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/InstructionSimplify.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;
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), 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>();
116 virtual void releaseMemory() {
117 UnswitchedVals.clear();
120 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
122 void RemoveLoopFromWorklist(Loop *L) {
123 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
124 LoopProcessWorklist.end(), L);
125 if (I != LoopProcessWorklist.end())
126 LoopProcessWorklist.erase(I);
129 void initLoopData() {
130 loopHeader = currentLoop->getHeader();
131 loopPreheader = currentLoop->getLoopPreheader();
134 /// Split all of the edges from inside the loop to their exit blocks.
135 /// Update the appropriate Phi nodes as we do so.
136 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
138 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
139 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
140 BasicBlock *ExitBlock);
141 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
143 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
144 Constant *Val, bool isEqual);
146 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
147 BasicBlock *TrueDest,
148 BasicBlock *FalseDest,
149 Instruction *InsertPt);
151 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
152 void RemoveBlockIfDead(BasicBlock *BB,
153 std::vector<Instruction*> &Worklist, Loop *l);
154 void RemoveLoopFromHierarchy(Loop *L);
155 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
156 BasicBlock **LoopExit = 0);
160 char LoopUnswitch::ID = 0;
161 INITIALIZE_PASS(LoopUnswitch, "loop-unswitch", "Unswitch loops", false, false);
163 Pass *llvm::createLoopUnswitchPass(bool Os) {
164 return new LoopUnswitch(Os);
167 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
168 /// invariant in the loop, or has an invariant piece, return the invariant.
169 /// Otherwise, return null.
170 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
171 // We can never unswitch on vector conditions.
172 if (Cond->getType()->isVectorTy())
175 // Constants should be folded, not unswitched on!
176 if (isa<Constant>(Cond)) return 0;
178 // TODO: Handle: br (VARIANT|INVARIANT).
180 // Hoist simple values out.
181 if (L->makeLoopInvariant(Cond, Changed))
184 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
185 if (BO->getOpcode() == Instruction::And ||
186 BO->getOpcode() == Instruction::Or) {
187 // If either the left or right side is invariant, we can unswitch on this,
188 // which will cause the branch to go away in one loop and the condition to
189 // simplify in the other one.
190 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
192 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
199 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
200 LI = &getAnalysis<LoopInfo>();
202 DT = getAnalysisIfAvailable<DominatorTree>();
204 Function *F = currentLoop->getHeader()->getParent();
205 bool Changed = false;
207 assert(currentLoop->isLCSSAForm(*DT));
209 Changed |= processCurrentLoop();
213 // FIXME: Reconstruct dom info, because it is not preserved properly.
215 DT->runOnFunction(*F);
220 /// processCurrentLoop - Do actual work and unswitch loop if possible
222 bool LoopUnswitch::processCurrentLoop() {
223 bool Changed = false;
224 LLVMContext &Context = currentLoop->getHeader()->getContext();
226 // Loop over all of the basic blocks in the loop. If we find an interior
227 // block that is branching on a loop-invariant condition, we can unswitch this
229 for (Loop::block_iterator I = currentLoop->block_begin(),
230 E = currentLoop->block_end(); I != E; ++I) {
231 TerminatorInst *TI = (*I)->getTerminator();
232 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
233 // If this isn't branching on an invariant condition, we can't unswitch
235 if (BI->isConditional()) {
236 // See if this, or some part of it, is loop invariant. If so, we can
237 // unswitch on it if we desire.
238 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
239 currentLoop, Changed);
240 if (LoopCond && UnswitchIfProfitable(LoopCond,
241 ConstantInt::getTrue(Context))) {
246 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
247 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
248 currentLoop, Changed);
249 if (LoopCond && SI->getNumCases() > 1) {
250 // Find a value to unswitch on:
251 // FIXME: this should chose the most expensive case!
252 Constant *UnswitchVal = SI->getCaseValue(1);
253 // Do not process same value again and again.
254 if (!UnswitchedVals.insert(UnswitchVal))
257 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
264 // Scan the instructions to check for unswitchable values.
265 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
267 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
268 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
269 currentLoop, Changed);
270 if (LoopCond && UnswitchIfProfitable(LoopCond,
271 ConstantInt::getTrue(Context))) {
280 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
281 /// 1. Exit the loop with no side effects.
282 /// 2. Branch to the latch block with no side-effects.
