Basic infrastructure for implementing a prototype moving gc.

- Refactor the stack logic into a separate data structure ChunkedStack,
  keep the push function that deals with marking as functionality specific
  to a subclass of ChunkedStack (no virtual, just reuse code).
- Add an additional (disabled) phase that maps out the reference graph
  reachable objects after the sweep phase. Prepare moving all objects in
  the small heap.

Also, rename TraceStack to TraversalWorklist because the fact
that we're using a stack is an implementation detail.

src/gc/collector.cpp

@@ -30,6 +30,7 @@
 #include "runtime/hiddenclass.h"
 #include "runtime/objmodel.h"
 #include "runtime/types.h"
+#include "runtime/util.h"
 
 #ifndef NVALGRIND
 #include "valgrind.h"
@@ -67,14 +68,12 @@ static int ncollections = 0;
 static bool gc_enabled = true;
 static bool should_not_reenter_gc = false;
 
-enum TraceStackType {
-    MarkPhase,
-    FinalizationOrderingFindReachable,
-    FinalizationOrderingRemoveTemporaries,
-};
-
-class TraceStack {
-private:
+// This is basically a stack. However, for optimization purposes,
+// blocks of memory are allocated at once when things need to be pushed.
+//
+// For performance, this should not have virtual methods.
+class ChunkedStack {
+protected:
     const int CHUNK_SIZE = 256;
     const int MAX_FREE_CHUNKS = 50;
 
@@ -85,8 +84,6 @@ private:
     void** start;
     void** end;
 
-    TraceStackType visit_type;
-
     void get_chunk() {
         if (free_chunks.size()) {
             start = free_chunks.back();
@@ -98,12 +95,14 @@ private:
         cur = start;
         end = start + CHUNK_SIZE;
     }
+
     void release_chunk(void** chunk) {
         if (free_chunks.size() == MAX_FREE_CHUNKS)
             free(chunk);
         else
             free_chunks.push_back(chunk);
     }
+
     void pop_chunk() {
         start = chunks.back();
         chunks.pop_back();
@@ -111,29 +110,79 @@ private:
         cur = end;
     }
 
-public:
-    TraceStack(TraceStackType type) : visit_type(type) { get_chunk(); }
-    TraceStack(TraceStackType type, const std::unordered_set<void*>& roots) : visit_type(type) {
-        get_chunk();
-        for (void* p : roots) {
-            ASSERT(!isMarked(GCAllocation::fromUserData(p)), "");
-            push(p);
+    void* pop_chunk_and_item() {
+        release_chunk(start);
+        if (chunks.size()) {
+            pop_chunk();
+            assert(cur == end);
+            return *--cur; // no need for any bounds checks here since we're guaranteed we're CHUNK_SIZE from the start
+        } else {
+            // We emptied the stack, but we should prepare a new chunk in case another item
+            // gets added onto the stack.
+            get_chunk();
+            return NULL;
         }
     }
-    ~TraceStack() {
-        RELEASE_ASSERT(end - cur == CHUNK_SIZE, "destroying non-empty TraceStack");
 
-        // We always have a block available in case we want to push items onto the TraceStack,
+public:
+    ChunkedStack() { get_chunk(); }
+    ~ChunkedStack() {
+        RELEASE_ASSERT(end - cur == CHUNK_SIZE, "destroying non-empty ChunkedStack");
+        // We always have a block available in case we want to push items onto the TraversalWorklist,
         // but that chunk needs to be released after use to avoid a memory leak.
         release_chunk(start);
     }
 
