support type annotations on global variables

This is an initial proposal for supporting typed globals based on #43455.
The syntax for this is either `global x::T` or just doing `x::T = 1` in
global scope. It is even supported to add these annotations to globals
from inside functions. The type declaration will then be applied when
the method is defined and inside the function body conversion to the
specified type will happen automatically, similar to type annotations
for local variables.

This conversion will not be applied if the assignment is not done inside
the same scope of the type annotation however. This could potentially be
supported as well, but the problem is that this would mean any
assignment to a global variable - typed or not - would need to go
through a call to `convert` first, since lowering can't know if a given
binding already has a type annotation. This probably wouldn't be a good
thing for latency and sysimage size and would add an invalidation risk.

It is allowed to refine a type annotation, but there will be a warning.
Widening the type or changing it to any other type that is not a subtype
of the previous one is an error, since that could cause cached compiled
code to become invalid.

replaces #43455
closes #964

Author: simeonschaub

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 ccall(:jl_get_binding_type, Any, (Any, Any), M, s)
 end
 
 function abstract_eval_ssavalue(s::SSAValue, src::CodeInfo)

src/builtin_proto.h

@@ -65,6 +65,7 @@ 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__set_typeof);
 
 #ifdef __cplusplus
 }

src/builtins.c

@@ -1655,6 +1655,36 @@ JL_CALLABLE(jl_f__equiv_typedef)
     return equiv_type(args[0], args[1]) ? jl_true : jl_false;
 }
 
+JL_CALLABLE(jl_f__set_typeof)
+{
+    JL_NARGS(_set_typeof!, 3, 3);
+    JL_TYPECHK(_set_typeof!, module, args[0]);
+    JL_TYPECHK(_set_typeof!, symbol, args[1]);
+    jl_value_t *ty = args[2];
+    JL_TYPECHK(_set_typeof!, 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));
+    }
+    if (b->value && !jl_isa(b->value, 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 +1863,7 @@ 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("_set_typeof!", jl_f__set_typeof);
 
     // builtin types
     add_builtin("Any", (jl_value_t*)jl_any_type);

src/julia-syntax.scm

@@ -260,6 +260,12 @@
   (or (overlay? e)
       (sym-ref? e)))
 
