Re: [Caml-list] NaN reresentations

[Andreas Rossberg <rossberg@mpi-sws.org>]

On Sep 2, 2015, at 21:14 , Xavier Leroy <Xavier.Leroy@inria.fr> wrote:

> 2015-09-02 10:56 GMT-07:00 Andreas Rossberg <rossberg@mpi-sws.org>:
> > Does the Ocaml implementation make guarantees about the stable representation of floats? In particular, if I use Int64.float_of_bits to create a particular NaN representation, am I guaranteed that its bit pattern is maintained no matter where the value is stored or passed?
> 
> It depends on the underlying hardware.  For instance, with x86-32 bits, some FP moves go through the x87 FP stack, undergoing a double -> extended -> double conversion.   These conversions turn signaling NaNs into quiet NaNs, and I'm not sure they preserve the other bits of the NaN payload.
> 
> On other platforms, esp. x86-64 bits, I'm pretty confident that NaN bits are preserved by copying and parameter passing.

I see. Okay, in that case it seems safer not to rely on it, for the sake of portability.

> > We are currently in the process of implementing a reference interpreter for a little low-level language, and that tries to be as accurate as possible about float representations.
> 
> One possibility would be to represent your floats as int64 values (= their bit-level representation), and convert only when you operate over them, e.g.
> 
> let fp_add x y = Int64.bits_of_float (Int64.float_of_bits x +. Int64.float_of_bits y)

Right, that is the approach we are currently discussing. but I was wondering whether it is necessary.

Thanks!

/Andreas