Remove dead stores

Once we know the type, for Node3 and higher we know the first child
exists anyway

CMakeLists.txt

@@ -32,7 +32,12 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
 endif()
 
 add_compile_options(
-  -Werror=switch-enum -Wswitch-enum -fPIC -fdata-sections -ffunction-sections
+  -Werror=switch-enum
+  -Wswitch-enum
+  -Wunused-variable
+  -fPIC
+  -fdata-sections
+  -ffunction-sections
   -fno-jump-tables # https://github.com/llvm/llvm-project/issues/54247
 )
 
@@ -104,19 +109,11 @@ else()
   add_link_options(-Wl,--gc-sections)
 endif()
 
-if(NOT USE_SIMD_FALLBACK)
+if(USE_SIMD_FALLBACK)
-  cmake_push_check_state()
+  add_compile_definitions(USE_SIMD_FALLBACK)
-  list(APPEND CMAKE_REQUIRED_FLAGS -mavx)
+else()
-  check_include_file_cxx("immintrin.h" HAS_AVX)
+  if(CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64)
-  if(HAS_AVX)
     add_compile_options(-mavx)
-    add_compile_definitions(HAS_AVX)
-  endif()
-  cmake_pop_check_state()
-
-  check_include_file_cxx("arm_neon.h" HAS_ARM_NEON)
-  if(HAS_ARM_NEON)
-    add_compile_definitions(HAS_ARM_NEON)
   endif()
 endif()
 
@@ -346,8 +343,7 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   # c++98
   add_executable(conflict_set_cxx_api_test conflict_set_cxx_api_test.cpp)
   target_compile_options(conflict_set_cxx_api_test PRIVATE ${TEST_FLAGS})
-  target_link_libraries(conflict_set_cxx_api_test
+  target_link_libraries(conflict_set_cxx_api_test PRIVATE ${PROJECT_NAME})
-                        PRIVATE ${PROJECT_NAME}-static)
   set_target_properties(conflict_set_cxx_api_test PROPERTIES CXX_STANDARD 98)
   set_target_properties(conflict_set_cxx_api_test
                         PROPERTIES CXX_STANDARD_REQUIRED ON)

ConflictSet.cpp

@@ -14,6 +14,16 @@ See the License for the specific language governing permissions and
 limitations under the License.
 */
 
+#if !defined(USE_SIMD_FALLBACK) && defined(__has_include)
+#if defined(__x86_64__) && __has_include("immintrin.h")
+#define HAS_AVX 1
+#include <immintrin.h>
+#elif __has_include("arm_neon.h")
+#define HAS_ARM_NEON 1
+#include <arm_neon.h>
+#endif
+#endif
+
 #include "ConflictSet.h"
 #include "Internal.h"
 #include "LongestCommonPrefix.h"
@@ -34,12 +44,6 @@ limitations under the License.
 #include <type_traits>
 #include <utility>
 
-#ifdef HAS_AVX
-#include <immintrin.h>
-#elif defined(HAS_ARM_NEON)
-#include <arm_neon.h>
-#endif
-
 #ifndef __SANITIZE_THREAD__
 #if defined(__has_feature)
 #if __has_feature(thread_sanitizer)
@@ -341,8 +345,8 @@ struct Node3 : Node {
   // Sorted
   uint8_t index[kMaxNodes];
 
-  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
 
   uint8_t *partialKey() {
     assert(!releaseDeferred);
@@ -696,8 +700,6 @@ constexpr int64_t kMaxFreeListBytes = 1 << 20;
 // doesn't meet the capacity constraints, it's freed and a new node is allocated
 // with the minimum capacity. The hope is that "unfit" nodes don't get stuck in
 // the free list.
-//
-// TODO valgrind annotations
 template <class T> struct NodeAllocator {
 
   static_assert(std::derived_from<T, Node>);
@@ -734,6 +736,7 @@ template <class T> struct NodeAllocator {
     p->parent = freeList;
     freeList = p;
     freeListSize += sizeof(T) + p->partialKeyCapacity;
+    VALGRIND_MAKE_MEM_NOACCESS(p, sizeof(T) + p->partialKeyCapacity);
   }
 
   void deferRelease(T *p, Node *forwardTo) {
@@ -755,6 +758,13 @@ template <class T> struct NodeAllocator {
   void releaseDeferred() {
     if (deferredList != nullptr) {
       deferredListFront->parent = freeList;
+#ifndef NVALGRIND
+      for (auto *iter = deferredList; iter != freeList;) {
+        auto *tmp = iter;
+        iter = (T *)iter->parent;
+        VALGRIND_MAKE_MEM_NOACCESS(tmp, sizeof(T) + tmp->partialKeyCapacity);
+      }
+#endif
       freeList = std::exchange(deferredList, nullptr);
     }
     for (T *n = std::exchange(deferredListOverflow, nullptr); n != nullptr;) {
@@ -775,6 +785,7 @@ template <class T> struct NodeAllocator {
     assert(deferredList == nullptr);
     assert(deferredListOverflow == nullptr);
     for (T *iter = freeList; iter != nullptr;) {
+      VALGRIND_MAKE_MEM_DEFINED(iter, sizeof(T));
       auto *tmp = iter;
       iter = (T *)iter->parent;
       removeNode(tmp);
@@ -792,6 +803,7 @@ private:
 
