1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #define DEBUG_TYPE "loop-unswitch"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Constants.h"
32 #include "llvm/DerivedTypes.h"
33 #include "llvm/Function.h"
34 #include "llvm/Instructions.h"
35 #include "llvm/Analysis/ConstantFolding.h"
36 #include "llvm/Analysis/LoopInfo.h"
37 #include "llvm/Analysis/LoopPass.h"
38 #include "llvm/Analysis/Dominators.h"
39 #include "llvm/Transforms/Utils/Cloning.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/Support/Debug.h"
51 STATISTIC(NumBranches, "Number of branches unswitched");
52 STATISTIC(NumSwitches, "Number of switches unswitched");
53 STATISTIC(NumSelects , "Number of selects unswitched");
54 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
55 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
57 static cl::opt<unsigned>
58 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
59 cl::init(10), cl::Hidden);
62 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
63 LoopInfo *LI; // Loop information
66 // LoopProcessWorklist - Used to check if second loop needs processing
67 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
68 std::vector<Loop*> LoopProcessWorklist;
69 SmallPtrSet<Value *,8> UnswitchedVals;
75 DominanceFrontier *DF;
77 BasicBlock *loopHeader;
78 BasicBlock *loopPreheader;
80 // LoopBlocks contains all of the basic blocks of the loop, including the
81 // preheader of the loop, the body of the loop, and the exit blocks of the
82 // loop, in that order.
83 std::vector<BasicBlock*> LoopBlocks;
84 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
85 std::vector<BasicBlock*> NewBlocks;
88 static char ID; // Pass ID, replacement for typeid
89 explicit LoopUnswitch(bool Os = false) :
90 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
91 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
92 loopPreheader(NULL) {}
94 bool runOnLoop(Loop *L, LPPassManager &LPM);
95 bool processCurrentLoop();
97 /// This transformation requires natural loop information & requires that
98 /// loop preheaders be inserted into the CFG...
100 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
101 AU.addRequiredID(LoopSimplifyID);
102 AU.addPreservedID(LoopSimplifyID);
103 AU.addRequired<LoopInfo>();
104 AU.addPreserved<LoopInfo>();
105 AU.addRequiredID(LCSSAID);
106 AU.addPreservedID(LCSSAID);
107 AU.addPreserved<DominatorTree>();
108 AU.addPreserved<DominanceFrontier>();
113 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
115 void RemoveLoopFromWorklist(Loop *L) {
116 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
117 LoopProcessWorklist.end(), L);
118 if (I != LoopProcessWorklist.end())
119 LoopProcessWorklist.erase(I);
122 void initLoopData() {
123 loopHeader = currentLoop->getHeader();
124 loopPreheader = currentLoop->getLoopPreheader();
127 /// Split all of the edges from inside the loop to their exit blocks.
128 /// Update the appropriate Phi nodes as we do so.
129 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
131 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
132 unsigned getLoopUnswitchCost(Value *LIC);
133 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
134 BasicBlock *ExitBlock);
135 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
137 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
138 Constant *Val, bool isEqual);
140 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
141 BasicBlock *TrueDest,
142 BasicBlock *FalseDest,
143 Instruction *InsertPt);
145 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
146 void RemoveBlockIfDead(BasicBlock *BB,
147 std::vector<Instruction*> &Worklist, Loop *l);
148 void RemoveLoopFromHierarchy(Loop *L);
149 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
150 BasicBlock **LoopExit = 0);
154 char LoopUnswitch::ID = 0;
155 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
157 Pass *llvm::createLoopUnswitchPass(bool Os) {
158 return new LoopUnswitch(Os);
161 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
162 /// invariant in the loop, or has an invariant piece, return the invariant.
163 /// Otherwise, return null.
164 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
165 // Constants should be folded, not unswitched on!
166 if (isa<Constant>(Cond)) return 0;
168 // TODO: Handle: br (VARIANT|INVARIANT).
169 // TODO: Hoist simple expressions out of loops.
170 if (Instruction *I = dyn_cast<Instruction>(Cond))
171 if (!L->contains(I->getParent()))
174 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
175 if (BO->getOpcode() == Instruction::And ||
176 BO->getOpcode() == Instruction::Or) {
177 // If either the left or right side is invariant, we can unswitch on this,
178 // which will cause the branch to go away in one loop and the condition to
179 // simplify in the other one.
