Summary: public
The context leaks were reported multiple times. If a leaks was found on method `f()` and `g()` calls `f()`, then the same leak was report both in `f()` and in `g()`.
Reviewed By: sblackshear
Differential Revision: D2598110
fb-gh-sync-id: ca90b57
Summary: public
Extends the current activity leak checker to all sort of context leaks.
Reviewed By: sblackshear
Differential Revision: D2572548
fb-gh-sync-id: 9da18e4
Summary: public
The harness was created all the time but we do not report issues comming from the harness right now. The option:
infer --android-harness -- <build command>
Allows to create the harness
Reviewed By: sblackshear
Differential Revision: D2595211
fb-gh-sync-id: c7a6dc5
Summary: public
This allow to tell Infer to skip the translation of some files. This is especially useful to skip the translation of some generated files following the syntax:
> cat .inferconfig
{
"skip_translation": [
{
"language": "Java",
"source_contains": "_SHOULD_BE_SKIPPED_"
}
]
}
Reviewed By: cristianoc
Differential Revision: D2588095
fb-gh-sync-id: 3fda816
Summary: public
In C pre-increment/decrement returns rvalue, but in C++ it returns lvalue.
Make translation aware of the difference and treat these cases differently.
Reviewed By: ddino
Differential Revision: D2575136
fb-gh-sync-id: 952c095
Summary: public This diff fixes incorrect mangling of captured variables in blocks. Because they are formals,
they shouldn't be mangled, but this case was not taken into account. This caused an assert false in the
example infer/tests/codetoanalyze/objc/frontend/block/block.m which wasn't caught before because there
wasn't an endtoend test for it.
Reviewed By: akotulski
Differential Revision: D2560379
fb-gh-sync-id: db500b6
Summary: public
Adds incomplete translation of constructor bodies. Treat constructors as
methods with something 'extra'.
We still don't translate initializer lists, just pass the information to cTrans
where it's ignored
Reviewed By: dulmarod
Differential Revision: D2550214
fb-gh-sync-id: 102c13a
Summary: public anonymous types have file:line in its name.
Since file is relative path, type name can have '/' in its name.
This is very fragile since we might create file wiht typename in its name (for example for methods).
Replacing '/' with '_' should make frontend more resilient to failure.
Translation of anonymous structs is still pretty fragile (due to relative path in its name),
but at least it doesn't crash frontend
Reviewed By: dulmarod
Differential Revision: D2559936
fb-gh-sync-id: 647fd7f
Summary: public
C++ assignment operation result is lvalue, while in C it was rvalue.
This leads to different AST produced by clang for then same code!
Use language information from clang (`-x` flag) to distinguish these cases.
More specifically, let's look at following code:
int r;
int f = (r = 3);
// type of (r = 3) expression:
// C/objC -> int rvalue
// C++/objC++ -> int lvalue
Existing code did extra dereference because it was rvalue in C and there was no cast afterwards
in C++ there will be extra LValueToRvalue cast when neccesary so we don't have to do extra dereference manually
Reference:
http://en.cppreference.com/w/c/language/value_category (search for 'assignment and compound assignment operators')
NOTE: AST output doesn't change when something is hidden behind `extern "C"`, so we should use global language information
Reviewed By: ddino
Differential Revision: D2549866
fb-gh-sync-id: b193b11
Summary: public Two cases were not handled properly so far:
1. Declaration of a reference variable missed reference bit in type
2. Parameters to a function expecting T& had type T.
The way to distinguish reference types from value types is to look
whether parameter is type 'T rvalue' or type 'T lvalue' (xvalue probably as well)
Unfortunately, we can't just say 'T lvalue' = 'T&' because it would break
a lot of things in our frontend.
However, we know that when parameter to a function call has type 'T lvalue', it has to be 'T&' type.
Same applies when init_expression type is lvalue.
So, the solution is to add wrapper function that looks at results of `instruction` function and
expected expression type. Then if it's lvalue, wrap the type in reference.
Do this wrapping magic only when we know that lvalue mean reference type.
The rest of the changes is to make frontend tests pass - since we use different fields
in the AST, some of them were incorrectly set before and no one noticed.
Reviewed By: cristianoc
Differential Revision: D2549991
fb-gh-sync-id: 067f5d5
Summary: public
Dictionary literals are normally implemented using
`+dictionaryWithObjects:forKeys:count:` but were modeled as
`+dictionaryWithObjectsAndKeys:`
In particular, `@{@"aaa": nil}` would trigger a sentinel error instead of an NPE.
This models dictionary literals as a special infer builtin that the backend
interprets so as to give NPEs when passed nil objects or keys.
Reviewed By: dulmarod
Differential Revision: D2550039
fb-gh-sync-id: 1a10656
Summary: @public
This diff changes following things:
1. expression_info.type_ptr has type than decl_ref_info.type_ptr for reference types. Use type from decl_ref_info as a source of truth
2. reference types need to have one extra dereference that is not in AST. Add handling for this.
