Re: [Caml-list] Possible ephemeron bug?

[Lindsay Errington <lindsay.errington@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1853 bytes --]

Great answer! Thanks to you both.

I have a followup ephemeron question which I'll post with a different
subject.

On Thu, Dec 20, 2018 at 1:46 AM Stephen Dolan <stephen.dolan@cl.cam.ac.uk>
wrote:

> On Thu, 20 Dec 2018 at 00:30, Lindsay Errington <
> lindsay.errington@gmail.com> wrote:
>
>> If however, I use the standalone bytecode compiler or the native compiler
>> (4.07.1), then the entries are not nullified. Is this a bug or is there
>> another way to persuade the garbage collector to clobber the entries?
>>
>
> Thanks for the detailed example!
>
> The issue is with the test code at the end:
>
>     let (root,map) =
>       let map = [] in
>       let (k0,map) = intern 0 (str 1) map in
>       let (k1,map) = intern 1 (Link k0) map in
>       let (k2,map) = intern 2 (Link k1) map in
>       (k2,map);;
>
>     Fmt.printf "root=%a, map=@[<v>%a@]@." pp_key root pp_map map;;
>
>     let map = upd root (str 2) map;;
>
>     Fmt.printf "root=%a map1=@[<v>%a@]@." pp_key root pp_map map;;
>
> Both the bytecode and native code compilers take any value bound to a
> global to be reachable forever, even if that global is later shadowed by
> another global of the same name. The ephemerons don't get cleared, because
> the original map is still alive.
>
> You can change the code to avoid putting the original map in a global
> constant:
>
>     let (root, map) =
>       let map = [] in
>       let (k0,map) = intern 0 (str 1) map in
>       let (k1,map) = intern 1 (Link k0) map in
>       let (k2,map) = intern 2 (Link k1) map in
>       Fmt.printf "root=%a, map=@[<v>%a@]@." pp_key k2 pp_map map;
>       let map = upd k2 (str 2) map in
>       (k2, map);;
>
>     Fmt.printf "root=%a map1=@[<v>%a@]@." pp_key root pp_map map;;
>
> With this version, the original map becomes unreachable, so the GC can
> clear the ephemerons.
>
> Stephen
>

[-- Attachment #2: Type: text/html, Size: 3837 bytes --]