Restrict wrapped types to reduce invalidations

src/array.jl

@@ -19,15 +19,41 @@ function reftype(sz::Int)
     end
 end
 
+# This check is only there to print a user-friendly warning before
+# a TypeError is thrown due to restrictions in the type signature
+function check_supported_eltype(::Type{T}, ::Type{U}) where {T, U}
+    T === Symbol &&
+    throw(ArgumentError("CategoricalArray no longer supports Symbol as element type "*
+                        "as that forces recompiling too many Julia Base methods: " *
+                        "use strings instead, e.g. via categorical(string.(x))"))
+    T <: Union{SupportedTypes, Missing} ||
+        throw(ArgumentError("CategoricalArray only supports " *
+                            "AbstractString, AbstractChar and Number element types " *
+                            "(got element type $U)"))
+end
+
 fixstringtype(T::Type) = T <: SubString || T === AbstractString ? String : T
 fixstringtype(T::Union) = Union{fixstringtype(T.a), fixstringtype(T.b)}
 fixstringtype(::Type{Union{}}) = Union{}
 
+# Find a narrow type that is supported to hold all elements if possible
+function fixtype(A::AbstractArray{T}) where T
+    if T <: Union{SupportedTypes, Missing}
+        return fixstringtype(T)
+    else
+        U = fixstringtype(mapreduce(typeof, Base.promote_typejoin, A))
+        check_supported_eltype(U, T)
+        return U
+    end
+end
+
 """
     CategoricalArray{T}(undef, dims::Dims; levels=nothing, ordered=false)
     CategoricalArray{T}(undef, dims::Int...; levels=nothing, ordered=false)
 
-Construct an uninitialized `CategoricalArray` with levels of type `T` and dimensions `dim`.
+Construct an uninitialized `CategoricalArray` with levels of type
+`T <: $SupportedTypes` and dimensions `dims`.
+
 The `levels` keyword argument can be a vector specifying possible values for the data
 (this is equivalent to but more efficient than calling [`levels!`](@ref)
 on the resulting array).
@@ -52,8 +78,6 @@ in ascending order; else, they are kept in their order of appearance in `A`.
 The `ordered` keyword argument determines whether the array values can be compared
 according to the ordering of levels or not (see [`isordered`](@ref)).
 
-    CategoricalArray(A::CategoricalArray; levels=nothing, ordered=false)
-
 If `A` is already a `CategoricalArray`, its levels, orderedness and reference type
 are preserved unless explicitly overriden.
 """
@@ -62,7 +86,8 @@ function CategoricalArray end
 """
     CategoricalVector{T}(undef, m::Int; levels=nothing, ordered=false)
 
-Construct an uninitialized `CategoricalVector` with levels of type `T` and dimensions `dim`.
+Construct an uninitialized `CategoricalVector` with levels of type
+`T <: $SupportedTypes` and dimensions `dim`.
 
 The `levels` keyword argument can be a vector specifying possible values for the data
 (this is equivalent to but more efficient than calling [`levels!`](@ref)
@@ -87,8 +112,6 @@ in ascending order; else, they are kept in their order of appearance in `A`.
 The `ordered` keyword argument determines whether the array values can be compared
 according to the ordering of levels or not (see [`isordered`](@ref)).
 
-    CategoricalVector(A::CategoricalVector; levels=nothing, ordered=false)
-
 If `A` is already a `CategoricalVector`, its levels, orderedness and reference type
 are preserved unless explicitly overriden.
 """
@@ -97,7 +120,8 @@ function CategoricalVector end
 """
     CategoricalMatrix{T}(undef, m::Int, n::Int; levels=nothing, ordered=false)
 
-Construct an uninitialized `CategoricalMatrix` with levels of type `T` and dimensions `dim`.
+Construct an uninitialized `CategoricalMatrix` with levels of type
+`T <: $SupportedTypes` and dimensions `dim`.
 The `ordered` keyword argument determines whether the array values can be compared
 according to the ordering of levels or not (see [`isordered`](@ref)).
 
@@ -118,8 +142,6 @@ in ascending order; else, they are kept in their order of appearance in `A`.
 The `ordered` keyword argument determines whether the array values can be compared
 according to the ordering of levels or not (see [`isordered`](@ref)).
 
