datatype: handle concrete type intersections that happen (JuliaLang#4…

…9017)

This is actually very similar to the current might_intersect_concrete,
but for subtyping and memoized. It replaces cached_by_hash, which was a
confusingly-named incomplete prior work towards this.

This gives NamedTuple{(:names)} hashes, which lets them go into the
faster type lookup tables. This is a fairly common type for some
packages to create, so we need this to avoid polluting our cache tables.

Reverts efafd83, since these types have
no intersection, the morespecific algorithm is no longer required to
have any opinion on them.

src/datatype.c

@@ -103,7 +103,7 @@ jl_datatype_t *jl_new_uninitialized_datatype(void)
     t->isprimitivetype = 0;
     t->zeroinit = 0;
     t->has_concrete_subtype = 1;
-    t->cached_by_hash = 0;
+    t->maybe_subtype_of_cache = 1;
     t->ismutationfree = 0;
     t->isidentityfree = 0;
     t->name = NULL;

src/jl_exported_funcs.inc

@@ -487,7 +487,6 @@
     XX(jl_typename_str) \
     XX(jl_typeof_str) \
     XX(jl_types_equal) \
-    XX(jl_type_equality_is_identity) \
     XX(jl_type_error) \
     XX(jl_type_error_rt) \
     XX(jl_type_intersection) \

src/julia.h

@@ -548,7 +548,7 @@ typedef struct _jl_datatype_t {
     uint16_t isbitstype:1; // relevant query for C-api and type-parameters
     uint16_t zeroinit:1; // if one or more fields requires zero-initialization
     uint16_t has_concrete_subtype:1; // If clear, no value will have this datatype
-    uint16_t cached_by_hash:1; // stored in hash-based set cache (instead of linear cache)
+    uint16_t maybe_subtype_of_cache:1; // Computational bit for has_concrete_supertype. See description in jltypes.c.
     uint16_t isprimitivetype:1; // whether this is declared with 'primitive type' keyword (sized, no fields, and immutable)
     uint16_t ismutationfree:1; // whether any mutable memory is reachable through this type (in the type or via fields)
     uint16_t isidentityfree:1; // whether this type or any object reachable through its fields has non-content-based identity
@@ -1414,7 +1414,6 @@ STATIC_INLINE int jl_egal_(const jl_value_t *a JL_MAYBE_UNROOTED, const jl_value
 #define jl_egal(a, b) jl_egal_((a), (b))
 
 // type predicates and basic operations
-JL_DLLEXPORT int jl_type_equality_is_identity(jl_value_t *t1, jl_value_t *t2) JL_NOTSAFEPOINT;
 JL_DLLEXPORT int jl_has_free_typevars(jl_value_t *v) JL_NOTSAFEPOINT;
 JL_DLLEXPORT int jl_has_typevar(jl_value_t *t, jl_tvar_t *v) JL_NOTSAFEPOINT;
 JL_DLLEXPORT int jl_has_typevar_from_unionall(jl_value_t *t, jl_unionall_t *ua);

src/julia_internal.h

@@ -722,6 +722,7 @@ void jl_init_main_module(void);
 JL_DLLEXPORT int jl_is_submodule(jl_module_t *child, jl_module_t *parent) JL_NOTSAFEPOINT;
 jl_array_t *jl_get_loaded_modules(void);
 JL_DLLEXPORT int jl_datatype_isinlinealloc(jl_datatype_t *ty, int pointerfree);
+int jl_type_equality_is_identity(jl_value_t *t1, jl_value_t *t2) JL_NOTSAFEPOINT;
 
 void jl_eval_global_expr(jl_module_t *m, jl_expr_t *ex, int set_type);
 jl_value_t *jl_toplevel_eval_flex(jl_module_t *m, jl_value_t *e, int fast, int expanded);

src/subtype.c

@@ -306,11 +306,8 @@ static int obviously_unequal(jl_value_t *a, jl_value_t *b)
             if (ad->name != bd->name)
                 return 1;
             int istuple = (ad->name == jl_tuple_typename);
-            if ((jl_is_concrete_type(a) || jl_is_concrete_type(b)) &&
-                jl_type_equality_is_identity(a, b)) {
-                if (!istuple && ad->name != jl_type_typename) // HACK: can't properly normalize Tuple{Float64} == Tuple{<:Float64} like types or Type{T} types
-                    return 1;
-            }
+            if (jl_type_equality_is_identity(a, b))
+                return 1;
             size_t i, np;
             if (istuple) {
                 size_t na = jl_nparams(ad), nb = jl_nparams(bd);
@@ -395,10 +392,7 @@ static int obviously_disjoint(jl_value_t *a, jl_value_t *b, int specificity)
         return 0;
     if (specificity && a == (jl_value_t*)jl_typeofbottom_type)
         return 0;
-    if (jl_is_concrete_type(a) && jl_is_concrete_type(b) &&
-        jl_type_equality_is_identity(a, b) &&
-        (((jl_datatype_t*)a)->name != jl_tuple_typename ||
-         ((jl_datatype_t*)b)->name != jl_tuple_typename))
+    if (jl_is_concrete_type(a) && jl_is_concrete_type(b) && jl_type_equality_is_identity(a, b))
         return 1;
     if (jl_is_unionall(a)) a = jl_unwrap_unionall(a);
     if (jl_is_unionall(b)) b = jl_unwrap_unionall(b);
@@ -1766,7 +1760,7 @@ static int obvious_subtype(jl_value_t *x, jl_value_t *y, jl_value_t *y0, int *su
     if (jl_is_datatype(y)) {
         int istuple = (((jl_datatype_t*)y)->name == jl_tuple_typename);
         int iscov = istuple;
-        // TODO: this would be a nice fast-path to have, unfortuanately,
+        // TODO: this would be a nice fast-path to have, unfortunately,
         //       datatype allocation fails to correctly hash-cons them
         //       and the subtyping tests include tests for this case
         //if (!iscov && ((jl_datatype_t*)y)->isconcretetype && !jl_is_type_type(x)) {
@@ -2253,7 +2247,7 @@ JL_DLLEXPORT int jl_isa(jl_value_t *x, jl_value_t *t)
             return 0;
         }
     }
-    if (jl_is_concrete_type(t) && jl_type_equality_is_identity(jl_typeof(x), t))
+    if (jl_is_concrete_type(t))
         return 0;
     return jl_subtype(jl_typeof(x), t);
 }
@@ -3735,6 +3729,7 @@ static jl_value_t *switch_union_tuple(jl_value_t *a, jl_value_t *b)
 
 // `a` might have a non-empty intersection with some concrete type b even if !(a<:b) and !(b<:a)
 // For example a=`Tuple{Type{<:Vector}}` and b=`Tuple{DataType}`
+// TODO: this query is partly available memoized as jl_type_equality_is_identity
 static int might_intersect_concrete(jl_value_t *a)
 {
     if (jl_is_unionall(a))
@@ -3784,9 +3779,9 @@ jl_value_t *jl_type_intersection_env_s(jl_value_t *a, jl_value_t *b, jl_svec_t *
         *ans = a; sz = szb;
         if (issubty) *issubty = 1;
     }
-    else if (lta && ltb) {
-        goto bot;
-    }
+    // else if (lta && ltb) { // !jl_type_equality_is_identity known in this case because obviously_disjoint returned false
+    //     goto bot;
+    // }
     else if (jl_subtype(b, a)) {
         *ans = b;
     }

test/specificity.jl

@@ -214,7 +214,7 @@ f27361(::M) where M <: Tuple{3} = nothing
 @test length(methods(f27361)) == 2
 
 # specificity of TypeofBottom
-@test args_morespecific(Tuple{Core.TypeofBottom}, Tuple{DataType})
+@test !args_morespecific(Tuple{DataType}, Tuple{Core.TypeofBottom})
 @test args_morespecific(Tuple{Core.TypeofBottom}, Tuple{Type{<:Tuple}})
 
 @test  args_morespecific(Tuple{Type{Any}, Type}, Tuple{Type{T}, Type{T}} where T)