1. Refactor files so that all implementations are in the multiarch
directory
- Moved the implementation portion of memmove sse2 from memmove.S
to multiarch/memmove-sse2.S
- The non-multiarch file now only includes one of the
implementations in the multiarch directory based on the compiled
ISA level (only used for non-multiarch builds. Otherwise we go
through the ifunc selector).
2. Add ISA level build guards to different implementations.
- I.e memmove-avx2-unaligned-erms.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 (memmove-evex-unaligned-erms.S).
3. Add new multiarch/rtld-memmove.S that just include the
non-multiarch memmove.S which will in turn select the best
implementation based on the compiled ISA level.
4. 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.
isa raising memmove
No bug.
The optimizations are as follows:
1) Always align entry to 64 bytes. This makes behavior more
predictable and makes other frontend optimizations easier.
2) Make the L(more_8x_vec) cases 4k aliasing aware. This can have
significant benefits in the case that:
0 < (dst - src) < [256, 512]
3) Align before `rep movsb`. For ERMS this is roughly a [0, 30%]
improvement and for FSRM [-10%, 25%].
In addition to these primary changes there is general cleanup
throughout to optimize the aligning routines and control flow logic.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
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.
If assembler doesn't support AVX512DQ, _dl_runtime_resolve_avx is used
to save the first 8 vector registers, which only saves the lower 256
bits of vector register, for lazy binding. When it is called on AVX512
platform, the upper 256 bits of ZMM registers are clobbered. Parameters
passed in ZMM registers will be wrong when the function is called the
first time. This patch requires binutils 2.24, whose assembler can store
and load ZMM registers, to build x86-64 glibc. Since mathvec library
needs assembler support for AVX512DQ, we disable mathvec if assembler
doesn't support AVX512DQ.
[BZ #20139]
* config.h.in (HAVE_AVX512_ASM_SUPPORT): Renamed to ...
(HAVE_AVX512DQ_ASM_SUPPORT): This.
* sysdeps/x86_64/configure.ac: Require assembler from binutils
2.24 or above.
(HAVE_AVX512_ASM_SUPPORT): Removed.
(HAVE_AVX512DQ_ASM_SUPPORT): New.
* sysdeps/x86_64/configure: Regenerated.
* sysdeps/x86_64/dl-trampoline.S: Make HAVE_AVX512_ASM_SUPPORT
check unconditional.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c: Likewise.
* sysdeps/x86_64/multiarch/memcpy.S: Likewise.
* sysdeps/x86_64/multiarch/memcpy_chk.S: Likewise.
* sysdeps/x86_64/multiarch/memmove-avx512-no-vzeroupper.S:
Likewise.
* sysdeps/x86_64/multiarch/memmove-avx512-unaligned-erms.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/memset-avx512-no-vzeroupper.S:
Likewise.
* sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S:
Likewise.
* sysdeps/x86_64/multiarch/memset.S: Likewise.
* sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_d_cos8_core_avx512.S: Check
HAVE_AVX512DQ_ASM_SUPPORT instead of HAVE_AVX512_ASM_SUPPORT.
* sysdeps/x86_64/fpu/multiarch/svml_d_exp8_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_d_log8_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_d_pow8_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_d_sin8_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_d_sincos8_core_avx512.:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_cosf16_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_expf16_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_logf16_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_powf16_core_avx512.S:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_sincosf16_core_avx51:
Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_sinf16_core_avx512.S:
Likewise.
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.
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.
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.
