[R-pkg-devel] Possible false negative for compiled C++ code in CRAN checks

[Ivan Krylov]

Dear Mauricio Vargas Sepulveda,

Welcome to R-package-devel!

В Sat, 9 Nov 2024 17:34:07 +0000
Mauricio Vargas Sepulveda <m.sepulveda using mail.utoronto.ca> пишет:

> CRAN reported memory leaks for:
> 
> CLAN/ASAN:
> https://www.stats.ox.ac.uk/pub/bdr/memtests/clang-ASAN/redatam/00check.log
> CLANG/UBSAN:
> https://www.stats.ox.ac.uk/pub/bdr/memtests/gcc-UBSAN/redatam/00check.log

These include container-overflow at
src/redatamlib/ByteArrayReader.cpp:170:23 and method calls with `this`
being a null pointer at src/redatamlib/FuzzyVariableParser.cpp:47 and
src/redatamlib/Entity.cpp:52.

A container-overflow in ByteArrayReader::ReadByte is dangerous because
the code is reading past the end of the vector populated with known
data. While the code doesn't crash, the remainder of the vector could
contain anything. This could invalidate the conclusions of a scientific
work if they end up being based on random contents of uninitialised
memory. You need to find out why m_currPos exceeds the number of bytes
in the array and prevent it from happening.

Calling methods on a null pointer is also an error in the code. This is
quite confusing, but I have a guess for what might be causing this:

  for (size_t i = 0; i < entities.size() - 1; ++i) {
    ret.push_back(
        {entities[i].GetBounds().second, entities[i + 1].GetBounds().first});
  }

If entities.size() is 0, entities.size() - 1 will overflow to a large
positive number, due to std::vector<T>::size_type being unsigned. The
loop will proceed accessing out-of-bounds elements of 'entities' until
something will cause it to crash. The guess could be completely wrong,
unfortunately.

Have you fixed these problems since v2.0.0?

> After asking on Stack Overflow
> (https://stackoverflow.com/q/79171799/3720258), it was suggested that
> I set 'CXXFLAGS="-stdlib=libc++"' in 'configure'. The question is
> very long and provides all the details that I skip here.

I am not seeing alloc-dealloc-mismatch in the CRAN checks, neither in
the links you provided, nor in the latest pretest results at
<https://win-builder.r-project.org/incoming_pretest/redatam_2.0.3_20241107_154750/>.
I think that these indicate a problem with the way your GitHub actions
are set up. Have you received a CRAN check report with
alloc-dealloc-mismatch in it?

I've tried to compare the versions visible in the CRAN archive but
couldn't get far because there was a lot of formatting changes. What
exactly are the problems for which your latest submission of v2.0.3 has
been archived? Do you need help reproducing the container-overflow
and/or null object pointer errors?

-- 
Best regards,
Ivan