WIP

.pre-commit-config.yaml

@@ -24,6 +24,14 @@ repos:
         entry: "^#define SHOW_MEMORY 1$"
         language: pygrep
         types: [c++]
+  - repo: local
+    hooks:
+      - id: sim cache misses check
+        name: disallow checking in SIM_CACHE_MISSES=1
+        description: disallow checking in SIM_CACHE_MISSES=1
+        entry: "^#define SIM_CACHE_MISSES 1$"
+        language: pygrep
+        types: [c++]
   - repo: https://github.com/shellcheck-py/shellcheck-py
     rev: a23f6b85d0fdd5bb9d564e2579e678033debbdff # frozen: v0.10.0.1
     hooks:

CMakeLists.txt

@@ -350,6 +350,11 @@ 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

@@ -48,6 +48,17 @@ limitations under the License.
 #endif
 #endif
 
+#define SIM_CACHE_MISSES 0
+#if SIM_CACHE_MISSES
+constexpr void simCacheMiss(void *x) {
+  if (x) {
+    _mm_clflush(x);
+  }
+}
+#else
+constexpr void simCacheMiss(void *) {}
+#endif
+
 #include <memcheck.h>
 
 using namespace weaselab;
@@ -836,21 +847,29 @@ int getNodeIndex(Node16 *self, uint8_t index) {
 
 // Precondition - an entry for index must exist in the node
 Node *&getChildExists(Node3 *self, uint8_t index) {
-  return self->children[getNodeIndex(self, index)];
+  auto &result = self->children[getNodeIndex(self, index)];
+  simCacheMiss(result);
+  return result;
 }
 // Precondition - an entry for index must exist in the node
 Node *&getChildExists(Node16 *self, uint8_t index) {
-  return self->children[getNodeIndex(self, index)];
+  auto &result = self->children[getNodeIndex(self, index)];
+  simCacheMiss(result);
+  return result;
 }
 // Precondition - an entry for index must exist in the node
 Node *&getChildExists(Node48 *self, uint8_t index) {
   assert(self->bitSet.test(index));
-  return self->children[self->index[index]];
+  auto &result = self->children[self->index[index]];
+  simCacheMiss(result);
+  return result;
 }
 // Precondition - an entry for index must exist in the node
 Node *&getChildExists(Node256 *self, uint8_t index) {
   assert(self->bitSet.test(index));
-  return self->children[index];
+  auto &result = self->children[index];
+  simCacheMiss(result);
+  return result;
 }
 
 // Precondition - an entry for index must exist in the node
@@ -1030,6 +1049,7 @@ ChildAndMaxVersion getChildAndMaxVersion(Node3 *self, uint8_t index) {
   if (i < 0) {
     return {};
   }
+  simCacheMiss(self->children[i]);
   return {self->children[i], self->childMaxVersion[i]};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node16 *self, uint8_t index) {
@@ -1037,6 +1057,7 @@ ChildAndMaxVersion getChildAndMaxVersion(Node16 *self, uint8_t index) {
   if (i < 0) {
     return {};
   }
+  simCacheMiss(self->children[i]);
   return {self->children[i], self->childMaxVersion[i]};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node48 *self, uint8_t index) {
@@ -1044,9 +1065,11 @@ ChildAndMaxVersion getChildAndMaxVersion(Node48 *self, uint8_t index) {
   if (i < 0) {
     return {};
   }
+  simCacheMiss(self->children[i]);
   return {self->children[i], self->childMaxVersion[i]};
 }
 ChildAndMaxVersion getChildAndMaxVersion(Node256 *self, uint8_t index) {
+  simCacheMiss(self->children[index]);
   return {self->children[index], self->childMaxVersion[index]};
 }
 
