Re: [Sbcl-devel] Two modular arithmetic issues

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

Hi,

"Nathan Froyd" <fr...@gm...> writes:

> On 8/14/07, Nikodemus Siivola <nik...@ra...> wrote:
>> 2. 32 bit modular arithmetic not working on x86-64. Is this to be expected, is
>> this a regression, or am I doing something wrong? How hard would it be to
>> make it work?
>
> This is expected and somewhat unfortunate.  I posted a patch to make
> this work several months ago:
>
> http://thread.gmane.org/gmane.lisp.steel-bank.devel/9155/
>
> Bill and Christophe expressed reservations about the patch, so I did
> not commit it.

I've resurrected my work towards making modular arithmetic choose
which width to use more cleverly than before, from patches I sent to
sbcl-devel in September.  In particular, I have got

* better support for signed modular arithmetic: all LOGFOO functions
  are treated as "good";
* preference for tagged modular arithmetic over untagged if the width
  requested can be satisfied by both, unless the width matches
  exactly.

Test case for before/after comparison:
  (defun foom (x y)
    (declare (fixnum x y))
    (logand (logxor (+ x y) x) most-positive-fixnum))

This should allow two further pieces of work: 

* support for signed word-sized (i.e. untagged) arithmetic;
* support for sub-word-sized untagged arithmetic, as in Nathan's
  original work.

You can see my work-in-progress (and my inability to drive git
usefully) at
<http://www-jcsu.jesus.cam.ac.uk/~csr21/cgi-bin/gitweb.cgi>; I'm
attaching a patch for those who, like me, are not quite comfortable
with this particular VCS yet.

Nathan (or Nikodemus, or anyone else for that matter): are you in a
position to retry the x86-64 32-bit modular arithmetic on top of this?

Best,

Christophe