Support multiple return values into variables and stores, resolve #41.

Author: chriselrod

src/LoopVectorization.jl

@@ -45,6 +45,7 @@ If you want good performance, DO NOT use a 32-bit build of Julia if you don't ha
 """
 const REGISTER_COUNT = Sys.ARCH === :i686 ? 8 : VectorizationBase.REGISTER_COUNT
 
+include("getconstindexes.jl")
 include("vectorizationbase_extensions.jl")
 include("predicates.jl")
 include("map.jl")

src/costs.jl

@@ -233,6 +233,15 @@ const COST = Dict{Symbol,InstructionCost}(
     :adjoint => InstructionCost(0,0.0,0.0,0),
     :conj => InstructionCost(0,0.0,0.0,0),
     :transpose => InstructionCost(0,0.0,0.0,0),
+    :first => InstructionCost(0,0.0,0.0,0),
+    :second => InstructionCost(0,0.0,0.0,0),
+    :third => InstructionCost(0,0.0,0.0,0),
+    :fourth => InstructionCost(0,0.0,0.0,0),
+    :fifth => InstructionCost(0,0.0,0.0,0),
+    :sixth => InstructionCost(0,0.0,0.0,0),
+    :seventh => InstructionCost(0,0.0,0.0,0),
+    :eigth => InstructionCost(0,0.0,0.0,0),
+    :ninth => InstructionCost(0,0.0,0.0,0),
     :prefetch => InstructionCost(0,0.0,0.0,0),
     :prefetch0 => InstructionCost(0,0.0,0.0,0),
     :prefetch1 => InstructionCost(0,0.0,0.0,0),

src/graphs.jl

@@ -669,16 +669,39 @@ function Base.push!(ls::LoopSet, ex::Expr, elementbytes::Int, position::Int)
                 add_constant!(ls, RHS, ls.loopsymbols[1:position], LHS, elementbytes)
             end
         elseif LHS isa Expr
-            @assert LHS.head === :ref
-            if RHS isa Symbol
-                add_store_ref!(ls, RHS, LHS, elementbytes)
-            elseif RHS isa Expr
-                # need to check if LHS appears in RHS
-                # assign RHS to lrhs
-                array, rawindices = ref_from_expr!(ls, LHS)
-                prepare_rhs_for_storage!(ls, RHS, array, rawindices, elementbytes, position)
+            if LHS.head === :ref
+                if RHS isa Symbol
+                    add_store_ref!(ls, RHS, LHS, elementbytes)
+                elseif RHS isa Expr
+                    # need to check if LHS appears in RHS
+                    # assign RHS to lrhs
+                    array, rawindices = ref_from_expr!(ls, LHS)
+                    prepare_rhs_for_storage!(ls, RHS, array, rawindices, elementbytes, position)
+                else
+                    add_store_ref!(ls, RHS, LHS, elementbytes)
+                end
+            elseif LHS.head === :tuple
+                @assert length(LHS.args) ≤ 9 "Functions returning more than 9 values aren't currently supported."
+                lhstemp = gensym(:lhstuple)
+                vparents = Operation[maybe_const_compute!(ls, add_operation!(ls, lhstemp, RHS, elementbytes, position), elementbytes, position)]
+                for i ∈ eachindex(LHS.args)
+                    f = (:first,:second,:third,:fourth,:fifth,:sixth,:seventh,:eigth,:ninth)[i]
+                    lhsi = LHS.args[i]
+                    if lhsi isa Symbol
+                        add_compute!(ls, lhsi, f, vparents, elementbytes)
+                    elseif lhsi isa Expr && lhsi.head === :ref
+                        tempunpacksym = gensym(:tempunpack)
+                        add_compute!(ls, tempunpacksym, f, vparents, elementbytes)
+                        add_store_ref!(ls, tempunpacksym, lhsi, elementbytes)
+                    else
+                        println(lhsi)
+                        throw("Unpacking the above expression in the left hand side was not understood/supported.")
+                    end
+                end
+                first(vparents)
             else
-                add_store_ref!(ls, RHS, LHS, elementbytes)
+                println(LHS)
+                throw("LHS not understood; only `:ref`s and `:tuple`s are currently supported.")
             end
         else
             println(LHS)

test/special.jl

@@ -303,6 +303,37 @@
         end
         ent
     end
+
+    function vsincos(x)
+        y = Matrix{eltype(x)}(undef, length(x), 2)
+        for i = eachindex(x)
+            y[i,1], y[i,2] = sincos(x[i])
+        end
+        return y
+    end
+    function vsincosavx(x)
+        y = Matrix{eltype(x)}(undef, length(x), 2)
+        @avx for i = eachindex(x)
+            y[i,1], y[i,2] = sincos(x[i])
+        end
+        return y
+    end
+    function sincosdot(x)
+        a = zero(eltype(x))
+        for i ∈ eachindex(x)
+            s, c = sincos(x[i])
+            a += s * c
+        end
+        a
+    end
+    function sincosdotavx(x)
+        a = zero(eltype(x))
+        @avx for i ∈ eachindex(x)
+            s, c = sincos(x[i])
+            a += s * c
+        end
+        a
+    end
 
     for T ∈ (Float32, Float64)
         @show T, @__LINE__
@@ -320,6 +351,9 @@
         @test s ≈ myvexp_avx(a)
         @test b1 ≈ @avx exp.(a)
 
+        @test vsincos(a) ≈ vsincosavx(a)
+        @test sincosdot(a) ≈ sincosdotavx(a)
+        
         A = rand(T, 73, 73);
         ld = logdet(UpperTriangular(A))
         @test ld ≈ trianglelogdetavx(A)