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/DiagnosticInfo.h"
68 #include "llvm/IR/Function.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/IntrinsicInst.h"
71 #include "llvm/IR/Module.h"
72 #include "llvm/IR/ValueHandle.h"
73 #include "llvm/Pass.h"
74 #include "llvm/Support/Debug.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
80 #define DEBUG_TYPE "tailcallelim"
82 STATISTIC(NumEliminated, "Number of tail calls removed");
83 STATISTIC(NumRetDuped, "Number of return duplicated");
84 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
87 struct TailCallElim : public FunctionPass {
88 const TargetTransformInfo *TTI;
90 static char ID; // Pass identification, replacement for typeid
91 TailCallElim() : FunctionPass(ID) {
92 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
95 void getAnalysisUsage(AnalysisUsage &AU) const override;
97 bool runOnFunction(Function &F) override;
100 bool runTRE(Function &F);
101 bool markTails(Function &F, bool &AllCallsAreTailCalls);
103 CallInst *FindTRECandidate(Instruction *I,
104 bool CannotTailCallElimCallsMarkedTail);
105 bool EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
106 BasicBlock *&OldEntry,
107 bool &TailCallsAreMarkedTail,
108 SmallVectorImpl<PHINode *> &ArgumentPHIs,
109 bool CannotTailCallElimCallsMarkedTail);
110 bool FoldReturnAndProcessPred(BasicBlock *BB,
111 ReturnInst *Ret, BasicBlock *&OldEntry,
112 bool &TailCallsAreMarkedTail,
113 SmallVectorImpl<PHINode *> &ArgumentPHIs,
114 bool CannotTailCallElimCallsMarkedTail);
115 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
116 bool &TailCallsAreMarkedTail,
117 SmallVectorImpl<PHINode *> &ArgumentPHIs,
118 bool CannotTailCallElimCallsMarkedTail);
119 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
120 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
124 char TailCallElim::ID = 0;
125 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim",
126 "Tail Call Elimination", false, false)
127 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
128 INITIALIZE_PASS_END(TailCallElim, "tailcallelim",
129 "Tail Call Elimination", false, false)
131 // Public interface to the TailCallElimination pass
132 FunctionPass *llvm::createTailCallEliminationPass() {
133 return new TailCallElim();
136 void TailCallElim::getAnalysisUsage(AnalysisUsage &AU) const {
137 AU.addRequired<TargetTransformInfo>();
140 /// \brief Scan the specified function for alloca instructions.
141 /// If it contains any dynamic allocas, returns false.
142 static bool CanTRE(Function &F) {
143 // Because of PR962, we don't TRE dynamic allocas.
146 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I)) {
147 if (!AI->isStaticAlloca())
156 bool TailCallElim::runOnFunction(Function &F) {
157 if (skipOptnoneFunction(F))
160 bool AllCallsAreTailCalls = false;
161 bool Modified = markTails(F, AllCallsAreTailCalls);
162 if (AllCallsAreTailCalls)
163 Modified |= runTRE(F);
168 struct AllocaDerivedValueTracker {
169 // Start at a root value and walk its use-def chain to mark calls that use the
170 // value or a derived value in AllocaUsers, and places where it may escape in
172 void walk(Value *Root) {
173 SmallVector<Use *, 32> Worklist;
174 SmallPtrSet<Use *, 32> Visited;
176 auto AddUsesToWorklist = [&](Value *V) {
177 for (auto &U : V->uses()) {
178 if (!Visited.insert(&U))
180 Worklist.push_back(&U);
184 AddUsesToWorklist(Root);
186 while (!Worklist.empty()) {
187 Use *U = Worklist.pop_back_val();
188 Instruction *I = cast<Instruction>(U->getUser());
190 switch (I->getOpcode()) {
191 case Instruction::Call:
192 case Instruction::Invoke: {
194 bool IsNocapture = !CS.isCallee(U) &&
195 CS.doesNotCapture(CS.getArgumentNo(U));
196 callUsesLocalStack(CS, IsNocapture);
198 // If the alloca-derived argument is passed in as nocapture, then it
199 // can't propagate to the call's return. That would be capturing.
204 case Instruction::Load: {
205 // The result of a load is not alloca-derived (unless an alloca has
206 // otherwise escaped, but this is a local analysis).
209 case Instruction::Store: {
210 if (U->getOperandNo() == 0)
211 EscapePoints.insert(I);
212 continue; // Stores have no users to analyze.
