerminator()->getNumSuccessors() == 1 &&
454 "Should have a single succ!");
455 return SplitBlock(BB, BB->getTerminator());
461 // RemapInstruction - Convert the instruction operands from referencing the
462 // current values into those specified by ValueMap.
464 static inline void RemapInstruction(Instruction *I,
465 DenseMap<const Value *, Value*> &ValueMap) {
466 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
467 Value *Op = I->getOperand(op);
468 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
469 if (It != ValueMap.end()) Op = It->second;
470 I->setOperand(op, Op);
474 // CloneDomInfo - NewBB is cloned from Orig basic block. Now clone Dominator Info.
475 // If Orig is in Loop then find and use Orig dominator's cloned block as NewBB
477 void CloneDomInfo(BasicBlock *NewBB, BasicBlock *Orig, Loop *L,
478 DominatorTree *DT, DominanceFrontier *DF,
479 DenseMap<const Value*, Value*> &VM) {
481 DomTreeNode *OrigNode = DT->getNode(Orig);
484 BasicBlock *OrigIDom = OrigNode->getBlock();
485 BasicBlock *NewIDom = OrigIDom;
486 if (L->contains(OrigIDom)) {
487 if (!DT->getNode(OrigIDom))
488 CloneDomInfo(NewIDom, OrigIDom, L, DT, DF, VM);
489 NewIDom = cast<BasicBlock>(VM[OrigIDom]);
491 if (NewBB == NewIDom) {
492 DT->addNewBlock(NewBB, OrigIDom);
493 DT->changeImmediateDominator(NewBB, NewIDom);
495 DT->addNewBlock(NewBB, NewIDom);
497 DominanceFrontier::DomSetType NewDFSet;
499 DominanceFrontier::iterator DFI = DF->find(Orig);
500 if ( DFI != DF->end()) {
501 DominanceFrontier::DomSetType S = DFI->second;
502 for (DominanceFrontier::DomSetType::iterator I = S.begin(), E = S.end();
506 NewDFSet.insert(cast<BasicBlock>(VM[Orig]));
511 DF->addBasicBlock(NewBB, NewDFSet);
515 /// CloneLoop - Recursively clone the specified loop and all of its children,
516 /// mapping the blocks with the specified map.
517 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
518 LoopInfo *LI, LPPassManager *LPM) {
519 Loop *New = new Loop();
521 LPM->insertLoop(New, PL);
523 // Add all of the blocks in L to the new loop.
524 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
526 if (LI->getLoopFor(*I) == L)
527 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
529 // Add all of the subloops to the new loop.
530 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
531 CloneLoop(*I, New, VM, LI, LPM);
536 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
537 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
538 /// code immediately before InsertPt.
539 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
540 BasicBlock *TrueDest,
541 BasicBlock *FalseDest,
542 Instruction *InsertPt) {
543 // Insert a conditional branch on LIC to the two preheaders. The original
544 // code is the true version and the new code is the false version.
545 Value *BranchVal = LIC;
546 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
547 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
548 else if (Val != ConstantInt::getTrue())
549 // We want to enter the new loop when the condition is true.
550 std::swap(TrueDest, FalseDest);
552 // Insert the new branch.
553 BranchInst *BRI = new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt);
555 // Update dominator info.
556 // BranchVal is a new preheader so it dominates true and false destination
558 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>()) {
559 DT->changeImmediateDominator(TrueDest, BRI->getParent());
560 DT->changeImmediateDominator(FalseDest, BRI->getParent());
562 // No need to update DominanceFrontier. BRI->getParent() dominated TrueDest
563 // and FalseDest anyway. Now it immediately dominates them.
567 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
568 /// condition in it (a cond branch from its header block to its latch block,
569 /// where the path through the loop that doesn't execute its body has no
570 /// side-effects), unswitch it. This doesn't involve any code duplication, just
571 /// moving the conditional branch outside of the loop and updating loop info.
