Re: [Sbcl-devel] [Sbcl-commits] master: Derive that (INTERN x "KEYWORD") produces a KEYWORD

[Nikodemus S. <nik...@ra...>]

On 24 June 2014 07:56, Stas Boukarev <sta...@gm...> wrote:

> I don't think that it's totally different, before,
> (let ((x (intern "x" :keyword))) (unintern x :keyword) (keywordp x))
> returned NIL, now it returns T.

I'm trying to decide if I think this is somehow different from

  (let ((x :foo)) (unintern x) (keywordp x)) => T

which has been the case for a long time... Still, I suspect KEYWORD is
a dangerous type to use in inference because like dimensions for
non-simple arrays, or CONS-types it can change over the lifetime of
the object.

ir1opt.lisp has working but hacky code that allows using some CONS and
array types safely under the nomer of LVAR-CONSERVATIVE-TYPE.

...but instead of adding KEYWORD there as well (which would definitely
be doable), I would suggest re-architecting it a bit:

1. Call these "handle with care" types LVAR-VOLATILE-TYPE.

2. Make sure such types don't end up in LVAR-TYPE unless explicitly DECLAREd.

3. Use LVAR-VOLATILE-TYPE only (a) if it all LVAR-USES are "known
harmless", or (b) if it can be proven that none of the "possibly
harmful" LVAR-USES is a predecessor of the current node.

4. Add clobbers-arg-types annotations to fndb for functions that do.

...A is should be easy to implement and covers many more cases than
the current code. B is left as an exercise to reader. :)

Cheers,

 -- nikodemus