fix chroma mc2 bug, this is based on a patch by (Oleg Metelitsa oleg hitron co kr)

and does slow the mc2 chroma put down, avg interrestingly seems unaffected speedwise on duron
this of course should be rather done in a way which doesnt slow it down but its better a few %
slower but correct then incorrect

Originally committed as revision 8093 to svn://svn.ffmpeg.org/ffmpeg/trunk

diff --git a/libavcodec/i386/dsputil_h264_template_mmx.c b/libavcodec/i386/dsputil_h264_template_mmx.c
index 8fa9eac09e2d01d46cb8d11577240bab128d9579..48cb4802288693bc0f6906e7ae1919167a424588 100644 (file)
--- a/libavcodec/i386/dsputil_h264_template_mmx.c
+++ b/libavcodec/i386/dsputil_h264_template_mmx.c
@@ -293,11 +293,14 @@ static void H264_CHROMA_MC2_TMPL(uint8_t *dst/*align 2*/, uint8_t *src/*align 1*
         /* writes garbage to the right of dst.
             * ok because partitions are processed from left to right. */
         H264_CHROMA_OP4((%0), %%mm1, %%mm3)
-        "movd %%mm1, (%0)\n\t"
+        "movd %%mm1, %%esi\n\t"
+        "movw %%si, (%0)\n\t"
         "add %4, %0\n\t"
         "sub $1, %2\n\t"
         "jnz 1b\n\t"
-        : "+r" (dst), "+r"(src), "+r"(h) : "m" (ff_pw_32), "r"(stride));
+        : "+r" (dst), "+r"(src), "+r"(h)
+        : "m" (ff_pw_32), "r"(stride)
+        : "%esi");
 
 }
 #endif