Re: [Axiom-developer] differences

[Gabriel Dos Reis]

Waldek Hebisch <address@hidden> writes:

| > Ok, the problem isn't in axiom and it appears that it isn't in GCL.
| > I built the same GCL sources (no axiom) on redhat 9 and fedora 5.
| > The result of
| > 
| > (cos (acos 0.7)) on Redhat is 0.69999999999....96
| > (cos (acos 0.7)) on Fedora is 0.70000000000...07
| > 
| > so the problem change is likely somewhere in a system library
| > that gets linked into GCL. Since every other combination of
| > sins, cos, logs, exps, etc ends up with 0.69999.... I have
| > to conclude that there is a very minor bug in the Fedora lib.
| > 
| 
| I see the first result on 32-bit machines and the second one
| on 64-bit ones.  AFAICS the difference is in least significant bit.
| Since Lisp does not specify accuracy it is hard to call such difference
| a bug.  To repeatable results one have to sacrifce some speed.
| The are long flame wars related to this, see for example:
| 
| http://gcc.gnu.org/bugzilla/show_bug.cgi?id=323

That is only remotely related to this.  Assuming that is fixed, you're
not going to repeat the same digiton Sparc and other architectures
that don't use 80 bits for long double.  Please, please, use something
that is not common across all major plaftorms: C's double which is
mapped to 64 bits, with clear semantics described by IEEE-754.

-- Gaby