ARC math test regressions (was Re: [PATCH v7 00/13] glibc port to ARC processors)

[Vineet Gupta]

On 7/2/20 9:27 AM, Joseph Myers wrote:
> On Thu, 2 Jul 2020, Vineet Gupta via Libc-alpha wrote:
> 
>> Now acos can come from 2 places depending on LIBM_SVID_COMPAT
>>  - math/w_acos_template.c	<-- sets errno only
> 
> This wrapper, unlike the compat one, always calls __ieee754_acos 
> (appropriately suffixed), relying on it to raise appropriate exceptions.  
> The implementation in sysdeps/ieee754/dbl-64/e_asin.c raises the exception 
> by doing an Inf / Inf computation.
> 
>     u.i[HIGH_HALF]=0x7ff00000;
>     v.i[HIGH_HALF]=0x7ff00000;
>     u.i[LOW_HALF]=0;
>     v.i[LOW_HALF]=0;
>     return u.x/v.x;
> 
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95115

Aha. I almost forgot that this was upstream gcc-10 for ARC too (and problem
doesn't happen with gcc-9). So my glibc updates didn't trigger this, phew !

> The right place to fix this is in GCC, to stop it constant-folding Inf / 
> Inf when flag_trapping_math.  That should be a straightforward fix.
> 
> It would also be reasonable to change the above code to just "return 0.0 / 
> 0.0;", as a simpler, more idiomatic way of returning a NaN with "invalid" 
> raised.  (Any code constructing infinities / NaNs via bit patterns likely 
> predates built-in functions that can be used to represent such values 
> directly.  Though the GCC bug means a literal replacement with "return 
> __builtin_inf () / __builtin_inf ();" wouldn't fix the present problem.)