180 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
182 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
189 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
190 LI = &getAnalysis<LoopInfo>();
192 DF = getAnalysisToUpdate<DominanceFrontier>();
193 DT = getAnalysisToUpdate<DominatorTree>();
195 Function *F = currentLoop->getHeader()->getParent();
196 bool Changed = false;
198 assert(currentLoop->isLCSSAForm());
200 Changed |= processCurrentLoop();
204 // FIXME: Reconstruct dom info, because it is not preserved properly.
206 DT->runOnFunction(*F);
208 DF->runOnFunction(*F);
213 /// processCurrentLoop - Do actual work and unswitch loop if possible
215 bool LoopUnswitch::processCurrentLoop() {
216 bool Changed = false;
218 // Loop over all of the basic blocks in the loop. If we find an interior
219 // block that is branching on a loop-invariant condition, we can unswitch this
221 for (Loop::block_iterator I = currentLoop->block_begin(),
222 E = currentLoop->block_end();
224 TerminatorInst *TI = (*I)->getTerminator();
225 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
226 // If this isn't branching on an invariant condition, we can't unswitch
228 if (BI->isConditional()) {
229 // See if this, or some part of it, is loop invariant. If so, we can
230 // unswitch on it if we desire.
231 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
232 currentLoop, Changed);
233 if (LoopCond && UnswitchIfProfitable(LoopCond,
234 ConstantInt::getTrue())) {
239 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
240 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
241 currentLoop, Changed);
242 if (LoopCond && SI->getNumCases() > 1) {
243 // Find a value to unswitch on:
244 // FIXME: this should chose the most expensive case!
245 Constant *UnswitchVal = SI->getCaseValue(1);
246 // Do not process same value again and again.
247 if (!UnswitchedVals.insert(UnswitchVal))
250 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
257 // Scan the instructions to check for unswitchable values.
258 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
260 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
261 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
262 currentLoop, Changed);
263 if (LoopCond && UnswitchIfProfitable(LoopCond,
264 ConstantInt::getTrue())) {
273 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
274 /// 1. Exit the loop with no side effects.
275 /// 2. Branch to the latch block with no side-effects.
277 /// If these conditions are true, we return true and set ExitBB to the block we
280 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
282 std::set<BasicBlock*> &Visited) {
283 if (!Visited.insert(BB).second) {
284 // Already visited and Ok, end of recursion.
286 } else if (!L->contains(BB)) {
287 // Otherwise, this is a loop exit, this is fine so long as this is the
289 if (ExitBB != 0) return false;
294 // Otherwise, this is an unvisited intra-loop node. Check all successors.
295 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
296 // Check to see if the successor is a trivial loop exit.
297 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
301 // Okay, everything after this looks good, check to make sure that this block
302 // doesn't include any side effects.
303 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
304 if (I->mayWriteToMemory())
310 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
311 /// leads to an exit from the specified loop, and has no side-effects in the
312 /// process. If so, return the block that is exited to, otherwise return null.
313 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
314 std::set<BasicBlock*> Visited;
315 Visited.insert(L->getHeader()); // Branches to header are ok.
316 BasicBlock *ExitBB = 0;
317 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
322 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
323 /// trivial: that is, that the condition controls whether or not the loop does
324 /// anything at all. If this is a trivial condition, unswitching produces no
325 /// code duplications (equivalently, it produces a simpler loop and a new empty
326 /// loop, which gets deleted).
328 /// If this is a trivial condition, return true, otherwise return false. When
329 /// returning true, this sets Cond and Val to the condition that controls the
330 /// trivial condition: when Cond dynamically equals Val, the loop is known to
331 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
334 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
335 BasicBlock **LoopExit) {
336 BasicBlock *Header = currentLoop->getHeader();
337 TerminatorInst *HeaderTerm = Header->getTerminator();
339 BasicBlock *LoopExitBB = 0;
340 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
341 // If the header block doesn't end with a conditional branch on Cond, we
343 if (!BI->isConditional() || BI->getCondition() != Cond)
346 // Check to see if a successor of the branch is guaranteed to go to the
347 // latch block or exit through a one exit block without having any
348 // side-effects. If so, determine the value of Cond that causes it to do
350 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
351 BI->getSuccessor(0)))) {
352 if (Val) *Val = ConstantInt::getTrue();
353 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
354 BI->getSuccessor(1)))) {
355 if (Val) *Val = ConstantInt::getFalse();
357 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
358 // If this isn't a switch on Cond, we can't handle it.
