Use one canonical ELTYPE

Author: chriselrod

src/codegen/lowering.jl

@@ -426,7 +426,8 @@ end
 
 function outer_reduction_zero(op::Operation, u₁u::Bool, Umax::Int, reduct_class::Float64, rs::Expr)
   reduct_zero = reduction_zero(reduct_class)
-  Tsym = outer_reduct_init_typename(op)
+  # Tsym = outer_reduct_init_typename(op)
+  Tsym = ELTYPESYMBOL
   if isvectorized(op)
     if Umax == 1 || !u₁u
       if reduct_zero === :zero
@@ -614,7 +615,16 @@ function gc_preserve(ls::LoopSet, q::Expr)
     q2
     # Expr(:block, gcp)
 end
-
+function push_outer_reduct_types!(pt::Expr, ls::LoopSet, ortypdefined::Bool)
+  for j ∈ ls.outer_reductions
+    oreducop = ls.operations[j]
+    if ortypdefined
+      push!(pt.args, typeof_expr(oreducop))
+    else
+      push!(pt.args, outer_reduct_init_typename(oreducop))
+    end
+  end
+end
 function determine_eltype(ls::LoopSet, ortypdefined::Bool)::Union{Symbol,Expr}
   narrays = length(ls.includedactualarrays)
   noreduc = length(ls.outer_reductions)
@@ -637,14 +647,7 @@ function determine_eltype(ls::LoopSet, ortypdefined::Bool)::Union{Symbol,Expr}
   for array ∈ ls.includedactualarrays
     push!(pt.args, Expr(:call, lv(:eltype), array))
   end
-  for j ∈ ls.outer_reductions
-    oreducop = ls.operations[j]
-    if ortypdefined
-      push!(pt.args, typeof_expr(oreducop))
-    else
-      push!(pt.args, outer_reduct_init_typename(oreducop))
-    end
-  end
+  push_outer_reduct_types!(pt, ls, ortypdefined)
   return pt
 end
 @inline _eltype(x) = eltype(x)
@@ -653,20 +656,21 @@ function determine_width(
     ls::LoopSet, vectorized::Union{Symbol,Nothing}
 )
     vwidth_q = Expr(:call, lv(:pick_vector_width))
-    if !(vectorized === nothing)
+    if vectorized ≢ nothing
         vloop = getloop(ls, vectorized)
         if isstaticloop(vloop)
             push!(vwidth_q.args, Expr(:call, Expr(:curly, :Val, length(vloop))))
         end
     end
     # push!(vwidth_q.args, ls.T)
-    if length(ls.includedactualarrays) < 2
-        push!(vwidth_q.args, ELTYPESYMBOL)
-    else
-        for array ∈ ls.includedactualarrays
-            push!(vwidth_q.args, Expr(:call, lv(:_eltype), array))
-        end
-    end
+    # if length(ls.includedactualarrays) < 2
+    push!(vwidth_q.args, ELTYPESYMBOL)
+    # else
+    # for array ∈ ls.includedactualarrays
+    #     push!(vwidth_q.args, Expr(:call, lv(:_eltype), array))
+    # end
+    # push_outer_reduct_types!(vwidth_q, 
+    # end
     vwidth_q
 end
 function init_remblock(unrolledloop::Loop, lssm::LoopStartStopManager, n::Int)#u₁loop::Symbol = unrolledloop.itersymbol)

src/constructors.jl

@@ -217,6 +217,12 @@ macro tvectorize(args...)
     vectorize_macro(__module__, __source__, last(args), :(thread=true), Base.front(args)...)
 end
 
+function def_outer_reduct_types!(ls::LoopSet)
+  for or ∈ ls.outer_reductions
+    op = operations(ls)[or]
+    pushpreamble!(ls, Expr(:(=), outer_reduct_init_typename(op), typeof_expr(op)))
+  end
+end
 """
     @_vectorize
 
@@ -230,19 +236,17 @@ macro _vectorize(q)
   q = macroexpand(__module__, q)
   ls = LoopSet(q, __module__)
   set_hw!(ls)
-  for or ∈ ls.outer_reductions
-    op = operations(ls)[or]
-    pushpreamble!(ls, Expr(:(=), outer_reduct_init_typename(op), typeof_expr(op)))
-  end
+  def_outer_reduct_types!(ls)
   esc(Expr(:block, ls.prepreamble, lower_and_split_loops(ls, -1)))
 end
 macro _vectorize(arg, q)
-    @assert q.head === :for
-    q = macroexpand(__module__, q)
-    inline, check_empty, u₁, u₂ = check_macro_kwarg(arg, false, false, zero(Int8), zero(Int8), 1)
-    ls = LoopSet(q, __module__)
-    set_hw!(ls)
-    esc(Expr(:block, ls.prepreamble, lower(ls, u₁ % Int, u₂ % Int, -1)))
+  @assert q.head === :for
+  q = macroexpand(__module__, q)
+  inline, check_empty, u₁, u₂ = check_macro_kwarg(arg, false, false, zero(Int8), zero(Int8), 1)
+  ls = LoopSet(q, __module__)
+  set_hw!(ls)
+  def_outer_reduct_types!(ls)
+  esc(Expr(:block, ls.prepreamble, lower(ls, u₁ % Int, u₂ % Int, -1)))
 end
 
 macro vectorize_debug(q)

src/reconstruct_loopset.jl

@@ -607,8 +607,8 @@ function avx_loopset!(
     extractind = add_ops!(ls, instr, ops, mrefs, opsymbols, 1, nopsv, expandedv, elementbytes)
     extractind = process_metadata!(ls, AM, extractind)
     extractind = add_array_symbols!(ls, arraysymbolinds, extractind)
-    extract_external_functions!(ls, extractind, vargs)
-    extract_outerreduct_types!(ls, extractind, vargs)
+    extractind = extract_external_functions!(ls, extractind, vargs)
+    extractind = extract_outerreduct_types!(ls, extractind, vargs)
     ls
 end
 function avx_body(ls::LoopSet, UNROLL::Tuple{Bool,Int8,Int8,Bool,Int,Int,Int,Int,Int,Int,Int,UInt})