base/indices.jl - master - githubmirror/julia

base/indices.jl
579 строк21 KB
Nicholas Bauer
Fix performance regression in hvcat of simple matrices (#57422)
08 апр 2026, 06:14
08 апр 2026, 06:14211af65
# This file is a part of Julia. License is MIT: https://julialang.org/license

"""
    Dims{N}

An `NTuple` of `N` `Int`s used to represent the dimensions
of an [`AbstractArray`](@ref).
"""
Dims{N} = NTuple{N,Int}
DimsInteger{N} = NTuple{N,Integer}
Indices{N} = NTuple{N,AbstractUnitRange}

## Traits for array types ##

abstract type IndexStyle end
"""
    IndexLinear()

Subtype of [`IndexStyle`](@ref) used to describe arrays which
are optimally indexed by one linear index.

A linear indexing style uses one integer index to describe the position in the array
(even if it's a multidimensional array) and column-major
ordering is used to efficiently access the elements. This means that
requesting [`eachindex`](@ref) from an array that is `IndexLinear` will return
a simple one-dimensional range, even if it is multidimensional.

A custom array that reports its `IndexStyle` as `IndexLinear` only needs
to implement indexing (and indexed assignment) with a single `Int` index;
all other indexing expressions — including multidimensional accesses — will
be recomputed to the linear index.  For example, if `A` were a `2×3` custom
matrix with linear indexing, and we referenced `A[1, 3]`, this would be
recomputed to the equivalent linear index and call `A[5]` since `1 + 2*(3 - 1) = 5`.

See also [`IndexCartesian`](@ref).
"""
struct IndexLinear <: IndexStyle end

"""
    IndexCartesian()

Subtype of [`IndexStyle`](@ref) used to describe arrays which
are optimally indexed by a Cartesian index. This is the default
for new custom [`AbstractArray`](@ref) subtypes.

A Cartesian indexing style uses multiple integer indices to describe the position in
a multidimensional array, with exactly one index per dimension. This means that
requesting [`eachindex`](@ref) from an array that is `IndexCartesian` will return
a range of [`CartesianIndices`](@ref).

A `N`-dimensional custom array that reports its `IndexStyle` as `IndexCartesian` needs
to implement indexing (and indexed assignment) with exactly `N` `Int` indices;
all other indexing expressions — including linear indexing — will
be recomputed to the equivalent Cartesian location.  For example, if `A` were a `2×3` custom
matrix with cartesian indexing, and we referenced `A[5]`, this would be
recomputed to the equivalent Cartesian index and call `A[1, 3]` since `5 = 1 + 2*(3 - 1)`.

It is significantly more expensive to compute Cartesian indices from a linear index than it is
to go the other way.  The former operation requires division — a very costly operation — whereas
the latter only uses multiplication and addition and is essentially free. This asymmetry means it
is far more costly to use linear indexing with an `IndexCartesian` array than it is to use
Cartesian indexing with an `IndexLinear` array.

See also [`IndexLinear`](@ref).
"""
struct IndexCartesian <: IndexStyle end

"""
    IndexStyle(A)
    IndexStyle(typeof(A))

`IndexStyle` specifies the "native indexing style" for array `A`. When
you define a new [`AbstractArray`](@ref) type, you can choose to implement
either linear indexing (with [`IndexLinear`](@ref)) or cartesian indexing.
If you decide to only implement linear indexing, then you must set this trait for your array
type:

    Base.IndexStyle(::Type{<:MyArray}) = IndexLinear()

The default is [`IndexCartesian()`](@ref).

Julia's internal indexing machinery will automatically (and invisibly)
recompute all indexing operations into the preferred style. This allows users
to access elements of your array using any indexing style, even when explicit
methods have not been provided.

