Re: master: x86-64: TRUNCATE using the Lemire, Kaser, Kurz transform

Maybe that transform can just do all that.
But you can override transforms with a... vop-transform!
e.g.

(deftransform %ldb ((size posn integer) (:or (((constant-arg (integer
1 #.(1- n-word-bits)))
                                               (constant-arg integer)
                                               word) unsigned-byte)
                                             (((constant-arg (integer
1 #.(1- n-word-bits)))
                                               (constant-arg integer)
                                               signed-word) unsigned-byte)) *
                    :vop t)
  (and (plusp (lvar-value posn))
       (or (= (lvar-value size) 1)
           (<= (+ (lvar-value size)
                  (lvar-value posn))
               n-word-bits))))

On Sun, Apr 19, 2026 at 5:52 AM Douglas Katzman <[email protected]> wrote:
>
>
>
> On Sat, Apr 18, 2026 at 10:40 PM Stas Boukarev <[email protected]> wrote:
>>
>> The VOP gets defeated by one of the 23 transforms, when (speed 3):
>> https://github.com/sbcl/sbcl/blob/master/src/compiler/srctran.lisp#L5135
>>
> Good catch.
> Can we make that transform give-up if it knows that the quotient isn't used, and that the backend supports fastrem and that the argument types would be acceptable ?  Or should we try to just add a 24th transform that (miraculously) is squeezed in between two other transforms?


_______________________________________________
Sbcl-commits mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/sbcl-commits