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/InlineCost.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/LoopPass.h"
40 #include "llvm/Analysis/Dominators.h"
41 #include "llvm/Transforms/Utils/Cloning.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/STLExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
54 STATISTIC(NumBranches, "Number of branches unswitched");
55 STATISTIC(NumSwitches, "Number of switches unswitched");
56 STATISTIC(NumSelects , "Number of selects unswitched");
57 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
58 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
60 // The specific value of 50 here was chosen based only on intuition and a
61 // few specific examples.
62 static cl::opt<unsigned>
63 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
64 cl::init(50), cl::Hidden);
67 class LoopUnswitch : public LoopPass {
68 LoopInfo *LI; // Loop information
71 // LoopProcessWorklist - Used to check if second loop needs processing
72 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
73 std::vector<Loop*> LoopProcessWorklist;
74 SmallPtrSet<Value *,8> UnswitchedVals;
80 DominanceFrontier *DF;
82 BasicBlock *loopHeader;
83 BasicBlock *loopPreheader;
85 // LoopBlocks contains all of the basic blocks of the loop, including the
86 // preheader of the loop, the body of the loop, and the exit blocks of the
87 // loop, in that order.
88 std::vector<BasicBlock*> LoopBlocks;
89 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
90 std::vector<BasicBlock*> NewBlocks;
93 static char ID; // Pass ID, replacement for typeid
94 explicit LoopUnswitch(bool Os = false) :
95 LoopPass(&ID), OptimizeForSize(Os), redoLoop(false),
96 currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL),
97 loopPreheader(NULL) {}
99 bool runOnLoop(Loop *L, LPPassManager &LPM);
100 bool processCurrentLoop();
102 /// This transformation requires natural loop information & requires that
103 /// loop preheaders be inserted into the CFG...
105 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
106 AU.addRequiredID(LoopSimplifyID);
107 AU.addPreservedID(LoopSimplifyID);
108 AU.addRequired<LoopInfo>();
109 AU.addPreserved<LoopInfo>();
110 AU.addRequiredID(LCSSAID);
111 AU.addPreservedID(LCSSAID);
112 AU.addPreserved<DominatorTree>();
113 AU.addPreserved<DominanceFrontier>();
118 virtual void releaseMemory() {
119 UnswitchedVals.clear();
122 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
124 void RemoveLoopFromWorklist(Loop *L) {
125 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
126 LoopProcessWorklist.end(), L);
127 if (I != LoopProcessWorklist.end())
128 LoopProcessWorklist.erase(I);
131 void initLoopData() {
132 loopHeader = currentLoop->getHeader();
133 loopPreheader = currentLoop->getLoopPreheader();
136 /// Split all of the edges from inside the loop to their exit blocks.
137 /// Update the appropriate Phi nodes as we do so.
138 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
140 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
141 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
142 BasicBlock *ExitBlock);
143 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
145 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
146 Constant *Val, bool isEqual);
148 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
149 BasicBlock *TrueDest,
150 BasicBlock *FalseDest,
151 Instruction *InsertPt);
153 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
154 void RemoveBlockIfDead(BasicBlock *BB,
155 std::vector<Instruction*> &Worklist, Loop *l);
156 void RemoveLoopFromHierarchy(Loop *L);
157 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
158 BasicBlock **LoopExit = 0);
162 char LoopUnswitch::ID = 0;
163 static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
165 Pass *llvm::createLoopUnswitchPass(bool Os) {
166 return new LoopUnswitch(Os);
169 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
170 /// invariant in the loop, or has an invariant piece, return the invariant.
171 /// Otherwise, return null.
172 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
173 // Constants should be folded, not unswitched on!
174 if (isa<Constant>(Cond)) return 0;
176 // TODO: Handle: br (VARIANT|INVARIANT).
178 // Hoist simple values out.
179 if (L->makeLoopInvariant(Cond, Changed))
182 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
183 if (BO->getOpcode() == Instruction::And ||
184 BO->getOpcode() == Instruction::Or) {
185 // If either the left or right side is invariant, we can unswitch on this,
186 // which will cause the branch to go away in one loop and the condition to
187 // simplify in the other one.