If you define both styles of indexing for your `AbstractArray`, this
trait can be used to select the most performant indexing style. Some
methods check this trait on their inputs, and dispatch to different
algorithms depending on the most efficient access pattern. In
particular, [`eachindex`](@ref) creates an iterator whose type depends
on the setting of this trait.
"""
IndexStyle(A::AbstractArray) = IndexStyle(typeof(A))
IndexStyle(::Type{Union{}}, slurp...) = IndexLinear()
IndexStyle(::Type{<:AbstractArray}) = IndexCartesian()
IndexStyle(::Type{<:Array}) = IndexLinear()
IndexStyle(::Type{<:AbstractRange}) = IndexLinear()

IndexStyle(A::AbstractArray, B::AbstractArray) = IndexStyle(IndexStyle(A), IndexStyle(B))
IndexStyle(A::AbstractArray, B::AbstractArray...) = IndexStyle(IndexStyle(A), IndexStyle(B...))
IndexStyle(::IndexLinear, ::IndexLinear) = IndexLinear()
IndexStyle(::IndexStyle, ::IndexStyle) = IndexCartesian()

# array shape rules

promote_shape(::Tuple{}, ::Tuple{}) = ()

# Consistent error message for promote_shape mismatch, hiding type details like
# OneTo. When b ≡ nothing, it is omitted; i can be supplied for an index.
function throw_promote_shape_mismatch(a::Tuple, b::Union{Nothing,Tuple}, i = nothing)
    if a isa Tuple{Vararg{Base.OneTo}} && (b === nothing || b isa Tuple{Vararg{Base.OneTo}})
        a = map(lastindex, a)::Dims
        b === nothing || (b = map(lastindex, b)::Dims)
    end
    _has_axes = !(a isa Dims && (b === nothing || b isa Dims))
    if _has_axes
        _normalize(d) = map(x -> firstindex(x):lastindex(x), d)
        a = _normalize(a)
        b === nothing || (b = _normalize(b))
        _things = "axes "
    else
        _things = "size "
    end
    msg = IOBuffer()
    print(msg, "a has ", _things)
    print(msg, a)
    if b ≢ nothing
        print(msg, ", b has ", _things)
        print(msg, b)
    end
    if i ≢ nothing
        print(msg, ", mismatch at dim ", i)
    end
    throw(DimensionMismatch(takestring!(msg)))
end

function promote_shape(a::Tuple{Int,}, b::Tuple{Int,})
    a[1] != b[1] && throw_promote_shape_mismatch(a, b)
    return a
end

function promote_shape(a::Tuple{Int,Int}, b::Tuple{Int,})
    (a[1] != b[1] || a[2] != 1) && throw_promote_shape_mismatch(a, b)
    return a
end

promote_shape(a::Tuple{Int,}, b::Tuple{Int,Int}) = promote_shape(b, a)

function promote_shape(a::Tuple{Int, Int}, b::Tuple{Int, Int})
    (a[1] != b[1] || a[2] != b[2]) && throw_promote_shape_mismatch(a, b)
    return a
end

"""
    promote_shape(s1, s2)

Check two array shapes for compatibility, allowing trailing singleton dimensions, and return
whichever shape has more dimensions.

# Examples
```jldoctest
julia> a = fill(1, (3,4,1,1,1));

julia> b = fill(1, (3,4));

julia> promote_shape(a,b)
(Base.OneTo(3), Base.OneTo(4), Base.OneTo(1), Base.OneTo(1), Base.OneTo(1))

julia> promote_shape((2,3,1,4), (2, 3, 1, 4, 1))
(2, 3, 1, 4, 1)
```
"""
function promote_shape(a::Dims, b::Dims)
    if length(a) < length(b)
        return promote_shape(b, a)
    end
    for i=1:length(b)
        a[i] != b[i] && throw_promote_shape_mismatch(a, b, i)
    end
    for i=length(b)+1:length(a)
        a[i] != 1 && throw_promote_shape_mismatch(a, nothing, i)
    end
    return a
end

function promote_shape(a::AbstractArray, b::AbstractArray)
    promote_shape(axes(a), axes(b))
end

function promote_shape(a::Indices, b::Indices)
    if length(a) < length(b)
        return promote_shape(b, a)
    end
    for i=1:length(b)
        a[i] != b[i] && throw_promote_shape_mismatch(a, b, i)
    end
    for i=length(b)+1:length(a)
        a[i] != 1:1 && throw_promote_shape_mismatch(a, nothing, i)
    end
    return a
end

function throw_setindex_mismatch(X, I)
    if length(I) == 1
        throw(DimensionMismatch("tried to assign $(length(X)) elements to $(I[1]) destinations"))
    else
        throw(DimensionMismatch("tried to assign $(dims2string(size(X))) array to $(dims2string(I)) destination"))
    end
end

