Decouple sum/prod promotion from reduce

Author: TotalVerb

NEWS.md

@@ -95,6 +95,9 @@ Language changes
   * Prefix `&` for by-reference arguments to `ccall` has been deprecated in favor of
     `Ref` argument types ([#6080]).
 
+  * `reduce(+, [...])` and `reduce(*, [...])` no longer widen the iterated over arguments to
+    system word size. `sum` and `prod` still preserve this behavior. ([#22825])
+
 Breaking changes
 ----------------
 

base/process.jl

@@ -230,8 +230,6 @@ setenv(cmd::Cmd; dir="") = Cmd(cmd; dir=dir)
 (&)(left::AbstractCmd, right::AbstractCmd) = AndCmds(left, right)
 redir_out(src::AbstractCmd, dest::AbstractCmd) = OrCmds(src, dest)
 redir_err(src::AbstractCmd, dest::AbstractCmd) = ErrOrCmds(src, dest)
-Base.mr_empty(f, op::typeof(&), T::Type{<:Base.AbstractCmd}) =
-    throw(ArgumentError("reducing over an empty collection of type $T with operator & is not allowed"))
 
 # Stream Redirects
 redir_out(dest::Redirectable, src::AbstractCmd) = CmdRedirect(src, dest, STDIN_NO)

base/reduce.jl

@@ -5,45 +5,29 @@
 ###### Generic (map)reduce functions ######
 
 if Int === Int32
-const SmallSigned = Union{Int8,Int16}
-const SmallUnsigned = Union{UInt8,UInt16}
+    const SmallSigned = Union{Int8,Int16}
+    const SmallUnsigned = Union{UInt8,UInt16}
 else
-const SmallSigned = Union{Int8,Int16,Int32}
-const SmallUnsigned = Union{UInt8,UInt16,UInt32}
+    const SmallSigned = Union{Int8,Int16,Int32}
+    const SmallUnsigned = Union{UInt8,UInt16,UInt32}
 end
 
-const CommonReduceResult = Union{UInt64,UInt128,Int64,Int128,Float16,Float32,Float64}
-const WidenReduceResult = Union{SmallSigned, SmallUnsigned}
-
-promote_sys_size{T}(::Type{T}) = T
-promote_sys_size{T<:SmallSigned}(::Type{T}) = Int
-promote_sys_size{T<:SmallUnsigned}(::Type{T}) = UInt
-# r_promote_type: promote T to the type of reduce(op, ::Array{T})
-# (some "extra" methods are required here to avoid ambiguity warnings)
-r_promote_type(op, ::Type{T}) where {T} = T
-r_promote_type(op, ::Type{T}) where {T<:WidenReduceResult} = promote_sys_size(T)
-r_promote_type(::typeof(+), ::Type{T}) where {T<:WidenReduceResult} = promote_sys_size(T)
-r_promote_type(::typeof(*), ::Type{T}) where {T<:WidenReduceResult} = promote_sys_size(T)
-r_promote_type(::typeof(+), ::Type{T}) where {T<:Number} = typeof(zero(T)+zero(T))
-r_promote_type(::typeof(*), ::Type{T}) where {T<:Number} = typeof(one(T)*one(T))
-r_promote_type(::typeof(scalarmax), ::Type{T}) where {T<:WidenReduceResult} = T
-r_promote_type(::typeof(scalarmin), ::Type{T}) where {T<:WidenReduceResult} = T
-r_promote_type(::typeof(max), ::Type{T}) where {T<:WidenReduceResult} = T
-r_promote_type(::typeof(min), ::Type{T}) where {T<:WidenReduceResult} = T
-
-# r_promote: promote x to the type of reduce(op, [x])
-r_promote(op, x::T) where {T} = convert(r_promote_type(op, T), x)
+# Certain reductions like sum and prod may wish to promote the items being
+# reduced over to an appropriate size.
+promote_sys_size(x) = x
+promote_sys_size(x::SmallSigned) = Int(x)
+promote_sys_size(x::SmallUnsigned) = UInt(x)
 
