Remove bogus assert

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

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.13
+  VERSION 0.0.14
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"
@@ -72,12 +72,6 @@ else()
   add_link_options(-Wl,--gc-sections)
 endif()
 
-if(EMSCRIPTEN)
-  # https://github.com/emscripten-core/emscripten/issues/15377#issuecomment-1285167486
-  add_link_options(-lnodefs.js -lnoderawfs.js)
-  add_link_options(-s ALLOW_MEMORY_GROWTH)
-endif()
-
 if(NOT USE_SIMD_FALLBACK)
   cmake_push_check_state()
   list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
@@ -144,6 +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
@@ -191,6 +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
@@ -336,7 +333,7 @@ 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})
@@ -351,6 +348,9 @@ 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)
+  target_link_libraries(interleaving_test PRIVATE nanobench)
 endif()
 
 # packaging

ConflictSet.cpp

@@ -17,9 +17,9 @@ limitations under the License.
 #include "ConflictSet.h"
 #include "Internal.h"
 #include "LongestCommonPrefix.h"
+#include "Metrics.h"
 
 #include <algorithm>
-#include <atomic>
 #include <bit>
 #include <cassert>
 #include <cstddef>
@@ -578,39 +578,6 @@ static_assert(kBytesPerKey - sizeof(Node0) >= kMinNodeSurplus);
 
 constexpr int64_t kFreeListMaxMemory = 1 << 20;
 
-struct Metric {
-  Metric *prev;
-  const char *name;
-  const char *help;
-  ConflictSet::MetricsV1::Type type;
-  std::atomic<int64_t> value;
-
-protected:
-  Metric(ConflictSet::Impl *impl, const char *name, const char *help,
-         ConflictSet::MetricsV1::Type type);
-};
-
-struct Gauge : private Metric {
-  Gauge(ConflictSet::Impl *impl, const char *name, const char *help)
-      : Metric(impl, name, help, ConflictSet::MetricsV1::Gauge) {}
-
-  void set(int64_t value) {
-    this->value.store(value, std::memory_order_relaxed);
-  }
-};
-
-struct Counter : private Metric {
-  Counter(ConflictSet::Impl *impl, const char *name, const char *help)
-      : Metric(impl, name, help, ConflictSet::MetricsV1::Counter) {}
-  // Expensive. Accumulate locally and then call add instead of repeatedly
-  // calling add.
-  void add(int64_t value) {
-    assert(value >= 0);
-    static_assert(std::atomic<int64_t>::is_always_lock_free);
-    this->value.fetch_add(value, std::memory_order_relaxed);
-  }
-};
-
 template <class T> struct BoundedFreeListAllocator {
 
   static_assert(sizeof(T) >= sizeof(void *));
@@ -799,6 +766,8 @@ 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;
@@ -807,6 +776,18 @@ 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
@@ -867,13 +848,52 @@ 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)];
+  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)];
+  return self->children[getNodeIndexExists(self, index)];
 }
 // Precondition - an entry for index must exist in the node
 Node *&getChildExists(Node48 *self, uint8_t index) {
@@ -918,12 +938,12 @@ 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: {
@@ -951,12 +971,12 @@ 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: {
@@ -985,13 +1005,13 @@ 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;
   }
@@ -1103,6 +1123,8 @@ 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];
@@ -1582,7 +1604,6 @@ __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
 
@@ -2449,6 +2470,7 @@ 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);
@@ -2543,38 +2565,19 @@ downLeftSpine:
 namespace {
 // Return true if the max version among all keys that start with key[:prefixLen]
 // that are >= key is <= readVersion
-struct CheckRangeLeftSide {
-  CheckRangeLeftSide(Node *n, std::span<const uint8_t> key, int prefixLen,
-                     InternalVersionT readVersion, ReadContext *tls)
-      : n(n), remaining(key), prefixLen(prefixLen), readVersion(readVersion),
-        impl(tls->impl), tls(tls) {
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-    fprintf(stderr, "Check range left side from %s for keys starting with %s\n",
-            printable(key).c_str(),
-            printable(key.subspan(0, prefixLen)).c_str());
-#endif
-  }
-
-  Node *n;
-  std::span<const uint8_t> remaining;
-  int prefixLen;
-  InternalVersionT readVersion;
-  ConflictSet::Impl *impl;
-  ReadContext *tls;
+bool checkRangeLeftSide(Node *n, std::span<const uint8_t> key, int prefixLen,
+                        InternalVersionT readVersion, ReadContext *tls) {
+  auto remaining = key;
   int searchPathLen = 0;
-  bool ok;
-
-  bool step() {
+  for (;; ++tls->range_read_iterations_accum) {
     if (remaining.size() == 0) {
       assert(searchPathLen >= prefixLen);
-      ok = maxVersion(n) <= readVersion;
-      return true;
+      return maxVersion(n) <= readVersion;
     }
 
     if (searchPathLen >= prefixLen) {
       if (!checkMaxBetweenExclusive(n, remaining[0], 256, readVersion, tls)) {
-        ok = false;
-        return true;
+        return false;
       }
     }
 
