I’m assuming code supported by clang 5 will be supported by clang 6 but not necessarily the reverse. But I’m happy to limit our support to the more recent one if that’s what two people are clamoring for.
I think I like @ChrisChiasson’s proposal, @syclik do you like it? I think it fits @bgoodri’s criteria though in some sense we had already resolved this in a different way and moved on to the organizational issues over who is responsible for fixing bugs for which compilers :P And I’m sticking with gcc 5 for now until someone shows me gcc 7 compiling or the RTools people actually successfully complete the change to gcc 7 in the next couple of years.
To flesh out the proposal a little:
Math repo
PRs:
- mac / clang 6 / libc++ Threading
- linux / gcc 5 / stdlibc++ OpenMPI
- mac clang OR linux clang distribution tests (without row vectors)
Additional tests on merge to develop [0]
- linux / gcc 5 / stdlibc++ Threading
- mac / clang OR linux / gcc GPU
- mac clang OR linux clang distribution tests (with row vectors)
I don’t think we have any Windows Math repo tests right now - I think they have been broken for a while? I think the only place the math library gets implicitly tested on Windows is during the CmdStan integration tests. But we’ll be sticking with the RTools compiler (gcc 4.9.3) for those tests.
[0] we can add an interstitial branch if this causes develop to become unstable.