Update README benchmarks

150% pessimized the "real data" benchmark

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.18)
 project(
   conflict-set
-  VERSION 0.0.11
+  VERSION 0.0.12
   DESCRIPTION
     "A data structure for optimistic concurrency control on ranges of bitwise-lexicographically-ordered keys."
   HOMEPAGE_URL "https://git.weaselab.dev/weaselab/conflict-set"

ConflictSet.cpp

@@ -16,12 +16,12 @@ limitations under the License.
 
 #include "ConflictSet.h"
 #include "Internal.h"
+#include "LongestCommonPrefix.h"
 
 #include <algorithm>
 #include <atomic>
 #include <bit>
 #include <cassert>
-#include <compare>
 #include <cstddef>
 #include <cstdint>
 #include <cstring>
@@ -40,11 +40,13 @@ limitations under the License.
 #include <arm_neon.h>
 #endif
 
+#ifndef __SANITIZE_THREAD__
 #if defined(__has_feature)
 #if __has_feature(thread_sanitizer)
 #define __SANITIZE_THREAD__
 #endif
 #endif
+#endif
 
 #include <memcheck.h>
 
@@ -1123,12 +1125,68 @@ Node *getFirstChildExists(Node *self) {
   }
 }
 
-// Caller is responsible for assigning a non-null pointer to the returned
-// reference if null. Updates child's max version to `newMaxVersion` if child
-// exists but does not have a partial key.
-Node *&getOrCreateChild(Node *&self, uint8_t index,
+void consumePartialKeyFull(Node *&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;
+    // Since root cannot have a partial key
+    assert(old->parent != nullptr);
+    InternalVersionT oldMaxVersion = exchangeMaxVersion(old, writeVersion);
+
+    // *self will have one child (old)
+    auto *newSelf = tls->allocate<Node3>(partialKeyIndex);
+
+    newSelf->parent = old->parent;
+    newSelf->parentsIndex = old->parentsIndex;
+    newSelf->partialKeyLen = partialKeyIndex;
+    newSelf->entryPresent = false;
+    newSelf->numChildren = 1;
+
+    memcpy(newSelf->partialKey(), old->partialKey(), newSelf->partialKeyLen);
+
+    uint8_t oldDistinguishingByte = old->partialKey()[partialKeyIndex];
+    old->parent = newSelf;
+    old->parentsIndex = oldDistinguishingByte;
+    newSelf->index[0] = oldDistinguishingByte;
+    newSelf->children[0] = old;
+    newSelf->childMaxVersion[0] = oldMaxVersion;
+    self = newSelf;
+
+    memmove(old->partialKey(), old->partialKey() + partialKeyIndex + 1,
+            old->partialKeyLen - (partialKeyIndex + 1));
+    old->partialKeyLen -= partialKeyIndex + 1;
+
+    // We would consider decreasing capacity here, but we can't invalidate
+    // old since it's not on the search path. setOldestVersion will clean it
+    // up.
+  }
+  key = key.subspan(partialKeyIndex, key.size() - partialKeyIndex);
+}
+
+// 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,
+                  InternalVersionT writeVersion, WriteContext *tls) {
+  if (self->partialKeyLen > 0) {
+    consumePartialKeyFull(self, key, writeVersion, tls);
+  }
+}
+
+// Return the next node along the search path of key, consuming bytes of 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) {
 
