Remove generic tlsdesc code related to lazy tlsdesc processing since
lazy tlsdesc relocation is no longer supported. This includes removing
GL(dl_load_lock) from _dl_make_tlsdesc_dynamic which is only called at
load time when that lock is already held.
Added a documentation comment too.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
No bug. This commit optimizes strlen-avx2.S. The optimizations are
mostly small things but they add up to roughly 10-30% performance
improvement for strlen. The results for strnlen are bit more
ambiguous. test-strlen, test-strnlen, test-wcslen, and test-wcsnlen
are all passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
No bug. This commit optimizes strlen-evex.S. The
optimizations are mostly small things but they add up to roughly
10-30% performance improvement for strlen. The results for strnlen are
bit more ambiguous. test-strlen, test-strnlen, test-wcslen, and
test-wcsnlen are all passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
No bug. This commit adds optimized cased for less_vec memset case that
uses the avx512vl/avx512bw mask store avoiding the excessive
branches. test-memset and test-wmemset are passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Since strchr-avx2.S updated by
commit 1f745ecc21
Author: noah <goldstein.w.n@gmail.com>
Date: Wed Feb 3 00:38:59 2021 -0500
x86-64: Refactor and improve performance of strchr-avx2.S
uses sarx:
c4 e2 72 f7 c0 sarx %ecx,%eax,%eax
for strchr-avx2 family functions, require BMI2 in ifunc-impl-list.c and
ifunc-avx2.h.
Since __strlen_evex and __strnlen_evex added by
commit 1fd8c163a8
Author: H.J. Lu <hjl.tools@gmail.com>
Date: Fri Mar 5 06:24:52 2021 -0800
x86-64: Add ifunc-avx2.h functions with 256-bit EVEX
use sarx:
c4 e2 6a f7 c0 sarx %edx,%eax,%eax
require BMI2 for __strlen_evex and __strnlen_evex in ifunc-impl-list.c.
ifunc-avx2.h already requires BMI2 for EVEX implementation.
No Bug. This commit updates the large memcpy case (no overlap). The
update is to perform memcpy on either 2 or 4 contiguous pages at
once. This 1) helps to alleviate the affects of false memory aliasing
when destination and source have a close 4k alignment and 2) In most
cases and for most DRAM units is a modestly more efficient access
pattern. These changes are a clear performance improvement for
VEC_SIZE =16/32, though more ambiguous for VEC_SIZE=64. test-memcpy,
test-memccpy, test-mempcpy, test-memmove, and tst-memmove-overflow all
pass.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Some registers that can be clobbered by the kernel during a syscall are not
listed on the clobbers list in sysdeps/unix/sysv/linux/powerpc/sysdep.h.
For syscalls using sc:
- XER is zeroed by the kernel on exit
For syscalls using scv:
- XER is zeroed by the kernel on exit
- Different from the sc case, most CR fields can be clobbered (according to
the ELF ABI and the Linux kernel's syscall ABI for powerpc
(linux/Documentation/powerpc/syscall64-abi.rst)
The same should apply to vsyscalls, which effectively execute a function call
but are not currently adding these registers as clobbers either.
These are likely not causing issues today, but they should be added to the
clobbers list just in case things change on the kernel side in the future.
Reported-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Reviewed-by: Raphael M Zinsly <rzinsly@linux.ibm.com>
MSG_NOSIGNAL was added on POSIX 2008 and Hurd seems to support it.
The SIGPIPE handling also makes the implementation not thread-safe
(due the sigaction usage).
Checked on x86_64-linux-gnu.
Now that libsupport abstract Linux possible missing support (either
due FS limitation that can't handle 64 bit timestamp or architectures
that do not handle values larger than unsigned 32 bit values) the
tests can be turned generic.
Checked on x86_64-linux-gnu and i686-linux-gnu. I also built the
tests for i686-gnu.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Lazy tlsdesc relocation is racy because the static tls optimization and
tlsdesc management operations are done without holding the dlopen lock.
This similar to the commit b7cf203b5c
for aarch64, but it fixes a different race: bug 27137.
On i386 the code is a bit more complicated than on x86_64 because both
rel and rela relocs are supported.
Lazy tlsdesc relocation is racy because the static tls optimization and
tlsdesc management operations are done without holding the dlopen lock.
This similar to the commit b7cf203b5c
for aarch64, but it fixes a different race: bug 27137.
Another issue is that ld auditing ignores DT_BIND_NOW and thus tries to
relocate tlsdesc lazily, but that does not work in a BIND_NOW module
due to missing DT_TLSDESC_PLT. Unconditionally relocating tlsdesc at
load time fixes this bug 27721 too.
