https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71050
--- Comment #11 from Segher Boessenkool <segher at gcc dot gnu.org> --- No, please go ahead, I couldn't find an easy way out. The generic code is hell-bent on using a subreg of the DF reg. The backend won't necessarily use any nop here btw, but the testcase should be pretty stable with just checking for any nop, it is small enough generated code. Might want to add a comment saying exactly what it tries to test (i.e., there should be enough nops to force the load into a different issue group than the store is in).