+    void* pop() {
+        if (cur > start)
+            return *--cur;
+
+        return pop_chunk_and_item();
+    }
+
     void push(void* p) {
+        *cur++ = p;
+        if (cur == end) {
+            chunks.push_back(start);
+            get_chunk();
+        }
+    }
+};
+std::vector<void**> ChunkedStack::free_chunks;
+
+enum TraversalType {
+    MarkPhase,
+    FinalizationOrderingFindReachable,
+    FinalizationOrderingRemoveTemporaries,
+    MapReferencesPhase,
+};
+
+class Worklist {
+protected:
+    ChunkedStack stack;
+
+public:
+    void* next() { return stack.pop(); }
+};
+
+class TraversalWorklist : public Worklist {
+    TraversalType visit_type;
+
+public:
+    TraversalWorklist(TraversalType type) : visit_type(type) {}
+    TraversalWorklist(TraversalType type, const std::unordered_set<void*>& roots) : TraversalWorklist(type) {
+        for (void* p : roots) {
+            ASSERT(!isMarked(GCAllocation::fromUserData(p)), "");
+            addWork(p);
+        }
+    }
+
+    void addWork(void* p) {
         GC_TRACE_LOG("Pushing %p\n", p);
         GCAllocation* al = GCAllocation::fromUserData(p);
 
         switch (visit_type) {
-            case TraceStackType::MarkPhase:
+            case TraversalType::MarkPhase:
 // Use this to print the directed edges of the GC graph traversal.
 // i.e. print every a -> b where a is a pointer and b is something a references
 #if 0
@@ -162,7 +211,7 @@ public:
                 break;
             // See PyPy's finalization ordering algorithm:
             // http://pypy.readthedocs.org/en/latest/discussion/finalizer-order.html
-            case TraceStackType::FinalizationOrderingFindReachable:
+            case TraversalType::FinalizationOrderingFindReachable:
                 if (orderingState(al) == FinalizationState::UNREACHABLE) {
                     setOrderingState(al, FinalizationState::TEMPORARY);
                 } else if (orderingState(al) == FinalizationState::REACHABLE_FROM_FINALIZER) {
@@ -171,7 +220,7 @@ public:
                     return;
                 }
                 break;
-            case TraceStackType::FinalizationOrderingRemoveTemporaries:
+            case TraversalType::FinalizationOrderingRemoveTemporaries:
                 if (orderingState(al) == FinalizationState::TEMPORARY) {
                     setOrderingState(al, FinalizationState::REACHABLE_FROM_FINALIZER);
                 } else {
@@ -182,36 +231,62 @@ public:
                 assert(false);
         }
 
-        *cur++ = p;
-        if (cur == end) {
-            chunks.push_back(start);
-            get_chunk();
-        }
+        stack.push(p);
     }
+};
 
-    void* pop_chunk_and_item() {
-        release_chunk(start);
-        if (chunks.size()) {
-            pop_chunk();
-            assert(cur == end);
-            return *--cur; // no need for any bounds checks here since we're guaranteed we're CHUNK_SIZE from the start
-        } else {
-            // We emptied the stack, but we should prepare a new chunk in case another item
-            // gets added onto the stack.
-            get_chunk();
-            return NULL;
+class ReferenceMapWorklist : public Worklist {
+    ReferenceMap* refmap;
+
+public:
+    ReferenceMapWorklist(ReferenceMap* refmap) : refmap(refmap) {}
+    ReferenceMapWorklist(ReferenceMap* refmap, const std::unordered_set<void*>& roots) : refmap(refmap) {
+        for (void* p : roots) {
+            addWork(GCAllocation::fromUserData(p), NULL);
         }
     }
 
+    void addWork(GCAllocation* al, GCAllocation* source) {
+        assert(refmap);
 
-    void* pop() {
-        if (cur > start)
-            return *--cur;
+        auto it = refmap->references.find(al);
+        if (it == refmap->references.end()) {
+            refmap->references.emplace(al, std::vector<GCAllocation*>());
+            auto& vec = refmap->references[al];
 
