1 //===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===//
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 file transforms calls of the current function (self recursion) followed
11 // by a return instruction with a branch to the entry of the function, creating
12 // a loop. This pass also implements the following extensions to the basic
15 // 1. Trivial instructions between the call and return do not prevent the
16 // transformation from taking place, though currently the analysis cannot
17 // support moving any really useful instructions (only dead ones).
18 // 2. This pass transforms functions that are prevented from being tail
19 // recursive by an associative and commutative expression to use an
20 // accumulator variable, thus compiling the typical naive factorial or
21 // 'fib' implementation into efficient code.
22 // 3. TRE is performed if the function returns void, if the return
23 // returns the result returned by the call, or if the function returns a
24 // run-time constant on all exits from the function. It is possible, though
25 // unlikely, that the return returns something else (like constant 0), and
26 // can still be TRE'd. It can be TRE'd if ALL OTHER return instructions in
27 // the function return the exact same value.
28 // 4. If it can prove that callees do not access their caller stack frame,
29 // they are marked as eligible for tail call elimination (by the code
32 // There are several improvements that could be made:
34 // 1. If the function has any alloca instructions, these instructions will be
35 // moved out of the entry block of the function, causing them to be
36 // evaluated each time through the tail recursion. Safely keeping allocas
37 // in the entry block requires analysis to proves that the tail-called
38 // function does not read or write the stack object.
39 // 2. Tail recursion is only performed if the call immediately precedes the
40 // return instruction. It's possible that there could be a jump between
41 // the call and the return.
42 // 3. There can be intervening operations between the call and the return that
43 // prevent the TRE from occurring. For example, there could be GEP's and
44 // stores to memory that will not be read or written by the call. This
45 // requires some substantial analysis (such as with DSA) to prove safe to
46 // move ahead of the call, but doing so could allow many more TREs to be
47 // performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark.
48 // 4. The algorithm we use to detect if callees access their caller stack
49 // frames is very primitive.
51 //===----------------------------------------------------------------------===//
53 #include "llvm/Transforms/Scalar.h"
54 #include "llvm/ADT/STLExtras.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/Analysis/CaptureTracking.h"
58 #include "llvm/Analysis/CFG.h"
59 #include "llvm/Analysis/InlineCost.h"
60 #include "llvm/Analysis/InstructionSimplify.h"
61 #include "llvm/Analysis/Loads.h"
62 #include "llvm/Analysis/TargetTransformInfo.h"
63 #include "llvm/IR/CFG.h"
64 #include "llvm/IR/CallSite.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DerivedTypes.h"
67 #include "llvm/IR/Function.h"
68 #include "llvm/IR/Instructions.h"
69 #include "llvm/IR/IntrinsicInst.h"
70 #include "llvm/IR/Module.h"
71 #include "llvm/IR/ValueHandle.h"
72 #include "llvm/Pass.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
76 #include "llvm/Transforms/Utils/Local.h"
79 #define DEBUG_TYPE "tailcallelim"
81 STATISTIC(NumEliminated, "Number of tail calls removed");
82 STATISTIC(NumRetDuped, "Number of return duplicated");
83 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
86 struct TailCallElim : public FunctionPass {
87 const TargetTransformInfo *TTI;
89 static char ID; // Pass identification, replacement for typeid
90 TailCallElim() : FunctionPass(ID) {
91 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
94 void getAnalysisUsage(AnalysisUsage &AU) const override;
96 bool runOnFunction(Function &F) override;
99 bool runTRE(Function &F);
100 bool markTails(Function &F, bool &AllCallsAreTailCalls);
102 CallInst *FindTRECandidate(Instruction *I,
103 bool CannotTailCallElimCallsMarkedTail);
104 bool EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
105 BasicBlock *&OldEntry,
106 bool &TailCallsAreMarkedTail,
107 SmallVectorImpl<PHINode *> &ArgumentPHIs,
108 bool CannotTailCallElimCallsMarkedTail);
109 bool FoldReturnAndProcessPred(BasicBlock *BB,
110 ReturnInst *Ret, BasicBlock *&OldEntry,
111 bool &TailCallsAreMarkedTail,
112 SmallVectorImpl<PHINode *> &ArgumentPHIs,
113 bool CannotTailCallElimCallsMarkedTail);
114 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
115 bool &TailCallsAreMarkedTail,
116 SmallVectorImpl<PHINode *> &ArgumentPHIs,
117 bool CannotTailCallElimCallsMarkedTail);
118 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
119 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
123 char TailCallElim::ID = 0;
124 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim",
125 "Tail Call Elimination", false, false)
126 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
127 INITIALIZE_PASS_END(TailCallElim, "tailcallelim",
128 "Tail Call Elimination", false, false)
130 // Public interface to the TailCallElimination pass
131 FunctionPass *llvm::createTailCallEliminationPass() {
132 return new TailCallElim();
135 void TailCallElim::getAnalysisUsage(AnalysisUsage &AU) const {
136 AU.addRequired<TargetTransformInfo>();
139 /// \brief Scan the specified function for alloca instructions.
