Update GCOVR annotations now that jenkins agent has avx512

CMakeLists.txt

@@ -276,9 +276,15 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
   find_program(VALGRIND_EXE valgrind)
   if(VALGRIND_EXE AND NOT CMAKE_CROSSCOMPILING)
-    add_test(NAME conflict_set_blackbox_valgrind
-             COMMAND ${VALGRIND_EXE} --error-exitcode=99 --
-                     $<TARGET_FILE:driver> ${CORPUS_TESTS})
+    list(LENGTH CORPUS_TESTS len)
+    math(EXPR last "${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()
 
   # api smoke tests

ConflictSet.cpp

@@ -197,7 +197,6 @@ struct Node {
   /* end section that's copied to the next node */
 
   uint8_t *partialKey();
-
   Type getType() const { return type; }
   int32_t getCapacity() const { return partialKeyCapacity; }
 
@@ -221,84 +220,83 @@ constexpr int kNodeCopySize =
 
 struct Node0 : Node {
   constexpr static auto kType = Type_Node0;
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const struct Node3 &other);
-
   size_t size() const { return sizeof(Node0) + getCapacity(); }
 };
 
 struct Node3 : Node {
   constexpr static auto kMaxNodes = 3;
   constexpr static auto kType = Type_Node3;
-  // Sorted
-  uint8_t index[kMaxNodes];
+
   Node *children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
+  // Sorted
+  uint8_t index[kMaxNodes];
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node0 &other);
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const struct Node16 &other);
-
   size_t size() const { return sizeof(Node3) + getCapacity(); }
 };
 
 struct Node16 : Node {
   constexpr static auto kType = Type_Node16;
   constexpr static auto kMaxNodes = 16;
-  // Sorted
-  uint8_t index[kMaxNodes];
+
   Node *children[kMaxNodes];
   InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
+  // Sorted
+  uint8_t index[kMaxNodes];
 
+  uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node3 &other);
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const struct Node48 &other);
-
   size_t size() const { return sizeof(Node16) + getCapacity(); }
 };
 
 struct Node48 : Node {
   constexpr static auto kType = Type_Node48;
   constexpr static auto kMaxNodes = 48;
-  BitSet bitSet;
-  int8_t nextFree;
-  int8_t index[256];
-  Node *children[kMaxNodes];
-  InternalVersionT childMaxVersion[kMaxNodes];
-  uint8_t reverseIndex[kMaxNodes];
   constexpr static int kMaxOfMaxPageSize = 16;
   constexpr static int kMaxOfMaxShift =
       std::countr_zero(uint32_t(kMaxOfMaxPageSize));
   constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
+
+  BitSet bitSet;
+  Node *children[kMaxNodes];
+  InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
+  uint8_t reverseIndex[kMaxNodes];
+  int8_t index[256];
 
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
-
   void copyChildrenAndKeyFrom(const Node16 &other);
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const struct Node256 &other);
-
   size_t size() const { return sizeof(Node48) + getCapacity(); }
 };
 
 struct Node256 : Node {
   constexpr static auto kType = Type_Node256;
-  BitSet bitSet;
-  Node *children[256];
-  InternalVersionT childMaxVersion[256];
+  constexpr static auto kMaxNodes = 256;
   constexpr static int kMaxOfMaxPageSize = 16;
   constexpr static int kMaxOfMaxShift =
       std::countr_zero(uint32_t(kMaxOfMaxPageSize));
-  constexpr static int kMaxOfMaxTotalPages = 256 / kMaxOfMaxPageSize;
+  constexpr static int kMaxOfMaxTotalPages = kMaxNodes / kMaxOfMaxPageSize;
+
+  BitSet bitSet;
+  Node *children[kMaxNodes];
+  InternalVersionT childMaxVersion[kMaxNodes];
   InternalVersionT maxOfMax[kMaxOfMaxTotalPages];
+
   uint8_t *partialKey() { return (uint8_t *)(this + 1); }
   void copyChildrenAndKeyFrom(const Node48 &other);
   void copyChildrenAndKeyFrom(const Node256 &other);
-
   size_t size() const { return sizeof(Node256) + getCapacity(); }
 };
 
