Add benchmark

Bench.cpp

@@ -1,5 +1,7 @@
 #include "ConflictSet.h"
 #include "Internal.h"
+#include <byteswap.h>
+#include <cstdint>
 
 #define ANKERL_NANOBENCH_IMPLEMENT
 #include "third_party/nanobench.h"
@@ -15,94 +17,165 @@ std::string toKey(int n) {
 }
 
 constexpr int kNumKeys = 100000;
+// A range read, a point read, and a point write. Range writes can erase
+// keys, and we don't want to change the number of keys stored in the
+// conflict set.
+constexpr int kOpsPerTx = 100;
 
-void benchPointRead() {
+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], 4};
+  index = __builtin_bswap32(index);
+  memcpy(result.data(), &index, 4);
+  return result;
+}
+
+template <class ConflictSet_> void benchConflictSet() {
   ankerl::nanobench::Bench bench;
+  ConflictSet_ cs{0};
 
-  int readKey = kNumKeys / 2;
+  bench.batch(kOpsPerTx);
 
-  ConflictSet cs(0);
+  int64_t version = 0;
+
+  // Populate conflict set
+  Arena arena;
+  {
+    std::vector<ConflictSet::WriteRange> writes;
+    writes.reserve(kNumKeys);
     for (int i = 0; i < kNumKeys; ++i) {
-    if (i != readKey) {
+      auto key = makeKey(arena, i);
+      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;
+  }
+
+  // I don't know why std::less didn't work /shrug
+  struct Less {
+    bool operator()(const std::span<const uint8_t> &lhs,
+                    const std::span<const uint8_t> &rhs) const {
+      return lhs < rhs;
+    }
+  };
+  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));
+  }
+
+  // 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 key = toKey(i);
-    ConflictSet::WriteRange w;
-    w.begin.p = (const uint8_t *)key.data();
-    w.begin.len = key.size();
-    w.end.len = 0;
-    w.writeVersion = 1;
-    cs.addWrites(&w, 1);
+      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;
   }
 
-  auto key = toKey(readKey);
+  {
+    std::vector<ConflictSet::ReadRange> reads;
+    auto iter = points.begin();
+    for (int i = 0; i < kOpsPerTx; ++i) {
       ConflictSet::ReadRange r;
-  r.begin.p = (const uint8_t *)key.data();
-  r.begin.len = key.size();
+      r.begin.p = iter->data();
+      r.begin.len = iter->size();
       r.end.len = 0;
-  r.readVersion = 0;
-  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));
+      r.readVersion = version - 1;
+      reads.push_back(r);
+      ++iter;
     }
 
-  bench.run("Create and destroy Arena hashset", [&]() {
-    Arena arena;
-    auto h = hashSet<void *>(arena);
-    for (auto p : pointers) {
-      h.insert(p);
-    }
-  });
+    auto *results = new (arena) ConflictSet::Result[kOpsPerTx];
 
-  bench.run("Create and destroy malloc hashset", [&]() {
-    std::unordered_set<void *, MyHash<void *>> h;
-    for (auto p : pointers) {
-      h.insert(p);
+    bench.run("radix tree (point reads)",
+              [&]() { cs.check(reads.data(), results, kOpsPerTx); });
   }
-  });
-
-  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));
+    std::vector<ConflictSet::ReadRange> reads;
+    auto iter = points.begin();
+    for (int i = 0; i < kOpsPerTx; ++i) {
+      auto begin = *iter++;
+      auto end = *iter++;
+      ConflictSet::ReadRange r;
+      r.begin.p = begin.data();
+      r.begin.len = begin.size();
+      r.end.p = end.data();
+      r.end.len = end.size();
+      r.readVersion = version - 1;
+      reads.push_back(r);
     }
+
+    auto *results = new (arena) ConflictSet::Result[kOpsPerTx];
+
+    bench.run("radix tree (range reads)",
+              [&]() { cs.check(reads.data(), results, kOpsPerTx); });
+  }
+
+  {
+    std::vector<ConflictSet::WriteRange> writes;
+    auto iter = points.begin();
+    for (int i = 0; i < kOpsPerTx; ++i) {
+      ConflictSet::WriteRange w;
+      w.begin.p = iter->data();
+      w.begin.len = iter->size();
+      w.end.len = 0;
+      writes.push_back(w);
+      ++iter;
+    }
+
+    bench.run("radix tree (point writes)", [&]() {
+      auto v = ++version;
+      for (auto &w : writes) {
+        w.writeVersion = v;
+      }
+      cs.addWrites(writes.data(), writes.size());
     });
   }
 
