remove bad eltype methods

Author: bvdmitri

src/helpers/algebra/companion_matrix.jl

@@ -20,9 +20,9 @@ struct CompanionMatrix{R <: Real, T <: AbstractVector{R}} <: AbstractMatrix{R}
     θ::T
 end
 
-Base.eltype(::CompanionMatrix{R}) where {R} = R
-Base.size(cmatrix::CompanionMatrix)         = (length(cmatrix.θ), length(cmatrix.θ))
-Base.length(cmatrix::CompanionMatrix)       = prod(size(cmatrix))
+Base.eltype(::Type{<:CompanionMatrix{R}}) where {R} = R
+Base.size(cmatrix::CompanionMatrix) = (length(cmatrix.θ), length(cmatrix.θ))
+Base.length(cmatrix::CompanionMatrix) = prod(size(cmatrix))
 
 Base.getindex(cmatrix::CompanionMatrix, i::Int) = getindex(cmatrix, map(r -> r + 1, reverse(divrem(i - 1, first(size(cmatrix)))))...)
 
@@ -40,9 +40,9 @@ struct CompanionMatrixTransposed{R <: Real, T <: AbstractVector{R}} <: AbstractM
     θ::T
 end
 
-Base.eltype(::CompanionMatrixTransposed{R}) where {R} = R
-Base.size(cmatrix::CompanionMatrixTransposed)         = (length(cmatrix.θ), length(cmatrix.θ))
-Base.length(cmatrix::CompanionMatrixTransposed)       = prod(size(cmatrix))
+Base.eltype(::Type{<:CompanionMatrixTransposed{R}}) where {R} = R
+Base.size(cmatrix::CompanionMatrixTransposed) = (length(cmatrix.θ), length(cmatrix.θ))
+Base.length(cmatrix::CompanionMatrixTransposed) = prod(size(cmatrix))
 
 Base.getindex(cmatrix::CompanionMatrixTransposed, i::Int) = getindex(cmatrix, map(r -> r + 1, reverse(divrem(i - 1, first(size(cmatrix)))))...)
 

src/helpers/algebra/permutation_matrix.jl

@@ -40,7 +40,8 @@ function PermutationMatrix(dim::T; switch_first::Bool = true) where {T <: Intege
 end
 
 # extensions of base functionality
-Base.eltype(::PermutationMatrix{T}) where {T} = T
+Base.eltype(::Type{<:PermutationMatrix{T}}) where {T} = T
+
 function Base.size(mat::PermutationMatrix)
     nr_elements = length(mat.ind)
     return (nr_elements, nr_elements)

src/helpers/algebra/standard_basis_vector.jl

@@ -28,7 +28,6 @@ function StandardBasisVector(length::Int, index::Int, scale::T = one(Int)) where
     return StandardBasisVector{T}(length, index, scale)
 end
 
-Base.eltype(::StandardBasisVector{T}) where {T}       = T
 Base.eltype(::Type{StandardBasisVector{T}}) where {T} = T
 
 # extensions of base functionality

src/marginal.jl

@@ -133,10 +133,13 @@ MacroHelpers.@proxy_methods Marginal getdata [
     Base.precision,
     Base.length,
     Base.ndims,
-    Base.size,
-    Base.eltype
+    Base.size
 ]
 
+# Eltype is special here, because it should be only defined on types
+# Otherwise it causes invalidations and slower compile times
+Base.eltype(::Type{<:Marginal{D}}) where {D} = Base.eltype(D)
+
 Distributions.mean(fn::Function, marginal::Marginal) = mean(fn, getdata(marginal))
 
 """

src/message.jl

@@ -136,9 +136,8 @@ function multiply_messages(prod_strategy, left::Message, right::Message)
     return Message(new_dist, is_prod_clamped, is_prod_initial, new_addons)
 end
 
-constrain_form_as_message(message::Message, form_constraint) = Message(
-    constrain_form(form_constraint, getdata(message)), is_clamped(message), is_initial(message), getaddons(message)
-)
+constrain_form_as_message(message::Message, form_constraint) =
+    Message(constrain_form(form_constraint, getdata(message)), is_clamped(message), is_initial(message), getaddons(message))
 
