Re: [Caml-list] Specify the default hash function for a type

[Alain Frisch <alain.frisch@lexifi.com>]

FWIW, I was not seriously suggesting this.  That said, ..., the case 
should be relatively safe, since the code generated by OCaml never reads 
the tag of record values (it is assumed to be 0).  The tag should only 
be read by polymorphic primitives, which don't receive any information 
from the OCaml optimizer.  So I don't see how the optimizer could break 
the (not-really-)suggested hack, except on purpose.

Alain



On 13/05/2016 15:50, Pierre Chambart wrote:
> For those who don't see that as a joke, please forget everything about that.
> Obj.set_tag is an extremely dangerous function as the compiler assumes that
> value's tag never change, even for completely unknown values. Hence, it
> might
> be legal for the compiler to do something that you might not expect on:
>
> let f x =
>   let x' = Obj.repr x in
>   let tag = Obj.tag x' in
>   Obj.set_tag x' 0
>   let ... = match x with ... in
>   Obj.set_tag x' tag;
>   match x with ...
>
> The access to the tag can be shared for instance. And lots of
> 'surprising' stuff might happen.
> This is of course a quite brutal example but this could apply to a lot
> of less obvious situation !
>
> So Obj.magic and Obj.set_field : bad
> Obj.set_tag: really really bad !
>