netbsd-7 i386 floating point troubles

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

(I know this maybe belongs on tech-toolchain but want to reach a wider
audience to test.)

While testing an update to proj 6 (geospatial coordinate conversions), I
noticed that the tests failed on NetBSD 7 i386, after passing on NetBSD
8 amd64.

I then ran paranoia (pkgsrc/benchmarks/paranoia), which reported
multiple serious problems:

  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.

  The arithmetic diagnosed has unacceptable Serious Defects.
  Potentially fatal FAILURE may have spoiled this program's subsequent diagnoses.
  

paranoia passss on NetBSD 8 amd64 and NetBSD 8 evbearmv7hf-el.  By pass,
I mean:

  No failures, defects nor flaws have been discovered.
  Rounding appears to conform to the proposed IEEE standard P754.
  The arithmetic diagnosed appears to be Excellent!

I tried a second physical machine under netbsd-7 i386, and got the same
failures.  My two machines have:

cpu0 at mainbus0 apid 0: Intel(R) Core(TM)2 Duo CPU     T8100  @ 2.10GHz, id 0x10676
cpu0 at mainbus0: Geode(TM) Integrated Processor by AMD PCS, id 0x5a2

which do not seem likely to have common bugs, so I am suspecting this is
a NetBSD issue.

I am wondering if anyone else is seeing issues.

paranoia is easy to run; just run the binary and hit return a lot and
'y<cr>' when asked about computing 1/0 and 0/0.

Long ago I noted some issues on i386, and that sparc and sparc64 where
ok.

I'd be interested in reports on 6/7/8/current, i386, amd64, RPI, plus of
course any failures regardless.