188 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
190 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
197 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
198 LI = &getAnalysis<LoopInfo>();
200 DF = getAnalysisIfAvailable<DominanceFrontier>();
201 DT = getAnalysisIfAvailable<DominatorTree>();
203 Function *F = currentLoop->getHeader()->getParent();
204 bool Changed = false;
206 assert(currentLoop->isLCSSAForm());
208 Changed |= processCurrentLoop();
212 // FIXME: Reconstruct dom info, because it is not preserved properly.
214 DT->runOnFunction(*F);
216 DF->runOnFunction(*F);
221 /// processCurrentLoop - Do actual work and unswitch loop if possible
223 bool LoopUnswitch::processCurrentLoop() {
224 bool Changed = false;
225 LLVMContext &Context = currentLoop->getHeader()->getContext();
227 // Loop over all of the basic blocks in the loop. If we find an interior
228 // block that is branching on a loop-invariant condition, we can unswitch this
230 for (Loop::block_iterator I = currentLoop->block_begin(),
231 E = currentLoop->block_end();
233 TerminatorInst *TI = (*I)->getTerminator();
234 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
235 // If this isn't branching on an invariant condition, we can't unswitch
237 if (BI->isConditional()) {
238 // See if this, or some part of it, is loop invariant. If so, we can
239 // unswitch on it if we desire.
240 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
241 currentLoop, Changed);
242 if (LoopCond && UnswitchIfProfitable(LoopCond,
243 ConstantInt::getTrue(Context))) {
248 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
249 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
250 currentLoop, Changed);
251 if (LoopCond && SI->getNumCases() > 1) {
252 // Find a value to unswitch on:
253 // FIXME: this should chose the most expensive case!
254 Constant *UnswitchVal = SI->getCaseValue(1);
255 // Do not process same value again and again.
256 if (!UnswitchedVals.insert(UnswitchVal))
259 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
266 // Scan the instructions to check for unswitchable values.
267 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
269 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
270 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
271 currentLoop, Changed);
272 if (LoopCond && UnswitchIfProfitable(LoopCond,
273 ConstantInt::getTrue(Context))) {
282 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
283 /// 1. Exit the loop with no side effects.
284 /// 2. Branch to the latch block with no side-effects.
286 /// If these conditions are true, we return true and set ExitBB to the block we
289 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
291 std::set<BasicBlock*> &Visited) {
292 if (!Visited.insert(BB).second) {
293 // Already visited and Ok, end of recursion.
295 } else if (!L->contains(BB)) {
296 // Otherwise, this is a loop exit, this is fine so long as this is the
298 if (ExitBB != 0) return false;
303 // Otherwise, this is an unvisited intra-loop node. Check all successors.
304 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
305 // Check to see if the successor is a trivial loop exit.
306 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
310 // Okay, everything after this looks good, check to make sure that this block
311 // doesn't include any side effects.
312 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
313 if (I->mayHaveSideEffects())
319 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
320 /// leads to an exit from the specified loop, and has no side-effects in the
321 /// process. If so, return the block that is exited to, otherwise return null.
322 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
323 std::set<BasicBlock*> Visited;
324 Visited.insert(L->getHeader()); // Branches to header are ok.
325 BasicBlock *ExitBB = 0;
326 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
331 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
332 /// trivial: that is, that the condition controls whether or not the loop does
333 /// anything at all. If this is a trivial condition, unswitching produces no
334 /// code duplications (equivalently, it produces a simpler loop and a new empty
335 /// loop, which gets deleted).
337 /// If this is a trivial condition, return true, otherwise return false. When
338 /// returning true, this sets Cond and Val to the condition that controls the
339 /// trivial condition: when Cond dynamically equals Val, the loop is known to
340 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
343 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
344 BasicBlock **LoopExit) {
345 BasicBlock *Header = currentLoop->getHeader();
346 TerminatorInst *HeaderTerm = Header->getTerminator();
347 LLVMContext &Context = Header->getContext();
349 BasicBlock *LoopExitBB = 0;
350 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
351 // If the header block doesn't end with a conditional branch on Cond, we
353 if (!BI->isConditional() || BI->getCondition() != Cond)
356 // Check to see if a successor of the branch is guaranteed to go to the
357 // latch block or exit through a one exit block without having any
358 // side-effects. If so, determine the value of Cond that causes it to do
360 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
361 BI->getSuccessor(0)))) {
362 if (Val) *Val = ConstantInt::getTrue(Context);
363 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
364 BI->getSuccessor(1)))) {
365 if (Val) *Val = ConstantInt::getFalse(Context);
367 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
368 // If this isn't a switch on Cond, we can't handle it.
