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/LLVMContext.h"
36 #include "llvm/Analysis/ConstantFolding.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/LoopPass.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Transforms/Utils/Cloning.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/SmallPtrSet.h"
45 #include "llvm/ADT/STLExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.h"
53 STATISTIC(NumBranches, "Number of branches unswitched");
54 STATISTIC(NumSwitches, "Number of switches unswitched");
55 STATISTIC(NumSelects , "Number of selects unswitched");
56 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
57 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
59 static cl::opt<unsigned>
60 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
61 cl::init(10), cl::Hidden);
64 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
65 LoopInfo *LI; // Loop information
68 // LoopProcessWorklist - Used to check if second loop needs processing
69 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
70 std::vector<Loop*> LoopProcessWorklist;
71 SmallPtrSet<Value *,8> UnswitchedVals;
77 DominanceFrontier *DF;
79 BasicBlock *loopHeader;
80 BasicBlock *loopPreheader;
82 // LoopBlocks contains all of the basic blocks of the loop, including the
83 // preheader of the loop, the body of the loop, and the exit blocks of the
84 // loop, in that order.
85 std::vector<BasicBlock*> LoopBlocks;
86 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
87 std::vector<BasicBlock*> NewBlocks;
90 static char ID; // Pass ID, replacement for typeid
91 explicit LoopUnswitch(bool Os = false) :
92 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
93 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
94 loopPreheader(NULL) {}
96 bool runOnLoop(Loop *L, LPPassManager &LPM);
97 bool processCurrentLoop();
99 /// This transformation requires natural loop information & requires that
100 /// loop preheaders be inserted into the CFG...
102 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
103 AU.addRequiredID(LoopSimplifyID);
104 AU.addPreservedID(LoopSimplifyID);
105 AU.addRequired<LoopInfo>();
106 AU.addPreserved<LoopInfo>();
107 AU.addRequiredID(LCSSAID);
108 AU.addPreservedID(LCSSAID);
109 AU.addPreserved<DominatorTree>();
110 AU.addPreserved<DominanceFrontier>();
115 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
117 void RemoveLoopFromWorklist(Loop *L) {
118 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
119 LoopProcessWorklist.end(), L);
120 if (I != LoopProcessWorklist.end())
121 LoopProcessWorklist.erase(I);
124 void initLoopData() {
125 loopHeader = currentLoop->getHeader();
126 loopPreheader = currentLoop->getLoopPreheader();
129 /// Split all of the edges from inside the loop to their exit blocks.
130 /// Update the appropriate Phi nodes as we do so.
131 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
133 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
134 unsigned getLoopUnswitchCost(Value *LIC);
135 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
136 BasicBlock *ExitBlock);
137 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
139 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
140 Constant *Val, bool isEqual);
142 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
143 BasicBlock *TrueDest,
144 BasicBlock *FalseDest,
145 Instruction *InsertPt);
147 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
148 void RemoveBlockIfDead(BasicBlock *BB,
149 std::vector<Instruction*> &Worklist, Loop *l);
150 void RemoveLoopFromHierarchy(Loop *L);
151 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
152 BasicBlock **LoopExit = 0);
156 char LoopUnswitch::ID = 0;
157 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
159 Pass *llvm::createLoopUnswitchPass(bool Os) {
160 return new LoopUnswitch(Os);
163 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
164 /// invariant in the loop, or has an invariant piece, return the invariant.
165 /// Otherwise, return null.
166 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
167 // Constants should be folded, not unswitched on!
168 if (isa<Constant>(Cond)) return 0;
170 // TODO: Handle: br (VARIANT|INVARIANT).
171 if (L->isLoopInvariant(Cond)) return Cond;
173 // Hoist simple values out.
174 if (L->makeLoopInvariant(Cond)) {
179 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
180 if (BO->getOpcode() == Instruction::And ||
181 BO->getOpcode() == Instruction::Or) {
182 // If either the left or right side is invariant, we can unswitch on this,
183 // which will cause the branch to go away in one loop and the condition to
184 // simplify in the other one.