140 /// If it contains any dynamic allocas, returns false.
141 static bool CanTRE(Function &F) {
142 // Because of PR962, we don't TRE dynamic allocas.
145 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) {
146 if (!AI->isStaticAlloca())
155 bool TailCallElim::runOnFunction(Function &F) {
156 if (skipOptnoneFunction(F))
159 bool AllCallsAreTailCalls = false;
160 bool Modified = markTails(F, AllCallsAreTailCalls);
161 if (AllCallsAreTailCalls)
162 Modified |= runTRE(F);
167 struct AllocaDerivedValueTracker {
168 // Start at a root value and walk its use-def chain to mark calls that use the
169 // value or a derived value in AllocaUsers, and places where it may escape in
171 void walk(Value *Root) {
172 SmallVector<Use *, 32> Worklist;
173 SmallPtrSet<Use *, 32> Visited;
175 auto AddUsesToWorklist = [&](Value *V) {
176 for (auto &U : V->uses()) {
177 if (!Visited.insert(&U))
179 Worklist.push_back(&U);
183 AddUsesToWorklist(Root);
185 while (!Worklist.empty()) {
186 Use *U = Worklist.pop_back_val();
187 Instruction *I = cast<Instruction>(U->getUser());
189 switch (I->getOpcode()) {
190 case Instruction::Call:
191 case Instruction::Invoke: {
193 bool IsNocapture = !CS.isCallee(U) &&
194 CS.doesNotCapture(CS.getArgumentNo(U));
195 callUsesLocalStack(CS, IsNocapture);
197 // If the alloca-derived argument is passed in as nocapture, then it
198 // can't propagate to the call's return. That would be capturing.
203 case Instruction::Load: {
204 // The result of a load is not alloca-derived (unless an alloca has
205 // otherwise escaped, but this is a local analysis).
208 case Instruction::Store: {
209 if (U->getOperandNo() == 0)
210 EscapePoints.insert(I);
211 continue; // Stores have no users to analyze.
213 case Instruction::BitCast:
214 case Instruction::GetElementPtr:
215 case Instruction::PHI:
216 case Instruction::Select:
217 case Instruction::AddrSpaceCast:
220 EscapePoints.insert(I);
224 AddUsesToWorklist(I);
228 void callUsesLocalStack(CallSite CS, bool IsNocapture) {
229 // Add it to the list of alloca users. If it's already there, skip further
231 if (!AllocaUsers.insert(CS.getInstruction()))
234 // If it's nocapture then it can't capture the alloca.
238 // If it can write to memory, it can leak the alloca value.
