Re: [Sbcl-devel] food for constraint propagation, maybe

[Christophe R. <cs...@ca...>]

Nikodemus Siivola <nik...@ra...> writes:

> Though constraint propagation is already an unholy timesink, some of
> these look potentially interesting: eg. propagating BOUNDP would make
> for nicer special variable access code, and ARRAY-HEADER-P is also
> almost certainly worth looking at. EQUAL seems like a good candidate
> too...

> COMMON-LISP:ARRAY-HAS-FILL-POINTER-P
> COMMON-LISP:ARRAY-IN-BOUNDS-P
> SB!KERNEL:%ARRAY-DISPLACED-P

Based on extensive user and developer feedback (read: a conversation
with Juho and Paul on IRC), it seems that there is demand for a little
bit of type system and compiler hackery, to teach the system about
particular kinds of non-simple arrays.  The comment that motivates this
is "adjustable arrays are total crap in sbcl in general;
vector-push-extend is particularly crap".

I think I can do the type system modifications; a preliminary, which is
to teach the type system that (and <array-type> (not simple-array))
means that the array is definitely complex, seems to be working, and
incidentally fixes #309129.  The full modifications will have to wait
until after sbcl-1.0.33; this is just a heads-up to say that it might
soon become possible to exploit specific bits of array properties rather
more than at present.

Cheers,

Christophe