[prev in list] [next in list] [prev in thread] [next in thread] 

List:       sbcl-devel
Subject:    Re: [Sbcl-devel] self-hosted failure on float math discrepancy
From:       Christophe Rhodes <csr21 () cantab ! net>
Date:       2019-04-10 22:43:29
Message-ID: 8736mp7a7i.fsf () cantab ! net
[Download RAW message or body]

Christophe Rhodes <csr21@cantab.net> writes:

> In this instance, this is an intermediate value in a calculation, which
> is even worse; what if we have currently baked in the wrong answer, and
> the subsequent log2 is off on various platforms by 2ulp?  I guess,
> thanks to the new floating point oracle, we will at least have the same
> wrong answer everywhere...

For what it's worth, I believe that (log (exp $1.0d0) $2.0d0) should be
spelt as (make-flonum #x3FF71547652B82FE 'double-float).  (It could have
a name, maybe LOG2E$ ?)

While I'm looking at things, I'm squinting at the cross-compiler
most-fooative float definitions in src/code/cross-float, and I'm not
convinced.  Primarily I am not convinced because I remember at one point
having an SBCL with a bug in its floating point reader (off by 1ulp),
and you can imagine the hilarity in tracking down the cross-compilation
artifacts that that produced.  Since we know what the bit patterns will
be for our most-fooative floats, shouldn't we use those directly?
Likewise pi.  (Or else, convince me that what we have is right!)

Cheers,

Christophe



_______________________________________________
Sbcl-devel mailing list
Sbcl-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sbcl-devel
[prev in list] [next in list] [prev in thread] [next in thread]