I think Mark's original pointer to issue1762561 was right on.  The last
two cases are failing due to the mixed-endian format (mentioned
in that issue) used in OABI.

You can see this in the output of 'test_endian_double':

   '182D4454FB210940' != 'FB210940182D4454'

Note that the values are the same except the two 32-bit words are
swapped.

Similarly, in 'test_unaligned_native_struct_fields':

   '123412007856341200B81E09401F85EB51' !=
   '1234120078563412001F85EB51B81E0940'

The first 8 bytes in each case are the same.   The last 8 bytes of
each (which represent floating-point doubles) are the same except,
again, the words are swapped.

I am going to close this out as "won't fix".  As mentioned in issue1762561, supporting OABI will involve taking on another
host platform.  EABI is surely more predominant these days anyway.