The optimization is achieved by following techniques:
> data alignment [gain from aligned memory access on read/write]
> POWER7 gains performance with loop unrolling/unwinding
[gain by reduction of branch penalty].
> zero padding done by calling optimized memset
This patch changes de default symbol redirection for internal call of
memcpy, memset, memchr, and strlen to the IFUNC resolved ones. The
performance improvement is noticeable in algorithms that uses these
symbols extensible, like the regex functions.
In 7447ccd98e, the XDR currency was
removed from locale/iso-4217.def, despite the fact that it's both
still a part of the standard, according to the official table:
http://www.currency-iso.org/dam/downloads/table_a1.xml
... and, more importantly, is referenced from localedata/i18n, so
any quick-and-dirty locale definition that uses "copy i18n" for
LC_MONETARY wouldn't work anymore.
Define FEATURE_INDEX_1 and FEATURE_INDEX_MAX as macros
for use by both assembly and C code. This fixes the
-Wundef error for cases where FEATURE_INDEX_1 was not
defined but used the correct value of 0 for an undefined
macro.
[BZ #16885]
* sysdeps/sparc/sparc64/strcmp.S: Fix end comparison handling when
multiple zero bytes exist at the end of a string.
Reported by Aurelien Jarno <aurelien@aurel32.net>
* string/test-strcmp.c (check): Add explicit test for situations where
there are multiple zero bytes after the first.
lowlevellock.c for arm differs from the generic lowlevellock.c only in
insignificant ways, so can be removed. Happily, this fixes BZ 15119
(unnecessary busy loop in __lll_timedlock_wait on arm).
The notable differences between the arm and generic implementations are:
1) arm __lll_timedlock_wait has a fast path out if futex has been set
to 0 between since the function was called. This seems unlikely to
happen very often, so it seems at worst harmless to lose this fast
path.
2) Some function in arm's lowlevellock.c set futex to 2 if it was 1.
The generic version always sets the futex to 2. As futex can only be
0, 1 or 2 on entry into these functions, the behaviour is equivalent.
(If the futex manages to be 0 on entry then we've just lost another
unlikely fast path out.)
There are no test suite regressions.
Note that hppa and sparc also have their own lowlevellock.c. I believe
hppa can also be removed, so I'll send a separate patch for that
shortly. sparc's seems to be genuinely needed as it uses a different
locking structure.
Also note that the analysis at
https://sourceware.org/ml/libc-ports/2013-02/msg00021.html indicates a
further locking performance bug to fix - I've got a partial patch for
that which I can submit once I've finished testing.
2014-05-01 Bernard Ogden <bernie.ogden@linaro.org>
[BZ #15119]
* sysdeps/unix/sysv/linux/arm/nptl/lowlevellock.c: Remove file.
This patch fixes what I believe to be a bug in the handling of
R_ARM_IRELATIVE RELA relocations. At present, these are handled the
same as REL relocations: i.e. the addend is loaded from the relocation
address. Most of the time this isn't a problem because RELA relocations
aren't used on ARM (GNU/Linux at least) anyway, but it causes problems
with prelink, which uses RELA on all targets for its conflict table.
(Support for ifunc prelinking requires a prelink patch, not yet posted.)
Anyway, this patch works, though I'm not 100% sure if it is correct: I
notice that this code path received attention last year:
https://sourceware.org/ml/libc-ports/2013-07/msg00000.html
I'm not sure under what circumstances that patch would have had an
effect, nor if my patch conflicts with that case.
No regressions using Mentor's usual glibc cross-testing infrastructure.
[BZ #16888]
* sysdeps/arm/dl-machine.h (elf_machine_rela): Fix R_ARM_IRELATIVE
handling.
This patch increases the minimum Linux kernel version for glibc to
2.6.32, as discussed in the thread starting at
<https://sourceware.org/ml/libc-alpha/2014-01/msg00511.html>.
This patch just does the minimal change to arch_minimum_kernel
settings (and LIBC_LINUX_VERSION, which determines the minimum kernel
headers version, as it doesn't make sense for that to be older than
the minimum kernel that can be used at runtime). Followups would be
expected to do, roughly and not necessarily precisely in this order:
* Remove __LINUX_KERNEL_VERSION checks in kernel-features.h files
where those checks are always true / always false for kernels 2.6.32
and above.
* Otherwise simplify/improve conditionals in those files (for example,
where defining once in the main file then undefining in
architecture-specific files makes things clearer than having lots of
separate definitions of the same macro), possibly fixing in the
process cases where a macro should optimally have been defined for a
given architecture but wasn't. (In the review in preparation for
this version increase I checked what the right conditions should be
for all macros in the main kernel-features.h whose definitions there
would have been affected by the increase - but I only fixed that
subset of the issues found where --enable-kernel=2.6.32 would have
caused a kernel feature to be wrongly assumed to be present, not any
cases where a feature is not assumed but could be assumed.)
* Remove conditionals on __ASSUME_* where they can now be taken to be
always-true, and the definitions when the macros are only used in
Linux-specific files.
* Split more architectures out of the main kernel-features.h (like
ex-ports architectures), once various of the architecture
conditionals there have been eliminated so the new
architecture-specific files are no larger than actually necessary.
Tested x86_64.
2014-03-27 Joseph Myers <joseph@codesourcery.com>
[BZ #9894]
* sysdeps/unix/sysv/linux/configure.ac (LIBC_LINUX_VERSION):
Change to 2.6.32.
(arch_minimum_kernel): Change all 2.6.16 settings to 2.6.32.
* sysdeps/unix/sysv/linux/configure: Regenerated.
* sysdeps/unix/sysv/linux/microblaze/configure.ac: Remove file.
* sysdeps/unix/sysv/linux/microblaze/configure: Likewise.
* sysdeps/unix/sysv/linux/tile/configure.ac: Likewise.
* sysdeps/unix/sysv/linux/tile/configure: Likewise.
* README: Update reference to required Linux kernel version.
* manual/install.texi (Linux): Update reference to required Linux
kernel headers version.
* INSTALL: Regenerated.
Continuing the series of patches to clean up conformtest expectations
for "POSIX" (1995/6) based on review of the expectations against the
standard, this patch cleans up expectations for stdlib.h and
string.h. Tested x86_64; no new XFAILs needed.
