diff options
author | Jussi Kivilinna <jussi.kivilinna@iki.fi> | 2013-09-01 16:50:55 +0300 |
---|---|---|
committer | Jussi Kivilinna <jussi.kivilinna@iki.fi> | 2013-09-01 16:50:55 +0300 |
commit | d12828cd821a4b4428eae19de5aee02cf536e536 (patch) | |
tree | d91b173619c4cc23260ccd1cd918ea7ecc419937 /cipher/serpent-avx2-amd64.S | |
parent | fd6721c235a5bdcb332c8eb708fbd4f96e52e824 (diff) | |
download | libgcrypt-d12828cd821a4b4428eae19de5aee02cf536e536.tar.gz |
serpent-avx2-amd64: Move register clearing to assembly
* cipher/serpent-avx2-amd64.S (_gcry_serpent_avx2_ctr_enc)
(_gcry_serpent_avx2_cbc_dec, _gcry_serpent_avx2_cfb_dec): Change last
'vzeroupper' to 'vzeroall'.
* cipher/serpent.c (_gcry_serpent_ctr_enc, _gcry_serpent_cbc_dec)
(_gcry_serpent_avx2_cfb_dec) [USE_AVX2]: Remove register clearing with
'vzeroall'.
--
AVX2 implementation was already clearing upper halfs of YMM registers at end of
assembly functions to prevent long SSE<->AVX transition stalls present on Intel
CPUs. Patch changes these 'vzeroupper' instructions to 'vzeroall' to fully
clear YMM registers. After this change register clearing in serpent.c in not
needed.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Diffstat (limited to 'cipher/serpent-avx2-amd64.S')
-rw-r--r-- | cipher/serpent-avx2-amd64.S | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/cipher/serpent-avx2-amd64.S b/cipher/serpent-avx2-amd64.S index 7586c0c5..c726e7ba 100644 --- a/cipher/serpent-avx2-amd64.S +++ b/cipher/serpent-avx2-amd64.S @@ -730,7 +730,7 @@ _gcry_serpent_avx2_ctr_enc: vmovdqu RB2, (6 * 32)(%rsi); vmovdqu RB3, (7 * 32)(%rsi); - vzeroupper; + vzeroall; ret .size _gcry_serpent_avx2_ctr_enc,.-_gcry_serpent_avx2_ctr_enc; @@ -799,7 +799,7 @@ _gcry_serpent_avx2_cbc_dec: vmovdqu RB2, (6 * 32)(%rsi); vmovdqu RB3, (7 * 32)(%rsi); - vzeroupper; + vzeroall; ret .size _gcry_serpent_avx2_cbc_dec,.-_gcry_serpent_avx2_cbc_dec; @@ -870,7 +870,7 @@ _gcry_serpent_avx2_cfb_dec: vmovdqu RB2, (6 * 32)(%rsi); vmovdqu RB3, (7 * 32)(%rsi); - vzeroupper; + vzeroall; ret .size _gcry_serpent_avx2_cfb_dec,.-_gcry_serpent_avx2_cfb_dec; |