Eric Blake [Fri, 30 Jul 2010 22:01:41 +0000 (16:01 -0600)]
strtod: next round of AIX fixes
* lib/strtod.c (strtod): Work around AIX bug of parsing p with no
exponent.
* tests/test-strtod.c (main): Enhance tests.
* doc/posix-functions/strtod.texi (strtod): Document next bug.
Reported by Rainer Tammer.
Paul R. Eggert [Mon, 26 Jul 2010 17:56:32 +0000 (10:56 -0700)]
timespec: use cast and not conditional, as truncation isn't possible
* lib/timespec.h (timespec_cmp): Use cast to pacify gcc -Wconversion
instead of a conditional. Comment about the situation in more detail.
This undoes most of the 2009-10-29 patch.
Paolo Bonzini [Tue, 20 Jul 2010 14:23:13 +0000 (16:23 +0200)]
unistr/u*-chr: test multibyte sequences more
* tests/unistr/test-chr.h: Do complete testing of the characters in the
test vector.
* tests/unistr/test-u8-chr.c, tests/unistr/test-u32-chr.c,
tests/unistr/test-u16-chr.c (U_UCTOMB): Define.
Paolo Bonzini [Tue, 20 Jul 2010 14:06:33 +0000 (16:06 +0200)]
unistr/u*-chr: prepare for multibyte tests
* modules/unistr/u16-chr-tests: Depend on u32-to-u16.
* modules/unistr/u8-chr-tests: Depend on u32-to-u8.
* tests/unistr/test-chr.h: Build initial version as UCS-4 then convert.
* tests/unistr/test-u16-chr.c, tests/unistr/test-u32-chr.c,
tests/unistr/test-u8-chr.c: Define U32_TO_U.
Paul R. Eggert [Wed, 14 Jul 2010 07:06:03 +0000 (00:06 -0700)]
getcwd: on Solaris, work better if ancestors are inaccessible
* lib/getcwd.c (__getcwd): If getcwd returns EINVAL for zero
buffer and size, try again with a large buffer. This works better
on Solaris, since its getcwd succeeds even if the path to the root
is inaccessible, and this is helpful in common cases such as .zfs
hidden directories. Problem reported by J Chapman Flack in
http://lists.gnu.org/archive/html/bug-tar/2010-06/msg00000.html
Use system getcwd if it's declared, not merely if it's partly
working; use the partly-working test only to avoid needless effort
if the system getcwd fails.
* m4/getcwd-path-max.m4 (gl_FUNC_GETCWD_PATH_MAX): Omit
comment that was already obsolete and is now even more obsolete.
* modules/getcwd (Depends-on): Depend on strdup, since __getcwd
now might call strdup.
Paul R. Eggert [Tue, 13 Jul 2010 22:55:36 +0000 (15:55 -0700)]
pthread: Add enough so that coreutils/src/sort.c compiles.
* lib/pthread.in.h: Add self to author comment. Conditionalize on
_GL_PTHREAD_H, not PTHREAD_H_, for consistency with the rest of
gnulib. Include <sched.h> and <time.h>, as per POSIX.
Include <sys/types.h>, in case it defines pthread_t.
(pthread_t, pthread_attr_t, pthread_barrier_t, pthread_barrierattr_t):
(pthread_cond_t, pthread_condattr_t, pthread_key_t, pthread_mutex_t):
(pthread_mutexattr_t, pthread_once_t, pthread_rwlock_t):
(pthread_rwlockattr_t, pthread_spinlock_t):
New typedefs, if HAVE_PTHREAD_T is not defined.
