Add a statement-cost printer for analyzing inlining decisions

[timholy]

While it now has a doctest, https://docs.julialang.org/en/latest/devdocs/inference/#The-inlining-algorithm-(inline_worthy)-1 has bitrotted more than once. And it's sufficiently laborious that it discourages rapid investigation. This makes printing the costs easy.

Here's a demo:

julia> Base.print_statement_costs(map, (typeof(sqrt), Tuple{Int}))
map(f, t::Tuple{Any}) in Base at tuple.jl:169
  0 1 ─ %1  = Base.getfield(_3, 1, true)::Int64
  1 │   %2  = Base.sitofp(Float64, %1)::Float64
  2 │   %3  = Base.lt_float(%2, 0.0)::Bool
  0 └──       goto #3 if not %3
  0 2 ─       Base.Math.throw_complex_domainerror(:sqrt, %2)::Union{}
  0 └──       unreachable
 20 3 ─ %7  = Base.Math.sqrt_llvm(%2)::Float64
  0 └──       goto #4
  0 4 ─       goto #5
  0 5 ─ %10 = Core.tuple(%7)::Tuple{Float64}
  0 └──       return %10

The numbers in the first column are the inliner's assigned cost.

If folks like this, I can add a test to make sure this stays working. (done)

[KristofferC]

Maybe it makes sense to add a @code_ version of this to InteractiveUtils. Could always be done later of course.

[kimikage]

No way, is that @code_costs?:joy:

[KristofferC]

Or an argument to one of the others (cost = true)

[kimikage]

FYI, I'm providing @code_costs and it's exactly @timholy who gave me the motivation to do so. 😄

I have some objections about the display style, but I agree that Julia supports that feature.

[KristofferC]

Oh, nice!

[timholy]

I like your printing quite a lot better. But if you do it as separate code-display (print(io, src); lines = readlines(io)) then one risks having the two get misaligned, whereas this version ensures that the two will always stay aligned.

Demo:

julia> function buildexpr()
           items = [7, 3]
           ex = quote
               X = $items
               for x in X
                   println(x)
               end
           end
           return ex
       end
buildexpr (generic function with 1 method)

julia> @code_typed buildexpr()
CodeInfo(
1 ─ %1  = Core.tuple(7, 3)::Core.Const((7, 3), false)
│   %2  = $(Expr(:foreigncall, :(:jl_alloc_array_1d), Vector{Int64}, svec(Any, Int64), 0, :(:ccall), Vector{Int64}, 2, 2))::Vector{Int64}
│         Base.arraysize(%2, 1)::Int64
└──       goto #7 if not true
2 ┄ %5  = φ (#1 => 1, #6 => %10)::Int64
│   %6  = φ (#1 => 1, #6 => %16)::Int64
│   %7  = φ (#1 => 1, #6 => %17)::Int64
│   %8  = Base.getfield(%1, %6, true)::Int64
│         Base.arrayset(false, %2, %8, %5)::Vector{Int64}
│   %10 = Base.add_int(%5, 1)::Int64
│   %11 = (%7 === 2)::Bool
└──       goto #4 if not %11
3 ─       goto #5
4 ─ %14 = Base.add_int(%7, 1)::Int64
└──       goto #5
5 ┄ %16 = φ (#4 => %14)::Int64
│   %17 = φ (#4 => %14)::Int64
│   %18 = φ (#3 => true, #4 => false)::Bool
│   %19 = Base.not_int(%18)::Bool
└──       goto #7 if not %19
6 ─       goto #2
7 ┄       goto #8
8 ─ %23 = Core._expr(:(=), :X, %2)::Expr
│   %24 = $(Expr(:copyast, :($(QuoteNode(:(for x = X
      #= REPL[4]:6 =#
      println(x)
  end))))))::Expr
│   %25 = Core._expr(:block, $(QuoteNode(:(#= REPL[4]:4 =#))), %23, $(QuoteNode(:(#= REPL[4]:5 =#))), %24)::Expr
└──       return %25
) => Expr

You can see statement %24 took more than one line to print.

[kimikage]

I just made a change that will make it easier to fix that problem. Perhaps that will be fixed in a few hours. 😃
Edit: Done. CodeCosts.jl v0.2.0 is available.

[timholy]

I switched the order of the printing so the cost comes first...I just needed to look a little harder at how these functions were working. I edited the top post to show the current output.

[kimikage]

  0 2 ─       Base.Math.throw_complex_domainerror(:sqrt, %2)::Union{}
  0 └──       unreachable

BTW, is this some kind of magic?

[timholy]

#35982 and #30222 (the latter should be closed once I split out a couple of nuggets into a separate PR)

[kimikage]

What I'm suggesting is the cost "zero". 😅

Edit:
Sorry. WRT the former I was aware of the one in May (returning a flat cost, not zero), so I had not understood the relationship with the latter.

[timholy]

I'm not sure I understand your point, but the zero for the throw comes from

julia/base/compiler/optimize.jl

Lines 329 to 331 in 3285ee0

	extyp = line == -1 ? Any : src.ssavaluetypes[line]
	if extyp === Union{}
	return 0

The old demo at https://docs.julialang.org/en/v1.6-dev/devdocs/inference/ set line to -1 (because the simple approach using map doesn't allow it), so it didn't return 0.