+  int index = key.front();
+  key = key.subspan(1, key.size() - 1);
+
   // Fast path for if it exists already
   switch (self->getType()) {
   case Type_Node0:
@@ -1137,9 +1195,8 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     auto *self3 = static_cast<Node3 *>(self);
     int i = getNodeIndex(self3, index);
     if (i >= 0) {
-      if (self3->children[i]->partialKeyLen == 0) {
-        self3->childMaxVersion[i] = newMaxVersion;
-      }
+      consumePartialKey(self3->children[i], key, newMaxVersion, tls);
+      self3->childMaxVersion[i] = newMaxVersion;
       return self3->children[i];
     }
   } break;
@@ -1147,9 +1204,8 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     auto *self16 = static_cast<Node16 *>(self);
     int i = getNodeIndex(self16, index);
     if (i >= 0) {
-      if (self16->children[i]->partialKeyLen == 0) {
-        self16->childMaxVersion[i] = newMaxVersion;
-      }
+      consumePartialKey(self16->children[i], key, newMaxVersion, tls);
+      self16->childMaxVersion[i] = newMaxVersion;
       return self16->children[i];
     }
   } break;
@@ -1157,23 +1213,21 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     auto *self48 = static_cast<Node48 *>(self);
     int secondIndex = self48->index[index];
     if (secondIndex >= 0) {
-      if (self48->children[secondIndex]->partialKeyLen == 0) {
-        self48->childMaxVersion[secondIndex] = newMaxVersion;
-        self48->maxOfMax[secondIndex >> Node48::kMaxOfMaxShift] =
-            std::max(self48->maxOfMax[secondIndex >> Node48::kMaxOfMaxShift],
-                     newMaxVersion);
-      }
+      consumePartialKey(self48->children[secondIndex], key, newMaxVersion, tls);
+      self48->childMaxVersion[secondIndex] = newMaxVersion;
+      self48->maxOfMax[secondIndex >> Node48::kMaxOfMaxShift] =
+          std::max(self48->maxOfMax[secondIndex >> Node48::kMaxOfMaxShift],
+                   newMaxVersion);
       return self48->children[secondIndex];
     }
   } break;
   case Type_Node256: {
     auto *self256 = static_cast<Node256 *>(self);
     if (auto &result = self256->children[index]; result != nullptr) {
-      if (self256->children[index]->partialKeyLen == 0) {
-        self256->childMaxVersion[index] = newMaxVersion;
-        self256->maxOfMax[index >> Node256::kMaxOfMaxShift] = std::max(
-            self256->maxOfMax[index >> Node256::kMaxOfMaxShift], newMaxVersion);
-      }
+      consumePartialKey(result, key, newMaxVersion, tls);
+      self256->childMaxVersion[index] = newMaxVersion;
+      self256->maxOfMax[index >> Node256::kMaxOfMaxShift] = std::max(
+          self256->maxOfMax[index >> Node256::kMaxOfMaxShift], newMaxVersion);
       return result;
     }
   } break;
@@ -1181,6 +1235,14 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
 
+  auto *newChild = tls->allocate<Node0>(key.size());
+  newChild->numChildren = 0;
+  newChild->entryPresent = false;
+  newChild->partialKeyLen = key.size();
+  newChild->parentsIndex = index;
+  memcpy(newChild->partialKey(), key.data(), key.size());
+  key = {};
+
   switch (self->getType()) {
   case Type_Node0: {
     auto *self0 = static_cast<Node0 *>(self);
@@ -1215,8 +1277,10 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     }
     self3->index[i + 1] = index;
     auto &result = self3->children[i + 1];
-    result = nullptr;
+    self3->childMaxVersion[i + 1] = newMaxVersion;
+    result = newChild;
     ++self->numChildren;
+    newChild->parent = self;
     return result;
   }
   case Type_Node16: {
@@ -1243,8 +1307,10 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     }
     self16->index[i + 1] = index;
     auto &result = self16->children[i + 1];
-    result = nullptr;
+    self16->childMaxVersion[i + 1] = newMaxVersion;
+    result = newChild;
     ++self->numChildren;
+    newChild->parent = self;
     return result;
   }
   case Type_Node48: {
@@ -1267,7 +1333,11 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     self48->index[index] = nextFree;
     self48->reverseIndex[nextFree] = index;
     auto &result = self48->children[nextFree];
-    result = nullptr;
+    self48->childMaxVersion[nextFree] = newMaxVersion;
+    self48->maxOfMax[nextFree >> Node48::kMaxOfMaxShift] = std::max(
+        newMaxVersion, self48->maxOfMax[nextFree >> Node48::kMaxOfMaxShift]);
+    result = newChild;
+    newChild->parent = self;
     return result;
   }
   case Type_Node256: {
@@ -1276,7 +1346,13 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
     auto *self256 = static_cast<Node256 *>(self);
     ++self->numChildren;
     self256->bitSet.set(index);
-    return self256->children[index];
+    auto &result = self256->children[index];
+    self256->childMaxVersion[index] = newMaxVersion;
+    self256->maxOfMax[index >> Node256::kMaxOfMaxShift] = std::max(
+        newMaxVersion, self256->maxOfMax[index >> Node256::kMaxOfMaxShift]);
+    result = newChild;
+    newChild->parent = self;
+    return result;
   }
   default:                   // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -1687,167 +1763,6 @@ Node *nextSibling(Node *node) {
   }
 }
 
