> Implementing any of these options means the test will only work with the new gettext, it won't work with the old one anymore.

This is true for the po2test.awk suggestion. But the first of my 3 suggestions (use a modified de.po file, without POT-Creation-Date: header line) should work with both the old and new msgfmt, and does not require modifications of the C code.

> This change breaks the assumption that any msgstr entry from the po file is put unchanged into the mo file.

You are right, this assumption does not hold any more for the header entry.

> This breaks at least the glibc intl/tst-gettext test which checks that.

The glibc/intl/tst-gettext test needs to be updated.
- Either use as input instead of de.po a modified PO file, possibly "grep -v 'POT-Creation-Date:' de.po"
- Or modify the po2test.awk script so that it eliminates the POT-Creation-Date: header line.
- Or modify the positive_gettext_test / negative_gettext_test functions in tst-gettext.c to allow for the difference in POT-Creation-Date line.

Thanks a lot! I am adding your patch, with the necessary corrections (mostly comments, and message_list_delete_header_field could produce a non-NUL-terminated string in one case).

Daiki,

> Bruno, are you willing to provide a patch for this?

No, I'd leave that to you. I'm trying to focus on libiconv and libsigsegv in the moment.

> I think it could also be worked around by disabling the auto-msgmerge rule in po/Makefile.in.in

There are several ways in which PO files can be integrated into a release, compiled, and then installed on a user's system. For example:
- The maintainer installs PO files manually, "make" merges them with the POT file, "make install" installs the .mo file. (This is what Makefile.in.in currently does.)
- "make" fetches the PO file from the Translation project and msgmerges it. "make install" installs the .mo file.
- You may omit msgmerge and thereby put old, unused messages into the .mo file.
- Some of the gettext 'hello' examples use a Makefile.am instead of Makefile.in.in.

In summary, there is considerable variation in the way the PO files are transmitted and msgmerged. But the common thing is that they use 'xgettext' (or some similar tool), 'msgmerge' (needed to filter out old, unused messages) and 'msgfmt'. I would therefore not try to fix a general problem with the tools by modifying Makefile.in.in.

I agree with Bruno's analysis.

The flow of this timestamp is as follows:
(1) xgettext puts it in, when it creates the .pot file.
(2) msgmerge preserves it, when it merges with a translator's .po file.
(3) msgfmt preserves it, when it compiles a .po file to a .mo file.

The POT-Creation-Date is important information for a translator, so that she knows whether to review the entire file or not.

Therefore IMO it's not in the steps (1) or (2) that we should change something, but in step (3): msgfmt should eliminate the POT-Creation-Date from the binary output that it produces (regardless whether it is .mo or other).

This would be work in
  write-mo.c
  write-java.c
  write-csharp.c
  write-resources.c
  write-tcl.c
  write-qt.c
  write-desktop.c
  write-xml.c

Use of msgunfmt on a given .mo file would then not be able to restore the lost POT-Creation-Date of course; the test suite of msgunfmt will have to be adjusted.

Note: A first step towards reproducible builds was already taken in version 0.19.8 (see NEWS):
  - msgfmt now produces little-endian .mo files by default.