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/InlineCost.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/Debug.h"
48 #include "llvm/Support/raw_ostream.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 // The specific value of 50 here was chosen based only on intuition and a
60 // few specific examples.
61 static cl::opt<unsigned>
62 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
63 cl::init(50), cl::Hidden);
66 class LoopUnswitch : public LoopPass {
67 LoopInfo *LI; // Loop information
70 // LoopProcessWorklist - Used to check if second loop needs processing
71 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
72 std::vector<Loop*> LoopProcessWorklist;
73 SmallPtrSet<Value *,8> UnswitchedVals;
79 DominanceFrontier *DF;
81 BasicBlock *loopHeader;
82 BasicBlock *loopPreheader;
84 // LoopBlocks contains all of the basic blocks of the loop, including the
85 // preheader of the loop, the body of the loop, and the exit blocks of the
86 // loop, in that order.
87 std::vector<BasicBlock*> LoopBlocks;
88 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
89 std::vector<BasicBlock*> NewBlocks;
92 static char ID; // Pass ID, replacement for typeid
93 explicit LoopUnswitch(bool Os = false) :
94 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
95 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
96 loopPreheader(NULL) {}
98 bool runOnLoop(Loop *L, LPPassManager &LPM);
99 bool processCurrentLoop();
101 /// This transformation requires natural loop information & requires that
102 /// loop preheaders be inserted into the CFG...
104 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
105 AU.addRequiredID(LoopSimplifyID);
106 AU.addPreservedID(LoopSimplifyID);
107 AU.addRequired<LoopInfo>();
108 AU.addPreserved<LoopInfo>();
109 AU.addRequiredID(LCSSAID);
110 AU.addPreservedID(LCSSAID);
111 AU.addPreserved<DominatorTree>();
112 AU.addPreserved<DominanceFrontier>();
117 virtual void releaseMemory() {
118 UnswitchedVals.clear();
121 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
123 void RemoveLoopFromWorklist(Loop *L) {
124 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
125 LoopProcessWorklist.end(), L);
126 if (I != LoopProcessWorklist.end())
127 LoopProcessWorklist.erase(I);
130 void initLoopData() {
131 loopHeader = currentLoop->getHeader();
132 loopPreheader = currentLoop->getLoopPreheader();
135 /// Split all of the edges from inside the loop to their exit blocks.
136 /// Update the appropriate Phi nodes as we do so.
137 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
139 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
140 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
141 BasicBlock *ExitBlock);
142 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
144 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
145 Constant *Val, bool isEqual);
147 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
148 BasicBlock *TrueDest,
149 BasicBlock *FalseDest,
150 Instruction *InsertPt);
152 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
153 void RemoveBlockIfDead(BasicBlock *BB,
154 std::vector<Instruction*> &Worklist, Loop *l);
155 void RemoveLoopFromHierarchy(Loop *L);
156 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
157 BasicBlock **LoopExit = 0);
161 char LoopUnswitch::ID = 0;
162 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
164 Pass *llvm::createLoopUnswitchPass(bool Os) {
165 return new LoopUnswitch(Os);
168 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
169 /// invariant in the loop, or has an invariant piece, return the invariant.
170 /// Otherwise, return null.
171 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
172 // We can never unswitch on vector conditions.
173 if (Cond->getType()->isVectorTy())
176 // Constants should be folded, not unswitched on!
177 if (isa<Constant>(Cond)) return 0;
179 // TODO: Handle: br (VARIANT|INVARIANT).
181 // Hoist simple values out.
182 if (L->makeLoopInvariant(Cond, Changed))
185 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
186 if (BO->getOpcode() == Instruction::And ||
187 BO->getOpcode() == Instruction::Or) {
188 // If either the left or right side is invariant, we can unswitch on this,
189 // which will cause the branch to go away in one loop and the condition to
190 // simplify in the other one.
