Fix a couple choose_order tests, new precompile.jl.

Author: chriselrod

src/broadcast.jl

@@ -203,7 +203,7 @@ function add_broadcast!(
 end
 function add_broadcast!(
     ls::LoopSet, destname::Symbol, bcname::Symbol, loopsyms::Vector{Symbol},
-    ::Type{SubArray{T,N,A,S,B}}, elementbytes::Int
+    @nospecialize(::Type{SubArray{T,N,A,S,B}}), elementbytes::Int
 ) where {T,N,N2,A<:AbstractArray{T,N2},B,N3,S <: Tuple{Int,Vararg{Any,N3}}}
     inds = Vector{Symbol}(undef, N+1)
     inds[1] = Symbol("##DISCONTIGUOUSSUBARRAY##")

src/filter.jl

@@ -25,11 +25,11 @@ if (Base.libllvm_version ≥ v"7" && VectorizationBase.AVX512F) || Base.libllvm_
         end
         x
     end
-    vfilter!(f, x::Vector{T}) where {T<:NativeTypes} = vfilter!(f, x, x)
-    vfilter(f, y::AbstractArray{T}) where {T<:NativeTypes} = vfilter!(f, Vector{T}(undef, length(y)), y)
+    vfilter!(f::F, x::Vector{T}) where {F, T<:NativeTypes} = vfilter!(f, x, x)
+    vfilter(f::F, y::AbstractArray{T}) where {F, T<:NativeTypes} = vfilter!(f, Vector{T}(undef, length(y)), y)
 end
-vfilter(f, y) = filter(f, y)
-vfilter!(f, y) = filter!(f, y)
+vfilter(f::F, y) where {F} = filter(f, y)
+vfilter!(f::F, y) where {F} = filter!(f, y)
 
 """
     vfilter(f, a::AbstractArray)

src/precompile.jl

@@ -1,30 +1,33 @@
 function _precompile_()
     ccall(:jl_generating_output, Cint, ()) == 1 || return nothing
 
