// evaluated each time through the tail recursion. Safely keeping allocas
// in the entry block requires analysis to proves that the tail-called
// function does not read or write the stack object.
-// 2. Tail recursion is only performed if the call immediately preceeds the
+// 2. Tail recursion is only performed if the call immediately precedes the
// return instruction. It's possible that there could be a jump between
// the call and the return.
// 3. There can be intervening operations between the call and the return that
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
+#include "llvm/Module.h"
#include "llvm/Pass.h"
#include "llvm/Analysis/CaptureTracking.h"
#include "llvm/Analysis/InlineCost.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/ValueHandle.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
using namespace llvm;
return new TailCallElim();
}
-/// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by
-/// callees of this function. We only do very simple analysis right now, this
-/// could be expanded in the future to use mod/ref information for particular
-/// call sites if desired.
-static bool AllocaMightEscapeToCalls(AllocaInst *AI) {
- // FIXME: do simple 'address taken' analysis.
- return true;
+/// CanTRE - Scan the specified basic block for alloca instructions.
+/// If it contains any that are variable-sized or not in the entry block,
+/// returns false.
+static bool CanTRE(AllocaInst *AI) {
+ // Because of PR962, we don't TRE allocas outside the entry block.
+
+ // If this alloca is in the body of the function, or if it is a variable
+ // sized allocation, we cannot tail call eliminate calls marked 'tail'
+ // with this mechanism.
+ BasicBlock *BB = AI->getParent();
+ return BB == &BB->getParent()->getEntryBlock() &&
+ isa<ConstantInt>(AI->getArraySize());
}
-/// CheckForEscapingAllocas - Scan the specified basic block for alloca
-/// instructions. If it contains any that might be accessed by calls, return
-/// true.
-static bool CheckForEscapingAllocas(BasicBlock *BB,
- bool &CannotTCETailMarkedCall) {
- bool RetVal = false;
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
- if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
- RetVal |= AllocaMightEscapeToCalls(AI);
-
- // If this alloca is in the body of the function, or if it is a variable
- // sized allocation, we cannot tail call eliminate calls marked 'tail'
- // with this mechanism.
- if (BB != &BB->getParent()->getEntryBlock() ||
- !isa<ConstantInt>(AI->getArraySize()))
- CannotTCETailMarkedCall = true;
- }
- return RetVal;
-}
+struct AllocaCaptureTracker : public CaptureTracker {
+ AllocaCaptureTracker() : Captured(false) {}
+
+ void tooManyUses() { Captured = true; }
+
+ bool shouldExplore(Use *U) {
+ Value *V = U->getUser();
+ if (isa<CallInst>(V) || isa<InvokeInst>(V))
+ UsesAlloca.push_back(V);
+
+ return true;
+ }
+
+ bool captured(Use *U) {
+ if (isa<ReturnInst>(U->getUser()))
+ return false;
+
+ Captured = true;
+ return true;
+ }
+
+ SmallVector<WeakVH, 64> UsesAlloca;
+
+ bool Captured;
+};
bool TailCallElim::runOnFunction(Function &F) {
// If this function is a varargs function, we won't be able to PHI the args
bool MadeChange = false;
bool FunctionContainsEscapingAllocas = false;
- // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls
+ // CanTRETailMarkedCall - If false, we cannot perform TRE on tail calls
// marked with the 'tail' attribute, because doing so would cause the stack
- // size to increase (real TCE would deallocate variable sized allocas, TCE
+ // size to increase (real TRE would deallocate variable sized allocas, TRE
// doesn't).
- bool CannotTCETailMarkedCall = false;
-
- // Loop over the function, looking for any returning blocks, and keeping track
- // of whether this function has any non-trivially used allocas.
- for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
- if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall)
- break;
-
- FunctionContainsEscapingAllocas |=
- CheckForEscapingAllocas(BB, CannotTCETailMarkedCall);
+ bool CanTRETailMarkedCall = true;
+
+ // Find calls that can be marked tail.
+ AllocaCaptureTracker ACT;
+ for (Function::iterator BB = F.begin(), EE = F.end(); BB != EE; ++BB) {
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
+ CanTRETailMarkedCall &= CanTRE(AI);
+ PointerMayBeCaptured(AI, &ACT);
+ if (ACT.Captured)
+ return false;
+ }
+ }
}
-
- /// FIXME: The code generator produces really bad code when an 'escaping
- /// alloca' is changed from being a static alloca to being a dynamic alloca.
- /// Until this is resolved, disable this transformation if that would ever
- /// happen. This bug is PR962.
- if (FunctionContainsEscapingAllocas)
- return false;
- // Second pass, change any tail calls to loops.
