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).
172 // Hoist simple values out.
173 if (L->makeLoopInvariant(Cond, Changed))
176 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
177 if (BO->getOpcode() == Instruction::And ||
178 BO->getOpcode() == Instruction::Or) {
179 // If either the left or right side is invariant, we can unswitch on this,
180 // which will cause the branch to go away in one loop and the condition to
181 // simplify in the other one.
182 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
184 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
191 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
192 LI = &getAnalysis<LoopInfo>();
194 DF = getAnalysisIfAvailable<DominanceFrontier>();
195 DT = getAnalysisIfAvailable<DominatorTree>();
197 Function *F = currentLoop->getHeader()->getParent();
198 bool Changed = false;
200 assert(currentLoop->isLCSSAForm());
202 Changed |= processCurrentLoop();
206 // FIXME: Reconstruct dom info, because it is not preserved properly.
208 DT->runOnFunction(*F);
210 DF->runOnFunction(*F);
215 /// processCurrentLoop - Do actual work and unswitch loop if possible
217 bool LoopUnswitch::processCurrentLoop() {
218 bool Changed = false;
220 // Loop over all of the basic blocks in the loop. If we find an interior
221 // block that is branching on a loop-invariant condition, we can unswitch this
223 for (Loop::block_iterator I = currentLoop->block_begin(),
224 E = currentLoop->block_end();
226 TerminatorInst *TI = (*I)->getTerminator();
227 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
228 // If this isn't branching on an invariant condition, we can't unswitch
230 if (BI->isConditional()) {
231 // See if this, or some part of it, is loop invariant. If so, we can
232 // unswitch on it if we desire.
233 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
234 currentLoop, Changed);
235 if (LoopCond && UnswitchIfProfitable(LoopCond,
236 Context->getTrue())) {
241 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
242 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
243 currentLoop, Changed);
244 if (LoopCond && SI->getNumCases() > 1) {
245 // Find a value to unswitch on:
246 // FIXME: this should chose the most expensive case!
247 Constant *UnswitchVal = SI->getCaseValue(1);
248 // Do not process same value again and again.
249 if (!UnswitchedVals.insert(UnswitchVal))
252 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
259 // Scan the instructions to check for unswitchable values.
260 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
262 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
263 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
264 currentLoop, Changed);
265 if (LoopCond && UnswitchIfProfitable(LoopCond,
266 Context->getTrue())) {
275 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
276 /// 1. Exit the loop with no side effects.
277 /// 2. Branch to the latch block with no side-effects.
279 /// If these conditions are true, we return true and set ExitBB to the block we
282 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
284 std::set<BasicBlock*> &Visited) {
285 if (!Visited.insert(BB).second) {
286 // Already visited and Ok, end of recursion.
288 } else if (!L->contains(BB)) {
289 // Otherwise, this is a loop exit, this is fine so long as this is the
291 if (ExitBB != 0) return false;
296 // Otherwise, this is an unvisited intra-loop node. Check all successors.
297 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
298 // Check to see if the successor is a trivial loop exit.
299 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
303 // Okay, everything after this looks good, check to make sure that this block
304 // doesn't include any side effects.
305 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
306 if (I->mayHaveSideEffects())
312 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
313 /// leads to an exit from the specified loop, and has no side-effects in the
314 /// process. If so, return the block that is exited to, otherwise return null.
315 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
316 std::set<BasicBlock*> Visited;
317 Visited.insert(L->getHeader()); // Branches to header are ok.
318 BasicBlock *ExitBB = 0;
319 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
324 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
325 /// trivial: that is, that the condition controls whether or not the loop does
326 /// anything at all. If this is a trivial condition, unswitching produces no
327 /// code duplications (equivalently, it produces a simpler loop and a new empty
328 /// loop, which gets deleted).
330 /// If this is a trivial condition, return true, otherwise return false. When
331 /// returning true, this sets Cond and Val to the condition that controls the
332 /// trivial condition: when Cond dynamically equals Val, the loop is known to
333 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
336 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
337 BasicBlock **LoopExit) {
338 BasicBlock *Header = currentLoop->getHeader();
339 TerminatorInst *HeaderTerm = Header->getTerminator();
341 BasicBlock *LoopExitBB = 0;
342 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
343 // If the header block doesn't end with a conditional branch on Cond, we
345 if (!BI->isConditional() || BI->getCondition() != Cond)
348 // Check to see if a successor of the branch is guaranteed to go to the
349 // latch block or exit through a one exit block without having any
350 // side-effects. If so, determine the value of Cond that causes it to do
352 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
353 BI->getSuccessor(0)))) {
354 if (Val) *Val = Context->getTrue();
355 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
356 BI->getSuccessor(1)))) {
357 if (Val) *Val = Context->getFalse();
359 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
360 // If this isn't a switch on Cond, we can't handle it.
