[llvm-commits] patch: CXAGuardElimination pass.

[Duncan Sands]

Hi Nick,

> +    CallInst *A = dyn_cast<CallInst>(I);
> +    if (!A) continue;
> +
> +    // TODO: __cxa_guard_acquire returns a value, then makes a conditional
> +    // branch off of it. We need to handle this.

the acquire function may be being passed as a parameter to some other
function rather than being called.

> +      if (I->mayReadFromMemory() || I->mayWriteToMemory())

Maybe you should use mayHaveSideEffects rather than mayWriteToMemory,
since a "const" function can still throw an exception.  On the other
hand, since such a function doesn't read or write memory it shouldn't
need to be wrapped in guards anyway.

> +BasicBlock *CXAGuardElimination::getUniqueSuccessor(TerminatorInst *TI,
> +                                                    CallInst *Acquire) {
> +  if (TI->getNumSuccessors() == 1)
> +    return TI->getSuccessor(0);

What about this:

bb1: acquire
      br bbr
bb2: acquire
      br bbr
bbr: release

Aren't you going to eliminate the first acquire and the release, but
not the second acquire?

> +  addPass(PM, createCXAGuardEliminationPass());  // Remove __cxa_guard_* calls.

Are you going to add this to llvm-gcc too?

> +  This pass deletes dead calls to __cxa_guard_acquire and __cxa_guard_release
> +  which are part of the Itanium C++ ABI. T hese are used to prevent a function

T hese -> These

Ciao,

Duncan.