Closure and ccall lowering causes unneeded boxing

[yuyichao]

Another special case of #15276 though also related to ccall lowering.

The ccall and the normal function call should have the same effect on closure lowering but apparently not. The ccall causes the second range to be boxed but not when it's a normal function call.

julia> function test0()
           range1 = 1
           threadsfor_fun1() = range1
           ccall(:jl_breakpoint, Void, (Any,), threadsfor_fun1)
           range2 = 2
           threadsfor_fun2() = range2
           ccall(:jl_breakpoint, Void, (Any,), threadsfor_fun2)
       end
test0 (generic function with 1 method)

julia>

julia> function test1()
           range1 = 1
           threadsfor_fun1() = range1
           identity(threadsfor_fun1)
           range2 = 2
           threadsfor_fun2() = range2
           identity(threadsfor_fun2)
       end
test1 (generic function with 1 method)

julia>

julia> @code_lowered test0()
CodeInfo(:(begin
        nothing
        range2 = (Core.Box)()
        range1 = 1 # line 3:
        threadsfor_fun1 = $(Expr(:new, :((Core.apply_type)(Main.#threadsfor_fun1#1, (Core.typeof)(range1))), :(range1))) # line 4:
        $(Expr(:foreigncall, :(:jl_breakpoint), Void, svec(Any), :(threadsfor_fun1), 0)) # line 5:
        (Core.setfield!)(range2, :contents, 2) # line 6:
        threadsfor_fun2 = $(Expr(:new, :(Main.#threadsfor_fun2#2), :(range2))) # line 7:
        return $(Expr(:foreigncall, :(:jl_breakpoint), Void, svec(Any), :(threadsfor_fun2), 0))
    end))

julia> @code_lowered test1()
CodeInfo(:(begin
        nothing
        range1 = 1 # line 3:
        threadsfor_fun1 = $(Expr(:new, :((Core.apply_type)(Main.#threadsfor_fun1#3, (Core.typeof)(range1))), :(range1))) # line 4:
        (Main.identity)(threadsfor_fun1) # line 5:
        range2 = 2 # line 6:
        threadsfor_fun2 = $(Expr(:new, :((Core.apply_type)(Main.#threadsfor_fun2#4, (Core.typeof)(range2))), :(range2))) # line 7:
        return (Main.identity)(threadsfor_fun2)
    end))

This is a regression compare to 0.5.

[yuyichao]

Reported here https://discourse.julialang.org/t/multithreading-performance-regressions-in-0-6/3748