+(define (binding-to-globalref e)
+  (and (nodot-sym-ref? e)
+       (let ((mod (if (globalref? e) (cadr e) '(thismodule)))
+             (sym (if (symbol? e) e (last e))))
+         `(globalref ,mod ,sym))))
+
 ;; convert final (... x) to (curly Vararg x)
 (define (dots->vararg a)
   (if (null? a) a
@@ -3383,7 +3389,7 @@ f(x) = yt(x)
 ;; declared types.
 ;; when doing this, the original value needs to be preserved, to
 ;; ensure the expression `a=b` always returns exactly `b`.
-(define (convert-assignment var rhs0 fname lam interp opaq)
+(define (convert-assignment var rhs0 fname lam interp opaq globals)
   (cond
     ((symbol? var)
      (let* ((vi (assq var (car  (lam:vinfo lam))))
@@ -3393,32 +3399,39 @@ f(x) = yt(x)
                      '(core Any)))
             (closed (and cv (vinfo:asgn cv) (vinfo:capt cv)))
             (capt   (and vi (vinfo:asgn vi) (vinfo:capt vi))))
-       (if (and (not closed) (not capt) (equal? vt '(core Any)))
-           `(= ,var ,rhs0)
-           (let* ((rhs1 (if (or (simple-atom? rhs0)
-                                (equal? rhs0 '(the_exception)))
-                            rhs0
-                            (make-ssavalue)))
-                  (rhs  (if (equal? vt '(core Any))
-                            rhs1
-                            (convert-for-type-decl rhs1 (cl-convert vt fname lam #f #f #f interp opaq))))
-                  (ex (cond (closed `(call (core setfield!)
-                                           ,(if interp
-                                                `($ ,var)
-                                                (capt-var-access var fname opaq))
-                                           (inert contents)
-                                           ,rhs))
-                            (capt `(call (core setfield!) ,var (inert contents) ,rhs))
-                            (else `(= ,var ,rhs)))))
-             (if (eq? rhs1 rhs0)
-                 `(block ,ex ,rhs0)
-                 `(block (= ,rhs1 ,rhs0)
-                         ,ex
-                         ,rhs1))))))
-     ((and (pair? var) (or (eq? (car var) 'outerref)
-                           (eq? (car var) 'globalref)))
-
-      `(= ,var ,rhs0))
+       (let* ((rhs1 (if (or (simple-atom? rhs0)
+                            (equal? rhs0 '(the_exception)))
+                        rhs0
+                        (make-ssavalue)))
+              (rhs  (if (equal? vt '(core Any))
+                        (if (and (not closed) (not capt))
+                            (convert-for-type-decl rhs1 (get globals (binding-to-globalref var) '(core Any)))
+                            rhs1)
+                        (convert-for-type-decl rhs1 (cl-convert vt fname lam #f #f #f interp opaq globals))))
+              (ex (cond (closed `(call (core setfield!)
+                                       ,(if interp
+                                            `($ ,var)
+                                            (capt-var-access var fname opaq))
+                                       (inert contents)
+                                       ,rhs))
+                        (capt `(call (core setfield!) ,var (inert contents) ,rhs))
+                        (else `(= ,var ,rhs)))))
+        (if (eq? rhs1 rhs0)
+            `(block ,ex ,rhs0)
+            `(block (= ,rhs1 ,rhs0)
+                    ,ex
+                    ,rhs1)))))
+     ((or (outerref? var) (globalref? var))
+      (let* ((rhs1 (if (or (simple-atom? rhs0)
+                           (equal? rhs0 '(the_exception)))
+                       rhs0
+                       (make-ssavalue)))
+             (ex   `(= ,var ,(convert-for-type-decl rhs1 (get globals (binding-to-globalref var) '(core Any))))))
+        (if (eq? rhs1 rhs0)
+            `(block ,ex ,rhs0)
+            `(block (= ,rhs1 ,rhs0)
+                    ,ex
+                    ,rhs1))))
      ((ssavalue? var)
       `(= ,var ,rhs0))
      (else
@@ -3678,17 +3691,17 @@ f(x) = yt(x)
 (define (toplevel-preserving? e)
   (and (pair? e) (memq (car e) '(if elseif block trycatch tryfinally trycatchelse))))
 
-(define (map-cl-convert exprs fname lam namemap defined toplevel interp opaq)
+(define (map-cl-convert exprs fname lam namemap defined toplevel interp opaq (globals (table)))
   (if toplevel
       (map (lambda (x)
              (let ((tl (lift-toplevel (cl-convert x fname lam namemap defined
                                                   (and toplevel (toplevel-preserving? x))
-                                                  interp opaq))))
+                                                  interp opaq globals))))
                (if (null? (cdr tl))
                    (car tl)
                    `(block ,@(cdr tl) ,(car tl)))))
            exprs)
-      (map (lambda (x) (cl-convert x fname lam namemap defined #f interp opaq)) exprs)))
+      (map (lambda (x) (cl-convert x fname lam namemap defined #f interp opaq globals)) exprs)))
 
 (define (prepare-lambda! lam)
   ;; mark all non-arguments as assigned, since locals that are never assigned
@@ -3697,11 +3710,11 @@ f(x) = yt(x)
             (list-tail (car (lam:vinfo lam)) (length (lam:args lam))))
   (lambda-optimize-vars! lam))
 
