Clean up some old/removed code, add tests for first-ninth.

Author: chriselrod

src/broadcast.jl

@@ -236,10 +236,6 @@ function add_broadcast!(
         parent = add_broadcast!(ls, gensym(:temp), argname, loopsyms, arg, elementbytes)::Operation
         push!(parents, parent)
         mergesetdiffv!(deps, loopdependencies(parent), reduceddependencies(parent))
-        # if !(isload(parent) || isconstant(parent))# && parent.instruction.instr ∉ (:reduced_add, :reduced_prod, :reduce_to_add, :reduce_to_prod)
-            # mergesetv!(reduceddeps, reduceddependencies(parent))
-        # end
-        # pushparent!(parents, deps, reduceddeps, parent)
     end
     op = Operation(
         length(operations(ls)), destname, elementbytes, instr, compute, deps, NOPARENTS, parents

src/costs.jl

@@ -244,7 +244,7 @@ const COST = Dict{Symbol,InstructionCost}(
     :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),
+    :eighth => 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),

src/graphs.jl

@@ -178,10 +178,10 @@ end
 use_expect() = true
 function looplengthexpr(loop, n)
     le = looplengthexpr(loop)
-    if false && use_expect() && isone(n) && !isstaticloop(loop)
-        le = expect(le)
-        push!(le.args, Expr(:call, Expr(:curly, :Val, length(loop))))
-    end
+    # if false && use_expect() && isone(n) && !isstaticloop(loop)
+    #     le = expect(le)
+    #     push!(le.args, Expr(:call, Expr(:curly, :Val, length(loop))))
+    # end
     le
 end
 
@@ -251,7 +251,6 @@ struct LoopSet
     loadelimination::Base.RefValue{Bool}
     lssm::Base.RefValue{LoopStartStopManager}
     vector_width::Base.RefValue{Int}
-    # align_loops::Base.RefValue{Int}
     isbroadcast::Base.RefValue{Bool}
     mod::Symbol
 end
@@ -341,7 +340,7 @@ function LoopSet(mod::Symbol)
         Matrix{Float64}(undef, 5, 2),
         Bool[], Bool[], Ref{UnrollSpecification}(),
         Ref(false), Ref{LoopStartStopManager}(),
-        Ref(0), #=Ref(0),=# Ref(false), mod
+        Ref(0), Ref(false), mod
     )
 end
 
@@ -393,16 +392,6 @@ function Operation(ls::LoopSet, variable, elementbytes, instr, optype, mpref::Ar
     Operation(length(operations(ls)), variable, elementbytes, instr, optype, mpref)
 end
 
-# load_operations(ls::LoopSet) = ls.loadops
-# compute_operations(ls::LoopSet) = ls.computeops
-# store_operations(ls::LoopSet) = ls.storeops
-# function operations(ls::LoopSet)
-    # Base.Iterators.flatten((
-        # load_operations(ls),
-        # compute_operations(ls),
-        # store_operations(ls)
-    # ))
-# end
 operations(ls::LoopSet) = ls.operations
 function pushop!(ls::LoopSet, op::Operation, var::Symbol = name(op))
     for opp ∈ operations(ls)
@@ -686,7 +675,7 @@ function Base.push!(ls::LoopSet, ex::Expr, elementbytes::Int, position::Int)
                 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]
+                    f = (:first,:second,:third,:fourth,:fifth,:sixth,:seventh,:eighth,:ninth)[i]
                     lhsi = LHS.args[i]
                     if lhsi isa Symbol
                         add_compute!(ls, lhsi, f, vparents, elementbytes)

test/miscellaneous.jl

@@ -721,7 +721,23 @@ function findreducedparentfornonvecstore!(U::AbstractMatrix{T}, E1::AbstractVect
     U,E1
 end
 
-
+ninereturns(x) = (0.25x, 0.5x, 0.75, 1.0x, 1.25x, 1.5x, 1.75x, 2.0x, 2.25x)
+function manyreturntest(x)
+    s = zero(eltype(x))
+    for j ∈ eachindex(x)
+        a, b, c, d, e, f, g, h, i = ninereturns(x[j])
+        s += a * i + b * h + c * g - d
+    end
+    s
+end
+function manyreturntestavx(x)
+    s = zero(eltype(x))
+    @avx for j ∈ eachindex(x)
+        a, b, c, d, e, f, g, h, i = ninereturns(x[j])
+        s += a * i + b * h + c * g - d
+    end
+    s
+end
 
 
     for T ∈ (Float32, Float64)
@@ -927,7 +943,9 @@ end
         R .+= randn.(T); Rc = copy(R);
         @test maxavx!(R, Q, true) == max.(vec(maximum(Q, dims=(2,3))), Rc)
 
-        U0 = randn(5,7); E0 = randn(7);
+        @test manyreturntest(R) ≈ manyreturntestavx(R)
+        
+        U0 = randn(T, 5, 7); E0 = randn(T, 7);
         U1, E1 = splitintonoloop_reference(copy(U0), copy(E0));
         U2, E2 = splitintonoloop(copy(U0), copy(E0));
         @test U1 ≈ U2