1 //===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file transforms calls of the current function (self recursion) followed
11 // by a return instruction with a branch to the entry of the function, creating
12 // a loop. This pass also implements the following extensions to the basic
15 // 1. Trivial instructions between the call and return do not prevent the
16 // transformation from taking place, though currently the analysis cannot
17 // support moving any really useful instructions (only dead ones).
18 // 2. This pass transforms functions that are prevented from being tail
19 // recursive by an associative and commutative expression to use an
20 // accumulator variable, thus compiling the typical naive factorial or
21 // 'fib' implementation into efficient code.
22 // 3. TRE is performed if the function returns void, if the return
23 // returns the result returned by the call, or if the function returns a
24 // run-time constant on all exits from the function. It is possible, though
25 // unlikely, that the return returns something else (like constant 0), and
26 // can still be TRE'd. It can be TRE'd if ALL OTHER return instructions in
27 // the function return the exact same value.
28 // 4. If it can prove that callees do not access their caller stack frame,
29 // they are marked as eligible for tail call elimination (by the code
32 // There are several improvements that could be made:
34 // 1. If the function has any alloca instructions, these instructions will be
35 // moved out of the entry block of the function, causing them to be
36 // evaluated each time through the tail recursion. Safely keeping allocas
37 // in the entry block requires analysis to proves that the tail-called
38 // function does not read or write the stack object.
39 // 2. Tail recursion is only performed if the call immediately precedes the
40 // return instruction. It's possible that there could be a jump between
41 // the call and the return.
42 // 3. There can be intervening operations between the call and the return that
43 // prevent the TRE from occurring. For example, there could be GEP's and
44 // stores to memory that will not be read or written by the call. This
45 // requires some substantial analysis (such as with DSA) to prove safe to
46 // move ahead of the call, but doing so could allow many more TREs to be
47 // performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark.
48 // 4. The algorithm we use to detect if callees access their caller stack
49 // frames is very primitive.
51 //===----------------------------------------------------------------------===//
53 #include "llvm/Transforms/Scalar.h"
54 #include "llvm/ADT/STLExtras.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/Analysis/CaptureTracking.h"
58 #include "llvm/Analysis/CFG.h"
59 #include "llvm/Analysis/InlineCost.h"
60 #include "llvm/Analysis/InstructionSimplify.h"
61 #include "llvm/Analysis/Loads.h"
62 #include "llvm/Analysis/TargetTransformInfo.h"
63 #include "llvm/IR/CFG.h"
64 #include "llvm/IR/CallSite.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DerivedTypes.h"
67 #include "llvm/IR/Function.h"
68 #include "llvm/IR/Instructions.h"
69 #include "llvm/IR/IntrinsicInst.h"
70 #include "llvm/IR/Module.h"
71 #include "llvm/IR/ValueHandle.h"
72 #include "llvm/Pass.h"
73 #include "llvm/Support/Debug.h"
74 #include "llvm/Support/raw_ostream.h"
75 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
76 #include "llvm/Transforms/Utils/Local.h"
79 #define DEBUG_TYPE "tailcallelim"
81 STATISTIC(NumEliminated, "Number of tail calls removed");
82 STATISTIC(NumRetDuped, "Number of return duplicated");
83 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
86 struct TailCallElim : public FunctionPass {
87 const TargetTransformInfo *TTI;
89 static char ID; // Pass identification, replacement for typeid
90 TailCallElim() : FunctionPass(ID) {
91 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
94 void getAnalysisUsage(AnalysisUsage &AU) const override;
96 bool runOnFunction(Function &F) override;
99 bool runTRE(Function &F);
100 bool markTails(Function &F, bool &AllCallsAreTailCalls);
102 CallInst *FindTRECandidate(Instruction *I,
103 bool CannotTailCallElimCallsMarkedTail);
104 bool EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
105 BasicBlock *&OldEntry,
106 bool &TailCallsAreMarkedTail,
107 SmallVectorImpl<PHINode *> &ArgumentPHIs,
108 bool CannotTailCallElimCallsMarkedTail);
109 bool FoldReturnAndProcessPred(BasicBlock *BB,
110 ReturnInst *Ret, BasicBlock *&OldEntry,
111 bool &TailCallsAreMarkedTail,
112 SmallVectorImpl<PHINode *> &ArgumentPHIs,
113 bool CannotTailCallElimCallsMarkedTail);
114 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
115 bool &TailCallsAreMarkedTail,
116 SmallVectorImpl<PHINode *> &ArgumentPHIs,
117 bool CannotTailCallElimCallsMarkedTail);
118 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
119 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
123 char TailCallElim::ID = 0;
124 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim",
125 "Tail Call Elimination", false, false)
126 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
127 INITIALIZE_PASS_END(TailCallElim, "tailcallelim",
128 "Tail Call Elimination", false, false)
130 // Public interface to the TailCallElimination pass
131 FunctionPass *llvm::createTailCallEliminationPass() {
132 return new TailCallElim();
135 void TailCallElim::getAnalysisUsage(AnalysisUsage &AU) const {
136 AU.addRequired<TargetTransformInfo>();
139 /// CanTRE - Scan the specified basic block for alloca instructions.
