Update for new mask API in VectoriationBase.

Author: chriselrod

src/add_loads.jl

@@ -74,7 +74,8 @@ end
 struct LoopValue end
 @inline VectorizationBase.stridedpointer(::LoopValue) = LoopValue()
 @inline VectorizationBase.vload(::LoopValue, i::Tuple{_MM{W}}) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
-@inline VectorizationBase.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
+# @inline VectorizationBase.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
+@inline VectorizationBase.vload(::LoopValue, i::Tuple{_MM{W}}, ::Mask) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
 @inline VectorizationBase.vload(::LoopValue, i::Integer) = i + one(i)
 @inline VectorizationBase.vload(::LoopValue, i::Tuple{I}) where {I<:Integer} = @inbounds(i[1]) + one(I)
 @inline Base.eltype(::LoopValue) = Int8

src/lower_store.jl

@@ -70,10 +70,6 @@ function lower_conditionalstore_scalar!(
     end
     nothing
 end
-@inline combinemasks(a::Unsigned, b::Unsigned) = a & b
-@inline combinemasks(a::Unsigned, b::Bool) = b ? a : zero(a)
-@inline combinemasks(a::Bool, b::Unsigned) = a ? b : zero(b)
-@inline combinemasks(a::Bool, b::Bool) = a & b
 function lower_conditionalstore_vectorized!(
     q::Expr, op::Operation, vectorized::Symbol, W::Symbol, unrolled::Symbol, tiled::Symbol, U::Int,
     suffix::Union{Nothing,Int}, mask::Union{Nothing,Symbol,Unsigned}, isunrolled::Bool
@@ -105,7 +101,7 @@ function lower_conditionalstore_vectorized!(
         condvarname = varassignname(condvar, u, condunrolled)
         instrcall = Expr(:call, lv(:vstore!), ptr, name, mo)
         if mask !== nothing && (vecnotunrolled || u == U - 1)
-            push!(instrcall.args, Expr(:call, lv(:combinemasks), condvarname, mask))
+            push!(instrcall.args, Expr(:call, :&, condvarname, mask))
         else
             push!(instrcall.args, condvarname)
         end

src/lowering.jl

@@ -232,14 +232,14 @@ function determine_width(ls::LoopSet, vectorized::Symbol)
     if isstaticloop(vloop)
         push!(vwidth_q.args, Expr(:call, Expr(:curly, :Val, length(vloop))))
     end
-    push!(vwidth_q.args, ls.T)
-    # if length(ls.includedactualarrays) < 2
-    #     push!(vwidth_q.args, ls.T)
-    # else
-    #     for array ∈ ls.includedactualarrays
-    #         push!(vwidth_q.args, Expr(:call, :eltype, array))
-    #     end
-    # end
+    # push!(vwidth_q.args, ls.T)
+    if length(ls.includedactualarrays) < 2
+        push!(vwidth_q.args, ls.T)
+    else
+        for array ∈ ls.includedactualarrays
+            push!(vwidth_q.args, Expr(:call, :eltype, array))
+        end
+    end
     vwidth_q
 end
 function lower_unrolled!(
@@ -330,9 +330,9 @@ function lower_unrolled_dynamic!(
     end
     q
 end
-function maskexpr(W::Symbol, looplimit, allon::Bool)
+function maskexpr(W::Symbol, looplimit, all_on::Bool)
     rem = Expr(:call, lv(:valrem), W, looplimit)
-    Expr(:(=), Symbol("##mask##"), Expr(:call, lv(allon ? :masktable : :mask), W, rem))
+    Expr(:(=), Symbol("##mask##"), Expr(:call, lv(all_on ? :masktable : :mask), W, rem))
 end
 function definemask(loop::Loop, W::Symbol, allon::Bool)
     if isstaticloop(loop)