Include childType in ChildAndMaxVersion

Unsigned compare is incorrect here

Bench.cpp

@@ -7,7 +7,6 @@
 void showMemory(const ConflictSet &cs);
 #endif
 
-#define ANKERL_NANOBENCH_IMPLEMENT
 #include "third_party/nanobench.h"
 
 constexpr int kNumKeys = 1000000;

CMakeLists.txt

@@ -138,6 +138,8 @@ include(CTest)
 # disable tests if this is being used through e.g. FetchContent
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
+  add_library(nanobench ${CMAKE_CURRENT_SOURCE_DIR}/nanobench.cpp)
+
   set(TEST_FLAGS -Wall -Wextra -Wunreachable-code -Wpedantic -UNDEBUG)
 
   # corpus tests, which are tests curated by libfuzzer. The goal is to get broad
@@ -185,6 +187,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   target_include_directories(conflict_set_main
                              PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
   target_compile_definitions(conflict_set_main PRIVATE ENABLE_MAIN)
+  target_link_libraries(conflict_set_main PRIVATE nanobench)
 
   if(NOT APPLE)
     # libfuzzer target, to generate/manage corpus
@@ -245,6 +248,19 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
     add_test(NAME conflict_set_blackbox_${hash} COMMAND driver ${TEST})
   endforeach()
 
+  find_program(VALGRIND_EXE valgrind)
+  if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
+    list(LENGTH CORPUS_TESTS len)
+    math(EXPR last "${len} - 1")
+    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()
+
   # scripted tests. Written manually to fill in anything libfuzzer couldn't
   # find.
   if(NOT CMAKE_CROSSCOMPILING)
@@ -265,19 +281,14 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
           ${Python3_EXECUTABLE}
           ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
           --build-dir ${CMAKE_CURRENT_BINARY_DIR})
-    endforeach()
-  endif()
-
-  find_program(VALGRIND_EXE valgrind)
-  if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
-    list(LENGTH CORPUS_TESTS len)
-    math(EXPR last "${len} - 1")
-    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> ${partition})
+      if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
+        add_test(
+          NAME script_test_${TEST}_valgrind
+          COMMAND
+            ${VALGRIND_EXE} ${Python3_EXECUTABLE}
+            ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
+            --build-dir ${CMAKE_CURRENT_BINARY_DIR})
+      endif()
     endforeach()
   endif()
 
@@ -330,7 +341,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
   # bench
   add_executable(conflict_set_bench Bench.cpp)
-  target_link_libraries(conflict_set_bench PRIVATE ${PROJECT_NAME})
+  target_link_libraries(conflict_set_bench PRIVATE ${PROJECT_NAME} nanobench)
   set_target_properties(conflict_set_bench PROPERTIES SKIP_BUILD_RPATH ON)
   add_executable(real_data_bench RealDataBench.cpp)
   target_link_libraries(real_data_bench PRIVATE ${PROJECT_NAME})
@@ -345,6 +356,14 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   add_executable(server_bench ServerBench.cpp)
   target_link_libraries(server_bench PRIVATE ${PROJECT_NAME})
   set_target_properties(server_bench PROPERTIES SKIP_BUILD_RPATH ON)
+
+  add_executable(interleaving_test InterleavingTest.cpp)
+  # work around lack of musttail for gcc
+  if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_BUILD_TYPE STREQUAL "Debug")
+    target_compile_options(interleaving_test PRIVATE -Og
+                                                     -foptimize-sibling-calls)
+  endif()
+  target_link_libraries(interleaving_test PRIVATE nanobench)
 endif()
 
 # packaging

ConflictSet.cpp

@@ -203,6 +203,60 @@ enum Type : int8_t {
 
 template <class T> struct BoundedFreeListAllocator;
 
+struct TaggedNodePointer {
+  TaggedNodePointer() = default;
+  operator struct Node *() { return (struct Node *)(uintptr_t)p; }
+  operator struct Node0 *() {
+    assert(t == Type_Node0);
+    return (struct Node0 *)(uintptr_t)p;
+  }
+  operator struct Node3 *() {
+    assert(t == Type_Node3);
+    return (struct Node3 *)(uintptr_t)p;
+  }
+  operator struct Node16 *() {
+    assert(t == Type_Node16);
+    return (struct Node16 *)(uintptr_t)p;
+  }
+  operator struct Node48 *() {
+    assert(t == Type_Node48);
+    return (struct Node48 *)(uintptr_t)p;
+  }
+  operator struct Node256 *() {
+    assert(t == Type_Node256);
+    return (struct Node256 *)(uintptr_t)p;
+  }
+  /*implicit*/ TaggedNodePointer(std::nullptr_t) : p(0), t(Type_Node0) {}
+  /*implicit*/ TaggedNodePointer(Node0 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node0) {}
+  /*implicit*/ TaggedNodePointer(Node3 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node3) {}
+  /*implicit*/ TaggedNodePointer(Node16 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node16) {}
+  /*implicit*/ TaggedNodePointer(Node48 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node48) {}
+  /*implicit*/ TaggedNodePointer(Node256 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node256) {}
+
+  bool operator!=(std::nullptr_t) { return p != 0 || t != Type_Node0; }
+  bool operator==(std::nullptr_t) { return p == 0 && t == Type_Node0; }
+  bool operator==(const TaggedNodePointer &) const = default;
+  bool operator==(Node *n) const { return (uintptr_t)n == p; }
+  Node *operator->() { return (Node *)(uintptr_t)p; }
+  Type getType() { return t; }
+
+  TaggedNodePointer(const TaggedNodePointer &) = default;
+  TaggedNodePointer &operator=(const TaggedNodePointer &) = default;
+  /*implicit*/ TaggedNodePointer(Node *n);
+
+private:
+  TaggedNodePointer(struct Node *p, Type t) : p((uintptr_t)p), t(t) {
+    assume(p != 0);
+  }
+  uintptr_t p : 56;
+  Type t : 8;
+};
+
 struct Node {
 
