[ET Trac] [Einstein Toolkit] #1276: Intel 2013.1.117 mis-compiles NewRad

Einstein Toolkit trac-noreply at einsteintoolkit.org
Sun Mar 3 16:38:23 CST 2013


#1276: Intel 2013.1.117 mis-compiles NewRad
------------------------------------+---------------------------------------
  Reporter:  rhaas                  |       Owner:        
      Type:  defect                 |      Status:  review
  Priority:  minor                  |   Milestone:        
 Component:  EinsteinToolkit thorn  |     Version:        
Resolution:                         |    Keywords:  NewRad
------------------------------------+---------------------------------------

Comment (by rhaas):

 I do not fully understand what triggers the compiler bug. Unfortunately I
 do not have a short test case that reproduces it. Only NewRad. The error
 manifested as a segfault when the code was run, it was easy to detect. To
 diagnose I added printf statements at various places and ran the code in
 gdb. Eventually I found that the error is triggered in the very first
 iteration of the loop in extrap_kernel at which point i,j,k had the values
 that bmin was initialized with :-INT_MAX,-INT_MAX/2,-INT_MAX/3 in my case.
 Adding printf statements at the wrong spot removes the error eg adding a
 printf(%d %d %d\n", bmin[0],...) removes it, indep. of whether it is added
 in the driver or in the extrap routine, which in itself is odd since it
 would hint at some kind of cross-procedure optimization.

 I cannot really lower the optimization settings below -O2 (for the whole
 code) and thus will use a different compiler version. As far are removing
 the statements goes: we'd never be able to remove them since this compiler
 version will always be buggy. Only when no one uses the compiler anymore
 could we do so.

 I am unfamiliar with the procedure of submitting a bug report to intel.
 Can one submit all of newrad as the failing routine or is a stand-alone
 code that demonstrates the issue required?

-- 
Ticket URL: <https://trac.einsteintoolkit.org/ticket/1276#comment:5>
Einstein Toolkit <http://einsteintoolkit.org>
The Einstein Toolkit


More information about the Trac mailing list