-        return pop_chunk_and_item();
+            if (source) {
+                // We found that there exists a pointer from `source` to `al`
+                vec.push_back(source);
+            } else {
+                // No source => this is a root. We should pin roots.
+                refmap->pinned.emplace(al);
+            }
+
+            // Pin these types of objects - they are likely to be untracked at
+            // this time.
+            if (al->kind_id == GCKind::RUNTIME) {
+                pin(al);
+            } else if (al->kind_id == GCKind::PYTHON) {
+                Box* b = (Box*)al->user_data;
+                if (b->cls == type_cls || b->cls == module_cls) {
+                    pin(al);
+                }
+            }
+
+            stack.push(al->user_data);
+        } else {
+            if (source) {
+                // We found that there exists a pointer from `source` to `al`
+                it->second.push_back(source);
+            } else {
+                // No source => this is a root. We should pin roots.
+                pin(al);
+            }
+        }
     }
+
+    void pin(GCAllocation* al) { refmap->pinned.emplace(al); }
 };
-std::vector<void**> TraceStack::free_chunks;
 
 void registerPermanentRoot(void* obj, bool allow_duplicates) {
     assert(global_heap.getAllocationFromInteriorPointer(obj));
@@ -328,7 +403,7 @@ void GCVisitor::_visit(void** ptr_address) {
     }
 
     ASSERT(global_heap.getAllocationFromInteriorPointer(p)->user_data == p, "%p", p);
-    stack->push(p);
+    worklist->addWork(p);
 }
 
 void GCVisitor::_visitRange(void** start, void** end) {
@@ -347,7 +422,7 @@ void GCVisitor::_visitRange(void** start, void** end) {
 void GCVisitor::visitPotential(void* p) {
     GCAllocation* a = global_heap.getAllocationFromInteriorPointer(p);
     if (a) {
-        stack->push(a->user_data);
+        worklist->addWork(a->user_data);
     }
 }
 
@@ -375,10 +450,31 @@ void GCVisitor::visitPotentialRange(void** start, void** end) {
     }
 }
 
+void GCVisitorPinning::_visit(void** ptr_address) {
+    void* p = *ptr_address;
+    if ((uintptr_t)p < SMALL_ARENA_START || (uintptr_t)p >= HUGE_ARENA_START + ARENA_SIZE) {
+        ASSERT(!p || isNonheapRoot(p), "%p", p);
+        return;
+    }
+
+    GCAllocation* al = global_heap.getAllocationFromInteriorPointer(p);
+    ASSERT(al->user_data == p, "%p", p);
+    worklist->addWork(al, source);
+}
+
+void GCVisitorPinning::visitPotential(void* p) {
+    GCAllocation* a = global_heap.getAllocationFromInteriorPointer(p);
+    if (a) {
+        worklist->pin(a);
+        worklist->addWork(a, source);
+    }
+}
+
 static __attribute__((always_inline)) void visitByGCKind(void* p, GCVisitor& visitor) {
     assert(((intptr_t)p) % 8 == 0);
 
     GCAllocation* al = GCAllocation::fromUserData(p);
+    visitor.setSource(al);
 
     GCKind kind_id = al->kind_id;
     if (kind_id == GCKind::UNTRACKED) {
@@ -409,7 +505,7 @@ static __attribute__((always_inline)) void visitByGCKind(void* p, GCVisitor& vis
     }
 }
 
-static void markRoots(GCVisitor& visitor) {
+static void visitRoots(GCVisitor& visitor) {
     GC_TRACE_LOG("Looking at the stack\n");
     threading::visitAllStacks(&visitor);
 
@@ -445,11 +541,11 @@ static void finalizationOrderingFindReachable(Box* obj) {
     static StatCounter sc_us("us_gc_mark_finalizer_ordering_1");
     Timer _t("finalizationOrderingFindReachable", /*min_usec=*/10000);
 
-    TraceStack stack(TraceStackType::FinalizationOrderingFindReachable);
-    GCVisitor visitor(&stack);
+    TraversalWorklist worklist(TraversalType::FinalizationOrderingFindReachable);
+    GCVisitor visitor(&worklist);
 
-    stack.push(obj);
-    while (void* p = stack.pop()) {
+    worklist.addWork(obj);
+    while (void* p = worklist.next()) {
         sc_marked_objs.log();
 
         visitByGCKind(p, visitor);
@@ -463,11 +559,11 @@ static void finalizationOrderingRemoveTemporaries(Box* obj) {
     static StatCounter sc_us("us_gc_mark_finalizer_ordering_2");
     Timer _t("finalizationOrderingRemoveTemporaries", /*min_usec=*/10000);
 
