Re: [PATCH 8/9] i386: Move hypot implementation to C

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 06/10/2021 16:19, Adhemerval Zanella wrote:
> 
> 
> On 06/10/2021 15:37, Joseph Myers wrote:
>> On Wed, 6 Oct 2021, Adhemerval Zanella via Libc-alpha wrote:
>>
>>> +/* The i386 allows ot use the default excess of precision to optimize the
>>
>> s/ot/to/
> 
> Ack.
> 
>>
>>> +   hypot implementation, since internal multiplication and sqrt is carried
>>> +   with 80-bit FP type.  */
>>> +double
>>> +__ieee754_hypot (double x, double y)
>>> +{
>>> +  if ((isinf (x) || isinf (y))
>>> +      && !issignaling (x) && !issignaling (y))
>>> +    return INFINITY;
>>> +  if (isnan (x) || isnan (y))
>>> +    return x + y;
>>> +
>>> +  double r = math_narrow_eval (sqrt (x * x + y * y));
>>
>> There is no guarantee of when excess precision might or might not be used 
>> for intermediate computations, or whether it might not be used (so 
>> resulting in spurious overflows).  I'd expect explicit conversions of x 
>> and y to long double here, use of sqrtl and conversion of the result to 
>> double.
>>
> 
> Something like the below?
> 
>   long double lx = x;
>   long double ly = y;
>   double r = sqrtl (lx * lx + ly * ly);
>   return r;
> 

In fact we need to handle underflow:

  long double lx = x;
  long double ly = y;
  double r = sqrtl (lx * lx + ly * ly);
  math_check_force_underflow_nonneg (r);
  return r;