Edi Weitz <[EMAIL PROTECTED]> writes:

> The good thing: It's not Gerd's PCL that's responsible for the
> slowdown.

That's good news!

> The bad thing: Something that's relatively new in CMUCL[1] seems to be
> responsible for a significant slowdown in GC times.

That's bad news, because most of the stuff that's currently new in
HEAD is is from me, too ;).

> After fiddling around with various test cases I finally saw that the
> difference between the old "fast" code and the new "slow" code was
> always exactly the GC time. CL-PPCRE was just a good lackmus test for
> this because it has to cons a lot to create scanners.
> 
> I then switched to the 18e-pre1 binaries from 2003-03-23 and found the
> same results - GC is slower by a factor of about 10 to 15! I think
> that's quite a lot and I wonder if that's considered acceptable. 

Certainly not acceptable.  I just haven't noticed so far that there is
a bug, although I'd thought I'd noticed a 10x GC slowdown.  Strange.

I'll investigate this further.  Thanks for tracking it down.
Actually, there are only two candidates that could cause a GC
slowdown, I think: the control stack checking or the new code for
weak hash tables.  Let's see.  I'll come back to you when I find
something.

Reply via email to