Add a demo measuring all poorly-inferred backedges

These are the scripts used for JuliaLang/julia#36393

demos/abstract_analyze_versions.jl

@@ -0,0 +1,107 @@
+using PyPlot
+
+function parseline(line)
+    m = match(r"(.*) => (.*)", line)
+    sigstr, count = m.captures[1], parse(Int, m.captures[2])
+    sig = try
+        ex = Meta.parse(sigstr)
+        eval(ex)
+    catch
+        return nothing
+    end
+    return sig, count
+end
+
+function parsedata(filename)
+    lines = readlines(filename)
+    sigcount = IdDict{Any,Int}()
+    for line in lines
+        ret = parseline(line)
+        ret === nothing && continue
+        sig, count = ret
+        sigcount[sig] = count
+    end
+    return sigcount
+end
+
+function split_comparable(sigc1, sigc2)
+    c1, c2, sigs = Int[], Int[], Any[]
+    for (sig, c) in sigc1
+        push!(sigs, sig)
+        push!(c1, sigc1[sig])
+        push!(c2, get(sigc2, sig, 0))
+    end
+    for (sig, c) in sigc2
+        if !haskey(sigc1, sig)
+            push!(sigs, sig)
+            push!(c1, 0)
+            push!(c2, c)
+        end
+    end
+    return sigs, c1, c2
+end
+
+sigc16 = parsedata("/tmp/methdata_$VERSION.log")
+sigc14 = parsedata("/tmp/methdata_1.4.3-pre.0.log")
+
+sigs, c1, c2 = split_comparable(sigc14, sigc16)
+mx1, mx2 = maximum(c1), maximum(c2)
+isexported(sig) = (ft = Base.unwrap_unionall(sig).parameters[1]; isdefined(Main, ft.name.mt.name))
+colors = [isexported(sig) ? "magenta" : "green" for sig in sigs]
+
+function on_click(event)
+    x, y = event.xdata, event.ydata
+    normsqrdist(pr) = ((pr[1]-x)/mx1)^2 + ((pr[2]-y)/mx2)^2
+    idx = argmin(normsqrdist.(zip(c1, c2)))
+    println(sigs[idx])
+end
+begin
+    fig, ax = plt.subplots()
+    ax.scatter(c1 .+ 1, c2 .+ 1, c=colors)  # + 1 for the log-scaling
+    ax.set_xlabel("# backedges + 1, 1.4")
+    ax.set_ylabel("# backedges + 1, 1.6")
+    ax.set_xscale("log")
+    ax.set_yscale("log")
+    fig.canvas.callbacks.connect("button_press_event", on_click)
+    fig
+end
+
+# Ones we've made progress on:
+# ==(::Any, Symbol)
+# ==(::Symbol, ::Any)
+# ==(::Any, ::Nothing)
+# ==(::UUID, ::Any)
+# ==(::AbstractString, ::String)
+# isequal(::Symbol, ::Any)
+# isequal(::Any, ::Symbol)
+# isequal(::Any, ::Nothing)
+# isequal(::UUID, ::Any)
+# cmp(::AbstractString, ::String)
+# convert(::Type{Int}, ::Integer)
+# convert(::Type{UInt}, ::Integer)
+# convert(::Type{Union{Nothing,Module}}, ::Any)
+# Base.to_index(::Integer)
+# iterate(::Base.OneTo, ::Any)
+# repr(::Any)
+# thisind(::AbstractString, ::Int)
+# getindex(::String, ::Any)
+# string(::String, ::Integer, ::String)
+# ^(::String, ::Integer)
+# repeat(::String, ::Integer)
+# Base.isidentifier(::AbstractString)
+# +(::Ptr{UInt8}, ::Integer)
+# Base._show_default(::Base.GenericIOBuffer{Array{UInt8,1}}, ::Any)
+
+# Ones that are better but I don't remember helping with
+# isconcretetype(::Any)
+# pointer(::String, ::Integer)
+
+# Regressions:
+# basename(::AbstractString)
+# splitdir(::AbstractString)
+# isfile(::Any)
+# joinpath(::AbstractString, ::String)
+# sizeof(::Unsigned)
+# +(::Int, ::Any, ::Any)
+# Base.split_sign(::Integer)
+# in(::Any, ::Tuple{Symbol})

