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 #define DEBUG_TYPE "tailcallelim"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/ADT/STLExtras.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/Analysis/CaptureTracking.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 STATISTIC(NumEliminated, "Number of tail calls removed");
80 STATISTIC(NumRetDuped, "Number of return duplicated");
81 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
84 struct TailCallElim : public FunctionPass {
85 const TargetTransformInfo *TTI;
87 static char ID; // Pass identification, replacement for typeid
88 TailCallElim() : FunctionPass(ID) {
89 initializeTailCallElimPass(*PassRegistry::getPassRegistry());
92 void getAnalysisUsage(AnalysisUsage &AU) const override;
94 bool runOnFunction(Function &F) override;
97 CallInst *FindTRECandidate(Instruction *I,
98 bool CannotTailCallElimCallsMarkedTail);
99 bool EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
100 BasicBlock *&OldEntry,
101 bool &TailCallsAreMarkedTail,
102 SmallVectorImpl<PHINode *> &ArgumentPHIs,
103 bool CannotTailCallElimCallsMarkedTail);
104 bool FoldReturnAndProcessPred(BasicBlock *BB,
105 ReturnInst *Ret, BasicBlock *&OldEntry,
106 bool &TailCallsAreMarkedTail,
107 SmallVectorImpl<PHINode *> &ArgumentPHIs,
108 bool CannotTailCallElimCallsMarkedTail);
109 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
110 bool &TailCallsAreMarkedTail,
111 SmallVectorImpl<PHINode *> &ArgumentPHIs,
112 bool CannotTailCallElimCallsMarkedTail);
113 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
114 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
118 char TailCallElim::ID = 0;
119 INITIALIZE_PASS_BEGIN(TailCallElim, "tailcallelim",
120 "Tail Call Elimination", false, false)
121 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
122 INITIALIZE_PASS_END(TailCallElim, "tailcallelim",
123 "Tail Call Elimination", false, false)
125 // Public interface to the TailCallElimination pass
126 FunctionPass *llvm::createTailCallEliminationPass() {
127 return new TailCallElim();
130 void TailCallElim::getAnalysisUsage(AnalysisUsage &AU) const {
131 AU.addRequired<TargetTransformInfo>();
134 /// CanTRE - Scan the specified basic block for alloca instructions.
135 /// If it contains any that are variable-sized or not in the entry block,
137 static bool CanTRE(AllocaInst *AI) {
138 // Because of PR962, we don't TRE allocas outside the entry block.
140 // If this alloca is in the body of the function, or if it is a variable
141 // sized allocation, we cannot tail call eliminate calls marked 'tail'
142 // with this mechanism.
143 BasicBlock *BB = AI->getParent();
144 return BB == &BB->getParent()->getEntryBlock() &&
145 isa<ConstantInt>(AI->getArraySize());
149 struct AllocaCaptureTracker : public CaptureTracker {
150 AllocaCaptureTracker() : Captured(false) {}
152 void tooManyUses() override { Captured = true; }
154 bool shouldExplore(const Use *U) override {
155 Value *V = U->getUser();
156 if (isa<CallInst>(V) || isa<InvokeInst>(V))
157 UsesAlloca.insert(V);
161 bool captured(const Use *U) override {
162 if (isa<ReturnInst>(U->getUser()))
169 SmallPtrSet<const Value *, 16> UsesAlloca;
171 } // end anonymous namespace
173 bool TailCallElim::runOnFunction(Function &F) {
174 if (skipOptnoneFunction(F))
177 // If this function is a varargs function, we won't be able to PHI the args
178 // right, so don't even try to convert it...
179 if (F.getFunctionType()->isVarArg()) return false;
181 TTI = &getAnalysis<TargetTransformInfo>();
182 BasicBlock *OldEntry = 0;
183 bool TailCallsAreMarkedTail = false;
184 SmallVector<PHINode*, 8> ArgumentPHIs;
185 bool MadeChange = false;
187 // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
188 // marked with the 'tail' attribute, because doing so would cause the stack
189 // size to increase (real TRE would deallocate variable sized allocas, TRE
191 bool CanTRETailMarkedCall = true;
193 // Find calls that can be marked tail.
