Optimisation for amd64

Increased major number as the ABI was broken
by insertion of padding into the BLAKE2
parameter structures (except for BLAKE2Xs)

Signed-off-by: Mattias Andrée <maandree@kth.se>

1 files changed, 70 insertions, 40 deletions

diff --git a/libblake_blake2b_init.c b/libblake_blake2b_init.c
index b520a87..f07d1ef 100644
--- a/libblake_blake2b_init.c
+++ b/libblake_blake2b_init.c
@@ -1,6 +1,25 @@
 /* See LICENSE file for copyright and license details. */
 #include "common.h"
 
+#if defined(LITTLE_ENDIAN)
+# define le64(X) X
+#else
+static uint_least64_t
+le64(uint_least64_t h)
+{
+	unsigned char r[8];
+	r[0] = (unsigned char)((h >>  0) & 255);
+	r[1] = (unsigned char)((h >>  8) & 255);
+	r[2] = (unsigned char)((h >> 16) & 255);
+	r[3] = (unsigned char)((h >> 24) & 255);
+	r[4] = (unsigned char)((h >> 32) & 255);
+	r[5] = (unsigned char)((h >> 40) & 255);
+	r[6] = (unsigned char)((h >> 48) & 255);
+	r[7] = (unsigned char)((h >> 56) & 255);
+	return *(uint_least64_t *)r;
+}
+#endif
+
 void
 libblake_blake2b_init(struct libblake_blake2b_state *state, const struct libblake_blake2b_params *params, const unsigned char *key)
 {
@@ -18,46 +37,57 @@ libblake_blake2b_init(struct libblake_blake2b_state *state, const struct libblak
 	state->f[0] = 0;
 	state->f[1] = 0;
 
