On 1/1/19 12:20 PM, Richard W.M. Jones wrote:
Should we use attribute((nonnull)) at all? There's a very interesting history of this in libvirt -- try looking at commit eefb881 plus the commits referencing eefb881 -- but it does seem to work for me using recent GCC and Clang.
I only did a few functions because annotating them gets old quickly...
For internal usage, I think the annotations help a working compiler do a better job. I don't know if modern gcc does better than the old one that you mention being the reason that libvirt doesn't use it. I also haven't researched if libvirt could change their policy now (that is, are we stuck with avoiding something because of having been burned in the past, even though we would no longer be burned by it?). For installed headers, where the annotations are used by clients compiling a plugin, we need to be a bit more careful - while nbdkit itself requires compilation with gcc or clang (for things like our automatic variable cleanups) and therefore has unconditional support for attributes, I think we should still aim for a generic plugin being compiled with a standards-compliant compiler rather than requiring clients to use just gcc or clang. It's not impossible to make the use of attributes in a public header be properly guarded by the preprocessor, as long as we remember to do it right. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org