@@ -2584,18 +2587,16 @@ struct CheckRangeLeftSide {
       if (c != nullptr) {
         if (searchPathLen < prefixLen) {
           n = c;
-          return downLeftSpine();
+          goto downLeftSpine;
         }
         n = c;
-        ok = maxVersion(n) <= readVersion;
-        return true;
+        return maxVersion(n) <= readVersion;
       } else {
         n = nextSibling(n);
         if (n == nullptr) {
-          ok = true;
           return true;
         }
-        return downLeftSpine();
+        goto downLeftSpine;
       }
     }
 
@@ -2611,21 +2612,18 @@ struct CheckRangeLeftSide {
         auto c = n->partialKey()[i] <=> remaining[i];
         if (c > 0) {
           if (searchPathLen < prefixLen) {
-            return downLeftSpine();
+            goto downLeftSpine;
           }
           if (n->entryPresent && n->entry.rangeVersion > readVersion) {
-            ok = false;
-            return true;
+            return false;
           }
-          ok = maxVersion(n) <= readVersion;
-          return true;
+          return maxVersion(n) <= readVersion;
         } else {
           n = nextSibling(n);
           if (n == nullptr) {
-            ok = true;
             return true;
           }
-          return downLeftSpine();
+          goto downLeftSpine;
         }
       }
       if (commonLen == n->partialKeyLen) {
@@ -2634,83 +2632,47 @@ struct CheckRangeLeftSide {
       } else if (n->partialKeyLen > int(remaining.size())) {
         assert(searchPathLen >= prefixLen);
         if (n->entryPresent && n->entry.rangeVersion > readVersion) {
-          ok = false;
-          return true;
+          return false;
         }
-        ok = maxVersion(n) <= readVersion;
-        return true;
+        return maxVersion(n) <= readVersion;
       }
     }
     if (maxV <= readVersion) {
-      ok = true;
       return true;
     }
-    return false;
   }
-
-  bool downLeftSpine() {
-    for (; !n->entryPresent; n = getFirstChildExists(n)) {
-    }
-    ok = n->entry.rangeVersion <= readVersion;
-    return true;
+downLeftSpine:
+  for (; !n->entryPresent; n = getFirstChildExists(n)) {
   }
-};
+  return n->entry.rangeVersion <= readVersion;
+}
 
 // Return true if the max version among all keys that start with key[:prefixLen]
 // that are < key is <= readVersion
-struct CheckRangeRightSide {
-  CheckRangeRightSide(Node *n, std::span<const uint8_t> key, int prefixLen,
-                      InternalVersionT readVersion, ReadContext *tls)
-      : n(n), key(key), remaining(key), prefixLen(prefixLen),
-        readVersion(readVersion), impl(tls->impl), tls(tls) {
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-    fprintf(stderr, "Check range right side to %s for keys starting with %s\n",
-            printable(key).c_str(),
-            printable(key.subspan(0, prefixLen)).c_str());
-#endif
-  }
-
-  Node *n;
-  std::span<const uint8_t> key;
-  std::span<const uint8_t> remaining;
-  int prefixLen;
-  InternalVersionT readVersion;
-  ConflictSet::Impl *impl;
-  ReadContext *tls;
+bool checkRangeRightSide(Node *n, std::span<const uint8_t> key, int prefixLen,
+                         InternalVersionT readVersion, ReadContext *tls) {
+  auto remaining = key;
   int searchPathLen = 0;
-  bool ok;
-
-  bool step() {
-#if DEBUG_VERBOSE && !defined(NDEBUG)
-    fprintf(stderr,
-            "Search path: %s, searchPathLen: %d, prefixLen: %d, remaining: "
-            "%s\n",
-            getSearchPathPrintable(n).c_str(), searchPathLen, prefixLen,
-            printable(remaining).c_str());
-#endif
 
