jenkins hash: optimize last memcpy

compiler cannot optimize variable-size memcpy.
use a inlined simple version then.

That makes this version always faster than
Jenkins version on Core Duo.  On other cpu-s
its +3% win although still slower than Jenkins'.

diff --git a/src/hash.c b/src/hash.c
index d87e656aeec9965ab089440ad1766f53af7e54bc..c1531291edeae2e366b73a139d6aac9210137f99 100644 (file)
--- a/src/hash.c
+++ b/src/hash.c
@@ -56,7 +56,17 @@
        c ^= b; c -= rot(b,24); \
 } while (0)
 
-/* simple version - let compiler worry about memory access */
+/* for a small amount of bytes the call to libc is a loss */
+static inline void simple_memcpy(void *dst, const void *src, size_t n)
+{
+       const uint8_t *s = src;
+       uint8_t *d = dst;
+
+       while (n--)
+               *d++ = *s++;
+}
+
+/* short version - let compiler worry about memory access */
 uint32_t lookup3_hash(const void *data, size_t len)
 {
        uint32_t a, b, c;
@@ -78,7 +88,7 @@ uint32_t lookup3_hash(const void *data, size_t len)
        }
 
        buf[0] = buf[1] = buf[2] = 0;
-       memcpy(buf, p, len);
+       simple_memcpy(buf, p, len);
        a += buf[0];
        b += buf[1];
        c += buf[2];