1. Add ISA level build guards to different implementations.
- wcscpy-ssse3.S is used as ISA level 2/3/4.
- wcscpy-generic.c is only used at ISA level 1 and will
only build if compiled with ISA level == 1. Otherwise
there is no reason to include it as we will always use
wcscpy-ssse3.S
2. Refactor the ifunc selector and ifunc implementation list to use
the ISA level aware wrapper macros that allow functions below the
compiled ISA level (with a guranteed replacement) to be skipped.
Tested with and without multiarch on x86_64 for ISA levels:
{generic, x86-64-v2, x86-64-v3, x86-64-v4}
And m32 with and without multiarch.
1. Add default ISA level selection in non-multiarch/rtld
implementations.
2. Add ISA level build guards to different implementations.
- I.e strcmp-avx2.S which is ISA level 3 will only build if
compiled ISA level <= 3. Otherwise there is no reason to
include it as we will always use one of the ISA level 4
implementations (strcmp-evex.S).
3. Refactor the ifunc selector and ifunc implementation list to use
the ISA level aware wrapper macros that allow functions below the
compiled ISA level (with a guranteed replacement) to be skipped.
Tested with and without multiarch on x86_64 for ISA levels:
{generic, x86-64-v2, x86-64-v3, x86-64-v4}
And m32 with and without multiarch.
These functions all have optimized versions:
__strncat_sse2_unaligned, __strncpy_sse2_unaligned, and
stpncpy_sse2_unaligned which are faster than their respective generic
implementations. Since the sse2 versions can run on baseline x86_64,
we should use these as the baseline implementation and can remove the
generic implementations.
Geometric mean of N=20 runs of the entire benchmark suite on:
11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz (Tigerlake)
__strncat_sse2_unaligned / __strncat_generic: .944
__strncpy_sse2_unaligned / __strncpy_generic: .726
__stpncpy_sse2_unaligned / __stpncpy_generic: .650
Tested build with and without multiarch and full check with multiarch.
Remove redundant strcspn-generic, strpbrk-generic and strspn-generic
from sysdep_routines in sysdeps/x86_64/multiarch/Makefile added by
commit c69f960b01
Author: Noah Goldstein <goldstein.w.n@gmail.com>
Date: Sun Jul 3 21:28:07 2022 -0700
x86: Add support for building str{c|p}{brk|spn} with explicit ISA level
since they have been added to sysdep_routines in sysdeps/x86_64/Makefile.
The primary memmove_{impl}_unaligned_erms implementations don't
interact with this function. Putting them in same file both
wastes space and unnecessarily bloats a hot code section.
Implementation wise:
1. Remove the VZEROUPPER as memset_{impl}_unaligned_erms does not
use the L(stosb) label that was previously defined.
2. Don't give the hotpath (fallthrough) to zero size.
Code positioning wise:
Move memset_{chk}_erms to its own file. Leaving it in between the
memset_{impl}_unaligned both adds unnecessary complexity to the
file and wastes space in a relatively hot cache section.
No functions are changed. It just renames generic implementations from
'{func}_sse2' to '{func}_generic'. This is just because the postfix
"_sse2" was overloaded and was used for files that had hand-optimized
sse2 assembly implementations and files that just redirected back
to the generic implementation.
Full xcheck passed on x86_64.
Adding a 512-bit EVEX version of strstr. The algorithm works as follows:
(1) We spend a few cycles at the begining to peek into the needle. We
locate an edge in the needle (first occurance of 2 consequent distinct
characters) and also store the first 64-bytes into a zmm register.
(2) We search for the edge in the haystack by looking into one cache
line of the haystack at a time. This avoids having to read past a page
boundary which can cause a seg fault.
(3) If an edge is found in the haystack we first compare the first
64-bytes of the needle (already stored in a zmm register) before we
proceed with a full string compare performed byte by byte.
Benchmarking results: (old = strstr_sse2_unaligned, new = strstr_avx512)
Geometric mean of all benchmarks: new / old = 0.66
Difficult skiptable(0) : new / old = 0.02
Difficult skiptable(1) : new / old = 0.01
Difficult 2-way : new / old = 0.25
Difficult testing first 2 : new / old = 1.26
Difficult skiptable(0) : new / old = 0.05
Difficult skiptable(1) : new / old = 0.06
Difficult 2-way : new / old = 0.26
Difficult testing first 2 : new / old = 1.05
Difficult skiptable(0) : new / old = 0.42
Difficult skiptable(1) : new / old = 0.24
Difficult 2-way : new / old = 0.21
Difficult testing first 2 : new / old = 1.04
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
This patch implements following evex512 version of string functions.
Perf gain for evex512 version is up to 50% as compared to evex,
depending on length and alignment.
Placeholder function, not used by any processor at the moment.
