Better varargs are actually worse about cassette.jl

[jrevels]

That issue name 😂 Yeah, the "better" in that commit referred to a correctness improvement rather than a performance improvement 😛

Thanks for finding/cataloging these regressions, it's extremely useful.

[jrevels]

With #40 + JuliaLang/julia#26826 + a hack to also ignore Overdub{Execute} in the recursion limiting heuristic [1], this goes back to your "before" case...which is still inferring as Any.

Sigh.

[1] This is a separate hack from the spoofing we're using to ignore Overdub{Transform}...we haven't decided how to do it in a not hacky way yet.

[jrevels]

Okay, JuliaLang/julia#26826 now implements the recursion limiting spoofing in a way that fixes the problem we were hacking around, so with #40 and JuliaLang/julia#26826 this becomes:

julia> code_warntype(Cassette.overdub(Ctx, unsafe_load), Tuple{Ptr{Float32}})
Variables:
  o<optimized out>
  x1::Ptr{Float32}
  i<optimized out>

Body:
  begin
      Core.SSAValue(119) = (typeassert)(1, Int64)::Int64
      Core.SSAValue(192) = (pointerref)(x1::Ptr{Float32}, Core.SSAValue(119), 1)::Any
      return Core.SSAValue(192)
  end::Any

[maleadt]

Yay, looking good. Any idea which limit/deficiency prevents pointerref from being inferred?

[jrevels]

Any idea which limit/deficiency prevents pointerref from being inferred?

The problem was that intrinsics don't each have their own type - they're all of type Core.IntrinsicFunction, so inference relies on constant propagation in order to infer calls to them. However, Cassette's splatting was thwarting the compiler's ability to constant propagate. JuliaLang/julia#26826 enables varargs constant propagation (and adds another patch on top of the recursion limiting heuristic), thus allowing this case to be inferred.

Leveraging this required refactoring Cassette's overdub implementation a bit (#41), because storing the function in an intermediate Overdub struct was also getting in the way of constant propagation (and is no longer necessary, anyway, since I've been told that other compiler patches have since fixed the problems that struct was hacking around).

[jrevels]

closed by #41