Misattributed profile information

[vchuravy]

Reproducer:

using Base.Cartesian
using Profile

data = zeros(10_000, 10_000);

@generated function ohmy(data, op::Op, ::Val{N}) where {Op, N}
  return quote
    Base.@_propagate_inbounds_meta
      @nloops $N i n -> 1:size(data, n) begin
         I = CartesianIndex(@ntuple $N i)
         data[I] = op(data[I])
       end
    end
end

N=1000; @profile for i in 1:N; ohmy(data, +, Val(2)) end

Produces something like:

julia> Profile.print()
27160 ./task.jl:268; (::getfield(REPL, Symbol("##26#27")){REPL.REPLBackend})()
 27160 ...churavy/builds/julia/usr/share/julia/stdlib/v1.3/REPL/src/REPL.jl:118; macro expansion
  27160 ...churavy/builds/julia/usr/share/julia/stdlib/v1.3/REPL/src/REPL.jl:86; eval_user_input(::Any, ::REPL.REPLBackend)
   27160 ./boot.jl:330; eval(::Module, ::Any)
    27160 ...y/builds/julia/usr/share/julia/stdlib/v1.3/Profile/src/Profile.jl:25; top-level scope
     27159 ./REPL[25]:1; macro expansion
      2553  ./REPL[4]:2; ohmy(::Array{Float64,2}, ::typeof(+), ::Val{2})
       2553 ./REPL[4]:4; macro expansion
        26   ./array.jl:729; macro expansion
        2348 ./cartesian.jl:64; macro expansion
         2348 ./REPL[4]:6; macro expansion
          2348 ./multidimensional.jl:486; getindex
           2348 ./array.jl:729; getindex
        177  ./cartesian.jl:65; macro expansion
         1   ./range.jl:595; iterate
          1 ./promotion.jl:403; ==
         176 ./range.jl:596; iterate
          176 ./int.jl:53; +
        2    ./int.jl:53; macro expansion
      3159  ./array.jl:0; ohmy(::Array{Float64,2}, ::typeof(+), ::Val{2})
      21428 ./cartesian.jl:65; ohmy(::Array{Float64,2}, ::typeof(+), ::Val{2})

The last line is attributed ./cartesian.jl instead of [REPL]. So the nesting of the information got broken and so the ticks are missattributed.

cc: @peterahrens

[vtjnash]

Do you know if the DWARF info (@code_native) is correct? I’ve seen LLVM 6 do some crazy things. For instance, look at how many files setup_stdio and link_pipe gets splatted across in: code_native(Base.setup_stdio, (Base.PipeEndpoint, Bool)). It looks like my LLVM 7 build is better on it.

[vchuravy]

I see what your saying. I checked @code_typed, but yeah @code_llvm looks okay, but @code_native looks wrong.