-    precompile(Tuple{Type{LoopVectorization.LoopSet},Array{Symbol,1},Array{Int64,1},Array{LoopVectorization.Loop,1},Dict{Symbol,LoopVectorization.Operation},Array{LoopVectorization.Operation,1},Array{Int64,1},Array{Int64,1},LoopVectorization.LoopOrder,Expr,Expr,Array{Tuple{Int64,Symbol},1},Array{Tuple{Int64,Int64},1},Array{Tuple{Int64,Float64},1},Array{Int64,1},Array{Int64,1},Array{Symbol,1},Array{Symbol,1},Array{Symbol,1},Array{LoopVectorization.ArrayReferenceMeta,1},Array{Float64,2},Array{Float64,2},Array{Bool,1},Array{Bool,1},Base.RefValue{LoopVectorization.UnrollSpecification},Base.RefValue{Bool},Base.RefValue{LoopVectorization.LoopStartStopManager},Base.RefValue{Int64},Base.RefValue{Bool},Symbol})
+    precompile(Tuple{Type{LoopVectorization.ArrayRefStruct},LoopVectorization.LoopSet,LoopVectorization.ArrayReferenceMeta,Vector{Symbol},Vector{Int64}})
     precompile(Tuple{Type{LoopVectorization.LoopSet},Expr,Symbol})
-    precompile(Tuple{Type{LoopVectorization.Loop},LoopVectorization.LoopSet,Expr,Symbol,Type{VectorizationBase.StaticLowerUnitRange{1}}})
+    precompile(Tuple{Type{LoopVectorization.LoopSet},Symbol})
+    precompile(Tuple{Type{Vector{Tuple{Int64,LoopVectorization.NumberType}}},Vector{Int64}})
+    precompile(Tuple{typeof(Base._mapreduce),Base.ComposedFunction{Type{Float64},typeof(length)},typeof(Base.mul_prod),IndexLinear,Vector{LoopVectorization.Loop}})
     precompile(Tuple{typeof(LoopVectorization._avx_loopset),Core.SimpleVector,Core.SimpleVector,Core.SimpleVector,Core.SimpleVector,Core.SimpleVector,Any})
+    precompile(Tuple{typeof(LoopVectorization.add_broadcast!),LoopVectorization.LoopSet,Symbol,Symbol,Vector{Symbol},Type{Vector{Float64}},Int64})
     precompile(Tuple{typeof(LoopVectorization.add_compute!),LoopVectorization.LoopSet,Symbol,Expr,Int64,Int64,LoopVectorization.ArrayReferenceMetaPosition})
     precompile(Tuple{typeof(LoopVectorization.add_compute!),LoopVectorization.LoopSet,Symbol,Expr,Int64,Int64,Nothing})
-    precompile(Tuple{typeof(LoopVectorization.add_constant!),LoopVectorization.LoopSet,Float64,Array{Symbol,1},Symbol,Int64})
-    precompile(Tuple{typeof(LoopVectorization.add_constant_offset_load_elmination_cost!),SubArray{Float64,1,Array{Float64,2},Tuple{Base.Slice{Base.OneTo{Int64}},Int64},true},SubArray{Float64,1,Array{Float64,2},Tuple{Base.Slice{Base.OneTo{Int64}},Int64},true},Base.RefValue{Bool},LoopVectorization.LoopSet,LoopVectorization.Operation,Float64,LoopVectorization.UnrollSymbols,Bool,Bool,Int64,Int64,Bool})
+    precompile(Tuple{typeof(LoopVectorization.add_constant!),LoopVectorization.LoopSet,Float64,Vector{Symbol},Symbol,Int64})
     precompile(Tuple{typeof(LoopVectorization.add_if!),LoopVectorization.LoopSet,Symbol,Expr,Int64,Int64,LoopVectorization.ArrayReferenceMetaPosition})
-    precompile(Tuple{typeof(LoopVectorization.add_if!),LoopVectorization.LoopSet,Symbol,Expr,Int64,Int64,Nothing})
-    precompile(Tuple{typeof(LoopVectorization.add_mref!),LoopVectorization.LoopSet,LoopVectorization.ArrayReferenceMeta,Int64,Type{VectorizationBase.PermutedDimsStridedPointer{(2, 1),(2, 1),Float64,VectorizationBase.PackedStridedPointer{Float64,1}}},Symbol})
-    precompile(Tuple{typeof(LoopVectorization.add_pow!),LoopVectorization.LoopSet,Symbol,Symbol,Int64,Int64,Int64})
-    precompile(Tuple{typeof(LoopVectorization.argmeta_and_consts_description),LoopVectorization.LoopSet,Array{Symbol,1}})
-    precompile(Tuple{typeof(LoopVectorization.array_reference_meta!),LoopVectorization.LoopSet,Symbol,SubArray{Any,1,Array{Any,1},Tuple{UnitRange{Int64}},true},Int64,Nothing})
-    precompile(Tuple{typeof(LoopVectorization.choose_tile),LoopVectorization.LoopSet})
+    precompile(Tuple{typeof(LoopVectorization.add_load!),LoopVectorization.LoopSet,LoopVectorization.Operation,Bool,Bool})
+    precompile(Tuple{typeof(LoopVectorization.add_mref!),LoopVectorization.LoopSet,LoopVectorization.ArrayReferenceMeta,Int64,Type{VectorizationBase.PermutedDimsStridedPointer{(4, 2, 3, 1),(4, 2, 3, 1),Float64,VectorizationBase.PermutedDimsStridedPointer{(3, 1, 4, 2),(2, 4, 1, 3),Float64,VectorizationBase.PackedStridedPointer{Float64,3}}}},Symbol})
+    precompile(Tuple{typeof(LoopVectorization.add_pow!),LoopVectorization.LoopSet,Symbol,Expr,Float64,Int64,Int64})
+    precompile(Tuple{typeof(LoopVectorization.add_pow!),LoopVectorization.LoopSet,Symbol,Expr,Int64,Int64,Int64})
+    precompile(Tuple{typeof(LoopVectorization.argmeta_and_consts_description),LoopVectorization.LoopSet,Vector{Symbol}})
+    precompile(Tuple{typeof(LoopVectorization.array_reference_meta!),LoopVectorization.LoopSet,Symbol,SubArray{Any,1,Vector{Any},Tuple{UnitRange{Int64}},true},Int64,Nothing})
+    precompile(Tuple{typeof(LoopVectorization.avx_loopset),Vector{LoopVectorization.Instruction},Vector{LoopVectorization.OperationStruct},Vector{LoopVectorization.ArrayRefStruct},Core.SimpleVector,Core.SimpleVector,Core.SimpleVector,Any})
+    precompile(Tuple{typeof(LoopVectorization.check_macro_kwarg),Expr})
     precompile(Tuple{typeof(LoopVectorization.cost_vec_buf),LoopVectorization.LoopSet})
-    precompile(Tuple{typeof(LoopVectorization.determine_unroll_factor),LoopVectorization.LoopSet,Array{Symbol,1},Symbol,Symbol})
-    precompile(Tuple{typeof(LoopVectorization.evaluate_cost_tile),LoopVectorization.LoopSet,Array{Symbol,1},LoopVectorization.UnrollSymbols})
-    precompile(Tuple{typeof(LoopVectorization.evaluate_cost_unroll),LoopVectorization.LoopSet,Array{Symbol,1},Symbol,Float64})
+    precompile(Tuple{typeof(LoopVectorization.evaluate_cost_tile),LoopVectorization.LoopSet,Vector{Symbol},LoopVectorization.UnrollSymbols})
+    precompile(Tuple{typeof(LoopVectorization.evaluate_cost_unroll),LoopVectorization.LoopSet,Vector{Symbol},Symbol,Float64})
     precompile(Tuple{typeof(LoopVectorization.generate_call),LoopVectorization.LoopSet,Tuple{Int8,Int8,Int8},Bool})
     precompile(Tuple{typeof(LoopVectorization.pushop!),LoopVectorization.LoopSet,LoopVectorization.Operation,Symbol})
-    precompile(Tuple{typeof(LoopVectorization.stride_penalty),LoopVectorization.LoopSet,LoopVectorization.Operation,Array{Symbol,1},Array{Int64,1}})
+    precompile(Tuple{typeof(LoopVectorization.reject_candidate),LoopVectorization.Operation,Symbol,Symbol})
+    precompile(Tuple{typeof(LoopVectorization.stride_penalty),LoopVectorization.LoopSet,Vector{Symbol}})
     precompile(Tuple{typeof(LoopVectorization.substitute_broadcast),Expr,Symbol})
-    precompile(Tuple{typeof(println),Base.GenericIOBuffer{Array{UInt8,1}},Array{LoopVectorization.Operation,1}})
+    precompile(Tuple{typeof(println),IOBuffer,Vector{LoopVectorization.Operation}})
     precompile(Tuple{typeof(push!),LoopVectorization.LoopSet,Expr,Int64,Int64})
 end
-

