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 expression to use an accumulator variable,
20 // thus compiling the typical naive factorial or 'fib' implementation into
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 preceeds 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/Transforms/Utils/Local.h"
56 #include "llvm/Constants.h"
57 #include "llvm/DerivedTypes.h"
58 #include "llvm/Function.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/Pass.h"
61 #include "llvm/Analysis/CaptureTracking.h"
62 #include "llvm/Analysis/InlineCost.h"
63 #include "llvm/Analysis/Loads.h"
64 #include "llvm/Support/CallSite.h"
65 #include "llvm/Support/CFG.h"
66 #include "llvm/ADT/Statistic.h"
69 STATISTIC(NumEliminated, "Number of tail calls removed");
70 STATISTIC(NumAccumAdded, "Number of accumulators introduced");
73 struct TailCallElim : public FunctionPass {
74 static char ID; // Pass identification, replacement for typeid
75 TailCallElim() : FunctionPass(&ID) {}
77 virtual bool runOnFunction(Function &F);
80 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
81 bool &TailCallsAreMarkedTail,
82 SmallVector<PHINode*, 8> &ArgumentPHIs,
83 bool CannotTailCallElimCallsMarkedTail);
84 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
85 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
89 char TailCallElim::ID = 0;
90 static RegisterPass<TailCallElim> X("tailcallelim", "Tail Call Elimination");
92 // Public interface to the TailCallElimination pass
93 FunctionPass *llvm::createTailCallEliminationPass() {
94 return new TailCallElim();
97 /// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by
98 /// callees of this function. We only do very simple analysis right now, this
99 /// could be expanded in the future to use mod/ref information for particular
100 /// call sites if desired.
101 static bool AllocaMightEscapeToCalls(AllocaInst *AI) {
102 // FIXME: do simple 'address taken' analysis.
106 /// CheckForEscapingAllocas - Scan the specified basic block for alloca
107 /// instructions. If it contains any that might be accessed by calls, return
109 static bool CheckForEscapingAllocas(BasicBlock *BB,
110 bool &CannotTCETailMarkedCall) {
112 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
113 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
114 RetVal |= AllocaMightEscapeToCalls(AI);
116 // If this alloca is in the body of the function, or if it is a variable
117 // sized allocation, we cannot tail call eliminate calls marked 'tail'
118 // with this mechanism.
119 if (BB != &BB->getParent()->getEntryBlock() ||
120 !isa<ConstantInt>(AI->getArraySize()))
121 CannotTCETailMarkedCall = true;
126 bool TailCallElim::runOnFunction(Function &F) {
127 // If this function is a varargs function, we won't be able to PHI the args
128 // right, so don't even try to convert it...
129 if (F.getFunctionType()->isVarArg()) return false;
131 BasicBlock *OldEntry = 0;
132 bool TailCallsAreMarkedTail = false;
133 SmallVector<PHINode*, 8> ArgumentPHIs;
134 bool MadeChange = false;
136 bool FunctionContainsEscapingAllocas = false;
138 // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls
139 // marked with the 'tail' attribute, because doing so would cause the stack
140 // size to increase (real TCE would deallocate variable sized allocas, TCE
142 bool CannotTCETailMarkedCall = false;
144 // Loop over the function, looking for any returning blocks, and keeping track
145 // of whether this function has any non-trivially used allocas.
146 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
147 if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall)
150 FunctionContainsEscapingAllocas |=
151 CheckForEscapingAllocas(BB, CannotTCETailMarkedCall);
154 /// FIXME: The code generator produces really bad code when an 'escaping
155 /// alloca' is changed from being a static alloca to being a dynamic alloca.
156 /// Until this is resolved, disable this transformation if that would ever
157 /// happen. This bug is PR962.
158 if (FunctionContainsEscapingAllocas)
161 // Second pass, change any tail calls to loops.