* conform/data/stdlib.h-data [POSIX] (stddef.h): Do not allow
header inclusion.
[POSIX] (limits.h): Likewise.
[POSIX] (math.h): Likewise.
[POSIX] (sys/wait.h): Likewise.
* conform/data/string.h-data [POSIX || UNIX98] (strtok_r): Require
function.
[POSIX] (stddef.h): Do not allow header inclusion.
lll_unlock() will be called again if it goes to "wake_all" in
pthread_cond_broadcast(). This may make another thread which is
waiting for lock in pthread_cond_timedwait() unlock. So there are
more than one threads get the lock, it will break the shared data.
It's introduced by commit 8313cb997d2d("FUTEX_*_REQUEUE_PI support for
non-x86 code")
The datahead structure has an unused padding field that remains
uninitialized. Valgrind prints out a warning for it on querying a
netgroups entry. This is harmless, but is a potential data leak since
it would result in writing out an uninitialized byte to the cache
file. Besides, this happens only when there is a cache miss, so we're
not adding computation to any fast path.
This patch consolidates the code to initialize the header of a dataset
into a single set of functions (one for positive and another for
negative datasets) primarily to reduce repetition of code. The
secondary reason is to simplify Patch 2/2 which fixes the problem of
an uninitialized byte in the header by initializing an unused field in
the structure and hence preventing a possible data leak into the cache
file.
[Fixes BZ #14308, #12994, #13651]
AF_UNSPEC results in sending two queries in parallel, one for the A
record and the other for the AAAA record. If one of these is a
referral, then the query fails, which is wrong. It should return at
least the one successful response.
The fix has two parts. The first part makes the referral fall back to
the SERVFAIL path, which results in using the successful response.
There is a bug in that path however, due to which the second part is
necessary. The bug here is that if the first response is a failure
and the second succeeds, __libc_res_nsearch does not detect that and
assumes a failure. The case where the first response is a success and
the second fails, works correctly.
This condition is produced by buggy routers, so here's a crude
interposable library that can simulate such a condition. The library
overrides the recvfrom syscall and modifies the header of the packet
received to reproduce this scenario. It has two key variables:
mod_packet and first_error.
The mod_packet variable when set to 0, results in odd packets being
modified to be a referral. When set to 1, even packets are modified
to be a referral.
The first_error causes the first response to be a failure so that a
domain-appended search is performed to test the second part of the
__libc_nsearch fix.
The driver for this fix is a simple getaddrinfo program that does an
AF_UNSPEC query. I have omitted this since it should be easy to
implement.
I have tested this on x86_64.
The interceptor library source:
/* Override recvfrom and modify the header of the first DNS response to make it
a referral and reproduce bz #845218. We have to resort to this ugly hack
because we cannot make bind return the buggy response of a referral for the
AAAA record and an authoritative response for the A record. */
#define _GNU_SOURCE
#include <sys/types.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include <stdio.h>
#include <stdbool.h>
#include <endian.h>
#include <dlfcn.h>
#include <stdlib.h>
/* Lifted from resolv/arpa/nameser_compat.h. */
typedef struct {
unsigned id :16; /*%< query identification number */
#if BYTE_ORDER == BIG_ENDIAN
/* fields in third byte */
unsigned qr: 1; /*%< response flag */
unsigned opcode: 4; /*%< purpose of message */
unsigned aa: 1; /*%< authoritive answer */
unsigned tc: 1; /*%< truncated message */
unsigned rd: 1; /*%< recursion desired */
/* fields
* in
* fourth
* byte
* */
unsigned ra: 1; /*%< recursion available */
unsigned unused :1; /*%< unused bits (MBZ as of 4.9.3a3) */
unsigned ad: 1; /*%< authentic data from named */
unsigned cd: 1; /*%< checking disabled by resolver */
unsigned rcode :4; /*%< response code */
#endif
#if BYTE_ORDER == LITTLE_ENDIAN || BYTE_ORDER == PDP_ENDIAN
/* fields
* in
* third
* byte
* */
unsigned rd :1; /*%< recursion desired */
unsigned tc :1; /*%< truncated message */
unsigned aa :1; /*%< authoritive answer */
unsigned opcode :4; /*%< purpose of message */
unsigned qr :1; /*%< response flag */
/* fields
* in
* fourth
* byte
* */
unsigned rcode :4; /*%< response code */
unsigned cd: 1; /*%< checking disabled by resolver */
unsigned ad: 1; /*%< authentic data from named */
unsigned unused :1; /*%< unused bits (MBZ as of 4.9.3a3) */
unsigned ra :1; /*%< recursion available */
#endif
/* remaining
* bytes
* */
unsigned qdcount :16; /*%< number of question entries */
unsigned ancount :16; /*%< number of answer entries */
unsigned nscount :16; /*%< number of authority entries */
unsigned arcount :16; /*%< number of resource entries */
} HEADER;
static int done = 0;
/* Packets to modify. 0 for the odd packets and 1 for even packets. */
static const int mod_packet = 0;
/* Set to true if the first request should result in an error, resulting in a
search query. */
static bool first_error = true;
static ssize_t (*real_recvfrom) (int sockfd, void *buf, size_t len, int flags,
struct sockaddr *src_addr, socklen_t *addrlen);
void
__attribute__ ((constructor))
init (void)
{
real_recvfrom = dlsym (RTLD_NEXT, "recvfrom");
if (real_recvfrom == NULL)
{
printf ("Failed to get reference to recvfrom: %s\n", dlerror ());
printf ("Cannot simulate test\n");
abort ();
}
}
/* Modify the second packet that we receive to set the header in a manner as to
reproduce BZ #845218. */
static void
mod_buf (HEADER *h, int port)
{
if (done % 2 == mod_packet || (first_error && done == 1))
{
printf ("(Modifying header)");
if (first_error && done == 1)
h->rcode = 3;
else
h->rcode = 0; /* NOERROR == 0. */
h->ancount = 0;
h->aa = 0;
h->ra = 0;
h->arcount = 0;
}
done++;
}
ssize_t
recvfrom (int sockfd, void *buf, size_t len, int flags,
struct sockaddr *src_addr, socklen_t *addrlen)
{
ssize_t ret = real_recvfrom (sockfd, buf, len, flags, src_addr, addrlen);
int port = htons (((struct sockaddr_in *) src_addr)->sin_port);
struct in_addr addr = ((struct sockaddr_in *) src_addr)->sin_addr;
const char *host = inet_ntoa (addr);
printf ("\n*** From %s:%d: ", host, port);
mod_buf (buf, port);
printf ("returned %zd\n", ret);
return ret;
}
This patch optimizes the FPSCR update on exception and rounding change
functions by just updating its value if new value if different from
current one. It also optimizes fedisableexcept and feenableexcept by
removing an unecessary FPSCR read.
