support type annotations on global variables

NEWS.md

@@ -20,6 +20,7 @@ New language features
  them after construction, providing for greater clarity and optimization
  ability of these objects ([#43305]).
 * Empty n-dimensional arrays can now be created using multiple semicolons inside square brackets, i.e. `[;;;]` creates a 0×0×0 `Array`. ([#41618])
+* Type annotations can now be added to global variables to make accessing the variable type stable. ([#43671])
 
 Language changes
 ----------------
@@ -35,6 +36,10 @@ Language changes
  to mitigate the ["trojan source"](https://www.trojansource.codes) vulnerability ([#42918]).
 * `Base.ifelse` is now defined as a generic function rather than a builtin one, allowing packages to
  extend its definition ([#37343]).
+* Every assignment to a global variable now first goes through a call to `convert(Any, x)` (or `convert(T, x)`
+ respectively if a type `T` has already been declared for the global). This means great care should be taken
+ to ensure the invariant `convert(Any, x) === x` always holds, as this change could otherwise lead to
+ unexpected behavior. ([#43671])
 
 Compiler/Runtime improvements
 -----------------------------

base/Base.jl

@@ -57,7 +57,7 @@ modifyproperty!(x, f::Symbol, op, v, order::Symbol=:notatomic) =
 replaceproperty!(x, f::Symbol, expected, desired, success_order::Symbol=:notatomic, fail_order::Symbol=success_order) =
  (@inline; Core.replacefield!(x, f, expected, convert(fieldtype(typeof(x), f), desired), success_order, fail_order))
 
-
+convert(::Type{Any}, Core.@nospecialize x) = x
 include("coreio.jl")
 
 eval(x) = Core.eval(Base, x)

base/compiler/abstractinterpretation.jl

@@ -1805,7 +1805,7 @@ function abstract_eval_global(M::Module, s::Symbol)
  if isdefined(M,s) && isconst(M,s)
  return Const(getfield(M,s))
  end
- return Any
+ return Core.get_binding_type(M, s)
 end
 
 function abstract_eval_ssavalue(s::SSAValue, src::CodeInfo)

base/compiler/optimize.jl

@@ -156,7 +156,7 @@ const _PURE_BUILTINS = Any[tuple, svec, ===, typeof, nfields]
 const _PURE_OR_ERROR_BUILTINS = [
  fieldtype, apply_type, isa, UnionAll,
  getfield, arrayref, const_arrayref, arraysize, isdefined, Core.sizeof,
- Core.kwfunc, Core.ifelse, Core._typevar, (<:)
+ Core.kwfunc, Core.ifelse, Core._typevar, (<:), Core.get_binding_type,
 ]
 
 const TOP_TUPLE = GlobalRef(Core, :tuple)

base/compiler/tfuncs.jl

@@ -1695,6 +1695,9 @@ function _builtin_nothrow(@nospecialize(f), argtypes::Array{Any,1}, @nospecializ
  return true
  end
  return false
+ elseif f === Core.get_binding_type
+ length(argtypes) == 2 || return false
+ return argtypes[1] ⊑ Module && argtypes[2] ⊑ Symbol
  end
  return false
 end
@@ -1898,4 +1901,13 @@ function typename_static(@nospecialize(t))
  return isType(t) ? _typename(t.parameters[1]) : Core.TypeName
 end
 
+function get_binding_type_tfunc(@nospecialize(M), @nospecialize(s))
+ if M isa Const && widenconst(M) === Module &&
+ s isa Const && widenconst(s) === Symbol
+ return Const(Core.get_binding_type(M.val, s.val))
+ end
+ return Type
+end
+add_tfunc(Core.get_binding_type, 2, 2, get_binding_type_tfunc, 0)
+
 @specialize

base/essentials.jl

@@ -161,6 +161,11 @@ end
 
 function iterate end
 
+if @isdefined(convert)
+else
+ convert(::Type{Any}, @nospecialize x) = x
+end
+
 """
  convert(T, x)
 
@@ -211,7 +216,6 @@ See also: [`round`](@ref), [`trunc`](@ref), [`oftype`](@ref), [`reinterpret`](@r
 function convert end
 
 convert(::Type{Union{}}, @nospecialize x) = throw(MethodError(convert, (Union{}, x)))
-convert(::Type{Any}, @nospecialize x) = x
 convert(::Type{T}, x::T) where {T} = x
 convert(::Type{Type}, x::Type) = x # the ssair optimizer is strongly dependent on this method existing to avoid over-specialization
  # in the absence of inlining-enabled

doc/src/manual/performance-tips.md

@@ -11,9 +11,9 @@ The use of functions is not only important for performance: functions are more r
 
 The functions should take arguments, instead of operating directly on global variables, see the next point.
 
-## Avoid global variables
+## Avoid untyped global variables
 
-A global variable might have its value, and therefore possibly its type, changed at any point. This makes
+An untyped global variable might have its value, and therefore possibly its type, changed at any point. This makes
 it difficult for the compiler to optimize code using global variables. This also applies to type-valued variables,
 i.e. type aliases on the global level. Variables should be local, or passed as arguments to functions, whenever possible.
 