239 if (!CS.onlyReadsMemory())
240 EscapePoints.insert(CS.getInstruction());
243 SmallPtrSet<Instruction *, 32> AllocaUsers;
244 SmallPtrSet<Instruction *, 32> EscapePoints;
248 bool TailCallElim::markTails(Function &F, bool &AllCallsAreTailCalls) {
249 if (F.callsFunctionThatReturnsTwice())
251 AllCallsAreTailCalls = true;
253 // The local stack holds all alloca instructions and all byval arguments.
254 AllocaDerivedValueTracker Tracker;
255 for (Argument &Arg : F.args()) {
256 if (Arg.hasByValAttr())
261 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I))
265 bool Modified = false;
267 // Track whether a block is reachable after an alloca has escaped. Blocks that
268 // contain the escaping instruction will be marked as being visited without an
269 // escaped alloca, since that is how the block began.
275 DenseMap<BasicBlock *, VisitType> Visited;
277 // We propagate the fact that an alloca has escaped from block to successor.
278 // Visit the blocks that are propagating the escapedness first. To do this, we
279 // maintain two worklists.
280 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
282 // We may enter a block and visit it thinking that no alloca has escaped yet,
283 // then see an escape point and go back around a loop edge and come back to
284 // the same block twice. Because of this, we defer setting tail on calls when
285 // we first encounter them in a block. Every entry in this list does not
286 // statically use an alloca via use-def chain analysis, but may find an alloca
287 // through other means if the block turns out to be reachable after an escape
289 SmallVector<CallInst *, 32> DeferredTails;
291 BasicBlock *BB = &F.getEntryBlock();
292 VisitType Escaped = UNESCAPED;
294 for (auto &I : *BB) {
295 if (Tracker.EscapePoints.count(&I))
298 CallInst *CI = dyn_cast<CallInst>(&I);
299 if (!CI || CI->isTailCall())
302 if (CI->doesNotAccessMemory()) {
303 // A call to a readnone function whose arguments are all things computed
304 // outside this function can be marked tail. Even if you stored the
305 // alloca address into a global, a readnone function can't load the
308 // Note that this runs whether we know an alloca has escaped or not. If
309 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
310 bool SafeToTail = true;
311 for (auto &Arg : CI->arg_operands()) {
312 if (isa<Constant>(Arg.getUser()))
314 if (Argument *A = dyn_cast<Argument>(Arg.getUser()))
315 if (!A->hasByValAttr())
321 F.getContext().emitOptimizationRemark(
322 "tailcallelim", F, CI->getDebugLoc(),
323 "marked this readnone call a tail call candidate");
330 if (Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) {
331 DeferredTails.push_back(CI);
333 AllCallsAreTailCalls = false;
337 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) {
338 auto &State = Visited[SuccBB];
339 if (State < Escaped) {
341 if (State == ESCAPED)
342 WorklistEscaped.push_back(SuccBB);
344 WorklistUnescaped.push_back(SuccBB);
348 if (!WorklistEscaped.empty()) {
349 BB = WorklistEscaped.pop_back_val();
353 while (!WorklistUnescaped.empty()) {
354 auto *NextBB = WorklistUnescaped.pop_back_val();
355 if (Visited[NextBB] == UNESCAPED) {
364 for (CallInst *CI : DeferredTails) {
365 if (Visited[CI->getParent()] != ESCAPED) {
366 // If the escape point was part way through the block, calls after the
367 // escape point wouldn't have been put into DeferredTails.
368 F.getContext().emitOptimizationRemark(
369 "tailcallelim", F, CI->getDebugLoc(),
370 "marked this call a tail call candidate");
374 AllCallsAreTailCalls = false;
381 bool TailCallElim::runTRE(Function &F) {
382 // If this function is a varargs function, we won't be able to PHI the args
383 // right, so don't even try to convert it...
384 if (F.getFunctionType()->isVarArg()) return false;
386 TTI = &getAnalysis<TargetTransformInfo>();
387 BasicBlock *OldEntry = nullptr;
388 bool TailCallsAreMarkedTail = false;
389 SmallVector<PHINode*, 8> ArgumentPHIs;
390 bool MadeChange = false;
392 // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
393 // marked with the 'tail' attribute, because doing so would cause the stack
394 // size to increase (real TRE would deallocate variable sized allocas, TRE
396 bool CanTRETailMarkedCall = CanTRE(F);
398 // Change any tail recursive calls to loops.
