[OpenMP][Tools] Add omp_all_memory support for Archer

[protze.joachim]

Summary

The semantic of depend(out:omp_all_memory) is quite similar to taskwait in that it separates all tasks (with dependency) created before an all_memory-task from all tasks (with dependency) created after an all_memory-task.
Only a single of such tasks can execute at a time. Similar to taskwait, we have a CV (AllMemory[1]) in the generating task to express the dependency sink semantic of an all_memory-task. In addition, AllMemory[0] describes the dependency source semantic of an all_memory-task. All tasks with dependency create an HB-arc towards the sink and terminate an HB-arc from the source.

Since we expect that not many applications will use such dependency, the support for handling the synchronization semantic is off by default and can be turned on using ARCHER_OPTION="all_memory=1". The most costly part is the precautionary posting of an HB-arc towards the sink, which represents a potentially contentious write from all concurrently executing sibling tasks. A warning is printed at runtime, when the option is off while such dependency is observed. In most cases the lazy activation will still lead to false alerts.

[Hahnfeld]

Sorry for the delay, looks good to me!

openmp/tools/archer/README.md

 the sequential part of an OpenMP program. (Only effective when OpenMP
 runtime is initialized. In doubt, insert omp_get_max_threads() as first
 statement in main!)</td>
 </tr>
 </tbody>
 
 <tbody>
 <tr>
+<td class="org-left">all&#95;memory</td>
+<td class="org-right">0</td>
+<td class="org-left">Turn on tracking and analysis of omp_all_memory
+dependencies. Archer will activate the support automatically when
+such dependency is seen during execution. At this time the analysis
+already missed synchronization semantics, which will lead to false
+reports in most cases.</td>
+</tr>
+</tbody>
+
+<tbody>
+<tr>
 <td class="org-left">report&#95;data&#95;leak</td>
 <td class="org-right">0</td>
 <td class="org-left">Report leaking OMPT data for execution under
 Archer. Used for testing and debugging Archer if errors occur.</td>
 </tr>
 </tbody>
 
 <tbody>

openmp/tools/archer/ompt-tsan.cpp

 #if (LLVM_VERSION) >= 40
   int flush_shadow{0};
 #endif
   int print_max_rss{0};
   int verbose{0};
   int enabled{1};
   int report_data_leak{0};
   int ignore_serial{0};
+  std::atomic<int> all_memory{0};
 
   ArcherFlags(const char *env) {
     if (env) {
       std::vector<std::string> tokens;
       std::string token;
       std::string str(env);
       std::istringstream iss(str);
+      int tmp_int;
       while (std::getline(iss, token, ' '))
         tokens.push_back(token);
 
       for (std::vector<std::string>::iterator it = tokens.begin();
            it != tokens.end(); ++it) {
 #if (LLVM_VERSION) >= 40
         if (sscanf(it->c_str(), "flush_shadow=%d", &flush_shadow))
           continue;
 #endif
         if (sscanf(it->c_str(), "print_max_rss=%d", &print_max_rss))
           continue;
         if (sscanf(it->c_str(), "verbose=%d", &verbose))
           continue;
         if (sscanf(it->c_str(), "report_data_leak=%d", &report_data_leak))
           continue;
         if (sscanf(it->c_str(), "enable=%d", &enabled))
           continue;
         if (sscanf(it->c_str(), "ignore_serial=%d", &ignore_serial))
           continue;
+        if (sscanf(it->c_str(), "all_memory=%d", &tmp_int)) {
+          all_memory = tmp_int;
+          continue;
+        }
         std::cerr << "Illegal values for ARCHER_OPTIONS variable: " << token
                   << std::endl;
       }
     }
   }
 };
 
 class TsanFlags {

 struct TaskData final : DataPoolEntry<TaskData> {
   /// Its address is used for relationships of this task.
   ompt_tsan_clockid Task{0};
 
   /// Child tasks use its address to declare a relationship to a taskwait in
   /// this task.
   ompt_tsan_clockid Taskwait{0};
 
+  /// Child tasks use its address to model omp_all_memory dependencies
+  ompt_tsan_clockid AllMemory[2]{0};
+
   /// Whether this task is currently executing a barrier.
   bool InBarrier{false};
 
   /// Whether this task is an included task.
   int TaskType{0};
 
   /// count execution phase
   int execution{0};
#endif
   bool isMergable() { return TaskType & ompt_task_mergeable; }
   bool isMerged() { return TaskType & ompt_task_merged; }
 
   bool isExplicit() { return TaskType & ompt_task_explicit; }
   bool isImplicit() { return TaskType & ompt_task_implicit; }
   bool isInitial() { return TaskType & ompt_task_initial; }
   bool isTarget() { return TaskType & ompt_task_target; }
 
