Include childType in ChildAndMaxVersion

CMakeLists.txt

@@ -248,6 +248,19 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
     add_test(NAME conflict_set_blackbox_${hash} COMMAND driver ${TEST})
   endforeach()
 
+  find_program(VALGRIND_EXE valgrind)
+  if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
+    list(LENGTH CORPUS_TESTS len)
+    math(EXPR last "${len} - 1")
+    set(partition_size 100)
+    foreach(i RANGE 0 ${last} ${partition_size})
+      list(SUBLIST CORPUS_TESTS ${i} ${partition_size} partition)
+      add_test(NAME conflict_set_blackbox_valgrind_${i}
+               COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
+                       $<TARGET_FILE:driver> ${partition})
+    endforeach()
+  endif()
+
   # scripted tests. Written manually to fill in anything libfuzzer couldn't
   # find.
   if(NOT CMAKE_CROSSCOMPILING)
@@ -268,19 +281,14 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
           ${Python3_EXECUTABLE}
           ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
           --build-dir ${CMAKE_CURRENT_BINARY_DIR})
-    endforeach()
-  endif()
-
-  find_program(VALGRIND_EXE valgrind)
-  if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
-    list(LENGTH CORPUS_TESTS len)
-    math(EXPR last "${len} - 1")
-    set(partition_size 100)
-    foreach(i RANGE 0 ${last} ${partition_size})
-      list(SUBLIST CORPUS_TESTS ${i} ${partition_size} partition)
-      add_test(NAME conflict_set_blackbox_valgrind_${i}
-               COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
-                       $<TARGET_FILE:driver> ${partition})
+      if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
+        add_test(
+          NAME script_test_${TEST}_valgrind
+          COMMAND
+            ${VALGRIND_EXE} ${Python3_EXECUTABLE}
+            ${CMAKE_CURRENT_SOURCE_DIR}/test_conflict_set.py test ${TEST}
+            --build-dir ${CMAKE_CURRENT_BINARY_DIR})
+      endif()
     endforeach()
   endif()
 
@@ -350,6 +358,11 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   set_target_properties(server_bench PROPERTIES SKIP_BUILD_RPATH ON)
 
   add_executable(interleaving_test InterleavingTest.cpp)
+  # work around lack of musttail for gcc
+  if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_BUILD_TYPE STREQUAL "Debug")
+    target_compile_options(interleaving_test PRIVATE -Og
+                                                     -foptimize-sibling-calls)
+  endif()
   target_link_libraries(interleaving_test PRIVATE nanobench)
 endif()
 

ConflictSet.cpp

@@ -203,6 +203,60 @@ enum Type : int8_t {
 
 template <class T> struct BoundedFreeListAllocator;
 
+struct TaggedNodePointer {
+  TaggedNodePointer() = default;
+  operator struct Node *() { return (struct Node *)(uintptr_t)p; }
+  operator struct Node0 *() {
+    assert(t == Type_Node0);
+    return (struct Node0 *)(uintptr_t)p;
+  }
+  operator struct Node3 *() {
+    assert(t == Type_Node3);
+    return (struct Node3 *)(uintptr_t)p;
+  }
+  operator struct Node16 *() {
+    assert(t == Type_Node16);
+    return (struct Node16 *)(uintptr_t)p;
+  }
+  operator struct Node48 *() {
+    assert(t == Type_Node48);
+    return (struct Node48 *)(uintptr_t)p;
+  }
+  operator struct Node256 *() {
+    assert(t == Type_Node256);
+    return (struct Node256 *)(uintptr_t)p;
+  }
+  /*implicit*/ TaggedNodePointer(std::nullptr_t) : p(0), t(Type_Node0) {}
+  /*implicit*/ TaggedNodePointer(Node0 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node0) {}
+  /*implicit*/ TaggedNodePointer(Node3 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node3) {}
+  /*implicit*/ TaggedNodePointer(Node16 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node16) {}
+  /*implicit*/ TaggedNodePointer(Node48 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node48) {}
+  /*implicit*/ TaggedNodePointer(Node256 *x)
+      : TaggedNodePointer((struct Node *)x, Type_Node256) {}
+
+  bool operator!=(std::nullptr_t) { return p != 0 || t != Type_Node0; }
+  bool operator==(std::nullptr_t) { return p == 0 && t == Type_Node0; }
+  bool operator==(const TaggedNodePointer &) const = default;
+  bool operator==(Node *n) const { return (uintptr_t)n == p; }
+  Node *operator->() { return (Node *)(uintptr_t)p; }
+  Type getType() { return t; }
+
+  TaggedNodePointer(const TaggedNodePointer &) = default;
+  TaggedNodePointer &operator=(const TaggedNodePointer &) = default;
+  /*implicit*/ TaggedNodePointer(Node *n);
+
+private:
+  TaggedNodePointer(struct Node *p, Type t) : p((uintptr_t)p), t(t) {
+    assume(p != 0);
+  }
+  uintptr_t p : 56;
+  Type t : 8;
+};
+
 struct Node {
 
