Cache subexpressions when building MOI.ScalarNonlinearExpression

.vale.ini

@@ -15,6 +15,7 @@ BasedOnStyles = Vale, Google
 
 # TODO(odow): fix all of these
 Google.Ellipses = OFF
+Google.EmDash = OF
 Google.Exclamation = OFF
 Google.FirstPerson = OFF
 Google.OptionalPlurals = OFF

src/JuMP.jl

@@ -162,6 +162,8 @@ mutable struct GenericModel{T<:Real} <: AbstractModel
     # A dictionary to store timing information from the JuMP macros.
     enable_macro_timing::Bool
     macro_times::Dict{Tuple{LineNumberNode,String},Float64}
+    # A cache to track common subexpressions based on their `objectid`.
+    subexpressions::Dict{UInt64,MOI.ScalarNonlinearFunction}
 end
 
 value_type(::Type{GenericModel{T}}) where {T} = T
@@ -276,6 +278,7 @@ function direct_generic_model(
         Dict{Any,MOI.ConstraintIndex}(),
         false,
         Dict{Tuple{LineNumberNode,String},Float64}(),
+        Dict{UInt64,MOI.ScalarNonlinearFunction}(),
     )
 end
 

src/constraints.jl

@@ -769,6 +769,10 @@ function moi_function(constraint::AbstractConstraint)
     return moi_function(jump_function(constraint))
 end
 
+function moi_function(model, constraint::AbstractConstraint)
+    return moi_function(model, jump_function(constraint))
+end
+
 """
     moi_set(constraint::AbstractConstraint)
 
@@ -1028,6 +1032,18 @@ function _moi_add_constraint(
     return MOI.add_constraint(model, f, s)
 end
 
+function check_belongs_to_model(f::Vector, model)
+    for func in f
+        check_belongs_to_model(func, model)
+    end
+    return
+end
+
+function moi_function(model, f)
+    check_belongs_to_model(f, model)
+    return moi_function(f)
+end
+
 """
     add_constraint(
         model::GenericModel,
@@ -1044,10 +1060,9 @@ function add_constraint(
     name::String = "",
 )
     con = model_convert(model, con)
+    func, set = moi_function(model, con), moi_set(con)
     # The type of backend(model) is unknown so we directly redirect to another
     # function.
-    check_belongs_to_model(con, model)
-    func, set = moi_function(con), moi_set(con)
     cindex = _moi_add_constraint(
         backend(model),
         func,

src/nlp_expr.jl

@@ -596,18 +596,29 @@ moi_function(x::Number) = x
 # `moi_function(AbstractArray{<:AbstractVariableRef})`
 moi_function(x::AbstractArray) = moi_function.(x)
 
-function moi_function(f::GenericNonlinearExpr{V}) where {V}
+function moi_function(model::GenericModel, f::GenericNonlinearExpr{V}) where {V}
+    key = objectid(f)
+    if haskey(model.subexpressions, key)
+        return model.subexpressions[key]
+    end
     ret = MOI.ScalarNonlinearFunction(f.head, similar(f.args))
     stack = Tuple{MOI.ScalarNonlinearFunction,Int,GenericNonlinearExpr{V}}[]
     for i in length(f.args):-1:1
         if f.args[i] isa GenericNonlinearExpr{V}
             push!(stack, (ret, i, f.args[i]))
+        elseif f.args[i] isa AbstractJuMPScalar
+            ret.args[i] = moi_function(model, f.args[i])
         else
             ret.args[i] = moi_function(f.args[i])
         end
     end
     while !isempty(stack)
         parent, i, arg = pop!(stack)
+        arg_key = objectid(arg)
+        if haskey(model.subexpressions, arg_key)
+            parent.args[i] = model.subexpressions[arg_key]
+            continue
+        end
         child = MOI.ScalarNonlinearFunction(arg.head, similar(arg.args))
         parent.args[i] = child
         for j in length(arg.args):-1:1
@@ -617,7 +628,9 @@ function moi_function(f::GenericNonlinearExpr{V}) where {V}
                 child.args[j] = moi_function(arg.args[j])
             end
         end
+        model.subexpressions[arg_key] = child
     end
+    model.subexpressions[key] = ret
     return ret
 end
 
@@ -1239,7 +1252,8 @@ function moi_function(f::AbstractVector{<:GenericNonlinearExpr})
 end
 
 function MOI.VectorNonlinearFunction(f::Vector{<:AbstractJuMPScalar})
-    return MOI.VectorNonlinearFunction(map(moi_function, f))
+    model = owner_model(first(f))
+    return MOI.VectorNonlinearFunction(moi_function.(model, f))
 end
 
 """
@@ -1285,7 +1299,7 @@ x
 ```
 """
 function simplify(model::GenericModel, f::AbstractJuMPScalar)
-    g = MOI.Nonlinear.SymbolicAD.simplify(moi_function(f))
+    g = MOI.Nonlinear.SymbolicAD.simplify(moi_function(model, f))
     return jump_function(model, g)
 end
 
@@ -1336,7 +1350,8 @@ function derivative(
     f::AbstractJuMPScalar,
     x::GenericVariableRef{T},
 ) where {T}
-    df_dx = MOI.Nonlinear.SymbolicAD.derivative(moi_function(f), index(x))
+    df_dx =
+        MOI.Nonlinear.SymbolicAD.derivative(moi_function(model, f), index(x))
     return jump_function(model, MOI.Nonlinear.SymbolicAD.simplify!(df_dx))
 end
 
@@ -1381,7 +1396,7 @@ julia> ∇f[y]
 ```
 """
 function gradient(model::GenericModel{T}, f::AbstractJuMPScalar) where {T}
-    g = moi_function(f)
+    g = moi_function(model, f)
     ∇f = Dict{GenericVariableRef{T},Any}()
     for xi in MOI.Nonlinear.SymbolicAD.variables(g)
         df_dx = MOI.Nonlinear.SymbolicAD.simplify!(

src/objective.jl

@@ -280,7 +280,7 @@ end
 
 function set_objective_function(model::GenericModel, func::AbstractJuMPScalar)
     check_belongs_to_model(func, model)
-    set_objective_function(model, moi_function(func))
+    set_objective_function(model, moi_function(model, func))
     return
 end
 
@@ -299,7 +299,7 @@ function set_objective_function(
     for f in func
         check_belongs_to_model(f, model)
     end
-    set_objective_function(model, moi_function(func))
+    set_objective_function(model, moi_function(model, func))
     return
 end