   /* begin section that's copied to the next node */
@@ -228,6 +282,9 @@ private:
   int32_t partialKeyCapacity;
 };
 
+TaggedNodePointer::TaggedNodePointer(Node *n)
+    : TaggedNodePointer(n, n->getType()) {}
+
 constexpr int kNodeCopyBegin = offsetof(Node, entry);
 constexpr int kNodeCopySize =
     offsetof(Node, parentsIndex) + sizeof(Node::parentsIndex) - kNodeCopyBegin;
@@ -251,7 +308,7 @@ struct Node3 : Node {
   constexpr static auto kMaxNodes = 3;
   constexpr static auto kType = Type_Node3;
 
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   // Sorted
   uint8_t index[kMaxNodes];
@@ -267,7 +324,7 @@ struct Node16 : Node {
   constexpr static auto kType = Type_Node16;
   constexpr static auto kMaxNodes = 16;
 
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   // Sorted
   uint8_t index[kMaxNodes];
@@ -288,7 +345,7 @@ struct Node48 : Node {
   constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
 
   BitSet bitSet;
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
   uint8_t reverseIndex[kMaxNodes];
@@ -310,7 +367,7 @@ struct Node256 : Node {
   constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
 
   BitSet bitSet;
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
 
@@ -766,6 +823,8 @@ private:
 
 int getNodeIndex(Node3 *self, uint8_t index) {
   Node3 *n = (Node3 *)self;
+  assume(n->numChildren >= 1);
+  assume(n->numChildren <= 3);
   for (int i = 0; i < n->numChildren; ++i) {
     if (n->index[i] == index) {
       return i;
@@ -774,6 +833,18 @@ int getNodeIndex(Node3 *self, uint8_t index) {
   return -1;
 }
 
+int getNodeIndexExists(Node3 *self, uint8_t index) {
+  Node3 *n = (Node3 *)self;
+  assume(n->numChildren >= 1);
+  assume(n->numChildren <= 3);
+  for (int i = 0; i < n->numChildren; ++i) {
+    if (n->index[i] == index) {
+      return i;
+    }
+  }
+  __builtin_unreachable(); // GCOVR_EXCL_LINE
+}
+
 int getNodeIndex(Node16 *self, uint8_t index) {
 
 #ifdef HAS_AVX
@@ -834,27 +905,66 @@ int getNodeIndex(Node16 *self, uint8_t index) {
 #endif
 }
 
+int getNodeIndexExists(Node16 *self, uint8_t index) {
+
+#ifdef HAS_AVX
+  __m128i key_vec = _mm_set1_epi8(index);
+  __m128i indices;
+  memcpy(&indices, self->index, Node16::kMaxNodes);
+  __m128i results = _mm_cmpeq_epi8(key_vec, indices);
+  uint32_t mask = (1 << self->numChildren) - 1;
+  uint32_t bitfield = _mm_movemask_epi8(results) & mask;
+  assume(bitfield != 0);
+  return std::countr_zero(bitfield);
+#elif defined(HAS_ARM_NEON)
+  // Based on
+  // 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, Node16::kMaxNodes);
+  // 0xff for each match
+  uint16x8_t results =
+      vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(index), indices));
+  assume(self->numChildren <= Node16::kMaxNodes);
+  uint64_t mask = self->numChildren == 16
+                      ? uint64_t(-1)
+                      : (uint64_t(1) << (self->numChildren * 4)) - 1;
+  // 0xf for each match in valid range
+  uint64_t bitfield =
+      vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0) & mask;
+  assume(bitfield != 0);
+  return std::countr_zero(bitfield) / 4;
+#else
+  for (int i = 0; i < self->numChildren; ++i) {
+    if (self->index[i] == index) {
+      return i;
+    }
+  }
+  __builtin_unreachable(); // GCOVR_EXCL_LINE
+#endif
+}
+
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node3 *self, uint8_t index) {
-  return self->children[getNodeIndex(self, index)];
+TaggedNodePointer &getChildExists(Node3 *self, uint8_t index) {
+  return self->children[getNodeIndexExists(self, index)];
 }
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node16 *self, uint8_t index) {
-  return self->children[getNodeIndex(self, index)];
+TaggedNodePointer &getChildExists(Node16 *self, uint8_t index) {
+  return self->children[getNodeIndexExists(self, index)];
 }
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node48 *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node48 *self, uint8_t index) {
   assert(self->bitSet.test(index));
   return self->children[self->index[index]];
 }
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node256 *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node256 *self, uint8_t index) {
   assert(self->bitSet.test(index));
   return self->children[index];
 }
 
