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.
231 AllocaUsers.insert(CS.getInstruction());
233 // If it's nocapture then it can't capture this alloca.
237 // If it can write to memory, it can leak the alloca value.
238 if (!CS.onlyReadsMemory())
239 EscapePoints.insert(CS.getInstruction());
242 SmallPtrSet<Instruction *, 32> AllocaUsers;
243 SmallPtrSet<Instruction *, 32> EscapePoints;
247 bool TailCallElim::markTails(Function &F, bool &AllCallsAreTailCalls) {
248 if (F.callsFunctionThatReturnsTwice())
250 AllCallsAreTailCalls = true;
252 // The local stack holds all alloca instructions and all byval arguments.
253 AllocaDerivedValueTracker Tracker;
254 for (Argument &Arg : F.args()) {
255 if (Arg.hasByValAttr())
260 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I))
264 bool Modified = false;
266 // Track whether a block is reachable after an alloca has escaped. Blocks that
267 // contain the escaping instruction will be marked as being visited without an
268 // escaped alloca, since that is how the block began.
274 DenseMap<BasicBlock *, VisitType> Visited;
276 // We propagate the fact that an alloca has escaped from block to successor.
277 // Visit the blocks that are propagating the escapedness first. To do this, we
278 // maintain two worklists.
279 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
281 // We may enter a block and visit it thinking that no alloca has escaped yet,
282 // then see an escape point and go back around a loop edge and come back to
283 // the same block twice. Because of this, we defer setting tail on calls when
284 // we first encounter them in a block. Every entry in this list does not
285 // statically use an alloca via use-def chain analysis, but may find an alloca
286 // through other means if the block turns out to be reachable after an escape
288 SmallVector<CallInst *, 32> DeferredTails;
290 BasicBlock *BB = &F.getEntryBlock();
291 VisitType Escaped = UNESCAPED;
293 for (auto &I : *BB) {
294 if (Tracker.EscapePoints.count(&I))
297 CallInst *CI = dyn_cast<CallInst>(&I);
298 if (!CI || CI->isTailCall())
301 if (CI->doesNotAccessMemory()) {
302 // A call to a readnone function whose arguments are all things computed
303 // outside this function can be marked tail. Even if you stored the
304 // alloca address into a global, a readnone function can't load the
307 // Note that this runs whether we know an alloca has escaped or not. If
308 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
309 bool SafeToTail = true;
310 for (auto &Arg : CI->arg_operands()) {
311 if (isa<Constant>(Arg.getUser()))
313 if (Argument *A = dyn_cast<Argument>(Arg.getUser()))
314 if (!A->hasByValAttr())
320 emitOptimizationRemark(
321 F.getContext(), "tailcallelim", F, CI->getDebugLoc(),
322 "marked this readnone call a tail call candidate");
329 if (Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) {
330 DeferredTails.push_back(CI);
332 AllCallsAreTailCalls = false;
336 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) {
337 auto &State = Visited[SuccBB];
338 if (State < Escaped) {
340 if (State == ESCAPED)
341 WorklistEscaped.push_back(SuccBB);
343 WorklistUnescaped.push_back(SuccBB);
347 if (!WorklistEscaped.empty()) {
348 BB = WorklistEscaped.pop_back_val();
352 while (!WorklistUnescaped.empty()) {
353 auto *NextBB = WorklistUnescaped.pop_back_val();
354 if (Visited[NextBB] == UNESCAPED) {
363 for (CallInst *CI : DeferredTails) {
364 if (Visited[CI->getParent()] != ESCAPED) {
365 // If the escape point was part way through the block, calls after the
366 // escape point wouldn't have been put into DeferredTails.
367 emitOptimizationRemark(F.getContext(), "tailcallelim", F,
369 "marked this call a tail call candidate");
373 AllCallsAreTailCalls = false;
380 bool TailCallElim::runTRE(Function &F) {
381 // If this function is a varargs function, we won't be able to PHI the args
382 // right, so don't even try to convert it...