@@ -323,7 +321,7 @@ inline void Node3::copyChildrenAndKeyFrom(const Node0 &other) {
 inline void Node3::copyChildrenAndKeyFrom(const Node3 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
-  memcpy(index, other.index, sizeof(*this) - sizeof(Node));
+  memcpy(children, other.children, sizeof(*this) - sizeof(Node));
   memcpy(partialKey(), &other + 1, partialKeyLen);
   for (int i = 0; i < numChildren; ++i) {
     assert(children[i]->parent == &other);
@@ -404,7 +402,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node16 &other) {
   }
   memcpy(partialKey(), &other + 1, partialKeyLen);
   bitSet.init();
-  nextFree = Node16::kMaxNodes;
   int i = 0;
   for (auto x : other.index) {
     bitSet.set(x);
@@ -424,7 +421,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node48 &other) {
   memcpy((char *)this + kNodeCopyBegin, (char *)&other + kNodeCopyBegin,
          kNodeCopySize);
   bitSet = other.bitSet;
-  nextFree = other.nextFree;
   memcpy(index, other.index, sizeof(index));
   memset(children, 0, sizeof(children));
   const auto z = InternalVersionT::zero;
@@ -451,7 +447,6 @@ inline void Node48::copyChildrenAndKeyFrom(const Node256 &other) {
   for (auto &v : childMaxVersion) {
     v = z;
   }
-  nextFree = other.numChildren;
   bitSet = other.bitSet;
   int i = 0;
   bitSet.forEachSet([&](int c) {
@@ -773,22 +768,17 @@ private:
   BoundedFreeListAllocator<Node256> node256;
 };
 
-template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
-  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
-
-  // cachegrind says the plain loop is fewer instructions and more mis-predicted
-  // branches. Microbenchmark says plain loop is faster. It's written in this
-  // weird "generic" way though in case someday we can use the simd
-  // implementation easily if we want.
-  if constexpr (std::is_same_v<NodeT, Node3>) {
-    Node3 *n = (Node3 *)self;
-    for (int i = 0; i < n->numChildren; ++i) {
-      if (n->index[i] == index) {
-        return i;
-      }
+int getNodeIndex(Node3 *self, uint8_t index) {
+  Node3 *n = (Node3 *)self;
+  for (int i = 0; i < n->numChildren; ++i) {
+    if (n->index[i] == index) {
+      return i;
     }
-    return -1;
   }
+  return -1;
+}
+
+int getNodeIndex(Node16 *self, uint8_t index) {
 
 #ifdef HAS_AVX
   // Based on https://www.the-paper-trail.org/post/art-paper-notes/
@@ -801,7 +791,7 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   // keys aren't valid, we'll mask the results to only consider the valid ones
   // below.
   __m128i indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, indices);
 
   // Build a mask to select only the first node->num_children values from the
@@ -824,12 +814,11 @@ template <class NodeT> int getNodeIndex(NodeT *self, uint8_t index) {
   // https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
 
   uint8x16_t indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   // 0xff for each match
   uint16x8_t results =
       vreinterpretq_u16_u8(vceqq_u8(vdupq_n_u8(index), indices));
-  static_assert(NodeT::kMaxNodes <= 16);
-  assume(self->numChildren <= NodeT::kMaxNodes);
+  assume(self->numChildren <= Node16::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -1082,22 +1071,18 @@ ChildAndMaxVersion getChildAndMaxVersion(Node *self, uint8_t index) {
   }
 }
 
-template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
-  static_assert(std::is_same_v<NodeT, Node3> || std::is_same_v<NodeT, Node16>);
+Node *getChildGeq(Node0 *, int) { return nullptr; }
 
-  // cachegrind says the plain loop is fewer instructions and more mis-predicted
-  // branches. Microbenchmark says plain loop is faster. It's written in this
-  // weird "generic" way though so that someday we can use the simd
-  // implementation easily if we want.
-  if constexpr (std::is_same_v<NodeT, Node3>) {
-    Node3 *n = (Node3 *)self;
-    for (int i = 0; i < n->numChildren; ++i) {
-      if (n->index[i] >= child) {
-        return n->children[i];
-      }
+Node *getChildGeq(Node3 *n, int child) {
+  for (int i = 0; i < n->numChildren; ++i) {
+    if (n->index[i] >= child) {
+      return n->children[i];
     }
-    return nullptr;
   }
+  return nullptr;
+}
+
+Node *getChildGeq(Node16 *self, int child) {
   if (child > 255) {
     return nullptr;
   }
@@ -1105,7 +1090,7 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
 #ifdef HAS_AVX
   __m128i key_vec = _mm_set1_epi8(child);
   __m128i indices;
-  memcpy(&indices, self->index, NodeT::kMaxNodes);
+  memcpy(&indices, self->index, Node16::kMaxNodes);
   __m128i results = _mm_cmpeq_epi8(key_vec, _mm_min_epu8(key_vec, indices));
   int mask = (1 << self->numChildren) - 1;
   uint32_t bitfield = _mm_movemask_epi8(results) & mask;
@@ -1115,8 +1100,7 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
   memcpy(&indices, self->index, sizeof(self->index));
   // 0xff for each leq
   auto results = vcleq_u8(vdupq_n_u8(child), indices);
-  static_assert(NodeT::kMaxNodes <= 16);
-  assume(self->numChildren <= NodeT::kMaxNodes);
+  assume(self->numChildren <= Node16::kMaxNodes);
   uint64_t mask = self->numChildren == 16
                       ? uint64_t(-1)
                       : (uint64_t(1) << (self->numChildren * 4)) - 1;
@@ -1141,13 +1125,6 @@ template <class NodeT> Node *getChildGeqSimd(NodeT *self, int child) {
 #endif
 }
 
