diff --git a/ssh.h b/ssh.h
index 4039b00b..2f946e83 100644
--- a/ssh.h
+++ b/ssh.h
@@ -258,11 +258,12 @@ void hmacmd5_key(void *handle, void const *key, int len);
 void hmacmd5_do_hmac(void *handle, unsigned char const *blk, int len,
 		     unsigned char *hmac);
 
-typedef struct {
+typedef struct SHA_State {
     uint32 h[5];
     unsigned char block[64];
     int blkused;
     uint32 lenhi, lenlo;
+    void (*sha1)(struct SHA_State * s, const unsigned char *p, int len);
 } SHA_State;
 void SHA_Init(SHA_State * s);
 void SHA_Bytes(SHA_State * s, const void *p, int len);
@@ -271,11 +272,12 @@ void SHA_Simple(const void *p, int len, unsigned char *output);
 
 void hmac_sha1_simple(void *key, int keylen, void *data, int datalen,
 		      unsigned char *output);
-typedef struct {
+typedef struct SHA256_State {
     uint32 h[8];
     unsigned char block[64];
     int blkused;
     uint32 lenhi, lenlo;
+    void (*sha256)(struct SHA256_State * s, const unsigned char *p, int len);
 } SHA256_State;
 void SHA256_Init(SHA256_State * s);
 void SHA256_Bytes(SHA256_State * s, const void *p, int len);
diff --git a/sshsh256.c b/sshsh256.c
index 4186f3e8..e83266b5 100644
--- a/sshsh256.c
+++ b/sshsh256.c
@@ -19,6 +19,8 @@
 #define smallsigma0(x) ( ror((x),7) ^ ror((x),18) ^ shr((x),3) )
 #define smallsigma1(x) ( ror((x),17) ^ ror((x),19) ^ shr((x),10) )
 
+static void SHA256_sw(SHA256_State *s, const unsigned char *q, int len);
+
 void SHA256_Core_Init(SHA256_State *s) {
     s->h[0] = 0x6a09e667;
     s->h[1] = 0xbb67ae85;
@@ -97,19 +99,24 @@ void SHA256_Init(SHA256_State *s) {
     SHA256_Core_Init(s);
     s->blkused = 0;
     s->lenhi = s->lenlo = 0;
+    s->sha256 = &SHA256_sw;
 }
 
 void SHA256_Bytes(SHA256_State *s, const void *p, int len) {
     unsigned char *q = (unsigned char *)p;
-    uint32 wordblock[16];
     uint32 lenw = len;
-    int i;
 
     /*
      * Update the length field.
      */
     s->lenlo += lenw;
     s->lenhi += (s->lenlo < lenw);
+    (*(s->sha256))(s, q, len);
+}
+
+static void SHA256_sw(SHA256_State *s, const unsigned char *q, int len) {
+    uint32 wordblock[16];
+    int i;
 
     if (s->blkused && s->blkused+len < BLKSIZE) {
         /*
diff --git a/sshsha.c b/sshsha.c
index c10a8217..711b08a6 100644
--- a/sshsha.c
+++ b/sshsha.c
@@ -13,6 +13,8 @@
 
 #define rol(x,y) ( ((x) << (y)) | (((uint32)x) >> (32-y)) )
 
+static void sha1_sw(SHA_State * s, const unsigned char *q, int len);
+
 static void SHA_Core_Init(uint32 h[5])
 {
     h[0] = 0x67452301;
@@ -124,20 +126,26 @@ void SHA_Init(SHA_State * s)
     SHA_Core_Init(s->h);
     s->blkused = 0;
     s->lenhi = s->lenlo = 0;
+    s->sha1 = &sha1_sw;
 }
 
 void SHA_Bytes(SHA_State * s, const void *p, int len)
 {
     const unsigned char *q = (const unsigned char *) p;
-    uint32 wordblock[16];
     uint32 lenw = len;
-    int i;
 
     /*
      * Update the length field.
      */
     s->lenlo += lenw;
     s->lenhi += (s->lenlo < lenw);
+    (*(s->sha1))(s, q, len);
+}
+
+static void sha1_sw(SHA_State * s, const unsigned char *q, int len)
+{
+    uint32 wordblock[16];
+    int i;
 
     if (s->blkused && s->blkused + len < 64) {
 	/*