better handling of missing backedges (#58636)

Manage a single dictionary (keyed by TypeName) instead of scattering
this info into each TypeName scattered across the system. This makes it
much easier to scan the whole table when required and to split it up
better, so that all kwcalls and all constructors don't end up stuck into
just one table. While not enormous (or even the largest) just using the
REPL and Pkg, they are clearly larger than intended for a linear scan:

```
julia> length(Type.body.name.backedges)
1024

julia> length(typeof(Core.kwcall).name.backedges)
196

julia> length(typeof(convert).name.backedges)
1510
```

Author: vtjnash

src/datatype.c

@@ -62,6 +62,7 @@ JL_DLLEXPORT jl_methtable_t *jl_new_method_table(jl_sym_t *name, jl_module_t *mo
     mt->cache = mc;
     mt->name = name;
     mt->module = module;
+    mt->backedges = (jl_genericmemory_t*)jl_an_empty_memory_any;
     JL_GC_POP();
     return mt;
 }
@@ -88,7 +89,6 @@ JL_DLLEXPORT jl_typename_t *jl_new_typename_in(jl_sym_t *name, jl_module_t *modu
     tn->partial = NULL;
     tn->atomicfields = NULL;
     tn->constfields = NULL;
-    tn->backedges = NULL;
     tn->max_methods = 0;
     jl_atomic_store_relaxed(&tn->max_args, 0);
     jl_atomic_store_relaxed(&tn->cache_entry_count, 0);

src/gf.c

@@ -3004,23 +3004,22 @@ void jl_init_types(void) JL_GC_DISABLED
     jl_typename_type->name->wrapper = (jl_value_t*)jl_typename_type;
     jl_typename_type->super = jl_any_type;
     jl_typename_type->parameters = jl_emptysvec;
-    jl_typename_type->name->n_uninitialized = 19 - 2;
-    jl_typename_type->name->names = jl_perm_symsvec(19, "name", "module", "singletonname",
+    jl_typename_type->name->n_uninitialized = 18 - 2;
+    jl_typename_type->name->names = jl_perm_symsvec(18, "name", "module", "singletonname",
                                                     "names", "atomicfields", "constfields",
                                                     "wrapper", "Typeofwrapper", "cache", "linearcache",
-                                                    "backedges", "partial",
-                                                    "hash", "max_args", "n_uninitialized",
+                                                    "partial", "hash", "max_args", "n_uninitialized",
                                                     "flags", // "abstract", "mutable", "mayinlinealloc",
                                                     "cache_entry_count", "max_methods", "constprop_heuristic");
-    const static uint32_t typename_constfields[1]  = { 0b0001101000001001011 }; // TODO: put back atomicfields and constfields in this list
-    const static uint32_t typename_atomicfields[1] = { 0b0010010001110000000 };
+    const static uint32_t typename_constfields[1]  = { 0b000110100001001011 }; // TODO: put back atomicfields and constfields in this list
+    const static uint32_t typename_atomicfields[1] = { 0b001001001110000000 };
     jl_typename_type->name->constfields = typename_constfields;
     jl_typename_type->name->atomicfields = typename_atomicfields;
     jl_precompute_memoized_dt(jl_typename_type, 1);
-    jl_typename_type->types = jl_svec(19, jl_symbol_type, jl_any_type /*jl_module_type*/, jl_symbol_type,
+    jl_typename_type->types = jl_svec(18, jl_symbol_type, jl_any_type /*jl_module_type*/, jl_symbol_type,
                                       jl_simplevector_type,
                                       jl_any_type/*jl_voidpointer_type*/, jl_any_type/*jl_voidpointer_type*/,
-                                      jl_type_type, jl_type_type, jl_simplevector_type, jl_simplevector_type,
+                                      jl_type_type, jl_simplevector_type, jl_simplevector_type,
                                       jl_methcache_type, jl_any_type,
                                       jl_any_type /*jl_long_type*/,
                                       jl_any_type /*jl_int32_type*/,
@@ -3046,13 +3045,13 @@ void jl_init_types(void) JL_GC_DISABLED
     jl_methtable_type->super = jl_any_type;
     jl_methtable_type->parameters = jl_emptysvec;
     jl_methtable_type->name->n_uninitialized = 0;
-    jl_methtable_type->name->names = jl_perm_symsvec(4, "defs", "cache", "name", "module");
-    const static uint32_t methtable_constfields[1] = { 0b1110 };
-    const static uint32_t methtable_atomicfields[1] = { 0b0001 };
+    jl_methtable_type->name->names = jl_perm_symsvec(5, "defs", "cache", "name", "module", "backedges");
+    const static uint32_t methtable_constfields[1] = { 0b01110 };
+    const static uint32_t methtable_atomicfields[1] = { 0b00001 };
     jl_methtable_type->name->constfields = methtable_constfields;
     jl_methtable_type->name->atomicfields = methtable_atomicfields;
     jl_precompute_memoized_dt(jl_methtable_type, 1);
-    jl_methtable_type->types = jl_svec(4, jl_any_type, jl_methcache_type, jl_symbol_type, jl_any_type /*jl_module_type*/);
+    jl_methtable_type->types = jl_svec(5, jl_any_type, jl_methcache_type, jl_symbol_type, jl_any_type /*jl_module_type*/, jl_any_type);
 
