diff --git a/src/nditeration.jl b/src/nditeration.jl
index 32d52e26..b933c2be 100644
--- a/src/nditeration.jl
+++ b/src/nditeration.jl
@@ -36,12 +36,12 @@ end
 
 _ind2sub_recuse(::Tuple{}, ind) = (ind+1,)
 function _ind2sub_recurse(indslast::NTuple{1}, ind)
-    @inline
+    Base.@_inline_meta
     (_lookup(ind, indslast[1]),)
 end
 
 function _ind2sub_recurse(inds, ind)
-    @inline
+    Base.@_inline_meta
     inv = inds[1]
     indnext, f, l = _div(ind, inv)
     (ind-l*indnext+f, _ind2sub_recurse(Base.tail(inds), indnext)...)
@@ -54,7 +54,7 @@ function _div(ind, inv::SignedMultiplicativeInverse)
 end
 
 function Base._ind2sub(inv::FastCartesianIndices, ind)
-    @inline
+    Base.@_inline_meta
     _ind2sub_recurse(inv.inverses, ind-1)
 end
 
@@ -151,24 +151,7 @@ Base.length(range::NDRange) = length(blocks(range))
 end
 
 Base.@propagate_inbounds function expand(ndrange::NDRange{N}, groupidx::Integer, idx::Integer) where {N}
-    # This causes two sdiv operations, one for each Linear to CartesianIndex
     return expand(ndrange, blocks(ndrange)[groupidx], workitems(ndrange)[idx])
-
-    # The formulation below saves one sdiv
-    # but leads to a different index order...
-    # previous: julia> expand(ndrange, 1, 32*32)
-    #                  CartesianIndex(32, 32)
-    # now:      julia> expand(ndrange, 1, 32*32)
-    #                  CartesianIndex(1024, 1)
-    # B = blocks(ndrange)::CartesianIndices
-    # W = workitems(ndrange)::CartesianIndices
-    # Ind = ntuple(Val(N)) do I
-    #     Base.@_inline_meta
-    #     b = B.indices[I]
-    #     w = W.indices[I]
-    #     length(b) * length(w)
-    # end
-    # CartesianIndices(Ind)[(groupidx-1)* prod(size(W)) + idx]
 end
 
 Base.@propagate_inbounds function expand(ndrange::NDRange{N}, groupidx::CartesianIndex{N}, idx::Integer) where {N}