gosthash2012: Optimize out temporary `buf' from `stage3'

[openssl-gost/engine.git] / gosthash2012.c

diff --git a/gosthash2012.c b/gosthash2012.c
index 0868d6c1c2113d90ac920cf8e76a24adce8edb2d..ddd35fad45c7cf32faefc58d933f88378e098a13 100644 (file)
--- a/gosthash2012.c
+++ b/gosthash2012.c
@@ -158,28 +158,19 @@ static INLINE void stage2(gost2012_hash_ctx * CTX, const union uint512_u *data)
 
 static INLINE void stage3(gost2012_hash_ctx * CTX)
 {
-    ALIGN(16) union uint512_u buf;
-
-    memset(&buf, 0x00, sizeof buf);
-    memcpy(&buf, &(CTX->buffer), CTX->bufsize);
-    memcpy(&(CTX->buffer), &buf, sizeof(uint512_u));
+    pad(CTX);
+    g(&(CTX->h), &(CTX->N), &(CTX->buffer));
+    add512(&(CTX->Sigma), &CTX->buffer);
 
-    memset(&buf, 0x00, sizeof buf);
+    memset(&(CTX->buffer.B[0]), 0, sizeof(uint512_u));
 #ifndef __GOST3411_BIG_ENDIAN__
-    buf.QWORD[0] = CTX->bufsize << 3;
+    CTX->buffer.QWORD[0] = CTX->bufsize << 3;
 #else
-    buf.QWORD[0] = BSWAP64(CTX->bufsize << 3);
+    CTX->buffer.QWORD[0] = BSWAP64(CTX->bufsize << 3);
 #endif
-
-    pad(CTX);
-
-    g(&(CTX->h), &(CTX->N), &(CTX->buffer));
-
-    add512(&(CTX->N), &buf);
-    add512(&(CTX->Sigma), &CTX->buffer);
+    add512(&(CTX->N), &(CTX->buffer));
 
     g(&(CTX->h), &buffer0, &(CTX->N));
-
     g(&(CTX->h), &buffer0, &(CTX->Sigma));
 }
 
@@ -190,33 +181,35 @@ static INLINE void stage3(gost2012_hash_ctx * CTX)
 void gost2012_hash_block(gost2012_hash_ctx * CTX,
                          const unsigned char *data, size_t len)
 {
-    size_t chunksize;
-
-    while (len > 63 && CTX->bufsize == 0) {
-        memcpy(CTX->buffer.B, data, 64);
-        stage2(CTX, &(CTX->buffer));
+    register size_t chunksize;
+    register size_t bufsize = CTX->bufsize;
 
-        data += 64;
-        len -= 64;
+    if (bufsize == 0) {
+        while (len >= 64) {
+            memcpy(&CTX->buffer.B[0], data, 64);
+            stage2(CTX, &(CTX->buffer));
+            data += 64;
+            len -= 64;
+        }
     }
 
     while (len) {
-        chunksize = 64 - CTX->bufsize;
+        chunksize = 64 - bufsize;
         if (chunksize > len)
             chunksize = len;
 
-        memcpy(&CTX->buffer.B[CTX->bufsize], data, chunksize);
+        memcpy(&CTX->buffer.B[bufsize], data, chunksize);
 
-        CTX->bufsize += chunksize;
+        bufsize += chunksize;
         len -= chunksize;
         data += chunksize;
 
-        if (CTX->bufsize == 64) {
-            stage2(CTX, &(CTX->buffer));
-
-            CTX->bufsize = 0;
+        if (bufsize == 64) {
+            stage2(CTX, &(CTX->buffer) );
+            bufsize = 0;
         }
     }
+    CTX->bufsize = bufsize;
 }
 
 /*