 ## foldl && mapfoldl
 
 @noinline function mapfoldl_impl(f, op, v0, itr, i)
     # Unroll the while loop once; if v0 is known, the call to op may
     # be evaluated at compile time
     if done(itr, i)
-        return r_promote(op, v0)
+        return v0
     else
         (x, i) = next(itr, i)
-        v = op(r_promote(op, v0), f(x))
+        v = op(v0, f(x))
         while !done(itr, i)
             @inbounds (x, i) = next(itr, i)
             v = op(v, f(x))
@@ -69,7 +53,7 @@ this cannot be used with empty collections (see `reduce(op, itr)`).
 function mapfoldl(f, op, itr)
     i = start(itr)
     if done(itr, i)
-        return Base.mr_empty_iter(f, op, itr, iteratoreltype(itr))
+        return Base.mapreduce_empty_iter(f, op, itr, iteratoreltype(itr))
     end
     (x, i) = next(itr, i)
     v0 = f(x)
@@ -108,10 +92,10 @@ function mapfoldr_impl(f, op, v0, itr, i::Integer)
     # Unroll the while loop once; if v0 is known, the call to op may
     # be evaluated at compile time
     if isempty(itr) || i == 0
-        return r_promote(op, v0)
+        return v0
     else
         x = itr[i]
-        v  = op(f(x), r_promote(op, v0))
+        v  = op(f(x), v0)
         while i > 1
             x = itr[i -= 1]
             v = op(f(x), v)
@@ -137,7 +121,7 @@ this cannot be used with empty collections (see `reduce(op, itr)`).
 function mapfoldr(f, op, itr)
     i = endof(itr)
     if isempty(itr)
-        return Base.mr_empty_iter(f, op, itr, iteratoreltype(itr))
+        return Base.mapreduce_empty_iter(f, op, itr, iteratoreltype(itr))
     end
     return mapfoldr_impl(f, op, f(itr[i]), itr, i-1)
 end
@@ -180,12 +164,12 @@ foldr(op, itr) = mapfoldr(identity, op, itr)
 @noinline function mapreduce_impl(f, op, A::AbstractArray, ifirst::Integer, ilast::Integer, blksize::Int)
     if ifirst == ilast
         @inbounds a1 = A[ifirst]
-        return r_promote(op, f(a1))
+        return f(a1)
     elseif ifirst + blksize > ilast
         # sequential portion
         @inbounds a1 = A[ifirst]
         @inbounds a2 = A[ifirst+1]
-        v = op(r_promote(op, f(a1)), r_promote(op, f(a2)))
+        v = op(f(a1), f(a2))
         @simd for i = ifirst + 2 : ilast
             @inbounds ai = A[i]
             v = op(v, f(ai))
@@ -244,39 +228,49 @@ pairwise_blocksize(::typeof(abs2), ::typeof(+)) = 4096
 
