Added some (currently unused) infrastructure for tracking arrays as offsets.

Author: chriselrod

src/add_loads.jl

@@ -73,8 +73,10 @@ end
 
 struct LoopValue end
 @inline VectorizationBase.stridedpointer(::LoopValue) = LoopValue()
-@inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}) where {W} = SVec(SIMDPirates.vrangeincr(Val{W}(), @inbounds(i[1].i), Val{1}()))
-@inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = SVec(SIMDPirates.vrangeincr(Val{W}(), @inbounds(i[1].i), Val{1}()))
+@inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
+@inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
+# @inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}) where {W} = SVec(SIMDPirates.vrangeincr(Val{W}(), @inbounds(i[1].i), Val{1}()))
+# @inline SIMDPirates.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = SVec(SIMDPirates.vrangeincr(Val{W}(), @inbounds(i[1].i), Val{1}()))
 @inline VectorizationBase.load(::LoopValue, i::Integer) = i + one(i)
 @inline VectorizationBase.load(::LoopValue, i::Tuple{I}) where {I<:Integer} = @inbounds(i[1]) + one(I)
 @inline Base.eltype(::LoopValue) = Int8

src/condense_loopset.jl

@@ -7,6 +7,7 @@ Base.:(==)(u::Unsigned, it::IndexType) = (u % UInt8) == UInt8(it)
 struct ArrayRefStruct
     index_types::UInt64
     indices::UInt64
+    offsets::UInt64
 end
 
 function findindoradd!(v::Vector{T}, s::T) where {T}
@@ -18,12 +19,16 @@ end
 function ArrayRefStruct(ls::LoopSet, mref::ArrayReferenceMeta, arraysymbolinds::Vector{Symbol})
     index_types = zero(UInt64)
     indices = zero(UInt64)
+    offsets = zero(UInt64)
     indv = mref.ref.indices
+    offv = mref.ref.offsets
     # we can discard that the array was considered discontiguous, as it should be recovered from type information
     start = 1 + (first(indv) === Symbol("##DISCONTIGUOUSSUBARRAY##"))
     for (n,ind) ∈ enumerate(@view(indv[start:end]))
         index_types <<= 8
         indices <<= 8
+        offsets <<= 8
+        offsets |= offv[n]
         if mref.loopedindex[n]
             index_types |= LoopIndex
             indices |= getloopid(ls, ind)
@@ -38,7 +43,7 @@ function ArrayRefStruct(ls::LoopSet, mref::ArrayReferenceMeta, arraysymbolinds::
             end
         end
     end
-    ArrayRefStruct( index_types, indices )
+    ArrayRefStruct( index_types, indices, offsets )
 end
 
 struct OperationStruct

src/operations.jl

@@ -2,7 +2,9 @@
 struct ArrayReference
     array::Symbol
     indices::Vector{Symbol}
+    offsets::Vector{Int8}
 end
+ArrayReference(array, indices) = ArrayReference(array, indices, similar(indices, Int8))
 function Base.isequal(x::ArrayReference, y::ArrayReference)
     x.array === y.array || return false
     xinds = x.indices

src/reconstruct_loopset.jl

@@ -30,8 +30,10 @@ function ArrayReferenceMeta(
 )
     index_types = ar.index_types
     indices = ar.indices
+    offsets = ar.offsets
     ni = filled_8byte_chunks(index_types)
     index_vec = Vector{Symbol}(undef, ni)
+    offset_vec = Vector{Int8}(undef, ni)
     loopedindex = fill(false, ni)
     while index_types != zero(UInt64)
         ind = indices % UInt8
@@ -45,12 +47,14 @@ function ArrayReferenceMeta(
             arraysymbolinds[ind] 
         end
         index_vec[ni] = symind
+        offset_vec[ni] = offsets % Int8
         index_types >>>= 8
         indices >>>= 8
+        offsets >>>= 8
         ni -= 1
     end
     ArrayReferenceMeta(
-        ArrayReference(array, index_vec),
+        ArrayReference(array, index_vec, offset_vec),
         loopedindex, vp
     )
 end