214 case Instruction::BitCast:
215 case Instruction::GetElementPtr:
216 case Instruction::PHI:
217 case Instruction::Select:
218 case Instruction::AddrSpaceCast:
221 EscapePoints.insert(I);
225 AddUsesToWorklist(I);
229 void callUsesLocalStack(CallSite CS, bool IsNocapture) {
230 // Add it to the list of alloca users. If it's already there, skip further
232 if (!AllocaUsers.insert(CS.getInstruction()))
235 // If it's nocapture then it can't capture the alloca.
239 // If it can write to memory, it can leak the alloca value.
240 if (!CS.onlyReadsMemory())
241 EscapePoints.insert(CS.getInstruction());
244 SmallPtrSet<Instruction *, 32> AllocaUsers;
245 SmallPtrSet<Instruction *, 32> EscapePoints;
249 bool TailCallElim::markTails(Function &F, bool &AllCallsAreTailCalls) {
250 if (F.callsFunctionThatReturnsTwice())
252 AllCallsAreTailCalls = true;
254 // The local stack holds all alloca instructions and all byval arguments.
255 AllocaDerivedValueTracker Tracker;
256 for (Argument &Arg : F.args()) {
257 if (Arg.hasByValAttr())
262 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I))
266 bool Modified = false;
268 // Track whether a block is reachable after an alloca has escaped. Blocks that
269 // contain the escaping instruction will be marked as being visited without an
270 // escaped alloca, since that is how the block began.
276 DenseMap<BasicBlock *, VisitType> Visited;
278 // We propagate the fact that an alloca has escaped from block to successor.
279 // Visit the blocks that are propagating the escapedness first. To do this, we
280 // maintain two worklists.
281 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
283 // We may enter a block and visit it thinking that no alloca has escaped yet,
284 // then see an escape point and go back around a loop edge and come back to
285 // the same block twice. Because of this, we defer setting tail on calls when
286 // we first encounter them in a block. Every entry in this list does not
287 // statically use an alloca via use-def chain analysis, but may find an alloca
288 // through other means if the block turns out to be reachable after an escape
290 SmallVector<CallInst *, 32> DeferredTails;
292 BasicBlock *BB = &F.getEntryBlock();
293 VisitType Escaped = UNESCAPED;
295 for (auto &I : *BB) {
296 if (Tracker.EscapePoints.count(&I))
299 CallInst *CI = dyn_cast<CallInst>(&I);
300 if (!CI || CI->isTailCall())
303 if (CI->doesNotAccessMemory()) {
304 // A call to a readnone function whose arguments are all things computed
305 // outside this function can be marked tail. Even if you stored the
306 // alloca address into a global, a readnone function can't load the
309 // Note that this runs whether we know an alloca has escaped or not. If
310 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
311 bool SafeToTail = true;
312 for (auto &Arg : CI->arg_operands()) {
313 if (isa<Constant>(Arg.getUser()))
315 if (Argument *A = dyn_cast<Argument>(Arg.getUser()))
316 if (!A->hasByValAttr())
322 emitOptimizationRemark(
323 F.getContext(), "tailcallelim", F, CI->getDebugLoc(),
324 "marked this readnone call a tail call candidate");
331 if (Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) {
332 DeferredTails.push_back(CI);
334 AllCallsAreTailCalls = false;
338 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) {
339 auto &State = Visited[SuccBB];
340 if (State < Escaped) {
342 if (State == ESCAPED)
343 WorklistEscaped.push_back(SuccBB);
345 WorklistUnescaped.push_back(SuccBB);
349 if (!WorklistEscaped.empty()) {
350 BB = WorklistEscaped.pop_back_val();
354 while (!WorklistUnescaped.empty()) {
355 auto *NextBB = WorklistUnescaped.pop_back_val();
356 if (Visited[NextBB] == UNESCAPED) {
365 for (CallInst *CI : DeferredTails) {
366 if (Visited[CI->getParent()] != ESCAPED) {
367 // If the escape point was part way through the block, calls after the
368 // escape point wouldn't have been put into DeferredTails.
369 emitOptimizationRemark(F.getContext(), "tailcallelim", F,
371 "marked this call a tail call candidate");
375 AllCallsAreTailCalls = false;
382 bool TailCallElim::runTRE(Function &F) {
383 // If this function is a varargs function, we won't be able to PHI the args
384 // right, so don't even try to convert it...
385 if (F.getFunctionType()->isVarArg()) return false;
387 TTI = &getAnalysis<TargetTransformInfo>();
388 BasicBlock *OldEntry = nullptr;
389 bool TailCallsAreMarkedTail = false;
390 SmallVector<PHINode*, 8> ArgumentPHIs;
391 bool MadeChange = false;
393 // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
394 // marked with the 'tail' attribute, because doing so would cause the stack
395 // size to increase (real TRE would deallocate variable sized allocas, TRE
397 bool CanTRETailMarkedCall = CanTRE(F);
399 // Change any tail recursive calls to loops.