-#if defined(HAS_AVX) || defined(HAS_ARM_NEON)
-constexpr int kStride = 64;
-#else
-constexpr int kStride = 16;
-#endif
-
-constexpr int kUnrollFactor = 4;
-
-bool compareStride(const uint8_t *ap, const uint8_t *bp) {
-#if defined(HAS_ARM_NEON)
-  static_assert(kStride == 64);
-  uint8x16_t x[4]; // GCOVR_EXCL_LINE
-  for (int i = 0; i < 4; ++i) {
-    x[i] = vceqq_u8(vld1q_u8(ap + i * 16), vld1q_u8(bp + i * 16));
-  }
-  auto results = vreinterpretq_u16_u8(
-      vandq_u8(vandq_u8(x[0], x[1]), vandq_u8(x[2], x[3])));
-  bool eq = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0) ==
-            uint64_t(-1);
-#elif defined(HAS_AVX)
-  static_assert(kStride == 64);
-  __m128i x[4]; // GCOVR_EXCL_LINE
-  for (int i = 0; i < 4; ++i) {
-    x[i] = _mm_cmpeq_epi8(_mm_loadu_si128((__m128i *)(ap + i * 16)),
-                          _mm_loadu_si128((__m128i *)(bp + i * 16)));
-  }
-  auto eq =
-      _mm_movemask_epi8(_mm_and_si128(_mm_and_si128(x[0], x[1]),
-                                      _mm_and_si128(x[2], x[3]))) == 0xffff;
-#else
-  // Hope it gets vectorized
-  auto eq = memcmp(ap, bp, kStride) == 0;
-#endif
-  return eq;
-}
-
-// Precondition: ap[:kStride] != bp[:kStride]
-int firstNeqStride(const uint8_t *ap, const uint8_t *bp) {
-#if defined(HAS_AVX)
-  static_assert(kStride == 64);
-  uint64_t c[kStride / 16]; // GCOVR_EXCL_LINE
-  for (int i = 0; i < kStride; i += 16) {
-    const auto a = _mm_loadu_si128((__m128i *)(ap + i));
-    const auto b = _mm_loadu_si128((__m128i *)(bp + i));
-    const auto compared = _mm_cmpeq_epi8(a, b);
-    c[i / 16] = _mm_movemask_epi8(compared) & 0xffff;
-  }
-  return std::countr_zero(~(c[0] | c[1] << 16 | c[2] << 32 | c[3] << 48));
-#elif defined(HAS_ARM_NEON)
-  static_assert(kStride == 64);
-  for (int i = 0; i < kStride; i += 16) {
-    // 0xff for each match
-    uint16x8_t results =
-        vreinterpretq_u16_u8(vceqq_u8(vld1q_u8(ap + i), vld1q_u8(bp + i)));
-    // 0xf for each mismatch
-    uint64_t bitfield =
-        ~vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0);
-    if (bitfield) {
-      return i + (std::countr_zero(bitfield) >> 2);
-    }
-  }
-  __builtin_unreachable(); // GCOVR_EXCL_LINE
-#else
-  int i = 0;
-  for (; i < kStride - 1; ++i) {
-    if (*ap++ != *bp++) {
-      break;
-    }
-  }
-  return i;
-#endif
-}
-
-// This gets covered in local development
-// GCOVR_EXCL_START
-#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-__attribute__((target("avx512f,avx512bw"))) int
-longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
-  int i = 0;
-  int end = cl & ~63;
-  while (i < end) {
-    const uint64_t eq =
-        _mm512_cmpeq_epi8_mask(_mm512_loadu_epi8(ap), _mm512_loadu_epi8(bp));
-    if (eq != uint64_t(-1)) {
-      return i + std::countr_one(eq);
-    }
-    i += 64;
-    ap += 64;
-    bp += 64;
-  }
-  if (i < cl) {
-    const uint64_t mask = (uint64_t(1) << (cl - i)) - 1;
-    const uint64_t eq = _mm512_cmpeq_epi8_mask(
-        _mm512_maskz_loadu_epi8(mask, ap), _mm512_maskz_loadu_epi8(mask, bp));
-    return i + std::countr_one(eq & mask);
-  }
-  assert(i == cl);
-  return i;
-}
-__attribute__((target("default")))
-#endif
-// GCOVR_EXCL_STOP
-
-int longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
-  assume(cl >= 0);
-  int i = 0;
-  int end;
-
-  // kStride * kUnrollCount at a time
-  end = cl & ~(kStride * kUnrollFactor - 1);
-  while (i < end) {
-    for (int j = 0; j < kUnrollFactor; ++j) {
-      if (!compareStride(ap, bp)) {
-        return i + firstNeqStride(ap, bp);
-      }
-      i += kStride;
-      ap += kStride;
-      bp += kStride;
-    }
-  }
-
-  // kStride at a time
-  end = cl & ~(kStride - 1);
-  while (i < end) {
-    if (!compareStride(ap, bp)) {
-      return i + firstNeqStride(ap, bp);
-    }
-    i += kStride;
-    ap += kStride;
-    bp += kStride;
-  }
-
-  // word at a time
-  end = cl & ~(sizeof(uint64_t) - 1);
-  while (i < end) {
-    uint64_t a; // GCOVR_EXCL_LINE
-    uint64_t b; // GCOVR_EXCL_LINE
-    memcpy(&a, ap, 8);
-    memcpy(&b, bp, 8);
-    const auto mismatched = a ^ b;
-    if (mismatched) {
-      return i + std::countr_zero(mismatched) / 8;
-    }
-    i += 8;
-    ap += 8;
-    bp += 8;
-  }
-
-  // byte at a time
-  while (i < cl) {
-    if (*ap != *bp) {
-      break;
-    }
-    ++ap;
-    ++bp;
-    ++i;
-  }
-
-  return i;
-}
-
 // Logically this is the same as performing firstGeq and then checking against
 // point or range version according to cmp, but this version short circuits as
 // soon as it can prove that there's no conflict.
@@ -2877,94 +2792,22 @@ checkMaxBetweenExclusiveImpl<true>(Node *n, int begin, int end,
                                    InternalVersionT readVersion, ReadContext *);
 #endif
 