185 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
187 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
194 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
195 LI = &getAnalysis<LoopInfo>();
197 DF = getAnalysisIfAvailable<DominanceFrontier>();
198 DT = getAnalysisIfAvailable<DominatorTree>();
200 Function *F = currentLoop->getHeader()->getParent();
201 bool Changed = false;
203 assert(currentLoop->isLCSSAForm());
205 Changed |= processCurrentLoop();
209 // FIXME: Reconstruct dom info, because it is not preserved properly.
211 DT->runOnFunction(*F);
213 DF->runOnFunction(*F);
218 /// processCurrentLoop - Do actual work and unswitch loop if possible
220 bool LoopUnswitch::processCurrentLoop() {
221 bool Changed = false;
223 // Loop over all of the basic blocks in the loop. If we find an interior
224 // block that is branching on a loop-invariant condition, we can unswitch this
226 for (Loop::block_iterator I = currentLoop->block_begin(),
227 E = currentLoop->block_end();
229 TerminatorInst *TI = (*I)->getTerminator();
230 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
231 // If this isn't branching on an invariant condition, we can't unswitch
233 if (BI->isConditional()) {
234 // See if this, or some part of it, is loop invariant. If so, we can
235 // unswitch on it if we desire.
236 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
237 currentLoop, Changed);
238 if (LoopCond && UnswitchIfProfitable(LoopCond,
239 Context->getConstantIntTrue())) {
244 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
245 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
246 currentLoop, Changed);
247 if (LoopCond && SI->getNumCases() > 1) {
248 // Find a value to unswitch on:
249 // FIXME: this should chose the most expensive case!
250 Constant *UnswitchVal = SI->getCaseValue(1);
251 // Do not process same value again and again.
252 if (!UnswitchedVals.insert(UnswitchVal))
255 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
262 // Scan the instructions to check for unswitchable values.
263 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
265 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
266 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
267 currentLoop, Changed);
268 if (LoopCond && UnswitchIfProfitable(LoopCond,
269 Context->getConstantIntTrue())) {
278 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
279 /// 1. Exit the loop with no side effects.
280 /// 2. Branch to the latch block with no side-effects.
282 /// If these conditions are true, we return true and set ExitBB to the block we
285 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
287 std::set<BasicBlock*> &Visited) {
288 if (!Visited.insert(BB).second) {
289 // Already visited and Ok, end of recursion.
291 } else if (!L->contains(BB)) {
292 // Otherwise, this is a loop exit, this is fine so long as this is the
294 if (ExitBB != 0) return false;
299 // Otherwise, this is an unvisited intra-loop node. Check all successors.
300 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
301 // Check to see if the successor is a trivial loop exit.
302 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
306 // Okay, everything after this looks good, check to make sure that this block
307 // doesn't include any side effects.
308 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
309 if (I->mayHaveSideEffects())
315 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
316 /// leads to an exit from the specified loop, and has no side-effects in the
317 /// process. If so, return the block that is exited to, otherwise return null.
318 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
319 std::set<BasicBlock*> Visited;
320 Visited.insert(L->getHeader()); // Branches to header are ok.
321 BasicBlock *ExitBB = 0;
322 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
327 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
328 /// trivial: that is, that the condition controls whether or not the loop does
329 /// anything at all. If this is a trivial condition, unswitching produces no
330 /// code duplications (equivalently, it produces a simpler loop and a new empty
331 /// loop, which gets deleted).
333 /// If this is a trivial condition, return true, otherwise return false. When
334 /// returning true, this sets Cond and Val to the condition that controls the
335 /// trivial condition: when Cond dynamically equals Val, the loop is known to
336 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
339 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
340 BasicBlock **LoopExit) {
341 BasicBlock *Header = currentLoop->getHeader();
342 TerminatorInst *HeaderTerm = Header->getTerminator();
344 BasicBlock *LoopExitBB = 0;
345 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
346 // If the header block doesn't end with a conditional branch on Cond, we
348 if (!BI->isConditional() || BI->getCondition() != Cond)
351 // Check to see if a successor of the branch is guaranteed to go to the
352 // latch block or exit through a one exit block without having any
353 // side-effects. If so, determine the value of Cond that causes it to do
355 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
356 BI->getSuccessor(0)))) {
357 if (Val) *Val = Context->getConstantIntTrue();
358 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
359 BI->getSuccessor(1)))) {
360 if (Val) *Val = Context->getConstantIntFalse();
362 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
363 // If this isn't a switch on Cond, we can't handle it.
