Fix reduce_expr! call in lower_computer, and define static_step on CartesianIndices for Julia 1.5

Author: chriselrod

src/LoopVectorization.jl

@@ -36,7 +36,13 @@ using Base.FastMath: add_fast, sub_fast, mul_fast, div_fast, inv_fast, abs2_fast
 
 
 using ArrayInterface
-using ArrayInterface: OptionallyStaticUnitRange, OptionallyStaticRange, Zero, One, StaticBool, True, False, reduce_tup, indices, static_step
+using ArrayInterface: OptionallyStaticUnitRange, OptionallyStaticRange, Zero, One, StaticBool, True, False, reduce_tup, indices
+@static if VERSION ≥ v"1.6.0-rc1" #TODO: delete `else` when dropping 1.5 support
+    using ArrayInterface: static_step
+else # Julia 1.5 did not define `step` on CartesianIndices
+    @inline static_step(x) = ArrayInterface.static_step
+    static_step(x::CartesianIndices) = CartesianIndex(map(step, x.indices))
+end
 const Static = ArrayInterface.StaticInt
 
 using Requires
@@ -50,6 +56,7 @@ export LowDimArray, stridedpointer, indices,
 
 const VECTORWIDTHSYMBOL, ELTYPESYMBOL, MASKSYMBOL = Symbol("##Wvecwidth##"), Symbol("##Tloopeltype##"), Symbol("##mask##")
 
+
 include("vectorizationbase_compat/contract_pass.jl")
 include("vectorizationbase_compat/subsetview.jl")
 include("closeopen.jl")

src/codegen/lower_compute.jl

@@ -351,7 +351,7 @@ function lower_compute!(
             else
                 newpname = Symbol(newparentname, '_', u₁)
                 push!(q.args, Expr(:(=), newpname, Symbol(parentname, '_', u₁)))
-                reduce_expr!(q, newparentname, instruction(newparentop), u₁, -1, true)
+                reduce_expr!(q, newparentname, instruction(newparentop), u₁, -1, true, false)
                 push!(q.args, Expr(:(=), Symbol(newparentname, '_', 1), Symbol(newparentname, "##onevec##")))
             end
         end

src/reconstruct_loopset.jl

@@ -619,7 +619,7 @@ Execute an `@avx` block. The block's code is represented via the arguments:
 @generated function _avx_!(
     ::Val{UNROLL}, ::Val{OPS}, ::Val{ARF}, ::Val{AM}, ::Val{LPSYM}, var"#lv#tuple#args#"::Tuple{LB,V}
 ) where {UNROLL, OPS, ARF, AM, LPSYM, LB, V}
-    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, ARF, AM, LPSYM, LB.parameters, V.parameters, UNROLL)
     # return @show avx_body(ls, UNROLL)
     if last(UNROLL) > 1