Add benchmarks for getChild{L,G}eq

.clangd

@@ -1,2 +1,2 @@
 CompileFlags:
-  Add: [-DENABLE_MAIN, -UNDEBUG, -DENABLE_FUZZ]
+  Add: [-DENABLE_MAIN, -UNDEBUG, -DENABLE_FUZZ, -fexceptions]

CMakeLists.txt

@@ -82,8 +82,8 @@ target_compile_definitions(conflict_set_main PRIVATE ENABLE_MAIN)
 set(FUZZ_FLAGS "-fsanitize=fuzzer-no-link,address,undefined")
 include(CheckCXXCompilerFlag)
 cmake_push_check_state()
-set(CMAKE_REQUIRED_LINK_OPTIONS -fsanitize=fuzzer)
-check_cxx_compiler_flag(-fsanitize=fuzzer HAS_LIB_FUZZER)
+set(CMAKE_REQUIRED_LINK_OPTIONS -fsanitize=fuzzer-no-link)
+check_cxx_compiler_flag(-fsanitize=fuzzer-no-link HAS_LIB_FUZZER)
 cmake_pop_check_state()
 
 if(HAS_LIB_FUZZER)

ConflictSet.cpp

@@ -411,8 +411,7 @@ int getChildLeq(Node *self, int child) {
 #endif
   } else if (self->type == Type::Node48) {
     auto *self48 = static_cast<Node48 *>(self);
-    // TODO the plain loop is faster?
-#if 0 && (defined(HAS_AVX) || defined(HAS_ARM_NEON))
+#if defined(HAS_AVX) || defined(HAS_ARM_NEON)
     int i = child;
     if (i < 0) {
       return -1;
@@ -736,6 +735,24 @@ void insert(Node **self_, std::span<const uint8_t> key, int64_t writeVersion) {
   }
 }
 
+void destroyTree(Node *root) {
+  Arena arena;
+  auto toFree = vector<Node *>(arena);
+  toFree.push_back(root);
+  while (toFree.size() > 0) {
+    auto *n = toFree.back();
+    toFree.pop_back();
+    // Add all children to toFree
+    for (int child = getChildGeq(n, 0); child >= 0;
+         child = getChildGeq(n, child + 1)) {
+      auto *c = getChildExists(n, child);
+      assert(c != nullptr);
+      toFree.push_back(c);
+    }
+    free(n);
+  }
+}
+
 struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   void check(const ReadRange *reads, Result *result, int count) const {
     for (int i = 0; i < count; ++i) {
@@ -781,23 +798,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     root->entry.rangeVersion = oldestVersion;
     root->entryPresent = true;
   }
-  ~Impl() {
-    Arena arena;
-    auto toFree = vector<Node *>(arena);
-    toFree.push_back(root);
-    while (toFree.size() > 0) {
-      auto *n = toFree.back();
-      toFree.pop_back();
-      // Add all children to toFree
-      for (int child = getChildGeq(n, 0); child >= 0;
-           child = getChildGeq(n, child + 1)) {
-        auto *c = getChildExists(n, child);
-        assert(c != nullptr);
-        toFree.push_back(c);
-      }
-      free(n);
-    }
-  }
+  ~Impl() { destroyTree(root); }
 
   Node *root;
   int64_t oldestVersion;
@@ -991,21 +992,33 @@ void __throw_length_error(const char *) { __builtin_unreachable(); }
 } // namespace std
 
 #ifdef ENABLE_MAIN
-int main(void) {
-  int64_t writeVersion = 0;
-  ConflictSet::Impl cs{writeVersion};
-  ReferenceImpl refImpl{writeVersion};
-  Arena arena;
-  constexpr int kNumKeys = 10;
-  auto *write = new (arena) ConflictSet::WriteRange[kNumKeys];
-  for (int i = 0; i < kNumKeys; ++i) {
-    write[i].begin = toKey(arena, i);
-    write[i].end.len = 0;
-    write[i].writeVersion = ++writeVersion;
+#define ANKERL_NANOBENCH_IMPLEMENT
+#include "third_party/nanobench.h"
+
+void bench() {
+  ankerl::nanobench::Bench bench;
+  {
+    auto *n = newNode();
+    for (int i = 0; i < 64; ++i) {
+      getOrCreateChild(n, i) = newNode();
+      bench.run("getChildLeq" + std::to_string(i),
+                [&]() { bench.doNotOptimizeAway(getChildLeq(n, 255)); });
+    }
+    destroyTree(n);
   }
-  cs.addWrites(write, kNumKeys);
-  refImpl.addWrites(write, kNumKeys);
-  debugPrintDot(stdout, cs.root);
+  {
+    auto *n = newNode();
+    for (int i = 255; i >= 3 * 64; --i) {
+      getOrCreateChild(n, i) = newNode();
+      bench.run("getChildGeq" + std::to_string(i),
+                [&]() { bench.doNotOptimizeAway(getChildGeq(n, 0)); });
+    }
+    destroyTree(n);
+  }
+}
+
+int main(void) {
+  bench();
   return 0;
 }
 #endif