Question to GC experts: NO_SCAN for a part of the block

[ag0aep6g via Digitalmars-d]

On 06/04/2017 03:08 AM, Stanislav Blinov wrote:
> ---
> 
> import core.memory : GC;
> import std.stdio;
> import std.typecons;
> 
> class C {
>      int i;
>      this(int i) {
>          this.i = i;
>      }
> 
>      ~this() {
>          writeln("C(", i, ") dtor");
>      }
> }
[...]
> auto selective(int numCs, int numBytes) {
[...]
>      auto memory = GC.malloc(numCs*C.sizeof + numBytes*byte.sizeof,
>              GC.BlkAttr.NO_SCAN);
>      auto cs = (cast(C*)memory)[0..numCs];
>      cs[] = C.init;
>      // add scanning range for references
>      GC.addRange(cs.ptr, cs.length*C.sizeof, typeid(C));
>      auto bytes = (cast(byte*)(memory + numCs*C.sizeof))[0..numBytes];
>      bytes[] = byte.init;
>      return tuple!("Cs", "bytes")(cs, bytes);
> }
> 
> void main() {
> 
>      int numCs = 4; // comes at runtime from elsewhere
>      int numBytes = 32; // comes at runtime from elsewhere
[...]
>      int counter;
[...]
>      auto arrays3 = selective(numCs, numBytes);
>      foreach (ref e; arrays3.Cs)
>          e = new C(counter++); // dtors will not be called
> }
> 
> ---
> 
> Should this work, and if not, why?

As far as I can tell, the `addRange` call works correctly, but maybe too 
well in a sense. It keeps the `new`ed `C`s alive as long as `arrays3.Cs` 
has pointers to them. And `arrays3.Cs` has those pointers until the very 
end.

If you add `GC.removeRange(arrays3.Cs.ptr);` at the of `main`, the dtors 
show up. Overwriting `arrays3.Cs`'s elements with `null`s also works. My 
guess is that the `removeRange` call isn't done automatically before the 
final run of the GC. Maybe it should be?

But I have a vague memory that the GC isn't required to call destructors 
on everything in the final run. Or maybe it's not guaranteed that there 
is a final run when the program ends? Anyway, that would mean 
everything's working as intended, and you just can't rely on destructors 
like that.