Re: [hatari-devel] FPU update

[ Thread Index | Date Index | More Archives ]

Am 18.01.2017 um 21:50 schrieb Christian Zietz <czietz@xxxxxxx>:

> Andreas Grabher schrieb:
>> Did you ever hear about
>> this difference between x87/IEEE and 6888x? It seems this behavior
>> does not comply with the IEEE standard. 
> It doesn't? From what I understand, IEEE 754-1985 only specifies very
> little about extended formats which are said to be
> "implementation-dependent". I can't find neither the bias nor the
> specific encoding of (de-normalized) numbers in the standard document
> for extended formats.

I didn't read through the standard. So this formula: (-1) ^ s x 2 ^ (1 - b) x 0.f for denormals might be only true for single and double precision.
> In particular, in Motorola's implementation with the explicit integer
> bit it would lead to redundant/ambiguous encoding of some numbers if the
> bias changed either for de-normalized numbers or generally when the
> unbiased exponent becomes 0. So I can see why Motorola has implemented
> it the way they did.

I agree that this implementation would make sense. But x87 obviously has the other implementation, which is also supported by SoftFloat, which claims to be IEEE compliant. That confuses me.
>> Or might this even be a
>> mistake in the data sheets (i checked different versions of it, they
>> all tell the same)?
> Wouldn't the easiest way be to check this on actual hardware?

Absolutely. But i don't have any 68k hardware. So if anyone wants to test this, that would be very helpful.
Some changes to SoftFloat's core functions would be required. I don't want to touch them without some reliable information about how this should work. 
> Regards
> Christian
> -- 
> Christian Zietz  -  CHZ-Soft  -  czietz@xxxxxxx
> WWW:
> PGP/GnuPG-Key-ID: 0x52CB97F66DA025CA / 0x6DA025CA
Thank you for the reply. Best wishes, Andreas

Mail converted by MHonArc 2.6.19+