[MLton-user] minor 32bit vs. 64bit differences in floating-point calculations with large numbers

[David Hansel]

Hi Wesley,

Thanks for that information.  Just one more question
(more out of curiosity than anything else):  is there
a technical reason that 64-bit MLton does NOT use the
FPU?  Or is this just a case of "nobody has implemented
that yet"?  If it is the latter,  do you have any (rough)
estimate on how much effort it would take to implement?

Thanks,

David


On 10/15/2010 6:50 AM, Wesley W. Terpstra wrote:
> On Thu, Oct 14, 2010 at 4:18 PM, David Hansel <hansel at reactive-systems.com <mailto:hansel at reactive-systems.com>> wrote:
> 
>     For example:
> 
>     (1E308 * 10.0) / 10.0 = 1E308 when compiled as 32-bit
>     (1E308 * 10.0) / 10.0 = inf   when compiled as 64-bit
> 
>  
> This is documented in the man page. 32-bit MLton uses (by default) the FPU and 64-bit uses SSE registers. This means that 32-bit MLton has slightly higher intermediate precision than a "normal" IEEE 754 point registers, but performs nicely. You can override this behaviour with "-ieee-fp true" which will force MLton to lose this precision for intermediate calculations.
> 
> Note: on i386, the results can also differ depending on when the compiler can no longer fit the intermediate floating point values in the FPU registers. YMMV.
> 
> If you care to know the precision MLton is using for a given platform, check regression/real-basic.sml You'll see that MLton behaves the same on every supported architecture except i386. (but don't depend on this!)

-- 
  ----------------------------------------------------------
  David Hansel
  Chief Technology Officer -- Reactive Systems, Inc.
  http://www.reactive-systems.com/
  (919) 324-3507 ext. 102 -- hansel at reactive-systems.com
  OpenPGP (GnuPG) public key file:
  http://www.reactive-systems.com/~hansel/pgp_public_key.txt
  ----------------------------------------------------------