(PTHREAD_COND_INITIALIZER, PTHREAD_MUTEX_INITIALIZER):
(PTHREAD_ONCE_INIT, PTHREAD_RWLOCK_INITIALIZER):
(PTHREAD_BARRIER_SERIAL_THREAD, PTHREAD_CANCEL_DEFERRED):
(PTHREAD_CANCEL_ASYNCHRONOUS, PTHREAD_CANCEL_ENABLE):
(PTHREAD_CANCEL_DISABLE, PTHREAD_CANCELED, PTHREAD_CREATE_JOINABLE):
(PTHREAD_CREATE_DETACHED, PTHREAD_INHERIT_SCHED):
(PTHREAD_EXPLICIT_SCHED, PTHREAD_MUTEX_DEFAULT, PTHREAD_MUTEX_NORMAL):
(PTHREAD_MUTEX_ERRORCHECK, PTHREAD_MUTEX_RECURSIVE):
(PTHREAD_MUTEX_STALLED, PTHREAD_MUTEX_ROBUST, PTHREAD_PRIO_NONE):
(PTHREAD_PRIO_INHERIT, PTHREAD_PRIO_PROTECT, PTHREAD_PROCESS_PRIVATE):
(PTHREAD_PROCESS_SHARED, PTHREAD_SCOPE_SYSTEM, PTHREAD_SCOPE_PROCESS):
New macros.
(pthread_cond_destroy, pthread_cond_init, pthread_cond_signal):
(pthread_cond_wait, pthread_exit, pthread_mutex_destroy):
(pthread_mutex_init, pthread_mutex_lock, pthread_mutex_unlock):
(pthread_spin_init, pthread_spin_lock, pthread_spin_trylock);
(pthread_spin_unlock): New dummy functions.
(pthread_create): Return EAGAIN; don't set errno.
* m4/pthread.m4 (gl_PTHREAD_CHECK): Check for pthread_t, and
require AC_C_INLINE.
* modules/pthread (Depends-on): Add sched, time.
(pthread.h): Use AM_V_GEN.
Paul R. Eggert [Mon, 12 Jul 2010 17:58:35 +0000 (10:58 -0700)]
memcoll: clarify sizes versus lengths, document better, and tweak perf
* lib/memcoll.c (strcoll_loop, memcoll0):
Improve quality of descriptive comments. Name variables
consistently as to whether they are lengths (which do not include
terminating null) versus sizes (which do).
* lib/xmemcoll.c (xmemcoll0): Likewise.
* lib/memcoll.c (strcoll_loop): Tweak the way that the diff is
returned when s1size == 0; this is easier to compile and saves
about 17% of memcoll's code space on x86-64 with GCC 4.1.2.
Paul R. Eggert [Mon, 12 Jul 2010 16:14:10 +0000 (09:14 -0700)]
strtod: make it more-accurate typically, and don't require libm
* lib/strtod.c (_GL_ARG_NONNULL): Remove; no longer needed.
Include limits.h. Don't include string.h.
(HAVE_LDEXP_IN_LIBC, HAVE_RAW_DECL_STRTOD): Define to 0 if not defined.
(locale_isspace): New function, so that no casts are needed to
check whether *s is a space.
(ldexp): Provide an unused dummy if not available.
(scale_radix_exp, parse_number, underlying_strtod): New functions.
(strtod): Use them. This implementation prefers to use the
underlying strtod if available, falling back on our own code
only to fix known bugs. This is more likely to produce an
accurate result. Also, it avoids the use of libm functions.
* m4/strtod.m4 (gl_FUNC_STRTOD): Don't invoke _AC_LIBOBJ_STRTOD;
no longer needed. Invoke AC_LIBOBJ([strtod]); don't know why this
was absent, but it caused a test failure with coreutils.
(gl_PREREQ_STRTOD): Check wither ldexp can be used without linking
with libm.
* modules/strtod (Makefile.am, Link): libm is no longer needed.
* modules/strtod-tests (Makefile.am): Likewise.
Paul Eggert [Fri, 9 Jul 2010 01:16:40 +0000 (18:16 -0700)]
(x)memcoll: minor tweaks
* lib/memcoll.c (strcoll_loop): Prefer the style where 'const'
is after the type that it qualifies.
(memcoll0): Likewise.
* lib/memcoll.h (memcoll0): Likewise.
* lib/xmemcoll.c (collate_error, xmemcoll0): Likewise.
* lib/xmemcoll.h (xmemcoll0): Likewise.