- String length function using 512 bit vectors.
- String N length using 512 bit vectors.
- Wide string length using 512 bit vectors.
- Wide string N length using 512 bit vectors.
Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
Both symbols are marked as legacy in POSIX.1-2001 and removed on
POSIX.1-2008, although the prototypes are defined for _GNU_SOURCE
or _DEFAULT_SOURCE.
GCC also replaces bcopy with a memmove and bzero with memset on default
configuration (to actually get a bzero libc call the code requires
to omit string.h inclusion and built with -fno-builtin), so it is
highly unlikely programs are actually calling libc bzero symbol.
On a recent Linux distro (Ubuntu 22.04), there is no bzero calls
by the installed binaries.
$ cat count_bstring.sh
#!/bin/bash
files=`IFS=':';for i in $PATH; do test -d "$i" && find "$i" -maxdepth 1 -executable -type f; done`
total=0
for file in $files; do
symbols=`objdump -R $file 2>&1`
if [ $? -eq 0 ]; then
ncalls=`echo $symbols | grep -w $1 | wc -l`
((total=total+ncalls))
if [ $ncalls -gt 0 ]; then
echo "$file: $ncalls"
fi
fi
done
echo "TOTAL=$total"
$ ./count_bstring.sh bzero
TOTAL=0
Checked on x86_64-linux-gnu.
Code didn't actually use any sse4 instructions since `ptest` was
removed in:
commit 2f9062d717
Author: Noah Goldstein <goldstein.w.n@gmail.com>
Date: Wed Nov 10 16:18:56 2021 -0600
x86: Shrink memcmp-sse4.S code size
The new memcmp-sse2 implementation is also faster.
geometric_mean(N=20) of page cross cases SSE2 / SSE4: 0.905
Note there are two regressions preferring SSE2 for Size = 1 and Size =
65.
Size = 1:
size, align0, align1, ret, New Time/Old Time
1, 1, 1, 0, 1.2
1, 1, 1, 1, 1.197
1, 1, 1, -1, 1.2
This is intentional. Size == 1 is significantly less hot based on
profiles of GCC11 and Python3 than sizes [4, 8] (which is made
hotter).
Python3 Size = 1 -> 13.64%
Python3 Size = [4, 8] -> 60.92%
GCC11 Size = 1 -> 1.29%
GCC11 Size = [4, 8] -> 33.86%
size, align0, align1, ret, New Time/Old Time
4, 4, 4, 0, 0.622
4, 4, 4, 1, 0.797
4, 4, 4, -1, 0.805
5, 5, 5, 0, 0.623
5, 5, 5, 1, 0.777
5, 5, 5, -1, 0.802
6, 6, 6, 0, 0.625
6, 6, 6, 1, 0.813
6, 6, 6, -1, 0.788
7, 7, 7, 0, 0.625
7, 7, 7, 1, 0.799
7, 7, 7, -1, 0.795
8, 8, 8, 0, 0.625
8, 8, 8, 1, 0.848
8, 8, 8, -1, 0.914
9, 9, 9, 0, 0.625
Size = 65:
size, align0, align1, ret, New Time/Old Time
65, 0, 0, 0, 1.103
65, 0, 0, 1, 1.216
65, 0, 0, -1, 1.227
65, 65, 0, 0, 1.091
65, 0, 65, 1, 1.19
65, 65, 65, -1, 1.215
This is because A) the checks in range [65, 96] are now unrolled 2x
and B) because smaller values <= 16 are now given a hotter path. By
contrast the SSE4 version has a branch for Size = 80. The unrolled
version has get better performance for returns which need both
comparisons.
size, align0, align1, ret, New Time/Old Time
128, 4, 8, 0, 0.858
128, 4, 8, 1, 0.879
128, 4, 8, -1, 0.888
As well, out of microbenchmark environments that are not full
predictable the branch will have a real-cost.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
New code save size (-303 bytes) and has significantly better
performance.
geometric_mean(N=20) of page cross cases New / Original: 0.634
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
The goal is to remove most SSSE3 function as SSE4, AVX2, and EVEX are
generally preferable. memcpy/memmove is one exception where avoiding
unaligned loads with `palignr` is important for some targets.
This commit replaces memmove-ssse3 with a better optimized are lower
code footprint verion. As well it aliases memcpy to memmove.
Aside from this function all other SSSE3 functions should be safe to
remove.
The performance is not changed drastically although shows overall
improvements without any major regressions or gains.
bench-memcpy geometric_mean(N=50) New / Original: 0.957
bench-memcpy-random geometric_mean(N=50) New / Original: 0.912
bench-memcpy-large geometric_mean(N=50) New / Original: 0.892
Benchmarks where run on Zhaoxin KX-6840@2000MHz See attached numbers
for all results.