284 /// If these conditions are true, we return true and set ExitBB to the block we
287 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
289 std::set<BasicBlock*> &Visited) {
290 if (!Visited.insert(BB).second) {
291 // Already visited and Ok, end of recursion.
293 } else if (!L->contains(BB)) {
294 // Otherwise, this is a loop exit, this is fine so long as this is the
296 if (ExitBB != 0) return false;
301 // Otherwise, this is an unvisited intra-loop node. Check all successors.
302 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
303 // Check to see if the successor is a trivial loop exit.
304 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
308 // Okay, everything after this looks good, check to make sure that this block
309 // doesn't include any side effects.
310 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
311 if (I->mayHaveSideEffects())
317 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
318 /// leads to an exit from the specified loop, and has no side-effects in the
319 /// process. If so, return the block that is exited to, otherwise return null.
320 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
321 std::set<BasicBlock*> Visited;
322 Visited.insert(L->getHeader()); // Branches to header are ok.
323 BasicBlock *ExitBB = 0;
324 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
329 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
330 /// trivial: that is, that the condition controls whether or not the loop does
331 /// anything at all. If this is a trivial condition, unswitching produces no
332 /// code duplications (equivalently, it produces a simpler loop and a new empty
333 /// loop, which gets deleted).
335 /// If this is a trivial condition, return true, otherwise return false. When
336 /// returning true, this sets Cond and Val to the condition that controls the
337 /// trivial condition: when Cond dynamically equals Val, the loop is known to
338 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
341 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
342 BasicBlock **LoopExit) {
343 BasicBlock *Header = currentLoop->getHeader();
344 TerminatorInst *HeaderTerm = Header->getTerminator();
345 LLVMContext &Context = Header->getContext();
347 BasicBlock *LoopExitBB = 0;
348 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
349 // If the header block doesn't end with a conditional branch on Cond, we
351 if (!BI->isConditional() || BI->getCondition() != Cond)
354 // Check to see if a successor of the branch is guaranteed to go to the
355 // latch block or exit through a one exit block without having any
356 // side-effects. If so, determine the value of Cond that causes it to do
358 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
359 BI->getSuccessor(0)))) {
360 if (Val) *Val = ConstantInt::getTrue(Context);
361 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
362 BI->getSuccessor(1)))) {
363 if (Val) *Val = ConstantInt::getFalse(Context);
365 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
366 // If this isn't a switch on Cond, we can't handle it.
367 if (SI->getCondition() != Cond) return false;
369 // Check to see if a successor of the switch is guaranteed to go to the
370 // latch block or exit through a one exit block without having any
371 // side-effects. If so, determine the value of Cond that causes it to do
372 // this. Note that we can't trivially unswitch on the default case.
373 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
374 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
375 SI->getSuccessor(i)))) {
376 // Okay, we found a trivial case, remember the value that is trivial.
377 if (Val) *Val = SI->getCaseValue(i);
382 // If we didn't find a single unique LoopExit block, or if the loop exit block
383 // contains phi nodes, this isn't trivial.
384 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
385 return false; // Can't handle this.
387 if (LoopExit) *LoopExit = LoopExitBB;
389 // We already know that nothing uses any scalar values defined inside of this
390 // loop. As such, we just have to check to see if this loop will execute any
391 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
392 // part of the loop that the code *would* execute. We already checked the
393 // tail, check the header now.
394 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
395 if (I->mayHaveSideEffects())
400 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
401 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
402 /// unswitch the loop, reprocess the pieces, then return true.
403 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
407 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
411 Function *F = loopHeader->getParent();
413 Constant *CondVal = 0;
414 BasicBlock *ExitBlock = 0;
415 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
416 // If the condition is trivial, always unswitch. There is no code growth
418 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
422 // Check to see if it would be profitable to unswitch current loop.
424 // Do not do non-trivial unswitch while optimizing for size.
425 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
428 // FIXME: This is overly conservative because it does not take into
429 // consideration code simplification opportunities and code that can
430 // be shared by the resultant unswitched loops.
432 for (Loop::block_iterator I = currentLoop->block_begin(),
433 E = currentLoop->block_end();
435 Metrics.analyzeBasicBlock(*I);
437 // Limit the number of instructions to avoid causing significant code
438 // expansion, and the number of basic blocks, to avoid loops with
439 // large numbers of branches which cause loop unswitching to go crazy.
440 // This is a very ad-hoc heuristic.