demos/abstract_gen_data.jl

@@ -0,0 +1,129 @@
+using MethodAnalysis
+
+# Analyze MethodInstance signatures and select those that seem at risk for being invalidated.
+function atrisktype(@nospecialize(typ))
+    # signatures like `convert(Vector, a)`, `foo(::Vararg{Synbol,N}) where N` do not seem to pose a problem
+    isa(typ, TypeVar) && return false
+    # isbits parameters are not a problem
+    isa(typ, Type) || return false
+    if isa(typ, UnionAll)
+        typ = Base.unwrap_unionall(typ)
+    end
+    # Exclude signatures with Union{}
+    typ === Union{} && return false
+    isa(typ, Union) && return atrisktype(typ.a) | atrisktype(typ.b)
+    # Type{T}: signatures like `convert(::Type{AbstractString}, ::String)` are not problematic, so mark Type as OK
+    typ <: Type && return false
+    if typ <: Tuple && length(typ.parameters) >= 1
+        p1 = typ.parameters[1]
+        # Constructor calls are not themselves a problem (any `convert`s they trigger might be, but those are covered)
+        isa(p1, Type) && p1 <: Type && return false
+        # convert(::Type{T}, ::S) where S<:T is not problematic
+        if p1 === typeof(Base.convert) || p1 === typeof(Core.convert) || p1 === typeof(Core.Compiler.convert)
+            p2, p3 = typ.parameters[2], typ.parameters[3]
+            if isa(p2, Type)
+                p2 = Base.unwrap_unionall(p2)
+                if isa(p2, DataType) && length(p2.parameters) === 1
+                    T = p2.parameters[1]
+                    isa(p3, Type) && isa(T, Type) && p3 <: T && return false
+                end
+            end
+        # `getindex`, `length`, etc are OK for various Tuple{T1,T2,...}
+        elseif (p1 === typeof(Base.getindex) || p1 === typeof(Core.Compiler.getindex)) ||
+               (p1 === typeof(Base.length)  || p1 === typeof(Core.Compiler.length)) ||
+               (p1 === typeof(Base.isempty) || p1 === typeof(Core.Compiler.isempty)) ||
+               (p1 === typeof(Base.iterate) || p1 === typeof(Core.iterate) || p1 === typeof(Core.Compiler.iterate))
+            p2 = typ.parameters[2]
+            if isa(p2, Type)
+                p2 = Base.unwrap_unionall(p2)
+                p2 <: Tuple && return false
+            end
+        # show(io::IO, x) is OK as long as typeof(x) is safe
+        elseif p1 === typeof(Base.show)
+            atrisktype(typ.parameters[2]) && return true
+            length(typ.parameters) == 3 && atrisktype(typ.parameters[3]) && return true
+            return false
+        end
+    end
+    # Standard DataTypes
+    isconcretetype(typ) && return false
+    # ::Function args are excluded
+    typ === Function && return false
+    !isempty(typ.parameters) && (any(atrisktype, typ.parameters) || return false)
+    return true
+end
+
+# A few tests
+@assert  atrisktype(Tuple{typeof(==),Any,Any})
+@assert  atrisktype(Tuple{typeof(==),Symbol,Any})
+@assert  atrisktype(Tuple{typeof(==),Any,Symbol})
+@assert !atrisktype(Tuple{typeof(==),Symbol,Symbol})
+@assert !atrisktype(Tuple{typeof(convert),Type{Any},Any})
+@assert !atrisktype(Tuple{typeof(convert),Type{AbstractString},AbstractString})
+@assert !atrisktype(Tuple{typeof(convert),Type{AbstractString},String})
+@assert  atrisktype(Tuple{typeof(convert),Type{String},AbstractString})
+@assert !atrisktype(Tuple{typeof(map),Function,Vector{Any}})
+@assert !atrisktype(Tuple{typeof(getindex),Dict{Union{String,Int},Any},Union{String,Int}})
+@assert  atrisktype(Tuple{typeof(getindex),Dict{Union{String,Int},Any},Any})
+@assert !atrisktype(Tuple{Type{BoundsError},Any,Any})
+@assert  atrisktype(Tuple{typeof(sin),Any})
+@assert !atrisktype(Tuple{typeof(length),Tuple{Any,Any}})
+
+function collect_mis(m::Method)
+    list = Core.MethodInstance[]
+    visit(m) do item
+        if isa(item, Core.MethodInstance)
+            push!(list, item)
+            return false
+        end
+        return true
+    end
+    return list
+end
+
+const mis = Dict{Method,Vector{Core.MethodInstance}}()
+visit() do item
+    if item isa Method
+        m = item
+        mis[m] = collect_mis(m)
+        return false
+    end
+    return true
+end
+
+# Count # of backedges for MethodInstances with abstract types
+const becounter = Dict{Core.MethodInstance,Int}()
+visit() do item
+    if item isa Core.MethodInstance
+        if atrisktype(item.specTypes)
+            becounter[item] = length(all_backedges(item))
+        end
+        return false
+    end
+    return true
+end
+
+prs = sort!(collect(becounter); by=last)
+
+# Organize them by method instead
+
+const mcounter = Dict{Method,Int}()
+for (mi, c) in becounter
+    oc = get(mcounter, mi.def, 0)
+    mcounter[mi.def] = oc + c
+end
+
+mprs = sort!(collect(mcounter); by=last)
+
+open("/tmp/methinstdata_$VERSION.log", "w") do io
+    for (mi, c) in prs
+        c == 0 && continue
+        println(io, mi.specTypes=>c)
+    end
+end
+open("/tmp/methdata_$VERSION.log", "w") do io
+    for (m, c) in mprs
+        c == 0 && continue
+        println(io, m.sig=>c)
+    end
+end