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 Metrics.NeverInline) {
435 DEBUG(errs() << "NOT unswitching loop %"
436 << currentLoop->getHeader()->getName() << ", cost too high: "
437 << currentLoop->getBlocks().size() << "\n");
443 BasicBlock *ExitBlock;
444 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
445 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
447 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
453 // RemapInstruction - Convert the instruction operands from referencing the
454 // current values into those specified by ValueMap.
456 static inline void RemapInstruction(Instruction *I,
457 DenseMap<const Value *, Value*> &ValueMap) {
458 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
459 Value *Op = I->getOperand(op);
460 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
461 if (It != ValueMap.end()) Op = It->second;
462 I->setOperand(op, Op);
466 /// CloneLoop - Recursively clone the specified loop and all of its children,
467 /// mapping the blocks with the specified map.
468 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
469 LoopInfo *LI, LPPassManager *LPM) {
470 Loop *New = new Loop();
472 LPM->insertLoop(New, PL);
474 // Add all of the blocks in L to the new loop.
475 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
477 if (LI->getLoopFor(*I) == L)
478 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
480 // Add all of the subloops to the new loop.
481 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
482 CloneLoop(*I, New, VM, LI, LPM);
487 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
488 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
489 /// code immediately before InsertPt.
490 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
491 BasicBlock *TrueDest,
492 BasicBlock *FalseDest,
493 Instruction *InsertPt) {
494 // Insert a conditional branch on LIC to the two preheaders. The original
495 // code is the true version and the new code is the false version.
496 Value *BranchVal = LIC;
497 if (!isa<ConstantInt>(Val) ||
498 Val->getType() != Type::getInt1Ty(LIC->getContext()))
499 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
500 else if (Val != ConstantInt::getTrue(Val->getContext()))
501 // We want to enter the new loop when the condition is true.
502 std::swap(TrueDest, FalseDest);
504 // Insert the new branch.
505 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
507 // If either edge is critical, split it. This helps preserve LoopSimplify
508 // form for enclosing loops.
509 SplitCriticalEdge(BI, 0, this);
510 SplitCriticalEdge(BI, 1, this);
513 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
514 /// condition in it (a cond branch from its header block to its latch block,
515 /// where the path through the loop that doesn't execute its body has no
516 /// side-effects), unswitch it. This doesn't involve any code duplication, just
517 /// moving the conditional branch outside of the loop and updating loop info.
518 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
520 BasicBlock *ExitBlock) {
521 DEBUG(errs() << "loop-unswitch: Trivial-Unswitch loop %"
522 << loopHeader->getName() << " [" << L->getBlocks().size()
523 << " blocks] in Function " << L->getHeader()->getParent()->getName()
524 << " on cond: " << *Val << " == " << *Cond << "\n");
526 // First step, split the preheader, so that we know that there is a safe place
527 // to insert the conditional branch. We will change loopPreheader to have a
528 // conditional branch on Cond.
529 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
531 // Now that we have a place to insert the conditional branch, create a place
532 // to branch to: this is the exit block out of the loop that we should
535 // Split this block now, so that the loop maintains its exit block, and so
536 // that the jump from the preheader can execute the contents of the exit block
537 // without actually branching to it (the exit block should be dominated by the
538 // loop header, not the preheader).
539 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
540 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
542 // Okay, now we have a position to branch from and a position to branch to,
543 // insert the new conditional branch.
544 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
545 loopPreheader->getTerminator());
546 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
547 loopPreheader->getTerminator()->eraseFromParent();
549 // We need to reprocess this loop, it could be unswitched again.
552 // Now that we know that the loop is never entered when this condition is a
553 // particular value, rewrite the loop with this info. We know that this will
554 // at least eliminate the old branch.
555 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
559 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
560 /// blocks. Update the appropriate Phi nodes as we do so.
561 void LoopUnswitch::SplitExitEdges(Loop *L,
562 const SmallVector<BasicBlock *, 8> &ExitBlocks)
565 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
566 BasicBlock *ExitBlock = ExitBlocks[i];
567 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
568 pred_end(ExitBlock));
569 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
574 /// UnswitchNontrivialCondition - We determined that the loop is profitable
575 /// to unswitch when LIC equal Val. Split it into loop versions and test the
576 /// condition outside of either loop. Return the loops created as Out1/Out2.
577 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
579 Function *F = loopHeader->getParent();
580 DEBUG(errs() << "loop-unswitch: Unswitching loop %"
581 << loopHeader->getName() << " [" << L->getBlocks().size()
582 << " blocks] in Function " << F->getName()
583 << " when '" << *Val << "' == " << *LIC << "\n");
588 // First step, split the preheader and exit blocks, and add these blocks to
589 // the LoopBlocks list.