+  for (;; ++tls->range_read_iterations_accum) {
     assert(searchPathLen <= int(key.size()));
-
     if (remaining.size() == 0) {
-      return downLeftSpine();
+      goto downLeftSpine;
     }
 
     if (searchPathLen >= prefixLen) {
       if (n->entryPresent && n->entry.pointVersion > readVersion) {
-        ok = false;
-        return true;
+        return false;
       }
 
       if (!checkMaxBetweenExclusive(n, -1, remaining[0], readVersion, tls)) {
-        ok = false;
-        return true;
+        return false;
       }
     }
 
     if (searchPathLen > prefixLen && n->entryPresent &&
         n->entry.rangeVersion > readVersion) {
-      ok = false;
-      return true;
+      return false;
     }
 
     auto *child = getChild(n, remaining[0]);
@@ -2718,9 +2680,9 @@ struct CheckRangeRightSide {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
         n = c;
-        return downLeftSpine();
+        goto downLeftSpine;
       } else {
-        return backtrack();
+        goto backtrack;
       }
     }
 
@@ -2736,57 +2698,48 @@ struct CheckRangeRightSide {
         ++searchPathLen;
         auto c = n->partialKey()[i] <=> remaining[i];
         if (c > 0) {
-          return downLeftSpine();
+          goto downLeftSpine;
         } else {
           if (searchPathLen > prefixLen && n->entryPresent &&
               n->entry.rangeVersion > readVersion) {
-            ok = false;
-            return true;
+            return false;
           }
-          return backtrack();
+          goto backtrack;
         }
       }
       if (commonLen == n->partialKeyLen) {
         // partial key matches
         remaining = remaining.subspan(commonLen, remaining.size() - commonLen);
       } else if (n->partialKeyLen > int(remaining.size())) {
-        return downLeftSpine();
-      }
-    }
-    return false;
-  }
-
-  bool backtrack() {
-    for (;;) {
-      // searchPathLen > prefixLen implies n is not the root
-      if (searchPathLen > prefixLen && maxVersion(n) > readVersion) {
-        ok = false;
-        return true;
-      }
-      if (n->parent == nullptr) {
-        ok = true;
-        return true;
-      }
-      auto next = getChildGeq(n->parent, n->parentsIndex + 1);
-      if (next == nullptr) {
-        searchPathLen -= 1 + n->partialKeyLen;
-        n = n->parent;
-      } else {
-        searchPathLen -= n->partialKeyLen;
-        n = next;
-        searchPathLen += n->partialKeyLen;
-        return downLeftSpine();
+        goto downLeftSpine;
       }
     }
   }
-
-  bool downLeftSpine() {
-    for (; !n->entryPresent; n = getFirstChildExists(n)) {
+backtrack:
+  for (;;) {
+    // searchPathLen > prefixLen implies n is not the root
+    if (searchPathLen > prefixLen && maxVersion(n) > readVersion) {
+      return false;
+    }
+    if (n->parent == nullptr) {
+      return true;
+    }
+    auto next = getChildGeq(n->parent, n->parentsIndex + 1);
+    if (next == nullptr) {
+      searchPathLen -= 1 + n->partialKeyLen;
+      n = n->parent;
+    } else {
+      searchPathLen -= n->partialKeyLen;
+      n = next;
+      searchPathLen += n->partialKeyLen;
+      goto downLeftSpine;
     }
-    ok = n->entry.rangeVersion <= readVersion;
-    return true;
   }
-};
+downLeftSpine:
+  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  }
+  return n->entry.rangeVersion <= readVersion;
+}
 } // namespace
 
 bool checkRangeRead(Node *n, std::span<const uint8_t> begin,
@@ -2807,8 +2760,8 @@ bool checkRangeRead(Node *n, std::span<const uint8_t> begin,
 
   auto remaining = begin.subspan(0, lcp);
   Arena arena;
-  // If the common prefix isn't a prefix of any physical entry in the tree, we
-  // can go to "downLeftSpine"
+
+  // Advance down common prefix, but stay on a physical path in the tree
   for (;; ++tls->range_read_iterations_accum) {
     assert(getSearchPath(arena, n) <=>
                begin.subspan(0, lcp - remaining.size()) ==
@@ -2849,47 +2802,17 @@ bool checkRangeRead(Node *n, std::span<const uint8_t> begin,
   lcp -= consumed;
 
   if (lcp == int(begin.size())) {
-    CheckRangeRightSide checkRangeRightSide{n, end, lcp, readVersion, tls};
-    while (!checkRangeRightSide.step())
-      ;
-    return checkRangeRightSide.ok;
+    return checkRangeRightSide(n, end, lcp, readVersion, tls);
   }
 
