https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116977
--- Comment #3 from David Malcolm <dmalcolm at gcc dot gnu.org> --- (In reply to Thomas Schwinge from comment #1) > David, re (a) and (c), do we teach directly the Analyzer about specific > functions (like 'acc_malloc', 'acc_free' if '-fopenacc' is enabled), or do > we add GCC function attributes to designate return value and formal > parameters as "device" instead of "host" pointer, or something else? I've done both. Using the specific hardcoded functions approach (see known_function logic, and the various hacks in e.g. sm-malloc.cc), is easier to get started with, and allows for much more complicated semantics, such as "realloc", which has its own complicated special-case behaviors and is thus a pain to deal with. Using function attributes is much more flexible in that the end-user can use them, but perhaps isn't as expressive.
