[PATCH 13/23] math: Fix the expected atan2f (inf) results

[Adhemerval Zanella Netto]

On 03/12/24 16:03, Adhemerval Zanella Netto wrote:
> 
> 
> On 29/11/24 13:40, Joseph Myers wrote:
>> On Fri, 29 Nov 2024, Adhemerval Zanella wrote:
>>
>>> The pi defined constants are not the expected value for atan2
>>> on non-default rounding modes.  Instead use the autogenerated value.
>>
>> The same principle presumably also applies to (the imaginary part of) clog 
>> and clog10.  (And indeed expected errors for the imaginary part of clogf 
>> could also be removed from libm-test-ulps files if clog is changed to 
>> defer to atan2 for the imaginary part when called on a zero argument.)
>>
> 
> I will check clog10, and I think we can apply it to carg as well.

It turns out that although the new atan2f does improve clog10 imaginary
precision on some archs (on x86_64 max ulps now show 2 instead of 4 for
the tested inputs); it is still not correctly rounded (I guess because it
does not calculate in the expected range, since it requires a 
M_LOG10E multiplication after the atan2 call).

It does make the imaginary part of clog and the carg result correctly
rounded.  I will update the patch.