@@ -24,7 +24,9 @@ performance:
 const DEFAULT_VAL = 0
 ```
 
-Uses of non-constant globals can be optimized by annotating their types at the point of use:
+If a global is known to always be of the same type, [the type should be annotated](@ref man-typed-globals).
+
+Uses of untyped globals can be optimized by annotating their types at the point of use:
 
 ```julia
 global x = rand(1000)

doc/src/manual/variables-and-scoping.md

@@ -799,3 +799,67 @@ WARNING: redefinition of constant x. This may fail, cause incorrect answers, or
 julia> f()
 1
 ```
+
+## [Typed Globals](@id man-typed-globals)
+
+!!! compat "Julia 1.8"
+ Support for typed globals was added in Julia 1.8
+
+Similar to being declared as constants, global bindings can also be declared to always be of a
+constant type. This can either be done without assigning an actual value using the syntax
+`global x::T` or upon assignment as `x::T = 123`.
+
+```jldoctest
+julia> x::Float64 = 2.718
+2.718
+
+julia> f() = x
+f (generic function with 1 method)
+
+julia> Base.return_types(f)
+1-element Vector{Any}:
+ Float64
+```
+
+For any assignment to a global, Julia will first try to convert it to the appropriate type using
+[`convert`](@ref):
+
+```jldoctest
+julia> global y::Int
+
+julia> y = 1.0
+1.0
+
+julia> y
+1
+
+julia> y = 3.14
+ERROR: InexactError: Int64(3.14)
+Stacktrace:
+[...]
+```
+
+The type does not need to be concrete, but annotations with abstract types typically have little
+performance benefit.
+
+As for constants, the type of a global is not typically meant to be redefined. In interactive usage,
+changing the annotation to a narrower type (the new type is a subtype of the previous type) is
+allowed, but can result in unexpected behavior in code that has already been compiled:
+
+```jldoctest
+julia> global a::Integer
+
+julia> g() = a
+g (generic function with 1 method)
+
+julia> Base.return_types(g)
+1-element Vector{Any}:
+ Integer
+
+julia> global a::Int
+WARNING: redefinition of type of a. This may fail, cause incorrect answers, or produce other errors.
+
+julia> Base.return_types(g)
+1-element Vector{Any}:
+ Integer
+```

src/builtin_proto.h

@@ -65,6 +65,8 @@ JL_CALLABLE(jl_f__abstracttype);
 JL_CALLABLE(jl_f__primitivetype);
 JL_CALLABLE(jl_f__setsuper);
 JL_CALLABLE(jl_f__equiv_typedef);
+JL_CALLABLE(jl_f_get_binding_type);
+JL_CALLABLE(jl_f_set_binding_type);
 
 #ifdef __cplusplus
 }

src/builtins.c

@@ -1655,6 +1655,46 @@ JL_CALLABLE(jl_f__equiv_typedef)
  return equiv_type(args[0], args[1]) ? jl_true : jl_false;
 }
 
+JL_CALLABLE(jl_f_get_binding_type)
+{
+ JL_NARGS(get_binding_type, 2, 2);
+ JL_TYPECHK(get_binding_type, module, args[0]);
+ JL_TYPECHK(get_binding_type, symbol, args[1]);
+ jl_value_t* ty = jl_binding_type((jl_module_t*)args[0], (jl_sym_t*)args[1]);
+ return ty;
+}
+
+JL_CALLABLE(jl_f_set_binding_type)
+{
+ JL_NARGS(set_binding_type!, 3, 3);
+ JL_TYPECHK(set_binding_type!, module, args[0]);
+ JL_TYPECHK(set_binding_type!, symbol, args[1]);
+ jl_value_t *ty = args[2];
+ JL_TYPECHK(set_binding_type!, type, ty);
+ jl_binding_t *b = jl_get_binding_wr((jl_module_t*)args[0], (jl_sym_t*)args[1], 1);
+ if (b->constp && ty != (jl_value_t*)jl_any_type) {
+ jl_errorf("cannot set type for constant %s", jl_symbol_name(b->name));
+ }
+ jl_value_t *val = jl_atomic_load_relaxed(&b->value);
+ if (val && !jl_isa(val, ty)) {
+ jl_errorf("cannot set type for global %s. It already has a value of a different type.",
+ jl_symbol_name(b->name));
+ }
+ if (b->ty && ty != b->ty) {
+ if (jl_subtype(ty, b->ty)) {
+ jl_safe_printf("WARNING: redefinition of type of %s. This may fail, cause incorrect answers, or produce other errors.\n",
+ jl_symbol_name(b->name));
+ }
+ else {
+ jl_errorf("invalid redefinition of type of %s", jl_symbol_name(b->name));
+ }
+ }
+ if (ty != (jl_value_t*)jl_any_type) {
+ b->ty = ty;
+ }
+ return jl_nothing;
+}
+
 // IntrinsicFunctions ---------------------------------------------------------
 
 static void (*runtime_fp[num_intrinsics])(void);
@@ -1833,6 +1873,8 @@ void jl_init_primitives(void) JL_GC_DISABLED
  add_builtin_func("_setsuper!", jl_f__setsuper);
  jl_builtin__typebody = add_builtin_func("_typebody!", jl_f__typebody);
  add_builtin_func("_equiv_typedef", jl_f__equiv_typedef);
+ add_builtin_func("get_binding_type", jl_f_get_binding_type);
+ add_builtin_func("set_binding_type!", jl_f_set_binding_type);
 
  // builtin types
  add_builtin("Any", (jl_value_t*)jl_any_type);