369 if (SI->getCondition() != Cond) return false;
371 // Check to see if a successor of the switch is guaranteed to go to the
372 // latch block or exit through a one exit block without having any
373 // side-effects. If so, determine the value of Cond that causes it to do
374 // this. Note that we can't trivially unswitch on the default case.
375 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
376 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
377 SI->getSuccessor(i)))) {
378 // Okay, we found a trivial case, remember the value that is trivial.
379 if (Val) *Val = SI->getCaseValue(i);
384 // If we didn't find a single unique LoopExit block, or if the loop exit block
385 // contains phi nodes, this isn't trivial.
386 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
387 return false; // Can't handle this.
389 if (LoopExit) *LoopExit = LoopExitBB;
391 // We already know that nothing uses any scalar values defined inside of this
392 // loop. As such, we just have to check to see if this loop will execute any
393 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
394 // part of the loop that the code *would* execute. We already checked the
395 // tail, check the header now.
396 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
397 if (I->mayHaveSideEffects())
402 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
403 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
404 /// unswitch the loop, reprocess the pieces, then return true.
405 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
408 Function *F = loopHeader->getParent();
410 // If the condition is trivial, always unswitch. There is no code growth for
412 if (!IsTrivialUnswitchCondition(LoopCond)) {
413 // Check to see if it would be profitable to unswitch current loop.
415 // Do not do non-trivial unswitch while optimizing for size.
416 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
419 // FIXME: This is overly conservative because it does not take into
420 // consideration code simplification opportunities and code that can
421 // be shared by the resultant unswitched loops.
423 for (Loop::block_iterator I = currentLoop->block_begin(),
424 E = currentLoop->block_end();
426 Metrics.analyzeBasicBlock(*I);
428 // Limit the number of instructions to avoid causing significant code
429 // expansion, and the number of basic blocks, to avoid loops with
430 // large numbers of branches which cause loop unswitching to go crazy.
431 // This is a very ad-hoc heuristic.
432 if (Metrics.NumInsts > Threshold ||
433 Metrics.NumBlocks * 5 > Threshold) {
434 DEBUG(errs() << "NOT unswitching loop %"
435 << currentLoop->getHeader()->getName() << ", cost too high: "
436 << currentLoop->getBlocks().size() << "\n");
442 BasicBlock *ExitBlock;
443 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
444 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
446 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
452 // RemapInstruction - Convert the instruction operands from referencing the
453 // current values into those specified by ValueMap.
455 static inline void RemapInstruction(Instruction *I,
456 DenseMap<const Value *, Value*> &ValueMap) {
457 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
458 Value *Op = I->getOperand(op);
459 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
460 if (It != ValueMap.end()) Op = It->second;
461 I->setOperand(op, Op);
465 /// CloneLoop - Recursively clone the specified loop and all of its children,
466 /// mapping the blocks with the specified map.
467 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
468 LoopInfo *LI, LPPassManager *LPM) {
469 Loop *New = new Loop();
471 LPM->insertLoop(New, PL);
473 // Add all of the blocks in L to the new loop.
474 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
476 if (LI->getLoopFor(*I) == L)
477 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
479 // Add all of the subloops to the new loop.
480 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
481 CloneLoop(*I, New, VM, LI, LPM);
486 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
487 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
488 /// code immediately before InsertPt.
489 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
490 BasicBlock *TrueDest,
491 BasicBlock *FalseDest,
492 Instruction *InsertPt) {
493 // Insert a conditional branch on LIC to the two preheaders. The original
494 // code is the true version and the new code is the false version.
495 Value *BranchVal = LIC;
496 if (!isa<ConstantInt>(Val) ||
497 Val->getType() != Type::getInt1Ty(LIC->getContext()))
498 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
499 else if (Val != ConstantInt::getTrue(Val->getContext()))
500 // We want to enter the new loop when the condition is true.
501 std::swap(TrueDest, FalseDest);
503 // Insert the new branch.
504 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
506 // If either edge is critical, split it. This helps preserve LoopSimplify
507 // form for enclosing loops.