The value of PR_TAGGED_ADDR_ENABLE was incorrect in the installed
headers and the prctl command macros were missing that are needed
for it to be useful (PR_SET_TAGGED_ADDR_CTRL). Linux headers have
the definitions since 5.4 so it's widely available, we don't need
to repeat these definitions. The remaining definitions are from
Linux 5.10.
To build glibc with --enable-memory-tagging, Linux 5.4 headers and
binutils 2.33.1 or newer is needed.
Reviewed-by: DJ Delorie <dj@redhat.com>
Since Linux 4.13, kernel limits the maximum command line arguments
length to 6 MiB [1]. Normally the limit is still quarter of the maximum
stack size but if that limit exceeds 6 MiB it's clamped down.
glibc's __sysconf implementation for Linux platform is not aware of
this limitation and for stack sizes of over 24 MiB it returns higher
ARG_MAX than Linux will actually accept. This can be verified by
executing the following application on Linux 4.13 or newer:
#include <stdio.h>
#include <string.h>
#include <sys/resource.h>
#include <sys/time.h>
#include <unistd.h>
int main(void) {
const struct rlimit rlim = { 40 * 1024 * 1024,
40 * 1024 * 1024 };
if (setrlimit(RLIMIT_STACK, &rlim) < 0) {
perror("setrlimit: RLIMIT_STACK");
return 1;
}
printf("ARG_MAX : %8ld\n", sysconf(_SC_ARG_MAX));
printf("63 * 100 KiB: %8ld\n", 63L * 100 * 1024);
printf("6 MiB : %8ld\n", 6L * 1024 * 1024);
char str[100 * 1024], *argv[64], *envp[1];
memset(&str, 'A', sizeof str);
str[sizeof str - 1] = '\0';
for (size_t i = 0; i < sizeof argv / sizeof *argv - 1; ++i) {
argv[i] = str;
}
argv[sizeof argv / sizeof *argv - 1] = envp[0] = 0;
execve("/bin/true", argv, envp);
perror("execve");
return 1;
}
On affected systems the program will report ARG_MAX as 10 MiB but
despite that executing /bin/true with a bit over 6 MiB of command line
arguments will fail with E2BIG error. Expected result is that ARG_MAX
is reported as 6 MiB.
Update the __sysconf function to clamp ARG_MAX value to 6 MiB if it
would otherwise exceed it. This resolves bug #25305 which was market
WONTFIX as suggested solution was to cap ARG_MAX at 128 KiB.
As an aside and point of comparison, bionic (a libc implementation for
Android systems) decided to resolve this issue by always returning 128
KiB ignoring any potential xargs regressions [2].
On older kernels this results in returning overly conservative value
but that's a safer option than being aggressive and returning invalid
value on recent systems. It's also worth noting that at this point
all supported Linux releases have the 6 MiB barrier so only someone
running an unsupported kernel version would get incorrectly truncated
result.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
[1] See https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=da029c11e6b12f321f36dac8771e833b65cec962
[2] See baed51ee3a
The commit 2433d39b69, which added time64 support to select, changed
the function to use __NR_pselect6 (or __NR_pelect6_time64) on all
architectures. However, on architectures where the symbol was
implemented with __NR_select the kernel normalizes the passed timeout
instead of return EINVAL. For instance, the input timeval
{ 0, 5000000 } is interpreted as { 5, 0 }.
And as indicated by BZ #27651, this semantic seems to be expected
and changing it results in some performance issues (most likely
the program does not check the return code and keeps issuing
select with unormalized tv_usec argument).
To avoid a different semantic depending whether which syscall the
architecture used to issue, select now always normalize the timeout
input. This is a slight change for some ABIs (for instance aarch64).
Checked on x86_64-linux-gnu and i686-linux-gnu.
An incorrect check in __longjmp_chk could fail on valid code causing
FAIL: debug/tst-longjmp_chk2
The original check was
altstack_sp + altstack_size - setjmp_sp > altstack_size
i.e. sp at setjmp was outside of the altstack range. Here we know that
longjmp is called from a signal handler on the altstack (SS_ONSTACK),
and that it jumps in the wrong direction (sp decreases), so the check
wants to ensure the jump goes to another stack.
The check is wrong when altstack_sp == setjmp_sp which can happen
when the altstack is a local buffer in the function that calls setjmp,
so the patch allows == too. This fixes bug 27709.
Note that the generic __longjmp_chk check seems to be different.
(it checks if longjmp was on the altstack but does not check setjmp,
so it would not catch incorrect longjmp use within the signal handler).