400 // FIXME: The code generator produces really bad code when an 'escaping
401 // alloca' is changed from being a static alloca to being a dynamic alloca.
402 // Until this is resolved, disable this transformation if that would ever
403 // happen. This bug is PR962.
404 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
405 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
406 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
407 ArgumentPHIs, !CanTRETailMarkedCall);
408 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
409 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
410 TailCallsAreMarkedTail, ArgumentPHIs,
411 !CanTRETailMarkedCall);
412 MadeChange |= Change;
416 // If we eliminated any tail recursions, it's possible that we inserted some
417 // silly PHI nodes which just merge an initial value (the incoming operand)
418 // with themselves. Check to see if we did and clean up our mess if so. This
419 // occurs when a function passes an argument straight through to its tail
421 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
422 PHINode *PN = ArgumentPHIs[i];
424 // If the PHI Node is a dynamic constant, replace it with the value it is.
425 if (Value *PNV = SimplifyInstruction(PN)) {
426 PN->replaceAllUsesWith(PNV);
427 PN->eraseFromParent();
435 /// CanMoveAboveCall - Return true if it is safe to move the specified
436 /// instruction from after the call to before the call, assuming that all
437 /// instructions between the call and this instruction are movable.
439 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
440 // FIXME: We can move load/store/call/free instructions above the call if the
441 // call does not mod/ref the memory location being processed.
442 if (I->mayHaveSideEffects()) // This also handles volatile loads.
445 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
446 // Loads may always be moved above calls without side effects.
447 if (CI->mayHaveSideEffects()) {
448 // Non-volatile loads may be moved above a call with side effects if it
449 // does not write to memory and the load provably won't trap.
450 // FIXME: Writes to memory only matter if they may alias the pointer
451 // being loaded from.
452 if (CI->mayWriteToMemory() ||
453 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
459 // Otherwise, if this is a side-effect free instruction, check to make sure
460 // that it does not use the return value of the call. If it doesn't use the
461 // return value of the call, it must only use things that are defined before
462 // the call, or movable instructions between the call and the instruction
464 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
465 if (I->getOperand(i) == CI)
470 // isDynamicConstant - Return true if the specified value is the same when the
471 // return would exit as it was when the initial iteration of the recursive
472 // function was executed.
474 // We currently handle static constants and arguments that are not modified as
475 // part of the recursion.
477 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
478 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
480 // Check to see if this is an immutable argument, if so, the value
481 // will be available to initialize the accumulator.
482 if (Argument *Arg = dyn_cast<Argument>(V)) {
483 // Figure out which argument number this is...
485 Function *F = CI->getParent()->getParent();
486 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
489 // If we are passing this argument into call as the corresponding
490 // argument operand, then the argument is dynamically constant.
491 // Otherwise, we cannot transform this function safely.
492 if (CI->getArgOperand(ArgNo) == Arg)
496 // Switch cases are always constant integers. If the value is being switched
497 // on and the return is only reachable from one of its cases, it's
498 // effectively constant.
499 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
500 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
501 if (SI->getCondition() == V)
502 return SI->getDefaultDest() != RI->getParent();
504 // Not a constant or immutable argument, we can't safely transform.
508 // getCommonReturnValue - Check to see if the function containing the specified
509 // tail call consistently returns the same runtime-constant value at all exit
510 // points except for IgnoreRI. If so, return the returned value.
512 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
513 Function *F = CI->getParent()->getParent();
514 Value *ReturnedValue = nullptr;
516 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) {
517 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator());
518 if (RI == nullptr || RI == IgnoreRI) continue;
520 // We can only perform this transformation if the value returned is
521 // evaluatable at the start of the initial invocation of the function,
522 // instead of at the end of the evaluation.
