https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105297
--- Comment #15 from Jiu Fu Guo <guojiufu at gcc dot gnu.org> ---
(In reply to Patrick Palka from comment #13)
> (In reply to Jiu Fu Guo from comment #11)
> > (In reply to Patrick Palka from comment #10)
> > >
> > > Interestingly that doesn't seem to make a difference. What seems to
> > > matter
> > > is whether the constexpr function modifies the CONSTRUCTOR that it
> > > returns:
> > >
> > > constexpr auto foo() {
> > > struct S { int d; } t = {};
> > > t.d = 0; // doesn't ICE if this line is commented out
> > > return t;
> > > }
> > >
> > > template<int>
> > > int bar() {
> > > constexpr auto t = foo();
> > > return 0;
> > > }
> >
> > Right, it is weird. Some PRs on Xtreme-* failure (including ICE) were also
> > reported before. e.g. PR100052, PR101853, PR99910. As commented in those
> > PRs, these may be random failures, and changes in headers that could expose
> > the ICE.
> > I'm also wondering if this may be an issue hidden inside somewhere (GC?).
>
> In this case I suspect it's just a bug in the modules code, I opened
> PR105322 to track it.
Oh, thanks! This failure seems only about the module code on 'struct member
cross functions'.
