diff options
author | Joel E. Denny <jdenny@clemson.edu> | 2010-02-22 18:09:19 -0500 |
---|---|---|
committer | Joel E. Denny <jdenny@clemson.edu> | 2010-02-22 18:32:47 -0500 |
commit | 47fa574761319b0a422691223c9b8a9a72f36aa2 (patch) | |
tree | 293bebddb72a6e8ca06acc6da36d6d5a9e60f767 /po | |
parent | d67e37a74bc91e2fe254e5b7a28c62537a20e206 (diff) | |
download | bison-47fa574761319b0a422691223c9b8a9a72f36aa2.tar.gz |
portability: fix several issues with M4 subprocess.
M4's output pipe was not being drained upon fatal errors during
scan_skel. As a result, broken-pipe messages from M4 were seen
on at least AIX, HP-UX, Solaris, and RHEL4, and this caused a
failure in the test suite. The problem was that, on platforms
where the default disposition for SIGPIPE is ignore instead of
terminate, M4 sometimes saw fwrite fail with errno=EPIPE and
then reported it. However, there's some sort of race condition,
because the new test group occasionally succeeded.
Reported by Albert Chin at
<http://lists.gnu.org/archive/html/bug-bison/2010-02/msg00004.html>.
There were also problems with the test suite livelocking on
Tru64 5.1b. Reported by Didier Godefroy at
<http://lists.gnu.org/archive/html/bug-bison/2009-05/msg00005.html>.
Switching to create_pipe_bidi suggested by Akim Demaille.
To attempt to solve both of these problems, switch to gnulib's
create_pipe_bidi and register M4 process as a slave. Along the
way, clean up file name conflict handling, which was affected by
the broken-pipe problem before the switch.
* NEWS (2.4.2): Document.
* THANKS (Didier Godefroy): Add.
* bootstrap.conf (gnulib_modules): Add pipe.
* gnulib: Update to latest to make sure we have all the latest
fixes.
* lib/Makefile.am (libbison_a_SOURCES): Remove subpipe.h and
subpipe.c.
* po/POTFILES.in (lib/subpipe.c): Remove.
* src/files.c (compute_output_file_names): Update invocations
of output_file_name_check.
(output_file_name_check): In the case that the grammar file
would be overwritten, use complain instead of fatal, but replace
the output file name with /dev/null. Use the /dev/null solution
for the case of two conflicting output files as well because it
seems safer in case Bison one day tries to open both files at
the same time.
* src/files.h (output_file_name_check): Update prototype.
* src/output.c (output_skeleton): Use create_pipe_bidi and
wait_subprocess. Assert that scan_skel completely drains the
pipe.
* src/scan-skel.l (at_directive_perform): Update
output_file_name_check invocation.
* tests/output.at (AT_CHECK_CONFLICTING_OUTPUT): Check that the
grammar file actually isn't overwritten.
(Conflicting output files: -o foo.y): Update expected output.
* tests/skeletons.at (Fatal errors but M4 continues producing
output): New test group.
(cherry picked from commit 22cc8d813ee57c9631e527a31010ab138f9b7e06)
Conflicts:
NEWS
bootstrap.conf
lib/.cvsignore
lib/.gitignore
m4/.cvsignore
m4/.gitignore
src/output.c
Diffstat (limited to 'po')
-rw-r--r-- | po/POTFILES.in | 1 |
1 files changed, 0 insertions, 1 deletions
diff --git a/po/POTFILES.in b/po/POTFILES.in index 59f7701c..a9519338 100644 --- a/po/POTFILES.in +++ b/po/POTFILES.in @@ -22,6 +22,5 @@ lib/error.c lib/getopt.c lib/obstack.c lib/quotearg.c -lib/subpipe.c lib/timevar.c lib/xalloc-die.c |