replace incorrect Method.deleted_world with more useful Method.dispatch_status enum (#58291)

The original purpose of this field was to manage quickly detecting if a
method was replaced, but that stopped being correct after #53415. It
was a fairly heavy-weight description of that single bit of
information. This bit of information allows quickly bypassing some
method lookups from pkgimages, since it can quickly detect that the
result is trivially correct (such as single-argument functions).

Also fixes #58215

Author: vtjnash

Compiler/src/typeinfer.jl

@@ -1540,7 +1540,7 @@ function compile!(codeinfos::Vector{Any}, workqueue::CompilationQueue;
             # if this method is generally visible to the current compilation world,
             # and this is either the primary world, or not applicable in the primary world
             # then we want to compile and emit this
-            if item.def.primary_world <= world <= item.def.deleted_world
+            if item.def.primary_world <= world
                 ci = typeinf_ext(interp, item, SOURCE_MODE_GET_SOURCE)
                 ci isa CodeInstance && push!(workqueue, ci)
             end

base/errorshow.jl

@@ -592,8 +592,6 @@ function show_method_candidates(io::IO, ex::MethodError, kwargs=[])
             end
             if ex.world < reinterpret(UInt, method.primary_world)
                 print(iob, " (method too new to be called from this world context.)")
-            elseif ex.world > reinterpret(UInt, method.deleted_world)
-                print(iob, " (method deleted before this world age.)")
             end
             println(iob)
 

base/staticdata.jl

@@ -288,6 +288,30 @@ end
 
 function verify_call(@nospecialize(sig), expecteds::Core.SimpleVector, i::Int, n::Int, world::UInt)
     # verify that these edges intersect with the same methods as before
+    if n == 1
+        # first, fast-path a check if the expected method simply dominates its sig anyways
+        # so the result of ml_matches is already simply known
+        let t = expecteds[i], meth, minworld, maxworld, result
+            if t isa Method
+                meth = t
+            else
+                if t isa CodeInstance
+                    t = get_ci_mi(t)
+                else
+                    t = t::MethodInstance
+                end
+                meth = t.def::Method
+            end
+            if !iszero(meth.dispatch_status & METHOD_SIG_LATEST_ONLY)
+                minworld = meth.primary_world
+                @assert minworld ≤ world
+                maxworld = typemax(UInt)
+                result = Any[] # result is unused
+                return minworld, maxworld, result
+            end
+        end
+    end
+    # next, compare the current result of ml_matches to the old result
     lim = _jl_debug_method_invalidation[] !== nothing ? Int(typemax(Int32)) : n
     minworld = Ref{UInt}(1)
     maxworld = Ref{UInt}(typemax(UInt))
@@ -340,24 +364,25 @@ function verify_call(@nospecialize(sig), expecteds::Core.SimpleVector, i::Int, n
     return minworld[], maxworld[], result
 end
 
+# fast-path dispatch_status bit definitions (false indicates unknown)
+# true indicates this method would be returned as the result from `which` when invoking `method.sig` in the current latest world
+const METHOD_SIG_LATEST_WHICH = 0x1
+# true indicates this method would be returned as the only result from `methods` when calling `method.sig` in the current latest world
+const METHOD_SIG_LATEST_ONLY = 0x2
+
 function verify_invokesig(@nospecialize(invokesig), expected::Method, world::UInt)
     @assert invokesig isa Type
     local minworld::UInt, maxworld::UInt
     matched = nothing
-    if invokesig === expected.sig
-        # the invoke match is `expected` for `expected->sig`, unless `expected` is invalid
-        # TODO: this is broken since PR #53415
+    if invokesig === expected.sig && !iszero(expected.dispatch_status & METHOD_SIG_LATEST_WHICH)
+        # the invoke match is `expected` for `expected->sig`, unless `expected` is replaced
         minworld = expected.primary_world
-        maxworld = expected.deleted_world
         @assert minworld ≤ world
-        if maxworld < world
-            maxworld = 0
-        end
-    else
-        minworld = 1
         maxworld = typemax(UInt)
+    else
         mt = get_methodtable(expected)
         if mt === nothing
+            minworld = 1
             maxworld = 0
         else
             matched, valid_worlds = Compiler._findsup(invokesig, mt, world)

