1 //===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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 theier 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 #include "llvm/Transforms/Scalar.h"
54 #include "llvm/Constants.h"
55 #include "llvm/DerivedTypes.h"
56 #include "llvm/Function.h"
57 #include "llvm/Instructions.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/CFG.h"
60 #include "llvm/ADT/Statistic.h"
64 Statistic NumEliminated("tailcallelim", "Number of tail calls removed");
65 Statistic NumAccumAdded("tailcallelim","Number of accumulators introduced");
67 struct TailCallElim : public FunctionPass {
68 virtual bool runOnFunction(Function &F);
71 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
72 bool &TailCallsAreMarkedTail,
73 std::vector<PHINode*> &ArgumentPHIs,
74 bool CannotTailCallElimCallsMarkedTail);
75 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
76 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
78 RegisterPass<TailCallElim> X("tailcallelim", "Tail Call Elimination");
81 // Public interface to the TailCallElimination pass
82 FunctionPass *llvm::createTailCallEliminationPass() {
83 return new TailCallElim();
87 /// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by
88 /// callees of this function. We only do very simple analysis right now, this
89 /// could be expanded in the future to use mod/ref information for particular
90 /// call sites if desired.
91 static bool AllocaMightEscapeToCalls(AllocaInst *AI) {
92 // FIXME: do simple 'address taken' analysis.
96 /// FunctionContainsAllocas - Scan the specified basic block for alloca
97 /// instructions. If it contains any that might be accessed by calls, return
99 static bool CheckForEscapingAllocas(BasicBlock *BB,
100 bool &CannotTCETailMarkedCall) {
102 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
103 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
104 RetVal |= AllocaMightEscapeToCalls(AI);
106 // If this alloca is in the body of the function, or if it is a variable
107 // sized allocation, we cannot tail call eliminate calls marked 'tail'
108 // with this mechanism.
109 if (BB != &BB->getParent()->front() ||
110 !isa<ConstantInt>(AI->getArraySize()))
111 CannotTCETailMarkedCall = true;
116 bool TailCallElim::runOnFunction(Function &F) {
117 // If this function is a varargs function, we won't be able to PHI the args
118 // right, so don't even try to convert it...
119 if (F.getFunctionType()->isVarArg()) return false;
121 BasicBlock *OldEntry = 0;
122 bool TailCallsAreMarkedTail = false;
123 std::vector<PHINode*> ArgumentPHIs;
124 bool MadeChange = false;
126 bool FunctionContainsEscapingAllocas = false;
128 // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls
129 // marked with the 'tail' attribute, because doing so would cause the stack
130 // size to increase (real TCE would deallocate variable sized allocas, TCE
132 bool CannotTCETailMarkedCall = false;
134 // Loop over the function, looking for any returning blocks, and keeping track
135 // of whether this function has any non-trivially used allocas.
136 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
137 if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall)
140 FunctionContainsEscapingAllocas |=
141 CheckForEscapingAllocas(BB, CannotTCETailMarkedCall);
144 /// FIXME: The code generator produces really bad code when an 'escaping
145 /// alloca' is changed from being a static alloca to being a dynamic alloca.
146 /// Until this is resolved, disable this transformation if that would ever
147 /// happen. This bug is PR962.
148 if (FunctionContainsEscapingAllocas)
152 // Second pass, change any tail calls to loops.
153 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
154 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator()))
155 MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
156 ArgumentPHIs,CannotTCETailMarkedCall);
158 // If we eliminated any tail recursions, it's possible that we inserted some
159 // silly PHI nodes which just merge an initial value (the incoming operand)
160 // with themselves. Check to see if we did and clean up our mess if so. This
161 // occurs when a function passes an argument straight through to its tail
163 if (!ArgumentPHIs.empty()) {
164 for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) {
165 PHINode *PN = ArgumentPHIs[i];
167 // If the PHI Node is a dynamic constant, replace it with the value it is.
168 if (Value *PNV = PN->hasConstantValue()) {
169 PN->replaceAllUsesWith(PNV);
170 PN->eraseFromParent();
175 // Finally, if this function contains no non-escaping allocas, mark all calls
176 // in the function as eligible for tail calls (there is no stack memory for
178 if (!FunctionContainsEscapingAllocas)
179 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
180 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
181 if (CallInst *CI = dyn_cast<CallInst>(I))
188 /// CanMoveAboveCall - Return true if it is safe to move the specified
189 /// instruction from after the call to before the call, assuming that all
190 /// instructions between the call and this instruction are movable.
192 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
193 // FIXME: We can move load/store/call/free instructions above the call if the
194 // call does not mod/ref the memory location being processed.