359 if (SI->getCondition() != Cond) return false;
361 // Check to see if a successor of the switch is guaranteed to go to the
362 // latch block or exit through a one exit block without having any
363 // side-effects. If so, determine the value of Cond that causes it to do
364 // this. Note that we can't trivially unswitch on the default case.
365 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
366 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
367 SI->getSuccessor(i)))) {
368 // Okay, we found a trivial case, remember the value that is trivial.
369 if (Val) *Val = SI->getCaseValue(i);
374 // If we didn't find a single unique LoopExit block, or if the loop exit block
375 // contains phi nodes, this isn't trivial.
376 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
377 return false; // Can't handle this.
379 if (LoopExit) *LoopExit = LoopExitBB;
381 // We already know that nothing uses any scalar values defined inside of this
382 // loop. As such, we just have to check to see if this loop will execute any
383 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
384 // part of the loop that the code *would* execute. We already checked the
385 // tail, check the header now.
386 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
387 if (I->mayWriteToMemory())
392 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
393 /// we choose to unswitch current loop on the specified value.
395 unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
396 // If the condition is trivial, always unswitch. There is no code growth for
398 if (IsTrivialUnswitchCondition(LIC))
401 // FIXME: This is really overly conservative. However, more liberal
402 // estimations have thus far resulted in excessive unswitching, which is bad
403 // both in compile time and in code size. This should be replaced once
404 // someone figures out how a good estimation.
405 return currentLoop->getBlocks().size();
408 // FIXME: this is brain dead. It should take into consideration code
410 for (Loop::block_iterator I = currentLoop->block_begin(),
411 E = currentLoop->block_end();
414 // Do not include empty blocks in the cost calculation. This happen due to
415 // loop canonicalization and will be removed.
416 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
419 // Count basic blocks.
426 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
427 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
428 /// unswitch the loop, reprocess the pieces, then return true.
429 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
432 Function *F = loopHeader->getParent();
434 // Do not unswitch if the function is optimized for size.
435 if (!F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
438 // Check to see if it would be profitable to unswitch current loop.
439 unsigned Cost = getLoopUnswitchCost(LoopCond);
441 // Do not do non-trivial unswitch while optimizing for size.
442 if (Cost && OptimizeForSize)
445 if (Cost > Threshold) {
446 // FIXME: this should estimate growth by the amount of code shared by the
447 // resultant unswitched loops.
449 DOUT << "NOT unswitching loop %"
450 << currentLoop->getHeader()->getName() << ", cost too high: "
451 << currentLoop->getBlocks().size() << "\n";
456 BasicBlock *ExitBlock;
457 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
458 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
460 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
466 // RemapInstruction - Convert the instruction operands from referencing the
467 // current values into those specified by ValueMap.
469 static inline void RemapInstruction(Instruction *I,
470 DenseMap<const Value *, Value*> &ValueMap) {
471 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
472 Value *Op = I->getOperand(op);
473 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
474 if (It != ValueMap.end()) Op = It->second;
475 I->setOperand(op, Op);
479 /// CloneLoop - Recursively clone the specified loop and all of its children,
480 /// mapping the blocks with the specified map.
481 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
482 LoopInfo *LI, LPPassManager *LPM) {
483 Loop *New = new Loop();
485 LPM->insertLoop(New, PL);
487 // Add all of the blocks in L to the new loop.
488 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
490 if (LI->getLoopFor(*I) == L)
491 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
493 // Add all of the subloops to the new loop.
494 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
495 CloneLoop(*I, New, VM, LI, LPM);
500 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
501 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
502 /// code immediately before InsertPt.
503 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
504 BasicBlock *TrueDest,
505 BasicBlock *FalseDest,
506 Instruction *InsertPt) {
507 // Insert a conditional branch on LIC to the two preheaders. The original
508 // code is the true version and the new code is the false version.
509 Value *BranchVal = LIC;
510 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
511 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
512 else if (Val != ConstantInt::getTrue())
513 // We want to enter the new loop when the condition is true.