More important this saves 7246 bytes of code size in memmove an
additional 10741 bytes by reusing memmove code for memcpy (total 17987
bytes saves). As well an additional 896 bytes of rodata for the jump
table entries.
With SSE2, SSE4.1, AVX2, and EVEX versions very few targets prefer
SSSE3. As a result it is no longer worth it to keep the SSSE3
versions given the code size cost.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
With SSE2, SSE4.1, AVX2, and EVEX versions very few targets prefer
SSSE3. As a result it is no longer worth it to keep the SSSE3
versions given the code size cost.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
With SSE2, SSE4.1, AVX2, and EVEX versions very few targets prefer
SSSE3. As a result it is no longer worth it to keep the SSSE3
versions given the code size cost.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
With SSE2, SSE4.1, AVX2, and EVEX versions very few targets prefer
SSSE3. As a result it is no longer worth it to keep the SSSE3
versions given the code size cost.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
With SSE2, SSE4.1, AVX2, and EVEX versions very few targets prefer
SSSE3. As a result it is no longer worth it to keep the SSSE3
versions given the code size cost.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
The rational is:
1. SSE42 has nearly identical logic so any benefit is minimal (3.4%
regression on Tigerlake using SSE42 versus AVX across the
benchtest suite).
2. AVX2 version covers the majority of targets that previously
prefered it.
3. The targets where AVX would still be best (SnB and IVB) are
becoming outdated.
All in all the saving the code size is worth it.
All string/memory tests pass.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
memset with zero as the value to set is by far the majority value (99%+
for Python3 and GCC).
bzero can be slightly more optimized for this case by using a zero-idiom
xor for broadcasting the set value to a register (vector or GPR).
Co-developed-by: Noah Goldstein <goldstein.w.n@gmail.com>
No bug. This commit adds support for __memcmpeq to be implemented
seperately from memcmp. Support is added for versions optimized with
sse2, avx2, and evex.
No bug. This comment adds the ifunc / build infrastructure
necessary for wcslen to prefer the sse4.1 implementation
in strlen-vec.S. test-wcslen.c is passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
No bug.
This commit adds a new implementation for EVEX memchr that is not safe
for RTM because it uses vzeroupper. The benefit is that by using
ymm0-ymm15 it can use vpcmpeq and vpternlogd in the 4x loop which is
faster than the RTM safe version which cannot use vpcmpeq because
there is no EVEX encoding for the instruction. All parts of the
implementation aside from the 4x loop are the same for the two
versions and the optimization is only relevant for large sizes.
Tigerlake:
size , algn , Pos , Cur T , New T , Win , Dif
512 , 6 , 192 , 9.2 , 9.04 , no-RTM , 0.16
512 , 7 , 224 , 9.19 , 8.98 , no-RTM , 0.21
2048 , 0 , 256 , 10.74 , 10.54 , no-RTM , 0.2
2048 , 0 , 512 , 14.81 , 14.87 , RTM , 0.06
2048 , 0 , 1024 , 22.97 , 22.57 , no-RTM , 0.4
2048 , 0 , 2048 , 37.49 , 34.51 , no-RTM , 2.98 <--
Icelake:
size , algn , Pos , Cur T , New T , Win , Dif
512 , 6 , 192 , 7.6 , 7.3 , no-RTM , 0.3
512 , 7 , 224 , 7.63 , 7.27 , no-RTM , 0.36
2048 , 0 , 256 , 8.48 , 8.38 , no-RTM , 0.1
2048 , 0 , 512 , 11.57 , 11.42 , no-RTM , 0.15
2048 , 0 , 1024 , 17.92 , 17.38 , no-RTM , 0.54
2048 , 0 , 2048 , 30.37 , 27.34 , no-RTM , 3.03 <--
test-memchr, test-wmemchr, and test-rawmemchr are all passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
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.
Update ifunc-memcmp.h to select the function optimized with 256-bit EVEX
instructions using YMM16-YMM31 registers to avoid RTM abort with usable
AVX512VL, AVX512BW and MOVBE since VZEROUPPER isn't needed at function
exit.
Update ifunc-memset.h/ifunc-wmemset.h to select the function optimized
with 256-bit EVEX instructions using YMM16-YMM31 registers to avoid RTM
abort with usable AVX512VL and AVX512BW since VZEROUPPER isn't needed at
function exit.
Update ifunc-memmove.h to select the function optimized with 256-bit EVEX
instructions using YMM16-YMM31 registers to avoid RTM abort with usable
AVX512VL since VZEROUPPER isn't needed at function exit.
Update ifunc-strcpy.h to select the function optimized with 256-bit EVEX
instructions using YMM16-YMM31 registers to avoid RTM abort with usable
AVX512VL and AVX512BW since VZEROUPPER isn't needed at function exit.