361 if (SI->getCondition() != Cond) return false;
363 // Check to see if a successor of the switch is guaranteed to go to the
364 // latch block or exit through a one exit block without having any
365 // side-effects. If so, determine the value of Cond that causes it to do
366 // this. Note that we can't trivially unswitch on the default case.
367 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
368 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
369 SI->getSuccessor(i)))) {
370 // Okay, we found a trivial case, remember the value that is trivial.
371 if (Val) *Val = SI->getCaseValue(i);
376 // If we didn't find a single unique LoopExit block, or if the loop exit block
377 // contains phi nodes, this isn't trivial.
378 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
379 return false; // Can't handle this.
381 if (LoopExit) *LoopExit = LoopExitBB;
383 // We already know that nothing uses any scalar values defined inside of this
384 // loop. As such, we just have to check to see if this loop will execute any
385 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
386 // part of the loop that the code *would* execute. We already checked the
387 // tail, check the header now.
388 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
389 if (I->mayHaveSideEffects())
394 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
395 /// we choose to unswitch current loop on the specified value.
397 unsigned LoopUnswitch::getLoopUnswitchCost(Value *LIC) {
398 // If the condition is trivial, always unswitch. There is no code growth for
400 if (IsTrivialUnswitchCondition(LIC))
403 // FIXME: This is really overly conservative. However, more liberal
404 // estimations have thus far resulted in excessive unswitching, which is bad
405 // both in compile time and in code size. This should be replaced once
406 // someone figures out how a good estimation.
407 return currentLoop->getBlocks().size();
410 // FIXME: this is brain dead. It should take into consideration code
412 for (Loop::block_iterator I = currentLoop->block_begin(),
413 E = currentLoop->block_end();
416 // Do not include empty blocks in the cost calculation. This happen due to
417 // loop canonicalization and will be removed.
418 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
421 // Count basic blocks.
428 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
429 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
430 /// unswitch the loop, reprocess the pieces, then return true.
431 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
434 Function *F = loopHeader->getParent();
437 // Check to see if it would be profitable to unswitch current loop.
438 unsigned Cost = getLoopUnswitchCost(LoopCond);
440 // Do not do non-trivial unswitch while optimizing for size.
441 if (Cost && OptimizeForSize)
443 if (Cost && !F->isDeclaration() && F->hasFnAttr(Attribute::OptimizeForSize))
446 if (Cost > Threshold) {
447 // FIXME: this should estimate growth by the amount of code shared by the
448 // resultant unswitched loops.
450 DOUT << "NOT unswitching loop %"
451 << currentLoop->getHeader()->getName() << ", cost too high: "
452 << currentLoop->getBlocks().size() << "\n";
457 BasicBlock *ExitBlock;
458 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
459 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
461 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
467 // RemapInstruction - Convert the instruction operands from referencing the
468 // current values into those specified by ValueMap.
470 static inline void RemapInstruction(Instruction *I,
471 DenseMap<const Value *, Value*> &ValueMap) {
472 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
473 Value *Op = I->getOperand(op);
474 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
475 if (It != ValueMap.end()) Op = It->second;
476 I->setOperand(op, Op);
480 /// CloneLoop - Recursively clone the specified loop and all of its children,
481 /// mapping the blocks with the specified map.
482 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
483 LoopInfo *LI, LPPassManager *LPM) {
484 Loop *New = new Loop();
486 LPM->insertLoop(New, PL);
488 // Add all of the blocks in L to the new loop.
489 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
491 if (LI->getLoopFor(*I) == L)
492 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
494 // Add all of the subloops to the new loop.
495 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
496 CloneLoop(*I, New, VM, LI, LPM);
501 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
502 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
503 /// code immediately before InsertPt.
504 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
505 BasicBlock *TrueDest,
506 BasicBlock *FalseDest,
507 Instruction *InsertPt) {
508 // Insert a conditional branch on LIC to the two preheaders. The original
509 // code is the true version and the new code is the false version.