-  if (!checkRangeStartsWith(n, begin.subspan(0, lcp), begin[lcp], end[lcp],
-                            readVersion, tls)) {
-    return false;
-  }
-
-  // This makes it safe to check maxVersion within CheckRangeLeftSide. If this
+  // This makes it safe to check maxVersion within checkRangeLeftSide. If this
   // were false, then we would have returned above since lcp == begin.size().
   assert(!(n->parent == nullptr && begin.size() == 0));
-  CheckRangeLeftSide checkRangeLeftSide{n, begin, lcp + 1, readVersion, tls};
-  CheckRangeRightSide checkRangeRightSide{n, end, lcp + 1, readVersion, tls};
 
-  for (;;) {
-    bool leftDone = checkRangeLeftSide.step();
-    bool rightDone = checkRangeRightSide.step();
-    if (!leftDone && !rightDone) {
-      tls->range_read_iterations_accum += 2;
-      continue;
-    }
-    if (leftDone && rightDone) {
-      break;
-    } else if (leftDone) {
-      while (!checkRangeRightSide.step()) {
-        ++tls->range_read_iterations_accum;
-      }
-      break;
-    } else {
-      assert(rightDone);
-      while (!checkRangeLeftSide.step()) {
-        ++tls->range_read_iterations_accum;
-      }
-    }
-    break;
-  }
-
-  return checkRangeLeftSide.ok && checkRangeRightSide.ok;
+  return checkRangeStartsWith(n, begin.subspan(0, lcp), begin[lcp], end[lcp],
+                              readVersion, tls) &&
+         checkRangeLeftSide(n, begin, lcp + 1, readVersion, tls) &&
+         checkRangeRightSide(n, end, lcp + 1, readVersion, tls);
 }
 
 #ifdef __x86_64__
@@ -2987,6 +2910,72 @@ void addPointWrite(Node *&root, std::span<const uint8_t> key,
   }
 }
 
+#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
+__attribute__((target("avx512f"))) InternalVersionT
+horizontalMaxUpTo16(InternalVersionT *vs, 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, 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;
+  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);
@@ -2998,15 +2987,13 @@ 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);
@@ -3016,9 +3003,7 @@ 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
@@ -3242,7 +3227,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
                           0) *
         2;
 
-    memory_bytes.set(totalBytes);
     point_writes_total.add(tls.accum.point_writes);
     range_writes_total.add(tls.accum.range_writes);
     nodes_allocated_total.add(tls.accum.nodes_allocated);
@@ -3331,7 +3315,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 #endif
     keyUpdates = gcScanStep(keyUpdates);
 
-    memory_bytes.set(totalBytes);
     nodes_allocated_total.add(std::exchange(tls.accum.nodes_allocated, 0));
     nodes_released_total.add(std::exchange(tls.accum.nodes_released, 0));
     entries_inserted_total.add(std::exchange(tls.accum.entries_inserted, 0));
@@ -3379,7 +3362,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   explicit Impl(int64_t oldestVersion) {
     assert(oldestVersion >= 0);
     init(oldestVersion);
-    initMetrics();
+    metrics = initMetrics(metricsList, metricsCount);
   }
   ~Impl() {
     eraseTree(root, &tls);
@@ -3402,23 +3385,12 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   MetricsV1 *metrics;
   int metricsCount = 0;
-  void initMetrics() {
-    metrics = (MetricsV1 *)safe_malloc(metricsCount * sizeof(metrics[0]));
-    for (auto [i, m] = std::make_tuple(metricsCount - 1, metricList); i >= 0;
-         --i, m = m->prev) {
-      metrics[i].name = m->name;
-      metrics[i].help = m->help;
-      metrics[i].p = m;
-      metrics[i].type = m->type;
-    }
-  }
-
-  Metric *metricList = nullptr;
+  Metric *metricsList = nullptr;
 
 #define GAUGE(name, help)                                                      \
-  Gauge name { this, #name, help }
+  Gauge name { metricsList, metricsCount, #name, help }
 #define COUNTER(name, help)                                                    \
-  Counter name { this, #name, help }
+  Counter name { metricsList, metricsCount, #name, help }
   // ==================== METRICS DEFINITIONS ====================
   COUNTER(point_read_total, "Total number of point reads checked");
   COUNTER(point_read_short_circuit_total,
@@ -3484,13 +3456,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 };
 
