https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118478
--- Comment #3 from Tobias Burnus <burnus at gcc dot gnu.org> --- > I don't believe we should do two copy constructors for target, > one on the host and one on the device. My understanding is that the (first)privatization happens twice. I am not saying that the implementation should actually do it twice. Indeed, I think it would be a missed-optimization doing it twice. The OpenMP 6.0 specification has for "firstprivate": "For each variable of class type: * If the firstprivate clause is *not* on a *target* construct then a copy constructor is invoked to perform the initialization; and * If the firstprivate clause is on a *target* construct then how many copy constructors, if any, are invoked is unspecified." [For Fortran, some similar wording is planned, cf. OpenMP spec Issue 4371.] Thus, if 'firstprivate' is not on target, we have to use the C++ copy constructor and Fortran defined assignments. For 'firstprivate' on target, the simplest is probably to keep using the current code, i.e. not invoking any copy constructor/defined assignment - albeit at least for 'omp requires self_maps' (and, possibly, unified_shared_memory) we could invoke it on the host - as the memory is device accessible, albeit possibly slower (topic of shared memory controller vs. remote access or memory-page migration).