524 Value *RetOp = RI->getOperand(0);
525 if (!isDynamicConstant(RetOp, CI, RI))
528 if (ReturnedValue && RetOp != ReturnedValue)
529 return nullptr; // Cannot transform if differing values are returned.
530 ReturnedValue = RetOp;
532 return ReturnedValue;
535 /// CanTransformAccumulatorRecursion - If the specified instruction can be
536 /// transformed using accumulator recursion elimination, return the constant
537 /// which is the start of the accumulator value. Otherwise return null.
539 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
541 if (!I->isAssociative() || !I->isCommutative()) return nullptr;
542 assert(I->getNumOperands() == 2 &&
543 "Associative/commutative operations should have 2 args!");
545 // Exactly one operand should be the result of the call instruction.
546 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
547 (I->getOperand(0) != CI && I->getOperand(1) != CI))
550 // The only user of this instruction we allow is a single return instruction.
551 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back()))
554 // Ok, now we have to check all of the other return instructions in this
555 // function. If they return non-constants or differing values, then we cannot
556 // transform the function safely.
557 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI);
560 static Instruction *FirstNonDbg(BasicBlock::iterator I) {
561 while (isa<DbgInfoIntrinsic>(I))
567 TailCallElim::FindTRECandidate(Instruction *TI,
568 bool CannotTailCallElimCallsMarkedTail) {
569 BasicBlock *BB = TI->getParent();
570 Function *F = BB->getParent();
572 if (&BB->front() == TI) // Make sure there is something before the terminator.
575 // Scan backwards from the return, checking to see if there is a tail call in
576 // this block. If so, set CI to it.
577 CallInst *CI = nullptr;
578 BasicBlock::iterator BBI = TI;
580 CI = dyn_cast<CallInst>(BBI);
581 if (CI && CI->getCalledFunction() == F)
584 if (BBI == BB->begin())
585 return nullptr; // Didn't find a potential tail call.
589 // If this call is marked as a tail call, and if there are dynamic allocas in
590 // the function, we cannot perform this optimization.
591 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
594 // As a special case, detect code like this:
595 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
596 // and disable this xform in this case, because the code generator will
597 // lower the call to fabs into inline code.
598 if (BB == &F->getEntryBlock() &&
599 FirstNonDbg(BB->front()) == CI &&
600 FirstNonDbg(std::next(BB->begin())) == TI &&
601 CI->getCalledFunction() &&
602 !TTI->isLoweredToCall(CI->getCalledFunction())) {
603 // A single-block function with just a call and a return. Check that
604 // the arguments match.
605 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
606 E = CallSite(CI).arg_end();
607 Function::arg_iterator FI = F->arg_begin(),
609 for (; I != E && FI != FE; ++I, ++FI)
610 if (*I != &*FI) break;
611 if (I == E && FI == FE)
618 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
619 BasicBlock *&OldEntry,
620 bool &TailCallsAreMarkedTail,
621 SmallVectorImpl<PHINode *> &ArgumentPHIs,
622 bool CannotTailCallElimCallsMarkedTail) {
623 // If we are introducing accumulator recursion to eliminate operations after
624 // the call instruction that are both associative and commutative, the initial
625 // value for the accumulator is placed in this variable. If this value is set
626 // then we actually perform accumulator recursion elimination instead of
627 // simple tail recursion elimination. If the operation is an LLVM instruction
628 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
629 // we are handling the case when the return instruction returns a constant C
630 // which is different to the constant returned by other return instructions
631 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
632 // special case of accumulator recursion, the operation being "return C".
633 Value *AccumulatorRecursionEliminationInitVal = nullptr;
634 Instruction *AccumulatorRecursionInstr = nullptr;
636 // Ok, we found a potential tail call. We can currently only transform the
637 // tail call if all of the instructions between the call and the return are
638 // movable to above the call itself, leaving the call next to the return.
