Merge pull request 'Improve perf/codegen' (#94) from perf into main

Reviewed-on: #94

CMakeLists.txt

@@ -182,6 +182,8 @@ endif()
 
 include(CTest)
 
+option(TEST_SKIP_LIST "Run the slow skip list cross-check tests" OFF)
+
 # disable tests if this is being used through e.g. FetchContent
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
 
@@ -224,8 +226,9 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
   target_compile_options(driver_skip_list PRIVATE ${TEST_FLAGS})
   target_link_libraries(driver_skip_list PRIVATE skip_list)
 
-  # enable to test skip list
-  if(0)
+  # Optional cross-check of SkipList.cpp against ReferenceImpl. Opt-in via
+  # -DTEST_SKIP_LIST=ON, since these are slow.
+  if(TEST_SKIP_LIST)
     foreach(TEST ${CORPUS_TESTS})
       get_filename_component(hash ${TEST} NAME)
       add_test(NAME skip_list_${hash} COMMAND driver_skip_list ${TEST})
@@ -390,6 +393,24 @@ if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR AND BUILD_TESTING)
                         PROPERTIES CXX_STANDARD_REQUIRED ON)
   add_test(NAME conflict_set_cxx_api_test COMMAND conflict_set_cxx_api_test)
 
+  # SHOW_MEMORY instrumentation, which the rest of the test suite doesn't
+  # compile with
+  add_executable(conflict_set_show_memory_test ConflictSet.cpp
+                                               ShowMemoryTest.cpp)
+  target_compile_options(conflict_set_show_memory_test PRIVATE ${TEST_FLAGS})
+  target_compile_definitions(conflict_set_show_memory_test
+                             PRIVATE SHOW_MEMORY=1)
+  target_include_directories(conflict_set_show_memory_test
+                             PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
+  if(NOT CMAKE_CROSSCOMPILING)
+    target_compile_options(conflict_set_show_memory_test
+                           PRIVATE -fsanitize=address,undefined)
+    target_link_options(conflict_set_show_memory_test PRIVATE
+                        -fsanitize=address,undefined)
+  endif()
+  add_test(NAME conflict_set_show_memory_test
+           COMMAND conflict_set_show_memory_test)
+
   # symbol visibility tests
   if(NOT CMAKE_BUILD_TYPE STREQUAL Debug)
     if(APPLE)

ConflictSet.cpp

@@ -80,17 +80,19 @@ using namespace weaselab;
 #define assume assert
 #endif
 
-#if SHOW_MEMORY
+// The SHOW_MEMORY instrumentation. Declared unconditionally so that each of
+// these has a single definition in every configuration (see the definitions at
+// the end of this file), rather than a no-op stub standing in for the real
+// thing.
 void addNode(struct Node *);
 void removeNode(struct Node *);
 void addKey(struct Node *);
 void removeKey(struct Node *);
-#else
-constexpr void addNode(struct Node *) {}
-constexpr void removeNode(struct Node *) {}
-constexpr void addKey(struct Node *) {}
-constexpr void removeKey(struct Node *) {}
-#endif
+// Like `removeKey`, but with a caller-provided search path length. Used when
+// the node's ancestors may already have been released, which makes walking
+// `parent` pointers invalid.
+void removeKey(struct Node *, int64_t searchPathLength);
+int64_t getSearchPathLength(struct Node *);
 
 // ==================== BEGIN IMPLEMENTATION ====================
 
@@ -116,13 +118,40 @@ struct InternalVersionT {
   constexpr InternalVersionT() = default;
   constexpr explicit InternalVersionT(int64_t value) : value(value) {}
   constexpr int64_t toInt64() const { return value; } // GCOVR_EXCL_LINE
-  constexpr auto operator<=>(const InternalVersionT &rhs) const {
+  constexpr bool operator<(const InternalVersionT &rhs) const {
 #if USE_64_BIT
-    return value <=> rhs.value;
+    return value < rhs.value;
 #else
     // Maintains ordering after overflow, as long as the full-precision versions
     // are within `kMaxCorrectVersionWindow` of eachother.
-    return int32_t(value - rhs.value) <=> 0;
+    return int32_t(value - rhs.value) < 0;
+#endif
+  }
+  constexpr bool operator<=(const InternalVersionT &rhs) const {
+#if USE_64_BIT
+    return value <= rhs.value;
+#else
+    // Maintains ordering after overflow, as long as the full-precision versions
+    // are within `kMaxCorrectVersionWindow` of eachother.
+    return int32_t(value - rhs.value) <= 0;
+#endif
+  }
+  constexpr bool operator>=(const InternalVersionT &rhs) const {
+#if USE_64_BIT
+    return value >= rhs.value;
+#else
+    // Maintains ordering after overflow, as long as the full-precision versions
+    // are within `kMaxCorrectVersionWindow` of eachother.
+    return int32_t(value - rhs.value) >= 0;
+#endif
+  }
+  constexpr bool operator>(const InternalVersionT &rhs) const {
+#if USE_64_BIT
+    return value > rhs.value;
+#else
+    // Maintains ordering after overflow, as long as the full-precision versions
+    // are within `kMaxCorrectVersionWindow` of eachother.
+    return int32_t(value - rhs.value) > 0;
 #endif
   }
   constexpr bool operator==(const InternalVersionT &) const = default;
@@ -307,7 +336,9 @@ struct Node {
   uint8_t *partialKey();
   Type getType() const { return type; }
   int32_t getCapacity() const {
-    assert(!releaseDeferred);
+    // `partialKeyCapacity` is not repurposed by `deferRelease`, so this is also
+    // valid for nodes with `releaseDeferred` set. It's needed to account for
+    // the capacity of a deferred node when it's finally released.
     return partialKeyCapacity;
   }
 