140 /// If it contains any that are variable-sized or not in the entry block,
142 static bool CanTRE(AllocaInst *AI) {
143 // Because of PR962, we don't TRE allocas outside the entry block.
145 // If this alloca is in the body of the function, or if it is a variable
146 // sized allocation, we cannot tail call eliminate calls marked 'tail'
147 // with this mechanism.
148 BasicBlock *BB = AI->getParent();
149 return BB == &BB->getParent()->getEntryBlock() &&
150 isa<ConstantInt>(AI->getArraySize());
153 bool TailCallElim::runOnFunction(Function &F) {
154 if (skipOptnoneFunction(F))
157 bool AllCallsAreTailCalls = false;
158 bool Modified = markTails(F, AllCallsAreTailCalls);
159 if (AllCallsAreTailCalls)
160 Modified |= runTRE(F);
165 struct AllocaDerivedValueTracker {
166 // Start at a root value and walk its use-def chain to mark calls that use the
167 // value or a derived value in AllocaUsers, and places where it may escape in
169 void walk(Value *Root) {
170 SmallVector<Use *, 32> Worklist;
171 SmallPtrSet<Use *, 32> Visited;
173 auto AddUsesToWorklist = [&](Value *V) {
174 for (auto &U : V->uses()) {
175 if (!Visited.insert(&U))
177 Worklist.push_back(&U);
181 AddUsesToWorklist(Root);
183 while (!Worklist.empty()) {
184 Use *U = Worklist.pop_back_val();
185 Instruction *I = cast<Instruction>(U->getUser());
187 switch (I->getOpcode()) {
188 case Instruction::Call:
189 case Instruction::Invoke: {
191 bool IsNocapture = !CS.isCallee(U) &&
192 CS.doesNotCapture(CS.getArgumentNo(U));
193 callUsesLocalStack(CS, IsNocapture);
195 // If the alloca-derived argument is passed in as nocapture, then it
196 // can't propagate to the call's return. That would be capturing.
201 case Instruction::Load: {
202 // The result of a load is not alloca-derived (unless an alloca has
203 // otherwise escaped, but this is a local analysis).
206 case Instruction::Store: {
207 if (U->getOperandNo() == 0)
208 EscapePoints.insert(I);
209 continue; // Stores have no users to analyze.
211 case Instruction::BitCast:
212 case Instruction::GetElementPtr:
213 case Instruction::PHI:
214 case Instruction::Select:
215 case Instruction::AddrSpaceCast:
218 EscapePoints.insert(I);
222 AddUsesToWorklist(I);
226 void callUsesLocalStack(CallSite CS, bool IsNocapture) {
227 // Add it to the list of alloca users. If it's already there, skip further
229 if (!AllocaUsers.insert(CS.getInstruction()))
232 // If it's nocapture then it can't capture the alloca.
236 // If it can write to memory, it can leak the alloca value.
