Re: [PR69315] enable finish_function to recurse for constexpr functions

[Jakub Jelinek <jakub@redhat.com>]

On Tue, Mar 22, 2016 at 05:28:10PM -0400, Jason Merrill wrote:
> Jakub, you added defer_mark_used_calls for BZ 37189, do you think it's still
> needed?  The testcase passes without it now.

That's a question.  Digging through history, I found:
1) r149750 aka gimplification unit-at-a-time
2) PR48869 that fixed the regression caused by that; so we now should be
   trying to get_copy_ctor/get_dtor for parallel/task already during
   cp-genericize.c
But, looking at this function, e.g. taskloop construct isn't handled,
will try to construct a testcase where it could matter.  Dunno about OpenACC
or Cilk+ constructs, those aren't handled during cp-gimplify.c either.
Though, as we have gimplification unit-at-a-time, I'd say that even if we
synthetize_method during gimplification of something, we'd still enqueue it
in cgraph for later analysis instead of gimplification right away.
The gimplifier still asserts that it isn't called recursively, so I'd hope
the patch can be reverted (test kept).

	Jakub