[fix] Null-guard the cling.printValue lookup in op_str - #56
Conversation
cc6b1ab to
06c8899
Compare
| PyObject* cl = PyObject_GetAttrString(gbl, (char*)"cling"); | ||
| printValue = PyObject_GetAttrString(cl, (char*)"printValue"); | ||
| Py_DECREF(cl); | ||
| // no cling namespace exists unless user code declares one |
There was a problem hiding this comment.
Perhaps we can just short-circuit this path completely if we know the interpreter is clang-repl, since that lookup would always return null. We could use the compile-time definition CPPJIT_USE_CLING, or a runtime check like in test/support.py
There was a problem hiding this comment.
Good call — done. The whole pretty-printing block now sits behind #ifdef CPPJIT_USE_CLING, so clang-repl builds compile it out and str() falls straight through to the generic repr; strings on the clang-repl libcppjit.so shows the printValue lookup gone, and the object shrinks slightly. One thing the gate needed: CPPJIT_USE_CLING was only ever a CMake option and never reached the sources, so I added the matching target_compile_definitions entry — without it the #ifdef would have been false on the cling build too. The null-guards stay inside the cling branch, in case an odd cling state leaves the lookup empty. Worth naming the trade-off: a user-declared namespace cling { printValue } is no longer honored on clang-repl, which I think is right given the native value-printing direction in compiler-research/CppInterOp#1100.
06c8899 to
40e8086
Compare
The in-process test52 crashed the llvm22 and cling CI lanes with a SIGSEGV, hiding the rest of the suite. Match compiler-research#56's head: the cling pretty-print path in op_str now compiles out entirely behind CPPJIT_USE_CLING on clang-repl, with the matching CMake compile-definition wiring, and the regression test runs its repro in a subprocess so a crash fails the assertion instead of the runner. Co-developed-with-the-help-of: Claude Code (Sonnet 5, human in the loop)
Commit 5a33027 dropped the faked cling::runtime::gCling and with it the last cling namespace in the interpreter. The pretty-print fallback in op_str then dereferenced the failed cppjit.gbl.cling lookup. str() of any instance without an ostream inserter crashed with SIGSEGV. Guard both lookups, clear the AttributeError, and fall back to the generic repr. clang-repl has no cling namespace to look up, so the whole pretty-print path now compiles out behind CPPJIT_USE_CLING. CMake gains the matching compile definition: the option existed but never reached the sources. The null-guards stay on the cling side. This drops the old escape hatch where a user-declared namespace cling { printValue } was honored on clang-repl, which matches the native value-printing direction in compiler-research/CppInterOp#1100. The regression test runs the repro in a subprocess, so a return of the crash cannot kill the test runner. The subprocess gets this process's sys.path through PYTHONPATH, because a build system can put cppjit on the path in-process instead. A clean exit passes, whichever repr form the backend prints. Output that names the upstream toString stub xfails. Anything else fails and reports the child's output. Co-developed-with-the-help-of: Claude Code (Fable 5, human in the loop)
40e8086 to
e043129
Compare
| // reasons) | ||
| // reasons). Cling only: clang-repl has no cling namespace to look up, so the | ||
| // whole path compiles out and str() falls through to the generic repr. | ||
| #ifdef CPPJIT_USE_CLING |
There was a problem hiding this comment.
This should probably be required to CppInterOp’s toString interface which will still crash for clang-repl but this time we can fix it there.
There was a problem hiding this comment.
Happy to make that change here: route str() through Cpp::ObjToString and drop the cling-only block. clang-repl would then keep failing in toString until compiler-research/CppInterOp#1100 lands the fix there. Do you want that in this PR, or should we keep the short-circuit and switch once #1100 is fixed?
3 participants
str()on any instance without anostreaminserter segfaults on currentmain. Commit5a33027dropped the fakedcling::runtime::gCling, and thepretty-print fallback in
op_strthen dereferenced the failedcppjit.gbl.clinglookup. This guards both lookups, clears theAttributeError, and falls back to the generic repr.Repro:
A regression test is included. The crash kills the interpreter, so the test
asserts the fixed behaviour rather than the crash.