Re: Yet another bootstrap failure: Required feature `esh-groups' was not provided

[Stephen J. Turnbull]

Alan Mackenzie writes:

 alan> It can't be that difficult; lisp is designed for this type of
 alan> manipulation.

It's not easy.  Several people at XEmacs have tried over a period of
years.  What hasn't been tried yet is adapting the byte-compiler to
the task.  It's been suggested, but people who hack byte generally
don't have a problem with debugging the makefile problem.

 alan> The dependencies absolutely must be generated automatically.
 alan> Whether this should be done by the byte-compiler or a separate
 alan> script isn't clear (yet).  Such a separate script could
 alan> probably be run in temacs, creating the dependencies very early
 alan> in the build process.

I don't think that makes a lot of sense.  The problem is that to
generate the Lisp dependencies you have to do this parsing of *all*
the Lisp files.  If you're going to do that, why not just do "find
. -name '*.elc' -exec rm '{}'" and recompile?  (Especially on Windows
just reading that many files costs a lot of time, although the
byte-optimizer is quite time-consuming, too.)

Also, there's no real reason why (non-dependency-breaking) users and
3rd party developers should have to run this.  Rather, have the
committers run it as part of their precommit testing (although they
mostly won't, at least you have an appropriate person to take out your
frustration on).

 alan> One kind of "peer review" we could do is doing a build test as
 alan> part of the commission process.  This might be a bit heavy on
 alan> server CPU time.

That's not really helpful, it's only one machine.  Eg, I misdoubt that
Glenn sees very many broken builds, that's why he committed those
changes.  The useful aspect of "one-time" build testing is to insist
that committers do a "make" just to ensure that anything they touched
has no syntax errors in it, that's about the maximum effect you can
ask for.

"Build-bot" has been mentioned; that would be a much more useful
infrastructure improvement.  I suspect it would cost about as much
hacker effort to set up a build-bot master as to hack the commit-hook
to run a build, it needn't be redone when (medatashi, medetashi!) CVS
bites the dust, and it won't involve repo downtime at all.