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node *self, uint8_t index) {
   switch (self->getType()) {
   case Type_Node0:           // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -885,12 +995,12 @@ InternalVersionT maxVersion(Node *n) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   case Type_Node3: {
     auto *n3 = static_cast<Node3 *>(n);
-    int i = getNodeIndex(n3, index);
+    int i = getNodeIndexExists(n3, index);
     return n3->childMaxVersion[i];
   }
   case Type_Node16: {
     auto *n16 = static_cast<Node16 *>(n);
-    int i = getNodeIndex(n16, index);
+    int i = getNodeIndexExists(n16, index);
     return n16->childMaxVersion[i];
   }
   case Type_Node48: {
@@ -918,12 +1028,12 @@ InternalVersionT exchangeMaxVersion(Node *n, InternalVersionT newMax) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   case Type_Node3: {
     auto *n3 = static_cast<Node3 *>(n);
-    int i = getNodeIndex(n3, index);
+    int i = getNodeIndexExists(n3, index);
     return std::exchange(n3->childMaxVersion[i], newMax);
   }
   case Type_Node16: {
     auto *n16 = static_cast<Node16 *>(n);
-    int i = getNodeIndex(n16, index);
+    int i = getNodeIndexExists(n16, index);
     return std::exchange(n16->childMaxVersion[i], newMax);
   }
   case Type_Node48: {
@@ -952,13 +1062,13 @@ void setMaxVersion(Node *n, InternalVersionT newMax) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   case Type_Node3: {
     auto *n3 = static_cast<Node3 *>(n);
-    int i = getNodeIndex(n3, index);
+    int i = getNodeIndexExists(n3, index);
     n3->childMaxVersion[i] = newMax;
     return;
   }
   case Type_Node16: {
     auto *n16 = static_cast<Node16 *>(n);
-    int i = getNodeIndex(n16, index);
+    int i = getNodeIndexExists(n16, index);
     n16->childMaxVersion[i] = newMax;
     return;
   }
@@ -985,7 +1095,7 @@ void setMaxVersion(Node *n, InternalVersionT newMax) {
   }
 }
 
-Node *&getInTree(Node *n, ConflictSet::Impl *);
+TaggedNodePointer &getInTree(Node *n, ConflictSet::Impl *);
 
 Node *getChild(Node0 *, uint8_t) { return nullptr; }
 Node *getChild(Node3 *self, uint8_t index) {
@@ -1022,6 +1132,7 @@ Node *getChild(Node *self, uint8_t index) {
 struct ChildAndMaxVersion {
   Node *child;
   InternalVersionT maxVersion;
+  Type childType;
 };
 
 ChildAndMaxVersion getChildAndMaxVersion(Node0 *, uint8_t) { return {}; }
@@ -1030,24 +1141,28 @@ ChildAndMaxVersion getChildAndMaxVersion(Node3 *self, uint8_t index) {
   if (i < 0) {
     return {};
   }
-  return {self->children[i], self->childMaxVersion[i]};
+  return {self->children[i], self->childMaxVersion[i],
+          self->children[i].getType()};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node16 *self, uint8_t index) {
   int i = getNodeIndex(self, index);
   if (i < 0) {
     return {};
   }
-  return {self->children[i], self->childMaxVersion[i]};
+  return {self->children[i], self->childMaxVersion[i],
+          self->children[i].getType()};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node48 *self, uint8_t index) {
   int i = self->index[index];
   if (i < 0) {
     return {};
   }
-  return {self->children[i], self->childMaxVersion[i]};
+  return {self->children[i], self->childMaxVersion[i],
+          self->children[i].getType()};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node256 *self, uint8_t index) {
-  return {self->children[index], self->childMaxVersion[index]};
+  return {self->children[index], self->childMaxVersion[index],
+          self->children[index].getType()};
 }
 
 ChildAndMaxVersion getChildAndMaxVersion(Node *self, uint8_t index) {
@@ -1070,6 +1185,8 @@ ChildAndMaxVersion getChildAndMaxVersion(Node *self, uint8_t index) {
 Node *getChildGeq(Node0 *, int) { return nullptr; }
 
