From 65429ebd0af689be2d72b9b3b4510dc10f43234f Mon Sep 17 00:00:00 2001
From: Andrew Noyes <andrew@weaselab.dev>
Date: Thu, 2 May 2024 15:34:19 -0700
Subject: [PATCH] Implement `insert`

---
 VersionedMap.cpp | 110 +++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 101 insertions(+), 9 deletions(-)

diff --git a/VersionedMap.cpp b/VersionedMap.cpp
index fc3059b..1ef3a5a 100644
--- a/VersionedMap.cpp
+++ b/VersionedMap.cpp
@@ -59,6 +59,11 @@ void munmapSafe(void *ptr, size_t size) {
 
 namespace weaselab {
 
+// 96 is enough for an entire search path in a tree with a size that
+// overflows int. See
+// https://en.wikipedia.org/wiki/Random_binary_tree#The_longest_path
+constexpr int kPathLengthUpperBound = 96;
+
 struct Entry {
   int64_t insertVersion;
   int keyLen;
@@ -216,7 +221,9 @@ struct MemManager {
   void gc(const uint32_t *roots, int numRoots, int64_t oldestVersion) {
     // Calculate reachable set
     BitSet reachable{next};
-    uint32_t stack[1000]; // Much more than bound imposed by max height of tree
+    // Each node has at most 3 children and nodes along the search path aren't
+    // in the stack, so we need 2 * kPathLengthUpperBound
+    uint32_t stack[2 * kPathLengthUpperBound];
     int stackIndex = 0;
     auto tryPush = [&](uint32_t p) {
       if (!reachable.set(p)) {
@@ -391,11 +398,37 @@ private:
   uint32_t end;
 };
 
+auto operator<=>(const VersionedMap::Mutation &lhs, const Node &rhs) {
+  int cl = std::min(lhs.param1Length, rhs.entry->keyLen);
+  if (cl > 0) {
+    int c = memcmp(lhs.param1, rhs.entry->getKey(), cl);
+    if (c != 0) {
+      return c <=> 0;
+    }
+  }
+  return lhs.param1Length <=> rhs.entry->keyLen;
+}
+
+struct Finger {
+  void push(uint32_t node) { searchPath[searchPathSize_++] = node; }
+  void pop() { --searchPathSize_; }
+  uint32_t back() const {
+    assert(searchPathSize_ > 0);
+    return searchPath[searchPathSize_ - 1];
+  }
+  uint32_t searchPathSize() const { return searchPathSize_; }
+
+private:
+  uint32_t searchPath[kPathLengthUpperBound];
+  int searchPathSize_ = 0;
+};
+
 struct VersionedMap::Impl {
 
   template <std::memory_order kOrder>
   uint32_t child(uint32_t node, bool which, int64_t at) {
-	  static_assert(kOrder == std::memory_order_acquire || kOrder == std::memory_order_relaxed);
+    static_assert(kOrder == std::memory_order_acquire ||
+                  kOrder == std::memory_order_relaxed);
     auto &n = mm.base[node];
     if (n.updated.load(kOrder) && n.updateVersion <= at &&
         which == n.replacedPointer) {
@@ -469,6 +502,60 @@ struct VersionedMap::Impl {
         at);
   }
 
+  void insert(const Mutation &m) {
+    Finger finger;
+    finger.push(latestRoot);
+    bool directionStack[kPathLengthUpperBound];
+    int directionStackSize = 0;
+
+    // Initialize finger to the search path of `m`
+    for (;;) {
+      auto n = finger.back();
+      if (n == 0) {
+        break;
+      }
+      auto c = m <=> mm.base[n];
+      if (c == 0) {
+        // No duplicates
+        break;
+      }
+      finger.push(child<std::memory_order_relaxed>(n, c > 0, latestVersion));
+      directionStack[directionStackSize++] = c > 0;
+    }
+
+    // Prepare new node
+    uint32_t node = newNode(latestVersion, m.param1, m.param1Length, m.param2,
+                            m.param2Length, false /* TODO set correctly */);
+    if (finger.back() != 0) {
+      auto &n = mm.base[node];
+      n.pointer[0] =
+          child<std::memory_order_relaxed>(finger.back(), false, latestVersion);
+      n.pointer[1] =
+          child<std::memory_order_relaxed>(finger.back(), true, latestVersion);
+    }
+
+    // Rotate and propagate up the search path
+    for (;;) {
+      if (finger.searchPathSize() == 1) {
+        // Made it to the root
+        latestRoot = node;
+        break;
+      }
+      finger.pop();
+      auto parent = finger.back();
+      const bool direction = directionStack[--directionStackSize];
+      parent = update(parent, latestVersion, direction, node);
+      if (mm.base[node].entry->priority > mm.base[parent].entry->priority) {
+        rotate(parent, latestVersion, direction);
+      } else {
+        if (parent == finger.back()) {
+          break;
+        }
+      }
+      node = parent;
+    }
+  }
+
   uint32_t newNode(int64_t version, const uint8_t *key, int keyLen,
                    const uint8_t *val, int valLen, bool clearTo) {
     auto result = mm.allocate();
@@ -513,6 +600,9 @@ struct VersionedMap::Impl {
 
   MemManager mm;
   RootSet roots;
+  // Only meaningful within the callstack of `addMutations`
+  uint32_t latestRoot;
+  int64_t latestVersion = 0;
 };
 } // namespace weaselab
 
@@ -523,19 +613,21 @@ int main() {
 
   {
     weaselab::VersionedMap::Impl impl;
-    impl.roots.add(impl.newNode(1, (const uint8_t *)"a", 1, nullptr, 0, true),
-                   1);
-    impl.roots.add(impl.newNode(2, (const uint8_t *)"b", 1, nullptr, -1, false),
-                   2);
-    impl.roots.add(impl.newNode(3, (const uint8_t *)"c", 1, nullptr, -1, false),
-                   3);
+    impl.latestRoot = impl.roots.roots()[impl.roots.rootCount() - 1];
+    impl.latestVersion = 1;
+    impl.insert(weaselab::VersionedMap::Mutation{
+        (const uint8_t *)"c", nullptr, 1, 0, weaselab::VersionedMap::Set});
+    impl.insert(weaselab::VersionedMap::Mutation{
+        (const uint8_t *)"b", nullptr, 1, 0, weaselab::VersionedMap::Set});
+    impl.insert(weaselab::VersionedMap::Mutation{
+        (const uint8_t *)"a", nullptr, 1, 0, weaselab::VersionedMap::Set});
+    impl.roots.add(impl.latestRoot, impl.latestVersion);
     impl.printInOrder(0);
     impl.printInOrder(1);
     impl.printInOrder(2);
     impl.printInOrder(3);
     impl.setOldestVersion(3);
   }
-  return 0;
 
   ankerl::nanobench::Bench bench;
   bench.minEpochIterations(5000);