Re: [PATCH v2] c++: ICE with VEC_INIT_EXPR and defarg [PR106925]

[Jason Merrill <jason@redhat.com>]

On 10/12/22 14:23, Marek Polacek wrote:
> On Wed, Oct 12, 2022 at 01:12:57PM -0400, Marek Polacek wrote:
>> On Wed, Oct 12, 2022 at 12:47:21PM -0400, Jason Merrill wrote:
>>> On 10/12/22 12:27, Marek Polacek wrote:
>>>> On Tue, Oct 11, 2022 at 04:28:11PM -0400, Jason Merrill wrote:
>>>>> On 10/11/22 16:00, Marek Polacek wrote:
>>>>>> Since r12-8066, in cxx_eval_vec_init we perform expand_vec_init_expr
>>>>>> while processing the default argument in this test.
>>>>>
>>>>> Hmm, why are we calling cxx_eval_vec_init during parsing of the default
>>>>> argument?  In particular, any expansion that depends on the enclosing
>>>>> function context should be deferred until the default arg is used by a call.
>>>>
>>>> I think this is part of the semantic constraints checking [dcl.fct.default]/5
>>>> talks about, as in, this doesn't compile even though the default argument is
>>>> not executed:
>>>>
>>>> struct S {
>>>>     S() = delete;
>>>> };
>>>> void foo (S = S()) { }
>>>> In the test below we parse '= MyVector<1>()' and end up calling mark_used
>>>> on the implicit "constexpr MyVector<1>::MyVector() noexcept (<uninstantiated>)"
>>>> ctor.  mark_used calls maybe_instantiate_noexcept.  Since the ctor has
>>>> a DEFERRED_NOEXCEPT, we have to figure out if the ctor should be noexcept
>>>> or not using get_defaulted_eh_spec.  That means walking the members of
>>>> MyVector.  Thus we reach
>>>>     /* Core 1351: If the field has an NSDMI that could throw, the
>>>>        default constructor is noexcept(false).  */
>>>
>>> Maybe we need a cp_unevaluated here?  The operand of noexcept should be
>>> unevaluated.
>>
>> That wouldn't help since get_nsdmi specifically does "cp_evaluated ev;",
>> so...
>>   
>>>> and call get_nsdmi on 'data'.  There we digest its initializer which is {}.
>>>> massage_init_elt calls digest_init_r on the {} and produces
>>>>     TARGET_EXPR <D.2518, <<< Unknown tree: vec_init_expr
>>>>       D.2518
>>>>       {} >>>>
>>>> and the subsequent fold_non_dependent_init leads to cxx_eval_vec_init
>>>> -> expand_vec_init_expr.
>>>>
>>>> I think this is all correct except that the fold_non_dependent_init is
>>>> somewhat questionable to me; do we really have to fold in order to say
>>>> if the NSDMI init can throw?  Sure, we need to digest the {}, maybe
>>>> the field's ctors can throw, but I don't know about the folding.
>>>
>>> And we can check cp_unevaluated_operand to avoid the
>>> fold_non_dependent_init?
>>
>> ...we'd still fold.  I'm not sure if we want a LOOKUP_ flag that says
>> "we're just checking if we can throw, don't fold".
> 
> Eh, a new flag is overkill.  Maybe don't do cp_evaluated in get_nsdmi if
> we're called from walk_field_subobs would be worth a try?

It seems that we treat DMI instantiations as evaluated even if they're 
triggered from unevaluated context so sharing lambdas between different 
uses of the DMI works properly.  I don't think this is worth messing 
with at this point; thanks for satisfying my curiosity.

Jason