Make Vararg not a DataType (JuliaLang#38136)

Currently `Vararg` is a DataType, but is special cased in a
bunch of places to give it special behavior (e.g. in subtyping
and of course in tuple construction). However, unlike all other
DataTypes, it cannot appear as a type parameter, which caused trouble in
PR JuliaLang#38071. Having it be a DataType is a bit of a pun of convenience
in the first place - it's a lot more similar to a tvar (which can
be considered an implementation detail of UnionAll in the same way
Vararg is an implementation detail of Tuple), which has its own
non-type object. This PR does the same to Vararg, and moves it
from being an abstract DataType with special cased behavior to
its own custom type (called `Core.TypeofVararg`). There are a
few small behavior differences, but they are mostly internal.
In particular, we no longer have `Vararg <: Type` and Vararg
objects no longer have the .parameters field. Also, things like
`Vararg{T} where T` are technically illegal now since Vararg is
not a type. However, since a lot of people are using that pattern,
I've brought it back with a deprecation (which is of course off
by default). The only things that's disallowed is `Vararg{N, N} where N`,
but I haven't seen anybody use that.

Author: Keno

base/boot.jl

@@ -242,14 +242,23 @@ ccall(:jl_toplevel_eval_in, Any, (Any, Any),
       (f::typeof(Typeof))(x) = ($(_expr(:meta,:nospecialize,:x)); isa(x,Type) ? Type{x} : typeof(x))
       end)
 
-# let the compiler assume that calling Union{} as a constructor does not need
-# to be considered ever (which comes up often as Type{<:T})
-Union{}(a...) = throw(MethodError(Union{}, a))
 
 macro nospecialize(x)
     _expr(:meta, :nospecialize, x)
 end
 
+TypeVar(n::Symbol) = _typevar(n, Union{}, Any)
+TypeVar(n::Symbol, @nospecialize(ub)) = _typevar(n, Union{}, ub)
+TypeVar(n::Symbol, @nospecialize(lb), @nospecialize(ub)) = _typevar(n, lb, ub)
+
+UnionAll(v::TypeVar, @nospecialize(t)) = ccall(:jl_type_unionall, Any, (Any, Any), v, t)
+
+const Vararg = ccall(:jl_toplevel_eval_in, Any, (Any, Any), Core, _expr(:new, TypeofVararg))
+
+# let the compiler assume that calling Union{} as a constructor does not need
+# to be considered ever (which comes up often as Type{<:T})
+Union{}(a...) = throw(MethodError(Union{}, a))
+
 Expr(@nospecialize args...) = _expr(args...)
 
 abstract type Exception end
@@ -378,12 +387,6 @@ mutable struct WeakRef
                                       (Ptr{Cvoid}, Any), getptls(), v)
 end
 
-TypeVar(n::Symbol) = _typevar(n, Union{}, Any)
-TypeVar(n::Symbol, @nospecialize(ub)) = _typevar(n, Union{}, ub)
-TypeVar(n::Symbol, @nospecialize(lb), @nospecialize(ub)) = _typevar(n, lb, ub)
-
-UnionAll(v::TypeVar, @nospecialize(t)) = ccall(:jl_type_unionall, Any, (Any, Any), v, t)
-
 Tuple{}() = ()
 
 struct VecElement{T}

base/broadcast.jl

@@ -727,8 +727,7 @@ end
             ci = U
         end
         if i == lr && Core.Compiler.isvarargtype(pi)
-            N = (Base.unwrap_unionall(pi)::DataType).parameters[2]
-            c[i] = Base.rewrap_unionall(Vararg{ci, N}, pi)
+            c[i] = isdefined(pi, :N) ? Vararg{ci, pi.N} : Vararg{ci}
         else
             c[i] = ci
         end

base/combinatorics.jl

@@ -307,7 +307,7 @@ julia> 2^2 * 3^3
 !!! compat "Julia 1.6"
     The method that accepts a tuple requires Julia 1.6 or later.
 """
-function nextprod(a::Union{Tuple{Vararg{<:Integer}},AbstractVector{<:Integer}}, x::Real)
+function nextprod(a::Union{Tuple{Vararg{Integer}},AbstractVector{<:Integer}}, x::Real)
     if x > typemax(Int)
         throw(ArgumentError("unsafe for x > typemax(Int), got $x"))
     end