* lib/memcoll.c (memcoll0): Correct the comment. This function
differs from memcoll in that the NUL byte is part of the argument.
Omit the abort-checks, as performance is a real issue here. Plus,
the checks were wrong anyway (an off-by-one error). Omit local
variable 'diff', as it's a bit clearer that way.
* m4/memcoll.m4 (gl_MEMCOLL): Omit AC_FUNC_STRCOLL, as it's
no longer needed.
(x)memcoll: speedup when input is known to be NUL delimited
* lib/memcoll.c: Include stdlib.
(memcoll0) New function.
(strcoll_loop) New function, refactored for use in both memcoll
and memcoll0.
* lib/memcoll.h: Add prototype for memcoll0.
* lib/xmemcoll.c: (xmemcoll0) New function.
(collate_error) New function, refactored for use in both xmemcoll
and xmemcoll0.
* lib/xmemcoll.h: Add prototype for xmemcoll0.
* m4/memcoll.m4: add inline invocation.
bootstrap: discard non translation project po files
* build-aux/bootstrap (update_po_files): Remove any
local only files from the $po_dir/.reference directory.
The LINGUAS index is generated from this directory
and so those translations will not be distributed.
Note, local only translations are not removed from the
$po_dir directory.
Jim Meyering [Sun, 4 Jul 2010 08:54:38 +0000 (10:54 +0200)]
hash: once again explicitly disallow insertion of NULL
* lib/hash.c (hash_insert0): Reinstate just-removed test:
inserting a NULL pointer cannot work with these functions.
Add a comment with details.
This reverts part of the 2010-07-01 commit, 5bef1a35
"hash: extend module to deal with non-pointer keys".
Jim Meyering [Thu, 1 Jul 2010 21:17:25 +0000 (23:17 +0200)]
hash: extend module to deal with non-pointer keys
* lib/hash.c (hash_insert0): New interface, much like hash_insert
but that allows insertion of non-pointer entries.
Do not disallow an ENTRY value of NULL.
(hash_insert): This is now just a thin wrapper. Call hash_insert0.
* lib/hash.h (hash_insert0): Declare.
Eric Blake [Mon, 21 Jun 2010 22:16:44 +0000 (16:16 -0600)]
memmem: slight optimization
For any needle, the factorization 0/n has a local period of 1, so it
is a critical factorization iff the entire needle consists only of a
single repeated byte, in which case 1/n-1 would also be critical.
Starting with a comparison of x[0] and x[1] in the maximal suffix
check will either find the 0/n case or move on to something else, so
we can optimize and start with the x[1] vs. x[2] case to begin with.
To avoid out-of-bounds references, we must then special case needles
of length two or less. However, for these cases, we can determine a
critical factorization without any probes of the needle (we already
require a non-empty needle; a 1-byte needle can factor as either 0/1
or 1/0 but the rest of our code assumes a non-empty suffix; and of the
two 2-byte needle patterns, "aa" can factor as either 0/2 or 1/1 but
with best performance for 1/1, and "ab" must be factored as 1/1).
* lib/str-two-way.h (critical_factorization): Update comments.
Reduce work during factorization phase.
Reported by Carlos Bueno <carlos@bueno.org>.
Paul Eggert [Mon, 14 Jun 2010 19:20:22 +0000 (12:20 -0700)]
priv-set: Don't assume that priv.h exists merely because getppriv does.
See Jan Andersen's bug report about AIX 5L in
http://lists.gnu.org/archive/html/bug-tar/2010-06/msg00019.html
* m4/priv-set.m4 (gl_PRIV_SET): Check for priv.h.
* lib/priv-set.c: Do nothing unless HAVE_PRIV_H.
* lib/priv-set.h: Likewise.
* tests/test-priv-set.c: Likewise.
Jim Meyering [Fri, 11 Jun 2010 20:09:27 +0000 (22:09 +0200)]
test-sys_socket: mark variables as used more readably
* tests/test-sys_socket.c (main): Mark otherwise unused variables
as "used" explicitly via (void) statement casts. This is more
readable than using them in an artificial return expression.
Suggestion from Bruno Haible.