create a big test, collect the fallout

[Ralf Wildenhues]

One of the cheapest (ahem) ways to catch a number of gnulib bugs is to
create one tree with all modules and all tests in it.  If it works, that
is.  So I started doing this manually, again.  Here's some fallout on
the way to make
    path/to/gnulib-tool --with-tests $l --dir=foo --megatest

work.  Apologies in advance for this being a huge blob.


0) I used Autoconf-2.60 and CVS Automake.  YMMV.  (I'll denote spots
   where things may need to be handled differently with older versions.)

1) $destdir may not contain slashes: gnulib-tool does `mkdir "$destdir"'
   to create it, and `cd "$destdir"; $cmds; cd ..' to enter/exit.

2) Outside the gnulib directory, the addition of `-tests' modules fails,
   thus limiting actual test exposure somewhat.  :-)

3) Not all instances of `build-aux' in ALL/tests/configure.ac are
   rewritten to `../build-aux'.

The following patch fixes (1)-(3).  OK to apply?
More below...

        * gnulib-tool (Command-line option processing): Ensure that
        $destdir contains no slash.
        (func_get_tests_module): Make work outside the gnulib source
        directory.
        (func_create_testdir): Rewrite all occurrences of `build-aux'.

Index: gnulib-tool
===================================================================
RCS file: /cvsroot/gnulib/gnulib/gnulib-tool,v
retrieving revision 1.113
diff -u -r1.113 gnulib-tool
--- gnulib-tool 23 Jun 2006 19:27:17 -0000      1.113
+++ gnulib-tool 6 Jul 2006 20:39:45 -0000
@@ -431,6 +431,10 @@
     do_changelog=false
   fi
 
+  case $destdir in
+  */*) func_fatal_error "argument to --dir may not contain slashes" ;;
+  esac
+
   # Remove trailing slashes from the directory names. This is necessary for
   # m4base (to avoid an error in func_import) and optional for the others.
   sed_trimtrailingslashes='s,\([^/]\)//*$,\1,'
@@ -595,7 +599,7 @@
 func_get_tests_module ()
 {
   # The naming convention for tests modules is hardwired: ${module}-tests.
-  if test -f modules/"$1"-tests; then
+  if (cd "$gnulib_dir" && test -f modules/"$1"-tests); then
     echo "$1"-tests
   fi
 }
@@ -1611,17 +1616,13 @@
      else
        echo "AM_CONDITIONAL([GL_COND_LIBTOOL], [true])"
      fi
-     if test "$auxdir" != "build-aux"; then
-       sed_replace_build_aux='
+     sed_replace_build_aux='
          :a
          /AC_CONFIG_FILES(.*:build-aux\/.*)/{
            
s|AC_CONFIG_FILES(\(.*\):build-aux/\(.*\))|AC_CONFIG_FILES(\1:../'"$auxdir"'/\2)|
            ba
          }'
-       sed_replace_build_aux=`echo "$sed_replace_build_aux" | sed -e 1d -e 
's/^ *//'`
-     else
-       sed_replace_build_aux=
-     fi
+     sed_replace_build_aux=`echo "$sed_replace_build_aux" | sed -e 1d -e 's/^ 
*//'`
      # We don't have explicit ordering constraints between the various
      # autoconf snippets. It's cleanest to put those of the library before
      # those of the tests.



4) Documentation should not be copied into the source tree (bug in
   getdate module; I'm inferring from other modules here) or the
   gnulib-tool should 'mkdir doc' to avoid a cp failure.  OK?

        * modules/getdate: Do not list doc/getdate.texi.

Index: modules/getdate
===================================================================
RCS file: /cvsroot/gnulib/gnulib/modules/getdate,v
retrieving revision 1.13
diff -u -r1.13 getdate
--- modules/getdate     12 May 2005 15:22:34 -0000      1.13
+++ modules/getdate     6 Jul 2006 20:18:43 -0000
@@ -2,7 +2,6 @@
 Convert a date/time string to linear time.
 