-    CategoricalMatrix(A::CategoricalMatrix; levels=nothing, ordered=isordered(A))
-
 If `A` is already a `CategoricalMatrix`, its levels, orderedness and reference type
 are preserved unless explicitly overriden.
 """
@@ -137,6 +159,7 @@ function CategoricalArray{T, N, R}(::UndefInitializer, dims::NTuple{N,Int};
                                    ordered::Bool=false) where {T, N, R}
     U = leveltype(nonmissingtype(T))
     S = T >: Missing ? Union{U, Missing} : U
+    check_supported_eltype(S, T)
     V = CategoricalValue{U, R}
     levs = levels === nothing ? U[] : collect(U, levels)
     CategoricalArray{S, N}(zeros(R, dims), CategoricalPool{U, R, V}(levs, ordered))
@@ -231,6 +254,7 @@ function CategoricalArray{T, N, R}(A::AbstractArray;
 end
 
 # From AbstractArray
+
 CategoricalArray{T, N}(A::AbstractArray{S, N};
                        levels::Union{AbstractVector, Nothing}=nothing,
                        ordered::Bool=_isordered(A)) where {S, T, N} =
@@ -242,17 +266,17 @@ CategoricalArray{T}(A::AbstractArray{S, N};
 CategoricalArray(A::AbstractArray{T, N};
                  levels::Union{AbstractVector, Nothing}=nothing,
                  ordered::Bool=_isordered(A)) where {T, N} =
-    CategoricalArray{fixstringtype(T), N}(A, levels=levels, ordered=ordered)
+    CategoricalArray{fixtype(A), N}(A, levels=levels, ordered=ordered)
 
 CategoricalVector(A::AbstractVector{T};
                   levels::Union{AbstractVector, Nothing}=nothing,
                   ordered::Bool=_isordered(A)) where {T} =
-    CategoricalArray{fixstringtype(T), 1}(A, levels=levels, ordered=ordered)
+    CategoricalArray{fixtype(A), 1}(A, levels=levels, ordered=ordered)
 
 CategoricalMatrix(A::AbstractMatrix{T};
                   levels::Union{AbstractVector, Nothing}=nothing,
                   ordered::Bool=_isordered(A)) where {T} =
-    CategoricalArray{fixstringtype(T), 2}(A, levels=levels, ordered=ordered)
+    CategoricalArray{fixtype(A), 2}(A, levels=levels, ordered=ordered)
 
 # From CategoricalArray (preserve R)
 CategoricalArray{T, N}(A::CategoricalArray{S, N, R};
@@ -286,12 +310,12 @@ convert(::Type{CategoricalArray{T, N}}, A::AbstractArray{S, N}) where {S, T, N}
 convert(::Type{CategoricalArray{T}}, A::AbstractArray{S, N}) where {S, T, N} =
     convert(CategoricalArray{T, N}, A)
 convert(::Type{CategoricalArray}, A::AbstractArray{T, N}) where {T, N} =
-    convert(CategoricalArray{T, N}, A)
+    convert(CategoricalArray{fixtype(A), N}, A)
 
 convert(::Type{CategoricalVector{T}}, A::AbstractVector) where {T} =
     convert(CategoricalVector{T, DefaultRefType}, A)
 convert(::Type{CategoricalVector}, A::AbstractVector{T}) where {T} =
-    convert(CategoricalVector{T}, A)
+    convert(CategoricalVector{fixtype(A)}, A)
 convert(::Type{CategoricalVector{T}},
         A::CategoricalVector{S, R}) where {S, T, R <: Integer} =
     convert(CategoricalVector{T, R}, A)
@@ -301,7 +325,7 @@ convert(::Type{CategoricalVector}, A::CategoricalVector) = A
 convert(::Type{CategoricalMatrix{T}}, A::AbstractMatrix) where {T} =
     convert(CategoricalMatrix{T, DefaultRefType}, A)
 convert(::Type{CategoricalMatrix}, A::AbstractMatrix{T}) where {T} =
-    convert(CategoricalMatrix{T}, A)
+    convert(CategoricalMatrix{fixtype(A)}, A)
 convert(::Type{CategoricalMatrix{T}},
         A::CategoricalMatrix{S, R}) where {S, T, R <: Integer} =
     convert(CategoricalMatrix{T, R}, A)
@@ -313,6 +337,8 @@ convert(::Type{CategoricalArray{T, N, R}}, A::AbstractArray{S, N}) where {S, T,
 
 function _convert(::Type{CategoricalArray{T, N, R}}, A::AbstractArray{S, N};
                   levels::Union{AbstractVector, Nothing}=nothing) where {S, T, N, R}
+    check_supported_eltype(T, T)
+
     res = CategoricalArray{T, N, R}(undef, size(A), levels=levels)
     copyto!(res, A)
 
@@ -698,7 +724,7 @@ function vcat(A::CategoricalArray...)
         [x==0 ? 0 : ii[x] for x in a.refs]::Array{Int,ndims(a)}
     end
 
