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, 53 insertions, 25 deletions

diff --git a/libblake_blake2s_init.c b/libblake_blake2s_init.c
index c4b126c..fae7e0c 100644
--- a/libblake_blake2s_init.c
+++ b/libblake_blake2s_init.c
@@ -1,6 +1,21 @@
 /* See LICENSE file for copyright and license details. */
 #include "common.h"
 
+#if defined(LITTLE_ENDIAN)
+# define le32(X) X
+#else
+static uint_least32_t
+le32(uint_least32_t h)
+{
+	unsigned char r[4];
+	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);
+	return *(uint_least32_t *)r;
+}
+#endif
+
 void
 libblake_blake2s_init(struct libblake_blake2s_state *state, const struct libblake_blake2s_params *params, const unsigned char *key)
 {
@@ -18,31 +33,44 @@ libblake_blake2s_init(struct libblake_blake2s_state *state, const struct libblak
 	state->f[0] = 0;
 	state->f[1] = 0;
 
-	state->h[0] ^= ((uint_least32_t)params->digest_len & 255) << 0;
-	state->h[0] ^= ((uint_least32_t)params->key_len & 255) << 8;
-	state->h[0] ^= ((uint_least32_t)params->fanout & 255) << 16;
-	state->h[0] ^= ((uint_least32_t)params->depth & 255) << 24;
-	state->h[1] ^= params->leaf_len & UINT_LEAST32_C(0xFFFFffff);
-	state->h[2] ^= (uint_least32_t)((params->node_offset >> 0) & UINT_LEAST64_C(0xFFFFffff));
-	state->h[3] ^= (uint_least32_t)((params->node_offset >> 32) & UINT_LEAST64_C(0xFFFF)) << 0;
-	state->h[3] ^= ((uint_least32_t)params->node_depth & 255) << 16;
-	state->h[3] ^= ((uint_least32_t)params->inner_len & 255) << 24;
-	state->h[4] ^= ((uint_least32_t)params->salt[0] & 255) << 0;
-	state->h[4] ^= ((uint_least32_t)params->salt[1] & 255) << 8;
-	state->h[4] ^= ((uint_least32_t)params->salt[2] & 255) << 16;
-	state->h[4] ^= ((uint_least32_t)params->salt[3] & 255) << 24;
-	state->h[5] ^= ((uint_least32_t)params->salt[4] & 255) << 0;
-	state->h[5] ^= ((uint_least32_t)params->salt[5] & 255) << 8;
-	state->h[5] ^= ((uint_least32_t)params->salt[6] & 255) << 16;
-	state->h[5] ^= ((uint_least32_t)params->salt[7] & 255) << 24;
-	state->h[6] ^= ((uint_least32_t)params->pepper[0] & 255) << 0;
-	state->h[6] ^= ((uint_least32_t)params->pepper[1] & 255) << 8;
-	state->h[6] ^= ((uint_least32_t)params->pepper[2] & 255) << 16;
-	state->h[6] ^= ((uint_least32_t)params->pepper[3] & 255) << 24;
-	state->h[7] ^= ((uint_least32_t)params->pepper[4] & 255) << 0;
-	state->h[7] ^= ((uint_least32_t)params->pepper[5] & 255) << 8;
-	state->h[7] ^= ((uint_least32_t)params->pepper[6] & 255) << 16;
-	state->h[7] ^= ((uint_least32_t)params->pepper[7] & 255) << 24;
+	if (offsetof(struct libblake_blake2s_params, inner_len) == 17) {
+		state->h[0] ^= le32(((uint_least32_t *)params)[0]);
+		state->h[1] ^= le32(((uint_least32_t *)params)[1]);
+		state->h[2] ^= le32((uint_least32_t)(params->node_offset >> 0));
+		state->h[3] ^= le32(((uint_least32_t)(params->node_offset >> 32) & UINT_LEAST64_C(0xFFFF)) |
+		                    ((uint_least32_t)params->node_depth << 16) |
+		                    ((uint_least32_t)params->inner_len << 24));
+		state->h[4] ^= le32(*(uint_least32_t *)&params->salt[0]);
+		state->h[5] ^= le32(*(uint_least32_t *)&params->salt[4]);
+		state->h[6] ^= le32(*(uint_least32_t *)&params->pepper[0]);
+		state->h[7] ^= le32(*(uint_least32_t *)&params->pepper[4]);
+	} else {
+		state->h[0] ^= ((uint_least32_t)params->digest_len & 255) << 0;
+		state->h[0] ^= ((uint_least32_t)params->key_len & 255) << 8;
+		state->h[0] ^= ((uint_least32_t)params->fanout & 255) << 16;
+		state->h[0] ^= ((uint_least32_t)params->depth & 255) << 24;
+		state->h[1] ^= params->leaf_len & UINT_LEAST32_C(0xFFFFffff);
+		state->h[2] ^= (uint_least32_t)((params->node_offset >> 0) & UINT_LEAST64_C(0xFFFFffff));
+		state->h[3] ^= (uint_least32_t)((params->node_offset >> 32) & UINT_LEAST64_C(0xFFFF)) << 0;
+		state->h[3] ^= ((uint_least32_t)params->node_depth & 255) << 16;
+		state->h[3] ^= ((uint_least32_t)params->inner_len & 255) << 24;
+		state->h[4] ^= ((uint_least32_t)params->salt[0] & 255) << 0;
+		state->h[4] ^= ((uint_least32_t)params->salt[1] & 255) << 8;
+		state->h[4] ^= ((uint_least32_t)params->salt[2] & 255) << 16;
+		state->h[4] ^= ((uint_least32_t)params->salt[3] & 255) << 24;
+		state->h[5] ^= ((uint_least32_t)params->salt[4] & 255) << 0;
+		state->h[5] ^= ((uint_least32_t)params->salt[5] & 255) << 8;
+		state->h[5] ^= ((uint_least32_t)params->salt[6] & 255) << 16;
+		state->h[5] ^= ((uint_least32_t)params->salt[7] & 255) << 24;
+		state->h[6] ^= ((uint_least32_t)params->pepper[0] & 255) << 0;
+		state->h[6] ^= ((uint_least32_t)params->pepper[1] & 255) << 8;
+		state->h[6] ^= ((uint_least32_t)params->pepper[2] & 255) << 16;
+		state->h[6] ^= ((uint_least32_t)params->pepper[3] & 255) << 24;
+		state->h[7] ^= ((uint_least32_t)params->pepper[4] & 255) << 0;
+		state->h[7] ^= ((uint_least32_t)params->pepper[5] & 255) << 8;
+		state->h[7] ^= ((uint_least32_t)params->pepper[6] & 255) << 16;
+		state->h[7] ^= ((uint_least32_t)params->pepper[7] & 255) << 24;
+	}
 
 	if (params->key_len) {
 		state->t[0] = 32;