Variant_test fails with boost 1.73.0

[hirschsn]

The unit test Variant_test fails with the most recent version of boost 1.73.0 in the unpack_pair test. The const std::vector<Variant> &v passed to unpack_pair (Variant.hpp:79) contains garbage. :)

The reason is some fancy template gone wrong in the script interface. Compiler warning:

[...]
/home/hirschsn/esp/espresso-add/src/script_interface/constraints/initialize.cpp:67:79:   required from here
/home/hirschsn/esp/espresso-add/src/script_interface/Variant.hpp:75:29: warning: narrowing conversion of ‘(double)pair.std::pair<int, double>::second’ from ‘double’ to ‘boost::enable_if_c<true, bool>::type’ {aka ‘bool’} [-Wnarrowing]
   75 |   return {{pair.first, pair.second}};
      |                        ~~~~~^~~~~~

Latest ESPResSo python branch.
Edit: Forgot to mention: Of course, with boost 1.72.0 everything works fine.

[jngrad]

Thanks for bringing our attention to this issue. Boost 1.73 will be shipped with Fedora 33, which will start rebuilding all packages including espressomd on July 22. I can maybe have a look next week.

[jngrad]

Just tried it with boost 1.73 on Ubuntu 20.04 with espresso 4.2-dev and 4.1.2. The Variant test fails because the pair is serialized into a vector of size 1. I originally also got random values as I forgot to add boost::get<T>() when printing the first value of the vector. The second value is uninitialized memory, so it can only be random.

The double brace expression return {{pair.first, pair.second}}; instantiates a vector of 1 Variant with underlying type int instead of a vector of 2 Variants with underlying types int and double, therefore unpack_pair() accesses memory out of bounds when fetching the second element. That's an easy fix, PR is coming.