+ // Second pass, change any tail recursive calls to loops.
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) {
bool Change = ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail,
- ArgumentPHIs,CannotTCETailMarkedCall);
+ ArgumentPHIs, !CanTRETailMarkedCall);
if (!Change && BB->getFirstNonPHIOrDbg() == Ret)
Change = FoldReturnAndProcessPred(BB, Ret, OldEntry,
TailCallsAreMarkedTail, ArgumentPHIs,
- CannotTCETailMarkedCall);
+ !CanTRETailMarkedCall);
MadeChange |= Change;
}
}
}
}
- // Finally, if this function contains no non-escaping allocas, mark all calls
- // in the function as eligible for tail calls (there is no stack memory for
- // them to access).
- if (!FunctionContainsEscapingAllocas)
+ // Finally, if this function contains no non-escaping allocas and doesn't
+ // call setjmp, mark all calls in the function as eligible for tail calls
+ // (there is no stack memory for them to access).
+ std::sort(ACT.UsesAlloca.begin(), ACT.UsesAlloca.end());
+
+ if (!FunctionContainsEscapingAllocas && !F.callsFunctionThatReturnsTwice())
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
- if (CallInst *CI = dyn_cast<CallInst>(I)) {
- CI->setTailCall();
- MadeChange = true;
- }
+ if (CallInst *CI = dyn_cast<CallInst>(I))
+ if (!std::binary_search(ACT.UsesAlloca.begin(), ACT.UsesAlloca.end(),
+ CI)) {
+ CI->setTailCall();
+ MadeChange = true;
+ }
return MadeChange;
}
-
/// CanMoveAboveCall - Return true if it is safe to move the specified
/// instruction from after the call to before the call, assuming that all
/// instructions between the call and this instruction are movable.
// call does not mod/ref the memory location being processed.
if (I->mayHaveSideEffects()) // This also handles volatile loads.
return false;
-
+
if (LoadInst *L = dyn_cast<LoadInst>(I)) {
// Loads may always be moved above calls without side effects.
if (CI->mayHaveSideEffects()) {
if (&BB->front() == TI) // Make sure there is something before the terminator.
return 0;
-
+
// Scan backwards from the return, checking to see if there is a tail call in
// this block. If so, set CI to it.
CallInst *CI = 0;
// double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call
// and disable this xform in this case, because the code generator will
// lower the call to fabs into inline code.
- if (BB == &F->getEntryBlock() &&
+ if (BB == &F->getEntryBlock() &&
FirstNonDbg(BB->front()) == CI &&
FirstNonDbg(llvm::next(BB->begin())) == TI &&
- callIsSmall(F)) {
+ callIsSmall(CI)) {
// A single-block function with just a call and a return. Check that
// the arguments match.
CallSite::arg_iterator I = CallSite(CI).arg_begin(),
BasicBlock::iterator BBI = CI;
for (++BBI; &*BBI != Ret; ++BBI) {
if (CanMoveAboveCall(BBI, CI)) continue;
-
+
// If we can't move the instruction above the call, it might be because it
- // is an associative and commutative operation that could be tranformed
+ // is an associative and commutative operation that could be transformed
// using accumulator recursion elimination. Check to see if this is the
// case, and if so, remember the initial accumulator value for later.
if ((AccumulatorRecursionEliminationInitVal =
Instruction *InsertPos = OldEntry->begin();
for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
I != E; ++I) {
- PHINode *PN = PHINode::Create(I->getType(),
+ PHINode *PN = PHINode::Create(I->getType(), 2,
I->getName() + ".tr", InsertPos);
- PN->reserveOperandSpace(2);
I->replaceAllUsesWith(PN); // Everyone use the PHI node now!
PN->addIncoming(I, NewEntry);
ArgumentPHIs.push_back(PN);
pred_iterator PB = pred_begin(OldEntry), PE = pred_end(OldEntry);
PHINode *AccPN =
PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(),
+ std::distance(PB, PE) + 1,
"accumulator.tr", OldEntry->begin());
- AccPN->reserveOperandSpace(std::distance(PB, PE) + 1);
// Loop over all of the predecessors of the tail recursion block. For the
// real entry into the function we seed the PHI with the initial value,
// Now that all of the PHI nodes are in place, remove the call and
// ret instructions, replacing them with an unconditional branch.
- BranchInst::Create(OldEntry, Ret);
+ BranchInst *NewBI = BranchInst::Create(OldEntry, Ret);
+ NewBI->setDebugLoc(CI->getDebugLoc());
+
BB->getInstList().erase(Ret); // Remove return.
BB->getInstList().erase(CI); // Remove call.
++NumEliminated;