639 // Check that this is the case now.
640 BasicBlock::iterator BBI = CI;
641 for (++BBI; &*BBI != Ret; ++BBI) {
642 if (CanMoveAboveCall(BBI, CI)) continue;
644 // If we can't move the instruction above the call, it might be because it
645 // is an associative and commutative operation that could be transformed
646 // using accumulator recursion elimination. Check to see if this is the
647 // case, and if so, remember the initial accumulator value for later.
648 if ((AccumulatorRecursionEliminationInitVal =
649 CanTransformAccumulatorRecursion(BBI, CI))) {
650 // Yes, this is accumulator recursion. Remember which instruction
652 AccumulatorRecursionInstr = BBI;
654 return false; // Otherwise, we cannot eliminate the tail recursion!
658 // We can only transform call/return pairs that either ignore the return value
659 // of the call and return void, ignore the value of the call and return a
660 // constant, return the value returned by the tail call, or that are being
661 // accumulator recursion variable eliminated.
662 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
663 !isa<UndefValue>(Ret->getReturnValue()) &&
664 AccumulatorRecursionEliminationInitVal == nullptr &&
665 !getCommonReturnValue(nullptr, CI)) {
666 // One case remains that we are able to handle: the current return
667 // instruction returns a constant, and all other return instructions
668 // return a different constant.
669 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
670 return false; // Current return instruction does not return a constant.
671 // Check that all other return instructions return a common constant. If
672 // so, record it in AccumulatorRecursionEliminationInitVal.
673 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
674 if (!AccumulatorRecursionEliminationInitVal)
678 BasicBlock *BB = Ret->getParent();
679 Function *F = BB->getParent();
681 F->getContext().emitOptimizationRemark(
682 "tailcallelim", *F, CI->getDebugLoc(),
683 "transforming tail recursion to loop");
685 // OK! We can transform this tail call. If this is the first one found,
686 // create the new entry block, allowing us to branch back to the old entry.
688 OldEntry = &F->getEntryBlock();
689 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
690 NewEntry->takeName(OldEntry);
691 OldEntry->setName("tailrecurse");
692 BranchInst::Create(OldEntry, NewEntry);
694 // If this tail call is marked 'tail' and if there are any allocas in the
695 // entry block, move them up to the new entry block.
696 TailCallsAreMarkedTail = CI->isTailCall();
697 if (TailCallsAreMarkedTail)
698 // Move all fixed sized allocas from OldEntry to NewEntry.
699 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
700 NEBI = NewEntry->begin(); OEBI != E; )
701 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
702 if (isa<ConstantInt>(AI->getArraySize()))
703 AI->moveBefore(NEBI);
705 // Now that we have created a new block, which jumps to the entry
706 // block, insert a PHI node for each argument of the function.
707 // For now, we initialize each PHI to only have the real arguments
708 // which are passed in.
709 Instruction *InsertPos = OldEntry->begin();
710 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
712 PHINode *PN = PHINode::Create(I->getType(), 2,
713 I->getName() + ".tr", InsertPos);
714 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
715 PN->addIncoming(I, NewEntry);
716 ArgumentPHIs.push_back(PN);
720 // If this function has self recursive calls in the tail position where some
721 // are marked tail and some are not, only transform one flavor or another. We
722 // have to choose whether we move allocas in the entry block to the new entry
723 // block or not, so we can't make a good choice for both. NOTE: We could do
724 // slightly better here in the case that the function has no entry block
726 if (TailCallsAreMarkedTail && !CI->isTailCall())
729 // Ok, now that we know we have a pseudo-entry block WITH all of the
730 // required PHI nodes, add entries into the PHI node for the actual
731 // parameters passed into the tail-recursive call.
732 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
733 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
735 // If we are introducing an accumulator variable to eliminate the recursion,
736 // do so now. Note that we _know_ that no subsequent tail recursion
737 // eliminations will happen on this function because of the way the
738 // accumulator recursion predicate is set up.
