[PATCH] powerpc: Add optimized ilogbf128 for POWER9

[Paul E Murphy]

On 1/4/21 5:20 PM, Paul E Murphy via Libc-alpha wrote:
> 
> 
> On 12/22/20 9:30 AM, Raphael Moreira Zinsly via Libc-alpha wrote:
>> The instruction xsxexpqp introduced on POWER9 extracts the exponent
>> from a quad-precision floating-point, thus it can be used to improve
>> ilogbf128 and llogbf128.
>> ---
>>   .../powerpc/powerpc64/le/fpu/e_ilogbf128.c    | 22 +++++++++++++++++++
>>   1 file changed, 22 insertions(+)
>>   create mode 100644 sysdeps/powerpc/powerpc64/le/fpu/e_ilogbf128.c
>>
>> diff --git a/sysdeps/powerpc/powerpc64/le/fpu/e_ilogbf128.c 
>> b/sysdeps/powerpc/powerpc64/le/fpu/e_ilogbf128.c
>> new file mode 100644
>> index 0000000000..47558bbadc
>> --- /dev/null
>> +++ b/sysdeps/powerpc/powerpc64/le/fpu/e_ilogbf128.c
>> @@ -0,0 +1,22 @@
>> +#ifdef _ARCH_PWR9
>> +int _ilogbf128 (_Float128 __x);
> 
> This should be a locally (static) scoped function.
> 
>> +
>> +int
>> +#if defined(_F128_ENABLE_IFUNC)
>> +__ieee754_ilogbf128_power9 (_Float128 __x)
>> +#else
>> +__ieee754_ilogbf128 (_Float128 __x)
>> +#endif
>> +{
>> +  /* Check for exceptional cases.  */
>> +  if (!__builtin_vsx_scalar_test_data_class_qp (__x, 0x7f))
>> +    return __builtin_vsx_scalar_extract_expq (__x) - 0x3fff;
>> +  else
>> +    /* Fallback to the generic ilogb if __x is NaN, Inf or 
>> subnormal.  */
>> +    return _ilogbf128(__x);
>> +}
>> +
>> +#define __ieee754_ilogbf128 _ilogbf128
>> +#endif
>> +
>> +#include<sysdeps/ieee754/float128/e_ilogbf128.c>
> 
> A space seems to be missing between include and <.
> 
> Otherwise, LGTM.
> 
> As a side note, I think the benchtests are not too impressive. I am 
> surprised normal values don't show better results.

After spending a little time looking at this, the call overhead of the 
wrapper is hiding most of the improvement.  Similarly, power9 adds 
similar instructions for float32/float64.

I would recommend refactoring this patch to provide an override to 
w_ilogb_template.c so all three formats can use these new instructions 
without the call overhead for normal numbers.