This is for Gawk, which does not use malloc-gnu.
* lib/regex.c (_GL_USE_STDLIB_ALLOC) [!_LIBC]:
Define, since this module is now safe for AIX-like malloc.
* lib/regex_internal.h (re_malloc) [!_LIBC && !HAVE_MALLOC_0_NONNULL]:
Don’t pass 0 to malloc.
* lib/regex_internal.h (IDX_MAX) [_REGEX_LARGE_OFFSETS]:
Port to non-POSIX platforms where limits.h does not define SSIZE_MAX.
This is for Gawk, which does not use Gnulib limits.h.
* lib/idx.h: Include <stddef.h>, <stdint.h> only if needed.
(idx_t, IDX_MAX): Rely on builtin macros __PTRDIFF_TYPE__,
__PTRDIFF_MAX__ if present; this avoids polluting the namespace on
GNUish systems.
Respect --disable-threads on native MS-Windows.
* lib/localename-unsafe.c (GLWTHREAD_MUTEX_INIT, glwthread_mutex_t)
(glwthread_mutex_lock, glwthread_mutex_unlock)
[AVOID_ANY_THREADS && _WIN32 && !__CYGWIN]:
Instead of including "windows-mutex.h", provide no-op substitutes
for its symbols used here.
* lib/glthread/cond.h: Revert my 2026-03-31 namespace cleanup
patch, which was incorrect because this header’s
gl_cond_timedwait_func function uses lock_t even when threading is
disabled.
* lib/glthread/thread.c (gl_thread_create):
Define as a function only if multithreading.
This is cleaner than my previous patch today that sometimes made
it _Noreturn, as the function’s definition and declaration now always
agree on noreturnedness, and it pushes the noreturnedness issue up
to the caller. Also, it suppresses GCC’s “warning: function
declared 'noreturn' has a 'return' statement”.
* lib/glthread/thread.h (glthread_atfork, glthread_sigmask)
(glthread_create, glthread_join): Evaluate arguments even when
these macros are no-ops. Type-check the arguments too. This is
cleaner anyway, in case the args have the wrong types (or have
side effects!).
In the regex code, use the char32_t functions instead of the
wchar_t functions, so that regex stays in sync with dfa.
This should fix a bug in Gnu grep reported by Dennis Clarke for
OpenBSD <https://debbugs.gnu.org/cgi/bugreport.cgi?bug=80774#47>.
A similar bug occurs in some macOS releases.
* modules/regex (Depends-on): Replace btowc, iswctype, mbrtowc,
wchar-h, wcrtomb, wctype-h, and wctype with btoc32,
c32_apply_type_test, c32_get_type_test, c32isalnum, c32rtomb,
c32tolower, c32toupper, mbrtoc32-regular, uchar-h.
* lib/regex_internal.h [!_LIBC]: Do not include <wchar.h>, <wctype.h>.
Instead, include <uchar.h> and #define wchar_t, wctype_t,
__wctype, __iswalnum, __iswctype, __towlower, __towupper, __btowc,
__mbrtowc, and __wcrtomb to their char32_t counterparts.
This should help merges changes from Gawk, which always uses the
char32_t API though that’s sometimes implemented with the wchar_t
API even on platforms where wchar_t and char32_t act differently.
The idea is to use char32_t uniformly in both the dfa and regex
modules, so that they get consistent answers on all platforms.
* lib/dfa.c, lib/localeinfo.c, lib/localeinfo.h: If GAWK, do not
include <wctype.h> or redefine the Gnulib char32_t types and
functions to be wchar.h and wctype.h functions or define mbszero
and streq, as I think I have a better way to do this with Gawk
that is less intrusive here; instead, always include <uchar.h>.
* lib/dfa.c: Do not include <wchar.h>. Include "getext.h" before
including "xalloc.h" and "localinfo.h", as Gnulib doesn’t care
about the order and this works better with Gawk’s way of overriding Gnulib.
(parse_bracket_exp): Use && instead of &; either is correct and
both are equally fast nowadays but && triggers a warning in some
Gawk compiles.
* lib/dfa.h (_GL_ATTRIBUTE_MALLOC, _GL_ATTRIBUTE_DEALLOC)
(_GL_ATTRIBUTE_DEALLOC_FREE)
(_GL_ATTRIBUTE_RETURNS_NONNULL) [!_GL_ATTRIBUTE_MALLOC]:
Remove, as Gawk’s custom.h can define them.
* lib/localeinfo.c: Go back to using <verify.h> and ‘verify’
instead of using static_assert which Gawk can’t easily use because
it does not use Gnulib’s assert-h module.
* lib/localeinfo.h: Do not include <limits.h>, avoding some
namespace pollution.
(struct localeinfo): Use (unsigned char) -1 instead of UCHAR_MAX
to avoid the need to include <limits.h>.
* modules/dfa (Depends-on): Remove wchar-h.