3. [small refactor] create function that creates temporary variable from res_trans expression and returns new res_trans.
Some caveats:
1. types are not quite right yet (see .dot files).
2. decl_ref_info might not be set for DeclRefExpr, make frontend crash in that case to catch when this happens
This is high risk change since it changes behavior of every translation on very widely used expr.
Reviewed By: @dulmarod
Differential Revision: D2540632
fb-gh-sync-id: aa28936
Summary: @public
This removes the old way of finding variable declarations to create sil variables and replaces it with
a a new way based on the map from pointers to declarations.
Basically, every variable dereference contains a pointer to the variable declaration, with that we can
build the corresponding sil variable.
Reviewed By: @akotulski
Differential Revision: D2536000
fb-gh-sync-id: dd29cf9
Summary: @public Add some basic tests to make sure that there is no
regression afterwards
Reviewed By: @dulmarod
Differential Revision: D2521887
fb-gh-sync-id: 1b8a15c
Summary: @public
Remove some of the duplicated code, move .dot files to new locations
Reviewed By: @dulmarod
Differential Revision: D2521709
fb-gh-sync-id: 0cc333d
Summary: @public
1. Add support for `this` keyword. It will allow to access fields/methods of the object from the method body.
2. Fix problem with method formals to add pointer to type of first parameter (which corresponds to `this`)
Reviewed By: @dulmarod
Differential Revision: D2484882
fb-gh-sync-id: c318619
Summary: @public
1. Factor out some of the common code for comparing C++ dot files
2. Create new directory with smaller .cpp files to translate
Reviewed By: @dulmarod
Differential Revision: D2507757
Summary: @public
First diff to give better language information in the frontend.
This information is necessary to understand when 'self' is objc keyword,
when 'this' is C++ keyword and when they are not.
Reviewed By: @ddino, @dulmarod
Differential Revision: D2489252
Summary: @public Infer previously did not work correctly when a function returns the result of a skip function:
```
retUndef() {
x = undefined();
return x;
}
derefUndef() {
y = retUndef();
y.doSomething(); // Symexec_memory_error here, prevents spec inference
}
```
The problem is that angelic mode did not know to add the return value of `retUndef()` to the footprint.
This diff fixes the problem by adding return values marked with the `Aundef` attribute to the footprint.
This is done lazily (e.g., a value only gets added to the footprint when you try to deref it).
Reviewed By: @jvillard
Differential Revision: D2444929
Summary:
frontend and backend assume that Sil.Struct
doesn't have methods nor inheritance. In order to plug C++
classes we probably need them to be Sil.Class
Summary:
In the new clang the parameters to these functions have notnull annotations, because of that infer tests fail. More concretely, the tests say there would be a memory leak. In the symbolic execution of those functions though, an inconsistency is created, because the parameter was nil, and the constraint argument should not be nil was also there, which leads to an error in the execution and no object is created, hence, no memory leak.
Summary:
Added two annotations @TrueOnNull and @FalseOnNull to be used for boolean functions to specify what value is returned when the argument is null.
Added model for TextUtils.isEmpty, which corresponds to the annotation
@TrueOnNull
static boolean isEmpty(@Nullable java.lang.CharSequence s)
Summary:
Pass inheritance information to the backend
It also changes some functions in cTypes_decl and we are using type and decl maps to resolve these types
Summary:
This is the second of 3 stack diffs to deal with replacing the parser of types.
This diff is about changes to translate record types, as well as class types and enum
types. For class types and enum types we store the declaration pointer in the map of
types to find the type easier later.
For record declarations, we change the way we build record names.
Moreover, we don't translate typedefs anymore, because when we have a pointer to a typedef,
we can find the actual type it points to.
Summary:
System.getProperty can return null when the property is not found, and expects a non-null argument.
Add models for Infer and Eradicate to reflect that.
Summary:
Fix the issue with Linux, introduced in 721cc1957c and temporarily fixed in aef13134bf
Now clang gets the `x86_64-apple-darwin14` target only when running on ObjCXX source code with a specific `isysroot` flag.
Summary:
each procedure has a different scope, so we can restart the fresh name generator and have more stable instructions in the cfg, that don't change when other procedures are changed
Summary:
Errors arising from overriding methods defined in other files were not reported, because during parallel analysis the clusters did not have access to overridden methods, so could not load their annotation.
Changed cluster generation to add location information for the methods overridden by the procedures defined in the current cluster.
