optimizer: cancel finalizer registration rather than inserting finalize

This seems to perform well.

```julia
mutable struct AtomicCounter
    @atomic count::Int
end
const counter = AtomicCounter(0)

function withfinalizer(x)
    xs = finalizer(Ref(x)) do obj
        Base.@assume_effects :nothrow :notaskstate
        @atomic counter.count += obj[]
    end
    println(devnull, xs[])
    xs[] = xs[]
    println(devnull, xs[])
    return xs[]
end

@benchmark withfinalizer(100)
```

> master
```
BenchmarkTools.Trial: 10000 samples with 867 evaluations.
 Range (min … max):  135.717 ns …  18.617 μs  ┊ GC (min … max):  0.00% … 47.51%
 Time  (median):     144.992 ns               ┊ GC (median):     0.00%
 Time  (mean ± σ):   192.949 ns ± 687.513 ns  ┊ GC (mean ± σ):  11.57% ±  3.23%

     ▇  ▃█▅▂
  ▁▁▂█▇▂████▇▆▅▄▄▄▅▄▄▂▂▂▂▂▁▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▂
  136 ns           Histogram: frequency by time          201 ns <

 Memory estimate: 208 bytes, allocs estimate: 7.
```

> this commit
```
BenchmarkTools.Trial: 10000 samples with 964 evaluations.
 Range (min … max):   83.506 ns …   3.408 μs  ┊ GC (min … max):  0.00% … 96.85%
 Time  (median):      88.780 ns               ┊ GC (median):     0.00%
 Time  (mean ± σ):   101.797 ns ± 137.067 ns  ┊ GC (mean ± σ):  10.26% ±  7.31%

    ▄█▃▅
  ▄▃████▅▅▆▅▄▄▄▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▂▁▂▂▁▁▁▁▁▁▂▁▂▂▁▂▂ ▃
  83.5 ns          Histogram: frequency by time          145 ns <

 Memory estimate: 208 bytes, allocs estimate: 7.
```

Author: aviatesk

base/boot.jl

@@ -214,7 +214,7 @@ export
     nfields, throw, tuple, ===, isdefined, eval,
     # access to globals
     getglobal, setglobal!, swapglobal!, modifyglobal!, replaceglobal!, setglobalonce!,
-    # ifelse, sizeof, finalize # not exported, to avoid conflicting with Base
+    # ifelse, sizeof    # not exported, to avoid conflicting with Base
     # type reflection
     <:, typeof, isa, typeassert,
     # method reflection

base/compiler/abstractinterpretation.jl

@@ -2280,10 +2280,10 @@ function abstract_finalizer(interp::AbstractInterpreter, argtypes::Vector{Any},
         finalizer_argvec = Any[argtypes[2], argtypes[3]]
         call = abstract_call(interp, ArgInfo(nothing, finalizer_argvec), StmtInfo(false), sv, #=max_methods=#1)::Future
         return Future{CallMeta}(call, interp, sv) do call, interp, sv
-            return CallMeta(Nothing, Any, Effects(), FinalizerInfo(call.info, call.effects))
+            return CallMeta(Int, Any, Effects(), FinalizerInfo(call.info, call.effects))
         end
     end
-    return Future(CallMeta(Nothing, Any, Effects(), NoCallInfo()))
+    return Future(CallMeta(Int, Any, Effects(), NoCallInfo()))
 end
 
 function abstract_throw(interp::AbstractInterpreter, argtypes::Vector{Any}, ::AbsIntState)

base/compiler/ssair/passes.jl

@@ -1626,11 +1626,9 @@ function try_resolve_finalizer!(ir::IRCode, alloc_idx::Int, finalizer_idx::Int,
     insert_bb != 0 || return nothing # verify post-dominator of all uses exists
 
     # Figure out the exact statement where we're going to inline the finalizer.
-    loc = insert_idx === nothing ? first(ir.cfg.blocks[insert_bb].stmts) : insert_idx::Int
-    attach_after = insert_idx !== nothing
-    flag = info isa FinalizerInfo ? flags_for_effects(info.effects) : IR_FLAG_NULL
     finalizer_stmt = ir[SSAValue(finalizer_idx)][:stmt]
 
+    current_task_ssa = nothing
     if !OptimizationParams(inlining.interp).assume_fatal_throw
         # Collect all reachable blocks between the finalizer registration and the
         # insertion point
@@ -1655,15 +1653,25 @@ function try_resolve_finalizer!(ir::IRCode, alloc_idx::Int, finalizer_idx::Int,
 
             # An exception may be thrown between the finalizer registration and the point
             # where the object’s lifetime ends (`insert_idx`): In such cases, we can’t
-            # remove the finalizer registration, but we can still insert a `Core.finalize`
-            # call at `insert_idx` while leaving the registration intact.
-            newinst = add_flag(NewInstruction(Expr(:call, GlobalRef(Core, :finalize), finalizer_stmt.args[3]), Nothing), flag)
-            insert_node!(ir, loc, newinst, attach_after)
-            return nothing
+            # remove the finalizer registration, but we can still inline the finalizer
+            # with inserting `Core._cancel_finalizer` at the end.
+            # Here, prepare a reference to the current task object that should be passed to
+            # `Core._cancel_finalizer` and insert it into `Core.finalizer` so that the
+            # finalizer is added to the ptls of the current task.
+            current_task_stmt = Expr(:foreigncall, QuoteNode(:jl_get_current_task),
+                Core.Ref{Core.Task}, Core.svec(), 0, QuoteNode(:ccall))
+            newinst = NewInstruction(current_task_stmt, Core.Task)
+            current_task_ssa = insert_node!(ir, finalizer_idx, newinst)
+            push!(finalizer_stmt.args, current_task_ssa)
+            break
         end
     end
 
-    argexprs = Any[finalizer_stmt.args[2], finalizer_stmt.args[3]]
+    loc = insert_idx === nothing ? first(ir.cfg.blocks[insert_bb].stmts) : insert_idx::Int
+    attach_after = insert_idx !== nothing
+    flag = info isa FinalizerInfo ? flags_for_effects(info.effects) : IR_FLAG_NULL
+    alloc_obj = finalizer_stmt.args[3]
+    argexprs = Any[finalizer_stmt.args[2], alloc_obj]
     if length(finalizer_stmt.args) >= 4
         inline = finalizer_stmt.args[4]
         if inline === nothing
@@ -1681,8 +1689,16 @@ function try_resolve_finalizer!(ir::IRCode, alloc_idx::Int, finalizer_idx::Int,
         newinst = add_flag(NewInstruction(Expr(:call, argexprs...), Nothing), flag)
         insert_node!(ir, loc, newinst, attach_after)
     end
-    # Erase the call to `finalizer`
-    ir[SSAValue(finalizer_idx)][:stmt] = nothing
+    cancel_registration = current_task_ssa !== nothing
+    if cancel_registration
+        lookup_idx_ssa = SSAValue(finalizer_idx)
+        finalize_call = Expr(:call, GlobalRef(Core, :_cancel_finalizer), alloc_obj, current_task_ssa, lookup_idx_ssa)
+        newinst = add_flag(NewInstruction(finalize_call, Nothing), flag)
+        insert_node!(ir, loc, newinst, #=attach_after=#true)
+    else
+        # Erase the call to `finalizer`
+        ir[SSAValue(finalizer_idx)][:stmt] = nothing
+    end
     return nothing
 end
 

base/compiler/tfuncs.jl

@@ -742,8 +742,13 @@ add_tfunc(donotdelete, 0, INT_INF, @nospecs((𝕃::AbstractLattice, args...)->No
     end
 end
 add_tfunc(compilerbarrier, 2, 2, compilerbarrier_tfunc, 5)
-add_tfunc(Core.finalizer, 2, 4, @nospecs((𝕃::AbstractLattice, args...)->Nothing), 5)
-add_tfunc(Core.finalize, 1, 1, @nospecs((𝕃::AbstractLattice, o)->Nothing), 100)
+add_tfunc(Core.finalizer, 2, 5, @nospecs((𝕃::AbstractLattice, args...)->Int), 5)
+@nospecs function _cancel_finalizer_tfunc(𝕃::AbstractLattice, o, ct, lookup_idx)
+    hasintersect(widenconst(ct), Task) || return Bottom
+    hasintersect(widenconst(lookup_idx), Int) || return Bottom
+    return Nothing
+end
+add_tfunc(Core._cancel_finalizer, 3, 3, _cancel_finalizer_tfunc, 5)
 
 @nospecs function compilerbarrier_nothrow(setting, val)
     return isa(setting, Const) && contains_is((:type, :const, :conditional), setting.val)
@@ -2288,12 +2293,12 @@ function _builtin_nothrow(𝕃::AbstractLattice, @nospecialize(f::Builtin), argt
     elseif f === donotdelete
         return true
     elseif f === Core.finalizer
-        2 <= na <= 4 || return false
+        2 <= na <= 5 || return false
         # `Core.finalizer` does no error checking - that's done in Base.finalizer
         return true
-    elseif f === Core.finalize
-        na == 2 || return false
-        return true # `Core.finalize` does no error checking
+    elseif f === Core._cancel_finalizer
+        na == 3 || return false
+        return argtypes[2] ⊑ Task && argtypes[3] ⊑ Int
     elseif f === Core.compilerbarrier
         na == 2 || return false
         return compilerbarrier_nothrow(argtypes[1], nothing)

base/gcutils.jl

@@ -98,7 +98,7 @@ end
 
 Immediately run finalizers registered for object `x`.
 """
-finalize(@nospecialize(o)) = Core.finalize(o)
+finalize(@nospecialize(o)) = ccall(:jl_finalize_th, Cvoid, (Any, Any,), current_task(), o)
 
 """
     Base.GC

src/builtin_proto.h

@@ -26,6 +26,7 @@ DECLARE_BUILTIN(_apply_pure);
 DECLARE_BUILTIN(_call_in_world);
 DECLARE_BUILTIN(_call_in_world_total);
 DECLARE_BUILTIN(_call_latest);
+DECLARE_BUILTIN(_cancel_finalizer);
 DECLARE_BUILTIN(_compute_sparams);
 DECLARE_BUILTIN(_expr);
 DECLARE_BUILTIN(_svec_ref);
@@ -37,7 +38,6 @@ DECLARE_BUILTIN(compilerbarrier);
 DECLARE_BUILTIN(current_scope);
 DECLARE_BUILTIN(donotdelete);
 DECLARE_BUILTIN(fieldtype);
-DECLARE_BUILTIN(finalize);
 DECLARE_BUILTIN(finalizer);
 DECLARE_BUILTIN(getfield);
 DECLARE_BUILTIN(getglobal);

src/builtins.c

@@ -2004,16 +2004,25 @@ JL_CALLABLE(jl_f_compilerbarrier)
 JL_CALLABLE(jl_f_finalizer)
 {
     // NOTE the compiler may temporarily insert additional argument for the later inlining pass
-    JL_NARGS(finalizer, 2, 4);
+    JL_NARGS(finalizer, 2, 5);
+    if (nargs == 5 && jl_is_task(args[4]))
+        // There are cases where the compiler inserts `current_task` as the 5th argument,
+        // and in such cases, the finalizer is added to the `ptls` of that task.
+        // This task is later referenced by `_cancel_finalizer`.
+        return jl_box_long(jl_gc_add_finalizer_(((jl_task_t*)args[4])->ptls, args[1], args[0]));
     jl_task_t *ct = jl_current_task;
-    jl_gc_add_finalizer_(ct->ptls, args[1], args[0]);
-    return jl_nothing;
+    return jl_box_long(jl_gc_add_finalizer_(ct->ptls, args[1], args[0]));
 }
 
-JL_CALLABLE(jl_f_finalize)
+JL_CALLABLE(jl_f__cancel_finalizer)
 {
-    JL_NARGS(finalize, 1, 1);
-    jl_finalize(args[0]);
+    JL_NARGS(_cancel_finalizer, 3, 3);
+    JL_TYPECHK(_cancel_finalizer, task, args[1])
+    JL_TYPECHK(_cancel_finalizer, long, args[2])
+    jl_value_t *o = args[0];
+    jl_task_t *ct = (jl_task_t*)args[1];
+    size_t lookup_idx = jl_unbox_long(args[2]);
+    jl_cancel_finalizer(o, ct, lookup_idx);
     return jl_nothing;
 }
 
@@ -2449,7 +2458,7 @@ void jl_init_primitives(void) JL_GC_DISABLED
     jl_builtin_donotdelete = add_builtin_func("donotdelete", jl_f_donotdelete);
     jl_builtin_compilerbarrier = add_builtin_func("compilerbarrier", jl_f_compilerbarrier);
     add_builtin_func("finalizer", jl_f_finalizer);
-    add_builtin_func("finalize", jl_f_finalize);
+    add_builtin_func("_cancel_finalizer", jl_f__cancel_finalizer);
     add_builtin_func("_compute_sparams", jl_f__compute_sparams);
     add_builtin_func("_svec_ref", jl_f__svec_ref);
     jl_builtin_current_scope = add_builtin_func("current_scope", jl_f_current_scope);

src/codegen.cpp

@@ -1646,7 +1646,7 @@ static const auto &builtin_func_map() {
           { jl_f_donotdelete_addr,          new JuliaFunction<>{XSTR(jl_f_donotdelete), get_donotdelete_sig, get_donotdelete_func_attrs} },
           { jl_f_compilerbarrier_addr,      new JuliaFunction<>{XSTR(jl_f_compilerbarrier), get_func_sig, get_func_attrs} },
           { jl_f_finalizer_addr,            new JuliaFunction<>{XSTR(jl_f_finalizer), get_func_sig, get_func_attrs} },
-          { jl_f_finalize_addr,             new JuliaFunction<>{XSTR(jl_f_finalize), get_func_sig, get_func_attrs} },
+          { jl_f__cancel_finalizer_addr,    new JuliaFunction<>{XSTR(jl_f__cancel_finalizer), get_func_sig, get_func_attrs} },
           { jl_f__svec_ref_addr,            new JuliaFunction<>{XSTR(jl_f__svec_ref), get_func_sig, get_func_attrs} },
           { jl_f_current_scope_addr,        new JuliaFunction<>{XSTR(jl_f_current_scope), get_func_sig, get_func_attrs} },
         };

src/gc-common.c

@@ -387,7 +387,7 @@ void jl_gc_run_all_finalizers(jl_task_t *ct)
     run_finalizers(ct, 1);
 }
 
-void jl_gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT
+size_t jl_gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT
 {
     assert(jl_atomic_load_relaxed(&ptls->gc_state) == JL_GC_STATE_UNSAFE);
     arraylist_t *a = &ptls->finalizers;
@@ -413,6 +413,7 @@ void jl_gc_add_finalizer_(jl_ptls_t ptls, void *v, void *f) JL_NOTSAFEPOINT
     items[oldlen] = v;
     items[oldlen + 1] = f;
     jl_atomic_store_release((_Atomic(size_t)*)&a->len, oldlen + 2);
+    return oldlen;
 }
 
 JL_DLLEXPORT void jl_gc_add_ptr_finalizer(jl_ptls_t ptls, jl_value_t *v, void *f) JL_NOTSAFEPOINT
@@ -463,13 +464,11 @@ JL_DLLEXPORT void jl_finalize_th(jl_task_t *ct, jl_value_t *o)
             finalize_object(&ptls2->finalizers, o, &copied_list, jl_atomic_load_relaxed(&ct->tid) != i);
     }
     finalize_object(&finalizer_list_marked, o, &copied_list, 0);
-    if (copied_list.len > 0) {
+    if (copied_list.len > 0)
         // This releases the finalizers lock.
         jl_gc_run_finalizers_in_list(ct, &copied_list);
-    }
-    else {
+    else
         JL_UNLOCK_NOGC(&finalizers_lock);
-    }
     arraylist_free(&copied_list);
 }
 
@@ -478,6 +477,46 @@ JL_DLLEXPORT void jl_finalize(jl_value_t *o)
     jl_finalize_th(jl_current_task, o);
 }
 
+int erase_finalizer_at(arraylist_t *list, jl_value_t *o, size_t idx)
+{
+    void **items = list->items;
+    void *v = items[idx];
+    if (o == (jl_value_t*)gc_ptr_clear_tag(v, 1)) {
+        for (size_t j = idx + 2; j < list->len; j += 2) {
+            items[j-2] = items[j];
+            items[j-1] = items[j+1];
+        }
+        list->len = list->len - 2;
+        return 1;
+    }
+    return 0;
+}
+
+int erase_finalizer(arraylist_t *list, jl_value_t *o)
+{
+    for (size_t i = 0; i < list->len; i += 2) {
+        if (erase_finalizer_at(list, o, i))
+            return 1;
+    }
+    return 0;
+}
+
+// Remove the finalizer for `o` from `ct->ptls->finalizers` and cancel that finalizer.
+// `lookup_idx` is the index at which the finalizer was registered in `ct->ptls->finalizers`
+// by `jl_gc_add_finalizer_` (used for the fast path).
+// Note that it is assumed that only a single finalizer exists for `o`.
+void jl_cancel_finalizer(jl_value_t *o, jl_task_t *ct, size_t lookup_idx)
+{
+    arraylist_t *list = &ct->ptls->finalizers;
+    // fast path
+    if (lookup_idx < list->len && erase_finalizer_at(list, o, lookup_idx))
+        return;
+    // slow path
+    if (erase_finalizer(list, o) || erase_finalizer(&finalizer_list_marked, o))
+        return;
+    assert(0 && "finalizer not found");
+}
+
 // =========================================================================== //
 // Threading
 // =========================================================================== //

src/julia.h

@@ -2326,6 +2326,8 @@ JL_DLLEXPORT void JL_NORETURN jl_no_exc_handler(jl_value_t *e, jl_task_t *ct);
 JL_DLLEXPORT JL_CONST_FUNC jl_gcframe_t **(jl_get_pgcstack)(void) JL_GLOBALLY_ROOTED JL_NOTSAFEPOINT;
 #define jl_current_task (container_of(jl_get_pgcstack(), jl_task_t, gcstack))
 
+void jl_cancel_finalizer(jl_value_t *o, jl_task_t *ct, size_t lookup_idx);
+
 extern JL_DLLIMPORT int jl_task_gcstack_offset;
 extern JL_DLLIMPORT int jl_task_ptls_offset;