The generic implementation is shows only slight worse performance:
POWER10 reciprocal-throughput latency
master 8.28478 13.7253
new hypot 7.21945 13.1933
POWER9 reciprocal-throughput latency
master 13.4024 14.0967
new hypot 14.8479 15.8061
POWER8 reciprocal-throughput latency
master 15.5767 16.8885
new hypot 16.5371 18.4057
One way to improve might to make gcc generate xsmaxdp/xsmindp for
fmax/fmin (it onl does for -ffast-math, clang does for default
options).
Checked on powerpc64-linux-gnu (power8) and powerpc64le-linux-gnu
(power9).
The generic hypotf is slight slower, mostly due the tricks the assembly
does to optimize the isinf/isnan/issignaling. The generic hypot is way
slower, since the optimized implementation uses the i386 default
excessive precision to issue the operation directly. A similar
implementation is provided instead of using the generic implementation:
Checked on i686-linux-gnu.
This implementation is based on 'An Improved Algorithm for hypot(a,b)'
by Carlos F. Borges [1] using the MyHypot3 with the following changes:
- Handle qNaN and sNaN.
- Tune the 'widely varying operands' to avoid spurious underflow
due the multiplication and fix the return value for upwards
rounding mode.
- Handle required underflow exception for subnormal results.
The main advantage of the new algorithm is its precision. With a
random 1e9 input pairs in the range of [LDBL_MIN, LDBL_MAX], glibc
current implementation shows around 0.05% results with an error of
1 ulp (453266 results) while the new implementation only shows
0.0001% of total (1280).
Checked on aarch64-linux-gnu and x86_64-linux-gnu.
[1] https://arxiv.org/pdf/1904.09481.pdf
This implementation is based on 'An Improved Algorithm for hypot(a,b)'
by Carlos F. Borges [1] using the MyHypot3 with the following changes:
- Handle qNaN and sNaN.
- Tune the 'widely varying operands' to avoid spurious underflow
due the multiplication and fix the return value for upwards
rounding mode.
- Handle required underflow exception for subnormal results.
The main advantage of the new algorithm is its precision. With a
random 1e8 input pairs in the range of [LDBL_MIN, LDBL_MAX], glibc
current implementation shows around 0.02% results with an error of
1 ulp (23158 results) while the new implementation only shows
0.0001% of total (111).
[1] https://arxiv.org/pdf/1904.09481.pdf
Improve hypot performance significantly by using fma when available. The
fma version has twice the throughput of the previous version and 70% of
the latency. The non-fma version has 30% higher throughput and 10%
higher latency.
Max ULP error is 0.949 with fma and 0.792 without fma.
Passes GLIBC testsuite.
This implementation is based on the 'An Improved Algorithm for
hypot(a,b)' by Carlos F. Borges [1] using the MyHypot3 with the
following changes:
- Handle qNaN and sNaN.
- Tune the 'widely varying operands' to avoid spurious underflow
due the multiplication and fix the return value for upwards
rounding mode.
- Handle required underflow exception for denormal results.
The main advantage of the new algorithm is its precision: with a
random 1e9 input pairs in the range of [DBL_MIN, DBL_MAX], glibc
current implementation shows around 0.34% results with an error of
1 ulp (3424869 results) while the new implementation only shows
0.002% of total (18851).
The performance result are also only slight worse than current
implementation. On x86_64 (Ryzen 5900X) with gcc 12:
Before:
"hypot": {
"workload-random": {
"duration": 3.73319e+09,
"iterations": 1.12e+08,
"reciprocal-throughput": 22.8737,
"latency": 43.7904,
"max-throughput": 4.37184e+07,
"min-throughput": 2.28361e+07
}
}
After:
"hypot": {
"workload-random": {
"duration": 3.7597e+09,
"iterations": 9.8e+07,
"reciprocal-throughput": 23.7547,
"latency": 52.9739,
"max-throughput": 4.2097e+07,
"min-throughput": 1.88772e+07
}
}
Co-Authored-By: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Checked on x86_64-linux-gnu and aarch64-linux-gnu.
[1] https://arxiv.org/pdf/1904.09481.pdf
Use a more optimized comparison for check for NaN and infinite and
add an inlined issignaling implementation for float. With gcc it
results in 2 FP comparisons.
The file Copyright is also changed to use GPL, the implementation was
completely changed by 7c10fd3515 to use double precision instead of
scaling and this change removes all the GET_FLOAT_WORD usage.
Checked on x86_64-linux-gnu.
Replace non-UTF-8 and non-ASCII characters in comments with their UTF-8
equivalents so that files don't end up with mixed encodings. With this,
all files (except tests that actually test different encodings) have a
single encoding.
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Build glibc programs and tests as PIE by default and enable static-pie
automatically if the architecture and toolchain supports it.
Also add a new configuration option --disable-default-pie to prevent
building programs as PIE.
