https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101544
--- Comment #3 from Andrew Stubbs <ams at gcc dot gnu.org> --- The standalone amdgcn configuration does not support C++. There are a number of technical reasons why it doesn't Just Work, but basically it comes down to no-one ever working on it. Our customers were primarily interested in Fortran with C second. C++ offloading works fine provided that there are no library calls or exceptions. Ignoring unsupported C++ language features, for now, I don't think there's any reason why libstdc++ would need to be cut down. We already build the full libgfortran for amdgcn. System calls that make no sense on the GPU were implemented as stubs in Newlib (mostly returning some reasonable errno value), and it would be straight-forward to implement more the same way. I believe static constructors work (libgfortran uses some), but exception handling does not. I'm not sure what other exotica C++ might need? As for exceptions, set-jump-long-jump is not implemented because there was no call for it and I didn't know how to handle the GCN register files properly. Not only are they variable-sized, they're also potentially very large: ranging from ~6KB up to ~65KB, I think (102 32-bit scalar, and 256 2048-bit vector registers, for single-threaded mode, but only 80 scalar and 24 vector registers in maximum occupancy mode, in which case per-thread stack space is also quite limited). I'm not sure now the other exception implementations work.