590 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
591 LoopBlocks.push_back(NewPreheader);
593 // We want the loop to come after the preheader, but before the exit blocks.
594 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
596 SmallVector<BasicBlock*, 8> ExitBlocks;
597 L->getUniqueExitBlocks(ExitBlocks);
599 // Split all of the edges from inside the loop to their exit blocks. Update
600 // the appropriate Phi nodes as we do so.
601 SplitExitEdges(L, ExitBlocks);
603 // The exit blocks may have been changed due to edge splitting, recompute.
605 L->getUniqueExitBlocks(ExitBlocks);
607 // Add exit blocks to the loop blocks.
608 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
610 // Next step, clone all of the basic blocks that make up the loop (including
611 // the loop preheader and exit blocks), keeping track of the mapping between
612 // the instructions and blocks.
613 NewBlocks.reserve(LoopBlocks.size());
614 DenseMap<const Value*, Value*> ValueMap;
615 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
616 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
617 NewBlocks.push_back(New);
618 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
619 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
622 // Splice the newly inserted blocks into the function right before the
623 // original preheader.
624 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
625 NewBlocks[0], F->end());
627 // Now we create the new Loop object for the versioned loop.
628 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
629 Loop *ParentLoop = L->getParentLoop();
631 // Make sure to add the cloned preheader and exit blocks to the parent loop
633 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
636 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
637 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
638 // The new exit block should be in the same loop as the old one.
639 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
640 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
642 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
643 "Exit block should have been split to have one successor!");
644 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
646 // If the successor of the exit block had PHI nodes, add an entry for
649 for (BasicBlock::iterator I = ExitSucc->begin();
650 (PN = dyn_cast<PHINode>(I)); ++I) {
651 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
652 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
653 if (It != ValueMap.end()) V = It->second;
654 PN->addIncoming(V, NewExit);
658 // Rewrite the code to refer to itself.
659 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
660 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
661 E = NewBlocks[i]->end(); I != E; ++I)
662 RemapInstruction(I, ValueMap);
664 // Rewrite the original preheader to select between versions of the loop.
665 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
666 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
667 "Preheader splitting did not work correctly!");
669 // Emit the new branch that selects between the two versions of this loop.
670 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
671 LPM->deleteSimpleAnalysisValue(OldBR, L);
672 OldBR->eraseFromParent();
674 LoopProcessWorklist.push_back(NewLoop);
677 // Now we rewrite the original code to know that the condition is true and the
678 // new code to know that the condition is false.
679 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
681 // It's possible that simplifying one loop could cause the other to be
682 // deleted. If so, don't simplify it.
683 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
684 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
688 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
690 static void RemoveFromWorklist(Instruction *I,
691 std::vector<Instruction*> &Worklist) {
692 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
694 while (WI != Worklist.end()) {
695 unsigned Offset = WI-Worklist.begin();
697 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
701 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
702 /// program, replacing all uses with V and update the worklist.
703 static void ReplaceUsesOfWith(Instruction *I, Value *V,
704 std::vector<Instruction*> &Worklist,
705 Loop *L, LPPassManager *LPM) {
706 DEBUG(errs() << "Replace with '" << *V << "': " << *I);
708 // Add uses to the worklist, which may be dead now.
709 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
710 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
711 Worklist.push_back(Use);
713 // Add users to the worklist which may be simplified now.
714 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
716 Worklist.push_back(cast<Instruction>(*UI));
717 LPM->deleteSimpleAnalysisValue(I, L);
718 RemoveFromWorklist(I, Worklist);
719 I->replaceAllUsesWith(V);
720 I->eraseFromParent();
724 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
725 /// information, and remove any dead successors it has.
727 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
728 std::vector<Instruction*> &Worklist,
730 if (pred_begin(BB) != pred_end(BB)) {
731 // This block isn't dead, since an edge to BB was just removed, see if there
732 // are any easy simplifications we can do now.
733 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
734 // If it has one pred, fold phi nodes in BB.
735 while (isa<PHINode>(BB->begin()))
736 ReplaceUsesOfWith(BB->begin(),
737 cast<PHINode>(BB->begin())->getIncomingValue(0),
740 // If this is the header of a loop and the only pred is the latch, we now
741 // have an unreachable loop.
742 if (Loop *L = LI->getLoopFor(BB))
743 if (loopHeader == BB && L->contains(Pred)) {
744 // Remove the branch from the latch to the header block, this makes
745 // the header dead, which will make the latch dead (because the header
746 // dominates the latch).
747 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
748 Pred->getTerminator()->eraseFromParent();
749 new UnreachableInst(BB->getContext(), Pred);
751 // The loop is now broken, remove it from LI.
752 RemoveLoopFromHierarchy(L);
754 // Reprocess the header, which now IS dead.
