Update GCOVR annotations now that jenkins agent has avx512

Events in the same group should be associated with the same set of
instructions

CMakeLists.txt

@@ -276,9 +276,15 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
   find_program(VALGRIND_EXE valgrind)
   if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
-    add_test(NAME conflict_set_blackbox_valgrind
+    list(LENGTH CORPUS_TESTS len)
-             COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
+    math(EXPR last "${len} - 1")
-                     $<TARGET_FILE:driver> ${CORPUS_TESTS})
+    set(partition_size 100)
+    foreach(i RANGE 0 ${last} ${partition_size})
+      list(SUBLIST CORPUS_TESTS ${i} ${partition_size} partition)
+      add_test(NAME conflict_set_blackbox_valgrind_${i}
+               COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
+                       $<TARGET_FILE:driver> ${partition})
+    endforeach()
   endif()
 
   # api smoke tests

ConflictSet.cpp

@@ -173,7 +173,7 @@ int BitSet::firstSetGeq(int i) const {
   return -1;
 }
 
-enum Type {
+enum Type : int8_t {
   Type_Node0,
   Type_Node3,
   Type_Node16,
@@ -191,12 +191,12 @@ struct Node {
   int32_t partialKeyLen;
   int16_t numChildren;
   bool entryPresent;
+  // Temp variable used to signal the end of the range during addWriteRange
+  bool endOfRange;
   uint8_t parentsIndex;
   /* end section that's copied to the next node */
 
   uint8_t *partialKey();
-  size_t size() const;
-
   Type getType() const { return type; }
   int32_t getCapacity() const { return partialKeyCapacity; }
 
@@ -220,84 +220,83 @@ constexpr int kNodeCopySize =
 
 struct Node0 : Node {
   constexpr static auto kType = Type_Node0;
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const struct Node3 &other);
-
   size_t size() const { return sizeof(Node0) + getCapacity(); }
 };
 
 struct Node3 : Node {
   constexpr static auto kMaxNodes = 3;
   constexpr static auto kType = Type_Node3;
-  // Sorted
+
-  uint8_t index[kMaxNodes];
   Node *children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
+  // Sorted
+  uint8_t index[kMaxNodes];
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const struct Node16 &other);
-
   size_t size() const { return sizeof(Node3) + getCapacity(); }
 };
 
 struct Node16 : Node {
   constexpr static auto kType = Type_Node16;
   constexpr static auto kMaxNodes = 16;
-  // Sorted
+
-  uint8_t index[kMaxNodes];
   Node *children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
+  // Sorted
+  uint8_t index[kMaxNodes];
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const struct Node48 &other);
-
   size_t size() const { return sizeof(Node16) + getCapacity(); }
 };
 
 struct Node48 : Node {
   constexpr static auto kType = Type_Node48;
   constexpr static auto kMaxNodes = 48;
-  BitSet bitSet;
-  int8_t nextFree;
-  int8_t index[256];
-  Node *children[kMaxNodes];
-  InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t reverseIndex[kMaxNodes];
   constexpr static int kMaxOfMaxPageSize = 16;
   constexpr static int kMaxOfMaxShift =
       std::countr_zero(uint32_t(kMaxOfMaxPageSize));
   constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
+
+  BitSet bitSet;
+  Node *children[kMaxNodes];
+  InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
+  uint8_t reverseIndex[kMaxNodes];
+  int8_t index[256];
 
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
-
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const struct Node256 &other);
-
   size_t size() const { return sizeof(Node48) + getCapacity(); }
 };
 
 struct Node256 : Node {
   constexpr static auto kType = Type_Node256;
-  BitSet bitSet;
+  constexpr static auto kMaxNodes = 256;
-  Node *children[256];
-  InternalVersionT childMaxVersion[256];
   constexpr static int kMaxOfMaxPageSize = 16;
   constexpr static int kMaxOfMaxShift =
       std::countr_zero(uint32_t(kMaxOfMaxPageSize));
-  constexpr static int kMaxOfMaxTotalPages = 256 / kMaxOfMaxPageSize;
+  constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
+
+  BitSet bitSet;
+  Node *children[kMaxNodes];
+  InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
+
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const Node256 &other);
-
   size_t size() const { return sizeof(Node256) + getCapacity(); }
 };
 