740 if (AccumulatorRecursionEliminationInitVal) {
741 Instruction *AccRecInstr = AccumulatorRecursionInstr;
742 // Start by inserting a new PHI node for the accumulator.
743 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
745 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(),
746 std::distance(PB, PE) + 1,
747 "accumulator.tr", OldEntry->begin());
749 // Loop over all of the predecessors of the tail recursion block. For the
750 // real entry into the function we seed the PHI with the initial value,
751 // computed earlier. For any other existing branches to this block (due to
752 // other tail recursions eliminated) the accumulator is not modified.
753 // Because we haven't added the branch in the current block to OldEntry yet,
754 // it will not show up as a predecessor.
755 for (pred_iterator PI = PB; PI != PE; ++PI) {
757 if (P == &F->getEntryBlock())
758 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
760 AccPN->addIncoming(AccPN, P);
764 // Add an incoming argument for the current block, which is computed by
765 // our associative and commutative accumulator instruction.
766 AccPN->addIncoming(AccRecInstr, BB);
768 // Next, rewrite the accumulator recursion instruction so that it does not
769 // use the result of the call anymore, instead, use the PHI node we just
771 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
773 // Add an incoming argument for the current block, which is just the
774 // constant returned by the current return instruction.
775 AccPN->addIncoming(Ret->getReturnValue(), BB);
778 // Finally, rewrite any return instructions in the program to return the PHI
779 // node instead of the "initval" that they do currently. This loop will
780 // actually rewrite the return value we are destroying, but that's ok.
781 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
782 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
783 RI->setOperand(0, AccPN);
787 // Now that all of the PHI nodes are in place, remove the call and
788 // ret instructions, replacing them with an unconditional branch.
789 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
790 NewBI->setDebugLoc(CI->getDebugLoc());
792 BB->getInstList().erase(Ret); // Remove return.
793 BB->getInstList().erase(CI); // Remove call.
798 bool TailCallElim::FoldReturnAndProcessPred(BasicBlock *BB,
799 ReturnInst *Ret, BasicBlock *&OldEntry,
800 bool &TailCallsAreMarkedTail,
801 SmallVectorImpl<PHINode *> &ArgumentPHIs,
802 bool CannotTailCallElimCallsMarkedTail) {
805 // If the return block contains nothing but the return and PHI's,
806 // there might be an opportunity to duplicate the return in its
807 // predecessors and perform TRC there. Look for predecessors that end
808 // in unconditional branch and recursive call(s).
809 SmallVector<BranchInst*, 8> UncondBranchPreds;
810 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
811 BasicBlock *Pred = *PI;
812 TerminatorInst *PTI = Pred->getTerminator();
813 if (BranchInst *BI = dyn_cast<BranchInst>(PTI))
814 if (BI->isUnconditional())
815 UncondBranchPreds.push_back(BI);
818 while (!UncondBranchPreds.empty()) {
819 BranchInst *BI = UncondBranchPreds.pop_back_val();
820 BasicBlock *Pred = BI->getParent();
821 if (CallInst *CI = FindTRECandidate(BI, CannotTailCallElimCallsMarkedTail)){
822 DEBUG(dbgs() << "FOLDING: " << *BB
823 << "INTO UNCOND BRANCH PRED: " << *Pred);
824 EliminateRecursiveTailCall(CI, FoldReturnIntoUncondBranch(Ret, BB, Pred),
825 OldEntry, TailCallsAreMarkedTail, ArgumentPHIs,
826 CannotTailCallElimCallsMarkedTail);
836 TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
837 bool &TailCallsAreMarkedTail,
838 SmallVectorImpl<PHINode *> &ArgumentPHIs,
839 bool CannotTailCallElimCallsMarkedTail) {
840 CallInst *CI = FindTRECandidate(Ret, CannotTailCallElimCallsMarkedTail);
844 return EliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail,
846 CannotTailCallElimCallsMarkedTail);