+  void setAllMemoryDep() { AllMemory[0] = 1; }
+  bool hasAllMemoryDep() { return AllMemory[0]; }
+
   void *GetTaskPtr() { return &Task; }
 
   void *GetTaskwaitPtr() { return &Taskwait; }
 
+  void *GetLastAllMemoryPtr() { return AllMemory; }
+  void *GetNextAllMemoryPtr() { return AllMemory + 1; }
+
   TaskData *Init(TaskData *parent, int taskType) {
     TaskType = taskType;
     Parent = parent;
     Team = Parent->Team;
     if (Parent != nullptr) {
       Parent->RefCount++;
       // Copy over pointer to taskgroup. This task may set up its own stack
       // but for now belongs to its parent's taskgroup.

 static void freeTask(TaskData *task) {
   while (task != nullptr && --task->RefCount == 0) {
     TaskData *Parent = task->Parent;
     task->Delete();
     task = Parent;
   }
 }
 
+// LastAllMemoryPtr marks the beginning of an all_memory epoch
+// NextAllMemoryPtr marks the end of an all_memory epoch
+// All tasks with depend begin execution after LastAllMemoryPtr
+// and end before NextAllMemoryPtr
 static void releaseDependencies(TaskData *task) {
+  if (archer_flags->all_memory) {
+    if (task->hasAllMemoryDep()) {
+      TsanHappensBefore(task->Parent->GetLastAllMemoryPtr());
+      TsanHappensBefore(task->Parent->GetNextAllMemoryPtr());
+    } else if (task->DependencyCount)
+      TsanHappensBefore(task->Parent->GetNextAllMemoryPtr());
+  }
   for (unsigned i = 0; i < task->DependencyCount; i++) {
     task->Dependencies[i].AnnotateEnd();
   }
 }
 
 static void acquireDependencies(TaskData *task) {
+  if (archer_flags->all_memory) {
+    if (task->hasAllMemoryDep())
+      TsanHappensAfter(task->Parent->GetNextAllMemoryPtr());
+    else if (task->DependencyCount)
+      TsanHappensAfter(task->Parent->GetLastAllMemoryPtr());
+  }
   for (unsigned i = 0; i < task->DependencyCount; i++) {
     task->Dependencies[i].AnnotateBegin();
   }
 }
 
 static void ompt_tsan_task_schedule(ompt_data_t *first_task_data,
                                     ompt_task_status_t prior_task_status,
                                     ompt_data_t *second_task_data) {
if (!Data->Parent) {
       return;
     }
     if (!Data->Parent->DependencyMap)
       Data->Parent->DependencyMap =
           new std::unordered_map<void *, DependencyData *>();
     Data->Dependencies =
         (TaskDependency *)malloc(sizeof(TaskDependency) * ndeps);
     Data->DependencyCount = ndeps;
-    for (int i = 0; i < ndeps; i++) {
+    for (int i = 0, d = 0; i < ndeps; i++, d++) {
+      if (deps[i].dependence_type == ompt_dependence_type_out_all_memory ||
+          deps[i].dependence_type == ompt_dependence_type_inout_all_memory) {
+        Data->setAllMemoryDep();
+        Data->DependencyCount--;
+        if (!archer_flags->all_memory) {
+          printf("The application uses omp_all_memory, but Archer was\n"
+                 "started to not consider omp_all_memory. This can lead\n"
+                 "to false data race alerts.\n"
+                 "Include all_memory=1 in ARCHER_OPTIONS to consider\n"
+                 "omp_all_memory from the beginning.\n");
+          archer_flags->all_memory = 1;
+        }
+        d--;
+        continue;
+      }
       auto ret = Data->Parent->DependencyMap->insert(
           std::make_pair(deps[i].variable.ptr, nullptr));
       if (ret.second) {
         ret.first->second = DependencyData::New();
       }
-      new ((void *)(Data->Dependencies + i))
+      new ((void *)(Data->Dependencies + d))
           TaskDependency(ret.first->second, deps[i].dependence_type);
     }
 