441 if (Metrics.NumInsts > Threshold ||
442 Metrics.NumBlocks * 5 > Threshold ||
443 Metrics.containsIndirectBr || Metrics.isRecursive) {
444 DEBUG(dbgs() << "NOT unswitching loop %"
445 << currentLoop->getHeader()->getName() << ", cost too high: "
446 << currentLoop->getBlocks().size() << "\n");
450 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
454 // RemapInstruction - Convert the instruction operands from referencing the
455 // current values into those specified by VMap.
457 static inline void RemapInstruction(Instruction *I,
458 ValueMap<const Value *, Value*> &VMap) {
459 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
460 Value *Op = I->getOperand(op);
461 ValueMap<const Value *, Value*>::iterator It = VMap.find(Op);
462 if (It != VMap.end()) Op = It->second;
463 I->setOperand(op, Op);
467 /// CloneLoop - Recursively clone the specified loop and all of its children,
468 /// mapping the blocks with the specified map.
469 static Loop *CloneLoop(Loop *L, Loop *PL, ValueMap<const Value*, Value*> &VM,
470 LoopInfo *LI, LPPassManager *LPM) {
471 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(dbgs() << "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){
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(dbgs() << "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 ValueMap<const Value*, Value*> VMap;
614 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
615 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
616 NewBlocks.push_back(NewBB);
617 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
618 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
621 // Splice the newly inserted blocks into the function right before the
622 // original preheader.
623 F->getBasicBlockList().splice(NewPreheader, 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(), VMap, 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>(VMap[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(); isa<PHINode>(I); ++I) {
649 PN = cast<PHINode>(I);
650 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
651 ValueMap<const Value *, Value*>::iterator It = VMap.find(V);
652 if (It != VMap.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, VMap);
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 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
677 // deletes the instruction (for example by simplifying a PHI that feeds into
678 // the condition that we're unswitching on), we don't rewrite the second
680 WeakVH LICHandle(LIC);
682 // Now we rewrite the original code to know that the condition is true and the
683 // new code to know that the condition is false.
684 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
686 // It's possible that simplifying one loop could cause the other to be
687 // changed to another value or a constant. If its a constant, don't simplify
689 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
690 LICHandle && !isa<Constant>(LICHandle))
691 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
694 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
696 static void RemoveFromWorklist(Instruction *I,
697 std::vector<Instruction*> &Worklist) {
698 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
700 while (WI != Worklist.end()) {
701 unsigned Offset = WI-Worklist.begin();
703 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
707 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
708 /// program, replacing all uses with V and update the worklist.
709 static void ReplaceUsesOfWith(Instruction *I, Value *V,
710 std::vector<Instruction*> &Worklist,
711 Loop *L, LPPassManager *LPM) {
712 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
714 // Add uses to the worklist, which may be dead now.
715 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
716 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
717 Worklist.push_back(Use);
719 // Add users to the worklist which may be simplified now.
720 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
722 Worklist.push_back(cast<Instruction>(*UI));
723 LPM->deleteSimpleAnalysisValue(I, L);
724 RemoveFromWorklist(I, Worklist);
725 I->replaceAllUsesWith(V);
726 I->eraseFromParent();
730 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
731 /// information, and remove any dead successors it has.
733 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
734 std::vector<Instruction*> &Worklist,
736 if (pred_begin(BB) != pred_end(BB)) {
737 // This block isn't dead, since an edge to BB was just removed, see if there
738 // are any easy simplifications we can do now.
739 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
740 // If it has one pred, fold phi nodes in BB.
741 while (isa<PHINode>(BB->begin()))
742 ReplaceUsesOfWith(BB->begin(),
743 cast<PHINode>(BB->begin())->getIncomingValue(0),
746 // If this is the header of a loop and the only pred is the latch, we now
747 // have an unreachable loop.
748 if (Loop *L = LI->getLoopFor(BB))
749 if (loopHeader == BB && L->contains(Pred)) {
750 // Remove the branch from the latch to the header block, this makes
751 // the header dead, which will make the latch dead (because the header
752 // dominates the latch).
753 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
754 Pred->getTerminator()->eraseFromParent();
755 new UnreachableInst(BB->getContext(), Pred);
757 // The loop is now broken, remove it from LI.
758 RemoveLoopFromHierarchy(L);
760 // Reprocess the header, which now IS dead.
761 RemoveBlockIfDead(BB, Worklist, L);
765 // If pred ends in a uncond branch, add uncond branch to worklist so that
766 // the two blocks will get merged.