 Files:
-doc/getdate.texi
 lib/getdate.h
 lib/getdate.y
 m4/bison.m4



5) Autoconf's getloadavg (still) needs getloadavg.c, which it only finds
   if AC_CONFIG_LIBOBJ_DIR is set correctly.  This patch is Autoconf 2.60
   and CVS Automake only.  I'm showing it for completeness.
   (Dunno, maybe we should add
      m4_ifndef([AC_CONFIG_LIBOBJ_DIR],
                [AC_DEFUN([AC_CONFIG_LIBOBJ_DIR])])
    just so the configure.ac works with older Autoconf for all other
    modules at least?)

Index: gnulib-tool
===================================================================
RCS file: /cvsroot/gnulib/gnulib/gnulib-tool,v
retrieving revision 1.113
diff -u -r1.113 gnulib-tool
--- gnulib-tool 23 Jun 2006 19:27:17 -0000      1.113
+++ gnulib-tool 6 Jul 2006 20:39:45 -0000
@@ -1585,6 +1589,7 @@
     (echo "# Process this file with autoconf to produce a configure script."
      echo "AC_INIT([dummy], [0])"
      echo "AC_CONFIG_AUX_DIR([../$auxdir])"
+     echo "AC_CONFIG_LIBOBJ_DIR([../lib])"
      echo "AM_INIT_AUTOMAKE"
      echo
      echo "AM_CONFIG_HEADER([config.h])"
@@ -1666,6 +1667,7 @@
    if test "$auxdir" != "."; then
      echo "AC_CONFIG_AUX_DIR([$auxdir])"
    fi
+   echo "AC_CONFIG_LIBOBJ_DIR([lib])"
    echo "AM_INIT_AUTOMAKE"
    echo
    echo "AM_CONFIG_HEADER([config.h])"


6) I'd like a megatest that does not test each individual module but
   just all of them (I may be patient, but not _that_ patient ;-).
   Also I'd like failed checks to make the thingy fail, and configure to
   use a cache file whenever possible (due to the fact that the tests/
   directory uses a separate configure script, this is a large speedup).
   The patch needs more work (maybe a new command line argument?),
   but shows where this could go: it would make --megatest feasible
   enough to be used more regularly.  (Note that
     gnulib-tool --with-tests --test `gnulib-tool --list`

   is not equivalent, as it does not take into account that some modules
   may not be used together.)

Index: gnulib-tool
===================================================================
RCS file: /cvsroot/gnulib/gnulib/gnulib-tool,v
retrieving revision 1.113
diff -u -r1.113 gnulib-tool
--- gnulib-tool 23 Jun 2006 19:27:17 -0000      1.113
+++ gnulib-tool 6 Jul 2006 20:39:45 -0000
@@ -1756,10 +1758,10 @@
 
   megasubdirs=
   # First, all modules one by one.
-  for onemodule in $allmodules; do
-    func_create_testdir "$megatestdir/$onemodule" $onemodule
-    megasubdirs="${megasubdirs}$onemodule "
-  done
+  #for onemodule in $allmodules; do
+  #  func_create_testdir "$megatestdir/$onemodule" $onemodule
+  #  megasubdirs="${megasubdirs}$onemodule "
+  #done
   # Then, all modules all together.
   # Except fnmatch-posix, which conflicts with fnmatch-gnu. FIXME.
   allmodules=`for m in $allmodules; do if test $m != fnmatch-posix; then echo 
$m; fi; done`
@@ -1955,9 +1957,9 @@
     cd "$destdir"
       mkdir build
       cd build
-        ../configure
+        ../configure -C
         make
-        make check
+        make check || exit 1
         make distclean
         remaining=`find . -type f -print`
         if test -n "$remaining"; then
@@ -1978,9 +1980,9 @@
     cd "$destdir"
       mkdir build
       cd build
-        ../configure
+        ../configure -C
         make
-        make check
+        make check || exit 1
         make distclean
         remaining=`find . -type f -print`
         if test -n "$remaining"; then


7) Annoying warning from automake in tests/ due to extra directory
   (I don't know what to do here):
| Makefile.am:16: AM_GNU_GETTEXT used but `po' not in SUBDIRS


