Burn stack in transform functions for SHA1 AMD64 implementations

* cipher/sha1-avx-amd64.S: Burn stack inside transform functions.
* cipher/sha1-avx-bmi2-amd64.S: Ditto.
* cipher/sha1-avx2-bmi2-amd64.S: Ditto.
* cipher/sha1-ssse3-amd64.S: Ditto.
--

This change reduces per call overhead for SHA1.

Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>

4 files changed, 52 insertions, 40 deletions

diff --git a/cipher/sha1-avx-amd64.S b/cipher/sha1-avx-amd64.S
index 143e4066..5d674c15 100644
--- a/cipher/sha1-avx-amd64.S
+++ b/cipher/sha1-avx-amd64.S
@@ -380,7 +380,7 @@ _gcry_sha1_transform_amd64_avx:
 .Lend:
   vzeroall;
 
-  /* Transform 64-79. */
+  /* Transform 64-79 + burn stack */
   R( b, c, d, e, a, F4, 64 );
   R( a, b, c, d, e, F4, 65 );
   R( e, a, b, c, d, F4, 66 );
@@ -393,12 +393,15 @@ _gcry_sha1_transform_amd64_avx:
   R( c, d, e, a, b, F4, 73 );
   R( b, c, d, e, a, F4, 74 );
   R( a, b, c, d, e, F4, 75 );
-  R( e, a, b, c, d, F4, 76 );
-  R( d, e, a, b, c, F4, 77 );
-  R( c, d, e, a, b, F4, 78 );
+  R( e, a, b, c, d, F4, 76 ); vmovdqa %xmm0, (0*16)(%rsp);
+  R( d, e, a, b, c, F4, 77 ); vmovdqa %xmm0, (1*16)(%rsp);
+  R( c, d, e, a, b, F4, 78 ); vmovdqa %xmm0, (2*16)(%rsp);
   addl state_h0(RSTATE), a;
   R( b, c, d, e, a, F4, 79 );
 
+  /* 16*4/16-1 = 3 */
+  vmovdqa %xmm0, (3*16)(%rsp);
+
   /* Update the chaining variables. */
   addl state_h3(RSTATE), d;
   addl state_h2(RSTATE), c;
@@ -416,8 +419,8 @@ _gcry_sha1_transform_amd64_avx:
   popq %rbp;
   popq %rbx;
 
-  /* burn_stack */
-  movl $(16*4 + 2*8 + 31), %eax;
+  /* stack already burned */
+  xorl %eax, %eax;
 
 .Lret:
   ret;
diff --git a/cipher/sha1-avx-bmi2-amd64.S b/cipher/sha1-avx-bmi2-amd64.S
index 79ea24ef..fe8901ef 100644
--- a/cipher/sha1-avx-bmi2-amd64.S
+++ b/cipher/sha1-avx-bmi2-amd64.S
@@ -387,7 +387,7 @@ _gcry_sha1_transform_amd64_avx_bmi2:
 .Lend:
   vzeroall;
 
-  /* Transform 64-79. */
+  /* Transform 64-79 + burn stack */
   R( b, c, d, e, a, F4, 64 );
   R( a, b, c, d, e, F4, 65 );
   R( e, a, b, c, d, F4, 66 );
@@ -400,14 +400,17 @@ _gcry_sha1_transform_amd64_avx_bmi2:
   R( c, d, e, a, b, F4, 73 );
   R( b, c, d, e, a, F4, 74 );
   R( a, b, c, d, e, F4, 75 );
-  R( e, a, b, c, d, F4, 76 );
-  R( d, e, a, b, c, F4, 77 );
-  R( c, d, e, a, b, F4, 78 );
+  R( e, a, b, c, d, F4, 76 ); vmovdqa %xmm0, (0*16)(%rsp);
+  R( d, e, a, b, c, F4, 77 ); vmovdqa %xmm0, (1*16)(%rsp);
+  R( c, d, e, a, b, F4, 78 ); vmovdqa %xmm0, (2*16)(%rsp);
   addl state_h0(RSTATE), a;
   R( b, c, d, e, a, F4, 79 );
   addl ne, a;
   xorl ne, ne;
 