Update ifunc-avx2.h, strchr.c, strcmp.c, strncmp.c and wcsnlen.c to
select the function optimized with 256-bit EVEX instructions using
YMM16-YMM31 registers to avoid RTM abort with usable AVX512VL, AVX512BW
and BMI2 since VZEROUPPER isn't needed at function exit.
For strcmp/strncmp, prefer AVX2 strcmp/strncmp if Prefer_AVX2_STRCMP
is set.
Install <sys/platform/x86.h> so that programmers can do
#if __has_include(<sys/platform/x86.h>)
#include <sys/platform/x86.h>
#endif
...
if (CPU_FEATURE_USABLE (SSE2))
...
if (CPU_FEATURE_USABLE (AVX2))
...
<sys/platform/x86.h> exports only:
enum
{
COMMON_CPUID_INDEX_1 = 0,
COMMON_CPUID_INDEX_7,
COMMON_CPUID_INDEX_80000001,
COMMON_CPUID_INDEX_D_ECX_1,
COMMON_CPUID_INDEX_80000007,
COMMON_CPUID_INDEX_80000008,
COMMON_CPUID_INDEX_7_ECX_1,
/* Keep the following line at the end. */
COMMON_CPUID_INDEX_MAX
};
struct cpuid_features
{
struct cpuid_registers cpuid;
struct cpuid_registers usable;
};
struct cpu_features
{
struct cpu_features_basic basic;
struct cpuid_features features[COMMON_CPUID_INDEX_MAX];
};
/* Get a pointer to the CPU features structure. */
extern const struct cpu_features *__x86_get_cpu_features
(unsigned int max) __attribute__ ((const));
Since all feature checks are done through macros, programs compiled with
a newer <sys/platform/x86.h> are compatible with the older glibc binaries
as long as the layout of struct cpu_features is identical. The features
array can be expanded with backward binary compatibility for both .o and
.so files. When COMMON_CPUID_INDEX_MAX is increased to support new
processor features, __x86_get_cpu_features in the older glibc binaries
returns NULL and HAS_CPU_FEATURE/CPU_FEATURE_USABLE return false on the
new processor feature. No new symbol version is neeeded.
Both CPU_FEATURE_USABLE and HAS_CPU_FEATURE are provided. HAS_CPU_FEATURE
can be used to identify processor features.
Note: Although GCC has __builtin_cpu_supports, it only supports a subset
of <sys/platform/x86.h> and it is equivalent to CPU_FEATURE_USABLE. It
doesn't support HAS_CPU_FEATURE.
Optimize x86-64 strcat/strncat, strcpy/strncpy and stpcpy/stpncpy with AVX2.
It uses vector comparison as much as possible. In general, the larger the
source string, the greater performance gain observed, reaching speedups of
1.6x compared to SSE2 unaligned routines. Select AVX2 strcat/strncat,
strcpy/strncpy and stpcpy/stpncpy on AVX2 machines where vzeroupper is
preferred and AVX unaligned load is fast.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strcat-avx2, strncat-avx2, strcpy-avx2, strncpy-avx2,
stpcpy-avx2 and stpncpy-avx2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c:
(__libc_ifunc_impl_list): Add tests for __strcat_avx2,
__strncat_avx2, __strcpy_avx2, __strncpy_avx2, __stpcpy_avx2
and __stpncpy_avx2.
* sysdeps/x86_64/multiarch/{ifunc-unaligned-ssse3.h =>
ifunc-strcpy.h}: rename header for a more generic name.
* sysdeps/x86_64/multiarch/ifunc-strcpy.h:
(IFUNC_SELECTOR): Return OPTIMIZE (avx2) on AVX 2 machines if
AVX unaligned load is fast and vzeroupper is preferred.
* sysdeps/x86_64/multiarch/stpcpy-avx2.S: New file
* sysdeps/x86_64/multiarch/stpncpy-avx2.S: Likewise
* sysdeps/x86_64/multiarch/strcat-avx2.S: Likewise
* sysdeps/x86_64/multiarch/strcpy-avx2.S: Likewise
* sysdeps/x86_64/multiarch/strncat-avx2.S: Likewise
* sysdeps/x86_64/multiarch/strncpy-avx2.S: Likewise
Optimize x86-64 strcmp/wcscmp and strncmp/wcsncmp with AVX2. It uses vector
comparison as much as possible. Peak performance observed on a SkyLake
machine: 9x, 3x, 2.5x and 5.5x for strcmp, strncmp, wcscmp and wcsncmp,
respectively. The larger the comparison length, the more benefit using
avx2 functions, except on the strcmp, where peak is observed at length
== 32 bytes. Select AVX2 strcmp/wcscmp on AVX2 machines where vzeroupper
is preferred and AVX unaligned load is fast.
