Summary:
At function calls, it copies a subset of heap memory that is newly
allocated by callees and is reachable from the return value.
Reviewed By: mbouaziz
Differential Revision: D7081425
fbshipit-source-id: 1ce777a
Summary:
This is to fix the conflicts between Eradicate and the Biabduction when reporting the same kind of errors: when Eradicate is on, the Eradicate warnings will have priority over the null deference reported by the biabduction.
If this approach proved to be successful in prod, I will refactor the reporting mechanism in the analysis itself to simply not report the null dereference in this case at all. For the codebases that aren't yet fully consistently using `Nullable`, this combined approach looks like a good way to deploy Infer toward full null safety.
Reviewed By: mbouaziz
Differential Revision: D7102119
fbshipit-source-id: 35d3add
Summary:
Before D7100561, the frontend translated capture-by-ref and capture-by-value in the same way.
Now we can tell the difference and report bugs in the capture-by-value case.
Reviewed By: jeremydubreil
Differential Revision: D7102214
fbshipit-source-id: e9d3ac7
Summary:
The `may_last_field` boolean value in the `decl_sym_val` function presents that the location *may* (not *must*) be a flexible array member.
By the modular analysis nature, it is impossible to determine whether a given argument is a flexible array member or not---because of lack of calling context. For example, there are two function calls of `foo` below: (2) passes a flexible array member as an argument and (1) passes a non-flexible array, however it is hard to notice when analyzing the `foo` function.
```
struct T {
int c[1];
};
struct S {
struct T a;
struct T b;
};
void foo(struct T x) { ... }
void goo () {
struct S* x = (struct S*)malloc(sizeof(struct S) + 10 * sizeof(int));
foo(&(x->a)); // (1)
foo(&(x->b)); // (2)
}
```
We assume that any given arguments may stem from the last field of struct, i.e., flexible array member. (This is why `decl_sym_val` is called with `may_last_field:true` at the first time.) With some tests, we noticed that the assumption does not harm the analysis precision, because whether regarding a parameter as a flexible array member or not is about using a symbolic array size instead of a constant array size written in the type during the analysis of callee. Therefore still it can raise correct alarms if the actual parameter is given in its caller.
Reviewed By: mbouaziz
Differential Revision: D7081295
fbshipit-source-id: a4d57a0
Summary:
Switch to the current stable branch for clang.
update-submodule: facebook-clang-plugins
Reviewed By: mbouaziz
Differential Revision: D7067890
fbshipit-source-id: aedff90
Summary:
You can capture a variable by reference in a lambda, assign to it, and then invoke the lambda.
This looks like a dead store from the perspective of the current analysis.
This diff mitigates the problem by computing an additional analysis that tracks variables captured by ref at each program point.
It refuses to report a dead store on a variable that has already been captured by reference.
Later, we might want to incorporate the results of this analysis directly into the liveness analysis instead of just using it to gate reporting.
Reviewed By: jeremydubreil
Differential Revision: D7090291
fbshipit-source-id: 25eeffa
Summary:
Ran into this issue on Debian Testing, in which assert.h is probably different
due to a more recent toolchain. Without this change I get the following CFG
for `assert(e)`:
```
start
|-> prune (e) -> join
|-> prune (!e) -> __infer_fail("ASSERTION_FAILURE") -> exit
```
Notice that the first branch does not get to the exit, so infer must think that
the assertion is *always* violated, and reports `error: ASSERTION_FAILURE`.
This is broken.
Reviewed By: dulmarod
Differential Revision: D7067822
fbshipit-source-id: a2bf5ac
Summary:
It supports flexible array member using the following heuristic:
- a memory for a class is allocated by `malloc(sizeof(C) + n * sizeof(T))` format
- the last field of the class is an array
- the static size of the last field is one, i.e., `T field_name[1]`
When allocating and initializing members of classes, it sets the size of flexible array to `n+1` if the above conditions are met.