 Node *getChildGeq(Node3 *n, int child) {
+  assume(n->numChildren >= 1);
+  assume(n->numChildren <= 3);
   for (int i = 0; i < n->numChildren; ++i) {
     if (n->index[i] >= child) {
       return n->children[i];
@@ -1190,14 +1307,15 @@ Node *getFirstChildExists(Node *self) {
   }
 }
 
-void consumePartialKeyFull(Node *&self, std::span<const uint8_t> &key,
+void consumePartialKeyFull(TaggedNodePointer &self,
+                           std::span<const uint8_t> &key,
                            InternalVersionT writeVersion, WriteContext *tls) {
   // Handle an existing partial key
   int commonLen = std::min<int>(self->partialKeyLen, key.size());
   int partialKeyIndex =
       longestCommonPrefix(self->partialKey(), key.data(), commonLen);
   if (partialKeyIndex < self->partialKeyLen) {
-    auto *old = self;
+    Node *old = self;
     // Since root cannot have a partial key
     assert(old->parent != nullptr);
     InternalVersionT oldMaxVersion = exchangeMaxVersion(old, writeVersion);
@@ -1235,7 +1353,7 @@ void consumePartialKeyFull(Node *&self, std::span<const uint8_t> &key,
 // Consume any partial key of `self`, and update `self` and
 // `key` such that `self` is along the search path of `key`
 inline __attribute__((always_inline)) void
-consumePartialKey(Node *&self, std::span<const uint8_t> &key,
+consumePartialKey(TaggedNodePointer &self, std::span<const uint8_t> &key,
                   InternalVersionT writeVersion, WriteContext *tls) {
   if (self->partialKeyLen > 0) {
     consumePartialKeyFull(self, key, writeVersion, tls);
@@ -1246,8 +1364,10 @@ consumePartialKey(Node *&self, std::span<const uint8_t> &key,
 // such that the search path of the result + key is the same as the search path
 // of self + key before the call. Creates a node if necessary. Updates
 // `maxVersion` for result.
-Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
-                        InternalVersionT newMaxVersion, WriteContext *tls) {
+TaggedNodePointer &getOrCreateChild(TaggedNodePointer &self,
+                                    std::span<const uint8_t> &key,
+                                    InternalVersionT newMaxVersion,
+                                    WriteContext *tls) {
 
   int index = key.front();
   key = key.subspan(1, key.size() - 1);
@@ -1549,7 +1669,6 @@ __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);
 }
-
 __attribute__((target("default")))
 #endif
 
@@ -1606,10 +1725,11 @@ void rezero(Node *n, InternalVersionT z) {
 }
 #endif
 
-void mergeWithChild(Node *&self, WriteContext *tls, ConflictSet::Impl *impl,
-                    Node *&dontInvalidate, Node3 *self3) {
+void mergeWithChild(TaggedNodePointer &self, WriteContext *tls,
+                    ConflictSet::Impl *impl, Node *&dontInvalidate,
+                    Node3 *self3) {
   assert(!self3->entryPresent);
-  auto *child = self3->children[0];
+  Node *child = self3->children[0];
   int minCapacity = self3->partialKeyLen + 1 + child->partialKeyLen;
 
   if (minCapacity > child->getCapacity()) {
@@ -1858,7 +1978,7 @@ bool checkPointRead(Node *n, const std::span<const uint8_t> key,
       goto downLeftSpine;
     }
 
-    auto [child, maxV] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, maxV, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
@@ -1928,7 +2048,7 @@ bool checkPrefixRead(Node *n, const std::span<const uint8_t> key,
       return maxVersion(n) <= readVersion;
     }
 
-    auto [child, maxV] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, maxV, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
@@ -2201,7 +2321,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
     if (!mask) {
       return true;
     }
-    auto *child = self->children[std::countr_zero(mask)];
+    Node *child = self->children[std::countr_zero(mask)];
     const bool firstRangeOk =
         !child->entryPresent || child->entry.rangeVersion <= readVersion;
     uint32_t compared = 0;
@@ -2276,7 +2396,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
     if (!mask) {
       return true;
     }
-    auto *child = self->children[std::countr_zero(mask)];
+    Node *child = self->children[std::countr_zero(mask)];
     const bool firstRangeOk =
         !child->entryPresent || child->entry.rangeVersion <= readVersion;
 
@@ -2321,7 +2441,7 @@ 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]];
+        Node *child = self->children[self->index[c]];
         if (child->entryPresent && child->entry.rangeVersion > readVersion) {
           return false;
         }
@@ -2355,7 +2475,7 @@ 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];
+        Node *child = self->children[c];
         if (child->entryPresent && child->entry.rangeVersion > readVersion) {
           return false;
         }
@@ -2416,6 +2536,7 @@ checkMaxBetweenExclusive(Node *n, int begin, int end,
 }
 __attribute__((target("default")))
 #endif
+
 bool checkMaxBetweenExclusive(Node *n, int begin, int end,
                          InternalVersionT readVersion, ReadContext *tls) {
   return checkMaxBetweenExclusiveImpl<false>(n, begin, end, readVersion, tls);
@@ -2526,7 +2647,7 @@ bool checkRangeLeftSide(Node *n, std::span<const uint8_t> key, int prefixLen,
       }
     }
 