NB: It uses TZCNT instead of BSF since TZCNT produces the same result
as BSF for non-zero input. TZCNT is faster than BSF and is executed
as BSF if machine doesn't support TZCNT.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strcmp-avx2, strncmp-avx2, wcscmp-avx2, wcscmp-sse2, wcsncmp-avx2 and
wcsncmp-sse2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add tests for __strcmp_avx2,
__strncmp_avx2, __wcscmp_avx2, __wcsncmp_avx2, __wcscmp_sse2
and __wcsncmp_sse2.
* sysdeps/x86_64/multiarch/strcmp.c (OPTIMIZE (avx2)):
(IFUNC_SELECTOR): Return OPTIMIZE (avx2) on AVX 2 machines if
AVX unaligned load is fast and vzeroupper is preferred.
* sysdeps/x86_64/multiarch/strncmp.c: Likewise.
* sysdeps/x86_64/multiarch/strcmp-avx2.S: New file.
* sysdeps/x86_64/multiarch/strncmp-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcscmp-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcscmp-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/wcscmp.c: Likewise.
* sysdeps/x86_64/multiarch/wcsncmp-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcsncmp-sse2.c: Likewise.
* sysdeps/x86_64/multiarch/wcsncmp.c: Likewise.
* sysdeps/x86_64/wcscmp.S (__wcscmp): Add alias only if __wcscmp
is undefined.
Implement strcmp family IFUNC selectors in C.
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for strcmp family functions within libc.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strcmp-sse2, strcmp-sse4_2, strncmp-sse2, strncmp-sse4_2,
strcasecmp_l-sse2, strcasecmp_l-sse4_2, strcasecmp_l-avx,
strncase_l-sse2, strncase_l-sse4_2 and strncase_l-avx.
* sysdeps/x86_64/multiarch/ifunc-strcasecmp.h: New file.
* sysdeps/x86_64/multiarch/strcasecmp.c: Likewise.
* sysdeps/x86_64/multiarch/strcasecmp_l-avx.S: Likewise.
* sysdeps/x86_64/multiarch/strcasecmp_l-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strcasecmp_l-sse4_2.S: Likewise.
* sysdeps/x86_64/multiarch/strcasecmp_l.c: Likewise.
* sysdeps/x86_64/multiarch/strcmp-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strcmp-sse4_2.S: Likewise.
* sysdeps/x86_64/multiarch/strcmp.c: Likewise.
* sysdeps/x86_64/multiarch/strncase.c: Likewise.
* sysdeps/x86_64/multiarch/strncase_l-avx.S : Likewise.
* sysdeps/x86_64/multiarch/strncase_l-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strncase_l-sse4_2.S: Likewise.
* sysdeps/x86_64/multiarch/strncase_l.c: Likewise.
* sysdeps/x86_64/multiarch/strncmp-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strncmp-sse4_2.S: Likewise.
* sysdeps/x86_64/multiarch/strncmp.c: Likewise.
* sysdeps/x86_64/multiarch/strcasecmp_l.S: Removed.
* sysdeps/x86_64/multiarch/strcmp.S: Likewise.
* sysdeps/x86_64/multiarch/strncase_l.S: Likewise.
* sysdeps/x86_64/multiarch/strncmp.S: Likewise.
* sysdeps/x86_64/multiarch/strcmp-sse42.S: Include <sysdep.h>.
(STRCMP_SSE42): New. Defined to __strcmp_sse42 if not defined.
[USE_AS_STRCASECMP_L || USE_AS_STRNCASECMP_L]: Include
"locale-defines.h".
(UPDATE_STRNCMP_COUNTER): New.
(SECTION): Likewise.
(GLABEL): Likewise.
(LABEL): Likewise.
* sysdeps/x86_64/multiarch/strncmp-ssse3.S: Rewrite and enable
for libc.a.
Implement strcspn/strpbrk/strspn IFUNC selectors in C
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for strcspn/strpbrk/strspn functions within libc.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strcspn-sse2, strpbrk-sse2 and strspn-sse2.
* sysdeps/x86_64/strcspn.S (STRPBRK_P): Removed.
Check USE_AS_STRPBRK instead of STRPBRK_P.
* sysdeps/x86_64/strpbrk.S (USE_AS_STRPBRK): New.
* sysdeps/x86_64/multiarch/ifunc-sse4_2.h: New file.
* sysdeps/x86_64/multiarch/strcspn-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strcspn.c: Likewise.
* sysdeps/x86_64/multiarch/strpbrk-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strpbrk.c: Likewise.
* sysdeps/x86_64/multiarch/strspn-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strspn.c: Likewise.
* sysdeps/x86_64/multiarch/strcspn.S: Removed.