508 SplitCriticalEdge(BI, 0, this);
509 SplitCriticalEdge(BI, 1, this);
512 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
513 /// condition in it (a cond branch from its header block to its latch block,
514 /// where the path through the loop that doesn't execute its body has no
515 /// side-effects), unswitch it. This doesn't involve any code duplication, just
516 /// moving the conditional branch outside of the loop and updating loop info.
517 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
519 BasicBlock *ExitBlock) {
520 DEBUG(errs() << "loop-unswitch: Trivial-Unswitch loop %"
521 << loopHeader->getName() << " [" << L->getBlocks().size()
522 << " blocks] in Function " << L->getHeader()->getParent()->getName()
523 << " on cond: " << *Val << " == " << *Cond << "\n");
525 // First step, split the preheader, so that we know that there is a safe place
526 // to insert the conditional branch. We will change loopPreheader to have a
527 // conditional branch on Cond.
528 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
530 // Now that we have a place to insert the conditional branch, create a place
531 // to branch to: this is the exit block out of the loop that we should
534 // Split this block now, so that the loop maintains its exit block, and so
535 // that the jump from the preheader can execute the contents of the exit block
536 // without actually branching to it (the exit block should be dominated by the
537 // loop header, not the preheader).
538 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
539 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
541 // Okay, now we have a position to branch from and a position to branch to,
542 // insert the new conditional branch.
543 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
544 loopPreheader->getTerminator());
545 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
546 loopPreheader->getTerminator()->eraseFromParent();
548 // We need to reprocess this loop, it could be unswitched again.
551 // Now that we know that the loop is never entered when this condition is a
552 // particular value, rewrite the loop with this info. We know that this will
553 // at least eliminate the old branch.
554 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
558 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
559 /// blocks. Update the appropriate Phi nodes as we do so.
560 void LoopUnswitch::SplitExitEdges(Loop *L,
561 const SmallVector<BasicBlock *, 8> &ExitBlocks)
564 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
565 BasicBlock *ExitBlock = ExitBlocks[i];
566 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
567 pred_end(ExitBlock));
568 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
573 /// UnswitchNontrivialCondition - We determined that the loop is profitable
574 /// to unswitch when LIC equal Val. Split it into loop versions and test the
575 /// condition outside of either loop. Return the loops created as Out1/Out2.
576 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
578 Function *F = loopHeader->getParent();
579 DEBUG(errs() << "loop-unswitch: Unswitching loop %"
580 << loopHeader->getName() << " [" << L->getBlocks().size()
581 << " blocks] in Function " << F->getName()
582 << " when '" << *Val << "' == " << *LIC << "\n");
587 // First step, split the preheader and exit blocks, and add these blocks to
588 // the LoopBlocks list.
589 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
590 LoopBlocks.push_back(NewPreheader);
592 // We want the loop to come after the preheader, but before the exit blocks.
593 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
595 SmallVector<BasicBlock*, 8> ExitBlocks;
596 L->getUniqueExitBlocks(ExitBlocks);
598 // Split all of the edges from inside the loop to their exit blocks. Update
599 // the appropriate Phi nodes as we do so.
600 SplitExitEdges(L, ExitBlocks);
602 // The exit blocks may have been changed due to edge splitting, recompute.
604 L->getUniqueExitBlocks(ExitBlocks);
606 // Add exit blocks to the loop blocks.
607 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
609 // Next step, clone all of the basic blocks that make up the loop (including
610 // the loop preheader and exit blocks), keeping track of the mapping between
611 // the instructions and blocks.
612 NewBlocks.reserve(LoopBlocks.size());
613 DenseMap<const Value*, Value*> ValueMap;
614 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
615 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
616 NewBlocks.push_back(New);
617 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
618 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
621 // Splice the newly inserted blocks into the function right before the
622 // original preheader.
623 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
624 NewBlocks[0], F->end());
626 // Now we create the new Loop object for the versioned loop.
627 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
628 Loop *ParentLoop = L->getParentLoop();
630 // Make sure to add the cloned preheader and exit blocks to the parent loop
632 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
635 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
636 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
637 // The new exit block should be in the same loop as the old one.
