Fix #24914.

Reimplement and generalize all-scalar optimization.

Add documentation.

Explicitly return dest in various broadcast!-related methods. This is to make things easier on inference. Found by @timholy.

Collapse spbroadcast_args! into broadcast! as suggested by @Sacha0.

Also add nothing argument to broadcast! calls in sparse broadcast as suggested by @timholy.

base/broadcast.jl

@@ -239,12 +239,6 @@ broadcast_indices
 # special cases defined for performance
 broadcast(f, x::Number...) = f(x...)
 @inline broadcast(f, t::NTuple{N,Any}, ts::Vararg{NTuple{N,Any}}) where {N} = map(f, t, ts...)
-@inline broadcast!(::typeof(identity), x::AbstractArray{T,N}, y::AbstractArray{S,N}) where {T,S,N} =
-    Base.axes(x) == Base.axes(y) ? copyto!(x, y) : _broadcast!(identity, x, y)
-
-# special cases for "X .= ..." (broadcast!) assignments
-broadcast!(::typeof(identity), X::AbstractArray, x::Number) = fill!(X, x)
-broadcast!(f, X::AbstractArray, x::Number...) = (@inbounds for I in eachindex(X); X[I] = f(x...); end; X)
 
 ## logic for deciding the BroadcastStyle
 # Dimensionality: computing max(M,N) in the type domain so we preserve inferrability
@@ -261,7 +255,7 @@ longest(::Tuple{}, ::Tuple{}) = ()
 # combine_styles operates on values (arbitrarily many)
 combine_styles(c) = result_style(BroadcastStyle(typeof(c)))
 combine_styles(c1, c2) = result_style(combine_styles(c1), combine_styles(c2))
-combine_styles(c1, c2, cs...) = result_style(combine_styles(c1), combine_styles(c2, cs...))
+@inline combine_styles(c1, c2, cs...) = result_style(combine_styles(c1), combine_styles(c2, cs...))
 
 # result_style works on types (singletons and pairs), and leverages `BroadcastStyle`
 result_style(s::BroadcastStyle) = s
@@ -397,6 +391,7 @@ Base.@propagate_inbounds _broadcast_getindex(::Style{Tuple}, A::Tuple{Any}, I) =
             result = @ncall $nargs f val
             @inbounds B[I] = result
         end
+        return B
     end
 end
 
@@ -433,6 +428,7 @@ end
             @inbounds C[ind:bitcache_size] = false
             dumpbitcache(Bc, cind, C)
         end
+        return B
     end
 end
 
@@ -445,11 +441,38 @@ Note that `dest` is only used to store the result, and does not supply
 arguments to `f` unless it is also listed in the `As`,
 as in `broadcast!(f, A, A, B)` to perform `A[:] = broadcast(f, A, B)`.
 """
-@inline broadcast!(f, C::AbstractArray, A, Bs::Vararg{Any,N}) where {N} =
-    _broadcast!(f, C, A, Bs...)
+@inline broadcast!(f::Tf, dest, As::Vararg{Any,N}) where {Tf,N} = broadcast!(f, dest, combine_styles(As...), As...)
+@inline broadcast!(f::Tf, dest, ::BroadcastStyle, As::Vararg{Any,N}) where {Tf,N} = broadcast!(f, dest, nothing, As...)
+
+# Default behavior (separated out so that it can be called by users who want to extend broadcast!).
+@inline function broadcast!(f, dest, ::Nothing, As::Vararg{Any, N}) where N
+    if f isa typeof(identity) && N == 1
+        A = As[1]
+        if A isa AbstractArray && Base.axes(dest) == Base.axes(A)
+            return copyto!(dest, A)
+        end
+    end
+    _broadcast!(f, dest, As...)
+    return dest
+end
+
+# Optimization for the all-Scalar case.
+@inline function broadcast!(f, dest, ::Scalar, As::Vararg{Any, N}) where N
+    if dest isa AbstractArray
+        if f isa typeof(identity) && N == 1
+            return fill!(dest, As[1])
+        else
+            @inbounds for I in eachindex(dest)
+                dest[I] = f(As...)
+            end
+            return dest
+        end
+    end
+    _broadcast!(f, dest, As...)
+    return dest
+end
 
