Masks weren't being defined for short static length single loops.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <elrodc@gmail.com>"]
-version = "0.8.18"
+version = "0.8.19"
 
 [deps]
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"

src/lowering.jl

@@ -322,7 +322,7 @@ function lower_no_unroll(ls::LoopSet, us::UnrollSpecification, n::Int, inclmask:
 end
 function lower_unrolled_dynamic(ls::LoopSet, us::UnrollSpecification, n::Int, inclmask::Bool)
     UF = unrollfactor(us, n)
-    UF == 1 && return lower_no_unroll(ls, us, n, inclmask)
+    isone(UF) && return lower_no_unroll(ls, us, n, inclmask)
     @unpack u₁loopnum, vectorizedloopnum, u₁, u₂ = us
     order = names(ls)
     loopsym = order[n]
@@ -354,6 +354,7 @@ function lower_unrolled_dynamic(ls::LoopSet, us::UnrollSpecification, n::Int, in
             q = Expr(:while, tc, body)
         end
         remblock = Expr(:block)
+        (nisvectorized && (UFt > 0) && isone(num_loops(ls))) && push!(remblock.args, definemask(loop))
     else
         remblock = init_remblock(loop, ls.lssm[], n)#loopsym)
         q = Expr(:while, tc, body)

test/copy.jl

@@ -108,7 +108,13 @@ using LoopVectorization, OffsetArrays, Test
         end
         B
     end
-
+    function copy3!(B, A)
+        @assert (length(B) ≥ 3) && (length(A) ≥ 3)
+        @avx for i in 1:3
+            B[i] = A[i]
+        end
+        B
+    end
 
     for T ∈ (Float32, Float64, Int32, Int64)
         @show T, @__LINE__
@@ -179,5 +185,8 @@ using LoopVectorization, OffsetArrays, Test
 
         @test reversecopy1!(zeros(T, 10), collect(1:10)) == reversecopy1avx!(zeros(T, 10), collect(1:10))
         @test reversecopy2!(zeros(T, 10), OffsetArray(collect(1:10), -10:-1)) == reversecopy2avx!(zeros(T, 10), OffsetArray(collect(1:10), -10:-1))
+
+        x = rand(R, 3); y = similar(x);
+        @test copy3!(y, x) == x
     end
 end