mark julia.write_barrier readonly on argument

[vchuravy]

While it is marked InaccessibleMemOnly already, that seems to still be only partially implemented in LLVM.
IIUC readonly on the argument is correct since write_barrier will escape the pointer but not write to it.

@wsmoses and I noticed this when looking at IR for Enzyme.

[Keno]

Hmm, this is tricky, and might require some LLVM IR reference language lawyering. The write barrier does twiddle bits in the object header. However, we treat those as inaccessible to the program and always mask them out on access. The IR reference says "If a readonly function writes memory visible to the program, or has other side-effects, the behavior is undefined. If a function writes to a readonly pointer argument, the behavior is undefined.". I think because of that we are technically in the clear since this memory isn't supposed to be "visible to the program"

[vchuravy]

@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @christopher-dG

[vchuravy]

@Keno ok if I land this? Well I need to look into the performance regression first

[Keno]

Yeah, why not