* sysdeps/x86_64/multiarch/strpbrk.S: Likewise.
* sysdeps/x86_64/multiarch/strspn.S: Likewise.
* sysdeps/x86_64/multiarch/strpbrk-c.c: Remove "#ifdef SHARED"
and "#endif".
Implement strcat family IFUNC selectors in C.
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for strcat family functions within libc.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strcat-sse2.
* sysdeps/x86_64/multiarch/strcat-sse2.S: New file.
* sysdeps/x86_64/multiarch/strcat.c: Likewise.
* sysdeps/x86_64/multiarch/strncat.c: Likewise.
* sysdeps/x86_64/multiarch/strcat.S: Removed.
* sysdeps/x86_64/multiarch/strncat.S: Likewise.
Implement memcmp family IFUNC selectors in C.
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for memcmp family functions within libc.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memcmp-sse2.
* sysdeps/x86_64/multiarch/ifunc-memcmp.h: New file.
* sysdeps/x86_64/multiarch/memcmp-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/memcmp.c: Likewise.
* sysdeps/x86_64/multiarch/wmemcmp.c: Likewise.
* sysdeps/x86_64/multiarch/memcmp.S: Removed.
* sysdeps/x86_64/multiarch/wmemcmp.S: Likewise.
Implement memset family IFUNC selectors in C.
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for memset functions within libc.
2017-06-07 H.J. Lu <hongjiu.lu@intel.com>
Erich Elsen <eriche@google.com>
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memset-sse2-unaligned-erms, and memset_chk-nonshared.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add test for __memset_chk_erms.
Update comments.
* sysdeps/x86_64/multiarch/ifunc-memset.h: New file.
* sysdeps/x86_64/multiarch/memset-sse2-unaligned-erms.S: Likewise.
* sysdeps/x86_64/multiarch/memset.c: Likewise.
* sysdeps/x86_64/multiarch/memset_chk-nonshared.S: Likewise.
* sysdeps/x86_64/multiarch/memset_chk.c: Likewise.
* sysdeps/x86_64/multiarch/memset.S: Removed.
* sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
* sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
(__memset_chk_erms): New function.
Implement memmove family IFUNC selectors in C.
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for memmove family functions within libc.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memmove-sse2-unaligned-erms, memcpy_chk-nonshared,
mempcpy_chk-nonshared and memmove_chk-nonshared.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add tests for __memmove_chk_erms,
__memcpy_chk_erms and __mempcpy_chk_erms. Update comments.
* sysdeps/x86_64/multiarch/ifunc-memmove.h: New file.
* sysdeps/x86_64/multiarch/memcpy.c: Likewise.
* sysdeps/x86_64/multiarch/memcpy_chk-nonshared.S: Likewise.
* sysdeps/x86_64/multiarch/memcpy_chk.c: Likewise.
* sysdeps/x86_64/multiarch/memmove-sse2-unaligned-erms.S: Likewise.
* sysdeps/x86_64/multiarch/memmove.c: Likewise.
* sysdeps/x86_64/multiarch/memmove_chk-nonshared.S: Likewise.
* sysdeps/x86_64/multiarch/memmove_chk.c: Likewise.
* sysdeps/x86_64/multiarch/mempcpy.c: Likewise.
* sysdeps/x86_64/multiarch/mempcpy_chk-nonshared.S: Likewise.
* sysdeps/x86_64/multiarch/mempcpy_chk.c: Likewise.
* sysdeps/x86_64/multiarch/memcpy.S: Removed.
* sysdeps/x86_64/multiarch/memcpy_chk.S: Likewise.
* sysdeps/x86_64/multiarch/memmove.S: Likewise.
* sysdeps/x86_64/multiarch/memmove_chk.S: Likewise.
* sysdeps/x86_64/multiarch/mempcpy.S: Likewise.
* sysdeps/x86_64/multiarch/mempcpy_chk.S: Likewise.
* sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
(__mempcpy_chk_erms): New function.
(__memmove_chk_erms): Likewise.
(__memcpy_chk_erms): New alias.
Implement strcpy family IFUNC selectors in C.
All internal calls within libc.so can use IFUNC on x86-64 since unlike
x86, x86-64 supports PC-relative addressing to access the GOT entry so
that it can call via PLT without using an extra register. For libc.a,
we can't use IFUNC for functions which are called before IFUNC has been
initialized. Use IFUNC internally reduces the icache footprint since
libc.so and other codes in the process use the same implementations.
This patch uses IFUNC for strcpy family functions within libc.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strcpy-sse2 and stpcpy-sse2.
* sysdeps/x86_64/multiarch/ifunc-unaligned-ssse3.h: New file.
* sysdeps/x86_64/multiarch/stpcpy-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/stpcpy.c: Likewise.
