Allow querying if the current GC may move objects. (#1128)

The major changes include:

-   Adding a method `Plan::current_gc_may_move_object`.
-   Immix-based plans now reset the in-defrag state at the end of GC.

The main change of this PR is allowing querying if the current GC may
move any object.

Whether a plan moves objects or not is not a simple "yes or no"
question. It may vary in each GC.

-   Some plans, such as MarkSweep, never moves any object.
- Some plans, such as SemiSpace and MarkCompact, moves objects in every
GC.
- Some plans, such as Immix, moves object in defrag GC, while "fast" GCs
are non-moving.
- Some plans, such as Immix and StickyImmix, depends on Cargo features.
"immix_non_moving" prevents all movement, and
"sticky_immix_non_moving_nursery" prevents movement in nursery GCs.

The information of whether the current GC is useful inside mmtk-core.
For example, we can skip clearing on-the-side forwarding bits in
non-moving GCs because no objects will be forwarded. Because this should
happen during the Release stage, we postponed the time for Immix-based
plans to reset the in-defrag state to the end of GC so that
`Plan::current_gc_may_move_object` remains callable during the Release
stage. *(Note that this PR does not fix
#1118, but this mechanism
introduced in this PR can be used to implement the clearing of side
forwarding bits efficiently.)*

The information is also useful for VM bindings. Therefore,
`Plan::current_gc_may_move_object` is part of the public API.

- For VMs that have "potential pinning parents" (PPP, i.e. non-root
objects that have fields that cannot be updated), the VM binding must
pin the children of those fields before starting tracing. If the VM
binding knows the current GC never moves any object, it can skip pinning
the children of PPPs.
- For VMs that maintain weak tables (tables that contain weak references
to heap objects) in the runtime, the VM binding needs to scan the weak
tables, updating entries for objects that are moved, and deleting
entries for unreachable objects. (Note the cost of updating entries may
be high if the table is hashed by address.) If the VM binding knows the
current GC never moves any object, it will not need to update the
entries for moved live objects. And if the current GC is also a nursery
GC, the VM binding will only need to scan entries added before the last
GC.

This new method complements existing mechanisms for checking if objects
can be moved. They are useful for different purposes and cannot replace
each other.

- The scope of the new `Plan::current_gc_may_move_object` is one GC. It
is useful for the VM to decide whether to pin PPPs objects before a GC
and how to process weak references or weak tables.
- `PlanConstraints::moves_objects` is a per-plan constant. It is useful
for the VM to determine the object layout w.r.t. address-based hashing.
- The scope of `ObjectReference::is_movable` is a single object. It is
useful for the VM to skip pinning certain objects when interacting with
native code.
- The scope of `PlanTraceObject::may_move_object` is a trace. It is an
internal mechanism for MMTk core to specialize work packets for
different kinds of traces.

Author: wks

docs/userguide/src/tutorial/code/mygc_semispace/global.rs

@@ -71,6 +71,13 @@ impl<VM: VMBinding> Plan for MyGC<VM> {
     }
     // ANCHOR_END: create_copy_config
 
+    // Modify
+    // ANCHOR: current_gc_may_move_object
+    fn current_gc_may_move_object(&self) -> bool {
+        true
+    }
+    // ANCHOR_END: current_gc_may_move_object
+
     // Modify
     // ANCHOR: schedule_collection
     fn schedule_collection(&'static self, scheduler: &GCWorkScheduler<VM>) {