-// Consume the partial key of `self` (which must exist), and update `self` and
-// `key` such that `self` is along the search path of `key`
-void consumePartialKey(Node *&self, std::span<const uint8_t> &key,
-                       InternalVersionT writeVersion, WriteContext *tls) {
-  assert(self->partialKeyLen > 0);
-  // 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;
-    // Since root cannot have a partial key
-    assert(old->parent != nullptr);
-    InternalVersionT oldMaxVersion = exchangeMaxVersion(old, writeVersion);
-
-    // *self will have one child (old)
-    auto *newSelf = tls->allocate<Node3>(partialKeyIndex);
-
-    newSelf->parent = old->parent;
-    newSelf->parentsIndex = old->parentsIndex;
-    newSelf->partialKeyLen = partialKeyIndex;
-    newSelf->entryPresent = false;
-    newSelf->numChildren = 1;
-
-    memcpy(newSelf->partialKey(), old->partialKey(), newSelf->partialKeyLen);
-
-    uint8_t oldDistinguishingByte = old->partialKey()[partialKeyIndex];
-    old->parent = newSelf;
-    old->parentsIndex = oldDistinguishingByte;
-    newSelf->index[0] = oldDistinguishingByte;
-    newSelf->children[0] = old;
-    newSelf->childMaxVersion[0] = oldMaxVersion;
-    self = newSelf;
-
-    memmove(old->partialKey(), old->partialKey() + partialKeyIndex + 1,
-            old->partialKeyLen - (partialKeyIndex + 1));
-    old->partialKeyLen -= partialKeyIndex + 1;
-
-    // We would consider decreasing capacity here, but we can't invalidate
-    // old since it's not on the search path. setOldestVersion will clean it
-    // up.
-  }
-  key = key.subspan(partialKeyIndex, key.size() - partialKeyIndex);
-}
-
-// Returns a pointer to the newly inserted node.  Caller must set
-// `entryPresent`, and `entry` fields.  All nodes along the search path of the
-// result will have `maxVersion` set to `writeVersion` as a postcondition. Nodes
-// along the search path may be invalidated.
+// Returns a pointer the pointer to the newly inserted node in the tree.  Caller
+// must set `entryPresent`, and `entry` fields.  All nodes along the search path
+// of the result will have `maxVersion` set to `writeVersion` as a
+// postcondition. Nodes along the search path may be invalidated.
 [[nodiscard]]
-Node *insert(Node **self, std::span<const uint8_t> key,
-             InternalVersionT writeVersion, WriteContext *tls,
-             ConflictSet::Impl *impl) {
+Node **insert(Node **self, std::span<const uint8_t> key,
+              InternalVersionT writeVersion, WriteContext *tls,
+              ConflictSet::Impl *impl) {
 