* sysdeps/x86_64/multiarch/stpncpy.c: Likewise.
* sysdeps/x86_64/multiarch/strcpy-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strcpy.c: Likewise.
* sysdeps/x86_64/multiarch/strncpy.c: Likewise.
* sysdeps/x86_64/multiarch/stpcpy.S: Removed.
* sysdeps/x86_64/multiarch/stpncpy.S: Likewise.
* sysdeps/x86_64/multiarch/strcpy.S: Likewise.
* sysdeps/x86_64/multiarch/strncpy.S: Likewise.
* sysdeps/x86_64/multiarch/stpncpy-c.c (weak_alias): New.
(libc_hidden_def): Always defined as empty.
* sysdeps/x86_64/multiarch/strncpy-c.c (libc_hidden_builtin_def):
Always Defined as empty.
Optimize strrchr/wcsrchr with AVX2 to check 32 bytes with vector
instructions. It is as fast as SSE2 version for small data sizes
and up to 1X faster for large data sizes on Haswell. Select AVX2
version on AVX2 machines where vzeroupper is preferred and AVX
unaligned load is fast.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strrchr-sse2, strrchr-avx2, wcsrchr-sse2 and wcsrchr-avx2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add tests for __strrchr_avx2,
__strrchr_sse2, __wcsrchr_avx2 and __wcsrchr_sse2.
* sysdeps/x86_64/multiarch/strrchr-avx2.S: New file.
* sysdeps/x86_64/multiarch/strrchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strrchr.c: Likewise.
* sysdeps/x86_64/multiarch/wcsrchr-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcsrchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/wcsrchr.c: Likewise.
Optimize memrchr with AVX2 to search 32 bytes with a single vector
compare instruction. It is as fast as SSE2 memrchr for small data
sizes and up to 1X faster for large data sizes on Haswell. Select
AVX2 memrchr on AVX2 machines where vzeroupper is preferred and AVX
unaligned load is fast.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memrchr-sse2 and memrchr-avx2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add tests for __memrchr_avx2 and
__memrchr_sse2.
* sysdeps/x86_64/multiarch/memrchr-avx2.S: New file.
* sysdeps/x86_64/multiarch/memrchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/memrchr.c: Likewise.
Optimize strchr/strchrnul/wcschr with AVX2 to search 32 bytes with vector
instructions. It is as fast as SSE2 versions for size <= 16 bytes and up
to 1X faster for or size > 16 bytes on Haswell. Select AVX2 version on
AVX2 machines where vzeroupper is preferred and AVX unaligned load is fast.
NB: It uses TZCNT instead of BSF since TZCNT produces the same result
as BSF for non-zero input. TZCNT is faster than BSF and is executed
as BSF if machine doesn't support TZCNT.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strchr-sse2, strchrnul-sse2, strchr-avx2, strchrnul-avx2,
wcschr-sse2 and wcschr-avx2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add tests for __strchr_avx2,
__strchrnul_avx2, __strchrnul_sse2, __wcschr_avx2 and
__wcschr_sse2.
* sysdeps/x86_64/multiarch/strchr-avx2.S: New file.
* sysdeps/x86_64/multiarch/strchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strchr.c: Likewise.
* sysdeps/x86_64/multiarch/strchrnul-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/strchrnul-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strchrnul.c: Likewise.
* sysdeps/x86_64/multiarch/wcschr-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcschr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/wcschr.c: Likewise.
* sysdeps/x86_64/multiarch/strchr.S: Removed.
Optimize strlen/strnlen/wcslen/wcsnlen with AVX2 to check 32 bytes with
a single vector compare instruction. It is as fast as SSE2 versions for
size <= 16 bytes and up to 1X faster for or size > 16 bytes on Haswell.
Select AVX2 version on AVX2 machines where vzeroupper is preferred and
AVX unaligned load is fast.
NB: It uses TZCNT instead of BSF since TZCNT produces the same result
as BSF for non-zero input. TZCNT is faster than BSF and is executed
as BSF if machine doesn't support TZCNT.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
strlen-sse2, strnlen-sse2, strlen-avx2, strnlen-avx2,
wcslen-sse2, wcslen-avx2 and wcsnlen-avx2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add tests for __strlen_avx2,
__strlen_sse2, __strnlen_avx2, __strnlen_sse2, __wcslen_avx2,
__wcslen_sse2 and __wcsnlen_avx2.
* sysdeps/x86_64/multiarch/strlen-avx2.S: New file.
* sysdeps/x86_64/multiarch/strlen-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strlen.c: Likewise.
* sysdeps/x86_64/multiarch/strnlen-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/strnlen-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/strnlen.c: Likewise.
* sysdeps/x86_64/multiarch/wcslen-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcslen-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/wcslen.c: Likewise.