383 if (F.getFunctionType()->isVarArg()) return false;
385 TTI = &getAnalysis<TargetTransformInfo>();
386 BasicBlock *OldEntry = nullptr;
387 bool TailCallsAreMarkedTail = false;
388 SmallVector<PHINode*, 8> ArgumentPHIs;
389 bool MadeChange = false;
391 // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
392 // marked with the 'tail' attribute, because doing so would cause the stack
393 // size to increase (real TRE would deallocate variable sized allocas, TRE
395 bool CanTRETailMarkedCall = CanTRE(F);
397 // Change any tail recursive calls to loops.
399 // FIXME: The code generator produces really bad code when an 'escaping
400 // alloca' is changed from being a static alloca to being a dynamic alloca.
401 // Until this is resolved, disable this transformation if that would ever
402 // happen. This bug is PR962.
403 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
404 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
405 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
406 ArgumentPHIs, !CanTRETailMarkedCall);
407 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
408 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
409 TailCallsAreMarkedTail, ArgumentPHIs,
410 !CanTRETailMarkedCall);
411 MadeChange |= Change;
415 // If we eliminated any tail recursions, it's possible that we inserted some
416 // silly PHI nodes which just merge an initial value (the incoming operand)
417 // with themselves. Check to see if we did and clean up our mess if so. This
418 // occurs when a function passes an argument straight through to its tail
420 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
421 PHINode *PN = ArgumentPHIs[i];
423 // If the PHI Node is a dynamic constant, replace it with the value it is.
424 if (Value *PNV = SimplifyInstruction(PN)) {
425 PN->replaceAllUsesWith(PNV);
426 PN->eraseFromParent();
434 /// CanMoveAboveCall - Return true if it is safe to move the specified
435 /// instruction from after the call to before the call, assuming that all
436 /// instructions between the call and this instruction are movable.
438 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
439 // FIXME: We can move load/store/call/free instructions above the call if the
440 // call does not mod/ref the memory location being processed.
441 if (I->mayHaveSideEffects()) // This also handles volatile loads.
444 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
445 // Loads may always be moved above calls without side effects.
446 if (CI->mayHaveSideEffects()) {
447 // Non-volatile loads may be moved above a call with side effects if it
448 // does not write to memory and the load provably won't trap.
449 // FIXME: Writes to memory only matter if they may alias the pointer
450 // being loaded from.
451 if (CI->mayWriteToMemory() ||
452 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
458 // Otherwise, if this is a side-effect free instruction, check to make sure
459 // that it does not use the return value of the call. If it doesn't use the
460 // return value of the call, it must only use things that are defined before
461 // the call, or movable instructions between the call and the instruction
463 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
464 if (I->getOperand(i) == CI)
469 // isDynamicConstant - Return true if the specified value is the same when the
470 // return would exit as it was when the initial iteration of the recursive
471 // function was executed.
473 // We currently handle static constants and arguments that are not modified as
474 // part of the recursion.
476 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
477 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
479 // Check to see if this is an immutable argument, if so, the value
480 // will be available to initialize the accumulator.
481 if (Argument *Arg = dyn_cast<Argument>(V)) {
482 // Figure out which argument number this is...
484 Function *F = CI->getParent()->getParent();
485 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
488 // If we are passing this argument into call as the corresponding
489 // argument operand, then the argument is dynamically constant.
490 // Otherwise, we cannot transform this function safely.
491 if (CI->getArgOperand(ArgNo) == Arg)
495 // Switch cases are always constant integers. If the value is being switched
496 // on and the return is only reachable from one of its cases, it's
497 // effectively constant.
498 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
499 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
500 if (SI->getCondition() == V)
501 return SI->getDefaultDest() != RI->getParent();
503 // Not a constant or immutable argument, we can't safely transform.
507 // getCommonReturnValue - Check to see if the function containing the specified
508 // tail call consistently returns the same runtime-constant value at all exit
509 // points except for IgnoreRI. If so, return the returned value.