__int128 was added in GCC 4.6 and __int128_t was added before x86-64
was supported. This patch replaces __int128 with __int128_t so that
the installed bits/link.h can be used with older GCC.
* sysdeps/x86/bits/link.h (La_x86_64_regs): Replace __int128
with __int128_t.
(La_x86_64_retval): Likewise.
In the glibc manual we have a "Roadmap to the manual" section at
the end of the "Introduction" chapter.
The introductory text says "Here is an overview of the contents
of the remaining chapters of this manual.", but then proceeds to
list chapters out of order and some chapter are never referenced.
This commit reorders the overview to correctly match the manual
order.
See:
https://sourceware.org/ml/libc-alpha/2014-02/msg00823.html
This patch removes the arch specific powerpc implementation and instead
uses the linux default one. Although the current powerpc implementation
already constains the required memory barriers for correct
initialization, the default implementation shows a better performance on
newer chips.
Calling setcontext from a signal handler can be done safely so
it is sufficient to note that it is not recommended.
Also mention in setcontext documentation that the behaviour of
setcontext when restoring a context created by a call to a signal
handler is unspecified.
2014-04-17 Will Newton <will.newton@linaro.org>
* manual/setjmp.texi (System V contexts): Add note that
calling setcontext on a context created by a call to a
signal handler is undefined. Update text to note that
setcontext from a signal handler is possible but not
recommended.
On aarch64 calling swapcontext clobbers the state of the signal
stack (BZ #16629). Check that the address and size of the signal
stack before and after the call to swapcontext remains the same.
ChangeLog:
2014-04-17 Will Newton <will.newton@linaro.org>
[BZ #16629]
* stdlib/tst-setcontext.c: Include signal.h.
(main): Check that the signal stack before and
after swapcontext is the same.
The current implementation of setcontext uses rt_sigreturn to restore
the contents of registers. This contrasts with the way most other
architectures implement setcontext:
powerpc64, mips, tile:
Call rt_sigreturn if context was created by a call to a signal handler,
otherwise restore in user code.
powerpc32:
Call swapcontext system call and don't call sigreturn or rt_sigreturn.
x86_64, sparc, hppa, sh, ia64, m68k, s390, arm:
Only support restoring "synchronous" contexts, that is contexts
created by getcontext, and restoring in user code and don't call
sigreturn or rt_sigreturn.
alpha:
Call sigreturn (but not rt_sigreturn) in all cases to do the restore.
The text of the setcontext manpage suggests that the requirement to be
able to restore a signal handler created context has been dropped from
SUSv2:
If the context was obtained by a call to a signal handler, then old
standard text says that "program execution continues with the program
instruction following the instruction interrupted by the signal".
However, this sentence was removed in SUSv2, and the present verdict
is "the result is unspecified".
Implementing setcontext by calling rt_sigreturn unconditionally causes
problems when used with sigaltstack as in BZ #16629. On this basis it
seems that aarch64 is broken and that new ports should only support
restoring contexts created with getcontext and do not need to call
rt_sigreturn at all.
This patch re-implements the aarch64 setcontext function to restore
the context in user code in a similar manner to x86_64 and other ports.
ChangeLog:
2014-04-17 Will Newton <will.newton@linaro.org>
[BZ #16629]
* sysdeps/unix/sysv/linux/aarch64/setcontext.S (__setcontext):
Re-implement to restore registers in user code and avoid
rt_sigreturn system call.
Besides fixing the bugzilla, this also fixes corner-cases where the high
and low double differ greatly in magnitude, and handles a denormal
input without resorting to a fp rescale.
[BZ #16740]
[BZ #16619]
* sysdeps/ieee754/ldbl-128ibm/s_frexpl.c (__frexpl): Rewrite.
* math/libm-test.inc (frexp_test_data): Add tests.
Using -lm and -lpthread results in the shared objects in the system
being used to link against. This happened to work for libm because
there haven't been any changes to the libm ABI recently that could
break the existing benchmarks. This doesn't always work for the
pthread benchmarks. The correct way to build against libraries in the
build directory is to have the binaries explicitly depend on them so
that $(+link) can pick them up.
We initialize _r_debug for static binaries to allows debug
agents to treat static binaries a little more like dyanmic
ones. This simplifies the work a debug agent has to do to
access TLS in a static binary via libthread_db.
Tested on x86_64.
See:
https://sourceware.org/ml/libc-alpha/2014-04/msg00183.html
[BZ #16831]
* csu/libc-start.c (LIBC_START_MAIN) [!SHARED]: Call
_dl_debug_initialize.
The SELinux team has indicated to me that glibc's SELinux checks
in nscd are not being carried out as they would expect the API
to be used today. They would like to move away from static header
defines for class and permissions and instead use dynamic checks
at runtime that provide an answer which is dependent on the runtime
status of SELinux i.e. more dynamic.
The following patch is a minimal change that moves us forward in
this direction.
It does the following:
* Stop checking for SELinux headers that define NSCD__SHMEMHOST.
Check only for the presence or absence of the library.
* Don't encode the specific SELinux permission constants into a
table at build time, and instead use the symbolic name for the
permission as expected.
* Lookup the "What do we do if we don't know this permission?"
policy and use that if we find SELinux's policy is older than
the glibc policy e.g. we make a request for a permission that
SELinux doesn't know about.
* Lastly, translate the class and permission and then make
the permission check. This is done every time we lookup
a permission, and this is the expected way to use the API.
SELinux will optimize this for us, and we expect the network
latencies to hide these extra library calls.
Tested on x86, x86-64, and via Fedora Rawhide since November 2013.
See:
https://sourceware.org/ml/libc-alpha/2014-04/msg00179.html
Add a small library to print JSON values and use it to improve the
readability of the benchmark output and the readability of the
benchmark code.