237 if (!CS.onlyReadsMemory())
238 EscapePoints.insert(CS.getInstruction());
241 SmallPtrSet<Instruction *, 32> AllocaUsers;
242 SmallPtrSet<Instruction *, 32> EscapePoints;
246 bool TailCallElim::markTails(Function &F, bool &AllCallsAreTailCalls) {
247 if (F.callsFunctionThatReturnsTwice())
249 AllCallsAreTailCalls = true;
251 // The local stack holds all alloca instructions and all byval arguments.
252 AllocaDerivedValueTracker Tracker;
253 for (Argument &Arg : F.args()) {
254 if (Arg.hasByValAttr())
259 if (AllocaInst *AI = dyn_cast<AllocaInst>(&I))
263 bool Modified = false;
265 // Track whether a block is reachable after an alloca has escaped. Blocks that
266 // contain the escaping instruction will be marked as being visited without an
267 // escaped alloca, since that is how the block began.
273 DenseMap<BasicBlock *, VisitType> Visited;
275 // We propagate the fact that an alloca has escaped from block to successor.
276 // Visit the blocks that are propagating the escapedness first. To do this, we
277 // maintain two worklists.
278 SmallVector<BasicBlock *, 32> WorklistUnescaped, WorklistEscaped;
280 // We may enter a block and visit it thinking that no alloca has escaped yet,
281 // then see an escape point and go back around a loop edge and come back to
282 // the same block twice. Because of this, we defer setting tail on calls when
283 // we first encounter them in a block. Every entry in this list does not
284 // statically use an alloca via use-def chain analysis, but may find an alloca
285 // through other means if the block turns out to be reachable after an escape
287 SmallVector<CallInst *, 32> DeferredTails;
289 BasicBlock *BB = &F.getEntryBlock();
290 VisitType Escaped = UNESCAPED;
292 for (auto &I : *BB) {
293 if (Tracker.EscapePoints.count(&I))
296 CallInst *CI = dyn_cast<CallInst>(&I);
297 if (!CI || CI->isTailCall())
300 if (CI->doesNotAccessMemory()) {
301 // A call to a readnone function whose arguments are all things computed
302 // outside this function can be marked tail. Even if you stored the
303 // alloca address into a global, a readnone function can't load the
306 // Note that this runs whether we know an alloca has escaped or not. If
307 // it has, then we can't trust Tracker.AllocaUsers to be accurate.
308 bool SafeToTail = true;
309 for (auto &Arg : CI->arg_operands()) {
310 if (isa<Constant>(Arg.getUser()))
312 if (Argument *A = dyn_cast<Argument>(Arg.getUser()))
313 if (!A->hasByValAttr())
319 F.getContext().emitOptimizationRemark(
320 "tailcallelim", F, CI->getDebugLoc(),
321 "found readnone tail call candidate");
328 if (Escaped == UNESCAPED && !Tracker.AllocaUsers.count(CI)) {
329 DeferredTails.push_back(CI);
331 AllCallsAreTailCalls = false;
335 for (auto *SuccBB : make_range(succ_begin(BB), succ_end(BB))) {
336 auto &State = Visited[SuccBB];
337 if (State < Escaped) {
339 if (State == ESCAPED)
340 WorklistEscaped.push_back(SuccBB);
342 WorklistUnescaped.push_back(SuccBB);
346 if (!WorklistEscaped.empty()) {
347 BB = WorklistEscaped.pop_back_val();
351 while (!WorklistUnescaped.empty()) {
352 auto *NextBB = WorklistUnescaped.pop_back_val();
353 if (Visited[NextBB] == UNESCAPED) {
362 for (CallInst *CI : DeferredTails) {
363 if (Visited[CI->getParent()] != ESCAPED) {
364 // If the escape point was part way through the block, calls after the
365 // escape point wouldn't have been put into DeferredTails.
366 F.getContext().emitOptimizationRemark(
367 "tailcallelim", F, CI->getDebugLoc(), "found tail call candidate");
371 AllCallsAreTailCalls = false;
378 bool TailCallElim::runTRE(Function &F) {
379 // If this function is a varargs function, we won't be able to PHI the args
380 // right, so don't even try to convert it...