-(define (cl-convert e fname lam namemap defined toplevel interp opaq)
+(define (cl-convert e fname lam namemap defined toplevel interp opaq (globals (table)))
   (if (and (not lam)
            (not (and (pair? e) (memq (car e) '(lambda method macro opaque_closure)))))
       (if (atom? e) e
-          (cons (car e) (map-cl-convert (cdr e) fname lam namemap defined toplevel interp opaq)))
+          (cons (car e) (map-cl-convert (cdr e) fname lam namemap defined toplevel interp opaq globals)))
       (cond
        ((symbol? e)
         (define (new-undef-var name)
@@ -3720,7 +3733,7 @@ f(x) = yt(x)
                  (val (if (equal? typ '(core Any))
                           val
                           `(call (core typeassert) ,val
-                                 ,(cl-convert typ fname lam namemap defined toplevel interp opaq)))))
+                                 ,(cl-convert typ fname lam namemap defined toplevel interp opaq globals)))))
             `(block
                ,@(if (eq? box access) '() `((= ,access ,box)))
                ,undefcheck
@@ -3752,8 +3765,8 @@ f(x) = yt(x)
            e)
           ((=)
            (let ((var (cadr e))
-                 (rhs (cl-convert (caddr e) fname lam namemap defined toplevel interp opaq)))
-             (convert-assignment var rhs fname lam interp opaq)))
+                 (rhs (cl-convert (caddr e) fname lam namemap defined toplevel interp opaq globals)))
+             (convert-assignment var rhs fname lam interp opaq globals)))
           ((local-def) ;; make new Box for local declaration of defined variable
            (let ((vi (assq (cadr e) (car (lam:vinfo lam)))))
              (if (and vi (vinfo:asgn vi) (vinfo:capt vi))
@@ -3819,7 +3832,7 @@ f(x) = yt(x)
                   (sp-inits (if (or short (not (eq? (car sig) 'block)))
                                 '()
                                 (map-cl-convert (butlast (cdr sig))
-                                                fname lam namemap defined toplevel interp opaq)))
+                                                fname lam namemap defined toplevel interp opaq globals)))
                   (sig      (and sig (if (eq? (car sig) 'block)
                                          (last sig)
                                          sig))))
@@ -3846,7 +3859,7 @@ f(x) = yt(x)
                                           ;; anonymous functions with keyword args generate global
                                           ;; functions that refer to the type of a local function
                                           (rename-sig-types sig namemap)
-                                          fname lam namemap defined toplevel interp opaq)
+                                          fname lam namemap defined toplevel interp opaq globals)
                                   ,(let ((body (add-box-inits-to-body
                                                 lam2
                                                 (cl-convert (cadddr lam2) 'anon lam2 (table) (table) #f interp opaq))))
@@ -3860,7 +3873,7 @@ f(x) = yt(x)
                                (newlam    (compact-and-renumber (linearize (car exprs)) 'none 0)))
                           `(toplevel-butfirst
                             (block ,@sp-inits
-                                   (method ,name ,(cl-convert sig fname lam namemap defined toplevel interp opaq)
+                                   (method ,name ,(cl-convert sig fname lam namemap defined toplevel interp opaq globals)
                                            ,(julia-bq-macro newlam)))
                             ,@top-stmts))))
 
@@ -3963,7 +3976,7 @@ f(x) = yt(x)
                                (append (map (lambda (gs tvar)
                                               (make-assignment gs `(call (core TypeVar) ',tvar (core Any))))
                                             closure-param-syms closure-param-names)
-                                       `((method #f ,(cl-convert arg-defs fname lam namemap defined toplevel interp opaq)
+                                       `((method #f ,(cl-convert arg-defs fname lam namemap defined toplevel interp opaq globals)
                                                  ,(convert-lambda lam2
                                                                   (if iskw
                                                                       (caddr (lam:args lam2))
@@ -4002,7 +4015,7 @@ f(x) = yt(x)
                        (begin
                          (put! defined name #t)
                          `(toplevel-butfirst
-                           ,(convert-assignment name mk-closure fname lam interp opaq)
+                           ,(convert-assignment name mk-closure fname lam interp opaq globals)
                            ,@typedef
                            ,@(map (lambda (v) `(moved-local ,v)) moved-vars)
                            ,@sp-inits
@@ -4015,30 +4028,42 @@ f(x) = yt(x)
                                        (table)
                                        (table)
                                        (null? (cadr e)) ;; only toplevel thunks have 0 args
-                                       interp opaq)))
+                                       interp opaq globals)))
              `(lambda ,(cadr e)
                 (,(clear-capture-bits (car (lam:vinfo e)))
                  () ,@(cddr (lam:vinfo e)))
                 (block ,@body))))
           ;; remaining `::` expressions are type assertions
           ((|::|)
-           (cl-convert `(call (core typeassert) ,@(cdr e)) fname lam namemap defined toplevel interp opaq))
+           (cl-convert `(call (core typeassert) ,@(cdr e)) fname lam namemap defined toplevel interp opaq globals))
           ;; remaining `decl` expressions are only type assertions if the
           ;; argument is global or a non-symbol.
           ((decl)
            (cond ((and (symbol? (cadr e))
                        (local-in? (cadr e) lam))
                   '(null))
                  (else
-                  (if (or (symbol? (cadr e)) (and (pair? (cadr e)) (eq? (caadr e) 'outerref)))
-                      (error "type declarations on global variables are not yet supported"))
-                  (cl-convert `(call (core typeassert) ,@(cdr e)) fname lam namemap defined toplevel interp opaq))))
+                  (cl-convert
+                    (let ((ref (binding-to-globalref (cadr e))))
+                      (and ref (get globals ref #f)
+                           (error (string "multiple type annotations for global \""
+                                         (deparse (cadr e)) "\".")))
+                      (if ref
+                          (let* ((rhs0 (caddr e))
+                                 (temp (and (not (effect-free? rhs0)) (make-ssavalue)))
+                                 (rhs  (or temp rhs0)))
+                            (put! globals ref rhs)
+                            `(toplevel-butfirst
+                               ,(if temp `(block (= ,temp ,rhs0) (null)) '(null))
+                               (call (core _set_typeof!) ,(cadr ref) (inert ,(caddr ref)) ,(caddr e))))
+                          `(call (core typeassert) ,@(cdr e))))
+                    fname lam namemap defined toplevel interp opaq globals))))
           ;; `with-static-parameters` expressions can be removed now; used only by analyze-vars
           ((with-static-parameters)
-           (cl-convert (cadr e) fname lam namemap defined toplevel interp opaq))
+           (cl-convert (cadr e) fname lam namemap defined toplevel interp opaq globals))
           (else
            (cons (car e)
-                 (map-cl-convert (cdr e) fname lam namemap defined toplevel interp opaq))))))))
+                 (map-cl-convert (cdr e) fname lam namemap defined toplevel interp opaq globals))))))))
 
 (define (closure-convert e) (cl-convert e #f #f #f #f #f #f #f))
 

src/julia.h

@@ -503,12 +503,12 @@ typedef struct {
     jl_sym_t *name;
     _Atomic(jl_value_t*) value;
     _Atomic(jl_value_t*) globalref;  // cached GlobalRef for this binding
-    jl_value_t *ty                   // add type for globals
     struct _jl_module_t* owner;  // for individual imported bindings -- TODO: make _Atomic
     uint8_t constp:1;
     uint8_t exportp:1;
     uint8_t imported:1;
     uint8_t deprecated:2; // 0=not deprecated, 1=renamed, 2=moved to another package
+    jl_value_t *ty;                  // add type for globals
 } jl_binding_t;
 