-    auto [child, maxV] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, maxV, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
@@ -2714,7 +2835,7 @@ bool checkRangeRead(Node *n, std::span<const uint8_t> begin,
     if (remaining.size() == 0) {
       break;
     }
-    auto [child, v] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, v, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       break;
     }
@@ -2780,8 +2901,10 @@ checkMaxBetweenExclusiveImpl<true>(Node *n, int begin, int end,
 // of the result will have `maxVersion` set to `writeVersion` as a
 // postcondition. Nodes along the search path may be invalidated. Callers must
 // ensure that the max version of the self argument is updated.
-[[nodiscard]] Node **insert(Node **self, std::span<const uint8_t> key,
-                            InternalVersionT writeVersion, WriteContext *tls) {
+[[nodiscard]] TaggedNodePointer *insert(TaggedNodePointer *self,
+                                        std::span<const uint8_t> key,
+                                        InternalVersionT writeVersion,
+                                        WriteContext *tls) {
 
   for (; key.size() != 0; ++tls->accum.insert_iterations) {
     self = &getOrCreateChild(*self, key, writeVersion, tls);
@@ -2809,17 +2932,23 @@ void eraseTree(Node *root, WriteContext *tls) {
     } break;
     case Type_Node3: {
       auto *n3 = static_cast<Node3 *>(n);
-      toFree.append(std::span<Node *>(n3->children, n3->numChildren));
+      for (int i = 0; i < n3->numChildren; ++i) {
+        toFree.push_back(n3->children[i]);
+      }
       tls->release(n3);
     } break;
     case Type_Node16: {
       auto *n16 = static_cast<Node16 *>(n);
-      toFree.append(std::span<Node *>(n16->children, n16->numChildren));
+      for (int i = 0; i < n16->numChildren; ++i) {
+        toFree.push_back(n16->children[i]);
+      }
       tls->release(n16);
     } break;
     case Type_Node48: {
       auto *n48 = static_cast<Node48 *>(n);
-      toFree.append(std::span<Node *>(n48->children, n48->numChildren));
+      for (int i = 0; i < n48->numChildren; ++i) {
+        toFree.push_back(n48->children[i]);
+      }
       tls->release(n48);
     } break;
     case Type_Node256: {
@@ -2835,10 +2964,10 @@ void eraseTree(Node *root, WriteContext *tls) {
   }
 }
 
-void addPointWrite(Node *&root, std::span<const uint8_t> key,
+void addPointWrite(TaggedNodePointer &root, std::span<const uint8_t> key,
                    InternalVersionT writeVersion, WriteContext *tls) {
   ++tls->accum.point_writes;
-  auto *n = *insert(&root, key, writeVersion, tls);
+  auto n = *insert(&root, key, writeVersion, tls);
   if (!n->entryPresent) {
     ++tls->accum.entries_inserted;
     auto *p = nextLogical(n);
@@ -2855,6 +2984,72 @@ void addPointWrite(Node *&root, std::span<const uint8_t> key,
   }
 }
 