     // This callback is executed before this task is first started.
     TsanHappensBefore(Data->GetTaskPtr());
   }
 }
 

openmp/tools/archer/tests/task/omp_task_depend_all.c

+// RUN--: %libarcher-compile-and-run | FileCheck %s --check-prefix=NOENV
+// RUN: %libarcher-compile && env ARCHER_OPTIONS="all_memory=1" \
+// RUN:   %libarcher-run | FileCheck %s --check-prefix=ENV
+// REQUIRES: tsan
+
+// The runtime currently does not get dependency information from GCC.
+// UNSUPPORTED: gcc
+
+// Tests OMP 5.x task dependence "omp_all_memory",
+// emulates compiler codegen versions for new dep kind
+//
+// Task tree created:
+//      task0 - task1 (in: i1, i2)
+//             \
+//        task2 (inoutset: i2), (in: i1)
+//             /
+//        task3 (omp_all_memory) via flag=0x80
+//             /
+//      task4 - task5 (in: i1, i2)
+//           /
+//       task6 (omp_all_memory) via addr=-1
+//           /
+//       task7 (omp_all_memory) via flag=0x80
+//           /
+//       task8 (in: i3)
+//
+
+#include <omp.h>
+#include <stdio.h>
+
+#ifdef _WIN32
+#include <windows.h>
+#define mysleep(n) Sleep(n)
+#else
+#include <unistd.h>
+#define mysleep(n) usleep((n)*1000)
+#endif
+
+// to check the # of concurrent tasks (must be 1 for MTX, <3 for other kinds)
+static int checker = 0;
+static int err = 0;
+#ifndef DELAY
+#define DELAY 100
+#endif
+
+// ---------------------------------------------------------------------------
+// internal data to emulate compiler codegen
+typedef struct DEP {
+  size_t addr;
+  size_t len;
+  unsigned char flags;
+} dep;
+#define DEP_ALL_MEM 0x80
+typedef struct task {
+  void **shareds;
+  void *entry;
+  int part_id;
+  void *destr_thunk;
+  int priority;
+  long long device_id;
+  int f_priv;
+} task_t;
+#define TIED 1
+typedef int (*entry_t)(int, task_t *);
+typedef struct ID {
+  int reserved_1;
+  int flags;
+  int reserved_2;
+  int reserved_3;
+  char *psource;
+} id;
+// thunk routine for tasks with ALL dependency
+int thunk_m(int gtid, task_t *ptask) {
+  int lcheck, th;
+#pragma omp atomic capture
+  lcheck = ++checker;
+  th = omp_get_thread_num();
+  printf("task m_%d, th %d, checker %d\n", ptask->f_priv, th, lcheck);
+  if (lcheck != 1) { // no more than 1 task at a time
+    err++;
+    printf("Error m1, checker %d != 1\n", lcheck);
+  }
+  mysleep(DELAY);
+#pragma omp atomic read
+  lcheck = checker; // must still be equal to 1
+  if (lcheck != 1) {
+    err++;
+    printf("Error m2, checker %d != 1\n", lcheck);
+  }
+#pragma omp atomic
+  --checker;
+  return 0;
+}
+// thunk routine for tasks with inoutset dependency
+int thunk_s(int gtid, task_t *ptask) {
+  int lcheck, th;
+#pragma omp atomic capture
+  lcheck = ++checker; // 1
+  th = omp_get_thread_num();
+  printf("task 2_%d, th %d, checker %d\n", ptask->f_priv, th, lcheck);
+  if (lcheck != 1) { // no more than 1 task at a time
+    err++;
+    printf("Error s1, checker %d != 1\n", lcheck);
+  }
+  mysleep(DELAY);
+#pragma omp atomic read
+  lcheck = checker; // must still be equal to 1
+  if (lcheck != 1) {
+    err++;
+    printf("Error s2, checker %d != 1\n", lcheck);
+  }
+#pragma omp atomic
+  --checker;
+  return 0;
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+int __kmpc_global_thread_num(id *);
+task_t *__kmpc_omp_task_alloc(id *loc, int gtid, int flags, size_t sz,
+                              size_t shar, entry_t rtn);
+int __kmpc_omp_task_with_deps(id *loc, int gtid, task_t *task, int ndeps,
+                              dep *dep_lst, int nd_noalias, dep *noalias_lst);
+static id loc = {0, 2, 0, 0, ";file;func;0;0;;"};
+#ifdef __cplusplus
+} // extern "C"
+#endif
+// End of internal data
+// ---------------------------------------------------------------------------
+
+int main() {
+  char *ompx_all_memory = (void *)0xffffffffffffffff;
+  int i1, i2, i3;
+  omp_set_num_threads(8);
+  omp_set_dynamic(0);
+#pragma omp parallel
+  {
+#pragma omp single nowait
+    {
+      dep sdep[2];
+      task_t *ptr;
+      int gtid = __kmpc_global_thread_num(&loc);
+      int t = omp_get_thread_num();
+#pragma omp task depend(in : i1, i2)
+      { // task 0
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker; // 1 or 2
+        th = omp_get_thread_num();
+        printf("task 0_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck > 2 || lcheck < 1) {
+          err++; // no more than 2 tasks concurrently
+          printf("Error1, checker %d, not 1 or 2\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // 1 or 2
+        if (lcheck > 2 || lcheck < 1) {
+#pragma omp atomic
+          err++;
+          printf("Error2, checker %d, not 1 or 2\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+#pragma omp task depend(in : i1, i2)
+      { // task 1
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker; // 1 or 2
+        th = omp_get_thread_num();
+        printf("task 1_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck > 2 || lcheck < 1) {
