The commit 985fc132f2
"strfmon_l: Use specified locale for number formatting [BZ #19633]"
introduced an elf/check-abi-libc testfailure due to __printf_fp_l
on architectures which use sysdeps/ieee754/ldbl-opt/math_ldbl_opt.h.
This patch uses libc_hidden_def instead of ldbl_hidden_def.
The ldbl_strong_alias is removed due to the rename of ___printf_fp_l
to __printf_fp_l.
ChangeLog:
* stdio-common/printf_fp.c (__printf_fp_l):
Rename ___printf_fp_l to __printf_fp_l and
remove strong alias. Use libc_hidden_def instead
of ldbl_hidden_def macro.
The fork handler now runs so late that there is no risk anymore that
other fork handlers in the same thread use malloc, so it is no
longer necessary to install malloc hooks which made a subset
of malloc functionality available to the thread that called fork.
Previously, a thread M invoking fork would acquire locks in this order:
(M1) malloc arena locks (in the registered fork handler)
(M2) libio list lock
A thread F invoking flush (NULL) would acquire locks in this order:
(F1) libio list lock
(F2) individual _IO_FILE locks
A thread G running getdelim would use this order:
(G1) _IO_FILE lock
(G2) malloc arena lock
After executing (M1), (F1), (G1), none of the threads can make progress.
This commit changes the fork lock order to:
(M'1) libio list lock
(M'2) malloc arena locks
It explicitly encodes the lock order in the implementations of fork,
and does not rely on the registration order, thus avoiding the deadlock.
The overloading approach in the W* macros was incompatible with
integer expressions of a type different from int. Applications
using union wait and these macros will have to migrate to the
POSIX-specified int status type.
The large memcpy micro benchmark in glibc shows that there is a
regression with large data on Haswell machine. non-temporal store in
memcpy on large data can improve performance significantly. This
patch adds a threshold to use non temporal store which is 6 times of
shared cache size. When size is above the threshold, non temporal
store will be used, but avoid non-temporal store if there is overlap
between destination and source since destination may be in cache when
source is loaded.
For size below 8 vector register width, we load all data into registers
and store them together. Only forward and backward loops, which move 4
vector registers at a time, are used to support overlapping addresses.
For forward loop, we load the last 4 vector register width of data and
the first vector register width of data into vector registers before the
loop and store them after the loop. For backward loop, we load the first
4 vector register width of data and the last vector register width of
data into vector registers before the loop and store them after the loop.
[BZ #19928]
* sysdeps/x86_64/cacheinfo.c (__x86_shared_non_temporal_threshold):
New.
(init_cacheinfo): Set __x86_shared_non_temporal_threshold to 6
times of shared cache size.
* sysdeps/x86_64/multiarch/memmove-avx-unaligned-erms.S
(VMOVNT): New.
* sysdeps/x86_64/multiarch/memmove-avx512-unaligned-erms.S
(VMOVNT): Likewise.
* sysdeps/x86_64/multiarch/memmove-sse2-unaligned-erms.S
(VMOVNT): Likewise.
(VMOVU): Changed to movups for smaller code sizes.
(VMOVA): Changed to movaps for smaller code sizes.
* sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S: Update
comments.
(PREFETCH): New.
(PREFETCH_SIZE): Likewise.
(PREFETCHED_LOAD_SIZE): Likewise.
(PREFETCH_ONE_SET): Likewise.
Rewrite to use forward and backward loops, which move 4 vector
registers at a time, to support overlapping addresses and use
non temporal store if size is above the threshold and there is
no overlap between destination and source.
This patch adds support for using the implementations of gettimeofday()
and clock_gettime() provided by the kernel in the VDSO. The VDSO will
always provide clock_gettime() as CLOCK_{REALTIME,MONOTONIC}_COARSE can
be implemented regardless of platform. CLOCK_{REALTIME,MONOTONIC}, along
with gettimeofday(), are only implemented on platforms which make use of
either the CP0 count or GIC as their clocksource. On other platforms,
the VDSO does not provide the __vdso_gettimeofday symbol, as it is
never useful.
The VDSO functions return ENOSYS when they encounter an unsupported
request, in which case glibc should fall back to the standard syscall.
Tested with upstream kernel 4.5 and QEMU emulating Malta.