base/compiler/abstractinterpretation.jl

@@ -1057,10 +1057,11 @@ function sp_type_rewrap(@nospecialize(T), linfo::MethodInstance, isreturn::Bool)
         spsig = linfo.def.sig
         if isa(spsig, UnionAll)
             if !isempty(linfo.sparam_vals)
-                env = pointer_from_objref(linfo.sparam_vals) + sizeof(Ptr{Cvoid})
-                T = ccall(:jl_instantiate_type_in_env, Any, (Any, Any, Ptr{Any}), T, spsig, env)
+                sparam_vals = Any[isa(v, Core.TypeofVararg) ? TypeVar(:N, Union{}, Any) :
+                                  v for v in  linfo.sparam_vals]
+                T = ccall(:jl_instantiate_type_in_env, Any, (Any, Any, Ptr{Any}), T, spsig, sparam_vals)
                 isref && isreturn && T === Any && return Bottom # catch invalid return Ref{T} where T = Any
-                for v in linfo.sparam_vals
+                for v in sparam_vals
                     if isa(v, TypeVar)
                         T = UnionAll(v, T)
                     end

base/compiler/inferencestate.jl

@@ -192,6 +192,8 @@ function sptypes_from_meth_instance(linfo::MethodInstance)
                     ty = UnionAll(tv, Type{tv})
                 end
             end
+        elseif isa(v, Core.TypeofVararg)
+            ty = Int
         else
             ty = Const(v)
         end

base/compiler/ssair/inlining.jl

@@ -738,7 +738,7 @@ function analyze_method!(match::MethodMatch, atypes::Vector{Any},
     # Bail out if any static parameters are left as TypeVar
     ok = true
     for i = 1:length(match.sparams)
-        isa(match.sparams[i], TypeVar) && return nothing
+        (isa(match.sparams[i], TypeVar) || isa(match.sparams[i], Core.TypeofVararg)) && return nothing
     end
 
     if !params.inlining

base/compiler/tfuncs.jl

@@ -447,9 +447,7 @@ function typebound_nothrow(b)
     b = widenconst(b)
     (b ⊑ TypeVar) && return true
     if isType(b)
-        b = unwrap_unionall(b.parameters[1])
-        b === Union{} && return true
-        return !isa(b, DataType) || b.name != _va_typename
+        return true
     end
     return false
 end
@@ -489,9 +487,8 @@ function typeof_concrete_vararg(t::DataType)
     for i = 1:np
         p = t.parameters[i]
         if i == np && isvarargtype(p)
-            pp = unwrap_unionall(p)
-            if isconcretetype(pp.parameters[1]) && pp.parameters[2] isa TypeVar
-                return rewrap_unionall(Type{Tuple{t.parameters[1:np-1]..., pp}}, p)
+            if isdefined(p, :T) && !isdefined(p, :N) && isconcretetype(p.T)
+                return Type{Tuple{t.parameters[1:np-1]..., Vararg{p.T, N}}} where N
             end
         elseif !isconcretetype(p)
             break
@@ -555,15 +552,21 @@ function typeassert_type_instance(@nospecialize(v), @nospecialize(t))
         widev = widenconst(v)
         if widev <: t
             return v
-        elseif typeintersect(widev, t) === Bottom
+        end
+        ti = typeintersect(widev, t)
+        if ti === Bottom
             return Bottom
         end
         @assert widev <: Tuple
         new_fields = Vector{Any}(undef, length(v.fields))
         for i = 1:length(new_fields)
-            new_fields[i] = typeassert_type_instance(v.fields[i], getfield_tfunc(t, Const(i)))
-            if new_fields[i] === Bottom
-                return Bottom
+            if isa(v.fields[i], Core.TypeofVararg)
+                new_fields[i] = v.fields[i]
+            else
+                new_fields[i] = typeassert_type_instance(v.fields[i], getfield_tfunc(t, Const(i)))
+                if new_fields[i] === Bottom
+                    return Bottom
+                end
             end
         end
         return tuple_tfunc(new_fields)
@@ -1130,10 +1133,10 @@ function apply_type_tfunc(@nospecialize(headtypetype), @nospecialize args...)
     elseif isconstType(headtypetype)
         headtype = headtypetype.parameters[1]
     else
-        return Type
+        return Any
     end
     if !isempty(args) && isvarargtype(args[end])
-        return Type
+        return isvarargtype(headtype) ? Core.TypeofVararg : Type
     end
     largs = length(args)
     if headtype === Union
@@ -1175,8 +1178,8 @@ function apply_type_tfunc(@nospecialize(headtypetype), @nospecialize args...)
         end
         return allconst ? Const(ty) : Type{ty}
     end
-    istuple = (headtype == Tuple)
-    if !istuple && !isa(headtype, UnionAll)
+    istuple = isa(headtype, Type) && (headtype == Tuple)
+    if !istuple && !isa(headtype, UnionAll) && !isvarargtype(headtype)
         return Union{}
     end
     uw = unwrap_unionall(headtype)
@@ -1193,7 +1196,8 @@ function apply_type_tfunc(@nospecialize(headtypetype), @nospecialize args...)
             aip1 = ai.parameters[1]
             canconst &= !has_free_typevars(aip1)
             push!(tparams, aip1)
-        elseif isa(ai, Const) && (isa(ai.val, Type) || isa(ai.val, TypeVar) || valid_tparam(ai.val))
+        elseif isa(ai, Const) && (isa(ai.val, Type) || isa(ai.val, TypeVar) ||
+                                  valid_tparam(ai.val) || (istuple && isa(ai.val, Core.TypeofVararg)))
             push!(tparams, ai.val)
         elseif isa(ai, PartialTypeVar)
             canconst = false
@@ -1259,11 +1263,11 @@ function apply_type_tfunc(@nospecialize(headtypetype), @nospecialize args...)
     catch ex
         # type instantiation might fail if one of the type parameters
         # doesn't match, which could happen if a type estimate is too coarse
-        return Type{<:headtype}
+        return isvarargtype(headtype) ? Core.TypeofVararg : Type{<:headtype}
     end
     !uncertain && canconst && return Const(appl)
-    if isvarargtype(headtype)
-        return Type
+    if isvarargtype(appl)
+        return Core.TypeofVararg
     end
     if istuple
         return Type{<:appl}
@@ -1315,7 +1319,7 @@ function tuple_tfunc(atypes::Vector{Any})
         x = atypes[i]
         # TODO ignore singleton Const (don't forget to update cache logic if you implement this)
         if !anyinfo
-            anyinfo = !isa(x, Type) || isType(x)
+            anyinfo = (!isa(x, Type) && !isvarargtype(x)) || isType(x)
         end
         if isa(x, Const)
             params[i] = typeof(x.val)