@@ -322,7 +321,7 @@ inline void Node3::copyChildrenAndKeyFrom(const Node0 &other) {
 inline void Node3::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  memcpy(index, other.index, sizeof(*this) - sizeof(Node));
+  memcpy(children, other.children, sizeof(*this) - sizeof(Node));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   for (int i = 0; i < numChildren; ++i) {
     assert(children[i]->parent == &other);
@@ -403,7 +402,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
   }
   memcpy(partialKey(), &other + 1, partialKeyLen);
   bitSet.init();
-  nextFree = Node16::kMaxNodes;
   int i = 0;
   for (auto x : other.index) {
     bitSet.set(x);
@@ -423,7 +421,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   bitSet = other.bitSet;
-  nextFree = other.nextFree;
   memcpy(index, other.index, sizeof(index));
   memset(children, 0, sizeof(children));
   const auto z = InternalVersionT::zero;
@@ -450,7 +447,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
   for (auto &v : childMaxVersion) {
     v = z;
   }
-  nextFree = other.numChildren;
   bitSet = other.bitSet;
   int i = 0;
   bitSet.forEachSet([&](int c) {
@@ -626,6 +622,7 @@ template <class T> struct BoundedFreeListAllocator {
 
   T *allocate(int partialKeyCapacity) {
     T *result = allocate_helper(partialKeyCapacity);
+    result->endOfRange = false;
     if constexpr (!std::is_same_v<T, Node0>) {
       memset(result->children, 0, sizeof(result->children));
       const auto z = InternalVersionT::zero;
@@ -693,23 +690,6 @@ uint8_t *Node::partialKey() {
   }
 }
 
-size_t Node::size() const {
-  switch (type) {
-  case Type_Node0:
-    return ((Node0 *)this)->size();
-  case Type_Node3:
-    return ((Node3 *)this)->size();
-  case Type_Node16:
-    return ((Node16 *)this)->size();
-  case Type_Node48:
-    return ((Node48 *)this)->size();
-  case Type_Node256:
-    return ((Node256 *)this)->size();
-  default:                   // GCOVR_EXCL_LINE
-    __builtin_unreachable(); // GCOVR_EXCL_LINE
-  }
-}
-
 // A type that's plumbed along the check call tree. Lifetime ends after each
 // check call.
 struct ReadContext {
@@ -788,22 +768,17 @@ private:
   BoundedFreeListAllocator<Node256> node256;
 };
 
-template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
+int getNodeIndex(Node3 *self, uint8_t index) {
-  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
+  Node3 *n = (Node3 *)self;
-
+  for (int i = 0; i < n->numChildren; ++i) {
-  // cachegrind says the plain loop is fewer instructions and more mis-predicted
+    if (n->index[i] == index) {
-  // branches. Microbenchmark says plain loop is faster. It's written in this
+      return i;
-  // weird "generic" way though in case someday we can use the simd
-  // implementation easily if we want.
-  if constexpr (std::is_same_v<NodeT, Node3>) {
-    Node3 *n = (Node3 *)self;
-    for (int i = 0; i < n->numChildren; ++i) {
-      if (n->index[i] == index) {
-        return i;
-      }
     }
-    return -1;
   }
+  return -1;
+}
+
+int getNodeIndex(Node16 *self, uint8_t index) {
 
 #ifdef HAS_AVX
   // Based on https://www.the-paper-trail.org/post/art-paper-notes/
@@ -816,7 +791,7 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   // keys aren't valid, we'll mask the results to only consider the valid ones
   // below.
   __m128i indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, indices);
 
   // Build a mask to select only the first node->num_children values from the
@@ -839,12 +814,11 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   // https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
 
   uint8x16_t indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   // 0xff for each match
   uint16x8_t results =
       vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(index), indices));
-  static_assert(NodeT::kMaxNodes <= 16);
+  assume(self->numChildren <= Node16::kMaxNodes);
-  assume(self->numChildren <= NodeT::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -1097,22 +1071,18 @@ ChildAndMaxVersion getChildAndMaxVersion(Node *self, uint8_t index) {
   }
 }
 
-template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
+Node *getChildGeq(Node0 *, int) { return nullptr; }
-  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
 
-  // cachegrind says the plain loop is fewer instructions and more mis-predicted
+Node *getChildGeq(Node3 *n, int child) {
-  // branches. Microbenchmark says plain loop is faster. It's written in this
+  for (int i = 0; i < n->numChildren; ++i) {
-  // weird "generic" way though so that someday we can use the simd
+    if (n->index[i] >= child) {
-  // implementation easily if we want.
+      return n->children[i];
-  if constexpr (std::is_same_v<NodeT, Node3>) {
-    Node3 *n = (Node3 *)self;
-    for (int i = 0; i < n->numChildren; ++i) {
-      if (n->index[i] >= child) {
-        return n->children[i];
-      }
     }
-    return nullptr;
   }
+  return nullptr;
+}
+
+Node *getChildGeq(Node16 *self, int child) {
   if (child > 255) {
     return nullptr;
   }
@@ -1120,7 +1090,7 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
 #ifdef HAS_AVX
   __m128i key_vec = _mm_set1_epi8(child);
   __m128i indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, _mm_min_epu8(key_vec, indices));
   int mask = (1 << self->numChildren) - 1;
   uint32_t bitfield = _mm_movemask_epi8(results) & mask;
@@ -1130,8 +1100,7 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
   memcpy(&indices, self->index, sizeof(self->index));
   // 0xff for each leq
   auto results = vcleq_u8(vdupq_n_u8(child), indices);
-  static_assert(NodeT::kMaxNodes <= 16);
+  assume(self->numChildren <= Node16::kMaxNodes);
-  assume(self->numChildren <= NodeT::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -1156,13 +1125,6 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
 #endif
 }
 
