i#6938 sched migrate: Make initial_cur_time atomic (#7043)

Fixes a race in the drmemtrace scheduler by making
output_info_t.initial_cur_time atomic.

Tested on ThreadSanitizer on the internal test where this was first
reported.

Issue: #6938

clients/drcachesim/scheduler/scheduler.cpp

@@ -2784,7 +2784,8 @@ scheduler_tmpl_t<RecordType, ReaderType>::pop_from_ready_queue_hold_locks(
                         // For never-executed inputs we consider their last execution
                         // to be the very first simulation time, which we can't
                         // easily initialize until here.
-                        res->last_run_time = outputs_[from_output].initial_cur_time;
+                        res->last_run_time = outputs_[from_output].initial_cur_time->load(
+                            std::memory_order_acquire);
                     }
                     VPRINT(this, 5,
                            "migration check %d to %d: cur=%" PRIu64 " last=%" PRIu64
@@ -3800,8 +3801,8 @@ scheduler_tmpl_t<RecordType, ReaderType>::next_record(output_ordinal_t output,
         cur_time = 1 + outputs_[output].stream->get_output_instruction_ordinal() +
             outputs_[output].idle_count;
     }
-    if (outputs_[output].initial_cur_time == 0) {
-        outputs_[output].initial_cur_time = cur_time;
+    if (outputs_[output].initial_cur_time->load(std::memory_order_acquire) == 0) {
+        outputs_[output].initial_cur_time->store(cur_time, std::memory_order_release);
     }
     // Invalid values for cur_time are checked below.
     outputs_[output].cur_time->store(cur_time, std::memory_order_release);

clients/drcachesim/scheduler/scheduler.h

@@ -1645,6 +1645,9 @@ template <typename RecordType, typename ReaderType> class scheduler_tmpl_t {
             cur_time =
                 std::unique_ptr<std::atomic<uint64_t>>(new std::atomic<uint64_t>());
             cur_time->store(0, std::memory_order_relaxed);
+            initial_cur_time =
+                std::unique_ptr<std::atomic<uint64_t>>(new std::atomic<uint64_t>());
+            initial_cur_time->store(0, std::memory_order_relaxed);
             record_index = std::unique_ptr<std::atomic<int>>(new std::atomic<int>());
             record_index->store(0, std::memory_order_relaxed);
         }
@@ -1693,7 +1696,9 @@ template <typename RecordType, typename ReaderType> class scheduler_tmpl_t {
         // Indirected so we can store it in our vector.
         std::unique_ptr<std::atomic<uint64_t>> cur_time;
         // The first simulation time passed to this output.
-        uint64_t initial_cur_time = 0;
+        // This is accessed by other outputs for stealing and rebalancing.
+        // Indirected so we can store it in our vector.
+        std::unique_ptr<std::atomic<uint64_t>> initial_cur_time;
         // Used for MAP_TO_RECORDED_OUTPUT get_output_cpuid().
         int64_t as_traced_cpuid = -1;
         // Used for MAP_AS_PREVIOUSLY with live_replay_output_count_.