     jl_symbol_type->name = jl_new_typename_in(jl_symbol("Symbol"), core, 0, 1);
     jl_symbol_type->name->wrapper = (jl_value_t*)jl_symbol_type;
@@ -3378,6 +3377,7 @@ void jl_init_types(void) JL_GC_DISABLED
     core = jl_core_module;
     jl_method_table->module = core;
     jl_atomic_store_relaxed(&jl_method_table->cache->leafcache, (jl_genericmemory_t*)jl_an_empty_memory_any);
+    jl_method_table->backedges = (jl_genericmemory_t*)jl_an_empty_memory_any;
     jl_atomic_store_relaxed(&core->bindingkeyset, (jl_genericmemory_t*)jl_an_empty_memory_any);
     // export own name, so "using Foo" makes "Foo" itself visible
     jl_set_initial_const(core, core->name, (jl_value_t*)core, 1);
@@ -3842,13 +3842,13 @@ void jl_init_types(void) JL_GC_DISABLED
     jl_svecset(jl_typename_type->types, 5, jl_voidpointer_type);
     jl_svecset(jl_typename_type->types, 6, jl_type_type);
     jl_svecset(jl_typename_type->types, 7, jl_type_type);
-    jl_svecset(jl_typename_type->types, 12, jl_long_type);
+    jl_svecset(jl_typename_type->types, 11, jl_long_type);
+    jl_svecset(jl_typename_type->types, 12, jl_int32_type);
     jl_svecset(jl_typename_type->types, 13, jl_int32_type);
-    jl_svecset(jl_typename_type->types, 14, jl_int32_type);
+    jl_svecset(jl_typename_type->types, 14, jl_uint8_type);
     jl_svecset(jl_typename_type->types, 15, jl_uint8_type);
     jl_svecset(jl_typename_type->types, 16, jl_uint8_type);
     jl_svecset(jl_typename_type->types, 17, jl_uint8_type);
-    jl_svecset(jl_typename_type->types, 18, jl_uint8_type);
     jl_svecset(jl_methcache_type->types, 2, jl_long_type); // voidpointer
     jl_svecset(jl_methcache_type->types, 3, jl_long_type); // uint32_t plus alignment
     jl_svecset(jl_methtable_type->types, 3, jl_module_type);

src/jltypes.c

@@ -520,7 +520,6 @@ typedef struct {
     _Atomic(jl_value_t*) Typeofwrapper;  // cache for Type{wrapper}
     _Atomic(jl_svec_t*) cache;        // sorted array
     _Atomic(jl_svec_t*) linearcache;  // unsorted array
-    jl_array_t *backedges; // uncovered (sig => caller::CodeInstance) pairs with this type as the function
     jl_array_t *partial;     // incomplete instantiations of this type
     intptr_t hash;
     _Atomic(int32_t) max_args;  // max # of non-vararg arguments in a signature with this type as the function
@@ -882,6 +881,7 @@ typedef struct _jl_methtable_t {
     jl_methcache_t *cache;
     jl_sym_t *name; // sometimes used for debug printing
     jl_module_t *module; // sometimes used for debug printing
+    jl_genericmemory_t *backedges; // IdDict{top typenames, Vector{uncovered (sig => caller::CodeInstance)}}
 } jl_methtable_t;
 
 typedef struct {

src/julia.h

@@ -857,7 +857,6 @@ jl_expr_t *jl_exprn(jl_sym_t *head, size_t n);
 jl_function_t *jl_new_generic_function(jl_sym_t *name, jl_module_t *module, size_t new_world);
 jl_function_t *jl_new_generic_function_with_supertype(jl_sym_t *name, jl_module_t *module, jl_datatype_t *st, size_t new_world);
 int jl_foreach_reachable_mtable(int (*visit)(jl_methtable_t *mt, void *env), jl_array_t *mod_array, void *env);
-int foreach_mtable_in_module(jl_module_t *m, int (*visit)(jl_methtable_t *mt, void *env), void *env);
 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);
@@ -939,7 +938,6 @@ JL_DLLEXPORT jl_methtable_t *jl_method_get_table(
     jl_method_t *method JL_PROPAGATES_ROOT) JL_NOTSAFEPOINT;
 JL_DLLEXPORT jl_methcache_t *jl_method_get_cache(
     jl_method_t *method JL_PROPAGATES_ROOT) JL_NOTSAFEPOINT;
-void jl_foreach_top_typename_for(void (*f)(jl_typename_t*, void*), jl_value_t *argtypes JL_PROPAGATES_ROOT, void *env);
 