./vdsotest gettimeofday bench
gettimeofday: syscall: 1021 nsec/call
gettimeofday: libc: 262 nsec/call
gettimeofday: vdso: 174 nsec/call
* sysdeps/unix/sysv/linux/mips/Makefile (sysdep_routines):
Include dl-vdso.
* sysdeps/unix/sysv/linux/mips/Versions: Add
__vdso_clock_gettime.
* sysdeps/unix/sysv/linux/mips/init-first.c: New file.
* sysdeps/unix/sysv/linux/mips/libc-vdso.h: New file.
* sysdeps/unix/sysv/linux/mips/mips32/sysdep.h:
(INTERNAL_VSYSCALL_CALL): Define to be compatible with MIPS
definitions of INTERNAL_SYSCALL_{ERROR_P,ERRNO}.
(HAVE_CLOCK_GETTIME_VSYSCALL): Define.
(HAVE_GETTIMEOFDAY_VSYSCALL): Define.
* sysdeps/unix/sysv/linux/mips/mips64/n32/sysdep.h: Likewise.
* sysdeps/unix/sysv/linux/mips/mips64/n64/sysdep.h: Likewise.
This patch consolidates all the pwrite/pwrite64 implementation for Linux
in only one (sysdeps/unix/sysv/linux/pwrite{64}.c). It also removes the
syscall from the auto-generation using assembly macros.
For pwrite{64} offset argument placement the new SYSCALL_LL{64} macro
is used. For pwrite ports that do not define __NR_pwrite will use
__NR_pwrite64 and for pwrite64 ports that dot define __NR_pwrite64 will
use __NR_pwrite for the syscall.
Checked on x86_64, x32, i386, aarch64, and ppc64le.
* sysdeps/unix/sysv/linux/arm/pwrite.c: Remove file.
* sysdeps/unix/sysv/linux/arm/pwrite64.c: Likewise.
* sysdeps/unix/sysv/linux/generic/wordsize-32/pwrite.c: Likewise.
* sysdeps/unix/sysv/linux/generic/wordsize-32/pwrite64.c: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/pwrite.c: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/pwrite64.c: Likewise.
* sysdeps/unix/sysv/linux/wordsize-64/pwrite64.c: Likewise.
* sysdeps/unix/sysv/linux/wordsize-64/syscalls.list (prite): Remove
syscalls generation.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/sysdep.h
[__NR_pwrite64] (__NR_write): Remove define.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/sysdep.h
[__NR_pwrite64] (__NR_write): Remove define.
* sysdeps/unix/sysv/linux/pwrite.c [__NR_pwrite64] (__NR_pwrite):
Remove define.
(__libc_pwrite): Use SYSCALL_LL macro on offset argument.
* sysdeps/unix/sysv/linux/pwrite64.c [__NR_pwrite64] (__NR_pwrite):
Remove define.
(__libc_pwrite64): Use SYSCALL_LL64 macro on offset argument.
* sysdeps/unix/sysv/linux/sh/pwrite.c: Rewrite using default
Linux implementation as base.
* sysdeps/unix/sysv/linux/sh/pwrite64.c: Likewise.
* sysdeps/unix/sysv/linux/mips/pwrite.c: Likewise.
* sysdeps/unix/sysv/linux/mips/pwrite64.c: Likewise.
This patch consolidates all the pread/pread64 implementation for Linux
in only one (sysdeps/unix/sysv/linux/pread.c). It also removes the
syscall from the auto-generation using assembly macros.
For pread{64} offset argument placement the new SYSCALL_LL{64} macro
is used. For pread ports that do not define __NR_pread will use
__NR_pread64 and for pread64 ports that dot define __NR_pread64 will
use __NR_pread for the syscall.
Checked on x86_64, x32, i386, aarch64, and ppc64le.
* sysdeps/unix/sysv/linux/arm/pread.c: Remove file.
* sysdeps/unix/sysv/linux/arm/pread64.c: Likewise.
* sysdeps/unix/sysv/linux/generic/wordsize-32/pread.c: Likewise.
* sysdeps/unix/sysv/linux/generic/wordsize-32/pread64.c: Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/pread.c: Likewise,
* sysdeps/unix/sysv/linux/powerpc/powerpc32/pread64.c: Likewise.
* sysdeps/unix/sysv/linux/wordsize-64/pread64.c: Likewise.
* sysdeps/unix/sysv/linux/wordsize-64/syscalls.list (pread): Remove
syscall generation.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/sysdep.h
[__NR_pread64] (__NR_pread): Remove define.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/sysdep.h:
[__NR_pread64] (__NR_pread): Likewise.
* sysdeps/unix/sysv/linux/pread.c [__NR_pread64] (__NR_pread): Remove
define.
(__libc_pread): Use SYSCALL_LL macro on offset argument.
* sysdeps/unix/sysv/linux/pread64.c [__NR_pread64] (__NR_pread):
Remove define.
(__libc_pread64): Use SYSCALL_LL64 macro on offset argument.
* sysdeps/unix/sysv/linux/sh/pread.c: Rewrite using default
Linux implementation as base.
* sysdeps/unix/sysv/linux/sh/pread64.c: Likewise.
* sysdeps/unix/sysv/linux/mips/pread.c: Likewise.
* sysdeps/unix/sysv/linux/mips/pread64.c: Likewise.
This patch add three new macros (SYSCALL_LL, SYSCALL_LL64, and
__ASSUME_WORDSIZE64_ILP32) to use along with off_t and off64_t argument
syscalls. The rationale for this change is:
1. Remove multiple implementations for the same syscall for different
architectures (for instance, pread have 6 different implementations).
2. Also remove the requirement to use syscall wrappers for cancellable
entrypoints.
The macro usage should be used along __ALIGNMENT_ARG to follow ABI constrains
for architecture where it applies. For instance, pread can be rewritten as:
return SYSCALL_CANCEL (pread, fd, buf, count,
__ALIGNMENT_ARG SYSCALL_LL (offset));
Another macro, SYSCALL_LL64, is provided for off64_t. The macro
__ASSUME_WORDSIZE64_ILP32 is used by the ABI to define is uses 64-bit register
even if ABI is ILP32 (for instance x32 and mips64-n32).
The changes itself are not currently used in any implementation, so no
code change is expected.
* sysdeps/unix/sysv/linux/generic/sysdep.h (__ALIGNMENT_ARG): Move
definition.
(__ALIGNMENT_COUNT): Likewise.
* sysdeps/unix/sysv/linux/sysdep.h (__ALIGNMENT_ARG): To here.
(__ALIGNMENT_COUNT): Likewise.
(SYSCALL_LL): New define.
(SYSCALL_LL64): Likewise.
* sysdeps/unix/sysv/linux/mips/kernel-features.h:
[_MIPS_SIM == _ABIO32] (__ASSUME_WORDSIZE64_ILP32): Define.
* sysdeps/unix/sysv/linux/x86_64/kernel-features.h:
[ILP32] (__ASUME_WORDSIZE64_ILP32): Likewise.
This patch defines __ASSUME_ALIGNED_REGISTER_PAIRS for the missing
ports that require 64-bit value (e.g., long long) to be aligned to
an even register pair in argument passing.
No code change is expected, tested with builds for powerpc32,
mips-o32, and armhf.
* sysdeps/unix/sysv/linux/arm/kernel-features.h
(__ASSUME_ALIGNED_REGISTER_PAIRS): Define.
* sysdeps/unix/sysv/linux/mips/kernel-features.h
[_MIPS_SIM == _ABIO32] (__ASSUME_ALIGNED_REGISTER_PAIRS): Likewise.
* sysdeps/unix/sysv/linux/powerpc/kernel-features.h
[!__powerpc64__] (__ASSUME_ALIGNED_REGISTER_PAIRS): Likewise.
* sysdeps/generic/ldsodefs.h (struct rtld_global_ro)
[!HAVE_AUX_VECTOR]: Do not define _dl_auxv field.
* misc/getauxval.c (__getauxval) [!HAVE_AUX_VECTOR]: Do not go through
GLRO(dl_auxv) list.
This fixes errors when we inject sse options through CFLAGS and now
that we have -Werror turned on by default this warning turns into an
error on x86:
$ gcc -m32 -march=core2 -mtune=core2 -msse3 -mfpmath=sse -x c /dev/null -S -mno-sse -mno-mmx
/dev/null:1:0: warning: SSE instruction set disabled, using 387 arithmetics
Where as:
$ gcc -m32 -march=core2 -mtune=core2 -msse3 -mfpmath=sse -x c /dev/null -S -mno-sse -mno-mmx -mfpmath=387
Generates no warnings.
