aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorDenys Vlasenko <vda.linux@googlemail.com>2013-01-15 15:22:30 +0100
committerDenys Vlasenko <vda.linux@googlemail.com>2013-01-15 15:22:30 +0100
commita55df2793660941f42589182537d02ce54eaed66 (patch)
treefe4d2df5b05d5f78798c368ab77a5298d98af21b
parent07a54e21dd08bcd752a23095fdedc904eb7127fb (diff)
downloadbusybox-a55df2793660941f42589182537d02ce54eaed66.tar.gz
sha3: code shrink
function old new delta KeccakF 1064 1053 -11 Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
-rw-r--r--libbb/hash_md5_sha.c31
1 files changed, 18 insertions, 13 deletions
diff --git a/libbb/hash_md5_sha.c b/libbb/hash_md5_sha.c
index 3b1366762..a0eec7789 100644
--- a/libbb/hash_md5_sha.c
+++ b/libbb/hash_md5_sha.c
@@ -988,24 +988,29 @@ static void KeccakF(uint64_t *state)
for (round = 0; round < cKeccakNumberOfRounds; ++round) {
/* Theta */
{
- uint64_t BC[5];
+ uint64_t BC[10];
for (x = 0; x < 5; ++x) {
- BC[x] = state[x] ^ state[5 + x] ^ state[10 + x] ^
- state[15 + x] ^ state[20 + x];
+ BC[x + 5] = BC[x] = state[x]
+ ^ state[x + 5] ^ state[x + 10]
+ ^ state[x + 15] ^ state[x + 20];
}
+ /* Using 2x5 vector above eliminates the need to use
+ * [Mod5[x+N]] index trick below to calculate (x+N) % 5,
+ * and the code is a bit _smaller_.
+ */
for (x = 0; x < 5; ++x) {
- uint64_t temp = BC[KeccakF_Mod5[x + 4]] ^
- rotl64(BC[KeccakF_Mod5[x + 1]], 1);
+ uint64_t temp = BC[x + 4] ^ rotl64(BC[x + 1], 1);
if (SHA3_SMALL && !ARCH_IS_64BIT) {
for (y = 0; y <= 20; y += 5)
- state[y + x] ^= temp;
+ state[x + y] ^= temp;
} else {
- /* on 64-bit arch, this is actually smaller too */
- state[0 + x] ^= temp;
- state[5 + x] ^= temp;
- state[10 + x] ^= temp;
- state[15 + x] ^= temp;
- state[20 + x] ^= temp;
+ /* On 64-bit, this is also smaller,
+ * not only faster, than loop */
+ state[x] ^= temp;
+ state[x + 5] ^= temp;
+ state[x + 10] ^= temp;
+ state[x + 15] ^= temp;
+ state[x + 20] ^= temp;
}
}
}
@@ -1019,7 +1024,7 @@ static void KeccakF(uint64_t *state)
t1 = t0;
}
} else {
- /* Especially large benefit for 32-bit arch:
+ /* Especially large benefit for 32-bit arch (75% faster):
* 64-bit rotations by non-constant usually are SLOW on those.
* We resort to unrolling here.
* This optimizes out KeccakF_PiLane[] and KeccakF_RotationConstants[],