@@ -2280,6 +2311,8 @@ bool checkMaxBetweenExclusiveImpl(Node3 *n, int begin, int end,
   assert(!(begin == -1 && end == 256));
 
   auto *self = static_cast<Node3 *>(n);
+  assume(1 <= self->numChildren);
+  assume(self->numChildren <= 3);
 
   ++begin;
 
@@ -2753,10 +2786,39 @@ checkMaxBetweenExclusiveImpl<true>(Node256 *n, int begin, int end,
   return self;
 }
 
+#if SHOW_MEMORY
+// Length of the key leading to `child`, given the length of the key leading to
+// its parent. Consistent with `getSearchPathLength`.
+int64_t childSearchPathLength(int64_t parentSearchPathLength, Node *child) {
+  return parentSearchPathLength + 1 + child->partialKeyLen;
+}
+#endif
+
 void eraseTree(Node *root, WriteContext *writeContext) {
   Arena arena;
   auto toFree = vector<Node *>(arena);
   toFree.push_back(root);
+  // Length of the key leading to each node in `toFree`, on a parallel stack.
+  // A node is released before its children are visited, so the `parent`
+  // pointers that `getSearchPathLength` walks are dangling by the time a queued
+  // node is popped. Carry the lengths along instead. Only the SHOW_MEMORY
+  // accounting needs them.
+#if SHOW_MEMORY
+  auto searchPathLengths = vector<int64_t>(arena);
+  searchPathLengths.push_back(getSearchPathLength(root));
+#endif
+
+  // Queue `child` for release, where `parentSearchPathLength` is the length of
+  // the key leading to its parent.
+  auto queueChild = [&](Node *child, int64_t parentSearchPathLength) {
+    toFree.push_back(child);
+#if SHOW_MEMORY
+    searchPathLengths.push_back(
+        childSearchPathLength(parentSearchPathLength, child));
+#else
+    (void)parentSearchPathLength;
+#endif
+  };
 
   while (toFree.size() > 0) {
     auto *n = toFree.back();
@@ -2764,7 +2826,13 @@ void eraseTree(Node *root, WriteContext *writeContext) {
     writeContext->accum.entries_erased += n->entryPresent;
     ++writeContext->accum.nodes_released;
 
-    removeKey(n);
+#if SHOW_MEMORY
+    const int64_t searchPathLength = searchPathLengths.back();
+    searchPathLengths.pop_back();
+#else
+    const int64_t searchPathLength = 0;
+#endif
+    removeKey(n, searchPathLength);
 
