Re: [PATCH] Update sparc ULPs.

[David Miller <davem at davemloft dot net>]

From: "Joseph S. Myers" <joseph@codesourcery.com>
Date: Sun, 18 Nov 2012 23:33:41 +0000

> On Sun, 18 Nov 2012, David Miller wrote:
> 
>> I've done some investigation meanwhile and we end up down in
>> atanl(), via atan2l().
>> 
>> atan2l: x[-2.083218e+306L] y[1.755781e+174L]
>> 
>> atanl: x[-7.672687e+306L]
>> 
>> And the atanl() is what generates the spurious underflow.
> 
> I don't know if this is an "atan of large argument" issue, but I see we 
> don't have any direct tests in the testsuite for atan / atan2 of large 
> finite arguments at all.  Maybe we should have tests of atan (+/- 
> max_value) and atan2 (+/- max_value, +/-min_value) - expected results all 
> being +/- M_PI_2l.  And for that matter, atan2 (+/- min_value, max_value), 
> expected results +/- M_PIl, to verify that that doesn't produce any 
> spurious underflows either (as it would for a naive conversion from atan2 
> to atan).

Yes, this is exactly what the issue is here, and I have a patch
I'm testing which returns {+,-}pi/2 immediately for large values
instead of trying to compute the polynomial.

I'll add some tests too.