   T *allocate_helper(int minCapacity, int maxCapacity) {
     if (freeList != nullptr) {
+      VALGRIND_MAKE_MEM_DEFINED(freeList, sizeof(T));
       freeListSize -= sizeof(T) + freeList->partialKeyCapacity;
       assume(freeList->partialKeyCapacity >= 0);
       assume(minCapacity >= 0);
@@ -800,6 +812,11 @@ private:
           freeList->partialKeyCapacity <= maxCapacity) {
         auto *result = freeList;
         freeList = (T *)freeList->parent;
+        VALGRIND_MAKE_MEM_UNDEFINED(result,
+                                    sizeof(T) + result->partialKeyCapacity);
+        VALGRIND_MAKE_MEM_DEFINED(&result->partialKeyCapacity,
+                                  sizeof(result->partialKeyCapacity));
+        VALGRIND_MAKE_MEM_DEFINED(&result->type, sizeof(result->type));
         return result;
       } else {
         auto *p = freeList;
@@ -943,8 +960,7 @@ private:
   NodeAllocator<Node256> node256;
 };
 
-int getNodeIndex(Node3 *self, uint8_t index) {
+int getNodeIndex(Node3 *n, uint8_t index) {
-  Node3 *n = (Node3 *)self;
   assume(n->numChildren >= 1);
   assume(n->numChildren <= 3);
   for (int i = 0; i < n->numChildren; ++i) {
@@ -955,8 +971,7 @@ int getNodeIndex(Node3 *self, uint8_t index) {
   return -1;
 }
 
-int getNodeIndexExists(Node3 *self, uint8_t index) {
+int getNodeIndexExists(Node3 *n, uint8_t index) {
-  Node3 *n = (Node3 *)self;
   assume(n->numChildren >= 1);
   assume(n->numChildren <= 3);
   for (int i = 0; i < n->numChildren; ++i) {
@@ -1257,33 +1272,32 @@ TaggedNodePointer getChild(Node *self, uint8_t index) {
 struct ChildAndMaxVersion {
   TaggedNodePointer child;
   InternalVersionT maxVersion;
+  static ChildAndMaxVersion empty() {
+    ChildAndMaxVersion result;
+    result.child = nullptr;
+    return result;
+  }
 };
 
 ChildAndMaxVersion getChildAndMaxVersion(Node0 *, uint8_t) { return {}; }
 ChildAndMaxVersion getChildAndMaxVersion(Node3 *self, uint8_t index) {
   int i = getNodeIndex(self, index);
   if (i < 0) {
-    ChildAndMaxVersion result;
+    return ChildAndMaxVersion::empty();
-    result.child = nullptr;
-    return result;
   }
   return {self->children[i], self->childMaxVersion[i]};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node16 *self, uint8_t index) {
   int i = getNodeIndex(self, index);
   if (i < 0) {
-    ChildAndMaxVersion result;
+    return ChildAndMaxVersion::empty();
-    result.child = nullptr;
-    return result;
   }
   return {self->children[i], self->childMaxVersion[i]};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node48 *self, uint8_t index) {
   int i = self->index[index];
   if (i < 0) {
-    ChildAndMaxVersion result;
+    return ChildAndMaxVersion::empty();
-    result.child = nullptr;
-    return result;
   }
   return {self->children[i], self->childMaxVersion[i]};
 }
@@ -1366,17 +1380,11 @@ TaggedNodePointer getChildGeq(Node16 *self, int child) {
 
 TaggedNodePointer getChildGeq(Node48 *self, int child) {
   int c = self->bitSet.firstSetGeq(child);
-  if (c < 0) {
+  return c < 0 ? nullptr : self->children[self->index[c]];
-    return nullptr;
-  }
-  return self->children[self->index[c]];
 }
 TaggedNodePointer getChildGeq(Node256 *self, int child) {
   int c = self->bitSet.firstSetGeq(child);
-  if (c < 0) {
+  return c < 0 ? nullptr : self->children[c];
-    return nullptr;
-  }
-  return self->children[c];
 }
 
 TaggedNodePointer getChildGeq(Node *self, int child) {
@@ -1396,22 +1404,25 @@ TaggedNodePointer getChildGeq(Node *self, int child) {
   }
 }
 
-Node *getFirstChild(Node0 *) { return nullptr; }
+TaggedNodePointer getFirstChild(Node0 *) { return nullptr; }
-Node *getFirstChild(Node3 *self) {
+TaggedNodePointer getFirstChild(Node3 *self) {
-  return self->numChildren == 0 ? nullptr : self->children[0];
+  // Improves scan performance
+  __builtin_prefetch(self->children[1]);
+  return self->children[0];
 }
-Node *getFirstChild(Node16 *self) {
+TaggedNodePointer getFirstChild(Node16 *self) {
-  return self->numChildren == 0 ? nullptr : self->children[0];
+  // Improves scan performance
+  __builtin_prefetch(self->children[1]);
+  return self->children[0];
 }
-Node *getFirstChild(Node48 *self) {
+TaggedNodePointer getFirstChild(Node48 *self) {
-  int index = self->index[self->bitSet.firstSetGeq(0)];
+  return self->children[self->index[self->bitSet.firstSetGeq(0)]];
-  return index < 0 ? nullptr : self->children[index];
 }
-Node *getFirstChild(Node256 *self) {
+TaggedNodePointer getFirstChild(Node256 *self) {
   return self->children[self->bitSet.firstSetGeq(0)];
 }
 