401 // FIXME: The code generator produces really bad code when an 'escaping
402 // alloca' is changed from being a static alloca to being a dynamic alloca.
403 // Until this is resolved, disable this transformation if that would ever
404 // happen. This bug is PR962.
405 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
406 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
407 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
408 ArgumentPHIs, !CanTRETailMarkedCall);
409 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
410 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
411 TailCallsAreMarkedTail, ArgumentPHIs,
412 !CanTRETailMarkedCall);
413 MadeChange |= Change;
417 // If we eliminated any tail recursions, it's possible that we inserted some
418 // silly PHI nodes which just merge an initial value (the incoming operand)
419 // with themselves. Check to see if we did and clean up our mess if so. This
420 // occurs when a function passes an argument straight through to its tail
422 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
423 PHINode *PN = ArgumentPHIs[i];
425 // If the PHI Node is a dynamic constant, replace it with the value it is.
426 if (Value *PNV = SimplifyInstruction(PN)) {
427 PN->replaceAllUsesWith(PNV);
428 PN->eraseFromParent();
436 /// CanMoveAboveCall - Return true if it is safe to move the specified
437 /// instruction from after the call to before the call, assuming that all
438 /// instructions between the call and this instruction are movable.
440 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
441 // FIXME: We can move load/store/call/free instructions above the call if the
442 // call does not mod/ref the memory location being processed.
443 if (I->mayHaveSideEffects()) // This also handles volatile loads.
446 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
447 // Loads may always be moved above calls without side effects.
448 if (CI->mayHaveSideEffects()) {
449 // Non-volatile loads may be moved above a call with side effects if it
450 // does not write to memory and the load provably won't trap.
451 // FIXME: Writes to memory only matter if they may alias the pointer
452 // being loaded from.
453 if (CI->mayWriteToMemory() ||
454 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
460 // Otherwise, if this is a side-effect free instruction, check to make sure
461 // that it does not use the return value of the call. If it doesn't use the
462 // return value of the call, it must only use things that are defined before
463 // the call, or movable instructions between the call and the instruction
465 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
466 if (I->getOperand(i) == CI)
471 // isDynamicConstant - Return true if the specified value is the same when the
472 // return would exit as it was when the initial iteration of the recursive
473 // function was executed.
475 // We currently handle static constants and arguments that are not modified as
476 // part of the recursion.
478 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
479 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
481 // Check to see if this is an immutable argument, if so, the value
482 // will be available to initialize the accumulator.
483 if (Argument *Arg = dyn_cast<Argument>(V)) {
484 // Figure out which argument number this is...
486 Function *F = CI->getParent()->getParent();
487 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
490 // If we are passing this argument into call as the corresponding
491 // argument operand, then the argument is dynamically constant.
492 // Otherwise, we cannot transform this function safely.
493 if (CI->getArgOperand(ArgNo) == Arg)
497 // Switch cases are always constant integers. If the value is being switched
498 // on and the return is only reachable from one of its cases, it's
499 // effectively constant.
500 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
501 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
502 if (SI->getCondition() == V)
503 return SI->getDefaultDest() != RI->getParent();
505 // Not a constant or immutable argument, we can't safely transform.
509 // getCommonReturnValue - Check to see if the function containing the specified
510 // tail call consistently returns the same runtime-constant value at all exit
511 // points except for IgnoreRI. If so, return the returned value.
513 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
514 Function *F = CI->getParent()->getParent();
515 Value *ReturnedValue = nullptr;
517 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) {
518 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator());
519 if (RI == nullptr || RI == IgnoreRI) continue;
521 // We can only perform this transformation if the value returned is
522 // evaluatable at the start of the initial invocation of the function,
523 // instead of at the end of the evaluation.
525 Value *RetOp = RI->getOperand(0);
526 if (!isDynamicConstant(RetOp, CI, RI))
529 if (ReturnedValue && RetOp != ReturnedValue)
530 return nullptr; // Cannot transform if differing values are returned.
531 ReturnedValue = RetOp;
533 return ReturnedValue;
536 /// CanTransformAccumulatorRecursion - If the specified instruction can be
537 /// transformed using accumulator recursion elimination, return the constant
538 /// which is the start of the accumulator value. Otherwise return null.
540 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
542 if (!I->isAssociative() || !I->isCommutative()) return nullptr;
543 assert(I->getNumOperands() == 2 &&
544 "Associative/commutative operations should have 2 args!");
546 // Exactly one operand should be the result of the call instruction.
547 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
548 (I->getOperand(0) != CI && I->getOperand(1) != CI))
551 // The only user of this instruction we allow is a single return instruction.
