There is also a case where some of our “patches” are hacks, workarounds and otherwise unsuitable for upstreaming. LLVM upstream might be working on a design of some sort that would solve the problem in a more general manner or otherwise a better way, but we often cannot afford to wait for the feature to get designed, implemented and then 6 more months for the next LLVM release.
I can’t see ourselves refraining from patching our LLVM.