ChangeLog:
2014-04-11 Will Newton <will.newton@linaro.org>
* benchtests/Makefile (extra-objs): Add json-lib.o.
(bench-func): Tidy up JSON output.
* benchtests/bench-skeleton.c: Include json-lib.h.
(main): Use JSON library functions to do output of
benchmark results.
* benchtests/bench-timing-type.c (main): Output the
timing type simply, leaving formatting to the user.
* benchtests/json-lib.c: New file.
* benchtests/json-lib.h: Likewise.
[BZ #15215] This unifies various pthread_once architecture-specific
implementations which were using the same algorithm with slightly different
implementations. It also adds missing memory barriers that are required for
correctness.
MALLOC_DEBUG is set optionally on the command line. Default the value
to zero if it is not set on the command line, and test its value
with #if rather than #ifdef. Verified the code is identical before
and after this change apart from line numbers.
ChangeLog:
2014-04-11 Will Newton <will.newton@linaro.org>
* malloc/malloc.c [!MALLOC_DEBUG]: #define MALLOC_DEBUG
to zero if it is not defined elsewhere. (mtrim): Test
the value of MALLOC_DEBUG with #if rather than #ifdef.
We have a single thread that runs a no-op initialization once and then
repeatedly runs checks of the initialization (i.e., an acquire load and
conditional jump) in a tight loop. This gives us, on average, the
best-case latency of pthread_once (the initialization is the
exactly-once slow path, and we're not looking at initialization-related
synchronization overheads in this case).
This patch saves and restores bound registers in symbol lookup for x86-64:
1. Branches without BND prefix clear bound registers.
2. x86-64 pass bounds in bound registers as specified in MPX psABI
extension on hjl/mpx/master branch at
https://github.com/hjl-tools/x86-64-psABIhttps://groups.google.com/forum/#!topic/x86-64-abi/KFsB0XTgWYc
Binutils has been updated to create an alternate PLT to add BND prefix
when branching to ld.so.
* config.h.in (HAVE_MPX_SUPPORT): New #undef.
* sysdeps/x86_64/configure.ac: Set HAVE_MPX_SUPPORT.
* sysdeps/x86_64/configure: Regenerated.
* sysdeps/x86_64/dl-trampoline.S (REGISTER_SAVE_AREA): New
macro.
(REGISTER_SAVE_RAX): Likewise.
(REGISTER_SAVE_RCX): Likewise.
(REGISTER_SAVE_RDX): Likewise.
(REGISTER_SAVE_RSI): Likewise.
(REGISTER_SAVE_RDI): Likewise.
(REGISTER_SAVE_R8): Likewise.
(REGISTER_SAVE_R9): Likewise.
(REGISTER_SAVE_BND0): Likewise.
(REGISTER_SAVE_BND1): Likewise.
(REGISTER_SAVE_BND2): Likewise.
(_dl_runtime_resolve): Use them. Save and restore Intel MPX
bound registers when calling _dl_fixup.
This patch defines _STRING_ARCH_unaligned to 0 on default bits/string.h
header to avoid undefined compiler warnings on platforms that do not
define it. It also make adjustments in code where tests checked if macro
existed or not.
pathconf(_PC_NAME_MAX) was implemented on top of statfs(). The 32bit
version therefore fails EOVERFLOW if the filesystem blockcount is
sufficiently large.
Most pathconf() queries use statvfs64(), which avoids this issue. This
patch modifies pathconf(_PC_NAME_MAX) to do likewise.
This patch moves the __PTHREAD_SPINS definition to arch specific header
since pthread_mutex_t layout is also arch specific. This leads to no
need to defining __PTHREAD_MUTEX_HAVE_ELISION and thus removing of the
undefined compiler warning.
This patch fixes some powerpc32 and powerpc64 builds with
--disable-multi-arch option along with different --with-cpu=powerN.
It cleanups the Implies directories by removing the multiarch
folder for non multiarch config and also fixing two assembly
implementations: powerpc64/power7/strncat.S that is calling the
wrong strlen; and power8/fpu/s_isnan.S that misses the hidden_def and
weak_alias directives.
Clean up string functions that do not have a version in gnulib on
the assumption that glibc is the canonical upstream copy of this
code. basename has a copy in gnulib but it is largely written to
handle Windows paths so merging it is not really viable. The changes
mostly consist of switching to ANSI function prototypes and removing
unused includes.
As many of these functions do not get built in a typical build due
to architecture optimized versions being used instead I built these
by hand to verify there were no build warnings and the code was
identical.
2014-04-07 Will Newton <will.newton@linaro.org>
* string/basename.c [HAVE_CONFIG_H]: Remove #ifdef and
and contents. [!_LIBC] Remove #ifndef and contents.
(basename): Use ANSI prototype. [_LIBC] Remove #idef.
* string/memccpy.c (__memccpy): Use ANSI prototype.
* string/memfrob.c (memfrob): Likewise.
* string/strcoll.c (STRCOLL): Likewise.
* string/strlen.c (strlen): Likewise.
* string/strtok.c (STRTOK): Likewise.
* string/strcat.c: Remove unused #include of memcopy.h.
(strcat): Use ANSI prototype.
* string/strchr.c: Remove unused #include of memcopy.h.
(strchr): Use ANSI prototype.
* string/strcmp.c: Remove unused #include of memcopy.h.
(strcmp): Use ANSI prototype.
* string/strcpy.c: Remove unused #include of memcopy.h.
(strcpy): Use ANSI prototype.
This patch makes the configure adds -D_CALL_ELF=1 when compiler does
not define _CALL_ELF (versions before powerpc64le support). It cleans
up compiler warnings on old compiler where _CALL_ELF is not defined
on powerpc64(be) builds.
It does by add a new config.make variable for configure-deduced
CPPFLAGS and accumulate into that (confix-extra-cppflags). It also
generalizes libc_extra_cflags so it accumulates in sysdeps configure
fragmenets.
This patch fixes the powerpc32 optimized nearbyint/nearbyintf bogus
results for FE_DOWNWARD rounding mode. This is due wrong instructions
sequence used in the rounding calculation (two subtractions instead of
adition and a subtraction).
Fixes BZ#16815.
If the user has requested automatic buffer creation, getline may create
it and not free things when an error occurs. That means the user is
always responsible for calling free() regardless of the return value.