191 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
193 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
200 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
201 LI = &getAnalysis<LoopInfo>();
203 DF = getAnalysisIfAvailable<DominanceFrontier>();
204 DT = getAnalysisIfAvailable<DominatorTree>();
206 Function *F = currentLoop->getHeader()->getParent();
207 bool Changed = false;
209 assert(currentLoop->isLCSSAForm());
211 Changed |= processCurrentLoop();
215 // FIXME: Reconstruct dom info, because it is not preserved properly.
217 DT->runOnFunction(*F);
219 DF->runOnFunction(*F);
224 /// processCurrentLoop - Do actual work and unswitch loop if possible
226 bool LoopUnswitch::processCurrentLoop() {
227 bool Changed = false;
228 LLVMContext &Context = currentLoop->getHeader()->getContext();
230 // Loop over all of the basic blocks in the loop. If we find an interior
231 // block that is branching on a loop-invariant condition, we can unswitch this
233 for (Loop::block_iterator I = currentLoop->block_begin(),
234 E = currentLoop->block_end();
236 TerminatorInst *TI = (*I)->getTerminator();
237 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
238 // If this isn't branching on an invariant condition, we can't unswitch
240 if (BI->isConditional()) {
241 // See if this, or some part of it, is loop invariant. If so, we can
242 // unswitch on it if we desire.
243 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
244 currentLoop, Changed);
245 if (LoopCond && UnswitchIfProfitable(LoopCond,
246 ConstantInt::getTrue(Context))) {
251 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
252 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
253 currentLoop, Changed);
254 if (LoopCond && SI->getNumCases() > 1) {
255 // Find a value to unswitch on:
256 // FIXME: this should chose the most expensive case!
257 Constant *UnswitchVal = SI->getCaseValue(1);
258 // Do not process same value again and again.
259 if (!UnswitchedVals.insert(UnswitchVal))
262 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
269 // Scan the instructions to check for unswitchable values.
270 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
272 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
273 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
274 currentLoop, Changed);
275 if (LoopCond && UnswitchIfProfitable(LoopCond,
276 ConstantInt::getTrue(Context))) {
285 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
286 /// 1. Exit the loop with no side effects.
287 /// 2. Branch to the latch block with no side-effects.
289 /// If these conditions are true, we return true and set ExitBB to the block we
292 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
294 std::set<BasicBlock*> &Visited) {
295 if (!Visited.insert(BB).second) {
296 // Already visited and Ok, end of recursion.
298 } else if (!L->contains(BB)) {
299 // Otherwise, this is a loop exit, this is fine so long as this is the
301 if (ExitBB != 0) return false;
306 // Otherwise, this is an unvisited intra-loop node. Check all successors.
307 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
308 // Check to see if the successor is a trivial loop exit.
309 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
313 // Okay, everything after this looks good, check to make sure that this block
314 // doesn't include any side effects.
315 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
316 if (I->mayHaveSideEffects())
322 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
323 /// leads to an exit from the specified loop, and has no side-effects in the
324 /// process. If so, return the block that is exited to, otherwise return null.
325 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
326 std::set<BasicBlock*> Visited;
327 Visited.insert(L->getHeader()); // Branches to header are ok.
328 BasicBlock *ExitBB = 0;
329 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
334 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
335 /// trivial: that is, that the condition controls whether or not the loop does
336 /// anything at all. If this is a trivial condition, unswitching produces no
337 /// code duplications (equivalently, it produces a simpler loop and a new empty
338 /// loop, which gets deleted).
