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/TailRecursionElimination.h"
54 #include "llvm/ADT/STLExtras.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/Analysis/CFG.h"
58 #include "llvm/Analysis/CaptureTracking.h"
59 #include "llvm/Analysis/GlobalsModRef.h"
60 #include "llvm/Analysis/InlineCost.h"
61 #include "llvm/Analysis/InstructionSimplify.h"
62 #include "llvm/Analysis/Loads.h"
63 #include "llvm/Analysis/TargetTransformInfo.h"
64 #include "llvm/IR/CFG.h"
65 #include "llvm/IR/CallSite.h"
66 #include "llvm/IR/Constants.h"
67 #include "llvm/IR/DataLayout.h"
68 #include "llvm/IR/DerivedTypes.h"
69 #include "llvm/IR/DiagnosticInfo.h"
70 #include "llvm/IR/Function.h"
71 #include "llvm/IR/InstIterator.h"
72 #include "llvm/IR/Instructions.h"
73 #include "llvm/IR/IntrinsicInst.h"
74 #include "llvm/IR/Module.h"
75 #include "llvm/IR/ValueHandle.h"
76 #include "llvm/Pass.h"
77 #include "llvm/Support/Debug.h"
78 #include "llvm/Support/raw_ostream.h"
79 #include "llvm/Transforms/Scalar.h"
80 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
81 #include "llvm/Transforms/Utils/Local.h"
84 #define DEBUG_TYPE "tailcallelim"
86 STATISTIC(NumEliminated, "Number of tail calls removed");
87 STATISTIC(NumRetDuped, "Number of return duplicated");
88 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
90 /// \brief Scan the specified function for alloca instructions.
91 /// If it contains any dynamic allocas, returns false.
92 static bool canTRE(Function &F) {
93 // Because of PR962, we don't TRE dynamic allocas.
94 return llvm::all_of(instructions(F), [](Instruction &I) {
95 auto *AI = dyn_cast<AllocaInst>(&I);
96 return !AI || AI->isStaticAlloca();
101 struct AllocaDerivedValueTracker {
102 // Start at a root value and walk its use-def chain to mark calls that use the
103 // value or a derived value in AllocaUsers, and places where it may escape in
105 void walk(Value *Root) {
106 SmallVector<Use *, 32> Worklist;
107 SmallPtrSet<Use *, 32> Visited;
109 auto AddUsesToWorklist = [&](Value *V) {
110 for (auto &U : V->uses()) {
111 if (!Visited.insert(&U).second)
113 Worklist.push_back(&U);
117 AddUsesToWorklist(Root);
119 while (!Worklist.empty()) {
120 Use *U = Worklist.pop_back_val();
121 Instruction *I = cast<Instruction>(U->getUser());
123 switch (I->getOpcode()) {
124 case Instruction::Call:
125 case Instruction::Invoke: {
128 CS.isDataOperand(U) && CS.doesNotCapture(CS.getDataOperandNo(U));
129 callUsesLocalStack(CS, IsNocapture);
131 // If the alloca-derived argument is passed in as nocapture, then it
132 // can't propagate to the call's return. That would be capturing.
137 case Instruction::Load: {
138 // The result of a load is not alloca-derived (unless an alloca has
139 // otherwise escaped, but this is a local analysis).
142 case Instruction::Store: {
143 if (U->getOperandNo() == 0)
144 EscapePoints.insert(I);
145 continue; // Stores have no users to analyze.
147 case Instruction::BitCast:
148 case Instruction::GetElementPtr:
149 case Instruction::PHI:
150 case Instruction::Select:
151 case Instruction::AddrSpaceCast:
154 EscapePoints.insert(I);
158 AddUsesToWorklist(I);
162 void callUsesLocalStack(CallSite CS, bool IsNocapture) {
163 // Add it to the list of alloca users.
164 AllocaUsers.insert(CS.getInstruction());
166 // If it's nocapture then it can't capture this alloca.
170 // If it can write to memory, it can leak the alloca value.
171 if (!CS.onlyReadsMemory())
172 EscapePoints.insert(CS.getInstruction());
175 SmallPtrSet<Instruction *, 32> AllocaUsers;
176 SmallPtrSet<Instruction *, 32> EscapePoints;
180 static bool markTails(Function &F, bool &AllCallsAreTailCalls) {
181 if (F.callsFunctionThatReturnsTwice())
183 AllCallsAreTailCalls = true;
185 // The local stack holds all alloca instructions and all byval arguments.