195 if (I->mayWriteToMemory() || isa<LoadInst>(I))
198 // Otherwise, if this is a side-effect free instruction, check to make sure
199 // that it does not use the return value of the call. If it doesn't use the
200 // return value of the call, it must only use things that are defined before
201 // the call, or movable instructions between the call and the instruction
203 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
204 if (I->getOperand(i) == CI)
209 // isDynamicConstant - Return true if the specified value is the same when the
210 // return would exit as it was when the initial iteration of the recursive
211 // function was executed.
213 // We currently handle static constants and arguments that are not modified as
214 // part of the recursion.
216 static bool isDynamicConstant(Value *V, CallInst *CI) {
217 if (isa<Constant>(V)) return true; // Static constants are always dyn consts
219 // Check to see if this is an immutable argument, if so, the value
220 // will be available to initialize the accumulator.
221 if (Argument *Arg = dyn_cast<Argument>(V)) {
222 // Figure out which argument number this is...
224 Function *F = CI->getParent()->getParent();
225 for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI)
228 // If we are passing this argument into call as the corresponding
229 // argument operand, then the argument is dynamically constant.
230 // Otherwise, we cannot transform this function safely.
231 if (CI->getOperand(ArgNo+1) == Arg)
234 // Not a constant or immutable argument, we can't safely transform.
238 // getCommonReturnValue - Check to see if the function containing the specified
239 // return instruction and tail call consistently returns the same
240 // runtime-constant value at all exit points. If so, return the returned value.
242 static Value *getCommonReturnValue(ReturnInst *TheRI, CallInst *CI) {
243 Function *F = TheRI->getParent()->getParent();
244 Value *ReturnedValue = 0;
246 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
247 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
249 Value *RetOp = RI->getOperand(0);
251 // We can only perform this transformation if the value returned is
252 // evaluatable at the start of the initial invocation of the function,
253 // instead of at the end of the evaluation.
255 if (!isDynamicConstant(RetOp, CI))
258 if (ReturnedValue && RetOp != ReturnedValue)
259 return 0; // Cannot transform if differing values are returned.
260 ReturnedValue = RetOp;
262 return ReturnedValue;
265 /// CanTransformAccumulatorRecursion - If the specified instruction can be
266 /// transformed using accumulator recursion elimination, return the constant
267 /// which is the start of the accumulator value. Otherwise return null.
269 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
271 if (!I->isAssociative()) return 0;
272 assert(I->getNumOperands() == 2 &&
273 "Associative operations should have 2 args!");
275 // Exactly one operand should be the result of the call instruction...
276 if (I->getOperand(0) == CI && I->getOperand(1) == CI ||
277 I->getOperand(0) != CI && I->getOperand(1) != CI)
280 // The only user of this instruction we allow is a single return instruction.
281 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back()))
284 // Ok, now we have to check all of the other return instructions in this
285 // function. If they return non-constants or differing values, then we cannot
286 // transform the function safely.
287 return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI);
290 bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
291 bool &TailCallsAreMarkedTail,
292 std::vector<PHINode*> &ArgumentPHIs,
293 bool CannotTailCallElimCallsMarkedTail) {
294 BasicBlock *BB = Ret->getParent();
295 Function *F = BB->getParent();
297 if (&BB->front() == Ret) // Make sure there is something before the ret...
300 // Scan backwards from the return, checking to see if there is a tail call in
301 // this block. If so, set CI to it.
303 BasicBlock::iterator BBI = Ret;
305 CI = dyn_cast<CallInst>(BBI);
306 if (CI && CI->getCalledFunction() == F)
309 if (BBI == BB->begin())
310 return false; // Didn't find a potential tail call.
314 // If this call is marked as a tail call, and if there are dynamic allocas in
315 // the function, we cannot perform this optimization.
316 if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail)
319 // If we are introducing accumulator recursion to eliminate associative
320 // operations after the call instruction, this variable contains the initial
321 // value for the accumulator. If this value is set, we actually perform
322 // accumulator recursion elimination instead of simple tail recursion
324 Value *AccumulatorRecursionEliminationInitVal = 0;
325 Instruction *AccumulatorRecursionInstr = 0;
327 // Ok, we found a potential tail call. We can currently only transform the
328 // tail call if all of the instructions between the call and the return are
329 // movable to above the call itself, leaving the call next to the return.
330 // Check that this is the case now.
331 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
332 if (!CanMoveAboveCall(BBI, CI)) {
333 // If we can't move the instruction above the call, it might be because it
334 // is an associative operation that could be tranformed using accumulator
335 // recursion elimination. Check to see if this is the case, and if so,
336 // remember the initial accumulator value for later.
