Add x86-64 memset with unaligned store and rep stosb
commit830566307f038387ca0af3fd327706a8d1a2f595
authorH.J. Lu <hjl.tools@gmail.com>
Thu, 31 Mar 2016 17:05:51 +0000 (31 10:05 -0700)
committerH.J. Lu <hjl.tools@gmail.com>
Thu, 31 Mar 2016 17:06:07 +0000 (31 10:06 -0700)
tree22d89ebf426a8799ec13913fd6591a53d4663973
parent88b57b8ed41d5ecf2e1bdfc19556f9246a665ebb
Add x86-64 memset with unaligned store and rep stosb

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.
ChangeLog
sysdeps/x86_64/multiarch/Makefile
sysdeps/x86_64/multiarch/ifunc-impl-list.c
sysdeps/x86_64/multiarch/memset-avx2-unaligned-erms.S [new file with mode: 0644]
sysdeps/x86_64/multiarch/memset-avx512-unaligned-erms.S [new file with mode: 0644]
sysdeps/x86_64/multiarch/memset-sse2-unaligned-erms.S [new file with mode: 0644]
sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S [new file with mode: 0644]