Packed first-class module allocation is not hoisted to top-level

[ncik-roberts]

Repro:

(* a.ml *)
module Make (T : sig val x : int ref end) = struct
  module L1 = struct
    let x = 3
    let y = T.x
  end
end

include Make [@inlined never] (struct
  let x = ref 1
end)

(* b.ml *)
module type S = sig
  module L1 : sig
    val y : int ref
  end
end

let run () = Sys.opaque_identity (module A : S)

Compile ocamlopt a.ml b.ml -dflambda (with or without -O3). The result is that run allocates the packed first-class module each time:

   (λ〈k32〉《k31》⟅my_region/88N⟆
      (param/90UV ∷ 𝕍=tagged_ℕ𝕚) (my_closure/89N ∷ 𝕍) my_depth/87N .
    Pfield/91N =
     (((Block_load (Values (tag ⊤) (size ⊤) (field_kind Any_value))
       Immutable) A.camlA 1) b.ml:7,41--42)
    Pfield/92N =
     (((Block_load (Values (tag ⊤) (size ⊤) (field_kind Any_value))
       Immutable) Pfield/91N 1) b.ml:7,41--42)
    Pmakeblock/93N =
     (((Make_block (Values (tag 0) (shape (𝕍))) Immutable Heap) Pfield/92N)
      b.ml:7,41--42)
    Pmakeblock/94N =
     (((Make_block (Values (tag 0) (shape (𝕍))) Immutable Heap)
       Pmakeblock/93N) b.ml:7,41--42)
    Popaque/95N =
     (((Opaque_identity (middle_end_only false) (kind 𝕍)) Pmakeblock/94N)
      b.ml:7,13--47)
    return k32 Popaque/95N)))

The expected behavior is that the packed module is hoisted to a top-level allocation.

This test is fairly sensitive to a lot of things:

• S has to be a restriction on A's inferred signature; it can't be exactly A's inferred signature
• S has to involve a nested module (here, L1) that with a binding that's some mutable value (here, an int ref).