Experimental Physics and
Industrial Control System

mdavidsaver via Core-talk <[email protected]> · Thu, 10 Oct 2019 22:38:12 -0000

I'll try to summarize the outcome of the core meeting last week.  For
epicsCalcTest the decision was to remove the cross-check part of the
test, and only check the the CALC express has the expected.  For
epicsMathTest, if was felt that trying to trick the MSVC optimizer just
to make the tests pass wasn't worthwhile, so these are marked TODO for
the compiler.

** Changed in: epics-base
    Milestone: None => 7.0.3.1

** Changed in: epics-base
     Assignee: (unassigned) => mdavidsaver (mdavidsaver)

** Changed in: epics-base
   Importance: Undecided => Low

** Changed in: epics-base
       Status: New => Confirmed

-- 
You received this bug notification because you are a member of EPICS
Core Developers, which is subscribed to EPICS Base.
Matching subscriptions: epics-core-list-subscription
https://bugs.launchpad.net/bugs/1845854

Title:
  EPICS Math tests failing on Windows with MSVC

Status in EPICS Base:
  Confirmed

Bug description:
  Some epics base tests involving NaN and Inf have been failing for most
  windows MSVC builds, I have written a small test program
  https://github.com/FreddieAkeroyd/Testing/blob/master/src/mytests/inftest.cpp
  with output at https://ci.appveyor.com/project/FreddieAkeroyd/testing
  which illustrates the underlying cause.

  The general summary is that:

  VS2010, VS2017 and VS2019 are OK in all configurations
  For other Visual studio versions:
      Debug builds are often OK  
      All release builds fail.

  The failure mode is as follows (Inf shown here but applies to NaN
  too):

  Inf – Inf is evaluated correctly
  Inf + -Inf is evaluated incorrectly
  Inf + (-Inf) is evaluated correctly on VS 11.0 and 12.0, but not 14.0

  The test code is using expressions like “a + -a” with a = Inf, if it
  instead tests “a + -b” with a = b = Inf then release mode works in all
  cases. So it looks like the repeated variable name is causing an
  undesired optimisation. I feel expressions of the form “a + -b” rather
  than “a + -a” are a better reflection of what is  likely in practice,
  if that change is acceptable then I would be happy to submit a PR to
  use different variable names within binary operation tests

  Regards,

  Freddie

To manage notifications about this bug go to:
https://bugs.launchpad.net/epics-base/+bug/1845854/+subscriptions

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System