-Node *getChildGeq(Node0 *, int) { return nullptr; }
-Node *getChildGeq(Node3 *self, int child) {
-  return getChildGeqSimd(self, child);
-}
-Node *getChildGeq(Node16 *self, int child) {
-  return getChildGeqSimd(self, child);
-}
 Node *getChildGeq(Node48 *self, int child) {
   int c = self->bitSet.firstSetGeq(child);
   if (c < 0) {
@@ -1360,7 +1337,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
     auto *self3 = static_cast<Node3 *>(self);
     int i = self->numChildren - 1;
     for (; i >= 0; --i) {
-      if (int(self3->index[i]) < int(index)) {
+      if (self3->index[i] < index) {
         break;
       }
       self3->index[i + 1] = self3->index[i];
@@ -1390,7 +1367,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
     auto *self16 = static_cast<Node16 *>(self);
     int i = self->numChildren - 1;
     for (; i >= 0; --i) {
-      if (int(self16->index[i]) < int(index)) {
+      if (self16->index[i] < index) {
         break;
       }
       self16->index[i + 1] = self16->index[i];
@@ -1419,9 +1396,7 @@ Node *&getOrCreateChild(Node *&self, std::span<const uint8_t> &key,
   insert48:
     auto *self48 = static_cast<Node48 *>(self);
     self48->bitSet.set(index);
-    ++self->numChildren;
-    assert(self48->nextFree < 48);
-    int nextFree = self48->nextFree++;
+    auto nextFree = self48->numChildren++;
     self48->index[index] = nextFree;
     self48->reverseIndex[nextFree] = index;
     auto &result = self48->children[nextFree];
@@ -1569,8 +1544,6 @@ void maybeDecreaseCapacity(Node *&self, WriteContext *tls,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
                                                  InternalVersionT zero) {
   uint32_t z;
@@ -1580,7 +1553,6 @@ __attribute__((target("avx512f"))) void rezero16(InternalVersionT *vs,
       _mm512_sub_epi32(_mm512_loadu_epi32(vs), zvec), _mm512_setzero_epi32());
   _mm512_mask_storeu_epi32(vs, m, zvec);
 }
-// GCOVR_EXCL_STOP
 
 __attribute__((target("default")))
 #endif
@@ -1812,7 +1784,7 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
     parent48->bitSet.reset(parentsIndex);
     int8_t toRemoveChildrenIndex =
         std::exchange(parent48->index[parentsIndex], -1);
-    int8_t lastChildrenIndex = --parent48->nextFree;
+    auto lastChildrenIndex = --parent48->numChildren;
     assert(toRemoveChildrenIndex >= 0);
     assert(lastChildrenIndex >= 0);
     if (toRemoveChildrenIndex != lastChildrenIndex) {
@@ -1831,8 +1803,6 @@ Node *erase(Node *self, WriteContext *tls, ConflictSet::Impl *impl,
     }
     parent48->childMaxVersion[lastChildrenIndex] = tls->zero;
 
-    --parent->numChildren;
-
     if (needsDownsize(parent48)) {
       downsize(parent48, tls, impl, result);
     }
@@ -2019,9 +1989,9 @@ downLeftSpine:
 #ifdef HAS_AVX
 uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
   uint32_t compared = 0;
-  __m128i w[4];
+  __m128i w[4]; // GCOVR_EXCL_LINE
   memcpy(w, vs, sizeof(w));
-  uint32_t r;
+  uint32_t r; // GCOVR_EXCL_LINE
   memcpy(&r, &rv, sizeof(r));
   const auto rvVec = _mm_set1_epi32(r);
   const auto zero = _mm_setzero_si128();
@@ -2033,8 +2003,6 @@ uint32_t compare16_32bit(const InternalVersionT *vs, InternalVersionT rv) {
   return compared;
 }
 
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) uint32_t
 compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
   uint32_t r;
@@ -2044,7 +2012,6 @@ compare16_32bit_avx512(const InternalVersionT *vs, InternalVersionT rv) {
       _mm512_setzero_epi32());
 }
 #endif
-// GCOVR_EXCL_STOP
 