From the bug:
Obsolete locale. The ISO-639 code for Hebrew was changed from 'iw'
to 'he' in 1989, according to Bruno Haible on libc-alpha 2003-09-01.
Reported-by: Chris Leonard <cjlhomeaddress@gmail.com>
bits/xopen_lim.h (included by limits.h if __USE_XOPEN) defines
NL_NMAX, but this constant was removed in the 2008 edition of POSIX so
should not be defined in that case. This patch duly disables that
define for __USE_XOPEN2K8. It remains enabled for __USE_GNU to avoid
affecting sysconf (_SC_NL_NMAX), the implementation of which uses
"#ifdef NL_NMAX".
Tested for x86_64 and x86 (testsuite, and that installed stripped
shared libraries are unchanged by the patch).
[BZ #19929]
* include/bits/xopen_lim.h (NL_NMAX): Do not define if
[__USE_XOPEN2K8 && !__USE_GNU].
* conform/Makefile (test-xfail-XOPEN2K8/limits.h/conform): Remove
variable.
bits/termios.h (various versions under sysdeps/unix/sysv/linux)
defines XCASE if defined __USE_MISC || defined __USE_XOPEN. This
macro was removed in the 2001 edition of POSIX, and is not otherwise
reserved, so should not be defined for 2001 and later versions of
POSIX. This patch fixes the conditions accordingly (leaving the macro
defined for __USE_MISC, so still in the default namespace).
Tested for x86_64 and x86 (testsuite, and that installed shared
libraries are unchanged by the patch).
[BZ #19925]
* sysdeps/unix/sysv/linux/alpha/bits/termios.h (XCASE): Do not
define if [!__USE_MISC && __USE_XOPEN2K].
* sysdeps/unix/sysv/linux/bits/termios.h (XCASE): Likewise.
* sysdeps/unix/sysv/linux/mips/bits/termios.h (XCASE): Likewise.
* sysdeps/unix/sysv/linux/powerpc/bits/termios.h (XCASE):
Likewise.
* sysdeps/unix/sysv/linux/sparc/bits/termios.h (XCASE): Likewise.
* conform/Makefile (test-xfail-XOPEN2K/termios.h/conform): Remove
variable.
(test-xfail-XOPEN2K8/termios.h/conform): Likewise.
This utilizes vectors and bitmasks. For small needle, large
haystack, the performance improvement is upto 8x. For short
strings (0-4B), the cost of computing the bitmask dominates,
and is a tad slower.
Prepare memmove-vec-unaligned-erms.S to make the SSE2 version as the
default memcpy, mempcpy and memmove.
* sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
(MEMCPY_SYMBOL): New.
(MEMPCPY_SYMBOL): Likewise.
(MEMMOVE_CHK_SYMBOL): Likewise.
Replace MEMMOVE_SYMBOL with MEMMOVE_CHK_SYMBOL on __mempcpy_chk
symbols. Replace MEMMOVE_SYMBOL with MEMPCPY_SYMBOL on
__mempcpy symbols. Provide alias for __memcpy_chk in libc.a.
Provide alias for memcpy in libc.a and ld.so.
Prepare memset-vec-unaligned-erms.S to make the SSE2 version as the
default memset.
* sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
(MEMSET_CHK_SYMBOL): New. Define if not defined.
(__bzero): Check VEC_SIZE == 16 instead of USE_MULTIARCH.
Disabled fro now.
Replace MEMSET_SYMBOL with MEMSET_CHK_SYMBOL on __memset_chk
symbols. Properly check USE_MULTIARCH on __memset symbols.
Since memmove and memset in ld.so don't use IFUNC, don't put SSE2, AVX
and AVX512 memmove and memset in ld.so.
* sysdeps/x86_64/multiarch/memmove-avx-unaligned-erms.S: Skip
if not in libc.
* sysdeps/x86_64/multiarch/memmove-avx512-unaligned-erms.S:
Likewise.
* sysdeps/x86_64/multiarch/memset-avx2-unaligned-erms.S:
Likewise.
* sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S:
Likewise.
__mempcpy_erms and __memmove_erms can't be placed between __memmove_chk
and __memmove it breaks __memmove_chk.
Don't check source == destination first since it is less common.
* sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S:
(__mempcpy_erms, __memmove_erms): Moved before __mempcpy_chk
with unaligned_erms.
(__memmove_erms): Skip if source == destination.
(__memmove_unaligned_erms): Don't check source == destination
first.
Intel Core i3, i5 and i7 processors have fast unaligned copy and
copy backward is ignored. Remove Fast_Copy_Backward from Intel Core
processors to avoid confusion.
* sysdeps/x86/cpu-features.c (init_cpu_features): Don't set
bit_arch_Fast_Copy_Backward for Intel Core proessors.
This patch removes the powerpc64 optimized strspn, strcspn, and
strpbrk assembly implementation now that the default C one
implements the same strategy. On internal glibc benchtests
current implementations shows similar performance with -O2.
Tested on powerpc64le (POWER8).
* sysdeps/powerpc/powerpc64/strcspn.S: Remove file.
* sysdeps/powerpc/powerpc64/strpbrk.S: Remove file.
* sysdeps/powerpc/powerpc64/strspn.S: Remove file.
With now a faster strcspn implementation, it is faster to just use
it with some return tests than reimplementing strpbrk itself.
As for strcspn optimization, it is generally at least 10 times faster
than the existing implementation on bench-strspn on a few AArch64
implementations.
Also the string/bits/string2.h inlines make no longer sense, as current
implementation will already implement most of the optimizations.
Tested on x86_64, i386, and aarch64.
* string/strpbrk.c (strpbrk): Rewrite function.
* string/bits/string2.h (strpbrk): Use __builtin_strpbrk.
(__strpbrk_c2): Likewise.
(__strpbrk_c3): Likewise.
* string/string-inlines.c
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strpbrk_c2):
Likewise.
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strpbrk_c3):
Likewise.
As for strcspn, this patch improves strspn performance using a much
faster algorithm. It first constructs a 256-entry table based on
the accept string and then uses it as a lookup table for the
input string. As for strcspn optimization, it is generally at least
10 times faster than the existing implementation on bench-strspn
on a few AArch64 implementations.
Also the string/bits/string2.h inlines make no longer sense, as current
implementation will already implement most of the optimizations.
Tested on x86_64, i686, and aarch64.
* string/strspn.c (strcspn): Rewrite function.
* string/bits/string2.h (strspn): Use __builtin_strcspn.
(__strspn_c1): Remove inline function.
(__strspn_c2): Likewise.
(__strspn_c3): Likewise.
* string/string-inlines.c
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strspn_c1): Add
compatibility symbol.
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strspn_c2):
Likewise.
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strspn_c3):
Likewise.
Improve strcspn performance using a much faster algorithm. It is kept simple
so it works well on most targets. It is generally at least 10 times faster
than the existing implementation on bench-strcspn on a few AArch64
implementations, and for some tests 100 times as fast (repeatedly calling
strchr on a small string is extremely slow...).
In fact the string/bits/string2.h inlines make no longer sense, as GCC
already uses strlen if reject is an empty string, strchrnul is 5 times as
fast as __strcspn_c1, while __strcspn_c2 and __strcspn_c3 are slower than
the strcspn main loop for large strings (though reject length 2-4 could be
special cased in the future to gain even more performance).
Tested on x86_64, i686, and aarch64.
* string/Version (libc): Add GLIBC_2.24.
* string/strcspn.c (strcspn): Rewrite function.
* string/bits/string2.h (strcspn): Use __builtin_strcspn.
(__strcspn_c1): Remove inline function.
(__strcspn_c2): Likewise.
(__strcspn_c3): Likewise.
* string/string-inline.c
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strcspn_c1): Add
compatibility symbol.
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strcspn_c2):
Likewise.
[SHLIB_COMPAT(libc, GLIBC_2_1_1, GLIBC_2_24)] (__strcspn_c3):
Likewise.
* sysdeps/i386/string-inlines.c: Include generic string-inlines.c.
This patch uses ahi instead of aghi in 32bit _dl_runtime_resolve
to adjust the stack pointer. This is no functional change,
but a cosmetic one.
ChangeLog:
* sysdeps/s390/s390-32/dl-trampoline.h (_dl_runtime_resolve):
Use ahi instead of aghi to adjust stack pointer.
When the signs differ, the precision of the conversion sometimes
drops below 106 bits. This strategy is identical to the
hexadecimal variant.