@@ -1072,6 +1095,7 @@ Node *getChildGeq(Node0 *, int) { return nullptr; }
 Node *getChildGeq(Node3 *n, int child) {
   for (int i = 0; i < n->numChildren; ++i) {
     if (n->index[i] >= child) {
+      simCacheMiss(n->children[i]);
       return n->children[i];
     }
   }
@@ -1090,7 +1114,10 @@ Node *getChildGeq(Node16 *self, int child) {
   __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;
-  return bitfield == 0 ? nullptr : self->children[std::countr_zero(bitfield)];
+  auto *result =
+      bitfield == 0 ? nullptr : self->children[std::countr_zero(bitfield)];
+  simCacheMiss(result);
+  return result;
 #elif defined(HAS_ARM_NEON)
   uint8x16_t indices;
   memcpy(&indices, self->index, sizeof(self->index));
@@ -1126,13 +1153,16 @@ Node *getChildGeq(Node48 *self, int child) {
   if (c < 0) {
     return nullptr;
   }
-  return self->children[self->index[c]];
+  auto *result = self->children[self->index[c]];
+  simCacheMiss(result);
+  return result;
 }
 Node *getChildGeq(Node256 *self, int child) {
   int c = self->bitSet.firstSetGeq(child);
   if (c < 0) {
     return nullptr;
   }
+  simCacheMiss(self->children[c]);
   return self->children[c];
 }
 
@@ -1156,20 +1186,26 @@ Node *getChildGeq(Node *self, int child) {
 // Precondition: self has a child
 Node *getFirstChildExists(Node3 *self) {
   assert(self->numChildren > 0);
+  simCacheMiss(self->children[0]);
   return self->children[0];
 }
 // Precondition: self has a child
 Node *getFirstChildExists(Node16 *self) {
   assert(self->numChildren > 0);
+  simCacheMiss(self->children[0]);
   return self->children[0];
 }
 // Precondition: self has a child
 Node *getFirstChildExists(Node48 *self) {
-  return self->children[self->index[self->bitSet.firstSetGeq(0)]];
+  auto *result = self->children[self->index[self->bitSet.firstSetGeq(0)]];
+  simCacheMiss(result);
+  return result;
 }
 // Precondition: self has a child
 Node *getFirstChildExists(Node256 *self) {
-  return self->children[self->bitSet.firstSetGeq(0)];
+  auto *result = self->children[self->bitSet.firstSetGeq(0)];
+  simCacheMiss(result);
+  return result;
 }
 
 // Precondition: self has a child
@@ -3009,143 +3045,243 @@ Node *firstGeqPhysical(Node *n, const std::span<const uint8_t> key) {
   }
 }
 
-struct CheckJob {
+#ifndef __has_attribute
-  Node *n;
+#define __has_attribute(x) 0
-  std::span<const uint8_t> begin;
+#endif
-  InternalVersionT readVersion;
-  ReadContext *tls;
-  ConflictSet::Result *result;
 
+#if __has_attribute(musttail)
+#define MUSTTAIL __attribute__((musttail))
+#else
+#define MUSTTAIL
+#endif
+
+#if __has_attribute(preserve_none)
+#define CONTINUATION_CALLING_CONVENTION __attribute__((preserve_none))
+#else
+#define CONTINUATION_CALLING_CONVENTION
+#endif
+
+typedef CONTINUATION_CALLING_CONVENTION void (*continuation)(struct CheckAll *,
+                                                             int64_t prevJob,
+                                                             int64_t job,
+                                                             int64_t started,
+                                                             int64_t count);
+
+// State relevant to a particular query
+struct CheckJob {
   void setResult(bool ok) {
     *result = ok ? ConflictSet::Commit : ConflictSet::Conflict;
   }
 
-  typedef void (*typeErasedContinuation)(void *);
+  [[nodiscard]] continuation init(const ConflictSet::ReadRange *read,
+                                  ConflictSet::Result *result, Node *root,
+                                  int64_t oldestVersionFullPrecision,
+                                  ReadContext *tls);
 
-  // The type of a function that takes a CheckJob* and returns its own type
+  Node *n;
-  struct continuation {
+  ChildAndMaxVersion childAndVersion;
-    typedef continuation (*functionPtrType)(CheckJob *);
+  std::span<const uint8_t> begin;
-    functionPtrType func;
+  InternalVersionT readVersion;
-    continuation operator()(CheckJob *job) { return func(job); }
+  ConflictSet::Result *result;
-    /*implicit*/ continuation(functionPtrType func) : func(func) {}
-    continuation() = default;
-    operator bool() { return func != nullptr; }
-  };
-
-  continuation next;
-  void init(const ConflictSet::ReadRange *read, ConflictSet::Result *result,
-            Node *root, int64_t oldestVersionFullPrecision, ReadContext *tls);
 };
 