Only the following architectures now have PIE disabled by default
because they do not work at the moment. hppa, ia64, alpha and csky
don't work because the linker is unable to handle a pcrel relocation
generated from PIE objects. The microblaze compiler is currently
failing with an ICE. GNU hurd tries to enable static-pie, which does
not work and hence fails. All these targets have default PIE disabled
at the moment and I have left it to the target maintainers to enable PIE
on their targets.
build-many-glibcs runs clean for all targets. I also tested x86_64 on
Fedora and Ubuntu, to verify that the default build as well as
--disable-default-pie work as expected with both system toolchains.
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Remove the LD_PREFER_MAP_32BIT_EXEC environment variable support since
the first PT_LOAD segment is no longer executable due to defaulting to
-z separate-code.
This fixes [BZ #28656].
Reviewed-by: Florian Weimer <fweimer@redhat.com>
When PT_LOAD segment alignment > the page size, allocate enough space to
ensure that the segment can be properly aligned. This change helps code
segments use huge pages become simple and available.
This fixes [BZ #28676].
Signed-off-by: Xu Yu <xuyu@linux.alibaba.com>
Signed-off-by: Rongwei Wang <rongwei.wang@linux.alibaba.com>
This makes ld.so features such as --preload, --audit,
and --list-diagnostics more accessible to end users because they
do not need to know the ABI name of the dynamic loader.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
Without the bar_ctor_finish barrier, it was possible that thread2
re-locked user_lock before ctor had a chance to lock it. ctor then
blocked in its locking operation, xdlopen from the main thread
did not return, and thread2 was stuck waiting in bar_dtor:
thread 1: started.
thread 2: started.
thread 2: locked user_lock.
constructor started: 0.
thread 1: in ctor: started.
thread 3: started.
thread 3: done.
thread 2: unlocked user_lock.
thread 2: locked user_lock.
Fixes the test in commit 83b5323261
("elf: Avoid deadlock between pthread_create and ctors [BZ #28357]").
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
TLS_INIT_TCB_ALIGN is not actually used. TLS_TCB_ALIGN was likely
introduced to support a configuration where the thread pointer
has not the same alignment as THREAD_SELF. Only ia64 seems to use
that, but for the stack/pointer guard, not for storing tcbhead_t.
Some ports use TLS_TCB_OFFSET and TLS_PRE_TCB_SIZE to shift
the thread pointer, potentially landing in a different residue class
modulo the alignment, but the changes should not impact that.
In general, given that TLS variables have their own alignment
requirements, having different alignment for the (unshifted) thread
pointer and struct pthread would potentially result in dynamic
offsets, leading to more complexity.
hppa had different values before: __alignof__ (tcbhead_t), which
seems to be 4, and __alignof__ (struct pthread), which was 8
(old default) and is now 32. However, it defines THREAD_SELF as:
/* Return the thread descriptor for the current thread. */
# define THREAD_SELF \
({ struct pthread *__self; \
__self = __get_cr27(); \
__self - 1; \
})
So the thread pointer points after struct pthread (hence __self - 1),
and they have to have the same alignment on hppa as well.
Similarly, on ia64, the definitions were different. We have:
# define TLS_PRE_TCB_SIZE \
(sizeof (struct pthread) \
+ (PTHREAD_STRUCT_END_PADDING < 2 * sizeof (uintptr_t) \
? ((2 * sizeof (uintptr_t) + __alignof__ (struct pthread) - 1) \
& ~(__alignof__ (struct pthread) - 1)) \
: 0))
# define THREAD_SELF \
((struct pthread *) ((char *) __thread_self - TLS_PRE_TCB_SIZE))
And TLS_PRE_TCB_SIZE is a multiple of the struct pthread alignment
(confirmed by the new _Static_assert in sysdeps/ia64/libc-tls.c).
On m68k, we have a larger gap between tcbhead_t and struct pthread.
But as far as I can tell, the port is fine with that. The definition
of TCB_OFFSET is sufficient to handle the shifted TCB scenario.
This fixes commit 23c77f6018
("nptl: Increase default TCB alignment to 32").
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
The relationship between the thread pointer and the rseq area
is made explicit. The constant offset can be used by JIT compilers
to optimize rseq access (e.g., for really fast sched_getcpu).
Extensibility is provided through __rseq_size and __rseq_flags.
(In the future, the kernel could request a different rseq size
via the auxiliary vector.)
Co-Authored-By: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
This tunable allows applications to register the rseq area instead
of glibc.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
The rseq area is placed directly into struct pthread. rseq
registration failure is not treated as an error, so it is possible
that threads run with inconsistent registration status.
<sys/rseq.h> is not yet installed as a public header.
Co-Authored-By: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
<tls.h> already contains a definition that is quite similar,
but it is not consistent across architectures.
