inference: simplify is_const_prop_profitable_arg

[aviatesk]

We have actually forwarded PartialStruct always.

[vtjnash]

Seems a whole lot more aggressive now, since it now ignores the type of the fields?

[aviatesk]

We have this code that returns true for all PartialStructs anyway:

julia/base/compiler/abstractinterpretation.jl

Line 1153 in 490fdce

isa(arg, Const) || return true

[Keno]

Should that just have been return false? Ordinarily this function isn't called on things for which has_nontrivial_const_info is false.

[Keno]

I think the idea of the return true was to allow lattice elements that may be introduced by external abstract interpreters.

[aviatesk]

Should that just have been return false?

I tried the same on #47851 and found it causes some inference failures (e.g. https://buildkite.com/julialang/julia-master/builds/19102#0184f9d9-def9-4e65-8af2-6fd79714e996).

[aviatesk]

I think I can come up with the better heuristic though. So maybe we want to comment out these lines for now and try to enable it properly in a separate PR.

[aviatesk]

diff --git a/base/compiler/abstractinterpretation.jl b/base/compiler/abstractinterpretation.jl
index 5b5aec520f..5c23068b42 100644
--- a/base/compiler/abstractinterpretation.jl
+++ b/base/compiler/abstractinterpretation.jl
@@ -1143,7 +1143,16 @@ end
 
 function is_const_prop_profitable_arg(@nospecialize(arg))
     # have new information from argtypes that wasn't available from the signature
-    isa(arg, PartialStruct) && return true
+    if isa(arg, PartialStruct)
+        for i = 1:length(arg.fields)
+            fld = arg.fields[i]
+            isconstType(fld) && return true
+            is_const_prop_profitable_arg(fld) && return true
+            fld ⊏ fieldtype(arg.typ, i) && return true
+        end
+        println("[is_const_prop_profitable_arg] ", arg)
+        return false
+    end
     isa(arg, PartialOpaque) && return true
     isa(arg, Const) || return true
     val = arg.val

This change is sufficient to pass the existing inference test cases, but it turns out that then we never hit the return false case during sysimg build and test runs.

If we exclude the fld ⊏ fieldtype(arg.typ, i) && return true check, then the following test cases would fail:

• julia/test/compiler/inference.jl

  Lines 2422 to 2428 in 05328b8

  	getf(a) = a.f
  	@test Base.return_types((AliasableField,); interp=MustAliasInterpreter()) do a
  	if isa(getf(a), Int)
  	return getf(a)
  	end
  	return 0
  	end |> only === Int

• julia/test/compiler/inference.jl

  Lines 4355 to 4359 in 05328b8

  	@test Base.return_types((Int,)) do a
  	s1 = Some{Any}(a)
  	s2 = Some{Any}(s1)
  	s2.value.value
  	end |> only == Int

[aviatesk]

So I'd like to go with return true for now (it was at least as aggressive as we have been).