base/compiler/typelattice.jl

@@ -220,6 +220,7 @@ widenconst(c::PartialTypeVar) = TypeVar
 widenconst(t::PartialStruct) = t.typ
 widenconst(t::Type) = t
 widenconst(t::TypeVar) = t
+widenconst(t::Core.TypeofVararg) = t
 
 issubstate(a::VarState, b::VarState) = (a.typ ⊑ b.typ && a.undef <= b.undef)
 

base/compiler/typelimits.jl

@@ -46,6 +46,8 @@ function is_derived_type(@nospecialize(t), @nospecialize(c), mindepth::Int)
         # see if it is derived from the body
         # also handle the var here, since this construct bounds the mindepth to the smallest possible value
         return is_derived_type(t, c.var.ub, mindepth) || is_derived_type(t, c.body, mindepth)
+    elseif isa(c, Core.TypeofVararg)
+        return is_derived_type(t, unwrapva(c), mindepth)
     elseif isa(c, DataType)
         if mindepth > 0
             mindepth -= 1
@@ -85,7 +87,7 @@ function _limit_type_size(@nospecialize(t), @nospecialize(c), sources::SimpleVec
         return t # fast path: unparameterized are always simple
     else
         ut = unwrap_unionall(t)
-        if isa(ut, DataType) && ut.name !== _va_typename && isa(c, Type) && c !== Union{} && c <: t
+        if isa(ut, DataType) && isa(c, Type) && c !== Union{} && c <: t
             # TODO: need to check that the UnionAll bounds on t are limited enough too
             return t # t is already wider than the comparison in the type lattice
         elseif is_derived_type_from_any(ut, sources, depth)
@@ -118,19 +120,19 @@ function _limit_type_size(@nospecialize(t), @nospecialize(c), sources::SimpleVec
             b = _limit_type_size(t.b, c.b, sources, depth, allowed_tuplelen)
             return Union{a, b}
         end
+    elseif isa(t, Core.TypeofVararg)
+        isa(c, Core.TypeofVararg) || return Vararg
+        VaT = _limit_type_size(unwrapva(t), unwrapva(c), sources, depth + 1, 0)
+        if isdefined(t, :N) && (isa(t.N, TypeVar) || (isdefined(c, :N) && t.N === c.N))
+            return Vararg{VaT, t.N}
+        end
+        return Vararg{VaT}
     elseif isa(t, DataType)
         if isa(c, DataType)
             tP = t.parameters
             cP = c.parameters
             if t.name === c.name && !isempty(cP)
-                if isvarargtype(t)
-                    VaT = _limit_type_size(tP[1], cP[1], sources, depth + 1, 0)
-                    N = tP[2]
-                    if isa(N, TypeVar) || N === cP[2]
-                        return Vararg{VaT, N}
-                    end
-                    return Vararg{VaT}
-                elseif t.name === Tuple.name
+                if t.name === Tuple.name
                     # for covariant datatypes (Tuple),
                     # apply type-size limit element-wise
                     ltP = length(tP)
@@ -155,21 +157,17 @@ function _limit_type_size(@nospecialize(t), @nospecialize(c), sources::SimpleVec
                     end
                     return Tuple{Q...}
                 end
-            elseif isvarargtype(c)
-                # Tuple{Vararg{T}} --> Tuple{T} is OK
-                return _limit_type_size(t, cP[1], sources, depth, 0)
             end
+        elseif isa(c, Core.TypeofVararg)
+            # Tuple{Vararg{T}} --> Tuple{T} is OK
+            return _limit_type_size(t, c.T, sources, depth, 0)
         end
         if isType(t) # allow taking typeof as Type{...}, but ensure it doesn't start nesting
             tt = unwrap_unionall(t.parameters[1])
             if isa(tt, DataType) && !isType(tt)
                 is_derived_type_from_any(tt, sources, depth) && return t
             end
         end
-        if isvarargtype(t)
-            # never replace Vararg with non-Vararg
-            return Vararg
-        end
         if allowed_tuplelen < 1 && t.name === Tuple.name
             return Any
         end

