[cfe-dev] how to tolerate the assertion failures in llvm and clang

[George Karpenkov via cfe-dev]

> On Jun 27, 2018, at 7:54 PM, Ella Oikawa <alansnape3058 at gmail.com> wrote:
> 
> Hi George,
> 
> > You can compile LLVM with assertions disabled, but I don’t think it would solve your problem.
> 
> I think so.
> 
> > By the time the assertion is hit, your program is already in an inconsistent state it can not recover from.
> 
> Not the entire program, only the process analyzing that function. The data of analyzing each function is isolated from each other, if the analysis of a function crashes, these bad data will not affect other functions and the entire program as far as I think.
> 
> So I think the forking method is at least an acceptable bypass for the problem I encountered.

OK then indeed analyzing each function in a separate process is a good approach for you.

> 
> Regards,
> Ella
> 
> George Karpenkov <ekarpenkov at apple.com <mailto:ekarpenkov at apple.com>> 于2018年6月28日周四 上午10:43写道：
> Hi Ella,
> 
> I still don’t understand your question.
> You can compile LLVM with assertions disabled, but I don’t think it would solve your problem.
> 
>> On Jun 27, 2018, at 7:24 PM, Ella Oikawa <alansnape3058 at gmail.com <mailto:alansnape3058 at gmail.com>> wrote:
>> 
>> Hi George,
>> 
>> As you mentioned above, assertions are good for developers. But for users, they are annoyed, especially when the scan process is nearly finished
> 
> What would you like to do once this happens? By the time the assertion is hit, your program is already in an inconsistent state it can not recover from.
> 
> One possibility here is to run clang once to ask it to dump analyzer entry points in a file [you would have to implement that],
> and then fork a separate process per each entry point.
> Then if the process crashes, all other functions could still be analyzed.
> 
>> . My team is developing the tool, but the users are from other teams, so the "production" version for these teams should be "stable" enough, even though all the mistakes are hidden.
>> The errors vary from project to project, and only a few open source projects in our benchmark will trigger the assertions in clang, so they are hard to model. And my team has very few developers, we do not have enough time to model these errors, neither the user team (they even do not want to provide their code to us).
>> My team is using the old version (3.3) of the llvm/clang project
>> (the oldest code may be written originally by Zhongxing Xu, the writer of the paper which the clang static analyzer is based on, even older than the clang static analyzer itself), I am not sure these "bugs" in clang and llvm still exists in the latest versions. And the count of developers limits us from updating the clang/llvm version, as we have so many features to develop and have limited time to do so.
>> 
>> Regards,
>> Ella
>> 
>> George Karpenkov <ekarpenkov at apple.com <mailto:ekarpenkov at apple.com>> 于2018年6月26日周二 上午2:38写道：
>> Hi Ella,
>> 
>>> On Jun 24, 2018, at 1:23 AM, alan snape via cfe-dev <cfe-dev at lists.llvm.org <mailto:cfe-dev at lists.llvm.org>> wrote:
>>> 
>>> My team is developing a static analysis tool based on clang and llvm, but the assertion failures in the source code of llvm and clang will always crash the program execution, which is not acceptable in a **stable product**.
>> 
>> Each assertion failure is a bug you can report at bugs.llvm.org <http://bugs.llvm.org/>. Do you really get that many?
>> We at Apple regularly run the clang static analyzer on huge chunks of the internal codebase, and in my understanding Google does the same,
>> so I am quite surprised this is a problem for you.
>>  
>>> The tool analyzes the functions one by one in the Call Graph SCC order, so is there any way to tolerate the assertion failures and continue the analysis on the next function when assertion failures occur on calling some APIs of clang and llvm? (crashes only the analysis of the function (the analysis methods of the FunctionDecl) being analyzed, not the entire program)
>> 
>> I would really like to challenge your assumption here that assertions are unacceptable.
>> 
>> A clean crash with an understandable stack trace means that the problem can be fixed.
>> Adding layers of indirection which hide those failures means that the bug gets unnoticed,
>> and the analyzer probably ends up doing something wrong.
>> 
>> Could you clarify what is the common error mode for you?
>> 
>> Regards,
>> George
>> 
>>> 
>>> Thanks,
>>> Ella
>>> _______________________________________________
>>> cfe-dev mailing list
>>> cfe-dev at lists.llvm.org <mailto:cfe-dev at lists.llvm.org>
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev <http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev>
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180627/bf38385c/attachment.html>