From: Denis Vlasenko <vda@port.imtp.ilyichevsk.odessa.ua>
Switches sha512 to asm-optimized 64bit rotation. Reduces sha512 code size
by ~300 insns, ~1K:
# size sha512.o.old sha512.o
text data bss dec hex filename
6642 364 0 7006 1b5e sha512.o.old
5587 364 0 5951 173f sha512.o
Run-tested with tcrypt module.
I also looked into optimizing cast[56] and sha256 by replacing their 32bit
rotation functions/macros, but it had no measurable effect on i386. gcc
seems to be on par with asm there.
Signed-off-by: Andrew Morton <akpm@osdl.org>
---
25-akpm/crypto/sha512.c | 13 ++++---------
1 files changed, 4 insertions(+), 9 deletions(-)
diff -puN crypto/sha512.c~sha512-use-asm-optimized-bit-rotation crypto/sha512.c
--- 25/crypto/sha512.c~sha512-use-asm-optimized-bit-rotation 2004-10-03 16:33:29.703029872 -0700
+++ 25-akpm/crypto/sha512.c 2004-10-03 16:33:29.708029112 -0700
@@ -43,11 +43,6 @@ static inline u64 Maj(u64 x, u64 y, u64
return (x & y) | (z & (x | y));
}
-static inline u64 RORu64(u64 x, u64 y)
-{
- return (x >> y) | (x << (64 - y));
-}
-
const u64 sha512_K[80] = {
0x428a2f98d728ae22ULL, 0x7137449123ef65cdULL, 0xb5c0fbcfec4d3b2fULL,
0xe9b5dba58189dbbcULL, 0x3956c25bf348b538ULL, 0x59f111f1b605d019ULL,
@@ -78,10 +73,10 @@ const u64 sha512_K[80] = {
0x5fcb6fab3ad6faecULL, 0x6c44198c4a475817ULL,
};
-#define e0(x) (RORu64(x,28) ^ RORu64(x,34) ^ RORu64(x,39))
-#define e1(x) (RORu64(x,14) ^ RORu64(x,18) ^ RORu64(x,41))
-#define s0(x) (RORu64(x, 1) ^ RORu64(x, 8) ^ (x >> 7))
-#define s1(x) (RORu64(x,19) ^ RORu64(x,61) ^ (x >> 6))
+#define e0(x) (ror64(x,28) ^ ror64(x,34) ^ ror64(x,39))
+#define e1(x) (ror64(x,14) ^ ror64(x,18) ^ ror64(x,41))
+#define s0(x) (ror64(x, 1) ^ ror64(x, 8) ^ (x >> 7))
+#define s1(x) (ror64(x,19) ^ ror64(x,61) ^ (x >> 6))
/* H* initial state for SHA-512 */
#define H0 0x6a09e667f3bcc908ULL
_