-Node *getFirstChild(Node *self) {
+TaggedNodePointer getFirstChild(Node *self) {
   // Only require that the node-specific overloads are covered
   // GCOVR_EXCL_START
   switch (self->getType()) {
@@ -1431,46 +1442,6 @@ Node *getFirstChild(Node *self) {
   // GCOVR_EXCL_STOP
 }
 
-// Precondition: self has a child
-TaggedNodePointer getFirstChildExists(Node3 *self) {
-  assert(self->numChildren > 0);
-  return self->children[0];
-}
-// Precondition: self has a child
-TaggedNodePointer getFirstChildExists(Node16 *self) {
-  assert(self->numChildren > 0);
-  return self->children[0];
-}
-// Precondition: self has a child
-TaggedNodePointer getFirstChildExists(Node48 *self) {
-  return self->children[self->index[self->bitSet.firstSetGeq(0)]];
-}
-// Precondition: self has a child
-TaggedNodePointer getFirstChildExists(Node256 *self) {
-  return self->children[self->bitSet.firstSetGeq(0)];
-}
-
-// Precondition: self has a child
-TaggedNodePointer getFirstChildExists(Node *self) {
-  // Only require that the node-specific overloads are covered
-  // GCOVR_EXCL_START
-  switch (self->getType()) {
-  case Type_Node0:
-    __builtin_unreachable();
-  case Type_Node3:
-    return getFirstChildExists(static_cast<Node3 *>(self));
-  case Type_Node16:
-    return getFirstChildExists(static_cast<Node16 *>(self));
-  case Type_Node48:
-    return getFirstChildExists(static_cast<Node48 *>(self));
-  case Type_Node256:
-    return getFirstChildExists(static_cast<Node256 *>(self));
-  default:
-    __builtin_unreachable();
-  }
-  // GCOVR_EXCL_STOP
-}
-
 // self must not be the root
 void maybeDecreaseCapacity(Node *&self, WriteContext *writeContext,
                            ConflictSet::Impl *impl);
@@ -1724,7 +1695,7 @@ TaggedNodePointer &getOrCreateChild(TaggedNodePointer &self, TrivialSpan &key,
 }
 
 Node *nextPhysical(Node *node) {
-  auto nextChild = getFirstChild(node);
+  Node *nextChild = getFirstChild(node);
   if (nextChild != nullptr) {
     return nextChild;
   }
@@ -1742,7 +1713,7 @@ Node *nextPhysical(Node *node) {
 }
 
 Node *nextLogical(Node *node) {
-  auto nextChild = getFirstChild(node);
+  Node *nextChild = getFirstChild(node);
   if (nextChild != nullptr) {
     node = nextChild;
     goto downLeftSpine;
@@ -1760,7 +1731,7 @@ Node *nextLogical(Node *node) {
     }
   }
 downLeftSpine:
-  for (; !node->entryPresent; node = getFirstChildExists(node)) {
+  for (; !node->entryPresent; node = getFirstChild(node)) {
   }
   return node;
 }
@@ -1930,7 +1901,7 @@ void mergeWithChild(TaggedNodePointer &self, WriteContext *writeContext,
   child->parentsIndex = self->parentsIndex;
 
   // Max versions are stored in the parent, so we need to update it now
-  // that we have a new parent. Safe we call since the root never has a partial
+  // that we have a new parent. Safe to call since the root never has a partial
   // key.
   setMaxVersion(child, std::max(childMaxVersion, writeContext->zero));
 
@@ -2808,7 +2779,7 @@ bool checkRangeStartsWith(NodeT *nTyped, TrivialSpan key, int begin, int end,
   __builtin_unreachable(); // GCOVR_EXCL_LINE
 
 downLeftSpine:
-  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  for (; !n->entryPresent; n = getFirstChild(n)) {
   }
   return n->entry.rangeVersion <= readVersion;
 }
@@ -3265,7 +3236,7 @@ PRESERVE_NONE void down_left_spine(Job *job, Context *context) {
     job->setResult(n->entry.rangeVersion <= job->readVersion);
     MUSTTAIL return complete(job, context);
   }
-  auto child = getFirstChildExists(n);
+  auto child = getFirstChild(n);
   job->n = child;
   __builtin_prefetch(job->n);
   job->continuation = downLeftSpineTable[child.getType()];
@@ -3360,7 +3331,7 @@ template <class NodeT> void iter(Job *job, Context *context) {
       job->setResult(n->entry.pointVersion <= job->readVersion);
       MUSTTAIL return complete(job, context);
     }
-    auto c = getFirstChildExists(n);
+    auto c = getFirstChild(n);
     job->n = c;
     job->continuation = downLeftSpineTable[c.getType()];
     __builtin_prefetch(job->n);
@@ -3884,7 +3855,7 @@ void left_side_down_left_spine(Job *job, Context *context) {
     }
     MUSTTAIL return done_left_side_iter(job, context);
   }
-  auto c = getFirstChildExists(n);
+  auto c = getFirstChild(n);
   job->n = c;
   job->continuation = leftSideDownLeftSpineTable[c.getType()];
   __builtin_prefetch(job->n);
@@ -4547,7 +4518,7 @@ bool checkPointRead(Node *n, const TrivialSpan key,
       if (n->entryPresent) {
         return n->entry.pointVersion <= readVersion;
       }
-      n = getFirstChildExists(n);
+      n = getFirstChild(n);
       goto downLeftSpine;
     }
 
