[Bug 2095501] [NEW] valgrind can compute incorrectly rounded results on arm64

[tom stevens]

Public bug reported:

I encountered regression failures running the gpsbabel test suite.  The
computed results differ if I run gpsbabel directly versus running
gpsbabel under valgrind.  I have only observed this problem on arm64.
Note that this is not a problem with the output of valgrind, it is a
problem with the output of the program run by valgrind.

I expect the output of a program to be identical if I run it with or
without valgrind.

I backported valgrind_3.23.0-0ubuntu2 to noble.  It did not resolve the
issue.

I then created a ppa with the latest upstream version of valgrind,
3.24.0.  This resolved the issue.  The ppa is at
https://launchpad.net/~tsteven4/+archive/ubuntu/valgrind-backports,
version 1:3.24.0-0ubuntu1~noble11.  As noted in the changelog I deleted
the patch 11_arm64-cache-flush.patch because it conflicted with the
upstream fix for the issue that necessitated the patch.  I did not
review the remaining  patches to see if they have been resolved
upstream, although I note several of them originated with the vendor and
may or may not have been fixed upstream.

The original regression failures were on:

Description:	Ubuntu 24.04.1 LTS
Release:	24.04
valgrind:
  Installed: 1:3.22.0-0ubuntu3
  Candidate: 1:3.22.0-0ubuntu3
  Version table:
 *** 1:3.22.0-0ubuntu3 500
        500 http://ports.ubuntu.com/ubuntu-ports noble/main arm64 Packages
        100 /var/lib/dpkg/status

** Affects: valgrind (Ubuntu)
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Ubuntu
Foundations Bugs, which is subscribed to valgrind in Ubuntu.
https://bugs.launchpad.net/bugs/2095501

Title:
  valgrind can compute incorrectly rounded results on arm64

Status in valgrind package in Ubuntu:
  New

Bug description:
  I encountered regression failures running the gpsbabel test suite.
  The computed results differ if I run gpsbabel directly versus running
  gpsbabel under valgrind.  I have only observed this problem on arm64.
  Note that this is not a problem with the output of valgrind, it is a
  problem with the output of the program run by valgrind.

  I expect the output of a program to be identical if I run it with or
  without valgrind.

  I backported valgrind_3.23.0-0ubuntu2 to noble.  It did not resolve
  the issue.

  I then created a ppa with the latest upstream version of valgrind,
  3.24.0.  This resolved the issue.  The ppa is at
  https://launchpad.net/~tsteven4/+archive/ubuntu/valgrind-backports,
  version 1:3.24.0-0ubuntu1~noble11.  As noted in the changelog I
  deleted the patch 11_arm64-cache-flush.patch because it conflicted
  with the upstream fix for the issue that necessitated the patch.  I
  did not review the remaining  patches to see if they have been
  resolved upstream, although I note several of them originated with the
  vendor and may or may not have been fixed upstream.

  The original regression failures were on:

  Description:	Ubuntu 24.04.1 LTS
  Release:	24.04
  valgrind:
    Installed: 1:3.22.0-0ubuntu3
    Candidate: 1:3.22.0-0ubuntu3
    Version table:
   *** 1:3.22.0-0ubuntu3 500
          500 http://ports.ubuntu.com/ubuntu-ports noble/main arm64 Packages
          100 /var/lib/dpkg/status

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/valgrind/+bug/2095501/+subscriptions