+  /* 16*4/16-1 = 3 */
+  vmovdqa %xmm0, (3*16)(%rsp);
+
   /* Update the chaining variables. */
   addl state_h3(RSTATE), d;
   addl state_h2(RSTATE), c;
@@ -426,8 +429,8 @@ _gcry_sha1_transform_amd64_avx_bmi2:
   popq %rbp;
   popq %rbx;
 
-  /* burn_stack */
-  movl $(16*4 + 3*8 + 31), %eax;
+  /* stack already burned */
+  xorl %eax, %eax;
 
 .Lret:
   ret;
diff --git a/cipher/sha1-avx2-bmi2-amd64.S b/cipher/sha1-avx2-bmi2-amd64.S
index c666290f..2a2f21a5 100644
--- a/cipher/sha1-avx2-bmi2-amd64.S
+++ b/cipher/sha1-avx2-bmi2-amd64.S
@@ -504,7 +504,7 @@ _gcry_sha1_transform_amd64_avx2_bmi2:
 .Lend:
   vzeroall;
 
-  /* Transform 48-79 for block 2. */
+  /* Transform 48-79 for block 2 + burn stack */
   R( c, d, e, a, b, F3, 48, 1 );
   R( b, c, d, e, a, F3, 49, 1 );
   R( a, b, c, d, e, F3, 50, 1 );
@@ -517,30 +517,33 @@ _gcry_sha1_transform_amd64_avx2_bmi2:
   R( d, e, a, b, c, F3, 57, 1 );
   R( c, d, e, a, b, F3, 58, 1 );
   R( b, c, d, e, a, F3, 59, 1 );
-  R( a, b, c, d, e, F4, 60, 1 );
-  R( e, a, b, c, d, F4, 61, 1 );
-  R( d, e, a, b, c, F4, 62, 1 );
-  R( c, d, e, a, b, F4, 63, 1 );
-  R( b, c, d, e, a, F4, 64, 1 );
-  R( a, b, c, d, e, F4, 65, 1 );
-  R( e, a, b, c, d, F4, 66, 1 );
-  R( d, e, a, b, c, F4, 67, 1 );
-  R( c, d, e, a, b, F4, 68, 1 );
-  R( b, c, d, e, a, F4, 69, 1 );
-  R( a, b, c, d, e, F4, 70, 1 );
-  R( e, a, b, c, d, F4, 71, 1 );
-  R( d, e, a, b, c, F4, 72, 1 );
-  R( c, d, e, a, b, F4, 73, 1 );
-  R( b, c, d, e, a, F4, 74, 1 );
-  R( a, b, c, d, e, F4, 75, 1 );
-  R( e, a, b, c, d, F4, 76, 1 );
-  R( d, e, a, b, c, F4, 77, 1 );
-  R( c, d, e, a, b, F4, 78, 1 );
+  R( a, b, c, d, e, F4, 60, 1 ); vmovdqa %ymm0, (0*32)(%rsp);
+  R( e, a, b, c, d, F4, 61, 1 ); vmovdqa %ymm0, (1*32)(%rsp);
+  R( d, e, a, b, c, F4, 62, 1 ); vmovdqa %ymm0, (2*32)(%rsp);
+  R( c, d, e, a, b, F4, 63, 1 ); vmovdqa %ymm0, (3*32)(%rsp);
+  R( b, c, d, e, a, F4, 64, 1 ); vmovdqa %ymm0, (4*32)(%rsp);
+  R( a, b, c, d, e, F4, 65, 1 ); vmovdqa %ymm0, (5*32)(%rsp);
+  R( e, a, b, c, d, F4, 66, 1 ); vmovdqa %ymm0, (6*32)(%rsp);
+  R( d, e, a, b, c, F4, 67, 1 ); vmovdqa %ymm0, (7*32)(%rsp);
+  R( c, d, e, a, b, F4, 68, 1 ); vmovdqa %ymm0, (8*32)(%rsp);
+  R( b, c, d, e, a, F4, 69, 1 ); vmovdqa %ymm0, (9*32)(%rsp);
+  R( a, b, c, d, e, F4, 70, 1 ); vmovdqa %ymm0, (10*32)(%rsp);
+  R( e, a, b, c, d, F4, 71, 1 ); vmovdqa %ymm0, (11*32)(%rsp);
+  R( d, e, a, b, c, F4, 72, 1 ); vmovdqa %ymm0, (12*32)(%rsp);
+  R( c, d, e, a, b, F4, 73, 1 ); vmovdqa %ymm0, (13*32)(%rsp);
+  R( b, c, d, e, a, F4, 74, 1 ); vmovdqa %ymm0, (14*32)(%rsp);
+  R( a, b, c, d, e, F4, 75, 1 ); vmovdqa %ymm0, (15*32)(%rsp);
+  R( e, a, b, c, d, F4, 76, 1 ); vmovdqa %ymm0, (16*32)(%rsp);
+  R( d, e, a, b, c, F4, 77, 1 ); vmovdqa %ymm0, (17*32)(%rsp);
+  R( c, d, e, a, b, F4, 78, 1 ); vmovdqa %ymm0, (18*32)(%rsp);
   addl state_h0(RSTATE), a;
   R( b, c, d, e, a, F4, 79, 1 );
   addl ne, a;
   xorl ne, ne;
 