     switch (n->getType()) {
     case Type_Node0: {
@@ -2774,28 +2842,39 @@ void eraseTree(Node *root, WriteContext *writeContext) {
     case Type_Node3: {
       auto *n3 = static_cast<Node3 *>(n);
       for (int i = 0; i < n3->numChildren; ++i) {
-        toFree.push_back(n3->children[i]);
+        queueChild(n3->children[i], searchPathLength);
       }
       writeContext->release(n3);
     } break;
     case Type_Node16: {
       auto *n16 = static_cast<Node16 *>(n);
       for (int i = 0; i < n16->numChildren; ++i) {
-        toFree.push_back(n16->children[i]);
+        queueChild(n16->children[i], searchPathLength);
       }
       writeContext->release(n16);
     } break;
     case Type_Node48: {
       auto *n48 = static_cast<Node48 *>(n);
       for (int i = 0; i < n48->numChildren; ++i) {
-        toFree.push_back(n48->children[i]);
+        queueChild(n48->children[i], searchPathLength);
       }
       writeContext->release(n48);
     } break;
     case Type_Node256: {
       auto *n256 = static_cast<Node256 *>(n);
       auto *out = toFree.unsafePrepareAppend(n256->numChildren).data();
-      n256->bitSet.forEachSet([&](int i) { *out++ = n256->children[i]; });
+#if SHOW_MEMORY
+      auto *outSearchPathLengths =
+          searchPathLengths.unsafePrepareAppend(n256->numChildren).data();
+#endif
+      n256->bitSet.forEachSet([&](int i) {
+        Node *child = n256->children[i];
+        *out++ = child;
+#if SHOW_MEMORY
+        *outSearchPathLengths++ =
+            childSearchPathLength(searchPathLength, child);
+#endif
+      });
       assert(out == toFree.end());
       writeContext->release(n256);
     } break;
@@ -2907,9 +2986,13 @@ void fixupMaxVersion(Node *node, WriteContext *writeContext) {
     break;
   case Type_Node3: {
     auto *self3 = static_cast<Node3 *>(node);
-    max = std::max(max,
-                   horizontalMaxUpTo16(self3->childMaxVersion,
-                                       writeContext->zero, self3->numChildren));
+    assume(1 <= self3->numChildren);
+    assume(self3->numChildren <= 3);
+    InternalVersionT m = self3->childMaxVersion[0];
+    for (int i = 1; i < self3->numChildren; ++i) {
+      m = std::max(m, self3->childMaxVersion[i]);
+    }
+    max = std::max(max, m);
   } break;
   case Type_Node16: {
     auto *self16 = static_cast<Node16 *>(node);
@@ -5463,8 +5546,9 @@ void internal_addWrites(ConflictSet::Impl *impl,
   impl->totalBytes += mallocBytesDelta;
   impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
+  // Testing-only cross-check of the accounting.
   if (impl->totalBytes != mallocBytes) {
-    abort();
+    abort(); // GCOVR_EXCL_LINE
   }
 #endif
 }
@@ -5475,8 +5559,9 @@ void internal_setOldestVersion(ConflictSet::Impl *impl, int64_t oldestVersion) {
   impl->totalBytes += mallocBytesDelta;
   impl->memory_bytes.set(impl->totalBytes);
 #if SHOW_MEMORY
+  // Testing-only cross-check of the accounting.
   if (impl->totalBytes != mallocBytes) {
-    abort();
+    abort(); // GCOVR_EXCL_LINE
   }
 #endif
 }
@@ -5942,6 +6027,14 @@ checkMaxVersion(Node *root, Node *node, InternalVersionT oldestVersion,
 
 } // namespace
 
+// The SHOW_MEMORY instrumentation. Each of these functions is defined exactly
+// once, with the accounting compiled out when SHOW_MEMORY is disabled. Defining
+// them once (rather than defining no-op stubs when SHOW_MEMORY is disabled)
+// keeps every configuration of this file reporting each function on the same
+// line, which merging coverage from objects compiled with different options
+// requires: gcovr, merging the SHOW_MEMORY test with the rest of the build,
+// otherwise aborts with "Got function ... on multiple lines". The optimizer
+// removes the resulting empty calls.
 #if SHOW_MEMORY
 
 int64_t nodeBytes = 0;
@@ -5970,7 +6063,10 @@ int64_t getNodeSize(struct Node *n) {
   }
 }
 
+#endif
+
 int64_t getSearchPathLength(Node *n) {
+#if SHOW_MEMORY
   assert(n != nullptr);
   int64_t result = 0;
   for (;;) {
@@ -5982,9 +6078,14 @@ int64_t getSearchPathLength(Node *n) {
     n = n->parent;
   }
   return result;
+#else
+  (void)n;
+  return 0;
+#endif
 }
 
 void addNode(Node *n) {
+#if SHOW_MEMORY
   nodeBytes += getNodeSize(n);
   partialCapacityBytes += n->getCapacity();
   if (nodeBytes > peakNodeBytes) {
@@ -5993,14 +6094,22 @@ void addNode(Node *n) {
   if (partialCapacityBytes > peakPartialCapacityBytes) {
     peakPartialCapacityBytes = partialCapacityBytes;
   }
+#else
+  (void)n;
+#endif
 }
 
 void removeNode(Node *n) {
+#if SHOW_MEMORY
   nodeBytes -= getNodeSize(n);
   partialCapacityBytes -= n->getCapacity();
+#else
+  (void)n;
+#endif
 }
 
 void addKey(Node *n) {
+#if SHOW_MEMORY
   if (!n->entryPresent) {
     ++totalKeys;
     keyBytes += getSearchPathLength(n);
@@ -6011,15 +6120,36 @@ void addKey(Node *n) {
       peakKeyBytes = keyBytes;
     }
   }
+#else
+  (void)n;
+#endif
 }
 
 void removeKey(Node *n) {
+#if SHOW_MEMORY
   if (n->entryPresent) {
     --totalKeys;
     keyBytes -= getSearchPathLength(n);
   }
+#else
+  (void)n;
+#endif
 }
 