514 std::swap(TrueDest, FalseDest);
516 // Insert the new branch.
517 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
520 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
521 /// condition in it (a cond branch from its header block to its latch block,
522 /// where the path through the loop that doesn't execute its body has no
523 /// side-effects), unswitch it. This doesn't involve any code duplication, just
524 /// moving the conditional branch outside of the loop and updating loop info.
525 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
527 BasicBlock *ExitBlock) {
528 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
529 << loopHeader->getName() << " [" << L->getBlocks().size()
530 << " blocks] in Function " << L->getHeader()->getParent()->getName()
531 << " on cond: " << *Val << " == " << *Cond << "\n";
533 // First step, split the preheader, so that we know that there is a safe place
534 // to insert the conditional branch. We will change loopPreheader to have a
535 // conditional branch on Cond.
536 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
538 // Now that we have a place to insert the conditional branch, create a place
539 // to branch to: this is the exit block out of the loop that we should
542 // Split this block now, so that the loop maintains its exit block, and so
543 // that the jump from the preheader can execute the contents of the exit block
544 // without actually branching to it (the exit block should be dominated by the
545 // loop header, not the preheader).
546 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
547 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
549 // Okay, now we have a position to branch from and a position to branch to,
550 // insert the new conditional branch.
551 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
552 loopPreheader->getTerminator());
553 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
554 loopPreheader->getTerminator()->eraseFromParent();
556 // We need to reprocess this loop, it could be unswitched again.
559 // Now that we know that the loop is never entered when this condition is a
560 // particular value, rewrite the loop with this info. We know that this will
561 // at least eliminate the old branch.
562 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
566 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
567 /// blocks. Update the appropriate Phi nodes as we do so.
568 void LoopUnswitch::SplitExitEdges(Loop *L,
569 const SmallVector<BasicBlock *, 8> &ExitBlocks)
572 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
573 BasicBlock *ExitBlock = ExitBlocks[i];
574 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
576 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
577 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
578 BasicBlock* StartBlock = Preds[j];
579 BasicBlock* EndBlock;
580 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
581 EndBlock = NewExitBlock;
582 NewExitBlock = EndBlock->getSinglePredecessor();;
584 EndBlock = ExitBlock;
587 std::set<PHINode*> InsertedPHIs;
588 PHINode* OldLCSSA = 0;
589 for (BasicBlock::iterator I = EndBlock->begin();
590 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
591 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
592 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
593 OldLCSSA->getName() + ".us-lcssa",
594 NewExitBlock->getTerminator());
595 NewLCSSA->addIncoming(OldValue, StartBlock);
596 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
598 InsertedPHIs.insert(NewLCSSA);
601 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
602 for (BasicBlock::iterator I = NewExitBlock->begin();
603 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
605 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
606 OldLCSSA->getName() + ".us-lcssa",
608 OldLCSSA->replaceAllUsesWith(NewLCSSA);
609 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
617 /// UnswitchNontrivialCondition - We determined that the loop is profitable
618 /// to unswitch when LIC equal Val. Split it into loop versions and test the
619 /// condition outside of either loop. Return the loops created as Out1/Out2.
620 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
622 Function *F = loopHeader->getParent();
623 DOUT << "loop-unswitch: Unswitching loop %"
624 << loopHeader->getName() << " [" << L->getBlocks().size()
625 << " blocks] in Function " << F->getName()
626 << " when '" << *Val << "' == " << *LIC << "\n";
631 // First step, split the preheader and exit blocks, and add these blocks to
632 // the LoopBlocks list.
633 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
634 LoopBlocks.push_back(NewPreheader);
636 // We want the loop to come after the preheader, but before the exit blocks.
637 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
639 SmallVector<BasicBlock*, 8> ExitBlocks;
640 L->getUniqueExitBlocks(ExitBlocks);
642 // Split all of the edges from inside the loop to their exit blocks. Update
643 // the appropriate Phi nodes as we do so.
644 SplitExitEdges(L, ExitBlocks);
646 // The exit blocks may have been changed due to edge splitting, recompute.
648 L->getUniqueExitBlocks(ExitBlocks);
650 // Add exit blocks to the loop blocks.
