Tue, 26 Sep 2000 14:41:55 -0700 (PDT)
> I was wondering if you had put in the hack to keep the limit a bit behind so
> that you don't have to subtract. Note, the old verion of MLton (20000906) that
> we used for the ICFP contest did NOT run out of space on the same program with
> the same input, and as I recall, the GC wasn't using much of the time. Thus
> perhaps there is another bug with the liveness analysis which is keeping too
> much alive.
Maybe, but I bet the problem is just due to the major changes in the
implementation of TextIO. I.E., no compiler bug, just (very) poor performance.
I'll continue to look into the performance.
> I still think that the notion of checking to make sure that the frontier didn't
> move more than the checked amount is cheap and easy to add (only in -g mode).
I'll think about it. In the meantime, send me the mods for
GC_EVERY_PROGRAM_POINT (or whatever you called it).