The current documentation does not explicitly cover this which leaves it
slightly ambiguous to the reader. So clarify things.
URL: https://sourceware.org/bugzilla/show_bug.cgi?id=5666
The nested function referred to has gone away so remove the
comment. Also move the variable declaration down to where other
variables of a similar lifetime are declared for clarity.
2014-04-03 Will Newton <will.newton@linaro.org>
* elf/dl-lookup.c (do_lookup_x): Remove comment
referring to nested function and move variable
declarations down to before first use.
This patch fixes incorrect results from catan and catanh of certain
special inputs in round-downward mode (bug 16799), and incorrect
results of __ieee754_logf (+/-0) in round-downward mode (bug 16800)
that show up through catan/catanh when tested in all rounding modes,
but not directly in the testing for logf because the bug gets hidden
by the wrappers.
Both bugs involve a zero that should be +0 being -0 instead: one
computed as (1-x)*(1+x) in the catan/catanh case, and one as (x-x) in
the logf case. The fixes ensure positive zero is used. Testing of
catan and catanh in all rounding modes is duly enabled.
I expect there are various other bugs in special cases in __ieee754_*
functions that are normally hidden by the wrappers but would show up
for testing with -lieee (or in future with -fno-math-errno if we
replace -lieee and _LIB_VERSION with compile-time redirection to new
*_noerrno symbol names).
Tested x86_64 and x86 and ulps updated accordingly.
[BZ #16799]
[BZ #16800]
* math/s_catan.c (__catan): Avoid passing -0 denominator to atan2
with 0 numerator.
* math/s_catanf.c (__catanf): Likewise.
* math/s_catanh.c (__catanh): Likewise.
* math/s_catanhf.c (__catanhf): Likewise.
* math/s_catanhl.c (__catanhl): Likewise.
* math/s_catanl.c (__catanl): Likewise.
* sysdeps/ieee754/flt-32/e_logf.c (__ieee754_logf): Always divide
by positive zero when computing -Inf result.
* math/libm-test.inc (catan_test): Use ALL_RM_TEST.
(catanh_test): Likewise.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
This patch fixes bug 16789, incorrect sign of (real part) zero result
from clog and clog10 in round-downward mode, arising from that real
part being computed as 0 - 0. To ensure that an underflow exception
occurred, the code used an underflowing value (the next term in the
series for log1p) in arithmetic computing the real part of the result,
yielding the problematic 0 - 0 computation in some cases even when the
mathematical result would be small but positive. The patch changes
this code to use the math_force_eval approach to ensuring that an
underflowing computation actually occurs. Tests of clog and clog10
are enabled in all rounding modes.
Tested x86_64 and x86 and ulps updated accordingly.
[BZ #16789]
* math/s_clog.c (__clog): Use math_force_eval to ensure underflow
instead of using underflowing value in computing result.
* math/s_clog10.c (__clog10): Likewise.
* math/s_clog10f.c (__clog10f): Likewise.
* math/s_clog10l.c (__clog10l): Likewise.
* math/s_clogf.c (__clogf): Likewise.
* math/s_clogl.c (__clogl): Likewise.
* math/libm-test.inc (clog_test): Use ALL_RM_TEST.
(clog10_test): Likewise.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
Fix for values near a power of two, and some tidies.
[BZ #16739]
* sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c (__nextafterl): Correct
output when value is near a power of two. Use int64_t for lx and
remove casts. Use decimal rather than hex exponent constants.
Don't use long double multiplication when double will suffice.
* math/libm-test.inc (nextafter_test_data): Add tests.
* NEWS: Add 16739 and 16786 to bug list.
Without this flag it is possible that the compiler will optimize
away the calls to ffs/ffsll.
ChangeLog:
2014-04-01 Will Newton <will.newton@linaro.org>
* benchtests/Makefile (CFLAGS-bench-ffs.c): Add
-fno-builtin. (CFLAGS-bench-ffsll.c): Likewise.
At this point, we can only abort the process because we have already
switched credentials on other threads. Returning an error would still
leave the process in an inconsistent state.
The new xtest needs root privileges to run.
This patch fixes the default mode of scalb to set errno (bugs 6803 and
6804).
Previously, the _LIB_VERSION == _SVID_ mode would set errno but only
in some relevant cases, and with various peculiarities (such as errno
setting when an exact infinity or zero result arises with an argument
to scalb being an infinity). This patch leaves this mode
bug-compatible, while making the default mode set errno in accordance
with normal practice (so an exact infinity from an infinite argument
is not an error, and nor is an exact zero result). gen-libm-test.pl
is taught new notation such as ERRNO_PLUS_OFLOW to facilitate writing
the tests of errno setting for underflow / overflow in libm-test.inc.
Note that bug 6803 also covers scalbn and scalbln, but this patch only
addresses the scalb parts of that bug (along with the whole of bug
6804).
Tested x86_64 and x86.
[BZ #6803]
[BZ #6804]
* math/w_scalb.c (__scalb): For non-SVID mode, check result and
set errno as appropriate.
* math/w_scalbf.c (__scalbf): Likewise.
* math/w_scalbl.c (__scalbl): Likewise.
* math/gen-libm-test.pl (parse_args): Handle ERRNO_PLUS_OFLOW,
ERRNO_MINUS_OFLOW, ERRNO_PLUS_UFLOW and ERRNO_MINUS_UFLOW.
* math/libm-test.inc (scalb_test_data): Add errno expectations.
Add more NaN tests.
This patch fixes bug 16349, missing errno setting for atan2 underflow,
by adding appropriate checks to the existing wrappers. (As in other
cases, the __kernel_standard support for calling matherr is considered
to be for existing code expecting existing rules for what's considered
an error, even if those don't correspond to a general logical scheme
for what counts as what kind of error, so __set_errno calls are added
directly without any changes to __kernel_standard.)
Tested x86_64 and x86.
[BZ #16349]
* math/w_atan2.c: Include <errno.h>.
(__atan2): Set errno for result underflowing to zero.
* math/w_atan2f.c: Include <errno.h>.
(__atan2f): Set errno for result underflowing to zero.
* math/w_atan2l.c: Include <errno.h>.
(__atan2l): Set errno for result underflowing to zero.