 # handling empty arrays
 _empty_reduce_error() = throw(ArgumentError("reducing over an empty collection is not allowed"))
-mr_empty(f, op, T) = _empty_reduce_error()
-# use zero(T)::T to improve type information when zero(T) is not defined
-mr_empty(::typeof(identity), op::typeof(+), T) = r_promote(op, zero(T)::T)
-mr_empty(::typeof(abs), op::typeof(+), T) = r_promote(op, abs(zero(T)::T))
-mr_empty(::typeof(abs2), op::typeof(+), T) = r_promote(op, abs2(zero(T)::T))
-mr_empty(::typeof(identity), op::typeof(*), T) = r_promote(op, one(T)::T)
-mr_empty(::typeof(abs), op::typeof(scalarmax), T) = abs(zero(T)::T)
-mr_empty(::typeof(abs2), op::typeof(scalarmax), T) = abs2(zero(T)::T)
-mr_empty(::typeof(abs), op::typeof(max), T) = mr_empty(abs, scalarmax, T)
-mr_empty(::typeof(abs2), op::typeof(max), T) = mr_empty(abs2, scalarmax, T)
-mr_empty(f, op::typeof(&), T) = true
-mr_empty(f, op::typeof(|), T) = false
-
-mr_empty_iter(f, op, itr, ::HasEltype) = mr_empty(f, op, eltype(itr))
-mr_empty_iter(f, op::typeof(&), itr, ::EltypeUnknown) = true
-mr_empty_iter(f, op::typeof(|), itr, ::EltypeUnknown) = false
-mr_empty_iter(f, op, itr, ::EltypeUnknown) = _empty_reduce_error()
+reduce_empty(op, T) = _empty_reduce_error()
+reduce_empty(::typeof(+), T) = zero(T)
+reduce_empty(::typeof(*), T) = one(T)
+reduce_empty(::typeof(&), ::Type{Bool}) = true
+reduce_empty(::typeof(|), ::Type{Bool}) = false
+
+mapreduce_empty(f, op, T) = _empty_reduce_error()
+mapreduce_empty(::typeof(identity), op, T) = reduce_empty(op, T)
+mapreduce_empty(::typeof(promote_sys_size), op, T) =
+    promote_sys_size(mapreduce_empty(identity, op, T))
+mapreduce_empty(::typeof(abs), ::typeof(+), T) = abs(zero(T))
+mapreduce_empty(::typeof(abs2), ::typeof(+), T) = abs2(zero(T))
+mapreduce_empty(::typeof(abs), ::Union{typeof(scalarmax), typeof(max)}, T) =
+    abs(zero(T))
+mapreduce_empty(::typeof(abs2), ::Union{typeof(scalarmax), typeof(max)}, T) =
+    abs2(zero(T))
+
+# Allow mapreduce_empty to “see through” promote_sys_size
+let ComposedFunction = typename(typeof(identity ∘ identity)).wrapper
+    global mapreduce_empty(f::ComposedFunction{typeof(promote_sys_size)}, op, T) =
+        promote_sys_size(mapreduce_empty(f.g, op, T))
+end
+
+mapreduce_empty_iter(f, op, itr, ::HasEltype) = mapreduce_empty(f, op, eltype(itr))
+mapreduce_empty_iter(f, op::typeof(&), itr, ::EltypeUnknown) = true
+mapreduce_empty_iter(f, op::typeof(|), itr, ::EltypeUnknown) = false
+mapreduce_empty_iter(f, op, itr, ::EltypeUnknown) = _empty_reduce_error()
 
 _mapreduce(f, op, A::AbstractArray) = _mapreduce(f, op, IndexStyle(A), A)
 
 function _mapreduce(f, op, ::IndexLinear, A::AbstractArray{T}) where T
     inds = linearindices(A)
     n = length(inds)
     if n == 0
-        return mr_empty(f, op, T)
+        return mapreduce_empty(f, op, T)
     elseif n == 1
         @inbounds a1 = A[inds[1]]
-        return r_promote(op, f(a1))
+        return f(a1)
     elseif n < 16 # process short array here, avoid mapreduce_impl() compilation
         @inbounds i = inds[1]
         @inbounds a1 = A[i]
         @inbounds a2 = A[i+=1]
-        s = op(r_promote(op, f(a1)), r_promote(op, f(a2)))
+        s = op(f(a1), f(a2))
         while i < last(inds)
             @inbounds Ai = A[i+=1]
             s = op(s, f(Ai))
@@ -299,9 +293,6 @@ Reduce the given collection `itr` with the given binary operator `op`. `v0` must
 neutral element for `op` that will be returned for empty collections. It is unspecified
 whether `v0` is used for non-empty collections.
 