-Node *getChildGeq(Node0 *, int) { return nullptr; }
-Node *getChildGeq(Node3 *self, int child) {
-  return getChildGeqSimd(self, child);
-}
-Node *getChildGeq(Node16 *self, int child) {
-  return getChildGeqSimd(self, child);
-}
 Node *getChildGeq(Node48 *self, int child) {
   int c = self->bitSet.firstSetGeq(child);
   if (c < 0) {
@@ -1375,7 +1337,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
     auto *self3 = static_cast<Node3 *>(self);
     int i = self->numChildren - 1;
     for (; i >= 0; --i) {
-      if (int(self3->index[i]) < int(index)) {
+      if (self3->index[i] < index) {
         break;
       }
       self3->index[i + 1] = self3->index[i];
@@ -1405,7 +1367,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
     auto *self16 = static_cast<Node16 *>(self);
     int i = self->numChildren - 1;
     for (; i >= 0; --i) {
-      if (int(self16->index[i]) < int(index)) {
+      if (self16->index[i] < index) {
         break;
       }
       self16->index[i + 1] = self16->index[i];
@@ -1434,9 +1396,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
   insert48:
     auto *self48 = static_cast<Node48 *>(self);
     self48->bitSet.set(index);
-    ++self->numChildren;
+    auto nextFree = self48->numChildren++;
-    assert(self48->nextFree < 48);
-    int nextFree = self48->nextFree++;
     self48->index[index] = nextFree;
     self48->reverseIndex[nextFree] = index;
     auto &result = self48->children[nextFree];
@@ -1584,8 +1544,6 @@ void maybeDecreaseCapacity(Node *&self, WriteContext *tls,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
                                                  InternalVersionT zero) {
   uint32_t z;
@@ -1595,7 +1553,6 @@ __attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
       _mm512_sub_epi32(_mm512_loadu_epi32(vs), zvec), _mm512_setzero_epi32());
   _mm512_mask_storeu_epi32(vs, m, zvec);
 }
-// GCOVR_EXCL_STOP
 
 __attribute__((target("default")))
 #endif
@@ -1690,53 +1647,67 @@ void mergeWithChild(Node *&self, WriteContext *tls, ConflictSet::Impl *impl,
   tls->release(self3);
 }
 
-void maybeDownsize(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
+bool needsDownsize(Node *n) {
-                   Node *&dontInvalidate) {
+  static int minTable[] = {0, kMinChildrenNode3, kMinChildrenNode16,
+                           kMinChildrenNode48, kMinChildrenNode256};
+  return n->numChildren + n->entryPresent < minTable[n->getType()];
+}
 
-#if DEBUG_VERBOSE && !defined(NDEBUG)
+void downsize(Node3 *self, WriteContext *tls, ConflictSet::Impl *impl,
-  fprintf(stderr, "maybeDownsize: %s\n", getSearchPathPrintable(self).c_str());
+              Node *&dontInvalidate) {
-#endif
+  if (self->numChildren == 0) {
+    auto *newSelf = tls->allocate<Node0>(self->partialKeyLen);
+    newSelf->copyChildrenAndKeyFrom(*self);
+    getInTree(self, impl) = newSelf;
+    tls->release(self);
+  } else {
+    assert(self->numChildren == 1 && !self->entryPresent);
+    mergeWithChild(getInTree(self, impl), tls, impl, dontInvalidate, self);
+  }
+}
+
+void downsize(Node16 *self, WriteContext *tls, ConflictSet::Impl *impl) {
+  assert(self->numChildren + int(self->entryPresent) < kMinChildrenNode16);
+  auto *newSelf = tls->allocate<Node3>(self->partialKeyLen);
+  newSelf->copyChildrenAndKeyFrom(*self);
+  getInTree(self, impl) = newSelf;
+  tls->release(self);
+}
+
+void downsize(Node48 *self, WriteContext *tls, ConflictSet::Impl *impl) {
+  assert(self->numChildren + int(self->entryPresent) < kMinChildrenNode48);
+  auto *newSelf = tls->allocate<Node16>(self->partialKeyLen);
+  newSelf->copyChildrenAndKeyFrom(*self);
+  getInTree(self, impl) = newSelf;
+  tls->release(self);
+}
+
+void downsize(Node256 *self, WriteContext *tls, ConflictSet::Impl *impl) {
+  assert(self->numChildren + int(self->entryPresent) < kMinChildrenNode256);
+  auto *self256 = (Node256 *)self;
+  auto *newSelf = tls->allocate<Node48>(self->partialKeyLen);
+  newSelf->copyChildrenAndKeyFrom(*self256);
+  getInTree(self, impl) = newSelf;
+  tls->release(self256);
+}
+
+void downsize(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
+              Node *&dontInvalidate) {
 