# check for valid sizes in A[I...] = X where X <: AbstractArray
# we want to allow dimensions that are equal up to permutation, but only
# for permutations that leave array elements in the same linear order.
# those are the permutations that preserve the order of the non-singleton
# dimensions.
function setindex_shape_check(X::AbstractArray, I::Integer...)
    @inline
    li = ndims(X)
    lj = length(I)
    i = j = 1
    while true
        ii = length(axes(X,i))
        jj = I[j]
        if i == li || j == lj
            while i < li
                i += 1
                ii *= length(axes(X,i))
            end
            while j < lj
                j += 1
                jj *= I[j]
            end
            if ii != jj
                throw_setindex_mismatch(X, I)
            end
            return
        end
        if ii == jj
            i += 1
            j += 1
        elseif ii == 1
            i += 1
        elseif jj == 1
            j += 1
        else
            throw_setindex_mismatch(X, I)
        end
    end
end

setindex_shape_check(X::AbstractArray) =
    (length(X)==1 || throw_setindex_mismatch(X,()))

setindex_shape_check(X::AbstractArray, i::Integer) =
    (length(X)==i || throw_setindex_mismatch(X, (i,)))

setindex_shape_check(X::AbstractArray{<:Any, 0}, i::Integer...) =
    (length(X) == prod(i) || throw_setindex_mismatch(X, i))

setindex_shape_check(X::AbstractArray{<:Any,1}, i::Integer) =
    (length(X)==i || throw_setindex_mismatch(X, (i,)))

setindex_shape_check(X::AbstractArray{<:Any,1}, i::Integer, j::Integer) =
    (length(X)==i*j || throw_setindex_mismatch(X, (i,j)))

function setindex_shape_check(X::AbstractArray{<:Any,2}, i::Integer, j::Integer)
    if length(X) != i*j
        throw_setindex_mismatch(X, (i,j))
    end
    sx1 = length(axes(X,1))
    if !(i == 1 || i == sx1 || sx1 == 1)
        throw_setindex_mismatch(X, (i,j))
    end
end

setindex_shape_check(::Any...) =
    throw(ArgumentError("indexed assignment with a single value to possibly many locations is not supported; perhaps use broadcasting `.=` instead?"))

# convert to a supported index type (array or Int)
"""
    to_index(A, i)

Convert index `i` to an `Int` or array of indices to be used as an index into array `A`.

Custom array types may specialize `to_index(::CustomArray, i)` to provide
special indexing behaviors. Note that some index types (like `Colon`) require
more context in order to transform them into an array of indices; those get
converted in the more complicated `to_indices` function. By default, this
simply calls the generic `to_index(i)`. This must return either an `Int` or an
`AbstractArray` of scalar indices that are supported by `A`.
"""
to_index(A, i) = to_index(i)

# This is ok for Array because values larger than
# typemax(Int) will BoundsError anyway
to_index(A::Array, i::UInt) = reinterpret(Int, i)

"""
    to_index(i)

Convert index `i` to an `Int` or array of `Int`s to be used as an index for all arrays.

Custom index types may specialize `to_index(::CustomIndex)` to provide special
indexing behaviors. This must return either an `Int` or an `AbstractArray` of
`Int`s.
"""
to_index(i::Integer) = convert(Int,i)::Int
to_index(i::Bool) = throw(ArgumentError("invalid index: $i of type Bool"))
to_index(I::AbstractArray{Bool}) = LogicalIndex(I)
to_index(I::AbstractArray) = I
to_index(I::AbstractArray{Union{}}) = I
to_index(I::AbstractArray{<:Union{AbstractArray, Colon}}) =
    throw(ArgumentError(LazyString("invalid index: ", limitrepr(I), " of type ", typeof(I))))
to_index(::Colon) = throw(ArgumentError("colons must be converted by to_indices(...)"))
to_index(i) = throw(ArgumentError(LazyString("invalid index: ", limitrepr(i), " of type ", typeof(i))))

# The general to_indices is mostly defined in multidimensional.jl, but this
# definition is required for bootstrap:
"""
    to_indices(A, I::Tuple)

Convert the tuple `I` to a tuple of indices for use in indexing into array `A`.

