C++ PATCH to fix rejects-valid with constexpr ctor in C++17 (PR c++/83692)

[Jason Merrill]

On Mon, Feb 5, 2018 at 8:37 AM, Marek Polacek <polacek@redhat.com> wrote:
> On Fri, Feb 02, 2018 at 02:11:27PM -0500, Jason Merrill wrote:
>> On Thu, Jan 25, 2018 at 4:16 PM, Marek Polacek <polacek@redhat.com> wrote:
>> > This is a similar problem to 83116: we'd cached a constexpr call, but after a
>> > store the result had become invalid, yet we used the wrong result again when
>> > encountering the same call later.  This resulted in evaluating a THROW_EXPR
>> > which doesn't work.  Details in
>> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83692#c5
>> >
>> > The fix for 83116 didn't work here, because when evaluating the body of the
>> > ctor via store_init_value -> cxx_constant_value we are in STRICT, so we do
>> > cache.
>>
>> > It seems that we may no longer rely on the constexpr call table when we
>> > do cxx_eval_store_expression, because that just rewrites *valp, i.e. the
>> > value of an object.  Might be too big a hammer again, but I couldn't think
>> > of how I could guard the caching of a constexpr call.
>>
>> > This doesn't manifest in C++14 because build_special_member_call in C++17 is
>> > more aggressive with copy elisions (as required by P0135 which changed how we
>> > view prvalues).  In C++14 build_special_member_call produces a CALL_EXPR, so
>> > expand_default_init calls maybe_constant_init, for which STRICT is false, so
>> > we avoid caching as per 83116.
>>
>> So it sounds like the problem is using cxx_constant_value for the
>> diagnostic when it has different semantics from the
>> maybe_constant_init that follows right after.  I guess we want a
>> cxx_constant_init function that is a hybrid of the two.
>
> So like the following?  Thanks,
>
> Bootstrapped/regtested on x86_64-linux, ok for trunk?
>
> 2018-02-04  Marek Polacek  <polacek@redhat.com>
>
>         PR c++/83692
>         * constexpr.c (cxx_constant_init): New function.
>         * cp-tree.h (cxx_constant_init): Declare.
>         * typeck2.c (store_init_value): Call cxx_constant_init instead of
>         cxx_constant_value.
>
> +/* Like cxx_constant_value, but non-strict mode.  */
> +
> +tree
> +cxx_constant_init (tree t, tree decl)
> +{
> +  return cxx_eval_outermost_constant_expr (t, false, false, decl);
> +}

Hmm, that doesn't do the TARGET_EXPR stripping that
maybe_constant_init does.  I was thinking of a version of
maybe_constant_init that passes false to allow_non_constant.  Probably
by making "maybe_constant_init" and cxx_constant_init both call the
current function with an additional parameter.  And then the existing
call to maybe_constant_init can move under an 'else' to avoid
redundant constexpr evaluation.

Jason