340 /// If this is a trivial condition, return true, otherwise return false. When
341 /// returning true, this sets Cond and Val to the condition that controls the
342 /// trivial condition: when Cond dynamically equals Val, the loop is known to
343 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
346 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
347 BasicBlock **LoopExit) {
348 BasicBlock *Header = currentLoop->getHeader();
349 TerminatorInst *HeaderTerm = Header->getTerminator();
350 LLVMContext &Context = Header->getContext();
352 BasicBlock *LoopExitBB = 0;
353 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
354 // If the header block doesn't end with a conditional branch on Cond, we
356 if (!BI->isConditional() || BI->getCondition() != Cond)
359 // Check to see if a successor of the branch is guaranteed to go to the
360 // latch block or exit through a one exit block without having any
361 // side-effects. If so, determine the value of Cond that causes it to do
363 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
364 BI->getSuccessor(0)))) {
365 if (Val) *Val = ConstantInt::getTrue(Context);
366 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
367 BI->getSuccessor(1)))) {
368 if (Val) *Val = ConstantInt::getFalse(Context);
370 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
371 // If this isn't a switch on Cond, we can't handle it.
372 if (SI->getCondition() != Cond) return false;
374 // Check to see if a successor of the switch is guaranteed to go to the
375 // latch block or exit through a one exit block without having any
376 // side-effects. If so, determine the value of Cond that causes it to do
377 // this. Note that we can't trivially unswitch on the default case.
378 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
379 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
380 SI->getSuccessor(i)))) {
381 // Okay, we found a trivial case, remember the value that is trivial.
382 if (Val) *Val = SI->getCaseValue(i);
387 // If we didn't find a single unique LoopExit block, or if the loop exit block
388 // contains phi nodes, this isn't trivial.
389 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
390 return false; // Can't handle this.
392 if (LoopExit) *LoopExit = LoopExitBB;
394 // We already know that nothing uses any scalar values defined inside of this
395 // loop. As such, we just have to check to see if this loop will execute any
396 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
397 // part of the loop that the code *would* execute. We already checked the
398 // tail, check the header now.
399 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
400 if (I->mayHaveSideEffects())
405 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
406 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
407 /// unswitch the loop, reprocess the pieces, then return true.
408 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
412 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
416 Function *F = loopHeader->getParent();
418 // If the condition is trivial, always unswitch. There is no code growth for
420 if (!IsTrivialUnswitchCondition(LoopCond)) {
421 // Check to see if it would be profitable to unswitch current loop.
423 // Do not do non-trivial unswitch while optimizing for size.
424 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
427 // FIXME: This is overly conservative because it does not take into
428 // consideration code simplification opportunities and code that can
429 // be shared by the resultant unswitched loops.
431 for (Loop::block_iterator I = currentLoop->block_begin(),
432 E = currentLoop->block_end();
434 Metrics.analyzeBasicBlock(*I);
436 // Limit the number of instructions to avoid causing significant code
437 // expansion, and the number of basic blocks, to avoid loops with
438 // large numbers of branches which cause loop unswitching to go crazy.
439 // This is a very ad-hoc heuristic.
440 if (Metrics.NumInsts > Threshold ||
441 Metrics.NumBlocks * 5 > Threshold ||
442 Metrics.NeverInline) {
443 DEBUG(dbgs() << "NOT unswitching loop %"
444 << currentLoop->getHeader()->getName() << ", cost too high: "
445 << currentLoop->getBlocks().size() << "\n");
451 BasicBlock *ExitBlock;
452 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
453 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
455 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
461 // RemapInstruction - Convert the instruction operands from referencing the
462 // current values into those specified by ValueMap.
464 static inline void RemapInstruction(Instruction *I,
465 DenseMap<const Value *, Value*> &ValueMap) {
466 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
467 Value *Op = I->getOperand(op);
468 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
469 if (It != ValueMap.end()) Op = It->second;
470 I->setOperand(op, Op);
474 /// CloneLoop - Recursively clone the specified loop and all of its children,
475 /// mapping the blocks with the specified map.
476 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
477 LoopInfo *LI, LPPassManager *LPM) {
478 Loop *New = new Loop();
480 LPM->insertLoop(New, PL);
482 // Add all of the blocks in L to the new loop.
483 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
485 if (LI->getLoopFor(*I) == L)
486 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
488 // Add all of the subloops to the new loop.
489 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
490 CloneLoop(*I, New, VM, LI, LPM);
495 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
496 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
497 /// code immediately before InsertPt.