I've refactored tst-sprintf3 to enable testing a value with more
than 30 significant digits in order to demonstrate this failure
and its solution.
Additionally, this implicitly fixes a typo in the shift
quantities when subtracting from the high mantissa to compute
the difference.
Implement x86-64 memset with unaligned store and rep movsb. Support
16-byte, 32-byte and 64-byte vector register sizes. A single file
provides 2 implementations of memset, one with rep stosb and the other
without rep stosb. They share the same codes when size is between 2
times of vector register size and REP_STOSB_THRESHOLD which defaults
to 2KB.
Key features:
1. Use overlapping store to avoid branch.
2. For size <= 4 times of vector register size, fully unroll the loop.
3. For size > 4 times of vector register size, store 4 times of vector
register size at a time.
[BZ #19881]
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memset-sse2-unaligned-erms, memset-avx2-unaligned-erms and
memset-avx512-unaligned-erms.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Test __memset_chk_sse2_unaligned,
__memset_chk_sse2_unaligned_erms, __memset_chk_avx2_unaligned,
__memset_chk_avx2_unaligned_erms, __memset_chk_avx512_unaligned,
__memset_chk_avx512_unaligned_erms, __memset_sse2_unaligned,
__memset_sse2_unaligned_erms, __memset_erms,
__memset_avx2_unaligned, __memset_avx2_unaligned_erms,
__memset_avx512_unaligned_erms and __memset_avx512_unaligned.
* sysdeps/x86_64/multiarch/memset-avx2-unaligned-erms.S: New
file.
* sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S:
Likewise.
* sysdeps/x86_64/multiarch/memset-sse2-unaligned-erms.S:
Likewise.
* sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S:
Likewise.
Implement x86-64 memmove with unaligned load/store and rep movsb.
Support 16-byte, 32-byte and 64-byte vector register sizes. When
size <= 8 times of vector register size, there is no check for
address overlap bewteen source and destination. Since overhead for
overlap check is small when size > 8 times of vector register size,
memcpy is an alias of memmove.
A single file provides 2 implementations of memmove, one with rep movsb
and the other without rep movsb. They share the same codes when size is
between 2 times of vector register size and REP_MOVSB_THRESHOLD which
is 2KB for 16-byte vector register size and scaled up by large vector
register size.
Key features:
1. Use overlapping load and store to avoid branch.
2. For size <= 8 times of vector register size, load all sources into
registers and store them together.
3. If there is no address overlap bewteen source and destination, copy
from both ends with 4 times of vector register size at a time.
4. If address of destination > address of source, backward copy 8 times
of vector register size at a time.
5. Otherwise, forward copy 8 times of vector register size at a time.
6. Use rep movsb only for forward copy. Avoid slow backward rep movsb
by fallbacking to backward copy 8 times of vector register size at a
time.
7. Skip when address of destination == address of source.
[BZ #19776]
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memmove-sse2-unaligned-erms, memmove-avx-unaligned-erms and
memmove-avx512-unaligned-erms.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Test
