diff --git a/libavcodec/x86/fmtconvert.asm b/libavcodec/x86/fmtconvert.asm
index b4220e21ea..8f62a0a093 100644
--- a/libavcodec/x86/fmtconvert.asm
+++ b/libavcodec/x86/fmtconvert.asm
@@ -107,7 +107,14 @@ cglobal int32_to_float_fmul_array8, 5, 5, 5, c, dst, src, mul, len
     add     mulq, 4
     add     lenq, 32
     jl .loop
-    REP_RET
+%if notcpuflag(sse2)
+    ;; cvtpi2ps switches to MMX even if the source is a memory location
+    ;; possible an error in documentation since every tested CPU disagrees with
+    ;; that. Use emms anyway since the vast majority of machines will use the
+    ;; SSE2 variant
+    emms
+%endif
+    RET
 %endmacro
 
 INIT_XMM sse