-Metric::Metric(ConflictSet::Impl *impl, const char *name, const char *help,
-               ConflictSet::MetricsV1::Type type)
-    : prev(std::exchange(impl->metricList, this)), name(name), help(help),
-      type(type), value(0) {
-  ++impl->metricsCount;
-}
-
 Node *&getInTree(Node *n, ConflictSet::Impl *impl) {
   return n->parent == nullptr ? impl->root
                               : getChildExists(n->parent, n->parentsIndex);
@@ -3509,6 +3474,7 @@ void internal_addWrites(ConflictSet::Impl *impl,
   mallocBytesDelta = 0;
   impl->addWrites(writes, count, writeVersion);
   impl->totalBytes += mallocBytesDelta;
+  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
   if (impl->totalBytes != mallocBytes) {
     abort();
@@ -3520,6 +3486,7 @@ void internal_setOldestVersion(ConflictSet::Impl *impl, int64_t oldestVersion) {
   mallocBytesDelta = 0;
   impl->setOldestVersion(oldestVersion);
   impl->totalBytes += mallocBytesDelta;
+  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
   if (impl->totalBytes != mallocBytes) {
     abort();
@@ -4073,6 +4040,90 @@ struct __attribute__((visibility("default"))) PeakPrinter {
 
 #ifdef ENABLE_MAIN
 
+#include "third_party/nanobench.h"
+
+template <int kN> void benchRezero() {
+  static_assert(kN % 16 == 0);
+  ankerl::nanobench::Bench bench;
+  InternalVersionT vs[kN];
+  InternalVersionT zero;
+  bench.run("rezero" + std::to_string(kN), [&]() {
+    bench.doNotOptimizeAway(vs);
+    bench.doNotOptimizeAway(zero);
+    for (int i = 0; i < kN; i += 16) {
+      rezero16(vs + i, zero);
+    }
+  });
+}
+
+template <int kN> void benchScan1() {
+  static_assert(kN % 16 == 0);
+  ankerl::nanobench::Bench bench;
+  InternalVersionT vs[kN];
+  uint8_t is[kN];
+  uint8_t begin;
+  uint8_t end;
+  InternalVersionT v;
+  bench.run("scan" + std::to_string(kN), [&]() {
+    bench.doNotOptimizeAway(vs);
+    bench.doNotOptimizeAway(is);
+    bench.doNotOptimizeAway(begin);
+    bench.doNotOptimizeAway(end);
+    bench.doNotOptimizeAway(v);
+    for (int i = 0; i < kN; i += 16) {
+      scan16</*kAVX512=*/true>(vs + i, is + i, begin, end, v);
+    }
+  });
+}
+
+template <int kN> void benchScan2() {
+  static_assert(kN % 16 == 0);
+  ankerl::nanobench::Bench bench;
+  InternalVersionT vs[kN];
+  uint8_t is[kN];
+  uint8_t begin;
+  uint8_t end;
+  InternalVersionT v;
+  bench.run("scan" + std::to_string(kN), [&]() {
+    bench.doNotOptimizeAway(vs);
+    bench.doNotOptimizeAway(begin);
+    bench.doNotOptimizeAway(end);
+    bench.doNotOptimizeAway(v);
+    for (int i = 0; i < kN; i += 16) {
+      scan16</*kAVX512=*/true>(vs + i, begin, end, v);
+    }
+  });
+}
+
+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);
+  std::vector<uint8_t> rhs(len);
+  bench.run("lcp " + std::to_string(len), [&]() {
+    bench.doNotOptimizeAway(lhs);
+    bench.doNotOptimizeAway(rhs);
+    bench.doNotOptimizeAway(longestCommonPrefix(lhs.data(), rhs.data(), len));
+  });
+}
+
 void printTree() {
   int64_t writeVersion = 0;
   ConflictSet::Impl cs{writeVersion};
@@ -4094,7 +4145,7 @@ void printTree() {
   debugPrintDot(stdout, cs.root, &cs);
 }
 
-int main(void) { printTree(); }
+int main(void) { benchHorizontal16(); }
 #endif
 
 #ifdef ENABLE_FUZZ

InterleavingTest.cpp

@@ -0,0 +1,178 @@
+#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;
+}
+
+// So we can look at the disassembly more easily
+
+extern "C" {
+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];
+  }
+}
+
+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 = 100;
+  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(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;
+  }
+}

