Where to report a bug?

[etienne-forest]

We have some problems with Intel Composer on Windows 7 and also on LINUX. Here I will report the problem on Windows 7 (Ultimate).

I put the full MS_developer project at http://www.takafumi.org/intel/

There is a jpeg too.

If one compiles debug using the latest composer .175, one gets the compiling error shown on the jpeg
If one compiles using release, then all is fine.

All is fine with the version .154. both release and debug.

It seems that Composer mixes up two different overloading of operators. This code runs fine on g95, Compaq CVF and older Intel.

Thanks

E. Forest

[mecej4 · Accepted solution]

There have been a couple of bug reports related to the ambitious /warn:interfaces option. For example:
Compiler errors with /warn:interfaces diagnostics option enabled .

If this option has served its purpose, i.e., if you can compile your project in the Release configuration without interface-related errors, simply deselecting this option (Configuration Properties->Fortran->Diagnostics->check routine interfaces) will enable your project to be built in debug mode.

[etienne-forest]

Thanks mecej4,

It did work. The question is whether this was a real bug or not..... Scary.

Etienne

[mecej4]

Certainly, it's a real bug in the compiler. Such bugs (i.e., those related to the complex interactions between compiler options) usually get fixed with the next update of the compiler (issued almost every month), but it's always nice to have a workaround so that one's work can proceed in the interim.

[Steven_L_Intel1]

As for "where to report a bug", you can do it here, as you did. Other options are shown here. This forum has a "Tech Support" link you can use, and the compiler release notes have more information on support.

I verified that this problem is fixed in the upcoming Update 4.

[Alexander_Molodozhen]

I work with Etienne Forest who reported a window problem in this thread. We posted a Composer XE 3.175 problem on LINUX at the same place:

www.takafumi.org/intel/linux_problem.gif

This is a huge program using Forest library (on which he reported a problem in this thread) linking with some C++ bigger multi-processor program.

We had to revert to our old version Intel 9 to get that to work. Intel composer 154 gave us the same problem.

Incidentally Composer complains at link time on intrisic sine and cosine used in this d_lielib.f90....

Is this a related problem: obviously interface mess up.

(all files available on request)


Sasha

[mecej4]

This new post (#5) seems to be about a different issue -- in fact, the GIF file shows the results of a failure to list SVML libraries on the link line.

I am puzzled by the note about having to revert to IFort 9. First of all, if view of the tendency of -warn xxx options to give false positives when run on old code, and their limited applicability to large, mixed language production code, I question why such options are used in your project at all. Secondly, IFort 9 was much less ambitious as to the kind of inconsistencies it tried to catch, so its failure to report errors is not necessarily a virtue.

I assume that you have many test results and benchmark cases. I think that compiling your code with at least -O2 and running in sequential mode would be a good starting point, to be followed up with parallelization, while ensuring that the results match the test results. Leave -warn, -C, etc. for the preliminary developmental stages, which it seems have been completed already.

But then, I know little about your code, and I may be projecting thoughts farther than I should, so take what I say with a good pinch of salt.

[etienne-forest]

Thanks mecej4,

Incidentally, we just compiled the same thing on LINUX on a CERN cluster where these guys are running INTEL 11.3.084 (Not sure about the 084). But anyway no SVML errors with 11.XXX either. Sasha and I will ask the CERN people to help us..... definitely scary if they upgrade to composer.

Etienne

[mecej4]

The latest update of IFort 11.1 for Linux is L_CPROF_P_11.1.075.TGZ; Intel has not issued any compiler with the second part of the version number other than 0 or 1; there was a 11.1.084, but that was a MAC OSX version.

One supposes that you use the Intel compiler for its high optimization capabilities. Although its diagnostic and checking capabilities are being improved, these features were not its strong points in earlier versions, and checking the results using other compilers is to be advised in the development/debugging stage.

Be aware that versions earlier than 11.1.073 (for Linux ; the third part of the version number may be different for other platforms) had a nasty optimizer bug, reported in

http://software.intel.com/en-us/forums/showthread.php?t=73629&o=a&s=lr

Unless you can satisfy yourself that your code is not affected by that bug, I think that you should steer clear of any version older than 11.1.073 on Linux. Using 11.1.075 would be my choice, since it has fixes to a number of earlier bugs, and does not contain several Fortran 2003 features that were introduced in the 12.0.xxx versions, some of which may have caused the introduction of new bugs that were not present in 11.1.075.