sha: fix thinko in sha512; add FAST_FUNC to sha1/sha256

function                                             old     new   delta
sha512_process_block128                             1334    1336      +2
sha1_hash                                            112     111      -1
sha256_process_block64                               446     440      -6
sha1_process_block64                                 494     484     -10
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 1/3 up/down: 2/-17)             Total: -15 bytes

2 files changed, 5 insertions, 5 deletions

diff --git a/include/libbb.h b/include/libbb.h
index b6eab6f24..bc47ce7ba 100644
--- a/include/libbb.h
+++ b/include/libbb.h
@@ -1324,7 +1324,7 @@ typedef struct sha1_ctx_t {
         uint32_t hash[8];    /* 5, +3 elements for sha256 */
         uint64_t total64;
         uint8_t wbuffer[64]; /* NB: always correctly aligned for uint64_t */
-        void (*process_block)(struct sha1_ctx_t*);
+        void (*process_block)(struct sha1_ctx_t*) FAST_FUNC;
 } sha1_ctx_t;
 void sha1_begin(sha1_ctx_t *ctx) FAST_FUNC;
 void sha1_hash(const void *data, size_t length, sha1_ctx_t *ctx) FAST_FUNC;
diff --git a/libbb/sha1.c b/libbb/sha1.c
index 8ecf71a9d..ccc317667 100644
--- a/libbb/sha1.c
+++ b/libbb/sha1.c
@@ -56,7 +56,7 @@ static inline uint64_t hton64(uint64_t v)
 #define SHA1_BLOCK_SIZE  64
 #define SHA1_MASK        (SHA1_BLOCK_SIZE - 1)
 
-static void sha1_process_block64(sha1_ctx_t *ctx)
+static void FAST_FUNC sha1_process_block64(sha1_ctx_t *ctx)
 {
         unsigned i;
         uint32_t w[80], a, b, c, d, e, t;
@@ -203,7 +203,7 @@ static const uint32_t K512_lo[80] = {
 
 /* Process LEN bytes of BUFFER, accumulating context into CTX.
    LEN is rounded _down_ to 64.  */
-static void sha256_process_block64(sha256_ctx_t *ctx)
+static void FAST_FUNC sha256_process_block64(sha256_ctx_t *ctx)
 {
         unsigned t;
         uint32_t W[64];
@@ -266,7 +266,7 @@ static void sha256_process_block64(sha256_ctx_t *ctx)
 }
 /* Process LEN bytes of BUFFER, accumulating context into CTX.
    LEN is rounded _down_ to 128.  */
-static void sha512_process_block128(sha512_ctx_t *ctx)
+static void FAST_FUNC sha512_process_block128(sha512_ctx_t *ctx)
 {
         unsigned t;
         uint64_t W[80];
@@ -278,7 +278,7 @@ static void sha512_process_block128(sha512_ctx_t *ctx)
         uint64_t f = ctx->hash[5];
         uint64_t g = ctx->hash[6];
         uint64_t h = ctx->hash[7];
-        const uint32_t *words = (uint32_t*) ctx->wbuffer;
+        const uint64_t *words = (uint64_t*) ctx->wbuffer;
 
         /* Operators defined in FIPS 180-2:4.1.2.  */
 #define Ch(x, y, z) ((x & y) ^ (~x & z))