Cycle collector. (#175)

backend.native/compiler/ir/backend.native/src/org/jetbrains/kotlin/backend/konan/LinkStage.kt

@@ -33,7 +33,7 @@ internal class MacOSPlatform(distrib: Distribution,
 
     override val linkerOptimizationFlags = 
         properties.propertyList("linkerOptimizationFlags.osx")
-    override val linkerKonanFlags = listOf(distrib.libCppAbi)
+    override val linkerKonanFlags = properties.propertyList("linkerKonanFlags.osx")
     override val linker = "${distrib.sysRoot}/usr/bin/ld" 
 
     override fun linkCommand(objectFiles: List<String>, executable: String, optimize: Boolean): List<String> {

backend.native/konan.properties

@@ -10,12 +10,13 @@ libGcc.linux = target-gcc-toolchain-3-linux-x86-64/lib/gcc/x86_64-unknown-linux-
 
 llvmLtoFlags.osx = -O3 -function-sections -exported-symbol=_main
 llvmLlcFlags.osx = -mtriple=x86_64-apple-macosx10.10.0
+linkerKonanFlags.osx = -lc++
 linkerOptimizationFlags.osx = -dead_strip
 macosVersionMin.osx = 10.10.0
 
 llvmLtoFlags.linux = -O3 -function-sections -exported-symbol=main
 llvmLlcFlags.linux = -march=x86-64
-linkerKonanFlags.linux = -Bstatic -lc++abi -Bdynamic -ldl -lm -lpthread
+linkerKonanFlags.linux = -Bstatic -lstdc++ -Bdynamic -ldl -lm -lpthread
 linkerOptimizationFlags.linux = --gc-sections
 pluginOptimizationFlags.linux = -plugin-opt=mcpu=x86-64 -plugin-opt=O3
 

backend.native/tests/build.gradle

@@ -41,11 +41,11 @@ abstract class KonanTest extends DefaultTask {
         project.javaexec {
             main = 'org.jetbrains.kotlin.cli.bc.K2NativeKt'
             classpath = project.configurations.cli_bc
-            jvmArgs "-ea", 
+            jvmArgs "-ea",
                 "-Dkonan.home=${dist.canonicalPath}",
                 "-Djava.library.path=${dist.canonicalPath}/konan/nativelib"
             args("-output", output,
-                 source, 
+                 source,
                  *moreArgs,
                  *project.globalArgs)
         }
@@ -88,7 +88,7 @@ abstract class KonanTest extends DefaultTask {
 
 class RunKonanTest extends KonanTest {
     void compileTest(String source, String exe) {
-        runCompiler(source, exe, []) 
+        runCompiler(source, exe, [])
     }
 }
 
@@ -98,7 +98,7 @@ class LinkKonanTest extends KonanTest {
     void compileTest(String source, String exe) {
         def libDir = project.file(lib).absolutePath
         def libBc = "${libDir}.bc"
-        
+
         runCompiler(lib, libBc, ['-nolink', '-nostdlib'])
         runCompiler(source, exe, ['-library', libBc])
     }
@@ -246,8 +246,8 @@ task tostring3(type: RunKonanTest) {
             "1.17549E-38\n3.40282E+38\n-INF\nINF\n" +
             // Linux version prints -NAN.
             // "NAN\n" +
-            "4.94066E-324\n1.79769E+308\n-INF\nINF\n" 
+            "4.94066E-324\n1.79769E+308\n-INF\nINF\n"
-            // "NAN\n" 
+            // "NAN\n"
     source = "runtime/basic/tostring3.kt"
 }
 
@@ -359,7 +359,7 @@ task intrinsic(type: RunKonanTest) {
 }
 
 /*
-    Disabled until we extract the classes that should be 
+    Disabled until we extract the classes that should be
     always present from stdlib.kt.bc into a separate binary.
 
 task link(type: LinkKonanTest) {
@@ -805,6 +805,11 @@ task memory_throw_cleanup(type: RunKonanTest) {
     source = "runtime/memory/throw_cleanup.kt"
 }
 