Only architectures for which rseq support is added are covered.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
commit d585ba47fc
Author: Noah Goldstein <goldstein.w.n@gmail.com>
Date: Mon Nov 1 00:49:48 2021 -0500
string: Make tests birdirectional test-memcpy.c
Add tests that had src/dst non 4-byte aligned. Since src/dst are
initialized/compared as uint32_t type which is 4-byte aligned this can
break on some targets.
Fix the issue by specifying a new non-aligned 4-byte
`unaligned_uint32_t` for src/dst.
Another alternative is to rely on memcpy/memcmp for
initializing/testing src/dst. Using memcpy for initializing in memcpy
tests, however, could lead to future bugs.
Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
localedef currently blindly trust the archive header. When passed an
archive file with the wrong endianess, this leads to a segmentation
fault:
$ localedef --big-endian --list-archive /usr/lib/locale/locale-archive
Segmentation fault (core dumped)
When passed non-archive files, asserts are reported on the best case,
but sometimes it can lead to a segmentation fault:
$ localedef --list-archive /bin/true
localedef: programs/locarchive.c:1643: show_archive_content: Assertion `used < GET (head->namehash_used)' failed.
Aborted (core dumped)
$ localedef --list-archive /usr/lib/locale/C.utf8/LC_COLLATE
Segmentation fault (core dumped)
This patch improves the user experience by looking at the magic value,
which is always written, but never checked. It should still be possible
to trigger a segmentation fault with crafted files, but this already
catch many cases.
The function was renamed to __atomic_wide_counter_load_relaxed
in commit 8bd336a00a ("nptl: Extract
<bits/atomic_wide_counter.h> from pthread_cond_common.c").
Current binutils defines __executable_start as the lowest text
address, so using the entry point address as a fallback is no
longer necessary. As a result, overriding <entry.h> is only
necessary if the entry point is not called _start.
The previous approach to define __ASSEMBLY__ to suppress the
declaration breaks if headers included by <entry.h> are not
compatible with __ASSEMBLY__. This happens with rseq integration
because it is necessary to include kernel headers in more places.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Programs without dynamic dependencies and without a program
interpreter are now run via execve.
Previously, the dynamic linker either crashed while attempting to
read a non-existing dynamic segment (looking for DT_AUDIT/DT_DEPAUDIT
data), or the self-relocated in the static PIE executable crashed
because the outer dynamic linker had already applied RELRO protection.
<dl-execve.h> is needed because execve is not available in the
dynamic loader on Hurd.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
On Ice Lake and Tiger Lake laptops, some test programs timeout when there
are 3 "make check -j8" runs in parallel. Add --with-timeoutfactor=NUM to
specify an integer to scale the timeout of test programs, which can be
overriden by TIMEOUTFACTOR environment variable.
Reviewed-by: Florian Weimer <fweimer@redhat.com>
Must use notl %edi here as lower bits are for CHAR comparisons
potentially out of range thus can be 0 without indicating mismatch.
This fixes BZ #28646.
Co-Authored-By: H.J. Lu <hjl.tools@gmail.com>
rseq support will use a 32-byte aligned field in struct pthread,
so the whole struct needs to have at least that alignment.
nptl/tst-tls3mod.c uses TCB_ALIGNMENT, therefore include <descr.h>
to obtain the fallback definition.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
v2 is a complete rewrite of the A64FX memcpy. Performance is improved
by streamlining the code, aligning all large copies and using a single
unrolled loop for all sizes. The code size for memcpy and memmove goes
down from 1796 bytes to 868 bytes. Performance is better in all cases:
bench-memcpy-random is 2.3% faster overall, bench-memcpy-large is ~33%
faster for large sizes, bench-memcpy-walk is 25% faster for small sizes
and 20% for the largest sizes. The geomean of all tests in bench-memcpy
is 5.1% faster, and total time is reduced by 4%.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Rewrite memcmp to improve performance. On small and medium inputs performance
is 10-20% better. Large inputs use a SIMD loop processing 64 bytes per
iteration, which is 30-50% faster depending on the size.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
The syscall pipe2 was added in linux 2.6.27 and glibc requires linux
3.2.0. The patch removes the arch-specific implementation for alpha,
ia64, mips, sh, and sparc which requires a different kernel ABI
than the usual one.
Checked on x86_64-linux-gnu and with a build for the affected ABIs.
Variadic function calls in syscalls.list does not work for all ABIs
(for instance where the argument are passed on the stack instead of
registers) and might have underlying issues depending of the variadic
type (for instance if a 64-bit argument is used).
Checked on x86_64-linux-gnu.
The LFS prlimit64 requires a arch-specific implementation in
syscalls.list. Instead add a generic one that handles the
required symbol alias for __RLIM_T_MATCHES_RLIM64_T.
HPPA is the only outlier which requires a different default
symbol.
Checked on x86_64-linux-gnu and with build for the affected ABIs.
