Pointer alignment errors occurring inconsistently

[GoldenStack]

Zig Version

0.15.0-dev.149+2b57f6b71

Steps to Reproduce and Observed Behavior

I've come across a few inconsistencies with aligned fields in structs. Primarily, the following code produces error: struct layout depends on being pointer aligned (pointing to the definition of A):

pub fn main() !void {
    _ = struct { a: A }; // Line A
    _ = A; // Line B
}

pub const A = struct {
    b: struct {
        a: *align(@alignOf(struct { a: A })) void,
    },
    a: *A,
    c: struct {
        a: *align(@alignOf(struct { a: A })) void,
    },
    d: u128,
};

but if you swap the order of line A and line B, no such error occurs. Furthermore, commenting out line B makes it error once again, but an error is not produced with solely line B.

Finally, if a: *A is moved up as to become the first field, with just line B (so line A is commented out), the error occurs once again.

However, if struct { a: *align(@alignOf(struct { a: A })) void } is extracted to a separate variable (e.g. B) and is referred to by A (b: B, c: B), the error does not occur in any case that I have tested.

My best guess as to why this is occurring is that the alignment of the struct is being retrieved before the alignment has been correctly set (since A depends on the alignment of itself) so the comptime evaluation of the types of fields of A is inconsistently wrong, dependent on the order of how the expressions are evaluated.

Expected Behavior

To my understanding, none of this code should error, but at a minimum, reordering the two lines in main should not change whether or not there is a compile error.

[GoldenStack]

This doesn't occur on 0.13.0 (as well as 0.12.0, 0.11.0, 0.10.0), and occurs on 0.14.0 and master, so I assume this is a regression and was introduced somewhere between 0.13.0 and 0.14.0. I haven't bisected individual commits though.