+task memory_collect_cycles(type: RunKonanTest) {
+    goldValue = "42\n"
+    source = "runtime/memory/cycles0.kt"
+}
+
 task unit1(type: RunKonanTest) {
     goldValue = "First\nkotlin.Unit\n"
     source = "codegen/basics/unit1.kt"

backend.native/tests/runtime/memory/cycles0.kt

@@ -0,0 +1,45 @@
+data class Node(val data: Int, var next: Node?, var prev: Node?, val outer: Node?)
+
+fun makeCycle(len: Int, outer: Node?): Node {
+    val start = Node(0, null, null, outer)
+    var prev = start
+    for (i in 1 .. len - 1) {
+        prev = Node(i, prev, null, outer)
+    }
+    start.next = prev
+    return start
+}
+
+fun makeDoubleCycle(len: Int): Node {
+    val start = makeCycle(len, null)
+    var prev = start
+    var cur = prev.next
+    while (cur != start) {
+        cur!!.prev = prev
+        prev = cur
+        cur = cur.next
+    }
+    start.prev = prev
+    return start
+}
+
+fun createCycles(junk: Node) {
+    val cycle1 = makeCycle(1, junk)
+    val cycle2 = makeCycle(2, junk)
+    val cycle10 = makeCycle(10, junk)
+    val cycle100 = makeCycle(100, junk)
+    val dcycle1 = makeDoubleCycle(1)
+    val dcycle2 = makeDoubleCycle(2)
+    val dcycle10 = makeDoubleCycle(10)
+    val dcycle100 = makeDoubleCycle(100)
+
+}
+
+fun main(args : Array<String>) {
+    // Create outer link from cyclic garbage.
+    val outer = Node(42, null, null, null)
+    createCycles(outer)
+    konan.internal.GC.collect()
+    // Ensure outer is not collected.
+    println(outer.data)
+}

runtime/src/main/cpp/Memory.cpp

@@ -3,7 +3,7 @@
 #include <stdio.h>
 
 #include <cstddef> // for offsetof
-#include <set> // only for memory tracing.
+#include <unordered_set>
 #include <vector>
 
 #include "Assert.h"
@@ -11,44 +11,186 @@
 #include "Memory.h"
 #include "Natives.h"
 
-// Define to 1 to see all memory operations.
+// Define to 1 to use in the multithreaded environment.
-#define TRACE_MEMORY 0
-// Define to 1 to use in multithreaded environment.
 #define CONCURRENT 0
+// If garbage collection algorithm for cyclic garbage to be used.
+#define USE_GC 1
+// Define to 1 to print all memory operations.
+#define TRACE_MEMORY 0
+// Trace garbage collection phases.
+#define TRACE_GC_PHASES 0
 
-ContainerHeader ObjHeader::theStaticObjectsContainer = { CONTAINER_TAG_NOCOUNT };
+ContainerHeader ObjHeader::theStaticObjectsContainer = {
+  CONTAINER_TAG_NOCOUNT | CONTAINER_TAG_INCREMENT
+};
 
 namespace {
 
-// Current number of allocated containers.
+#if USE_GC
-int allocCount = 0;
+// Collection threshold (collect after having so many elements in the
-#if TRACE_MEMORY
+// release candidates set).
-// List of all global objects addresses.
+constexpr size_t kGcThreshold = 10000;
-std::vector<KRef*>* globalObjects = nullptr;
-// Set of all containers.
-std::set<ContainerHeader*>* containers = nullptr;
 #endif
 
