fix type-predicate queries

expand the set of them to be more "complete",
and document them more fully

Author: vtjnash

NEWS.md

@@ -819,10 +819,10 @@ Deprecated or removed
     been deprecated due to inconsistency with linear algebra. Use `.+` and `.-` for these operations
     instead ([#22880], [#22932]).
 
-  * `isleaftype` is deprecated in favor of a simpler predicate `isconcrete`. Concrete types are
-    those that might equal `typeof(x)` for some `x`; `isleaftype` includes some types for which
-    this is not true. If you are certain you need the old behavior, it is temporarily available
-    as `Base._isleaftype` ([#17086]).
+  * `isleaftype` is deprecated in favor of the simpler predicates `isconcretetype` and `isdispatchtuple`.
+    Concrete types are those that might equal `typeof(x)` for some `x`;
+    `isleaftype` included some types for which this is not true. Those are now categorized more precisely
+    as "dispatch tuple types" and "!has_free_typevars" (not exported). ([#17086], [#25496])
 
   * `contains(eq, itr, item)` is deprecated in favor of `any` with a predicate ([#23716]).
 

base/abstractarray.jl

@@ -1985,7 +1985,7 @@ function hash(a::AbstractArray{T}, h::UInt) where T
             # to a range with more than two elements because more extreme values
             # cannot be represented. We must still hash the two first values as a
             # range since they can always be considered as such (in a wider type)
-            if isconcrete(T)
+            if isconcretetype(T)
                 try
                     step = x2 - x1
                 catch err

base/broadcast.jl

@@ -622,7 +622,7 @@ julia> string.(("one","two","three","four"), ": ", 1:4)
 const NonleafHandlingTypes = Union{DefaultArrayStyle,ArrayConflict,VectorStyle,MatrixStyle}
 
 @inline function broadcast(f, s::NonleafHandlingTypes, ::Type{ElType}, inds::Indices, As...) where ElType
-    if !Base._isleaftype(ElType)
+    if !Base.isconcretetype(ElType)
         return broadcast_nonleaf(f, s, ElType, inds, As...)
     end
     dest = broadcast_similar(f, s, ElType, inds, As...)

base/compiler/abstractinterpretation.jl

@@ -118,12 +118,10 @@ function abstract_call_method_with_const_args(@nospecialize(f), argtypes::Vector
     haveconst = false
     for i in 1:nargs
         a = argtypes[i]
-        if isa(a, Const) && !isdefined(typeof(a.val), :instance)
-            if !isleaftype(a.val) # alternately: !isa(a.val, DataType) || !isconstType(Type{a.val})
-                # have new information from argtypes that wasn't available from the signature
-                haveconst = true
-                break
-            end
+        if isa(a, Const) && !isdefined(typeof(a.val), :instance) && !(isa(a.val, Type) && issingletontype(a.val))
+            # have new information from argtypes that wasn't available from the signature
+            haveconst = true
+            break
         end
     end
     haveconst || return Any
@@ -378,11 +376,13 @@ end
 