 JL_DLLEXPORT int jl_pointer_egal(jl_value_t *t);
 JL_DLLEXPORT jl_value_t *jl_nth_slot_type(jl_value_t *sig JL_PROPAGATES_ROOT, size_t i) JL_NOTSAFEPOINT;

src/julia_internal.h

@@ -1154,39 +1154,6 @@ JL_DLLEXPORT jl_value_t *jl_declare_const_gf(jl_module_t *mod, jl_sym_t *name)
     return gf;
 }
 
-static void foreach_top_nth_typename(void (*f)(jl_typename_t*, void*), jl_value_t *a JL_PROPAGATES_ROOT, int n, void *env)
-{
-    if (jl_is_datatype(a)) {
-        if (n == 0) {
-            jl_datatype_t *dt = ((jl_datatype_t*)a);
-            jl_typename_t *tn = NULL;
-            while (1) {
-                if (dt != jl_any_type && dt != jl_function_type)
-                    tn = dt->name;
-                if (dt->super == dt)
-                    break;
-                dt = dt->super;
-            }
-            if (tn)
-                f(tn, env);
-        }
-        else if (jl_is_tuple_type(a)) {
-            if (jl_nparams(a) >= n)
-                foreach_top_nth_typename(f, jl_tparam(a, n - 1), 0, env);
-        }
-    }
-    else if (jl_is_typevar(a)) {
-        foreach_top_nth_typename(f, ((jl_tvar_t*)a)->ub, n, env);
-    }
-    else if (jl_is_unionall(a)) {
-        foreach_top_nth_typename(f, ((jl_unionall_t*)a)->body, n, env);
-    }
-    else if (jl_is_uniontype(a)) {
-        jl_uniontype_t *u = (jl_uniontype_t*)a;
-        foreach_top_nth_typename(f, u->a, n, env);
-        foreach_top_nth_typename(f, u->b, n, env);
-    }
-}
 
 // get the MethodTable for dispatch, or `nothing` if cannot be determined
 JL_DLLEXPORT jl_methtable_t *jl_method_table_for(jl_value_t *argtypes JL_PROPAGATES_ROOT) JL_NOTSAFEPOINT
@@ -1200,11 +1167,6 @@ JL_DLLEXPORT jl_methcache_t *jl_method_cache_for(jl_value_t *argtypes JL_PROPAGA
     return jl_method_table->cache;
 }
 
-void jl_foreach_top_typename_for(void (*f)(jl_typename_t*, void*), jl_value_t *argtypes JL_PROPAGATES_ROOT, void *env)
-{
-    foreach_top_nth_typename(f, argtypes, 1, env);
-}
-
 jl_methcache_t *jl_kwmethod_cache_for(jl_value_t *argtypes JL_PROPAGATES_ROOT) JL_NOTSAFEPOINT
 {
     return jl_method_table->cache;