+void removeKey(Node *n, int64_t searchPathLength) {
+#if SHOW_MEMORY
+  if (n->entryPresent) {
+    --totalKeys;
+    keyBytes -= searchPathLength;
+  }
+#else
+  (void)n;
+  (void)searchPathLength;
+#endif
+}
+
+#if SHOW_MEMORY
+
 struct __attribute__((visibility("default"))) PeakPrinter {
   ~PeakPrinter() {
     printf("--- radix_tree ---\n");

Internal.h

@@ -23,7 +23,9 @@ using namespace weaselab;
 #include <callgrind.h>
 
 #define DEBUG_VERBOSE 0
+#ifndef SHOW_MEMORY
 #define SHOW_MEMORY 0
+#endif
 
 // std::span is not trivially constructible. We want a span that leaves its
 // members uninitialized for performance reasons.
@@ -96,6 +98,11 @@ static_assert(std::is_trivial_v<TrivialSpan>);
 #if SHOW_MEMORY
 inline int64_t mallocBytes = 0;
 inline int64_t peakMallocBytes = 0;
+// Defined in ConflictSet.cpp
+extern int64_t nodeBytes;
+extern int64_t partialCapacityBytes;
+extern int64_t totalKeys;
+extern int64_t keyBytes;
 #endif
 
 inline thread_local int64_t mallocBytesDelta = 0;

LongestCommonPrefix.h

@@ -96,7 +96,7 @@ inline int firstNeqStride(const uint8_t *ap, const uint8_t *bp) {
 // GCOVR_EXCL_START
 #if defined(HAS_AVX) && !defined(__SANITIZE_THREAD__)
 __attribute__((target("avx512f,avx512bw"))) inline int
-longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {
   int i = 0;
   int end = cl & ~63;
   while (i < end) {
@@ -123,7 +123,7 @@ __attribute__((target("default")))
 // GCOVR_EXCL_STOP
 
 inline int
-longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+longestCommonPrefixLong(const uint8_t *ap, const uint8_t *bp, int cl) {
   if (!(cl >= 0)) {
     __builtin_unreachable(); // GCOVR_EXCL_LINE
   }
@@ -182,4 +182,72 @@ longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
   }
 
   return i;
-}
+}
+// Use in-bounds native-word loads for short common lengths. The second load
+// overlaps the first; because the first load matched, any mismatch found in
+// the overlap is impossible and the returned absolute index is still correct.
+// Precondition: cl <= 2 * sizeof(uint64_t)
+__attribute__((always_inline)) static inline int
+longestCommonPrefixShort(const uint8_t *ap, const uint8_t *bp, int cl) {
+  assert(cl <= 2 * int(sizeof(uint64_t)));
+  if (cl >= 8) {
+    uint64_t a0;
+    uint64_t b0;
+    memcpy(&a0, ap, sizeof(a0));
+    memcpy(&b0, bp, sizeof(b0));
+    if (uint64_t mismatched = a0 ^ b0) {
+      return std::countr_zero(mismatched) / 8;
+    }
+    if (cl == 8) {
+      return 8;
+    }
+
+    uint64_t a1;
+    uint64_t b1;
+    memcpy(&a1, ap + cl - 8, sizeof(a1));
+    memcpy(&b1, bp + cl - 8, sizeof(b1));
+    if (uint64_t mismatched = a1 ^ b1) {
+      return cl - 8 + std::countr_zero(mismatched) / 8;
+    }
+    return cl;
+  }
+
+  if (cl >= 4) {
+    uint32_t a0;
+    uint32_t b0;
+    memcpy(&a0, ap, sizeof(a0));
+    memcpy(&b0, bp, sizeof(b0));
+    if (uint32_t mismatched = a0 ^ b0) {
+      return std::countr_zero(mismatched) / 8;
+    }
+    if (cl == 4) {
+      return 4;
+    }
+
+    uint32_t a1;
+    uint32_t b1;
+    memcpy(&a1, ap + cl - 4, sizeof(a1));
+    memcpy(&b1, bp + cl - 4, sizeof(b1));
+    if (uint32_t mismatched = a1 ^ b1) {
+      return cl - 4 + std::countr_zero(mismatched) / 8;
+    }
+    return cl;
+  }
+
+  int i = 0;
+  while (i < cl && ap[i] == bp[i]) {
+    ++i;
+  }
+  return i;
+}
+
+__attribute__((always_inline)) static inline int
+longestCommonPrefix(const uint8_t *ap, const uint8_t *bp, int cl) {
+  if (!(cl >= 0)) {
+    __builtin_unreachable(); // GCOVR_EXCL_LINE
+  }
+  if (cl <= 2 * int(sizeof(uint64_t))) {
+    return longestCommonPrefixShort(ap, bp, cl);
+  }
+  return longestCommonPrefixLong(ap, bp, cl);
+}