* math/auto-libm-test-in: Don't allow missing errno for some atan2
tests.
* math/auto-libm-test-out: Regenerated.
This patch continues fixing __ASSUME_* issues in preparation for
moving to a 2.6.32 minimum kernel version by addressing assumptions on
robust mutex and PI futex support availability. Those assumptions are
bug 9894, but to be clear this patch does not address all the issues
from that bug about wrong version assumptions, only those still
applicable for --enable-kernel=2.6.32 or later (with the expectation
that the move to that minimum kernel will obsolete the other parts of
the bug). The patch is independent of
<https://sourceware.org/ml/libc-alpha/2014-03/msg00585.html>, my other
pending-review patch preparing for the kernel version change; the two
together complete all the changes I believe are needed in preparation
regarding any macro in sysdeps/unix/sysv/linux/kernel-features.h that
would be affected by such a change. (I have not checked the
correctness of macros whose conditions are unaffected by such a
change, or macros only defined in other kernel-features.h files.)
As discussed in that bug, robust mutexes and PI futexes need
futex_atomic_cmpxchg_inatomic to be implemented, in addition to
certain syscalls needed for robust mutexes (and
architecture-independent kernel pieces for all the features in
question). That is, as I understand it, they need
futex_atomic_cmpxchg_inatomic to *work* (not return an ENOSYS error).
The issues identified in my analysis relate to ARM, M68K, MicroBlaze,
MIPS and SPARC.
On ARM, whether futex_atomic_cmpxchg_inatomic works depends on the
kernel configuration. As of 3.13, the condition for *not* working is
CONFIG_CPU_USE_DOMAINS && CONFIG_SMP. As of 2.6.32 it was simply
CONFIG_SMP that meant the feature was not implemented. I don't know
if there are any circumstances in which we can say "we can assume a
userspace glibc binary built with these options will never run on a
kernel with the problematic configuration", but at least for now I'm
just undefining the relevant __ASSUME_* macros for ARM.
On M68K, two of the three macros are undefined for kernels before
3.10, but as far as I can see __ASSUME_FUTEX_LOCK_PI is in the same
group needing futex_atomic_cmpxchg_inatomic support and so should be
undefined as well.
On MicroBlaze the required support was added in 2.6.33.
On MIPS, the support depends on cpu_has_llsc in the kernel - that is,
actual hardware LL/SC support (GCC and glibc for MIPS GNU/Linux rely
on the instructions being supported in some way, but it may be kernel
emulation; futex_atomic_cmpxchg_inatomic doesn't work with that
emulation). The same condition as in GCC for indicating LL/SC support
may not be available is used for undefining the macros in glibc,
__mips == 1 || defined _MIPS_ARCH_R5900. (Maybe we could in fact
desupport MIPS processors without the hardware support in glibc.)
On SPARC, 32-bit kernels don't support futex_atomic_cmpxchg_inatomic;
__arch64__ || __sparc_v9__ is used as the condition for binaries that
won't run on 32-bit kernels.
This patch is not tested beyond the sanity check of an x86_64 build.
[BZ #9894]
* sysdeps/unix/sysv/linux/kernel-features.h
[__sparc__ && !__arch64__ && !__sparc_v9__]
(__ASSUME_SET_ROBUST_LIST): Do not define.
[__sparc__ && !__arch64__ && !__sparc_v9__]
(__ASSUME_FUTEX_LOCK_PI): Likewise.
[__sparc__ && !__arch64__ && !__sparc_v9__] (__ASSUME_REQUEUE_PI):
Likewise.
* sysdeps/unix/sysv/linux/arm/kernel-features.h
(__ASSUME_FUTEX_LOCK_PI): Undefine.
(__ASSUME_REQUEUE_PI): Likewise.
(__ASSUME_SET_ROBUST_LIST): Likewise.
* sysdeps/unix/sysv/linux/m68k/kernel-features.h
[__LINUX_KERNEL_VERSION < 0x030a00] (__ASSUME_FUTEX_LOCK_PI):
Undefine.
* sysdeps/unix/sysv/linux/microblaze/kernel-features.h
[__LINUX_KERNEL_VERSION < 0x020621] (__ASSUME_FUTEX_LOCK_PI):
Likewise.
[__LINUX_KERNEL_VERSION < 0x020621] (__ASSUME_REQUEUE_PI):
Likewise.
[__LINUX_KERNEL_VERSION < 0x020621] (__ASSUME_SET_ROBUST_LIST):
Likewise.
* sysdeps/unix/sysv/linux/mips/kernel-features.h
[__mips == 1 || _MIPS_ARCH_R5900] (__ASSUME_FUTEX_LOCK_PI):
Undefine.
[__mips == 1 || _MIPS_ARCH_R5900] (__ASSUME_REQUEUE_PI): Likewise.
[__mips == 1 || _MIPS_ARCH_R5900] (__ASSUME_SET_ROBUST_LIST):
Likewise.
Continuing the fixes for __ASSUME_* issues in preparation for moving
to a 2.6.32 minimum kernel version, this *untested* patch fixes bug
16648, the definition of __ASSUME_ATFCTS meaning that the futimesat
syscall is assumed for all MicroBlaze kernels despite not being
present until 2.6.33.
__ASSUME_ATFCTS controls conditionals relating to a lot of different
syscalls in Linux-specific code (fstatat64 faccessat fchmodat fchownat
futimesat newfstatat linkat mkdirat openat readlinkat renameat
symlinkat unlinkat mknodat), where whether newfstatat fstatat64
futimesat are used depends on the architecture, as well as controlling
whether openat64_not_cancel_3 is expected to work in
sysdeps/posix/getcwd.c. The assumptions are all OK as of 2.6.32
except for this MicroBlaze case, and it's generally desirable to get
rid of as many of the __ASSUME_ATFCTS conditionals as possible, to
simplify the code (the fallbacks include potential unbounded dynamic
stack allocations). Thus, rather than the simplest approach of
undefining __ASSUME_ATFCTS for older kernels on MicroBlaze, this patch
takes the approach of using the linux-generic implementation of
futimesat for MicroBlaze kernels before 2.6.33 (all such kernels have
the utimensat syscall).
[BZ #16648]
* sysdeps/unix/sysv/linux/microblaze/kernel-features.h
[__LINUX_KERNEL_VERSION >= 0x020621] (__ASSUME_FUTIMESAT): Define.