186 AllocaDerivedValueTracker Tracker;
187 for (Argument &Arg : F.args()) {
188 if (Arg.hasByValAttr())
193 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I))
197 bool Modified = false;
199 // Track whether a block is reachable after an alloca has escaped. Blocks that
200 // contain the escaping instruction will be marked as being visited without an
201 // escaped alloca, since that is how the block began.
207 DenseMap<BasicBlock *, VisitType> Visited;
209 // We propagate the fact that an alloca has escaped from block to successor.
210 // Visit the blocks that are propagating the escapedness first. To do this, we
211 // maintain two worklists.
212 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
214 // We may enter a block and visit it thinking that no alloca has escaped yet,
215 // then see an escape point and go back around a loop edge and come back to
216 // the same block twice. Because of this, we defer setting tail on calls when
217 // we first encounter them in a block. Every entry in this list does not
218 // statically use an alloca via use-def chain analysis, but may find an alloca
219 // through other means if the block turns out to be reachable after an escape
221 SmallVector<CallInst *, 32> DeferredTails;
223 BasicBlock *BB = &F.getEntryBlock();
224 VisitType Escaped = UNESCAPED;
226 for (auto &I : *BB) {
227 if (Tracker.EscapePoints.count(&I))
230 CallInst *CI = dyn_cast<CallInst>(&I);
231 if (!CI || CI->isTailCall())
234 bool IsNoTail = CI->isNoTailCall() || CI->hasOperandBundles();
236 if (!IsNoTail && CI->doesNotAccessMemory()) {
237 // A call to a readnone function whose arguments are all things computed
238 // outside this function can be marked tail. Even if you stored the
239 // alloca address into a global, a readnone function can't load the
242 // Note that this runs whether we know an alloca has escaped or not. If
243 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
244 bool SafeToTail = true;
245 for (auto &Arg : CI->arg_operands()) {
246 if (isa<Constant>(Arg.getUser()))
248 if (Argument *A = dyn_cast<Argument>(Arg.getUser()))
249 if (!A->hasByValAttr())
255 emitOptimizationRemark(
256 F.getContext(), "tailcallelim", F, CI->getDebugLoc(),
257 "marked this readnone call a tail call candidate");
264 if (!IsNoTail && Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) {
265 DeferredTails.push_back(CI);
267 AllCallsAreTailCalls = false;
271 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) {
272 auto &State = Visited[SuccBB];
273 if (State < Escaped) {
275 if (State == ESCAPED)
276 WorklistEscaped.push_back(SuccBB);
278 WorklistUnescaped.push_back(SuccBB);
282 if (!WorklistEscaped.empty()) {
283 BB = WorklistEscaped.pop_back_val();
287 while (!WorklistUnescaped.empty()) {
288 auto *NextBB = WorklistUnescaped.pop_back_val();
289 if (Visited[NextBB] == UNESCAPED) {
298 for (CallInst *CI : DeferredTails) {
299 if (Visited[CI->getParent()] != ESCAPED) {
300 // If the escape point was part way through the block, calls after the
301 // escape point wouldn't have been put into DeferredTails.
302 emitOptimizationRemark(F.getContext(), "tailcallelim", F,
304 "marked this call a tail call candidate");
308 AllCallsAreTailCalls = false;
315 /// Return true if it is safe to move the specified
316 /// instruction from after the call to before the call, assuming that all
317 /// instructions between the call and this instruction are movable.
319 static bool canMoveAboveCall(Instruction *I, CallInst *CI, AliasAnalysis *AA) {
320 // FIXME: We can move load/store/call/free instructions above the call if the
321 // call does not mod/ref the memory location being processed.
322 if (I->mayHaveSideEffects()) // This also handles volatile loads.
325 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
326 // Loads may always be moved above calls without side effects.
327 if (CI->mayHaveSideEffects()) {
328 // Non-volatile loads may be moved above a call with side effects if it
329 // does not write to memory and the load provably won't trap.
330 // Writes to memory only matter if they may alias the pointer
331 // being loaded from.
332 const DataLayout &DL = L->getModule()->getDataLayout();
333 if ((AA->getModRefInfo(CI, MemoryLocation::get(L)) & MRI_Mod) ||
334 !isSafeToLoadUnconditionally(L->getPointerOperand(),
335 L->getAlignment(), DL, L))
340 // Otherwise, if this is a side-effect free instruction, check to make sure
341 // that it does not use the return value of the call. If it doesn't use the
342 // return value of the call, it must only use things that are defined before
343 // the call, or movable instructions between the call and the instruction
345 return !is_contained(I->operands(), CI);
348 /// Return true if the specified value is the same when the return would exit
349 /// as it was when the initial iteration of the recursive function was executed.