498 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
499 BasicBlock *TrueDest,
500 BasicBlock *FalseDest,
501 Instruction *InsertPt) {
502 // Insert a conditional branch on LIC to the two preheaders. The original
503 // code is the true version and the new code is the false version.
504 Value *BranchVal = LIC;
505 if (!isa<ConstantInt>(Val) ||
506 Val->getType() != Type::getInt1Ty(LIC->getContext()))
507 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
508 else if (Val != ConstantInt::getTrue(Val->getContext()))
509 // We want to enter the new loop when the condition is true.
510 std::swap(TrueDest, FalseDest);
512 // Insert the new branch.
513 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
515 // If either edge is critical, split it. This helps preserve LoopSimplify
516 // form for enclosing loops.
517 SplitCriticalEdge(BI, 0, this);
518 SplitCriticalEdge(BI, 1, this);
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 DEBUG(dbgs() << "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 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
576 pred_end(ExitBlock));
577 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
582 /// UnswitchNontrivialCondition - We determined that the loop is profitable
583 /// to unswitch when LIC equal Val. Split it into loop versions and test the
584 /// condition outside of either loop. Return the loops created as Out1/Out2.
585 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
587 Function *F = loopHeader->getParent();
588 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
589 << loopHeader->getName() << " [" << L->getBlocks().size()
590 << " blocks] in Function " << F->getName()
591 << " when '" << *Val << "' == " << *LIC << "\n");
596 // First step, split the preheader and exit blocks, and add these blocks to
597 // the LoopBlocks list.
598 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
599 LoopBlocks.push_back(NewPreheader);
601 // We want the loop to come after the preheader, but before the exit blocks.
602 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
604 SmallVector<BasicBlock*, 8> ExitBlocks;
605 L->getUniqueExitBlocks(ExitBlocks);
607 // Split all of the edges from inside the loop to their exit blocks. Update
608 // the appropriate Phi nodes as we do so.
609 SplitExitEdges(L, ExitBlocks);
611 // The exit blocks may have been changed due to edge splitting, recompute.
613 L->getUniqueExitBlocks(ExitBlocks);
615 // Add exit blocks to the loop blocks.
616 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
618 // Next step, clone all of the basic blocks that make up the loop (including
619 // the loop preheader and exit blocks), keeping track of the mapping between
620 // the instructions and blocks.
621 NewBlocks.reserve(LoopBlocks.size());
622 DenseMap<const Value*, Value*> ValueMap;
623 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
624 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
625 NewBlocks.push_back(New);
626 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
627 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
630 // Splice the newly inserted blocks into the function right before the
631 // original preheader.
632 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
633 NewBlocks[0], F->end());
635 // Now we create the new Loop object for the versioned loop.
636 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
637 Loop *ParentLoop = L->getParentLoop();
639 // Make sure to add the cloned preheader and exit blocks to the parent loop
641 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
644 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
645 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
646 // The new exit block should be in the same loop as the old one.
647 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
648 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
650 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
651 "Exit block should have been split to have one successor!");
652 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
654 // If the successor of the exit block had PHI nodes, add an entry for
657 for (BasicBlock::iterator I = ExitSucc->begin();
658 (PN = dyn_cast<PHINode>(I)); ++I) {
659 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
660 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
661 if (It != ValueMap.end()) V = It->second;
662 PN->addIncoming(V, NewExit);
666 // Rewrite the code to refer to itself.
667 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
668 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
669 E = NewBlocks[i]->end(); I != E; ++I)
670 RemapInstruction(I, ValueMap);
672 // Rewrite the original preheader to select between versions of the loop.
673 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
674 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
675 "Preheader splitting did not work correctly!");
677 // Emit the new branch that selects between the two versions of this loop.
678 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
679 LPM->deleteSimpleAnalysisValue(OldBR, L);
680 OldBR->eraseFromParent();
682 LoopProcessWorklist.push_back(NewLoop);
685 // Now we rewrite the original code to know that the condition is true and the
686 // new code to know that the condition is false.
