Still need a better way identify left/right sidedness

Bench.cpp

@@ -1,4 +1,5 @@
 #include "ConflictSet.h"
+#include "Internal.h"
 
 #define ANKERL_NANOBENCH_IMPLEMENT
 #include "third_party/nanobench.h"
@@ -15,7 +16,7 @@ std::string toKey(int n) {
 
 constexpr int kNumKeys = 100000;
 
-int main(void) {
+void benchPointRead() {
   ankerl::nanobench::Bench bench;
 
   int readKey = kNumKeys / 2;
@@ -43,3 +44,65 @@ int main(void) {
   ConflictSet::Result result;
   bench.run("point read", [&]() { cs.check(&r, &result, 1); });
 }
+
+void benchPointerSet() {
+  ankerl::nanobench::Bench bench;
+
+  constexpr int kNumPointers = 32;
+  bench.batch(kNumPointers);
+
+  std::vector<void *> pointers;
+  for (int i = 0; i < kNumPointers; ++i) {
+    pointers.push_back(malloc(1));
+  }
+
+  bench.run("Create and destroy Arena hashset", [&]() {
+    Arena arena;
+    auto h = hashSet<void *>(arena);
+    for (auto p : pointers) {
+      h.insert(p);
+    }
+  });
+
+  bench.run("Create and destroy malloc hashset", [&]() {
+    std::unordered_set<void *, MyHash<void *>> h;
+    for (auto p : pointers) {
+      h.insert(p);
+    }
+  });
+
+  bench.run("Create and destroy Arena vector", [&]() {
+    Arena arena;
+    auto h = vector<void *>(arena);
+    for (auto p : pointers) {
+      h.push_back(p);
+    }
+  });
+
+  {
+    Arena arena;
+    auto h = hashSet<void *>(arena);
+    bench.run("Find hashset", [&]() {
+      for (auto p : pointers) {
+        bench.doNotOptimizeAway(h.find(p));
+      }
+    });
+  }
+
+  {
+    bench.run("Find vector", [&]() {
+      for (auto p : pointers) {
+        bench.doNotOptimizeAway(std::find(pointers.begin(), pointers.end(), p));
+      }
+    });
+  }
+
+  for (auto p : pointers) {
+    free(p);
+  }
+}
+
+int main(void) {
+  // benchPointRead();
+  benchPointerSet();
+}

ConflictSet.cpp

@@ -43,7 +43,7 @@ struct Node {
   int16_t numChildren = 0;
   bool entryPresent = false;
   uint8_t parentsIndex = 0;
-  constexpr static auto kPartialKeyMaxLen = 10;
+  constexpr static auto kPartialKeyMaxLen = 26;
   uint8_t partialKey[kPartialKeyMaxLen];
   int8_t partialKeyLen = 0;
   /* end section that's copied to the next node */
@@ -730,25 +730,37 @@ bool checkRangeRead(Node *n, const std::span<const uint8_t> begin,
     while (!left.step())
       ;
   }
+
   if (!rightDone) {
     while (!right.step())
       ;
   }
 
-  Arena arena;
+  Arena arena{64 << 10};
-  auto leftPath = vector<Node *>(arena);
+  // auto leftBorder = hashSet<Node *>(arena);
-  auto rightPath = vector<Node *>(arena);
+  // auto rightBorder = hashSet<Node *>(arena);
-  for (auto *iter = left.n; iter != nullptr; iter = iter->parent) {
+  auto leftBorder = vector<Node *>(arena);
-    leftPath.push_back(iter);
+  auto rightBorder = vector<Node *>(arena);
-  }
+  {
-  for (auto *iter = right.n; iter != nullptr; iter = iter->parent) {
+    auto *l = left.n;
-    rightPath.push_back(iter);
+    auto *r = right.n;
+    for (; l != nullptr && r != nullptr; l = l->parent, r = r->parent) {
+      leftBorder.push_back(l);
+      rightBorder.push_back(r);
+    }
+    for (; l != nullptr; l = l->parent) {
+      leftBorder.push_back(l);
+    }
+    for (; r != nullptr; r = r->parent) {
+      rightBorder.push_back(r);
+    }
   }
+
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "firstGeq for `%s' got `%s'\n", printable(begin).c_str(),
-          getSearchPath(left.n).c_str());
+          getSearchPathPrintable(left.n).c_str());
   fprintf(stderr, "firstGeq for `%s' got `%s'\n", printable(end).c_str(),
-          getSearchPath(right.n).c_str());
+          getSearchPathPrintable(right.n).c_str());
 #endif
 
   if (left.n != nullptr && left.cmp != 0 &&
@@ -764,12 +776,14 @@ bool checkRangeRead(Node *n, const std::span<const uint8_t> begin,
   }
 
   // TODO better data structure for pointer set. Also collect during first
-  // traversal.
+  // traversal?
   for (auto *iter = nextPhysical(left.n); iter != right.n;) {
-    const bool onLeftPath =
+    // const bool onLeftPath = leftBorder.find(iter) != leftBorder.end();
-        std::find(leftPath.begin(), leftPath.end(), iter) != leftPath.end();
+    // const bool onRightPath = rightBorder.find(iter) != rightBorder.end();
-    const bool onRightPath =
+    const bool onLeftPath = std::find(leftBorder.begin(), leftBorder.end(),
-        std::find(rightPath.begin(), rightPath.end(), iter) != rightPath.end();
+                                      iter) != leftBorder.end();
+    const bool onRightPath = std::find(rightBorder.begin(), rightBorder.end(),
+                                       iter) != rightBorder.end();
     if (!onLeftPath && !onRightPath) {
       if (iter->maxVersion > readVersion) {
         return false;

Internal.h

@@ -4,6 +4,7 @@
 
 #include <bit>
 #include <cassert>
+#include <cstddef>
 #include <cstdint>
 #include <cstdlib>
 #include <cstring>
@@ -184,9 +185,20 @@ template <class T> using Set = std::set<T, std::less<T>, ArenaAlloc<T>>;
 template <class T> auto set(Arena &arena) {
   return Set<T>(ArenaAlloc<T>(&arena));
 }
+
+template <class T> struct MyHash;
+
+template <class T> struct MyHash<T *> {
+  size_t operator()(const T *t) const noexcept {
+    size_t result;
+    memcpy(&result, &t, sizeof(result));
+    return result;
+  }
+};
+
 template <class T>
 using HashSet =
-    std::unordered_set<T, std::hash<T>, std::equal_to<T>, ArenaAlloc<T>>;
+    std::unordered_set<T, MyHash<T>, std::equal_to<T>, ArenaAlloc<T>>;
 template <class T> auto hashSet(Arena &arena) {
   return HashSet<T>(ArenaAlloc<T>(&arena));
 }
@@ -449,7 +461,7 @@ template <class ConflictSetImpl> struct TestDriver {
   ConflictSetImpl cs{writeVersion};
   ReferenceImpl refImpl{writeVersion};
 
-  constexpr static auto kMaxKeyLen = 24;
+  constexpr static auto kMaxKeyLen = 32;
 
   bool ok = true;