Re: [RFA] Compact EH Patch

[Richard Henderson <rth@redhat.com>]

On 10/06/2015 10:14 AM, Moore, Catherine wrote:
>>> +  switch (pr_index) {
>>> +  case 0:
>>> +      p = read_encoded_value (context, bases->eh_encoding, p,
>> &personality);
>>> +      fs->personality = (_Unwind_Personality_Fn) personality;
>>> +      break;
>>> +  case 1:
>>> +      fs->personality = __gnu_compact_pr1;
>>> +      break;
>>> +  case 2:
>>> +      fs->personality = __gnu_compact_pr2;
>>> +      break;
>>> +  default:
>>> +      fs->personality = NULL;
>>> +  }
>>
>> This is the aspect of this spec about which I am least keen.  The existing
>> method whereby the personality function is explicit in each object file means
>> that we've got automatic version control on data that is private to the object
>> file.
>>
>> That is, if we should ever change the format of the LSDA -- as in fact you are
>> doing here -- then all we need do is change __gcc_personality_v0 to
>> __gcc_personality_v1, and all is well.  One can mix and match object files
>> from different compiler versions and all that is required for correctness is
>> that the runtime libraries must continue to provide all previous versions.
>>
>> What you're doing here doesn't allow the format of index {1,2,3} to *ever*
>> change.  You've fixed it forever, barring abandoning those indices and always
>> using index 0.
>>
>> I know that the ARM EH format made exactly the same mistake, but let's see
>> if we can find some way of not replicating it, eh?
>
> I'm not 100% sure what you are suggesting here.  Would you like to see __gnu_compact_pr2_v0,  etc?
> Are you also saying that you want to see a undefined reference to the personality routine in the object file?

I'm commenting on what I see as a defect in the spec, but not making a specific 
recommendation.

Using __gnu_compact_pr2_v0 etc doesn't help, because you're still using just 2 
bits to reference it.  The only real solution is a symbolic reference from the 
object file, which you get from pr_index == 0.


r~