510 Value *BranchVal = LIC;
511 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
512 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
513 else if (Val != Context->getTrue())
514 // We want to enter the new loop when the condition is true.
515 std::swap(TrueDest, FalseDest);
517 // Insert the new branch.
518 BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
521 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
522 /// condition in it (a cond branch from its header block to its latch block,
523 /// where the path through the loop that doesn't execute its body has no
524 /// side-effects), unswitch it. This doesn't involve any code duplication, just
525 /// moving the conditional branch outside of the loop and updating loop info.
526 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
528 BasicBlock *ExitBlock) {
529 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
530 << loopHeader->getName() << " [" << L->getBlocks().size()
531 << " blocks] in Function " << L->getHeader()->getParent()->getName()
532 << " on cond: " << *Val << " == " << *Cond << "\n";
534 // First step, split the preheader, so that we know that there is a safe place
535 // to insert the conditional branch. We will change loopPreheader to have a
536 // conditional branch on Cond.
537 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
539 // Now that we have a place to insert the conditional branch, create a place
540 // to branch to: this is the exit block out of the loop that we should
543 // Split this block now, so that the loop maintains its exit block, and so
544 // that the jump from the preheader can execute the contents of the exit block
545 // without actually branching to it (the exit block should be dominated by the
546 // loop header, not the preheader).
547 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
548 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
550 // Okay, now we have a position to branch from and a position to branch to,
551 // insert the new conditional branch.
552 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
553 loopPreheader->getTerminator());
554 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
555 loopPreheader->getTerminator()->eraseFromParent();
557 // We need to reprocess this loop, it could be unswitched again.
560 // Now that we know that the loop is never entered when this condition is a
561 // particular value, rewrite the loop with this info. We know that this will
562 // at least eliminate the old branch.
563 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
567 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
568 /// blocks. Update the appropriate Phi nodes as we do so.
569 void LoopUnswitch::SplitExitEdges(Loop *L,
570 const SmallVector<BasicBlock *, 8> &ExitBlocks)
573 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
574 BasicBlock *ExitBlock = ExitBlocks[i];
575 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
577 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
578 BasicBlock* NewExitBlock = SplitEdge(Preds[j], ExitBlock, this);
579 BasicBlock* StartBlock = Preds[j];
580 BasicBlock* EndBlock;
581 if (NewExitBlock->getSinglePredecessor() == ExitBlock) {
582 EndBlock = NewExitBlock;
583 NewExitBlock = EndBlock->getSinglePredecessor();
585 EndBlock = ExitBlock;
588 std::set<PHINode*> InsertedPHIs;
589 PHINode* OldLCSSA = 0;
590 for (BasicBlock::iterator I = EndBlock->begin();
591 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
592 Value* OldValue = OldLCSSA->getIncomingValueForBlock(NewExitBlock);
593 PHINode* NewLCSSA = PHINode::Create(OldLCSSA->getType(),
594 OldLCSSA->getName() + ".us-lcssa",
595 NewExitBlock->getTerminator());
596 NewLCSSA->addIncoming(OldValue, StartBlock);
597 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(NewExitBlock),
599 InsertedPHIs.insert(NewLCSSA);
602 BasicBlock::iterator InsertPt = EndBlock->getFirstNonPHI();
603 for (BasicBlock::iterator I = NewExitBlock->begin();
604 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
606 PHINode *NewLCSSA = PHINode::Create(OldLCSSA->getType(),
607 OldLCSSA->getName() + ".us-lcssa",
609 OldLCSSA->replaceAllUsesWith(NewLCSSA);
610 NewLCSSA->addIncoming(OldLCSSA, NewExitBlock);
618 /// UnswitchNontrivialCondition - We determined that the loop is profitable
619 /// to unswitch when LIC equal Val. Split it into loop versions and test the
620 /// condition outside of either loop. Return the loops created as Out1/Out2.
621 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
623 Function *F = loopHeader->getParent();
624 DOUT << "loop-unswitch: Unswitching loop %"
625 << loopHeader->getName() << " [" << L->getBlocks().size()
626 << " blocks] in Function " << F->getName()
627 << " when '" << *Val << "' == " << *LIC << "\n";
632 // First step, split the preheader and exit blocks, and add these blocks to
633 // the LoopBlocks list.
634 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
635 LoopBlocks.push_back(NewPreheader);
637 // We want the loop to come after the preheader, but before the exit blocks.
