[cfe-dev] Proposal: Integrate static analysis test suites

Anna Zaks via cfe-dev cfe-dev at lists.llvm.org
Wed Jan 27 23:31:35 PST 2016 

> On Jan 27, 2016, at 11:16 PM, <Alexander G. Riccio> <test35965 at gmail.com> wrote:
> 
>  
> This is not surprising, the static analyzer does not catch buffer overflows. We do have an experimental checker for it but it is not very strong. 
> 
> Personally, I think detecting stack overruns is a very valuable capability of a static analysis tool. Getting Clang to detect this issue with the default options should be a high priority.
> 
> One of the main issues is that the solver we use does not reason about relational constraints that involve 2 symbols (ex: i < n).
> 
> Stepping through the checker code over the past couple days, I can see this: it appears to "brute force" array accesses, evaluating loop conditions every single time. When checking the attached minimized case, enabling only the "-analyzer-checker=alpha.security.ArrayBoundV2", Clang seemed to bail out at around the third access. That makes sense, as the default value for analyzer-max-loop is 4. Indeed, if I bump "analyzer-max-loop" up 11 (no pun intended), then Clang catches the issue. Console command line attached.
> 
> As I think you alluded to, this sort of checking is best done "algebraically" instead of with brute force. I'm not really sure how to implement that sort of "algebraic" checker, but I've been curious about the problem for several years now. By discovering Clang's weaknesses in static analysis, and subsequently fixing them, I'll learn exactly that. I usually learn best when I learn "the hard way", and this seems like the perfect opportunity.
> 
> Side note: during any of the optimization passes does Clang/LLVM do any kind of loop bound access analysis? Perhaps we can use that info to evaluate relational constraints that govern array access?
> 
> Would those be caught with compiler warnings? (Try running clang on them with -Weverything.)
> 
> Actually, they do seem to be caught when compiler warnings are turned on... but they ONLY warn when NOT passing --analyze? Huh? 

This is by design. Many more people have compiler as part of their daily flow so it’s best to have such errors being reported by the compiler.
Having the analyzer produce all of the compiler warnings is likely to be too nosy for the users.

> 
> I think diagnosing format string misuse during normal compilation is a fantastic idea - MSVC until 2015 required that you run /analyze, which very few people actually do <https://www.youtube.com/watch?v=4zgYG-_ha28&feature=player_detailpage#t=57m42s>* - but I'm not used to the idea that I can't run analysis at the same time...
> 
> In the short term:
> 
> Now that I'm a bit familiar with the codebase I expect to finish manually running the SAMATE tests in the next couple days, and start work on getting them to run under LIT after that. 
> 

Have you considered adding the tests to be tested on the additional analyzer build bot (http://lab.llvm.org:8080/green/job/StaticAnalyzerBenchmarks/ <http://lab.llvm.org:8080/green/job/StaticAnalyzerBenchmarks/>) instead of adding them to lit? (I’ve suggested that earlier.)