-	state->h[0] ^= ((uint_least64_t)params->digest_len & 255) << 0;
-	state->h[0] ^= ((uint_least64_t)params->key_len & 255) << 8;
-	state->h[0] ^= ((uint_least64_t)params->fanout & 255) << 16;
-	state->h[0] ^= ((uint_least64_t)params->depth & 255) << 24;
-	state->h[0] ^= (uint_least64_t)(params->leaf_len & UINT_LEAST32_C(0xFFFFffff)) << 32;
-	state->h[1] ^= params->node_offset & UINT_LEAST64_C(0xFFFFffffFFFFffff);
-	state->h[2] ^= ((uint_least64_t)params->node_depth & 255) << 0;
-	state->h[2] ^= ((uint_least64_t)params->inner_len & 255) << 8;
-	state->h[4] ^= ((uint_least64_t)params->salt[0] & 255) << 0;
-	state->h[4] ^= ((uint_least64_t)params->salt[1] & 255) << 8;
-	state->h[4] ^= ((uint_least64_t)params->salt[2] & 255) << 16;
-	state->h[4] ^= ((uint_least64_t)params->salt[3] & 255) << 24;
-	state->h[4] ^= ((uint_least64_t)params->salt[4] & 255) << 32;
-	state->h[4] ^= ((uint_least64_t)params->salt[5] & 255) << 40;
-	state->h[4] ^= ((uint_least64_t)params->salt[6] & 255) << 48;
-	state->h[4] ^= ((uint_least64_t)params->salt[7] & 255) << 56;
-	state->h[5] ^= ((uint_least64_t)params->salt[8] & 255) << 0;
-	state->h[5] ^= ((uint_least64_t)params->salt[9] & 255) << 8;
-	state->h[5] ^= ((uint_least64_t)params->salt[A] & 255) << 16;
-	state->h[5] ^= ((uint_least64_t)params->salt[B] & 255) << 24;
-	state->h[5] ^= ((uint_least64_t)params->salt[C] & 255) << 32;
-	state->h[5] ^= ((uint_least64_t)params->salt[D] & 255) << 40;
-	state->h[5] ^= ((uint_least64_t)params->salt[E] & 255) << 48;
-	state->h[5] ^= ((uint_least64_t)params->salt[F] & 255) << 56;
-	state->h[6] ^= ((uint_least64_t)params->pepper[0] & 255) << 0;
-	state->h[6] ^= ((uint_least64_t)params->pepper[1] & 255) << 8;
-	state->h[6] ^= ((uint_least64_t)params->pepper[2] & 255) << 16;
-	state->h[6] ^= ((uint_least64_t)params->pepper[3] & 255) << 24;
-	state->h[6] ^= ((uint_least64_t)params->pepper[4] & 255) << 32;
-	state->h[6] ^= ((uint_least64_t)params->pepper[5] & 255) << 40;
-	state->h[6] ^= ((uint_least64_t)params->pepper[6] & 255) << 48;
-	state->h[6] ^= ((uint_least64_t)params->pepper[7] & 255) << 56;
-	state->h[7] ^= ((uint_least64_t)params->pepper[8] & 255) << 0;
-	state->h[7] ^= ((uint_least64_t)params->pepper[9] & 255) << 8;
-	state->h[7] ^= ((uint_least64_t)params->pepper[A] & 255) << 16;
-	state->h[7] ^= ((uint_least64_t)params->pepper[B] & 255) << 24;
-	state->h[7] ^= ((uint_least64_t)params->pepper[C] & 255) << 32;
-	state->h[7] ^= ((uint_least64_t)params->pepper[D] & 255) << 40;
-	state->h[7] ^= ((uint_least64_t)params->pepper[E] & 255) << 48;
-	state->h[7] ^= ((uint_least64_t)params->pepper[F] & 255) << 56;
+	if (offsetof(struct libblake_blake2b_params, inner_len) == 17) {
+		state->h[0] ^= le64(((uint_least64_t *)params)[0]);
+		state->h[1] ^= le64(((uint_least64_t *)params)[1]);
+		state->h[2] ^= le64(((uint_least64_t)params->node_depth << 0) |
+		                    ((uint_least64_t)params->inner_len << 8));
+		state->h[4] ^= le64(*(uint_least64_t *)&params->salt[0]);
+		state->h[5] ^= le64(*(uint_least64_t *)&params->salt[8]);
+		state->h[6] ^= le64(*(uint_least64_t *)&params->pepper[0]);
+		state->h[7] ^= le64(*(uint_least64_t *)&params->pepper[8]);
+	} else {
+		state->h[0] ^= ((uint_least64_t)params->digest_len & 255) << 0;
+		state->h[0] ^= ((uint_least64_t)params->key_len & 255) << 8;
+		state->h[0] ^= ((uint_least64_t)params->fanout & 255) << 16;
+		state->h[0] ^= ((uint_least64_t)params->depth & 255) << 24;
+		state->h[0] ^= (uint_least64_t)(params->leaf_len & UINT_LEAST32_C(0xFFFFffff)) << 32;
+		state->h[1] ^= params->node_offset & UINT_LEAST64_C(0xFFFFffffFFFFffff);
+		state->h[2] ^= ((uint_least64_t)params->node_depth & 255) << 0;
+		state->h[2] ^= ((uint_least64_t)params->inner_len & 255) << 8;
+		state->h[4] ^= ((uint_least64_t)params->salt[0] & 255) << 0;
+		state->h[4] ^= ((uint_least64_t)params->salt[1] & 255) << 8;
+		state->h[4] ^= ((uint_least64_t)params->salt[2] & 255) << 16;
+		state->h[4] ^= ((uint_least64_t)params->salt[3] & 255) << 24;
+		state->h[4] ^= ((uint_least64_t)params->salt[4] & 255) << 32;
+		state->h[4] ^= ((uint_least64_t)params->salt[5] & 255) << 40;
+		state->h[4] ^= ((uint_least64_t)params->salt[6] & 255) << 48;
+		state->h[4] ^= ((uint_least64_t)params->salt[7] & 255) << 56;
+		state->h[5] ^= ((uint_least64_t)params->salt[8] & 255) << 0;
+		state->h[5] ^= ((uint_least64_t)params->salt[9] & 255) << 8;
+		state->h[5] ^= ((uint_least64_t)params->salt[A] & 255) << 16;
+		state->h[5] ^= ((uint_least64_t)params->salt[B] & 255) << 24;
+		state->h[5] ^= ((uint_least64_t)params->salt[C] & 255) << 32;
+		state->h[5] ^= ((uint_least64_t)params->salt[D] & 255) << 40;
+		state->h[5] ^= ((uint_least64_t)params->salt[E] & 255) << 48;
+		state->h[5] ^= ((uint_least64_t)params->salt[F] & 255) << 56;
+		state->h[6] ^= ((uint_least64_t)params->pepper[0] & 255) << 0;
+		state->h[6] ^= ((uint_least64_t)params->pepper[1] & 255) << 8;
+		state->h[6] ^= ((uint_least64_t)params->pepper[2] & 255) << 16;
+		state->h[6] ^= ((uint_least64_t)params->pepper[3] & 255) << 24;
+		state->h[6] ^= ((uint_least64_t)params->pepper[4] & 255) << 32;
+		state->h[6] ^= ((uint_least64_t)params->pepper[5] & 255) << 40;
+		state->h[6] ^= ((uint_least64_t)params->pepper[6] & 255) << 48;
+		state->h[6] ^= ((uint_least64_t)params->pepper[7] & 255) << 56;
+		state->h[7] ^= ((uint_least64_t)params->pepper[8] & 255) << 0;
+		state->h[7] ^= ((uint_least64_t)params->pepper[9] & 255) << 8;
+		state->h[7] ^= ((uint_least64_t)params->pepper[A] & 255) << 16;
+		state->h[7] ^= ((uint_least64_t)params->pepper[B] & 255) << 24;
+		state->h[7] ^= ((uint_least64_t)params->pepper[C] & 255) << 32;
+		state->h[7] ^= ((uint_least64_t)params->pepper[D] & 255) << 40;
+		state->h[7] ^= ((uint_least64_t)params->pepper[E] & 255) << 48;
+		state->h[7] ^= ((uint_least64_t)params->pepper[F] & 255) << 56;
+	}
 
 	if (params->key_len) {
 		state->t[0] = 128;