https://sourceware.org/bugzilla/show_bug.cgi?id=34498
--- Comment #2 from Karan Kurani <karankurani3k at gmail dot com> --- Hey Frank Ch. Eigler Happy to explain. The research method is variant hunting from previously fixed bug classes: I trace a concrete invariant from an earlier fix, census nearby/current code for the same semantic mistake, and then independently validate any candidate at runtime before treating it as a finding. For this one I used LLM tooling during the research/analysis phase for source navigation and hypothesis generation. I then cross-built eu-stacktrace for aarch64 from origin/main fb5473ac7368bbbbc2ea9d86241bd6c882e3fcb5 and validated it under qemu-aarch64-static. The relevant runtime result was n_regs=14 while index 31 was consumed the tight-frame control showed the read crossing into the following frame, reproduced 3/3, while the ABI32 control stayed in bounds. I also used LLM assistance when structuring some of the prose in this Bugzilla report. I noticed afterward that the current CONTRIBUTING policy permits LLM use for research/analysis/debugging but says LLM-generated output should not be included in contributions. I don't want to misrepresent the provenance or work around that policy. If Bugzilla report prose is intended to fall under that restriction too, I'm happy to replace the attachment/description with a version I write directly from my raw validation notes and logs. The runtime validation itself and the decision to report the issue were mine, and I can provide the exact build commands, fixtures and raw output used to confirm it. -- You are receiving this mail because: You are on the CC list for the bug.