511 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
512 Function *F = CI->getParent()->getParent();
513 Value *ReturnedValue = nullptr;
515 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) {
516 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator());
517 if (RI == nullptr || RI == IgnoreRI) continue;
519 // We can only perform this transformation if the value returned is
520 // evaluatable at the start of the initial invocation of the function,
521 // instead of at the end of the evaluation.
523 Value *RetOp = RI->getOperand(0);
524 if (!isDynamicConstant(RetOp, CI, RI))
527 if (ReturnedValue && RetOp != ReturnedValue)
528 return nullptr; // Cannot transform if differing values are returned.
529 ReturnedValue = RetOp;
531 return ReturnedValue;
534 /// CanTransformAccumulatorRecursion - If the specified instruction can be
535 /// transformed using accumulator recursion elimination, return the constant
536 /// which is the start of the accumulator value. Otherwise return null.
538 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
540 if (!I->isAssociative() || !I->isCommutative()) return nullptr;
541 assert(I->getNumOperands() == 2 &&
542 "Associative/commutative operations should have 2 args!");
544 // Exactly one operand should be the result of the call instruction.
545 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
546 (I->getOperand(0) != CI && I->getOperand(1) != CI))
549 // The only user of this instruction we allow is a single return instruction.
550 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back()))
553 // Ok, now we have to check all of the other return instructions in this
554 // function. If they return non-constants or differing values, then we cannot
555 // transform the function safely.
556 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI);
559 static Instruction *FirstNonDbg(BasicBlock::iterator I) {
560 while (isa<DbgInfoIntrinsic>(I))
566 TailCallElim::FindTRECandidate(Instruction *TI,
567 bool CannotTailCallElimCallsMarkedTail) {
568 BasicBlock *BB = TI->getParent();
569 Function *F = BB->getParent();
571 if (&BB->front() == TI) // Make sure there is something before the terminator.
574 // Scan backwards from the return, checking to see if there is a tail call in
575 // this block. If so, set CI to it.
576 CallInst *CI = nullptr;
577 BasicBlock::iterator BBI = TI;
579 CI = dyn_cast<CallInst>(BBI);
580 if (CI && CI->getCalledFunction() == F)
583 if (BBI == BB->begin())
584 return nullptr; // Didn't find a potential tail call.
588 // If this call is marked as a tail call, and if there are dynamic allocas in
589 // the function, we cannot perform this optimization.
590 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
593 // As a special case, detect code like this:
594 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
595 // and disable this xform in this case, because the code generator will
596 // lower the call to fabs into inline code.
597 if (BB == &F->getEntryBlock() &&
598 FirstNonDbg(BB->front()) == CI &&
599 FirstNonDbg(std::next(BB->begin())) == TI &&
600 CI->getCalledFunction() &&
601 !TTI->isLoweredToCall(CI->getCalledFunction())) {
602 // A single-block function with just a call and a return. Check that
603 // the arguments match.
604 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
605 E = CallSite(CI).arg_end();
606 Function::arg_iterator FI = F->arg_begin(),
608 for (; I != E && FI != FE; ++I, ++FI)
609 if (*I != &*FI) break;
610 if (I == E && FI == FE)
617 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
618 BasicBlock *&OldEntry,
619 bool &TailCallsAreMarkedTail,
620 SmallVectorImpl<PHINode *> &ArgumentPHIs,
621 bool CannotTailCallElimCallsMarkedTail) {
622 // If we are introducing accumulator recursion to eliminate operations after
623 // the call instruction that are both associative and commutative, the initial
624 // value for the accumulator is placed in this variable. If this value is set
625 // then we actually perform accumulator recursion elimination instead of
626 // simple tail recursion elimination. If the operation is an LLVM instruction
627 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
628 // we are handling the case when the return instruction returns a constant C
629 // which is different to the constant returned by other return instructions
630 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
631 // special case of accumulator recursion, the operation being "return C".
632 Value *AccumulatorRecursionEliminationInitVal = nullptr;
633 Instruction *AccumulatorRecursionInstr = nullptr;
635 // Ok, we found a potential tail call. We can currently only transform the
636 // tail call if all of the instructions between the call and the return are
637 // movable to above the call itself, leaving the call next to the return.