381 if (F.getFunctionType()->isVarArg()) return false;
383 TTI = &getAnalysis<TargetTransformInfo>();
384 BasicBlock *OldEntry = nullptr;
385 bool TailCallsAreMarkedTail = false;
386 SmallVector<PHINode*, 8> ArgumentPHIs;
387 bool MadeChange = false;
389 // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
390 // marked with the 'tail' attribute, because doing so would cause the stack
391 // size to increase (real TRE would deallocate variable sized allocas, TRE
393 bool CanTRETailMarkedCall = true;
395 // Find dynamic allocas.
396 for (Function::iterator BB = F.begin(), EE = F.end(); BB != EE; ++BB) {
397 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
398 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
399 CanTRETailMarkedCall &= CanTRE(AI);
404 // Change any tail recursive calls to loops.
406 // FIXME: The code generator produces really bad code when an 'escaping
407 // alloca' is changed from being a static alloca to being a dynamic alloca.
408 // Until this is resolved, disable this transformation if that would ever
409 // happen. This bug is PR962.
410 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
411 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
412 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
413 ArgumentPHIs, !CanTRETailMarkedCall);
414 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
415 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
416 TailCallsAreMarkedTail, ArgumentPHIs,
417 !CanTRETailMarkedCall);
418 MadeChange |= Change;
422 // If we eliminated any tail recursions, it's possible that we inserted some
423 // silly PHI nodes which just merge an initial value (the incoming operand)
424 // with themselves. Check to see if we did and clean up our mess if so. This
425 // occurs when a function passes an argument straight through to its tail
427 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
428 PHINode *PN = ArgumentPHIs[i];
430 // If the PHI Node is a dynamic constant, replace it with the value it is.
431 if (Value *PNV = SimplifyInstruction(PN)) {
432 PN->replaceAllUsesWith(PNV);
433 PN->eraseFromParent();
441 /// CanMoveAboveCall - Return true if it is safe to move the specified
442 /// instruction from after the call to before the call, assuming that all
443 /// instructions between the call and this instruction are movable.
445 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
446 // FIXME: We can move load/store/call/free instructions above the call if the
447 // call does not mod/ref the memory location being processed.
448 if (I->mayHaveSideEffects()) // This also handles volatile loads.
451 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
452 // Loads may always be moved above calls without side effects.
453 if (CI->mayHaveSideEffects()) {
454 // Non-volatile loads may be moved above a call with side effects if it
455 // does not write to memory and the load provably won't trap.
456 // FIXME: Writes to memory only matter if they may alias the pointer
457 // being loaded from.
458 if (CI->mayWriteToMemory() ||
459 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
465 // Otherwise, if this is a side-effect free instruction, check to make sure
466 // that it does not use the return value of the call. If it doesn't use the
467 // return value of the call, it must only use things that are defined before
468 // the call, or movable instructions between the call and the instruction
470 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
471 if (I->getOperand(i) == CI)
476 // isDynamicConstant - Return true if the specified value is the same when the
477 // return would exit as it was when the initial iteration of the recursive
478 // function was executed.
480 // We currently handle static constants and arguments that are not modified as
481 // part of the recursion.
483 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
484 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
486 // Check to see if this is an immutable argument, if so, the value
487 // will be available to initialize the accumulator.
488 if (Argument *Arg = dyn_cast<Argument>(V)) {
489 // Figure out which argument number this is...
491 Function *F = CI->getParent()->getParent();
492 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
495 // If we are passing this argument into call as the corresponding
496 // argument operand, then the argument is dynamically constant.
497 // Otherwise, we cannot transform this function safely.
498 if (CI->getArgOperand(ArgNo) == Arg)
502 // Switch cases are always constant integers. If the value is being switched
503 // on and the return is only reachable from one of its cases, it's
504 // effectively constant.
505 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
506 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
507 if (SI->getCondition() == V)
508 return SI->getDefaultDest() != RI->getParent();
510 // Not a constant or immutable argument, we can't safely transform.
514 // getCommonReturnValue - Check to see if the function containing the specified
515 // tail call consistently returns the same runtime-constant value at all exit
516 // points except for IgnoreRI. If so, return the returned value.
