Clang insists that "cmp" is ambiguous with a memory destination,
requiring an explicit size suffix.

There was a true error in the use of .cfi_def_cfa_offset in the
epilogue, but changing to use the proper .cfi_adjust_cfa_offset
runs afoul of a clang bug wrt .cfi_restore_state.  Better to
fold the two epilogues so that we don't trigger the bug.
Signed-off-by: NRichard Henderson <rth@twiddle.net>