This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: Don't handle ulps for integer tests in libm-test.inc
- From: Andreas Jaeger <aj at suse dot com>
- To: libc-alpha at sourceware dot org
- Date: Fri, 17 May 2013 20:24:04 +0200
- Subject: Re: Don't handle ulps for integer tests in libm-test.inc
- References: <Pine dot LNX dot 4 dot 64 dot 1305171525240 dot 21342 at digraph dot polyomino dot org dot uk>
On 05/17/2013 05:27 PM, Joseph S. Myers wrote:
libm-test.inc handles "ulps" values not just for tests with
floating-point results, but for those with integer results as well.
However, this handling has various problems:
* The code computing differences and absolute values could have signed
integer overflows.
* All integer results should always be exact and any deviation from
expected results is a bug which should be filed in Bugzilla (and if
necessary result in the test being disabled with a comment referring
to the bug) rather than a quality-of-implementation accuracy issue
that it's appropriate to handle through ulps settings.
* The ulps values output, if they do not overflow, end up getting
converted to the FLOAT type, possibly changing the value in some
cases.
This patch changes libm-test.inc not to consider or output ulps in
these cases, so check_int, check_long and check_longlong more closely
correspond to check_bool which also does not consider ulps. Tested
x86_64 and x86.
Ok - and I look forward to your followup patches,
thanks,
Andreas
--
Andreas Jaeger aj@{suse.com,opensuse.org} Twitter/Identica: jaegerandi
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn,Jennifer Guild,Felix Imendörffer,HRB16746 (AG Nürnberg)
GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126