Jenkinsfile

@@ -129,7 +129,7 @@ pipeline {
             }
             steps {
                 script {
-                    filter_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h"
+                    filter_args = "-f ConflictSet.cpp -f LongestCommonPrefix.h -f Metrics.h"
                 }
                 CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_C_FLAGS=--coverage -DCMAKE_CXX_FLAGS=--coverage -DCMAKE_BUILD_TYPE=Debug -DDISABLE_TSAN=ON")
                 sh """

Metrics.h

@@ -0,0 +1,64 @@
+#pragma once
+
+#include "ConflictSet.h"
+#include "Internal.h"
+
+#include <assert.h>
+#include <atomic>
+#include <tuple>
+
+struct Metric {
+  Metric *prev;
+  const char *name;
+  const char *help;
+  weaselab::ConflictSet::MetricsV1::Type type;
+  std::atomic<int64_t> value;
+
+protected:
+  Metric(Metric *&metricList, int &metricsCount, const char *name,
+         const char *help, weaselab::ConflictSet::MetricsV1::Type type)
+      : prev(std::exchange(metricList, this)), name(name), help(help),
+        type(type), value(0) {
+    ++metricsCount;
+  }
+};
+
+struct Gauge : private Metric {
+  Gauge(Metric *&metricList, int &metricsCount, const char *name,
+        const char *help)
+      : Metric(metricList, metricsCount, name, help,
+               weaselab::ConflictSet::MetricsV1::Gauge) {}
+
+  void set(int64_t value) {
+    this->value.store(value, std::memory_order_relaxed);
+  }
+};
+
+struct Counter : private Metric {
+  Counter(Metric *&metricList, int &metricsCount, const char *name,
+          const char *help)
+      : Metric(metricList, metricsCount, name, help,
+               weaselab::ConflictSet::MetricsV1::Counter) {}
+  // Expensive. Accumulate locally and then call add instead of repeatedly
+  // calling add.
+  void add(int64_t value) {
+    assert(value >= 0);
+    static_assert(std::atomic<int64_t>::is_always_lock_free);
+    this->value.fetch_add(value, std::memory_order_relaxed);
+  }
+};
+
+inline weaselab::ConflictSet::MetricsV1 *initMetrics(Metric *metricsList,
+                                                     int metricsCount) {
+  weaselab::ConflictSet::MetricsV1 *metrics =
+      (weaselab::ConflictSet::MetricsV1 *)safe_malloc(metricsCount *
+                                                      sizeof(metrics[0]));
+  for (auto [i, m] = std::make_tuple(metricsCount - 1, metricsList); i >= 0;
+       --i, m = m->prev) {
+    metrics[i].name = m->name;
+    metrics[i].help = m->help;
+    metrics[i].p = m;
+    metrics[i].type = m->type;
+  }
+  return metrics;
+}

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
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               11.18 |       89,455,125.34 |    0.6% |          185.37 |           57.08 |  3.248 |          41.51 |    0.4% |      0.01 | `point reads`
-|               14.53 |       68,800,688.89 |    0.4% |          282.41 |           74.80 |  3.776 |          55.06 |    0.3% |      0.01 | `prefix reads`
-|               36.54 |       27,367,576.87 |    0.2% |          798.06 |          188.90 |  4.225 |         141.69 |    0.2% |      0.01 | `range reads`
-|               16.69 |       59,912,106.02 |    0.6% |          314.57 |           86.29 |  3.645 |          39.84 |    0.4% |      0.01 | `point writes`
-|               30.09 |       33,235,744.07 |    0.5% |          591.33 |          155.92 |  3.793 |          82.69 |    0.2% |      0.01 | `prefix writes`
-|               35.77 |       27,956,388.03 |    1.4% |          682.25 |          187.63 |  3.636 |          96.12 |    0.1% |      0.01 | `range 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`
-|          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`
-|               92.46 |       10,814,961.65 |    0.5% |        1,800.00 |          463.41 |  3.884 |         297.00 |    0.0% |      0.01 | `create and destroy`
+|               12.88 |       77,653,350.77 |    0.5% |          185.37 |           64.45 |  2.876 |          41.51 |    0.4% |      0.01 | `point reads`
+|               14.67 |       68,179,354.49 |    0.1% |          271.44 |           73.40 |  3.698 |          53.70 |    0.3% |      0.01 | `prefix reads`
+|               34.84 |       28,701,444.36 |    0.3% |          715.74 |          175.27 |  4.084 |         127.30 |    0.2% |      0.01 | `range reads`
+|               17.12 |       58,422,988.28 |    0.2% |          314.30 |           86.11 |  3.650 |          39.82 |    0.4% |      0.01 | `point writes`
+|               31.42 |       31,830,804.65 |    0.1% |          591.06 |          158.07 |  3.739 |          82.67 |    0.2% |      0.01 | `prefix writes`
+|               37.37 |       26,759,432.70 |    2.2% |          681.98 |          188.95 |  3.609 |          96.10 |    0.1% |      0.01 | `range writes`
+|               76.72 |       13,035,140.63 |    2.3% |        1,421.28 |          387.17 |  3.671 |         257.76 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          297,452.00 |            3,361.89 |    0.9% |    3,508,083.00 |    1,500,834.67 |  2.337 |     727,525.33 |    0.1% |      0.01 | `worst case for radix tree`
+|               87.70 |       11,402,490.60 |    1.0% |        1,795.00 |          442.09 |  4.060 |         297.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 