 # do apply(af, fargs...), where af is a function value
 function abstract_apply(@nospecialize(aft), fargs::Vector{Any}, aargtypes::Vector{Any}, vtypes::VarTable, sv::InferenceState)
-    if !isa(aft, Const) && !isconstType(aft)
-        if !(isleaftype(aft) || aft <: Type) || (aft <: Builtin) || (aft <: IntrinsicFunction)
+    if !isa(aft, Const) && (!isType(aft) || has_free_typevars(aft))
+        if !isconcretetype(aft) || (aft <: Builtin)
+            # non-constant function of unknown type: bail now,
+            # since it seems unlikely that abstract_call will be able to do any better after splitting
+            # this also ensures we don't call abstract_call_gf_by_type below on an IntrinsicFunction or Builtin
             return Any
         end
-        # non-constant function, but type is known
     end
     res = Union{}
     nargs = length(fargs)
@@ -394,7 +394,7 @@ function abstract_apply(@nospecialize(aft), fargs::Vector{Any}, aargtypes::Vecto
         for ti in (splitunions ? uniontypes(aargtypes[i]) : Any[aargtypes[i]])
             cti = precise_container_type(fargs[i], ti, vtypes, sv)
             for ct in ctypes
-                if !isempty(ct) && isvarargtype(ct[end])
+                if isvarargtype(ct[end])
                     tail = tuple_tail_elem(unwrapva(ct[end]), cti)
                     push!(ctypes´, push!(ct[1:(end - 1)], tail))
                 else
@@ -672,23 +672,22 @@ function abstract_eval_call(e::Expr, vtypes::VarTable, sv::InferenceState)
     ft = argtypes[1]
     if isa(ft, Const)
         f = ft.val
+    elseif isconstType(ft)
+        f = ft.parameters[1]
+    elseif isa(ft, DataType) && isdefined(ft, :instance)
+        f = ft.instance
     else
-        if isType(ft) && isleaftype(ft.parameters[1])
-            f = ft.parameters[1]
-        elseif isleaftype(ft) && isdefined(ft, :instance)
-            f = ft.instance
-        else
-            for i = 2:(length(argtypes)-1)
-                if isvarargtype(argtypes[i])
-                    return Any
-                end
-            end
-            # non-constant function, but type is known
-            if (isleaftype(ft) || ft <: Type) && !(ft <: Builtin) && !(ft <: IntrinsicFunction)
-                return abstract_call_gf_by_type(nothing, argtypes, argtypes_to_type(argtypes), sv)
+        for i = 2:(length(argtypes) - 1)
+            if isvarargtype(argtypes[i])
+                return Any
             end
+        end
+        # non-constant function, but the number of arguments is known
+        # and the ft is not a Builtin or IntrinsicFunction
+        if typeintersect(widenconst(ft), Builtin) != Union{}
             return Any
         end
+        return abstract_call_gf_by_type(nothing, argtypes, argtypes_to_type(argtypes), sv)
     end
     return abstract_call(f, e.args, argtypes, vtypes, sv)
 end

base/compiler/compiler.jl

@@ -76,8 +76,6 @@ include("docs/core.jl")
 inlining_enabled() = (JLOptions().can_inline == 1)
 coverage_enabled() = (JLOptions().code_coverage != 0)
 
-const isleaftype = _isleaftype
-
 include("compiler/utilities.jl")
 include("compiler/validation.jl")
 

base/compiler/inferencestate.jl

@@ -84,7 +84,7 @@ mutable struct InferenceState
             at = (i > nargs) ? Bottom : argtypes[i]
             if !toplevel && linfo.def.isva && i == nargs
                 if !(at == Tuple) # would just be a no-op
-                    vararg_type_container = limit_tuple_depth(params, unwrap_unionall(at)) # TODO: should be limiting tuple depth much earlier than here
+                    vararg_type_container = unwrap_unionall(at)
                     vararg_type = tuple_tfunc(vararg_type_container) # returns a Const object, if applicable
                     at = rewrap(vararg_type, linfo.specTypes)
                 end

base/compiler/optimize.jl

@@ -816,8 +816,8 @@ function effect_free(@nospecialize(e), src::CodeInfo, mod::Module, allow_volatil
             return true
         end
         if head === :static_parameter
-            # if we aren't certain about the type, it might be an UndefVarError at runtime
-            return (isa(e.typ, DataType) && isleaftype(e.typ)) || isa(e.typ, Const)
+            # if we aren't certain enough about the type, it might be an UndefVarError at runtime
+            return isa(e.typ, Const) || issingletontype(widenconst(e.typ))
         end
         if e.typ === Bottom
             return false
@@ -830,17 +830,17 @@ function effect_free(@nospecialize(e), src::CodeInfo, mod::Module, allow_volatil
                         return false
                     elseif is_known_call(e, getfield, src, mod)
                         nargs = length(ea)
-                        (2 < nargs < 5) || return false
+                        (3 < nargs < 4) || return false
                         et = exprtype(e, src, mod)
-                        if !isa(et, Const) && !(isType(et) && isleaftype(et))
+                        # TODO: check ninitialized
+                        if !isa(et, Const) && !isconstType(et)
                             # first argument must be immutable to ensure e is affect_free
                             a = ea[2]
-                            typ = widenconst(exprtype(a, src, mod))
-                            if isconstType(typ)
-                                if Const(:uid) ⊑ exprtype(ea[3], src, mod)
-                                    return false    # DataType uid field can change
-                                end
-                            elseif typ !== SimpleVector && (!isa(typ, DataType) || typ.mutable || typ.abstract)
+                            typ = unwrap_unionall(widenconst(exprtype(a, src, mod)))
+                            if isType(typ)
+                                # all fields of subtypes of Type are effect-free
+                                # (including the non-inferrable uid field)
+                            elseif !isa(typ, DataType) || typ.abstract || (typ.mutable && length(typ.types) > 0)
                                 return false
                             end
                         end
@@ -863,7 +863,8 @@ function effect_free(@nospecialize(e), src::CodeInfo, mod::Module, allow_volatil
             # `Expr(:new)` of unknown type could raise arbitrary TypeError.
             typ, isexact = instanceof_tfunc(typ)
             isexact || return false
-            (isleaftype(typ) && !iskindtype(typ)) || return false
+            isconcretetype(typ) || return false
+            !iskindtype(typ) || return false
             typ = typ::DataType
             if !allow_volatile && typ.mutable
                 return false
@@ -1075,11 +1076,11 @@ function inlineable(@nospecialize(f), @nospecialize(ft), e::Expr, atypes::Vector
                     sv::OptimizationState)
     argexprs = e.args
 
