ahatanak marked an inline comment as done.
ahatanak added inline comments.
================
Comment at: lib/CodeGen/MicrosoftCXXABI.cpp:833
// passed in registers, which is non-conforming.
if (RD->hasNonTrivialDestructor() &&
getContext().getTypeSize(RD->getTypeForDecl()) > 64)
----------------
rsmith wrote:
> Should we be checking `hasNonTrivialDestructorForCalls` here? What should
> happen if we have a small class whose copy constructor is trivial for calls
> but whose destructor is not? The existing machinations of this ABI only work
> because they can make an additional trivial copy of the function parameter
> when passing it direct, *even if* it has a non-trivial destructor. I don't
> think that's correct in general if the copy constructor is only
> trival-for-calls. Consider:
>
> ```
> struct X {
> shared_ptr<T> sp; // shared_ptr<T> is trivial-for-calls
> ~X();
> };
> ```
>
> In the absence of the destructor, X should be passed direct.
> In the presence of the destructor, though, passing direct would result in the
> X destructor running in both caller and callee, and X being passed through
> registers without running the copy constructor. Which results in a
> double-delete.
>
> So I think what we want here is this:
>
> * If the copy ctor and dtor are both trivial-for-calls, pass direct.
> * Otherwise, if the copy ctor is trivial (not just trivial-for-calls) and
> the object is small, pass direct (regardless of the triviality of the dtor --
> we'll make a copy, notionally using the trivial copy ctor, and destroy both
> the original and the copy).
> * Otherwise, pass indirect.
I'm not very familiar with microsoft's ABI, but I think that makes sense.
https://reviews.llvm.org/D41039
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