-    TraceStack stack(TraceStackType::FinalizationOrderingRemoveTemporaries);
-    GCVisitor visitor(&stack);
+    TraversalWorklist worklist(TraversalType::FinalizationOrderingRemoveTemporaries);
+    GCVisitor visitor(&worklist);
 
-    stack.push(obj);
-    while (void* p = stack.pop()) {
+    worklist.addWork(obj);
+    while (void* p = worklist.next()) {
         GCAllocation* al = GCAllocation::fromUserData(p);
         assert(orderingState(al) != FinalizationState::UNREACHABLE);
         visitByGCKind(p, visitor);
@@ -513,12 +609,12 @@ static void orderFinalizers() {
     sc_us.log(us);
 }
 
-static void graphTraversalMarking(TraceStack& stack, GCVisitor& visitor) {
+static void graphTraversalMarking(Worklist& worklist, GCVisitor& visitor) {
     static StatCounter sc_us("us_gc_mark_phase_graph_traversal");
     static StatCounter sc_marked_objs("gc_marked_object_count");
     Timer _t("traversing", /*min_usec=*/10000);
 
-    while (void* p = stack.pop()) {
+    while (void* p = worklist.next()) {
         sc_marked_objs.log();
 
         GCAllocation* al = GCAllocation::fromUserData(p);
@@ -530,7 +626,9 @@ static void graphTraversalMarking(TraceStack& stack, GCVisitor& visitor) {
             GC_TRACE_LOG("Looking at non-python allocation %p\n", p);
 #endif
 
-        assert(isMarked(al));
+        // Won't work once we visit objects in more ways than just marking them.
+        assert(isMarked(al) || MOVING_GC);
+
         visitByGCKind(p, visitor);
     }
 
@@ -646,12 +744,12 @@ static void markPhase() {
     GC_TRACE_LOG("Starting collection %d\n", ncollections);
 
     GC_TRACE_LOG("Looking at roots\n");
-    TraceStack stack(TraceStackType::MarkPhase, roots);
-    GCVisitor visitor(&stack);
+    TraversalWorklist worklist(TraversalType::MarkPhase, roots);
+    GCVisitor visitor(&worklist);
 
-    markRoots(visitor);
+    visitRoots(visitor);
 
-    graphTraversalMarking(stack, visitor);
+    graphTraversalMarking(worklist, visitor);
 
     // Objects with finalizers cannot be freed in any order. During the call to a finalizer
     // of an object, the finalizer expects the object's references to still point to valid
@@ -679,6 +777,55 @@ static void sweepPhase(std::vector<Box*>& weakly_referenced) {
     sc_us.log(us);
 }
 
+static void mapReferencesPhase(ReferenceMap& refmap) {
+    ReferenceMapWorklist worklist(&refmap, roots);
+    GCVisitorPinning visitor(&worklist);
+
+    visitRoots(visitor);
+
+    for (auto obj : objects_with_ordered_finalizers) {
+        visitor.visit((void**)&obj);
+    }
+
+    graphTraversalMarking(worklist, visitor);
+}
+
+static void move(ReferenceMap& refmap, GCAllocation* al, size_t size) {
+    // Only move objects that are in the reference map (unreachable objects
+    // won't be in the reference map).
+    if (refmap.pinned.count(al) == 0 && refmap.references.count(al) > 0) {
+        auto& referencing = refmap.references[al];
+        assert(referencing.size() > 0);
+        // GCAllocation* new_al = realloc(al, size);
+    } else if (refmap.pinned.count(al) == 0) {
+        // TODO: This probably should not happen.
+    }
+}
+
+// Move objects around memory randomly. The purpose is to test whether the rest
+// of the program is able to support a moving collector (e.g. if all pointers are
+// being properly scanned by the GC).
+//
+// The way it works is very simple.
+// 1) Perform a mark phase where for every object, make a list of the location of
+//    all pointers to that object (make a reference map).
+//    Pin certain types of objects as necessary (e.g. conservatively scanned).
+// 2) Reallocate all non-pinned object. Update the value for every pointer locations
+//    from the map built in (1)
+static void testMoving() {
+    global_heap.prepareForCollection();
+
+    ReferenceMap refmap;
+    mapReferencesPhase(refmap);
+
+    // Reallocate (aka 'move') all objects in the small heap to a different
+    // location. This is not useful in terms of performance, but it is useful
+    // to check if the rest of the program is able to support moving collectors.
+    global_heap.forEachSmallArenaReference([&refmap](GCAllocation* al, size_t size) { move(refmap, al, size); });
+
+    global_heap.cleanupAfterCollection();
+}
+
 bool gcIsEnabled() {
     return gc_enabled;
 }
@@ -762,6 +909,12 @@ void runCollection() {
         global_heap.free(GCAllocation::fromUserData(o));
     }
 