The returned tuple must only contain either `Int`s or `AbstractArray`s of
scalar indices that are supported by array `A`. It will error upon encountering
a novel index type that it does not know how to process.

For simple index types, it defers to the unexported `Base.to_index(A, i)` to
process each index `i`. While this internal function is not intended to be
called directly, `Base.to_index` may be extended by custom array or index types
to provide custom indexing behaviors.

More complicated index types may require more context about the dimension into
which they index. To support those cases, `to_indices(A, I)` calls
`to_indices(A, axes(A), I)`, which then recursively walks through both the
given tuple of indices and the dimensional indices of `A` in tandem. As such,
not all index types are guaranteed to propagate to `Base.to_index`.

# Examples
```jldoctest
julia> A = zeros(1,2,3,4);

julia> to_indices(A, (1,1,2,2))
(1, 1, 2, 2)

julia> to_indices(A, (1,1,2,20)) # no bounds checking
(1, 1, 2, 20)

julia> to_indices(A, (CartesianIndex((1,)), 2, CartesianIndex((3,4)))) # exotic index
(1, 2, 3, 4)

julia> to_indices(A, ([1,1], 1:2, 3, 4))
([1, 1], 1:2, 3, 4)

julia> to_indices(A, (1,2)) # no shape checking
(1, 2)
```
"""
to_indices(A, I::Tuple) = (@inline; to_indices(A, axes(A), I))
to_indices(A, I::Tuple{Any}) = (@inline; to_indices(A, (eachindex(IndexLinear(), A),), I))
# In simple cases, we know that we don't need to use axes(A), optimize those.
# Having this here avoids invalidations from multidimensional.jl: to_indices(A, I::Tuple{Vararg{Union{Integer, CartesianIndex}}})
to_indices(A, I::Tuple{}) = ()
to_indices(A, I::Tuple{Vararg{Int}}) = I
to_indices(A, I::Tuple{Vararg{Integer}}) = (@inline; to_indices(A, (), I))
to_indices(A, inds, ::Tuple{}) = ()
to_indices(A, inds, I::Tuple{Any, Vararg}) =
    (@inline; (to_index(A, I[1]), to_indices(A, safe_tail(inds), tail(I))...))

"""
    Slice(indices)

Represent an AbstractUnitRange of indices as a vector of the indices themselves,
with special handling to signal they represent a complete slice of a dimension (:).

Upon calling `to_indices`, Colons are converted to Slice objects to represent
the indices over which the Colon spans. Slice objects are themselves unit
ranges with the same indices as those they wrap. This means that indexing into
Slice objects with an integer always returns that exact integer, and they
iterate over all the wrapped indices, even supporting offset indices.
"""
struct Slice{T<:AbstractUnitRange} <: AbstractUnitRange{Int}
    indices::T
end
Slice(S::Slice) = S
Slice{T}(S::Slice) where {T<:AbstractUnitRange} = Slice{T}(T(S.indices))

axes(S::Slice) = (axes1(S),)
axes1(S::Slice) = IdentityUnitRange(S.indices)
axes1(S::Slice{<:AbstractOneTo{<:Integer}}) = S.indices

first(S::Slice) = first(S.indices)
last(S::Slice) = last(S.indices)
size(S::Slice) = (length(S.indices),)
getindex(S::Slice, i::Int) = (@inline; @boundscheck checkbounds(S, i); i)
getindex(S::Slice, i::AbstractUnitRange{<:Integer}) = (@inline; @boundscheck checkbounds(S, i); i)
getindex(S::Slice, i::StepRange{<:Integer}) = (@inline; @boundscheck checkbounds(S, i); i)
show(io::IO, r::Slice) = print(io, "Base.Slice(", r.indices, ")")
iterate(S::Slice, s...) = iterate(S.indices, s...)

"""
    IdentityUnitRange(range::AbstractUnitRange)

Represent an AbstractUnitRange `range` as an offset vector such that `range[i] == i`.