   switch (self->getType()) {
   case Type_Node0:           // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
-  case Type_Node3: {
+  case Type_Node3:
-    auto *self3 = (Node3 *)self;
+    downsize(static_cast<Node3 *>(self), tls, impl, dontInvalidate);
-    if (self->numChildren == 0) {
+    break;
-      auto *newSelf = tls->allocate<Node0>(self->partialKeyLen);
-      newSelf->copyChildrenAndKeyFrom(*self3);
-      getInTree(self, impl) = newSelf;
-      tls->release(self3);
-    } else if (self->numChildren == 1 && !self->entryPresent) {
-      mergeWithChild(getInTree(self, impl), tls, impl, dontInvalidate, self3);
-    }
-  } break;
   case Type_Node16:
-    if (self->numChildren + int(self->entryPresent) < kMinChildrenNode16) {
+    downsize(static_cast<Node16 *>(self), tls, impl);
-      auto *self16 = (Node16 *)self;
-      auto *newSelf = tls->allocate<Node3>(self->partialKeyLen);
-      newSelf->copyChildrenAndKeyFrom(*self16);
-      getInTree(self, impl) = newSelf;
-      tls->release(self16);
-    }
     break;
   case Type_Node48:
-    if (self->numChildren + int(self->entryPresent) < kMinChildrenNode48) {
+    downsize(static_cast<Node48 *>(self), tls, impl);
-      auto *self48 = (Node48 *)self;
-      auto *newSelf = tls->allocate<Node16>(self->partialKeyLen);
-      newSelf->copyChildrenAndKeyFrom(*self48);
-      getInTree(self, impl) = newSelf;
-      tls->release(self48);
-    }
     break;
   case Type_Node256:
-    if (self->numChildren + int(self->entryPresent) < kMinChildrenNode256) {
+    downsize(static_cast<Node256 *>(self), tls, impl);
-      auto *self256 = (Node256 *)self;
-      auto *newSelf = tls->allocate<Node48>(self->partialKeyLen);
-      newSelf->copyChildrenAndKeyFrom(*self256);
-      getInTree(self, impl) = newSelf;
-      tls->release(self256);
-    }
     break;
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -1745,10 +1716,10 @@ void maybeDownsize(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
 
 // Precondition: self is not the root. May invalidate nodes along the search
 // path to self. May invalidate children of self->parent. Returns a pointer to
-// the node after self. If erase invalidates the pointee of `dontInvalidate`, it
+// the node after self. Precondition: `self->entryPresent`
-// will update it to its new pointee as well. Precondition: `self->entryPresent`
+
 Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
-            bool logical, Node *&dontInvalidate) {
+            bool logical) {
   ++tls->accum.entries_erased;
   assert(self->parent != nullptr);
 
@@ -1766,10 +1737,8 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
   self->entryPresent = false;
 
   if (self->numChildren != 0) {
-    const bool update = result == dontInvalidate;
+    if (needsDownsize(self)) {
-    maybeDownsize(self, tls, impl, result);
+      downsize(self, tls, impl, result);
-    if (update) {
-      dontInvalidate = result;
     }
     return result;
   }
@@ -1790,7 +1759,10 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
       parent3->children[i] = parent3->children[i + 1];
       parent3->childMaxVersion[i] = parent3->childMaxVersion[i + 1];
     }
-    assert(parent->numChildren > 0 || parent->entryPresent);
+
+    if (needsDownsize(parent3)) {
+      downsize(parent3, tls, impl, result);
+    }
   } break;
   case Type_Node16: {
     auto *parent16 = static_cast<Node16 *>(parent);
@@ -1803,15 +1775,16 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
       parent16->childMaxVersion[i] = parent16->childMaxVersion[i + 1];
     }
 
-    // By kMinChildrenNode16
+    if (needsDownsize(parent16)) {
-    assert(parent->numChildren > 0);
+      downsize(parent16, tls, impl, result);
+    }
   } break;
   case Type_Node48: {
     auto *parent48 = static_cast<Node48 *>(parent);
     parent48->bitSet.reset(parentsIndex);
     int8_t toRemoveChildrenIndex =
         std::exchange(parent48->index[parentsIndex], -1);
-    int8_t lastChildrenIndex = --parent48->nextFree;
+    auto lastChildrenIndex = --parent48->numChildren;
     assert(toRemoveChildrenIndex >= 0);
     assert(lastChildrenIndex >= 0);
     if (toRemoveChildrenIndex != lastChildrenIndex) {
@@ -1830,10 +1803,9 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
     }
     parent48->childMaxVersion[lastChildrenIndex] = tls->zero;
 
