Overlapping field initialization in CTFE-generated union literal

[ibuclaw]

Reduced from dlang/phobos#9064

union U
{
    struct S
    {
        long l;
    }

    int i;
    S s;
}

enum ctfeInit = {
    U u;
    u.s.l = 1;
    return u;
}();

void main()
{
    static a = ctfeInit;
    auto b = ctfeInit;
    assert(a is b);
}

Assert fails because what CTFE generates is:

enum U ctfeInit = U(0, S(1L));

And dmd backend codegen for static initializers ignores overlapping fields, resulting in the value becoming U(0), while the runtime initializer sets each field individually as b.i = 0; b.s.l = 1;.

---

Workaround is to declare variable as being void initialized.

enum ctfeInit = {
    U u = void;
    u.s.l = 1;
    return u;
}();

This results in the correct value being generated.

enum U ctfeInit = U(, S(1L));

[kinke]

u.s.l = 1; - I think that's the problem, setting the u.s union field implicitly by setting its member (u.s.l).

In

dmd/compiler/src/dmd/dinterpret.d

Lines 3871 to 3872 in a4cbc08

	// If it's a union, set all other members of this union to void
	stompOverlappedFields(sle, v);

we make sure the direct overlapping fields are stomped. So if u.s was a union, setting u.s.l would stomp all existing initializers for other u.s members. But this code doesn't take into account that there might be overlaps with parent fields (u.i here).

In v2.111.0-beta.1, we end up with overlapping initializers for std.json.JSONValue.emptyObject: JSONValue(Store(null, , , , Object(false, null, ), ), JSONType.object) (wrong: the first null initializer for JSONValue.Store.str, overlapping with the JSONValue.Store.object field). ldc-developers/ldc#4374 (comment)

[kinke]

Another testcase from ldc-developers/ldc#4374, again by setting a union member indirectly/partially by setting a member of that member:

union U {
    int x;
    int[1] arr;
}

U make() {
    U u;
    u.arr[0] = 123;
    return u;
}

void main()
{
    enum u = make(); // overlapping initializers: `U(0, [123])`
    auto r = u;
}