test/gemm.jl

@@ -336,7 +336,7 @@
                end);
     lsr2amb = LoopVectorization.LoopSet(r2ambq);
     if LoopVectorization.REGISTER_COUNT == 32
-        @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :n, :m, :m, 7, 3)
+        @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :n, :m, :m, 7, 3)
     elseif LoopVectorization.REGISTER_COUNT == 16
         @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :n, :m, :m, 4, 2)
     end

test/miscellaneous.jl

@@ -144,7 +144,7 @@ using Test
     lsvar = LoopVectorization.LoopSet(varq);
     # LoopVectorization.choose_order(lsvar)
     # @test LoopVectorization.choose_order(lsvar) == (Symbol[:j,:i], :j, :i, :j, Unum, Tnum)
-    @test LoopVectorization.choose_order(lsvar) == (Symbol[:j,:i], :j, Symbol("##undefined##"), :j, 8, -1)
+    @test LoopVectorization.choose_order(lsvar) == (Symbol[:j,:i], :j, Symbol("##undefined##"), :j, 4, -1)
     # if LoopVectorization.REGISTER_COUNT == 32
     #     @test LoopVectorization.choose_order(lsvar) == (Symbol[:j,:i], :j, :i, :j, 2, 10)
     # elseif LoopVectorization.REGISTER_COUNT == 16