351 /// We currently handle static constants and arguments that are not modified as
352 /// part of the recursion.
353 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
354 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
356 // Check to see if this is an immutable argument, if so, the value
357 // will be available to initialize the accumulator.
358 if (Argument *Arg = dyn_cast<Argument>(V)) {
359 // Figure out which argument number this is...
361 Function *F = CI->getParent()->getParent();
362 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
365 // If we are passing this argument into call as the corresponding
366 // argument operand, then the argument is dynamically constant.
367 // Otherwise, we cannot transform this function safely.
368 if (CI->getArgOperand(ArgNo) == Arg)
372 // Switch cases are always constant integers. If the value is being switched
373 // on and the return is only reachable from one of its cases, it's
374 // effectively constant.
375 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
376 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
377 if (SI->getCondition() == V)
378 return SI->getDefaultDest() != RI->getParent();
380 // Not a constant or immutable argument, we can't safely transform.
384 /// Check to see if the function containing the specified tail call consistently
385 /// returns the same runtime-constant value at all exit points except for
386 /// IgnoreRI. If so, return the returned value.
387 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
388 Function *F = CI->getParent()->getParent();
389 Value *ReturnedValue = nullptr;
391 for (BasicBlock &BBI : *F) {
392 ReturnInst *RI = dyn_cast<ReturnInst>(BBI.getTerminator());
393 if (RI == nullptr || RI == IgnoreRI) continue;
395 // We can only perform this transformation if the value returned is
396 // evaluatable at the start of the initial invocation of the function,
397 // instead of at the end of the evaluation.
399 Value *RetOp = RI->getOperand(0);
400 if (!isDynamicConstant(RetOp, CI, RI))
403 if (ReturnedValue && RetOp != ReturnedValue)
404 return nullptr; // Cannot transform if differing values are returned.
405 ReturnedValue = RetOp;
407 return ReturnedValue;
410 /// If the specified instruction can be transformed using accumulator recursion
411 /// elimination, return the constant which is the start of the accumulator
412 /// value. Otherwise return null.
413 static Value *canTransformAccumulatorRecursion(Instruction *I, CallInst *CI) {
414 if (!I->isAssociative() || !I->isCommutative()) return nullptr;
415 assert(I->getNumOperands() == 2 &&
416 "Associative/commutative operations should have 2 args!");
418 // Exactly one operand should be the result of the call instruction.
419 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
420 (I->getOperand(0) != CI && I->getOperand(1) != CI))
423 // The only user of this instruction we allow is a single return instruction.
424 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back()))
427 // Ok, now we have to check all of the other return instructions in this
428 // function. If they return non-constants or differing values, then we cannot
429 // transform the function safely.
430 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI);
433 static Instruction *firstNonDbg(BasicBlock::iterator I) {
434 while (isa<DbgInfoIntrinsic>(I))
439 static CallInst *findTRECandidate(Instruction *TI,
440 bool CannotTailCallElimCallsMarkedTail,
441 const TargetTransformInfo *TTI) {
442 BasicBlock *BB = TI->getParent();
443 Function *F = BB->getParent();
445 if (&BB->front() == TI) // Make sure there is something before the terminator.
448 // Scan backwards from the return, checking to see if there is a tail call in
449 // this block. If so, set CI to it.
450 CallInst *CI = nullptr;
451 BasicBlock::iterator BBI(TI);
453 CI = dyn_cast<CallInst>(BBI);
454 if (CI && CI->getCalledFunction() == F)
457 if (BBI == BB->begin())
458 return nullptr; // Didn't find a potential tail call.
462 // If this call is marked as a tail call, and if there are dynamic allocas in
463 // the function, we cannot perform this optimization.
464 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
467 // As a special case, detect code like this:
468 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
469 // and disable this xform in this case, because the code generator will
470 // lower the call to fabs into inline code.
471 if (BB == &F->getEntryBlock() &&
472 firstNonDbg(BB->front().getIterator()) == CI &&
473 firstNonDbg(std::next(BB->begin())) == TI && CI->getCalledFunction() &&
474 !TTI->isLoweredToCall(CI->getCalledFunction())) {
475 // A single-block function with just a call and a return. Check that
476 // the arguments match.