> *back then it was only for Xbox 360 devs, now it's in all editions of Visual Studio
> 
> Sincerely,
> Alexander Riccio
> --
> "Change the world or go home."
> about.me/ariccio <http://about.me/ariccio>
> 
>  <http://about.me/ariccio>
> If left to my own devices, I will build more.
> ⁂
> 
> On Mon, Jan 25, 2016 at 11:48 AM, Anna Zaks <ganna at apple.com <mailto:ganna at apple.com>> wrote:
> 
>> On Jan 24, 2016, at 12:58 AM, <Alexander G. Riccio> <test35965 at gmail.com <mailto:test35965 at gmail.com>> wrote:
>> 
>> Since that patch landed, I've manually run ~30 of the SAMATE/SARD tests, and so far, Clang has missed 5 stack buffer overruns, 4 heap buffer overruns,
> 
> This is not surprising, the static analyzer does not catch buffer overflows. We do have an experimental checker for it but it is not very strong. One of the main issues is that the solver we use does not reason about relational constraints that involve 2 symbols (ex: i < n).
> 
>> and a couple of format string issues.
> 
> Would those be caught with compiler warnings? (Try running clang on them with -Weverything.)
> 
>> Clang seems a bit better with double-free/use-after-free issues, and leak issues.
>> 
>> So it looks like there's some good stuff here, and we'll have a pretty specific set of things to work on! 
> 
> Thanks!
> Anna.
> 
>> 
>> Pretty cool, eh?
>> 
>> Sincerely,
>> Alexander Riccio
>> --
>> "Change the world or go home."
>> about.me/ariccio <http://about.me/ariccio>
>> 
>>  <http://about.me/ariccio>
>> If left to my own devices, I will build more.
>> ⁂
>> 
>> On Wed, Jan 20, 2016 at 1:11 AM, <Alexander G. Riccio> <test35965 at gmail.com <mailto:test35965 at gmail.com>> wrote:
>> A quick update on this project:
>> 
>> I've been slowed by a technical issue, and I lost ~2 weeks as two family members were in the hospital (sorry!).
>> 
>> Since I develop on Windows, I quickly hit a testcase that clang didn't detect, as I discussed in "Clang on Windows fails to detect trivial double free in static analysis".
>> 
>> That resulted in D16245 <http://reviews.llvm.org/D16245>, which (when accepted) fixes that issue. I want to ensure that novice can simply pass "--analyze", and clang to "just work", so I've intentionally put off further testing work. Otherwise, I could hack around it, and subsequently forget about the workaround. Once that's dealt with, then I can resume work at a faster pace.
>> 
>> 
>> 
>> Sincerely,
>> Alexander Riccio
>> --
>> "Change the world or go home."
>> about.me/ariccio <http://about.me/ariccio>
>> 
>>  <http://about.me/ariccio>
>> If left to my own devices, I will build more.
>> ⁂
>> 
>> On Mon, Jan 4, 2016 at 3:05 PM, Anna Zaks <ganna at apple.com <mailto:ganna at apple.com>> wrote:
>> 
>>> On Jan 2, 2016, at 12:45 PM, <Alexander G. Riccio> <test35965 at gmail.com <mailto:test35965 at gmail.com>> <Alexander G. Riccio> wrote:
>>> 
>>> Devin has started writing scripts for running additional analyzer tests as described in this thread:
>>> 
>>> A buildbot sounds like the perfect idea! 
>>> 
>>> The idea was to check out the tests/projects from the existing repos instead of copying them. Would it be possible to do the same with these tests?
>>> 
>>> Eh? What do you mean? Would that stop someone from running them in the clang unit test infrastructure?
>>> 
>>> I believe that these tests WILL need to be modified to run in the Clang testing infrastructure. 
>>> 
>> 
>> Currently, the analyzer is only tested with the regression tests. However, those need to be fast (since they effect all clang developers) and they have limited coverage. Internally, we’ve been testing the analyzer with the test scripts Devin described in the email I referenced. We use that testing method to analyze whole projects and long running tests. Those tests can and should be executed separately as they take more than an hour to complete. The plan is to set up an external builedbot running those tests. 
>> 
>> How long would it take to analyze the tests you are planning to add? Depending on the answer to that question, adding your tests to the new builedbot might be a better fit than adding them to the regression tests.
>> 
>> I also prefer not to modify the externally written tests since it would allow us to update them more easily, for example, when a new version of the tests comes out.
>> 
>>> Is there any way to treat static analyzer warnings as plain old warnings/errors? Dumping them to a plist file from a command line compilation is a bit annoying, and I think is incompatible with the clang unit testing infrastructure?
>>> 
>> 
>> Plist output is one if the outputs that the clang static analyzer supports. It is a much richer format than the textual warning since it contains information about the path on which the error occurred. We did have some lit tests checking plist output as well.
>> 
>>> 