ServerBench.cpp

@@ -21,7 +21,7 @@
 
 std::atomic<int64_t> transactions;
 
-constexpr int kBaseSearchDepth = 32;
+constexpr int kBaseSearchDepth = 115;
 constexpr int kWindowSize = 10000000;
 
 std::string numToKey(int64_t num) {

SkipList.cpp

@@ -22,6 +22,7 @@
 
 #include "ConflictSet.h"
 #include "Internal.h"
+#include "Metrics.h"
 
 #include <algorithm>
 #include <span>
@@ -434,13 +435,14 @@ public:
     return result;
   }
 
-  void detectConflicts(ReadConflictRange *ranges, int count,
-                       ConflictSet::Result *transactionConflictStatus) const {
+  // Return number of iterations of main loop
+  int detectConflicts(ReadConflictRange *ranges, int count,
+                      ConflictSet::Result *transactionConflictStatus) const {
     const int M = 16;
     int nextJob[M];
     CheckMax inProgress[M];
     if (!count)
-      return;
+      return 0;
 
     int started = std::min(M, count);
     for (int i = 0; i < started; i++) {
@@ -451,8 +453,9 @@ public:
 
     int prevJob = started - 1;
     int job = 0;
+    int iters = 0;
     // vtune: 340 parts
-    while (true) {
+    for (;; ++iters) {
       if (inProgress[job].advance()) {
         if (started == count) {
           if (prevJob == job)
@@ -468,6 +471,7 @@ public:
       prevJob = job;
       job = nextJob[job];
     }
+    return iters;
   }
 
   void find(const StringRef *values, Finger *results, int *temp, int count) {
@@ -702,15 +706,27 @@ private:
   };
 };
 
+struct ReadContext {
+  int64_t commits_accum = 0;
+  int64_t conflicts_accum = 0;
+  int64_t too_olds_accum = 0;
+  int64_t check_bytes_accum = 0;
+};
+
 struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   Impl(int64_t oldestVersion)
       : oldestVersion(oldestVersion), newestVersion(oldestVersion),
-        skipList(oldestVersion) {}
+        skipList(oldestVersion) {
+    metrics = initMetrics(metricsList, metricsCount);
+  }
+  ~Impl() { safe_free(metrics, metricsCount * sizeof(metrics[0])); }
   void check(const ConflictSet::ReadRange *reads, ConflictSet::Result *results,
-             int count) const {
+             int count) {
+    ReadContext tls;
     Arena arena;
     auto *ranges = new (arena) ReadConflictRange[count];
     for (int i = 0; i < count; ++i) {
+      tls.check_bytes_accum += reads[i].begin.len + reads[i].end.len;
       ranges[i].begin = {reads[i].begin.p, size_t(reads[i].begin.len)};
       ranges[i].end = reads[i].end.len > 0
                           ? StringRef{reads[i].end.p, size_t(reads[i].end.len)}
@@ -718,13 +734,22 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       ranges[i].version = reads[i].readVersion;
       results[i] = ConflictSet::Commit;
     }
-    skipList.detectConflicts(ranges, count, results);
+    int iters = skipList.detectConflicts(ranges, count, results);
     for (int i = 0; i < count; ++i) {
       if (reads[i].readVersion < oldestVersion ||
           reads[i].readVersion < newestVersion - 2e9) {
         results[i] = TooOld;
       }
+      tls.commits_accum += results[i] == Commit;
+      tls.conflicts_accum += results[i] == Conflict;
+      tls.too_olds_accum += results[i] == TooOld;
     }
+    range_read_iterations_total.add(iters);
+    range_read_total.add(count);
+    commits_total.add(tls.commits_accum);
+    conflicts_total.add(tls.conflicts_accum);
+    too_olds_total.add(tls.too_olds_accum);
+    check_bytes_total.add(tls.check_bytes_accum);
   }
 