-    T = Base.promote_eltype(A...) >: Missing ?
+    T = cat_promote_eltype(A...) >: Missing ?
         Union{eltype(newlevels), Missing} : eltype(newlevels)
     refs = DefaultRefType[refsvec...;]
     pool = CategoricalPool(newlevels, ordered)
@@ -912,15 +938,15 @@ are preserved unless explicitly overriden.
                              compress::Bool=false) where {T, N}
     # @inline is needed so that return type is inferred when compress is not provided
     RefType = compress ? reftype(length(unique(A))) : DefaultRefType
-    CategoricalArray{fixstringtype(T), N, RefType}(A, levels=levels, ordered=ordered)
+    CategoricalArray{fixtype(A), N, RefType}(A, levels=levels, ordered=ordered)
 end
 @inline function categorical(A::CategoricalArray{T, N, R};
                              levels::Union{AbstractVector, Nothing}=nothing,
                              ordered=_isordered(A),
                              compress::Bool=false) where {T, N, R}
     # @inline is needed so that return type is inferred when compress is not provided
     RefType = compress ? reftype(length(CategoricalArrays.levels(A))) : R
-    CategoricalArray{fixstringtype(T), N, RefType}(A, levels=levels, ordered=ordered)
+    CategoricalArray{T, N, RefType}(A, levels=levels, ordered=ordered)
 end
 
 function in(x::Any, y::CategoricalArray{T, N, R}) where {T, N, R}
@@ -1046,21 +1072,11 @@ end
 
 StructTypes.construct(::Type{<:CategoricalArray{Union{Missing, T}}},
                       A::AbstractVector) where {T} =
-    categoricalmissing(T, A)
+    CategoricalArray{Union{Missing, T}}(replace(A, nothing=>missing))
 StructTypes.construct(::Type{<:CategoricalArray{Union{Missing, T}}},
                       A::Vector) where {T} =
-    categoricalmissing(T, A)
-categoricalmissing(T, A::AbstractVector) =
     CategoricalArray{Union{Missing, T}}(replace(A, nothing=>missing))
 
-StructTypes.construct(::Type{<:CategoricalArray{Union{Nothing, T}}},
-                      A::AbstractVector) where {T} =
-    categoricalnothing(T, A)
-StructTypes.construct(::Type{<:CategoricalArray{Union{Nothing, T}}},
-                      A::Vector) where {T} =
-    categoricalnothing(T, A)
-categoricalnothing(T, A::AbstractVector) = CategoricalArray{Union{Nothing, T}}(A)
-
 # DataAPI refarray/refvalue/refpool support
 struct CategoricalRefPool{T, P} <: AbstractVector{T}
     pool::P

src/pool.jl

@@ -62,7 +62,7 @@ avoid doing a dict lookup twice
 end
 
 function mergelevels(ordered, levels...)
-    T = Base.promote_eltype(levels...)
+    T = cat_promote_eltype(levels...)
     res = Vector{T}(undef, 0)
 
     nonempty_lv = findfirst(!isempty, levels)

src/recode.jl

@@ -274,8 +274,9 @@ recode!(a::AbstractArray, default::Any, pairs::Pair...) =
     recode!(a, a, default, pairs...)
 recode!(a::AbstractArray, pairs::Pair...) = recode!(a, a, nothing, pairs...)
 
-promote_valuetype(x::Pair{K, V}) where {K, V} = V
-promote_valuetype(x::Pair{K, V}, y::Pair...) where {K, V} = promote_type(V, promote_valuetype(y...))
+cat_promote_valuetype(x::Pair{K, V}) where {K, V} = V
+cat_promote_valuetype(x::Pair{K, V}, y::Pair...) where {K, V} =
+    cat_promote_type(V, cat_promote_valuetype(y...))
 
 keytype_hasmissing(x::Pair{K}) where {K} = K === Missing
 keytype_hasmissing(x::Pair{K}, y::Pair...) where {K} = K === Missing || keytype_hasmissing(y...)
@@ -350,11 +351,11 @@ recode(a::AbstractArray, pairs::Pair...) = recode(a, nothing, pairs...)
 recode(a::CategoricalArray, pairs::Pair...) = recode(a, nothing, pairs...)
 
