Fix missing line info in Jupyter notebooks

[jiel-nv]

Fixes #721.

The linecache scenario (which Jupyter uses) was missed during vendoring via PR #457. Port in the upstream Numba approach of adding inspect.getsourcelines() to handle this case (original PR numba/numba#9861) with modifications so that it serves as the main/only approach instead of fallback.

Jupyter notebook cells have virtual filenames (e.g., ) that don't exist on disk. IPython registers cell source in Python's linecache module, which inspect.getsourcelines() uses internally. The previous implementation only used open() to read source files, causing lookups to fail for notebook cells.

Using inspect.getsourcelines() as the main approach rather than a fallback simplifies the code since it handles both disk files and linecache entries uniformly.

Added test_linecache_source() in test_debuginfo.py to simulate the Jupyter scenario by registering source in linecache and verifying that no NumbaDebugInfoWarning is raised when compiling a kernel with debug=True.

[copy-pr-bot]

Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

Replaces file-based source reading with inspect.getsourcelines() to fix missing line info in Jupyter notebook cells. The previous implementation used open() to read source files directly, which failed for virtual filenames like <ipython-input-X-Y> that Jupyter uses. The new approach leverages inspect.getsourcelines() which internally uses linecache, handling both regular disk files and IPython's registered cell sources uniformly.

• Refactored FindDefFirstLine.__init__() to accept name and firstlineno parameters instead of code object, simplifying the interface
• Both get_func_body_first_lineno() and get_func_def_lineno() now use inspect.getsourcelines() as the primary approach rather than a fallback
• Adjusted line number calculations to account for the offset returned by inspect.getsourcelines()
• Added test_linecache_source() that simulates Jupyter's linecache registration to verify no NumbaDebugInfoWarning is raised

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The changes are well-structured and follow the upstream Numba approach. The refactoring simplifies the code by using a unified method for both disk files and linecache entries. The implementation correctly handles offset adjustments, includes proper error handling, and adds a comprehensive test that validates the fix for the reported issue.
• No files require special attention

Important Files Changed

Filename	Overview
numba_cuda/numba/cuda/misc/firstlinefinder.py	Replaced file-based source reading with inspect.getsourcelines() to support Jupyter notebooks via linecache
numba_cuda/numba/cuda/tests/cudapy/test_debuginfo.py	Added test_linecache_source() to verify debug info works with linecache-registered sources

[jiel-nv]

/ok to test ec4d8f3

[gmarkall]

I'm getting a really weird traceback when I try to run the reproducer from the bug report in a Jupyter notebook.

It would be surprising if this PR was the cause of the issue, but perhaps it exposes another strange issue.

Edit: my local environment needed an update for recent changes. There is no issue.

(traceback deleted)

[gmarkall]

It's OK, it turns out my local main was out of date. We have some other regression which I'll look into separately.

[gmarkall]

Many thanks!