162 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
163 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator()))
164 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
165 ArgumentPHIs,CannotTCETailMarkedCall);
167 // If we eliminated any tail recursions, it's possible that we inserted some
168 // silly PHI nodes which just merge an initial value (the incoming operand)
169 // with themselves. Check to see if we did and clean up our mess if so. This
170 // occurs when a function passes an argument straight through to its tail
172 if (!ArgumentPHIs.empty()) {
173 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
174 PHINode *PN = ArgumentPHIs[i];
176 // If the PHI Node is a dynamic constant, replace it with the value it is.
177 if (Value *PNV = PN->hasConstantValue()) {
178 PN->replaceAllUsesWith(PNV);
179 PN->eraseFromParent();
184 // Finally, if this function contains no non-escaping allocas, mark all calls
185 // in the function as eligible for tail calls (there is no stack memory for
187 if (!FunctionContainsEscapingAllocas)
188 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
189 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
190 if (CallInst *CI = dyn_cast<CallInst>(I)) {
199 /// CanMoveAboveCall - Return true if it is safe to move the specified
200 /// instruction from after the call to before the call, assuming that all
201 /// instructions between the call and this instruction are movable.
203 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
204 // FIXME: We can move load/store/call/free instructions above the call if the
205 // call does not mod/ref the memory location being processed.
206 if (I->mayHaveSideEffects()) // This also handles volatile loads.
209 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
210 // Loads may always be moved above calls without side effects.
211 if (CI->mayHaveSideEffects()) {
212 // Non-volatile loads may be moved above a call with side effects if it
213 // does not write to memory and the load provably won't trap.
214 // FIXME: Writes to memory only matter if they may alias the pointer
215 // being loaded from.
216 if (CI->mayWriteToMemory() ||
217 !isSafeToLoadUnconditionally(L->getPointerOperand(), L,
223 // Otherwise, if this is a side-effect free instruction, check to make sure
224 // that it does not use the return value of the call. If it doesn't use the
225 // return value of the call, it must only use things that are defined before
226 // the call, or movable instructions between the call and the instruction
228 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
229 if (I->getOperand(i) == CI)
234 // isDynamicConstant - Return true if the specified value is the same when the
235 // return would exit as it was when the initial iteration of the recursive
236 // function was executed.
238 // We currently handle static constants and arguments that are not modified as
239 // part of the recursion.
241 static bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) {
242 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
244 // Check to see if this is an immutable argument, if so, the value
245 // will be available to initialize the accumulator.
246 if (Argument *Arg = dyn_cast<Argument>(V)) {
247 // Figure out which argument number this is...
249 Function *F = CI->getParent()->getParent();
250 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
253 // If we are passing this argument into call as the corresponding
254 // argument operand, then the argument is dynamically constant.
255 // Otherwise, we cannot transform this function safely.
256 if (CI->getArgOperand(ArgNo) == Arg)
260 // Switch cases are always constant integers. If the value is being switched
261 // on and the return is only reachable from one of its cases, it's
262 // effectively constant.
263 if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor())
264 if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator()))
265 if (SI->getCondition() == V)
266 return SI->getDefaultDest() != RI->getParent();
268 // Not a constant or immutable argument, we can't safely transform.
272 // getCommonReturnValue - Check to see if the function containing the specified
273 // return instruction and tail call consistently returns the same
274 // runtime-constant value at all exit points. If so, return the returned value.
276 static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) {
277 Function *F = TheRI->getParent()->getParent();
278 Value *ReturnedValue = 0;
280 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
281 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
283 Value *RetOp = RI->getOperand(0);
285 // We can only perform this transformation if the value returned is
286 // evaluatable at the start of the initial invocation of the function,
287 // instead of at the end of the evaluation.
289 if (!isDynamicConstant(RetOp, CI, RI))
292 if (ReturnedValue && RetOp != ReturnedValue)
293 return 0; // Cannot transform if differing values are returned.