* sysdeps/unix/sysv/linux/microblaze/futimesat.c: New file.
Add benchtests for ffs and ffsll. There is no benchtest for ffsl as
it is identical to one of the other functions.
2014-03-31 Will Newton <will.newton@linaro.org>
* benchtests/Makefile (bench): Add ffs and ffsll to list
of tests.
* benchtests/ffs-inputs: New file.
* benchtests/ffsll-inputs: Likewise.
This patch fixes bug 16770, spurious "invalid" exceptions from scalb
when testing whether the second argument is an integer, by inserting
appropriate range checks to determine whether a cast to int is safe.
(Note that invalid_fn is a function that handles both nonintegers and
large integers, distinguishing them reliably using functions such as
__rint; note also that there are no issues with scalb needing to avoid
spurious "inexact" exceptions - it's an old-POSIX XSI function, not a
standard C function bound to an IEEE 754 operation - although the
return value is still fully determined.)
Tested x86_64 and x86.
[BZ #16770]
* math/e_scalb.c (__ieee754_scalb): Check second argument is not
too large before casting to int.
* math/e_scalbf.c (__ieee754_scalbf): Likewise.
* math/e_scalbl.c (__ieee754_scalbl): Likewise.
* math/libm-test.inc (scalb_test_data): Add more tests.
This patch adds an option to get detailed benchmark output for
functions. Invoking the benchmark with 'make DETAILED=1 bench' causes
each benchmark program to store a mean execution time for each input
it works on. This is useful to give a more comprehensive picture of
performance of functions compared to just the single mean figure.
This patch changes the output format of the main benchmark output file
(bench.out) to an extensible format. I chose JSON over XML because in
addition to being extensible, it is also not too verbose.
Additionally it has good support in python.
The significant change I have made in terms of functionality is to put
timing information as an attribute in JSON instead of a string and to
do that, there is a separate program that prints out a JSON snippet
mentioning the type of timing (hp_timing or clock_gettime). The mean
timing has now changed from iterations per unit to actual timing per
iteration.
This patch fixes the imaginary part of clog10 (-0 +/- 0i), which
should be +/-pi / log(10) by analogy with clog (the functions were
wrongly returning a result with imaginary part +/-pi, same as for
clog, and the tests matched the incorrect result, though both
functions and tests were correct for the similar case of clog10 (-inf
+/- 0i)). Tested x86_64 and x86.
[BZ #16362]
* math/s_clog10.c (M_PI_LOG10E): New macro.
(__clog10): Use M_PI_LOG10E instead of M_PI when real and
imaginary parts are 0.
* math/s_clog10f.c (M_PI_LOG10Ef): New macro.
(__clog10f): Use M_PI_LOG10Ef instead of M_PI when real and
imaginary parts are 0.
* math/s_clog10l.c (M_PI_LOG10El): New macro.
(__clog10l): Use M_PI_LOG10El instead of M_PIl when real and
imaginary parts are 0.
* math/libm-test.inc (clog10_test_data): Update expected results
for when real and imaginary parts are 0.
This patch fixes bug 16348, spurious underflows from x86/x86_64 expl
on arguments close to 0. These implementations effectively use expm1
(on the fractional part of the argument) internally, so resulting in
spurious underflows when the result is very close to 1. For arguments
small enough that the round-to-nearest correct result is 1, this patch
uses 1+x instead.
These implementations are also used for exp10l and so the patch fixes
similar issues there (the 0x1p-67 threshold being small enough to be
correct for exp10l as well as expl). But because of spurious
underflows in other exp10 implementations (bug 16560), the tests
aren't added for exp10 at this point - they can be added when the
other exp10 parts of that bug are fixed.
Tested x86_64 and x86; no ulps updates needed.
[BZ #16348]
* sysdeps/i386/fpu/e_expl.S (IEEE754_EXPL) [!USE_AS_EXPM1L]: Use
1+x for argument with exponent below -67.
* sysdeps/x86_64/fpu/e_expl.S (IEEE754_EXPL) [!USE_AS_EXPM1L]:
Likewise.
* math/auto-libm-test-in: Add more tests of exp.
* math/auto-libm-test-out: Regenerated.
getnetgrent is supposed to return NULL for values that are wildcards
in the (host, user, domain) triplet. This works correctly with nscd
disabled, but with it enabled, it returns a blank ("") instead of a
NULL. This is easily seen with the output of `getent netgroup foonet`
for a netgroup foonet defined as follows in /etc/netgroup:
foonet (,foo,)
The output with nscd disabled is:
foonet ( ,foo,)
while with nscd enabled, it is:
foonet (,foo,)
The extra space with nscd disabled is due to the fact that `getent
netgroup` adds it if the return value from getnetgrent is NULL for
either host or user.
Calls to stpcpy from nscd netgroups code will have overlapping source
and destination when all three values in the returned triplet are
non-NULL and in the expected (host,user,domain) order. This is seen
in valgrind as:
==3181== Source and destination overlap in stpcpy(0x19973b48, 0x19973b48)
==3181== at 0x4C2F30A: stpcpy (in /usr/lib64/valgrind/vgpreload_memcheck-amd64-linux.so)
==3181== by 0x12567A: addgetnetgrentX (string3.h:111)
==3181== by 0x12722D: addgetnetgrent (netgroupcache.c:665)
==3181== by 0x11114C: nscd_run_worker (connections.c:1338)
==3181== by 0x4E3C102: start_thread (pthread_create.c:309)
==3181== by 0x59B81AC: clone (clone.S:111)
==3181==
Fix this by using memmove instead of stpcpy.
When profiling programs with lock problems with perf record -g dwarf,
libunwind can currently not backtrace through the futex and unlock
functions in pthread. This is because they use out of line sections,
and those are not correctly described in dwarf2 (I believe needs
dwarf3 or 4).
This patch first removes the out of line sections. They only save a
single jump, but cause a lot of pain. Then it converts the now inline
lock code to use the now standard gas .cfi_* commands.
With these changes libunwind/perf can backtrace through the futex
functions now.
Longer term it would be likely better to just use C futex() functions
on x86 like all the other architectures. This would clean the code up
even more.
nscd works correctly when the request in innetgr is a wildcard,
i.e. when one or more of host, user or domain parameters is NULL.