>>> 
>>> Sincerely,
>>> Alexander Riccio
>>> --
>>> "Change the world or go home."
>>> about.me/ariccio <http://about.me/ariccio>
>>> 
>>>  <http://about.me/ariccio>
>>> If left to my own devices, I will build more.
>>> ⁂
>>> 
>>> On Mon, Dec 28, 2015 at 12:23 AM, Anna Zaks <ganna at apple.com <mailto:ganna at apple.com>> wrote:
>>> 
>>>> On Dec 17, 2015, at 11:01 AM, <Alexander G. Riccio> <alexander at riccio.com <mailto:alexander at riccio.com>> <Alexander G. Riccio> wrote:
>>>> 
>>>> However, if the goal is to have the tests
>>>> because you would like to make efforts to have the compiler diagnose
>>>> their cases properly, that's far more interesting and a good reason to
>>>> bring in the tests.
>>>> 
>>>> That's exactly my intention. Improving the static analyzer to detect these cases, that will be interesting.
>>>> placeholder text
>>>> 
>>>> If the other tests are not clearly licensed, we
>>>> should try to get NIST to clarify the license of them before
>>>> inclusion.
>>>> 
>>>> That sounds like the best idea, as a government agency, they almost certainly have lawyers.
>>>> 
>>>> I think the next step is to integrate the working (error correctly diagnosed) tests, only those that are obviously in the public domain, and propose them as a big batched patch. This shouldn't itself be controversial. 
>>>> 
>>>> How exactly do I submit a patch? I see that the LLVM developer policy says to send it to the mailing list (cfe-commits), but I also see that Phabricator comes into this somewhere <http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20151214/145026.html>?
>>>> 
>>> 
>>> Devin has started writing scripts for running additional analyzer tests as described in this thread:
>>> http://clang-developers.42468.n3.nabble.com/analyzer-Adding-build-bot-for-static-analyzer-reference-results-td4047770.html <http://clang-developers.42468.n3.nabble.com/analyzer-Adding-build-bot-for-static-analyzer-reference-results-td4047770.html>
>>> 
>>> The idea was to check out the tests/projects from the existing repos instead of copying them. Would it be possible to do the same with these tests?
>>> 
>>> Sorry for not replying sooner!
>>> Anna.
>>>> Sincerely,
>>>> Alexander Riccio
>>>> --
>>>> "Change the world or go home."
>>>> about.me/ariccio <http://about.me/ariccio>
>>>> 
>>>>  <http://about.me/ariccio>
>>>> If left to my own devices, I will build more.
>>>> ⁂
>>>> 
>>>> On Thu, Dec 10, 2015 at 4:04 PM, Aaron Ballman <aaron at aaronballman.com <mailto:aaron at aaronballman.com>> wrote:
>>>> On Mon, Dec 7, 2015 at 9:50 PM, <Alexander G. Riccio> via cfe-dev
>>>> <cfe-dev at lists.llvm.org <mailto:cfe-dev at lists.llvm.org>> wrote:
>>>> > First time Clang contributor here,
>>>> >
>>>> > I'd like to add the "C Test Suite for Source Code Analyzer v2", a
>>>> > relatively small test suite (102 cases/flaws), some of which Clang
>>>> > doesn't yet detect*. See link at bottom.
>>>> >
>>>> > Immediate questions:
>>>> > 0. Does the Clang community/project like the idea?
>>>> 
>>>> I've included a few other devs (CCed) to get further opinions.
>>>> 
>>>> I like the idea of being able to diagnose the issues covered by the
>>>> test suite, but I don't think including the test suite by itself is
>>>> particularly useful without that goal in mind. Also, one question I
>>>> would have has to do with the licensing of the tests themselves and
>>>> whether we would need to do anything special there.
>>>> 
>>>> > 1. What's the procedure for including new tests? (not the technical,
>>>> > but the community/project).
>>>> 
>>>> Getting the discussion going about the desired goal (as you are doing)
>>>> is the right first step.
>>>> 
>>>> > 2. How do I include failing tests without breaking things? Some of
>>>> > these tests will fail - that's why I'm proposing their inclusion - but
>>>> > they shouldn't yet cause the regression testing system to complain.
>>>> 
>>>> Agreed, any test cases that are failing would have to fail gracefully.
>>>> I assume that by failure, you mean "should diagnose in some way, but
>>>> currently does not". I would probably split the tests into two types:
>>>> one set of tests that properly diagnose the issue (can be checked with
>>>> FileCheck or -verify, depending on the kind of tests we're talking
>>>> about), and one set of tests where we do not diagnose, but want to see
>>>> them someday (which can be tested with expect-no-diagnostics, for
>>>> example). This way, we can ensure test cases continue to diagnose when
>>>> we want them to, and we can be alerted when new diagnostics start to
>>>> catch previously uncaught tests. This is assuming that it makes sense
>>>> to include all of the tests at once, which may not make sense in
>>>> practice.
>>>> 
>>>> > 3. How does Clang handle licensing of third party code? Some of these
>>>> > tests are clearly in the public domain (developed at NIST, says "in
