[PATCH] Modeling APIs in the Static Analyzer

[Ted Kremenek]

Hi Gábor,

Thanks for making progress on this very promising enhancement to the analyzer.  I have an assortment of comments, in no particular order:

- ModelInjector.h and ModelConsumer.h

There is a comment at the top of these files, but I think a bit more explanation is needed.  For example:

  MetaConsumer.cpp:

    +// "Meta" ASTConsumer for consuming model files.

That doesn't really explain anything.  What does "Meta" in quotes mean?  I think an explanation here on what this does is helpful when someone discovers this code for the first time.

Similarly, we should add some high-level comments for CodeInjector.h and ModelInjector.h.  We have a good start in ModelInjector.h:

+/// \file
+/// \brief Defines the clang::ento::ModelInjector class which implements the
+/// clang::CodeInjector interface. This class is responsible for injecting
+/// function definitions that were synthetized from model files.
+///

Let's consider expanding it:

 /// \brief This file defines the clang::ento::ModelInjector class which implements the
 /// clang::CodeInjector interface. This class is responsible for injecting
 /// function definitions that were synthesized from model files.

 /// Model files allow definitions of functions to be lazily constituted for functions
 /// which lack bodies in the original source code.  This allows the analyzer
 /// to more precisely analyze code that calls such functions, analyzing the
 /// artificial definitions (which typically approximate the semantics of the
 /// called function) when called by client code.  These definitions are
 /// reconstituted lazily, on-demand, by the static analyzer engine.

CodeInjector.h provides some information, but it is a bit vague:

+///
+/// \file
+/// \brief Defines the clang::CodeInjector interface which is responsible for
+/// injecting AST of function definitions from external source.
+///

It's a bit unclear how this gets used.  I think a bit of prose here would help clarify its role in the static analyzer.  I also think the CodeInjector interface is also more abstract than the prose describes.  There's nothing about CodeInjector's interface that requires the injected definitions to come from an external source.  That's an implementation detail of a concrete subclass.  Instead, all CodeInjector does is provide an interface that lazily provides definitions for functions and methods that may not be present in the original source.

I'm also looking at the change to StaticAnalyzer/Frontend/FrontendActions.cpp, and wonder if we can simplify things:

> +++ lib/StaticAnalyzer/Frontend/FrontendActions.cpp (working copy)
> @@ -7,9 +7,11 @@
>  //
>  //===----------------------------------------------------------------------===//
>  
> +#include "clang/Frontend/CompilerInstance.h"
>  #include "clang/StaticAnalyzer/Frontend/FrontendActions.h"
> -#include "clang/Frontend/CompilerInstance.h"
>  #include "clang/StaticAnalyzer/Frontend/AnalysisConsumer.h"
> +#include "clang/StaticAnalyzer/Frontend/ModelConsumer.h"
> +#include "ModelInjector.h"
>  using namespace clang;
>  using namespace ento;
>  
> @@ -18,6 +20,14 @@
>    return CreateAnalysisConsumer(CI.getPreprocessor(),
>                                  CI.getFrontendOpts().OutputFile,
>                                  CI.getAnalyzerOpts(),
> -                                CI.getFrontendOpts().Plugins);
> +                                CI.getFrontendOpts().Plugins,
> +                                new ModelInjector(CI));
>  }
>  


It looks like CreateAnalysisConsumer just continues to grow more arguments, all which derive from using 'CI'.  This seems silly, since this function is called in one place.  Instead of intro ducting a dependency on ModelInjector.h in this file, we can just sink these arguments into CreateAnalysisConsumer() itself, resulting in:

  return CreateAnalysisConsumer(CI);

and let CreateAnalysisConsumer() do all that boilerplate.

Next, let's look at the change to FrontendAction:

>  class FrontendAction {
> +  /// Is this action invoked on a model file? Model files are incomplete
> +  /// translation units that relies on type information from another translation
> +  /// unit. Check ParseModelFileAction for details.
> +  bool ModelFile;

Perhaps "IsModelFile"?  "ModelFile" sounds like it should be a reference to the file itself.

