Optimization ideas

[phipsgabler]

• Play with @inline and @nospecialize (see inlineable!)
• Type stability of record! and trackedcall
• Type stability of graph API, esp. referenced
• Add position field with a TapeReference to each node, and reuse this in subsequent references
• Make sure that IR objects are saved only once and cached
• Inline IRIndex values
• Inline TapeConstants where possible
• Use fixed parent node value, instead of calling getfield(..., :incomplete_node) every time.
• Replace Base.getfield(%5, :context) by using the context argument directly.
• Turn argument lists in nodes into NTuples
• Make canrecur generated again (and other defaults as well, if possible)
• Make forward use less than quadratic time by using references instead of dependencies?

[phipsgabler]

Actually, I should first start to investigate type stability and benchmark things properly.

See: https://juliaci.github.io/PkgBenchmark.jl/stable/

[phipsgabler]

One possible remaining problem is that f(args...), aka Core._apply(f, args), is not type-stable for f::Core.IntrinsicFunction. This breaks correct inference in trackedprimitive.

@generated function stable_apply(fn, args)
    :(fn($((:(args[$i]) for i in length(args.parameters))...)))
end

won't help.

[phipsgabler]

Ha, adding some concrete types wouldn't be the dumbest idea either:

struct TapeCall{T} <: TapeExpr{T}
    value::T
    f::TapeValue
    arguments::ArgumentTuple{TapeValue}
    varargs::Union{ArgumentTuple{TapeValue}, Nothing}
end

🤦