However, it does not work when the the triplet in the netgroup
definition has a wildcard. This is easy to reproduce for a triplet
defined as follows:
foonet (,foo,)
Here, an innetgr call that looks like this:
innetgr ("foonet", "foohost", "foo", NULL);
should succeed and so should:
innetgr ("foonet", NULL, "foo", "foodomain");
It does succeed with nscd disabled, but not with nscd enabled. This
fix adds this additional check for all three parts of the triplet so
that it gives the correct result.
[BZ #16758]
* nscd/netgroupcache.c (addinnetgrX): Succeed if triplet has
blank values.
Bug 16198 is x86_64 fegetenv wrongly masking exceptions for which
traps are enabled, because that's a side-effect of the fnstenv
instruction. This patch fixes it to use fldenv immediately after
fnstenv, like the i386 version. Tested x86_64 and x86.
[BZ #16198]
* sysdeps/x86_64/fpu/fegetenv.c (fegetenv): Use fldenv after
fnstenv.
* math/test-fenv-preserve.c: New file.
* math/Makefile (tests): Add test-fenv-preserve.
This benchmark can take longer than the default 2 seconds on slower
platforms, so increase it to 10 seconds.
ChangeLog:
2014-03-26 Will Newton <will.newton@linaro.org>
* benchtests/bench-strtod.c (TIMEOUT): Define to 10.
GNU_IFUNC are shown by readelf in 'Relocation section' value as
"symbol()" instead of expected hexadecimal value. This causes the
check-localplt script to ignore potential PLT stub begin generated
by wrong IFUNC usage. This patch changes the localplt script to
emit such PLT cases.
gen-auto-libm-tests presently allows but does not require underflow
exceptions for results with magnitude in the range (greatest
subnormal, least normal].
In some cases, the magnitude of the exact result is very slightly
above the least normal, but rounding in the implementation results in
it effectively computing an infinite-precision result that is slightly
below the least normal, so raising an underflow exception. This is in
accordance with the documented accuracy goals, but results in
testsuite failures.
This patch changes the logic to allow underflows when the mathematical
result is up to 0.5ulp above the least normal (so in any case where
the round-to-nearest result is the least normal). Ideally underflows
in all these cases would be accepted only when an underflow with the
actual result is consistent with the rounding mode (in FE_TOWARDZERO
mode, a return value of the least normal implies that the
infinite-precision result did not underflow so there should be no
underflow exception, for example), so as to match the documented goals
more precisely - whereas at present the tests for exceptions are
completely independent of the tests of the returned values. (The same
applies to overflow exceptions as well - they too should be checked
for consistency with the result, as in FE_TOWARDZERO mode a result
1ulp below the largest finite value should be inconsistent with an
overflow exception and cause a failure with overflow rather than
simply being considered a 1ulp error when overflow is expected.) But
the present patch at least deals with the cases causing spurious
failures so that (a) certain existing tests no longer need to be
marked as having spurious exceptions (such markings in
auto-libm-test-in end up applying to more cases than just those they
are needed for) and (b) log1p can be tested in all rounding modes
without introducing more such failures. This patch duly moves tests
of log1p to ALL_RM_TEST.
Tested x86_64 and x86 and ulps updated accordingly.
[BZ #16357]
[BZ #16599]
* math/gen-auto-libm-tests.c (fp_format_desc): Add field
min_plus_half.
(fp_formats): Update initializers.
(init_fp_formats): Initialize new field.
(output_for_one_input_case): Allow underflow for results up to
min_plus_half.
* math/libm-test.inc (log1p_test): Use ALL_RM_TEST.
* math/auto-libm-test-in: Don't mark some underflows from asin and
atanh as spurious.
* math/auto-libm-test-out: Regenerated.
* sysdeps/i386/fpu/libm-test-ulps: Update.
* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
My recent exp patch introduced warnings about implicit __isinf
declarations in exp because e_exp.c didn't include <math.h>. This
patch fixes this. Because <math.h> can't be included after
<math_private.h> (because of macro definitions of __nan*), it was
necessary to put an include in sysdeps/x86_64/fpu/multiarch/e_exp.c as
well.
Tested x86_64.
* sysdeps/ieee754/dbl-64/e_exp.c: Include <math.h>.
* sysdeps/x86_64/fpu/multiarch/e_exp.c
[HAVE_FMA4_SUPPORT || HAVE_AVX_SUPPORT]: Likewise.
An application that erroneously tries to repeatedly dlopen("a.out", ...)
may hit assertion failure:
Inconsistency detected by ld.so: dl-tls.c: 474: _dl_allocate_tls_init:
Assertion `listp != ((void *)0)' failed!
dlopen() actually fails with "./a.out: cannot dynamically load executable",
but it does so after incrementing dl_tls_max_dtv_idx.
Once we run out of TLS_SLOTINFO_SURPLUS (62), we exit with above assertion
failure.
2014-03-24 Paul Pluzhnikov <ppluzhnikov@google.com>
[BZ #16634]
* elf/dl-load.c (open_verify): Add mode parameter.
Error early when ET_EXEC and mode does not have __RTLD_OPENEXEC.
(open_path): Change from boolean 'secure' to complete flag 'mode'
(_dl_map_object): Adjust.
* elf/Makefile (tests): Add tst-dlopen-aout.
* elf/tst-dlopen-aout.c: New test.
Roland McGrath
Steve Ellcey
Andreas Schwab
Ondřej Bílka
Aurelien Jarno
Vidya Ranganathan
Adhemerval Zanella
Adam Conrad
Allan McRae
Carlos O'Donell
Andreas Schwab
David S. Miller
Will Newton
Julian Brown
Joseph Myers
Yang Yingliang
Siddhesh Poyarekar
Stefan Liebler
Carlos O'Donell
Wilco Dijkstra
H.J. Lu
Ian Bolton
Yufeng Zhang
Venkataramanan Kumar
Sihai Yao
Wilco
Richard Henderson
Marcus Shawcroft
Igor Zamyatin
Samuel Thibault
Alan Modra
Paul Pluzhnikov
Chris Metcalf
Torvald Riegel
Joseph Anthony Pasquale Holsten
Peter TB Brett
David Svoboda
Florian Weimer
Andi Kleen