Re: More on Lisp structures performance

On 25 Jan 2026, at 18:09, Martin Simmons <[email protected]> wrote:
> 
> You just make all objects use a multiple of 16 bytes of memory and waste a
> word for those that are actually 8 bytes smaller.

Exactly so.  And indeed for LW we do not need to theorise about how big objects are (neither for SBCL):

LW:

  ? (find-object-size 16.0d0)
  16

SBCL:

  * (primitive-object-size 1.0d0)
  16

A question to ask is: what are the single-word types that are interesting?  The two obvious ones are full-word integers and full-word floats.  Beyond that you're running on fumes: special tiny strings, perhaps, tiny arrays of very small integers?  I don't think you can easily do locatives this way (you likely want a pointer to the start of the object you're pointing into as well, so the GC knows not to collect it) and in any case they're not a standard feature of the language.

Well, the answer for the numeric types is that you really want the compiler to be able to deal with them directly: the performance difference between programs which cons floats and programs which don't is usually enormous.  If you only cons a tiny number of the things then you don't care about them being two words.

It's just much better, I think, to prevent these things being boxed at all than to worry about how big the boxes are.

--tim

_______________________________________________
Lisp Hug - the mailing list for LispWorks users
[email protected]
http://www.lispworks.com/support/lisp-hug.html