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.

Reply via email to