@@ -4601,7 +4572,7 @@ bool checkPointRead(Node *n, const TrivialSpan key,
     }
   }
 downLeftSpine:
-  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  for (; !n->entryPresent; n = getFirstChild(n)) {
   }
   return n->entry.rangeVersion <= readVersion;
 }
@@ -4676,7 +4647,7 @@ bool checkPrefixRead(Node *n, const TrivialSpan key,
     }
   }
 downLeftSpine:
-  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  for (; !n->entryPresent; n = getFirstChild(n)) {
   }
   return n->entry.rangeVersion <= readVersion;
 }
@@ -4763,7 +4734,7 @@ bool checkRangeLeftSide(Node *n, TrivialSpan key, int prefixLen,
     }
   }
 downLeftSpine:
-  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  for (; !n->entryPresent; n = getFirstChild(n)) {
   }
   return n->entry.rangeVersion <= readVersion;
 }
@@ -4852,14 +4823,12 @@ backtrack:
       searchPathLen -= 1 + n->partialKeyLen;
       n = n->parent;
     } else {
-      searchPathLen -= n->partialKeyLen;
       n = next;
-      searchPathLen += n->partialKeyLen;
       goto downLeftSpine;
     }
   }
 downLeftSpine:
-  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  for (; !n->entryPresent; n = getFirstChild(n)) {
   }
   return n->entry.rangeVersion <= readVersion;
 }
@@ -5330,6 +5299,11 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     gc_iterations_total.add(set_oldest_iterations_accum);
     if (n == nullptr) {
       removalKey = {};
+      if (removalBufferSize > kMaxRemovalBufferSize) {
+        safe_free(removalBuffer, removalBufferSize);
+        removalBufferSize = kMinRemovalBufferSize;
+        removalBuffer = (uint8_t *)safe_malloc(removalBufferSize);
+      }
       oldestExtantVersion = oldestVersionAtGcBegin;
       oldest_extant_version.set(oldestExtantVersion);
       oldestVersionAtGcBegin = oldestVersionFullPrecision;
@@ -5340,12 +5314,47 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
               oldestExtantVersion, oldestVersionAtGcBegin);
 #endif
     } else {
-      removalKeyArena = Arena();
+      // Store the current search path to resume the scan later
-      removalKey = getSearchPath(removalKeyArena, n);
+      saveRemovalKey(n);
     }
     return fuel;
   }
 
+  void saveRemovalKey(Node *n) {
+    uint8_t *cursor = removalBuffer + removalBufferSize;
+    int size = 0;
+    auto reserve = [&](int delta) {
+      if (size + delta > removalBufferSize) [[unlikely]] {
+        int newBufSize = std::max(removalBufferSize * 2, size + delta);
+        uint8_t *newBuf = (uint8_t *)safe_malloc(newBufSize);
+        memcpy(newBuf + newBufSize - size, cursor, size);
+        safe_free(removalBuffer, removalBufferSize);
+        removalBuffer = newBuf;
+        removalBufferSize = newBufSize;
+        cursor = newBuf + newBufSize - size;
+      }
+    };
+    for (;;) {
+      auto partialKey = TrivialSpan{n->partialKey(), n->partialKeyLen};
+      reserve(partialKey.size());
+      size += partialKey.size();
+      cursor -= partialKey.size();
+      memcpy(cursor, partialKey.data(), partialKey.size());
+
+      if (n->parent == nullptr) {
+        break;
+      }
+
+      reserve(1);
+      ++size;
+      --cursor;
+      *cursor = n->parentsIndex;
+
+      n = n->parent;
+    }
+    removalKey = {cursor, size};
+  }
+
   void setOldestVersion(int64_t newOldestVersion) {
     assert(newOldestVersion >= 0);
     assert(newOldestVersion <= newestVersionFullPrecision);
@@ -5396,7 +5405,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     writeContext.~WriteContext();
     new (&writeContext) WriteContext();
 
-    removalKeyArena = Arena{};
+    // Leave removalBuffer as is
     removalKey = {};
     keyUpdates = 10;
 
@@ -5428,11 +5437,16 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   ~Impl() {
     eraseTree(root, &writeContext);
     safe_free(metrics, metricsCount * sizeof(metrics[0]));
+    safe_free(removalBuffer, removalBufferSize);
   }
 
   WriteContext writeContext;
 
-  Arena removalKeyArena;
+  static constexpr int kMinRemovalBufferSize = 1 << 10;
+  // Eventually downsize if larger than this value
+  static constexpr int kMaxRemovalBufferSize = 1 << 16;
+  uint8_t *removalBuffer = (uint8_t *)safe_malloc(kMinRemovalBufferSize);
+  int removalBufferSize = kMinRemovalBufferSize;
   TrivialSpan removalKey;
   int64_t keyUpdates;
 
