Merge pull request #2 from N5N3/revert-1-patch__

Focus on copyto_unalias!, put SimdLoop in advance.

base/Base.jl

@@ -124,6 +124,11 @@ include("refpointer.jl")
 include("checked.jl")
 using .Checked
 
+# SIMD loops
+@pure sizeof(s::String) = Core.sizeof(s)  # needed by gensym as called from simdloop
+include("simdloop.jl")
+using .SimdLoop
+
 # array structures
 include("indices.jl")
 include("array.jl")
@@ -172,11 +177,6 @@ using .MultiplicativeInverses
 include("abstractarraymath.jl")
 include("arraymath.jl")
 
-# SIMD loops
-@pure sizeof(s::String) = Core.sizeof(s)  # needed by gensym as called from simdloop
-include("simdloop.jl")
-using .SimdLoop
-
 # map-reduce operators
 include("reduce.jl")
 

base/abstractarray.jl

@@ -1009,38 +1009,36 @@ end
 
 function copyto_unaliased!(deststyle::IndexStyle, dest::AbstractArray, srcstyle::IndexStyle, src::AbstractArray)
     isempty(src) && return dest
-    destinds, srcinds = LinearIndices(dest), LinearIndices(src)
-    idf, isf = first(destinds), first(srcinds)
-    Δi = idf - isf
-    (checkbounds(Bool, destinds, isf+Δi) & checkbounds(Bool, destinds, last(srcinds)+Δi)) ||
-        throw(BoundsError(dest, srcinds))
+    length(dest) >= length(src) || throw(BoundsError(dest, LinearIndices(src)))
     if deststyle isa IndexLinear
         if srcstyle isa IndexLinear
-            # Single-index implementation
-            @inbounds for i in srcinds
-                dest[i + Δi] = src[i]
+            Δi = firstindex(dest) - firstindex(src)
+            for i in eachindex(src)
+                @inbounds dest[i + Δi] = src[i]
             end
         else
-            # Dual-index implementation
-            i = idf - 1
-            @inbounds for a in src
-                dest[i+=1] = a
+            j = firstindex(dest) - 1
+            @inbounds @simd for I in eachindex(src)
+                dest[j+=1] = src[I]
             end
         end
     else
-        iterdest, itersrc = eachindex(dest), eachindex(src)
-        if iterdest == itersrc
-            # Shared-iterator implementation
-            for I in iterdest
-                @inbounds dest[I] = src[I]
+        if srcstyle isa IndexLinear
+            i = firstindex(src) - 1
+            @inbounds @simd for J in eachindex(dest)
+                dest[J] = src[i+=1]
             end
         else
-            # Dual-iterator implementation
-            ret = iterate(iterdest)
-            @inbounds for a in src
-                idx, state = ret
-                dest[idx] = a
-                ret = iterate(iterdest, state)
+            iterdest, itersrc = eachindex(dest), eachindex(src)
+            if iterdest == itersrc
+                # Shared-iterator implementation
+                @inbounds @simd for I in itersrc
+                    dest[I] = src[I]
+                end
+            else
+                for (I,J) in zip(itersrc, iterdest)
+                    @inbounds dest[J] = src[I]
+                end
             end
         end
     end

base/broadcast.jl

@@ -991,17 +991,11 @@ preprocess_args(dest, args::Tuple{}) = ()
 # Specialize this method if all you want to do is specialize on typeof(dest)
 @inline function copyto!(dest::AbstractArray, bc::Broadcasted{Nothing})
     axes(dest) == axes(bc) || throwdm(axes(dest), axes(bc))
-    # Performance optimization: broadcast!(identity, dest, A) is equivalent to copyto!(dest, A) if indices match.
-    # However copyto!(dest, A) is very slow in many cases, implement a faster version here.
+    # Performance optimization: broadcast!(identity, dest, A) is equivalent to copyto!(dest, A) if indices match
     if bc.f === identity && bc.args isa Tuple{AbstractArray} # only a single input argument to broadcast!
         A = bc.args[1]
         if axes(dest) == axes(A)
-            A′ = broadcast_unalias(dest, A)
-            iter = IndexStyle(dest) isa IndexCartesian ? dest : A′
-            @inbounds @simd for I in eachindex(iter)
-                dest[I] = A′[I]
-            end
-            return dest
+            return copyto!(dest, A)
         end
     end
     bc′ = preprocess(dest, bc)