diff --git a/ChangeLog b/ChangeLog
index d51898b..b0ce2e2 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -3,6 +3,7 @@
 	* src/sha512.h, src/sha512.c: New.
 
 	* src/sha256.c (initial_state): Don't export, it's internal.
+	(memcpy_output_bswap32): Rename and remove last argument.
 
 2014-01-28  Niibe Yutaka  <gniibe@fsij.org>
 
diff --git a/src/sha256.c b/src/sha256.c
index c948018..7267014 100644
--- a/src/sha256.c
+++ b/src/sha256.c
@@ -52,13 +52,17 @@
 
 #define SHA256_MASK (SHA256_BLOCK_SIZE - 1)
 
-static void memcpy_bswap32 (void *dst, const uint32_t *p, int n)
+static void memcpy_output_bswap32 (unsigned char *dst, const uint32_t *p)
 {
-  uint32_t *q = (uint32_t *)dst;
+  int i;
+  uint32_t q;
 
-  n >>= 2;
-  while (n--)
-    q[n] = __builtin_bswap32 (p[n]); /* bswap32 is GCC extention */
+  for (i = 0; i < 32; i++)
+    {
+      if ((i & 3) == 0)
+	q[i] = __builtin_bswap32 (p[i >> 2]); /* bswap32 is GCC extention */
+      dst[i] = q >> ((i & 3) * 8);
+    }
 }
 
 #define rotr32(x,n)   (((x) >> n) | ((x) << (32 - n)))
@@ -193,7 +197,7 @@ sha256_finish (sha256_context *ctx, unsigned char output[32])
   ctx->wbuf[15] = __builtin_bswap32 (ctx->total[0] << 3);
   sha256_process (ctx);
 
-  memcpy_bswap32 (output, ctx->state, SHA256_DIGEST_SIZE);
+  memcpy_output_bswap32 (output, ctx->state);
   memset (ctx, 0, sizeof (sha256_context));
 }
 
diff --git a/src/sha512.c b/src/sha512.c
index c9e3897..640d9c7 100644
--- a/src/sha512.c
+++ b/src/sha512.c
@@ -37,7 +37,7 @@
 
 #define SHA512_MASK (SHA512_BLOCK_SIZE - 1)
 
-static void memcpy_bswap64 (unsigned char dst[64], const uint64_t *p)
+static void memcpy_output_bswap64 (unsigned char dst[64], const uint64_t *p)
 {
   int i;
   uint64_t q;
@@ -187,7 +187,7 @@ sha512_finish (sha512_context *ctx, unsigned char output[64])
   ctx->wbuf[15] = __builtin_bswap64 (ctx->total[0] << 3);
   sha512_process (ctx);
 
-  memcpy_bswap64 (output, ctx->state);
+  memcpy_output_bswap64 (output, ctx->state);
   memset (ctx, 0, sizeof (sha512_context));
 }