+// State relevant to all queries
+struct CheckAll {
+  constexpr static int kConcurrent = 32;
+  const ConflictSet::ReadRange *queries;
+  ConflictSet::Result *results;
+  CheckJob inProgress[kConcurrent];
+  continuation next[kConcurrent];
+  int nextJob[kConcurrent];
+  Node *root;
+  int64_t oldestVersionFullPrecision;
+  ReadContext *tls;
+};
+
+CONTINUATION_CALLING_CONVENTION void keepGoing(CheckAll *context,
+                                               int64_t prevJob, int64_t job,
+                                               int64_t started, int64_t count) {
+  prevJob = job;
+  job = context->nextJob[job];
+  MUSTTAIL return context->next[job](context, prevJob, job, started, count);
+}
+
+CONTINUATION_CALLING_CONVENTION void complete(CheckAll *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 {
+    int temp = started++;
+    context->next[job] = context->inProgress[job].init(
+        context->queries + temp, context->results + temp, context->root,
+        context->oldestVersionFullPrecision, context->tls);
+  }
+  MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+}
+
 namespace check_point_read_state_machine {
 
-CheckJob::continuation down_left_spine(CheckJob *job);
+CONTINUATION_CALLING_CONVENTION void
+down_left_spine(struct CheckAll *, int64_t prevJob, int64_t job,
+                int64_t started, int64_t count);
+CONTINUATION_CALLING_CONVENTION void iter(struct CheckAll *, int64_t prevJob,
+                                          int64_t job, int64_t started,
+                                          int64_t count);
+CONTINUATION_CALLING_CONVENTION void begin(struct CheckAll *, int64_t prevJob,
+                                           int64_t job, int64_t started,
+                                           int64_t count);
 
-// Logically this is the same as performing firstGeq and then checking against
+void begin(struct CheckAll *context, int64_t prevJob, int64_t job,
-// point or range version according to cmp, but this version short circuits as
+           int64_t started, int64_t count) {
-// soon as it can prove that there's no conflict.
+  ++context->tls->point_read_accum;
-CheckJob::continuation begin(CheckJob *job) {
-  ++job->tls->point_read_accum;
 #if DEBUG_VERBOSE && !defined(NDEBUG)
   fprintf(stderr, "Check point read: %s\n", printable(key).c_str());
 #endif
-  for (;; ++job->tls->point_read_iterations_accum) {
+  auto *j = context->inProgress + job;
-    if (job->begin.size() == 0) {
+
-      if (job->n->entryPresent) {
+  if (j->begin.size() == 0) {
-        job->setResult(job->n->entry.pointVersion <= job->readVersion);
+    if (j->n->entryPresent) {
-        return nullptr; // Done
+      j->setResult(j->n->entry.pointVersion <= j->readVersion);
+      MUSTTAIL return complete(context, prevJob, job, started, count);
     }
-      job->n = getFirstChildExists(job->n);
+    j->n = getFirstChildExists(j->n);
-      return down_left_spine;
+    context->next[job] = down_left_spine;
+    __builtin_prefetch(j->n);
+    MUSTTAIL return keepGoing(context, prevJob, job, started, count);
   }
 
-    auto [child, maxV] = getChildAndMaxVersion(job->n, job->begin[0]);
+  j->childAndVersion = getChildAndMaxVersion(j->n, j->begin[0]);
-    if (child == nullptr) {
+  context->next[job] = iter;
-      auto c = getChildGeq(job->n, job->begin[0]);
+  __builtin_prefetch(j->childAndVersion.child);
-      if (c != nullptr) {
+  MUSTTAIL return keepGoing(context, prevJob, job, started, count);
-        job->n = c;
-        return down_left_spine;
-      } else {
-        job->n = nextSibling(job->n);
-        if (job->n == nullptr) {
-          job->setResult(true);
-          return nullptr; // Done
-        }
-        return down_left_spine;
-      }
-    }
-
-    job->n = child;
-    job->begin = job->begin.subspan(1, job->begin.size() - 1);
-
-    if (job->n->partialKeyLen > 0) {
-      int commonLen = std::min<int>(job->n->partialKeyLen, job->begin.size());
-      int i = longestCommonPrefix(job->n->partialKey(), job->begin.data(),
-                                  commonLen);
-      if (i < commonLen) {
-        auto c = job->n->partialKey()[i] <=> job->begin[i];
-        if (c > 0) {
-          return down_left_spine;
-        } else {
-          job->n = nextSibling(job->n);
-          if (job->n == nullptr) {
-            job->setResult(true);
-            return nullptr; // Done
-          }
-          return down_left_spine;
-        }
-      }
-      if (commonLen == job->n->partialKeyLen) {
-        // partial key matches
-        job->begin =
-            job->begin.subspan(commonLen, job->begin.size() - commonLen);
-      } else if (job->n->partialKeyLen > int(job->begin.size())) {
-        // n is the first physical node greater than remaining, and there's no
-        // eq node
-        return down_left_spine;
-      }
-    }
-
-    if (maxV <= job->readVersion) {
-      ++job->tls->point_read_short_circuit_accum;
-      job->setResult(true);
-      return nullptr; // Done
-    }
-  }
 }
 