+#if TRACE_MEMORY || USE_GC
+typedef std::unordered_set<ContainerHeader*> ContainerHeaderSet;
+typedef std::vector<ContainerHeader*> ContainerHeaderList;
+typedef std::vector<KRef*> KRefPtrList;
+#endif
+
+struct MemoryState {
+  // Current number of allocated containers.
+  int allocCount = 0;
+#if TRACE_MEMORY
+  // List of all global objects addresses.
+  KRefPtrList* globalObjects;
+  // Set of all containers.
+  ContainerHeaderSet* containers;
+#endif
+#if USE_GC
+  // Set of references to release.
+  ContainerHeaderSet* toFree;
+  bool gcInProgress;
+  size_t gcThreshold;
+#endif
+};
+
+MemoryState* memoryState = nullptr;
+
+#if USE_GC
+bool isPermanent(const ContainerHeader* header) {
+  return (header->ref_count_ & CONTAINER_TAG_MASK) == CONTAINER_TAG_NOCOUNT;
+}
+
+// Must be vector or map 'container -> number', to keep reference counters correct.
+ContainerHeaderList collectMutableReferred(ContainerHeader* header) {
+  ContainerHeaderList result;
+  ObjHeader* obj = reinterpret_cast<ObjHeader*>(header + 1);
+  const TypeInfo* typeInfo = obj->type_info();
+  // TODO: generalize iteration over all references.
+  // TODO: this code relies on single object per container assumption.
+  for (int index = 0; index < typeInfo->objOffsetsCount_; index++) {
+    ObjHeader** location = reinterpret_cast<ObjHeader**>(
+      reinterpret_cast<uintptr_t>(obj + 1) + typeInfo->objOffsets_[index]);
+    ObjHeader* obj = *location;
+    if (obj != nullptr && !isPermanent(obj->container())) {
+      result.push_back(obj->container());
+    }
+  }
+  if (typeInfo == theArrayTypeInfo) {
+    ArrayHeader* array = obj->array();
+    for (int index = 0; index < array->count_; index++) {
+      ObjHeader* obj = *ArrayAddressOfElementAt(array, index);
+      if (obj != nullptr && !isPermanent(obj->container())) {
+        result.push_back(obj->container());
+      }
+    }
+  }
+  return result;
+}
+
+void dumpWorker(const char* prefix, ContainerHeader* header, ContainerHeaderSet* seen) {
+  fprintf(stderr, "%s: %p (%08x): %d refs %s\n",
+          prefix,
+          header, header->ref_count_, header->ref_count_ >> CONTAINER_TAG_SHIFT,
+          (header->ref_count_ & CONTAINER_TAG_SEEN) != 0 ? "X" : "-");
+  seen->insert(header);
+  auto children = collectMutableReferred(header);
+  for (auto child : children) {
+    if (seen->count(child) == 0) {
+      dumpWorker(prefix, child, seen);
+    }
+  }
+}
+
+void dumpReachable(const char* prefix, const ContainerHeaderSet* roots) {
+  ContainerHeaderSet seen;
+  for (auto container : *roots) {
+    dumpWorker(prefix, container, &seen);
+  }
+}
+
+void phase1(ContainerHeader* header) {
+  if ((header->ref_count_ & CONTAINER_TAG_SEEN) != 0)
+    return;
+  header->ref_count_ |= CONTAINER_TAG_SEEN;
+  auto containers = collectMutableReferred(header);
+  for (auto container : containers) {
+      container->ref_count_ -= CONTAINER_TAG_INCREMENT;
+      phase1(container);
+  }
+}
+
+void phase2(ContainerHeader* header, ContainerHeaderSet* rootset) {
+  if ((header->ref_count_ & CONTAINER_TAG_SEEN) == 0)
+    return;
+  if ((header->ref_count_ >> CONTAINER_TAG_SHIFT) != 0)
+    rootset->insert(header);
+  header->ref_count_ &= ~CONTAINER_TAG_SEEN;
+  auto containers = collectMutableReferred(header);
+  for (auto container : containers) {
+    phase2(container, rootset);
+  }
+}
+
+void phase3(ContainerHeader* header) {
+  if ((header->ref_count_ & CONTAINER_TAG_SEEN) != 0) {
+    return;
+  }
+  header->ref_count_ |= CONTAINER_TAG_SEEN;
+  auto containers = collectMutableReferred(header);
+  for (auto container : containers) {
+    container->ref_count_ += CONTAINER_TAG_INCREMENT;
+    phase3(container);
+  }
+}
+
+void phase4(ContainerHeader* header, ContainerHeaderSet* toRemove) {
+  auto ref_count = header->ref_count_ >> CONTAINER_TAG_SHIFT;
+  bool seen = (ref_count > 0 && (header->ref_count_ & CONTAINER_TAG_SEEN) == 0) ||
+      (ref_count == 0 && (header->ref_count_ & CONTAINER_TAG_SEEN) != 0);
+  if (seen) return;
+
+  // Add to toRemove set.
+  if (ref_count == 0)
+    toRemove->insert(header);
+
+  // Update seen bit.
+  if (ref_count == 0)
+    header->ref_count_ |= CONTAINER_TAG_SEEN;
+  else
+    header->ref_count_ &= ~CONTAINER_TAG_SEEN;
+  auto containers = collectMutableReferred(header);
+  for (auto container : containers) {
+    phase4(container, toRemove);
+  }
+}
+
+#endif // USE_GC
+
 }  // namespace
 
