[Bug math/16447] erfc (0x6.a8p+4) ldbl-128 throws underflow exception

["joseph at codesourcery dot com" <sourceware-bugzilla at sourceware dot org>]

http://sourceware.org/bugzilla/show_bug.cgi?id=16447

--- Comment #4 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
On Thu, 6 Feb 2014, stli at linux dot vnet.ibm.com wrote:

> unsafe = abs(n_i) >= -16363; 

Well, abs >= (negative value) doesn't seem right.

To work out what's unsafe, you need to consider how small x22 can be in 
cases where ex2_u.d is small.  The answer is roughly that x22, if not 0, 
is at least 1ulp of the input (not exactly, because of the way values from 
a table get subtracted from the input) - so, roughly, you could have 
problems up to n_i around -16382 + 113.  Certainly saying unsafe = abs 
(n_i) >= 15000 should be safe, even allowing for the way x gets adjusted.

> How to handle the unsafe-check?
> Which platform uses this implementation of expl?
> Can we simply change the unsafe-check in sysdeps/iee754 or is a new
> platform-specific implementation required?

The same correction should apply to all platforms using ldbl-128; no 
platform-specific implementation should be needed.

-- 
You are receiving this mail because:
You are on the CC list for the bug.