optimization/6368: ia64 bad code for LAPACK with -O2 -funroll-all-loops

[janis187@us.ibm.com]

>Number:         6368
>Category:       optimization
>Synopsis:       ia64 bad code for LAPACK with -O2 -funroll-all-loops
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Thu Apr 18 17:26:01 PDT 2002
>Closed-Date:
>Last-Modified:
>Originator:     Janis Johnson
>Release:        gcc version 3.1 20020418 (prerelease)
>Organization:
>Environment:
ia64-unknown-linux-gnu
Itanium, Red Hat Linux release 7.1.94 (Roswell)
>Description:
LAPACK test output file csb.out reports 120 unexpected
failures for ia64-linux when the tests are compiled with
-O2 -funroll-all-loops by GCC-3.1-20020418.  There are many
other failures as well, possibly all related to the one
described here.

LAPACK takes a couple of hours to build and test on an
Itanium system and I haven't yet checked to see if this is
a regression from 3.0.4.

I can continue to try to come up with a small test case but
in the meantime, here's a way to reproduce the problem.  If
someone looks into this using the full LAPACK distribution,
please let me know if my further efforts are not necessary.
>How-To-Repeat:
Build and test LAPACK as shown by the LAPACK build and test
guide (http://gcc.gnu.org/gcc-3.1/testing-lapack.html).
Verify that TESTING/csb.out shows no failures.  Modify the
OPTS line in make.inc to use -O2 -funroll-all-loops and do:

  cd TESTING/EIG
  rm chbt21.f
  make complex
  cd ..
  make csb.out

This time csb.out will say that 120 out of 540 tests failed
to pass the threshold.
>Fix:
Unknown
>Release-Note:
>Audit-Trail:
>Unformatted: