memoize more type properties (#16320)

has-typevars is called pretty heavily during intersection and subtyping
so it can be a hot-path

fix lowering of TypeVar in TypeConstructor typealias expression:
instead have lowering emit the correct TypeVar expression instead of mutating (aka corrupting)
the immutable TypeVars in the TypeConstructor constructor

avoid corrupting immutable TypeVar after building a DataType for them:
instead have lowering emit the correct TypeVar expression
instead of modifying them in the DataType constructor
fix #12238
fix #16301

Author: vtjnash

base/inference.jl

@@ -374,13 +374,7 @@ function type_depth(t::ANY)
         t === Bottom && return 0
         return maximum(type_depth, t.types) + 1
     elseif isa(t, DataType)
-        t = t::DataType
-        P = t.parameters
-        isempty(P) && return 0
-        if t.depth == 0
-            t.depth = maximum(type_depth, P) + 1
-        end
-        return t.depth
+        return (t::DataType).depth
     end
     return 0
 end

src/alloc.c

@@ -816,6 +816,9 @@ JL_DLLEXPORT jl_datatype_t *jl_new_uninitialized_datatype(size_t nfields, int8_t
     t->haspadding = 0;
     t->pointerfree = 0;
     t->depth = 0;
+    t->hastypevars = 0;
+    t->haswildcard = 0;
+    t->isleaftype = 1;
     return t;
 }
 
@@ -983,6 +986,7 @@ JL_DLLEXPORT jl_datatype_t *jl_new_datatype(jl_sym_t *name, jl_datatype_t *super
         t->name->primary = (jl_value_t*)t;
         jl_gc_wb(t->name, t);
     }
+    jl_precompute_memoized_dt(t);
 
     if (abstract || jl_svec_len(parameters) > 0) {
         t->uid = 0;
@@ -1011,14 +1015,22 @@ JL_DLLEXPORT jl_datatype_t *jl_new_bitstype(jl_value_t *name, jl_datatype_t *sup
 
 // type constructor -----------------------------------------------------------
 
-jl_typector_t *jl_new_type_ctor(jl_svec_t *params, jl_value_t *body)
+JL_DLLEXPORT jl_value_t *jl_new_type_constructor(jl_svec_t *p, jl_value_t *body)
 {
-    jl_typector_t *tc = (jl_typector_t*)newobj((jl_value_t*)jl_typector_type,NWORDS(sizeof(jl_typector_t)));
-    tc->parameters = params;
+#ifndef NDEBUG
+    size_t i, np = jl_svec_len(p);
+    for (i = 0; i < np; i++) {
+        jl_tvar_t *tv = (jl_tvar_t*)jl_svecref(p, i);
+        assert(jl_is_typevar(tv) && !tv->bound);
+    }
+#endif
+    jl_typector_t *tc = (jl_typector_t*)newobj((jl_value_t*)jl_typector_type, NWORDS(sizeof(jl_typector_t)));
+    tc->parameters = p;
     tc->body = body;
-    return (jl_typector_t*)tc;
+    return (jl_value_t*)tc;
 }
 
+
 // bits constructors ----------------------------------------------------------
 
 #define BOXN_FUNC(nb,nw)                                                \

src/builtins.c

@@ -928,15 +928,6 @@ JL_CALLABLE(jl_f_apply_type)
     return jl_apply_type_(args[0], &args[1], nargs-1);
 }
 
-JL_DLLEXPORT jl_value_t *jl_new_type_constructor(jl_svec_t *p, jl_value_t *t)
-{
-    jl_value_t *tc = (jl_value_t*)jl_new_type_ctor(p, t);
-    int i;
-    for(i=0; i < jl_svec_len(p); i++)
-        ((jl_tvar_t*)jl_svecref(p,i))->bound = 0;
-    return tc;
-}
-
 // generic function reflection ------------------------------------------------
 
 static void jl_check_type_tuple(jl_value_t *t, jl_sym_t *name, const char *ctx)

src/dump.c

@@ -546,7 +546,9 @@ static void jl_serialize_datatype(ios_t *s, jl_datatype_t *dt)
     write_uint16(s, nf);
     write_int32(s, dt->size);
     int has_instance = !!(dt->instance != NULL);
-    write_uint8(s, dt->abstract | (dt->mutabl<<1) | (dt->pointerfree<<2) | (has_instance<<3));
+    write_uint8(s, dt->abstract | (dt->mutabl<<1) | (dt->pointerfree<<2) | (has_instance<<3) |
+            (dt->hastypevars<<4) | (dt->haswildcard<<5) | (dt->isleaftype<<6));
+    write_int32(s, dt->depth);
     write_int8(s, dt->fielddesc_type);
     if (!dt->abstract) {
         write_uint16(s, dt->ninitialized);
@@ -1173,6 +1175,7 @@ static jl_value_t *jl_deserialize_datatype(ios_t *s, int pos, jl_value_t **loc)
     uint16_t nf = read_uint16(s);
     size_t size = read_int32(s);
     uint8_t flags = read_uint8(s);
+    uint8_t depth = read_int32(s);
     uint8_t fielddesc_type = read_int8(s);
     jl_datatype_t *dt;
     if (tag == 2)
@@ -1194,6 +1197,10 @@ static jl_value_t *jl_deserialize_datatype(ios_t *s, int pos, jl_value_t **loc)
     dt->abstract = flags&1;
     dt->mutabl = (flags>>1)&1;
     dt->pointerfree = (flags>>2)&1;
+    dt->hastypevars = (flags>>4)&1;
+    dt->haswildcard = (flags>>5)&1;
+    dt->isleaftype = (flags>>6)&1;
+    dt->depth = depth;
     if (!dt->abstract) {
         dt->ninitialized = read_uint16(s);
         dt->uid = mode != MODE_MODULE && mode != MODE_MODULE_POSTWORK ? read_int32(s) : 0;

src/interpreter.c

@@ -345,6 +345,12 @@ static jl_value_t *eval(jl_value_t *e, jl_value_t **locals, jl_lambda_info_t *la
         jl_datatype_t *dt = NULL;
         JL_GC_PUSH4(&para, &super, &temp, &dt);
         temp = eval(args[2], locals, lam);  // field names
+#ifndef NDEBUG
+        size_t i, l = jl_svec_len(para);
+        for (i = 0; i < l; i++) {
+            assert(!((jl_tvar_t*)jl_svecref(para, i))->bound);
+        }
+#endif
         dt = jl_new_datatype((jl_sym_t*)name, jl_any_type, (jl_svec_t*)para,
                              (jl_svec_t*)temp, NULL,
                              0, args[5]==jl_true ? 1 : 0, jl_unbox_long(args[6]));
@@ -377,9 +383,6 @@ static jl_value_t *eval(jl_value_t *e, jl_value_t **locals, jl_lambda_info_t *la
             b->value = temp;
             jl_rethrow();
         }
-        for(size_t i=0; i < jl_svec_len(para); i++) {
-            ((jl_tvar_t*)jl_svecref(para,i))->bound = 0;
-        }
         jl_compute_field_offsets(dt);
         if (para == (jl_value_t*)jl_emptysvec && jl_is_datatype_singleton(dt)) {
             dt->instance = newstruct(dt);

src/jltypes.c

@@ -95,12 +95,22 @@ static int jl_has_typevars__(jl_value_t *v, int incl_wildcard, jl_value_t **p, s
     if (jl_is_typector(v))
         return incl_wildcard;
     jl_svec_t *t;
+    int expect = -1;
     if (jl_is_uniontype(v)) {
         t = ((jl_uniontype_t*)v)->types;
     }
     else if (jl_is_datatype(v)) {
         if (is_unspec((jl_datatype_t*)v))
             return 0;
+        if (p == NULL) {
+            if (incl_wildcard)
+                expect = ((jl_datatype_t*)v)->haswildcard;
+            else
+                expect = ((jl_datatype_t*)v)->hastypevars;
+#ifdef NDEBUG
+            return expect;
+#endif
+        }
         t = ((jl_datatype_t*)v)->parameters;
     }
     else {
@@ -110,14 +120,17 @@ static int jl_has_typevars__(jl_value_t *v, int incl_wildcard, jl_value_t **p, s
     for(i=0; i < l; i++) {
         jl_value_t *elt = jl_svecref(t, i);
         if (elt != v) {
-            if (jl_has_typevars__(elt, incl_wildcard, p, np))
+            if (jl_has_typevars__(elt, incl_wildcard, p, np)) {
+                if (expect >= 0) assert(expect);
                 return 1;
+            }
         }
     }
     // probably not necessary; no reason to use match() instead of subtype()
     // on the unconstrained version of a type
     //if (jl_is_typector(v))
     //    return jl_svec_len((((jl_typector_t*)v)->parameters) > 0);
+    if (expect >= 0) assert(!expect);
     return 0;
 }
 
@@ -148,26 +161,39 @@ JL_DLLEXPORT int jl_has_typevars(jl_value_t *v)
 JL_DLLEXPORT int jl_is_leaf_type(jl_value_t *v)
 {
     if (jl_is_datatype(v)) {
+        int isleaf = ((jl_datatype_t*)v)->isleaftype;
+#ifdef NDEBUG
+        return isleaf;
+#else
         if (((jl_datatype_t*)v)->abstract) {
-            if (jl_is_type_type(v))
-                return !jl_is_typevar(jl_tparam0(v));
-            return 0;
+            int x = 0;
+            if (jl_is_type_type(v)) {
+                x = !jl_is_typevar(jl_tparam0(v));
+            }
+            assert(x == isleaf);
+            return x;
         }
         jl_svec_t *t = ((jl_datatype_t*)v)->parameters;
         size_t l = jl_svec_len(t);
         if (((jl_datatype_t*)v)->name == jl_tuple_typename) {
             for(int i=0; i < l; i++) {
-                if (!jl_is_leaf_type(jl_svecref(t,i)))
+                if (!jl_is_leaf_type(jl_svecref(t,i))) {
+                    assert(!isleaf);
                     return 0;
+                }
             }
         }
         else {
             for(int i=0; i < l; i++) {
-                if (jl_is_typevar(jl_svecref(t,i)))
+                if (jl_is_typevar(jl_svecref(t,i))) {
+                    assert(!isleaf);
                     return 0;
+                }
             }
         }
+        assert(isleaf);
         return 1;
+#endif
     }
     return 0;
 }
@@ -2039,6 +2065,45 @@ static jl_value_t *lookup_type_stack(jl_typestack_t *stack, jl_datatype_t *tt, s
     return NULL;
 }
 
+static size_t jl_type_depth(jl_value_t *dt)
+{
+    if (jl_is_uniontype(dt)) {
+        jl_svec_t *t = ((jl_uniontype_t*)dt)->types;
+        size_t i, l = jl_svec_len(t);
+        size_t depth = 0;
+        for (i = 0; i < l; i++) {
+            jl_value_t *p = jl_svecref(t, i);
+            size_t d = jl_type_depth(p);
+            if (d > depth)
+                depth = d;
+        }
+        return depth;
+    }
+    else if (jl_is_datatype(dt)) {
+        return ((jl_datatype_t*)dt)->depth;
+    }
+    return 0;
+}
+
+void jl_precompute_memoized_dt(jl_datatype_t *dt)
+{
+    int istuple = dt->name == jl_tuple_typename;
+    size_t i, l = jl_nparams(dt);
+    dt->isleaftype = !dt->abstract || (jl_type_type != NULL && dt->name == jl_type_type->name);
+    for (i = 0; i < l; i++) {
+        jl_value_t *p = jl_tparam(dt, i);
+        size_t d = jl_type_depth(p) + 1;
+        if (d > dt->depth)
+            dt->depth = d;
+        if (!dt->hastypevars)
+           dt->hastypevars = jl_has_typevars__(p, 0, NULL, 0);
+        if (!dt->haswildcard)
+            dt->haswildcard = jl_has_typevars__(p, 1, NULL, 0);
+        if (dt->isleaftype)
+            dt->isleaftype = (istuple ? jl_is_leaf_type(p) : !jl_is_typevar(p));
+    }
+}
+
 static jl_value_t *inst_datatype(jl_datatype_t *dt, jl_svec_t *p, jl_value_t **iparams, size_t ntp,
                                  int cacheable, int isabstract, jl_typestack_t *stack,
                                  jl_value_t **env, size_t n)
@@ -2115,6 +2180,7 @@ static jl_value_t *inst_datatype(jl_datatype_t *dt, jl_svec_t *p, jl_value_t **i
     ndt->ditype = NULL;
     ndt->size = 0;
     ndt->alignment = 1;
+    jl_precompute_memoized_dt(ndt);
 
