[libclang/python] Properly report errors when test fails

[DeinAlptraum]

test_cdb.py's test_create_fail captures stderr to suppress output but did not release it in case the test fails.

[DeinAlptraum]

This issue has been raised in #142353 (review)

[llvmbot]

@llvm/pr-subscribers-clang

Author: Jannick Kremer (DeinAlptraum)

Changes

test_cdb.py's test_create_fail captures stderr to suppress output but did not release it in case the test fails.

---

Full diff: https://github.com/llvm/llvm-project/pull/142371.diff

1 Files Affected:

• (modified) clang/bindings/python/tests/cindex/test_cdb.py (+10-7)

diff --git a/clang/bindings/python/tests/cindex/test_cdb.py b/clang/bindings/python/tests/cindex/test_cdb.py
index 757a14fbaef2c..5abe56f0d65f8 100644
--- a/clang/bindings/python/tests/cindex/test_cdb.py
+++ b/clang/bindings/python/tests/cindex/test_cdb.py
@@ -21,13 +21,16 @@ def test_create_fail(self):
 
         # clang_CompilationDatabase_fromDirectory calls fprintf(stderr, ...)
         # Suppress its output.
-        stderr = os.dup(2)
-        with open(os.devnull, "wb") as null:
-            os.dup2(null.fileno(), 2)
-        with self.assertRaises(CompilationDatabaseError) as cm:
-            CompilationDatabase.fromDirectory(path)
-        os.dup2(stderr, 2)
-        os.close(stderr)
+        try:
+            stderr = os.dup(2)
+            with open(os.devnull, "wb") as null:
+                os.dup2(null.fileno(), 2)
+            with self.assertRaises(CompilationDatabaseError) as cm:
+                CompilationDatabase.fromDirectory(path)
+        # Ensures that stderr is reset even if the above code crashes
+        finally:
+            os.dup2(stderr, 2)
+            os.close(stderr)
 
         e = cm.exception
         self.assertEqual(e.cdb_error, CompilationDatabaseError.ERROR_CANNOTLOADDATABASE)

[Endilll]

Have you reproduced the issue and checked that this fix does what's it supposed to do?

[DeinAlptraum]

@Endilll
I was too lazy to reproduce the exact issue (installing 32bit library and rebuilding Clang etc.) but I've forced this with a different error. The symptom (EEEEEEE% output and no error message on testing runs) is not specific to that one issue, but occurs whenever the get_cindex_library function raises.
For my own testing, I've confirmed that raising at the start of that function causes the same output, and that after adding my change you get the proper output (including the actual error message and traceback)

(just to be clear: that doesn't fix any issues related to 32bit/64bit mixing, it just ensures this test doesn't swallow all output whenever it fails)