-# This indirection allows size-dependent implementations (e.g., see the copying `identity`
-# specialization above)
+# This indirection allows size-dependent implementations.
 @inline function _broadcast!(f, C, A, Bs::Vararg{Any,N}) where N
     shape = broadcast_indices(C)
     @boundscheck check_broadcast_indices(shape, A, Bs...)
@@ -630,7 +653,7 @@ function broadcast_nonleaf(f, s::NonleafHandlingTypes, ::Type{ElType}, shape::In
         dest = Base.similar(Array{typeof(val)}, shape)
     end
     dest[I] = val
-    return _broadcast!(f, dest, keeps, Idefaults, As, Val(nargs), iter, st, 1)
+    _broadcast!(f, dest, keeps, Idefaults, As, Val(nargs), iter, st, 1)
 end
 
 broadcast(f, ::Union{Scalar,Unknown}, ::Nothing, ::Nothing, a...) = f(a...)

base/sparse/higherorderfns.jl

@@ -93,7 +93,8 @@ end
 # (3) broadcast[!] entry points
 broadcast(f::Tf, A::SparseVector) where {Tf} = _noshapecheck_map(f, A)
 broadcast(f::Tf, A::SparseMatrixCSC) where {Tf} = _noshapecheck_map(f, A)
-function broadcast!(f::Tf, C::SparseVecOrMat) where Tf
+
+@inline function broadcast!(f::Tf, C::SparseVecOrMat, ::Nothing) where Tf
     isempty(C) && return _finishempty!(C)
     fofnoargs = f()
     if _iszero(fofnoargs) # f() is zero, so empty C
@@ -106,14 +107,7 @@ function broadcast!(f::Tf, C::SparseVecOrMat) where Tf
     end
     return C
 end
-function broadcast!(f::Tf, C::SparseVecOrMat, A::SparseVecOrMat, Bs::Vararg{SparseVecOrMat,N}) where {Tf,N}
-    _aresameshape(C, A, Bs...) && return _noshapecheck_map!(f, C, A, Bs...)
-    Base.Broadcast.check_broadcast_indices(axes(C), A, Bs...)
-    fofzeros = f(_zeros_eltypes(A, Bs...)...)
-    fpreszeros = _iszero(fofzeros)
-    return fpreszeros ? _broadcast_zeropres!(f, C, A, Bs...) :
-                        _broadcast_notzeropres!(f, fofzeros, C, A, Bs...)
-end
+
 # the following three similar defs are necessary for type stability in the mixed vector/matrix case
 broadcast(f::Tf, A::SparseVector, Bs::Vararg{SparseVector,N}) where {Tf,N} =
     _aresameshape(A, Bs...) ? _noshapecheck_map(f, A, Bs...) : _diffshape_broadcast(f, A, Bs...)
@@ -1006,28 +1000,30 @@ Broadcast.BroadcastStyle(::SparseMatStyle, ::Broadcast.DefaultArrayStyle{N}) whe
 broadcast(f, ::PromoteToSparse, ::Nothing, ::Nothing, As::Vararg{Any,N}) where {N} =
     broadcast(f, map(_sparsifystructured, As)...)
 
-# ambiguity resolution
-broadcast!(::typeof(identity), dest::SparseVecOrMat, x::Number) =
-    fill!(dest, x)
-broadcast!(f, dest::SparseVecOrMat, x::Number...) =
-    spbroadcast_args!(f, dest, SPVM, x...)
-
 # For broadcast! with ::Any inputs, we need a layer of indirection to determine whether
 # the inputs can be promoted to SparseVecOrMat. If it's just SparseVecOrMat and scalars,
 # we can handle it here, otherwise see below for the promotion machinery.
-broadcast!(f, dest::SparseVecOrMat, mixedsrcargs::Vararg{Any,N}) where N =
-    spbroadcast_args!(f, dest, Broadcast.combine_styles(mixedsrcargs...), mixedsrcargs...)
-function spbroadcast_args!(f, dest, ::Type{SPVM}, mixedsrcargs::Vararg{Any,N}) where N
+function broadcast!(f::Tf, dest::SparseVecOrMat, ::SPVM, A::SparseVecOrMat, Bs::Vararg{SparseVecOrMat,N}) where {Tf,N}
+    if f isa typeof(identity) && N == 0 && Base.axes(dest) == Base.axes(A)
+        return copyto!(dest, A)
+    end
+    _aresameshape(dest, A, Bs...) && return _noshapecheck_map!(f, dest, A, Bs...)
+    Base.Broadcast.check_broadcast_indices(axes(dest), A, Bs...)
+    fofzeros = f(_zeros_eltypes(A, Bs...)...)
+    fpreszeros = _iszero(fofzeros)
+    fpreszeros ? _broadcast_zeropres!(f, dest, A, Bs...) :
+                        _broadcast_notzeropres!(f, fofzeros, dest, A, Bs...)
+    return dest
+end
+function broadcast!(f::Tf, dest::SparseVecOrMat, ::SPVM, mixedsrcargs::Vararg{Any,N}) where {Tf,N}
     # mixedsrcargs contains nothing but SparseVecOrMat and scalars
     parevalf, passedsrcargstup = capturescalars(f, mixedsrcargs)
-    return broadcast!(parevalf, dest, passedsrcargstup...)
-end
-function spbroadcast_args!(f, dest, ::PromoteToSparse, mixedsrcargs::Vararg{Any,N}) where N
-    broadcast!(f, dest, map(_sparsifystructured, mixedsrcargs)...)
+    broadcast!(parevalf, dest, nothing, passedsrcargstup...)
+    return dest
 end
-function spbroadcast_args!(f, dest, ::Any, mixedsrcargs::Vararg{Any,N}) where N
-    # Fallback. From a performance perspective would it be best to densify?
-    Broadcast._broadcast!(f, dest, mixedsrcargs...)
+function broadcast!(f::Tf, dest::SparseVecOrMat, ::PromoteToSparse, mixedsrcargs::Vararg{Any,N}) where {Tf,N}
+    broadcast!(f, dest, nothing, map(_sparsifystructured, mixedsrcargs)...)
+    return dest
 end
 