-CheckJob::continuation down_left_spine(CheckJob *job) {
+void iter(struct CheckAll *context, int64_t prevJob, int64_t job,
-  if (job->n->entryPresent) {
+          int64_t started, int64_t count) {
-    job->setResult(job->n->entry.rangeVersion <= job->readVersion);
+  auto *j = context->inProgress + job;
-    return nullptr; // Done
+  if (j->childAndVersion.child == nullptr) {
+    auto c = getChildGeq(j->n, j->begin[0]);
+    if (c != nullptr) {
+      j->n = c;
+      context->next[job] = down_left_spine;
+      __builtin_prefetch(j->n);
+      MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+    } else {
+      j->n = nextSibling(j->n);
+      if (j->n == nullptr) {
+        j->setResult(true);
+        MUSTTAIL return complete(context, prevJob, job, started, count);
       }
-  job->n = getFirstChildExists(job->n);
+      context->next[job] = down_left_spine;
-  return down_left_spine;
+      __builtin_prefetch(j->n);
+      MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+    }
+  }
+
+  j->n = j->childAndVersion.child;
+  j->begin = j->begin.subspan(1, j->begin.size() - 1);
+
+  if (j->n->partialKeyLen > 0) {
+    int commonLen = std::min<int>(j->n->partialKeyLen, j->begin.size());
+    int i = longestCommonPrefix(j->n->partialKey(), j->begin.data(), commonLen);
+    if (i < commonLen) {
+      auto c = j->n->partialKey()[i] <=> j->begin[i];
+      if (c > 0) {
+        context->next[job] = down_left_spine;
+        MUSTTAIL return down_left_spine(context, prevJob, job, started, count);
+      } else {
+        j->n = nextSibling(j->n);
+        if (j->n == nullptr) {
+          j->setResult(true);
+          MUSTTAIL return complete(context, prevJob, job, started, count);
+        }
+        context->next[job] = down_left_spine;
+        __builtin_prefetch(j->n);
+        MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+      }
+    }
+    if (commonLen == j->n->partialKeyLen) {
+      // partial key matches
+      j->begin = j->begin.subspan(commonLen, j->begin.size() - commonLen);
+    } else if (j->n->partialKeyLen > int(j->begin.size())) {
+      // n is the first physical node greater than remaining, and there's no
+      // eq node
+      context->next[job] = down_left_spine;
+      MUSTTAIL return down_left_spine(context, prevJob, job, started, count);
+    }
+  }
+
+  if (j->childAndVersion.maxVersion <= j->readVersion) {
+    ++context->tls->point_read_short_circuit_accum;
+    j->setResult(true);
+    MUSTTAIL return complete(context, prevJob, job, started, count);
+  }
+
+  ++context->tls->point_read_iterations_accum;
+
+  if (j->begin.size() == 0) {
+    if (j->n->entryPresent) {
+      j->setResult(j->n->entry.pointVersion <= j->readVersion);
+      MUSTTAIL return complete(context, prevJob, job, started, count);
+    }
+    j->n = getFirstChildExists(j->n);
+    context->next[job] = down_left_spine;
+    __builtin_prefetch(j->n);
+    MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+  }
+
+  j->childAndVersion = getChildAndMaxVersion(j->n, j->begin[0]);
+  __builtin_prefetch(j->childAndVersion.child);
+  // j->next is already iter
+  MUSTTAIL return keepGoing(context, prevJob, job, started, count);
+}
+
+void down_left_spine(struct CheckAll *context, int64_t prevJob, int64_t job,
+                     int64_t started, int64_t count) {
+  auto *j = context->inProgress + job;
+  if (j->n->entryPresent) {
+    j->setResult(j->n->entry.rangeVersion <= j->readVersion);
+    MUSTTAIL return complete(context, prevJob, job, started, count);
+  }
+  j->n = getFirstChildExists(j->n);
+  __builtin_prefetch(j->n);
+  // j->next is already down_left_spine
+  MUSTTAIL return keepGoing(context, prevJob, job, started, count);
 }
 
