X86-64: Remove the previous SSE2/AVX2 memsets
authorH.J. Lu <hjl.tools@gmail.com>
Wed, 8 Jun 2016 20:55:45 +0000 (13:55 -0700)
committerH.J. Lu <hjl.tools@gmail.com>
Wed, 8 Jun 2016 20:56:14 +0000 (13:56 -0700)
commit5e8c5bb1ac83aa2577d64d82467a653fa413f7ce
treec713874b92fbc3f6858e48d71a883a990665a2f3
parent5188b973250523d3e9c80ea3ab4001f696e6fa1a
X86-64: Remove the previous SSE2/AVX2 memsets

Since the new SSE2/AVX2 memsets are faster than the previous ones, we
can remove the previous SSE2/AVX2 memsets and replace them with the
new ones.  This reduces the size of libc.so by about 900 bytes.

No change in IFUNC selection if SSE2 and AVX2 memsets weren't used
before.  If SSE2 or AVX2 memset was used, the new SSE2 or AVX2 memset
optimized with Enhanced REP STOSB will be used for processors with
ERMS.  The new AVX512 memset will be used for processors with AVX512
which prefer vzeroupper.

[BZ #19881]
* sysdeps/x86_64/multiarch/memset-sse2-unaligned-erms.S: Folded
into ...
* sysdeps/x86_64/memset.S: This.
(__bzero): Removed.
(__memset_tail): Likewise.
(__memset_chk): Likewise.
(memset): Likewise.
(MEMSET_CHK_SYMBOL): New. Define only if MEMSET_SYMBOL isn't
defined.
(MEMSET_SYMBOL): Define only if MEMSET_SYMBOL isn't defined.
* sysdeps/x86_64/multiarch/memset-avx2.S: Removed.
(__memset_zero_constant_len_parameter): Check SHARED instead of
PIC.
* sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Remove
memset-avx2 and memset-sse2-unaligned-erms.
* sysdeps/x86_64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Remove __memset_chk_sse2,
__memset_chk_avx2, __memset_sse2 and __memset_avx2_unaligned.
* sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
(__bzero): Enabled.
* sysdeps/x86_64/multiarch/memset.S (memset): Replace
__memset_sse2 and __memset_avx2 with __memset_sse2_unaligned
and __memset_avx2_unaligned.  Use __memset_sse2_unaligned_erms
or __memset_avx2_unaligned_erms if processor has ERMS.  Support
__memset_avx512_unaligned_erms and __memset_avx512_unaligned.
(memset): Removed.
(__memset_chk): Likewise.
(MEMSET_SYMBOL): New.
(libc_hidden_builtin_def): Replace __memset_sse2 with
__memset_sse2_unaligned.
* sysdeps/x86_64/multiarch/memset_chk.S (__memset_chk): Replace
__memset_chk_sse2 and __memset_chk_avx2 with
__memset_chk_sse2_unaligned and __memset_chk_avx2_unaligned_erms.
Use __memset_chk_sse2_unaligned_erms or
__memset_chk_avx2_unaligned_erms if processor has ERMS.  Support
__memset_chk_avx512_unaligned_erms and
__memset_chk_avx512_unaligned.
ChangeLog
sysdeps/x86_64/memset.S
sysdeps/x86_64/multiarch/Makefile
sysdeps/x86_64/multiarch/ifunc-impl-list.c
sysdeps/x86_64/multiarch/memset-avx2.S [deleted file]
sysdeps/x86_64/multiarch/memset-sse2-unaligned-erms.S [deleted file]
sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
sysdeps/x86_64/multiarch/memset.S
sysdeps/x86_64/multiarch/memset_chk.S