   /* begin section that's copied to the next node */
@@ -228,6 +282,9 @@ private:
   int32_t partialKeyCapacity;
 };
 
+TaggedNodePointer::TaggedNodePointer(Node *n)
+    : TaggedNodePointer(n, n->getType()) {}
+
 constexpr int kNodeCopyBegin = offsetof(Node, entry);
 constexpr int kNodeCopySize =
     offsetof(Node, parentsIndex) + sizeof(Node::parentsIndex) - kNodeCopyBegin;
@@ -251,7 +308,7 @@ struct Node3 : Node {
   constexpr static auto kMaxNodes = 3;
   constexpr static auto kType = Type_Node3;
 
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   // Sorted
   uint8_t index[kMaxNodes];
@@ -267,7 +324,7 @@ struct Node16 : Node {
   constexpr static auto kType = Type_Node16;
   constexpr static auto kMaxNodes = 16;
 
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   // Sorted
   uint8_t index[kMaxNodes];
@@ -288,7 +345,7 @@ struct Node48 : Node {
   constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
 
   BitSet bitSet;
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
   uint8_t reverseIndex[kMaxNodes];
@@ -310,7 +367,7 @@ struct Node256 : Node {
   constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
 
   BitSet bitSet;
-  Node *children[kMaxNodes];
+  TaggedNodePointer children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
 
@@ -888,26 +945,26 @@ int getNodeIndexExists(Node16 *self, uint8_t index) {
 }
 
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node3 *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node3 *self, uint8_t index) {
   return self->children[getNodeIndexExists(self, index)];
 }
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node16 *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node16 *self, uint8_t index) {
   return self->children[getNodeIndexExists(self, index)];
 }
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node48 *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node48 *self, uint8_t index) {
   assert(self->bitSet.test(index));
   return self->children[self->index[index]];
 }
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node256 *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node256 *self, uint8_t index) {
   assert(self->bitSet.test(index));
   return self->children[index];
 }
 
 // Precondition - an entry for index must exist in the node
-Node *&getChildExists(Node *self, uint8_t index) {
+TaggedNodePointer &getChildExists(Node *self, uint8_t index) {
   switch (self->getType()) {
   case Type_Node0:           // GCOVR_EXCL_LINE
     __builtin_unreachable(); // GCOVR_EXCL_LINE
@@ -1038,7 +1095,7 @@ void setMaxVersion(Node *n, InternalVersionT newMax) {
   }
 }
 
-Node *&getInTree(Node *n, ConflictSet::Impl *);
+TaggedNodePointer &getInTree(Node *n, ConflictSet::Impl *);
 
 Node *getChild(Node0 *, uint8_t) { return nullptr; }
 Node *getChild(Node3 *self, uint8_t index) {
@@ -1075,6 +1132,7 @@ Node *getChild(Node *self, uint8_t index) {
 struct ChildAndMaxVersion {
   Node *child;
   InternalVersionT maxVersion;
+  Type childType;
 };
 
