892 – Another bug in the new GC - pointers in mixins

This is actually with DMD version 1.003 (and I suspect 1.001, and 1.002).
This program destroys the object from the mixin even though it is still very much referenced by live objects in the program.
------
import std.stdio;
static import std.gc;
class SomeObject
{
 this() { writefln("SomeObject created"); }
 ~this() { writefln("SomeObject destroyed"); }
}
template Mix()
{
 void init() {
 ptr = new SomeObject;
 }
 SomeObject ptr;
}
class Container
{
 this() { init(); }
 mixin Mix;
}
void main()
{
 auto x = new Container;
 writefln("---Pre collect");
 std.gc.fullCollect();
 writefln("---Post collect");
}

OK, the compiler was being a bit too smart apparently, realizing that x would no longer be used. That makes the test appear to fail all the way back to DMD 0.176. So I should have put writeflns in at the end to make sure the data had a reason to stick around.
Upon doing that, the example works in DMD 1.0 and fails in 1.001--1.003 (I went ahead and tried them all).
Here's the modified test:
-----------------
import std.stdio;
static import std.gc;
class SomeObject
{
 this() { writefln("SomeObject created"); }
 ~this() { writefln("SomeObject destroyed"); }
}
template Mix()
{
 void init() {
 ptr = new SomeObject;
 }
 SomeObject ptr;
}
class Container
{
 this() { init(); }
 mixin Mix;
 int v = 10;
}
void main()
{
 auto x = new Container;
 writefln("---Pre collect");
 std.gc.fullCollect();
 writefln("---Post collect");
 writefln(x.v);
 writefln(x.ptr);
}

Gone in 1.004.