[cppjit] Do not throw out of callback wrappers on Apple ARM64

.github/workflows/root-ci-config/buildconfig/mac-beta.txt

@@ -22,3 +22,4 @@ cocoa=ON
 davix=OFF
 minuit2_omp=OFF
 x11=OFF
+ROOT_CTEST_CUSTOM_FLAGS=-R ^pyunittests-bindings-cppjit-

.github/workflows/root-ci.yml

@@ -120,11 +120,10 @@
       matrix:
         # Specify platform + optional build option overrides
         # Build options: https://root.cern/install/build_from_source/#all-build-options
-        platform: [mac26, mac27, mac-beta]
+        # TEMP: only mac-beta, to measure the sporadic cppjit test_datatypes crash
+        platform: [mac-beta]
         arch: [ARM64]
         include:
-          - platform: mac15
-            arch: X86
           - platform: mac-beta
             is_special: true
 
@@ -132,7 +131,7 @@
       - self-hosted
       - ${{ matrix.platform }}
 
     name: |
       ${{ matrix.platform }} ${{ matrix.arch }}
       ${{ (github.event_name != 'schedule' && github.event_name != 'workflow_dispatch' && join( matrix.overrides, ', ' )) || '' }}
 
@@ -151,8 +150,8 @@
         if: github.event_name == 'pull_request'
         env:
           GITHUB_PR_ORIGIN: ${{ github.event.pull_request.head.repo.clone_url }}
           OVERRIDES: ${{ join( matrix.overrides, ' ') }}
         run: |
           [ -d "${VIRTUAL_ENV_DIR}" ] && source ${VIRTUAL_ENV_DIR}/bin/activate
           echo "Python is now $(which python3) $(python3 --version)"
           src/.github/workflows/root-ci-config/build_root.py  \
@@ -184,7 +183,7 @@
       - name: Nightlies and push to branch
         if:   github.event_name == 'schedule' || github.event_name == 'push'
         env:
           OVERRIDES: ${{ github.event_name == 'push' && format('{0} {1} {2}', '--overrides', env.GLOBAL_OVERRIDES, join( matrix.overrides, ' ')) || '' }}
           BASE_REF:  ${{ github.event_name == 'push' && github.ref_name || inputs.ref_name }}
           # For pushes, build exactly the pushed commit, not wherever the
           # branch happens to be when the job starts (issue #14061).
@@ -246,13 +245,7 @@
   build-windows:
     # For any event that is not a PR, the CI will always run. In PRs, the CI
     # can be skipped if the tag [skip-ci] or [skip ci] is written in the title.
-    if: |
-        (github.repository_owner == 'root-project' && github.event_name != 'pull_request') ||
-        (github.event_name == 'pull_request' && !(
-          contains(github.event.pull_request.title, '[skip-ci]') ||
-          contains(github.event.pull_request.title, '[skip ci]') ||
-          contains(github.event.pull_request.labels.*.name, 'skip ci')
-        ))
+    if: false # TEMP: only mac-beta runs, see build-macos
 
     permissions:
       contents: read
@@ -372,13 +365,7 @@
   build-linux:
     # For any event that is not a PR, the CI will always run. In PRs, the CI
     # can be skipped if the tag [skip-ci] or [skip ci] is written in the title.
-    if: |
-        (github.repository_owner == 'root-project' && github.event_name != 'pull_request') ||
-        (github.event_name == 'pull_request' && !(
-          contains(github.event.pull_request.title, '[skip-ci]') ||
-          contains(github.event.pull_request.title, '[skip ci]') ||
-          contains(github.event.pull_request.labels.*.name, 'skip ci')
-        ))
+    if: false # TEMP: only mac-beta runs, see build-macos
 
     permissions:
       contents: read

bindings/pyroot/cppjit/src/cpyrt/CPPMethod.cxx

@@ -181,10 +181,12 @@ inline PyObject* cpyrt::CPPMethod::ExecuteFast(void* self, ptrdiff_t offset,
     result = nullptr;
   }
 
