Alexandros Mandravillis
1.8 KiB
Primary same-thread RFE register restoration
This is a new authored correction to the primary register restore at upstream 7cddab8237e1, by Alexandros Mandravillis (Alexbeav). A real handler's final jump register is architectural state; restoring the interrupted snapshot must not overwrite that actual result on the qualified same-thread RFE return.
The added local captures the original delivery EPC before handler dispatch. The primary copy omits only GPR26 when mode 1, real RFE, outermost delivery, stable nonzero TCB and both original-EPC comparisons qualify. It does not manufacture a register value or alter the return PC. Other restore modes, legacy/syscall policy, other GPRs, HI/LO and the secondary deferred copy retain their existing behavior. RestoreState is not separately disqualified: a redirect that ultimately returns to the original EPC with the same TCB can qualify.
The regression uses the real current interrupts.c and traps.c, the existing upstream inert link fixture, and a new synthetic handler with arbitrary register values. It contains no ROM-derived instructions, retail addresses/captures, generated BIOS/game code, reference-emulator body or passive observer. Source-specific scheduler/redirect policies, drive configuration and recording infrastructure are deliberately excluded.
The current base fails 64 qualifying K0 checks. The fixed source is checked in modes 0 through 3, with real-RFE/syscall/legacy reasons, null/stable/changed TCB, KSEG aliases, changed targets/global EPC, actual sentinel fallback and arbitrary live K0 values. The existing restore-edge IRQ regression is also run. Nested delivery and RestoreState execution are not exercised by these fixtures. No new retail replay or whole-title accuracy claim is made.