From mboxrd@z Thu Jan  1 00:00:00 1970
From: Dave Eckhardt <davide+p9@cs.cmu.edu>
Subject: Re: [9fans] First-timer help
To: 9fans@cse.psu.edu
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <1227.1121877529.1@piper.nectar.cs.cmu.edu>
Date: Wed, 20 Jul 2005 12:38:49 -0400
Message-ID: <1228.1121877529@piper.nectar.cs.cmu.edu>
Topicbox-Message-UUID: 6cb50708-ead0-11e9-9d60-3106f5b1d025

> it might be `ironic' if `autoconf' had ever met
> its notional specification.  `autoconf' is not
> even an `oxymoron', just a `moron'.

The old (pre-autoconf) idea was that each package would
try to encapsulate platform dependencies into a couple
of files, maybe SysV.c, BSD.c, etc.  When a new release
of one of those came out, somebody on that platform would
have to hack that file.  Often the code for SysV.c and
BSD.c would have lots of similar chunks in it.  If you
had N platforms some chunks might appear in N/3 of the
platform files.  More likely, *almost*-identical chunks
would be gratuitously different just because they were
maintained by different people.  And while the package
would build on M releases of N platforms, it would build
on *only* them--anything new was guaranteed to break
something.

The vision of autoconf was to focus on the chunks, not
the platforms, with probe code (in "portable /bin/sh")
choosing which chunks to use on each release of each
platform.  This way when a package finds itself on a new
platform it can automatically choose a locking method
from column A, a networking API from column B, etc., thus
*theoretically* working with zero porting effort by
anybody.

But this approach has a fatal flaw.  It turns out that
the *probe* code is non-portable.  Just one example: I
saw a piece of code designed to check whether to link
against Heimdal or MIT Kerberos libraries.  It "worked"
by testing for the existence of a particular function,
which at one moment in history was present in MIT Kerberos
but not Heimdal.  Six months later, of course, the Heimdal
guys caught up and the probe code broke.

The result of auto* is that a package will build on
M releases of N platforms, but will build on *only*
them--anything new is guaranteed to break something.
In other words, pretty much the same as before.

The bad news is that a large complex infrastructure
has been deployed against a problem and the problem
is still there.  The *awful* news is that now when
something goes wrong it isn't a matter of fixing a
snippet of C code inside BSD.c, but of finding,
decoding, and increasing the brittleness of some
probe code written in a punitively complex mixture
of sh and m4.

For example, maybe
  AC_MUMBLE(foo,[bletch blobble])
works on Linux but on Solaris it has to be
  AC_MUMBLE(foo,[[bletch blobble]])
Good luck figuring that out when the error you get is
  sh: configure.sh: 7225: syntax error
(yes, that's a seven thousand line shell script, and of
course it's machine-generated code for maximal readability).
Maybe the problem was gnu m4 versus m4, maybe it was bash
versus sh, more likely you'll never know.

Dave Eckhardt