Bound memory, and disable free list for now

CC #9

ConflictSet.cpp

@@ -275,8 +275,8 @@ static_assert(sizeof(Node16) < kMinChildrenNode16 * kBytesPerKey);
 static_assert(sizeof(Node4) < kMinChildrenNode4 * kBytesPerKey);
 static_assert(sizeof(Node0) < kBytesPerKey);
 
-template <class T, int64_t kMemoryBound = (1 << 20),
-          int64_t kMaxIndividual = (1 << 10)>
+// TODO revive freeList? It's making bounding memory usage tricky
+template <class T, int64_t kMemoryBound = 0, int64_t kMaxIndividual = (1 << 10)>
 struct BoundedFreeListAllocator {
   static_assert(sizeof(T) >= sizeof(void *));
   static_assert(std::derived_from<T, Node>);
@@ -464,6 +464,8 @@ int64_t getChildMaxVersion(Node *self, uint8_t index) {
 // Precondition - an entry for index must exist in the node
 int64_t &maxVersion(Node *n, ConflictSet::Impl *);
 
+Node *&getRoot(ConflictSet::Impl *);
+
 Node *getChild(Node *self, uint8_t index) {
   if (self->type <= Type::Node16) {
     auto *self16 = static_cast<Node16 *>(self);
@@ -711,8 +713,107 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
   }
 }
 
+namespace {
+std::string getSearchPathPrintable(Node *n);
+}
+
+void maybeDecreaseCapacity(Node *&self, Node *&root,
+                           NodeAllocators *allocators) {
+  if (self->numChildren > 0 &&
+      self->numChildren * self->partialKeyLen >= self->partialKeyCapacity) {
+    return;
+  }
+
+  int newCapacity =
+      std::max(self->partialKeyLen, self->numChildren * self->partialKeyLen);
+
+  switch (self->type) {
+  case Type::Node0: {
+    auto *self0 = (Node0 *)self;
+    auto *newSelf = allocators->node0.allocate(newCapacity);
+    memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
+           kNodeCopySize);
+    memcpy(newSelf->partialKey(), self0->partialKey(), self->partialKeyLen);
+    (self->parent ? getChildExists(self->parent, self->parentsIndex) : root) =
+        newSelf;
+    allocators->node0.release(self0);
+    self = newSelf;
+  } break;
+  case Type::Node4: {
+    auto *self4 = (Node4 *)self;
+    auto *newSelf = allocators->node4.allocate(newCapacity);
+    memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
+           kNodeCopySize);
+    memcpy(newSelf->partialKey(), self4->partialKey(), self->partialKeyLen);
+    // TODO replace with memcpy?
+    for (int i = 0; i < 4; ++i) {
+      newSelf->index[i] = self4->index[i];
+      newSelf->children[i] = self4->children[i];
+    }
+    (self->parent ? getChildExists(self->parent, self->parentsIndex) : root) =
+        newSelf;
+    setChildrenParents(newSelf);
+    allocators->node4.release(self4);
+    self = newSelf;
+  } break;
+  case Type::Node16: {
+    auto *self16 = (Node16 *)self;
+    auto *newSelf = allocators->node16.allocate(newCapacity);
+    memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
+           kNodeCopySize);
+    memcpy(newSelf->partialKey(), self16->partialKey(), self->partialKeyLen);
+    // TODO replace with memcpy?
+    for (int i = 0; i < 16; ++i) {
+      newSelf->index[i] = self16->index[i];
+      newSelf->children[i] = self16->children[i];
+    }
+    (self->parent ? getChildExists(self->parent, self->parentsIndex) : root) =
+        newSelf;
+    setChildrenParents(newSelf);
+    allocators->node16.release(self16);
+    self = newSelf;
+  } break;
+  case Type::Node48: {
+    auto *self48 = (Node48 *)self;
+    auto *newSelf = allocators->node48.allocate(newCapacity);
+    memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
+           kNodeCopySize);
+    memcpy(newSelf->partialKey(), self48->partialKey(), self->partialKeyLen);
+    newSelf->bitSet = self48->bitSet;
+    newSelf->bitSet.forEachInRange(
+        [&](int c) {
+          int index = newSelf->nextFree;
+          newSelf->index[c] = index;
+          newSelf->children[index] = self48->children[self48->index[c]];
+          ++newSelf->nextFree;
+        },
+        0, 256);
+    (self->parent ? getChildExists(self->parent, self->parentsIndex) : root) =
+        newSelf;
+    setChildrenParents(newSelf);
+    allocators->node48.release(self48);
+    self = newSelf;
+  } break;
+  case Type::Node256: {
+    auto *self256 = (Node256 *)self;
+    auto *newSelf = allocators->node256.allocate(newCapacity);
+    memcpy((char *)newSelf + kNodeCopyBegin, (char *)self + kNodeCopyBegin,
+           kNodeCopySize);
+    memcpy(newSelf->partialKey(), self256->partialKey(), self->partialKeyLen);
+    newSelf->bitSet = self256->bitSet;
+    newSelf->bitSet.forEachInRange(
+        [&](int c) { newSelf->children[c] = self256->children[c]; }, 0, 256);
+    (self->parent ? getChildExists(self->parent, self->parentsIndex) : root) =
+        newSelf;
+    setChildrenParents(newSelf);
+    allocators->node256.release(self256);
+    self = newSelf;
+  } break;
+  }
+}
+
 // TODO fuse into erase child so we don't need to repeat branches on type
-void maybeDownsize(Node *self, Node *&root, NodeAllocators *allocators,
+void maybeDownsize(Node *&self, Node *&root, NodeAllocators *allocators,
                    ConflictSet::Impl *impl) {
   switch (self->type) {
   case Type::Node0:
@@ -732,6 +833,7 @@ void maybeDownsize(Node *self, Node *&root, NodeAllocators *allocators,
       }
 