518 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
519 Function *F = CI->getParent()->getParent();
520 Value *ReturnedValue = nullptr;
522 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) {
523 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator());
524 if (RI == nullptr || RI == IgnoreRI) continue;
526 // We can only perform this transformation if the value returned is
527 // evaluatable at the start of the initial invocation of the function,
528 // instead of at the end of the evaluation.
530 Value *RetOp = RI->getOperand(0);
531 if (!isDynamicConstant(RetOp, CI, RI))
534 if (ReturnedValue && RetOp != ReturnedValue)
535 return nullptr; // Cannot transform if differing values are returned.
536 ReturnedValue = RetOp;
538 return ReturnedValue;
541 /// CanTransformAccumulatorRecursion - If the specified instruction can be
542 /// transformed using accumulator recursion elimination, return the constant
543 /// which is the start of the accumulator value. Otherwise return null.
545 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
547 if (!I->isAssociative() || !I->isCommutative()) return nullptr;
548 assert(I->getNumOperands() == 2 &&
549 "Associative/commutative operations should have 2 args!");
551 // Exactly one operand should be the result of the call instruction.
552 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
553 (I->getOperand(0) != CI && I->getOperand(1) != CI))
556 // The only user of this instruction we allow is a single return instruction.
557 if (!I->hasOneUse() || !isa<ReturnInst>(I->user_back()))
560 // Ok, now we have to check all of the other return instructions in this
561 // function. If they return non-constants or differing values, then we cannot
562 // transform the function safely.
563 return getCommonReturnValue(cast<ReturnInst>(I->user_back()), CI);
566 static Instruction *FirstNonDbg(BasicBlock::iterator I) {
567 while (isa<DbgInfoIntrinsic>(I))
573 TailCallElim::FindTRECandidate(Instruction *TI,
574 bool CannotTailCallElimCallsMarkedTail) {
575 BasicBlock *BB = TI->getParent();
576 Function *F = BB->getParent();
578 if (&BB->front() == TI) // Make sure there is something before the terminator.
581 // Scan backwards from the return, checking to see if there is a tail call in
582 // this block. If so, set CI to it.
583 CallInst *CI = nullptr;
584 BasicBlock::iterator BBI = TI;
586 CI = dyn_cast<CallInst>(BBI);
587 if (CI && CI->getCalledFunction() == F)
590 if (BBI == BB->begin())
591 return nullptr; // Didn't find a potential tail call.
595 // If this call is marked as a tail call, and if there are dynamic allocas in
596 // the function, we cannot perform this optimization.
597 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
600 // As a special case, detect code like this:
601 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
602 // and disable this xform in this case, because the code generator will
603 // lower the call to fabs into inline code.
604 if (BB == &F->getEntryBlock() &&
605 FirstNonDbg(BB->front()) == CI &&
606 FirstNonDbg(std::next(BB->begin())) == TI &&
607 CI->getCalledFunction() &&
608 !TTI->isLoweredToCall(CI->getCalledFunction())) {
609 // A single-block function with just a call and a return. Check that
610 // the arguments match.
611 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
612 E = CallSite(CI).arg_end();
613 Function::arg_iterator FI = F->arg_begin(),
615 for (; I != E && FI != FE; ++I, ++FI)
616 if (*I != &*FI) break;
617 if (I == E && FI == FE)
624 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
625 BasicBlock *&OldEntry,
626 bool &TailCallsAreMarkedTail,
627 SmallVectorImpl<PHINode *> &ArgumentPHIs,
628 bool CannotTailCallElimCallsMarkedTail) {
629 // If we are introducing accumulator recursion to eliminate operations after
630 // the call instruction that are both associative and commutative, the initial
631 // value for the accumulator is placed in this variable. If this value is set
632 // then we actually perform accumulator recursion elimination instead of
633 // simple tail recursion elimination. If the operation is an LLVM instruction
634 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
635 // we are handling the case when the return instruction returns a constant C
636 // which is different to the constant returned by other return instructions
637 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
638 // special case of accumulator recursion, the operation being "return C".
