[llvm-dev] RFC: Adding a staging branch (temporarily) to facilitate upstreaming (original) (raw)

Nicolai Hähnle via llvm-dev llvm-dev at lists.llvm.org
Wed Jul 1 02:04:52 PDT 2020

On Tue, Jun 30, 2020 at 7:48 PM Chandler Carruth <chandlerc at google.com> wrote:

I'm mostly asking because we also have a fork of LLVM that we continuously keep aligned with github.com/llvm/llvm-project/master with the intention of deltas being broken out to be moved back upstream. That branch is currently at https://github.com/GPUOpen-Drivers/llvm-project, but if the LLVM community decides that branches whose goal is to be contributed upstream can (and should?) live in github.com/llvm/llvm-project, we'd probably be interested in doing that and having a branch such as amd/gfx-dev in the github.com/llvm/llvm-project repository.

See above -- I personally think this is an effective strategy when it solves a practical problem for people in the LLVM community at different organizations who want to collaborate on upstreaming something. I'd hope that anything like that was only done contingent on genuine interest in the community and with a pretty clear "it goes away" plan when merging stuff is complete.

Right. What I'm saying is that this would solve practical problems for us as well, where "us" is AMD plus folks we collaborate with (which includes Google).

I really like Chris' idea: set up a github.com/llvm/llvm-project-staging fork where groups that are in a situation similar to Apple's or ours can put branches that are clearly intended to have contents to be merged upstream, but may deviate while experimental features are developed. This keeps github.com/llvm/llvm-project clean, but provides a place where "messier" things can happen on a neutral collaborative playing field.

FWIW, I agree that it would be nice for deviated branches to not be permanent, but in my experience that's not really feasible with LLVM's development flow at the moment. Even if all changes go upstream, by the time they do, new changes are bound to have come in. This is largely a function of the dysfunctions in our review process, but not entirely, and in any case that review process seems like a much harder problem to fix.

Cheers, Nicolai

-- Lerne, wie die Welt wirklich ist, aber vergiss niemals, wie sie sein sollte.

More information about the llvm-dev mailing list