Re: automatic construction of mli files

[Damien Doligez <Damien.Doligez@inria.fr>]

>From: Francois Rouaix <frouaix@liquidmarket.com>

>This has to be one of the most cryptic comment ever made to this list.
>And a "rather complex issue" coming from Damien, the mind boggles, especially 
>on this mysterious 8% figure.

>Care to give some details ?


OK.  It has to do with examining the roots at the beginning of each
minor collection.  The global variables are roots, but each global
variable is assigned only once, so we only need to examine it once
(after that, the value will be in the major heap, so it is not a root
for the minor collector).

We do it by remembering which modules have executed some
initialisation code since the last collection, and only examining
their globals.

The 8% figure comes from the speedup on Coq that we got when we
implemented the trick.


Actually, now that you force me to remember the complex part, I want
to take back my comment.  It's the fact that only the exported symbols
are roots, so using .mli files will speed up the first garbage
collections, and only by a small amount.  Using an empty .mli
file for your main module (the one that's linked last) will speed up
all garbage collections (because the initialisation of that module is
only complete when the program stops running), again by a very small
amount.

I have to apologize for not checking my facts before I posted to the
list.


Oh, and this applies only to programs compiled with ocamlopt.

-- Damien