Introduce temporary if necessary to inline _apply

[martinholters]

With

julia> function foo(x,y)
           println("x=$x, y=$y")
           return x, y
       end
foo (generic function with 1 method)

julia> bar(x,y) = +(foo(x,y)...)
bar (generic function with 1 method)

master:

julia> @code_typed bar(1, 2)
CodeInfo(:(begin 
        return (Core._apply)(Main.+, $(Expr(:invoke, MethodInstance for foo(::Int64, ::Int64), :(Main.foo), :(x), :(y))))::Int64
    end))=>Int64

this PR:

julia> @code_typed bar(1, 2)
CodeInfo(:(begin 
        #temp# = $(Expr(:invoke, MethodInstance for foo(::Int64, ::Int64), :(Main.foo), :(x), :(y)))
        return (Base.add_int)((Core.getfield)(#temp#, 1)::Int64, (Core.getfield)(#temp#, 2)::Int64)::Int64
    end))=>Int64

Fixes #21065 and makes #21067 (almost?) obsolete. (The "almost" being a matter of inlining promote, promote_type, and convert, where #21067 might still be advantageous.)

[martinholters]

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

[JeffBezanson]

I think this should use newvar! and make an SSAValue instead. (That function is badly named; should be e.g. new_ssavalue.)

[martinholters]

Updated accordingly.

[martinholters]

I wasn't sure whether the effect_free test is still necessary. Is there any reason the SSAValue should not be introduced unconditionally?

Also, what's a good way to test this PR? Inspect code_typed(bar, Tuple{Int,Int})[1].first.code and verify there is no :(Core._apply) in there? Do we have similar tests somewhere?

[martinholters]

Added a test, but it does look a bit messy...

[JeffBezanson]

Looks fine to me. There is a test kind of similar to this in test/inline.jl, but we usually just test performance and behavior.

Keeping the effect_free test is fine. We could introduce the ssavalue unconditionally and let it get removed by remove_redundant_temp_vars!, but it doesn't really matter.

[nanosoldier]

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

[tkelman]

would have been better to squash merge this