 # Note: we need extra Base.Generator(as_message, messages) step here, because some of the messages might be VMP messages
 # We want to cast it explicitly to a Message structure (which as_message does in case of DeferredMessage)
@@ -187,10 +186,13 @@ MacroHelpers.@proxy_methods Message getdata [
     Base.precision,
     Base.length,
     Base.ndims,
-    Base.size,
-    Base.eltype
+    Base.size
 ]
 
+# Eltype is special here, because it should be only defined on types
+# Otherwise it causes invalidations and slower compile times
+Base.eltype(::Type{<:Message{D}}) where {D} = Base.eltype(D)
+
 Distributions.mean(fn::Function, message::Message) = mean(fn, getdata(message))
 
 ## Deferred Message

src/nodes/predefined/autoregressive.jl

@@ -128,7 +128,6 @@ Base.size(precision::ARPrecisionMatrix) = (precision.order, precision.order)
 Base.getindex(precision::ARPrecisionMatrix, i::Int, j::Int) = (i === 1 && j === 1) ? precision.γ : ((i === j) ? convert(eltype(precision), huge) : zero(eltype(precision)))
 
 Base.eltype(::Type{<:ARPrecisionMatrix{T}}) where {T} = T
-Base.eltype(::ARPrecisionMatrix{T}) where {T}         = T
 
 Base.convert(::Type{AbstractArray{T}}, matrix::ARPrecisionMatrix{R}) where {T, R} = ARPrecisionMatrix(matrix.order, convert(T, matrix.γ))
 Base.convert(::Type{AbstractArray{T}}, matrix::ARPrecisionMatrix{T}) where {T} = matrix
@@ -165,7 +164,6 @@ Base.size(transition::ARTransitionMatrix) = (transition.order, transition.order)
 Base.getindex(transition::ARTransitionMatrix, i::Int, j::Int) = (i === 1 && j === 1) ? transition.inv_γ : zero(eltype(transition))
 
 Base.eltype(::Type{<:ARTransitionMatrix{T}}) where {T} = T
-Base.eltype(::ARTransitionMatrix{T}) where {T}         = T
 
 Base.convert(::Type{AbstractArray{T}}, matrix::ARTransitionMatrix{R}) where {T, R} = ARTransitionMatrix{T}(matrix.order, convert(T, matrix.inv_γ))
 Base.convert(::Type{AbstractArray{T}}, matrix::ARTransitionMatrix{T}) where {T} = matrix

src/nodes/predefined/gcv.jl

@@ -14,7 +14,6 @@ ExponentialLinearQuadratic(approximation, a::Real, b::Real, c::Real, d::Real)
 ExponentialLinearQuadratic(approximation, a::Integer, b::Integer, c::Integer, d::Integer) = ExponentialLinearQuadratic(approximation, float(a), float(b), float(c), float(d))
 
 Base.eltype(::Type{<:ExponentialLinearQuadratic{A, T}}) where {A, T} = T
-Base.eltype(::ExponentialLinearQuadratic{A, T}) where {A, T}         = T
 
 Base.precision(dist::ExponentialLinearQuadratic) = mean_invcov(dist)[2]
 
@@ -79,8 +78,8 @@ default_meta(::Type{GCV}) = DefaultGCVNodeMetadata
 @average_energy GCV (q_y_x::MultivariateNormalDistributionsFamily, q_z::NormalDistributionsFamily, q_κ::Any, q_ω::Any, meta::Union{<:GCVMetadata, Nothing}) = begin
     y_x_mean, y_x_cov = mean_cov(q_y_x)
     z_mean, z_var     = mean_var(q_z)
-    κ_mean, κ_var   = mean_var(q_κ)
-    ω_mean, ω_var   = mean_var(q_ω)
+    κ_mean, κ_var     = mean_var(q_κ)
+    ω_mean, ω_var     = mean_var(q_ω)
 
     ksi = (κ_mean^2) * z_var + κ_var * ((z_mean^2) + z_var)
     psi = @inbounds (y_x_mean[2] - y_x_mean[1])^2 + y_x_cov[1, 1] + y_x_cov[2, 2] - y_x_cov[1, 2] - y_x_cov[2, 1]