638 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
640 SmallVector<BasicBlock*, 8> ExitBlocks;
641 L->getUniqueExitBlocks(ExitBlocks);
643 // Split all of the edges from inside the loop to their exit blocks. Update
644 // the appropriate Phi nodes as we do so.
645 SplitExitEdges(L, ExitBlocks);
647 // The exit blocks may have been changed due to edge splitting, recompute.
649 L->getUniqueExitBlocks(ExitBlocks);
651 // Add exit blocks to the loop blocks.
652 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
654 // Next step, clone all of the basic blocks that make up the loop (including
655 // the loop preheader and exit blocks), keeping track of the mapping between
656 // the instructions and blocks.
657 NewBlocks.reserve(LoopBlocks.size());
658 DenseMap<const Value*, Value*> ValueMap;
659 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
660 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
661 NewBlocks.push_back(New);
662 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
663 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
666 // Splice the newly inserted blocks into the function right before the
667 // original preheader.
668 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
669 NewBlocks[0], F->end());
671 // Now we create the new Loop object for the versioned loop.
672 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
673 Loop *ParentLoop = L->getParentLoop();
675 // Make sure to add the cloned preheader and exit blocks to the parent loop
677 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
680 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
681 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
682 // The new exit block should be in the same loop as the old one.
683 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
684 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
686 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
687 "Exit block should have been split to have one successor!");
688 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
690 // If the successor of the exit block had PHI nodes, add an entry for
693 for (BasicBlock::iterator I = ExitSucc->begin();
694 (PN = dyn_cast<PHINode>(I)); ++I) {
695 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
696 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
697 if (It != ValueMap.end()) V = It->second;
698 PN->addIncoming(V, NewExit);
702 // Rewrite the code to refer to itself.
703 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
704 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
705 E = NewBlocks[i]->end(); I != E; ++I)
706 RemapInstruction(I, ValueMap);
708 // Rewrite the original preheader to select between versions of the loop.
709 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
710 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
711 "Preheader splitting did not work correctly!");
713 // Emit the new branch that selects between the two versions of this loop.
714 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
715 LPM->deleteSimpleAnalysisValue(OldBR, L);
716 OldBR->eraseFromParent();
718 LoopProcessWorklist.push_back(NewLoop);
721 // Now we rewrite the original code to know that the condition is true and the
722 // new code to know that the condition is false.
723 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
725 // It's possible that simplifying one loop could cause the other to be
726 // deleted. If so, don't simplify it.
727 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
728 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
732 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
734 static void RemoveFromWorklist(Instruction *I,
735 std::vector<Instruction*> &Worklist) {
736 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
738 while (WI != Worklist.end()) {
739 unsigned Offset = WI-Worklist.begin();
741 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
745 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
746 /// program, replacing all uses with V and update the worklist.
747 static void ReplaceUsesOfWith(Instruction *I, Value *V,
748 std::vector<Instruction*> &Worklist,
749 Loop *L, LPPassManager *LPM) {
750 DOUT << "Replace with '" << *V << "': " << *I;
752 // Add uses to the worklist, which may be dead now.
753 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
754 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
755 Worklist.push_back(Use);
757 // Add users to the worklist which may be simplified now.
758 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
760 Worklist.push_back(cast<Instruction>(*UI));
761 LPM->deleteSimpleAnalysisValue(I, L);
762 RemoveFromWorklist(I, Worklist);
763 I->replaceAllUsesWith(V);
764 I->eraseFromParent();
768 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
769 /// information, and remove any dead successors it has.
771 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
772 std::vector<Instruction*> &Worklist,
774 if (pred_begin(BB) != pred_end(BB)) {
775 // This block isn't dead, since an edge to BB was just removed, see if there
776 // are any easy simplifications we can do now.
777 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
778 // If it has one pred, fold phi nodes in BB.
779 while (isa<PHINode>(BB->begin()))
780 ReplaceUsesOfWith(BB->begin(),
781 cast<PHINode>(BB->begin())->getIncomingValue(0),
784 // If this is the header of a loop and the only pred is the latch, we now
785 // have an unreachable loop.
786 if (Loop *L = LI->getLoopFor(BB))
787 if (loopHeader == BB && L->contains(Pred)) {
788 // Remove the branch from the latch to the header block, this makes
789 // the header dead, which will make the latch dead (because the header
790 // dominates the latch).
791 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
792 Pred->getTerminator()->eraseFromParent();
793 new UnreachableInst(Pred);
795 // The loop is now broken, remove it from LI.
