[PATCH] c++: ICE with template NEW_EXPR [PR105803]

[Jason Merrill]

On 6/2/22 10:03, Marek Polacek wrote:
> On Thu, Jun 02, 2022 at 08:42:36AM -0400, Patrick Palka wrote:
>> On Wed, 1 Jun 2022, Marek Polacek via Gcc-patches wrote:
>>
>>> Here we ICE because value_dependent_expression_p gets a NEW_EXPR
>>> whose operand is a type, and we go to the default case which just
>>> calls v_d_e_p on each operand of the NEW_EXPR.  Since one of them
>>> is a type, we crash on the new assert in t_d_e_p.
>>
>> Looks like NEW_EXPR is considered to be not potentially constant
>> according to potential_constant_expression.  I thought we shouldn't
>> be calling value_dependent_expression_p on such exprs?

Except in C++20 new-expressions are potentially constant, so the patch 
is OK, and we should adjust pce1 accordingly.

I notice we currently fail to handle

struct A
{
   int i;
   constexpr A(int *p): i(*p) { delete p; }
};

constexpr int i = A(new int(42)).i;

though it does work inside a function.

> You're correct.  This is non-obvious: instantiation_dependent_expression_p
> calls p_c_e before v_d_e_p, but the expression is CAST_EXPR<[NEW_EXPR]>,
> where the [] denotes a TREE_LIST, created in cp_parser_functional_cast.
> This TREE_LIST has no type.  So p_c_e_1/CAST_EXPR goes to
> 9183           /* If this is a dependent type, it could end up being a class
> 9184              with conversions.  */
> 9185           if (type == NULL_TREE || WILDCARD_TYPE_P (type))
> 9186             return true;
> and returns true.
> 
> So we call v_d_e_p, which looks at the CAST_EXPR's op and sees a TREE_LIST,
> so it calls any_value_dependent_elements_p, and we end up with a NEW_EXPR.
> 
> An alternative/more proper fix would be to fix p_c_e_1/CAST_EXPR.  Maybe
> by calling any_type_dependent_elements_p (which currently has no uses).
> Thoughts?
> 
>>> --- /dev/null
>>> +++ b/gcc/testsuite/g++.dg/template/new13.C
>>> @@ -0,0 +1,11 @@
>>> +// PR c++/105803
>>> +// { dg-do compile }
>>> +// { dg-additional-options "-fchecking=2" }
>>> +
>>> +namespace std {
>>> +template <typename> class shared_ptr;
>>> +}
>>> +struct S {};
>>> +template <int> void build_matrices() {
>>> +  std::shared_ptr<S>(new S);
>>> +}
>>
>> I think this testcase might be IFNDR since shared_ptr<S> is incomplete
>> at the point of its non-dependent use.
> 
> Ah, overreduced.  I've made shared_ptr complete.
> 
> Marek
>