README.md

@@ -19,30 +19,30 @@ InstalledDir: /usr/lib/llvm-21/bin
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|              164.29 |        6,086,873.38 |    0.0% |        3,107.03 |          604.19 |  5.142 |         558.59 |    0.0% |      1.96 | `point reads`
-|              161.05 |        6,209,395.38 |    0.1% |        3,036.76 |          592.21 |  5.128 |         539.35 |    0.0% |      1.93 | `prefix reads`
-|              239.55 |        4,174,539.38 |    0.1% |        3,722.71 |          880.68 |  4.227 |         692.00 |    0.0% |      2.86 | `range reads`
-|              354.75 |        2,818,919.14 |    0.7% |        4,523.64 |        1,304.75 |  3.467 |         720.22 |    2.0% |      4.23 | `point writes`
-|              345.32 |        2,895,878.47 |    0.1% |        4,484.57 |        1,270.31 |  3.530 |         705.00 |    1.8% |      4.12 | `prefix writes`
-|              193.48 |        5,168,547.42 |    0.1% |        2,224.10 |          711.72 |  3.125 |         377.17 |    3.3% |      2.32 | `range writes`
-|              404.89 |        2,469,777.50 |    2.4% |        6,855.96 |        1,489.70 |  4.602 |       1,227.82 |    1.3% |      0.05 | `monotonic increasing point writes`
-|          134,231.80 |            7,449.80 |    1.9% |      812,045.25 |      495,770.40 |  1.638 |     151,246.50 |    0.9% |      0.01 | `worst case for radix tree`
-|               37.80 |       26,454,311.17 |    0.4% |          701.00 |          139.14 |  5.038 |         102.00 |    0.0% |      0.01 | `create and destroy`
+|              165.93 |        6,026,520.04 |    0.1% |        3,107.03 |          610.42 |  5.090 |         558.59 |    0.0% |      1.98 | `point reads`
+|              163.08 |        6,132,046.57 |    0.1% |        3,036.76 |          599.68 |  5.064 |         539.35 |    0.0% |      1.95 | `prefix reads`
+|              250.26 |        3,995,834.21 |    0.1% |        3,722.71 |          920.25 |  4.045 |         692.00 |    0.0% |      2.99 | `range reads`
+|              359.64 |        2,780,581.63 |    0.1% |        4,523.64 |        1,322.63 |  3.420 |         720.22 |    2.1% |      4.30 | `point writes`
+|              358.20 |        2,791,726.93 |    0.1% |        4,484.57 |        1,317.23 |  3.405 |         705.00 |    2.0% |      4.28 | `prefix writes`
+|              193.91 |        5,157,081.29 |    0.1% |        2,224.10 |          713.33 |  3.118 |         377.17 |    3.3% |      2.32 | `range writes`
+|              422.03 |        2,369,513.38 |    1.1% |        6,855.96 |        1,551.55 |  4.419 |       1,227.82 |    1.4% |      0.05 | `monotonic increasing point writes`
+|          142,467.00 |            7,019.17 |    3.8% |      812,047.33 |      508,535.40 |  1.597 |     151,246.50 |    1.0% |      0.01 | `worst case for radix tree`
+|               37.66 |       26,554,217.74 |    0.8% |          701.00 |          138.80 |  5.050 |         102.00 |    0.0% |      0.01 | `create and destroy`
 
