Re: [PATCH/i386newframe/RFC] DWARF CFI frame unwinder

[Andrew Cagney <ac131313@redhat.com>]

>    Date: Sun, 04 May 2003 23:35:27 -0400
>    From: Andrew Cagney <ac131313@redhat.com>
> 
>    Mark, fyi,
> 
>    > +    case REG_SAVED_REG:
>    > +      *optimizedp = 0;
>    > +      *lvalp = lval_register;
>    > +      *addrp = 0;
>    > +      *realnump = DWARF2_REG_TO_REGNUM (cache->reg[regnum].loc.reg);
>    > +      if (valuep)
>    > +	{
>    > +	  /* Read the value from the register.  */
>    > +	  frame_unwind_register (next_frame, *realnump, valuep);
>    > +	}
>    > +      break;
>    > +
> 
>    Set *addrp to the register offset hack (Otherwize something mysterious 
>    fails.  What? I don't remember).
> 
> Yes, findvar.c:value_of_register() and findvar.c:value_from_register()
> use this.  Worse, our whole value subsystem seems to rely on this.
> Ughh, that's really gross.  We should do something about that!

[tell me about it :-)]

> Anyway.  Thinking about it a bit more, I suspect the whole handling of
> REG_SAVED_REG is wrong.  Instead, we should just change the register
> number according to the DWARF CFI rule and let the unwinder for
> NEXT_FRAME handle the request:
> 
>     case REG_SAVED_REG:
>       regnum = DWARF2_REG_TO_REGNUM (cache->reg[regnum].loc.reg);
>       /* FALLTHROUGH */

Please, anything but a fallthrough :-)

>     case REG_UNMODIFIED:
>       frame_register_unwind (next_frame, regnum,
> 			     optimizedp, lvalp, addrp, realnump, valuep);
>       break;
> 
> Eventually this means that sentinel_frame_prev_register will provide
> the register offset hack.
> 
> Do you agree with my analysis?

Yes.  Stores could otherwize go to the wrong register (a really horrible 
bug to track down).

I should note that the part of ``info registers'' where it prints out 
where a register was saved needs a re-think.  Now that things are 
recursive it has the potential for reporting a register that was saved N 
levels further in on the stack (it does this already with the d10v and 
PC vs LR).

BTW, suggest adding a few name prefixes to register number variables - 
it is really hard to know which space (DWARF 2 or GDB) the register's 
value falls in.

Andrew