On Tue, Jul 14, 2026 at 11:35:04PM -0400, Jason Merrill wrote:
> On 7/14/26 7:06 PM, Marek Polacek wrote:
> > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk/16?
> >
> > -- >8 --
> > Like in PR110822, the attached test produces an ICE in verify_address:
> >
> > error: constant not recomputed when 'ADDR_EXPR' changed
> >
> > since r10-7718, but it wasn't fixed by r15-7762. The reason is that
> > here we have:
> >
> > {.D.3013={.D.3006={.a={.ptr=&<retval>.D.3013.D.3006.a}},
> > .D.3007={.b={.ptr=&<retval>.D.3013.D.3007.b}}}}
> >
> > and we replace '<retval>' with 'names', but we only call
> > recompute_tree_invariant_expr for the first ADDR_EXPR, not the latter.
> > For the second ADDR_EXPR d->changed will be false: the first replacement
> > changed the shared tree <retval>.D.3013. So I'm afraid we have to
> > recompute the flags unconditionally. It may be a bit slower, but it's
> > correct.
>
> OK for 16.
>
> But how do we end up with this sharing? It looks like
> replace_placeholders_r properly does unshare_expr.
There are no PLACEHOLDER_EXPR so the unshare_expr in replace_placeholders_r
won't take place. There's an unshare_expr in replace_decl_r but that also
doesn't unshare the COMPONENT_REF because the condition before the
unshare_expr is *tp == d->decl which won't be true. I'm not sure if I'm
answering your question.
Marek