 ContainerHeader* AllocContainer(size_t size) {
   ContainerHeader* result = reinterpret_cast<ContainerHeader*>(calloc(1, size));
 #if TRACE_MEMORY
-  printf(">>> alloc %d -> %p\n", (int)size, result);
+  fprintf(stderr, ">>> alloc %d -> %p\n", (int)size, result);
-  containers->insert(result);
+   memoryState->containers->insert(result);
 #endif
   // TODO: atomic increment in concurrent case.
-  allocCount++;
+  memoryState->allocCount++;
   return result;
 }
 
 void FreeContainer(ContainerHeader* header) {
+  RuntimeAssert(!isPermanent(header), "this kind of container shalln't be freed");
 #if TRACE_MEMORY
-  printf("<<< free %p\n", header);
+  fprintf(stderr, "<<< free %p\n", header);
-  containers->erase(header);
+  memoryState->containers->erase(header);
 #endif
   header->ref_count_ = CONTAINER_TAG_INVALID;
-
+#if USE_GC
+  memoryState->toFree->erase(header);
+#endif
   // Now let's clean all object's fields in this container.
   // TODO: this is gross hack, relying on the fact that we now only alloc
   // ArenaContainer and ObjectContainer, which both have single element.
@@ -69,10 +211,29 @@ void FreeContainer(ContainerHeader* header) {
 
   // And release underlying memory.
   // TODO: atomic decrement in concurrent case.
-  allocCount--;
+#if CONCURRENT
+  #error "Atomic update of allocCount"
+#endif
+  memoryState->allocCount--;
   free(header);
 }
 
+#if USE_GC
+void FreeContainerNoRef(ContainerHeader* header) {
+  RuntimeAssert(!isPermanent(header), "this kind of container shalln't be freed");
+#if TRACE_MEMORY
+  fprintf(stderr, "<<< free %p\n", header);
+  memoryState->containers->erase(header);
+#endif
+  header->ref_count_ = CONTAINER_TAG_INVALID;
+#if USE_GC
+  memoryState->toFree->erase(header);
+#endif
+  memoryState->allocCount--;
+  free(header);
+}
+#endif
+
 ArenaContainer::ArenaContainer(uint32_t size) {
     ArenaContainerHeader* header =
         static_cast<ArenaContainerHeader*>(AllocContainer(size + sizeof(ArenaContainerHeader)));
@@ -92,7 +253,7 @@ void ObjectContainer::Init(const TypeInfo* type_info) {
      // header->ref_count_ is zero initialized by AllocContainer().
     SetMeta(GetPlace(), type_info);
 #if TRACE_MEMORY
-    printf("object at %p\n", GetPlace());
+    fprintf(stderr, "object at %p\n", GetPlace());
 #endif
   }
 }
@@ -109,7 +270,7 @@ void ArrayContainer::Init(const TypeInfo* type_info, uint32_t elements) {
     GetPlace()->count_ = elements;
     SetMeta(GetPlace()->obj(), type_info);
 #if TRACE_MEMORY
-    printf("array at %p\n", GetPlace());
+    fprintf(stderr, "array at %p\n", GetPlace());
 #endif
   }
 }
@@ -139,16 +300,35 @@ ArrayHeader* ArenaContainer::PlaceArray(const TypeInfo* type_info, int count) {
 
 inline void AddRef(const ObjHeader* object) {
 #if TRACE_MEMORY
-  printf("AddRef on %p in %p\n", object, object->container());
+  fprintf(stderr, "AddRef on %p in %p\n", object, object->container());
 #endif
   AddRef(object->container());
+#if USE_GC
+  // TODO: one could remove from toFree set here, as now container is reachable
+  // from the rootset, so cannot be cycle collection candidate.
+  // memoryState->toFree->erase(object->container());
+#endif
 }
 
 inline void ReleaseRef(const ObjHeader* object) {
 #if TRACE_MEMORY
-  printf("ReleaseRef on %p in %p\n", object, object->container());
+  fprintf(stderr, "ReleaseRef on %p in %p\n", object, object->container());
 #endif
+#if USE_GC
+  // If object is not a cycle candidate - just return.
+  if (Release(object->container())) {
+    return;
+  }
+#if TRACE_MEMORY
+  fprintf(stderr, "%p is release candidate\n", object->container());
+#endif
+  memoryState->toFree->insert(object->container());
+  if (memoryState->toFree->size() > memoryState->gcThreshold) {
+    GarbageCollect();
+  }
+#else // !USE_GC
   Release(object->container());
+#endif // USE_GC
 }
 
