Re: Speed regression from 3.2 to 3.3

[Matt Austern <austern at apple dot com>]

On Wednesday, January 22, 2003, at 11:22 AM, Joe Buck wrote:

> On Wed, Jan 22, 2003 at 10:03:41AM -0800, Matt Austern wrote:
> > On Wednesday, January 22, 2003, at 09:42 AM, Andrea 'fwyzard' Bocci
> > wrote:
> >
> > > At 15.30 21/01/2003 -0800, Matt Austern wrote:
> > > > What I see from the time report is pretty much what I said before: 
> > > > the
> > > > parser and the gc have gone up dramatically, nothing else matters.
> > > > I'm attaching a graph, for those who are interested.
> > > From your graph, it looks like GC times went back to a good shape in
> > > 3.4.
> > > Suggestions ?
> > That's sure what it looks like.  I'm sure that's an important clue, 
> > but
> > I don't
> > know what it means either.
> Just a guess:
> It's probably because of Mark's new C++ parser.  If it's faster, that's
> great news.
I don't think that can account for it, since the big change from 3.2 to
3.3, as shown by -ftime-report, was in the gc phase.   Time spent in
gc shot up from 3.2 to 3.2, and went down again in 3.4.  So if Mark's
new parser has an effect, it would be that it's much more efficient in
memory usage.  Is that plausible?  (Honest question, not rhetorical.
I haven't looked at his code yet; quite possible the answer is yes.)

Parsing time dominates, in 3.2 through 3.4.  As far as parsing time
goes, the main thing I'm seeing with -ftime-report about the new
parser is that there's no longer a separate lexical analysis phase,
so the time spent there goes to zero, but mostly it just means that
the time has been reshuffled: the sum of the lexical analysis and
parse phases is almost identical in 3.3 and 3.4.

Possibly there's just some straightforward gc bug in 3.3 that didn't
happen to get into 3.4.  I can hope, anyway.

			--Matt