364 if (SI->getCondition() != Cond) return false;
366 // Check to see if a successor of the switch is guaranteed to go to the
367 // latch block or exit through a one exit block without having any
368 // side-effects. If so, determine the value of Cond that causes it to do
369 // this. Note that we can't trivially unswitch on the default case.
370 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
371 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
372 SI->getSuccessor(i)))) {
373 // Okay, we found a trivial case, remember the value that is trivial.
374 if (Val) *Val = SI->getCaseValue(i);
379 // If we didn't find a single unique LoopExit block, or if the loop exit block
380 // contains phi nodes, this isn't trivial.
381 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
382 return false; // Can't handle this.
384 if (LoopExit) *LoopExit = LoopExitBB;
386 // We already know that nothing uses any scalar values defined inside of this
387 // loop. As such, we just have to check to see if this loop will execute any
388 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
389 // part of the loop that the code *would* execute. We already checked the
390 // tail, check the header now.
391 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
392 if (I->mayHaveSideEffects())
397 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
398 /// we choose to unswitch current loop on the specified value.
400 unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
401 // If the condition is trivial, always unswitch. There is no code growth for
403 if (IsTrivialUnswitchCondition(LIC))
406 // FIXME: This is really overly conservative. However, more liberal
407 // estimations have thus far resulted in excessive unswitching, which is bad
408 // both in compile time and in code size. This should be replaced once
409 // someone figures out how a good estimation.
410 return currentLoop->getBlocks().size();
413 // FIXME: this is brain dead. It should take into consideration code
415 for (Loop::block_iterator I = currentLoop->block_begin(),
416 E = currentLoop->block_end();
419 // Do not include empty blocks in the cost calculation. This happen due to
420 // loop canonicalization and will be removed.
421 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
424 // Count basic blocks.
431 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
432 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
433 /// unswitch the loop, reprocess the pieces, then return true.
434 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
437 Function *F = loopHeader->getParent();
440 // Check to see if it would be profitable to unswitch current loop.
441 unsigned Cost = getLoopUnswitchCost(LoopCond);
443 // Do not do non-trivial unswitch while optimizing for size.
444 if (Cost && OptimizeForSize)
446 if (Cost && !F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
449 if (Cost > Threshold) {
450 // FIXME: this should estimate growth by the amount of code shared by the
451 // resultant unswitched loops.
453 DOUT << "NOT unswitching loop %"
454 << currentLoop->getHeader()->getName() << ", cost too high: "
455 << currentLoop->getBlocks().size() << "\n";
460 BasicBlock *ExitBlock;
461 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
462 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
464 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
470 // RemapInstruction - Convert the instruction operands from referencing the
471 // current values into those specified by ValueMap.
473 static inline void RemapInstruction(Instruction *I,
474 DenseMap<const Value *, Value*> &ValueMap) {
475 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
476 Value *Op = I->getOperand(op);
477 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
478 if (It != ValueMap.end()) Op = It->second;
479 I->setOperand(op, Op);
483 /// CloneLoop - Recursively clone the specified loop and all of its children,
484 /// mapping the blocks with the specified map.
485 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
486 LoopInfo *LI, LPPassManager *LPM) {
487 Loop *New = new Loop();
489 LPM->insertLoop(New, PL);
491 // Add all of the blocks in L to the new loop.
492 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
494 if (LI->getLoopFor(*I) == L)
495 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
497 // Add all of the subloops to the new loop.
498 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
499 CloneLoop(*I, New, VM, LI, LPM);
504 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
505 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
506 /// code immediately before InsertPt.
507 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
508 BasicBlock *TrueDest,
509 BasicBlock *FalseDest,
510 Instruction *InsertPt) {
511 // Insert a conditional branch on LIC to the two preheaders. The original
512 // code is the true version and the new code is the false version.
513 Value *BranchVal = LIC;
514 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
515 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
516 else if (Val != Context->getConstantIntTrue())
517 // We want to enter the new loop when the condition is true.
518 std::swap(TrueDest, FalseDest);
520 // Insert the new branch.
