ASSERT_VAL in member initializer of constexpr constructor causes ICE on GCC

[CrackedMatter]

Using ASSERT_VAL in a member initializer of a constexpr constructor causes an ICE on GCC:

// internal compiler error: in bot_manip, at cp/tree.cc:3280
struct a {
    constexpr a() : b{ASSERT_VAL(1)} {}
    int b;
};

The error appears to be triggered when a name is introduced within a statement expression (https://godbolt.org/z/38x1E1jY6):

struct a {
    int b;
    constexpr a() : b{({
        int c{}; // struct d; typedef int e; ...
        1;
    })} {}
};

[jeremy-rifkin]

Hi, thanks for the report. Have you already reported to gcc?

[CrackedMatter]

I have not.

[jeremy-rifkin]

Unfortunately since this is a gcc bug there's not a ton that can be done on the libassert side of things. I can file a report to gcc's bugtracker later unless you'd like to do so.

For now the following workaround can be used:

struct a {
    constexpr a() : b{[&]{return ASSERT_VAL(1);}()} {}
    int b;
};

This could be worked around by libassert with an IILE instead of a statement expression, however, I suspect there is a higher cost to lambdas than statement expressions.

[CrackedMatter]

I can file a report to gcc's bugtracker later

Yes, that would be great.

[jeremy-rifkin]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122730