 ## Radix tree (this implementation)
 
 
 |               ns/op |                op/s |    err% |          ins/op |          cyc/op |    IPC |         bra/op |   miss% |     total | benchmark
 |--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
-|               12.89 |       77,565,115.56 |    0.1% |          244.55 |           47.43 |  5.155 |          34.21 |    0.6% |      0.15 | `point reads`
-|               15.11 |       66,162,047.76 |    0.1% |          297.79 |           55.60 |  5.356 |          43.23 |    0.4% |      0.18 | `prefix reads`
-|               36.29 |       27,559,358.29 |    0.1% |          783.16 |          133.44 |  5.869 |         109.52 |    0.2% |      0.43 | `range reads`
-|               20.53 |       48,719,405.55 |    0.1% |          381.81 |           75.51 |  5.057 |          51.04 |    0.5% |      0.25 | `point writes`
-|               39.37 |       25,402,042.40 |    0.1% |          685.00 |          144.83 |  4.730 |         106.72 |    0.3% |      0.47 | `prefix writes`
-|               43.78 |       22,843,841.63 |    0.1% |          800.40 |          161.06 |  4.970 |         127.36 |    0.1% |      0.53 | `range writes`
-|               78.37 |       12,760,008.75 |    1.0% |        1,452.61 |          288.24 |  5.040 |         278.69 |    0.1% |      0.01 | `monotonic increasing point writes`
-|          322,885.50 |            3,097.07 |    1.5% |    4,362,382.00 |    1,183,852.00 |  3.685 |     765,301.00 |    0.1% |      0.01 | `worst case for radix tree`
-|               99.99 |       10,000,718.79 |    0.4% |        1,775.00 |          367.93 |  4.824 |         288.00 |    0.0% |      0.01 | `create and destroy`
+|               12.25 |       81,664,669.29 |    0.4% |          233.49 |           44.94 |  5.196 |          31.23 |    0.7% |      0.15 | `point reads`
+|               13.35 |       74,903,950.49 |    0.1% |          271.75 |           49.10 |  5.534 |          40.21 |    0.4% |      0.16 | `prefix reads`
+|               31.37 |       31,874,583.62 |    0.1% |          700.39 |          115.43 |  6.067 |         100.83 |    0.2% |      0.37 | `range reads`
+|               19.55 |       51,153,057.80 |    0.1% |          366.51 |           71.90 |  5.098 |          47.92 |    0.5% |      0.23 | `point writes`
+|               36.92 |       27,088,344.17 |    0.1% |          645.11 |          135.72 |  4.753 |         100.31 |    0.4% |      0.44 | `prefix writes`
+|               36.84 |       27,143,553.08 |    0.1% |          691.04 |          135.42 |  5.103 |         114.01 |    0.1% |      0.45 | `range writes`
+|               74.62 |       13,400,467.72 |    0.9% |        1,418.37 |          274.20 |  5.173 |         273.39 |    0.1% |      0.01 | `monotonic increasing point writes`
+|          352,671.00 |            2,835.50 |   15.3% |    3,687,535.00 |    1,287,119.00 |  2.865 |     712,030.00 |    0.1% |      0.01 | :wavy_dash: `worst case for radix tree` (Unstable with ~1.0 iters. Increase `minEpochIterations` to e.g. 10)
+|               96.63 |       10,348,729.42 |    1.2% |        1,766.00 |          355.53 |  4.967 |         283.00 |    0.0% |      0.01 | `create and destroy`
 
 # "Real data" test
 
@@ -51,13 +51,13 @@ Point queries only. Gc ratio is the ratio of time spent doing garbage collection
 ## skip list
 
 ```
-Check: 4.62967 seconds, 352.195 MB/s, Add: 3.34177 seconds, 167.771 MB/s, Gc ratio: 37.9399%, Peak idle memory: 5.51852e+06
+Check: 4.69875 seconds, 347.017 MB/s, Add: 3.39475 seconds, 165.153 MB/s, Gc ratio: 38.7414%, Peak idle memory: 5.51852e+06
 ```
 
 ## radix tree
 
 ```
-Check: 1.00477 seconds, 1622.8 MB/s, Add: 1.21142 seconds, 462.808 MB/s, Gc ratio: 39.4716%, Peak idle memory: 2.0226e+06
+Check: 0.976745 seconds, 1669.37 MB/s, Add: 1.20464 seconds, 465.411 MB/s, Gc ratio: 39.4158%, Peak idle memory: 2.0226e+06
 ```
 
 ## hash table
@@ -65,6 +65,6 @@ Check: 1.00477 seconds, 1622.8 MB/s, Add: 1.21142 seconds, 462.808 MB/s, Gc rati
 (The hash table implementation doesn't work on range queries, and its purpose is to provide an idea of how fast point queries can be)
 
 ```
-Check: 0.854254 seconds, 1908.74 MB/s, Add: 0.632626 seconds, 886.232 MB/s, Gc ratio: 41.0827%, Peak idle memory: 0
+Check: 0.864589 seconds, 1885.92 MB/s, Add: 0.634604 seconds, 883.47 MB/s, Gc ratio: 41.184%, Peak idle memory: 0
 ```
 