 extern "C" {
@@ -162,36 +342,50 @@ void InitMemory() {
                 ==
                 offsetof(ObjHeader  , container_offset_negative_),
                 "Layout mismatch");
+  RuntimeAssert(memoryState == nullptr, "memory state must be clear");
+  memoryState = new MemoryState();
   // TODO: initialize heap here.
-  allocCount = 0;
+  memoryState->allocCount = 0;
 #if TRACE_MEMORY
-  globalObjects = new std::vector<KRef*>();
+  memoryState->globalObjects = new KRefPtrList();
-  containers = new std::set<ContainerHeader*>();
+  memoryState->containers = new ContainerHeaderSet();
+#endif
+#if USE_GC
+#if CONCURRENT
+  #error "Concurrent GC is not yet implemented"
+#endif
+  memoryState->toFree = new ContainerHeaderSet();
+  memoryState->gcInProgress = false;
+  memoryState->gcThreshold = kGcThreshold;
 #endif
 }
 
 void DeinitMemory() {
 #if TRACE_MEMORY
   // Free all global objects, to ensure no memory leaks happens.
-  for (auto location: *globalObjects) {
+  for (auto location: *memoryState->globalObjects) {
-    printf("Release global in *%p: %p\n", location, *location);
+    fprintf(stderr, "Release global in *%p: %p\n", location, *location);
     UpdateGlobalRef(location, nullptr);
   }
-  delete globalObjects;
+  delete memoryState->globalObjects;
-  globalObjects = nullptr;
+  memoryState->globalObjects = nullptr;
 #endif
 
-  if (allocCount > 0) {
+#if USE_GC
+  GarbageCollect();
+#endif // USE_GC
+
+  if (memoryState->allocCount > 0) {
 #if TRACE_MEMORY
-    // TODO: move out of TRACE_MEMORY, once exceptions free memory.
+    fprintf(stderr, "*** Memory leaks, leaked %d containers ***\n", memoryState->allocCount);
-    printf("*** Memory leaks, leaked %d containers ***\n", allocCount);
+    dumpReachable("", memoryState->containers);
-    for (auto container: *containers) {
+    delete memoryState->containers;
-      printf("Unfreed container %p, count = %d\n", container, container->ref_count_);
+    memoryState->containers = nullptr;
-    }
-    delete containers;
-    containers = nullptr;
 #endif
   }
+
+  delete memoryState;
+  memoryState = nullptr;
 }
 
 // Now we ignore all placement hints and always allocate heap space for new object.
@@ -242,7 +436,7 @@ OBJ_GETTER(InitInstance,
     // TODO: locking or smth lock-free in MT case?
 #endif
 #if TRACE_MEMORY
-    globalObjects->push_back(location);
+    memoryState->globalObjects->push_back(location);
 #endif
     RETURN_OBJ_RESULT();
   } catch (...) {
@@ -254,7 +448,7 @@ OBJ_GETTER(InitInstance,
 
 void SetLocalRef(ObjHeader** location, const ObjHeader* object) {
 #if TRACE_MEMORY
-  printf("SetLocalRef *%p: %p\n", location, object);
+  fprintf(stderr, "SetLocalRef *%p: %p\n", location, object);
 #endif
   *const_cast<const ObjHeader**>(location) = object;
   if (object != nullptr) {
@@ -264,7 +458,7 @@ void SetLocalRef(ObjHeader** location, const ObjHeader* object) {
 
 void SetGlobalRef(ObjHeader** location, const ObjHeader* object) {
 #if TRACE_MEMORY
-  printf("SetGlobalRef *%p: %p\n", location, object);
+  fprintf(stderr, "SetGlobalRef *%p: %p\n", location, object);
 #endif
   *const_cast<const ObjHeader**>(location) = object;
    if (object != nullptr) {
@@ -278,16 +472,16 @@ void SetGlobalRef(ObjHeader** location, const ObjHeader* object) {
 void UpdateLocalRef(ObjHeader** location, const ObjHeader* object) {
   ObjHeader* old = *location;
 #if TRACE_MEMORY
-  printf("UpdateLocalRef *%p: %p -> %p\n", location, old, object);
+  fprintf(stderr, "UpdateLocalRef *%p: %p -> %p\n", location, old, object);
 #endif
   if (old != object) {
+    if (object != nullptr) {
+      AddRef(object);
+    }
     *const_cast<const ObjHeader**>(location) = object;
     if (old > reinterpret_cast<ObjHeader*>(1)) {
       ReleaseRef(old);
     }
-    if (object != nullptr) {
-      AddRef(object);
-    }
   }
 }
 