 _sparsifystructured(M::AbstractMatrix) = SparseMatrixCSC(M)

doc/src/manual/interfaces.md

@@ -404,6 +404,8 @@ perhaps range-types `Ind` of your own design. For more information, see [Arrays
 | `broadcast(f, As...)` | Complete bypass of broadcasting machinery |
 | `broadcast(f, ::DestStyle, ::Nothing, ::Nothing, As...)` | Bypass after container type is computed |
 | `broadcast(f, ::DestStyle, ::Type{ElType}, inds::Tuple, As...)` | Bypass after container type, eltype, and indices are computed |
+| `broadcast!(f, dest::DestType, ::Nothing, As...)` | Bypass in-place broadcast, specialization on destination type |
+| `broadcast!(f, dest, ::BroadcastStyle, As...)` | Bypass in-place broadcast, specialization on `BroadcastStyle` |
 
 [Broadcasting](@ref) is triggered by an explicit call to `broadcast` or `broadcast!`, or implicitly by
 "dot" operations like `A .+ b`. Any `AbstractArray` type supports broadcasting,
@@ -591,3 +593,37 @@ yields another `SparseVecStyle`, that its combination with a 2-dimensional array
 yields a `SparseMatStyle`, and anything of higher dimensionality falls back to the dense arbitrary-dimensional framework.
 These rules allow broadcasting to keep the sparse representation for operations that result
 in one or two dimensional outputs, but produce an `Array` for any other dimensionality.
+
+### [Extending `broadcast!`](@id extending-in-place-broadcast)
+
+Extending `broadcast!` (in-place broadcast) should be done with care, as it is easy to introduce
+ambiguities between packages. To avoid these ambiguities, we adhere to the following conventions.
+
+First, if you want to specialize on the destination type, say `DestType`, then you should
+define a method with the following signature:
+
+```julia
+broadcast!(f, dest::DestType, ::Nothing, As...)
+```
+
+Note that no bounds should be placed on the types of `f` and `As...`.
+
+Second, if specialized `broadcast!` behavior is desired depending on the input types,
+you should write [binary broadcasting rules](@ref writing-binary-broadcasting-rules) to
+determine a custom `BroadcastStyle` given the input types, say `MyBroadcastStyle`, and you should define a method with the following
+signature:
+
+```julia
+broadcast!(f, dest, ::MyBroadcastStyle, As...)
+```
+
+Note the lack of bounds on `f`, `dest`, and `As...`.
+
+Third, simultaneously specializing on both the type of `dest` and the `BroadcastStyle` is fine. In this case,
+it is also allowed to specialize on the types of the source arguments (`As...`). For example, these method signatures are OK:
+
+```julia
+broadcast!(f, dest::DestType, ::MyBroadcastStyle, As...)
+broadcast!(f, dest::DestType, ::MyBroadcastStyle, As::AbstractArray...)
+broadcast!(f, dest::DestType, ::Broadcast.Scalar, As::Number...)
+```