>>>> > the public domain"), but others are less clearly licensed.
>>>> 
>>>> Oh look, you asked the same question I asked. ;-) If the tests are in
>>>> the public domain and clearly state as such, I think we can go ahead
>>>> and include them. If the other tests are not clearly licensed, we
>>>> should try to get NIST to clarify the license of them before
>>>> inclusion. Depending on the license, we may be able to include them
>>>> under their original license. If we cannot clarify the license, I
>>>> would guess that we simply should not include those tests as part of
>>>> our test suite. Note: I could be totally wrong, IANAL. :-)
>>>> 
>>>> > Should the community accept that testsuite, and I successfully add
>>>> > that test suite, then I'd like to step it up a bit, and include the
>>>> > "Juliet Test Suite for C/C++". "Juliet" is a huge test suite by the
>>>> > NSA Center for Assured Software & NIST's Software Assurance Metrics
>>>> > And Tool Evaluation project, which has 25,477 test cases (!!) for 118
>>>> > CWEs. I don't think any other open source compiler could compete with
>>>> > Clang after this. There's a ton of literature on the "Juliet" suite,
>>>> > and listing it here is not necessary.
>>>> >
>>>> > This project would be my first Clang contribution :)
>>>> >
>>>> > Personally, I'm interested in static analysis, and this is the first
>>>> > step in understanding & improving Clang's static analysis
>>>> > capabilities.
>>>> >
>>>> > I have some ideas on how to detect the currently undetected bugs, and
>>>> > I'm curious to see where things lead.
>>>> 
>>>> Adding the tests by themselves is not necessarily interesting to the
>>>> project unless they exercise the compiler in ways it's not currently
>>>> being exercised. So just having tests for the sake of having the tests
>>>> is not too useful (IMO). However, if the goal is to have the tests
>>>> because you would like to make efforts to have the compiler diagnose
>>>> their cases properly, that's far more interesting and a good reason to
>>>> bring in the tests.
>>>> 
>>>> One possible approach if you are interested in having the compiler
>>>> diagnose the cases is to bring the tests in one at a time. Start with
>>>> the initial batch of "these are diagnosed properly", then move on to
>>>> "this test is diagnosed properly because of this patch." Eventually
>>>> we'll get to the stage where all of the tests are diagnosed properly.
>>>> 
>>>> > Secondary questions:
>>>> > 1. How should I break the new tests up into patches? Should I just
>>>> > whack the whole 102 case suite into a single patch, or a bunch of
>>>> > smaller ones?
>>>> 
>>>> See comments above.
>>>> 
>>>> > 2. How does the Clang/LLVM static analysis testing infrastructure
>>>> > work? I'm going to have to figure this out myself anyways, but where
>>>> > should I start? Any tips on adding new tests?
>>>> 
>>>> http://clang-analyzer.llvm.org/checker_dev_manual.html <http://clang-analyzer.llvm.org/checker_dev_manual.html>
>>>> 
>>>> Another good place for some of these checkers may be clang-tidy, or
>>>> the compiler frontend itself. It's likely to depend on case-by-case
>>>> code patterns.
>>>> 
>>>> http://clang.llvm.org/extra/clang-tidy/ <http://clang.llvm.org/extra/clang-tidy/>
>>>> 
>>>> Thank you for looking into this!
>>>> 
>>>> ~Aaron
>>>> 
>>>> >
>>>> > *If I remember correctly,
>>>> > https://samate.nist.gov/SRD/view_testcase.php?tID=149055 <https://samate.nist.gov/SRD/view_testcase.php?tID=149055> passes
>>>> > analysis without complaint. I manually spot checked a very small
>>>> > number of tests.
>>>> >
>>>> > "C Test Suite for Source Code Analyzer v2" (valid code):
>>>> > https://samate.nist.gov/SRD/view.php?tsID=101 <https://samate.nist.gov/SRD/view.php?tsID=101>
>>>> > "C Test Suite for Source Code Analyzer v2" (invalid code):
>>>> > https://samate.nist.gov/SRD/view.php?tsID=100 <https://samate.nist.gov/SRD/view.php?tsID=100>
>>>> >
>>>> > "Juliet Test Suite for C/C++" (files):
>>>> > https://samate.nist.gov/SRD/testsuites/juliet/Juliet_Test_Suite_v1.2_for_C_Cpp.zip <https://samate.nist.gov/SRD/testsuites/juliet/Juliet_Test_Suite_v1.2_for_C_Cpp.zip>
>>>> > "Juliet Test Suite for C/C++" (docs):
>>>> > https://samate.nist.gov/SRD/resources/Juliet_Test_Suite_v1.2_for_C_Cpp_-_User_Guide.pdf <https://samate.nist.gov/SRD/resources/Juliet_Test_Suite_v1.2_for_C_Cpp_-_User_Guide.pdf>
>>>> >
>>>> >
>>>> > Sincerely,
>>>> > Alexander Riccio
>>>> > _______________________________________________
>>>> > 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>
>>>> 
>>> 
>>> 
>> 
>> 
>> 
> 
> 
> <minimized-stack_overflow-bad_commandline.txt><minimized-stack_overflow-bad.c>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20160127/3dc42463/attachment.html>

More information about the cfe-dev mailing list