@@ -295,7 +489,7 @@ void UpdateGlobalRef(ObjHeader** location, const ObjHeader* object) {
 #if CONCURRENT
   ObjHeader* old = *location;
 #if TRACE_MEMORY
-  printf("UpdateGlobalRef *%p: %p -> %p\n", location, old, object);
+  fprintf(stderr, "UpdateGlobalRef *%p: %p -> %p\n", location, old, object);
 #endif
   if (old != object) {
     if (object != nullptr) {
@@ -320,7 +514,7 @@ void UpdateGlobalRef(ObjHeader** location, const ObjHeader* object) {
 
 void ReleaseLocalRefs(ObjHeader** start, int count) {
 #if TRACE_MEMORY
-  printf("ReleaseLocalRefs %p .. %p\n", start, start + count);
+  fprintf(stderr, "ReleaseLocalRefs %p .. %p\n", start, start + count);
 #endif
   ObjHeader** current = start;
   while (count-- > 0) {
@@ -336,7 +530,7 @@ void ReleaseLocalRefs(ObjHeader** start, int count) {
 
 void ReleaseGlobalRefs(ObjHeader** start, int count) {
 #if TRACE_MEMORY
-  printf("ReleaseGlobalRefs %p .. %p\n", start, start + count);
+  fprintf(stderr, "ReleaseGlobalRefs %p .. %p\n", start, start + count);
 #endif
 #if CONCURRENT
   ObjHeader** current = start;
@@ -364,4 +558,73 @@ void ReleaseGlobalRefs(ObjHeader** start, int count) {
 #endif
 }
 
+#if USE_GC
+void GarbageCollect() {
+  RuntimeAssert(!memoryState->gcInProgress, "Recursive GC is disallowed");
+  memoryState->gcInProgress = true;
+  // Traverse inner pointers in the closure of release candidates, and
+  // temporary decrement refs on them. Set CONTAINER_TAG_SEEN while traversing.
+#if TRACE_GC_PHASES
+  dumpReachable("P0", memoryState->toFree);
+#endif
+  for (auto container : *memoryState->toFree) {
+    phase1(container);
+  }
+#if TRACE_GC_PHASES
+  dumpReachable("P1", memoryState->toFree);
+#endif
+
+  // Collect rootset from containers with non-zero reference counter. Those must
+  // be referenced from outside of newly released object graph.
+  // Clear CONTAINER_TAG_SEEN while traversing.
+  ContainerHeaderSet rootset;
+  for (auto container : *memoryState->toFree) {
+    phase2(container, &rootset);
+  }
+#if TRACE_GC_PHASES
+  dumpReachable("P2", memoryState->toFree);
+#endif
+
+  // Increment references for all elements reachable from the rootset.
+  // Set CONTAINER_TAG_SEEN while traversing.
+  for (auto container : rootset) {
+#if TRACE_MEMORY
+    fprintf(stderr, "rootset %p\n", container);
+#endif
+    phase3(container);
+  }
+#if TRACE_GC_PHASES
+  dumpReachable("P3", memoryState->toFree);
+#endif
+
+  // Traverse all elements, and collect those not having CONTAINER_TAG_SEEN and zero RC.
+  // Clear CONTAINER_TAG_SEEN while traversing on live elements, set in on dead elements.
+  ContainerHeaderSet toRemove;
+  for (auto container : *memoryState->toFree) {
+    phase4(container, &toRemove);
+  }
+#if TRACE_GC_PHASES
+  dumpReachable("P4", memoryState->toFree);
+#endif
+
+  // Clear cycle candidates list.
+  memoryState->toFree->clear();
+
+  for (auto header : toRemove) {
+    RuntimeAssert((header->ref_count_ & CONTAINER_TAG_SEEN) != 0, "Must be not seen");
+    FreeContainerNoRef(header);
+  }
+
+  memoryState->gcInProgress = false;
+}
+
+#endif // USE_GC
+
+void Kotlin_konan_internal_GC_collect(KRef) {
+#if USE_GC
+  GarbageCollect();
+#endif
+}
+
+
 } // extern "C"