-    if (f === typeassert || ft ⊑ typeof(typeassert)) && length(atypes)==3
+    if (f === typeassert || ft ⊑ typeof(typeassert)) && length(atypes) == 3
         # typeassert(x::S, T) => x, when S<:T
         a3 = atypes[3]
-        if (isType(a3) && isleaftype(a3) && atypes[2] ⊑ a3.parameters[1]) ||
-            (isa(a3,Const) && isa(a3.val,Type) && atypes[2] ⊑ a3.val)
+        if (isType(a3) && !has_free_typevars(a3) && atypes[2] ⊑ a3.parameters[1]) ||
+            (isa(a3, Const) && isa(a3.val, Type) && atypes[2] ⊑ a3.val)
             return (argexprs[2], ())
         end
     end
@@ -1108,7 +1109,9 @@ function inlineable(@nospecialize(f), @nospecialize(ft), e::Expr, atypes::Vector
     if f === Core.invoke && length(atypes) >= 3
         ft = widenconst(atypes[2])
         invoke_tt = widenconst(atypes[3])
-        if !isleaftype(ft) || !isleaftype(invoke_tt) || !isType(invoke_tt)
+        if !(isconcretetype(ft) || ft <: Type) || !isType(invoke_tt) ||
+                has_free_typevars(invoke_tt) || has_free_typevars(ft) || (ft <: Builtin)
+            # TODO: this is really aggressive at preventing inlining of closures. maybe drop `isconcretetype` requirement?
             return NOT_FOUND
         end
         if !(isa(invoke_tt.parameters[1], Type) &&
@@ -1271,12 +1274,10 @@ function inlineable(@nospecialize(f), @nospecialize(ft), e::Expr, atypes::Vector
     haveconst = false
     for i in 1:length(atypes)
         a = atypes[i]
-        if isa(a, Const) && !isdefined(typeof(a.val), :instance)
-            if !isleaftype(a.val) # alternately: !isa(a.val, DataType) || !isconstType(Type{a.val})
-                # have new information from argtypes that wasn't available from the signature
-                haveconst = true
-                break
-            end
+        if isa(a, Const) && !isdefined(typeof(a.val), :instance) && !(isa(a.val, Type) && issingletontype(a.val))
+            # have new information from argtypes that wasn't available from the signature
+            haveconst = true
+            break
         end
     end
     if haveconst
@@ -1538,8 +1539,9 @@ end
 