194 AllocaCaptureTracker ACT;
195 for (Function::iterator BB = F.begin(), EE = F.end(); BB != EE; ++BB) {
196 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
197 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
198 CanTRETailMarkedCall &= CanTRE(AI);
199 PointerMayBeCaptured(AI, &ACT);
200 // If any allocas are captured, exit.
207 // Second pass, change any tail recursive calls to loops.
209 // FIXME: The code generator produces really bad code when an 'escaping
210 // alloca' is changed from being a static alloca to being a dynamic alloca.
211 // Until this is resolved, disable this transformation if that would ever
212 // happen. This bug is PR962.
213 if (ACT.UsesAlloca.empty()) {
214 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
215 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
216 bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
217 ArgumentPHIs, !CanTRETailMarkedCall);
218 if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
219 Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
220 TailCallsAreMarkedTail, ArgumentPHIs,
221 !CanTRETailMarkedCall);
222 MadeChange |= Change;
227 // If we eliminated any tail recursions, it's possible that we inserted some
228 // silly PHI nodes which just merge an initial value (the incoming operand)
229 // with themselves. Check to see if we did and clean up our mess if so. This
230 // occurs when a function passes an argument straight through to its tail
232 if (!ArgumentPHIs.empty()) {
233 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
234 PHINode *PN = ArgumentPHIs[i];
236 // If the PHI Node is a dynamic constant, replace it with the value it is.
237 if (Value *PNV = SimplifyInstruction(PN)) {
238 PN->replaceAllUsesWith(PNV);
239 PN->eraseFromParent();
244 // At this point, we know that the function does not have any captured
245 // allocas. If additionally the function does not call setjmp, mark all calls
246 // in the function that do not access stack memory with the tail keyword. This
247 // implies ensuring that there does not exist any path from a call that takes
248 // in an alloca but does not capture it and the call which we wish to mark
250 if (!F.callsFunctionThatReturnsTwice()) {
251 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
252 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
253 if (CallInst *CI = dyn_cast<CallInst>(I)) {
254 if (!ACT.UsesAlloca.count(CI)) {
267 /// CanMoveAboveCall - Return true if it is safe to move the specified
268 /// instruction from after the call to before the call, assuming that all
269 /// instructions between the call and this instruction are movable.
271 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
272 // FIXME: We can move load/store/call/free instructions above the call if the
273 // call does not mod/ref the memory location being processed.
274 if (I->mayHaveSideEffects()) // This also handles volatile loads.
277 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
278 // Loads may always be moved above calls without side effects.
279 if (CI->mayHaveSideEffects()) {
280 // Non-volatile loads may be moved above a call with side effects if it
281 // does not write to memory and the load provably won't trap.
282 // FIXME: Writes to memory only matter if they may alias the pointer
283 // being loaded from.
284 if (CI->mayWriteToMemory() ||
285 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
291 // Otherwise, if this is a side-effect free instruction, check to make sure
292 // that it does not use the return value of the call. If it doesn't use the
293 // return value of the call, it must only use things that are defined before
294 // the call, or movable instructions between the call and the instruction
296 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
297 if (I->getOperand(i) == CI)
302 // isDynamicConstant - Return true if the specified value is the same when the
303 // return would exit as it was when the initial iteration of the recursive
304 // function was executed.
306 // We currently handle static constants and arguments that are not modified as
307 // part of the recursion.
309 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
310 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
312 // Check to see if this is an immutable argument, if so, the value
313 // will be available to initialize the accumulator.
314 if (Argument *Arg = dyn_cast<Argument>(V)) {
315 // Figure out which argument number this is...
317 Function *F = CI->getParent()->getParent();
318 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
321 // If we are passing this argument into call as the corresponding
322 // argument operand, then the argument is dynamically constant.
323 // Otherwise, we cannot transform this function safely.
324 if (CI->getArgOperand(ArgNo) == Arg)
328 // Switch cases are always constant integers. If the value is being switched
329 // on and the return is only reachable from one of its cases, it's
330 // effectively constant.
