Range reads are slow with dense keys

Bench.cpp

@@ -543,8 +543,8 @@ private:
 };
 
 struct SkipListConflictSet {
-  int64_t oldestVersion;
-  SkipListConflictSet(int64_t oldestVersion) : oldestVersion(oldestVersion) {}
+  SkipListConflictSet(int64_t oldestVersion)
+      : oldestVersion(oldestVersion), skipList(oldestVersion) {}
   void check(const ConflictSet::ReadRange *reads, ConflictSet::Result *results,
              int count) const {
     Arena arena;
@@ -594,29 +594,37 @@ struct SkipListConflictSet {
   }
 
 private:
+  int64_t oldestVersion;
   SkipList skipList;
 };
 
 } // namespace
 
-constexpr int kNumKeys = 100000;
+constexpr int kNumKeys = 1000000;
 
 constexpr int kOpsPerTx = 100;
 
 constexpr int kPrefixLen = 0;
 
 std::span<const uint8_t> makeKey(Arena &arena, int index) {
+
   auto result =
       std::span<uint8_t>{new (arena) uint8_t[4 + kPrefixLen], 4 + kPrefixLen};
   index = __builtin_bswap32(index);
   memset(result.data(), 0, kPrefixLen);
   memcpy(result.data() + kPrefixLen, &index, 4);
+
+  // auto result =
+  //     std::span<uint8_t>{new (arena) uint8_t[32], 32};
+  // for (int i = 0; i < 32; ++i) {
+  //   result[i] = index & (1 << (31 - i)) ? '1' : '0';
+  // }
+
   return result;
 }
 
 template <class ConflictSet_> void benchConflictSet(const std::string &name) {
   ankerl::nanobench::Bench bench;
-  bench.minEpochIterations(1000);
   ConflictSet_ cs{0};
 
   bench.batch(kOpsPerTx);
@@ -650,8 +658,6 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
   };
   auto points = set<std::span<const uint8_t>, Less>(arena);
 
-  // Two points for each range read, one for each point read, and one for each
-  // point write
   while (points.size() < kOpsPerTx * 2 + 1) {
     // TODO don't use rand?
     points.insert(makeKey(arena, rand() % kNumKeys));
@@ -660,21 +666,21 @@ template <class ConflictSet_> void benchConflictSet(const std::string &name) {
   // Make short-circuiting non-trivial
   {
     std::vector<ConflictSet::WriteRange> writes;
-    writes.reserve(kNumKeys);
-    for (int i = 0; i < kNumKeys; ++i) {
-      auto key = makeKey(arena, i);
-      if (points.find(key) != points.end()) {
-        continue;
+    auto iter = points.begin();
+    ++iter; // Complement of the set we'll be reading with range reads. Almost.
+    for (int i = 0; i < kOpsPerTx; ++i) {
+      auto begin = *iter++;
+      auto end = *iter++;
+      ConflictSet::WriteRange w;
+      w.begin.p = begin.data();
+      w.begin.len = begin.size();
+      w.end.p = end.data();
+      w.end.len = end.size();
+      w.writeVersion = version + 1;
+      writes.push_back(w);
     }
-      ConflictSet::WriteRange conflict;
-      conflict.begin.p = key.data();
-      conflict.begin.len = key.size();
-      conflict.end.len = 0;
-      conflict.writeVersion = version + 1;
-      writes.push_back(conflict);
-    }
-    cs.addWrites(writes.data(), writes.size());
     ++version;
+    cs.addWrites(writes.data(), kOpsPerTx);
   }
 
   {

ConflictSet.cpp

@@ -894,9 +894,7 @@ bytes:
 bool checkRangeRead(Node *n, const std::span<const uint8_t> begin,
                     const std::span<const uint8_t> end, int64_t readVersion,
                     Arena &arena) {
-  auto left = FirstGeqStepwise{n, begin};
-  while (!left.step())
-    ;
+  auto left = firstGeq(n, begin);
 
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "firstGeq for `%s' got `%s'\n", printable(begin).c_str(),