review comments

numpy · charris · Feb 23, 2025 · Feb 20, 2025 · Feb 20, 2025 · Feb 20, 2025
commit ef3c7d08a64fd5be559859818f319739c2144c52
diff --git a/.github/workflows/compiler_sanitizers.yml b/.github/workflows/compiler_sanitizers.yml
@@ -121,7 +121,7 @@ jobs:
    - name: Test
      run: |
        # These tests are slow, so only run tests in files that do "import threading" to make them count
-        TSAN_OPTIONS="allocator_may_return_null=1:halt_on_error=1:suppressions=tools\ci\tsan_suppressions.txt" \
+        TSAN_OPTIONS="allocator_may_return_null=1:suppressions=/Users/runner/work/numpy/numpy/tools/ci/tsan_suppressions.txt" \
        python -m spin test \
        `find numpy -name "test*.py" | xargs grep -l "import threading" | tr '\n' ' '` \
        -- -v -s --timeout=600 --durations=10
diff --git a/numpy/_core/src/multiarray/item_selection.c b/numpy/_core/src/multiarray/item_selection.c
@@ -2893,10 +2893,10 @@ PyArray_Nonzero(PyArrayObject *self)
             * the fast bool count is followed by this sparse path is faster
             * than combining the two loops, even for larger arrays
             */
+            npy_intp * multi_index_end = multi_index + nonzero_count;
            if (((double)nonzero_count / count) <= 0.1) {
                npy_intp subsize;
                npy_intp j = 0;
-                npy_intp * multi_index_end = multi_index + nonzero_count;
                while (multi_index < multi_index_end) {
                    npy_memchr(data + j * stride, 0, stride, count - j,
                               &subsize, 1);
@@ -2912,7 +2912,6 @@ PyArray_Nonzero(PyArrayObject *self)
             * stalls that are very expensive on most modern processors.
             */
            else {
-                npy_intp *multi_index_end = multi_index + nonzero_count;
                npy_intp j = 0;

                /* Manually unroll for GCC and maybe other compilers */

diff --git a/numpy/_core/tests/test_multithreading.py b/numpy/_core/tests/test_multithreading.py
@@ -274,16 +274,21 @@ def closure(b):

 @pytest.mark.parametrize("dtype", [bool, int, float])
 def test_nonzero_bool(dtype):
+    # See: gh-28361
+    #
    # np.nonzero uses np.count_nonzero to determine the size of the output array
    # In a second pass the indices of the non-zero elements are determined, but they can have changed
+    #
+    # This test triggers a data race which is suppressed in the TSAN CI. The test is to ensure
+    # np.nonzero does not generate a segmentation fault
    x = np.random.randint(4, size=10_000).astype(dtype)

-    def func(seed):
+    def func():
        x[::2] = np.random.randint(2)
        try:
            _ = np.nonzero(x)
        except RuntimeError as ex:
            assert 'number of non-zero array elements changed during function execution' in str(ex)            

-    run_threaded(func, max_workers=10, pass_count=True, outer_iterations=50)
+    run_threaded(func, max_workers=10, pass_count=False, outer_iterations=50)