331 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
332 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
333 if (SI->getCondition() == V)
334 return SI->getDefaultDest() != RI->getParent();
336 // Not a constant or immutable argument, we can't safely transform.
340 // getCommonReturnValue - Check to see if the function containing the specified
341 // tail call consistently returns the same runtime-constant value at all exit
342 // points except for IgnoreRI. If so, return the returned value.
344 static Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) {
345 Function *F = CI->getParent()->getParent();
346 Value *ReturnedValue = 0;
348 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) {
349 ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator());
350 if (RI == 0 || RI == IgnoreRI) continue;
352 // We can only perform this transformation if the value returned is
353 // evaluatable at the start of the initial invocation of the function,
354 // instead of at the end of the evaluation.
356 Value *RetOp = RI->getOperand(0);
357 if (!isDynamicConstant(RetOp, CI, RI))
360 if (ReturnedValue && RetOp != ReturnedValue)
361 return 0; // Cannot transform if differing values are returned.
362 ReturnedValue = RetOp;
364 return ReturnedValue;
367 /// CanTransformAccumulatorRecursion - If the specified instruction can be
368 /// transformed using accumulator recursion elimination, return the constant
369 /// which is the start of the accumulator value. Otherwise return null.
371 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
373 if (!I->isAssociative() || !I->isCommutative()) return 0;
374 assert(I->getNumOperands() == 2 &&
375 "Associative/commutative operations should have 2 args!");
377 // Exactly one operand should be the result of the call instruction.
378 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
379 (I->getOperand(0) != CI && I->getOperand(1) != CI))
382 // The only user of this instruction we allow is a single return instruction.
383 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back()))
386 // Ok, now we have to check all of the other return instructions in this
387 // function. If they return non-constants or differing values, then we cannot
388 // transform the function safely.
389 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI);
392 static Instruction *FirstNonDbg(BasicBlock::iterator I) {
393 while (isa<DbgInfoIntrinsic>(I))
399 TailCallElim::FindTRECandidate(Instruction *TI,
400 bool CannotTailCallElimCallsMarkedTail) {
401 BasicBlock *BB = TI->getParent();
402 Function *F = BB->getParent();
404 if (&BB->front() == TI) // Make sure there is something before the terminator.
407 // Scan backwards from the return, checking to see if there is a tail call in
408 // this block. If so, set CI to it.
410 BasicBlock::iterator BBI = TI;
412 CI = dyn_cast<CallInst>(BBI);
413 if (CI && CI->getCalledFunction() == F)
416 if (BBI == BB->begin())
417 return 0; // Didn't find a potential tail call.
421 // If this call is marked as a tail call, and if there are dynamic allocas in
422 // the function, we cannot perform this optimization.
423 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
426 // As a special case, detect code like this:
427 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
428 // and disable this xform in this case, because the code generator will
429 // lower the call to fabs into inline code.
430 if (BB == &F->getEntryBlock() &&
431 FirstNonDbg(BB->front()) == CI &&
432 FirstNonDbg(std::next(BB->begin())) == TI &&
433 CI->getCalledFunction() &&
434 !TTI->isLoweredToCall(CI->getCalledFunction())) {
435 // A single-block function with just a call and a return. Check that
436 // the arguments match.
437 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
438 E = CallSite(CI).arg_end();
439 Function::arg_iterator FI = F->arg_begin(),
441 for (; I != E && FI != FE; ++I, ++FI)
442 if (*I != &*FI) break;
443 if (I == E && FI == FE)
450 bool TailCallElim::EliminateRecursiveTailCall(CallInst *CI, ReturnInst *Ret,
451 BasicBlock *&OldEntry,
452 bool &TailCallsAreMarkedTail,
453 SmallVectorImpl<PHINode *> &ArgumentPHIs,
454 bool CannotTailCallElimCallsMarkedTail) {
455 // If we are introducing accumulator recursion to eliminate operations after
456 // the call instruction that are both associative and commutative, the initial
457 // value for the accumulator is placed in this variable. If this value is set
458 // then we actually perform accumulator recursion elimination instead of
459 // simple tail recursion elimination. If the operation is an LLVM instruction
460 // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then
461 // we are handling the case when the return instruction returns a constant C
462 // which is different to the constant returned by other return instructions
463 // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a
464 // special case of accumulator recursion, the operation being "return C".
