floating point woes on i386

[Greg Troxel <gdt%lexort.com@localhost>]

I've been running regression tests for geography/proj.  They pass on
amd64 and aarch64.  On 9 i386 and 10 i386 two of them have errors of
about 5m compared to the expected "within 1 mm".

I ran paranoia (pkgsrc/benchmarks/paranoia) on 9 and 10.  On 9:

  The number of  FAILUREs  encountered =       3.
  The number of  SERIOUS DEFECTs  discovered = 4.
  The number of  DEFECTs  discovered =         3.
  The number of  FLAWs  discovered =           2.

which is shocking, especially as I remember earlier NetBSD i386 versions
being ok, like 5 or 7.

So, I am curious if anyone can run paranoia on any of

  5
  6
  7
  8
  9
  10
  11
  current

and to reduce traffic, let me know offlist if you get anything much
different from the above.  (If your results area disaster but not quite
the same numbers, that's not so critical for now.)

If you think paranoia is broken, that's interesting too and probably
best to be onlist.


(I'm going to run it in earmv7hf-el too, but haven't finished proj
testing yet.)