Re: [PATCH] libgcc: Fix uninitialized RA signing on AArch64 [PR107678]

[Richard Sandiford <richard.sandiford@arm.com>]

Jakub Jelinek <jakub@redhat.com> writes:
> On Thu, Jan 12, 2023 at 03:22:56PM +0000, Richard Sandiford wrote:
>> If we have a new enum, I think we should handle it explicitly.  The fact
>> that the information isn't propagated between frames is a key part of
>> the semantics.
>> 
>> >> Another option is to just define the arch dependent value for how field
>> >> in the arch code, right now it is unsigned char type, so using say
>> >> (unsigned char) ~0 or (unsigned char) ~0 and (unsigned char) ~1 as arch
>> >> specific values might be ok too.
>> >
>> > Yet another option would be to define 1-2 extra REG_ values in the generic
>> > unwind-dw2.h header, but name them
>> >   REG_ARCH_SPECIFIC_1,
>> >   REG_ARCH_SPECIFIC_2,
>> > or so, and then the machine specific code can
>> > #define REG_AARCH64_TOGGLE_ON REG_ARCH_SPECIFIC_1
>> > Of course, all this depends on whether the arch specific codes can be
>> > handled in uw_update_context_1 by doing break; there and nothing else.
>> 
>> Yeah, personally I'd prefer for target-independent code to provide
>> the toggle representation, even if it isn't widely used.
>
> I can live even with that, I just hope it won't make code generation worse
> on other targets.
> Anyway, I understood aarch64 needs 2 states for the signing, so one would
> be REG_TOGGLE_ON and the other anything else?

The other is the default (no signing), so it needs to be REG_UNSAVED.

> Users can always create (invalid?) unwind info where they save the magic
> register, make it undefined etc.
>
> And
> void bar (void), baz (void), boo (void), qux (void), corge (void);
> enum {
>   REG_UNSAVED,
>   REG_SAVED_OFFSET,
>   REG_SAVED_REG,
>   REG_SAVED_EXP,
>   REG_SAVED_VAL_OFFSET,
>   REG_SAVED_VAL_EXP,
>   REG_UNDEFINED
> #ifdef ANOTHER
>   , REG_TOGGLE_ON
> #endif
> };
>
> void
> foo (unsigned char c)
> {
>   switch (c)
>     {
>       case REG_UNSAVED:
>       case REG_UNDEFINED:
> #ifdef ANOTHER
>       case REG_TOGGLE_ON:
> #endif
>         break;
>           
>       case REG_SAVED_OFFSET:
>         bar ();
>         break;
>      
>       case REG_SAVED_REG:
>         baz ();
>         break;
>   
>       case REG_SAVED_EXP:
>         boo ();
>         break;
>   
>       case REG_SAVED_VAL_OFFSET:
>         qux ();
>         break;
>
>       case REG_SAVED_VAL_EXP:
>         corge ();
>         break;
>     }
> }
> suggests that it doesn't, already cfg pass turns the implicit default:
> into something that covers even the REG_UNSAVED, REG_UNDEFINED and maybe
> REG_TOGGLE_ON values into default:

OK, that's good.  Maybe having it behind a macro wouldn't be too bad though,
if it comes to that.

Thanks,
Richard