ShowMemoryTest.cpp

@@ -0,0 +1,115 @@
+// Exercises the SHOW_MEMORY instrumentation, which the other tests don't
+// compile with (this target passes -DSHOW_MEMORY=1). The instrumentation has
+// its own consistency checks, which abort on failure, and asserts are enabled
+// by TEST_FLAGS. Only one ConflictSet may exist at a time, since the
+// accounting is process-wide.
+#include "ConflictSet.h"
+#include "Internal.h"
+
+#include <cassert>
+#include <cstdint>
+#include <string>
+
+using namespace weaselab;
+
+namespace {
+
+void writePoint(ConflictSet &cs, const std::string &key, int64_t writeVersion) {
+  ConflictSet::WriteRange w{{(const uint8_t *)key.data(), int(key.size())},
+                            {nullptr, 0}};
+  cs.addWrites(&w, 1, writeVersion);
+}
+
+ConflictSet::Result checkPoint(const ConflictSet &cs, const std::string &key,
+                               int64_t readVersion) {
+  ConflictSet::ReadRange r{{(const uint8_t *)key.data(), int(key.size())},
+                           {nullptr, 0},
+                           readVersion};
+  ConflictSet::Result result;
+  cs.check(&r, &result, 1);
+  return result;
+}
+
+// Everything a ConflictSet allocated must have been accounted for and released
+// by the time it's destroyed.
+void checkFullyReleased() {
+  assert(nodeBytes == 0);
+  assert(partialCapacityBytes == 0);
+  assert(totalKeys == 0);
+  assert(keyBytes == 0);
+  assert(mallocBytes == 0);
+}
+
+} // namespace
+
+int main() {
+  // Splitting a node replaces it, which defers the release of the old node.
+  // Releasing a deferred node must account for its capacity, and the key it
+  // held must be accounted for without walking `parent` pointers of nodes that
+  // have already been released.
+  {
+    ConflictSet cs{0};
+    writePoint(cs, "a", 1);
+    writePoint(cs, "ab", 2);
+    assert(checkPoint(cs, "a", 0) == ConflictSet::Conflict);
+    assert(checkPoint(cs, "ab", 0) == ConflictSet::Conflict);
+    assert(checkPoint(cs, "b", 0) == ConflictSet::Commit);
+    assert(nodeBytes > 0);
+    assert(keyBytes > 0);
+  }
+  checkFullyReleased();
+
+  // Exceeding the version window erases the tree mid-workload.
+  {
+    ConflictSet cs{0};
+    writePoint(cs, "abc", 1);
+    const int64_t version = int64_t(1) << 40;
+    writePoint(cs, "abc", version);
+    assert(checkPoint(cs, "abc", 0) == ConflictSet::TooOld);
+    assert(checkPoint(cs, "abc", version) == ConflictSet::Commit);
+  }
+  checkFullyReleased();
+
+  // Erase a tree with many nodes, both mid-workload and at destruction.
+  {
+    ConflictSet cs{0};
+    for (int i = 0; i < 1000; ++i) {
+      writePoint(cs, "prefix/" + std::to_string(i), i + 1);
+    }
+    const int64_t version = int64_t(1) << 40;
+    writePoint(cs, "prefix/1", version);
+    assert(checkPoint(cs, "prefix/1", version) == ConflictSet::Commit);
+
+    for (int i = 0; i < 1000; ++i) {
+      writePoint(cs, "other/" + std::to_string(i), version + i + 1);
+    }
+    assert(checkPoint(cs, "other/999", version) == ConflictSet::Conflict);
+  }
+  checkFullyReleased();
+
+  // Grow a node through every node type, so that the accounting for each node
+  // type's size is exercised, then erase the entries one at a time so that
+  // nodes are released individually rather than as part of a whole tree.
+  {
+    ConflictSet cs{0};
+    std::string key(1, '\0');
+    for (int i = 0; i < 256; ++i) {
+      key[0] = char(i);
+      writePoint(cs, key, i + 1);
+    }
+    for (int i = 0; i < 256; ++i) {
+      key[0] = char(i);
+      assert(checkPoint(cs, key, i) == ConflictSet::Conflict);
+    }
+
+    // Garbage collection erases entries that can only fail with TooOld anyway.
+    // Doing it one version at a time walks the tree, erasing, downsizing and
+    // releasing nodes as it goes.
+    for (int i = 0; i < 256; ++i) {
+      cs.setOldestVersion(i + 1);
+    }
+    key[0] = '\0';
+    assert(checkPoint(cs, key, 0) == ConflictSet::TooOld);
+  }
+  checkFullyReleased();
+}