runtime/src/main/cpp/Memory.h

@@ -26,10 +26,14 @@ typedef enum {
   CONTAINER_TAG_SHARED = 2,
   // Container is no longer valid.
   CONTAINER_TAG_INVALID = 3,
+  // Container was seen during GC.
+  CONTAINER_TAG_SEEN = 4,
+  // Shift to get actual counter..
+  CONTAINER_TAG_SHIFT = 3,
   // Actual value to increment/decrement conatiner by. Tag is in lower bits.
-  CONTAINER_TAG_INCREMENT = 1 << 2,
+  CONTAINER_TAG_INCREMENT = 1 << CONTAINER_TAG_SHIFT,
-  // Mask for container type.
+  // Mask for container type, disregard seen bit.
-  CONTAINER_TAG_MASK = (CONTAINER_TAG_INCREMENT - 1)
+  CONTAINER_TAG_MASK = ((CONTAINER_TAG_INCREMENT >> 1) - 1)
 } ContainerTag;
 
 // Could be made 64-bit for large memory configs.
@@ -127,8 +131,8 @@ inline uint32_t ArrayDataSizeBytes(const ArrayHeader* obj) {
 
 void FreeContainer(ContainerHeader* header);
 
-// Those two operations are implemented by translator when storing references
+// TODO: those two operations can be implemented by translator when storing
-// to objects.
+// reference to an object.
 inline void AddRef(ContainerHeader* header) {
   // Looking at container type we may want to skip AddRef() totally
   // (non-escaping stack objects, constant objects).
@@ -144,18 +148,25 @@ inline void AddRef(ContainerHeader* header) {
     case CONTAINER_TAG_INVALID:
       RuntimeAssert(false, "trying to addref invalid container");
       break;
+    default:
+      RuntimeAssert(false, "unknown container type");
+      break;
   }
 }
 
-inline void Release(ContainerHeader* header) {
+// Release returns 'true' iff container cannot be part of cycle (either NOCOUNT
+// object or container was fully released and will be collected).
+inline bool Release(ContainerHeader* header) {
   switch (header->ref_count_ & CONTAINER_TAG_MASK) {
     case CONTAINER_TAG_NORMAL:
       if ((header->ref_count_ -= CONTAINER_TAG_INCREMENT) == CONTAINER_TAG_NORMAL) {
         FreeContainer(header);
+        return true;
       }
       break;
     case CONTAINER_TAG_NOCOUNT:
-      break;
+      // NOCOUNT containers aren't loop candidate.
+      return true;
     // Note that shared containers have potentially subtle race, if object holds a
     // reference to another object, stored in shorter living container. In this
     // case there's unlikely, but possible case, where one mutator takes reference,
@@ -173,12 +184,18 @@ inline void Release(ContainerHeader* header) {
       if (__sync_sub_and_fetch(
               &header->ref_count_, CONTAINER_TAG_INCREMENT) == CONTAINER_TAG_SHARED) {
         FreeContainer(header);
+        return true;
       }
       break;
     case CONTAINER_TAG_INVALID:
       RuntimeAssert(false, "trying to release invalid container");
       break;
+    default:
+      RuntimeAssert(false, "unknown container type");
+      break;
   }
+  // Object with non-zero counter after release are loop candidates.
+  return false;
 }
 
 // Class representing arbitrary placement container.
@@ -346,6 +363,8 @@ void UpdateGlobalRef(ObjHeader** location, const ObjHeader* object) RUNTIME_NOTH
 // Optimization: release all references in range.
 void ReleaseLocalRefs(ObjHeader** start, int count) RUNTIME_NOTHROW;
 void ReleaseGlobalRefs(ObjHeader** start, int count) RUNTIME_NOTHROW;
+// Collect garbage, which cannot be found by reference counting (cycles).
+void GarbageCollect() RUNTIME_NOTHROW;
 
 #ifdef __cplusplus
 }

runtime/src/main/kotlin/konan/internal/GC.kt

@@ -0,0 +1,8 @@
+package konan.internal
+
+// Garbage collector interface.
+// TODO: more functions to come.
+object GC {
+    @SymbolName("Kotlin_konan_internal_GC_collect")
+    external fun collect()
+}