🚀 go-pugleaf

Dean Schulze (schulze@lpl.arizona.edu) wrote:
:     The evidence against gcc 2.7.2 under Linux continues to mount.  I would
: appreciate it if some of you can try this test on your Linux machines.

:     I compiled the paranoia test program (netlib.att.com, netlib/paranoia)
: both with and without optimization using gcc 2.7.2 on Linux 1.2.13.
: With optimization off paranoia detected a flaw in the arithmetic
: rounding mode (perhaps controlled by a compiler switch).  This is not
: considered a serious flaw, and paranoia pronounced the arithmetic to
: be OK.  However, when I compiled with -O2 paranoia found some serious
: flaws and hung before completing the test.

Linux defaults to use "extended" precision, which is better than "double",
but paranoia doesn't seem to know about this. You can set the rounding
mode to double with a function whose prototype is defined in fpu_control.h
No flaw anymore.

As for the optimization switch, from the gcc manpage

     -ffloat-store
          Do not store floating  point  variables  in  registers.
          This  prevents undesirable excess precision on machines
          such as the 68000 where the floating registers (of  the
          68881) keep more precision than a double is supposed to
          have.

          For most programs, the excess precision does only good,
          but  a  few  programs rely on the precise definition of
          IEEE floating point.  Use `-ffloat-store' for such pro-
          grams.

Paranoia seems to be one of those programs.

:     I compiled paranoia on the same machine using Visual C++ under
: Windows 95 both with and without optimization and in both cases
: paranoia found no problems and pronounced the arithmetic to be perfect.

:     This doesn't prove that my hardware works correctly in all cases, but
: it does show that gcc's optimization is seriously flawed.

See above.

Carlos

--
Carlos Fonseca
C.Fonseca@shef.ac.uk