@@ -5587,7 +5601,7 @@ Node *firstGeqLogical(Node *n, const TrivialSpan key) {
       if (n->entryPresent) {
         return n;
       }
-      n = getFirstChildExists(n);
+      n = getFirstChild(n);
       goto downLeftSpine;
     }
 
@@ -5632,7 +5646,7 @@ Node *firstGeqLogical(Node *n, const TrivialSpan key) {
     }
   }
 downLeftSpine:
-  for (; !n->entryPresent; n = getFirstChildExists(n)) {
+  for (; !n->entryPresent; n = getFirstChild(n)) {
   }
   return n;
 }
@@ -5858,13 +5872,13 @@ void checkVersionsGeqOldestExtant(Node *n,
   case Type_Node0: {
   } break;
   case Type_Node3: {
-    auto *self = static_cast<Node3 *>(n);
+    [[maybe_unused]] auto *self = static_cast<Node3 *>(n);
     for (int i = 0; i < 3; ++i) {
       assert(self->childMaxVersion[i] >= oldestExtantVersion);
     }
   } break;
   case Type_Node16: {
-    auto *self = static_cast<Node16 *>(n);
+    [[maybe_unused]] auto *self = static_cast<Node16 *>(n);
     for (int i = 0; i < 16; ++i) {
       assert(self->childMaxVersion[i] >= oldestExtantVersion);
     }
@@ -5874,7 +5888,7 @@ void checkVersionsGeqOldestExtant(Node *n,
     for (int i = 0; i < 48; ++i) {
       assert(self->childMaxVersion[i] >= oldestExtantVersion);
     }
-    for (auto m : self->maxOfMax) {
+    for ([[maybe_unused]] auto m : self->maxOfMax) {
       assert(m >= oldestExtantVersion);
     }
   } break;
@@ -5883,7 +5897,7 @@ void checkVersionsGeqOldestExtant(Node *n,
     for (int i = 0; i < 256; ++i) {
       assert(self->childMaxVersion[i] >= oldestExtantVersion);
     }
-    for (auto m : self->maxOfMax) {
+    for ([[maybe_unused]] auto m : self->maxOfMax) {
       assert(m >= oldestExtantVersion);
     }
   } break;

Internal.h

@@ -18,7 +18,6 @@ using namespace weaselab;
 #include <span>
 #include <string>
 #include <thread>
-#include <unordered_set>
 #include <utility>
 
 #include <callgrind.h>
@@ -368,23 +367,6 @@ template <class T, class C = std::less<T>> auto set(Arena &arena) {
   return Set<T, C>(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, MyHash<T>, std::equal_to<T>, ArenaAlloc<T>>;
-template <class T> auto hashSet(Arena &arena) {
-  return HashSet<T>(ArenaAlloc<T>(&arena));
-}
-
 template <class T, class U>
 bool operator==(const ArenaAlloc<T> &lhs, const ArenaAlloc<U> &rhs) {
   return lhs.arena == rhs.arena;

Jenkinsfile

@@ -91,7 +91,7 @@ pipeline {
                 minio bucket: 'jenkins', credentialsId: 'jenkins-minio', excludes: '', host: 'minio.weaselab.dev', includes: 'build/*.deb,build/*.rpm,paper/*.pdf', targetFolder: '${JOB_NAME}/${BUILD_NUMBER}/${STAGE_NAME}/'
             }
         }
-        stage('Release [gcc]') {
+        stage('gcc') {
             agent {
                 dockerfile {
                     args '-v /home/jenkins/ccache:/ccache'
@@ -99,7 +99,7 @@ pipeline {
                 }
             }
             steps {
-                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_CXX_FLAGS=-DNVALGRIND")
+                CleanBuildAndTest("-DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++")
                 recordIssues(tools: [gcc()])
             }
         }