docs/userguide/src/tutorial/mygc/ss/collection.md

@@ -29,6 +29,13 @@ space here.
 {{#include ../../code/mygc_semispace/global.rs:create_copy_config}}
 ```
 
+Because the semispace GC copies objects in every single GC, we modify the method
+`current_gc_may_move_object()` in `MyGC` so that it always returns `true`.
+
+```rust
+{{#include ../../code/mygc_semispace/global.rs:current_gc_may_move_object}}
+```
+
 ## Introduce collection to MyGC plan
 
 Add a new method to `Plan for MyGC`, `schedule_collection()`. This function 

src/plan/generational/copying/global.rs

@@ -127,6 +127,10 @@ impl<VM: VMBinding> Plan for GenCopy<VM> {
         self.gen.get_used_pages() + self.tospace().reserved_pages()
     }
 
+    fn current_gc_may_move_object(&self) -> bool {
+        true
+    }
+
     /// Return the number of pages available for allocation. Assuming all future allocations goes to nursery.
     fn get_available_pages(&self) -> usize {
         // super.get_available_pages() / 2 to reserve pages for copying

src/plan/generational/immix/global.rs

@@ -137,10 +137,7 @@ impl<VM: VMBinding> Plan for GenImmix<VM> {
         let full_heap = !self.gen.is_current_gc_nursery();
         self.gen.release(tls);
         if full_heap {
-            let did_defrag = self.immix_space.release(full_heap);
-            self.last_gc_was_defrag.store(did_defrag, Ordering::Relaxed);
-        } else {
-            self.last_gc_was_defrag.store(false, Ordering::Relaxed);
+            self.immix_space.release(full_heap);
         }
         self.last_gc_was_full_heap
             .store(full_heap, Ordering::Relaxed);
@@ -149,6 +146,17 @@ impl<VM: VMBinding> Plan for GenImmix<VM> {
     fn end_of_gc(&mut self, _tls: VMWorkerThread) {
         self.gen
             .set_next_gc_full_heap(CommonGenPlan::should_next_gc_be_full_heap(self));
+
+        let did_defrag = self.immix_space.end_of_gc();
+        self.last_gc_was_defrag.store(did_defrag, Ordering::Relaxed);
+    }
+
+    fn current_gc_may_move_object(&self) -> bool {
+        if self.is_current_gc_nursery() {
+            true
+        } else {
+            self.immix_space.in_defrag()
+        }
     }
 
     fn get_collection_reserved_pages(&self) -> usize {

src/plan/global.rs

@@ -299,6 +299,16 @@ pub trait Plan: 'static + HasSpaces + Sync + Downcast {
         true
     }
 
+    /// Return whether the current GC may move any object.  The VM binding can make use of this
+    /// information and choose to or not to update some data structures that record the addresses
+    /// of objects.
+    ///
+    /// This function is callable during a GC.  From the VM binding's point of view, the information
+    /// of whether the current GC moves object or not is available since `Collection::stop_mutators`
+    /// is called, and remains available until (but not including) `resume_mutators` at which time
+    /// the current GC has just finished.
+    fn current_gc_may_move_object(&self) -> bool;
+
     /// An object is firstly reached by a sanity GC. So the object is reachable
     /// in the current GC, and all the GC work has been done for the object (such as
     /// tracing and releasing). A plan can implement this to

src/plan/immix/global.rs

@@ -93,8 +93,16 @@ impl<VM: VMBinding> Plan for Immix<VM> {
     fn release(&mut self, tls: VMWorkerThread) {
         self.common.release(tls, true);
         // release the collected region
+        self.immix_space.release(true);
+    }
+
+    fn end_of_gc(&mut self, _tls: VMWorkerThread) {
         self.last_gc_was_defrag
-            .store(self.immix_space.release(true), Ordering::Relaxed);
+            .store(self.immix_space.end_of_gc(), Ordering::Relaxed);
+    }
+
+    fn current_gc_may_move_object(&self) -> bool {
+        self.immix_space.in_defrag()
     }
 
     fn get_collection_reserved_pages(&self) -> usize {

src/plan/markcompact/global.rs

@@ -171,6 +171,10 @@ impl<VM: VMBinding> Plan for MarkCompact<VM> {
     fn get_collection_reserved_pages(&self) -> usize {
         0
     }
+
+    fn current_gc_may_move_object(&self) -> bool {
+        true
+    }
 }
 
 impl<VM: VMBinding> MarkCompact<VM> {

src/plan/marksweep/global.rs

@@ -73,6 +73,10 @@ impl<VM: VMBinding> Plan for MarkSweep<VM> {
         self.base().collection_required(self, space_full)
     }
 
+    fn current_gc_may_move_object(&self) -> bool {
+        false
+    }
+
     fn get_used_pages(&self) -> usize {
         self.common.get_used_pages() + self.ms.reserved_pages()
     }

src/plan/nogc/global.rs

@@ -74,6 +74,10 @@ impl<VM: VMBinding> Plan for NoGC<VM> {
         unreachable!("GC triggered in nogc")
     }
 
+    fn current_gc_may_move_object(&self) -> bool {
+        false
+    }
+
     fn get_used_pages(&self) -> usize {
         self.nogc_space.reserved_pages()
             + self.immortal.reserved_pages()

src/plan/pageprotect/global.rs

@@ -62,6 +62,10 @@ impl<VM: VMBinding> Plan for PageProtect<VM> {
         self.base().collection_required(self, space_full)
     }
 
+    fn current_gc_may_move_object(&self) -> bool {
+        false
+    }
+
     fn get_used_pages(&self) -> usize {
         self.space.reserved_pages() + self.common.get_used_pages()
     }