755 RemoveBlockIfDead(BB, Worklist, L);
759 // If pred ends in a uncond branch, add uncond branch to worklist so that
760 // the two blocks will get merged.
761 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
762 if (BI->isUnconditional())
763 Worklist.push_back(BI);
768 DEBUG(errs() << "Nuking dead block: " << *BB);
770 // Remove the instructions in the basic block from the worklist.
771 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
772 RemoveFromWorklist(I, Worklist);
774 // Anything that uses the instructions in this basic block should have their
775 // uses replaced with undefs.
776 // If I is not void type then replaceAllUsesWith undef.
777 // This allows ValueHandlers and custom metadata to adjust itself.
778 if (!I->getType()->isVoidTy())
779 I->replaceAllUsesWith(UndefValue::get(I->getType()));
782 // If this is the edge to the header block for a loop, remove the loop and
783 // promote all subloops.
784 if (Loop *BBLoop = LI->getLoopFor(BB)) {
785 if (BBLoop->getLoopLatch() == BB)
786 RemoveLoopFromHierarchy(BBLoop);
789 // Remove the block from the loop info, which removes it from any loops it
794 // Remove phi node entries in successors for this block.
795 TerminatorInst *TI = BB->getTerminator();
796 SmallVector<BasicBlock*, 4> Succs;
797 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
798 Succs.push_back(TI->getSuccessor(i));
799 TI->getSuccessor(i)->removePredecessor(BB);
802 // Unique the successors, remove anything with multiple uses.
803 array_pod_sort(Succs.begin(), Succs.end());
804 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
806 // Remove the basic block, including all of the instructions contained in it.
807 LPM->deleteSimpleAnalysisValue(BB, L);
808 BB->eraseFromParent();
809 // Remove successor blocks here that are not dead, so that we know we only
810 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
811 // then getting removed before we revisit them, which is badness.
813 for (unsigned i = 0; i != Succs.size(); ++i)
814 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
815 // One exception is loop headers. If this block was the preheader for a
816 // loop, then we DO want to visit the loop so the loop gets deleted.
817 // We know that if the successor is a loop header, that this loop had to
818 // be the preheader: the case where this was the latch block was handled
819 // above and headers can only have two predecessors.
820 if (!LI->isLoopHeader(Succs[i])) {
821 Succs.erase(Succs.begin()+i);
826 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
827 RemoveBlockIfDead(Succs[i], Worklist, L);
830 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
831 /// become unwrapped, either because the backedge was deleted, or because the
832 /// edge into the header was removed. If the edge into the header from the
833 /// latch block was removed, the loop is unwrapped but subloops are still alive,
834 /// so they just reparent loops. If the loops are actually dead, they will be
836 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
837 LPM->deleteLoopFromQueue(L);
838 RemoveLoopFromWorklist(L);
841 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
842 // the value specified by Val in the specified loop, or we know it does NOT have
843 // that value. Rewrite any uses of LIC or of properties correlated to it.
844 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
847 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
849 // FIXME: Support correlated properties, like:
856 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
857 // selects, switches.
858 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
859 std::vector<Instruction*> Worklist;
860 LLVMContext &Context = Val->getContext();
863 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
864 // in the loop with the appropriate one directly.
865 if (IsEqual || (isa<ConstantInt>(Val) &&
866 Val->getType() == Type::getInt1Ty(Val->getContext()))) {
871 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
872 !cast<ConstantInt>(Val)->getZExtValue());
874 for (unsigned i = 0, e = Users.size(); i != e; ++i)
875 if (Instruction *U = cast<Instruction>(Users[i])) {
876 if (!L->contains(U->getParent()))
878 U->replaceUsesOfWith(LIC, Replacement);
879 Worklist.push_back(U);
882 // Otherwise, we don't know the precise value of LIC, but we do know that it
883 // is certainly NOT "Val". As such, simplify any uses in the loop that we
884 // can. This case occurs when we unswitch switch statements.
885 for (unsigned i = 0, e = Users.size(); i != e; ++i)
886 if (Instruction *U = cast<Instruction>(Users[i])) {
887 if (!L->contains(U->getParent()))
890 Worklist.push_back(U);
892 // If we know that LIC is not Val, use this info to simplify code.
893 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
894 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
895 if (SI->getCaseValue(i) == Val) {
896 // Found a dead case value. Don't remove PHI nodes in the
897 // successor if they become single-entry, those PHI nodes may
898 // be in the Users list.
900 // FIXME: This is a hack. We need to keep the successor around
901 // and hooked up so as to preserve the loop structure, because
902 // trying to update it is complicated. So instead we preserve the
903 // loop structure and put the block on a dead code path.
904 BasicBlock *Switch = SI->getParent();
905 SplitEdge(Switch, SI->getSuccessor(i), this);
906 // Compute the successors instead of relying on the return value
907 // of SplitEdge, since it may have split the switch successor
909 BasicBlock *NewSISucc = SI->getSuccessor(i);
910 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
911 // Create an "unreachable" destination.