-  if ((*self)->partialKeyLen > 0) {
-    consumePartialKey(*self, key, writeVersion, tls);
-  }
   assert(maxVersion(*self, impl) <= writeVersion);
   setMaxVersion(*self, impl, writeVersion);
 
-  for (;; ++tls->accum.insert_iterations) {
-
-    if (key.size() == 0) {
-      return *self;
-    }
-
-    auto &child = getOrCreateChild(*self, key.front(), writeVersion, tls);
-    if (!child) {
-      child = tls->allocate<Node0>(key.size() - 1);
-      child->numChildren = 0;
-      child->entryPresent = false;
-      child->partialKeyLen = key.size() - 1;
-      child->parent = *self;
-      child->parentsIndex = key.front();
-      setMaxVersion(child, impl, writeVersion);
-      memcpy(child->partialKey(), key.data() + 1, child->partialKeyLen);
-      return child;
-    }
-
-    self = &child;
-    key = key.subspan(1, key.size() - 1);
-
-    if ((*self)->partialKeyLen > 0) {
-      consumePartialKey(*self, key, writeVersion, tls);
-      assert(maxVersion(*self, impl) <= writeVersion);
-      setMaxVersion(*self, impl, writeVersion);
-    }
+  for (; key.size() != 0; ++tls->accum.insert_iterations) {
+    self = &getOrCreateChild(*self, key, writeVersion, tls);
   }
+  return self;
 }
 
 void destroyTree(Node *root) {
@@ -2996,7 +2839,7 @@ void addPointWrite(Node *&root, std::span<const uint8_t> key,
                    InternalVersionT writeVersion, WriteContext *tls,
                    ConflictSet::Impl *impl) {
   ++tls->accum.point_writes;
-  auto *n = insert(&root, key, writeVersion, tls, impl);
+  auto *n = *insert(&root, key, writeVersion, tls, impl);
   if (!n->entryPresent) {
     ++tls->accum.entries_inserted;
     auto *p = nextLogical(n);
@@ -3063,41 +2906,16 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   }
   ++tls->accum.range_writes;
   const bool beginIsPrefix = lcp == int(begin.size());
-  auto remaining = begin.subspan(0, lcp);
 
-  auto *n = root;
+  Node **useAsRoot =
+      insert(&root, begin.subspan(0, lcp), writeVersion, tls, impl);
 
-  for (;;) {
-    if (int(remaining.size()) <= n->partialKeyLen) {
-      break;
-    }
-    int i = longestCommonPrefix(n->partialKey(), remaining.data(),
-                                n->partialKeyLen);
-    if (i != n->partialKeyLen) {
-      break;
-    }
-
-    auto *child = getChild(n, remaining[n->partialKeyLen]);
-    if (child == nullptr) {
-      break;
-    }
-
-    assert(maxVersion(n, impl) <= writeVersion);
-    setMaxVersion(n, impl, writeVersion);
-
-    remaining = remaining.subspan(n->partialKeyLen + 1,
-                                  remaining.size() - (n->partialKeyLen + 1));
-    n = child;
-  }
-
-  Node **useAsRoot = &getInTree(n, impl);
-
-  int consumed = lcp - remaining.size();
+  int consumed = lcp;
 
   begin = begin.subspan(consumed, begin.size() - consumed);
   end = end.subspan(consumed, end.size() - consumed);
 
-  auto *beginNode = insert(useAsRoot, begin, writeVersion, tls, impl);
+  auto *beginNode = *insert(useAsRoot, begin, writeVersion, tls, impl);
 
   const bool insertedBegin = !beginNode->entryPresent;
 
@@ -3114,7 +2932,7 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   assert(writeVersion >= beginNode->entry.pointVersion);
   beginNode->entry.pointVersion = writeVersion;
 
-  auto *endNode = insert(useAsRoot, end, writeVersion, tls, impl);
+  auto *endNode = *insert(useAsRoot, end, writeVersion, tls, impl);
 
   const bool insertedEnd = !endNode->entryPresent;
 
@@ -3132,7 +2950,7 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   if (beginIsPrefix && insertedEnd) {
     // beginNode may have been invalidated when inserting end. TODO can we do
     // better?
-    beginNode = insert(useAsRoot, begin, writeVersion, tls, impl);
+    beginNode = *insert(useAsRoot, begin, writeVersion, tls, impl);
     assert(beginNode->entryPresent);
   }
 