337 if ((AccumulatorRecursionEliminationInitVal =
338 CanTransformAccumulatorRecursion(BBI, CI))) {
339 // Yes, this is accumulator recursion. Remember which instruction
341 AccumulatorRecursionInstr = BBI;
343 return false; // Otherwise, we cannot eliminate the tail recursion!
347 // We can only transform call/return pairs that either ignore the return value
348 // of the call and return void, ignore the value of the call and return a
349 // constant, return the value returned by the tail call, or that are being
350 // accumulator recursion variable eliminated.
351 if (Ret->getNumOperands() != 0 && Ret->getReturnValue() != CI &&
352 !isa<UndefValue>(Ret->getReturnValue()) &&
353 AccumulatorRecursionEliminationInitVal == 0 &&
354 !getCommonReturnValue(Ret, CI))
357 // OK! We can transform this tail call. If this is the first one found,
358 // create the new entry block, allowing us to branch back to the old entry.
360 OldEntry = &F->getEntryBlock();
361 std::string OldName = OldEntry->getName(); OldEntry->setName("tailrecurse");
362 BasicBlock *NewEntry = new BasicBlock(OldName, F, OldEntry);
363 new BranchInst(OldEntry, NewEntry);
365 // If this tail call is marked 'tail' and if there are any allocas in the
366 // entry block, move them up to the new entry block.
367 TailCallsAreMarkedTail = CI->isTailCall();
368 if (TailCallsAreMarkedTail)
369 // Move all fixed sized allocas from OldEntry to NewEntry.
370 for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(),
371 NEBI = NewEntry->begin(); OEBI != E; )
372 if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++))
373 if (isa<ConstantInt>(AI->getArraySize()))
374 AI->moveBefore(NEBI);
376 // Now that we have created a new block, which jumps to the entry
377 // block, insert a PHI node for each argument of the function.
378 // For now, we initialize each PHI to only have the real arguments
379 // which are passed in.
380 Instruction *InsertPos = OldEntry->begin();
381 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
383 PHINode *PN = new PHINode(I->getType(), I->getName()+".tr", InsertPos);
384 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
385 PN->addIncoming(I, NewEntry);
386 ArgumentPHIs.push_back(PN);
390 // If this function has self recursive calls in the tail position where some
391 // are marked tail and some are not, only transform one flavor or another. We
392 // have to choose whether we move allocas in the entry block to the new entry
393 // block or not, so we can't make a good choice for both. NOTE: We could do
394 // slightly better here in the case that the function has no entry block
396 if (TailCallsAreMarkedTail && !CI->isTailCall())
399 // Ok, now that we know we have a pseudo-entry block WITH all of the
400 // required PHI nodes, add entries into the PHI node for the actual
401 // parameters passed into the tail-recursive call.
402 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
403 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB);
405 // If we are introducing an accumulator variable to eliminate the recursion,
406 // do so now. Note that we _know_ that no subsequent tail recursion
407 // eliminations will happen on this function because of the way the
408 // accumulator recursion predicate is set up.
410 if (AccumulatorRecursionEliminationInitVal) {
411 Instruction *AccRecInstr = AccumulatorRecursionInstr;
412 // Start by inserting a new PHI node for the accumulator.
413 PHINode *AccPN = new PHINode(AccRecInstr->getType(), "accumulator.tr",
416 // Loop over all of the predecessors of the tail recursion block. For the
417 // real entry into the function we seed the PHI with the initial value,
418 // computed earlier. For any other existing branches to this block (due to
419 // other tail recursions eliminated) the accumulator is not modified.
420 // Because we haven't added the branch in the current block to OldEntry yet,
421 // it will not show up as a predecessor.
422 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
424 if (*PI == &F->getEntryBlock())
425 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
427 AccPN->addIncoming(AccPN, *PI);
430 // Add an incoming argument for the current block, which is computed by our
431 // associative accumulator instruction.
432 AccPN->addIncoming(AccRecInstr, BB);
434 // Next, rewrite the accumulator recursion instruction so that it does not
435 // use the result of the call anymore, instead, use the PHI node we just
437 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
439 // Finally, rewrite any return instructions in the program to return the PHI
440 // node instead of the "initval" that they do currently. This loop will
441 // actually rewrite the return value we are destroying, but that's ok.
442 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
443 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
444 RI->setOperand(0, AccPN);
448 // Now that all of the PHI nodes are in place, remove the call and
449 // ret instructions, replacing them with an unconditional branch.
450 new BranchInst(OldEntry, Ret);
451 BB->getInstList().erase(Ret); // Remove return.
452 BB->getInstList().erase(CI); // Remove call.