README.md

@@ -26,15 +26,15 @@ Compiler is `Ubuntu clang version 20.0.0 (++20241029082144+7544d3af0e28-1~exp1~2
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.63 |       79,186,868.18 |    1.4% |          241.61 |           64.76 |  3.731 |          31.64 |    0.8% |      0.01 | `point reads`
+|               12.83 |       77,947,334.88 |    0.8% |          247.13 |           63.80 |  3.873 |          32.64 |    0.6% |      0.01 | `point reads`
-|               14.48 |       69,078,073.40 |    0.3% |          292.42 |           74.69 |  3.915 |          41.49 |    0.5% |      0.01 | `prefix reads`
+|               14.73 |       67,908,470.74 |    0.1% |          299.99 |           73.66 |  4.073 |          42.50 |    0.5% |      0.01 | `prefix reads`
-|               34.37 |       29,094,694.11 |    0.2% |          759.53 |          179.77 |  4.225 |         100.38 |    0.2% |      0.01 | `range reads`
+|               35.63 |       28,066,165.78 |    0.1% |          782.70 |          178.49 |  4.385 |         106.65 |    0.2% |      0.01 | `range reads`
-|               19.34 |       51,713,896.36 |    0.7% |          369.70 |          101.81 |  3.631 |          47.88 |    0.6% |      0.01 | `point writes`
+|               20.00 |       49,993,123.62 |    0.1% |          376.83 |          100.50 |  3.749 |          50.05 |    0.5% |      0.01 | `point writes`
-|               39.16 |       25,538,968.61 |    0.2% |          653.16 |          206.77 |  3.159 |          89.62 |    0.8% |      0.01 | `prefix writes`
+|               38.04 |       26,287,266.49 |    0.1% |          665.86 |          191.21 |  3.482 |         100.41 |    0.4% |      0.01 | `prefix writes`
-|               40.58 |       24,642,681.12 |    4.7% |          718.44 |          216.44 |  3.319 |          99.28 |    0.6% |      0.01 | `range writes`
+|               40.48 |       24,703,557.31 |    1.3% |          732.80 |          204.36 |  3.586 |         111.26 |    0.2% |      0.01 | `range writes`
-|               78.77 |       12,694,520.69 |    3.8% |        1,395.55 |          421.73 |  3.309 |         249.81 |    0.1% |      0.01 | `monotonic increasing point writes`
+|               81.01 |       12,343,591.64 |    1.4% |        1,551.57 |          409.23 |  3.791 |         292.66 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          287,760.50 |            3,475.11 |    0.5% |    3,929,266.50 |    1,550,225.50 |  2.535 |     639,064.00 |    0.0% |      0.01 | `worst case for radix tree`
+|          315,672.00 |            3,167.85 |    1.7% |    4,043,066.00 |    1,590,315.00 |  2.542 |     714,828.00 |    0.1% |      0.01 | `worst case for radix tree`
-|              104.76 |        9,545,250.65 |    3.1% |        2,000.00 |          552.82 |  3.618 |         342.00 |    0.0% |      0.01 | `create and destroy`
+|              114.81 |        8,710,164.86 |    0.7% |        2,178.00 |          578.69 |  3.764 |         345.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
@@ -49,7 +49,7 @@ Check: 4.39702 seconds, 370.83 MB/s, Add: 4.50025 seconds, 124.583 MB/s, Gc rati
 ## radix tree
 
 ```
-Check: 0.987757 seconds, 1650.76 MB/s, Add: 1.24815 seconds, 449.186 MB/s, Gc ratio: 41.4675%, Peak idle memory: 2.02872e+06
+Check: 0.975666 seconds, 1728.24 MB/s, Add: 1.19751 seconds, 480.444 MB/s, Gc ratio: 36.8478%, Peak idle memory: 2.39447e+06
 ```
 
 ## hash table

RealDataBench.cpp

@@ -5,7 +5,7 @@
 #include <cstdio>
 #include <cstring>
 #include <fcntl.h>
-#include <string_view>
+#include <span>
 #include <sys/mman.h>
 #include <sys/stat.h>
 #include <unistd.h>
@@ -64,7 +64,7 @@ int main(int argc, const char **argv) {
     auto *const mapOriginal = begin;
     const auto sizeOriginal = size;
 
-    using StringView = std::basic_string_view<uint8_t>;
+    using StringView = std::span<const uint8_t>;
 
     StringView write;
     std::vector<StringView> reads;
@@ -78,9 +78,9 @@ int main(int argc, const char **argv) {
       end = (uint8_t *)memchr(begin, '\n', size);
 
       if (line.size() > 0 && line[0] == 'P') {
-        write = line.substr(2, line.size());
+        write = line.subspan(2, line.size());
       } else if (line.size() > 0 && line[0] == 'L') {
-        reads.push_back(line.substr(2, line.size()));
+        reads.push_back(line.subspan(2, line.size()));
       } else if (line.empty()) {
         {
           readRanges.resize(reads.size());

ServerBench.cpp

@@ -23,6 +23,97 @@
 #include "Internal.h"
 #include "third_party/nadeau.h"
 
