Re: Partial success building the trunk on MSYS2/Windows platform

[Gaius Mulley]

Gaius Mulley <gaiusmod2@gmail.com> writes:

> "Fischlin  Andreas" <andreas.fischlin@env.ethz.ch> writes:
>
>> Dear Gaius,
>>
>> Follow up: For our portable floating point arithmetic module
>> FloatPortab (Dialogue Machine) we have also implemented checks in
>> terms of type lengths that may succeed or not across hard- and
>> software (compilers) by
>> running a little overhead checking code during module
>> initialisation. This needs only to be done once and then you can
>> rely on correct interpretation and any numerical use of routines
>> from that module have to pay no
>> performance penalty at all.
>
> Hi Andreas,
>
> sounds great!  Currently gm2 can be requested to check against a NaN
> after every floating point operation -fsoft-check-all, highlighting the
> operator in the source code which caused the result,

I just performed a test on a game engine (with/without -fsoft-check-all)
and there was no noticable drop in frames per second.  Interestingly all
the nan detection is being implemented inline even with -O0 :-)

regards,
Gaius