test broadcasted integer division

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.62"
+version = "0.12.63"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/broadcast.jl

@@ -408,10 +408,9 @@ end
   storeop = add_simple_store!(ls, :destination, ArrayReference(:dest, loopsyms), elementbytes)
   doaddref!(ls, storeop)
   resize!(ls.loop_order, num_loops(ls)) # num_loops may be greater than N, eg Product
-  # return ls
   sc = setup_call(ls, :(Base.Broadcast.materialize!(dest, bc)), LineNumberNode(0), inline, false, u₁, u₂, v, threads%Int, warncheckarg)
-  # return sc
   Expr(:block, Expr(:meta,:inline), sc, :dest)
+  # setup_call_debug(ls)
 end
 @generated function vmaterialize!(
     dest′::Union{Adjoint{T,A},Transpose{T,A}}, bc::BC, ::Val{Mod}, ::Val{UNROLL}
@@ -434,31 +433,31 @@ end
 end
 # these are marked `@inline` so the `@turbo` itself can choose whether or not to inline.
 @generated function vmaterialize!(
-    dest::AbstractArray{T,N}, bc::Broadcasted{Base.Broadcast.DefaultArrayStyle{0},Nothing,typeof(identity),Tuple{T2}}, ::Val{Mod}, ::Val{UNROLL}
-) where {T <: NativeTypes, N, T2 <: Number, Mod, UNROLL}
+  dest::AbstractArray{T,N}, bc::Broadcasted{Base.Broadcast.DefaultArrayStyle{0},Nothing,typeof(identity),Tuple{T2}}, ::Val{Mod}, ::Val{UNROLL}
+  ) where {T <: NativeTypes, N, T2 <: Number, Mod, UNROLL}
   inline, u₁, u₂, v, isbroadcast, W, rs, rc, cls, l1, l2, l3, threads = UNROLL
-    quote
-        $(Expr(:meta,:inline))
-        arg = T(first(bc.args))
-        @turbo inline=$inline unroll=($u₁,$u₂) thread=$threads vectorize=$v for i ∈ eachindex(dest)
-            dest[i] = arg
-        end
-        dest
+  quote
+    $(Expr(:meta,:inline))
+    arg = T(first(bc.args))
+    @turbo inline=$inline unroll=($u₁,$u₂) thread=$threads vectorize=$v for i ∈ eachindex(dest)
+      dest[i] = arg
     end
+    dest
+  end
 end
 @generated function vmaterialize!(
-    dest′::Union{Adjoint{T,A},Transpose{T,A}}, bc::Broadcasted{Base.Broadcast.DefaultArrayStyle{0},Nothing,typeof(identity),Tuple{T2}}, ::Val{Mod}, ::Val{UNROLL}
-) where {T <: NativeTypes, N, A <: AbstractArray{T,N}, T2 <: Number, Mod, UNROLL}
-    inline, u₁, u₂, v, isbroadcast, W, rs, rc, cls, l1, l2, l3, threads = UNROLL
-    quote
-        $(Expr(:meta,:inline))
-        arg = T(first(bc.args))
-        dest = parent(dest′)
-        @turbo inline=$inline unroll=($u₁,$u₂) thread=$threads vectorize=$v for i ∈ eachindex(dest)
-            dest[i] = arg
-        end
-        dest′
+  dest′::Union{Adjoint{T,A},Transpose{T,A}}, bc::Broadcasted{Base.Broadcast.DefaultArrayStyle{0},Nothing,typeof(identity),Tuple{T2}}, ::Val{Mod}, ::Val{UNROLL}
+  ) where {T <: NativeTypes, N, A <: AbstractArray{T,N}, T2 <: Number, Mod, UNROLL}
+  inline, u₁, u₂, v, isbroadcast, W, rs, rc, cls, l1, l2, l3, threads = UNROLL
+  quote
+    $(Expr(:meta,:inline))
+    arg = T(first(bc.args))
+    dest = parent(dest′)
+    @turbo inline=$inline unroll=($u₁,$u₂) thread=$threads vectorize=$v for i ∈ eachindex(dest)
+      dest[i] = arg
     end
+    dest′
+  end
 end
 
 @inline function vmaterialize(

src/modeling/costs.jl

@@ -595,7 +595,8 @@ const FUNCTIONSYMBOLS = IdDict{Type{<:Function},Instruction}(
     typeof(Base.ifelse) => :ifelse,
     typeof(ifelse) => :ifelse,
     typeof(identity) => :identity,
-    typeof(conj) => :identity#conj
+    typeof(conj) => :identity,#conj,
+    typeof(÷) => :div_fast
     # typeof(zero) => :zero,
     # typeof(one) => :one,
     # typeof(axes) => :axes,

test/broadcast.jl

@@ -130,6 +130,16 @@ function test_broadcast(::Type{T}) where {T}
   @test D1 ≈ D2
   D2 = @outline @turbo @. C + A *ˡ (b + x');
   @test D1 ≈ D2
+  if T === Int64
+    xd = rand(-1_000_000_000_000:1_000_000_000_000,89);
+  elseif T === Int32
+    xd = rand(-Int32(10_000_000):Int32(10_000_000),89);
+  else
+    xd = rand(R,89);
+  end;
+  yd = rand(R,89);
+  yd[yd .== 0] .= 77;
+  @test xd .÷ yd == @turbo xd .÷ yd
 
   if T <: Union{Float32,Float64}
     D3 = cos.(B');