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
23 // There are several improvements that could be made:
25 // 1. If the function has any alloca instructions, these instructions will be
26 // moved out of the entry block of the function, causing them to be
27 // evaluated each time through the tail recursion. Safely keeping allocas
28 // in the entry block requires analysis to proves that the tail-called
29 // function does not read or write the stack object.
30 // 2. Tail recursion is only performed if the call immediately preceeds the
31 // return instruction. It's possible that there could be a jump between
32 // the call and the return.
33 // 3. TRE is only performed if the function returns void or if the return
34 // returns the result returned by the call. It is possible, but unlikely,
35 // that the return returns something else (like constant 0), and can still
36 // be TRE'd. It can be TRE'd if ALL OTHER return instructions in the
37 // function return the exact same value.
38 // 4. There can be intervening operations between the call and the return that
39 // prevent the TRE from occurring. For example, there could be GEP's and
40 // stores to memory that will not be read or written by the call. This
41 // requires some substantial analysis (such as with DSA) to prove safe to
42 // move ahead of the call, but doing so could allow many more TREs to be
43 // performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark.
45 //===----------------------------------------------------------------------===//
47 #include "llvm/Transforms/Scalar.h"
48 #include "llvm/DerivedTypes.h"
49 #include "llvm/Function.h"
50 #include "llvm/Instructions.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/CFG.h"
53 #include "Support/Statistic.h"
57 Statistic<> NumEliminated("tailcallelim", "Number of tail calls removed");
58 Statistic<> NumAccumAdded("tailcallelim","Number of accumulators introduced");
60 struct TailCallElim : public FunctionPass {
61 virtual bool runOnFunction(Function &F);
64 bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry,
65 std::vector<PHINode*> &ArgumentPHIs);
66 bool CanMoveAboveCall(Instruction *I, CallInst *CI);
67 Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI);
69 RegisterOpt<TailCallElim> X("tailcallelim", "Tail Call Elimination");
72 // Public interface to the TailCallElimination pass
73 FunctionPass *llvm::createTailCallEliminationPass() {
74 return new TailCallElim();
78 bool TailCallElim::runOnFunction(Function &F) {
79 // If this function is a varargs function, we won't be able to PHI the args
80 // right, so don't even try to convert it...
81 if (F.getFunctionType()->isVarArg()) return false;
83 BasicBlock *OldEntry = 0;
84 std::vector<PHINode*> ArgumentPHIs;
85 bool MadeChange = false;
87 // Loop over the function, looking for any returning blocks...
88 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
89 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator()))
90 MadeChange |= ProcessReturningBlock(Ret, OldEntry, ArgumentPHIs);
96 /// CanMoveAboveCall - Return true if it is safe to move the specified
97 /// instruction from after the call to before the call, assuming that all
98 /// instructions between the call and this instruction are movable.
100 bool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) {
101 // FIXME: We can move load/store/call/free instructions above the call if the
102 // call does not mod/ref the memory location being processed.
103 if (I->mayWriteToMemory() || isa<LoadInst>(I))
106 // Otherwise, if this is a side-effect free instruction, check to make sure
107 // that it does not use the return value of the call. If it doesn't use the
108 // return value of the call, it must only use things that are defined before
109 // the call, or movable instructions between the call and the instruction
111 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
112 if (I->getOperand(i) == CI)
118 /// CanTransformAccumulatorRecursion - If the specified instruction can be
119 /// transformed using accumulator recursion elimination, return the constant
120 /// which is the start of the accumulator value. Otherwise return null.
122 Value *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I,
124 if (!I->isAssociative()) return 0;
125 assert(I->getNumOperands() == 2 &&
126 "Associative operations should have 2 args!");
128 // Exactly one operand should be the result of the call instruction...
129 if (I->getOperand(0) == CI && I->getOperand(1) == CI ||
130 I->getOperand(0) != CI && I->getOperand(1) != CI)
133 // The only user of this instruction we allow is a single return instruction.
134 if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back()))
137 // Ok, now we have to check all of the other return instructions in this
138 // function. If they return non-constants or differing values, then we cannot
139 // transform the function safely.
140 Value *ReturnedValue = 0;
141 Function *F = CI->getParent()->getParent();
143 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
144 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) {
145 Value *RetOp = RI->getOperand(0);
146 if (isa<Constant>(RetOp)) {
147 if (ReturnedValue && RetOp != ReturnedValue)
148 return 0; // Cannot transform if differing constants are returned.
149 ReturnedValue = RetOp;
151 } else if (RetOp != I) { // Ignore the one returning I.
152 return 0; // Not returning a constant, cannot transform.
156 // Ok, if we passed this battery of tests, we can perform accumulator
157 // recursion elimination.
158 return ReturnedValue;
161 bool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry,
162 std::vector<PHINode*> &ArgumentPHIs) {
163 BasicBlock *BB = Ret->getParent();
164 Function *F = BB->getParent();
166 if (&BB->front() == Ret) // Make sure there is something before the ret...
