Add vadd1 and simplify maybeaddref! for loop gradients example.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.8.21"
+version = "0.8.22"
 
 [deps]
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"

src/LoopVectorization.jl

@@ -16,7 +16,7 @@ using SIMDPirates: VECTOR_SYMBOLS, evadd, evsub, evmul, evfdiv, vrange,
     reduced_add, reduced_prod, reduce_to_add, reduced_max, reduced_min, vsum, vprod, vmaximum, vminimum,
     sizeequivalentfloat, sizeequivalentint, vadd!, vsub!, vmul!, vfdiv!, vfmadd!, vfnmadd!, vfmsub!, vfnmsub!,
     vfmadd231, vfmsub231, vfnmadd231, vfnmsub231, sizeequivalentfloat, sizeequivalentint, #prefetch,
-    vmullog2, vmullog10, vdivlog2, vdivlog10, vmullog2add!, vmullog10add!, vdivlog2add!, vdivlog10add!, vfmaddaddone
+    vmullog2, vmullog10, vdivlog2, vdivlog10, vmullog2add!, vmullog10add!, vdivlog2add!, vdivlog10add!, vfmaddaddone, vadd1
 using SLEEFPirates: pow
 using Base.Broadcast: Broadcasted, DefaultArrayStyle
 using LinearAlgebra: Adjoint, Transpose

src/add_loads.jl

@@ -1,4 +1,5 @@
-function maybeaddref!(ls::LoopSet, op, ref)
+function maybeaddref!(ls::LoopSet, op)
+    ref = op.ref
     id = findfirst(r -> r == ref, ls.refs_aliasing_syms)
     # try to CSE
     if isnothing(id)
@@ -12,8 +13,7 @@ end
 
 function add_load!(ls::LoopSet, op::Operation, actualarray::Bool = true, broadcast::Bool = false)
     @assert isload(op)
-    ref = op.ref
-    if (id = maybeaddref!(ls, op, ref)) > 0 # try to CSE
+    if (id = maybeaddref!(ls, op)) > 0 # try to CSE
         opp = ls.opdict[ls.syms_aliasing_refs[id]] # throw an error if not found.
         return isstore(opp) ? getop(ls, first(parents(opp))) : opp
     end

src/costs.jl

@@ -130,6 +130,7 @@ const COST = Dict{Symbol,InstructionCost}(
     :(*) => InstructionCost(4,0.5),
     :(/) => InstructionCost(13,4.0,-2.0),
     :vadd => InstructionCost(4,0.5),
+    :vadd1 => InstructionCost(4,0.5),
     :add_fast => InstructionCost(4,0.5),
     :vsub => InstructionCost(4,0.5),
     :sub_fast => InstructionCost(4,0.5),