638 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
639 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
641 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
642 "Exit block should have been split to have one successor!");
643 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
645 // If the successor of the exit block had PHI nodes, add an entry for
648 for (BasicBlock::iterator I = ExitSucc->begin();
649 (PN = dyn_cast<PHINode>(I)); ++I) {
650 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
651 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
652 if (It != ValueMap.end()) V = It->second;
653 PN->addIncoming(V, NewExit);
657 // Rewrite the code to refer to itself.
658 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
659 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
660 E = NewBlocks[i]->end(); I != E; ++I)
661 RemapInstruction(I, ValueMap);
663 // Rewrite the original preheader to select between versions of the loop.
664 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
665 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
666 "Preheader splitting did not work correctly!");
668 // Emit the new branch that selects between the two versions of this loop.
669 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
670 LPM->deleteSimpleAnalysisValue(OldBR, L);
671 OldBR->eraseFromParent();
673 LoopProcessWorklist.push_back(NewLoop);
676 // Now we rewrite the original code to know that the condition is true and the
677 // new code to know that the condition is false.
678 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
680 // It's possible that simplifying one loop could cause the other to be
681 // deleted. If so, don't simplify it.
682 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
683 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
687 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
689 static void RemoveFromWorklist(Instruction *I,
690 std::vector<Instruction*> &Worklist) {
691 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
693 while (WI != Worklist.end()) {
694 unsigned Offset = WI-Worklist.begin();
696 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
700 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
701 /// program, replacing all uses with V and update the worklist.
702 static void ReplaceUsesOfWith(Instruction *I, Value *V,
703 std::vector<Instruction*> &Worklist,
704 Loop *L, LPPassManager *LPM) {
705 DEBUG(errs() << "Replace with '" << *V << "': " << *I);
707 // Add uses to the worklist, which may be dead now.
708 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
709 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
710 Worklist.push_back(Use);
712 // Add users to the worklist which may be simplified now.
713 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
715 Worklist.push_back(cast<Instruction>(*UI));
716 LPM->deleteSimpleAnalysisValue(I, L);
717 RemoveFromWorklist(I, Worklist);
718 I->replaceAllUsesWith(V);
719 I->eraseFromParent();
723 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
724 /// information, and remove any dead successors it has.
726 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
727 std::vector<Instruction*> &Worklist,
729 if (pred_begin(BB) != pred_end(BB)) {
730 // This block isn't dead, since an edge to BB was just removed, see if there
731 // are any easy simplifications we can do now.
732 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
733 // If it has one pred, fold phi nodes in BB.
734 while (isa<PHINode>(BB->begin()))
735 ReplaceUsesOfWith(BB->begin(),
736 cast<PHINode>(BB->begin())->getIncomingValue(0),
739 // If this is the header of a loop and the only pred is the latch, we now
740 // have an unreachable loop.
741 if (Loop *L = LI->getLoopFor(BB))
742 if (loopHeader == BB && L->contains(Pred)) {
743 // Remove the branch from the latch to the header block, this makes
744 // the header dead, which will make the latch dead (because the header
745 // dominates the latch).
746 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
747 Pred->getTerminator()->eraseFromParent();
748 new UnreachableInst(BB->getContext(), Pred);
750 // The loop is now broken, remove it from LI.
751 RemoveLoopFromHierarchy(L);
753 // Reprocess the header, which now IS dead.
754 RemoveBlockIfDead(BB, Worklist, L);
758 // If pred ends in a uncond branch, add uncond branch to worklist so that
759 // the two blocks will get merged.
760 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
761 if (BI->isUnconditional())
762 Worklist.push_back(BI);
767 DEBUG(errs() << "Nuking dead block: " << *BB);
769 // Remove the instructions in the basic block from the worklist.
770 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
771 RemoveFromWorklist(I, Worklist);
773 // Anything that uses the instructions in this basic block should have their
774 // uses replaced with undefs.
775 // If I is not void type then replaceAllUsesWith undef.
776 // This allows ValueHandlers and custom metadata to adjust itself.
777 if (!I->getType()->isVoidTy())
778 I->replaceAllUsesWith(UndefValue::get(I->getType()));
781 // If this is the edge to the header block for a loop, remove the loop and
782 // promote all subloops.
783 if (Loop *BBLoop = LI->getLoopFor(BB)) {
784 if (BBLoop->getLoopLatch() == BB)
785 RemoveLoopFromHierarchy(BBLoop);
788 // Remove the block from the loop info, which removes it from any loops it
793 // Remove phi node entries in successors for this block.
