Fix setOldestVersion

We were not accumulating keyUpdates strictly faster than we were
spending them for a monotonically increasing workload.

.pre-commit-config.yaml

@@ -16,3 +16,11 @@ repos:
         entry: '^#define DEBUG_VERBOSE 1$'
         language: pygrep
         types: [c++]
+  - repo: local
+    hooks:
+      - id: debug verbose check
+        name: disallow checking in SHOW_MEMORY=1
+        description: disallow checking in SHOW_MEMORY=1
+        entry: '^#define SHOW_MEMORY 1$'
+        language: pygrep
+        types: [c++]

Bench.cpp

@@ -4,6 +4,10 @@
 #include <cstring>
 #include <string>
 
+#if SHOW_MEMORY
+void showMemory(const ConflictSet &cs);
+#endif
+
 #define ANKERL_NANOBENCH_IMPLEMENT
 #include "third_party/nanobench.h"
 

ConflictSet.cpp

@@ -51,6 +51,10 @@ struct BoundedFreeListAllocator {
   static_assert(sizeof(T) >= sizeof(void *));
 
   T *allocate() {
+#if SHOW_MEMORY
+    ++liveAllocations;
+    maxLiveAllocations = std::max(maxLiveAllocations, liveAllocations);
+#endif
     if (freeListSize == 0) {
       assert(freeList == nullptr);
       return new (safe_malloc(sizeof(T))) T;
@@ -65,6 +69,9 @@ struct BoundedFreeListAllocator {
   }
 
   void release(T *p) {
+#if SHOW_MEMORY
+    --liveAllocations;
+#endif
     p->~T();
     if (freeListSize == kMaxFreeListSize) {
       return free(p);
@@ -84,10 +91,18 @@ struct BoundedFreeListAllocator {
     }
   }
 
+#if SHOW_MEMORY
+  int64_t highWaterMarkBytes() const { return maxLiveAllocations * sizeof(T); }
+#endif
+
 private:
   static constexpr int kMaxFreeListSize = kMemoryBound / sizeof(T);
   int freeListSize = 0;
   void *freeList = nullptr;
+#if SHOW_MEMORY
+  int64_t maxLiveAllocations = 0;
+  int64_t liveAllocations = 0;
+#endif
 };
 
 struct BitSet {
@@ -116,14 +131,20 @@ struct BitSet {
     // Check begin partial word
     if (begin & 63) {
       uint64_t word = words[begin >> 6] & (uint64_t(-1) << (begin & 63));
-      while (word) {
-        uint64_t temp = word & -word;
-        int index = (begin & ~63) + std::countr_zero(word);
-        f(index);
-        word ^= temp;
+      if (std::popcount(word) + (begin & 63) == 64) {
+        while (begin & 63) {
+          f(begin++);
+        }
+      } else {
+        while (word) {
+          uint64_t temp = word & -word;
+          int index = (begin & ~63) + std::countr_zero(word);
+          f(index);
+          word ^= temp;
+        }
+        begin &= ~63;
+        begin += 64;
       }
-      begin &= ~63;
-      begin += 64;
     }
 
     // Check inner, full words
@@ -147,11 +168,17 @@ struct BitSet {
     if (end & 63) {
       // Check end partial word
       uint64_t word = words[end >> 6] & ~(uint64_t(-1) << (end & 63));
-      while (word) {
-        uint64_t temp = word & -word;
-        int index = begin + std::countr_zero(word);
-        f(index);
-        word ^= temp;
+      if (std::popcount(word) == (end & 63)) {
+        while (begin < end) {
+          f(begin++);
+        }
+      } else {
+        while (word) {
+          uint64_t temp = word & -word;
+          int index = begin + std::countr_zero(word);
+          f(index);
+          word ^= temp;
+        }
       }
     }
   }
@@ -193,6 +220,7 @@ int BitSet::firstSetGeq(int i) const {
 }
 
 enum class Type : int8_t {
+  Node1,
   Node4,
   Node16,
   Node48,
@@ -221,10 +249,18 @@ struct Child {
   Node *child;
 };
 