+constexpr int kCacheLine = 64; // TODO mac m1 is 128
+
+template <class T> struct TxQueue {
+
+  explicit TxQueue(int lgSlotCount)
+      : slotCount(1 << lgSlotCount), slotCountMask(slotCount - 1),
+        slots(new T[slotCount]) {
+    // Otherwise we can't tell the difference between full and empty.
+    assert(!(slotCountMask & 0x80000000));
+  }
+
+  /// Call from producer thread, after ensuring consumer is no longer accessing
+  /// it somehow
+  ~TxQueue() { delete[] slots; }
+
+  /// Must be called from the producer thread
+  void push(T t) {
+    if (wouldBlock()) {
+      // Wait for pops to change and try again
+      consumer.pops.wait(producer.lastPopRead, std::memory_order_relaxed);
+      producer.lastPopRead = consumer.pops.load(std::memory_order_acquire);
+    }
+    slots[producer.pushesNonAtomic++ & slotCountMask] = std::move(t);
+    // seq_cst so that the notify can't be ordered before the store
+    producer.pushes.store(producer.pushesNonAtomic, std::memory_order_seq_cst);
+    // We have to notify every time, since we don't know if this is the last
+    // push ever
+    producer.pushes.notify_one();
+  }
+
+  /// Must be called from the producer thread
+  uint32_t outstanding() {
+    return producer.pushesNonAtomic -
+           consumer.pops.load(std::memory_order_relaxed);
+  }
+
+  /// Returns true if a call to push might block. Must be called from the
+  /// producer thread.
+  bool wouldBlock() {
+    // See if we can determine that overflow won't happen entirely from state
+    // local to the producer
+    if (producer.pushesNonAtomic - producer.lastPopRead == slotCount - 1) {
+      // Re-read pops with memory order
+      producer.lastPopRead = consumer.pops.load(std::memory_order_acquire);
+      return producer.pushesNonAtomic - producer.lastPopRead == slotCount - 1;
+    }
+    return false;
+  }
+
+  /// Valid until the next pop, or until this queue is destroyed.
+  T *pop() {
+    // See if we can determine that there's an entry we can pop entirely from
+    // state local to the consumer
+    if (consumer.lastPushRead - consumer.popsNonAtomic == 0) {
+      // Re-read pushes with memory order and try again
+      consumer.lastPushRead = producer.pushes.load(std::memory_order_acquire);
+      if (consumer.lastPushRead - consumer.popsNonAtomic == 0) {
+        // Wait for pushes to change and try again
+        producer.pushes.wait(consumer.lastPushRead, std::memory_order_relaxed);
+        consumer.lastPushRead = producer.pushes.load(std::memory_order_acquire);
+      }
+    }
+    auto result = &slots[consumer.popsNonAtomic++ & slotCountMask];
+    // We only have to write pops with memory order if we've run out of items.
+    // We know that we'll eventually run out.
+    if (consumer.lastPushRead - consumer.popsNonAtomic == 0) {
+      // seq_cst so that the notify can't be ordered before the store
+      consumer.pops.store(consumer.popsNonAtomic, std::memory_order_seq_cst);
+      consumer.pops.notify_one();
+    }
+    return result;
+  }
+
+private:
+  const uint32_t slotCount;
+  const uint32_t slotCountMask;
+  T *slots;
+  struct alignas(kCacheLine) ProducerState {
+    std::atomic<uint32_t> pushes{0};
+    uint32_t pushesNonAtomic{0};
+    uint32_t lastPopRead{0};
+  };
+  struct alignas(kCacheLine) ConsumerState {
+    std::atomic<uint32_t> pops{0};
+    uint32_t popsNonAtomic{0};
+    uint32_t lastPushRead{0};
+  };
+  ProducerState producer;
+  ConsumerState consumer;
+};
+
 std::atomic<int64_t> transactions;
 
 int64_t safeUnaryMinus(int64_t x) {
@@ -47,6 +138,7 @@ void tupleAppend(std::string &output, int64_t value) {
 
 void tupleAppend(std::string &output, std::string_view value) {
   output.push_back('\x02');
+  if (memchr(value.data(), '\x00', value.size()) != nullptr) {
     for (auto c : value) {
       if (c == '\x00') {
         output.push_back('\x00');
@@ -55,6 +147,9 @@ void tupleAppend(std::string &output, std::string_view value) {
         output.push_back(c);
       }
     }
+  } else {
+    output.insert(output.end(), value.begin(), value.end());
+  }
   output.push_back('\x00');
 }
 
@@ -66,50 +161,70 @@ template <class... Ts> std::string tupleKey(const Ts &...ts) {
 
 constexpr int kTotalKeyRange = 1'000'000'000;
 constexpr int kWindowSize = 1'000'000;
-constexpr int kNumKeys = 10;
+constexpr int kNumReadKeysPerTx = 10;
+constexpr int kNumWriteKeysPerTx = 5;
 
-void workload(weaselab::ConflictSet *cs) {
+struct Transaction {
-  int64_t version = kWindowSize;
+  std::vector<std::string> keys;
-  for (;; transactions.fetch_add(1, std::memory_order_relaxed)) {
+  std::vector<weaselab::ConflictSet::ReadRange> reads;
+  std::vector<weaselab::ConflictSet::WriteRange> writes;
+  int64_t version;
+  int64_t oldestVersion;
+  Transaction() = default;
+  explicit Transaction(int64_t version)
+      : version(version), oldestVersion(version - kWindowSize) {
     std::vector<int64_t> keyIndices;
-    for (int i = 0; i < kNumKeys; ++i) {
+    for (int i = 0; i < std::max(kNumReadKeysPerTx, kNumWriteKeysPerTx); ++i) {
       keyIndices.push_back(rand() % kTotalKeyRange);
     }
     std::sort(keyIndices.begin(), keyIndices.end());
-    std::vector<std::string> keys;
     constexpr std::string_view fullString =
         "this is a string, where a prefix of it is used as an element of the "
         "tuple forming the key";
-    for (int i = 0; i < kNumKeys; ++i) {
+    for (int i = 0; i < int(keyIndices.size()); ++i) {
       keys.push_back(
           tupleKey(0x100, keyIndices[i] / fullString.size(),
                    fullString.substr(0, keyIndices[i] % fullString.size())));
       // printf("%s\n", printable(keys.back()).c_str());
     }
-
+    for (int i = 0; i < kNumWriteKeysPerTx; ++i) {
-    std::vector<weaselab::ConflictSet::ReadRange> reads;
-    std::vector<weaselab::ConflictSet::WriteRange> writes;
-    std::vector<weaselab::ConflictSet::Result> results;
-    for (int i = 0; i < kNumKeys; ++i) {
       writes.push_back({{(const uint8_t *)keys[i].data(), int(keys[i].size())},
                         {nullptr, 0}});
+    }
+    reads.push_back({{(const uint8_t *)keys[0].data(), int(keys[0].size())},
+                     {(const uint8_t *)keys[1].data(), int(keys[1].size())},
+                     version - std::min(10, kWindowSize)});
+    static_assert(kNumReadKeysPerTx >= 3);
+    for (int i = 2; i < kNumReadKeysPerTx; ++i) {
       reads.push_back({{(const uint8_t *)keys[i].data(), int(keys[i].size())},
                        {nullptr, 0},
                        version - kWindowSize});
     }
-    results.resize(reads.size());
+  }
 