src/method.c

@@ -867,20 +867,30 @@ static void jl_insert_into_serialization_queue(jl_serializer_state *s, jl_value_
             assert(!jl_object_in_image((jl_value_t*)tn->module));
             assert(!jl_object_in_image((jl_value_t*)tn->wrapper));
         }
+    }
+    if (jl_is_mtable(v)) {
+        jl_methtable_t *mt = (jl_methtable_t*)v;
         // Any back-edges will be re-validated and added by staticdata.jl, so
         // drop them from the image here
         if (s->incremental || jl_options.trim || jl_options.strip_ir) {
-            record_field_change((jl_value_t**)&tn->backedges, NULL);
+            record_field_change((jl_value_t**)&mt->backedges, jl_an_empty_memory_any);
         }
         else {
             // don't recurse into all backedges memory (yet)
-            jl_value_t *backedges = get_replaceable_field((jl_value_t**)&tn->backedges, 1);
-            if (backedges) {
-                jl_queue_for_serialization_(s, (jl_value_t*)((jl_array_t*)backedges)->ref.mem, 0, 1);
-                for (size_t i = 0, n = jl_array_nrows(backedges); i < n; i += 2) {
-                    jl_value_t *t = jl_array_ptr_ref(backedges, i);
-                    assert(!jl_is_code_instance(t));
-                    jl_queue_for_serialization(s, t);
+            jl_value_t *allbackedges = get_replaceable_field((jl_value_t**)&mt->backedges, 1);
+            jl_queue_for_serialization_(s, allbackedges, 0, 1);
+            for (size_t i = 0, n = ((jl_genericmemory_t*)allbackedges)->length; i < n; i += 2) {
+                jl_value_t *tn = jl_genericmemory_ptr_ref(allbackedges, i);
+                jl_queue_for_serialization(s, tn);
+                jl_value_t *backedges = jl_genericmemory_ptr_ref(allbackedges, i + 1);
+                if (backedges && backedges != jl_nothing) {
+                    jl_queue_for_serialization_(s, (jl_value_t*)((jl_array_t*)backedges)->ref.mem, 0, 1);
+                    jl_queue_for_serialization(s, backedges);
+                    for (size_t i = 0, n = jl_array_nrows(backedges); i < n; i += 2) {
+                        jl_value_t *t = jl_array_ptr_ref(backedges, i);
+                        assert(!jl_is_code_instance(t));
+                        jl_queue_for_serialization(s, t);
+                    }
                 }
             }
         }
@@ -2573,8 +2583,6 @@ static void jl_prune_mi_backedges(jl_array_t *backedges)
 
 static void jl_prune_tn_backedges(jl_array_t *backedges)
 {
-    if (backedges == NULL)
-        return;
     size_t i = 0, ins = 0, n = jl_array_nrows(backedges);
     for (i = 1; i < n; i += 2) {
         jl_value_t *ci = jl_array_ptr_ref(backedges, i);
@@ -2586,6 +2594,15 @@ static void jl_prune_tn_backedges(jl_array_t *backedges)
     jl_array_del_end(backedges, n - ins);
 }
 
+static void jl_prune_mt_backedges(jl_genericmemory_t *allbackedges)
+{
+    for (size_t i = 0, n = allbackedges->length; i < n; i += 2) {
+        jl_value_t *tn = jl_genericmemory_ptr_ref(allbackedges, i);
+        jl_value_t *backedges = jl_genericmemory_ptr_ref(allbackedges, i + 1);
+        if (tn && tn != jl_nothing && backedges)
+            jl_prune_tn_backedges((jl_array_t*)backedges);
+    }
+}
 
 static void jl_prune_binding_backedges(jl_array_t *backedges)
 {
@@ -3240,8 +3257,6 @@ static void jl_save_system_image_to_stream(ios_t *f, jl_array_t *mod_array,
                     jl_prune_type_cache_hash(jl_atomic_load_relaxed(&tn->cache)));
                 jl_gc_wb(tn, jl_atomic_load_relaxed(&tn->cache));
                 jl_prune_type_cache_linear(jl_atomic_load_relaxed(&tn->linearcache));
-                jl_value_t *backedges = get_replaceable_field((jl_value_t**)&tn->backedges, 1);
-                jl_prune_tn_backedges((jl_array_t*)backedges);
             }
             else if (jl_is_method_instance(v)) {
                 jl_method_instance_t *mi = (jl_method_instance_t*)v;
@@ -3253,6 +3268,11 @@ static void jl_save_system_image_to_stream(ios_t *f, jl_array_t *mod_array,
                 jl_value_t *backedges = get_replaceable_field((jl_value_t**)&b->backedges, 1);
                 jl_prune_binding_backedges((jl_array_t*)backedges);
             }
+            else if (jl_is_mtable(v)) {
+                jl_methtable_t *mt = (jl_methtable_t*)v;
+                jl_value_t *backedges = get_replaceable_field((jl_value_t**)&mt->backedges, 1);
+                jl_prune_mt_backedges((jl_genericmemory_t*)backedges);
+            }
         }
     }
 

src/staticdata.c

@@ -1632,10 +1632,10 @@ end
 let errs = IOBuffer()
     run(`$(Base.julia_cmd()) -e '
         using Test
-        @test isdefined(Type.body.name, :backedges)
+        @test !isempty(Core.GlobalMethods.backedges)
         Base.Experimental.disable_new_worlds()
         @test_throws "disable_new_worlds" @eval f() = 1
-        @test !isdefined(Type.body.name, :backedges)
+        @test isempty(Core.GlobalMethods.backedges)
         @test_throws "disable_new_worlds" Base.delete_method(which(+, (Int, Int)))
         @test 1+1 == 2
         using Dates