+    global_heap.cleanupAfterCollection();
+
+#if MOVING_GC
+    testMoving();
+#endif
+
 #if TRACE_GC_MARKING
     fclose(trace_fp);
     trace_fp = NULL;
@@ -769,8 +922,6 @@ void runCollection() {
 
     should_not_reenter_gc = false; // end non-reentrant section
 
-    global_heap.cleanupAfterCollection();
-
     if (VERBOSITY("gc") >= 2)
         printf("Collection #%d done\n\n", ncollections);
 

src/gc/collector.h

@@ -15,6 +15,10 @@
 #ifndef PYSTON_GC_COLLECTOR_H
 #define PYSTON_GC_COLLECTOR_H
 
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
 #include "gc/gc.h"
 
 namespace pyston {
@@ -72,6 +76,43 @@ public:
     void visitPotentialRedundant(void* p) override { visitPotential(p); }
     void visitPotentialRangeRedundant(void** start, void** end) override { visitPotentialRange(start, end); }
 };
+
+//
+// Code to prototype a moving GC.
+//
+
+class ReferenceMapWorklist;
+
+#if MOVING_GC
+#define MOVING_OVERRIDE override
+#else
+#define MOVING_OVERRIDE
+#endif
+
+// Bulds the reference map, and also determine which objects cannot be moved.
+class GCVisitorPinning : public GCVisitorNoRedundancy {
+private:
+    ReferenceMapWorklist* worklist;
+
+    void _visit(void** ptr_address) MOVING_OVERRIDE;
+
+public:
+    GCVisitorPinning(ReferenceMapWorklist* worklist) : worklist(worklist) {}
+    virtual ~GCVisitorPinning() {}
+
+    void visitPotential(void* p) MOVING_OVERRIDE;
+};
+
+class GCAllocation;
+class ReferenceMap {
+public:
+    // Pinned objects are objects that should not be moved (their pointer value should
+    // never change).
+    std::unordered_set<GCAllocation*> pinned;
+
+    // Map from objects O to all objects that contain a reference to O.
+    std::unordered_map<GCAllocation*, std::vector<GCAllocation*>> references;
+};
 }
 }
 

src/gc/gc.h

@@ -54,9 +54,10 @@ void popGCObject(gc::GCVisitable* obj);
 
 namespace gc {
 
-class TraceStack;
+class GCAllocation;
+class TraversalWorklist;
 
-// The base version of the GC visitor is used for marking, in conjuction with a TraceStack.
+// The base version of the GC visitor is used for marking, in conjuction with a TraversalWorklist.
 //
 // Conceptually, GCVisitor should be abstract and the 'marking' behavior should be specific
 // to a subclass of GCVisitor. However, that requires the use of virtual functions which
@@ -65,7 +66,11 @@ class TraceStack;
 // the virtualness property is #if'd out for the regular use case with only mark-and-sweep.
 class GCVisitor {
 private:
-    TraceStack* stack;
+    TraversalWorklist* worklist = NULL;
+
+protected:
+    // The origin object of the current visit calls.
+    GCAllocation* source = NULL;
 
 #if MOVING_GC
     virtual void _visit(void** ptr_address);
@@ -78,7 +83,8 @@ private:
     virtual void _visitRangeRedundant(void** start, void** end) {}
 
 public:
-    GCVisitor(TraceStack* stack) : stack(stack) {}
+    GCVisitor() {}
+    GCVisitor(TraversalWorklist* worklist) : worklist(worklist) {}
     virtual ~GCVisitor() {}
 
