From 9e10015337204c0b216011cad583d002574428ca Mon Sep 17 00:00:00 2001
From: Matt Bauman <mbauman@gmail.com>
Date: Fri, 11 Oct 2019 11:10:43 -0500
Subject: [PATCH] Simplify, simdify, and speed up BitArray broadcasts with
 partitions

---
 base/broadcast.jl | 21 ++++++++++-----------
 1 file changed, 10 insertions(+), 11 deletions(-)

diff --git a/base/broadcast.jl b/base/broadcast.jl
index 16c0653be515d..ebce0afaba6c1 100644
--- a/base/broadcast.jl
+++ b/base/broadcast.jl
@@ -919,20 +919,19 @@ end
     length(dest) < 256 && return invoke(copyto!, Tuple{AbstractArray, Broadcasted{Nothing}}, dest, bc)
     tmp = Vector{Bool}(undef, bitcache_size)
     destc = dest.chunks
-    ind = cind = 1
+    cind = 1
     bc′ = preprocess(dest, bc)
-    @simd for I in eachindex(bc′)
-        @inbounds tmp[ind] = bc′[I]
-        ind += 1
-        if ind > bitcache_size
-            dumpbitcache(destc, cind, tmp)
-            cind += bitcache_chunks
-            ind = 1
+    for P in Iterators.partition(eachindex(bc′), bitcache_size)
+        ind = 1
+        @simd for I in P
+            @inbounds tmp[ind] = bc′[I]
+            ind += 1
+        end
+        @simd for i in ind:bitcache_size
+            @inbounds tmp[i] = false
         end
-    end
-    if ind > 1
-        @inbounds tmp[ind:bitcache_size] .= false
         dumpbitcache(destc, cind, tmp)
+        cind += bitcache_chunks
     end
     return dest
 end