477 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
478 E = CallSite(CI).arg_end();
479 Function::arg_iterator FI = F->arg_begin(),
481 for (; I != E && FI != FE; ++I, ++FI)
482 if (*I != &*FI) break;
483 if (I == E && FI == FE)
490 static bool eliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
491 BasicBlock *&OldEntry,
492 bool &TailCallsAreMarkedTail,
493 SmallVectorImpl<PHINode *> &ArgumentPHIs,
495 // If we are introducing accumulator recursion to eliminate operations after
496 // the call instruction that are both associative and commutative, the initial
497 // value for the accumulator is placed in this variable. If this value is set
498 // then we actually perform accumulator recursion elimination instead of
499 // simple tail recursion elimination. If the operation is an LLVM instruction
500 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
501 // we are handling the case when the return instruction returns a constant C
502 // which is different to the constant returned by other return instructions
503 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
504 // special case of accumulator recursion, the operation being "return C".
505 Value *AccumulatorRecursionEliminationInitVal = nullptr;
506 Instruction *AccumulatorRecursionInstr = nullptr;
508 // Ok, we found a potential tail call. We can currently only transform the
509 // tail call if all of the instructions between the call and the return are
510 // movable to above the call itself, leaving the call next to the return.
511 // Check that this is the case now.
512 BasicBlock::iterator BBI(CI);
513 for (++BBI; &*BBI != Ret; ++BBI) {
514 if (canMoveAboveCall(&*BBI, CI, AA))
517 // If we can't move the instruction above the call, it might be because it
518 // is an associative and commutative operation that could be transformed
519 // using accumulator recursion elimination. Check to see if this is the
520 // case, and if so, remember the initial accumulator value for later.
521 if ((AccumulatorRecursionEliminationInitVal =
522 canTransformAccumulatorRecursion(&*BBI, CI))) {
523 // Yes, this is accumulator recursion. Remember which instruction
525 AccumulatorRecursionInstr = &*BBI;
527 return false; // Otherwise, we cannot eliminate the tail recursion!
531 // We can only transform call/return pairs that either ignore the return value
532 // of the call and return void, ignore the value of the call and return a
533 // constant, return the value returned by the tail call, or that are being
534 // accumulator recursion variable eliminated.
535 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
536 !isa<UndefValue>(Ret->getReturnValue()) &&
537 AccumulatorRecursionEliminationInitVal == nullptr &&
538 !getCommonReturnValue(nullptr, CI)) {
539 // One case remains that we are able to handle: the current return
540 // instruction returns a constant, and all other return instructions
541 // return a different constant.
542 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
543 return false; // Current return instruction does not return a constant.
544 // Check that all other return instructions return a common constant. If
545 // so, record it in AccumulatorRecursionEliminationInitVal.
546 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
547 if (!AccumulatorRecursionEliminationInitVal)
551 BasicBlock *BB = Ret->getParent();
552 Function *F = BB->getParent();
554 emitOptimizationRemark(F->getContext(), "tailcallelim", *F, CI->getDebugLoc(),
555 "transforming tail recursion to loop");
557 // OK! We can transform this tail call. If this is the first one found,
558 // create the new entry block, allowing us to branch back to the old entry.
560 OldEntry = &F->getEntryBlock();
561 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
562 NewEntry->takeName(OldEntry);
563 OldEntry->setName("tailrecurse");
564 BranchInst::Create(OldEntry, NewEntry);
566 // If this tail call is marked 'tail' and if there are any allocas in the
567 // entry block, move them up to the new entry block.
568 TailCallsAreMarkedTail = CI->isTailCall();
569 if (TailCallsAreMarkedTail)
570 // Move all fixed sized allocas from OldEntry to NewEntry.
571 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
572 NEBI = NewEntry->begin(); OEBI != E; )
573 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
574 if (isa<ConstantInt>(AI->getArraySize()))
575 AI->moveBefore(&*NEBI);
577 // Now that we have created a new block, which jumps to the entry
578 // block, insert a PHI node for each argument of the function.
579 // For now, we initialize each PHI to only have the real arguments
580 // which are passed in.
581 Instruction *InsertPos = &OldEntry->front();
582 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
584 PHINode *PN = PHINode::Create(I->getType(), 2,
585 I->getName() + ".tr", InsertPos);
586 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
587 PN->addIncoming(&*I, NewEntry);
588 ArgumentPHIs.push_back(PN);
592 // If this function has self recursive calls in the tail position where some
593 // are marked tail and some are not, only transform one flavor or another. We
594 // have to choose whether we move allocas in the entry block to the new entry
595 // block or not, so we can't make a good choice for both. NOTE: We could do
596 // slightly better here in the case that the function has no entry block
598 if (TailCallsAreMarkedTail && !CI->isTailCall())
601 // Ok, now that we know we have a pseudo-entry block WITH all of the
602 // required PHI nodes, add entries into the PHI node for the actual
603 // parameters passed into the tail-recursive call.