+#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
+__attribute__((target("avx512f"))) InternalVersionT horizontalMaxUpTo16(
+    InternalVersionT *vs, [[maybe_unused]] InternalVersionT z, int len) {
+  assume(len <= 16);
+#if USE_64_BIT
+  // Hope it gets vectorized
+  InternalVersionT max = vs[0];
+  for (int i = 1; i < len; ++i) {
+    max = std::max(vs[i], max);
+  }
+  return max;
+#else
+  uint32_t zero;
+  memcpy(&zero, &z, sizeof(zero));
+  auto zeroVec = _mm512_set1_epi32(zero);
+  auto max = InternalVersionT(
+      zero +
+      _mm512_reduce_max_epi32(_mm512_sub_epi32(
+          _mm512_mask_loadu_epi32(zeroVec, _mm512_int2mask((1 << len) - 1), vs),
+          zeroVec)));
+  return max;
+#endif
+}
+__attribute__((target("default")))
+#endif
+
+InternalVersionT
+horizontalMaxUpTo16(InternalVersionT *vs, InternalVersionT, int len) {
+  assume(len <= 16);
+  InternalVersionT max = vs[0];
+  for (int i = 1; i < len; ++i) {
+    max = std::max(vs[i], max);
+  }
+  return max;
+}
+
+#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
+__attribute__((target("avx512f"))) InternalVersionT
+horizontalMax16(InternalVersionT *vs, [[maybe_unused]] InternalVersionT z) {
+#if USE_64_BIT
+  // Hope it gets vectorized
+  InternalVersionT max = vs[0];
+  for (int i = 1; i < 16; ++i) {
+    max = std::max(vs[i], max);
+  }
+  return max;
+#else
+  uint32_t zero; // GCOVR_EXCL_LINE
+  memcpy(&zero, &z, sizeof(zero));
+  auto zeroVec = _mm512_set1_epi32(zero);
+  return InternalVersionT(zero + _mm512_reduce_max_epi32(_mm512_sub_epi32(
+                                     _mm512_loadu_epi32(vs), zeroVec)));
+#endif
+}
+__attribute__((target("default")))
+#endif
+
+InternalVersionT
+horizontalMax16(InternalVersionT *vs, InternalVersionT) {
+  InternalVersionT max = vs[0];
+  for (int i = 1; i < 16; ++i) {
+    max = std::max(vs[i], max);
+  }
+  return max;
+}
+
 // Precondition: `node->entryPresent`, and node is not the root
 void fixupMaxVersion(Node *node, WriteContext *tls) {
   assert(node->parent);
@@ -2866,15 +3061,13 @@ void fixupMaxVersion(Node *node, WriteContext *tls) {
     break;
   case Type_Node3: {
     auto *self3 = static_cast<Node3 *>(node);
-    for (int i = 0; i < self3->numChildren; ++i) {
-      max = std::max(self3->childMaxVersion[i], max);
-    }
+    max = std::max(max, horizontalMaxUpTo16(self3->childMaxVersion, tls->zero,
+                                            self3->numChildren));
   } break;
   case Type_Node16: {
     auto *self16 = static_cast<Node16 *>(node);
-    for (int i = 0; i < self16->numChildren; ++i) {
-      max = std::max(self16->childMaxVersion[i], max);
-    }
+    max = std::max(max, horizontalMaxUpTo16(self16->childMaxVersion, tls->zero,
+                                            self16->numChildren));
   } break;
   case Type_Node48: {
     auto *self48 = static_cast<Node48 *>(node);
@@ -2884,9 +3077,7 @@ void fixupMaxVersion(Node *node, WriteContext *tls) {
   } break;
   case Type_Node256: {
     auto *self256 = static_cast<Node256 *>(node);
-    for (auto v : self256->maxOfMax) {
-      max = std::max(v, max);
-    }
+    max = std::max(max, horizontalMax16(self256->childMaxVersion, tls->zero));
   } break;
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -2894,7 +3085,7 @@ void fixupMaxVersion(Node *node, WriteContext *tls) {
   setMaxVersion(node, max);
 }
 
-void addWriteRange(Node *&root, std::span<const uint8_t> begin,
+void addWriteRange(TaggedNodePointer &root, std::span<const uint8_t> begin,
                    std::span<const uint8_t> end, InternalVersionT writeVersion,
                    WriteContext *tls, ConflictSet::Impl *impl) {
 
@@ -2907,12 +3098,12 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   ++tls->accum.range_writes;
   const bool beginIsPrefix = lcp == int(begin.size());
 
-  Node **useAsRoot = insert(&root, begin.subspan(0, lcp), writeVersion, tls);
+  auto useAsRoot = insert(&root, begin.subspan(0, lcp), writeVersion, tls);
 
   begin = begin.subspan(lcp, begin.size() - lcp);
   end = end.subspan(lcp, end.size() - lcp);
 
-  auto *beginNode = *insert(useAsRoot, begin, writeVersion, tls);
+  Node *beginNode = *insert(useAsRoot, begin, writeVersion, tls);
   addKey(beginNode);
   if (!beginNode->entryPresent) {
     ++tls->accum.entries_inserted;
@@ -2923,7 +3114,7 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   }
   beginNode->entry.pointVersion = writeVersion;
 
-  auto *endNode = *insert(useAsRoot, end, writeVersion, tls);
+  Node *endNode = *insert(useAsRoot, end, writeVersion, tls);
   addKey(endNode);
   if (!endNode->entryPresent) {
     ++tls->accum.entries_inserted;
@@ -2943,7 +3134,7 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   assert(!beginNode->endOfRange);
   assert(!endNode->endOfRange);
   endNode->endOfRange = true;
-  auto *iter = beginNode;
+  Node *iter = beginNode;
   for (iter = nextLogical(iter); !iter->endOfRange;
        iter = erase(iter, tls, impl, /*logical*/ true)) {
     assert(!iter->endOfRange);
@@ -3258,7 +3449,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   std::span<const uint8_t> removalKey;
   int64_t keyUpdates;
 
-  Node *root;
+  TaggedNodePointer root;
   InternalVersionT oldestVersion;
   int64_t oldestVersionFullPrecision;
   int64_t oldestExtantVersion;
@@ -3339,7 +3530,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 };
 
-Node *&getInTree(Node *n, ConflictSet::Impl *impl) {
+TaggedNodePointer &getInTree(Node *n, ConflictSet::Impl *impl) {
   return n->parent == nullptr ? impl->root
                               : getChildExists(n->parent, n->parentsIndex);
 }
@@ -3923,7 +4114,6 @@ struct __attribute__((visibility("default"))) PeakPrinter {
 