@@ -3246,6 +3064,8 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     InternalVersionT::zero = tls.zero = oldestVersion;
 
     assert(writeVersion >= newestVersionFullPrecision);
+    assert(tls.accum.entries_erased == 0);
+    assert(tls.accum.entries_inserted == 0);
 
     if (oldestExtantVersion < writeVersion - kMaxCorrectVersionWindow)
         [[unlikely]] {
@@ -3273,15 +3093,19 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
       auto begin = std::span<const uint8_t>(w.begin.p, w.begin.len);
       auto end = std::span<const uint8_t>(w.end.p, w.end.len);
       if (w.end.len > 0) {
-        keyUpdates += 3;
         addWriteRange(root, begin, end, InternalVersionT(writeVersion), &tls,
                       this);
       } else {
-        keyUpdates += 2;
         addPointWrite(root, begin, InternalVersionT(writeVersion), &tls, this);
       }
     }
 
+    // Run gc at least 200% the rate we're inserting entries
+    keyUpdates +=
+        std::max<int64_t>(tls.accum.entries_inserted - tls.accum.entries_erased,
+                          0) *
+        2;
+
     memory_bytes.set(totalBytes);
     point_writes_total.add(tls.accum.point_writes);
     range_writes_total.add(tls.accum.range_writes);

Jenkinsfile

@@ -117,15 +117,18 @@ pipeline {
                 }
             }
             steps {
+                script {
+                    filter_args = "-f ConflictSet.cpp -f LongestCommonPrefix.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 '''
-                gcovr -f ConflictSet.cpp --cobertura > build/coverage.xml
-                '''
+                sh """
+                gcovr ${filter_args} --cobertura > build/coverage.xml
+                """
                 recordCoverage qualityGates: [[criticality: 'NOTE', metric: 'MODULE']], tools: [[parser: 'COBERTURA', pattern: 'build/coverage.xml']]
-                sh '''
-                gcovr -f ConflictSet.cpp
-                gcovr -f ConflictSet.cpp --fail-under-line 100 > /dev/null
-                '''
+                sh """
+                gcovr ${filter_args}
+                gcovr ${filter_args} --fail-under-line 100 > /dev/null
+                """
             }
         }
     }