794 TerminatorInst *TI = BB->getTerminator();
795 SmallVector<BasicBlock*, 4> Succs;
796 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
797 Succs.push_back(TI->getSuccessor(i));
798 TI->getSuccessor(i)->removePredecessor(BB);
801 // Unique the successors, remove anything with multiple uses.
802 array_pod_sort(Succs.begin(), Succs.end());
803 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
805 // Remove the basic block, including all of the instructions contained in it.
806 LPM->deleteSimpleAnalysisValue(BB, L);
807 BB->eraseFromParent();
808 // Remove successor blocks here that are not dead, so that we know we only
809 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
810 // then getting removed before we revisit them, which is badness.
812 for (unsigned i = 0; i != Succs.size(); ++i)
813 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
814 // One exception is loop headers. If this block was the preheader for a
815 // loop, then we DO want to visit the loop so the loop gets deleted.
816 // We know that if the successor is a loop header, that this loop had to
817 // be the preheader: the case where this was the latch block was handled
818 // above and headers can only have two predecessors.
819 if (!LI->isLoopHeader(Succs[i])) {
820 Succs.erase(Succs.begin()+i);
825 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
826 RemoveBlockIfDead(Succs[i], Worklist, L);
829 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
830 /// become unwrapped, either because the backedge was deleted, or because the
831 /// edge into the header was removed. If the edge into the header from the
832 /// latch block was removed, the loop is unwrapped but subloops are still alive,
833 /// so they just reparent loops. If the loops are actually dead, they will be
835 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
836 LPM->deleteLoopFromQueue(L);
837 RemoveLoopFromWorklist(L);
840 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
841 // the value specified by Val in the specified loop, or we know it does NOT have
842 // that value. Rewrite any uses of LIC or of properties correlated to it.
843 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
846 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
848 // FIXME: Support correlated properties, like:
855 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
856 // selects, switches.
857 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
858 std::vector<Instruction*> Worklist;
859 LLVMContext &Context = Val->getContext();
862 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
863 // in the loop with the appropriate one directly.
864 if (IsEqual || (isa<ConstantInt>(Val) &&
865 Val->getType() == Type::getInt1Ty(Val->getContext()))) {
870 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
871 !cast<ConstantInt>(Val)->getZExtValue());
873 for (unsigned i = 0, e = Users.size(); i != e; ++i)
874 if (Instruction *U = cast<Instruction>(Users[i])) {
875 if (!L->contains(U->getParent()))
877 U->replaceUsesOfWith(LIC, Replacement);
878 Worklist.push_back(U);
881 // Otherwise, we don't know the precise value of LIC, but we do know that it
882 // is certainly NOT "Val". As such, simplify any uses in the loop that we
883 // can. This case occurs when we unswitch switch statements.
884 for (unsigned i = 0, e = Users.size(); i != e; ++i)
885 if (Instruction *U = cast<Instruction>(Users[i])) {
886 if (!L->contains(U->getParent()))
889 Worklist.push_back(U);
891 // If we know that LIC is not Val, use this info to simplify code.
892 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
893 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
894 if (SI->getCaseValue(i) == Val) {
895 // Found a dead case value. Don't remove PHI nodes in the
896 // successor if they become single-entry, those PHI nodes may
897 // be in the Users list.
899 // FIXME: This is a hack. We need to keep the successor around
900 // and hooked up so as to preserve the loop structure, because
901 // trying to update it is complicated. So instead we preserve the
902 // loop structure and put the block on a dead code path.
903 BasicBlock *Switch = SI->getParent();
904 SplitEdge(Switch, SI->getSuccessor(i), this);
905 // Compute the successors instead of relying on the return value
906 // of SplitEdge, since it may have split the switch successor
908 BasicBlock *NewSISucc = SI->getSuccessor(i);
909 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
910 // Create an "unreachable" destination.
911 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
914 new UnreachableInst(Context, Abort);
915 // Force the new case destination to branch to the "unreachable"
916 // block while maintaining a (dead) CFG edge to the old block.
917 NewSISucc->getTerminator()->eraseFromParent();
918 BranchInst::Create(Abort, OldSISucc,
919 ConstantInt::getTrue(Context), NewSISucc);
920 // Release the PHI operands for this edge.