 typedef struct {

src/module.c

@@ -663,6 +663,10 @@ JL_DLLEXPORT void jl_set_global(jl_module_t *m JL_ROOTING_ARGUMENT, jl_sym_t *va
 JL_DLLEXPORT void jl_set_const(jl_module_t *m JL_ROOTING_ARGUMENT, jl_sym_t *var, jl_value_t *val JL_ROOTED_ARGUMENT)
 {
     jl_binding_t *bp = jl_get_binding_wr(m, var, 1);
+    if (bp->ty) {
+        jl_errorf("cannot declare %s constant; it was already declared as a typed global",
+                  jl_symbol_name(bp->name));
+    }
     if (bp->value == NULL) {
         uint8_t constp = 0;
         // if (jl_atomic_cmpswap(&bp->constp, &constp, 1)) {
@@ -684,6 +688,13 @@ JL_DLLEXPORT int jl_is_const(jl_module_t *m, jl_sym_t *var)
     return b && b->constp;
 }
 
+JL_DLLEXPORT jl_value_t *jl_get_binding_type(jl_module_t *m, jl_sym_t *var) {
+    jl_binding_t *b = jl_get_binding(m, var);
+    if (b == NULL || b->ty == NULL)
+        return (jl_value_t*)jl_any_type;
+    return b->ty;
+}
+
 // set the deprecated flag for a binding:
 //   0=not deprecated, 1=renamed, 2=moved to another package
 JL_DLLEXPORT void jl_deprecate_binding(jl_module_t *m, jl_sym_t *var, int flag)
@@ -791,6 +802,10 @@ void jl_binding_deprecation_warning(jl_module_t *m, jl_binding_t *b)
 
 JL_DLLEXPORT void jl_checked_assignment(jl_binding_t *b, jl_value_t *rhs) JL_NOTSAFEPOINT
 {
+    if (b->ty && !jl_isa(rhs, b->ty)) {
+        jl_errorf("Cannot assign a value of an incompatible type to the typed global %s.",
+                  jl_symbol_name(b->name));
+    }
     if (b->constp) {
         jl_value_t *old = NULL;
         if (jl_atomic_cmpswap(&b->value, &old, rhs)) {
@@ -807,11 +822,6 @@ JL_DLLEXPORT void jl_checked_assignment(jl_binding_t *b, jl_value_t *rhs) JL_NOT
         }
         jl_safe_printf("WARNING: redefinition of constant %s. This may fail, cause incorrect answers, or produce other errors.\n",
                        jl_symbol_name(b->name));
-        // Get and set type of global binding if not already there
-        if (b->ty == NULL) {
-            jl_value_t *type = jl_typeof(b->value);
-            jl_set_typeof(b->value, type);
-        }
     }
     jl_atomic_store_relaxed(&b->value, rhs);
     jl_gc_wb_binding(b, rhs);
@@ -823,6 +833,10 @@ JL_DLLEXPORT void jl_declare_constant(jl_binding_t *b)
         jl_errorf("cannot declare %s constant; it already has a value",
                   jl_symbol_name(b->name));
     }
+    if (b->ty) {
+        jl_errorf("cannot declare %s constant; it was already declared as a typed global",
+                  jl_symbol_name(b->name));
+    }
     b->constp = 1;
 }
 

src/staticdata.c

@@ -249,7 +249,8 @@ static const jl_fptr_args_t id_to_fptrs[] = {
     &jl_f_arrayref, &jl_f_const_arrayref, &jl_f_arrayset, &jl_f_arraysize, &jl_f_apply_type,
     &jl_f_applicable, &jl_f_invoke, &jl_f_sizeof, &jl_f__expr, &jl_f__typevar,
     &jl_f_ifelse, &jl_f__structtype, &jl_f__abstracttype, &jl_f__primitivetype,
-    &jl_f__typebody, &jl_f__setsuper, &jl_f__equiv_typedef, &jl_f_opaque_closure_call,
+    &jl_f__typebody, &jl_f__setsuper, &jl_f__equiv_typedef, &jl_f__set_typeof,
+    &jl_f_opaque_closure_call,
     NULL };
 
 typedef struct {