796 RemoveLoopFromHierarchy(L);
798 // Reprocess the header, which now IS dead.
799 RemoveBlockIfDead(BB, Worklist, L);
803 // If pred ends in a uncond branch, add uncond branch to worklist so that
804 // the two blocks will get merged.
805 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
806 if (BI->isUnconditional())
807 Worklist.push_back(BI);
812 DOUT << "Nuking dead block: " << *BB;
814 // Remove the instructions in the basic block from the worklist.
815 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
816 RemoveFromWorklist(I, Worklist);
818 // Anything that uses the instructions in this basic block should have their
819 // uses replaced with undefs.
821 I->replaceAllUsesWith(Context->getUndef(I->getType()));
824 // If this is the edge to the header block for a loop, remove the loop and
825 // promote all subloops.
826 if (Loop *BBLoop = LI->getLoopFor(BB)) {
827 if (BBLoop->getLoopLatch() == BB)
828 RemoveLoopFromHierarchy(BBLoop);
831 // Remove the block from the loop info, which removes it from any loops it
836 // Remove phi node entries in successors for this block.
837 TerminatorInst *TI = BB->getTerminator();
838 SmallVector<BasicBlock*, 4> Succs;
839 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
840 Succs.push_back(TI->getSuccessor(i));
841 TI->getSuccessor(i)->removePredecessor(BB);
844 // Unique the successors, remove anything with multiple uses.
845 array_pod_sort(Succs.begin(), Succs.end());
846 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
848 // Remove the basic block, including all of the instructions contained in it.
849 LPM->deleteSimpleAnalysisValue(BB, L);
850 BB->eraseFromParent();
851 // Remove successor blocks here that are not dead, so that we know we only
852 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
853 // then getting removed before we revisit them, which is badness.
855 for (unsigned i = 0; i != Succs.size(); ++i)
856 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
857 // One exception is loop headers. If this block was the preheader for a
858 // loop, then we DO want to visit the loop so the loop gets deleted.
859 // We know that if the successor is a loop header, that this loop had to
860 // be the preheader: the case where this was the latch block was handled
861 // above and headers can only have two predecessors.
862 if (!LI->isLoopHeader(Succs[i])) {
863 Succs.erase(Succs.begin()+i);
868 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
869 RemoveBlockIfDead(Succs[i], Worklist, L);
872 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
873 /// become unwrapped, either because the backedge was deleted, or because the
874 /// edge into the header was removed. If the edge into the header from the
875 /// latch block was removed, the loop is unwrapped but subloops are still alive,
876 /// so they just reparent loops. If the loops are actually dead, they will be
878 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
879 LPM->deleteLoopFromQueue(L);
880 RemoveLoopFromWorklist(L);
883 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
884 // the value specified by Val in the specified loop, or we know it does NOT have
885 // that value. Rewrite any uses of LIC or of properties correlated to it.
886 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
889 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
891 // FIXME: Support correlated properties, like:
898 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
899 // selects, switches.
900 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
901 std::vector<Instruction*> Worklist;
903 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
904 // in the loop with the appropriate one directly.
905 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
910 Replacement = Context->getConstantInt(Type::Int1Ty,
911 !cast<ConstantInt>(Val)->getZExtValue());
913 for (unsigned i = 0, e = Users.size(); i != e; ++i)
914 if (Instruction *U = cast<Instruction>(Users[i])) {
915 if (!L->contains(U->getParent()))
917 U->replaceUsesOfWith(LIC, Replacement);
918 Worklist.push_back(U);
921 // Otherwise, we don't know the precise value of LIC, but we do know that it
922 // is certainly NOT "Val". As such, simplify any uses in the loop that we
923 // can. This case occurs when we unswitch switch statements.
924 for (unsigned i = 0, e = Users.size(); i != e; ++i)
925 if (Instruction *U = cast<Instruction>(Users[i])) {
926 if (!L->contains(U->getParent()))
929 Worklist.push_back(U);
931 // If we know that LIC is not Val, use this info to simplify code.
932 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
933 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
934 if (SI->getCaseValue(i) == Val) {
935 // Found a dead case value. Don't remove PHI nodes in the
936 // successor if they become single-entry, those PHI nodes may
937 // be in the Users list.
939 // FIXME: This is a hack. We need to keep the successor around
940 // and hooked up so as to preserve the loop structure, because
941 // trying to update it is complicated. So instead we preserve the
942 // loop structure and put the block on an dead code path.