Reviewed By: mbouaziz
Differential Revision: D7056291
fbshipit-source-id: 31c5868
Summary:
The semantics of "placement new" is defined simply as an assignment.
For example, `C* x = new (y) C();` is analyzed as if `C* x = y;`.
Reviewed By: mbouaziz
Differential Revision: D7054007
fbshipit-source-id: 1c6754f
Summary:
The struct fields in Cil have been sorted for long time, however the
checkers do not seem to depend on the sortedness.
Reviewed By: sblackshear
Differential Revision: D7027858
fbshipit-source-id: 9e7ab96
Summary:
This commit improves precision of symbol instantiations.
When a return value of a callee is `[s1 + s2, _]` and if we want to
instantiate `s1` to `c3 + max(c4, s5)`, the lower bound was
substituted to `-oo` because our domain cannot express `c3 + max(c4,
s5) + s2`.
However, we can have instantiations that are preciser than `-oo`:
(1) `c3 + c4 + s2`
(2) or `c3 + s5 + s2`
because they are smaller than the ideal instantiation, `c3 + max(c4,
s5) + s2` and it is on the lower bound position.
For now, the implementation instantiates to (1) between the two ones,
because constant values introduced by `assert` or `assume`(`if`)
command are often used as safety conditions, e.g., `assert(index >=
0);` can place before array accesses. (We can change the stratege
later if we find that it doesn't work on some other cases.)
Reviewed By: mbouaziz
Differential Revision: D7020063
fbshipit-source-id: 62fb390
Summary:
A simple intraprocedural analysis that tracks when a storage location is read or deleted.
For now, this works only with local variable storage locations; field and array accesses are ignored.
In order to test this, I added a new "use-after-lifetime" warning. It complains when a variable is read or deleted after it has already been deleted.
Reviewed By: jeremydubreil
Differential Revision: D6961314
fbshipit-source-id: 75e95a2
Summary: We do not inject a destructor call if the destructor declaration does not contain a body in AST. We miss all the cases where the destructor is declared in `.h` file and defined in `.cpp` file as other files include `.h` file and do not contain the body of the destructor when destructor calls are being injected based on AST information. After this diff we inject destructor calls even if we do not have body for the destructor in AST.
Reviewed By: sblackshear
Differential Revision: D6796567
fbshipit-source-id: 1c187ec
Summary:
It prunes abstract memories on `assert` commands.
Problem: Since the assert command is sometimes translated to two
sequential `if` statments, it was not able to prune the memory
precisely at `assert` commands in Inferbo---the pruned memory at the
first branch was joined before the second branch.
Solution: To avoid losing the pruning information at the first branch,
now, it records which locations are pruned at the first branch and
applies the same pruning at the next branch if they have
semantically the same condition.
Reviewed By: mbouaziz
Differential Revision: D6895919
fbshipit-source-id: 15ac1cb
Summary:
See comment--`Prop(resType = blah) myProp` will generate `.myProp`, `.myPropRes`, and `.myPropAttr`, and any of them can be used to set the prop.
Because our annotation parameter parsing is a bit primitive, handle this by simply checking the `Res` and `Attr` suffixes for every `Prop`.
This shouldn't lead to false negatives because these methods will only exist if the `resType` annotation is specified anyway.
Reviewed By: jeremydubreil
Differential Revision: D6955362
fbshipit-source-id: ec59b21
Summary: In Obj-C blocks, we explicitly insert reads of the captured vars. This does the same thing for C++. For example, `foo() { int x = 1; [x]() { return x; } }` would previously not contain a read of `x` in `foo`. Now, we'll create a temporary that reads from `x` and pass it to the closure value.
Reviewed By: dulmarod
Differential Revision: D6939997
fbshipit-source-id: f218afc
Summary:
1) Fixes some false negatives when a method annotated with `nullable` in the header is not annotated in the implementation and the attribute lookup returns the implementation. In that case, we should follow the information given in the header.
