[dmd-internals] Mixin forward reference semantic runs and field declarations

[David Nadlinger]

I am currently working on ironing out the last few kinks in the LDC
merge of the 2.060 frontend. The following snippet, a test case for
Bugzilla issue 7974, currently doesn't work with LDC:

---
mixin template mix7974() {
  uint _x;
}

struct Foo7974 {
  immutable fa = Foo7974(0);

  this(uint x) {
    _x = x;
  }
  mixin mix7974!();
}
---

The problem is that the fix Walter applied to the frontend for 7974
causes the semantic pass, including the arrayCopy() of the mixin
contents into the structs, to run twice, which leads to two separate
VarDeclarations for Foo7974._x being present at codegen time: one
instance, the one from the »forward reference« semantic run, is
referred to by the struct constructor, the other instance from the
second, »proper« run is e.g. what ends up in
StructDeclaration::fields.

As we tuck on IR construction metadata to field declarations in LDC,
having two instances for the same field breaks our codegen in
interesting ways – is this an unforeseen consequence of the fix in DMD
and  possibly a bug, or do I have to find a way to work around this in
LDC?

Thanks,
David