687 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
689 // It's possible that simplifying one loop could cause the other to be
690 // deleted. If so, don't simplify it.
691 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
692 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
696 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
698 static void RemoveFromWorklist(Instruction *I,
699 std::vector<Instruction*> &Worklist) {
700 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
702 while (WI != Worklist.end()) {
703 unsigned Offset = WI-Worklist.begin();
705 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
709 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
710 /// program, replacing all uses with V and update the worklist.
711 static void ReplaceUsesOfWith(Instruction *I, Value *V,
712 std::vector<Instruction*> &Worklist,
713 Loop *L, LPPassManager *LPM) {
714 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
716 // Add uses to the worklist, which may be dead now.
717 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
718 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
719 Worklist.push_back(Use);
721 // Add users to the worklist which may be simplified now.
722 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
724 Worklist.push_back(cast<Instruction>(*UI));
725 LPM->deleteSimpleAnalysisValue(I, L);
726 RemoveFromWorklist(I, Worklist);
727 I->replaceAllUsesWith(V);
728 I->eraseFromParent();
732 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
733 /// information, and remove any dead successors it has.
735 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
736 std::vector<Instruction*> &Worklist,
738 if (pred_begin(BB) != pred_end(BB)) {
739 // This block isn't dead, since an edge to BB was just removed, see if there
740 // are any easy simplifications we can do now.
741 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
742 // If it has one pred, fold phi nodes in BB.
743 while (isa<PHINode>(BB->begin()))
744 ReplaceUsesOfWith(BB->begin(),
745 cast<PHINode>(BB->begin())->getIncomingValue(0),
748 // If this is the header of a loop and the only pred is the latch, we now
749 // have an unreachable loop.
750 if (Loop *L = LI->getLoopFor(BB))
751 if (loopHeader == BB && L->contains(Pred)) {
752 // Remove the branch from the latch to the header block, this makes
753 // the header dead, which will make the latch dead (because the header
754 // dominates the latch).
755 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
756 Pred->getTerminator()->eraseFromParent();
757 new UnreachableInst(BB->getContext(), Pred);
759 // The loop is now broken, remove it from LI.
760 RemoveLoopFromHierarchy(L);
762 // Reprocess the header, which now IS dead.
763 RemoveBlockIfDead(BB, Worklist, L);
767 // If pred ends in a uncond branch, add uncond branch to worklist so that
768 // the two blocks will get merged.
769 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
770 if (BI->isUnconditional())
771 Worklist.push_back(BI);
776 DEBUG(dbgs() << "Nuking dead block: " << *BB);
778 // Remove the instructions in the basic block from the worklist.
779 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
780 RemoveFromWorklist(I, Worklist);
782 // Anything that uses the instructions in this basic block should have their
783 // uses replaced with undefs.
784 // If I is not void type then replaceAllUsesWith undef.
785 // This allows ValueHandlers and custom metadata to adjust itself.
786 if (!I->getType()->isVoidTy())
787 I->replaceAllUsesWith(UndefValue::get(I->getType()));
790 // If this is the edge to the header block for a loop, remove the loop and
791 // promote all subloops.
792 if (Loop *BBLoop = LI->getLoopFor(BB)) {
793 if (BBLoop->getLoopLatch() == BB)
794 RemoveLoopFromHierarchy(BBLoop);
797 // Remove the block from the loop info, which removes it from any loops it
802 // Remove phi node entries in successors for this block.
803 TerminatorInst *TI = BB->getTerminator();
804 SmallVector<BasicBlock*, 4> Succs;
805 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
806 Succs.push_back(TI->getSuccessor(i));
807 TI->getSuccessor(i)->removePredecessor(BB);
810 // Unique the successors, remove anything with multiple uses.
811 array_pod_sort(Succs.begin(), Succs.end());
812 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
814 // Remove the basic block, including all of the instructions contained in it.