 #if MOVING_GC
@@ -122,6 +128,8 @@ public:
     // change that later for performance.
     void visitNonRelocatable(void* p) { visitPotential(p); }
     void visitNonRelocatableRange(void** start, void** end) { visitPotentialRange(start, end); }
+
+    void setSource(GCAllocation* al) { source = al; }
 };
 
 enum class GCKind : uint8_t {

src/gc/heap.cpp

@@ -414,6 +414,53 @@ void SmallArena::assertConsistent() {
 }
 #endif
 
+void SmallArena::getPointersInBlockChain(std::vector<GCAllocation*>& ptrs, Block** head) {
+    while (Block* b = *head) {
+        int num_objects = b->numObjects();
+        int first_obj = b->minObjIndex();
+        int atoms_per_obj = b->atomsPerObj();
+
+        for (int atom_idx = first_obj * atoms_per_obj; atom_idx < num_objects * atoms_per_obj;
+             atom_idx += atoms_per_obj) {
+
+            if (b->isfree.isSet(atom_idx))
+                continue;
+
+            void* p = &b->atoms[atom_idx];
+            GCAllocation* al = reinterpret_cast<GCAllocation*>(p);
+
+            ptrs.push_back(al);
+        }
+
+        head = &b->next;
+    }
+}
+
+void SmallArena::forEachReference(std::function<void(GCAllocation*, size_t)> f) {
+    thread_caches.forEachValue([this, &f](ThreadBlockCache* cache) {
+        for (int bidx = 0; bidx < NUM_BUCKETS; bidx++) {
+            Block* h = cache->cache_free_heads[bidx];
+            std::vector<GCAllocation*> ptrs;
+            getPointersInBlockChain(ptrs, &cache->cache_free_heads[bidx]);
+            getPointersInBlockChain(ptrs, &cache->cache_full_heads[bidx]);
+
+            for (GCAllocation* al : ptrs) {
+                f(al, sizes[bidx]);
+            }
+        }
+    });
+
+    for (int bidx = 0; bidx < NUM_BUCKETS; bidx++) {
+        std::vector<GCAllocation*> ptrs;
+        getPointersInBlockChain(ptrs, &heads[bidx]);
+        getPointersInBlockChain(ptrs, &full_heads[bidx]);
+
+        for (GCAllocation* al : ptrs) {
+            f(al, sizes[bidx]);
+        }
+    }
+}
+
 void SmallArena::freeUnmarked(std::vector<Box*>& weakly_referenced) {
     assertConsistent();
 

src/gc/heap.h

@@ -93,6 +93,7 @@ inline void registerGCManagedBytes(size_t bytes) {
 
 
 class Heap;
+class ReferenceMap;
 struct HeapStatistics;
 
 typedef uint8_t kindid_t;
@@ -263,6 +264,8 @@ public:
         }
     }
 
+    void forEachReference(std::function<void(GCAllocation*, size_t)>);
+
     GCAllocation* realloc(GCAllocation* alloc, size_t bytes);
     void free(GCAllocation* al);
 
@@ -405,6 +408,7 @@ private:
     // TODO only use thread caches if we're in GRWL mode?
     threading::PerThreadSet<ThreadBlockCache, Heap*, SmallArena*> thread_caches;
 
+    void getPointersInBlockChain(std::vector<GCAllocation*>& ptrs, Block** head);
     Block* _allocBlock(uint64_t size, Block** prev);
     GCAllocation* _allocFromBlock(Block* b);
     Block* _claimBlock(size_t rounded_size, Block** free_head);
@@ -625,6 +629,9 @@ public:
         return NULL;
     }
 
+    // Calls the function for every object in the small heap.
+    void forEachSmallArenaReference(std::function<void(GCAllocation*, size_t)> f) { small_arena.forEachReference(f); }
+
     // not thread safe:
     void freeUnmarked(std::vector<Box*>& weakly_referenced) {
         small_arena.freeUnmarked(weakly_referenced);