[illumos-Developer] Ugh... xdr_{float,double} problems

[Garrett D'Amore]

On 02/ 1/11 06:55 AM, Deano wrote:
> A compiler error would be preferred it I was porting, I rather have the
> compiler mark "Here be dragons" then find it after hours of hard debugging
> some obscure rounding error later on.
>
> Done a fair amount of platform porting and I always prefer the compiler to
> just stop than try and patch over things that aren't really patch over-able.
>    

Not disagreeing with that point, but don't we have the potential to have 
a reasonable "portable" version here that works for all platforms 
(albeit slower)?

Having such a portable version might also aid people who want to develop 
a CPU specific optimized version, since then they'll have a known good 
implementation to compare against.

     - Garrett

> Just My 2p,
> Deano
> deano at cloudpixies.com
>
> -----Original Message-----
> From: Jason King [mailto:jason.brian.king at gmail.com]
> Sent: 01 February 2011 14:26
> To: Joerg Schilling
> Cc: developer at lists.illumos.org
> Subject: Re: [illumos-Developer] Ugh... xdr_{float,double} problems
>
> On Tue, Feb 1, 2011 at 3:30 AM, Joerg Schilling
> <Joerg.Schilling at fokus.fraunhofer.de>  wrote:
>    
>> "Garrett D'Amore"<garrett at nexenta.com>  wrote:
>>
>>      
>>> It sounds like all the main CPU targets support IEEE 754.  The only
>>> reasonable (?) question IMO is S390.  Roland?  Neale?
>>>        
>> If I interpret the floating point code written by David M. Gay from AT&T
>> correctly, then IBM and VAX are the two mayor deviations from IEEE.
>>
>> But HPPA uses at least different parameters for long double than other
>>      
> IEEE
>    
>> architectures do.
>>
>> Jörg
>>      
> Just to get back to the question at hand.  Is it reasonable to assume
> if someone is going to port Illumos to a platform that doesn't support
> IEEE 754 to have them write conversion code (that can utilize
> knowledge of the native format during the conversion) instead of a
> generic routine that is both slow and very likely to introduce greater
> rounding errors?
>
> If so, I will just remove the portable code altogether and just emit a
> compiler error to notify porters.  If not, I will revert it to an
> #ifdef to select between the two implementations.
>
> _______________________________________________
> Developer mailing list
> Developer at lists.illumos.org
> http://lists.illumos.org/m/listinfo/developer
>
>
> _______________________________________________
> Developer mailing list
> Developer at lists.illumos.org
> http://lists.illumos.org/m/listinfo/developer
>