[Xorp-hackers] [PATCH] Support for IEEE754 binary64 format in XRLs

Ben Greear greearb at candelatech.com
Thu Sep 8 09:09:36 PDT 2011


On 09/08/2011 07:35 AM, Steven Simpson wrote:
> On 01/09/11 16:31, Ben Greear wrote:
>> On 09/01/2011 02:50 AM, Steven Simpson wrote:
>>> On 31/08/11 16:18, Ben Greear wrote:
>>
>>>> In general, I'm amazed at how much code this took, but from skimming
>>>> through
>>>> it quickly, I didn't notice any issues.  I guess floating-point
>>>> serialization
>>>> is just a pain in the arse to do right!
>>>
>>> It's a matter of recognizing that double has many implementation-defined
>>> characteristics, so it can't just be shifted and masked like an
>>> integer.  frexp, ldexp, etc allow it to be expressed more portably, and
>>> shouldn't be too costly on most systems.
>>
>> Could you just convert it to a string and pass it that way?
>
> It has been suggested, but advice around the web often seems to point to
> frexp etc, and you get a fixed-size representation out of it.
>
>> But, as long as your stuff compiles and works on stuff at least as
>> old as Fedora 5, then that's good enough for me.
>
> I seem to be hitting one brick wall after another, just trying to set up
> a fresh FC5 system that XORP can run on.
>
> However, I did try just compiling fp64serial.c (with fp64serial.h and
> fp64.h), with gcc -std=c99, and produced no errors.  It's not so
> surprising - the features I'm after are basically a few macros that
> ought to have been in place for some time when FC5 came out.

Ok, I have a build machine on FC5 running the automated regression
builds and tests, so if it's broken, it will show up quickly.

It is a pain to get FC5 working..had to manually upgrade python and/or scons
it seems.

Please repost your patch when you're happy and I'll apply it.  Also, be sure
to re-sync against upstream..I've been pushing a bunch of changes lately, some
of which touches the xrl code lightly.

Thanks,
Ben

-- 
Ben Greear <greearb at candelatech.com>
Candela Technologies Inc  http://www.candelatech.com



More information about the Xorp-hackers mailing list