+struct Node1 : Node {
+  // Sorted
+  uint8_t index[16]; // 16 so that we can use the same simd index search
+                     // implementation as Node16
+  Child children[1];
+  Node1() { this->type = Type::Node1; }
+};
+
 struct Node4 : Node {
   // Sorted
   uint8_t index[16]; // 16 so that we can use the same simd index search
-                     // implementation for Node4 as Node16
+                     // implementation as Node16
   Child children[4];
   Node4() { this->type = Type::Node4; }
 };
@@ -259,6 +295,7 @@ struct Node256 : Node {
 };
 
 struct NodeAllocators {
+  BoundedFreeListAllocator<Node1> node1;
   BoundedFreeListAllocator<Node4> node4;
   BoundedFreeListAllocator<Node16> node16;
   BoundedFreeListAllocator<Node48> node48;
@@ -334,6 +371,7 @@ Node *&getChildExists(Node *self, uint8_t index) {
     return self48->children[self48->index[index]].child;
   } else {
     auto *self256 = static_cast<Node256 *>(self);
+    assert(self256->bitSet.test(index));
     return self256->children[index].child;
   }
   __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -350,6 +388,7 @@ int64_t getChildMaxVersion(Node *self, uint8_t index) {
     return self48->children[self48->index[index]].childMaxVersion;
   } else {
     auto *self256 = static_cast<Node256 *>(self);
+    assert(self256->bitSet.test(index));
     return self256->children[index].childMaxVersion;
   }
   __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -445,6 +484,12 @@ int getChildGeq(Node *self, int child) {
   return -1;
 }
 
+void setChildrenParents(Node4 *n) {
+  for (int i = 0; i < n->numChildren; ++i) {
+    n->children[i].child->parent = n;
+  }
+}
+
 void setChildrenParents(Node16 *n) {
   for (int i = 0; i < n->numChildren; ++i) {
     n->children[i].child->parent = n;
@@ -452,21 +497,13 @@ void setChildrenParents(Node16 *n) {
 }
 
 void setChildrenParents(Node48 *n) {
-  for (int i = 0; i < 256; ++i) {
-    int c = n->index[i];
-    if (c != -1) {
-      n->children[c].child->parent = n;
-    }
-  }
+  n->bitSet.forEachInRange(
+      [&](int i) { n->children[n->index[i]].child->parent = n; }, 0, 256);
 }
 
 void setChildrenParents(Node256 *n) {
-  for (int i = 0; i < 256; ++i) {
-    auto *child = n->children[i].child;
-    if (child != nullptr) {
-      child->parent = n;
-    }
-  }
+  n->bitSet.forEachInRange([&](int i) { n->children[i].child->parent = n; }, 0,
+                           256);
 }
 
 // Caller is responsible for assigning a non-null pointer to the returned
@@ -494,7 +531,23 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     }
   }
 
-  if (self->type == Type::Node4) {
+  if (self->type == Type::Node1) {
+    auto *self1 = static_cast<Node1 *>(self);
+
+    if (self->numChildren == 1) {
+      auto *newSelf = allocators->node4.allocate();
+      memcpy((void *)newSelf, self, sizeof(Node1));
+      newSelf->type = Type::Node4;
+      allocators->node1.release(self1);
+      setChildrenParents(newSelf);
+      self = newSelf;
+    } else {
+      assert(self->numChildren == 0);
+    }
+
+    goto insert16;
+
+  } else if (self->type == Type::Node4) {
     auto *self4 = static_cast<Node4 *>(self);
 
     if (self->numChildren == 4) {
@@ -554,12 +607,11 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
       auto *newSelf = allocators->node256.allocate();
       memcpy((void *)newSelf, self, offsetof(Node, type));
       newSelf->bitSet = self48->bitSet;
-      for (int i = 0; i < 256; ++i) {
-        int c = self48->index[i];
-        if (c >= 0) {
-          newSelf->children[i] = self48->children[c];
-        }
-      }
+      newSelf->bitSet.forEachInRange(
+          [&](int i) {
+            newSelf->children[i] = self48->children[self48->index[i]];
+          },
+          0, 256);
       allocators->node48.release(self48);
       setChildrenParents(newSelf);
       self = newSelf;
@@ -577,6 +629,7 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     result = nullptr;
     return result;
   } else {
+    assert(self->type == Type::Node256);
   insert256:
     auto *self256 = static_cast<Node256 *>(self);
     ++self->numChildren;
@@ -589,6 +642,9 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
 void eraseChild(Node *self, uint8_t index, NodeAllocators *allocators) {
   auto *child = getChildExists(self, index);
   switch (child->type) {
+  case Type::Node1:
+    allocators->node1.release((Node1 *)child);
+    break;
   case Type::Node4:
     allocators->node4.release((Node4 *)child);
     break;
@@ -974,6 +1030,8 @@ int64_t maxBetweenExclusive(Node *n, int begin, int end) {
     }
   }
   switch (n->type) {
+  case Type::Node1:
+    [[fallthrough]];
   case Type::Node4:
     [[fallthrough]];
   case Type::Node16: {
@@ -1475,7 +1533,7 @@ template <bool kBegin>
         auto *old = *self;
         int64_t oldMaxVersion = maxVersion(old, impl);
 