521 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
524 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
525 /// condition in it (a cond branch from its header block to its latch block,
526 /// where the path through the loop that doesn't execute its body has no
527 /// side-effects), unswitch it. This doesn't involve any code duplication, just
528 /// moving the conditional branch outside of the loop and updating loop info.
529 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
531 BasicBlock *ExitBlock) {
532 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
533 << loopHeader->getName() << " [" << L->getBlocks().size()
534 << " blocks] in Function " << L->getHeader()->getParent()->getName()
535 << " on cond: " << *Val << " == " << *Cond << "\n";
537 // First step, split the preheader, so that we know that there is a safe place
538 // to insert the conditional branch. We will change loopPreheader to have a
539 // conditional branch on Cond.
540 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
542 // Now that we have a place to insert the conditional branch, create a place
543 // to branch to: this is the exit block out of the loop that we should
546 // Split this block now, so that the loop maintains its exit block, and so
547 // that the jump from the preheader can execute the contents of the exit block
548 // without actually branching to it (the exit block should be dominated by the
549 // loop header, not the preheader).
550 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
551 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
553 // Okay, now we have a position to branch from and a position to branch to,
554 // insert the new conditional branch.
555 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
556 loopPreheader->getTerminator());
557 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
558 loopPreheader->getTerminator()->eraseFromParent();
560 // We need to reprocess this loop, it could be unswitched again.
563 // Now that we know that the loop is never entered when this condition is a
564 // particular value, rewrite the loop with this info. We know that this will
565 // at least eliminate the old branch.
566 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
570 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
571 /// blocks. Update the appropriate Phi nodes as we do so.
572 void LoopUnswitch::SplitExitEdges(Loop *L,
573 const SmallVector<BasicBlock *, 8> &ExitBlocks)
576 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
577 BasicBlock *ExitBlock = ExitBlocks[i];
578 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
580 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
581 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
582 BasicBlock* StartBlock = Preds[j];
583 BasicBlock* EndBlock;
584 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
585 EndBlock = NewExitBlock;
586 NewExitBlock = EndBlock->getSinglePredecessor();
588 EndBlock = ExitBlock;
591 std::set<PHINode*> InsertedPHIs;
592 PHINode* OldLCSSA = 0;
593 for (BasicBlock::iterator I = EndBlock->begin();
594 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
595 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
596 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
597 OldLCSSA->getName() + ".us-lcssa",
598 NewExitBlock->getTerminator());
599 NewLCSSA->addIncoming(OldValue, StartBlock);
600 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
602 InsertedPHIs.insert(NewLCSSA);
605 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
606 for (BasicBlock::iterator I = NewExitBlock->begin();
607 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
609 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
610 OldLCSSA->getName() + ".us-lcssa",
612 OldLCSSA->replaceAllUsesWith(NewLCSSA);
613 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
621 /// UnswitchNontrivialCondition - We determined that the loop is profitable
622 /// to unswitch when LIC equal Val. Split it into loop versions and test the
623 /// condition outside of either loop. Return the loops created as Out1/Out2.
624 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
626 Function *F = loopHeader->getParent();
627 DOUT << "loop-unswitch: Unswitching loop %"
628 << loopHeader->getName() << " [" << L->getBlocks().size()
629 << " blocks] in Function " << F->getName()
630 << " when '" << *Val << "' == " << *LIC << "\n";
635 // First step, split the preheader and exit blocks, and add these blocks to
636 // the LoopBlocks list.
637 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
638 LoopBlocks.push_back(NewPreheader);
640 // We want the loop to come after the preheader, but before the exit blocks.
641 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
643 SmallVector<BasicBlock*, 8> ExitBlocks;
644 L->getUniqueExitBlocks(ExitBlocks);
646 // Split all of the edges from inside the loop to their exit blocks. Update
647 // the appropriate Phi nodes as we do so.
648 SplitExitEdges(L, ExitBlocks);
650 // The exit blocks may have been changed due to edge splitting, recompute.
652 L->getUniqueExitBlocks(ExitBlocks);
654 // Add exit blocks to the loop blocks.
655 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
657 // Next step, clone all of the basic blocks that make up the loop (including
658 // the loop preheader and exit blocks), keeping track of the mapping between
659 // the instructions and blocks.