552 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back()))
555 // Ok, now we have to check all of the other return instructions in this
556 // function. If they return non-constants or differing values, then we cannot
557 // transform the function safely.
558 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI);
561 static Instruction *FirstNonDbg(BasicBlock::iterator I) {
562 while (isa<DbgInfoIntrinsic>(I))
568 TailCallElim::FindTRECandidate(Instruction *TI,
569 bool CannotTailCallElimCallsMarkedTail) {
570 BasicBlock *BB = TI->getParent();
571 Function *F = BB->getParent();
573 if (&BB->front() == TI) // Make sure there is something before the terminator.
576 // Scan backwards from the return, checking to see if there is a tail call in
577 // this block. If so, set CI to it.
578 CallInst *CI = nullptr;
579 BasicBlock::iterator BBI = TI;
581 CI = dyn_cast<CallInst>(BBI);
582 if (CI && CI->getCalledFunction() == F)
585 if (BBI == BB->begin())
586 return nullptr; // Didn't find a potential tail call.
590 // If this call is marked as a tail call, and if there are dynamic allocas in
591 // the function, we cannot perform this optimization.
592 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
595 // As a special case, detect code like this:
596 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
597 // and disable this xform in this case, because the code generator will
598 // lower the call to fabs into inline code.
599 if (BB == &F->getEntryBlock() &&
600 FirstNonDbg(BB->front()) == CI &&
601 FirstNonDbg(std::next(BB->begin())) == TI &&
602 CI->getCalledFunction() &&
603 !TTI->isLoweredToCall(CI->getCalledFunction())) {
604 // A single-block function with just a call and a return. Check that
605 // the arguments match.
606 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
607 E = CallSite(CI).arg_end();
608 Function::arg_iterator FI = F->arg_begin(),
610 for (; I != E && FI != FE; ++I, ++FI)
611 if (*I != &*FI) break;
612 if (I == E && FI == FE)
619 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
620 BasicBlock *&OldEntry,
621 bool &TailCallsAreMarkedTail,
622 SmallVectorImpl<PHINode *> &ArgumentPHIs,
623 bool CannotTailCallElimCallsMarkedTail) {
624 // If we are introducing accumulator recursion to eliminate operations after
625 // the call instruction that are both associative and commutative, the initial
626 // value for the accumulator is placed in this variable. If this value is set
627 // then we actually perform accumulator recursion elimination instead of
628 // simple tail recursion elimination. If the operation is an LLVM instruction
629 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
630 // we are handling the case when the return instruction returns a constant C
631 // which is different to the constant returned by other return instructions
632 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
633 // special case of accumulator recursion, the operation being "return C".
634 Value *AccumulatorRecursionEliminationInitVal = nullptr;
635 Instruction *AccumulatorRecursionInstr = nullptr;
637 // Ok, we found a potential tail call. We can currently only transform the
638 // tail call if all of the instructions between the call and the return are
639 // movable to above the call itself, leaving the call next to the return.
640 // Check that this is the case now.
641 BasicBlock::iterator BBI = CI;
642 for (++BBI; &*BBI != Ret; ++BBI) {
643 if (CanMoveAboveCall(BBI, CI)) continue;
645 // If we can't move the instruction above the call, it might be because it
646 // is an associative and commutative operation that could be transformed
647 // using accumulator recursion elimination. Check to see if this is the
648 // case, and if so, remember the initial accumulator value for later.
649 if ((AccumulatorRecursionEliminationInitVal =
650 CanTransformAccumulatorRecursion(BBI, CI))) {
651 // Yes, this is accumulator recursion. Remember which instruction
653 AccumulatorRecursionInstr = BBI;
655 return false; // Otherwise, we cannot eliminate the tail recursion!
659 // We can only transform call/return pairs that either ignore the return value
660 // of the call and return void, ignore the value of the call and return a
661 // constant, return the value returned by the tail call, or that are being
662 // accumulator recursion variable eliminated.
663 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
664 !isa<UndefValue>(Ret->getReturnValue()) &&
665 AccumulatorRecursionEliminationInitVal == nullptr &&
666 !getCommonReturnValue(nullptr, CI)) {
667 // One case remains that we are able to handle: the current return
668 // instruction returns a constant, and all other return instructions
669 // return a different constant.
670 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
671 return false; // Current return instruction does not return a constant.
672 // Check that all other return instructions return a common constant. If
673 // so, record it in AccumulatorRecursionEliminationInitVal.
674 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
675 if (!AccumulatorRecursionEliminationInitVal)
679 BasicBlock *BB = Ret->getParent();
680 Function *F = BB->getParent();
682 emitOptimizationRemark(F->getContext(), "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);