651 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
653 // Next step, clone all of the basic blocks that make up the loop (including
654 // the loop preheader and exit blocks), keeping track of the mapping between
655 // the instructions and blocks.
656 NewBlocks.reserve(LoopBlocks.size());
657 DenseMap<const Value*, Value*> ValueMap;
658 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
659 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
660 NewBlocks.push_back(New);
661 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
662 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
665 // Splice the newly inserted blocks into the function right before the
666 // original preheader.
667 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
668 NewBlocks[0], F->end());
670 // Now we create the new Loop object for the versioned loop.
671 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
672 Loop *ParentLoop = L->getParentLoop();
674 // Make sure to add the cloned preheader and exit blocks to the parent loop
676 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
679 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
680 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
681 // The new exit block should be in the same loop as the old one.
682 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
683 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
685 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
686 "Exit block should have been split to have one successor!");
687 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
689 // If the successor of the exit block had PHI nodes, add an entry for
692 for (BasicBlock::iterator I = ExitSucc->begin();
693 (PN = dyn_cast<PHINode>(I)); ++I) {
694 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
695 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
696 if (It != ValueMap.end()) V = It->second;
697 PN->addIncoming(V, NewExit);
701 // Rewrite the code to refer to itself.
702 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
703 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
704 E = NewBlocks[i]->end(); I != E; ++I)
705 RemapInstruction(I, ValueMap);
707 // Rewrite the original preheader to select between versions of the loop.
708 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
709 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
710 "Preheader splitting did not work correctly!");
712 // Emit the new branch that selects between the two versions of this loop.
713 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
714 LPM->deleteSimpleAnalysisValue(OldBR, L);
715 OldBR->eraseFromParent();
717 LoopProcessWorklist.push_back(NewLoop);
720 // Now we rewrite the original code to know that the condition is true and the
721 // new code to know that the condition is false.
722 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
724 // It's possible that simplifying one loop could cause the other to be
725 // deleted. If so, don't simplify it.
726 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
727 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
731 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
733 static void RemoveFromWorklist(Instruction *I,
734 std::vector<Instruction*> &Worklist) {
735 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
737 while (WI != Worklist.end()) {
738 unsigned Offset = WI-Worklist.begin();
740 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
744 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
745 /// program, replacing all uses with V and update the worklist.
746 static void ReplaceUsesOfWith(Instruction *I, Value *V,
747 std::vector<Instruction*> &Worklist,
748 Loop *L, LPPassManager *LPM) {
749 DOUT << "Replace with '" << *V << "': " << *I;
751 // Add uses to the worklist, which may be dead now.
752 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
753 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
754 Worklist.push_back(Use);
756 // Add users to the worklist which may be simplified now.
757 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
759 Worklist.push_back(cast<Instruction>(*UI));
760 LPM->deleteSimpleAnalysisValue(I, L);
761 RemoveFromWorklist(I, Worklist);
762 I->replaceAllUsesWith(V);
763 I->eraseFromParent();
767 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
768 /// information, and remove any dead successors it has.
770 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
771 std::vector<Instruction*> &Worklist,
773 if (pred_begin(BB) != pred_end(BB)) {
774 // This block isn't dead, since an edge to BB was just removed, see if there
775 // are any easy simplifications we can do now.
776 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
777 // If it has one pred, fold phi nodes in BB.
778 while (isa<PHINode>(BB->begin()))
779 ReplaceUsesOfWith(BB->begin(),
780 cast<PHINode>(BB->begin())->getIncomingValue(0),
783 // If this is the header of a loop and the only pred is the latch, we now
784 // have an unreachable loop.
785 if (Loop *L = LI->getLoopFor(BB))
786 if (loopHeader == BB && L->contains(Pred)) {
787 // Remove the branch from the latch to the header block, this makes
788 // the header dead, which will make the latch dead (because the header
789 // dominates the latch).
790 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
791 Pred->getTerminator()->eraseFromParent();
792 new UnreachableInst(Pred);
794 // The loop is now broken, remove it from LI.
795 RemoveLoopFromHierarchy(L);
797 // Reprocess the header, which now IS dead.
798 RemoveBlockIfDead(BB, Worklist, L);
802 // If pred ends in a uncond branch, add uncond branch to worklist so that
803 // the two blocks will get merged.
