Re: [PATCH v2 5/5] math: Remove the error handling wrapper from fmod and fmodf

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

Hi Adhemerval,

> It shows an small improvement, the results for fmod:

>  x86_64 (Ryzen 9) | close-exponents | 16.0244  | 13.119
>  aarch64 (N1)     | close-exponents | 8.21306  | 5.76138

I wouldn't call gains of 22% on Ryzen and 42% on N1 for the common case small!

> @@ -56,11 +58,13 @@ __ieee754_fmod (double x, double y)
>... 
> -  if (__glibc_unlikely (hy == 0        || hx >= EXPONENT_MASK || hy > EXPONENT_MASK))
> +  if (__glibc_unlikely (is_nan (hx) || is_nan (hy)))
>     return (x * y) / (x * y);
> +  if (__glibc_unlikely (hy == 0 || is_inf (hx)))
> +    return __math_edom ((x * y) / (x * y));
 
This now requires 4 compares rather than 2 of the previous if statement (as GCC is able
to merge the 2 tests on hy). It's better to keep that and then handle the special cases.

 
> @@ -59,8 +61,10 @@ __ieee754_fmodf (float x, float y)
>...
>+  if (__glibc_unlikely (hy == 0 || is_inf (hx)))
>+    return __math_edom ((x * y) / (x * y));

__math_edomf?


> diff --git a/sysdeps/ieee754/flt-32/math_config.h b/sysdeps/ieee754/flt-32/math_config.h
>...
>+attribute_hidden float __math_edom (float x);

__math_edomf?

Cheers,
Wilco