diff --git a/postproc/swscale.c b/postproc/swscale.c
index 7afb61b2b4db01f2fe7617a208edf935c2aa3503..8073552a3cdfbe0356b00c3d43167a2d3f08466e 100644
--- a/postproc/swscale.c
+++ b/postproc/swscale.c
@@ -9,7 +9,6 @@
 #include <string.h>
 #include "../config.h"
 #include "swscale.h"
-#include "../mmx_defs.h"
 #include "../cpudetect.h"
 #undef MOVNTQ
 #undef PAVGB
diff --git a/postproc/swscale_template.c b/postproc/swscale_template.c
index 256d9333658e3daf7732c8c39c2f51f931ebf296..3445f92f302bac30c7637b2dd47ad5a0bc19b912 100644
--- a/postproc/swscale_template.c
+++ b/postproc/swscale_template.c
@@ -7,6 +7,34 @@
 
 #undef MOVNTQ
 #undef PAVGB
+#undef PREFETCH
+#undef PREFETCHW
+#undef EMMS
+#undef SFENCE
+
+#ifdef HAVE_3DNOW
+/* On K6 femms is faster of emms. On K7 femms is directly mapped on emms. */
+#define EMMS     "femms"
+#else
+#define EMMS     "emms"
+#endif
+
+#ifdef HAVE_3DNOW
+#define PREFETCH  "prefetch"
+#define PREFETCHW "prefetchw"
+#elif defined ( HAVE_MMX2 )
+#define PREFETCH "prefetchnta"
+#define PREFETCHW "prefetcht0"
+#else
+#define PREFETCH "/nop"
+#define PREFETCHW "/nop"
+#endif
+
+#ifdef HAVE_MMX2
+#define SFENCE "sfence"
+#else
+#define SFENCE "/nop"
+#endif
 
 #ifdef HAVE_MMX2
 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"