src/gf.c

@@ -296,7 +296,7 @@ void jl_mk_builtin_func(jl_datatype_t *dt, jl_sym_t *sname, jl_fptr_args_t fptr)
     m->isva = 1;
     m->nargs = 2;
     jl_atomic_store_relaxed(&m->primary_world, 1);
-    jl_atomic_store_relaxed(&m->deleted_world, ~(size_t)0);
+    jl_atomic_store_relaxed(&m->dispatch_status, METHOD_SIG_LATEST_ONLY | METHOD_SIG_LATEST_ONLY);
     m->sig = (jl_value_t*)jl_anytuple_type;
     m->slot_syms = jl_an_empty_string;
     m->nospecialize = 0;
@@ -307,7 +307,7 @@ void jl_mk_builtin_func(jl_datatype_t *dt, jl_sym_t *sname, jl_fptr_args_t fptr)
     JL_GC_PUSH2(&m, &newentry);
 
     newentry = jl_typemap_alloc(jl_anytuple_type, NULL, jl_emptysvec,
-            (jl_value_t*)m, jl_atomic_load_relaxed(&m->primary_world), jl_atomic_load_relaxed(&m->deleted_world));
+            (jl_value_t*)m, 1, ~(size_t)0);
     jl_typemap_insert(&mt->defs, (jl_value_t*)mt, newentry, jl_cachearg_offset(mt));
 
     jl_method_instance_t *mi = jl_get_specialized(m, (jl_value_t*)jl_anytuple_type, jl_emptysvec);
@@ -1908,7 +1908,7 @@ static int is_replacing(char ambig, jl_value_t *type, jl_method_t *m, jl_method_
         // since m2 was also a previous match over isect,
         // see if m was previously dominant over all m2
         // or if this was already ambiguous before
-        if (ambig != morespec_is && !jl_type_morespecific(m->sig, m2->sig)) {
+        if (ambig == morespec_is && !jl_type_morespecific(m->sig, m2->sig)) {
             // m and m2 were previously ambiguous over the full intersection of mi with type, and will still be ambiguous with addition of type
             return 0;
         }
@@ -2242,17 +2242,22 @@ JL_DLLEXPORT void jl_method_table_disable(jl_methtable_t *mt, jl_method_t *metho
     JL_LOCK(&world_counter_lock);
     if (!jl_atomic_load_relaxed(&allow_new_worlds))
         jl_error("Method changes have been disabled via a call to disable_new_worlds.");
-    JL_LOCK(&mt->writelock);
-    // Narrow the world age on the method to make it uncallable
-    size_t world = jl_atomic_load_relaxed(&jl_world_counter);
-    assert(method == methodentry->func.method);
-    assert(jl_atomic_load_relaxed(&method->deleted_world) == ~(size_t)0);
-    jl_atomic_store_relaxed(&method->deleted_world, world);
-    jl_atomic_store_relaxed(&methodentry->max_world, world);
-    jl_method_table_invalidate(mt, method, world);
-    jl_atomic_store_release(&jl_world_counter, world + 1);
-    JL_UNLOCK(&mt->writelock);
+    int enabled = jl_atomic_load_relaxed(&methodentry->max_world) == ~(size_t)0;
+    if (enabled) {
+        JL_LOCK(&mt->writelock);
+        // Narrow the world age on the method to make it uncallable
+        size_t world = jl_atomic_load_relaxed(&jl_world_counter);
+        assert(method == methodentry->func.method);
+        jl_atomic_store_relaxed(&method->dispatch_status, 0);
+        assert(jl_atomic_load_relaxed(&methodentry->max_world) == ~(size_t)0);
+        jl_atomic_store_relaxed(&methodentry->max_world, world);
+        jl_method_table_invalidate(mt, method, world);
+        jl_atomic_store_release(&jl_world_counter, world + 1);
+        JL_UNLOCK(&mt->writelock);
+    }
     JL_UNLOCK(&world_counter_lock);
+    if (!enabled)
+        jl_errorf("Method of %s already disabled", jl_symbol_name(method->name));
 }
 
