> That's closer to the definitions used in the specification, so go
> ahead.  I used REG_UNSAVED since that's what GCC's unwinder uses, so
> perhaps you could add a comment that says this?

Ah!  That also explains why GCC has problems.  It's combined two 
different states :-/  I've added comments to that effect.


>> - if it detects an unspecified CFI entry it complains
>> It isn't perfect though - since it doesn't know the full range of valid 
>> debug info register numbers it can't check every entry.  Instead it 
>> checks the range provided by CFI for unspecified holes and then 
>> complains about that.  The reality is that GCC at least gets that bit 
>> right (but consistently forgets the SP).
> 
> 
> Sorry, GCC gets what right?

On the platforms I tested (amd64 & i386) all [0 .. max "column") were 
specified.  The problems were with numbers outside of that range.  I'll 
re-word the comment.

> 
> Anyway, the full range of valid debug info registers, is defined by
> the DWARF register mapping that some ABI's define (at least the x86-64
> ABI provides such a definition).  GDB sort of has this information in
> that we have the dwarf2_reg_to_regnum function in the architecture
> vector.

Yes.  There isn't a way of iterating over it.  Oh, and btw, it can also 
be sparse - I believe that the e500 (ppc variant) has has dwarf2 numbers 
 >1000.  Something to address in a later iteration.

>> I'd like to commit the patch as is for the 6.0 branch.  For the mainline 
>> though, I'd like to make the additional changes:
>> 
>> - delete the SP_REGNUM hack from the REG_UNDEFINED rule (it's no longer 
>> needed, I think)
> 
> 
> I guess that if the stack pointer is explicitly marked as being
> "undefined", all will be lost on most architectures.

Yes, bit of a loose-loose situtation.  If it's really "undefined" GDB 
can't use "CFI" as that would mislead the user into thinking that a 
bogus backtrace is valid.

>> - add a check/complaint for the SP v CFA problem.
>> 
>> Anyway, the end result is that on x86-64 and i386 store.exp now passes.
>> 
>> ok to commit?
> 
> 
> Hmm, three "style" nits:
> 
> - The uppercase MEMSET in the comment on `enum dwarf2_reg_rule' is
>   wrong.  Uppercase is reserved for when we're talking about the value
>   of a symbol, not the symbol itself.  So please convert this into
>   lowercase.
> 
> - I find the name `cfinum' a bit cryptic.  I understand that you want
>   to avoid `reg', how about calling it `column'?  That's sort of CFI
>   speak for a register number.
> 
> - Please remove the extra blank line before `case REG_SAME_VALUE:'.
> 
> With those changes this is OK (or can't I approve changes to a file
> I've written in the first place?).  You can also leave it to me to
> make those changes ;-).

I think they are fixed.  I've attached the diff I'll commit.

Andrew