-    --parent->numChildren;
+    if (needsDownsize(parent48)) {
-
+      downsize(parent48, tls, impl, result);
-    // By kMinChildrenNode48
+    }
-    assert(parent->numChildren > 0);
   } break;
   case Type_Node256: {
     auto *parent256 = static_cast<Node256 *>(parent);
@@ -1842,20 +1814,14 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
 
     --parent->numChildren;
 
-    // By kMinChildrenNode256
+    if (needsDownsize(parent256)) {
-    assert(parent->numChildren > 0);
+      downsize(parent256, tls, impl, result);
-
+    }
   } break;
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
 
-  const bool update = result == dontInvalidate;
-  maybeDownsize(parent, tls, impl, result);
-  if (update) {
-    dontInvalidate = result;
-  }
-
   return result;
 }
 
@@ -2023,9 +1989,9 @@ downLeftSpine:
 #ifdef HAS_AVX
 uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
   uint32_t compared = 0;
-  __m128i w[4];
+  __m128i w[4]; // GCOVR_EXCL_LINE
   memcpy(w, vs, sizeof(w));
-  uint32_t r;
+  uint32_t r; // GCOVR_EXCL_LINE
   memcpy(&r, &rv, sizeof(r));
   const auto rvVec = _mm_set1_epi32(r);
   const auto zero = _mm_setzero_si128();
@@ -2037,8 +2003,6 @@ uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
   return compared;
 }
 
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) uint32_t
 compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
   uint32_t r;
@@ -2048,7 +2012,6 @@ compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
       _mm512_setzero_epi32());
 }
 #endif
-// GCOVR_EXCL_STOP
 
 // Returns true if v[i] <= readVersion for all i such that begin <= is[i] < end
 // Preconditions: begin <= end, end - begin < 256
@@ -2344,10 +2307,8 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       int c = self->bitSet.firstSetGeq(begin + 1);
       if (c >= 0 && c < end) {
         auto *child = self->children[self->index[c]];
-        if (child->entryPresent) {
+        if (child->entryPresent && child->entry.rangeVersion > readVersion) {
-          if (!(child->entry.rangeVersion <= readVersion)) {
+          return false;
-            return false;
-          };
         }
         begin = c;
       } else {
@@ -2380,10 +2341,8 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       int c = self->bitSet.firstSetGeq(begin + 1);
       if (c >= 0 && c < end) {
         auto *child = self->children[c];
-        if (child->entryPresent) {
+        if (child->entryPresent && child->entry.rangeVersion > readVersion) {
-          if (!(child->entry.rangeVersion <= readVersion)) {
+          return false;
-            return false;
-          };
         }
         begin = c;
       } else {
@@ -2426,9 +2385,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
       }
     }
     // Check inner pages
-    const int innerPageBegin = (begin >> Node256::kMaxOfMaxShift) + 1;
+    return scan16<kAVX512>(self->maxOfMax, firstPage + 1, lastPage,
-    const int innerPageEnd = (end - 1) >> Node256::kMaxOfMaxShift;
-    return scan16<kAVX512>(self->maxOfMax, innerPageBegin, innerPageEnd,
                            readVersion);
   }
   default:                   // GCOVR_EXCL_LINE
@@ -2437,14 +2394,11 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) bool
 checkMaxBetweenExclusive(Node *n, int begin, int end,
                          InternalVersionT readVersion, ReadContext *tls) {
   return checkMaxBetweenExclusiveImpl<true>(n, begin, end, readVersion, tls);
 }
-// GCOVR_EXCL_STOP
 __attribute__((target("default")))
 #endif
 bool checkMaxBetweenExclusive(Node *n, int begin, int end,
@@ -2931,7 +2885,6 @@ void eraseTree(Node *root, WriteContext *tls) {
     tls->accum.entries_erased += n->entryPresent;
     ++tls->accum.nodes_released;
 
-    removeNode(n);
     removeKey(n);
 
     switch (n->getType()) {
@@ -3071,13 +3024,21 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   }
   endNode->entry.rangeVersion = writeVersion;
 
