Bug 15536 - ulp (0x0.0p0) is not FP_SUBNORMAL!
Summary: ulp (0x0.0p0) is not FP_SUBNORMAL!
Status: RESOLVED FIXED
Alias: None
Product: glibc
Classification: Unclassified
Component: math (show other bugs)
Version: 2.18
: P2 normal
Target Milestone: 2.18
Assignee: Carlos O'Donell
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2013-05-26 09:29 UTC by Andreas Schwab
Modified: 2014-06-13 17:36 UTC (History)
2 users (show)

See Also:
Host:
Target: powerpc*-*-*
Build:
Last reconfirmed:
fweimer: security-

Attachments
Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this bug.
Description Andreas Schwab 2013-05-26 09:29:18 UTC 
The IBM long double version of ulp(0) returns 0, because scalbnl(1,-1127) is 0.
Comment 1 Carlos O'Donell 2013-05-27 17:31:31 UTC 
Fixing this ASAP.
Comment 2 Carlos O'Donell 2013-05-31 05:38:50 UTC 
Is there a publicly available standard that describes IBM extended precision long doubles?

The smallest value I can compute is 2^(-1075):
4.94065645841246544177e-324
0x0.00000000000010000000p-1022

The smallest computed value matches the AIX expectations for the, I assume, similar type.

I'm going to go with the smallest computed value as ulp of zero. 

Why? 

Because the first double represents the number, and if it's mantissa is zero and the exponent is zero, then it's zero. The extra precision from the second double is useless for subnormal values. Therefore the smallest subnormal is identical to that for double.
Comment 3 jsm-csl@polyomino.org.uk 2013-05-31 19:28:32 UTC 
On Fri, 31 May 2013, carlos at redhat dot com wrote:

> Is there a publicly available standard that describes IBM extended precision
> long doubles?

libgcc/config/rs6000/ibm-ldouble-format in a GCC source tree.

> The smallest value I can compute is 2^(-1075):
> 4.94065645841246544177e-324
> 0x0.00000000000010000000p-1022

You have an off-by-one error there.  It's 0x1p-1074.
Comment 4 Carlos O'Donell 2013-06-03 20:23:32 UTC 
Resolved.

commit 8b0ccb2d7fd1ec646a622a16bd64e356739ffca3
Author: Carlos O'Donell <carlos@redhat.com>
Date:   Mon Jun 3 14:49:48 2013 -0400

    BZ #15536: Fix ulp for 128-bit IBM long double.
    
    In 128-bit IBM long double the precision of the type
    decreases as you approach subnormal numbers, equaling
    that of a double for subnormal numbers. Therefore
    adjust the computation in ulp to use 2^(MIN_EXP - MANT_DIG)
    which is correct for FP_SUBNORMAL for all types.