[PATCH 2/3] camellia-aesni-avx-amd64: Move register clearing to assembly functions

[Jussi Kivilinna]

* cipher/camellia-aesni-avx-amd64.S (_gcry_camellia_aesni_avx_ctr_enc)
(_gcry_camellia_aesni_avx_cbc_dec)
(_gcry_camellia_aesni_avx_cfb_dec): Add 'vzeroupper' at head and
'vzeroall' at tail.
* cipher/camellia-glue.c (_gcry_serpent_ctr_enc, _gcry_serpent_cbc_dec)
(_gcry_serpent_avx2_cfb_dec) [USE_AESNI_AVX]: Remove register clearing.
--

Patch moves register clearing with 'vzeroall' to assembly functions and
adds missing 'vzeroupper' instructions at head of assembly functions.

Signed-off-by: Jussi Kivilinna <jussi.kivilinna at iki.fi>
---
 cipher/camellia-aesni-avx-amd64.S |   12 ++++++++++++
 cipher/camellia-glue.c            |    9 ---------
 2 files changed, 12 insertions(+), 9 deletions(-)

diff --git a/cipher/camellia-aesni-avx-amd64.S b/cipher/camellia-aesni-avx-amd64.S
index b0ef5fd..9873d98 100644
--- a/cipher/camellia-aesni-avx-amd64.S
+++ b/cipher/camellia-aesni-avx-amd64.S
@@ -958,6 +958,8 @@ _gcry_camellia_aesni_avx_ctr_enc:
 	 *	%rcx: iv (big endian, 128bit)
 	 */
 
+	vzeroupper;
+
 	subq $(16 * 16), %rsp;
 	movq %rsp, %rax;
 
@@ -1054,6 +1056,8 @@ _gcry_camellia_aesni_avx_ctr_enc:
 		     %xmm15, %xmm14, %xmm13, %xmm12, %xmm11, %xmm10, %xmm9,
 		     %xmm8, %rsi);
 
+	vzeroall;
+
 	ret;
 .size _gcry_camellia_aesni_avx_ctr_enc,.-_gcry_camellia_aesni_avx_ctr_enc;
 
@@ -1069,6 +1073,8 @@ _gcry_camellia_aesni_avx_cbc_dec:
 	 *	%rcx: iv
 	 */
 
+	vzeroupper;
+
 	movq %rcx, %r9;
 
 	cmpl $128, key_bitlength(CTX);
@@ -1114,6 +1120,8 @@ _gcry_camellia_aesni_avx_cbc_dec:
 	movq %r10, (0)(%r9);
 	movq %r11, (8)(%r9);
 
+	vzeroall;
+
 	ret;
 .size _gcry_camellia_aesni_avx_cbc_dec,.-_gcry_camellia_aesni_avx_cbc_dec;
 
@@ -1129,6 +1137,8 @@ _gcry_camellia_aesni_avx_cfb_dec:
 	 *	%rcx: iv
 	 */
 
+	vzeroupper;
+
 	subq $(16 * 16), %rsp;
 	movq %rsp, %rax;
 
@@ -1179,6 +1189,8 @@ _gcry_camellia_aesni_avx_cfb_dec:
 		     %xmm15, %xmm14, %xmm13, %xmm12, %xmm11, %xmm10, %xmm9,
 		     %xmm8, %rsi);
 
+	vzeroall;
+
 	ret;
 .size _gcry_camellia_aesni_avx_cfb_dec,.-_gcry_camellia_aesni_avx_cfb_dec;
 
diff --git a/cipher/camellia-glue.c b/cipher/camellia-glue.c
index 6e2319d..7d8a4cd 100644
--- a/cipher/camellia-glue.c
+++ b/cipher/camellia-glue.c
@@ -344,9 +344,6 @@ _gcry_camellia_ctr_enc(void *context, unsigned char *ctr,
 
       if (did_use_aesni_avx)
         {
-          /* clear AVX registers */
-          asm volatile ("vzeroall;\n":::);
-
           if (burn_stack_depth < 16 * CAMELLIA_BLOCK_SIZE + 2 * sizeof(void *))
             burn_stack_depth = 16 * CAMELLIA_BLOCK_SIZE + 2 * sizeof(void *);
         }
@@ -440,9 +437,6 @@ _gcry_camellia_cbc_dec(void *context, unsigned char *iv,
 
       if (did_use_aesni_avx)
         {
-          /* clear AVX registers */
-          asm volatile ("vzeroall;\n":::);
-
           if (burn_stack_depth < 16 * CAMELLIA_BLOCK_SIZE + 2 * sizeof(void *))
             burn_stack_depth = 16 * CAMELLIA_BLOCK_SIZE + 2 * sizeof(void *);
         }
@@ -531,9 +525,6 @@ _gcry_camellia_cfb_dec(void *context, unsigned char *iv,
 
       if (did_use_aesni_avx)
         {
-          /* clear AVX registers */
-          asm volatile ("vzeroall;\n":::);
-
           if (burn_stack_depth < 16 * CAMELLIA_BLOCK_SIZE + 2 * sizeof(void *))
             burn_stack_depth = 16 * CAMELLIA_BLOCK_SIZE + 2 * sizeof(void *);
         }