patch: make ubsan report error summaries

[Nick Lewycky]

On 23 October 2013 15:29, Richard Smith <richard at metafoo.co.uk> wrote:

> I'm not really happy about including the 'summary' in the normal output --
> it's ugly and redundant. If we want to provide this to people who want
> summaries, that's fine, but the default should either be that this output
> goes nowhere, or that the summary text is exactly the 'runtime error:' line.
>

Okay, how about this direction. Let's add a sanitizer_common flag which
controls whether summaries are printed. The sanitizers can check the flag
and decide not to call the API (ie., to avoid requiring a buffer) and the
default implementation of the API checks the flag again and does nothing if
called with the flag off. Then we turn that flag on with the RUN lines to
test our summary emission, but it defaults to off for users that don't want
it.

+cc Kostya. Does that work for both of you? I agree with Richard that (my
idea of "normal") normal users would find this output redundant, hence the
idea to default the flag to off.

Nick

On Wed, Oct 23, 2013 at 2:51 PM, Nick Lewycky <nlewycky at google.com> wrote:
>
>> On 23 October 2013 11:18, Richard Smith <richard at metafoo.co.uk> wrote:
>>
>>> On Wed, Oct 23, 2013 at 2:21 AM, Nick Lewycky <nlewycky at google.com>wrote:
>>>
>>>> On 22 October 2013 22:07, Nick Lewycky <nlewycky at google.com> wrote:
>>>>
>>>>> On 22 October 2013 21:18, Nick Lewycky <nlewycky at google.com> wrote:
>>>>>
>>>>>> The attached patch makes ubsan emit summaries of errors it
>>>>>> encounters. The format of these summaries is:
>>>>>>   UndefinedBehaviourSanitizer: signed-integer-overflow file:49:7
>>>>>> where the string is the flag name. Most of the patch is adding the
>>>>>> flag names to all the reports all over.
>>>>>>
>>>>>
>>>>> I've noticed a small bug, for load-invalid-value we always pick "enum"
>>>>> and never "bool". I would guess that's because
>>>>> ASTContext::getTypeSize(BoolTy) returns 8 instead of 1?
>>>>>
>>>>> Richard, thoughts?
>>>>>
>>>>
>>>> Updated patch attached. It now detects bool sanitizer by looking at the
>>>> Type as a string, and is otherwise updated for the changes in
>>>> sanitizer-common.
>>>>
>>>
>>> This does the wrong thing for typedefs of bool. Can we emit a flag as
>>> part of the static info to say whether this was the bool sanitizer or the
>>> enum sanitizer? Otherwise, I don't see how we can distinguish the
>>> typedef-for-bool case from the enum-with-underlying-type-bool case.
>>>
>>
>> Done. Patch attached!
>>
>> Nick
>>
>>
>>>  This patch is stacked on top of
>>>>>> http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20131021/091535.html ,
>>>>>> or else ubsan's tests will fail.
>>>>>>
>>>>>> Please review!
>>>>>>
>>>>>> Nick
>>>>>>
>>>>>
>>>>>
>>>>
>>>> _______________________________________________
>>>> cfe-commits mailing list
>>>> cfe-commits at cs.uiuc.edu
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>>>
>>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20131023/e5db887d/attachment.html>