       allocators->node4.release(self4);
+      self = newSelf;
     } else if (self->numChildren == 1) {
       if (!self->entryPresent) {
         auto *child = self4->children[0].child;
@@ -773,6 +875,7 @@ void maybeDownsize(Node *self, Node *&root, NodeAllocators *allocators,
         maxVersion(child, impl) = childMaxVersion;
 
         allocators->node4.release(self4);
+        self = child;
       }
     }
   } break;
@@ -795,6 +898,7 @@ void maybeDownsize(Node *self, Node *&root, NodeAllocators *allocators,
       } else {
         getChildExists(newSelf->parent, newSelf->parentsIndex) = newSelf;
       }
+      self = newSelf;
     }
     break;
   case Type::Node48:
@@ -821,6 +925,7 @@ void maybeDownsize(Node *self, Node *&root, NodeAllocators *allocators,
       } else {
         getChildExists(newSelf->parent, newSelf->parentsIndex) = newSelf;
       }
+      self = newSelf;
     }
     break;
   case Type::Node256:
@@ -847,6 +952,7 @@ void maybeDownsize(Node *self, Node *&root, NodeAllocators *allocators,
       } else {
         getChildExists(newSelf->parent, newSelf->parentsIndex) = newSelf;
       }
+      self = newSelf;
     }
     break;
   }
@@ -907,6 +1013,7 @@ void eraseChild(Node *self, uint8_t index, NodeAllocators *allocators,
     eraseChild(self->parent, self->parentsIndex, allocators, root, impl);
   } else {
     maybeDownsize(self, root, allocators, impl);
+    maybeDecreaseCapacity(self, root, allocators);
   }
 }
 
@@ -1131,10 +1238,6 @@ struct SearchStepWise {
   }
 };
 
-namespace {
-std::string getSearchPathPrintable(Node *n);
-}
-
 // Logically this is the same as performing firstGeq and then checking against
 // point or range version according to cmp, but this version short circuits as
 // soon as it can prove that there's no conflict.
@@ -1750,6 +1853,19 @@ template <bool kBegin>
         memmove(old->partialKey(), old->partialKey() + partialKeyIndex + 1,
                 old->partialKeyLen - (partialKeyIndex + 1));
         old->partialKeyLen -= partialKeyIndex + 1;
+
+        [[maybe_unused]] int oldCap = old->partialKeyCapacity;
+        maybeDecreaseCapacity(old, getRoot(impl), allocators);
+
+#if DEBUG_VERBOSE && !defined(NDEBUG)
+        if (old->partialKeyCapacity < oldCap) {
+          fprintf(stderr,
+                  "%s: Length: %d, capacity: %d, numChildren: %d, oldCapacity: "
+                  "%d\n",
+                  getSearchPathPrintable(old).c_str(), old->partialKeyLen,
+                  old->partialKeyCapacity, old->numChildren, oldCap);
+        }
+#endif
       }
       key = key.subspan(partialKeyIndex, key.size() - partialKeyIndex);
 
@@ -2106,6 +2222,8 @@ int64_t &maxVersion(Node *n, ConflictSet::Impl *impl) {
   }
 }
 
+Node *&getRoot(ConflictSet::Impl *impl) { return impl->root; }
+
 // ==================== END IMPLEMENTATION ====================
 
 // GCOVR_EXCL_START
@@ -2377,7 +2495,7 @@ Iterator firstGeq(Node *n, std::string_view key) {
   return total;
 }
 
-[[maybe_unused]] void checkMinChildCount(Node *node, bool &success) {
+[[maybe_unused]] void checkMemoryBoundInvariants(Node *node, bool &success) {
   int minNumChildren;
   switch (node->type) {
   case Type::Node0:
@@ -2397,16 +2515,25 @@ Iterator firstGeq(Node *n, std::string_view key) {
     break;
   }
   if (node->numChildren < minNumChildren) {
-    Arena arena;
     fprintf(stderr,
             "%s has %d children, which is less than the minimum required %d\n",
             getSearchPathPrintable(node).c_str(), node->numChildren,
             minNumChildren);
     success = false;
   }
+  if (node->numChildren > 0 &&
+      node->numChildren * node->partialKeyLen < node->partialKeyCapacity) {
+    fprintf(stderr,
+            "%s has %d children, partial key length %d, and partial key "
+            "capacity %d. It's required that nodes with children have children "
+            "* length >= capacity\n",
+            getSearchPathPrintable(node).c_str(), node->numChildren,
+            node->partialKeyLen, node->partialKeyCapacity);
+    success = false;
+  }
   for (int i = getChildGeq(node, 0); i >= 0; i = getChildGeq(node, i + 1)) {
     auto *child = getChildExists(node, i);
-    checkMinChildCount(child, success);
+    checkMemoryBoundInvariants(child, success);
   }
 }
 
@@ -2417,7 +2544,7 @@ bool checkCorrectness(Node *node, int64_t oldestVersion,
   checkParentPointers(node, success);
   checkMaxVersion(node, node, oldestVersion, success, impl);
   checkEntriesExist(node, success);
-  checkMinChildCount(node, success);
+  checkMemoryBoundInvariants(node, success);
 
   return success;
 }