Make Julia's optimization passes configurable

[jpsamaroo]

The AMDGPU target's default datalayout causes issues with the LLVM verifier on debug builds because the verifier is run so early in the optimization pipeline. We can fix this by allowing backends to override running Julia's optimization pass pipeline, and reordering the individual passes to better suit the backend's needs.

[vchuravy]

You can overload the optimize! function used:

function GPUCompiler.optimize!(job::EnzymeJob, mod::LLVM.Module, entry::LLVM.Function)

[jpsamaroo]

My argument against that is that I'd have to also re-implement the rest of optimize!, which I don't necessarily want to replace, because any improvements to the default optimize! function don't necessarily benefit backends which override it.

[maleadt]

I don't understand why the IR fails verification in the first place. Shouldn't that be the thing to fix? Even if it gets corrected later on, the fact that the verifier croaks means some assumption is violated, and might that not combine badly with optimizations? I mean, the assertion is there for a reason.

[jpsamaroo]

Hmm true, the issue seems to stem from Julia choosing to alloca in AS 10 instead of the datalayout's stack AS, which seems like the wrong thing to do for GPU usecases. However, if we decided to fix that by choosing the right AS to alloca into, how does Julia do its GC tracking with its custom addrspaces?

[maleadt]

I'm guessing the alloca in AS 10 comes from here: https://github.com/JuliaLang/julia/blob/07ff9a960ff37255cf25ab2c73652aac02408a0c/src/llvm-alloc-opt.cpp#L955-L961

The lowering happens here: https://github.com/JuliaLang/julia/blob/master/src/llvm-late-gc-lowering.cpp. @Keno probably knows best whether this code can deal with alloca+addrspacecasts in and from a non-Julia AS to a Julia AS. If so, it seems like the easiest approach to rewrite alloca's after irgen to happen in a target-compatible AS and emit a addrspacecast to the original Julia AS. If you then optimize the IR and run the AS removal pass you should end up with just an alloca in the target AS.

[jpsamaroo]

The specific issue that this feature was intended to solve for GCN was fixed in #47 , so I'm closing this (since replacing Julia's pipeline is probably not a great idea in most cases). @vchuravy 's suggestion is a valid alternative if this functionality is absolutely needed.