* sysdeps/x86_64/multiarch/wcsnlen-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wcsnlen.c (OPTIMIZE (avx2)): New.
(IFUNC_SELECTOR): Return OPTIMIZE (avx2) on AVX2 machines where
vzeroupper is preferred and AVX unaligned load is fast.
SSE2 memchr is extended to support wmemchr. AVX2 memchr/rawmemchr/wmemchr
are added to search 32 bytes with a single vector compare instruction.
AVX2 memchr/rawmemchr/wmemchr are as fast as SSE2 memchr/rawmemchr/wmemchr
for small sizes and up to 1.5X faster for larger sizes on Haswell and
Skylake. Select AVX2 memchr/rawmemchr/wmemchr on AVX2 machines where
vzeroupper is preferred and AVX unaligned load is fast.
NB: It uses TZCNT instead of BSF since TZCNT produces the same result
as BSF for non-zero input. TZCNT is faster than BSF and is executed
as BSF if machine doesn't support TZCNT.
* sysdeps/x86_64/memchr.S (MEMCHR): New. Depending on if
USE_AS_WMEMCHR is defined.
(PCMPEQ): Likewise.
(memchr): Renamed to ...
(MEMCHR): This. Support wmemchr if USE_AS_WMEMCHR is defined.
Replace pcmpeqb with PCMPEQ.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memchr-sse2, rawmemchr-sse2, memchr-avx2, rawmemchr-avx2,
wmemchr-sse4_1, wmemchr-avx2 and wmemchr-c.
* sysdeps/x86_64/multiarch/ifunc-avx2.h: New file.
* sysdeps/x86_64/multiarch/memchr-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/memchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/memchr.c: Likewise.
* sysdeps/x86_64/multiarch/rawmemchr-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/rawmemchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/rawmemchr.c: Likewise.
* sysdeps/x86_64/multiarch/wmemchr-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/wmemchr-sse2.S: Likewise.
* sysdeps/x86_64/multiarch/wmemchr.c: Likewise.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Test __memchr_avx2, __memchr_sse2,
__rawmemchr_avx2, __rawmemchr_sse2, __wmemchr_avx2 and
__wmemchr_sse2.
Since wcsnlen.S uses pminud which is the part of SSE4.1, move wcsnlen.S
to multiarch/wcsnlen-sse4_1.S.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
wcsnlen-sse4_1 and wcsnlen-c.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Test __wcsnlen_sse4_1 and
__wcsnlen_sse2.
* sysdeps/x86_64/multiarch/ifunc-sse4_1.h: New file.
* sysdeps/x86_64/multiarch/wcsnlen-c.c: Likewise.
* sysdeps/x86_64/multiarch/wcsnlen-sse4_1.S: Likewise.
* sysdeps/x86_64/multiarch/wcsnlen.c: Likewise.
* sysdeps/x86_64/wcsnlen.S: Removed.
Optimize x86-64 memcmp/wmemcmp with AVX2. It uses vector compare as
much as possible. It is as fast as SSE4 memcmp for size <= 16 bytes
and up to 2X faster for size > 16 bytes on Haswell and Skylake. Select
AVX2 memcmp/wmemcmp on AVX2 machines where vzeroupper is preferred and
AVX unaligned load is fast.
NB: It uses TZCNT instead of BSF since TZCNT produces the same result
as BSF for non-zero input. TZCNT is faster than BSF and is executed
as BSF if machine doesn't support TZCNT.
Key features:
1. For size from 2 to 7 bytes, load as big endian with movbe and bswap
to avoid branches.
2. Use overlapping compare to avoid branch.
3. Use vector compare when size >= 4 bytes for memcmp or size >= 8
bytes for wmemcmp.
4. If size is 8 * VEC_SIZE or less, unroll the loop.
5. Compare 4 * VEC_SIZE at a time with the aligned first memory area.
6. Use 2 vector compares when size is 2 * VEC_SIZE or less.
7. Use 4 vector compares when size is 4 * VEC_SIZE or less.
8. Use 8 vector compares when size is 8 * VEC_SIZE or less.
* sysdeps/x86/cpu-features.h (index_cpu_MOVBE): New.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Add
memcmp-avx2 and wmemcmp-avx2.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Test __memcmp_avx2 and __wmemcmp_avx2.
* sysdeps/x86_64/multiarch/memcmp-avx2.S: New file.
* sysdeps/x86_64/multiarch/wmemcmp-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/memcmp.S: Use __memcmp_avx2 on AVX
2 machines if AVX unaligned load is fast and vzeroupper is
preferred.
* sysdeps/x86_64/multiarch/wmemcmp.S: Use __wmemcmp_avx2 on AVX
2 machines if AVX unaligned load is fast and vzeroupper is
preferred.
