Re: [PATCH] Fix sign of input to bsloww1 (BZ #16623)

[Siddhesh Poyarekar <siddhesh at redhat dot com>]

On Tue, Feb 25, 2014 at 06:01:03PM +0000, Joseph S. Myers wrote:
> (I'm wondering how useful it really is to have libm-test-ulps files 
> containing details of the errors for individual test inputs, rather than 
> just the final bit with maximal errors of functions.  There would 
> certainly be a lot less churn to those files from addition of tests if 
> they only had the final bit.)

Could you please elaborate on how the final bit (I assume you mean the
final bit of the mantissa of the input) would be useful?  I would have
thought that *just* the maximal error for a function (maybe
per-architecture, but not necessarily) would be sufficient, given that
we already do such a manual check when one updates ULPs; in fact I
briefly pondered over such an approach when I was trying to generate
test outputs for the new test cases and it seemed like the right thing
to do.

Siddhesh

Attachment: pgp6TpO9k6XB5.pgp
Description: PGP signature