294 ReturnedValue = RetOp;
296 return ReturnedValue;
299 /// CanTransformAccumulatorRecursion - If the specified instruction can be
300 /// transformed using accumulator recursion elimination, return the constant
301 /// which is the start of the accumulator value. Otherwise return null.
303 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
305 if (!I->isAssociative()) return 0;
306 assert(I->getNumOperands() == 2 &&
307 "Associative operations should have 2 args!");
309 // Exactly one operand should be the result of the call instruction...
310 if ((I->getOperand(0) == CI && I->getOperand(1) == CI) ||
311 (I->getOperand(0) != CI && I->getOperand(1) != CI))
314 // The only user of this instruction we allow is a single return instruction.
315 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back()))
318 // Ok, now we have to check all of the other return instructions in this
319 // function. If they return non-constants or differing values, then we cannot
320 // transform the function safely.
321 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI);
324 bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
325 bool &TailCallsAreMarkedTail,
326 SmallVector<PHINode*, 8> &ArgumentPHIs,
327 bool CannotTailCallElimCallsMarkedTail) {
328 BasicBlock *BB = Ret->getParent();
329 Function *F = BB->getParent();
331 if (&BB->front() == Ret) // Make sure there is something before the ret...
334 // Scan backwards from the return, checking to see if there is a tail call in
335 // this block. If so, set CI to it.
337 BasicBlock::iterator BBI = Ret;
339 CI = dyn_cast<CallInst>(BBI);
340 if (CI && CI->getCalledFunction() == F)
343 if (BBI == BB->begin())
344 return false; // Didn't find a potential tail call.
348 // If this call is marked as a tail call, and if there are dynamic allocas in
349 // the function, we cannot perform this optimization.
350 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
353 // As a special case, detect code like this:
354 // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
355 // and disable this xform in this case, because the code generator will
356 // lower the call to fabs into inline code.
357 if (BB == &F->getEntryBlock() &&
358 &BB->front() == CI && &*++BB->begin() == Ret &&
360 // A single-block function with just a call and a return. Check that
361 // the arguments match.
362 CallSite::arg_iterator I = CallSite(CI).arg_begin(),
363 E = CallSite(CI).arg_end();
364 Function::arg_iterator FI = F->arg_begin(),
366 for (; I != E && FI != FE; ++I, ++FI)
367 if (*I != &*FI) break;
368 if (I == E && FI == FE)
372 // If we are introducing accumulator recursion to eliminate associative
373 // operations after the call instruction, this variable contains the initial
374 // value for the accumulator. If this value is set, we actually perform
375 // accumulator recursion elimination instead of simple tail recursion
377 Value *AccumulatorRecursionEliminationInitVal = 0;
378 Instruction *AccumulatorRecursionInstr = 0;
380 // Ok, we found a potential tail call. We can currently only transform the
381 // tail call if all of the instructions between the call and the return are
382 // movable to above the call itself, leaving the call next to the return.
383 // Check that this is the case now.
384 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
385 if (!CanMoveAboveCall(BBI, CI)) {
386 // If we can't move the instruction above the call, it might be because it
387 // is an associative operation that could be tranformed using accumulator
388 // recursion elimination. Check to see if this is the case, and if so,
389 // remember the initial accumulator value for later.
390 if ((AccumulatorRecursionEliminationInitVal =
391 CanTransformAccumulatorRecursion(BBI, CI))) {
392 // Yes, this is accumulator recursion. Remember which instruction
394 AccumulatorRecursionInstr = BBI;
396 return false; // Otherwise, we cannot eliminate the tail recursion!
400 // We can only transform call/return pairs that either ignore the return value
401 // of the call and return void, ignore the value of the call and return a
402 // constant, return the value returned by the tail call, or that are being
403 // accumulator recursion variable eliminated.
