@andrewrk I've just push a change that recursively checks for undefined in ir_resolve_const. This catches the undefined array case, but will also catch the more general "pass undefined to comptime parameter" case.

Before i look into the failing tests, i'd like to know if this like the right solution. The compiler currently already stops undefined from being passed at comptime, but maybe we want that to work?

@Hejsil I think it should probably work how Valgrind treats undefined. Values can be undefined and that's fine. You can set values to undefined or pass undefined to a comptime function call. But if control flow tries to use an undefined value to decide whether to branch, then it is a compile error. Various other operations would be compile errors too, such as trying to deref a pointer which is undefined, or null-unwrap an optional whose value is undefined.

What about this case then?

fn a(comptime i: u8) u8 { return i }

test "" {
    _ = a(undefined);
}

Currently, this fails at callsite, but I would assume this should be made to work then?

Yeah I think this should work. What do you think?

Btw I don't mean to make this pull request extra work. We can merge your stuff and open a new issue for the extra changes. (As long as tests are in a passing state)

The above example was mostly concerned with the caching of comptime functions and undefined, but I guess we just don't cache values that have some undefined value.

I'll probably revert the latest two commits and merge, then open a bug for passing undefined arrays to comptime parameters.