more precise types in keyword argument method lowering

[JeffBezanson]

We have a bad interaction between two features: (1) generated methods for keyword arguments need to pass the original function around in case it contains closure data, (2) functions that take a function argument but don't call it aren't specialized on that argument (to avoid excessive specialization). That can lead to slow calls in keyword method shims, but is a bit unpredictable. For example it caused the slowdown observed in #30830 (comment), and likely other intermittent performance changes we've seen.

Long story short, this is before:

julia> code_typed(string, (Int,))
1-element Array{Any,1}:
 CodeInfo(
1 ─ %1 = invoke Base.:(#string#322)(10::Int64, 1::Int64, _1::Function, _2::Int64)::String
└──      return %1
) => String

and this is after:

julia> code_typed(string, (Int,))
1-element Array{Any,1}:
 CodeInfo(
1 ─ %1 = invoke Base.:(#string#322)(10::Int64, 1::Int64, _1::typeof(string), _2::Int64)::String
└──      return %1
) => String

Now the front end puts a specific type on the argument slot used to forward the function. After all, it knows exactly what function we're dealing with...right? Well, it can be quite difficult to do for closures, where we need to ensure everything is defined in just the right order, since there are some circular dependencies (method A calls method B, but method B needs to mention A in its signature). I hope this works.

[timholy]

Having poked at these recently, I wondered about precisely this change, but I didn't understand where the difficulties and opportunities lay. Specifically, I had wondered about adding the type info so that one could guess where the #self# argument was, but then came to the conclusion that it would be defeated by nefarious methods like foo(x; f::typeof(foo)=foo) (which seems pointless and cruel). This isn't currently allowed due to circular dependencies, however.

[JeffBezanson]

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

[nanosoldier]

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

[JeffBezanson]

I ran NewPkgEval and identified some issues. Tests added. Also needs #30972, then this should be ready to go.