With this patch, the maximal known error for tgamma is now reduced to 9 ulps
for dbl-64, for all rounding modes. Since exhaustive testing is not possible
for dbl-64, it might be that there are still cases with an error larger than
9 ulps, but all known cases are fixed (intensive tests were done to find cases
with large errors).
Tested on x86_64 and powerpc (and by Adhemerval Zanella on aarch64, arm,
s390x, sparc, and i686).
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
DL_UNMAP_IS_SPECIAL and DL_UNMAP were not defined. The definitions are
now copied from arm, since the same is needed on aarch64. The cleanup
of tlsdesc data is handled by the custom _dl_unmap.
Fixes bug 27403.
For j0f/j1f/y0f/y1f, the largest error for all binary32
inputs is reduced to at most 9 ulps for all rounding modes.
The new code is enabled only when there is a cancellation at the very end of
the j0f/j1f/y0f/y1f computation, or for very large inputs, thus should not
give any visible slowdown on average. Two different algorithms are used:
* around the first 64 zeros of j0/j1/y0/y1, approximation polynomials of
degree 3 are used, computed using the Sollya tool (https://www.sollya.org/)
* for large inputs, an asymptotic formula from [1] is used
[1] Fast and Accurate Bessel Function Computation,
John Harrison, Proceedings of Arith 19, 2009.
Inputs yielding the new largest errors are added to auto-libm-test-in,
and ulps are regenerated for various targets (thanks Adhemerval Zanella).
Tested on x86_64 with --disable-multi-arch and on powerpc64le-linux-gnu.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
config/i386/constraints.md in GCC has
(define_constraint "e"
"32-bit signed integer constant, or a symbolic reference known
to fit that range (for immediate operands in sign-extending x86-64
instructions)."
(match_operand 0 "x86_64_immediate_operand"))
Since movq takes a signed 32-bit immediate or a register source operand,
use "er", instead of "nr"/"ir", constraint for 32-bit signed integer
constant or register on movq.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
This fixes missing definition of math functions in libc in a static link
that are no longer built for libm after commit 4898d9712b ("Avoid adding
duplicated symbols into static libraries").
If building on s390 / i686 with -Os, various conformance
tests are failing with e.g.
conform/ISO/assert.h/linknamespace.out:
[initial] __assert_fail -> [libc.a(assert.o)] __dcgettext -> [libc.a(dcgettext.o)] __dcigettext -> [libc.a(dcigettext.o)] __getcwd -> [libc.a(getcwd.o)] __fstatat64 -> [libc.a(fstatat64.o)] gnu_dev_makedev
The usage of gnu_dev_makedev was recently introduced by
usage of the makedev makro in commit:
5b980d4809
linux: Use statx for MIPSn64
This patch is now linking against __gnu_dev_makedev as
also done in commit:
8b4a118222
Fix -Os gnu_dev_* linknamespace, localplt issues (bug 15105, bug 19463).
All of the isnan functions are in libc.so due to printf_fp, so move
__isnanf128 there too for consistency.
Reviewed-by: Tulio Magno Quites Machado Filho <tuliom@ascii.art.br>
Reviewed-by: Florian Weimer <fweimer@redhat.com>
UNREGISTER_ATFORK is now defined for all ports in register-atfork.h, so most
previous includes of fork.h actually only need register-atfork.h now, and
cxa_finalize.c does not need an ifdef UNREGISTER_ATFORK any more.
The nptl-specific fork generation counters can then go to pthreadP.h, and
fork.h be removed.
Checked on x86_64-linux-gnu and i686-gnu.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Update ifunc-memmove.h to select the function optimized with AVX512
instructions using ZMM16-ZMM31 registers to avoid RTM abort with usable
AVX512VL since VZEROUPPER isn't needed at function exit.
Update ifunc-memset.h/ifunc-wmemset.h to select the function optimized
with AVX512 instructions using ZMM16-ZMM31 registers to avoid RTM abort
with usable AVX512VL and AVX512BW since VZEROUPPER isn't needed at
function exit.
At function exit, AVX optimized string/memory functions have VZEROUPPER
which triggers RTM abort. When such functions are called inside a
transactionally executing RTM region, RTM abort causes severe performance
degradation. Add tests to verify that string/memory functions won't
cause RTM abort in RTM region.
Since VZEROUPPER triggers RTM abort while VZEROALL won't, select AVX
optimized string/memory functions with
xtest
jz 1f
vzeroall
ret
1:
vzeroupper
ret
at function exit on processors with usable RTM, but without 256-bit EVEX
instructions to avoid VZEROUPPER inside a transactionally executing RTM
region.