 #ifdef ENABLE_MAIN
 
-#define ANKERL_NANOBENCH_IMPLEMENT
 #include "third_party/nanobench.h"
 
 template <int kN> void benchRezero() {
@@ -3979,6 +4169,24 @@ template <int kN> void benchScan2() {
   });
 }
 
+void benchHorizontal16() {
+  ankerl::nanobench::Bench bench;
+  InternalVersionT vs[16];
+  for (int i = 0; i < 16; ++i) {
+    vs[i] = InternalVersionT(rand() % 1000 + 1000);
+  }
+#if !USE_64_BIT
+  InternalVersionT::zero = InternalVersionT(rand() % 1000);
+#endif
+  bench.run("horizontal16", [&]() {
+    bench.doNotOptimizeAway(horizontalMax16(vs, InternalVersionT::zero));
+  });
+  int x = rand() % 15 + 1;
+  bench.run("horizontalUpTo16", [&]() {
+    bench.doNotOptimizeAway(horizontalMaxUpTo16(vs, InternalVersionT::zero, x));
+  });
+}
+
 void benchLCP(int len) {
   ankerl::nanobench::Bench bench;
   std::vector<uint8_t> lhs(len);
@@ -4011,11 +4219,7 @@ void printTree() {
   debugPrintDot(stdout, cs.root, &cs);
 }
 
-int main(void) {
-  for (int i = 0; i < 256; ++i) {
-    benchLCP(i);
-  }
-}
+int main(void) { benchHorizontal16(); }
 #endif
 