169 // Scan backwards from the return, checking to see if there is a tail call in
170 // this block. If so, set CI to it.
172 BasicBlock::iterator BBI = Ret;
174 CI = dyn_cast<CallInst>(BBI);
175 if (CI && CI->getCalledFunction() == F)
178 if (BBI == BB->begin())
179 return false; // Didn't find a potential tail call.
183 // If we are introducing accumulator recursion to eliminate associative
184 // operations after the call instruction, this variable contains the initial
185 // value for the accumulator. If this value is set, we actually perform
186 // accumulator recursion elimination instead of simple tail recursion
188 Value *AccumulatorRecursionEliminationInitVal = 0;
189 Instruction *AccumulatorRecursionInstr = 0;
191 // Ok, we found a potential tail call. We can currently only transform the
192 // tail call if all of the instructions between the call and the return are
193 // movable to above the call itself, leaving the call next to the return.
194 // Check that this is the case now.
195 for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI)
196 if (!CanMoveAboveCall(BBI, CI)) {
197 // If we can't move the instruction above the call, it might be because it
198 // is an associative operation that could be tranformed using accumulator
199 // recursion elimination. Check to see if this is the case, and if so,
200 // remember the initial accumulator value for later.
201 if ((AccumulatorRecursionEliminationInitVal =
202 CanTransformAccumulatorRecursion(BBI, CI))) {
203 // Yes, this is accumulator recursion. Remember which instruction
205 AccumulatorRecursionInstr = BBI;
207 return false; // Otherwise, we cannot eliminate the tail recursion!
211 // We can only transform call/return pairs that either ignore the return value
212 // of the call and return void, or return the value returned by the tail call.
213 if (Ret->getNumOperands() != 0 && Ret->getReturnValue() != CI &&
214 AccumulatorRecursionEliminationInitVal == 0)
217 // OK! We can transform this tail call. If this is the first one found,
218 // create the new entry block, allowing us to branch back to the old entry.
220 OldEntry = &F->getEntryBlock();
221 std::string OldName = OldEntry->getName(); OldEntry->setName("tailrecurse");
222 BasicBlock *NewEntry = new BasicBlock(OldName, OldEntry);
223 new BranchInst(OldEntry, NewEntry);
225 // Now that we have created a new block, which jumps to the entry
226 // block, insert a PHI node for each argument of the function.
227 // For now, we initialize each PHI to only have the real arguments
228 // which are passed in.
229 Instruction *InsertPos = OldEntry->begin();
230 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I) {
231 PHINode *PN = new PHINode(I->getType(), I->getName()+".tr", InsertPos);
232 I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
233 PN->addIncoming(I, NewEntry);
234 ArgumentPHIs.push_back(PN);
238 // Ok, now that we know we have a pseudo-entry block WITH all of the
239 // required PHI nodes, add entries into the PHI node for the actual
240 // parameters passed into the tail-recursive call.
241 for (unsigned i = 0, e = CI->getNumOperands()-1; i != e; ++i)
242 ArgumentPHIs[i]->addIncoming(CI->getOperand(i+1), BB);
244 // If we are introducing an accumulator variable to eliminate the recursion,
245 // do so now. Note that we _know_ that no subsequent tail recursion
246 // eliminations will happen on this function because of the way the
247 // accumulator recursion predicate is set up.
249 if (AccumulatorRecursionEliminationInitVal) {
250 Instruction *AccRecInstr = AccumulatorRecursionInstr;
251 // Start by inserting a new PHI node for the accumulator.
252 PHINode *AccPN = new PHINode(AccRecInstr->getType(), "accumulator.tr",
255 // Loop over all of the predecessors of the tail recursion block. For the
256 // real entry into the function we seed the PHI with the initial value,
257 // computed earlier. For any other existing branches to this block (due to
258 // other tail recursions eliminated) the accumulator is not modified.
259 // Because we haven't added the branch in the current block to OldEntry yet,
260 // it will not show up as a predecessor.
261 for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry);
263 if (*PI == &F->getEntryBlock())
264 AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, *PI);
266 AccPN->addIncoming(AccPN, *PI);
269 // Add an incoming argument for the current block, which is computed by our
270 // associative accumulator instruction.
271 AccPN->addIncoming(AccRecInstr, BB);
273 // Next, rewrite the accumulator recursion instruction so that it does not
274 // use the result of the call anymore, instead, use the PHI node we just
276 AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN);
278 // Finally, rewrite any return instructions in the program to return the PHI
279 // node instead of the "initval" that they do currently. This loop will
280 // actually rewrite the return value we are destroying, but that's ok.
281 for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI)
282 if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()))
283 RI->setOperand(0, AccPN);
287 // Now that all of the PHI nodes are in place, remove the call and
288 // ret instructions, replacing them with an unconditional branch.
289 new BranchInst(OldEntry, Ret);
290 BB->getInstList().erase(Ret); // Remove return.
291 BB->getInstList().erase(CI); // Remove call.