LongestCommonPrefix.h

@@ -0,0 +1,185 @@
+#pragma once
+
+#include <assert.h>
+#include <bit>
+#include <stdint.h>
+#include <string.h>
+
+#ifdef HAS_AVX
+#include <immintrin.h>
+#elif HAS_ARM_NEON
+#include <arm_neon.h>
+#endif
+
+#ifndef __SANITIZE_THREAD__
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define __SANITIZE_THREAD__
+#endif
+#endif
+#endif
+
+#if defined(HAS_AVX) || defined(HAS_ARM_NEON)
+constexpr int kStride = 64;
+#else
+constexpr int kStride = 16;
+#endif
+
+constexpr int kUnrollFactor = 4;
+
+inline bool compareStride(const uint8_t *ap, const uint8_t *bp) {
+#if defined(HAS_ARM_NEON)
+  static_assert(kStride == 64);
+  uint8x16_t x[4]; // GCOVR_EXCL_LINE
+  for (int i = 0; i < 4; ++i) {
+    x[i] = vceqq_u8(vld1q_u8(ap + i * 16), vld1q_u8(bp + i * 16));
+  }
+  auto results = vreinterpretq_u16_u8(
+      vandq_u8(vandq_u8(x[0], x[1]), vandq_u8(x[2], x[3])));
+  bool eq = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0) ==
+            uint64_t(-1);
+#elif defined(HAS_AVX)
+  static_assert(kStride == 64);
+  __m128i x[4]; // GCOVR_EXCL_LINE
+  for (int i = 0; i < 4; ++i) {
+    x[i] = _mm_cmpeq_epi8(_mm_loadu_si128((__m128i *)(ap + i * 16)),
+                          _mm_loadu_si128((__m128i *)(bp + i * 16)));
+  }
+  auto eq =
+      _mm_movemask_epi8(_mm_and_si128(_mm_and_si128(x[0], x[1]),
+                                      _mm_and_si128(x[2], x[3]))) == 0xffff;
+#else
+  // Hope it gets vectorized
+  auto eq = memcmp(ap, bp, kStride) == 0;
+#endif
+  return eq;
+}
+
+// Precondition: ap[:kStride] != bp[:kStride]
+inline int firstNeqStride(const uint8_t *ap, const uint8_t *bp) {
+#if defined(HAS_AVX)
+  static_assert(kStride == 64);
+  uint64_t c[kStride / 16]; // GCOVR_EXCL_LINE
+  for (int i = 0; i < kStride; i += 16) {
+    const auto a = _mm_loadu_si128((__m128i *)(ap + i));
+    const auto b = _mm_loadu_si128((__m128i *)(bp + i));
+    const auto compared = _mm_cmpeq_epi8(a, b);
+    c[i / 16] = _mm_movemask_epi8(compared) & 0xffff;
+  }
+  return std::countr_zero(~(c[0] | c[1] << 16 | c[2] << 32 | c[3] << 48));
+#elif defined(HAS_ARM_NEON)
+  static_assert(kStride == 64);
+  for (int i = 0; i < kStride; i += 16) {
+    // 0xff for each match
+    uint16x8_t results =
+        vreinterpretq_u16_u8(vceqq_u8(vld1q_u8(ap + i), vld1q_u8(bp + i)));
+    // 0xf for each mismatch
+    uint64_t bitfield =
+        ~vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(results, 4)), 0);
+    if (bitfield) {
+      return i + (std::countr_zero(bitfield) >> 2);
+    }
+  }
+  __builtin_unreachable(); // GCOVR_EXCL_LINE
+#else
+  int i = 0;
+  for (; i < kStride - 1; ++i) {
+    if (*ap++ != *bp++) {
+      break;
+    }
+  }
+  return i;
+#endif
+}
+
+// This gets covered in local development
+// GCOVR_EXCL_START
+#if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
+__attribute__((target("avx512f,avx512bw"))) inline int
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+  int i = 0;
+  int end = cl & ~63;
+  while (i < end) {
+    const uint64_t eq =
+        _mm512_cmpeq_epi8_mask(_mm512_loadu_epi8(ap), _mm512_loadu_epi8(bp));
+    if (eq != uint64_t(-1)) {
+      return i + std::countr_one(eq);
+    }
+    i += 64;
+    ap += 64;
+    bp += 64;
+  }
+  if (i < cl) {
+    const uint64_t mask = (uint64_t(1) << (cl - i)) - 1;
+    const uint64_t eq = _mm512_cmpeq_epi8_mask(
+        _mm512_maskz_loadu_epi8(mask, ap), _mm512_maskz_loadu_epi8(mask, bp));
+    return i + std::countr_one(eq & mask);
+  }
+  assert(i == cl);
+  return i;
+}
+__attribute__((target("default")))
+#endif
+// GCOVR_EXCL_STOP
+
+inline int
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+  if (!(cl >= 0)) {
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+
+  int i = 0;
+  int end;
+
+  // kStride * kUnrollCount at a time
+  end = cl & ~(kStride * kUnrollFactor - 1);
+  while (i < end) {
+    for (int j = 0; j < kUnrollFactor; ++j) {
+      if (!compareStride(ap, bp)) {
+        return i + firstNeqStride(ap, bp);
+      }
+      i += kStride;
+      ap += kStride;
+      bp += kStride;
+    }
+  }
+
+  // kStride at a time
+  end = cl & ~(kStride - 1);
+  while (i < end) {
+    if (!compareStride(ap, bp)) {
+      return i + firstNeqStride(ap, bp);
+    }
+    i += kStride;
+    ap += kStride;
+    bp += kStride;
+  }
+
+  // word at a time
+  end = cl & ~(sizeof(uint64_t) - 1);
+  while (i < end) {
+    uint64_t a; // GCOVR_EXCL_LINE
+    uint64_t b; // GCOVR_EXCL_LINE
+    memcpy(&a, ap, 8);
+    memcpy(&b, bp, 8);
+    const auto mismatched = a ^ b;
+    if (mismatched) {
+      return i + std::countr_zero(mismatched) / 8;
+    }
+    i += 8;
+    ap += 8;
+    bp += 8;
+  }
+
+  // byte at a time
+  while (i < cl) {
+    if (*ap != *bp) {
+      break;
+    }
+    ++ap;
+    ++bp;
+    ++i;
+  }
+
+  return i;
+}