>    FrontendInputFile CurrentInput;
>    std::unique_ptr<ASTUnit> CurrentASTUnit;
>    CompilerInstance *Instance;
> @@ -105,7 +109,11 @@
>    /// @}
>  
>  public:
> -  FrontendAction();
> +  /// \brief Constructor
> +  ///
> +  /// \param modelFile determines whether the source files this action invoked
> +  /// on should be treated as a model file. Defaults to false.
> +  FrontendAction(bool modelFile = false);

It seems suboptimal to modify the interface of FrontendAction just for this one edge case.  Instead of modifying the constructor arguments, we could default initialize "IsModelFile" to false, and have a setter to change it.  For example:

  ParseModelFileAction::ParseModelFileAction(llvm::StringMap<Stmt *> &Bodies)
    : ASTFrontendAction(/*ModelFile=*/true), Bodies(Bodies) {}

becomes:
 
  ParseModelFileAction::ParseModelFileAction(llvm::StringMap<Stmt *> &Bodies)
    : Bodies(Bodies)  {
    IsModelFile = true;
  }

Looking at this more, I wonder if we should modify FrontendAction at all.  The only place where isModelParsingAction() is called is in one spot in CompilerInstance.cpp:

   if (hasSourceManager() && !Act.isModelParsingAction())

It *might* be cleaner to just have a virtual member function in FrontendAction, which defaults to returning false, but is generic for all subclasses to override.  Then we don't need the "IsModelFile" field in FrontendAction at all, and we just have ParseModelFileAction override that single member function.  We could then name that method to be something a bit more generic.  That would allow us to not touch FrontendAction at all except for providing that single virtual method that can be overridden in subclasses.  I somewhat prefer this approach because it provides a cleaner separation of concerns between FrontendAction (which is defined libFrontend) and the static analyzer.  That would also allow you to get rid of isModelParsingAction() entirely (replacing it with something more generic).

As for the test case:

> +typedef int* intptr;
> +
> +void modelled(intptr p);
> +
> +int main() {
> + modelled(0);
> + return 0;
> +}

Please add some comments in this test file explaining what is happening.  Also, it would be great if this both used FileCheck (which it does now) but also verified the diagnostics so we get cross-checking of the output (we see this in some analyzer tests).  It also makes it easier to understand the test.

Also, is there a reason to break up the tests between model-suppress-falsepos.cpp and model-file.cpp?  It seems like one test file will do fine; just clearly comment on what is happening for each test.  I also recommend called the modeled function "modeledFunction" instead of "modelled" (which according to my spell checker has an additional 'l'). 

As for the model files themselves:

> Index: test/Analysis/modelled.model
> ===================================================================
> --- test/Analysis/modelled.model  (revision 0)
> +++ test/Analysis/modelled.model  (working copy)
> @@ -0,0 +1,3 @@
> +void modelled(intptr p) {
> + ++*p;
> +}
> \ No newline at end of file
> Index: test/Analysis/notzero.model
> ===================================================================
> --- test/Analysis/notzero.model (revision 0)
> +++ test/Analysis/notzero.model (working copy)

Let's put these in a separate subdirectory, for example, "models", instead of mixing them with the tests.  This way they really serve as "inputs" to the analyzer.

Overall this is looking good.  I think the explanatory comments will really help people understand what this is doing, and I think changing how we thread the information through FrontendAction will help not introduce an artificial tainting of FrontendAction with concepts specific to the static analyzer.

Cheers,
Ted


On Jul 16, 2014, at 2:45 AM, Gábor Horváth <xazax.hun at gmail.com> wrote:

> 
> 
> 
> On 14 July 2014 19:32, Anna Zaks <ganna at apple.com> wrote:
> 
>> On Jul 13, 2014, at 6:11 AM, Gábor Horváth <xazax.hun at gmail.com> wrote:
>> 
>> Hi Anna,
>> 
>> Thank you for the review. I have tweaked the test, so it no longer requires the error reporting tweak that is not done yet to pass. I have also added some high level comments to some files, if you think some information is lacking I will add them in the next iteration as well. The BugReporter patch is now separated into a different patch. 
>> 
>> 
>> On 11 July 2014 18:02, Anna Zaks <ganna at apple.com> wrote:
>> 
>> For example, modeling functions should allow you to find bugs and suppress false positives outside of those functions. I would suggest adding a few of those tests first.
>> 
>> 
>> How are the false positives suppressed? I did not find any resource on that. Found some analyzer attributes but I did not find them suitable for this purpuse at the first glance. But I think once the locations that are in a model file are omitted from the report path, the regular methods for suppressing false positives should work (and I will definitely add test case to ensure this once it is done).
>> 
> 
> What I meant is that it is possible to construct a test where ability to model a function would eliminate a false positive. This would be another way to test your patch without worrying about BugReporter.
> 
> I got it now, thansk. I have updated the patch with a test case where a false positive case is eliminated by a model file.
> 
> Thanks,
> Gábor
>  
>> Thanks,
>> Gábor
>> <api_modeling.patch><bugreporter.patch>
> 
> 
> <api_modeling.patch><bugreporter.patch>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20140729/7c8b5eee/attachment.html>