Bug 14732 – [2.068 beta] Failing unittest in std.math
Status
RESOLVED
Resolution
FIXED
Severity
blocker
Priority
P1
Component
phobos
Product
D
Version
D2
Platform
All
OS
Linux
Creation time
2015-06-25T09:14:07Z
Last change time
2017-12-12T23:18:07Z
Assigned to
No Owner
Creator
Iain Buclaw
Comments
Comment #0 by ibuclaw — 2015-06-25T09:14:07Z
https://github.com/D-Programming-Language/phobos/pull/3285
I tested this using -mlong-double-64 on GDC.
Test:
---
real x = exp(-0x1.64p+9L);
The assert checks that the returned result is 0x0.06f84920bb2d3p-1022L
However, the actual value returned is 0x0.06f84920bb2d4p-1022L, where it appears that the last digit is rounded up.
To be sure, I even tested this using libm, and found that it also agreed with me, the value the unittest expects is wrong by 1 digit.
Now, maybe testing using -mlong-double-64 on a x86_64 machine is contrived, but given that I found a discrepancy for this test (and this test only) strikes me as not very assuring, and other targets may fail in the same way.
Comment #1 by bugzilla — 2015-06-26T22:09:56Z
Is this a regression?
Comment #2 by ibuclaw — 2015-06-29T09:08:51Z
(In reply to Walter Bright from comment #1)
> Is this a regression?
A change was introduced that caused tests to start failing with GDC.
Granted this is the first response I've received regarding this issue. Despite highlighting problems in the PR 3285, and despite raising a new PR to address some of problems I raised. What is going on with the review process?
Comment #3 by ibuclaw — 2015-06-29T11:05:39Z
Changing to blocker as (imo) a release shouldn't happen until it has been probably vetted as working.
Comment #4 by lt.infiltrator — 2015-10-31T13:13:50Z
Considering that this bug was rated as a blocker and 2.068 is already out, is it safe to assume that it has been fixed?
Comment #5 by ibuclaw — 2017-12-12T23:18:07Z
I have no idea if anything was done about it. I'll have to check on ARM, or some other target that has 64bit reals.