639 Value *AccumulatorRecursionEliminationInitVal = nullptr;
640 Instruction *AccumulatorRecursionInstr = nullptr;
642 // Ok, we found a potential tail call. We can currently only transform the
643 // tail call if all of the instructions between the call and the return are
644 // movable to above the call itself, leaving the call next to the return.
645 // Check that this is the case now.
646 BasicBlock::iterator BBI = CI;
647 for (++BBI; &*BBI != Ret; ++BBI) {
648 if (CanMoveAboveCall(BBI, CI)) continue;
650 // If we can't move the instruction above the call, it might be because it
651 // is an associative and commutative operation that could be transformed
652 // using accumulator recursion elimination. Check to see if this is the
653 // case, and if so, remember the initial accumulator value for later.
654 if ((AccumulatorRecursionEliminationInitVal =
655 CanTransformAccumulatorRecursion(BBI, CI))) {
656 // Yes, this is accumulator recursion. Remember which instruction
658 AccumulatorRecursionInstr = BBI;
660 return false; // Otherwise, we cannot eliminate the tail recursion!
664 // We can only transform call/return pairs that either ignore the return value
665 // of the call and return void, ignore the value of the call and return a
666 // constant, return the value returned by the tail call, or that are being
667 // accumulator recursion variable eliminated.
668 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
669 !isa<UndefValue>(Ret->getReturnValue()) &&
670 AccumulatorRecursionEliminationInitVal == nullptr &&
671 !getCommonReturnValue(nullptr, CI)) {
672 // One case remains that we are able to handle: the current return
673 // instruction returns a constant, and all other return instructions
674 // return a different constant.
675 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
676 return false; // Current return instruction does not return a constant.
677 // Check that all other return instructions return a common constant. If
678 // so, record it in AccumulatorRecursionEliminationInitVal.
679 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
680 if (!AccumulatorRecursionEliminationInitVal)
684 BasicBlock *BB = Ret->getParent();
685 Function *F = BB->getParent();
687 F->getContext().emitOptimizationRemark(
688 "tailcallelim", *F, CI->getDebugLoc(),
689 "transforming tail recursion to loop");
691 // OK! We can transform this tail call. If this is the first one found,
692 // create the new entry block, allowing us to branch back to the old entry.
694 OldEntry = &F->getEntryBlock();
695 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
696 NewEntry->takeName(OldEntry);
697 OldEntry->setName("tailrecurse");
698 BranchInst::Create(OldEntry, NewEntry);
700 // If this tail call is marked 'tail' and if there are any allocas in the
701 // entry block, move them up to the new entry block.
702 TailCallsAreMarkedTail = CI->isTailCall();
703 if (TailCallsAreMarkedTail)
704 // Move all fixed sized allocas from OldEntry to NewEntry.
705 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
706 NEBI = NewEntry->begin(); OEBI != E; )
707 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
708 if (isa<ConstantInt>(AI->getArraySize()))
709 AI->moveBefore(NEBI);
711 // Now that we have created a new block, which jumps to the entry
712 // block, insert a PHI node for each argument of the function.
713 // For now, we initialize each PHI to only have the real arguments
714 // which are passed in.
715 Instruction *InsertPos = OldEntry->begin();
716 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
718 PHINode *PN = PHINode::Create(I->getType(), 2,
719 I->getName() + ".tr", InsertPos);
720 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
721 PN->addIncoming(I, NewEntry);
722 ArgumentPHIs.push_back(PN);
726 // If this function has self recursive calls in the tail position where some
727 // are marked tail and some are not, only transform one flavor or another. We
728 // have to choose whether we move allocas in the entry block to the new entry
729 // block or not, so we can't make a good choice for both. NOTE: We could do
730 // slightly better here in the case that the function has no entry block
732 if (TailCallsAreMarkedTail && !CI->isTailCall())
735 // Ok, now that we know we have a pseudo-entry block WITH all of the
736 // required PHI nodes, add entries into the PHI node for the actual
737 // parameters passed into the tail-recursive call.