572 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
574 BasicBlock *ExitBlock) {
575 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
576 << L->getHeader()->getName() << " [" << L->getBlocks().size()
577 << " blocks] in Function " << L->getHeader()->getParent()->getName()
578 << " on cond: " << *Val << " == " << *Cond << "\n";
580 // First step, split the preheader, so that we know that there is a safe place
581 // to insert the conditional branch. We will change 'OrigPH' to have a
582 // conditional branch on Cond.
583 BasicBlock *OrigPH = L->getLoopPreheader();
584 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader());
586 // Now that we have a place to insert the conditional branch, create a place
587 // to branch to: this is the exit block out of the loop that we should
590 // Split this block now, so that the loop maintains its exit block, and so
591 // that the jump from the preheader can execute the contents of the exit block
592 // without actually branching to it (the exit block should be dominated by the
593 // loop header, not the preheader).
594 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
595 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin());
597 // Okay, now we have a position to branch from and a position to branch to,
598 // insert the new conditional branch.
599 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
600 OrigPH->getTerminator());
601 OrigPH->getTerminator()->eraseFromParent();
603 // We need to reprocess this loop, it could be unswitched again.
606 // Now that we know that the loop is never entered when this condition is a
607 // particular value, rewrite the loop with this info. We know that this will
608 // at least eliminate the old branch.
609 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
613 /// VersionLoop - We determined that the loop is profitable to unswitch when LIC
614 /// equal Val. Split it into loop versions and test the condition outside of
615 /// either loop. Return the loops created as Out1/Out2.
616 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
618 Function *F = L->getHeader()->getParent();
619 DOUT << "loop-unswitch: Unswitching loop %"
620 << L->getHeader()->getName() << " [" << L->getBlocks().size()
621 << " blocks] in Function " << F->getName()
622 << " when '" << *Val << "' == " << *LIC << "\n";
624 // LoopBlocks contains all of the basic blocks of the loop, including the
625 // preheader of the loop, the body of the loop, and the exit blocks of the
626 // loop, in that order.
627 std::vector<BasicBlock*> LoopBlocks;
629 // First step, split the preheader and exit blocks, and add these blocks to
630 // the LoopBlocks list.
631 BasicBlock *OrigPreheader = L->getLoopPreheader();
632 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader()));
634 // We want the loop to come after the preheader, but before the exit blocks.
635 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
637 std::vector<BasicBlock*> ExitBlocks;
638 L->getUniqueExitBlocks(ExitBlocks);
640 // Split all of the edges from inside the loop to their exit blocks. Update
641 // the appropriate Phi nodes as we do so.
642 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
643 BasicBlock *ExitBlock = ExitBlocks[i];
644 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
646 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
647 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock);
648 BasicBlock* StartBlock = Preds[j];
649 BasicBlock* EndBlock;
650 if (MiddleBlock->getSinglePredecessor() == ExitBlock) {
651 EndBlock = MiddleBlock;
652 MiddleBlock = EndBlock->getSinglePredecessor();;
654 EndBlock = ExitBlock;
657 std::set<PHINode*> InsertedPHIs;
658 PHINode* OldLCSSA = 0;
659 for (BasicBlock::iterator I = EndBlock->begin();
660 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
661 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock);
662 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(),
663 OldLCSSA->getName() + ".us-lcssa",
664 MiddleBlock->getTerminator());
665 NewLCSSA->addIncoming(OldValue, StartBlock);
666 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock),
668 InsertedPHIs.insert(NewLCSSA);
671 BasicBlock::iterator InsertPt = EndBlock->begin();
672 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt;
673 for (BasicBlock::iterator I = MiddleBlock->begin();
674 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
676 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(),
677 OldLCSSA->getName() + ".us-lcssa",
679 OldLCSSA->replaceAllUsesWith(NewLCSSA);
680 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock);
685 // The exit blocks may have been changed due to edge splitting, recompute.
687 L->getUniqueExitBlocks(ExitBlocks);
689 // Add exit blocks to the loop blocks.
690 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
692 // Next step, clone all of the basic blocks that make up the loop (including
693 // the loop preheader and exit blocks), keeping track of the mapping between
694 // the instructions and blocks.
