Add more if/else tests, and a bugfix for when a conditional appears as an argument to another function..

Author: chriselrod

src/add_ifelse.jl

@@ -21,13 +21,13 @@ end
 
 function add_andblock!(ls::LoopSet, condop::Operation, LHS, rhsop::Operation, elementbytes::Int, position::Int)
     if LHS isa Symbol
-        altop = getop(ls, LHS)
+        altop = getop(ls, LHS, elementbytes)
         return add_compute!(ls, LHS, :vifelse, [condop, rhsop, altop], elementbytes)
     elseif LHS isa Expr && LHS.head === :ref
         return add_conditional_store!(ls, LHS, condop, rhsop, elementbytes)
     else
         throw("Don't know how to assign onto $LHS.")
-    end        
+    end
 end
 function add_andblock!(ls::LoopSet, condop::Operation, LHS, RHS::Expr, elementbytes::Int, position::Int)
     rhsop = add_compute!(ls, gensym(:iftruerhs), RHS, elementbytes, position)
@@ -56,11 +56,15 @@ function add_andblock!(ls::LoopSet, ex::Expr, elementbytes::Int, position::Int)
 end
 
 function add_orblock!(ls::LoopSet, condop::Operation, LHS, rhsop::Operation, elementbytes::Int, position::Int)
+    negatedcondop = add_compute!(ls, gensym(:negated_mask), :~, [condop], elementbytes)
     if LHS isa Symbol
-        altop = getop(ls, LHS)
-        return add_compute!(ls, LHS, :vifelse, [condop, altop, rhsop], elementbytes)
+        altop = getop(ls, LHS, elementbytes)
+        # return add_compute!(ls, LHS, :vifelse, [condop, altop, rhsop], elementbytes)
+        # Placing altop second seems to let LLVM fuse operations; but as of LLVM 9.0.1 it will not if altop is first
+        # therefore, we negate the condition and switch order so that the altop is second.
+        return add_compute!(ls, LHS, :vifelse, [negatedcondop, rhsop, altop], elementbytes)
     elseif LHS isa Expr && LHS.head === :ref
-        negatedcondop = add_compute!(ls, gensym(:negated_mask), :~, [condop], elementbytes)
+        # negatedcondop = add_compute!(ls, gensym(:negated_mask), :~, [condop], elementbytes)
         return add_conditional_store!(ls, LHS, negatedcondop, rhsop, elementbytes)
     else
         throw("Don't know how to assign onto $LHS.")

src/graphs.jl

@@ -1,6 +1,4 @@
 
-isdense(::Type{<:DenseArray}) = true
-
 # """
 # ShortVector{T} simply wraps a Vector{T}, but uses a different hash function that is faster for short vectors to support using it as the keys of a Dict.
 # This hash function scales O(N) with length of the vectors, so it is slow for long vectors.
@@ -464,7 +462,7 @@ function add_operation!(
             add_compute!(ls, LHS, RHS, elementbytes, position)
         end
     elseif RHS.head === :if
-        add_if!(ls, LHS, RHS, elementbytes)
+        add_if!(ls, LHS, RHS, elementbytes, position)
     else
         throw("Expression not recognized:\n$x")
     end

test/ifelsemasks.jl

@@ -14,6 +14,21 @@
             c[i] = a[i] > b[i] ? a[i] + b[i] : a[i] * b[i]
         end
     end
+    function addormulp1!(c, a, b)
+        for i ∈ eachindex(c,a,b)
+            c[i] = 1 + (a[i] > b[i] ? a[i] + b[i] : a[i] * b[i])
+        end
+    end
+    function addormulp1_avx!(c, a, b)
+        @_avx for i ∈ eachindex(c,a,b)
+            c[i] = 1 + (a[i] > b[i] ? a[i] + b[i] : a[i] * b[i])
+        end
+    end
+    function addormulp1avx!(c, a, b)
+        @avx for i ∈ eachindex(c,a,b)
+            c[i] = 1 + (a[i] > b[i] ? a[i] + b[i] : a[i] * b[i])
+        end
+    end
 
 
     function maybewriteand!(c, a, b)
@@ -155,7 +170,34 @@
             (x1 < 60) || setindex!(x, x3, i)
         end
     end
-
+    function andorassignment!(x, y, z)
+        @inbounds for i ∈ eachindex(x, y, z)
+            yᵢ = y[i]
+            zᵢ = z[i]
+            (yᵢ > 0.5) || (yᵢ *= 2)
+            (zᵢ < 0.5) && (zᵢ *= 2)
+            x[i] = yᵢ * zᵢ
+        end
+    end
+    function andorassignmentavx!(x, y, z)
+        @avx for i ∈ eachindex(x, y, z)
+            yᵢ = y[i]
+            zᵢ = z[i]
+            (yᵢ > 0.5) || (yᵢ *= 2)
+            (zᵢ < 0.5) && (zᵢ *= 2)
+            x[i] = yᵢ * zᵢ
+        end
+    end
+    function andorassignment_avx!(x, y, z)
+        @avx for i ∈ eachindex(x, y, z)
+            yᵢ = y[i]
+            zᵢ = z[i]
+            (yᵢ > 0.5) || (yᵢ *= 2)
+            (zᵢ < 0.5) && (zᵢ *= 2)
+            x[i] = yᵢ * zᵢ
+        end
+    end
+    
     N = 117
     for T ∈ (Float32, Float64, Int32, Int64)
         @show T, @__LINE__
@@ -170,6 +212,11 @@
         @test c1 ≈ c2
         fill!(c2, -999999999); addormulavx!(c2, a, b)
         @test c1 ≈ c2
+        addormulp1!(c1, a, b)
+        addormulp1_avx!(c2, a, b)
+        @test c1 ≈ c2
+        fill!(c2, -999999999); addormulp1avx!(c2, a, b)
+        @test c1 ≈ c2
 
         fill!(c1, -999999999); maybewriteand!(c1, a, b)
         fill!(c2, -999999999); maybewriteand_avx!(c2, a, b)
@@ -183,6 +230,12 @@
         fill!(c2, -999999999); maybewriteoravx!(c2, a, b)
         @test c1 ≈ c2
 
+        andorassignment!(c1, a, b)
+        andorassignmentavx!(c2, a, b)
+        @test c1 ≈ c2
+        fill!(c2, -999999999); andorassignment_avx!(c2, a, b)
+        @test c1 ≈ c2
+        
         if T <: Union{Float32,Float64}
             a .*= 100;
         end