 ChildAndMaxVersion getChildAndMaxVersion(Node0 *, uint8_t) { return {}; }
@@ -1083,24 +1141,28 @@ ChildAndMaxVersion getChildAndMaxVersion(Node3 *self, uint8_t index) {
   if (i < 0) {
     return {};
   }
-  return {self->children[i], self->childMaxVersion[i]};
+  return {self->children[i], self->childMaxVersion[i],
+          self->children[i].getType()};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node16 *self, uint8_t index) {
   int i = getNodeIndex(self, index);
   if (i < 0) {
     return {};
   }
-  return {self->children[i], self->childMaxVersion[i]};
+  return {self->children[i], self->childMaxVersion[i],
+          self->children[i].getType()};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node48 *self, uint8_t index) {
   int i = self->index[index];
   if (i < 0) {
     return {};
   }
-  return {self->children[i], self->childMaxVersion[i]};
+  return {self->children[i], self->childMaxVersion[i],
+          self->children[i].getType()};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node256 *self, uint8_t index) {
-  return {self->children[index], self->childMaxVersion[index]};
+  return {self->children[index], self->childMaxVersion[index],
+          self->children[index].getType()};
 }
 
 ChildAndMaxVersion getChildAndMaxVersion(Node *self, uint8_t index) {
@@ -1245,14 +1307,15 @@ Node *getFirstChildExists(Node *self) {
   }
 }
 
-void consumePartialKeyFull(Node *&self, std::span<const uint8_t> &key,
+void consumePartialKeyFull(TaggedNodePointer &self,
+                           std::span<const uint8_t> &key,
                            InternalVersionT writeVersion, WriteContext *tls) {
   // Handle an existing partial key
   int commonLen = std::min<int>(self->partialKeyLen, key.size());
   int partialKeyIndex =
       longestCommonPrefix(self->partialKey(), key.data(), commonLen);
   if (partialKeyIndex < self->partialKeyLen) {
-    auto *old = self;
+    Node *old = self;
     // Since root cannot have a partial key
     assert(old->parent != nullptr);
     InternalVersionT oldMaxVersion = exchangeMaxVersion(old, writeVersion);
@@ -1290,7 +1353,7 @@ void consumePartialKeyFull(Node *&self, std::span<const uint8_t> &key,
 // Consume any partial key of `self`, and update `self` and
 // `key` such that `self` is along the search path of `key`
 inline __attribute__((always_inline)) void
-consumePartialKey(Node *&self, std::span<const uint8_t> &key,
+consumePartialKey(TaggedNodePointer &self, std::span<const uint8_t> &key,
                   InternalVersionT writeVersion, WriteContext *tls) {
   if (self->partialKeyLen > 0) {
     consumePartialKeyFull(self, key, writeVersion, tls);
@@ -1301,8 +1364,10 @@ consumePartialKey(Node *&self, std::span<const uint8_t> &key,
 // such that the search path of the result + key is the same as the search path
 // of self + key before the call. Creates a node if necessary. Updates
 // `maxVersion` for result.
-Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
-                        InternalVersionT newMaxVersion, WriteContext *tls) {
+TaggedNodePointer &getOrCreateChild(TaggedNodePointer &self,
+                                    std::span<const uint8_t> &key,
+                                    InternalVersionT newMaxVersion,
+                                    WriteContext *tls) {
 
   int index = key.front();
   key = key.subspan(1, key.size() - 1);
@@ -1660,10 +1725,11 @@ void rezero(Node *n, InternalVersionT z) {
 }
 #endif
 
-void mergeWithChild(Node *&self, WriteContext *tls, ConflictSet::Impl *impl,
-                    Node *&dontInvalidate, Node3 *self3) {
+void mergeWithChild(TaggedNodePointer &self, WriteContext *tls,
+                    ConflictSet::Impl *impl, Node *&dontInvalidate,
+                    Node3 *self3) {
   assert(!self3->entryPresent);
-  auto *child = self3->children[0];
+  Node *child = self3->children[0];
   int minCapacity = self3->partialKeyLen + 1 + child->partialKeyLen;
 
   if (minCapacity > child->getCapacity()) {
@@ -1912,7 +1978,7 @@ bool checkPointRead(Node *n, const std::span<const uint8_t> key,
       goto downLeftSpine;
     }
 
-    auto [child, maxV] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, maxV, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
@@ -1982,7 +2048,7 @@ bool checkPrefixRead(Node *n, const std::span<const uint8_t> key,
       return maxVersion(n) <= readVersion;
     }
 