   {
-    bench.run("Find vector", [&]() {
-      for (auto p : pointers) {
-        bench.doNotOptimizeAway(std::find(pointers.begin(), pointers.end(), p));
+    std::vector<ConflictSet::WriteRange> writes;
+    auto iter = points.begin();
+    for (int i = 0; i < kOpsPerTx - 1; ++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();
+      writes.push_back(w);
     }
+
+    bench.run("radix tree (range writes)", [&]() {
+      auto v = ++version;
+      for (auto &w : writes) {
+        w.writeVersion = v;
+      }
+      cs.addWrites(writes.data(), writes.size());
     });
   }
-
-  for (auto p : pointers) {
-    free(p);
-  }
 }
 
-int main(void) {
-  // benchPointRead();
-  benchPointerSet();
-}
+int main(void) { benchConflictSet<ConflictSet>(); }

ConflictSet.cpp

@@ -4,6 +4,7 @@
 #include <algorithm>
 #include <bit>
 #include <cassert>
+#include <compare>
 #include <cstdint>
 #include <cstring>
 #include <inttypes.h>
@@ -21,15 +22,6 @@
 
 // ==================== BEGIN IMPLEMENTATION ====================
 
-int compare(std::span<const uint8_t> lhs, std::span<const uint8_t> rhs) {
-  int cl = std::min<int>(lhs.size(), rhs.size());
-  int c = cl == 0 ? 0 : memcmp(lhs.data(), rhs.data(), cl);
-  if (c != 0) {
-    return c;
-  }
-  return int(lhs.size()) - int(rhs.size());
-}
-
 std::span<uint8_t> strincMutate(std::span<uint8_t> str, bool &ok) {
   int index;
   for (index = str.size() - 1; index >= 0; index--)
@@ -793,7 +785,7 @@ bool checkRangeRead(Node *n, const std::span<const uint8_t> begin,
   }
 
   bool first = true;
-  for (auto *iter = left.n; iter != nullptr && compare(searchPath, end) < 0;
+  for (auto *iter = left.n; iter != nullptr && searchPath < end;
        first = false) {
     if (iter->entryPresent) {
       if (!first && iter->entry.rangeVersion > readVersion) {
@@ -811,11 +803,11 @@ bool checkRangeRead(Node *n, const std::span<const uint8_t> begin,
 #endif
     bool ok = true;
     auto rangeEnd = strincMutate(searchPath, ok);
-    int c;
+    auto c = std::strong_ordering::equal;
     if (!ok) {
       goto iterate;
     }
-    c = compare(rangeEnd, end);
+    c = rangeEnd <=> end;
     --rangeEnd.back();
 
     if (c == 0) {

Internal.h

@@ -4,6 +4,7 @@
 
 #include <bit>
 #include <cassert>
+#include <compare>
 #include <cstddef>
 #include <cstdint>
 #include <cstdlib>
@@ -19,6 +20,18 @@
 
 #define DEBUG_VERBOSE 0
 
+[[nodiscard]] inline auto
+operator<=>(const std::span<const uint8_t> &lhs,
+            const std::span<const uint8_t> &rhs) noexcept {
+  int cl = std::min<int>(lhs.size(), rhs.size());
+  if (cl > 0) {
+    if (auto c = memcmp(lhs.data(), rhs.data(), cl) <=> 0; c != 0) {
+      return c;
+    }
+  }
+  return lhs.size() <=> rhs.size();
+}
+
 // This header contains code that we want to reuse outside of ConflictSet.cpp or
 // want to exclude from coverage since it's only testing related.
 
@@ -181,9 +194,9 @@ template <class T> using Vector = std::vector<T, ArenaAlloc<T>>;
 template <class T> auto vector(Arena &arena) {
   return Vector<T>(ArenaAlloc<T>(&arena));
 }
-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, class C> using Set = std::set<T, C, ArenaAlloc<T>>;
+template <class T, class C = std::less<T>> auto set(Arena &arena) {
+  return Set<T, C>(ArenaAlloc<T>(&arena));
 }
 
 template <class T> struct MyHash;