804 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
805 if (BI->isUnconditional())
806 Worklist.push_back(BI);
811 DOUT << "Nuking dead block: " << *BB;
813 // Remove the instructions in the basic block from the worklist.
814 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
815 RemoveFromWorklist(I, Worklist);
817 // Anything that uses the instructions in this basic block should have their
818 // uses replaced with undefs.
820 I->replaceAllUsesWith(UndefValue::get(I->getType()));
823 // If this is the edge to the header block for a loop, remove the loop and
824 // promote all subloops.
825 if (Loop *BBLoop = LI->getLoopFor(BB)) {
826 if (BBLoop->getLoopLatch() == BB)
827 RemoveLoopFromHierarchy(BBLoop);
830 // Remove the block from the loop info, which removes it from any loops it
835 // Remove phi node entries in successors for this block.
836 TerminatorInst *TI = BB->getTerminator();
837 std::vector<BasicBlock*> Succs;
838 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
839 Succs.push_back(TI->getSuccessor(i));
840 TI->getSuccessor(i)->removePredecessor(BB);
843 // Unique the successors, remove anything with multiple uses.
844 std::sort(Succs.begin(), Succs.end());
845 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
847 // Remove the basic block, including all of the instructions contained in it.
848 LPM->deleteSimpleAnalysisValue(BB, L);
849 BB->eraseFromParent();
850 // Remove successor blocks here that are not dead, so that we know we only
851 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
852 // then getting removed before we revisit them, which is badness.
854 for (unsigned i = 0; i != Succs.size(); ++i)
855 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
856 // One exception is loop headers. If this block was the preheader for a
857 // loop, then we DO want to visit the loop so the loop gets deleted.
858 // We know that if the successor is a loop header, that this loop had to
859 // be the preheader: the case where this was the latch block was handled
860 // above and headers can only have two predecessors.
861 if (!LI->isLoopHeader(Succs[i])) {
862 Succs.erase(Succs.begin()+i);
867 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
868 RemoveBlockIfDead(Succs[i], Worklist, L);
871 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
872 /// become unwrapped, either because the backedge was deleted, or because the
873 /// edge into the header was removed. If the edge into the header from the
874 /// latch block was removed, the loop is unwrapped but subloops are still alive,
875 /// so they just reparent loops. If the loops are actually dead, they will be
877 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
878 LPM->deleteLoopFromQueue(L);
879 RemoveLoopFromWorklist(L);
882 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
883 // the value specified by Val in the specified loop, or we know it does NOT have
884 // that value. Rewrite any uses of LIC or of properties correlated to it.
885 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
888 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
890 // FIXME: Support correlated properties, like:
897 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
898 // selects, switches.
899 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
900 std::vector<Instruction*> Worklist;
902 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
903 // in the loop with the appropriate one directly.
904 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
909 Replacement = ConstantInt::get(Type::Int1Ty,
910 !cast<ConstantInt>(Val)->getZExtValue());
912 for (unsigned i = 0, e = Users.size(); i != e; ++i)
913 if (Instruction *U = cast<Instruction>(Users[i])) {
914 if (!L->contains(U->getParent()))
916 U->replaceUsesOfWith(LIC, Replacement);
917 Worklist.push_back(U);
920 // Otherwise, we don't know the precise value of LIC, but we do know that it
921 // is certainly NOT "Val". As such, simplify any uses in the loop that we
922 // can. This case occurs when we unswitch switch statements.
923 for (unsigned i = 0, e = Users.size(); i != e; ++i)
924 if (Instruction *U = cast<Instruction>(Users[i])) {
925 if (!L->contains(U->getParent()))
928 Worklist.push_back(U);
930 // If we know that LIC is not Val, use this info to simplify code.
931 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
932 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
933 if (SI->getCaseValue(i) == Val) {
934 // Found a dead case value. Don't remove PHI nodes in the
935 // successor if they become single-entry, those PHI nodes may
936 // be in the Users list.
938 // FIXME: This is a hack. We need to keep the successor around
939 // and hooked up so as to preserve the loop structure, because
940 // trying to update it is complicated. So instead we preserve the
941 // loop structure and put the block on an dead code path.
