Don't optimize by hand (Was Re: Lack of fabsf on Solaris(patch included))

Wed Jan 31 04:56:00 GMT 2001

Andrew Haley wrote:
> Anthony Green writes:
> > Kazuyuki-san wrote:
> > > I have tried only a little old GCJ, which says,
> > > gcc version 2.95.2 19991024 (release).
> > > It just does not implement the strictfp.
> >
> > Did you compile your test case on x86 with -ffloat-store?
>> I don't think it will make any difference to this test case.

Yes. Store-reload technique (I suppose -ffloat-store
does it) cannot accomplish the strict FP semantics by
itself. Scaling is needed. Using SSE2 instructions
instead of x86 FPU may be another solution.
The problem x86 FPU involves and solutions are described
in the following document:
 http://math.nist.gov/javanumerics/reports/jgfnwg-01.html
It was written by the Java Grande Forum Numerics Working Group.
See the section `Issue 4: Use of floating-point hardware'.
I have talked about the problem and implementation of
the strictfp in a workshop. There are transparencies of
that talk here:
 http://www.shudo.net/publications/java-hpc2000/
Kazuyuki SHUDO				Happy Hacking!
 Muraoka Laboratory, Waseda University