Split and parameterize Core.Compiler.optimize

[vchuravy]

I want to experiment with new compiler passes out-of-tree and for that I need to specialize optimize on the AbstractInterpreter.
I also split optimize into two functions to make re-use easier.

[vchuravy]

@Keno or @vtjnash any objections? Otherwise I will merge this in the coming days.

[Keno]

Ok by me, but I've been mostly holding off on compiler work until the 1.6 branch to avoid delaying that any further.

[Roger-luo]

I'm wondering if we could also have run_passes have a interp parameters too if optimize is dispatched on interp? I'm inserting some optimization passes after Julia compiler's (currently I copy-pasta typeinf) This PR is also useful on my side.

NVM. I just realized I can put it before finish…

[vchuravy]

Ok by me, but I've been mostly holding off on compiler work until the 1.6 branch to avoid delaying that any further.

Yeah I think Tim and I would like to use the AbstractInterpreter work in 1.6 to have a split inference and inference caches for the GPUCompiler work. (This PR is not required for that, but #38370 is)

[Roger-luo]

Ok by me, but I've been mostly holding off on compiler work until the 1.6 branch to avoid delaying that any further.

I'm copy-pasting the _typeinf to dispatch custom optimization pass after typeinf and optimize (https://github.com/QuantumBFS/YaoCompiler.jl/blob/master/src/compiler/optimize.jl), it's not urgent but it would be great if this PR gets merged in 1.6, so I can release a version without copying code from Julia compiler once 1.6 released.

[vchuravy]

I'm copy-pasting the

:) yeah I was to lazy to do that in https://github.com/vchuravy/KernelCompiler.jl/blob/main/src/optimize.jl