This is only a partial cleanup; to be cleaner we’d need to
move declarations of Gnulib extensions like c32isalpha
into a separate .h file. However, if no Gnulib modules that
extend <uchar.h> are used, Gnulib <uchar.h> is now pretty clean on
recent GNUish platforms.
* lib/uchar.in.h: On GNUish platforms, include <stdint.h>,
<wchar.h>, <wctype.h> only if needed. Do not include <string.h>,
as we never need it directly: even if we use <string.h>’s memset
via mbszero, <wchar.h> should include <string.h> if needed, as
<wchar.h> defines mbszero. Move a static_assert from here to tests,
as the static_assert uses a symbol that is no longer guaranteed
to be visible.
* modules/uchar-h (Depends-on): Do not depend on assert-h.
* tests/test-uchar-h.c: Move a static_assert here from lib/uchar.in.h,
and include <wchar.h> so that wchar_t is guaranteed to be visible.
Do not pollute <wchar.h> includers with <string.h> names
on GCC-compatible platforms.
* lib/wchar.in.h (_GL_WCHAR_MEMSET) [@GNULIB_MBSZERO@]: New macro.
(mbszero): Use it.
* lib/readutmp.c (read_utmp_from_systemd):
Do not merely treat a "manager*" class as a login process.
Instead, treat all non-"user*" processes as login processes.
With current systemd, this changes the treatment of "greeter",
"lock-screen", "background", "background-light", and "none"
classes so that they are now considered to be login processes, not
user processes.
* lib/regex_internal.c (re_node_set_insert): Remove the DEBUG_ASSERT
and instead return early for an attempt to insert an ELEM that is
already present in the set. Relax the function's comment that says
there should be no duplicate. This function is called from many
places and has been working fine. With its nontrivial backrefs,
the sample regexp apparently elicits enough backtracking retries
and state-set merges to trigger this duplicate insertion attempt.
Reported by Bruno Haible in
https://lists.gnu.org/r/bug-gnulib/2026-04/msg00138.html
* lib/fts.in.h (FTS_NOSTAT): Use a FIFO, not a directory,
in the commentary’s example. This is a better example
because directories never have FTS_NSOK.
* lib/stdbit.in.h: With the Intel icpx C++ compiler, include <stddef.h>
and <stdint.h>, and define the __STDC_ENDIAN_* macros.
* doc/posix-headers/stdbit.texi: Document the Intel icpx bug.
* lib/regexec.c (re_search_internal): Use only one copy
of the 5-line reg-initialization code, removing that else block,
at the tiny cost of moving a small test into the loop.
With a backref pattern like ^(.?)(.?).?\2\1 (no $), the engine
could miss valid short matches. For example, "ab" should match
via all-empty groups, yet regexec returned no-match because
set_regs failed at the longest structural match (match_last=2)
and never retried at a shorter match_last.
* lib/regexec.c (re_search_internal): When set_regs fails for a
backref pattern, retry prune_impossible_nodes and set_regs at
progressively shorter match lengths. Save a copy of state_log
before pruning so shorter retries can re-sift from the original
states.
* m4/regex.m4: Also reject system regex with this bug.
* tests/test-regex.c (main): Add a test for this bug.
Reported by Ed Morton in https://bugs.gnu.org/68725
Co-authored-by: Claude <noreply@anthropic.com>
Use of sed or grep with both backreferences and an end-of-line
anchor could get false matches. For example, this grep command
would falsely declare "ab" to be a palindrome:
grep -E '^(.?)(.?).?\2\1$' <<< ab
In prune_impossible_nodes, falling back to a shorter match
skipped the halt-state context check, so $ was not verified.
* lib/regexec.c (prune_impossible_nodes): Also require
check_halt_state_context to succeed in the loop that searches
for an earlier halt state.
* m4/regex.m4: Bump serial to 82. Reject any system regex
implementation with this bug.
* tests/test-regex.c (main): Add a test for this bug.
Reported by Ed Morton in https://bugs.gnu.org/68725
* lib/tempname.c (random_bits) [CLOCK_REALTIME]: Do not call
‘clock’, as an optimization. There is no need to call ‘clock’, as
it likely gives us less info than clock_gettime, and if
clock_gettime fails then ‘clock’ will likely fail too.
This patch is a simplified version of the patch made in glibc commit
5f62cf88c4530c11904482775b7582bd7f6d80d2 dated 2024-09-25,
and Gnulib lib/tempname.c should now be suitable as-is for
replacing Glibc sysdeps/posix/tempname.c.
In documentation and comments, be more like POSIX in terminology
involving multithreading. Explain the distinction between
multithreaded process vs multithreaded program. Change “program”
to “process” when the latter wording is more accurate or informative.
Simplify the wording for the constraints on processes that use
unlocked I/O. Change “multithread-safe” to “thread-safe”.
Change “thread-safety” to “thread safety”.