638 // Check that this is the case now.
639 BasicBlock::iterator BBI = CI;
640 for (++BBI; &*BBI != Ret; ++BBI) {
641 if (CanMoveAboveCall(BBI, CI)) continue;
643 // If we can't move the instruction above the call, it might be because it
644 // is an associative and commutative operation that could be transformed
645 // using accumulator recursion elimination. Check to see if this is the
646 // case, and if so, remember the initial accumulator value for later.
647 if ((AccumulatorRecursionEliminationInitVal =
648 CanTransformAccumulatorRecursion(BBI, CI))) {
649 // Yes, this is accumulator recursion. Remember which instruction
651 AccumulatorRecursionInstr = BBI;
653 return false; // Otherwise, we cannot eliminate the tail recursion!
657 // We can only transform call/return pairs that either ignore the return value
658 // of the call and return void, ignore the value of the call and return a
659 // constant, return the value returned by the tail call, or that are being
660 // accumulator recursion variable eliminated.
661 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
662 !isa<UndefValue>(Ret->getReturnValue()) &&
663 AccumulatorRecursionEliminationInitVal == nullptr &&
664 !getCommonReturnValue(nullptr, CI)) {
665 // One case remains that we are able to handle: the current return
666 // instruction returns a constant, and all other return instructions
667 // return a different constant.
668 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
669 return false; // Current return instruction does not return a constant.
670 // Check that all other return instructions return a common constant. If
671 // so, record it in AccumulatorRecursionEliminationInitVal.
672 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
673 if (!AccumulatorRecursionEliminationInitVal)
677 BasicBlock *BB = Ret->getParent();
678 Function *F = BB->getParent();
680 emitOptimizationRemark(F->getContext(), "tailcallelim", *F, CI->getDebugLoc(),
681 "transforming tail recursion to loop");
683 // OK! We can transform this tail call. If this is the first one found,
684 // create the new entry block, allowing us to branch back to the old entry.
686 OldEntry = &F->getEntryBlock();
687 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
688 NewEntry->takeName(OldEntry);
689 OldEntry->setName("tailrecurse");
690 BranchInst::Create(OldEntry, NewEntry);
692 // If this tail call is marked 'tail' and if there are any allocas in the
693 // entry block, move them up to the new entry block.
694 TailCallsAreMarkedTail = CI->isTailCall();
695 if (TailCallsAreMarkedTail)
696 // Move all fixed sized allocas from OldEntry to NewEntry.
697 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
698 NEBI = NewEntry->begin(); OEBI != E; )
699 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
700 if (isa<ConstantInt>(AI->getArraySize()))
701 AI->moveBefore(NEBI);
703 // Now that we have created a new block, which jumps to the entry
704 // block, insert a PHI node for each argument of the function.
705 // For now, we initialize each PHI to only have the real arguments
706 // which are passed in.
707 Instruction *InsertPos = OldEntry->begin();
708 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
710 PHINode *PN = PHINode::Create(I->getType(), 2,
711 I->getName() + ".tr", InsertPos);
712 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
713 PN->addIncoming(I, NewEntry);
714 ArgumentPHIs.push_back(PN);
718 // If this function has self recursive calls in the tail position where some
719 // are marked tail and some are not, only transform one flavor or another. We
720 // have to choose whether we move allocas in the entry block to the new entry
721 // block or not, so we can't make a good choice for both. NOTE: We could do
722 // slightly better here in the case that the function has no entry block
724 if (TailCallsAreMarkedTail && !CI->isTailCall())
727 // Ok, now that we know we have a pseudo-entry block WITH all of the
728 // required PHI nodes, add entries into the PHI node for the actual
729 // parameters passed into the tail-recursive call.
730 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
731 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
733 // If we are introducing an accumulator variable to eliminate the recursion,
734 // do so now. Note that we _know_ that no subsequent tail recursion
735 // eliminations will happen on this function because of the way the
736 // accumulator recursion predicate is set up.