 static int jl_type_intersection2(jl_value_t *t1, jl_value_t *t2, jl_value_t **isect JL_REQUIRE_ROOTED_SLOT, jl_value_t **isect2 JL_REQUIRE_ROOTED_SLOT)
@@ -2292,9 +2297,9 @@ jl_typemap_entry_t *jl_method_table_add(jl_methtable_t *mt, jl_method_t *method,
     JL_LOCK(&mt->writelock);
     // add our new entry
     assert(jl_atomic_load_relaxed(&method->primary_world) == ~(size_t)0); // min-world
-    assert(jl_atomic_load_relaxed(&method->deleted_world) == 1); // max-world
-    newentry = jl_typemap_alloc((jl_tupletype_t*)method->sig, simpletype, jl_emptysvec, (jl_value_t*)method,
-            jl_atomic_load_relaxed(&method->primary_world), jl_atomic_load_relaxed(&method->deleted_world));
+    assert((jl_atomic_load_relaxed(&method->dispatch_status) & METHOD_SIG_LATEST_WHICH) == 0);
+    assert((jl_atomic_load_relaxed(&method->dispatch_status) & METHOD_SIG_LATEST_ONLY) == 0);
+    newentry = jl_typemap_alloc((jl_tupletype_t*)method->sig, simpletype, jl_emptysvec, (jl_value_t*)method, ~(size_t)0, 1);
     jl_typemap_insert(&mt->defs, (jl_value_t*)mt, newentry, jl_cachearg_offset(mt));
     update_max_args(mt, method->sig);
     JL_UNLOCK(&mt->writelock);
@@ -2315,7 +2320,8 @@ void jl_method_table_activate(jl_methtable_t *mt, jl_typemap_entry_t *newentry)
     JL_LOCK(&mt->writelock);
     size_t world = jl_atomic_load_relaxed(&method->primary_world);
     assert(world == jl_atomic_load_relaxed(&jl_world_counter) + 1); // min-world
-    assert(jl_atomic_load_relaxed(&method->deleted_world) == ~(size_t)0); // max-world
+    assert((jl_atomic_load_relaxed(&method->dispatch_status) & METHOD_SIG_LATEST_WHICH) == 0);
+    assert((jl_atomic_load_relaxed(&method->dispatch_status) & METHOD_SIG_LATEST_ONLY) == 0);
     assert(jl_atomic_load_relaxed(&newentry->min_world) == ~(size_t)0);
     assert(jl_atomic_load_relaxed(&newentry->max_world) == 1);
     jl_atomic_store_relaxed(&newentry->min_world, world);
@@ -2330,12 +2336,17 @@ void jl_method_table_activate(jl_methtable_t *mt, jl_typemap_entry_t *newentry)
     // then check what entries we replaced
     oldvalue = get_intersect_matches(jl_atomic_load_relaxed(&mt->defs), newentry, &replaced, jl_cachearg_offset(mt), max_world);
     int invalidated = 0;
+    int only = !(jl_atomic_load_relaxed(&method->dispatch_status) & METHOD_SIG_PRECOMPILE_MANY); // will compute if this will be currently the only result that would returned from `ml_matches` given `sig`
     if (replaced) {
         oldvalue = (jl_value_t*)replaced;
+        jl_method_t *m = replaced->func.method;
         invalidated = 1;
-        method_overwrite(newentry, replaced->func.method);
+        method_overwrite(newentry, m);
         // this is an optimized version of below, given we know the type-intersection is exact
-        jl_method_table_invalidate(mt, replaced->func.method, max_world);
+        jl_method_table_invalidate(mt, m, max_world);
+        int m_dispatch = jl_atomic_load_relaxed(&m->dispatch_status);
+        jl_atomic_store_relaxed(&m->dispatch_status, 0);
+        only = m_dispatch & METHOD_SIG_LATEST_ONLY;
     }
     else {
         jl_method_t *const *d;
@@ -2407,8 +2418,10 @@ void jl_method_table_activate(jl_methtable_t *mt, jl_typemap_entry_t *newentry)
             memset(morespec, morespec_unknown, n);
             for (j = 0; j < n; j++) {
                 jl_method_t *m = d[j];
-                if (morespec[j] == (char)morespec_is)
+                if (morespec[j] == (char)morespec_is) {
+                    only = 0;
                     continue;
+                }
                 loctag = jl_atomic_load_relaxed(&m->specializations); // use loctag for a gcroot
                 _Atomic(jl_method_instance_t*) *data;
                 size_t l;
@@ -2438,7 +2451,7 @@ void jl_method_table_activate(jl_methtable_t *mt, jl_typemap_entry_t *newentry)
                             // not actually shadowing--the existing method is still better
                             break;
                         if (ambig == morespec_unknown)
-                            ambig = jl_type_morespecific(type, m->sig) ? morespec_is : morespec_isnot;