 // Returns true if v[i] <= readVersion for all i such that begin <= is[i] < end
 // Preconditions: begin <= end, end - begin < 256
@@ -2427,14 +2394,11 @@ bool checkMaxBetweenExclusiveImpl(Node *n, int begin, int end,
 }
 
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
-// This gets covered in local development
-// GCOVR_EXCL_START
 __attribute__((target("avx512f"))) bool
 checkMaxBetweenExclusive(Node *n, int begin, int end,
                          InternalVersionT readVersion, ReadContext *tls) {
   return checkMaxBetweenExclusiveImpl<true>(n, begin, end, readVersion, tls);
 }
-// GCOVR_EXCL_STOP
 __attribute__((target("default")))
 #endif
 bool checkMaxBetweenExclusive(Node *n, int begin, int end,

README.md

@@ -24,15 +24,15 @@ Hardware for all benchmarks is an AMD Ryzen 9 7900 with (2x32GB) 5600MT/s CL28-3
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               11.04 |       90,614,308.12 |    0.8% |          180.38 |           55.13 |  3.272 |          41.51 |    0.4% |      0.01 | `point reads`
-|               14.96 |       66,843,629.12 |    0.4% |          274.41 |           74.73 |  3.672 |          55.05 |    0.3% |      0.01 | `prefix reads`
-|               37.06 |       26,982,847.61 |    0.2% |          791.04 |          185.28 |  4.269 |         142.67 |    0.2% |      0.01 | `range reads`
-|               17.89 |       55,887,365.73 |    0.6% |          335.54 |           89.79 |  3.737 |          43.84 |    0.4% |      0.01 | `point writes`
-|               31.85 |       31,394,336.65 |    0.3% |          615.32 |          159.63 |  3.855 |          87.69 |    0.2% |      0.01 | `prefix writes`
-|               36.17 |       27,647,221.45 |    0.6% |          705.11 |          182.80 |  3.857 |         100.62 |    0.1% |      0.01 | `range writes`
-|               79.01 |       12,656,457.78 |    0.7% |        1,498.35 |          402.46 |  3.723 |         270.50 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          303,667.50 |            3,293.08 |    1.1% |    3,931,273.00 |    1,612,702.50 |  2.438 |     806,223.33 |    0.0% |      0.01 | `worst case for radix tree`
-|               83.70 |       11,947,443.83 |    0.7% |        1,738.03 |          429.06 |  4.051 |         270.01 |    0.0% |      0.01 | `create and destroy`
+|               11.18 |       89,455,125.34 |    0.6% |          185.37 |           57.08 |  3.248 |          41.51 |    0.4% |      0.01 | `point reads`
+|               14.53 |       68,800,688.89 |    0.4% |          282.41 |           74.80 |  3.776 |          55.06 |    0.3% |      0.01 | `prefix reads`
+|               36.54 |       27,367,576.87 |    0.2% |          798.06 |          188.90 |  4.225 |         141.69 |    0.2% |      0.01 | `range reads`
+|               16.69 |       59,912,106.02 |    0.6% |          314.57 |           86.29 |  3.645 |          39.84 |    0.4% |      0.01 | `point writes`
+|               30.09 |       33,235,744.07 |    0.5% |          591.33 |          155.92 |  3.793 |          82.69 |    0.2% |      0.01 | `prefix writes`
+|               35.77 |       27,956,388.03 |    1.4% |          682.25 |          187.63 |  3.636 |          96.12 |    0.1% |      0.01 | `range writes`
+|               74.04 |       13,505,408.41 |    2.7% |        1,448.95 |          392.10 |  3.695 |         260.53 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          330,984.50 |            3,021.29 |    1.9% |    3,994,153.50 |    1,667,309.00 |  2.396 |     806,019.50 |    0.0% |      0.01 | `worst case for radix tree`
+|               92.46 |       10,814,961.65 |    0.5% |        1,800.00 |          463.41 |  3.884 |         297.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
@@ -47,7 +47,7 @@ Check: 4.47891 seconds, 364.05 MB/s, Add: 4.55599 seconds, 123.058 MB/s, Gc rati
 ## radix tree
 
 ```
-Check: 0.958985 seconds, 1700.28 MB/s, Add: 1.35083 seconds, 415.044 MB/s, Gc ratio: 44.4768%, Peak idle memory: 2.33588e+06
+Check: 0.953012 seconds, 1710.94 MB/s, Add: 1.30025 seconds, 431.188 MB/s, Gc ratio: 43.9816%, Peak idle memory: 2.28375e+06
 ```
 
 ## hash table