695 std::vector<BasicBlock*> NewBlocks;
696 NewBlocks.reserve(LoopBlocks.size());
697 DenseMap<const Value*, Value*> ValueMap;
698 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
699 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
700 NewBlocks.push_back(New);
701 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
704 // Update dominator info
705 DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>();
706 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>())
707 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
708 BasicBlock *LBB = LoopBlocks[i];
709 BasicBlock *NBB = NewBlocks[i];
710 CloneDomInfo(NBB, LBB, L, DT, DF, ValueMap);
713 // Splice the newly inserted blocks into the function right before the
714 // original preheader.
715 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
716 NewBlocks[0], F->end());
718 // Now we create the new Loop object for the versioned loop.
719 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
720 Loop *ParentLoop = L->getParentLoop();
722 // Make sure to add the cloned preheader and exit blocks to the parent loop
724 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
727 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
728 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
729 // The new exit block should be in the same loop as the old one.
730 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
731 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
733 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
734 "Exit block should have been split to have one successor!");
735 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
737 // If the successor of the exit block had PHI nodes, add an entry for
740 for (BasicBlock::iterator I = ExitSucc->begin();
741 (PN = dyn_cast<PHINode>(I)); ++I) {
742 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
743 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
744 if (It != ValueMap.end()) V = It->second;
745 PN->addIncoming(V, NewExit);
749 // Rewrite the code to refer to itself.
750 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
751 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
752 E = NewBlocks[i]->end(); I != E; ++I)
753 RemapInstruction(I, ValueMap);
755 // Rewrite the original preheader to select between versions of the loop.
756 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator());
757 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
758 "Preheader splitting did not work correctly!");
760 // Emit the new branch that selects between the two versions of this loop.
761 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
762 OldBR->eraseFromParent();
764 LoopProcessWorklist.push_back(NewLoop);
767 // Now we rewrite the original code to know that the condition is true and the
768 // new code to know that the condition is false.
769 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
771 // It's possible that simplifying one loop could cause the other to be
772 // deleted. If so, don't simplify it.
773 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
774 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
777 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
779 static void RemoveFromWorklist(Instruction *I,
780 std::vector<Instruction*> &Worklist) {
781 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
783 while (WI != Worklist.end()) {
784 unsigned Offset = WI-Worklist.begin();
786 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
790 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
791 /// program, replacing all uses with V and update the worklist.
792 static void ReplaceUsesOfWith(Instruction *I, Value *V,
793 std::vector<Instruction*> &Worklist) {
794 DOUT << "Replace with '" << *V << "': " << *I;
796 // Add uses to the worklist, which may be dead now.
797 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
798 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
799 Worklist.push_back(Use);
801 // Add users to the worklist which may be simplified now.
802 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
804 Worklist.push_back(cast<Instruction>(*UI));
805 I->replaceAllUsesWith(V);
806 I->eraseFromParent();
807 RemoveFromWorklist(I, Worklist);
811 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
812 /// information, and remove any dead successors it has.
814 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
815 std::vector<Instruction*> &Worklist) {
816 if (pred_begin(BB) != pred_end(BB)) {
817 // This block isn't dead, since an edge to BB was just removed, see if there
818 // are any easy simplifications we can do now.
819 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
820 // If it has one pred, fold phi nodes in BB.
821 while (isa<PHINode>(BB->begin()))
822 ReplaceUsesOfWith(BB->begin(),
823 cast<PHINode>(BB->begin())->getIncomingValue(0),
826 // If this is the header of a loop and the only pred is the latch, we now
827 // have an unreachable loop.
828 if (Loop *L = LI->getLoopFor(BB))
829 if (L->getHeader() == BB && L->contains(Pred)) {
830 // Remove the branch from the latch to the header block, this makes
831 // the header dead, which will make the latch dead (because the header
832 // dominates the latch).
833 Pred->getTerminator()->eraseFromParent();
834 new UnreachableInst(Pred);
836 // The loop is now broken, remove it from LI.
837 RemoveLoopFromHierarchy(L);
839 // Reprocess the header, which now IS dead.
840 RemoveBlockIfDead(BB, Worklist);
844 // If pred ends in a uncond branch, add uncond branch to worklist so that
845 // the two blocks will get merged.
