Re: [PATCH 1/2] fort_dyn_array: add basic fortran dyn array support

[Walfred Tedeschi <walfred.tedeschi@intel.com>]

Am 2/9/2016 um 12:36 PM schrieb Joel Brobecker:
>> I don't think I will be able to fix the failures on the mentioned
>> hosts/targets before you
>> create the branch, as I need to replicate the environment on my end first to
>> start investigating.
>> Therefore I suggest to revert the change for now. Sorry if this caused any
>> inconvenience.
>
> We need a good reason for reverting (Eg: regressions in pre-existing
> features). From what i read below:
>
>> Bernhard and I we have investigated the issues reported. Issues were
>> reproduced by using the newest GCC.  Looks like the issues are independent
>> from the target at first sight.
>>
>> Issues could be divided into 3 categories:
>>
>> 1.  Issue root causes to GCC 5.x and Fortran "associated" intrinsic.
>> Compiler is not reporting the “Associated property correctly. (see test case
>> 1)
>> Affected tests:
>> gdb.fortran/vla-ptype.exp: ptype pvla not initialized
>> gdb.fortran/vla-value.exp: print undefined pvla
>> gdb.fortran/vla-value.exp: print non-associated &pvla
>> gdb.fortran/vla-value.exp: print undefined pvla(1,3,8)
>> gdb.mi/mi-vla-fortran.exp: evaluate not associated vla
>> gdb.mi/mi-vla-fortran.exp: create local variable pvla2_not_associated
>> gdb.mi/mi-vla-fortran.exp: info type variable pvla2_not_associated
>>
>> 2.  Test issues. Test take into account that array created is initialized by
>> default. Initialization depends on the compiler also Fortran standard does
>> not state the need for initialization.
>> Affected tests:
>> gdb.mi/mi-vla-fortran.exp: evaluate allocated vla
>>
>> 3.  GCC or GDB error. Breakpoint is completely off, breakpoint on return
>> line is triggered before the execution of the functions body. Deeper
>> investigation should be done to define the real issue.
>> Affected tests:
>> gdb.cp/vla-cxx.exp
>
> Are these regressions? Or just showing that the new feature are
> not working well with the newer GCC?
>
> I fear that, reverting central changes at this stage could
> destabilize the branch, and therefore we need to be very careful
> as to what we do.
>

Joel,

What we have seen is that the GCC 5.x is causing the issue.
On Fedora 22 we used a Vanilla GCC version 5.3.1 testing mainline GDB 
and Fedora GDB (7.9.1-20.fc22) we run only the Fortran testsuite.
Results were:

GDB mainline: 463 passes and 6 fails.
GDB Fedora: 449 passes and 20 fails.

A bit more about the categories:

1. Associate intrinsic does not work even from the compiler itself. I.e. 
When one adds the associated (pvla) in the code where test stops to 
evaluate it returns true. No chance for the debugger here.
(we will add a defect for GFORTRAN)

2. Test issue still there.  Testing for initialized array where it is 
not needed.

3. GDB Fedora has the same behavior when the breakpoint at the last line 
is hit, function's body was still not executed.  However, this has to be 
further investigated. Is really the Compiler which is generating bad 
information?

In this sense we do not see any of the issues as regressions from the 
debugger point of view.

Thanks and regards,
Fred & Bernhard.







Intel Deutschland GmbH
Registered Address: Am Campeon 10-12, 85579 Neubiberg, Germany
Tel: +49 89 99 8853-0, www.intel.de
Managing Directors: Christin Eisenschmid, Christian Lamprechter
Chairperson of the Supervisory Board: Nicole Lau
Registered Office: Munich
Commercial Register: Amtsgericht Muenchen HRB 186928