[R-pkg-devel] CRAN and packages bundling third party C/C++ code

[Kevin Ushey]

Hi,

This is mostly a question for the CRAN team, but I figured posting here
would be best since other package maintainers could see the responses as
well.

I'm currently in the process of preparing a submission of RcppParallel to
CRAN. RcppParallel bundles Intel's Threading Building Blocks (https://www.
threadingbuildingblocks.org/) and builds + installs it alongside the
package.

However, the TBB sources themselves do a lot of things that trip up R CMD
check + versions of R compiled with sanitizers, which are mostly out of my
control. For example, some UBSAN errors are emitted here:

https://www.stats.ox.ac.uk/pub/bdr/memtests/clang-UBSAN/
RcppParallel/tests/doRUnit.Rout

Executing test function test.distance  ... /data/gannet/ripley/R/
packages/tests-clang-SAN/RcppParallel.Rcheck/RcppParallel/include/tbb/task.h:721:30:
runtime error: member call on address 0x7fbf31e93d00 which does not point
to an object of type 'tbb::internal::scheduler'
0x7fbf31e93d00: note: object is of type 'tbb::internal::custom_
scheduler<tbb::internal::IntelSchedulerTraits>'
 00 00 00 00  a8 ab 76 32 bf 7f 00 00  00 00 00 00 00 00 00 00  00 00 00 00
00 00 00 00  80 ff c9 31
              ^~~~~~~~~~~~~~~~~~~~~~~
              vptr for 'tbb::internal::custom_scheduler<tbb::internal::
IntelSchedulerTraits>'
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior
/data/gannet/ripley/R/packages/tests-clang-SAN/RcppParallel.Rcheck/
RcppParallel/include/tbb/task.h:721:30 in

This occurs, as far as I can tell, because TBB sidesteps typical C++ object
construction for some objects: instead, it allocates some memory, then
casts it to some required type. As an example,

https://github.com/RcppCore/RcppParallel/blob/603d1aa6f340668e34f5f490438967
489edc9197/src/tbb/src/tbb/scheduler.cpp#L322-L328

While this is 'wrong', it also 'works'. What am I, as a package maintainer,
supposed to do in this situation? To make things worse, this UBSAN warning
unfortunately affects all client packages of RcppParallel as well, as this
warning is triggered by code ran by RcppParallel on load / startup. Do
other package maintainers just need to say, "sorry, that's RcppParallel's
fault"?

In addition, TBB itself suppresses a number of compiler diagnostics
deliberately -- for example,

https://github.com/RcppCore/RcppParallel/blob/603d1aa6f340668e34f5f490438967
489edc9197/inst/include/tbb/flow_graph_opencl_node.h#L1234-L1252

https://github.com/RcppCore/RcppParallel/blob/603d1aa6f340668e34f5f490438967
489edc9197/src/tbb/include/tbb/tbb_machine.h#L666-L674

Presumedly, TBB does this because they know what they're doing and the
compiler diagnostic will not be useful within this region of code. However,
this also triggers R CMD check warnings. What should I do here -- just
explain the case to CRAN? Patch TBB to allow the compiler diagnostics to
flow through?

The more general question here -- what should package authors do when
third-party code bundled in the package emits compiler warnings?

Thanks,
Kevin

	[[alternative HTML version deleted]]