912 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
915 new UnreachableInst(Context, Abort);
916 // Force the new case destination to branch to the "unreachable"
917 // block while maintaining a (dead) CFG edge to the old block.
918 NewSISucc->getTerminator()->eraseFromParent();
919 BranchInst::Create(Abort, OldSISucc,
920 ConstantInt::getTrue(Context), NewSISucc);
921 // Release the PHI operands for this edge.
922 for (BasicBlock::iterator II = NewSISucc->begin();
923 PHINode *PN = dyn_cast<PHINode>(II); ++II)
924 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
925 UndefValue::get(PN->getType()));
926 // Tell the domtree about the new block. We don't fully update the
927 // domtree here -- instead we force it to do a full recomputation
928 // after the pass is complete -- but we do need to inform it of
931 DT->addNewBlock(Abort, NewSISucc);
937 // TODO: We could do other simplifications, for example, turning
938 // LIC == Val -> false.
942 SimplifyCode(Worklist, L);
945 /// SimplifyCode - Okay, now that we have simplified some instructions in the
946 /// loop, walk over it and constant prop, dce, and fold control flow where
947 /// possible. Note that this is effectively a very simple loop-structure-aware
948 /// optimizer. During processing of this loop, L could very well be deleted, so
949 /// it must not be used.
951 /// FIXME: When the loop optimizer is more mature, separate this out to a new
954 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
955 while (!Worklist.empty()) {
956 Instruction *I = Worklist.back();
959 // Simple constant folding.
960 if (Constant *C = ConstantFoldInstruction(I, I->getContext())) {
961 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
966 if (isInstructionTriviallyDead(I)) {
967 DEBUG(errs() << "Remove dead instruction '" << *I);
969 // Add uses to the worklist, which may be dead now.
970 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
971 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
972 Worklist.push_back(Use);
973 LPM->deleteSimpleAnalysisValue(I, L);
974 RemoveFromWorklist(I, Worklist);
975 I->eraseFromParent();
980 // Special case hacks that appear commonly in unswitched code.
981 switch (I->getOpcode()) {
982 case Instruction::Select:
983 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
984 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
989 case Instruction::And:
990 if (isa<ConstantInt>(I->getOperand(0)) &&
992 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
993 cast<BinaryOperator>(I)->swapOperands();
994 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
995 if (CB->getType() == Type::getInt1Ty(I->getContext())) {
996 if (CB->isOne()) // X & 1 -> X
997 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
999 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1003 case Instruction::Or:
1004 if (isa<ConstantInt>(I->getOperand(0)) &&
1006 I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
1007 cast<BinaryOperator>(I)->swapOperands();
1008 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1009 if (CB->getType() == Type::getInt1Ty(I->getContext())) {
1010 if (CB->isOne()) // X | 1 -> 1
1011 ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1013 ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1017 case Instruction::Br: {
1018 BranchInst *BI = cast<BranchInst>(I);
1019 if (BI->isUnconditional()) {
1020 // If BI's parent is the only pred of the successor, fold the two blocks
1022 BasicBlock *Pred = BI->getParent();
1023 BasicBlock *Succ = BI->getSuccessor(0);
1024 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1025 if (!SinglePred) continue; // Nothing to do.
1026 assert(SinglePred == Pred && "CFG broken");
1028 DEBUG(errs() << "Merging blocks: " << Pred->getName() << " <- "
1029 << Succ->getName() << "\n");
1031 // Resolve any single entry PHI nodes in Succ.
1032 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1033 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1035 // Move all of the successor contents from Succ to Pred.
1036 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1038 LPM->deleteSimpleAnalysisValue(BI, L);
1039 BI->eraseFromParent();
1040 RemoveFromWorklist(BI, Worklist);
1042 // If Succ has any successors with PHI nodes, update them to have
1043 // entries coming from Pred instead of Succ.
1044 Succ->replaceAllUsesWith(Pred);
1046 // Remove Succ from the loop tree.
1047 LI->removeBlock(Succ);
1048 LPM->deleteSimpleAnalysisValue(Succ, L);
1049 Succ->eraseFromParent();
1051 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1052 // Conditional branch. Turn it into an unconditional branch, then
1053 // remove dead blocks.
1054 break; // FIXME: Enable.
1056 DEBUG(errs() << "Folded branch: " << *BI);
1057 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1058 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1059 DeadSucc->removePredecessor(BI->getParent(), true);
1060 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1061 LPM->deleteSimpleAnalysisValue(BI, L);
1062 BI->eraseFromParent();
1063 RemoveFromWorklist(BI, Worklist);
1066 RemoveBlockIfDead(DeadSucc, Worklist, L);