     // assign uid as early as possible
     if (cacheable && !ndt->abstract && ndt->uid==0)
@@ -3340,6 +3406,9 @@ void jl_init_types(void)
     //jl_anytuple_type->parameters = jl_svec(1, jl_wrap_vararg((jl_value_t*)NULL, (jl_value_t*)NULL));
     jl_anytuple_type->types = jl_anytuple_type->parameters;
     jl_anytuple_type->nfields = 1;
+    jl_anytuple_type->hastypevars = 1;
+    jl_anytuple_type->haswildcard = 1;
+    jl_anytuple_type->isleaftype = 0;
 
     jl_tvar_t *tttvar = jl_new_typevar(jl_symbol("T"),
                                        (jl_value_t*)jl_bottom_type,(jl_value_t*)jl_any_type);

src/julia-syntax.scm

@@ -791,7 +791,7 @@
         (block
          (global ,name) (const ,name)
          ,@(map (lambda (v) `(local ,v)) params)
-         ,@(map make-assignment params (symbols->typevars params bounds #t))
+         ,@(map make-assignment params (symbols->typevars params bounds #f))
          (composite_type ,name (call (core svec) ,@params)
                          (call (core svec) ,@(map (lambda (x) `',x) field-names))
                          ,super (call (core svec) ,@field-types) ,mut ,min-initialized)))
@@ -1174,7 +1174,7 @@
                 ,@(map (lambda (v) `(local ,v)) params)
                 ,@(map (lambda (l r) (make-assignment l (expand-forms r)))
                        params
-                       (symbols->typevars params bounds #t))
+                       (symbols->typevars params bounds #f))
                 (call (core TypeConstructor)
                       (call (core svec) ,@params)
                       ,(expand-forms type-ex))))))))
@@ -2487,7 +2487,7 @@ f(x) = yt(x)
         ((,@(map (lambda (p) `(,p Any 18)) P))
          () 0 ())
         (body (global ,name) (const ,name)
-              ,@(map (lambda (p) `(= ,p (call (core TypeVar) ',p (core Any) true))) P)
+              ,@(map (lambda (p) `(= ,p (call (core TypeVar) ',p (core Any) false))) P)
               (composite_type ,name (call (core svec) ,@P)
                               (call (core svec) ,@(map (lambda (v) `',v) fields))
                               ,super

src/julia.h

@@ -329,7 +329,11 @@ typedef struct _jl_datatype_t {
     uint8_t mutabl;
     uint8_t pointerfree;
     int32_t ninitialized;
+    // memoized properties
     int32_t depth;
+    int8_t hastypevars; // bound
+    int8_t haswildcard; // unbound
+    int8_t isleaftype;
     // hidden fields:
     uint32_t nfields;
     uint32_t alignment : 29;  // strictest alignment over all fields

src/julia_internal.h

@@ -182,11 +182,11 @@ STATIC_INLINE int jl_is_type(jl_value_t *v)
 }
 jl_value_t *jl_type_intersection_matching(jl_value_t *a, jl_value_t *b,
                                           jl_svec_t **penv, jl_svec_t *tvars);
-jl_typector_t *jl_new_type_ctor(jl_svec_t *params, jl_value_t *body);
 jl_value_t *jl_apply_type_(jl_value_t *tc, jl_value_t **params, size_t n);
 jl_value_t *jl_instantiate_type_with(jl_value_t *t, jl_value_t **env, size_t n);
 jl_datatype_t *jl_new_abstracttype(jl_value_t *name, jl_datatype_t *super,
                                    jl_svec_t *parameters);
+void jl_precompute_memoized_dt(jl_datatype_t *dt);
 jl_datatype_t *jl_wrap_Type(jl_value_t *t);  // x -> Type{x}
 jl_datatype_t *jl_wrap_vararg(jl_value_t *t, jl_value_t *n);
 void jl_assign_bits(void *dest, jl_value_t *bits);

test/core.jl

@@ -4113,6 +4113,14 @@ f16090() = typeof(undefined_x16090::Tuple{Type{Int}})
 undefined_x16090 = (Int,)
 @test_throws TypeError f16090()
 
+# issue #12238
+type A12238{T} end
+type B12238{T,S}
+    a::A12238{B12238{Int,S}}
+end
+@test B12238.types[1] === A12238{B12238{Int}}
+@test A12238{B12238{Int}}.instance === B12238.types[1].instance
+
 # issue #16315
 let a = Any[]
     @noinline f() = a[end]