+  /* WK_STACK_WORDS*4/32-1 = 19 */
+  vmovdqa %ymm0, (19*32)(%rsp);
+
   /* Update the chaining variables. */
   addl state_h3(RSTATE), d;
   addl state_h2(RSTATE), c;
@@ -559,8 +562,8 @@ _gcry_sha1_transform_amd64_avx2_bmi2:
   popq %rbp;
   popq %rbx;
 
-  /* burn_stack */
-  movl $((WK_STACK_WORDS)*4 + 3*8 + 31), %eax;
+  /* stack already burned */
+  xorl %eax, %eax;
 
   ret;
 ELF(.size _gcry_sha1_transform_amd64_avx2_bmi2,
diff --git a/cipher/sha1-ssse3-amd64.S b/cipher/sha1-ssse3-amd64.S
index 421bebec..fff14034 100644
--- a/cipher/sha1-ssse3-amd64.S
+++ b/cipher/sha1-ssse3-amd64.S
@@ -388,7 +388,7 @@ _gcry_sha1_transform_amd64_ssse3:
 
 .align 16
 .Lend:
-  /* Transform 64-79 + Clear XMM registers. */
+  /* Transform 64-79 + Clear XMM registers + Burn stack. */
   R( b, c, d, e, a, F4, 64 ); CLEAR_REG(BSWAP_REG);
   R( a, b, c, d, e, F4, 65 ); CLEAR_REG(Wtmp0);
   R( e, a, b, c, d, F4, 66 ); CLEAR_REG(Wtmp1);
@@ -401,12 +401,15 @@ _gcry_sha1_transform_amd64_ssse3:
   R( c, d, e, a, b, F4, 73 ); CLEAR_REG(W6);
   R( b, c, d, e, a, F4, 74 ); CLEAR_REG(W7);
   R( a, b, c, d, e, F4, 75 );
-  R( e, a, b, c, d, F4, 76 );
-  R( d, e, a, b, c, F4, 77 );
-  R( c, d, e, a, b, F4, 78 );
+  R( e, a, b, c, d, F4, 76 ); movdqa Wtmp0, (0*16)(%rsp);
+  R( d, e, a, b, c, F4, 77 ); movdqa Wtmp0, (1*16)(%rsp);
+  R( c, d, e, a, b, F4, 78 ); movdqa Wtmp0, (2*16)(%rsp);
   addl state_h0(RSTATE), a;
   R( b, c, d, e, a, F4, 79 );
 
+  /* 16*4/16-1 = 3 */
+  vmovdqa Wtmp0, (3*16)(%rsp);
+
   /* Update the chaining variables. */
   addl state_h3(RSTATE), d;
   addl state_h2(RSTATE), c;
@@ -424,8 +427,8 @@ _gcry_sha1_transform_amd64_ssse3:
   popq %rbp;
   popq %rbx;
 
-  /* burn_stack */
-  movl $(16*4 + 2*8 + 31), %eax;
+  /* stack already burned */
+  xorl %eax, %eax;
 
 .Lret:
   ret;