[cfe-dev] [ubsan] Add -fsanitize-warn-once, only emit runtime error once per check

[Richard Smith]

Have you measured the code size overhead from this extra flag? Did you
consider implementing this in the runtime library instead (by
suppressing duplicates based on return address or SourceLocation)?

On Thu, Dec 13, 2012 at 11:51 AM, Will Dietz <willdtz at gmail.com> wrote:
> On Thu, Dec 13, 2012 at 12:05 PM, Dmitri Gribenko <gribozavr at gmail.com> wrote:
>> On Thu, Dec 13, 2012 at 7:33 PM, Will Dietz <willdtz at gmail.com> wrote:
>>> This flag causes clang to emit a byte for each check that is used by the
>>> runtime to track whether we've already printed an error for that check.
>>>
>>>  Often failed checks are triggered many times dynamically, but a user
>>>  is only interested in which checks failed (with example dynamic values
>>>  to aid in debugging).  This flag lets the user make such runs much
>>>  more efficient and generate more manageable output.
>>
>> Hi Will,
>>
>> +  if (Checked) {
>> +    if (*Checked) return;
>> +    *Checked = true;
>> +  }
>>
>> Does it make sense to do the store atomically?  The user's program is
>> already buggy, but introducing a possible data race is unfortunate.
>>
>> Dmitri
>>
>
> Hi Dmitri,
>
> Glad you brought this up.  I wasn't sure which way to go on this and
> erred on simplicity.  Attached is an updated compiler-rt patch using
> __sync_val_compare_and_swap, which also simplifies the code a bit.  If
> this builtin is sufficiently portable (architectures and compiler
> recognition) then I would prefer this for the reasons you mention.
>
> Thanks!
>
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev
>