767 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
768 if (BI->isUnconditional())
769 Worklist.push_back(BI);
774 DEBUG(dbgs() << "Nuking dead block: " << *BB);
776 // Remove the instructions in the basic block from the worklist.
777 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
778 RemoveFromWorklist(I, Worklist);
780 // Anything that uses the instructions in this basic block should have their
781 // uses replaced with undefs.
782 // If I is not void type then replaceAllUsesWith undef.
783 // This allows ValueHandlers and custom metadata to adjust itself.
784 if (!I->getType()->isVoidTy())
785 I->replaceAllUsesWith(UndefValue::get(I->getType()));
788 // If this is the edge to the header block for a loop, remove the loop and
789 // promote all subloops.
790 if (Loop *BBLoop = LI->getLoopFor(BB)) {
791 if (BBLoop->getLoopLatch() == BB)
792 RemoveLoopFromHierarchy(BBLoop);
795 // Remove the block from the loop info, which removes it from any loops it
800 // Remove phi node entries in successors for this block.
801 TerminatorInst *TI = BB->getTerminator();
802 SmallVector<BasicBlock*, 4> Succs;
803 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
804 Succs.push_back(TI->getSuccessor(i));
805 TI->getSuccessor(i)->removePredecessor(BB);
808 // Unique the successors, remove anything with multiple uses.
809 array_pod_sort(Succs.begin(), Succs.end());
810 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
812 // Remove the basic block, including all of the instructions contained in it.
813 LPM->deleteSimpleAnalysisValue(BB, L);
814 BB->eraseFromParent();
815 // Remove successor blocks here that are not dead, so that we know we only
816 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
817 // then getting removed before we revisit them, which is badness.
819 for (unsigned i = 0; i != Succs.size(); ++i)
820 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
821 // One exception is loop headers. If this block was the preheader for a
822 // loop, then we DO want to visit the loop so the loop gets deleted.
823 // We know that if the successor is a loop header, that this loop had to
824 // be the preheader: the case where this was the latch block was handled
825 // above and headers can only have two predecessors.
826 if (!LI->isLoopHeader(Succs[i])) {
827 Succs.erase(Succs.begin()+i);
832 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
833 RemoveBlockIfDead(Succs[i], Worklist, L);
836 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
837 /// become unwrapped, either because the backedge was deleted, or because the
838 /// edge into the header was removed. If the edge into the header from the
839 /// latch block was removed, the loop is unwrapped but subloops are still alive,
840 /// so they just reparent loops. If the loops are actually dead, they will be
842 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
843 LPM->deleteLoopFromQueue(L);
844 RemoveLoopFromWorklist(L);
847 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
848 // the value specified by Val in the specified loop, or we know it does NOT have
849 // that value. Rewrite any uses of LIC or of properties correlated to it.
850 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
853 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
855 // FIXME: Support correlated properties, like:
862 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
863 // selects, switches.
864 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
865 std::vector<Instruction*> Worklist;
866 LLVMContext &Context = Val->getContext();
869 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
870 // in the loop with the appropriate one directly.
871 if (IsEqual || (isa<ConstantInt>(Val) &&
872 Val->getType()->isIntegerTy(1))) {
877 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
878 !cast<ConstantInt>(Val)->getZExtValue());
880 for (unsigned i = 0, e = Users.size(); i != e; ++i)
881 if (Instruction *U = cast<Instruction>(Users[i])) {
884 U->replaceUsesOfWith(LIC, Replacement);
885 Worklist.push_back(U);
887 SimplifyCode(Worklist, L);
891 // Otherwise, we don't know the precise value of LIC, but we do know that it
892 // is certainly NOT "Val". As such, simplify any uses in the loop that we
893 // can. This case occurs when we unswitch switch statements.
894 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
895 Instruction *U = cast<Instruction>(Users[i]);
899 Worklist.push_back(U);
901 // TODO: We could do other simplifications, for example, turning
902 // 'icmp eq LIC, Val' -> false.
904 // If we know that LIC is not Val, use this info to simplify code.
905 SwitchInst *SI = dyn_cast<SwitchInst>(U);
906 if (SI == 0 || !isa<ConstantInt>(Val)) continue;
908 unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
909 if (DeadCase == 0) continue; // Default case is live for multiple values.
911 // Found a dead case value. Don't remove PHI nodes in the
912 // successor if they become single-entry, those PHI nodes may
913 // be in the Users list.