404 if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI &&
405 !isa<UndefValue>(Ret->getReturnValue()) &&
406 AccumulatorRecursionEliminationInitVal == 0 &&
407 !getCommonReturnValue(Ret, CI))
410 // OK! We can transform this tail call. If this is the first one found,
411 // create the new entry block, allowing us to branch back to the old entry.
413 OldEntry = &F->getEntryBlock();
414 BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry);
415 NewEntry->takeName(OldEntry);
416 OldEntry->setName("tailrecurse");
417 BranchInst::Create(OldEntry, NewEntry);
419 // If this tail call is marked 'tail' and if there are any allocas in the
420 // entry block, move them up to the new entry block.
421 TailCallsAreMarkedTail = CI->isTailCall();
422 if (TailCallsAreMarkedTail)
423 // Move all fixed sized allocas from OldEntry to NewEntry.
424 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
425 NEBI = NewEntry->begin(); OEBI != E; )
426 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
427 if (isa<ConstantInt>(AI->getArraySize()))
428 AI->moveBefore(NEBI);
430 // Now that we have created a new block, which jumps to the entry
431 // block, insert a PHI node for each argument of the function.
432 // For now, we initialize each PHI to only have the real arguments
433 // which are passed in.
434 Instruction *InsertPos = OldEntry->begin();
435 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
437 PHINode *PN = PHINode::Create(I->getType(),
438 I->getName() + ".tr", InsertPos);
439 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
440 PN->addIncoming(I, NewEntry);
441 ArgumentPHIs.push_back(PN);
445 // If this function has self recursive calls in the tail position where some
446 // are marked tail and some are not, only transform one flavor or another. We
447 // have to choose whether we move allocas in the entry block to the new entry
448 // block or not, so we can't make a good choice for both. NOTE: We could do
449 // slightly better here in the case that the function has no entry block
451 if (TailCallsAreMarkedTail && !CI->isTailCall())
454 // Ok, now that we know we have a pseudo-entry block WITH all of the
455 // required PHI nodes, add entries into the PHI node for the actual
456 // parameters passed into the tail-recursive call.
457 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
458 ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB);
460 // If we are introducing an accumulator variable to eliminate the recursion,
461 // do so now. Note that we _know_ that no subsequent tail recursion
462 // eliminations will happen on this function because of the way the
463 // accumulator recursion predicate is set up.
465 if (AccumulatorRecursionEliminationInitVal) {
466 Instruction *AccRecInstr = AccumulatorRecursionInstr;
467 // Start by inserting a new PHI node for the accumulator.
468 PHINode *AccPN = PHINode::Create(AccRecInstr->getType(), "accumulator.tr",
471 // Loop over all of the predecessors of the tail recursion block. For the
472 // real entry into the function we seed the PHI with the initial value,
473 // computed earlier. For any other existing branches to this block (due to
474 // other tail recursions eliminated) the accumulator is not modified.
475 // Because we haven't added the branch in the current block to OldEntry yet,
476 // it will not show up as a predecessor.
477 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
479 if (*PI == &F->getEntryBlock())
480 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
482 AccPN->addIncoming(AccPN, *PI);
485 // Add an incoming argument for the current block, which is computed by our
486 // associative accumulator instruction.
487 AccPN->addIncoming(AccRecInstr, BB);
489 // Next, rewrite the accumulator recursion instruction so that it does not
490 // use the result of the call anymore, instead, use the PHI node we just
492 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
494 // Finally, rewrite any return instructions in the program to return the PHI
495 // node instead of the "initval" that they do currently. This loop will
496 // actually rewrite the return value we are destroying, but that's ok.
497 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
498 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
499 RI->setOperand(0, AccPN);
503 // Now that all of the PHI nodes are in place, remove the call and
504 // ret instructions, replacing them with an unconditional branch.
505 BranchInst::Create(OldEntry, Ret);
506 BB->getInstList().erase(Ret); // Remove return.
507 BB->getInstList().erase(CI); // Remove call.