738 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
739 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
741 // If we are introducing an accumulator variable to eliminate the recursion,
742 // do so now. Note that we _know_ that no subsequent tail recursion
743 // eliminations will happen on this function because of the way the
744 // accumulator recursion predicate is set up.
746 if (AccumulatorRecursionEliminationInitVal) {
747 Instruction *AccRecInstr = AccumulatorRecursionInstr;
748 // Start by inserting a new PHI node for the accumulator.
749 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
751 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(),
752 std::distance(PB, PE) + 1,
753 "accumulator.tr", OldEntry->begin());
755 // Loop over all of the predecessors of the tail recursion block. For the
756 // real entry into the function we seed the PHI with the initial value,
757 // computed earlier. For any other existing branches to this block (due to
758 // other tail recursions eliminated) the accumulator is not modified.
759 // Because we haven't added the branch in the current block to OldEntry yet,
760 // it will not show up as a predecessor.
761 for (pred_iterator PI = PB; PI != PE; ++PI) {
763 if (P == &F->getEntryBlock())
764 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
766 AccPN->addIncoming(AccPN, P);
770 // Add an incoming argument for the current block, which is computed by
771 // our associative and commutative accumulator instruction.
772 AccPN->addIncoming(AccRecInstr, BB);
774 // Next, rewrite the accumulator recursion instruction so that it does not
775 // use the result of the call anymore, instead, use the PHI node we just
777 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
779 // Add an incoming argument for the current block, which is just the
780 // constant returned by the current return instruction.
781 AccPN->addIncoming(Ret->getReturnValue(), BB);
784 // Finally, rewrite any return instructions in the program to return the PHI
785 // node instead of the "initval" that they do currently. This loop will
786 // actually rewrite the return value we are destroying, but that's ok.
787 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
788 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
789 RI->setOperand(0, AccPN);
793 // Now that all of the PHI nodes are in place, remove the call and
794 // ret instructions, replacing them with an unconditional branch.
795 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
796 NewBI->setDebugLoc(CI->getDebugLoc());
798 BB->getInstList().erase(Ret); // Remove return.
799 BB->getInstList().erase(CI); // Remove call.
804 bool TailCallElim::FoldReturnAndProcessPred(BasicBlock *BB,
805 ReturnInst *Ret, BasicBlock *&OldEntry,
806 bool &TailCallsAreMarkedTail,
807 SmallVectorImpl<PHINode *> &ArgumentPHIs,
808 bool CannotTailCallElimCallsMarkedTail) {
811 // If the return block contains nothing but the return and PHI's,
812 // there might be an opportunity to duplicate the return in its
813 // predecessors and perform TRC there. Look for predecessors that end
814 // in unconditional branch and recursive call(s).
815 SmallVector<BranchInst*, 8> UncondBranchPreds;
816 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
817 BasicBlock *Pred = *PI;
818 TerminatorInst *PTI = Pred->getTerminator();
819 if (BranchInst *BI = dyn_cast<BranchInst>(PTI))
820 if (BI->isUnconditional())
821 UncondBranchPreds.push_back(BI);
824 while (!UncondBranchPreds.empty()) {
825 BranchInst *BI = UncondBranchPreds.pop_back_val();
826 BasicBlock *Pred = BI->getParent();
827 if (CallInst *CI = FindTRECandidate(BI, CannotTailCallElimCallsMarkedTail)){
828 DEBUG(dbgs() << "FOLDING: " << *BB
829 << "INTO UNCOND BRANCH PRED: " << *Pred);
830 EliminateRecursiveTailCall(CI, FoldReturnIntoUncondBranch(Ret, BB, Pred),
831 OldEntry, TailCallsAreMarkedTail, ArgumentPHIs,
832 CannotTailCallElimCallsMarkedTail);
842 TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
843 bool &TailCallsAreMarkedTail,
844 SmallVectorImpl<PHINode *> &ArgumentPHIs,
845 bool CannotTailCallElimCallsMarkedTail) {
846 CallInst *CI = FindTRECandidate(Ret, CannotTailCallElimCallsMarkedTail);
850 return EliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail,
852 CannotTailCallElimCallsMarkedTail);