604 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
605 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
607 // If we are introducing an accumulator variable to eliminate the recursion,
608 // do so now. Note that we _know_ that no subsequent tail recursion
609 // eliminations will happen on this function because of the way the
610 // accumulator recursion predicate is set up.
612 if (AccumulatorRecursionEliminationInitVal) {
613 Instruction *AccRecInstr = AccumulatorRecursionInstr;
614 // Start by inserting a new PHI node for the accumulator.
615 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
616 PHINode *AccPN = PHINode::Create(
617 AccumulatorRecursionEliminationInitVal->getType(),
618 std::distance(PB, PE) + 1, "accumulator.tr", &OldEntry->front());
620 // Loop over all of the predecessors of the tail recursion block. For the
621 // real entry into the function we seed the PHI with the initial value,
622 // computed earlier. For any other existing branches to this block (due to
623 // other tail recursions eliminated) the accumulator is not modified.
624 // Because we haven't added the branch in the current block to OldEntry yet,
625 // it will not show up as a predecessor.
626 for (pred_iterator PI = PB; PI != PE; ++PI) {
628 if (P == &F->getEntryBlock())
629 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
631 AccPN->addIncoming(AccPN, P);
635 // Add an incoming argument for the current block, which is computed by
636 // our associative and commutative accumulator instruction.
637 AccPN->addIncoming(AccRecInstr, BB);
639 // Next, rewrite the accumulator recursion instruction so that it does not
640 // use the result of the call anymore, instead, use the PHI node we just
642 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
644 // Add an incoming argument for the current block, which is just the
645 // constant returned by the current return instruction.
646 AccPN->addIncoming(Ret->getReturnValue(), BB);
649 // Finally, rewrite any return instructions in the program to return the PHI
650 // node instead of the "initval" that they do currently. This loop will
651 // actually rewrite the return value we are destroying, but that's ok.
652 for (BasicBlock &BBI : *F)
653 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI.getTerminator()))
654 RI->setOperand(0, AccPN);
658 // Now that all of the PHI nodes are in place, remove the call and
659 // ret instructions, replacing them with an unconditional branch.
660 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
661 NewBI->setDebugLoc(CI->getDebugLoc());
663 BB->getInstList().erase(Ret); // Remove return.
664 BB->getInstList().erase(CI); // Remove call.
669 static bool foldReturnAndProcessPred(BasicBlock *BB, ReturnInst *Ret,
670 BasicBlock *&OldEntry,
671 bool &TailCallsAreMarkedTail,
672 SmallVectorImpl<PHINode *> &ArgumentPHIs,
673 bool CannotTailCallElimCallsMarkedTail,
674 const TargetTransformInfo *TTI,
678 // Make sure this block is a trivial return block.
679 assert(BB->getFirstNonPHIOrDbg() == Ret &&
680 "Trying to fold non-trivial return block");
682 // If the return block contains nothing but the return and PHI's,
683 // there might be an opportunity to duplicate the return in its
684 // predecessors and perform TRE there. Look for predecessors that end
685 // in unconditional branch and recursive call(s).
686 SmallVector<BranchInst*, 8> UncondBranchPreds;
687 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
688 BasicBlock *Pred = *PI;
689 TerminatorInst *PTI = Pred->getTerminator();
690 if (BranchInst *BI = dyn_cast<BranchInst>(PTI))
691 if (BI->isUnconditional())
692 UncondBranchPreds.push_back(BI);
695 while (!UncondBranchPreds.empty()) {
696 BranchInst *BI = UncondBranchPreds.pop_back_val();
697 BasicBlock *Pred = BI->getParent();
698 if (CallInst *CI = findTRECandidate(BI, CannotTailCallElimCallsMarkedTail, TTI)){
699 DEBUG(dbgs() << "FOLDING: " << *BB
700 << "INTO UNCOND BRANCH PRED: " << *Pred);
701 ReturnInst *RI = FoldReturnIntoUncondBranch(Ret, BB, Pred);
703 // Cleanup: if all predecessors of BB have been eliminated by
704 // FoldReturnIntoUncondBranch, delete it. It is important to empty it,
705 // because the ret instruction in there is still using a value which
706 // eliminateRecursiveTailCall will attempt to remove.