915 // FIXME: This is a hack. We need to keep the successor around
916 // and hooked up so as to preserve the loop structure, because
917 // trying to update it is complicated. So instead we preserve the
918 // loop structure and put the block on a dead code path.
919 BasicBlock *Switch = SI->getParent();
920 SplitEdge(Switch, SI->getSuccessor(DeadCase), this);
921 // Compute the successors instead of relying on the return value
922 // of SplitEdge, since it may have split the switch successor
924 BasicBlock *NewSISucc = SI->getSuccessor(DeadCase);
925 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
926 // Create an "unreachable" destination.
927 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
930 new UnreachableInst(Context, Abort);
931 // Force the new case destination to branch to the "unreachable"
932 // block while maintaining a (dead) CFG edge to the old block.
933 NewSISucc->getTerminator()->eraseFromParent();
934 BranchInst::Create(Abort, OldSISucc,
935 ConstantInt::getTrue(Context), NewSISucc);
936 // Release the PHI operands for this edge.
937 for (BasicBlock::iterator II = NewSISucc->begin();
938 PHINode *PN = dyn_cast<PHINode>(II); ++II)
939 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
940 UndefValue::get(PN->getType()));
941 // Tell the domtree about the new block. We don't fully update the
942 // domtree here -- instead we force it to do a full recomputation
943 // after the pass is complete -- but we do need to inform it of
946 DT->addNewBlock(Abort, NewSISucc);
949 SimplifyCode(Worklist, L);
952 /// SimplifyCode - Okay, now that we have simplified some instructions in the
953 /// loop, walk over it and constant prop, dce, and fold control flow where
954 /// possible. Note that this is effectively a very simple loop-structure-aware
955 /// optimizer. During processing of this loop, L could very well be deleted, so
956 /// it must not be used.
958 /// FIXME: When the loop optimizer is more mature, separate this out to a new
961 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
962 while (!Worklist.empty()) {
963 Instruction *I = Worklist.back();
966 // Simple constant folding.
967 if (Constant *C = ConstantFoldInstruction(I)) {
968 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
973 if (isInstructionTriviallyDead(I)) {
974 DEBUG(dbgs() << "Remove dead instruction '" << *I);
976 // Add uses to the worklist, which may be dead now.
977 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
978 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
979 Worklist.push_back(Use);
980 LPM->deleteSimpleAnalysisValue(I, L);
981 RemoveFromWorklist(I, Worklist);
982 I->eraseFromParent();
987 // See if instruction simplification can hack this up. This is common for
988 // things like "select false, X, Y" after unswitching made the condition be
990 if (Value *V = SimplifyInstruction(I)) {
991 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
995 // Special case hacks that appear commonly in unswitched code.
996 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
997 if (BI->isUnconditional()) {
998 // If BI's parent is the only pred of the successor, fold the two blocks
1000 BasicBlock *Pred = BI->getParent();
1001 BasicBlock *Succ = BI->getSuccessor(0);
1002 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1003 if (!SinglePred) continue; // Nothing to do.
1004 assert(SinglePred == Pred && "CFG broken");
1006 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1007 << Succ->getName() << "\n");
1009 // Resolve any single entry PHI nodes in Succ.
1010 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1011 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1013 // Move all of the successor contents from Succ to Pred.
1014 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1016 LPM->deleteSimpleAnalysisValue(BI, L);
1017 BI->eraseFromParent();
1018 RemoveFromWorklist(BI, Worklist);
1020 // If Succ has any successors with PHI nodes, update them to have
1021 // entries coming from Pred instead of Succ.
1022 Succ->replaceAllUsesWith(Pred);
1024 // Remove Succ from the loop tree.
1025 LI->removeBlock(Succ);
1026 LPM->deleteSimpleAnalysisValue(Succ, L);
1027 Succ->eraseFromParent();
1032 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1033 // Conditional branch. Turn it into an unconditional branch, then
1034 // remove dead blocks.
1035 continue; // FIXME: Enable.
1037 DEBUG(dbgs() << "Folded branch: " << *BI);
1038 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1039 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1040 DeadSucc->removePredecessor(BI->getParent(), true);
1041 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1042 LPM->deleteSimpleAnalysisValue(BI, L);
1043 BI->eraseFromParent();
1044 RemoveFromWorklist(BI, Worklist);
1047 RemoveBlockIfDead(DeadSucc, Worklist, L);