A few bug fixes and added tests.

Author: chriselrod

src/add_ifelse.jl

@@ -17,7 +17,7 @@ function add_if!(ls::LoopSet, LHS::Symbol, RHS::Expr, elementbytes::Int, positio
     end
     iftrue = RHS.args[2]
     if iftrue isa Expr
-        trueop = add_operation!(ls, Symbol(:iftrue), iftrue, elementbytes, position)
+        trueop = add_operation!(ls, gensym(:iftrue), iftrue, elementbytes, position)
         if iftrue.head === :ref && all(ld -> ld ∈ loopdependencies(trueop), loopdependencies(condop)) && !search_tree(parents(condop), trueop)
             trueop.instruction = Instruction(:conditionalload)
             push!(parents(trueop), condop)
@@ -27,7 +27,7 @@ function add_if!(ls::LoopSet, LHS::Symbol, RHS::Expr, elementbytes::Int, positio
     end
     iffalse = RHS.args[3]
     if iffalse isa Expr
-        falseop = add_operation!(ls, Symbol(:iffalse), iffalse, elementbytes, position)
+        falseop = add_operation!(ls, gensym(:iffalse), iffalse, elementbytes, position)
         if iffalse.head === :ref && all(ld -> ld ∈ loopdependencies(falseop), loopdependencies(condop)) && !search_tree(parents(condop), falseop)
             falseop.instruction = Instruction(:conditionalload)
             push!(parents(falseop), negateop!(ls, condop, elementbytes))

src/constructors.jl

@@ -1,9 +1,19 @@
 
 ### This file contains convenience functions for constructing LoopSets.
 
+# function strip_unneeded_const_deps!(ls::LoopSet)
+#     for op ∈ operations(ls)
+#         if isconstant(op) && iszero(length(reducedchildren(op)))
+#             op.dependencies = NODEPENDENCY
+#         end
+#     end
+#     ls
+# end
+
 function Base.copyto!(ls::LoopSet, q::Expr)
     q.head === :for || throw("Expression must be a for loop.")
     add_loop!(ls, q, 8)
+    # strip_unneeded_const_deps!(ls)
 end
 
 function add_ci_call!(q::Expr, @nospecialize(f), args, syms, i, mod = nothing)

src/graphs.jl

@@ -555,9 +555,10 @@ end
 instruction!(ls::LoopSet, x::Symbol) = instruction(x)
 
 
-function maybe_const_compute!(ls::LoopSet, op::Operation, elementbytes::Int, position::Int)
+function maybe_const_compute!(ls::LoopSet, LHS::Symbol, op::Operation, elementbytes::Int, position::Int)
+    # return op
     if iscompute(op) && iszero(length(loopdependencies(op)))
-        add_constant!(ls, mangledvar(op), ls.loopsymbols[1:position], gensym(instruction(op).instr), elementbytes, :numericconstant)
+        ls.opdict[LHS] = add_constant!(ls, LHS, ls.loopsymbols[1:position], gensym(instruction(op).instr), elementbytes, :numericconstant)
     else
         op
     end
@@ -572,6 +573,9 @@ function strip_op_linenumber_nodes(q::Expr)
     end
 end
 
+function add_operation!(ls::LoopSet, LHS::Symbol, RHS::Symbol, elementbytes::Int, position::Int)
+    add_constant!(ls, RHS, ls.loopsymbols[1:position], LHS, elementbytes)
+end
 function add_operation!(
     ls::LoopSet, LHS::Symbol, RHS::Expr, elementbytes::Int, position::Int
 )
@@ -670,7 +674,7 @@ function Base.push!(ls::LoopSet, ex::Expr, elementbytes::Int, position::Int)
         RHS = ex.args[2]
         if LHS isa Symbol
             if RHS isa Expr
-                maybe_const_compute!(ls, add_operation!(ls, LHS, RHS, elementbytes, position), elementbytes, position)
+                maybe_const_compute!(ls, LHS, add_operation!(ls, LHS, RHS, elementbytes, position), elementbytes, position)
             else
                 add_constant!(ls, RHS, ls.loopsymbols[1:position], LHS, elementbytes)
             end
@@ -689,7 +693,7 @@ function Base.push!(ls::LoopSet, ex::Expr, elementbytes::Int, position::Int)
             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)]
+                vparents = Operation[maybe_const_compute!(ls, lhstemp, add_operation!(ls, lhstemp, RHS, elementbytes, position), elementbytes, position)]
                 for i ∈ eachindex(LHS.args)
                     f = (:first,:second,:third,:fourth,:fifth,:sixth,:seventh,:eighth,:ninth)[i]
                     lhsi = LHS.args[i]

