Another option would be to just overload equal? in match.scm to transform into
eqv? when there are char literals or numbers, eq? on symbols, booleans, the
empty list and keywords and (@@ (guile) equal?) for everything else.
Considering that it in this case contributed a 25% overhead to code that was
performance critical I think it would be a pretty valid thing to do. If you,
ludo, an Andy thinks that would be a good idea I can make such a patch for
match.scm. That would have the benefit of not changing the upstream code (which
is (include ...)d in match.scm), nor fiddling around with guile optimisations.
--
Linus Björnstam
On Mon, 4 May 2020, at 22:50, Linus Björnstam wrote:
> You didn't see my other reply. The matching code isn't suboptimal. The
> equality predicate is The problem is that match compares using equal?
> even for literal chars (where eqv? is a lot faster). It would be a
> rather trivial optimization to do, either to match.scm (meaning:
> breaking with upstream and use syntax-case) or to the guile compiler in
> general (changing equal? to eqv, when there are character literals),
> which seems ok-ish for this use-case but at very little benefit in
> general.
>
> A long-term goal of mine is to write a pattern matcher with the
> optimisations that the racket matcher does (among other things: some
> serious list matching reordering!). That is a daunting task though.
>
> --
> Linus Björnstam
>
> On Mon, 4 May 2020, at 22:09, Ludovic Courtès wrote:
> > Hi,
> >
> > Linus Björnstam skribis:
> >
> > > On Mon, 4 May 2020, at 11:36, Ludovic Courtès wrote:
> > >
> > >> > One thing I found is that `match` is slow. The code looked nicer but
> > >> > had to change it back to lets and conds as the performance
> > >> > increase was ~2 seconds.
> > >>
> > >> Oh, in which case exactly? And are you sure your hand-written code is
> > >> equivalent to the ‘match’ code (it’s common for hand-written code to be
> > >> more lax than ‘match’)?
> > >>
> > >> One thing to pay attention to is the use of ‘list?’, which is O(N), and
> > >> is implied by ellipses in ‘match’. If you want to use ‘match’ in a way
> > >> that avoids ‘list?’, write patterns such as (a . b) instead of (a b ...).
> > >> It doesn’t have the same meaning, but often the end result is the same,
> > >> for instance because you’ll later match on ‘b’ anyway.
> > >>
> > >> (I wish we can one day have a proper list type disjoint from pairs…)
> > >
> > > The change is here: he is only matching against chars and predicates:
> > > https://github.com/aconchillo/guile-json/commit/ad4b06d86e4822466983d00f55474c8f664b538d
> >
> > It would be nice if you could pinpoint which one of these changes causes
> > a difference, because:
> >
> > --8<---cut here---start->8---
> > scheme@(guile-user)> ,optimize (match (peek-char port) ((? eof-object?)
> > x) ((? whitespace?) w) (_ e))
> > $84 = (let ((v (peek-char port)))
> > (cond ((eof-object? v) x)
> > ((whitespace? v) w)
> > (else e)))
> > --8<---cut here---end--->8---
> >
> > What might make a difference is the code bloat when using ‘or’:
> >
> > --8<---cut here---start->8---
> > scheme@(guile-user)> ,optimize (match (peek-char port) ((or #\a #\b #\c
> > #\d) x))
> > $86 = (let ((v (peek-char port)))
> > (cond ((equal? v #\a) x)
> > ((equal? v #\b) x)
> > ((equal? v #\c) x)
> > ((equal? v #\d) x)
> > (else
> > ((@@ (ice-9 match) error)
> > 'match
> > "no matching pattern"
> > v)
> > #f)))
> > --8<---cut here---end--->8---
> >
> > but even that sounds unlikely.
> >
> > You’re compiling with -O2, right?
> >
> > Thanks,
> > Ludo’.
> >
>
>