 } // namespace check_point_read_state_machine
 
-void CheckJob::init(const ConflictSet::ReadRange *read,
+continuation CheckJob::init(const ConflictSet::ReadRange *read,
                             ConflictSet::Result *result, Node *root,
-                    int64_t oldestVersionFullPrecision, ReadContext *tls) {
+                            int64_t oldestVersionFullPrecision,
+                            ReadContext *tls) {
   auto begin = std::span<const uint8_t>(read->begin.p, read->begin.len);
   auto end = std::span<const uint8_t>(read->end.p, read->end.len);
   if (read->readVersion < oldestVersionFullPrecision) {
     *result = ConflictSet::TooOld;
-    next = +[](CheckJob *) -> continuation { return nullptr; };
+    return complete;
   } else if (end.size() == 0) {
     this->begin = begin;
     this->n = root;
     this->readVersion = InternalVersionT(read->readVersion);
     this->result = result;
-    this->tls = tls;
+    return check_point_read_state_machine::begin;
-    this->next = check_point_read_state_machine::begin;
+    // *result =
+    //     checkPointRead(root, begin, InternalVersionT(read->readVersion), tls)
+    //         ? ConflictSet::Commit
+    //         : ConflictSet::Conflict;
+    // return complete;
   } else {
     *result = checkRangeRead(root, begin, end,
                              InternalVersionT(read->readVersion), tls)
                   ? ConflictSet::Commit
                   : ConflictSet::Conflict;
-    next = +[](CheckJob *) -> continuation { return nullptr; };
+    return complete;
   }
 }
 
@@ -3163,38 +3299,23 @@ struct __attribute__((visibility("hidden"))) ConflictSet::Impl {
     tls.impl = this;
     int64_t check_byte_accum = 0;
 
-    constexpr int kConcurrent = 32;
+    CheckAll context;
-    CheckJob inProgress[kConcurrent];
+    context.oldestVersionFullPrecision = oldestVersionFullPrecision;
-    int nextJob[kConcurrent];
+    context.queries = reads;
+    context.results = result;
+    context.root = root;
+    context.tls = &tls;
 
-    int started = std::min(kConcurrent, count);
+    int64_t started = std::min(context.kConcurrent, count);
     for (int i = 0; i < started; i++) {
-      inProgress[i].init(reads + i, result + i, root,
+      context.next[i] = context.inProgress[i].init(
-                         oldestVersionFullPrecision, &tls);
+          reads + i, result + i, root, oldestVersionFullPrecision, &tls);
-      nextJob[i] = i + 1;
+      context.nextJob[i] = i + 1;
     }
-    nextJob[started - 1] = 0;
+    context.nextJob[started - 1] = 0;
-
     int prevJob = started - 1;
     int job = 0;
-    for (;;) {
+    context.next[job](&context, prevJob, job, started, count);
-      auto next = inProgress[job].next(inProgress + job);
-      inProgress[job].next = next;
-      if (!next) {
-        if (started == count) {
-          if (prevJob == job)
-            break;
-          nextJob[prevJob] = nextJob[job];
-          job = prevJob;
-        } else {
-          int temp = started++;
-          inProgress[job].init(reads + temp, result + temp, root,
-                               oldestVersionFullPrecision, &tls);
-        }
-      }
-      prevJob = job;
-      job = nextJob[job];
-    }
 
     for (int i = 0; i < count; ++i) {
       assert(reads[i].readVersion >= 0);

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"),

ServerBench.cpp

@@ -1,4 +1,5 @@
 #include <atomic>
+#include <cstdint>
 #include <errno.h>
 #include <netdb.h>
 #include <stdio.h>
@@ -21,78 +22,55 @@
 
 std::atomic<int64_t> transactions;
 
-constexpr int kBaseSearchDepth = 115;
 constexpr int kWindowSize = 10000000;
 
