diff --git a/CHANGES b/CHANGES
index 69d6408..eecd219 100644
--- a/CHANGES
+++ b/CHANGES
@@ -18,6 +18,8 @@
     unless you #define _NETBSD_SOURCE, and then their CMSG_* macros use
     it.  This is horrible even by OpenBSD standards)
   remove support for ip6.int (it's no longer delegated)
+  add asm versions of imult64 and umult64 for x86_64
+    (22 cycles -> 12 cycles on my Athlon 64)
 
 0.24:
   fix scan_to_sa (Tim Lorenz)
diff --git a/mult/imult64.c b/mult/imult64.c
index be59d66..961d226 100644
--- a/mult/imult64.c
+++ b/mult/imult64.c
@@ -1,3 +1,24 @@
+#ifdef __x86_64__
+
+void imult64() {
+  asm volatile(
+    "xchgq %rdx,%rsi\n"
+    "movq %rdi,%rax\n"
+    "imulq %rdx\n"
+    "jc 1f\n"	/* overflow */
+    "movq %rax,(%rsi)\n"
+    "xorq %rax,%rax\n"
+    "inc %rax\n"
+    "ret\n"
+    "1:\n"
+    "xorq %rax,%rax\n"
+    /* the closing ret is renerated by gcc */
+    );
+}
+
+
+#else
+
 #include "safemult.h"
 
 int imult64(int64 a,int64 b,int64* c) {
@@ -11,3 +32,4 @@ int imult64(int64 a,int64 b,int64* c) {
   return 1;
 }
 
+#endif
diff --git a/mult/umult64.c b/mult/umult64.c
index 6d5e37a..5b644dd 100644
--- a/mult/umult64.c
+++ b/mult/umult64.c
@@ -1,3 +1,23 @@
+#ifdef __x86_64__
+
+void umult64() {
+  asm volatile(
+    "xchgq %rdx,%rsi\n"
+    "movq %rdi,%rax\n"
+    "mulq %rdx\n"
+    "jc 1f\n"	/* overflow */
+    "movq %rax,(%rsi)\n"
+    "xorq %rax,%rax\n"
+    "inc %rax\n"
+    "ret\n"
+    "1:\n"
+    "xorq %rax,%rax\n"
+    /* the closing ret is renerated by gcc */
+    );
+}
+
+#else
+
 #include "safemult.h"
 
 /* return 1 for overflow, 0 for ok */
@@ -20,3 +40,4 @@ int umult64(uint64 a,uint64 b,uint64* c) {
   return 1;
 }
 
+#endif