one dependency too much

[Bruno Haible]

Hi,

While preparing a new libunistring release, I noticed that gnulib-tool
brought in many more files than for the last release, in particular these:

gnulib-m4/cloexec.m4
gnulib-m4/dos.m4
gnulib-m4/fcntl.m4
gnulib-m4/getdtablesize.m4
gnulib-m4/lstat.m4
gnulib-m4/pathmax.m4
gnulib-m4/stat.m4
gnulib-m4/symlink.m4
gnulib-m4/sys_stat_h.m4
gnulib-m4/time_h.m4
tests/binary-io.h
tests/cloexec.c
tests/cloexec.h
tests/fcntl.c
tests/getdtablesize.c
tests/lstat.c
tests/pathmax.h
tests/same-inode.h
tests/stat.c
tests/symlink.c
tests/sys_stat.in.h
tests/test-binary-io.c
tests/test-binary-io.sh
tests/test-cloexec.c
tests/test-fcntl-h.c
tests/test-getdtablesize.c
tests/test-lstat.c
tests/test-lstat.h
tests/test-open.h
tests/test-stat.c
tests/test-stat.h
tests/test-symlink.c
tests/test-symlink.h
tests/test-sys_stat.c
tests/test-time.c
tests/time.in.h

This is quite surprising, because libunistring does not make a single libc
call that deals with files or file descriptors. I don't really want
libunistring's "make check" to be sensitive to lstat, binary-io, and cloexec
issues.

It is due to a single dependency too much, that triggers an avalanche of
imports due to other dependencies. Namely,

  libunistring includes unistdio/ulc-fprintf, unistdio/ulc-vfprintf.
  |
  +-- unistdio/ulc-fprintf depends on fseterr.
      |
      +-- fseterr depends on dup2.
          |
          +-- dup2-tests depends on binary-io, cloexec, open.
              |
              +-- cloexec depends on fcntl.
              |   |
              |   +-- fcntl-tests depends on getdtablesize.
              |
              +-- open depends on stat
              |   |
              |   +-- stat-tests depends on symlink, same-inode
              |
              +-- open-tests depends on symlink
                  |
                  +-- symlink depends on lstat

For my feeling, there are too many dependencies here.

* fseterr does not need dup2, because
  1. the call to dup2 is in a piece of code that is disabled with #if 0,
  2. even if it was enabled, it would be immune to the dup2 portability
     pitfalls currently mentioned in doc/posix/functions/dup2.texi.

* The dependency of dup2-tests on 'cloexec' could be removed if the
  relevant (small) part of the test would be enclosed in
     #ifdef GNULIB_CLOEXEC
  so that it is skipped if the cloexec module is not present. coreutils
  does include the 'cloexec' module, therefore at least the coreutils users
  will still have a 100% check of dup2.

* The dependency of dup2-tests on 'open' can safely be removed because the
  file to be opened is "test-dup2.tmp", which does not trigger any of the
  problems listed in doc/posix-functions/open.texi.


I'm handling the first part, by removing the dependency of fseterr, since
that's the most obviously redundant one.


2009-12-26  Bruno Haible  <address@hidden>

        Remove an unneeded dependency.
        * modules/fseterr (Depends-on): Remove dup2.

*** modules/fseterr.orig        Sat Dec 26 19:57:41 2009
--- modules/fseterr     Sat Dec 26 19:12:26 2009
***************
*** 7,13 ****
  lib/stdio-impl.h
  
  Depends-on:
- dup2
  
  configure.ac:
  
--- 7,12 ----