aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorTom Tromey <tromey@adacore.com>2022-12-13 12:03:34 -0700
committerTom Tromey <tromey@adacore.com>2023-01-17 07:06:56 -0700
commit6107546876df070688d75d8cd87d7c7a5f246091 (patch)
tree6092d20796d031fa4583af8c3f69dd4531626b8e
parent2ec9694617bc3a84ef89588d849eaef8893c0ec4 (diff)
downloadgdb-6107546876df070688d75d8cd87d7c7a5f246091.zip
gdb-6107546876df070688d75d8cd87d7c7a5f246091.tar.gz
gdb-6107546876df070688d75d8cd87d7c7a5f246091.tar.bz2
Avoid submitting empty tasks in parallel_for_each
I found that parallel_for_each would submit empty tasks to the thread pool. For example, this can happen if the number of tasks is smaller than the number of available threads. In the DWARF reader, this resulted in the cooked index containing empty sub-indices. This patch arranges to instead shrink the result vector and process the trailing entries in the calling thread. (cherry picked from commit 63078a04984b73e1fdeb4571a63605ee5c13f929)
-rw-r--r--gdb/unittests/parallel-for-selftests.c39
-rw-r--r--gdbsupport/parallel-for.h30
2 files changed, 69 insertions, 0 deletions
diff --git a/gdb/unittests/parallel-for-selftests.c b/gdb/unittests/parallel-for-selftests.c
index 3162db1..15a095a 100644
--- a/gdb/unittests/parallel-for-selftests.c
+++ b/gdb/unittests/parallel-for-selftests.c
@@ -149,6 +149,45 @@ TEST (int n_threads)
SELF_CHECK (counter == NUMBER);
#undef NUMBER
+
+ /* Check that if there are fewer tasks than threads, then we won't
+ end up with a null result. */
+ std::vector<std::unique_ptr<int>> intresults;
+ std::atomic<bool> any_empty_tasks (false);
+
+ FOR_EACH (1, 0, 1,
+ [&] (int start, int end)
+ {
+ if (start == end)
+ any_empty_tasks = true;
+ return std::unique_ptr<int> (new int (end - start));
+ });
+ SELF_CHECK (!any_empty_tasks);
+ SELF_CHECK (std::all_of (intresults.begin (),
+ intresults.end (),
+ [] (const std::unique_ptr<int> &entry)
+ {
+ return entry != nullptr;
+ }));
+
+ /* The same but using the task size parameter. */
+ intresults.clear ();
+ any_empty_tasks = false;
+ FOR_EACH (1, 0, 1,
+ [&] (int start, int end)
+ {
+ if (start == end)
+ any_empty_tasks = true;
+ return std::unique_ptr<int> (new int (end - start));
+ },
+ task_size_one);
+ SELF_CHECK (!any_empty_tasks);
+ SELF_CHECK (std::all_of (intresults.begin (),
+ intresults.end (),
+ [] (const std::unique_ptr<int> &entry)
+ {
+ return entry != nullptr;
+ }));
}
#endif /* FOR_EACH */
diff --git a/gdbsupport/parallel-for.h b/gdbsupport/parallel-for.h
index b565676..de9ebb1 100644
--- a/gdbsupport/parallel-for.h
+++ b/gdbsupport/parallel-for.h
@@ -70,6 +70,12 @@ public:
return result;
}
+ /* Resize the results to N. */
+ void resize (size_t n)
+ {
+ m_futures.resize (n);
+ }
+
private:
/* A vector of futures coming from the tasks run in the
@@ -108,6 +114,12 @@ public:
}
}
+ /* Resize the results to N. */
+ void resize (size_t n)
+ {
+ m_futures.resize (n);
+ }
+
private:
std::vector<gdb::future<void>> m_futures;
@@ -232,6 +244,24 @@ parallel_for_each (unsigned n, RandomIt first, RandomIt last,
end = j;
remaining_size -= chunk_size;
}
+
+ /* This case means we don't have enough elements to really
+ distribute them. Rather than ever submit a task that does
+ nothing, we short-circuit here. */
+ if (first == end)
+ end = last;
+
+ if (end == last)
+ {
+ /* We're about to dispatch the last batch of elements, which
+ we normally process in the main thread. So just truncate
+ the result list here. This avoids submitting empty tasks
+ to the thread pool. */
+ count = i;
+ results.resize (count);
+ break;
+ }
+
if (parallel_for_each_debug)
{
debug_printf (_("Parallel for: elements on worker thread %i\t: %zu"),