Fix and test transposed offset array access.

Author: chriselrod

src/loopstartstopmanager.jl

@@ -65,8 +65,19 @@ function indices_calculated_by_pointer_offsets(ls::LoopSet, ar::ArrayReferenceMe
     out
 end
 
-@inline onetozeroindexgephack(sptr::AbstractStridedPointer) = gesp(sptr, (Static{-1}(),)) # go backwords 
+@generated function set_first_stride(sptr::StridedPointer{T,N,C,B,R}) where {T,N,C,B,R}
+    minrank = argmin(R)
+    newC = C > 0 ? (C == minrank ? 1 : 0) : C
+    newB = C > 0 ? (C == minrank ? B : 0) : B #TODO: confirm correctness
+    quote
+        $(Expr(:meta,:inline))
+        VectorizationBase.StridedPointer{$T,1,$newC,$newB,$(R[minrank],)}(pointer(sptr), (sptr.strd[$minrank],), (Zero(),))
+    end
+end
+set_first_stride(x) = x # cross fingers that this works
+@inline onetozeroindexgephack(sptr::AbstractStridedPointer) = gesp(set_first_stride(sptr), (Static{-1}(),)) # go backwords 
 @inline onetozeroindexgephack(sptr::AbstractStridedPointer{T,1}) where {T} = sptr
+# @inline onetozeroindexgephack(sptr::StridedPointer{T,1}) where {T} = sptr
 @inline onetozeroindexgephack(x) = x
 
 """

src/reconstruct_loopset.jl

@@ -133,10 +133,11 @@ function add_mref!(
     sp = ArrayInterface.rank_to_sortperm(R)
     # maybe no change needed? -- optimize common case
     column_major = ntuple(identity, N)
-    if sp === column_major
+    li = ar.loopedindex;
+    if sp === column_major || isone(length(li))
         return pushvarg!(ls, ar, i, name)
     end
-    li = ar.loopedindex; lic = copy(li);
+    lic = copy(li);
     inds = getindices(ar); indsc = copy(inds); 
     offsets = ar.ref.offsets; offsetsc = copy(offsets);
 

test/dot.jl

@@ -30,6 +30,7 @@ using Test
         end
         s
     end
+    
 
     selfdotq = :(for i ∈ eachindex(a)
                  s += a[i]*a[i]
@@ -51,6 +52,13 @@ using Test
         end
         s
     end
+    function myselfdotavx_v2(a)
+        s = zero(eltype(a))
+        @avx for i ∈ 1:length(a)
+            s += a[i]*a[i]
+        end
+        s
+    end
     function myselfdotavx_range(a)
         s = zero(eltype(a))
         rng = axes(a, 1)
@@ -249,10 +257,17 @@ using Test
         @test dot_unroll3avx_inline(a,b) ≈ s
         s = myselfdot(a)
         @test myselfdotavx(a) ≈ s
+        @test myselfdotavx_v2(a) ≈ s
         @test myselfdotavx_range(a) ≈ s
         @test myselfdot_avx(a) ≈ s
         @test myselfdotavx(a) ≈ s
 
+        A = OffsetArray(rand(37, 61), -5, 10);
+        s = myselfdot(A);
+        @test myselfdotavx(A) ≈ myselfdotavx(A') ≈ s
+        @test myselfdotavx_v2(A) ≈ myselfdotavx_v2(A') ≈ s
+        # @test myselfdot_avx(A) ≈ myselfdot_avx(A') ≈ s
+
         @test dot17(a,b) ≈ @view(a[1:17])' * @view(b[1:17])
         @test dot33(a,b) ≈ @view(a[1:33])' * @view(b[1:33])