465 Value *AccumulatorRecursionEliminationInitVal = 0;
466 Instruction *AccumulatorRecursionInstr = 0;
468 // Ok, we found a potential tail call. We can currently only transform the
469 // tail call if all of the instructions between the call and the return are
470 // movable to above the call itself, leaving the call next to the return.
471 // Check that this is the case now.
472 BasicBlock::iterator BBI = CI;
473 for (++BBI; &*BBI != Ret; ++BBI) {
474 if (CanMoveAboveCall(BBI, CI)) continue;
476 // If we can't move the instruction above the call, it might be because it
477 // is an associative and commutative operation that could be transformed
478 // using accumulator recursion elimination. Check to see if this is the
479 // case, and if so, remember the initial accumulator value for later.
480 if ((AccumulatorRecursionEliminationInitVal =
481 CanTransformAccumulatorRecursion(BBI, CI))) {
482 // Yes, this is accumulator recursion. Remember which instruction
484 AccumulatorRecursionInstr = BBI;
486 return false; // Otherwise, we cannot eliminate the tail recursion!
490 // We can only transform call/return pairs that either ignore the return value
491 // of the call and return void, ignore the value of the call and return a
492 // constant, return the value returned by the tail call, or that are being
493 // accumulator recursion variable eliminated.
494 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
495 !isa<UndefValue>(Ret->getReturnValue()) &&
496 AccumulatorRecursionEliminationInitVal == 0 &&
497 !getCommonReturnValue(0, CI)) {
498 // One case remains that we are able to handle: the current return
499 // instruction returns a constant, and all other return instructions
500 // return a different constant.
501 if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret))
502 return false; // Current return instruction does not return a constant.
503 // Check that all other return instructions return a common constant. If
504 // so, record it in AccumulatorRecursionEliminationInitVal.
505 AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI);
506 if (!AccumulatorRecursionEliminationInitVal)
510 BasicBlock *BB = Ret->getParent();
511 Function *F = BB->getParent();
513 // OK! We can transform this tail call. If this is the first one found,
514 // create the new entry block, allowing us to branch back to the old entry.
516 OldEntry = &F->getEntryBlock();
517 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
518 NewEntry->takeName(OldEntry);
519 OldEntry->setName("tailrecurse");
520 BranchInst::Create(OldEntry, NewEntry);
522 // If this tail call is marked 'tail' and if there are any allocas in the
523 // entry block, move them up to the new entry block.
524 TailCallsAreMarkedTail = CI->isTailCall();
525 if (TailCallsAreMarkedTail)
526 // Move all fixed sized allocas from OldEntry to NewEntry.
527 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
528 NEBI = NewEntry->begin(); OEBI != E; )
529 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
530 if (isa<ConstantInt>(AI->getArraySize()))
531 AI->moveBefore(NEBI);
533 // Now that we have created a new block, which jumps to the entry
534 // block, insert a PHI node for each argument of the function.
535 // For now, we initialize each PHI to only have the real arguments
536 // which are passed in.
537 Instruction *InsertPos = OldEntry->begin();
538 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
540 PHINode *PN = PHINode::Create(I->getType(), 2,
541 I->getName() + ".tr", InsertPos);
542 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
543 PN->addIncoming(I, NewEntry);
544 ArgumentPHIs.push_back(PN);
548 // If this function has self recursive calls in the tail position where some
549 // are marked tail and some are not, only transform one flavor or another. We
550 // have to choose whether we move allocas in the entry block to the new entry
551 // block or not, so we can't make a good choice for both. NOTE: We could do
552 // slightly better here in the case that the function has no entry block
554 if (TailCallsAreMarkedTail && !CI->isTailCall())
557 // Ok, now that we know we have a pseudo-entry block WITH all of the
558 // required PHI nodes, add entries into the PHI node for the actual
559 // parameters passed into the tail-recursive call.
560 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i)
561 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
563 // If we are introducing an accumulator variable to eliminate the recursion,
564 // do so now. Note that we _know_ that no subsequent tail recursion
565 // eliminations will happen on this function because of the way the
566 // accumulator recursion predicate is set up.