943 BasicBlock *SISucc = SI->getSuccessor(i);
944 BasicBlock* Old = SI->getParent();
945 BasicBlock* Split = SplitBlock(Old, SI, this);
947 Instruction* OldTerm = Old->getTerminator();
948 BranchInst::Create(Split, SISucc,
949 ConstantInt::getTrue(), OldTerm);
951 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
952 Old->getTerminator()->eraseFromParent();
955 for (BasicBlock::iterator II = SISucc->begin();
956 (PN = dyn_cast<PHINode>(II)); ++II) {
957 Value *InVal = PN->removeIncomingValue(Split, false);
958 PN->addIncoming(InVal, Old);
967 // TODO: We could do other simplifications, for example, turning
968 // LIC == Val -> false.
972 SimplifyCode(Worklist, L);
975 /// SimplifyCode - Okay, now that we have simplified some instructions in the
976 /// loop, walk over it and constant prop, dce, and fold control flow where
977 /// possible. Note that this is effectively a very simple loop-structure-aware
978 /// optimizer. During processing of this loop, L could very well be deleted, so
979 /// it must not be used.
981 /// FIXME: When the loop optimizer is more mature, separate this out to a new
984 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
985 while (!Worklist.empty()) {
986 Instruction *I = Worklist.back();
989 // Simple constant folding.
990 if (Constant *C = ConstantFoldInstruction(I)) {
991 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
996 if (isInstructionTriviallyDead(I)) {
997 DOUT << "Remove dead instruction '" << *I;
999 // Add uses to the worklist, which may be dead now.
1000 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1001 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1002 Worklist.push_back(Use);
1003 LPM->deleteSimpleAnalysisValue(I, L);
1004 RemoveFromWorklist(I, Worklist);
1005 I->eraseFromParent();
1010 // Special case hacks that appear commonly in unswitched code.
1011 switch (I->getOpcode()) {
1012 case Instruction::Select:
1013 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1014 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1019 case Instruction::And:
1020 if (isa<ConstantInt>(I->getOperand(0)) &&
1021 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1022 cast<BinaryOperator>(I)->swapOperands();
1023 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1024 if (CB->getType() == Type::Int1Ty) {
1025 if (CB->isOne()) // X & 1 -> X
1026 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1028 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1032 case Instruction::Or:
1033 if (isa<ConstantInt>(I->getOperand(0)) &&
1034 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1035 cast<BinaryOperator>(I)->swapOperands();
1036 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1037 if (CB->getType() == Type::Int1Ty) {
1038 if (CB->isOne()) // X | 1 -> 1
1039 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1041 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1045 case Instruction::Br: {
1046 BranchInst *BI = cast<BranchInst>(I);
1047 if (BI->isUnconditional()) {
1048 // If BI's parent is the only pred of the successor, fold the two blocks
1050 BasicBlock *Pred = BI->getParent();
1051 BasicBlock *Succ = BI->getSuccessor(0);
1052 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1053 if (!SinglePred) continue; // Nothing to do.
1054 assert(SinglePred == Pred && "CFG broken");
1056 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1057 << Succ->getName() << "\n";
1059 // Resolve any single entry PHI nodes in Succ.
1060 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1061 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1063 // Move all of the successor contents from Succ to Pred.
1064 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1066 LPM->deleteSimpleAnalysisValue(BI, L);
1067 BI->eraseFromParent();
1068 RemoveFromWorklist(BI, Worklist);
1070 // If Succ has any successors with PHI nodes, update them to have
1071 // entries coming from Pred instead of Succ.
1072 Succ->replaceAllUsesWith(Pred);
1074 // Remove Succ from the loop tree.
1075 LI->removeBlock(Succ);
1076 LPM->deleteSimpleAnalysisValue(Succ, L);
1077 Succ->eraseFromParent();
1079 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1080 // Conditional branch. Turn it into an unconditional branch, then
1081 // remove dead blocks.
1082 break; // FIXME: Enable.
1084 DOUT << "Folded branch: " << *BI;
1085 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1086 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1087 DeadSucc->removePredecessor(BI->getParent(), true);
1088 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1089 LPM->deleteSimpleAnalysisValue(BI, L);
1090 BI->eraseFromParent();
1091 RemoveFromWorklist(BI, Worklist);
1094 RemoveBlockIfDead(DeadSucc, Worklist, L);