 # saturating sum (inputs are nonnegative), prevents overflow with typemax(Int) below
 plus_saturate(x, y) = max(x, y, x+y)
+
 # known return type
-isknowntype(T) = (T == Union{}) || isleaftype(T)
+isknowntype(@nospecialize T) = (T == Union{}) || isconcretetype(T)
 
 function statement_cost(ex::Expr, line::Int, src::CodeInfo, mod::Module, params::Params)
     head = ex.head
@@ -1770,7 +1772,8 @@ function inline_call(e::Expr, sv::OptimizationState, stmts::Vector{Any}, boundsc
         ft = Bool
     else
         f = nothing
-        if !( isleaftype(ft) || ft<:Type )
+        if !(isconcretetype(ft) || (widenconst(ft) <: Type)) || has_free_typevars(ft)
+            # TODO: this is really aggressive at preventing inlining of closures. maybe drop `isconcretetype` requirement?
             return e
         end
     end
@@ -1927,7 +1930,8 @@ function inline_call(e::Expr, sv::OptimizationState, stmts::Vector{Any}, boundsc
                 ft = Bool
             else
                 f = nothing
-                if !( isleaftype(ft) || ft<:Type )
+                if !(isconcretetype(ft) || (widenconst(ft) <: Type)) || has_free_typevars(ft)
+                    # TODO: this is really aggressive at preventing inlining of closures. maybe drop `isconcretetype` requirement?
                     return e
                 end
             end
@@ -2745,15 +2749,15 @@ end
 
 function split_disjoint_assign!(ctx::AllocOptContext, info, key)
     key.second && return false
-    isleaftype(ctx.sv.src.slottypes[key.first]) && return false
+    isdispatchelem(widenconst(ctx.sv.src.slottypes[key.first])) && return false # no splitting can be necessary
     alltypes = ObjectIdDict()
     ndefs = length(info.defs)
     deftypes = Vector{Any}(uninitialized, ndefs)
     for i in 1:ndefs
         def = info.defs[i]
         defex = (def.assign::Expr).args[2]
         rhstyp = widenconst(exprtype(defex, ctx.sv.src, ctx.sv.mod))
-        isleaftype(rhstyp) || return false
+        isdispatchelem(rhstyp) || return false
         alltypes[rhstyp] = nothing
         deftypes[i] = rhstyp
     end
@@ -2763,7 +2767,7 @@ function split_disjoint_assign!(ctx::AllocOptContext, info, key)
         slot = usex.args[use.exidx]
         if isa(slot, TypedSlot)
             usetyp = widenconst(slot.typ)
-            if isleaftype(usetyp)
+            if isdispatchelem(usetyp)
                 alltypes[usetyp] = nothing
                 continue
             end
@@ -2818,7 +2822,7 @@ function split_disjoint_assign!(ctx::AllocOptContext, info, key)
         slot = usex.args[use.exidx]
         if isa(slot, TypedSlot)
             usetyp = widenconst(slot.typ)
-            if isleaftype(usetyp)
+            if isdispatchelem(usetyp)
                 usetyp = widenconst(slot.typ)
                 new_slot = alltypes[usetyp]
                 if !isa(new_slot, SlotNumber)
@@ -2988,7 +2992,7 @@ function split_struct_alloc!(ctx::AllocOptContext, info, key)
             elseif defex.head === :new
                 typ = widenconst(exprtype(defex, ctx.sv.src, ctx.sv.mod))
                 # typ <: Tuple shouldn't happen but just in case someone generated invalid AST
-                if !isa(typ, DataType) || !isleaftype(typ) || typ <: Tuple
+                if !isa(typ, DataType) || !isdispatchelem(typ) || typ <: Tuple
                     return false
                 end
                 si = structinfo_new(ctx, defex, typ)

base/compiler/params.jl

@@ -28,7 +28,6 @@ struct Params
 
     # parameters limiting large types
     MAX_TUPLETYPE_LEN::Int
-    MAX_TUPLE_DEPTH::Int
 
     # when attempting to inlining _apply, abort the optimization if the tuple
     # contains more than this many elements
@@ -42,13 +41,12 @@ struct Params
                     inline_tupleret_bonus::Int = 400,
                     max_methods::Int = 4,
                     tupletype_len::Int = 15,
-                    tuple_depth::Int = 4,
                     tuple_splat::Int = 16,
                     union_splitting::Int = 4,
                     apply_union_enum::Int = 8)
         return new(Vector{InferenceResult}(),
                    world, inlining, true, false, inline_cost_threshold, inline_nonleaf_penalty,
                    inline_tupleret_bonus, max_methods, union_splitting, apply_union_enum,
-                   tupletype_len, tuple_depth, tuple_splat)
+                   tupletype_len, tuple_splat)
     end
 end