-        *self = allocators->node4.allocate();
+        *self = allocators->node1.allocate();
 
         memcpy((void *)*self, old, offsetof(Node, type));
         (*self)->partialKeyLen = partialKeyIndex;
@@ -1521,7 +1579,7 @@ template <bool kBegin>
 
     auto &child = getOrCreateChild(*self, key.front(), allocators);
     if (!child) {
-      child = allocators->node4.allocate();
+      child = allocators->node1.allocate();
       child->parent = *self;
       child->parentsIndex = key.front();
       maxVersion(child, impl) =
@@ -1793,11 +1851,11 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       auto begin = std::span<const uint8_t>(w.begin.p, w.begin.len);
       auto end = std::span<const uint8_t>(w.end.p, w.end.len);
       if (w.end.len > 0) {
-        keyUpdates += 2;
+        keyUpdates += 3;
         addWriteRange(root, oldestVersion, begin, end, w.writeVersion,
                       &allocators, this);
       } else {
-        keyUpdates += 1;
+        keyUpdates += 2;
         addPointWrite(root, oldestVersion, begin, w.writeVersion, &allocators,
                       this);
       }
@@ -1812,7 +1870,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     Node *prev = firstGeq(root, removalKey).n;
     // There's no way to erase removalKey without introducing a key after it
     assert(prev != nullptr);
-    while (keyUpdates-- > 0) {
+    for (; keyUpdates > 0; --keyUpdates) {
       Node *n = nextLogical(prev);
       if (n == nullptr) {
         removalKey = {};
@@ -1841,7 +1899,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   explicit Impl(int64_t oldestVersion) : oldestVersion(oldestVersion) {
     // Insert ""
-    root = allocators.node4.allocate();
+    root = allocators.node1.allocate();
     rootMaxVersion = oldestVersion;
     root->entry.pointVersion = oldestVersion;
     root->entry.rangeVersion = oldestVersion;
@@ -1877,6 +1935,7 @@ int64_t &maxVersion(Node *n, ConflictSet::Impl *impl) {
     return n48->children[n48->index[index]].childMaxVersion;
   } else {
     auto *n256 = static_cast<Node256 *>(n);
+    assert(n256->bitSet.test(index));
     return n256->children[index].childMaxVersion;
   }
 }
@@ -1908,6 +1967,23 @@ ConflictSet::~ConflictSet() {
   }
 }
 
+#if SHOW_MEMORY
+__attribute__((visibility("default"))) void showMemory(const ConflictSet &cs) {
+  ConflictSet::Impl *impl;
+  memcpy(&impl, &cs, sizeof(impl)); // NOLINT
+  fprintf(stderr, "Max Node1 memory usage: %" PRId64 "\n",
+          impl->allocators.node1.highWaterMarkBytes());
+  fprintf(stderr, "Max Node4 memory usage: %" PRId64 "\n",
+          impl->allocators.node4.highWaterMarkBytes());
+  fprintf(stderr, "Max Node16 memory usage: %" PRId64 "\n",
+          impl->allocators.node16.highWaterMarkBytes());
+  fprintf(stderr, "Max Node48 memory usage: %" PRId64 "\n",
+          impl->allocators.node48.highWaterMarkBytes());
+  fprintf(stderr, "Max Node256 memory usage: %" PRId64 "\n",
+          impl->allocators.node256.highWaterMarkBytes());
+}
+#endif
+
 ConflictSet::ConflictSet(ConflictSet &&other) noexcept
     : impl(std::exchange(other.impl, nullptr)) {}
 
@@ -2176,7 +2252,7 @@ int main(void) {
   ConflictSet::Impl cs{0};
   for (int j = 0; j < 256; ++j) {
     getOrCreateChild(cs.root, j, &cs.allocators) =
-        cs.allocators.node4.allocate();
+        cs.allocators.node1.allocate();
     if (j % 10 == 0) {
       bench.run("MaxExclusive " + std::to_string(j), [&]() {
         bench.doNotOptimizeAway(maxBetweenExclusive(cs.root, 0, 256));

Internal.h

@@ -21,6 +21,7 @@
 #include <callgrind.h>
 
 #define DEBUG_VERBOSE 0
+#define SHOW_MEMORY 0
 
 [[nodiscard]] inline auto
 operator<=>(const std::span<const uint8_t> &lhs,