9) Autoconf-2.60's AC_CHECK_DECLS_ONCE has two semantic differences
   with the one from onceonly_2_57.m4:
   - it accepts only one argument (why BTW?  This looks like a bug;
     I think it's been discussed here before, but don't remember the
     details):
| autoreconf: running: aclocal -I m4
| configure.ac:252: warning: _AC_Check_Decl_strtoimax strtoumax is m4_require'd 
but not m4_defun'd
| autoconf/general.m4:2586: AC_CHECK_DECLS_ONCE is expanded from...
| m4/_inttypes_h.m4:16: gl_INTTYPES_H is expanded from...
| configure.ac:252: the top level

   - it expands to nothing, i.e., it does not produce a valid shell
     list:
| checking for struct stat.st_blocks... (cached) yes
| ../../ALL/configure: line 20067: syntax error near unexpected token `fi'
| ../../ALL/configure: line 20067: `  fi'

   This patch should fix the two issues.  Note that since
   gl_PREREQ_GETPASS expands to nothing, there is little point in
   putting it in a shell conditional.  (Maybe there is incentive to
   instead not use AC_CHECK_DECLS_ONCE?)

   I've seen more intermediate failures, but not been able to reproduce
   them consistently.

        * m4/_inttypes_h.m4 (gl_INTTYPES_H): Use AC_CHECK_DECLS_ONCE
        with only one argument.
        * m4/fileblocks.m4 (gl_FILEBLOCKS): AC_CHECK_DECLS_ONCE may
        expand to nothing, so add a shell command to avoid syntax error.
        * m4/getpass.m4 (gl_FUNC_GETPASS): Likewise.

Index: m4/_inttypes_h.m4
===================================================================
RCS file: /cvsroot/gnulib/gnulib/m4/_inttypes_h.m4,v
retrieving revision 1.3
diff -u -r1.3 _inttypes_h.m4
--- m4/_inttypes_h.m4   4 Jul 2006 06:37:10 -0000       1.3
+++ m4/_inttypes_h.m4   6 Jul 2006 20:18:43 -0000
@@ -12,5 +12,6 @@
 if test $gl_cv_have_include_next = no; then
   gl_ABSOLUTE_HEADER([inttypes.h])
 fi
-AC_CHECK_DECLS_ONCE([strtoimax strtoumax])dnl
+AC_CHECK_DECLS_ONCE([strtoimax])dnl
+AC_CHECK_DECLS_ONCE([strtoumax])dnl
 ])
Index: m4/fileblocks.m4
===================================================================
RCS file: /cvsroot/gnulib/gnulib/m4/fileblocks.m4,v
retrieving revision 1.3
diff -u -r1.3 fileblocks.m4
--- m4/fileblocks.m4    23 Sep 2005 04:15:13 -0000      1.3
+++ m4/fileblocks.m4    6 Jul 2006 20:18:43 -0000
@@ -9,6 +9,7 @@
   AC_STRUCT_ST_BLOCKS
   dnl Note: AC_STRUCT_ST_BLOCKS does AC_LIBOBJ(fileblocks).
   if test $ac_cv_member_struct_stat_st_blocks = no; then
+    :
     gl_PREREQ_FILEBLOCKS
   fi
 ])
Index: m4/getpass.m4
===================================================================
RCS file: /cvsroot/gnulib/gnulib/m4/getpass.m4,v
retrieving revision 1.12
diff -u -r1.12 getpass.m4
--- m4/getpass.m4       20 Apr 2006 22:28:34 -0000      1.12
+++ m4/getpass.m4       6 Jul 2006 20:26:00 -0000
@@ -12,6 +12,7 @@
   AC_REPLACE_FUNCS(getpass)
   AC_CHECK_DECLS_ONCE(getpass)
   if test $ac_cv_func_getpass = no; then
+    :
     gl_PREREQ_GETPASS
   fi
 ])


9) With all these, I still get two test failures:
| - argp.25134 differ: char 557, line 20
| FAIL: test-argp-2.sh
| cu got b9a9a617
| FAIL: test-crc

10) making the command work for l=--libtool ... in another mail.

Cheers,
Ralf