Mutabilities: a type-level tool for ownership-by-convention

Mutabilities.jl is a type-level tool for describing mutabilities and ownership of objects in a composable manner.

See more in the documentation.

Summary

• readonly: create read-only view
• freeze, freezevalue, freezeindex, freezeproperties: create immutable copies
• melt, meltvalue, meltindex, meltproperties: create mutable copies
• move!: manually elides copies with freeze/melt APIs.

High-level interface

Read-only view

The most easy-to-use interface is readonly(x) which creates a read-only "view" to x:

julia> using Mutabilities

julia> x = [1, 2, 3];

julia> z = readonly(x)
3-element readonly(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

julia> z[1] = 111
ERROR: setindex! not defined for Mutabilities.ReadOnlyArray{Int64,1,Array{Int64,1}}

Note that changes in x would still be reflected to z:

julia> x[1] = 111;

julia> z
3-element readonly(::Array{Int64,1}) with eltype Int64:
 111
   2
   3

Freeze/melt

Use freeze(x) to get an independent immutable (shallow) copy of x:

julia> x = [1, 2, 3];

julia> z = freeze(x)
3-element freeze(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

julia> x[1] = 111;

julia> z
3-element freeze(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

freeze can be reverted by melt:

julia> y = melt(z)
3-element Array{Int64,1}:
 1
 2
 3

It returns an independent mutable (shallow) copy of y. Thus, y can be safely mutated:

julia> y[1] = 111;

julia> z
3-element freeze(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

Example usage

Julia's view is dangerous to use if the indices can be mutated after creating it:

idx = [1, 1, 1]
x = view([1], idx)
x[1]  # OK
idx[1] = 10_000_000_000
x[1]  # segfault

This can be avoided by freezing the index array:

view([1], freeze(idx))

Note that readonly is not enough.

Variants

freeze and melt work both on indices (keys) and values. It is possible to create an append-only vector by freezing the values:

julia> append_only = freezevalue([1, 2, 3]);

julia> push!(append_only, 4)
4-element freezevalue(::Array{Int64,1}) with eltype Int64:
 1
 2
 3
 4

julia> append_only[1] = 1
ERROR: setindex! not defined for Mutabilities.AppendOnlyVector{Int64,Array{Int64,1}}

It is possible to create a shape-frozen vector by freezing the indices:

julia> shape_frozen = freezeindex([1, 2, 3])
3-element freezeindex(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

julia> shape_frozen .*= 10
3-element freezeindex(::Array{Int64,1}) with eltype Int64:
 10
 20
 30

julia> push!(shape_frozen, 4)
ERROR: push! on freezeindex(::Array{Int64,1}) not allowed

Low-level interface

Using freeze and melt at API boundaries is a good way to ensure correctness of the programs. However, until the julia compiler gets a borrow checker and automatically elides such copies, it may be very expensive to use them in some situations. Until then, Mutabilities.jl provides an "escape hatch"; i.e., an API to let the programmer declare that there is no sharing of the given object:

julia> z = freeze(move!([1, 2, 3]))  # no copy
3-element freeze(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

julia> melt(move!(z))  # no copy
3-element Array{Int64,1}:
 1
 2
 3

This allows Julia programs to compose well, without defining immutable f and mutable f! variants of the API and without documenting the particular memory ownership for each function.

For example, melt is simply defined as

melt(x) = meltvalue(move!(meltindex(x)))

move! can be useful when, e.g., input values can be re-used for output values:

julia> function add(x, y)
           out = melt(x)
           out .+= y
           return freeze(move!(out))
       end;

julia> x = ones(3)
       y = ones(3);

julia> z = add(move!(x), y)  # no allocations
3-element freeze(::Array{Float64,1}) with eltype Float64:
 2.0
 2.0
 2.0

julia> melt(move!(z)) === x  # `x` is mutated
true

(Note: Above example intentionally violates the rule for using move! to show how it works. After add(move!(x), y), it is not allowed to use x, as done in the last statement.)

Supported collections and types

• AbstractArray
• AbstractDict
• AbstractSet
• Data types ("plain struct")

Interop

StaticArrays

Static arrays are converted to appropriate types instead of the wrapper arrays:

julia> using StaticArrays

julia> a = SA[1, 2, 3]
3-element SArray{Tuple{3},Int64,1,3} with indices SOneTo(3):
 1
 2
 3

julia> melt(a)
3-element Array{Int64,1}:
 1
 2
 3

julia> meltvalue(a)
3-element MArray{Tuple{3},Int64,1,3} with indices SOneTo(3):
 1
 2
 3

julia> freeze(MVector(1, 2, 3))  # or freezevalue
3-element SArray{Tuple{3},Int64,1,3} with indices SOneTo(3):
 1
 2
 3

StructArrays

Mutabilities.jl is aware of mutability of each field arrays wrapped in struct arrays:

julia> using StructArrays

julia> x = StructArray(a = 1:3);  # x.a is not mutable

julia> y = melt(x)
3-element StructArray(::Array{Int64,1}) with eltype NamedTuple{(:a,),Tuple{Int64}}:
 (a = 1,)
 (a = 2,)
 (a = 3,)

julia> y.a
3-element Array{Int64,1}:
 1
 2
 3

julia> z = freeze(StructArray(a = [1, 2, 3]))
3-element freeze(StructArray(::Array{Int64,1})) with eltype NamedTuple{(:a,),Tuple{Int64}}:
 (a = 1,)
 (a = 2,)
 (a = 3,)

julia> z.a
3-element freeze(::Array{Int64,1}) with eltype Int64:
 1
 2
 3

Related packages

• https://github.com/andyferris/Freeze.jl
• https://github.com/bkamins/ReadOnlyArrays.jl