On 2016/5/12 6:56 PM, Jakub Jelinek wrote: > On Thu, May 12, 2016 at 12:47:18PM +0200, Thomas Schwinge wrote: >> Hi! >> >> On Thu, 12 May 2016 12:02:58 +0200, Jakub Jelinek <ja...@redhat.com> wrote: >>> ABI incompatible change for the plugin >>> interface (affecting OpenACC capable plugins only), I think you just should >>> rename the plugin callback you add the argument to, so that >>> || !DLSYM_OPT (openacc.register_async_cleanup, >>> openacc_register_async_cleanup) >>> would fail when trying to load GCC 6.x nvptx plugin from GCC 7.x libgomp >>> or vice versa. >> >> Hmm, as far as I remember, we had previously agreed that libgomp plugin >> ABI changes are not of any concern, given that libgomp and its plugins >> will always be built from the same sources, at the same time, and so >> their ABIs will always correspond? Discussed before in >> <http://news.gmane.org/find-root.php?message_id=%3C56449E23.3050400%40acm.org%3E>, >> for example. > > I thought the agreement was that it is ok not to support mixing of > different libgomp and plugin versions, but we should make sure that we > refuse to load the plugin in case of mismatch, instead of silently crashing. > Of course, changes in unreleased compiler versions are fine. > > So, I'm not asking for compatibility in that 6.x nvptx plugin should still > work with 7.x libgomp and vice versa, but that it would be ignored or > diagnosed if somebody mixes it. > > Jakub >
I assume that bumping GOMP_VERSION in include/gomp-constants.h would be enough? Renaming is just really ugly, probably okay for the async patch which touches just one plugin hook, but the device lock deadlock fix audits several of them in a minor way; doesn't look like the right solution to rename them all. Chung-Lin