-    auto [child, maxV] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, maxV, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
@@ -2255,7 +2321,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
     if (!mask) {
       return true;
     }
-    auto *child = self->children[std::countr_zero(mask)];
+    Node *child = self->children[std::countr_zero(mask)];
     const bool firstRangeOk =
         !child->entryPresent || child->entry.rangeVersion <= readVersion;
     uint32_t compared = 0;
@@ -2330,7 +2396,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
     if (!mask) {
       return true;
     }
-    auto *child = self->children[std::countr_zero(mask)];
+    Node *child = self->children[std::countr_zero(mask)];
     const bool firstRangeOk =
         !child->entryPresent || child->entry.rangeVersion <= readVersion;
 
@@ -2375,7 +2441,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
     {
       int c = self->bitSet.firstSetGeq(begin + 1);
       if (c >= 0 && c < end) {
-        auto *child = self->children[self->index[c]];
+        Node *child = self->children[self->index[c]];
         if (child->entryPresent && child->entry.rangeVersion > readVersion) {
           return false;
         }
@@ -2409,7 +2475,7 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
     {
       int c = self->bitSet.firstSetGeq(begin + 1);
       if (c >= 0 && c < end) {
-        auto *child = self->children[c];
+        Node *child = self->children[c];
         if (child->entryPresent && child->entry.rangeVersion > readVersion) {
           return false;
         }
@@ -2581,7 +2647,7 @@ bool checkRangeLeftSide(Node *n, std::span<const uint8_t> key, int prefixLen,
       }
     }
 
-    auto [child, maxV] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, maxV, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       auto c = getChildGeq(n, remaining[0]);
       if (c != nullptr) {
@@ -2769,7 +2835,7 @@ bool checkRangeRead(Node *n, std::span<const uint8_t> begin,
     if (remaining.size() == 0) {
       break;
     }
-    auto [child, v] = getChildAndMaxVersion(n, remaining[0]);
+    auto [child, v, childT] = getChildAndMaxVersion(n, remaining[0]);
     if (child == nullptr) {
       break;
     }
@@ -2835,8 +2901,10 @@ checkMaxBetweenExclusiveImpl<true>(Node *n, int begin, int end,
 // of the result will have `maxVersion` set to `writeVersion` as a
 // postcondition. Nodes along the search path may be invalidated. Callers must
 // ensure that the max version of the self argument is updated.
-[[nodiscard]] Node **insert(Node **self, std::span<const uint8_t> key,
-                            InternalVersionT writeVersion, WriteContext *tls) {
+[[nodiscard]] TaggedNodePointer *insert(TaggedNodePointer *self,
+                                        std::span<const uint8_t> key,
+                                        InternalVersionT writeVersion,
+                                        WriteContext *tls) {
 
   for (; key.size() != 0; ++tls->accum.insert_iterations) {
     self = &getOrCreateChild(*self, key, writeVersion, tls);
@@ -2864,17 +2932,23 @@ void eraseTree(Node *root, WriteContext *tls) {
     } break;
     case Type_Node3: {
       auto *n3 = static_cast<Node3 *>(n);
-      toFree.append(std::span<Node *>(n3->children, n3->numChildren));
+      for (int i = 0; i < n3->numChildren; ++i) {
+        toFree.push_back(n3->children[i]);
+      }
       tls->release(n3);
     } break;
     case Type_Node16: {
       auto *n16 = static_cast<Node16 *>(n);
-      toFree.append(std::span<Node *>(n16->children, n16->numChildren));
+      for (int i = 0; i < n16->numChildren; ++i) {
+        toFree.push_back(n16->children[i]);
+      }
       tls->release(n16);
     } break;
     case Type_Node48: {
       auto *n48 = static_cast<Node48 *>(n);
-      toFree.append(std::span<Node *>(n48->children, n48->numChildren));
+      for (int i = 0; i < n48->numChildren; ++i) {
+        toFree.push_back(n48->children[i]);
+      }
       tls->release(n48);
     } break;
     case Type_Node256: {
@@ -2890,10 +2964,10 @@ void eraseTree(Node *root, WriteContext *tls) {
   }
 }
 
