Re: Preview: portable dumper

[Daniel Colascione]

On 12/02/2016 06:45 AM, Eli Zaretskii wrote:
> > From: Stefan Monnier <address@hidden>
> > Cc: address@hidden
> > Date: Fri, 02 Dec 2016 08:04:46 -0500
> >
> > > Granted, the proposed dumper is not very complicated.  But it isn't
> > > trivial either.  So if we can achieve a similar effect by using the
> > > "normal" loadup code, which is much simpler and doesn't really require
> > > understanding anything new, I think it's more beneficial for the
> > > project's future.
> >
> > Ken worked on speeding up the lread.c code, and it got to be
> > significantly faster, but not fast enough.  AFAIK it's got to the point
> > where it's not clear exactly how to speed it up further.  Not that it
> > can't be done, but that it's not obvious how, so it's likely going to
> > require some serious rethinking and maybe restructuring/rewrite of
> > the code.
> >
> > Is it going to happen if we don't merge the pdumper?  I'm not so sure.
>
> I'm willing to give that a chance.  I don't see any reason to make the
> decision today.
>
> > The main impetus behind speeding up lread.c is to replace unexec.c, so
> > I agree with you that merging the pdumper might mean that speeding up
> > lread.c will simply never happen.  But I think there's also a very
> > serious risk that even without the pdumper, speeding up lread.c will
> > still never happen: I have no intention on working at speeding up
> > lread.c, AFAICT Ken also gave up on it, anyone else?
>
> Judging by Ken's response, he didn't give up yet.
>
> In any case, it should be clear to anyone that code which isn't
> written cannot be used.  So a reality check will get us straight.
>
> > Personally, I think that maybe we should move in the other direction:
> > keep lread.c for "source code" and generalize the pdumper code so it can
> > also be used for the .elc files.
>
> You mean, move the byte compiler to C?

Ideally, yes: one way, we'll compile with libgccjit or something. But 
that's a long way away.

For now, nobody's talking about the byte compiler. Stefan's talking 
about the code that *reads* the byte compiler's output. Right now, this 
code is the normal lisp reader, but it doesn't have to be.

> If not, I don't understand the idea: pdump wants a file in a very
> specific format, because it doesn't really understand the entities it
> reads.

pdumper's loader is stupidly simple. It doesn't understand anything 
about what it's loading. That's what makes it fast. Stefan's idea is to 
make the elc format closer to the pdump format. Advantage: it'll be much 
faster that way. Disadvantage: elc files will be closely coupled to the 
precise Emacs builds that generated them.