[PATCH] Convert 0.0/0.0 to NAN

Craig Howland howland@LGSInnovations.com
Thu Apr 20 00:25:00 GMT 2017


On 04/19/2017 08:01 PM, Richard Allen wrote:
> Whoops, that was the wrong patch. Attached is correct patch.
>
> On Wed, Apr 19, 2017 at 6:50 PM, Richard Allen <rsaxvc@rsaxvc.net> wrote:
>> With newlib's default compiler flags(-O2 and whatnot),
>> the GCC I'm using creates calls to __aeabi_(f|d)div.
>> This causes a little extra register copying
>> and an unnecessary branch.
>>
>> Additionally, __aeabi_?div may throw exceptions,
>> which is probably not what we want a
>> simple assignment to do.
>>
>> Compiler version tested:
>> arm-none-eabi-gcc (15:5.4.1+svn241155-1) 5.4.1 20160919
>>
>> -Richard
- exc.retval = 0.0/0.0;
+ exc.retval = NAN;

Well, actually, an exception is exactly why it is coded that way; yes, this is 
what this simple assignment is intended to do.  The 0.0/0.0 construct is used so 
that at runtime the floating point flags gets set and yield the NAN result; a 
floating point exception is supposed to be raised.  (The compiler is not allowed 
to optimize it away because of the side effects--which is why you're seeing the 
divide even with -O2.  Although I would guess that in an implementation which 
does not have the side effects, the compiler would be able to substitute the NAN 
at compile time.)
Craig



More information about the Newlib mailing list