__memmove_chk_avx512_unaligned_2,
__memmove_chk_avx512_unaligned_erms,
__memmove_chk_avx_unaligned_2, __memmove_chk_avx_unaligned_erms,
__memmove_chk_sse2_unaligned_2,
__memmove_chk_sse2_unaligned_erms, __memmove_avx_unaligned_2,
__memmove_avx_unaligned_erms, __memmove_avx512_unaligned_2,
__memmove_avx512_unaligned_erms, __memmove_erms,
__memmove_sse2_unaligned_2, __memmove_sse2_unaligned_erms,
__memcpy_chk_avx512_unaligned_2,
__memcpy_chk_avx512_unaligned_erms,
__memcpy_chk_avx_unaligned_2, __memcpy_chk_avx_unaligned_erms,
__memcpy_chk_sse2_unaligned_2, __memcpy_chk_sse2_unaligned_erms,
__memcpy_avx_unaligned_2, __memcpy_avx_unaligned_erms,
__memcpy_avx512_unaligned_2, __memcpy_avx512_unaligned_erms,
__memcpy_sse2_unaligned_2, __memcpy_sse2_unaligned_erms,
__memcpy_erms, __mempcpy_chk_avx512_unaligned_2,
__mempcpy_chk_avx512_unaligned_erms,
__mempcpy_chk_avx_unaligned_2, __mempcpy_chk_avx_unaligned_erms,
__mempcpy_chk_sse2_unaligned_2, __mempcpy_chk_sse2_unaligned_erms,
__mempcpy_avx512_unaligned_2, __mempcpy_avx512_unaligned_erms,
__mempcpy_avx_unaligned_2, __mempcpy_avx_unaligned_erms,
__mempcpy_sse2_unaligned_2, __mempcpy_sse2_unaligned_erms and
__mempcpy_erms.
* sysdeps/x86_64/multiarch/memmove-avx-unaligned-erms.S: New
file.
* sysdeps/x86_64/multiarch/memmove-avx512-unaligned-erms.S:
Likwise.
* sysdeps/x86_64/multiarch/memmove-sse2-unaligned-erms.S:
Likwise.
* sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S:
Likwise.
Starting with z13, vector registers can also occur as argument registers.
Thus the passed input/output register structs for
la_s390_[32|64]_gnu_plt[enter|exit] functions should reflect those new
registers. This patch extends these structs La_s390_regs and La_s390_retval
and adjusts _dl_runtime_profile() to handle those fields in case of
running on a z13 machine.
ChangeLog:
* sysdeps/s390/bits/link.h: (La_s390_vr) New typedef.
(La_s390_32_regs): Append vector register lr_v24-lr_v31.
(La_s390_64_regs): Likewise.
(La_s390_32_retval): Append vector register lrv_v24.
(La_s390_64_retval): Likeweise.
* sysdeps/s390/s390-32/dl-trampoline.h (_dl_runtime_profile):
Handle extended structs La_s390_32_regs and La_s390_32_retval.
* sysdeps/s390/s390-64/dl-trampoline.h (_dl_runtime_profile):
Handle extended structs La_s390_64_regs and La_s390_64_retval.
On s390, no fpr/vrs were saved while resolving a symbol
via _dl_runtime_resolve/_dl_runtime_profile.
According to the abi, the fpr-arguments are defined as call clobbered.
In leaf-functions, gcc 4.9 and newer can use fprs for saving/restoring gprs
instead of saving them to the stack.
If gcc do this in one of the resolver-functions, then the floating point
arguments of a library-function are invalid for the first library-function-call.
Thus, this patch saves/restores the fprs around the resolving code.
The same could occur for vector registers. Furthermore an ifunc-resolver
could also clobber the vector/floating point argument registers.
Thus this patch provides the further variants _dl_runtime_resolve_vx/
_dl_runtime_profile_vx, which are used if the kernel claims, that
we run on a machine with vector registers.
Furthermore, if _dl_runtime_profile calls _dl_call_pltexit,
the pointers to inregs-/outregs-structs were setup invalid.
Now they point to the correct location in the stack-frame.
Before branching back to the caller, the return values are now
restored instead of containing the return values of the
_dl_call_pltexit() call.
On s390-32, an endless loop occurs if _dl_call_pltexit() should be called.
Now, this code-path branches to this function instead of just after the
preceding basr-instruction.
ChangeLog:
* sysdeps/s390/s390-32/dl-trampoline.S: Include dl-trampoline.h twice
to create a non-vector/vector version for _dl_runtime_resolve and
_dl_runtime_profile. Move implementation to ...
* sysdeps/s390/s390-32/dl-trampoline.h: ... here.
(_dl_runtime_resolve) Save and restore fpr/vrs.
(_dl_runtime_profile) Save and restore vrs and fix some issues
if _dl_call_pltexit is called.
* sysdeps/s390/s390-32/dl-machine.h (elf_machine_runtime_setup):
Choose the correct resolver function if running on a machine with vx.
* sysdeps/s390/s390-64/dl-trampoline.S: Include dl-trampoline.h twice
to create a non-vector/vector version for _dl_runtime_resolve and
_dl_runtime_profile. Move implementation to ...
* sysdeps/s390/s390-64/dl-trampoline.h: ... here.
(_dl_runtime_resolve) Save and restore fpr/vrs.
(_dl_runtime_profile) Save and restore vrs and fix some issues
* sysdeps/s390/s390-64/dl-machine.h: (elf_machine_runtime_setup):
Choose the correct resolver function if running on a machine with vx.