660 NewBlocks.reserve(LoopBlocks.size());
661 DenseMap<const Value*, Value*> ValueMap;
662 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
663 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
664 NewBlocks.push_back(New);
665 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
666 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
669 // Splice the newly inserted blocks into the function right before the
670 // original preheader.
671 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
672 NewBlocks[0], F->end());
674 // Now we create the new Loop object for the versioned loop.
675 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
676 Loop *ParentLoop = L->getParentLoop();
678 // Make sure to add the cloned preheader and exit blocks to the parent loop
680 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
683 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
684 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
685 // The new exit block should be in the same loop as the old one.
686 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
687 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
689 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
690 "Exit block should have been split to have one successor!");
691 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
693 // If the successor of the exit block had PHI nodes, add an entry for
696 for (BasicBlock::iterator I = ExitSucc->begin();
697 (PN = dyn_cast<PHINode>(I)); ++I) {
698 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
699 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
700 if (It != ValueMap.end()) V = It->second;
701 PN->addIncoming(V, NewExit);
705 // Rewrite the code to refer to itself.
706 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
707 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
708 E = NewBlocks[i]->end(); I != E; ++I)
709 RemapInstruction(I, ValueMap);
711 // Rewrite the original preheader to select between versions of the loop.
712 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
713 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
714 "Preheader splitting did not work correctly!");
716 // Emit the new branch that selects between the two versions of this loop.
717 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
718 LPM->deleteSimpleAnalysisValue(OldBR, L);
719 OldBR->eraseFromParent();
721 LoopProcessWorklist.push_back(NewLoop);
724 // Now we rewrite the original code to know that the condition is true and the
725 // new code to know that the condition is false.
726 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
728 // It's possible that simplifying one loop could cause the other to be
729 // deleted. If so, don't simplify it.
730 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
731 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
735 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
737 static void RemoveFromWorklist(Instruction *I,
738 std::vector<Instruction*> &Worklist) {
739 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
741 while (WI != Worklist.end()) {
742 unsigned Offset = WI-Worklist.begin();
744 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
748 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
749 /// program, replacing all uses with V and update the worklist.
750 static void ReplaceUsesOfWith(Instruction *I, Value *V,
751 std::vector<Instruction*> &Worklist,
752 Loop *L, LPPassManager *LPM) {
753 DOUT << "Replace with '" << *V << "': " << *I;
755 // Add uses to the worklist, which may be dead now.
756 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
757 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
758 Worklist.push_back(Use);
760 // Add users to the worklist which may be simplified now.
761 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
763 Worklist.push_back(cast<Instruction>(*UI));
764 LPM->deleteSimpleAnalysisValue(I, L);
765 RemoveFromWorklist(I, Worklist);
766 I->replaceAllUsesWith(V);
767 I->eraseFromParent();
771 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
772 /// information, and remove any dead successors it has.
774 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
775 std::vector<Instruction*> &Worklist,
777 if (pred_begin(BB) != pred_end(BB)) {
778 // This block isn't dead, since an edge to BB was just removed, see if there
779 // are any easy simplifications we can do now.
780 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
781 // If it has one pred, fold phi nodes in BB.
782 while (isa<PHINode>(BB->begin()))
783 ReplaceUsesOfWith(BB->begin(),
784 cast<PHINode>(BB->begin())->getIncomingValue(0),
787 // If this is the header of a loop and the only pred is the latch, we now
788 // have an unreachable loop.
789 if (Loop *L = LI->getLoopFor(BB))
790 if (loopHeader == BB && L->contains(Pred)) {
791 // Remove the branch from the latch to the header block, this makes
792 // the header dead, which will make the latch dead (because the header
793 // dominates the latch).
794 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
795 Pred->getTerminator()->eraseFromParent();
796 new UnreachableInst(Pred);
798 // The loop is now broken, remove it from LI.
799 RemoveLoopFromHierarchy(L);
801 // Reprocess the header, which now IS dead.
802 RemoveBlockIfDead(BB, Worklist, L);
806 // If pred ends in a uncond branch, add uncond branch to worklist so that
807 // the two blocks will get merged.
808 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
809 if (BI->isUnconditional())
810 Worklist.push_back(BI);
815 DOUT << "Nuking dead block: " << *BB;
817 // Remove the instructions in the basic block from the worklist.