Summary:
When someone runs --changed-only mode, there is a risk of corrupting the results
for future analyses. The problem is that changed-only mode does not analyze the callers of changed
procedures. If a subsequent analysis relies on the specs of one of these callers, they will be stale
and may give the wrong results. To be concrete, let's say we know `Parent.foo()` calls `Child.bar()` and we do the following rounds of analysis:
Analysis round 1: Analyze all files, including `Parent` and `Child`
Analysis round 2: Analyze `Child.bar()` only with `--changed-only flag`. `Parent.foo()` is now stale.
Analysis round 3: Add procedure `Parent.baz()` that calls `Parent.foo()`, analyze in (any) incremental mode.
The analysis will only analyze `Parent.baz()`. However, the specs for `Parent.foo()` are stale and may give us bad results for `Parent.baz()`. We want the analysis to re-analyze `Parent.baz()`, but before this diff it will not.
This diff fixes this problem by adding a `STALE` status bit to procedure summaries. In `--changed-only` mode,
the callers of a changed procedures are not re-analyzed, but their summaries are marked as stale. For both
`--changed-only` and regular incremental mode, callees of changed procedures that are marked as stale are
re-analyzed even if they have not changed. This is better than a more obvious solution like deleting stale
procedure summaries, since that would force the next analysis to re-analyze all stale procedures even if it
does not need the results for whatever analysis it is doing. This scheme implemented in this diff ensures
that each analysis only does the work that it needs to compute reliable results for its changed procedures.
Summary: Handler.postDelayed keeps a persistent reference to its Runnable argument that may cause a memory leak if an Activity is reachable from the Runnable.
Summary: The Nullable checker reported FP's when a Nullable field/param was reassigned to a non-Nullable value in the footprint. This diff fixes the problem.
Summary:
The @NonNull annotation, with camel case, can now be used to inform Eradicate that some fields that are not initialized by the constructor can be initialized by other means, e.g. via dependency injection.
Summary:
This test was actually testing: "at least one Field not initialized error is found" where we actualy want to test "exactly one Field not initialized error is found". The case of @Inject was also missing from the tests.
Summary:
Handle C++ method declarations and create cfgs for them.
Doesn't do:
Method calls (CXXMethodCall)
Using `this` expression in methods (including implicit ones)
Summary:
Add basic translation for C++ `new` keyword.
Currently, it's modeled as simple `malloc` call.
Following constructs are still not working properly:
- array new `new [size_expr]`
- run initializer attached to `new` (such as `new int(5)`)
- `delete[]`
Summary:
When detecting a resource leak, Infer used to raise an Leak exception and then prevent the specs to be computed for the paths containing a leak. This diff prevents resource leak to stop the analysis.
Summary:
Creating a persistent reference to an Activity leads to a nasty form of memory leaks (see http://android-developers.blogspot.com/2009/01/avoiding-memory-leaks.html, https://corner.squareup.com/2015/05/leak-canary.html). There are many ways to create a bad persistent reference to an Activity, but the most obvious one is via a static field.
This diff implements a very simple form of Activity leak checking by inspecting postconditions to see if a subtype of Activity is reachable from a static field (and it reports an error if so). This is a very simple and limited form of leak checking that does not understand the Android lifecycle at all. In particular, if one creates a persistent reference to an Activity and then nulls it out in `onDestroy` (a reasonably common pattern), this approach will wrongly report a bug.
Summary:
The methods in objc can have the same name in the same class, but one be instance and the other class,
so that we need to take the instance flag into account when defining unique names for ObjC methods.
Summary:
The symbolic execution was not stopping in case an unitialized dangling pointer was
passed to a function and then dereferenced inside the callee.
What would happen is that a wrong footprint would be added to the unititialized pointer
at the end of the function call in the caller proposition.
This checks that if we do:
frame * new_footprint
checks that we do not add heap predicates to the frame into uninitialized local variables.
If we can identify the variable then we raise a danglind pointer dereference. If instead
we cannot give a good explanation we give an internal error.
The latter case should be temporary. We should find a general way to raise dangling pointer
deref instead of the internal error.
I also fixed the model of getc that was the way I found the problem.
Summary:
This adds a sentinel check every time a function carrying a sentinel attribute
is called, regardless of whether we have a definition for that function or not.
Summary:
In objC we already prefix field names with classes.
It's better to make it consistent since it'll allow
us to share more code between C++ and objC
Summary:
Add a partial copy of TextUtils from Android source for commonly used TextUtils.isEmpty method.
Fixes#141
Closes https://github.com/facebook/infer/pull/143
Github Author: Deniz Türkoglu <deniz@spotify.com>
Summary:
The models for InputStreamReader and OutputStreamWriter are taking into consideration the charset passed as parameter in order to follow the exception branch when the charset is not valid. However, the previsous models were only considering encoding literals with uppercase letters. This diff adds the lowercase encoding names to the list.
Closes https://github.com/facebook/infer/issues/127
Summary:
Treat `arrayWithObjects` as a special case of a sentinel attribute check. This
will make it easier to extend to other variadic functions that use a sentinel
attribute.