815 LPM->deleteSimpleAnalysisValue(BB, L);
816 BB->eraseFromParent();
817 // Remove successor blocks here that are not dead, so that we know we only
818 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
819 // then getting removed before we revisit them, which is badness.
821 for (unsigned i = 0; i != Succs.size(); ++i)
822 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
823 // One exception is loop headers. If this block was the preheader for a
824 // loop, then we DO want to visit the loop so the loop gets deleted.
825 // We know that if the successor is a loop header, that this loop had to
826 // be the preheader: the case where this was the latch block was handled
827 // above and headers can only have two predecessors.
828 if (!LI->isLoopHeader(Succs[i])) {
829 Succs.erase(Succs.begin()+i);
834 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
835 RemoveBlockIfDead(Succs[i], Worklist, L);
838 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
839 /// become unwrapped, either because the backedge was deleted, or because the
840 /// edge into the header was removed. If the edge into the header from the
841 /// latch block was removed, the loop is unwrapped but subloops are still alive,
842 /// so they just reparent loops. If the loops are actually dead, they will be
844 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
845 LPM->deleteLoopFromQueue(L);
846 RemoveLoopFromWorklist(L);
849 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
850 // the value specified by Val in the specified loop, or we know it does NOT have
851 // that value. Rewrite any uses of LIC or of properties correlated to it.
852 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
855 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
857 // FIXME: Support correlated properties, like:
864 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
865 // selects, switches.
866 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
867 std::vector<Instruction*> Worklist;
868 LLVMContext &Context = Val->getContext();
871 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
872 // in the loop with the appropriate one directly.
873 if (IsEqual || (isa<ConstantInt>(Val) &&
874 Val->getType()->isIntegerTy(1))) {
879 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
880 !cast<ConstantInt>(Val)->getZExtValue());
882 for (unsigned i = 0, e = Users.size(); i != e; ++i)
883 if (Instruction *U = cast<Instruction>(Users[i])) {
886 U->replaceUsesOfWith(LIC, Replacement);
887 Worklist.push_back(U);
890 // Otherwise, we don't know the precise value of LIC, but we do know that it
891 // is certainly NOT "Val". As such, simplify any uses in the loop that we
892 // can. This case occurs when we unswitch switch statements.
893 for (unsigned i = 0, e = Users.size(); i != e; ++i)
894 if (Instruction *U = cast<Instruction>(Users[i])) {
898 Worklist.push_back(U);
900 // If we know that LIC is not Val, use this info to simplify code.
901 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
902 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
903 if (SI->getCaseValue(i) == Val) {
904 // Found a dead case value. Don't remove PHI nodes in the
905 // successor if they become single-entry, those PHI nodes may
906 // be in the Users list.
908 // FIXME: This is a hack. We need to keep the successor around
909 // and hooked up so as to preserve the loop structure, because
910 // trying to update it is complicated. So instead we preserve the
911 // loop structure and put the block on a dead code path.
912 BasicBlock *Switch = SI->getParent();
913 SplitEdge(Switch, SI->getSuccessor(i), this);
914 // Compute the successors instead of relying on the return value
915 // of SplitEdge, since it may have split the switch successor
917 BasicBlock *NewSISucc = SI->getSuccessor(i);
918 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
919 // Create an "unreachable" destination.
920 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
923 new UnreachableInst(Context, Abort);
924 // Force the new case destination to branch to the "unreachable"
925 // block while maintaining a (dead) CFG edge to the old block.
926 NewSISucc->getTerminator()->eraseFromParent();
927 BranchInst::Create(Abort, OldSISucc,
928 ConstantInt::getTrue(Context), NewSISucc);
929 // Release the PHI operands for this edge.
930 for (BasicBlock::iterator II = NewSISucc->begin();
931 PHINode *PN = dyn_cast<PHINode>(II); ++II)
932 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
933 UndefValue::get(PN->getType()));
934 // Tell the domtree about the new block. We don't fully update the
935 // domtree here -- instead we force it to do a full recomputation
936 // after the pass is complete -- but we do need to inform it of
939 DT->addNewBlock(Abort, NewSISucc);
945 // TODO: We could do other simplifications, for example, turning
946 // LIC == Val -> false.
