[Bug c++/87897] [9 Regression] ICE in maybe_constant_value, at cp/constexpr.c:5255 since r265788

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87897

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Slightly cleaned up testcase:
typedef struct A {} B;
A const a = B ();

The problem is that we have a TARGET_EXPR with B type and an empty CONSTRUCTOR
with B type, then cxx_eval_outermost_constant_expr is processing it,
5111      if (TREE_CODE (r) == CONSTRUCTOR && CLASS_TYPE_P (TREE_TYPE (r)))
5112        {
5113          r = adjust_temp_type (type, r);
5114          if (TREE_CODE (t) == TARGET_EXPR
5115              && TARGET_EXPR_INITIAL (t) == r)
5116            return t;
5117          else if (TREE_CODE (t) != CONSTRUCTOR)
5118            {
5119              r = get_target_expr (r);
5120              TREE_CONSTANT (r) = true;
5121            }
and then in the caller
5259      r = cxx_eval_outermost_constant_expr (t, true, true, false, decl);
5260      gcc_checking_assert (r == t
5261                           || CONVERT_EXPR_P (t)
5262                           || TREE_CODE (t) == VIEW_CONVERT_EXPR
5263                           || (TREE_CONSTANT (t) && !TREE_CONSTANT (r))
5264                           || !cp_tree_equal (r, t));

For A const a = A (); adjust_temp_type returns the passed r and thus r == t in
the caller, but for A const a = B (); adjust_temp_type creates a new
CONSTRUCTOR (with A type), also a new TARGET_EXPR because of that, but both are
cp_tree_equal, as it doesn't care about exact type as long as it has
same_type_p.

So, shall adjust_temp_type not bother if it is same_type_p, or shall the assert
be adjusted, or shall we adjust the type earlier already?  The assert is there
already since r166166.