Merge pull request #1900 from SciML/myb/morealias

Add one missing case in the complete alias generation

src/systems/alias_elimination.jl

@@ -708,17 +708,27 @@ function alias_eliminate_graph!(graph, var_to_diff, mm_orig::SparseMatrixCLIL)
         prev_r = -1
         for _ in 1:10_000 # just to make sure that we don't stuck in an infinite loop
             reach₌ = Pair{Int, Int}[]
+            # `r` is aliased to its equality aliases
             r === nothing || for n in neighbors(eqg, r)
                 (n == r || is_diff_edge(r, n)) && continue
                 c = get_weight(eqg, r, n)
                 push!(reach₌, c => n)
             end
+            # `r` is aliased to its previous differentiation level's aliases'
+            # derivative
             if (n = length(diff_aliases)) >= 1
                 as = diff_aliases[n]
                 for (c, a) in as
                     (da = var_to_diff[a]) === nothing && continue
                     da === r && continue
                     push!(reach₌, c => da)
+                    # `r` is aliased to its previous differentiation level's
+                    # aliases' derivative's equality aliases
+                    r === nothing || for n in neighbors(eqg, da)
+                        (n == da || n == prev_r || is_diff_edge(prev_r, n)) && continue
+                        c′ = get_weight(eqg, da, n)
+                        push!(reach₌, c * c′ => n)
+                    end
                 end
             end
             if r === nothing

test/reduction.jl

@@ -302,3 +302,25 @@ ss = alias_elimination(sys)
 @test length(equations(ss)) == length(states(ss)) == 1
 ss = structural_simplify(sys)
 @test length(equations(ss)) == length(states(ss)) == 2
+
+@variables t
+vars = @variables x(t) y(t) k(t) z(t) zₜ(t) ddx(t)
+D = Differential(t)
+eqs = [D(D(x)) ~ ddx
+       ddx ~ y
+       D(x) ~ z
+       D(z) ~ zₜ
+       D(zₜ) ~ sin(t)
+       D(x) ~ D(k)
+       D(D(D(x))) ~ sin(t)]
+@named sys = ODESystem(eqs, t, vars, [])
+state = TearingState(sys);
+ag, mm, complete_ag, complete_mm = ModelingToolkit.alias_eliminate_graph!(state)
+fullvars = state.fullvars
+aliases = []
+for (v, (c, a)) in complete_ag
+    push!(aliases, fullvars[v] => c == 0 ? 0 : c * fullvars[a])
+end
+ref_aliases = [D(k) => D(x); z => D(x); D(z) => D(D(x)); zₜ => D(D(x)); ddx => D(D(x));
+               y => D(D(x)); D(zₜ) => D(D(D(x)))]
+@test Set(aliases) == Set(ref_aliases)