+          err++; // no more than 2 tasks concurrently
+          printf("Error3, checker %d, not 1 or 2\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // 1 or 2
+        if (lcheck > 2 || lcheck < 1) {
+          err++;
+          printf("Error4, checker %d, not 1 or 2\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+      // compiler codegen start
+      // task2
+      ptr = __kmpc_omp_task_alloc(&loc, gtid, TIED, sizeof(task_t), 0, thunk_s);
+      sdep[0].addr = (size_t)&i1;
+      sdep[0].len = 0;   // not used
+      sdep[0].flags = 1; // IN
+      sdep[1].addr = (size_t)&i2;
+      sdep[1].len = 0;      // not used
+      sdep[1].flags = 8;    // INOUTSET
+      ptr->f_priv = t + 10; // init single first-private variable
+      __kmpc_omp_task_with_deps(&loc, gtid, ptr, 2, sdep, 0, 0);
+
+// task3
+#pragma omp task depend(in : i1) depend(inout : ompx_all_memory[0])
+      {
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker;
+        th = omp_get_thread_num();
+        printf("task 3_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck != 1) { // no more than 1 task at a time
+          err++;
+          printf("Error m1, checker %d != 1\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // must still be equal to 1
+        if (lcheck != 1) {
+          err++;
+          printf("Error m2, checker %d != 1\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+      // compiler codegen end
+#pragma omp task depend(in : i1, i2)
+      { // task 4
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker; // 1 or 2
+        th = omp_get_thread_num();
+        printf("task 4_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck > 2 || lcheck < 1) {
+          err++; // no more than 2 tasks concurrently
+          printf("Error5, checker %d, not 1 or 2\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // 1 or 2
+        if (lcheck > 2 || lcheck < 1) {
+          err++;
+          printf("Error6, checker %d, not 1 or 2\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+#pragma omp task depend(in : i1, i2)
+      { // task 5
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker; // 1 or 2
+        th = omp_get_thread_num();
+        printf("task 5_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck > 2 || lcheck < 1) {
+          err++; // no more than 2 tasks concurrently
+          printf("Error7, checker %d, not 1 or 2\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // 1 or 2
+        if (lcheck > 2 || lcheck < 1) {
+          err++;
+          printf("Error8, checker %d, not 1 or 2\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+// task6
+#pragma omp task depend(inout : ompx_all_memory[0])
+      {
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker;
+        th = omp_get_thread_num();
+        printf("task 6_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck != 1) { // no more than 1 task at a time
+          err++;
+          printf("Error m1, checker %d != 1\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // must still be equal to 1
+        if (lcheck != 1) {
+          err++;
+          printf("Error m2, checker %d != 1\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+// task7
+#pragma omp task depend(inout : ompx_all_memory[0]) depend(mutexinoutset : i3)
+      {
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker;
+        th = omp_get_thread_num();
+        printf("task 7_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck != 1) { // no more than 1 task at a time
+          err++;
+          printf("Error m1, checker %d != 1\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker; // must still be equal to 1
+        if (lcheck != 1) {
+          err++;
+          printf("Error m2, checker %d != 1\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+#pragma omp task depend(in : i3)
+      { // task 8
+        int lcheck, th;
+#pragma omp atomic capture
+        lcheck = ++checker; // 1
+        th = omp_get_thread_num();
+        printf("task 8_%d, th %d, checker %d\n", t, th, lcheck);
+        if (lcheck != 1) {
+          err++;
+          printf("Error9, checker %d, != 1\n", lcheck);
+        }
+        mysleep(DELAY);
+#pragma omp atomic read
+        lcheck = checker;
+        if (lcheck != 1) {
+          err++;
+          printf("Error10, checker %d, != 1\n", lcheck);
+        }
+#pragma omp atomic
+        --checker;
+      }
+    } // single
+  }   // parallel
+  if (err == 0 && checker == 0) {
+    printf("passed\n");
+    return 0;
+  } else {
+    printf("failed, err = %d, checker = %d\n", err, checker);
+    return 1;
+  }
+}
+
+// NOENV-NOT: ThreadSanitizer: data race
+// NOENV-NOT: ThreadSanitizer: reported
+// NOENV: omp_all_memory
+// NOENV-NOT: ThreadSanitizer: data race
+// NOENV-NOT: ThreadSanitizer: reported
+// NOENV: passed
+
+// ENV-NOT: ThreadSanitizer: data race
+// ENV-NOT: ThreadSanitizer: reported
+// ENV: passed