846 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
847 if (BI->isUnconditional())
848 Worklist.push_back(BI);
853 DOUT << "Nuking dead block: " << *BB;
855 // Remove the instructions in the basic block from the worklist.
856 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
857 RemoveFromWorklist(I, Worklist);
859 // Anything that uses the instructions in this basic block should have their
860 // uses replaced with undefs.
862 I->replaceAllUsesWith(UndefValue::get(I->getType()));
865 // If this is the edge to the header block for a loop, remove the loop and
866 // promote all subloops.
867 if (Loop *BBLoop = LI->getLoopFor(BB)) {
868 if (BBLoop->getLoopLatch() == BB)
869 RemoveLoopFromHierarchy(BBLoop);
872 // Remove the block from the loop info, which removes it from any loops it
877 // Remove phi node entries in successors for this block.
878 TerminatorInst *TI = BB->getTerminator();
879 std::vector<BasicBlock*> Succs;
880 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
881 Succs.push_back(TI->getSuccessor(i));
882 TI->getSuccessor(i)->removePredecessor(BB);
885 // Unique the successors, remove anything with multiple uses.
886 std::sort(Succs.begin(), Succs.end());
887 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
889 // Remove the basic block, including all of the instructions contained in it.
890 BB->eraseFromParent();
892 // Remove successor blocks here that are not dead, so that we know we only
893 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
894 // then getting removed before we revisit them, which is badness.
896 for (unsigned i = 0; i != Succs.size(); ++i)
897 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
898 // One exception is loop headers. If this block was the preheader for a
899 // loop, then we DO want to visit the loop so the loop gets deleted.
900 // We know that if the successor is a loop header, that this loop had to
901 // be the preheader: the case where this was the latch block was handled
902 // above and headers can only have two predecessors.
903 if (!LI->isLoopHeader(Succs[i])) {
904 Succs.erase(Succs.begin()+i);
909 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
910 RemoveBlockIfDead(Succs[i], Worklist);
913 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
914 /// become unwrapped, either because the backedge was deleted, or because the
915 /// edge into the header was removed. If the edge into the header from the
916 /// latch block was removed, the loop is unwrapped but subloops are still alive,
917 /// so they just reparent loops. If the loops are actually dead, they will be
919 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
920 LPM->deleteLoopFromQueue(L);
921 RemoveLoopFromWorklist(L);
926 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
927 // the value specified by Val in the specified loop, or we know it does NOT have
928 // that value. Rewrite any uses of LIC or of properties correlated to it.
929 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
932 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
934 // FIXME: Support correlated properties, like:
941 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
942 // selects, switches.
943 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
944 std::vector<Instruction*> Worklist;
946 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
947 // in the loop with the appropriate one directly.
948 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
953 Replacement = ConstantInt::get(Type::Int1Ty,
954 !cast<ConstantInt>(Val)->getZExtValue());
956 for (unsigned i = 0, e = Users.size(); i != e; ++i)
957 if (Instruction *U = cast<Instruction>(Users[i])) {
958 if (!L->contains(U->getParent()))
960 U->replaceUsesOfWith(LIC, Replacement);
961 Worklist.push_back(U);
964 // Otherwise, we don't know the precise value of LIC, but we do know that it
965 // is certainly NOT "Val". As such, simplify any uses in the loop that we
966 // can. This case occurs when we unswitch switch statements.
967 for (unsigned i = 0, e = Users.size(); i != e; ++i)
968 if (Instruction *U = cast<Instruction>(Users[i])) {
969 if (!L->contains(U->getParent()))
972 Worklist.push_back(U);
974 // If we know that LIC is not Val, use this info to simplify code.
975 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
976 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
977 if (SI->getCaseValue(i) == Val) {
978 // Found a dead case value. Don't remove PHI nodes in the
979 // successor if they become single-entry, those PHI nodes may
980 // be in the Users list.
982 // FIXME: This is a hack. We need to keep the successor around
983 // and hooked up so as to preserve the loop structure, because
984 // trying to update it is complicated. So instead we preserve the
985 // loop structure and put the block on an dead code path.