-The return type is `Int` (`UInt`) for (un)signed integers of less than system word size.
-For all other arguments, a common return type is found to which all arguments are promoted.
-
 Reductions for certain commonly-used operators may have special implementations, and
 should be used instead: `maximum(itr)`, `minimum(itr)`, `sum(itr)`, `prod(itr)`,
  `any(itr)`, `all(itr)`.
@@ -347,24 +338,47 @@ reduce(op, a::Number) = a
 
 Sum the results of calling function `f` on each element of `itr`.
 
+The return type is `Int` for signed integers of less than system word size, and
+`UInt` for unsigned integers of less than system word size.  For all other
+arguments, a common return type is found to which all arguments are promoted.
+
 ```jldoctest
 julia> sum(abs2, [2; 3; 4])
 29
 ```
+
+Note the important difference between `sum(A)` and `reduce(+, A)` for arrays
+with small integer eltype:
+
+```jldoctest
+julia> sum(Int8[100, 28])
+128
+
+julia> reduce(+, Int8[100, 28])
+-128
+```
+
+In the former case, the integers are widened to system word size and therefore
+the result is 128. In the latter case, no such widening happens and integer
+overflow results in -128.
 """
-sum(f::Callable, a) = mapreduce(f, +, a)
+sum(f::Callable, a) = mapreduce(promote_sys_size ∘ f, +, a)
 
 """
     sum(itr)
 
 Returns the sum of all elements in a collection.
 
+The return type is `Int` for signed integers of less than system word size, and
+`UInt` for unsigned integers of less than system word size.  For all other
+arguments, a common return type is found to which all arguments are promoted.
+
 ```jldoctest
 julia> sum(1:20)
 210
 ```
 """
-sum(a) = mapreduce(identity, +, a)
+sum(a) = mapreduce(promote_sys_size, +, a)
 sum(a::AbstractArray{Bool}) = count(a)
 
 
@@ -379,7 +393,7 @@ summation algorithm for additional accuracy.
 """
 function sum_kbn(A)
     T = _default_eltype(typeof(A))
-    c = r_promote(+, zero(T)::T)
+    c = promote_sys_size(zero(T)::T)
     i = start(A)
     if done(A, i)
         return c
@@ -405,24 +419,32 @@ end
 
 Returns the product of `f` applied to each element of `itr`.
 
+The return type is `Int` for signed integers of less than system word size, and
+`UInt` for unsigned integers of less than system word size.  For all other
+arguments, a common return type is found to which all arguments are promoted.
+
 ```jldoctest
 julia> prod(abs2, [2; 3; 4])
 576
 ```
 """
-prod(f::Callable, a) = mapreduce(f, *, a)
+prod(f::Callable, a) = mapreduce(promote_sys_size ∘ f, *, a)
 
 """
     prod(itr)
 
 Returns the product of all elements of a collection.
 
+The return type is `Int` for signed integers of less than system word size, and
+`UInt` for unsigned integers of less than system word size.  For all other
+arguments, a common return type is found to which all arguments are promoted.
+
 ```jldoctest
 julia> prod(1:20)
 2432902008176640000
 ```
 """
-prod(a) = mapreduce(identity, *, a)
+prod(a) = mapreduce(promote_sys_size, *, a)
 
 ## maximum & minimum
 

base/reducedim.jl

@@ -137,19 +137,22 @@ reducedim_init(f, op::typeof(|), A::AbstractArray, region) = reducedim_initarray
 