-std::string numToKey(int64_t num) {
+constexpr int kNumPrefixes = 250000;
+
+std::string makeKey(int64_t num, int suffixLen) {
   std::string result;
-  result.resize(kBaseSearchDepth + sizeof(int64_t));
+  result.resize(sizeof(int64_t) + suffixLen);
-  memset(result.data(), 0, kBaseSearchDepth);
   int64_t be = __builtin_bswap64(num);
-  memcpy(result.data() + kBaseSearchDepth, &be, sizeof(int64_t));
+  memcpy(result.data(), &be, sizeof(int64_t));
+  memset(result.data() + sizeof(int64_t), 0, suffixLen);
   return result;
 }
 
 void workload(weaselab::ConflictSet *cs) {
   int64_t version = kWindowSize;
-  cs->addWrites(nullptr, 0, version);
+  for (int i = 0; i < kNumPrefixes; ++i) {
-  for (;; transactions.fetch_add(1, std::memory_order_relaxed)) {
+    for (int j = 0; j < 50; ++j) {
-    // Reads
+      weaselab::ConflictSet::WriteRange wr;
-    {
+      auto k = makeKey(i, j);
-      auto beginK = numToKey(version - kWindowSize);
+      wr.begin.p = (const uint8_t *)k.data();
-      auto endK = numToKey(version - 1);
+      wr.begin.len = k.size();
-      auto pointRv = version - kWindowSize + rand() % kWindowSize + 1;
+      wr.end.len = 0;
-      auto pointK = numToKey(pointRv);
+      cs->addWrites(&wr, 1, version);
-      weaselab::ConflictSet::ReadRange reads[] = {
-          {
-              {(const uint8_t *)pointK.data(), int(pointK.size())},
-              {nullptr, 0},
-              pointRv,
-          },
-          {
-              {(const uint8_t *)beginK.data(), int(beginK.size())},
-              {(const uint8_t *)endK.data(), int(endK.size())},
-              version - 2,
-          },
-      };
-      weaselab::ConflictSet::Result result[sizeof(reads) / sizeof(reads[0])];
-      cs->check(reads, result, sizeof(reads) / sizeof(reads[0]));
-      // for (int i = 0; i < sizeof(reads) / sizeof(reads[0]); ++i) {
-      //   if (result[i] != weaselab::ConflictSet::Commit) {
-      //     fprintf(stderr, "Unexpected conflict: [%s, %s) @ %" PRId64 "\n",
-      //             printable(reads[i].begin).c_str(),
-      //             printable(reads[i].end).c_str(), reads[i].readVersion);
-      //     abort();
-      //   }
-      // }
-    }
-    // Writes
-    {
-      weaselab::ConflictSet::WriteRange w;
-      auto k = numToKey(version);
-      w.begin.p = (const uint8_t *)k.data();
-      w.end.len = 0;
-      if (version % (kWindowSize / 2) == 0) {
-        for (int l = 0; l <= k.size(); ++l) {
-          w.begin.len = l;
-          cs->addWrites(&w, 1, version);
-        }
-      } else {
-        w.begin.len = k.size();
-        cs->addWrites(&w, 1, version);
-        int64_t beginN = version - kWindowSize + rand() % kWindowSize;
-        auto b = numToKey(beginN);
-        auto e = numToKey(beginN + 1000);
-        w.begin.p = (const uint8_t *)b.data();
-        w.begin.len = b.size();
-        w.end.p = (const uint8_t *)e.data();
-        w.end.len = e.size();
-        cs->addWrites(&w, 1, version);
     }
   }
-    // GC
-    cs->setOldestVersion(version - kWindowSize);
   ++version;
+  for (int i = 0; i < kNumPrefixes; ++i) {
+    weaselab::ConflictSet::WriteRange wr;
+    auto k = makeKey(i, 50);
+    wr.begin.p = (const uint8_t *)k.data();
+    wr.begin.len = k.size();
+    wr.end.len = 0;
+    cs->addWrites(&wr, 1, version);
+  }
+
+  std::vector<weaselab::ConflictSet::Result> results(10);
+  for (;; transactions.fetch_add(1, std::memory_order_relaxed)) {
+    std::vector<std::string> keys(10);
+    for (auto &k : keys) {
+      k = makeKey(rand() % kNumPrefixes, 49);
+    }
+    std::vector<weaselab::ConflictSet::ReadRange> reads(10);
+    for (int i = 0; i < reads.size(); ++i) {
+      reads[i].begin.p = (const uint8_t *)(keys[i].data());
+      reads[i].begin.len = keys[i].size();
+      reads[i].end.len = 0;
+      reads[i].readVersion = version - 1;
+    }
+    cs->check(reads.data(), results.data(), 10);
   }
 }