+                            ambig = jl_type_morespecific(type, m->sig) ? morespec_isnot : morespec_is;
                         // replacing a method--see if this really was the selected method previously
                         // over the intersection (not ambiguous) and the new method will be selected now (morespec_is)
                         int replaced_dispatch = is_replacing(ambig, type, m, d, n, isect, isect2, morespec);
@@ -2455,6 +2468,20 @@ void jl_method_table_activate(jl_methtable_t *mt, jl_typemap_entry_t *newentry)
                         invalidated |= invalidatedmi;
                     }
                 }
+                // now compute and store updates to METHOD_SIG_LATEST_ONLY
+                int m_dispatch = jl_atomic_load_relaxed(&m->dispatch_status);
+                if (m_dispatch & METHOD_SIG_LATEST_ONLY) {
+                    if (morespec[j] == (char)morespec_unknown)
+                        morespec[j] = (char)(jl_type_morespecific(m->sig, type) ? morespec_is : morespec_isnot);
+                    if (morespec[j] == (char)morespec_isnot)
+                        jl_atomic_store_relaxed(&m->dispatch_status, ~METHOD_SIG_LATEST_ONLY & m_dispatch);
+                }
+                if (only) {
+                    if (morespec[j] == (char)morespec_is || ambig == morespec_is ||
+                        (ambig == morespec_unknown && !jl_type_morespecific(type, m->sig))) {
+                        only = 0;
+                    }
+                }
             }
             if (jl_array_nrows(oldmi)) {
                 // search mt->cache and leafcache and drop anything that might overlap with the new method
@@ -2485,7 +2512,8 @@ void jl_method_table_activate(jl_methtable_t *mt, jl_typemap_entry_t *newentry)
         loctag = jl_cstr_to_string("jl_method_table_insert");
         jl_array_ptr_1d_push(_jl_debug_method_invalidation, loctag);
     }
-    jl_atomic_store_relaxed(&newentry->max_world, jl_atomic_load_relaxed(&method->deleted_world));
+    jl_atomic_store_relaxed(&newentry->max_world, ~(size_t)0);
+    jl_atomic_store_relaxed(&method->dispatch_status, METHOD_SIG_LATEST_WHICH | (only ? METHOD_SIG_LATEST_ONLY : 0)); // TODO: this should be sequenced fully after the world counter store
     JL_UNLOCK(&mt->writelock);
     JL_GC_POP();
 }
@@ -2499,7 +2527,6 @@ JL_DLLEXPORT void jl_method_table_insert(jl_methtable_t *mt, jl_method_t *method
         jl_error("Method changes have been disabled via a call to disable_new_worlds.");
     size_t world = jl_atomic_load_relaxed(&jl_world_counter) + 1;
     jl_atomic_store_relaxed(&method->primary_world, world);
-    jl_atomic_store_relaxed(&method->deleted_world, ~(size_t)0);
     jl_method_table_activate(mt, newentry);
     jl_atomic_store_release(&jl_world_counter, world);
     JL_UNLOCK(&world_counter_lock);
@@ -3882,6 +3909,8 @@ static int ml_matches_visitor(jl_typemap_entry_t *ml, struct typemap_intersectio
             closure->match.min_valid = max_world + 1;
         return 1;
     }
+    if (closure->match.max_valid > max_world)
+        closure->match.max_valid = max_world;
     jl_method_t *meth = ml->func.method;
     if (closure->lim >= 0 && jl_is_dispatch_tupletype(meth->sig)) {
         int replaced = 0;
@@ -4554,12 +4583,9 @@ static jl_value_t *ml_matches(jl_methtable_t *mt,
         jl_method_t *m = matc->method;
         // method applicability is the same as typemapentry applicability
         size_t min_world = jl_atomic_load_relaxed(&m->primary_world);
-        size_t max_world = jl_atomic_load_relaxed(&m->deleted_world);
         // intersect the env valid range with method lookup's inclusive valid range
         if (env.match.min_valid < min_world)
             env.match.min_valid = min_world;
-        if (env.match.max_valid > max_world)
-            env.match.max_valid = max_world;
     }
     if (mt && cache_result && ((jl_datatype_t*)unw)->isdispatchtuple) { // cache_result parameter keeps this from being recursive
         if (len == 1 && !has_ambiguity) {

