From 116c79d3deeb191601fac70f46f75678a3ced511 Mon Sep 17 00:00:00 2001
From: Andrew Noyes <andrew@weaselab.dev>
Date: Fri, 23 Feb 2024 12:34:28 -0800
Subject: [PATCH] Optimize getOrCreateChild

Add fast path for if it exists

Avoid some unnecessary branches

Try to simplify some
---
 ConflictSet.cpp | 126 ++++++++++++++++++++++++------------------------
 1 file changed, 63 insertions(+), 63 deletions(-)

diff --git a/ConflictSet.cpp b/ConflictSet.cpp
index 9558e43..31247f8 100644
--- a/ConflictSet.cpp
+++ b/ConflictSet.cpp
@@ -382,11 +382,30 @@ void setChildrenParents(Node *node) {
 // reference if null
 Node *&getOrCreateChild(Node *&self, uint8_t index,
                         NodeAllocators *allocators) {
+
+  // Fast path for if it exists already
+  if (self->type <= Type::Node16) {
+    auto *self16 = static_cast<Node16 *>(self);
+    int i = getNodeIndex(self16, index);
+    if (i >= 0) {
+      return self16->children[i];
+    }
+  } else if (self->type == Type::Node48) {
+    auto *self48 = static_cast<Node48 *>(self);
+    int secondIndex = self48->index[index];
+    if (secondIndex >= 0) {
+      return self48->children[secondIndex];
+    }
+  } else {
+    auto *self256 = static_cast<Node256 *>(self);
+    if (auto &result = self256->children[index]; result != nullptr) {
+      return result;
+    }
+  }
+
   if (self->type == Type::Node4) {
     auto *self4 = static_cast<Node4 *>(self);
-    if (int i = getNodeIndex((Node16 *)self4, index); i >= 0) {
-      return self4->children[i];
-    }
+
     if (self->numChildren == 4) {
       auto *newSelf = allocators->node16.allocate();
       memcpy((void *)newSelf, self, offsetof(Node, type));
@@ -395,32 +414,14 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
       allocators->node4.release(self4);
       self = newSelf;
       setChildrenParents(self);
-      goto insert16;
-    } else {
-      ++self->numChildren;
-      for (int i = 0; i < int(self->numChildren) - 1; ++i) {
-        if (int(self4->index[i]) > int(index)) {
-          memmove(self4->index + i + 1, self4->index + i,
-                  self->numChildren - (i + 1));
-          memmove(self4->children + i + 1, self4->children + i,
-                  (self->numChildren - (i + 1)) * sizeof(void *));
-          self4->index[i] = index;
-          self4->children[i] = nullptr;
-          return self4->children[i];
-        }
-      }
-      self4->index[self->numChildren - 1] = index;
-      self4->children[self->numChildren - 1] = nullptr;
-      return self4->children[self->numChildren - 1];
     }
-  } else if (self->type == Type::Node16) {
-  insert16:
-    auto *self16 = static_cast<Node16 *>(self);
 
-    if (int i = getNodeIndex(self16, index); i >= 0) {
-      return self16->children[i];
-    }
+    goto insert16;
+
+  } else if (self->type == Type::Node16) {
+
     if (self->numChildren == 16) {
+      auto *self16 = static_cast<Node16 *>(self);
       auto *newSelf = allocators->node48.allocate();
       memcpy((void *)newSelf, self, offsetof(Node, type));
       newSelf->nextFree = 16;
@@ -436,62 +437,61 @@ Node *&getOrCreateChild(Node *&self, uint8_t index,
       self = newSelf;
       setChildrenParents(self);
       goto insert48;
-    } else {
-      ++self->numChildren;
-      for (int i = 0; i < int(self->numChildren) - 1; ++i) {
-        if (int(self16->index[i]) > int(index)) {
-          memmove(self16->index + i + 1, self16->index + i,
-                  self->numChildren - (i + 1));
-          memmove(self16->children + i + 1, self16->children + i,
-                  (self->numChildren - (i + 1)) * sizeof(void *));
-          self16->index[i] = index;
-          self16->children[i] = nullptr;
-          return self16->children[i];
-        }
-      }
-      self16->index[self->numChildren - 1] = index;
-      self16->children[self->numChildren - 1] = nullptr;
-      return self16->children[self->numChildren - 1];
     }
-  } else if (self->type == Type::Node48) {
-  insert48:
-    auto *self48 = static_cast<Node48 *>(self);
-    if (int c = self48->index[index];
 
-        c >= 0) {
-      return self48->children[c];
+  insert16:
+    auto *self16 = static_cast<Node16 *>(self);
+
+    ++self->numChildren;
+    int i = 0;
+    for (; i < int(self->numChildren) - 1; ++i) {
+      if (int(self16->index[i]) > int(index)) {
+        memmove(self16->index + i + 1, self16->index + i,
+                self->numChildren - (i + 1));
+        memmove(self16->children + i + 1, self16->children + i,
+                (self->numChildren - (i + 1)) * sizeof(void *));
+        break;
+      }
     }
+    self16->index[i] = index;
+    auto &result = self16->children[i];
+    result = nullptr;
+    return result;
+  } else if (self->type == Type::Node48) {
+
     if (self->numChildren == 48) {
+      auto *self48 = static_cast<Node48 *>(self);
       auto *newSelf = allocators->node256.allocate();
       memcpy((void *)newSelf, self, offsetof(Node, type));
+      newSelf->bitSet = self48->bitSet;
       for (int i = 0; i < 256; ++i) {
-        if (self48->bitSet.test(i)) {
-          newSelf->bitSet.set(i);
-          newSelf->children[i] = self48->children[self48->index[i]];
+        int c = self48->index[i];
+        if (c >= 0) {
+          newSelf->children[i] = self48->children[c];
         }
       }
       allocators->node48.release(self48);
       self = newSelf;
       setChildrenParents(self);
       goto insert256;
-    } else {
-      self48->bitSet.set(index);
-      ++self->numChildren;
-      assert(self48->nextFree < 48);
-      self48->index[index] = self48->nextFree;
-      self48->children[self48->nextFree] = nullptr;
-      return self48->children[self48->nextFree++];
     }
+  insert48:
+
+    auto *self48 = static_cast<Node48 *>(self);
+    self48->bitSet.set(index);
+    ++self->numChildren;
+    assert(self48->nextFree < 48);
+    int nextFree = self48->nextFree++;
+    self48->index[index] = nextFree;
+    auto &result = self48->children[nextFree];
+    result = nullptr;
+    return result;
   } else {
   insert256:
     auto *self256 = static_cast<Node256 *>(self);
-    auto *&result = self256->children[index];
-    if (result) {
-      return result;
-    }
     ++self->numChildren;
     self256->bitSet.set(index);
-    return result;
+    return self256->children[index];
   }
 }