-  for (beginNode = nextLogical(beginNode); beginNode != endNode;
+  // Erase nodes in range
-       beginNode = erase(beginNode, tls, impl, /*logical*/ true, endNode)) {
+  assert(!beginNode->endOfRange);
+  assert(!endNode->endOfRange);
+  endNode->endOfRange = true;
+  auto *iter = beginNode;
+  for (iter = nextLogical(iter); !iter->endOfRange;
+       iter = erase(iter, tls, impl, /*logical*/ true)) {
+    assert(!iter->endOfRange);
   }
+  assert(iter->endOfRange);
+  iter->endOfRange = false;
 
-  // Inserting end trashed endNode's maxVersion. Fix that. Safe to call since
+  // Inserting end trashed the last node's maxVersion. Fix that. Safe to call
-  // the end key always has non-zero size.
+  // since the end key always has non-zero size.
-  fixupMaxVersion(endNode, tls);
+  fixupMaxVersion(iter, tls);
 }
 
 Node *firstGeqPhysical(Node *n, const std::span<const uint8_t> key) {
@@ -3262,8 +3223,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
         // There's no way to insert a range such that range version of the right
         // node is greater than the point version of the left node
         assert(n->entry.rangeVersion <= oldestVersion);
-        Node *dummy = nullptr;
+        n = erase(n, &tls, this, /*logical*/ false);
-        n = erase(n, &tls, this, /*logical*/ false, dummy);
       } else {
         maybeDecreaseCapacity(n, &tls, this);
         n = nextPhysical(n);

Internal.h

@@ -748,7 +748,10 @@ struct TestDriver {
       fprintf(stderr, "%p Set oldest version: %" PRId64 "\n", this,
               oldestVersion);
 #endif
+      CALLGRIND_START_INSTRUMENTATION;
       cs.setOldestVersion(oldestVersion);
+      CALLGRIND_STOP_INSTRUMENTATION;
+
       if constexpr (kEnableAssertions) {
         refImpl.setOldestVersion(oldestVersion);
       }

README.md

@@ -24,15 +24,15 @@ Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-3
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               11.04 |       90,614,308.12 |    0.8% |          180.38 |           55.13 |  3.272 |          41.51 |    0.4% |      0.01 | `point reads`
+|               11.18 |       89,455,125.34 |    0.6% |          185.37 |           57.08 |  3.248 |          41.51 |    0.4% |      0.01 | `point reads`
-|               14.96 |       66,843,629.12 |    0.4% |          274.41 |           74.73 |  3.672 |          55.05 |    0.3% |      0.01 | `prefix reads`
+|               14.53 |       68,800,688.89 |    0.4% |          282.41 |           74.80 |  3.776 |          55.06 |    0.3% |      0.01 | `prefix reads`
-|               37.06 |       26,982,847.61 |    0.2% |          791.04 |          185.28 |  4.269 |         142.67 |    0.2% |      0.01 | `range reads`
+|               36.54 |       27,367,576.87 |    0.2% |          798.06 |          188.90 |  4.225 |         141.69 |    0.2% |      0.01 | `range reads`
-|               17.89 |       55,887,365.73 |    0.6% |          335.54 |           89.79 |  3.737 |          43.84 |    0.4% |      0.01 | `point writes`
+|               16.69 |       59,912,106.02 |    0.6% |          314.57 |           86.29 |  3.645 |          39.84 |    0.4% |      0.01 | `point writes`
-|               31.85 |       31,394,336.65 |    0.3% |          615.32 |          159.63 |  3.855 |          87.69 |    0.2% |      0.01 | `prefix writes`
+|               30.09 |       33,235,744.07 |    0.5% |          591.33 |          155.92 |  3.793 |          82.69 |    0.2% |      0.01 | `prefix writes`
-|               36.17 |       27,647,221.45 |    0.6% |          705.11 |          182.80 |  3.857 |         100.62 |    0.1% |      0.01 | `range writes`
+|               35.77 |       27,956,388.03 |    1.4% |          682.25 |          187.63 |  3.636 |          96.12 |    0.1% |      0.01 | `range writes`
-|               79.01 |       12,656,457.78 |    0.7% |        1,498.35 |          402.46 |  3.723 |         270.50 |    0.1% |      0.01 | `monotonic increasing point writes`
+|               74.04 |       13,505,408.41 |    2.7% |        1,448.95 |          392.10 |  3.695 |         260.53 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          303,667.50 |            3,293.08 |    1.1% |    3,931,273.00 |    1,612,702.50 |  2.438 |     806,223.33 |    0.0% |      0.01 | `worst case for radix tree`
+|          330,984.50 |            3,021.29 |    1.9% |    3,994,153.50 |    1,667,309.00 |  2.396 |     806,019.50 |    0.0% |      0.01 | `worst case for radix tree`
-|               83.70 |       11,947,443.83 |    0.7% |        1,738.03 |          429.06 |  4.051 |         270.01 |    0.0% |      0.01 | `create and destroy`
+|               92.46 |       10,814,961.65 |    0.5% |        1,800.00 |          463.41 |  3.884 |         297.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
@@ -47,7 +47,7 @@ Check: 4.47891 seconds, 364.05 MB/s, Add: 4.55599 seconds, 123.058 MB/s, Gc rati
 ## radix tree
 
 ```
-Check: 0.958985 seconds, 1700.28 MB/s, Add: 1.35083 seconds, 415.044 MB/s, Gc ratio: 44.4768%, Peak idle memory: 2.33588e+06
+Check: 0.953012 seconds, 1710.94 MB/s, Add: 1.30025 seconds, 431.188 MB/s, Gc ratio: 43.9816%, Peak idle memory: 2.28375e+06
 ```
 