src/jltypes.c

@@ -3545,8 +3545,8 @@ void jl_init_types(void) JL_GC_DISABLED
                             "module",
                             "file",
                             "line",
+                            "dispatch_status", // atomic
                             "primary_world", // atomic
-                            "deleted_world", // atomic
                             "sig",
                             "specializations", // !const
                             "speckeyset", // !const
@@ -3578,7 +3578,7 @@ void jl_init_types(void) JL_GC_DISABLED
                             jl_module_type,
                             jl_symbol_type,
                             jl_int32_type,
-                            jl_ulong_type,
+                            jl_int32_type,
                             jl_ulong_type,
                             jl_type_type,
                             jl_any_type, // union(jl_simplevector_type, jl_method_instance_type),

src/julia.h

@@ -331,8 +331,8 @@ typedef struct _jl_method_t {
     struct _jl_module_t *module;
     jl_sym_t *file;
     int32_t line;
+    _Atomic(int32_t) dispatch_status; // bits defined in staticdata.jl
     _Atomic(size_t) primary_world;
-    _Atomic(size_t) deleted_world;
 
     // method's type signature. redundant with TypeMapEntry->specTypes
     jl_value_t *sig;

src/julia_internal.h

@@ -677,6 +677,10 @@ typedef union {
 #define SOURCE_MODE_NOT_REQUIRED            0x0
 #define SOURCE_MODE_ABI                     0x1
 
+#define METHOD_SIG_LATEST_WHICH             0b0001
+#define METHOD_SIG_LATEST_ONLY              0b0010
+#define METHOD_SIG_PRECOMPILE_MANY          0b0100
+
 JL_DLLEXPORT jl_code_instance_t *jl_engine_reserve(jl_method_instance_t *m, jl_value_t *owner);
 JL_DLLEXPORT void jl_engine_fulfill(jl_code_instance_t *ci, jl_code_info_t *src);
 void jl_engine_sweep(jl_ptls_t *gc_all_tls_states) JL_NOTSAFEPOINT;

src/method.c

@@ -727,7 +727,7 @@ JL_DLLEXPORT jl_code_info_t *jl_code_for_staged(jl_method_instance_t *mi, size_t
     JL_TRY {
         ct->ptls->in_pure_callback = 1;
         ct->world_age = jl_atomic_load_relaxed(&def->primary_world);
-        if (ct->world_age > jl_atomic_load_acquire(&jl_world_counter) || jl_atomic_load_relaxed(&def->deleted_world) < ct->world_age)
+        if (ct->world_age > jl_atomic_load_acquire(&jl_world_counter))
             jl_error("The generator method cannot run until it is added to a method table.");
 
         // invoke code generator
@@ -1006,7 +1006,7 @@ JL_DLLEXPORT jl_method_t *jl_new_method_uninit(jl_module_t *module)
     m->isva = 0;
     m->nargs = 0;
     jl_atomic_store_relaxed(&m->primary_world, ~(size_t)0);
-    jl_atomic_store_relaxed(&m->deleted_world, 1);
+    jl_atomic_store_relaxed(&m->dispatch_status, 0);
     m->is_for_opaque_closure = 0;
     m->nospecializeinfer = 0;
     jl_atomic_store_relaxed(&m->did_scan_source, 0);

src/opaque_closure.c

@@ -159,7 +159,6 @@ JL_DLLEXPORT jl_opaque_closure_t *jl_new_opaque_closure_from_code_info(jl_tuplet
     size_t world = jl_current_task->world_age;
     // these are only legal in the current world since they are not in any tables
     jl_atomic_store_release(&meth->primary_world, world);
-    jl_atomic_store_release(&meth->deleted_world, world);
 
     if (isinferred) {
         jl_value_t *argslotty = jl_array_ptr_ref(ci->slottypes, 0);