 # specialize to make initialization more efficient for common cases
 
-for (IT, RT) in ((CommonReduceResult, :(eltype(A))), (SmallSigned, :Int), (SmallUnsigned, :UInt))
-    T = Union{[AbstractArray{t} for t in uniontypes(IT)]..., [AbstractArray{Complex{t}} for t in uniontypes(IT)]...}
-    @eval begin
-        reducedim_init(f::typeof(identity), op::typeof(+), A::$T, region) =
-            reducedim_initarray(A, region, zero($RT))
-        reducedim_init(f::typeof(identity), op::typeof(*), A::$T, region) =
-            reducedim_initarray(A, region, one($RT))
-        reducedim_init(f::Union{typeof(abs),typeof(abs2)}, op::typeof(+), A::$T, region) =
-            reducedim_initarray(A, region, real(zero($RT)))
-        reducedim_init(f::Union{typeof(abs),typeof(abs2)}, op::typeof(*), A::$T, region) =
-            reducedim_initarray(A, region, real(one($RT)))
-    end
+let
+    BitIntFloat = Union{BitInteger, Math.IEEEFloat}
+    T = Union{
+        [AbstractArray{t} for t in uniontypes(BitIntFloat)]...,
+        [AbstractArray{Complex{t}} for t in uniontypes(BitIntFloat)]...}
+
+    global reducedim_init(f::typeof(identity), op::typeof(+), A::T, region) =
+        reducedim_initarray(A, region, zero(eltype(A)))
+    global reducedim_init(f::typeof(identity), op::typeof(*), A::T, region) =
+        reducedim_initarray(A, region, one(eltype(A)))
+    global reducedim_init(f::Union{typeof(abs),typeof(abs2)}, op::typeof(+), A::T, region) =
+        reducedim_initarray(A, region, real(zero(eltype(A))))
+    global reducedim_init(f::Union{typeof(abs),typeof(abs2)}, op::typeof(*), A::T, region) =
+        reducedim_initarray(A, region, real(one(eltype(A))))
 end
+
 reducedim_init(f::Union{typeof(identity),typeof(abs),typeof(abs2)}, op::typeof(+), A::AbstractArray{Bool}, region) =
     reducedim_initarray(A, region, 0)
 
@@ -610,14 +613,21 @@ any!(r, A)
 for (fname, op) in [(:sum, :+), (:prod, :*),
                     (:maximum, :scalarmax), (:minimum, :scalarmin),
                     (:all, :&), (:any, :|)]
+    function compose_promote_sys_size(x)
+        if fname in [:sum, :prod]
+            :(promote_sys_size ∘ $x)
+        else
+            x
+        end
+    end
     fname! = Symbol(fname, '!')
     @eval begin
         $(fname!)(f::Function, r::AbstractArray, A::AbstractArray; init::Bool=true) =
             mapreducedim!(f, $(op), initarray!(r, $(op), init, A), A)
         $(fname!)(r::AbstractArray, A::AbstractArray; init::Bool=true) = $(fname!)(identity, r, A; init=init)
 
         $(fname)(f::Function, A::AbstractArray, region) =
-            mapreducedim(f, $(op), A, region)
+            mapreducedim($(compose_promote_sys_size(:f)), $(op), A, region)
         $(fname)(A::AbstractArray, region) = $(fname)(identity, A, region)
     end
 end

base/sparse/sparsematrix.jl

@@ -1660,7 +1660,7 @@ function Base._mapreduce(f, op, ::Base.IndexCartesian, A::SparseMatrixCSC{T}) wh
     n = length(A)
     if z == 0
         if n == 0
-            Base.mr_empty(f, op, T)
+            Base.mapreduce_empty(f, op, T)
         else
             _mapreducezeros(f, op, T, n-z-1, f(zero(T)))
         end

base/tuple.jl

@@ -308,9 +308,13 @@ reverse(t::Tuple) = revargs(t...)
 
 # TODO: these definitions cannot yet be combined, since +(x...)
 # where x might be any tuple matches too many methods.
+# TODO: this is inconsistent with the regular sum in cases where the arguments
+# require size promotion to system size.
 sum(x::Tuple{Any, Vararg{Any}}) = +(x...)
 
 # NOTE: should remove, but often used on array sizes
+# TODO: this is inconsistent with the regular prod in cases where the arguments
+# require size promotion to system size.
 prod(x::Tuple{}) = 1
 prod(x::Tuple{Any, Vararg{Any}}) = *(x...)