[PATCH v2 2/4] math: Optimize fma call on log2pf1

[Adhemerval Zanella Netto]

On 14/10/25 10:18, Wilco Dijkstra wrote:
> Hi Adhemerval,
> 
>> +++ b/sysdeps/ieee754/flt-32/s_log2p1f.c
>> @@ -231,7 +231,12 @@ __log2p1f (float x)
>>        int j = (m + ((int64_t) 1 << (52 - 8))) >> (52 - 7), k = j > 53;
>>        e += k;
>>        double xd = asdouble (m | (uint64_t) 0x3ff << 52);
>> -      z = fma (xd, ix[j], -1.0);
>> +#ifndef __FP_FAST_FMA
>> +      if (__glibc_unlikely (x == -0x1.da285cp-5f))
>> +       z = fma (xd, ix[j], -1.0);
>> +      else
>> +#endif
>> +        z = xd * ix[j] - 1.0;
> 
> This relies on fp contraction on targets with FMA. I think it's best to do
> the same as in previous patch, reverse the if condition and use fma in
> the else part.

Indeed, are you ok with the following:

#ifndef __FP_FAST_FMA
      /* The fma is required only for x == -0x1.da285cp-5f in FE_TONEAREST
         to provide correctly rounded results.  */
      if (__glibc_likely (x != -0x1.da285cp-5f))
        z = xd * ix[j] - 1.0;
      else
#endif
        z = fma (xd, ix[j], -1.0);

?