738 if (AccumulatorRecursionEliminationInitVal) {
739 Instruction *AccRecInstr = AccumulatorRecursionInstr;
740 // Start by inserting a new PHI node for the accumulator.
741 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
743 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(),
744 std::distance(PB, PE) + 1,
745 "accumulator.tr", OldEntry->begin());
747 // Loop over all of the predecessors of the tail recursion block. For the
748 // real entry into the function we seed the PHI with the initial value,
749 // computed earlier. For any other existing branches to this block (due to
750 // other tail recursions eliminated) the accumulator is not modified.
751 // Because we haven't added the branch in the current block to OldEntry yet,
752 // it will not show up as a predecessor.
753 for (pred_iterator PI = PB; PI != PE; ++PI) {
755 if (P == &F->getEntryBlock())
756 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
758 AccPN->addIncoming(AccPN, P);
762 // Add an incoming argument for the current block, which is computed by
763 // our associative and commutative accumulator instruction.
764 AccPN->addIncoming(AccRecInstr, BB);
766 // Next, rewrite the accumulator recursion instruction so that it does not
767 // use the result of the call anymore, instead, use the PHI node we just
769 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
771 // Add an incoming argument for the current block, which is just the
772 // constant returned by the current return instruction.
773 AccPN->addIncoming(Ret->getReturnValue(), BB);
776 // Finally, rewrite any return instructions in the program to return the PHI
777 // node instead of the "initval" that they do currently. This loop will
778 // actually rewrite the return value we are destroying, but that's ok.
779 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
780 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
781 RI->setOperand(0, AccPN);
785 // Now that all of the PHI nodes are in place, remove the call and
786 // ret instructions, replacing them with an unconditional branch.
787 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
788 NewBI->setDebugLoc(CI->getDebugLoc());
790 BB->getInstList().erase(Ret); // Remove return.
791 BB->getInstList().erase(CI); // Remove call.
796 bool TailCallElim::FoldReturnAndProcessPred(BasicBlock *BB,
797 ReturnInst *Ret, BasicBlock *&OldEntry,
798 bool &TailCallsAreMarkedTail,
799 SmallVectorImpl<PHINode *> &ArgumentPHIs,
800 bool CannotTailCallElimCallsMarkedTail) {
803 // If the return block contains nothing but the return and PHI's,
804 // there might be an opportunity to duplicate the return in its
805 // predecessors and perform TRC there. Look for predecessors that end
806 // in unconditional branch and recursive call(s).
807 SmallVector<BranchInst*, 8> UncondBranchPreds;
808 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
809 BasicBlock *Pred = *PI;
810 TerminatorInst *PTI = Pred->getTerminator();
811 if (BranchInst *BI = dyn_cast<BranchInst>(PTI))
812 if (BI->isUnconditional())
813 UncondBranchPreds.push_back(BI);
816 while (!UncondBranchPreds.empty()) {
817 BranchInst *BI = UncondBranchPreds.pop_back_val();
818 BasicBlock *Pred = BI->getParent();
819 if (CallInst *CI = FindTRECandidate(BI, CannotTailCallElimCallsMarkedTail)){
820 DEBUG(dbgs() << "FOLDING: " << *BB
821 << "INTO UNCOND BRANCH PRED: " << *Pred);
822 EliminateRecursiveTailCall(CI, FoldReturnIntoUncondBranch(Ret, BB, Pred),
823 OldEntry, TailCallsAreMarkedTail, ArgumentPHIs,
824 CannotTailCallElimCallsMarkedTail);
834 TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
835 bool &TailCallsAreMarkedTail,
836 SmallVectorImpl<PHINode *> &ArgumentPHIs,
837 bool CannotTailCallElimCallsMarkedTail) {
838 CallInst *CI = FindTRECandidate(Ret, CannotTailCallElimCallsMarkedTail);
842 return EliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail,
844 CannotTailCallElimCallsMarkedTail);