   void addWrites(const ConflictSet::WriteRange *writes, int count,
@@ -788,6 +813,9 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 
   void setOldestVersion(int64_t oldestVersion) {
+    // This isn't 100% accurate. It overcounts if you hit the end
+    gc_iterations_total.add(keyUpdates);
+
     assert(oldestVersion >= this->oldestVersion);
     this->oldestVersion = oldestVersion;
     SkipList::Finger finger;
@@ -802,6 +830,54 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
 
   int64_t totalBytes = 0;
 
+  MetricsV1 *metrics;
+  int metricsCount = 0;
+  Metric *metricsList = nullptr;
+
+#define GAUGE(name, help)                                                      \
+  Gauge name { metricsList, metricsCount, #name, help }
+#define COUNTER(name, help)                                                    \
+  Counter name { metricsList, metricsCount, #name, help }
+  // ==================== METRICS DEFINITIONS ====================
+  COUNTER(range_read_total, "Total number of range reads checked");
+  COUNTER(range_read_iterations_total,
+          "Total number of iterations of the main loops for range read checks");
+  COUNTER(commits_total,
+          "Total number of checks where the result is \"commit\"");
+  COUNTER(conflicts_total,
+          "Total number of checks where the result is \"conflict\"");
+  COUNTER(too_olds_total,
+          "Total number of checks where the result is \"too old\"");
+  COUNTER(check_bytes_total, "Total number of key bytes checked");
+  GAUGE(memory_bytes, "Total number of bytes in use");
+  COUNTER(nodes_allocated_total,
+          "The total number of physical tree nodes allocated");
+  COUNTER(nodes_released_total,
+          "The total number of physical tree nodes released");
+  COUNTER(insert_iterations_total,
+          "The total number of iterations of the main loop for insertion. "
+          "Includes searches where the entry already existed, and so insertion "
+          "did not take place");
+  COUNTER(entries_inserted_total,
+          "The total number of entries inserted in the tree");
+  COUNTER(entries_erased_total,
+          "The total number of entries erased from the tree");
+  COUNTER(
+      gc_iterations_total,
+      "The total number of iterations of the main loop for garbage collection");
+  COUNTER(write_bytes_total, "Total number of key bytes in calls to addWrites");
+  GAUGE(oldest_version,
+        "The lowest version that doesn't result in \"TooOld\" for checks");
+  GAUGE(newest_version, "The version of the most recent call to addWrites");
+  // ==================== END METRICS DEFINITIONS ====================
+#undef GAUGE
+#undef COUNTER
+
+  void getMetricsV1(MetricsV1 **metrics, int *count) {
+    *metrics = this->metrics;
+    *count = metricsCount;
+  }
+
 private:
   int64_t keyUpdates = 0;
   Arena removalArena;
@@ -824,6 +900,7 @@ void internal_addWrites(ConflictSet::Impl *impl,
   mallocBytesDelta = 0;
   impl->addWrites(writes, count, writeVersion);
   impl->totalBytes += mallocBytesDelta;
+  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
   if (impl->totalBytes != mallocBytes) {
     abort();
@@ -835,6 +912,7 @@ void internal_setOldestVersion(ConflictSet::Impl *impl, int64_t oldestVersion) {
   mallocBytesDelta = 0;
   impl->setOldestVersion(oldestVersion);
   impl->totalBytes += mallocBytesDelta;
+  impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
   if (impl->totalBytes != mallocBytes) {
     abort();
@@ -858,12 +936,11 @@ int64_t internal_getBytes(ConflictSet::Impl *impl) { return impl->totalBytes; }
 
 void internal_getMetricsV1(ConflictSet::Impl *impl,
                            ConflictSet::MetricsV1 **metrics, int *count) {
-  *metrics = nullptr;
-  *count = 0;
+  return impl->getMetricsV1(metrics, count);
 }
 
 double internal_getMetricValue(const ConflictSet::MetricsV1 *metric) {
-  return 0;
+  return ((Metric *)metric->p)->value.load(std::memory_order_relaxed);
 }
 
 void ConflictSet::check(const ReadRange *reads, Result *results,