-    cs->check(reads.data(), results.data(), reads.size());
+  Transaction(Transaction &&) = default;
-    bool ok = true;
+  Transaction &operator=(Transaction &&) = default;
-    for (auto result : results) {
+  Transaction(Transaction const &) = delete;
-      ok &= result == weaselab::ConflictSet::Commit;
+  Transaction const &operator=(Transaction const &) = delete;
-    }
+};
-    cs->addWrites(writes.data(), ok ? writes.size() : 0, version);
+
-    cs->setOldestVersion(version - kWindowSize);
+struct Resolver {
-    ++version;
+
-  }
+  void resolve(const weaselab::ConflictSet::ReadRange *reads, int readCount,
+               const weaselab::ConflictSet::WriteRange *writes, int writeCount,
+               int64_t newVersion, int64_t newOldestVersion) {
+    results.resize(readCount);
+    cs.check(reads, results.data(), readCount);
+    cs.addWrites(writes, writeCount, newVersion);
+    cs.setOldestVersion(newOldestVersion);
   }
 
+  ConflictSet cs{0};
+
+private:
+  std::vector<weaselab::ConflictSet::Result> results;
+};
+
 // Adapted from getaddrinfo man page
 int getListenFd(const char *node, const char *service) {
 
@@ -251,7 +366,8 @@ int main(int argc, char **argv) {
   {
     int listenFd = getListenFd(argv[1], argv[2]);
 
-    weaselab::ConflictSet cs{0};
+    Resolver resolver;
+    auto &cs = resolver.cs;
     weaselab::ConflictSet::MetricsV1 *metrics;
     int metricsCount;
     cs.getMetricsV1(&metrics, &metricsCount);
@@ -300,7 +416,22 @@ int main(int argc, char **argv) {
     }
 #endif
 
-    auto w = std::thread{workload, &cs};
+    TxQueue<Transaction> queue{10};
+
+    auto workloadThread = std::thread{[&]() {
+      for (int64_t version = kWindowSize;;
+           ++version, transactions.fetch_add(1, std::memory_order_relaxed)) {
+        queue.push(Transaction(version));
+      }
+    }};
+
+    auto resolverThread = std::thread{[&]() {
+      for (;;) {
+        auto tx = queue.pop();
+        resolver.resolve(tx->reads.data(), tx->reads.size(), tx->writes.data(),
+                         tx->writes.size(), tx->version, tx->oldestVersion);
+      }
+    }};
 
     for (;;) {
       struct sockaddr_storage peer_addr = {};

SkipList.cpp

@@ -796,7 +796,6 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     int temp[stripeSize];
     int stripes = (stringCount + stripeSize - 1) / stripeSize;
     StringRef values[stripeSize];
-    int64_t writeVersions[stripeSize / 2];
     int ss = stringCount - (stripes - 1) * stripeSize;
     int64_t entryDelta = 0;
     for (int s = stripes - 1; s >= 0; s--) {

apple-symbol-imports.txt

@@ -1,3 +1,4 @@
+___chkstk_darwin
 ___stack_chk_fail
 ___stack_chk_guard
 __tlv_bootstrap
@@ -5,6 +6,7 @@ _abort
 _bzero
 _free
 _malloc
+_memcmp
 _memcpy
 _memmove
 dyld_stub_binder

package_macos.sh

@@ -8,7 +8,7 @@ SRC_DIR="${0%/*}"
 BUILD_ARM="$(mktemp -d -t conflict-set-arm)"
 BUILD_X86="$(mktemp -d -t conflict-set-x86)"
 
-cmake_args=(-DCMAKE_CXX_FLAGS=-DNVALGRIND -DCPACK_PACKAGING_INSTALL_PREFIX=/usr/local)
+cmake_args=(-DCMAKE_CXX_FLAGS=-DNVALGRIND -DCPACK_PACKAGING_INSTALL_PREFIX=/usr/local -DCMAKE_CXX_COMPILER=/opt/homebrew/opt/llvm/bin/clang++)
 
 cmake -S"$SRC_DIR" -B"$BUILD_ARM" -DCMAKE_OSX_ARCHITECTURES=arm64 "${cmake_args[@]}"
 cmake --build "$BUILD_ARM" --target conflict-set --target conflict-set-static

test_conflict_set.py

@@ -1,3 +1,4 @@
+import struct
 from conflict_set import *
 
 
@@ -164,6 +165,16 @@ def test_fixup_256():
         cs.check(read(0, bytes([1]), bytes([2])))
 
 
+def test_large_removal_buffer():
+    with DebugConflictSet() as cs:
+        for i in range(1000):
+            # create extra gc work
+            for j in range(100):
+                cs.addWrites(1000 + i)
+            cs.addWrites(1000 + i, write(struct.pack(">l", i) + bytes([0] * 100000)))
+            cs.setOldestVersion(i)
+
+
 if __name__ == "__main__":
     # budget "pytest" for ctest integration without pulling in a dependency. You can of course still use pytest in local development.
     import argparse