-// TODO: covers the PyException throw case, which does not seem to work on
-// Windows, so instead leaves the error be
-#ifdef _WIN32
+#if CPYRT_CALLBACK_NOTHROW
+  // the callback wrappers leave the Python error set instead of throwing a
+  // PyException (see CommonDefs.h), so pick it up here, as the catch above
+  // would have, and flag it as the callee's error for overload resolution
   if (PyErr_Occurred()) {
+    ctxt->fFlags |= CallContext::kPyException;
     if (ResultIsPyObject())
       Py_XDECREF(result);
     result = nullptr;

bindings/pyroot/cppjit/src/cpyrt/CommonDefs.h

@@ -31,4 +31,23 @@
 
 #endif
 
+// How JIT-compiled callback wrappers (function pointers, std::function, and
+// virtual overrides implemented in Python) report a failed Python call. The
+// default is to throw a cpyrt::PyException from the wrapper, which unwinds
+// through the C++ frames in between (JIT-compiled, or built by another
+// compiler) up to the try block in CPPMethod::ExecuteFast. That unwind hangs
+// on Windows, and on Apple ARM64 exceptions through JIT-compiled frames are
+// known not to work (the test suite skips the tests that catch a PyException
+// in C++ there) and are the prime suspect for the sporadic crash of
+// test_datatypes on the macOS beta CI runner. On those platforms the wrapper
+// therefore leaves the Python error set and returns a default value instead;
+// the C++ code keeps running and ExecuteFast turns the pending error into a
+// failed call once the C++ call returns. The price is that C++ code can not
+// catch the Python error as a PyException.
+#if defined(_WIN32) || (defined(__APPLE__) && defined(__aarch64__))
+#define CPYRT_CALLBACK_NOTHROW 1
+#else
+#define CPYRT_CALLBACK_NOTHROW 0
+#endif
+
 #endif // !CPYRT_COMMONDEFS_H

bindings/pyroot/cppjit/src/cpyrt/Utility.cxx

@@ -1065,7 +1065,12 @@ void cpyrt::Utility::ConstructCallbackPreamble(
     }
     code << "    } catch(int) {\n"
          << "      for (auto pyarg : pyargs) Py_XDECREF(pyarg);\n"
+#if CPYRT_CALLBACK_NOTHROW
+         // error is set; see CPYRT_CALLBACK_NOTHROW in CommonDefs.h
+         << (isVoid ? "      return;\n" : "      return ret;\n")
+#else
          << "      cppjit::cpyrt::PyException pyexc; throw pyexc;\n"
+#endif
          << "    }\n";
   }
 }
@@ -1095,9 +1100,9 @@ void cpyrt::Utility::ConstructCallbackReturn(const std::string& retType,
   if (isPtr)
     code << "  }\n";
   code << "    if (!cOk) {" // assume error set when converter failed
-// TODO: On Windows, throwing a C++ exception here makes the code hang; leave
-// the error be which allows at least one layer of propagation
-#ifdef _WIN32
+// see CPYRT_CALLBACK_NOTHROW in CommonDefs.h: leave the error be, which allows
+// at least one layer of propagation
+#if CPYRT_CALLBACK_NOTHROW
           " /* do nothing */ }\n"
 #else
           " cppjit::cpyrt::PyException pyexc; throw pyexc; }\n"

bindings/pyroot/cppjit/test/CMakeLists.txt

@@ -37,7 +37,7 @@ endif()
 
 function(CPPJIT_ADD_PYTEST name)
   cmake_parse_arguments(ARG
-    "DROP_ROOTMAP" "" "DICTS;FIXTURES_SETUP;FIXTURES_REQUIRED" ${ARGN})
+    "DROP_ROOTMAP" "SUFFIX" "DICTS;FIXTURES_SETUP;FIXTURES_REQUIRED" ${ARGN})
 
   set(extra)
   if(ARG_DICTS)
@@ -60,7 +60,7 @@ function(CPPJIT_ADD_PYTEST name)
   endif()
 
   string(REGEX REPLACE "[_]" "-" good_name "${name}")
-  set(test_name pyunittests-bindings-cppjit-${good_name})
+  set(test_name pyunittests-bindings-cppjit-${good_name}${ARG_SUFFIX})
 
   ROOT_ADD_TEST(${test_name}
     # Upstream marks platform xfails without explicit strict=; the in-tree
@@ -115,6 +115,13 @@ CPPJIT_ADD_PYTEST(test_basic_api
   FIXTURES_REQUIRED cppjit_example01_dict_fixture)
 CPPJIT_ADD_PYTEST(test_datatypes
   FIXTURES_REQUIRED cppjit_datatypes_dict_fixture)
+# TEMP: 300 more copies of test_datatypes, run by ctest in parallel in this
+# directory like the real suite does, to measure its sporadic crash on the
+# mac-beta CI runner (about 3% of runs).
+foreach(i RANGE 1 300)
+  CPPJIT_ADD_PYTEST(test_datatypes SUFFIX -rep${i}
+    FIXTURES_REQUIRED cppjit_datatypes_dict_fixture)
+endforeach()
 
 CPPJIT_ADD_PYTEST(test_api)
 CPPJIT_ADD_PYTEST(test_boost)