SkipList.cpp

@@ -819,10 +819,17 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
   }
 
   void setOldestVersion(int64_t oldestVersion) {
+    // oldestVersion becomes the max of its previous value and the argument, so
+    // an older version is a no-op. We still fall through in the equal case so
+    // that garbage collection can make progress without advancing the oldest
+    // version.
+    if (oldestVersion < this->oldestVersion) {
+      return;
+    }
+
     // This isn't 100% accurate. It overcounts if you hit the end
     gc_iterations_total.add(keyUpdates);
 
-    assert(oldestVersion >= this->oldestVersion);
     this->oldestVersion = oldestVersion;
     SkipList::Finger finger;
     int temp;

test_conflict_set.py

@@ -68,6 +68,35 @@ def test_hash_table_getBytes():
         assert cs.check(read(0, b"key")) == [Result.CONFLICT]
 
 
+def test_setOldestVersion_does_not_decrease():
+    # Regression test for issue #87: `setOldestVersion` sets oldestVersion to
+    # the maximum of its previous value and its argument, so passing an older
+    # version is valid input and must be a no-op. skip_list used to assert in
+    # Debug builds and, in Release builds, lower its oldest version, which
+    # resurrected a window of versions that callers were told resulted in
+    # TooOld.
+    with DebugConflictSet() as cs:
+        cs.addWrites(5, write(b"k"))
+        cs.setOldestVersion(5)
+        # Valid: `2 >= 0` and `2 <= newestVersion == 5`. oldestVersion stays 5,
+        # so 3 and 4 are still TooOld.
+        cs.setOldestVersion(2)
+        assert cs.check(read(3, b"k")) == [Result.TOO_OLD]
+        assert cs.check(read(4, b"k")) == [Result.TOO_OLD]
+        assert cs.check(read(5, b"k")) == [Result.COMMIT]
+
+    # Same, but with the entries at the old oldestVersion already garbage
+    # collected, so lowering oldestVersion would report Commit for a version
+    # that the caller was told resulted in TooOld.
+    with DebugConflictSet() as cs:
+        cs.addWrites(1, write(b"a"))
+        cs.addWrites(2)  # No writes; advances newestVersion to 2
+        cs.setOldestVersion(2)  # Garbage collects the write at version 1
+        cs.setOldestVersion(1)
+        assert cs.check(read(1, b"a")) == [Result.TOO_OLD]
+        assert cs.check(read(2, b"a")) == [Result.COMMIT]
+
+
 def test_hash_table_setOldestVersion_empties_map():
     # Regression test for issue #78: setOldestVersion must terminate when
     # its garbage-collection pass erases every entry in the map. Each write
@@ -222,13 +251,18 @@ def test_merge_child_node48():
         cs.addWrites(1, write(b"\x00" * 8, b"\x00" * 10))
 
 
-def test_fixup_256():
+def test_fixup_256_heterogeneous_versions():
+    # Regression test for issue #89: recomputing the maxVersion of a Node256
+    # must consider all of its children, not just those with an index < 16.
+    # Here the highest-version child ("\x10\x50") has an index >= 16, and the
+    # range write ends at the Node256, which erases that child from the range
+    # and recomputes the Node256's maxVersion.
     with DebugConflictSet() as cs:
-        cs.addWrites(0, write(bytes([1])))
         for i in range(256):
-            cs.addWrites(1, write(bytes([1, i])))
-        cs.addWrites(2, write(bytes([0]), bytes([1])))
-        cs.check(read(0, bytes([1]), bytes([2])))
+            cs.addWrites(1, write(bytes([0x10, i])))
+        cs.addWrites(10, write(bytes([0x10, 0x50])))
+        cs.addWrites(11, write(bytes([0x05]), bytes([0x10])))
+        assert cs.check(read(5, bytes([0x10, 0x50]))) == [Result.CONFLICT]
 
 
 def test_large_removal_buffer():

update-readme.sh

@@ -1,7 +1,8 @@
 #!/bin/bash
 
 # Disable frequency scaling
-for i in $(seq "$(nproc)") ; do sudo cat /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_max_freq | sudo tee /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_min_freq >/dev/null ; done
+# shellcheck disable=SC2002
+for i in $(seq "$(nproc)") ; do cat /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_max_freq | sudo tee /sys/devices/system/cpu/cpu$((i-1))/cpufreq/scaling_min_freq >/dev/null ; done
 
 cmake -S. -Bbuild -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=clang++ -G Ninja -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_CXX_FLAGS=-DNVALGRIND > /dev/null
 cmake --build build > /dev/null