test/ifelsemasks.jl

@@ -317,6 +317,36 @@ T = Float32
             x[i] = yᵢ * zᵢ
         end
     end
+    
+    function twoifelses!(res, half, m, keep=nothing, final=true)
+        𝒶𝓍j=axes(half,1)
+        for j in 𝒶𝓍j
+            𝓇𝒽𝓈 = if isnothing(keep)
+                log(half[j]) + m[j]
+            else
+                res[j] + (log(half[j]) + m[j])
+            end
+            res[j] = isnothing(final) ? 𝓇𝒽𝓈 : exp(𝓇𝒽𝓈)
+        end
+        res
+    end
+    function twoifelses_avx!(res, half, m, keep=nothing, final=true)
+        𝒶𝓍j=axes(half,1)
+        @avx for j in 𝒶𝓍j
+            𝓇𝒽𝓈 = if isnothing(keep)
+                log(half[j]) + m[j]
+            else
+                res[j] + (log(half[j]) + m[j])
+            end
+            res[j] = if isnothing(final)
+                𝓇𝒽𝓈
+            else
+                exp(𝓇𝒽𝓈)
+            end
+        end
+        res
+    end
+    
     N = 117
     for T ∈ (Float32, Float64, Int32, Int64)
         @show T, @__LINE__
@@ -425,6 +455,25 @@ T = Float32
         condloadscalar!(C1, B, b, d)
         condloadscalaravx!(C2, B, b, d)
         @test C1 ≈ C2
+
+        if T <: Integer
+            half = rand(T(1):T(100), 7);
+            m = rand(T(-10):T(10), 7);
+        else
+            half = rand(T, 7); m = rand(T, 7);
+        end;
+        if sizeof(T) == 4
+            res1 = Vector{Float32}(undef, 7);
+            res2 = Vector{Float32}(undef, 7);
+        else
+            res1 = Vector{Float64}(undef, 7);
+            res2 = Vector{Float64}(undef, 7);
+        end
+
+        for keep ∈ (nothing,true), final ∈ (nothing,true)
+            @test twoifelses!(res1, half, m) ≈ twoifelses_avx!(res2, half, m)
+        end
+
     end
 
 

test/miscellaneous.jl

@@ -752,6 +752,38 @@ function manyreturntestavx(x)
     s
 end
 
+function maybe_const_issue144!(𝛥mat, 𝛥ℛ, mat, ℛ)
+    𝛥ℛ_value = 𝛥ℛ.value
+    for j in axes(mat,2)
+        for i in axes(mat,1)
+            ℰ𝓍1 = conj(𝛥ℛ_value) # could be outside both loops
+            ℰ𝓍2 = -(ℛ[j])        # could be outside i loop
+            ℰ𝓍3 = exp(ℰ𝓍2)       # could be outside i loop
+            ℰ𝓍4 = exp(mat[i, j])
+            ℰ𝓍5 = ℰ𝓍3 * ℰ𝓍4
+            ℰ𝓍6 = ℰ𝓍1 * ℰ𝓍5
+            ℰ𝓍7 = conj(ℰ𝓍6)
+            𝛥mat[i, j] = 𝛥mat[i, j] + ℰ𝓍7
+        end
+    end
+    𝛥mat
+end
+function maybe_const_issue144_avx!(𝛥mat, 𝛥ℛ, mat, ℛ)
+    𝛥ℛ_value = 𝛥ℛ.value
+    @avx for j in axes(mat,2)
+        for i in axes(mat,1)
+            ℰ𝓍1 = conj(𝛥ℛ_value)
+            ℰ𝓍2 = -(ℛ[j])
+            ℰ𝓍3 = exp(ℰ𝓍2)
+            ℰ𝓍4 = exp(mat[i, j])
+            ℰ𝓍5 = ℰ𝓍3 * ℰ𝓍4
+            ℰ𝓍6 = ℰ𝓍1 * ℰ𝓍5
+            ℰ𝓍7 = conj(ℰ𝓍6)
+            𝛥mat[i, j] = 𝛥mat[i, j] + ℰ𝓍7
+        end
+    end
+    𝛥mat
+end
 
     for T ∈ (Float32, Float64)
         @show T, @__LINE__
@@ -970,6 +1002,10 @@ end
 
         @test all(isequal(81), powcseliteral!(E0))
         @test all(isequal(81), powcsesymbol!(E3))
+
+        @test maybe_const_issue144!(zeros(T, 3,4), (value=one(T),), collect(reshape(1:12, 3,4)), ones(T, 4)) ≈ maybe_const_issue144_avx!(zeros(T,3,4), (value=one(T),), collect(reshape(1:12, 3,4)), ones(T,4))
+
+
     end
     for T ∈ [Int16, Int32, Int64]
         n = 8sizeof(T) - 1