X128R(xmm0, xmm2, xmm4, xmm6, xmm1, xmm3, xmm5, xmm7);
STORE(h, xmm0, xmm2, xmm4, xmm6);
-# ifndef __x86_64__
+# ifndef __i386__
/* Restore the Floating-point status on the CPU */
/* This is only required on MMX, but EXTRACT32 is using MMX */
_mm_empty();