This also removes the need for the `Sil.Avariadic_function_argument` attribute,
which will be removed in a subsequent diff.
Summary:
This reverts commit 306f5b71c24042c89f71848898402cbc9269c543.
Turns out that developers think that this bugs should be fixed. So leaving it in for now until I gather more information.
Summary:
@public
There are many FPs of the form init method that contains
if ((self = [super initWithFrame:frame])) {
...
}
return self;
then an object being initialised with that constructor and added to an array or dictionary.
There we flag NPE and very likely that won't be a bug. So I'm removing the option for self
to be nil in the constructor, which should solve the problem.
Test Plan: Changed the relevant test.
Summary:
@public
Previously, if the close() method was throwing an exception, then code overriding the file attribute with a mem attribute would be skipped, resulting in reporting a wrong resource leak. This diff fixes this.
Test Plan: Added new end-to-end tests which would previously have been failing
Summary:
@public
Add support for default function arguments.
As a side change - always create cmethod_signature for a function
Test Plan:
1. Call function with default parameter and confirm that it gets parsed and reports null dereference (B5 but still). It didn't before.
2. Created a test case
Summary: @publicThe first argument of builtin calls in C gets translated twice, which is bad if the argument is a side-effecting expression like a function call.
Test Plan: Attached test previously reported a memory leak because the translation introduces an extra call to malloc(), now reports nothing.
Summary:
@public
Translate CXXStaticCastExpr
Test Plan:
Add test, confirm that it gets translated.
Also create example to see that infer reports null dereference with this change:
struct X { int a; };
int main() {
X *x = static_cast<X*>(nullptr); // <- reports now
//X *x = (X*)nullptr; // <- reported before
return x->a;
}
Summary:
@public
Make c frontend understand CXXNullPtrLiteralExpr.
Note that the implementation differs from GNUNullExpr
Test Plan:
Create function that returns nullptr:
int* getPtr() {return nullptr;}
look at specs:
InferPrint infer-out/specs/getPtr\{831F\}.specs
Procedure: getPtr
int *getPtr()
Timestamp: 1
Status: INACTIVE
Phase: RE_EXECUTION
Dependency_map:
TIME:0.002853 s TIMEOUT:N SYMOPS:10 CALLS:1,0
ERRORS:
--------------------------- 1 of 1 [nvisited: 1] ---------------------------
PRE:
POST 1 of 1:
return = null:
----------------------------------------------------------------
Add test for it
Summary:
@public
Using InferBuiltins.assume previously caused an assertion failure in the analyzer. Fixed this, and fixed the implementation of the assume builtin to block when the assumed condition cannot hold.
Test Plan: Added several new tests.
Summary:
@public
Modeling bypasses the Closeable as resource assumption for `java.io.StringReader`, `java.io.ByteArrayInputStream` and `java.io.ByteArrayOutputStream`.
Test Plan: Infer CI. Some resource leak should also disappear on Instagram.
Summary:
@public
Sorting the fields in structs and classes. Was needed in the backend and forgotten.
Fixes the github issue https://github.com/facebook/infer/issues/90.
Test Plan: Added a new test that shows that we now get a spec for the example from the github issue.
Summary:
@public
The models for Java no longer require to keep the original fields since we now make the union of the fields from the models and the fields from the code to analyze.
Test Plan: Infer CI. No functional change intended.
Summary:
@public
Adds a small example of a Ant project in order to test that there is not regresssion when modifying the toplevel scripts
Test Plan:
cd infer/tests/codetoanalyze/java/infer
ant clean && infer -- ant compile
Summary:
@public
Attaching the resource attribute to the object allows to more easily remove this attribute during the symbolic execution when the resource is passed as a argument, e.g. with `res.close()` or when this resource is passed around via a skipped function.
Test Plan: Infer CI.
Summary:
@public
The clang location information is described in an incremental way: each location information is a delta with respect to the previous one in the AST. This is based on a the visit of the AST nodes which corresponds to the order in which the lines are printed with the standard clang AST dump:
clang -cc1 -ast-dump filename.c
This diff adds a preprocessing phase to the front-end so that location information is composed during a visit, and explicit location information is used instead.
In the case of include files, we report the last known location before including the file.
The current file for a function is the file where it is defined. So if a function is entirely defined in a .h file, then the location information will consistently be about the .h file. If instead a function is defined in the source file being analyzed, and some AST nodes come from macro expansion, line information will refer to the original file.
The front-end tests reveal that the location information was incorrect in a few dot files.
Test Plan: arc unit, after having fixed the wrong location in the existing .dot files
jrm
Andrzej Kotulski
Dulma Rodriguez
Jules Villard
Sam Blackshear
Cristiano Calcagno
martinoluca
Rohan Jacob-Rao
Dino Distefano
Deniz Türkoglu
Open Source Synchronization