However, do not change “multithreaded” to “multi-threaded” even
though there are some uses of both spellinga, as there are a whole
bunch of uses of “multithreaded”, also in identifier names;
perhaps Gnulib should even standardize on “multithreaded”
(not “multi-threaded”), contra POSIX.
* lib/sigprocmask.c (glwthread_spin_lock, glwthread_spin_unlock): Define
to empty if GNULIB_SIGPROCMASK_SINGLE_THREAD is defined.
(overrides_mt_lock, overrides_handler_lock): Don't define if
GNULIB_SIGPROCMASK_SINGLE_THREAD is defined.
* doc/multithread.texi (Multithreading Optimizations): Document
GNULIB_SIGPROCMASK_SINGLE_THREAD instead of
GNULIB_PTHREAD_SIGMASK_SINGLE_THREAD.
* lib/sigprocmask.c (overrides_handler_lock): Renamed from
overrides_lock.
(overrides_mt_lock): New variable.
(pthread_sigmask): Lock both locks.
(rpl_signal): Lock overrides_mt_lock, not overrides_lock.
* lib/signal.in.h (WIN_PTHREADS_SIGNAL_H): New macro.
* lib/sigprocmask.c: Include windows-spin.h.
(thread_local): New macro.
(blocked_set, pending_array): Mark as thread-local.
(blocked_handler): Remove function.
(struct override): New type.
(overrides, overrides_lock): New variables.
(override_handler): New function.
(pthread_sigmask): New function, borrowing from the previous sigprocmask
definition.
(sigprocmask): Now a wrapper around pthread_sigmask.
(rpl_signal): Use the overrides_lock to make it multithread-safe.
(_gl_raise_SIGPIPE): Add comments.
* modules/sigprocmask (Depends-on): Add windows-spin.
* lib/pthread_sigmask.c: Revert last change. On native Windows, don't
define pthread_sigmask here.
* modules/pthread_sigmask (Depends-on): Remove lock.
* NEWS: Remove the last entry.
* lib/term-style-control.c (block_relevant_signals)
(unblock_relevant_signals): Prefer pthread_sigmask to sigprocmask.
* modules/term-style-control (Depends-on):
Depend on pthread_sigmask, not on sigprocmask.
* lib/fatal-signal.c (block_fatal_signals)
(unblock_fatal_signals): Prefer pthread_sigmask to sigprocmask.
* modules/fatal-signal (Depends-on):
Depend on pthread_sigmask, not on sigprocmask.
* lib/sigaction.c (sigaction_handler, sigaction):
Use pthread_sigmask, not sigprocmask.
* modules/sigaction (Depends-on):
Depend on pthread_sigmask, not on sigprocmask.
Since sigprocmask should now be used only in single-threaded processes,
move the locking from lib/sigprocmask.c to lib/pthread_sigmask.c.
* lib/pthread_sigmask.c: If !HAVE_SIGPROCMASK &&
!GNULIB_PTHREAD_SIGMASK_SINGLE_THREAD, replace Gnulib sigprocmask
with a thread-safe subsitute sigprocmask_r.
Include glthread/lock.h to implement this.
* lib/sigprocmask.c (gl_lock_define_initialized, gl_lock_lock)
(gl_lock_unlock, sig_lock): Remove. All uses removed.
Do not include glthread/lock.h.
* m4/signalblocking.m4 (gl_SIGNALBLOCKING):
Define HAVE_SIGPROCMASK, for the benefit of pthread_sigmask.c.
* modules/pthread_sigmask (Depends-on): Add ‘lock’.
* modules/sigprocmask (Depends-on): Remove ‘lock’.
This is a performance win on GLIBC,
as tested on the coreutils multi-byte update to cut(1):
$ yes $(yes éééááé | head -n9 | paste -s -d，) |
head -n1M > mb.in
$ time LC_ALL=C.UTF-8 src/cut-before -c1 mb.in >/dev/null
real 0m1.582s
$ time LC_ALL=C.UTF-8 src/cut-after -c1 mb.in >/dev/null
real 0m0.592s
* lib/mcel.h: While GLIBC's mbrtoc32 is functional for mcel,
it is seen to be 2.6x slower than gnulib's implementation
due to GLIBC's per call locale handling.
Suggested by Paul Eggert.
* lib/asyncsafe-spin.c (asyncsafe_spin_lock, asyncsafe_spin_unlock): Use
pthread_sigmask instead of sigprocmask.
* modules/asyncsafe-spin (Depends-on): Add pthread_sigmask. Remove
sigprocmask.
(Link): New section.
* modules/asyncsafe-spin-tests (Makefile.am): Link test-asyncsafe-spin1
with $(PTHREAD_SIGMASK_LIB).
* modules/jit/cache-tests (Makefile.am): Link test-cache with
$(PTHREAD_SIGMASK_LIB).
Suggested by Paul Eggert.
* lib/spawn-pipe.c (execute): Use pthread_sigmask to get the set of
blocked signals.
* modules/spawn-pipe (Depends-on): Add pthread_sigmask.