818 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
819 RemoveFromWorklist(I, Worklist);
821 // Anything that uses the instructions in this basic block should have their
822 // uses replaced with undefs.
824 I->replaceAllUsesWith(Context->getUndef(I->getType()));
827 // If this is the edge to the header block for a loop, remove the loop and
828 // promote all subloops.
829 if (Loop *BBLoop = LI->getLoopFor(BB)) {
830 if (BBLoop->getLoopLatch() == BB)
831 RemoveLoopFromHierarchy(BBLoop);
834 // Remove the block from the loop info, which removes it from any loops it
839 // Remove phi node entries in successors for this block.
840 TerminatorInst *TI = BB->getTerminator();
841 SmallVector<BasicBlock*, 4> Succs;
842 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
843 Succs.push_back(TI->getSuccessor(i));
844 TI->getSuccessor(i)->removePredecessor(BB);
847 // Unique the successors, remove anything with multiple uses.
848 array_pod_sort(Succs.begin(), Succs.end());
849 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
851 // Remove the basic block, including all of the instructions contained in it.
852 LPM->deleteSimpleAnalysisValue(BB, L);
853 BB->eraseFromParent();
854 // Remove successor blocks here that are not dead, so that we know we only
855 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
856 // then getting removed before we revisit them, which is badness.
858 for (unsigned i = 0; i != Succs.size(); ++i)
859 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
860 // One exception is loop headers. If this block was the preheader for a
861 // loop, then we DO want to visit the loop so the loop gets deleted.
862 // We know that if the successor is a loop header, that this loop had to
863 // be the preheader: the case where this was the latch block was handled
864 // above and headers can only have two predecessors.
865 if (!LI->isLoopHeader(Succs[i])) {
866 Succs.erase(Succs.begin()+i);
871 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
872 RemoveBlockIfDead(Succs[i], Worklist, L);
875 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
876 /// become unwrapped, either because the backedge was deleted, or because the
877 /// edge into the header was removed. If the edge into the header from the
878 /// latch block was removed, the loop is unwrapped but subloops are still alive,
879 /// so they just reparent loops. If the loops are actually dead, they will be
881 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
882 LPM->deleteLoopFromQueue(L);
883 RemoveLoopFromWorklist(L);
886 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
887 // the value specified by Val in the specified loop, or we know it does NOT have
888 // that value. Rewrite any uses of LIC or of properties correlated to it.
889 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
892 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
894 // FIXME: Support correlated properties, like:
901 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
902 // selects, switches.
903 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
904 std::vector<Instruction*> Worklist;
906 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
907 // in the loop with the appropriate one directly.
908 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
913 Replacement = Context->getConstantInt(Type::Int1Ty,
914 !cast<ConstantInt>(Val)->getZExtValue());
916 for (unsigned i = 0, e = Users.size(); i != e; ++i)
917 if (Instruction *U = cast<Instruction>(Users[i])) {
918 if (!L->contains(U->getParent()))
920 U->replaceUsesOfWith(LIC, Replacement);
921 Worklist.push_back(U);
924 // Otherwise, we don't know the precise value of LIC, but we do know that it
925 // is certainly NOT "Val". As such, simplify any uses in the loop that we
926 // can. This case occurs when we unswitch switch statements.
927 for (unsigned i = 0, e = Users.size(); i != e; ++i)
928 if (Instruction *U = cast<Instruction>(Users[i])) {
929 if (!L->contains(U->getParent()))
932 Worklist.push_back(U);
934 // If we know that LIC is not Val, use this info to simplify code.
935 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
936 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
937 if (SI->getCaseValue(i) == Val) {
938 // Found a dead case value. Don't remove PHI nodes in the
939 // successor if they become single-entry, those PHI nodes may
940 // be in the Users list.
942 // FIXME: This is a hack. We need to keep the successor around
943 // and hooked up so as to preserve the loop structure, because
944 // trying to update it is complicated. So instead we preserve the
945 // loop structure and put the block on an dead code path.