707 if (!BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
708 BB->eraseFromParent();
710 eliminateRecursiveTailCall(CI, RI, OldEntry, TailCallsAreMarkedTail,
720 static bool processReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
721 bool &TailCallsAreMarkedTail,
722 SmallVectorImpl<PHINode *> &ArgumentPHIs,
723 bool CannotTailCallElimCallsMarkedTail,
724 const TargetTransformInfo *TTI,
726 CallInst *CI = findTRECandidate(Ret, CannotTailCallElimCallsMarkedTail, TTI);
730 return eliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail,
734 static bool eliminateTailRecursion(Function &F, const TargetTransformInfo *TTI,
736 if (F.getFnAttribute("disable-tail-calls").getValueAsString() == "true")
739 bool MadeChange = false;
740 bool AllCallsAreTailCalls = false;
741 MadeChange |= markTails(F, AllCallsAreTailCalls);
742 if (!AllCallsAreTailCalls)
745 // If this function is a varargs function, we won't be able to PHI the args
746 // right, so don't even try to convert it...
747 if (F.getFunctionType()->isVarArg())
750 BasicBlock *OldEntry = nullptr;
751 bool TailCallsAreMarkedTail = false;
752 SmallVector<PHINode*, 8> ArgumentPHIs;
754 // If false, we cannot perform TRE on tail calls marked with the 'tail'
755 // attribute, because doing so would cause the stack size to increase (real
756 // TRE would deallocate variable sized allocas, TRE doesn't).
757 bool CanTRETailMarkedCall = canTRE(F);
759 // Change any tail recursive calls to loops.
761 // FIXME: The code generator produces really bad code when an 'escaping
762 // alloca' is changed from being a static alloca to being a dynamic alloca.
763 // Until this is resolved, disable this transformation if that would ever
764 // happen. This bug is PR962.
765 for (Function::iterator BBI = F.begin(), E = F.end(); BBI != E; /*in loop*/) {
766 BasicBlock *BB = &*BBI++; // foldReturnAndProcessPred may delete BB.
767 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
769 processReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
770 ArgumentPHIs, !CanTRETailMarkedCall, TTI, AA);
771 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
772 Change = foldReturnAndProcessPred(BB, Ret, OldEntry,
773 TailCallsAreMarkedTail, ArgumentPHIs,
774 !CanTRETailMarkedCall, TTI, AA);
775 MadeChange |= Change;
779 // If we eliminated any tail recursions, it's possible that we inserted some
780 // silly PHI nodes which just merge an initial value (the incoming operand)
781 // with themselves. Check to see if we did and clean up our mess if so. This
782 // occurs when a function passes an argument straight through to its tail
784 for (PHINode *PN : ArgumentPHIs) {
785 // If the PHI Node is a dynamic constant, replace it with the value it is.
786 if (Value *PNV = SimplifyInstruction(PN, F.getParent()->getDataLayout())) {
787 PN->replaceAllUsesWith(PNV);
788 PN->eraseFromParent();
796 struct TailCallElim : public FunctionPass {
797 static char ID; // Pass identification, replacement for typeid
798 TailCallElim() : FunctionPass(ID) {
799 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
802 void getAnalysisUsage(AnalysisUsage &AU) const override {
803 AU.addRequired<TargetTransformInfoWrapperPass>();
804 AU.addRequired<AAResultsWrapperPass>();
805 AU.addPreserved<GlobalsAAWrapperPass>();
808 bool runOnFunction(Function &F) override {
812 return eliminateTailRecursion(
813 F, &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F),
814 &getAnalysis<AAResultsWrapperPass>().getAAResults());
819 char TailCallElim::ID = 0;
820 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim", "Tail Call Elimination",
822 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
823 INITIALIZE_PASS_END(TailCallElim, "tailcallelim", "Tail Call Elimination",
826 // Public interface to the TailCallElimination pass
827 FunctionPass *llvm::createTailCallEliminationPass() {
828 return new TailCallElim();
831 PreservedAnalyses TailCallElimPass::run(Function &F,
832 FunctionAnalysisManager &AM) {
834 TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F);
835 AliasAnalysis &AA = AM.getResult<AAManager>(F);
837 bool Changed = eliminateTailRecursion(F, &TTI, &AA);
840 return PreservedAnalyses::all();
841 PreservedAnalyses PA;
842 PA.preserve<GlobalsAA>();