987 BasicBlock* Old = SI->getParent();
988 BasicBlock* Split = SplitBlock(Old, SI);
990 Instruction* OldTerm = Old->getTerminator();
991 new BranchInst(Split, SI->getSuccessor(i),
992 ConstantInt::getTrue(), OldTerm);
994 Old->getTerminator()->eraseFromParent();
998 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin();
999 (PN = dyn_cast<PHINode>(II)); ++II) {
1000 Value *InVal = PN->removeIncomingValue(Split, false);
1001 PN->addIncoming(InVal, Old);
1010 // TODO: We could do other simplifications, for example, turning
1011 // LIC == Val -> false.
1015 SimplifyCode(Worklist);
1018 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1019 /// loop, walk over it and constant prop, dce, and fold control flow where
1020 /// possible. Note that this is effectively a very simple loop-structure-aware
1021 /// optimizer. During processing of this loop, L could very well be deleted, so
1022 /// it must not be used.
1024 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1027 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) {
1028 while (!Worklist.empty()) {
1029 Instruction *I = Worklist.back();
1030 Worklist.pop_back();
1032 // Simple constant folding.
1033 if (Constant *C = ConstantFoldInstruction(I)) {
1034 ReplaceUsesOfWith(I, C, Worklist);
1039 if (isInstructionTriviallyDead(I)) {
1040 DOUT << "Remove dead instruction '" << *I;
1042 // Add uses to the worklist, which may be dead now.
1043 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1044 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1045 Worklist.push_back(Use);
1046 I->eraseFromParent();
1047 RemoveFromWorklist(I, Worklist);
1052 // Special case hacks that appear commonly in unswitched code.
1053 switch (I->getOpcode()) {
1054 case Instruction::Select:
1055 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1056 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist);
1060 case Instruction::And:
1061 if (isa<ConstantInt>(I->getOperand(0)) &&
1062 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1063 cast<BinaryOperator>(I)->swapOperands();
1064 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1065 if (CB->getType() == Type::Int1Ty) {
1066 if (CB->isOne()) // X & 1 -> X
1067 ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
1069 ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
1073 case Instruction::Or:
1074 if (isa<ConstantInt>(I->getOperand(0)) &&
1075 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1076 cast<BinaryOperator>(I)->swapOperands();
1077 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1078 if (CB->getType() == Type::Int1Ty) {
1079 if (CB->isOne()) // X | 1 -> 1
1080 ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
1082 ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
1086 case Instruction::Br: {
1087 BranchInst *BI = cast<BranchInst>(I);
1088 if (BI->isUnconditional()) {
1089 // If BI's parent is the only pred of the successor, fold the two blocks
1091 BasicBlock *Pred = BI->getParent();
1092 BasicBlock *Succ = BI->getSuccessor(0);
1093 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1094 if (!SinglePred) continue; // Nothing to do.
1095 assert(SinglePred == Pred && "CFG broken");
1097 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1098 << Succ->getName() << "\n";
1100 // Resolve any single entry PHI nodes in Succ.
1101 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1102 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist);
1104 // Move all of the successor contents from Succ to Pred.
1105 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1107 BI->eraseFromParent();
1108 RemoveFromWorklist(BI, Worklist);
1110 // If Succ has any successors with PHI nodes, update them to have
1111 // entries coming from Pred instead of Succ.
1112 Succ->replaceAllUsesWith(Pred);
1114 // Remove Succ from the loop tree.
1115 LI->removeBlock(Succ);
1116 Succ->eraseFromParent();
1118 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1119 // Conditional branch. Turn it into an unconditional branch, then
1120 // remove dead blocks.
1121 break; // FIXME: Enable.
1123 DOUT << "Folded branch: " << *BI;
1124 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1125 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1126 DeadSucc->removePredecessor(BI->getParent(), true);
1127 Worklist.push_back(new BranchInst(LiveSucc, BI));
1128 BI->eraseFromParent();
1129 RemoveFromWorklist(BI, Worklist);
1132 RemoveBlockIfDead(DeadSucc, Worklist);