944 BasicBlock *SISucc = SI->getSuccessor(i);
945 BasicBlock* Old = SI->getParent();
946 BasicBlock* Split = SplitBlock(Old, SI, this);
948 Instruction* OldTerm = Old->getTerminator();
949 BranchInst::Create(Split, SISucc,
950 Context->getTrue(), OldTerm);
952 LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L);
953 Old->getTerminator()->eraseFromParent();
956 for (BasicBlock::iterator II = SISucc->begin();
957 (PN = dyn_cast<PHINode>(II)); ++II) {
958 Value *InVal = PN->removeIncomingValue(Split, false);
959 PN->addIncoming(InVal, Old);
968 // TODO: We could do other simplifications, for example, turning
969 // LIC == Val -> false.
973 SimplifyCode(Worklist, L);
976 /// SimplifyCode - Okay, now that we have simplified some instructions in the
977 /// loop, walk over it and constant prop, dce, and fold control flow where
978 /// possible. Note that this is effectively a very simple loop-structure-aware
979 /// optimizer. During processing of this loop, L could very well be deleted, so
980 /// it must not be used.
982 /// FIXME: When the loop optimizer is more mature, separate this out to a new
985 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
986 while (!Worklist.empty()) {
987 Instruction *I = Worklist.back();
990 // Simple constant folding.
991 if (Constant *C = ConstantFoldInstruction(I, Context)) {
992 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
997 if (isInstructionTriviallyDead(I)) {
998 DOUT << "Remove dead instruction '" << *I;
1000 // Add uses to the worklist, which may be dead now.
1001 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1002 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1003 Worklist.push_back(Use);
1004 LPM->deleteSimpleAnalysisValue(I, L);
1005 RemoveFromWorklist(I, Worklist);
1006 I->eraseFromParent();
1011 // Special case hacks that appear commonly in unswitched code.
1012 switch (I->getOpcode()) {
1013 case Instruction::Select:
1014 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1015 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
1020 case Instruction::And:
1021 if (isa<ConstantInt>(I->getOperand(0)) &&
1022 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1023 cast<BinaryOperator>(I)->swapOperands();
1024 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1025 if (CB->getType() == Type::Int1Ty) {
1026 if (CB->isOne()) // X & 1 -> X
1027 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1029 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1033 case Instruction::Or:
1034 if (isa<ConstantInt>(I->getOperand(0)) &&
1035 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1036 cast<BinaryOperator>(I)->swapOperands();
1037 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1038 if (CB->getType() == Type::Int1Ty) {
1039 if (CB->isOne()) // X | 1 -> 1
1040 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1042 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1046 case Instruction::Br: {
1047 BranchInst *BI = cast<BranchInst>(I);
1048 if (BI->isUnconditional()) {
1049 // If BI's parent is the only pred of the successor, fold the two blocks
1051 BasicBlock *Pred = BI->getParent();
1052 BasicBlock *Succ = BI->getSuccessor(0);
1053 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1054 if (!SinglePred) continue; // Nothing to do.
1055 assert(SinglePred == Pred && "CFG broken");
1057 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1058 << Succ->getName() << "\n";
1060 // Resolve any single entry PHI nodes in Succ.
1061 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1062 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1064 // Move all of the successor contents from Succ to Pred.
1065 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1067 LPM->deleteSimpleAnalysisValue(BI, L);
1068 BI->eraseFromParent();
1069 RemoveFromWorklist(BI, Worklist);
1071 // If Succ has any successors with PHI nodes, update them to have
1072 // entries coming from Pred instead of Succ.
1073 Succ->replaceAllUsesWith(Pred);
1075 // Remove Succ from the loop tree.
1076 LI->removeBlock(Succ);
1077 LPM->deleteSimpleAnalysisValue(Succ, L);
1078 Succ->eraseFromParent();
1080 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1081 // Conditional branch. Turn it into an unconditional branch, then
1082 // remove dead blocks.
1083 break; // FIXME: Enable.
1085 DOUT << "Folded branch: " << *BI;
1086 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1087 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1088 DeadSucc->removePredecessor(BI->getParent(), true);
1089 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1090 LPM->deleteSimpleAnalysisValue(BI, L);
1091 BI->eraseFromParent();
1092 RemoveFromWorklist(BI, Worklist);
1095 RemoveBlockIfDead(DeadSucc, Worklist, L);