`IdentityUnitRange`s are frequently used as axes for offset arrays.
"""
struct IdentityUnitRange{T<:AbstractUnitRange} <: AbstractUnitRange{Int}
    indices::T
end
IdentityUnitRange(S::IdentityUnitRange) = S
IdentityUnitRange{T}(S::IdentityUnitRange) where {T<:AbstractUnitRange} = IdentityUnitRange{T}(T(S.indices))

# IdentityUnitRanges are offset and thus have offset axes, so they are their own axes
axes(S::IdentityUnitRange) = (axes1(S),)
axes1(S::IdentityUnitRange) = S
axes1(S::IdentityUnitRange{<:AbstractOneTo{<:Integer}}) = S.indices

first(S::IdentityUnitRange) = first(S.indices)
last(S::IdentityUnitRange) = last(S.indices)
size(S::IdentityUnitRange) = (length(S.indices),)
unsafe_length(S::IdentityUnitRange) = unsafe_length(S.indices)
getindex(S::IdentityUnitRange, i::Integer) = (@inline; @boundscheck checkbounds(S, i); convert(eltype(S), i))
getindex(S::IdentityUnitRange, i::Bool) = throw(ArgumentError("invalid index: $i of type Bool"))
function getindex(S::IdentityUnitRange, i::AbstractUnitRange{<:Integer})
    @inline
    @boundscheck checkbounds(S, i)
    return convert(AbstractUnitRange{eltype(S)}, i)
end
function getindex(S::IdentityUnitRange, i::AbstractUnitRange{Bool})
    @inline
    @boundscheck checkbounds(S, i)
    range(first(i) ? first(S) : last(S), length = last(i))
end
function getindex(S::IdentityUnitRange, i::StepRange{<:Integer})
    @inline
    @boundscheck checkbounds(S, i)
    return convert(AbstractRange{eltype(S)}, i)
end
function getindex(S::IdentityUnitRange, i::StepRange{Bool})
    @inline
    @boundscheck checkbounds(S, i)
    range(first(i) ? first(S) : last(S), length = last(i), step = Int(step(i)))
end
# Indexing with offset ranges should preserve the axes of the indices
# however, this is only really possible in general with OffsetArrays.
# In some cases, though, we may obtain correct results using Base ranges
# the following methods are added to allow OffsetArrays to dispatch on the first argument without ambiguities
function getindex(S::IdentityUnitRange{<:AbstractUnitRange{<:Integer}},
                    i::IdentityUnitRange{<:AbstractUnitRange{<:Integer}})
    @inline
    @boundscheck checkbounds(S, i)
    return i
end
function getindex(S::Slice{<:AbstractUnitRange{<:Integer}},
                    i::IdentityUnitRange{<:AbstractUnitRange{<:Integer}})
    @inline
    @boundscheck checkbounds(S, i)
    return i
end
show(io::IO, r::IdentityUnitRange) = print(io, "Base.IdentityUnitRange(", r.indices, ")")
iterate(S::IdentityUnitRange, s...) = iterate(S.indices, s...)

# For AbstractOneTo, the values and indices of the values are identical, so this may be defined in Base.
# In general such an indexing operation would produce offset ranges
# This should also ideally return an AbstractUnitRange{eltype(S)}, but currently
# we're restricted to eltype(::IdentityUnitRange) == Int by definition
function getindex(S::AbstractOneTo{<:Integer}, I::IdentityUnitRange{<:AbstractUnitRange{<:Integer}})
    @inline
    @boundscheck checkbounds(S, I)
    return I
end

"""
    LinearIndices(A::AbstractArray)

Return a `LinearIndices` array with the same shape and [`axes`](@ref) as `A`,
holding the linear index of each entry in `A`. Indexing this array with
cartesian indices allows mapping them to linear indices.

For arrays with conventional indexing (indices start at 1), or any multidimensional
array, linear indices range from 1 to `length(A)`. However, for `AbstractVector`s
linear indices are `axes(A, 1)`, and therefore do not start at 1 for vectors with
unconventional indexing.

Calling this function is the "safe" way to write algorithms that
exploit linear indexing.