-void addPointWrite(Node *&root, std::span<const uint8_t> key,
+void addPointWrite(TaggedNodePointer &root, std::span<const uint8_t> key,
                    InternalVersionT writeVersion, WriteContext *tls) {
   ++tls->accum.point_writes;
-  auto *n = *insert(&root, key, writeVersion, tls);
+  auto n = *insert(&root, key, writeVersion, tls);
   if (!n->entryPresent) {
     ++tls->accum.entries_inserted;
     auto *p = nextLogical(n);
@@ -2911,8 +2985,8 @@ void addPointWrite(Node *&root, std::span<const uint8_t> key,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-__attribute__((target("avx512f"))) InternalVersionT
-horizontalMaxUpTo16(InternalVersionT *vs, InternalVersionT z, int len) {
+__attribute__((target("avx512f"))) InternalVersionT horizontalMaxUpTo16(
+    InternalVersionT *vs, [[maybe_unused]] InternalVersionT z, int len) {
   assume(len <= 16);
 #if USE_64_BIT
   // Hope it gets vectorized
@@ -2948,7 +3022,7 @@ horizontalMaxUpTo16(InternalVersionT *vs, InternalVersionT, int len) {
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
 __attribute__((target("avx512f"))) InternalVersionT
-horizontalMax16(InternalVersionT *vs, InternalVersionT z) {
+horizontalMax16(InternalVersionT *vs, [[maybe_unused]] InternalVersionT z) {
 #if USE_64_BIT
   // Hope it gets vectorized
   InternalVersionT max = vs[0];
@@ -2957,7 +3031,7 @@ horizontalMax16(InternalVersionT *vs, InternalVersionT z) {
   }
   return max;
 #else
-  uint32_t zero;
+  uint32_t zero; // GCOVR_EXCL_LINE
   memcpy(&zero, &z, sizeof(zero));
   auto zeroVec = _mm512_set1_epi32(zero);
   return InternalVersionT(zero + _mm512_reduce_max_epi32(_mm512_sub_epi32(
@@ -3011,7 +3085,7 @@ void fixupMaxVersion(Node *node, WriteContext *tls) {
   setMaxVersion(node, max);
 }
 
-void addWriteRange(Node *&root, std::span<const uint8_t> begin,
+void addWriteRange(TaggedNodePointer &root, std::span<const uint8_t> begin,
                    std::span<const uint8_t> end, InternalVersionT writeVersion,
                    WriteContext *tls, ConflictSet::Impl *impl) {
 
@@ -3024,12 +3098,12 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   ++tls->accum.range_writes;
   const bool beginIsPrefix = lcp == int(begin.size());
 
-  Node **useAsRoot = insert(&root, begin.subspan(0, lcp), writeVersion, tls);
+  auto useAsRoot = insert(&root, begin.subspan(0, lcp), writeVersion, tls);
 
   begin = begin.subspan(lcp, begin.size() - lcp);
   end = end.subspan(lcp, end.size() - lcp);
 
-  auto *beginNode = *insert(useAsRoot, begin, writeVersion, tls);
+  Node *beginNode = *insert(useAsRoot, begin, writeVersion, tls);
   addKey(beginNode);
   if (!beginNode->entryPresent) {
     ++tls->accum.entries_inserted;
@@ -3040,7 +3114,7 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   }
   beginNode->entry.pointVersion = writeVersion;
 
-  auto *endNode = *insert(useAsRoot, end, writeVersion, tls);
+  Node *endNode = *insert(useAsRoot, end, writeVersion, tls);
   addKey(endNode);
   if (!endNode->entryPresent) {
     ++tls->accum.entries_inserted;
@@ -3060,7 +3134,7 @@ void addWriteRange(Node *&root, std::span<const uint8_t> begin,
   assert(!beginNode->endOfRange);
   assert(!endNode->endOfRange);
   endNode->endOfRange = true;
-  auto *iter = beginNode;
+  Node *iter = beginNode;
   for (iter = nextLogical(iter); !iter->endOfRange;
        iter = erase(iter, tls, impl, /*logical*/ true)) {
     assert(!iter->endOfRange);
@@ -3375,7 +3449,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   std::span<const uint8_t> removalKey;
   int64_t keyUpdates;
 