2) Fixes some false positives when annotations are in the other way around, i.e. annotated in the implementation but not in the headers. For now, there should be no report in this case, but the analysis should be extended to report the inconsistency between the header and the implementation
3) Fixes some cases of weird reports caused by name conflicts where the method in the include has the same name has another method annotated differently.
Reviewed By: jvillard
Differential Revision: D6935379
fbshipit-source-id: 3577eb0
Summary:
Added a check for recursive calls not to add abduced reference parameters constraints. Abduced reference parameters constraints were causing assertion failure when renaming variables in specs, in particular, when transforming variables into callee variables.
A similar check is already in place for abduced retvals constraints.
Reviewed By: jeremydubreil
Differential Revision: D6856919
fbshipit-source-id: acfe840
Summary:
This allows Eradicate to lookup the annotations from the classpath and without requiring the code in the classpath to have been previously analyzed. The benefit is that source files can be analyzed independently of each other as long as the classpath is known.
The main goal is to run be able to run Eradicate as a linter without losing warnings.
We may have to add some more models of the standard libraries as no `Nullable` on a parameter does not necessarily mean that the method does not accept `null`.
Reviewed By: sblackshear
Differential Revision: D6921720
fbshipit-source-id: f525269
Summary:
- Combine two fields from ProcAttributes.t into a single field `method_kind` with more information
- New field details whether the procedure is an `OBJC_INSTANCE`, `CPP_INSTANCE`, `OBJ_CLASS`, `CPP_CLASS`, `BLOCK`, or `C_FUNCTION`
- `is_objc_instance_method` and `is_cpp_instance_method` fields no longer necessary
- Changed `is_instance` field in CMethod_signature to `method_kind` field of type ProcAttributes.method_kind
Reviewed By: dulmarod
Differential Revision: D6884402
fbshipit-source-id: 4b916c3
Summary:
Record "capture phases" in the runstate and in the source files table of the
database. Use this instead of filesystem timestamps to decide which files need
re-analyzing in the reactive analysis.
Reviewed By: jeremydubreil
Differential Revision: D6760833
fbshipit-source-id: 7955621
Summary:
The boolean lock domain is simple and surprisingly effective.
But it's starting to cause false positives in the case where locks are nested.
Releasing the inner lock also releases the outer lock.
This diff introduces a new locks domain: a map of locks (access paths) to a bounded count representing an underapproximation of the number of times the lock has been acquired.
For now, we just use a single dummy access path to represent all locks (and thus a count actually would have been sufficiently expressive; we don't need the map yet).
But I'm planning to remove this limitation in a follow-up by refactoring the lock models to give us an access path.
Knowing the names of locks could be useful for error messages and suggesting fixes.
Reviewed By: jberdine
Differential Revision: D6182006
fbshipit-source-id: 6624971
Summary:
Previously, we could understand than an access was safe either because it was possibly owned or protected by a thread/lock, but not both. If an access was both protected by a lock and rooted in a paramer (i.e., possibly owned), we would forget the ownership part of the precondition and remember only the lock bit. This leads to false positives in cases where an access protected by a lock is owned, but another unowned access to the same memory is not protected by a lock (see the new `unownedLockedWriteOk` E2E test for an example).
This diff makes access safety conditions disjunctive so we can simultaneously track whether an access is owned and whether an access is protected by a thread/lock. This will fix false positives like the one explained in T24015160.
Reviewed By: jberdine
Differential Revision: D6671489
fbshipit-source-id: d17715f
Summary:
We already knew not to warn when non-resource `Closeable`'s like `ByteArrayOutputStream` weren't closed, but we still warned on their subtypes.
This diff fixes that problem by checking subclasses in the frontend.
This also removes the need for Java source code models of non-resource types, so I removed them.
Reviewed By: jeremydubreil
Differential Revision: D6843413
fbshipit-source-id: 60fe7fb
Sungkeun Cho
Jeremy Dubreil
Sam Blackshear
Jérémie Marguerie
Dino Distefano
Jules Villard
Dulma Churchill
Daiva Naudziuniene
Varun Arora