DEBUG mode fails at assert(logger->f);

[cjlegg]

DEBUG mode fails at assert(logger->f); which is line 493 of mpi_alltoall.c. Geoff has noticed the same for alltoallv. Code works overall without debug mode.

[cjlegg]

Here are my notes of tracing through the call graph for where logger->f is referred to. This ends with a suggestion for a fix, which I shall try out.

logger->f is assigned to a new file at l536 in _log_data in logger.c (not ever ecluded by conditional compilation)

logger is passed to that call

_log_data is called (twice) by log_data, but inside #if ENABLE_RAW_DATA

logger is passed to that call too

log_data is called by log_profiling_data (no conditional compilation therein)

logger is passed to that call too.

log_profiling_data is called by _commit_data (no conditional comilation round it there)

calls to commit_data have no condtional compilation round them.

in mpi_alltoallv, for example) logger is intialsied at l789 in _mpi_init with logger_init()

logger_init() does not initialise the file

Therefore suggestion - wrap assert(logger->f) in #if ENABLE_RAW_DATA

[cjlegg]

wrapping assert(logger->f) in #if ENABLE_RAW_DATA did not work

note of investigation:
first had to fix some marcro calls and wrong variables in the call discussed below to make it compile - can do a pull request for those

first reporduced the error: set DEBUG =1 (in both alltoallv and alltoall configs)- thus:

ran run script ...-counts-variants,
got seg faults at _mpi_alltoall() l1328 using liballtoall_counts.so, and liballtoall_counts_unequal.so and liballtoall_counts_compact.so adn liballtoall_counts_unequal_compact.so
which prints to logger->f

so trying to mask out that call with ENABLE_RAW_DATA

recompiled and ran test again, as job 96232 - seg fault again at mpi_alltoall.c:1328

added an asstiopn to check that it is indeed logger->f that is seg faulting
aslo put some brackets around the test in #if, in case needed - ran as job 96233
alltoall: mpi_alltoall.c:1328: _mpi_alltoall: Assertion `logger->f' failed.

So why has logger-f not yet been initialised here?

[cjlegg]

from the comment above showing about the call graph down to where logger-f> is initiallised, that takes place in _commit_data, which does not happen until line 1392 (still in _mpi_alltoall()) or even later, when all the the mpi client program has called MPI_Finalize

temporary solution is to comment out these trace prints to logger->f, then a decision has to be made to make that permanent or to initialise earlier, if that is not detrimental in some way.