568 if (AccumulatorRecursionEliminationInitVal) {
569 Instruction *AccRecInstr = AccumulatorRecursionInstr;
570 // Start by inserting a new PHI node for the accumulator.
571 pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
573 PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(),
574 std::distance(PB, PE) + 1,
575 "accumulator.tr", OldEntry->begin());
577 // Loop over all of the predecessors of the tail recursion block. For the
578 // real entry into the function we seed the PHI with the initial value,
579 // computed earlier. For any other existing branches to this block (due to
580 // other tail recursions eliminated) the accumulator is not modified.
581 // Because we haven't added the branch in the current block to OldEntry yet,
582 // it will not show up as a predecessor.
583 for (pred_iterator PI = PB; PI != PE; ++PI) {
585 if (P == &F->getEntryBlock())
586 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P);
588 AccPN->addIncoming(AccPN, P);
592 // Add an incoming argument for the current block, which is computed by
593 // our associative and commutative accumulator instruction.
594 AccPN->addIncoming(AccRecInstr, BB);
596 // Next, rewrite the accumulator recursion instruction so that it does not
597 // use the result of the call anymore, instead, use the PHI node we just
599 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
601 // Add an incoming argument for the current block, which is just the
602 // constant returned by the current return instruction.
603 AccPN->addIncoming(Ret->getReturnValue(), BB);
606 // Finally, rewrite any return instructions in the program to return the PHI
607 // node instead of the "initval" that they do currently. This loop will
608 // actually rewrite the return value we are destroying, but that's ok.
609 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
610 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
611 RI->setOperand(0, AccPN);
615 // Now that all of the PHI nodes are in place, remove the call and
616 // ret instructions, replacing them with an unconditional branch.
617 BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
618 NewBI->setDebugLoc(CI->getDebugLoc());
620 BB->getInstList().erase(Ret); // Remove return.
621 BB->getInstList().erase(CI); // Remove call.
626 bool TailCallElim::FoldReturnAndProcessPred(BasicBlock *BB,
627 ReturnInst *Ret, BasicBlock *&OldEntry,
628 bool &TailCallsAreMarkedTail,
629 SmallVectorImpl<PHINode *> &ArgumentPHIs,
630 bool CannotTailCallElimCallsMarkedTail) {
633 // If the return block contains nothing but the return and PHI's,
634 // there might be an opportunity to duplicate the return in its
635 // predecessors and perform TRC there. Look for predecessors that end
636 // in unconditional branch and recursive call(s).
637 SmallVector<BranchInst*, 8> UncondBranchPreds;
638 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
639 BasicBlock *Pred = *PI;
640 TerminatorInst *PTI = Pred->getTerminator();
641 if (BranchInst *BI = dyn_cast<BranchInst>(PTI))
642 if (BI->isUnconditional())
643 UncondBranchPreds.push_back(BI);
646 while (!UncondBranchPreds.empty()) {
647 BranchInst *BI = UncondBranchPreds.pop_back_val();
648 BasicBlock *Pred = BI->getParent();
649 if (CallInst *CI = FindTRECandidate(BI, CannotTailCallElimCallsMarkedTail)){
650 DEBUG(dbgs() << "FOLDING: " << *BB
651 << "INTO UNCOND BRANCH PRED: " << *Pred);
652 EliminateRecursiveTailCall(CI, FoldReturnIntoUncondBranch(Ret, BB, Pred),
653 OldEntry, TailCallsAreMarkedTail, ArgumentPHIs,
654 CannotTailCallElimCallsMarkedTail);
664 TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
665 bool &TailCallsAreMarkedTail,
666 SmallVectorImpl<PHINode *> &ArgumentPHIs,
667 bool CannotTailCallElimCallsMarkedTail) {
668 CallInst *CI = FindTRECandidate(Ret, CannotTailCallElimCallsMarkedTail);
672 return EliminateRecursiveTailCall(CI, Ret, OldEntry, TailCallsAreMarkedTail,
674 CannotTailCallElimCallsMarkedTail);