README.md

@@ -24,15 +24,16 @@ 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
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.42 |       80,500,398.66 |    0.8% |          180.38 |           61.57 |  2.930 |          41.51 |    0.4% |      0.01 | `point reads`
-|               15.17 |       65,917,580.99 |    0.2% |          279.47 |           74.95 |  3.729 |          55.54 |    0.3% |      0.01 | `prefix reads`
-|               38.16 |       26,202,393.91 |    0.1% |          803.07 |          189.13 |  4.246 |         141.68 |    0.2% |      0.01 | `range reads`
-|               20.20 |       49,504,615.44 |    0.4% |          363.00 |          100.35 |  3.617 |          49.81 |    0.3% |      0.01 | `point writes`
-|               41.99 |       23,816,559.99 |    0.3% |          799.27 |          209.63 |  3.813 |         154.32 |    0.1% |      0.01 | `prefix writes`
-|               46.28 |       21,607,605.88 |    1.5% |          953.79 |          231.47 |  4.121 |         168.34 |    0.0% |      0.01 | `range writes`
-|               80.99 |       12,347,449.98 |    0.9% |        1,501.97 |          406.50 |  3.695 |         281.89 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          318,010.00 |            3,144.56 |    1.0% |    3,994,511.50 |    1,657,831.50 |  2.409 |     805,969.50 |    0.0% |      0.01 | `worst case for radix tree`
-|               75.85 |       13,183,612.56 |    0.5% |        1,590.01 |          385.64 |  4.123 |         258.00 |    0.0% |      0.01 | `create and destroy`
+|               10.80 |       92,600,541.52 |    0.6% |          180.38 |           54.49 |  3.310 |          41.51 |    0.4% |      0.01 | `point reads`
+|               15.00 |       66,687,691.68 |    0.4% |          278.44 |           76.44 |  3.642 |          55.56 |    0.3% |      0.01 | `prefix reads`
+|               36.81 |       27,163,394.61 |    0.4% |          795.06 |          187.91 |  4.231 |         142.67 |    0.2% |      0.01 | `range reads`
+|               18.14 |       55,137,674.01 |    1.2% |          338.19 |           92.86 |  3.642 |          42.81 |    0.4% |      0.01 | `point writes`
+|               33.19 |       30,127,119.71 |    0.1% |          681.03 |          170.05 |  4.005 |          98.68 |    0.2% |      0.01 | `prefix writes`
+|               37.37 |       26,759,432.70 |    1.9% |          779.70 |          195.45 |  3.989 |         114.21 |    0.0% |      0.01 | `range writes`
+|               74.36 |       13,448,582.47 |    1.9% |        1,425.68 |          389.08 |  3.664 |         258.88 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          316,928.00 |            3,155.29 |    1.5% |    3,992,986.00 |    1,699,813.00 |  2.349 |     806,226.50 |    0.0% |      0.01 | `worst case for radix tree`
+|               75.26 |       13,286,517.16 |    0.5% |        1,590.01 |          386.67 |  4.112 |         258.00 |    0.0% |      0.01 | `create and destroy`
+
 
 # "Real data" test
 
@@ -47,7 +48,7 @@ Check: 4.47891 seconds, 364.05 MB/s, Add: 4.55599 seconds, 123.058 MB/s, Gc rati
 ## radix tree
 
 ```
-Check: 0.963721 seconds, 1691.93 MB/s, Add: 1.3288 seconds, 421.924 MB/s, Gc ratio: 42.8819%
+Check: 0.910234 seconds, 1791.35 MB/s, Add: 1.25908 seconds, 445.287 MB/s, Gc ratio: 44.0415%
 ```
 
 ## hash table

ServerBench.cpp

@@ -76,6 +76,14 @@ void workload(weaselab::ConflictSet *cs) {
       } else {
         w.begin.len = k.size();
         cs->addWrites(&w, 1, version);
+        int64_t beginN = version - kWindowSize + rand() % kWindowSize;
+        auto b = numToKey(beginN);
+        auto e = numToKey(beginN + 1000);
+        w.begin.p = b.data();
+        w.begin.len = b.size();
+        w.end.p = e.data();
+        w.end.len = e.size();
+        cs->addWrites(&w, 1, version);
       }
     }
     // GC

corpus/2ae810daa467ba2bd354e835491a8b3026525764

@@ -0,0 +1 @@
+:ｩｩｩｩｩｩｩｩｩｩ:::::