https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113834
--- Comment #7 from Marek Polacek <mpolacek at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #6)
> (In reply to Marek Polacek from comment #5)
> > To fix the ICE we could do:
> >
> > --- a/gcc/cp/semantics.cc
> > +++ b/gcc/cp/semantics.cc
> > @@ -4644,7 +4644,9 @@ static tree
> > finish_type_pack_element (tree idx, tree types, tsubst_flags_t complain)
> > {
> > idx = maybe_constant_value (idx);
> > - if (TREE_CODE (idx) != INTEGER_CST || !INTEGRAL_TYPE_P (TREE_TYPE (idx)))
> > + if (TREE_CODE (idx) != INTEGER_CST
> > + || !INTEGRAL_TYPE_P (TREE_TYPE (idx))
> > + || !tree_fits_shwi_p (idx))
> > {
> > if (complain & tf_error)
> > error ("%<__type_pack_element%> index is not an integral constant");
>
> But then the diagnostics is confusing.
There's a lot of places where we do exactly that:
handle_assume_aligned_attribute, ...
> Perhaps use tree_int_cst_sgn (idx) < 0 instead of tree_to_shwi + val < 0,
> wi::to_widest (idx) >= TREE_VEC_LENGTH (types) for out of range and
> only use tree_to_shwi after those checks?
That sounds good too though, I can do that instead.