On 22/09/16 12:15 +0100, Jonathan Wakely wrote:
>On 22/09/16 11:16 +0100, Jonathan Wakely wrote:
>>(Somebody should fix PR58938 so exception_ptr is portable).
>
>Christophe, would you be able to test this patch?
>
>It uses a single global mutex for exception_ptr objects, which doesn't
>scale well but that probably isn't a problem for processors without
>lock-free atomics for single words.
>
>This also solves the problem of mismatched -march options, where the
>header is compiled for a CPU that supports the atomics but
>libstdc++.so was built for an older CPU that doesn't support them, and
>linking fails (as in https://gcc.gnu.org/PR58938#c13).

We'd also need something like this extra piece, to ensure we don't
leak exceptions. Currently __gxx_exception_cleanup assumes that if
ATOMIC_INT_LOCK_FREE < 2 the referenceCount can never be greater than
1, because there are not exception_ptr objects that could increase it.

If we enable exception_ptr unconditionally then that assumption
doesn't hold. This patch uses the exception_ptr code to do the
cleanup, under the same mutex as any other increments and decrements
of the reference count.

It's a bit of a hack though.