 function recode(a::AbstractArray, default::Any, pairs::Pair...)
-    V = promote_valuetype(pairs...)
+    V = cat_promote_valuetype(pairs...)
     # T cannot take into account eltype(src), since we can't know
     # whether it matters at compile time (all levels recoded or not)
     # and using a wider type than necessary would be annoying
-    T = default isa Nothing ? V : promote_type(typeof(default), V)
+    T = default isa Nothing ? V : cat_promote_type(typeof(default), V)
     # Exception 1: if T === Missing and default not missing,
     # assume the caller wants to recode only some values to missing,
     # but accept original values
@@ -371,11 +372,11 @@ function recode(a::AbstractArray, default::Any, pairs::Pair...)
 end
 
 function recode(a::CategoricalArray{S, N, R}, default::Any, pairs::Pair...) where {S, N, R}
-    V = promote_valuetype(pairs...)
+    V = cat_promote_valuetype(pairs...)
     # T cannot take into account eltype(src), since we can't know
     # whether it matters at compile time (all levels recoded or not)
     # and using a wider type than necessary would be annoying
-    T = default isa Nothing ? V : promote_type(typeof(default), V)
+    T = default isa Nothing ? V : cat_promote_type(typeof(default), V)
     # Exception 1: if T === Missing and default not missing,
     # assume the caller wants to recode only some values to missing,
     # but accept original values
@@ -396,13 +397,13 @@ end
 function Base.replace(a::CategoricalArray{S, N, R}, pairs::Pair...) where {S, N, R}
     # Base.replace(a::Array, pairs::Pair...) uses a wider type promotion than
     # recode. It promotes the source type S with the replaced types T.
-    T = promote_valuetype(pairs...)
+    T = cat_promote_valuetype(pairs...)
     # Exception: replacing missings
     # Example: replace(categorical([missing,1.5]), missing=>0)
     if keytype_hasmissing(pairs...)
-        dest = CategoricalArray{promote_type(nonmissingtype(S), T), N, R}(undef, size(a))
+        dest = CategoricalArray{cat_promote_type(nonmissingtype(S), T), N, R}(undef, size(a))
     else
-        dest = CategoricalArray{promote_type(S, T), N, R}(undef, size(a))
+        dest = CategoricalArray{cat_promote_type(S, T), N, R}(undef, size(a))
     end
     recode!(dest, a, nothing, pairs...)
 end

src/typedefs.jl

@@ -1,12 +1,13 @@
 const DefaultRefType = UInt32
+const SupportedTypes = Union{AbstractString, AbstractChar, Number}
 
 ## Pools
 
 # Type params:
 # * `T` type of categorized values
 # * `R` integer type for referencing category levels
 # * `V` categorical value type
-mutable struct CategoricalPool{T, R <: Integer, V}
+mutable struct CategoricalPool{T <: SupportedTypes, R <: Integer, V}
     levels::Vector{T}       # category levels ordered by their reference codes
     invindex::Dict{T, R}    # map from category levels to their reference codes
     valindex::Vector{V}     # "category value" objects 1-to-1 matching `index`
@@ -42,9 +43,6 @@ mutable struct CategoricalPool{T, R <: Integer, V}
     function CategoricalPool{T, R, V}(levels::Vector{T},
                                       invindex::Dict{T, R},
                                       ordered::Bool) where {T, R, V}
-        if T <: CategoricalValue && T !== Union{}
-            throw(ArgumentError("Level type $T cannot be a categorical value type"))
-        end
         if !(V <: CategoricalValue)
             throw(ArgumentError("Type $V is not a categorical value type"))
         end
@@ -70,7 +68,7 @@ end
 ## Values
 
 """
-    CategoricalValue{T, R <: Integer}
+    CategoricalValue{T <: $SupportedTypes, R <: Integer}
 
 A wrapper around a value of type `T` corresponding to a level
 in a `CategoricalPool`.
@@ -82,7 +80,7 @@ if [`isordered`](@ref) is `true` for the value's pool, and in that case
 the order of the pool's [`levels`](@ref DataAPI.levels) is used rather than the standard
 ordering of values of type `T`.
 """
-struct CategoricalValue{T, R <: Integer}
+struct CategoricalValue{T <: SupportedTypes, R <: Integer}
     level::R
     pool::CategoricalPool{T, R, CategoricalValue{T, R}}
 end
@@ -96,7 +94,9 @@ end
 # * `V` original type of elements (excluding Missing) before categorization
 # * `C` categorical value type
 # * `U` type of missing value, `Union{}` if missing values are not accepted
-abstract type AbstractCategoricalArray{T, N, R, V, C, U} <: AbstractArray{Union{C, U}, N} end
+abstract type AbstractCategoricalArray{T <: Union{CategoricalValue, SupportedTypes, Missing}, N,
+                                       R <: Integer, V, C <: CategoricalValue, U} <:
+    AbstractArray{Union{C, U}, N} end
 const AbstractCategoricalVector{T, R, V, C, U} = AbstractCategoricalArray{T, 1, R, V, C, U}
 const AbstractCategoricalMatrix{T, R, V, C, U} = AbstractCategoricalArray{T, 2, R, V, C, U}