julia

7a657c85 - signals: Scrub the x87 state when simulating longjmp on i686 (#62669)

signals: Scrub the x87 state when simulating longjmp on i686 (#62669) jl_simulate_longjmp edits only the general-purpose registers of the interrupted context; the kernel's sigreturn then restores the saved FPU state - possibly mid-computation with live x87 stack entries. The resumed code (an exception handler, or a cancellation reset point re-executing its check) expects the i386 ABI's function-boundary state, so the leaked entries overflow the x87 register stack and poison every subsequent float result with NaNs (and, through the GC's double-based heap heuristics, can corrupt the process outright). Scrub the saved x87 state to all-registers-empty on both the linux and windows x86 context layouts; the control word is preserved. Observed as NaN timing results and a SIGSEGV in the i686 cancellation tests, where the reset-point delivery exercises this path heavily. This pull request was written with the assistance of generative AI. Co-authored-by: Keno Fischer <Keno@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>