[Utilities] improve performance of ScalarNonlinearFunction utilities

[odow]

x-ref jump-dev/JuMP.jl#3106

[blegat]

It seems that having a special case when the index map is trivial in CleverDict could be helpful here if this becomes the bottleneck

[odow]

It was only a bottleneck because of the type stability problem. For the ::AbstractVector case, we were passing in a Vector{Any}, then trying to infer the return type, and then casting back to convert(Vector{Any}. Just lots of unnecessary stuff. The recursion also prevented deeply nested expressions.

One general improvement could be that if the CachingOptimizer knows that the mapping is one-to-one, then it could store nothing here:

MathOptInterface.jl/src/Utilities/cachingoptimizer.jl

Lines 82 to 83 in dd15549

	model_to_optimizer_map::IndexMap
	optimizer_to_model_map::IndexMap

and we could add a new overload to make all of these calls a no-op:

MathOptInterface.jl/src/Utilities/cachingoptimizer.jl

Line 525 in dd15549

map_indices(m.model_to_optimizer_map, func),

[blegat]

Yes I agree but I would make that logic inside CleverDict so that places that don't support nothing still work

[odow]

I don't understand what CleverDict has to do here?

[blegat]

The IndexMap is holding CleverDict's. CleverDict is already recognizing if the indices of src are contiguous so it seems natural to also recognize when the mapping is trivial. It's be discussed a few time with @joaquimg and I also reporting that it could improve performance in other situations also if we then don't even create the vector holding the mapping since it's trivial.