950 SimplifyCode(Worklist, L);
953 /// SimplifyCode - Okay, now that we have simplified some instructions in the
954 /// loop, walk over it and constant prop, dce, and fold control flow where
955 /// possible. Note that this is effectively a very simple loop-structure-aware
956 /// optimizer. During processing of this loop, L could very well be deleted, so
957 /// it must not be used.
959 /// FIXME: When the loop optimizer is more mature, separate this out to a new
962 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
963 while (!Worklist.empty()) {
964 Instruction *I = Worklist.back();
967 // Simple constant folding.
968 if (Constant *C = ConstantFoldInstruction(I)) {
969 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
974 if (isInstructionTriviallyDead(I)) {
975 DEBUG(dbgs() << "Remove dead instruction '" << *I);
977 // Add uses to the worklist, which may be dead now.
978 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
979 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
980 Worklist.push_back(Use);
981 LPM->deleteSimpleAnalysisValue(I, L);
982 RemoveFromWorklist(I, Worklist);
983 I->eraseFromParent();
988 // Special case hacks that appear commonly in unswitched code.
989 switch (I->getOpcode()) {
990 case Instruction::Select:
991 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
992 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
997 case Instruction::And:
998 if (isa<ConstantInt>(I->getOperand(0)) &&
1000 I->getOperand(0)->getType()->isIntegerTy(1))
1001 cast<BinaryOperator>(I)->swapOperands();
1002 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1003 if (CB->getType()->isIntegerTy(1)) {
1004 if (CB->isOne()) // X & 1 -> X
1005 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1007 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1011 case Instruction::Or:
1012 if (isa<ConstantInt>(I->getOperand(0)) &&
1014 I->getOperand(0)->getType()->isIntegerTy(1))
1015 cast<BinaryOperator>(I)->swapOperands();
1016 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1017 if (CB->getType()->isIntegerTy(1)) {
1018 if (CB->isOne()) // X | 1 -> 1
1019 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1021 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1025 case Instruction::Br: {
1026 BranchInst *BI = cast<BranchInst>(I);
1027 if (BI->isUnconditional()) {
1028 // If BI's parent is the only pred of the successor, fold the two blocks
1030 BasicBlock *Pred = BI->getParent();
1031 BasicBlock *Succ = BI->getSuccessor(0);
1032 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1033 if (!SinglePred) continue; // Nothing to do.
1034 assert(SinglePred == Pred && "CFG broken");
1036 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1037 << Succ->getName() << "\n");
1039 // Resolve any single entry PHI nodes in Succ.
1040 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1041 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1043 // Move all of the successor contents from Succ to Pred.
1044 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1046 LPM->deleteSimpleAnalysisValue(BI, L);
1047 BI->eraseFromParent();
1048 RemoveFromWorklist(BI, Worklist);
1050 // If Succ has any successors with PHI nodes, update them to have
1051 // entries coming from Pred instead of Succ.
1052 Succ->replaceAllUsesWith(Pred);
1054 // Remove Succ from the loop tree.
1055 LI->removeBlock(Succ);
1056 LPM->deleteSimpleAnalysisValue(Succ, L);
1057 Succ->eraseFromParent();
1059 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1060 // Conditional branch. Turn it into an unconditional branch, then
1061 // remove dead blocks.
1062 break; // FIXME: Enable.
1064 DEBUG(dbgs() << "Folded branch: " << *BI);
1065 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1066 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1067 DeadSucc->removePredecessor(BI->getParent(), true);
1068 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1069 LPM->deleteSimpleAnalysisValue(BI, L);
1070 BI->eraseFromParent();
1071 RemoveFromWorklist(BI, Worklist);
1074 RemoveBlockIfDead(DeadSucc, Worklist, L);