Small number of bug fixes.

Author: chriselrod

src/add_compute.jl

@@ -204,7 +204,7 @@ function add_compute!(
                 if mpref == argref
                     if varname(mpref) === var
                         reduction_ind = ind
-                        add_load!(ls, argref, elementbytes)
+                        mergesetv!(deps, loopdependencies(add_load!(ls, argref, elementbytes)))
                     else
                         pushparent!(vparents, deps, reduceddeps, add_load!(ls, argref, elementbytes))
                     end

src/add_constants.jl

@@ -39,7 +39,11 @@ function add_constant!(ls::LoopSet, mpref::ArrayReferenceMetaPosition, elementby
     op = Operation(length(operations(ls)), varname(mpref), elementbytes, LOOPCONSTANT, constant, NODEPENDENCY, Symbol[], NOPARENTS, mpref.mref)
     add_vptr!(ls, op)
     temp = gensym(:intermediateconstref)
-    pushpreamble!(ls, Expr(:(=), temp, Expr(:call, lv(:vload), mpref.mref.ptr, mem_offset(op, UnrollArgs(0, Symbol(""), Symbol(""), nothing)))))
+    vloadcall = Expr(:call, lv(:vload), mpref.mref.ptr)
+    if length(getindices(op)) > 0
+        push!(vloadcall.args, mem_offset(op, UnrollArgs(0, Symbol(""), Symbol(""), nothing)))
+    end
+    pushpreamble!(ls, Expr(:(=), temp, vloadcall))
     pushpreamble!(ls, op, temp)
     pushop!(ls, op, temp)
 end

src/reconstruct_loopset.jl

@@ -414,7 +414,7 @@ function _avx_loopset(OPSsv, ARFsv, AMsv, LPSYMsv, LBsv, vargs)
     )
 end
 @generated function _avx_!(::Val{UT}, ::Type{OPS}, ::Type{ARF}, ::Type{AM}, ::Type{LPSYM}, lb::LB, vargs...) where {UT, OPS, ARF, AM, LPSYM, LB}
-    1 + 1 # Irrelevant line you can comment out/in to force recompilation...
+    # 1 + 1 # Irrelevant line you can comment out/in to force recompilation...
     ls = _avx_loopset(OPS.parameters, ARF.parameters, AM.parameters, LPSYM.parameters, LB.parameters, vargs)
     avx_body(ls, UT)
 end

src/split_loops.jl

@@ -70,8 +70,9 @@ function lower_and_split_loops(ls::LoopSet)
         remaining_ops[1:ind-1] .= @view(split_candidates[1:ind-1]); remaining_ops[ind:end] .= @view(split_candidates[ind+1:end])
         ls_2 = split_loopset(ls, remaining_ops)
         order_2, unrolled_2, tiled_2, vectorized_2, U_2, T_2, cost_2 = choose_order_cost(ls_2)
-        U_1 = T_1 = U_2 = T_2 = 2
+        # U_1 = T_1 = U_2 = T_2 = 2
         if cost_1 + cost_2 ≤ cost_fused
+            # @show cost_1, cost_2 cost_fused
             ls_2_lowered = if length(remaining_ops) > 1
                 lower_and_split_loops(ls_2)
             else