 #ifdef ENABLE_FUZZ

InterleavingTest.cpp

@@ -0,0 +1,256 @@
+#include <alloca.h>
+#include <cassert>
+#ifdef __x86_64__
+#include <immintrin.h>
+#endif
+
+#include "third_party/nanobench.h"
+
+struct Job {
+  int *input;
+  // Returned void* is a function pointer to the next continuation. We have to
+  // use void* because otherwise the type would be recursive.
+  typedef void *(*continuation)(Job *);
+  continuation next;
+};
+
+void *stepJob(Job *j) {
+  auto done = --(*j->input) == 0;
+#ifdef __x86_64__
+  _mm_clflush(j->input);
+#endif
+  return done ? nullptr : (void *)stepJob;
+}
+
+void sequential(Job **jobs, int count) {
+  for (int i = 0; i < count; ++i) {
+    do {
+      jobs[i]->next = (Job::continuation)jobs[i]->next(jobs[i]);
+    } while (jobs[i]->next);
+  }
+}
+
+void sequentialNoFuncPtr(Job **jobs, int count) {
+  for (int i = 0; i < count; ++i) {
+    while (stepJob(jobs[i]))
+      ;
+  }
+}
+
+void interleaveSwapping(Job **jobs, int remaining) {
+  int current = 0;
+  while (remaining > 0) {
+    auto next = (Job::continuation)jobs[current]->next(jobs[current]);
+    jobs[current]->next = next;
+    if (next == nullptr) {
+      jobs[current] = jobs[remaining - 1];
+      --remaining;
+    } else {
+      ++current;
+    }
+    if (current == remaining) {
+      current = 0;
+    }
+  }
+}
+
+void interleaveBoundedCyclicList(Job **jobs, int count) {
+  if (count == 0) {
+    return;
+  }
+
+  constexpr int kConcurrent = 32;
+  Job *inProgress[kConcurrent];
+  int nextJob[kConcurrent];
+
+  int started = std::min(kConcurrent, count);
+  for (int i = 0; i < started; i++) {
+    inProgress[i] = jobs[i];
+    nextJob[i] = i + 1;
+  }
+  nextJob[started - 1] = 0;
+
+  int prevJob = started - 1;
+  int job = 0;
+  for (;;) {
+    auto next = (Job::continuation)inProgress[job]->next(inProgress[job]);
+    inProgress[job]->next = next;
+    if (next == nullptr) {
+      if (started == count) {
+        if (prevJob == job)
+          break;
+        nextJob[prevJob] = nextJob[job];
+        job = prevJob;
+      } else {
+        int temp = started++;
+        inProgress[job] = jobs[temp];
+      }
+    }
+    prevJob = job;
+    job = nextJob[job];
+  }
+}
+
+#ifndef __has_attribute
+#define __has_attribute(x) 0
+#endif
+
+#if __has_attribute(musttail)
+#define MUSTTAIL __attribute__((musttail))
+#else
+#define MUSTTAIL
+#endif
+
+struct Context {
+  constexpr static int kConcurrent = 32;
+  Job **jobs;
+  Job *inProgress[kConcurrent];
+  void (*continuation[kConcurrent])(Context *, int64_t prevJob, int64_t job,
+                                    int64_t started, int64_t count);
+  int nextJob[kConcurrent];
+};
+
+void keepGoing(Context *context, int64_t prevJob, int64_t job, int64_t started,
+               int64_t count) {
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->continuation[job](context, prevJob, job, started,
+                                             count);
+}
+
+void stepJobTailCall(Context *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count);
+
+void complete(Context *context, int64_t prevJob, int64_t job, int64_t started,
+              int64_t count) {
+  if (started == count) {
+    if (prevJob == job) {
+      return;
+    }
+    context->nextJob[prevJob] = context->nextJob[job];
+    job = prevJob;
+  } else {
+    context->inProgress[job] = context->jobs[started++];
+    context->continuation[job] = stepJobTailCall;
+  }
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->continuation[job](context, prevJob, job, started,
+                                             count);
+}
+
+void stepJobTailCall(Context *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count) {
+  auto *j = context->inProgress[job];
+  auto done = --(*j->input) == 0;
+#ifdef __x86_64__
+  _mm_clflush(j->input);
+#endif
+  if (done) {
+    MUSTTAIL return complete(context, prevJob, job, started, count);
+  } else {
+    context->continuation[job] = stepJobTailCall;
+    MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+  }
+}
+
+void useTailCalls(Job **jobs, int count) {
+  if (count == 0) {
+    return;
+  }
+  Context context;
+  context.jobs = jobs;
+  int64_t started = std::min(Context::kConcurrent, count);
+  for (int i = 0; i < started; i++) {
+    context.inProgress[i] = jobs[i];
+    context.nextJob[i] = i + 1;
+    context.continuation[i] = stepJobTailCall;
+  }
+  context.nextJob[started - 1] = 0;
+  int prevJob = started - 1;
+  int job = 0;
+  return context.continuation[job](&context, prevJob, job, started, count);
+}
+
+void interleaveCyclicList(Job **jobs, int count) {
+  auto *nextJob = (int *)alloca(sizeof(int) * count);
+
+  for (int i = 0; i < count - 1; ++i) {
+    nextJob[i] = i + 1;
+  }
+  nextJob[count - 1] = 0;
+
+  int prevJob = count - 1;
+  int job = 0;
+  for (;;) {
+    auto next = (Job::continuation)jobs[job]->next(jobs[job]);
+    jobs[job]->next = next;
+    if (next == nullptr) {
+      if (prevJob == job)
+        break;
+      nextJob[prevJob] = nextJob[job];
+      job = prevJob;
+    }
+    prevJob = job;
+    job = nextJob[job];
+  }
+}
+
+int main() {
+  ankerl::nanobench::Bench bench;
+
+  constexpr int kNumJobs = 10000;
+  bench.relative(true);
+
+  Job jobs[kNumJobs];
+  Job jobsCopy[kNumJobs];
+  int iters = 0;
+  int originalInput[kNumJobs];
+  for (int i = 0; i < kNumJobs; ++i) {
+    originalInput[i] = rand() % 5 + 3;
+    jobs[i].input = new int{originalInput[i]};
+    jobs[i].next = stepJob;
+    iters += *jobs[i].input;
+  }
+  bench.batch(iters);
+
+  for (auto [scheduler, name] :
+       {std::make_pair(sequentialNoFuncPtr, "sequentialNoFuncPtr"),
+        std::make_pair(sequential, "sequential"),
+        std::make_pair(useTailCalls, "useTailCalls"),
+        std::make_pair(interleaveSwapping, "interleavingSwapping"),
+        std::make_pair(interleaveBoundedCyclicList,
+                       "interleaveBoundedCyclicList"),
+        std::make_pair(interleaveCyclicList, "interleaveCyclicList")}) {
+    for (int i = 0; i < kNumJobs; ++i) {
+      *jobs[i].input = originalInput[i];
+    }
+    memcpy(jobsCopy, jobs, sizeof(jobs));
+    Job *ps[kNumJobs];
+    for (int i = 0; i < kNumJobs; ++i) {
+      ps[i] = jobsCopy + i;
+    }
+    scheduler(ps, kNumJobs);
+    for (int i = 0; i < kNumJobs; ++i) {
+      if (*jobsCopy[i].input != 0) {
+        fprintf(stderr, "%s failed\n", name);
+        abort();
+      }
+    }
+
+    bench.run(name, [&]() {
+      for (int i = 0; i < kNumJobs; ++i) {
+        *jobs[i].input = originalInput[i];
+      }
+      memcpy(jobsCopy, jobs, sizeof(jobs));
+      Job *ps[kNumJobs];
+      for (int i = 0; i < kNumJobs; ++i) {
+        ps[i] = jobsCopy + i;
+      }
+      scheduler(ps, kNumJobs);
+    });
+  }
+  for (int i = 0; i < kNumJobs; ++i) {
+    delete jobs[i].input;
+  }
+}

nanobench.cpp

@@ -0,0 +1,2 @@
+#define ANKERL_NANOBENCH_IMPLEMENT
+#include "third_party/nanobench.h"