make dist

[Zack Weinberg <zack@rabi.phys.columbia.edu>]

I'm looking at a rewrite of make dist.  The goal is to reduce the amount of
magic, which should make Roland happy :)  I have in mind four make targets:

make dist - generate all generated files that go in the distribution.
This target is run for each subdirectory before making the tarfile.

make snap-dist - same as make dist, but for a snapshot release; this leaves
out info and compiled message catalogs, etc.

make release - call make dist, then wrap up a tarfile named suitably for a
release (libc-2.x.y.tar.gz).  If CVS is available, generate a patch also.

make snapshot - call make snap-dist, then wrap up a snapshot tarfile named
libc-YYMMDD.tar.gz (maybe YYYYMMDD?)  If CVS, make a patch.

Currently, what to put into the tarfiles is a magic function of Makefile
variables.  There's ad-hockery in some of the subdir Makefiles to get things
distributed under odd configuration settings.

What I'd like to do is have a file MANIFEST at top level that lists every
file that goes into the distribution.  This file is generated essentially by
running 'find . -type f -print | clean-up-list' in the source tree. 
clean-up-list deletes initial `./' and weeds out files that *don't* go into
the distribution.  The issue is then how to know that, and the obvious way
is with files similar to .cvsignore -- I'd call it .nodist -- that list glob
patterns that don't belong.

The $64,000 question: Can people live with just one .nodist file at top
level, or should I support multiple such, each applying to the subtree it's
at the root of?

zw