-  Node *root;
+  TaggedNodePointer root;
   InternalVersionT oldestVersion;
   int64_t oldestVersionFullPrecision;
   int64_t oldestExtantVersion;
@@ -3456,7 +3530,7 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 };
 
-Node *&getInTree(Node *n, ConflictSet::Impl *impl) {
+TaggedNodePointer &getInTree(Node *n, ConflictSet::Impl *impl) {
   return n->parent == nullptr ? impl->root
                               : getChildExists(n->parent, n->parentsIndex);
 }

InterleavingTest.cpp

@@ -22,9 +22,6 @@ void *stepJob(Job *j) {
   return done ? nullptr : (void *)stepJob;
 }
 
-// So we can look at the disassembly more easily
-
-extern "C" {
 void sequential(Job **jobs, int count) {
   for (int i = 0; i < count; ++i) {
     do {
@@ -94,6 +91,87 @@ void interleaveBoundedCyclicList(Job **jobs, int count) {
   }
 }
 
+#ifndef __has_attribute
+#define __has_attribute(x) 0
+#endif
+
+#if __has_attribute(musttail)
+#define MUSTTAIL __attribute__((musttail))
+#else
+#define MUSTTAIL
+#endif
+
+struct Context {
+  constexpr static int kConcurrent = 32;
+  Job **jobs;
+  Job *inProgress[kConcurrent];
+  void (*continuation[kConcurrent])(Context *, int64_t prevJob, int64_t job,
+                                    int64_t started, int64_t count);
+  int nextJob[kConcurrent];
+};
+
+void keepGoing(Context *context, int64_t prevJob, int64_t job, int64_t started,
+               int64_t count) {
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->continuation[job](context, prevJob, job, started,
+                                             count);
+}
+
+void stepJobTailCall(Context *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count);
+
+void complete(Context *context, int64_t prevJob, int64_t job, int64_t started,
+              int64_t count) {
+  if (started == count) {
+    if (prevJob == job) {
+      return;
+    }
+    context->nextJob[prevJob] = context->nextJob[job];
+    job = prevJob;
+  } else {
+    context->inProgress[job] = context->jobs[started++];
+    context->continuation[job] = stepJobTailCall;
+  }
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->continuation[job](context, prevJob, job, started,
+                                             count);
+}
+
+void stepJobTailCall(Context *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count) {
+  auto *j = context->inProgress[job];
+  auto done = --(*j->input) == 0;
+#ifdef __x86_64__
+  _mm_clflush(j->input);
+#endif
+  if (done) {
+    MUSTTAIL return complete(context, prevJob, job, started, count);
+  } else {
+    context->continuation[job] = stepJobTailCall;
+    MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+  }
+}
+
+void useTailCalls(Job **jobs, int count) {
+  if (count == 0) {
+    return;
+  }
+  Context context;
+  context.jobs = jobs;
+  int64_t started = std::min(Context::kConcurrent, count);
+  for (int i = 0; i < started; i++) {
+    context.inProgress[i] = jobs[i];
+    context.nextJob[i] = i + 1;
+    context.continuation[i] = stepJobTailCall;
+  }
+  context.nextJob[started - 1] = 0;
+  int prevJob = started - 1;
+  int job = 0;
+  return context.continuation[job](&context, prevJob, job, started, count);
+}
+
 void interleaveCyclicList(Job **jobs, int count) {
   auto *nextJob = (int *)alloca(sizeof(int) * count);
 
@@ -117,12 +195,11 @@ void interleaveCyclicList(Job **jobs, int count) {
     job = nextJob[job];
   }
 }
-}
 
 int main() {
   ankerl::nanobench::Bench bench;
 
-  constexpr int kNumJobs = 100;
+  constexpr int kNumJobs = 10000;
   bench.relative(true);
 
   Job jobs[kNumJobs];
@@ -140,6 +217,7 @@ int main() {
   for (auto [scheduler, name] :
        {std::make_pair(sequentialNoFuncPtr, "sequentialNoFuncPtr"),
         std::make_pair(sequential, "sequential"),
+        std::make_pair(useTailCalls, "useTailCalls"),
         std::make_pair(interleaveSwapping, "interleavingSwapping"),
         std::make_pair(interleaveBoundedCyclicList,
                        "interleaveBoundedCyclicList"),