fix: fix codegen issues in preparation for CSE in MTK

src/variable.jl

@@ -291,6 +291,16 @@
 
 SymbolicIndexingInterface.symbolic_type(::Type{<:CallWithMetadata}) = ScalarSymbolic()
 
+# HACK:
+# A `DestructuredArgs` with `create_bindings = false` doesn't create a `Let` block, and
+# instead adds the assignments to the rewrites dictionary. This is problematic, because
+# if the `DestructuredArgs` contains a `CallWithMetadata` the key in the `Dict` will be
+# a `CallWithMetadata` which won't match against the operation of the called symbolic.
+# This is the _only_ hook we have and relies on the `DestructuredArgs` being converted
+# into a list of `Assignment`s before being addded to the `Dict` inside `toexpr(::Let, st)`.
+# The callable symbolic is unwrapped so it matches the operation of the called version.
+SymbolicUtils.Code.Assignment(f::CallWithMetadata, x) = SymbolicUtils.Code.Assignment(f.f, x)
+
 function Base.show(io::IO, c::CallWithMetadata)
     show(io, c.f)
     print(io, "⋆")

test/build_function.jl

@@ -1,7 +1,7 @@
 using Symbolics, SparseArrays, LinearAlgebra, Test
 using ReferenceTests
 using Symbolics: value
-using SymbolicUtils.Code: DestructuredArgs, Func, NameState
+using SymbolicUtils.Code: DestructuredArgs, Func, NameState, Let, cse
 @variables a b c1 c2 c3 d e g
 oop, iip = Symbolics.build_function([sqrt(a), sin(b)], [a, b], nanmath = true)
 @test all(isnan, eval(oop)([-1, Inf]))
@@ -301,3 +301,18 @@ end
         @test buf ≈ ones(2)
     end
 end
+
+@testset "cse with arrayops" begin
+    @variables x[1:3] y f(..)
+    t = x .+ y
+    t = t .* f(t)
+    res = cse(value(t))
+    @test res isa Let
+    @test !isempty(res.pairs)
+end
+
+@testset "`CallWithMetadata` in `DestructuredArgs` with `create_bindings = false`" begin
+    @variables x f(..)
+    fn = build_function(f(x), DestructuredArgs([f]; create_bindings = false), x; expression = Val{false})
+    @test fn([isodd], 3)
+end