947 BasicBlock *SISucc = SI->getSuccessor(i);
948 BasicBlock* Old = SI->getParent();
949 BasicBlock* Split = SplitBlock(Old, SI, this);
951 Instruction* OldTerm = Old->getTerminator();
952 BranchInst::Create(Split, SISucc,
953 Context->getConstantIntTrue(), OldTerm);
955 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
956 Old->getTerminator()->eraseFromParent();
959 for (BasicBlock::iterator II = SISucc->begin();
960 (PN = dyn_cast<PHINode>(II)); ++II) {
961 Value *InVal = PN->removeIncomingValue(Split, false);
962 PN->addIncoming(InVal, Old);
971 // TODO: We could do other simplifications, for example, turning
972 // LIC == Val -> false.
976 SimplifyCode(Worklist, L);
979 /// SimplifyCode - Okay, now that we have simplified some instructions in the
980 /// loop, walk over it and constant prop, dce, and fold control flow where
981 /// possible. Note that this is effectively a very simple loop-structure-aware
982 /// optimizer. During processing of this loop, L could very well be deleted, so
983 /// it must not be used.
985 /// FIXME: When the loop optimizer is more mature, separate this out to a new
988 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
989 while (!Worklist.empty()) {
990 Instruction *I = Worklist.back();
993 // Simple constant folding.
994 if (Constant *C = ConstantFoldInstruction(I, Context)) {
995 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
1000 if (isInstructionTriviallyDead(I)) {
1001 DOUT << "Remove dead instruction '" << *I;
1003 // Add uses to the worklist, which may be dead now.
1004 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1005 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1006 Worklist.push_back(Use);
1007 LPM->deleteSimpleAnalysisValue(I, L);
1008 RemoveFromWorklist(I, Worklist);
1009 I->eraseFromParent();
1014 // Special case hacks that appear commonly in unswitched code.
1015 switch (I->getOpcode()) {
1016 case Instruction::Select:
1017 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1018 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1023 case Instruction::And:
1024 if (isa<ConstantInt>(I->getOperand(0)) &&
1025 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1026 cast<BinaryOperator>(I)->swapOperands();
1027 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1028 if (CB->getType() == Type::Int1Ty) {
1029 if (CB->isOne()) // X & 1 -> X
1030 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1032 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1036 case Instruction::Or:
1037 if (isa<ConstantInt>(I->getOperand(0)) &&
1038 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1039 cast<BinaryOperator>(I)->swapOperands();
1040 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1041 if (CB->getType() == Type::Int1Ty) {
1042 if (CB->isOne()) // X | 1 -> 1
1043 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1045 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1049 case Instruction::Br: {
1050 BranchInst *BI = cast<BranchInst>(I);
1051 if (BI->isUnconditional()) {
1052 // If BI's parent is the only pred of the successor, fold the two blocks
1054 BasicBlock *Pred = BI->getParent();
1055 BasicBlock *Succ = BI->getSuccessor(0);
1056 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1057 if (!SinglePred) continue; // Nothing to do.
1058 assert(SinglePred == Pred && "CFG broken");
1060 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1061 << Succ->getName() << "\n";
1063 // Resolve any single entry PHI nodes in Succ.
1064 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1065 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1067 // Move all of the successor contents from Succ to Pred.
1068 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1070 LPM->deleteSimpleAnalysisValue(BI, L);
1071 BI->eraseFromParent();
1072 RemoveFromWorklist(BI, Worklist);
1074 // If Succ has any successors with PHI nodes, update them to have
1075 // entries coming from Pred instead of Succ.
1076 Succ->replaceAllUsesWith(Pred);
1078 // Remove Succ from the loop tree.
1079 LI->removeBlock(Succ);
1080 LPM->deleteSimpleAnalysisValue(Succ, L);
1081 Succ->eraseFromParent();
1083 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1084 // Conditional branch. Turn it into an unconditional branch, then
1085 // remove dead blocks.
1086 break; // FIXME: Enable.
1088 DOUT << "Folded branch: " << *BI;
1089 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1090 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1091 DeadSucc->removePredecessor(BI->getParent(), true);
1092 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1093 LPM->deleteSimpleAnalysisValue(BI, L);
1094 BI->eraseFromParent();
1095 RemoveFromWorklist(BI, Worklist);
1098 RemoveBlockIfDead(DeadSucc, Worklist, L);