[petsc-dev] Hijacked MPI calls...

[Barry Smith]

   I just had a bad experience with (void) so feel it is only a half-answer.


> On Sep 13, 2016, at 10:45 AM, Satish Balay <balay at mcs.anl.gov> wrote:
> 
> On Tue, 13 Sep 2016, Barry Smith wrote:
> 
>> 
>>> On Sep 13, 2016, at 10:13 AM, Satish Balay <balay at mcs.anl.gov> wrote:
>>> 
>>> There is stuff like:
>>> 
>>> src/mat/order/fn1wd.c
>>> 
>>> -    kstop = (i__2 = xadj[node + 1], (PetscInt)PetscAbsInt(i__2)) - 1;
>>> +    kstop = (PetscInt)PetscAbsInt(xadj[node + 1]) - 1;
>>> 
>>> Perhaps PetscAbsInt() type stuff should be switched over to static inline?
>>> Or use a different notation?
>>> 
>>> i__2  = xadj[node + 1];
>>> kstop = (PetscInt)PetscAbsInt(i__2) - 1;
>> 
>>   This is the way any normal person would write it :-)
> 
> I think the first one is what is more intutive. The second one is
> clearly encoding an optimization - as the code is aware PetscAbsInt()
> is a macro - not a proper function. Hence the thought about 'static inline'
> 
> The first change compiles fine [with clang - both c, complex - but
> don't know if there are problem compilers]
> 
> I'm now inclined to just preserve the current usage with void:
> 
>    kstop = ((void)(i__2 = xadj[node + 1]), (PetscInt)PetscAbsInt(i__2)) - 1;
> 
>> 
>>> 
>>> And for the following - use 'void' as compiler suggests?
>>> 
>>> a/src/mat/utils/pheap.c
>>> -  while (par = Parent(loc), Value(h,par) > val) {
>>> +  while ((void)(par = Parent(loc)), Value(h,par) > val) {
>>> 
>>> a/src/dm/impls/plex/plexdistribute.c
>>> -      for (q = 0; q < numAdj || (adj[numAdj++] = support[s],0); ++q) {
>>> +      for (q = 0; q < numAdj || ((void)(adj[numAdj++] = support[s]),0); ++q) {
>> 
>>   Does the (void) work to eliminate the warning?
> 
> yes.
> 
> Satish
> 
>> 
>>  Barry
>> 
>>> 
>>> 
>>> And this one is perhaps for Matt..
>>> 
>>>>>>>>>>>>>>>>>>>>>> 
>>>         CC arch-clang-cmplx/obj/src/dm/impls/composite/pack.o
>>> /home/balay/petsc/src/dm/impls/composite/pack.c:657:17: warning: passing an object that undergoes default argument promotion to 'va_start' has undefined behavior [-Wvarargs]
>>> va_start(Argp,imode);
>>>               ^
>>> /home/balay/petsc/src/dm/impls/composite/pack.c:641:61: note: parameter of type 'InsertMode' is declared here
>>> PetscErrorCode  DMCompositeGather(DM dm,Vec gvec,InsertMode imode,...)
>>>                                                           ^
>>> 1 warning generated.
>>> <<<<<<<<<<<<<<<<<
>>> 
>>> 
>>> 
>>> Satish
>>> 
>>> On Mon, 12 Sep 2016, Barry Smith wrote:
>>> 
>>>> 
>>>> Satish,
>>>> 
>>>>  Interesting. We can probably get rid of most of the other warnings easily; it looks like mostly frivolous use of the , notation.
>>>> 
>>>>  Barry
>>>> 
>>>>> On Sep 12, 2016, at 11:19 PM, Satish Balay <balay at mcs.anl.gov> wrote:
>>>>> 
>>>>> The attached patch gets rid of most of the warnings [esp MPI usage
>>>>> from PETSc code - with the logging wrappers].
>>>>> 
>>>>> [also added to balay/clang-Wcomma]
>>>>> 
>>>>> Satish
>>>>> 
>>>>> On Mon, 12 Sep 2016, Satish Balay wrote:
>>>>> 
>>>>>> ok - so you are using CFLAGS=-Wcomma on your build..
>>>>>> 
>>>>>> I see warnings now. Attaching make.log
>>>>>> 
>>>>>> Satish
>>>>>> 
>>>>>> On Mon, 12 Sep 2016, Eric Chamberland wrote:
>>>>>> 
>>>>>>> 
>>>>>>> 
>>>>>>> Le 2016-09-12 à 17:32, Satish Balay a écrit :
>>>>>>>> Do you get these warnings with PETSc library build aswell?
>>>>>>> I can't tell since I didn't tried to build PETSc with clang...
>>>>>>>> 
>>>>>>>> The logging code tries to log all messages in library and in
>>>>>>>> application - and prints a summary with -info.
>>>>>>>> 
>>>>>>>> You can disable logging in your build with configure option: --with-log=0
>>>>>>>> 
>>>>>>>> Or you can define the PETSC_HAVE_BROKEN_RECURSIVE_MACRO in your
>>>>>>>> code/compile - and the wrappers will be skipped..
>>>>>>> ok, so these may be good workaround for me to keep -Wcomma activated on our
>>>>>>> "push server" that filters incoming commits...
>>>>>>> 
>>>>>>> Thanks!!!
>>>>>>> 
>>>>>>> Eric
>>>>>>> 
>>>>>>> 
>>>>>>> 
>>>>>> 
>>>>> <clang-Wcomma.patch><make.log>