 ## hash table

ServerBench.cpp

@@ -6,11 +6,15 @@
 #include <string.h>
 #include <string>
 #include <string_view>
+#include <sys/ioctl.h>
 #include <sys/resource.h>
 #include <sys/socket.h>
+#include <sys/types.h>
 #include <sys/uio.h>
 #include <thread>
 #include <unistd.h>
+#include <utility>
+#include <vector>
 
 #include "ConflictSet.h"
 #include "third_party/nadeau.h"
@@ -20,8 +24,8 @@ std::atomic<int64_t> transactions;
 constexpr int kBaseSearchDepth = 32;
 constexpr int kWindowSize = 10000000;
 
-std::basic_string<uint8_t> numToKey(int64_t num) {
+std::string numToKey(int64_t num) {
-  std::basic_string<uint8_t> result;
+  std::string result;
   result.resize(kBaseSearchDepth + sizeof(int64_t));
   memset(result.data(), 0, kBaseSearchDepth);
   int64_t be = __builtin_bswap64(num);
@@ -41,13 +45,13 @@ void workload(weaselab::ConflictSet *cs) {
       auto pointK = numToKey(pointRv);
       weaselab::ConflictSet::ReadRange reads[] = {
           {
-              {pointK.data(), int(pointK.size())},
+              {(const uint8_t *)pointK.data(), int(pointK.size())},
               {nullptr, 0},
               pointRv,
           },
           {
-              {beginK.data(), int(beginK.size())},
+              {(const uint8_t *)beginK.data(), int(beginK.size())},
-              {endK.data(), int(endK.size())},
+              {(const uint8_t *)endK.data(), int(endK.size())},
               version - 2,
           },
       };
@@ -66,7 +70,7 @@ void workload(weaselab::ConflictSet *cs) {
     {
       weaselab::ConflictSet::WriteRange w;
       auto k = numToKey(version);
-      w.begin.p = k.data();
+      w.begin.p = (const uint8_t *)k.data();
       w.end.len = 0;
       if (version % (kWindowSize / 2) == 0) {
         for (int l = 0; l <= k.size(); ++l) {
@@ -79,9 +83,9 @@ void workload(weaselab::ConflictSet *cs) {
         int64_t beginN = version - kWindowSize + rand() % kWindowSize;
         auto b = numToKey(beginN);
         auto e = numToKey(beginN + 1000);
-        w.begin.p = b.data();
+        w.begin.p = (const uint8_t *)b.data();
         w.begin.len = b.size();
-        w.end.p = e.data();
+        w.end.p = (const uint8_t *)e.data();
         w.end.len = e.size();
         cs->addWrites(&w, 1, version);
       }
@@ -164,36 +168,29 @@ double toSeconds(timeval t) {
   return double(t.tv_sec) + double(t.tv_usec) * 1e-6;
 }
 
-#include <linux/perf_event.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <sys/ioctl.h>
-#include <sys/syscall.h>
-#include <sys/types.h>
-#include <unistd.h>
-
 #ifdef __linux__
+#include <linux/perf_event.h>
 struct PerfCounter {
-  explicit PerfCounter(int event) {
+  PerfCounter(int type, int config, const std::string &labels = {},
+              int groupLeaderFd = -1)
+      : labels(labels) {
     struct perf_event_attr pe;
 
     memset(&pe, 0, sizeof(pe));
-    pe.type = PERF_TYPE_HARDWARE;
+    pe.type = type;
     pe.size = sizeof(pe);
-    pe.config = event;
+    pe.config = config;
     pe.inherit = 1;
     pe.exclude_kernel = 1;
     pe.exclude_hv = 1;
 
-    fd = perf_event_open(&pe, 0, -1, -1, 0);
+    fd = perf_event_open(&pe, 0, -1, groupLeaderFd, 0);
-    if (fd == -1) {
+    if (fd < 0 && errno != ENOENT && errno != EINVAL) {
-      fprintf(stderr, "Error opening leader %llx\n", pe.config);
+      perror(labels.c_str());
-      exit(EXIT_FAILURE);
     }
   }
 