# Examples
```jldoctest
julia> A = fill(1, (5,6,7));

julia> b = LinearIndices(A);

julia> extrema(b)
(1, 210)
```

    LinearIndices(inds::CartesianIndices) -> R
    LinearIndices(sz::Dims) -> R
    LinearIndices((istart:istop, jstart:jstop, ...)) -> R

Return a `LinearIndices` array with the specified shape or [`axes`](@ref).

# Examples

The main purpose of this constructor is intuitive conversion
from cartesian to linear indexing:

```jldoctest
julia> linear = LinearIndices((1:3, 1:2))
3×2 LinearIndices{2, Tuple{UnitRange{Int64}, UnitRange{Int64}}}:
 1  4
 2  5
 3  6

julia> linear[1,2]
4
```
"""
struct LinearIndices{N,R<:NTuple{N,AbstractUnitRange{Int}}} <: AbstractArray{Int,N}
    indices::R
end
convert(::Type{LinearIndices{N,R}}, inds::LinearIndices{N}) where {N,R<:NTuple{N,AbstractUnitRange{Int}}} =
    LinearIndices{N,R}(convert(R, inds.indices))::LinearIndices{N,R}

LinearIndices(::Tuple{}) = LinearIndices{0,typeof(())}(())
LinearIndices(inds::NTuple{N,AbstractUnitRange{<:Integer}}) where {N} =
    LinearIndices(map(r->convert(AbstractUnitRange{Int}, r), inds))
LinearIndices(inds::NTuple{N,Union{<:Integer,AbstractUnitRange{<:Integer}}}) where {N} =
    LinearIndices(map(_convert2ind, inds))
LinearIndices(A::Union{AbstractArray,SimpleVector}) = LinearIndices(axes(A))

_convert2ind(i::Integer) = oneto(i)
_convert2ind(ind::AbstractUnitRange) = first(ind):last(ind)

function indices_promote_type(::Type{Tuple{R1,Vararg{R1,N}}}, ::Type{Tuple{R2,Vararg{R2,N}}}) where {R1,R2,N}
    R = promote_type(R1, R2)
    return Tuple{R, Vararg{R, N}}
end

promote_rule(::Type{LinearIndices{N,R1}}, ::Type{LinearIndices{N,R2}}) where {N,R1,R2} =
    LinearIndices{N,indices_promote_type(R1,R2)}
promote_rule(a::Type{Slice{T1}}, b::Type{Slice{T2}}) where {T1,T2} =
    el_same(promote_type(T1, T2), a, b)
promote_rule(a::Type{IdentityUnitRange{T1}}, b::Type{IdentityUnitRange{T2}}) where {T1,T2} =
    el_same(promote_type(T1, T2), a, b)

# AbstractArray implementation
IndexStyle(::Type{<:LinearIndices}) = IndexLinear()
axes(iter::LinearIndices) = map(axes1, iter.indices)
size(iter::LinearIndices) = map(length, iter.indices)
isassigned(iter::LinearIndices, i::Int) = checkbounds(Bool, iter, i)
function getindex(iter::LinearIndices, i::Int)
    @inline
    @boundscheck checkbounds(iter, i)
    i
end
function getindex(iter::LinearIndices, i::AbstractRange{<:Integer})
    @inline
    @boundscheck checkbounds(iter, i)
    @inbounds isa(iter, LinearIndices{1}) ? iter.indices[1][i] : (first(iter):last(iter))[i]
end
copy(iter::LinearIndices) = iter
# More efficient iteration — predominantly for non-vector LinearIndices
# but one-dimensional LinearIndices must be special-cased to support OffsetArrays
iterate(iter::LinearIndices{1}, s...) = iterate(axes1(iter.indices[1]), s...)
iterate(iter::LinearIndices, i=1) = i > length(iter) ? nothing : (i, i+1)

# Needed since firstindex and lastindex are defined in terms of LinearIndices
first(iter::LinearIndices) = 1
first(iter::LinearIndices{1}) = (@inline; first(axes1(iter.indices[1])))
last(iter::LinearIndices) = (@inline; length(iter))
last(iter::LinearIndices{1}) = (@inline; last(axes1(iter.indices[1])))

function show(io::IO, iter::LinearIndices)
    print(io, "LinearIndices(", iter.indices, ")")
end