try skipping teardown

Author: majosm

.github/workflows/ci.yml

@@ -38,6 +38,7 @@ jobs:
                 . ./ci-support-v0
                 build_py_project_in_conda_env
                 export NO_COLOR=1
+                export SKIP_TEARDOWN=1
                 python intel_crash_reproducer.py
 
 # vim: sw=4

intel_crash_reproducer.py

@@ -6,6 +6,9 @@
 group/local/chunk inames and tagged for an OpenCL-style execution.
 """
 
+import os
+import sys
+
 import numpy as np
 
 import loopy as lp
@@ -47,3 +50,15 @@
 
 _evt, (out,) = knl(queue)
 print(out.get())
+
+# The kernel result above is correct; the Intel CPU OpenCL runtime instead
+# aborts with "corrupted double-linked list" later, during process teardown
+# (its atexit/TBB thread-pool shutdown), not during the kernel launch.
+#
+# Setting SKIP_TEARDOWN=1 hard-exits before those finalizers run. If that makes
+# the crash disappear, it confirms the fault is in runtime/TBB shutdown rather
+# than the kernel.
+if os.environ.get("SKIP_TEARDOWN") == "1":
+    sys.stdout.flush()
+    sys.stderr.flush()
+    os._exit(0)