-  int64_t total() {
+  int64_t total() const {
     int64_t count;
     if (read(fd, &count, sizeof(count)) != sizeof(count)) {
       perror("read instructions from perf");
@@ -202,10 +199,27 @@ struct PerfCounter {
     return count;
   }
 
-  ~PerfCounter() { close(fd); }
+  PerfCounter(PerfCounter &&other)
+      : fd(std::exchange(other.fd, -1)), labels(std::move(other.labels)) {}
+  PerfCounter &operator=(PerfCounter &&other) {
+    fd = std::exchange(other.fd, -1);
+    labels = std::move(other.labels);
+    return *this;
+  }
+
+  ~PerfCounter() {
+    if (fd >= 0) {
+      close(fd);
+    }
+  }
+
+  bool ok() const { return fd >= 0; }
+  const std::string &getLabels() const { return labels; }
+  int getFd() const { return fd; }
 
 private:
   int fd;
+  std::string labels;
   static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
                               int cpu, int group_fd, unsigned long flags) {
     int ret;
@@ -214,11 +228,6 @@ private:
     return ret;
   }
 };
-#else
-struct PerfCounter {
-  explicit PerPerfCounter(int) {}
-  int64_t total() { return 0; }
-};
 #endif
 
 int main(int argc, char **argv) {
@@ -233,8 +242,50 @@ int main(int argc, char **argv) {
     int metricsCount;
     cs.getMetricsV1(&metrics, &metricsCount);
 
-    PerfCounter instructions{PERF_COUNT_HW_INSTRUCTIONS};
+#ifdef __linux__
-    PerfCounter cycles{PERF_COUNT_HW_CPU_CYCLES};
+    PerfCounter instructions{PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS};
+    PerfCounter cycles{PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES, "",
+                       instructions.getFd()};
+
+    std::vector<PerfCounter> cacheCounters;
+    for (auto [id, idStr] : std::initializer_list<std::pair<int, std::string>>{
+             {PERF_COUNT_HW_CACHE_L1D, "l1d"},
+             {PERF_COUNT_HW_CACHE_L1I, "l1i"},
+             {PERF_COUNT_HW_CACHE_LL, "ll"},
+             {PERF_COUNT_HW_CACHE_DTLB, "dtlb"},
+             {PERF_COUNT_HW_CACHE_ITLB, "itlb"},
+             {PERF_COUNT_HW_CACHE_BPU, "bpu"},
+             {PERF_COUNT_HW_CACHE_NODE, "node"},
+         }) {
+      for (auto [op, opStr] :
+           std::initializer_list<std::pair<int, std::string>>{
+               {PERF_COUNT_HW_CACHE_OP_READ, "read"},
+               {PERF_COUNT_HW_CACHE_OP_WRITE, "write"},
+               {PERF_COUNT_HW_CACHE_OP_PREFETCH, "prefetch"},
+           }) {
+        int groupLeaderFd = -1;
+        for (auto [result, resultStr] :
+             std::initializer_list<std::pair<int, std::string>>{
+                 {PERF_COUNT_HW_CACHE_RESULT_MISS, "miss"},
+                 {PERF_COUNT_HW_CACHE_RESULT_ACCESS, "access"},
+             }) {
+          auto labels = "{id=\"" + idStr + "\", op=\"" + opStr +
+                        "\", result=\"" + resultStr + "\"}";
+          cacheCounters.emplace_back(PERF_TYPE_HW_CACHE,
+                                     id | (op << 8) | (result << 16), labels,
+                                     groupLeaderFd);
+          if (!cacheCounters.back().ok()) {
+            cacheCounters.pop_back();
+          } else {
+            if (groupLeaderFd == -1) {
+              groupLeaderFd = cacheCounters.back().getFd();
+            }
+          }
+        }
+      }
+    }
+#endif
+
     auto w = std::thread{workload, &cs};
 
     for (;;) {
@@ -262,6 +313,7 @@ int main(int argc, char **argv) {
               "transactions_total ";
       body += std::to_string(transactions.load(std::memory_order_relaxed));
       body += "\n";
+#ifdef __linux__
       body += "# HELP instructions_total Total number of instructions\n"
               "# TYPE instructions_total counter\n"
               "instructions_total ";
@@ -272,6 +324,13 @@ int main(int argc, char **argv) {
               "cycles_total ";
       body += std::to_string(cycles.total());
       body += "\n";
+      body += "# HELP cache_event_total Total number of cache events\n"
+              "# TYPE cache_event_total counter\n";
+      for (const auto &counter : cacheCounters) {
+        body += "cache_event_total" + counter.getLabels() + " " +
+                std::to_string(counter.total()) + "\n";
+      }
+#endif
 
       for (int i = 0; i < metricsCount; ++i) {
         body += "# HELP ";