base/compiler/typeutils.jl

@@ -61,6 +61,17 @@ function valid_tparam(@nospecialize(x))
     return isa(x, Symbol) || isbits(x)
 end
 
+function compatible_vatuple(a::DataType, b::DataType)
+    vaa = a.parameters[end]
+    vab = a.parameters[end]
+    if !(isa(vaa, Core.TypeofVararg) && isa(vab, Core.TypeofVararg))
+        return isa(vaa, Core.TypeofVararg) == isa(vab, Core.TypeofVararg)
+    end
+    (isdefined(vaa, :N) == isdefined(vab, :N)) || return false
+    !isdefined(vaa, :N) && return true
+    return vaa.N === vab.N
+end
+
 # return an upper-bound on type `a` with type `b` removed
 # such that `return <: a` && `Union{return, b} == Union{a, b}`
 function typesubtract(@nospecialize(a), @nospecialize(b), MAX_UNION_SPLITTING::Int)
@@ -80,19 +91,23 @@ function typesubtract(@nospecialize(a), @nospecialize(b), MAX_UNION_SPLITTING::I
                     ta = switchtupleunion(a)
                     return typesubtract(Union{ta...}, b, 0)
                 elseif b isa DataType
+                    if !compatible_vatuple(a, b)
+                        return a
+                    end
                     # if exactly one element is not bottom after calling typesubtract
                     # then the result is all of the elements as normal except that one
                     notbottom = fill(false, length(a.parameters))
                     for i = 1:length(notbottom)
-                        ap = a.parameters[i]
-                        bp = b.parameters[i]
+                        ap = unwrapva(a.parameters[i])
+                        bp = unwrapva(b.parameters[i])
                         notbottom[i] = !(ap <: bp && isnotbrokensubtype(ap, bp))
                     end
                     let i = findfirst(notbottom)
                         if i !== nothing && findnext(notbottom, i + 1) === nothing
                             ta = collect(a.parameters)
                             ap = a.parameters[i]
                             bp = b.parameters[i]
+                            (isa(ap, Core.TypeofVararg) || isa(bp, Core.TypeofVararg)) && return a
                             ta[i] = typesubtract(ap, bp, min(2, MAX_UNION_SPLITTING))
                             return Tuple{ta...}
                         end
@@ -196,6 +211,7 @@ function unioncomplexity(t::DataType)
     return c
 end
 unioncomplexity(u::UnionAll) = max(unioncomplexity(u.body), unioncomplexity(u.var.ub))
+unioncomplexity(t::Core.TypeofVararg) = isdefined(t, :T) ? unioncomplexity(t.T) : 0
 unioncomplexity(@nospecialize(x)) = 0
 
 function improvable_via_constant_propagation(@nospecialize(t))