Alan Modra <amo...@gmail.com> writes: > On Mon, May 02, 2022 at 09:27:05AM -0500, Segher Boessenkool wrote: >> > 2) If a function changes LR or any non-volatile register, the save >> > location for those regs must be given. The cfi can be at any >> > instruction after the saves up to the point that the reg is >> > changed. (Exception: LR save should be described before a bl.) >> >> That isn't an exception? bl changes the current LR after all :-) > > The point is that in other cases the cfi can be as late as the > instruction that changes the reg. For calls it must be at least one > instruction before the call.
Got it. > Also, I'll note for the wider audience that delaying cfi is slightly > better than playing it safe as Michael has done in his patch in > describing the saves right at the save instruction. Register save cfi > can usually be grouped together, resulting in fewer CFI_advance codes > in .eh_frame. I didn't want to go overboard on combining them, because it's harder to read the source, especially with the #ifdefs we have for 64-bit. I was able to save one CFA_advance_loc by moving the LR restore down one line. The .eh_frame didn't shrink, I guess because it's padded to some alignment anyway. >> Alan proposed a larger patch that changed to a single stack frame, but it >> needs changes to >> take into account the red zone. > > Yes, now that you mention it, I see the obvious error in the patch I > wrote. I did say it was untested! No worries, identifying the source of the bug as the missing CFI is the key thing, it would have taken me a while to realise that. cheers