921 for (BasicBlock::iterator II = NewSISucc->begin();
922 PHINode *PN = dyn_cast<PHINode>(II); ++II)
923 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
924 UndefValue::get(PN->getType()));
925 // Tell the domtree about the new block. We don't fully update the
926 // domtree here -- instead we force it to do a full recomputation
927 // after the pass is complete -- but we do need to inform it of
930 DT->addNewBlock(Abort, NewSISucc);
936 // TODO: We could do other simplifications, for example, turning
937 // LIC == Val -> false.
941 SimplifyCode(Worklist, L);
944 /// SimplifyCode - Okay, now that we have simplified some instructions in the
945 /// loop, walk over it and constant prop, dce, and fold control flow where
946 /// possible. Note that this is effectively a very simple loop-structure-aware
947 /// optimizer. During processing of this loop, L could very well be deleted, so
948 /// it must not be used.
950 /// FIXME: When the loop optimizer is more mature, separate this out to a new
953 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
954 while (!Worklist.empty()) {
955 Instruction *I = Worklist.back();
958 // Simple constant folding.
959 if (Constant *C = ConstantFoldInstruction(I, I->getContext())) {
960 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
965 if (isInstructionTriviallyDead(I)) {
966 DEBUG(errs() << "Remove dead instruction '" << *I);
968 // Add uses to the worklist, which may be dead now.
969 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
970 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
971 Worklist.push_back(Use);
972 LPM->deleteSimpleAnalysisValue(I, L);
973 RemoveFromWorklist(I, Worklist);
974 I->eraseFromParent();
979 // Special case hacks that appear commonly in unswitched code.
980 switch (I->getOpcode()) {
981 case Instruction::Select:
982 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
983 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
988 case Instruction::And:
989 if (isa<ConstantInt>(I->getOperand(0)) &&
991 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
992 cast<BinaryOperator>(I)->swapOperands();
993 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
994 if (CB->getType() == Type::getInt1Ty(I->getContext())) {
995 if (CB->isOne()) // X & 1 -> X
996 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
998 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1002 case Instruction::Or:
1003 if (isa<ConstantInt>(I->getOperand(0)) &&
1005 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
1006 cast<BinaryOperator>(I)->swapOperands();
1007 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1008 if (CB->getType() == Type::getInt1Ty(I->getContext())) {
1009 if (CB->isOne()) // X | 1 -> 1
1010 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1012 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1016 case Instruction::Br: {
1017 BranchInst *BI = cast<BranchInst>(I);
1018 if (BI->isUnconditional()) {
1019 // If BI's parent is the only pred of the successor, fold the two blocks
1021 BasicBlock *Pred = BI->getParent();
1022 BasicBlock *Succ = BI->getSuccessor(0);
1023 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1024 if (!SinglePred) continue; // Nothing to do.
1025 assert(SinglePred == Pred && "CFG broken");
1027 DEBUG(errs() << "Merging blocks: " << Pred->getName() << " <- "
1028 << Succ->getName() << "\n");
1030 // Resolve any single entry PHI nodes in Succ.
1031 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1032 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1034 // Move all of the successor contents from Succ to Pred.
1035 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1037 LPM->deleteSimpleAnalysisValue(BI, L);
1038 BI->eraseFromParent();
1039 RemoveFromWorklist(BI, Worklist);
1041 // If Succ has any successors with PHI nodes, update them to have
1042 // entries coming from Pred instead of Succ.
1043 Succ->replaceAllUsesWith(Pred);
1045 // Remove Succ from the loop tree.
1046 LI->removeBlock(Succ);
1047 LPM->deleteSimpleAnalysisValue(Succ, L);
1048 Succ->eraseFromParent();
1050 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1051 // Conditional branch. Turn it into an unconditional branch, then
1052 // remove dead blocks.
1053 break; // FIXME: Enable.
1055 DEBUG(errs() << "Folded branch: " << *BI);
1056 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1057 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1058 DeadSucc->removePredecessor(BI->getParent(), true);
1059 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1060 LPM->deleteSimpleAnalysisValue(BI, L);
1061 BI->eraseFromParent();
1062 RemoveFromWorklist(BI, Worklist);
1065 RemoveBlockIfDead(DeadSucc, Worklist, L);