Denis PILAT wrote:
> Daniel Jacobowitz wrote:
>> On Tue, Jan 23, 2007 at 01:32:22PM +0100, Denis PILAT wrote:
>>  
>>> Hi,
>>>
>>> We have a bug in one of our gdb target, when the binary changed 
>>> while beeing debugged it appears that some of our varobj refers to 
>>> invalid symbols or type.
>>>
>>> I propose a patch that delete all varobj when symbols are reloaded. 
>>> May be there is a better place to do that but I think we must do 
>>> that somewhere, don't you ?
>>>     
>>
>> The right thing to do is probably to figure out where the
>> invalid references come from and fix them - probably by re-evaluating
>> expressions at the next -var-update.  Deleting things behind the front
>> end's back is a bad policy.
>>
>>
>>   
> Deleting all is a little bit radical, I can understand that.
> Re-evaluating expression would give the expected thing for globale 
> variables, but for varobj that has been set on local variable (so 
> within a "valid_block"), it should have side effects: the evaluation 
> of the expression might tied the varobj to a variable different than 
> the one the user originally decided.
>
> I propose to distinguish 2 cases
> - varobj on global variable (valid_block == null) must have their 
> expression re-evaluated.
> - other one must be put in a kind of invalid state that will gives 
> "in_scope="false" to the user.
> This last point will involve a new field in varobj structure, I 
> haven't seen something to mark varobj as invalid.
>
> Denis
>
Hi

Attached is a new patch proposal when I implemented what I explained above.

During the symbol re-loading I invalidate varobjs linked to locals 
variable and try to re-evaluate the expression for globals.
No more varobjs are deleted, and during varobj_update invalidated 
symbols are skipped so displayed as not in scope.

If you're fine with this implementation I'll write 2 more test cases in 
the testsuite that lead to segmentation fault in the current GDB:
- one where the binary file completely changed so any varobjs defined on 
global variable are invalidated.
- one where the binary file is just updated so varobjs on globals can 
still be evaluated.

I check the testsuite with i386-linux target with no regression and I 
passed a valgrind as well.

-- 
Denis PILAT
STMicroelectronics