Summary:
update-submodule: facebook-clang-plugins
Moving to a newer version of clang, see ffb5dd0114
Reviewed By: jvillard
Differential Revision: D5452529
fbshipit-source-id: 28bc215
Summary: Because making a diff which breaks the tests because it silently fails to create the right posts for the models is notoriously hard to debug
Reviewed By: sblackshear
Differential Revision: D5471611
fbshipit-source-id: ef04539
Summary:
:
because otherwise people would believe they can use the internal representation of these std lib but it fails for our models.
Reviewed By: jvillard
Differential Revision: D5368671
fbshipit-source-id: 4e53d5a
Summary:
We keep track of both `beginPtr` and `endPtr` but the modelling was mostly
about `beginPtr` as some sort of approximation I guess. This shouldn't change
much but will be useful later when doing more iterator stuff.
Reviewed By: mbouaziz
Differential Revision: D5255772
fbshipit-source-id: 0f6e3e8
Summary: This seems to move in the right direction. Also, `const operator[]` did not do an `access_at`, which I fixed.
Reviewed By: mbouaziz
Differential Revision: D5320427
fbshipit-source-id: c31c5ea
Summary:
:
There are throw wrapper functions like `std::__throw_bad_alloc()` defined in both libstdc++ (https://github.com/gcc-mirror/gcc/blob/master/libstdc%2B%2B-v3/include/bits/functexcept.h) and libc++ (e.g. 907c1196a7/include/new (L145)). Folly actually exports some of them as well (diffusion/FBS/browse/master/fbcode/folly/portability/BitsFunctexcept.h). The function body of those wrappers merely throws the corresponding exception. My understanding is that the primary purpose of the wrappers is to throw the exception if everything goes well and to fall back to something reasonable when exception is disabled (e.g. when `-fno-exceptions` is passed to the compiler).
The problem is that infer doesn't really understand what those functions do, and I've seen some false positives get reported as a result of it. So to remove those FPs we need to either model them or handle them specially. Modeling those wrappers by either whitelisting them or overriding the include files turns out to be difficult, as those wrappers are only declared but not defined in the STL headers. Their implementations are not available to Infer so whitelisting them does nothing, and if I provide custom implementations in the headers then normal compilation process will be disrupted because the linker would complain about duplicated implementation.
What I did here is to replace functions whose name matches one of the throw wrapper's name with a `BuiltinDecls.exit`. I have to admit that this is a bit hacky: initially I was trying to do something more general: replacing functions with `noreturn` attribute with `BuitinDecls.exit`. That did not work because, CMIIW, the current frontend only exports function attributes for functions with actual bodies, not declaration-only functions. I'd love to be informed if there are better ways to handle those wrappers.
Reviewed By: jeremydubreil
Differential Revision: D5266030
fbshipit-source-id: 4580227
Summary:
1. `noexcept` was missing from `unique_ptr` constructors leading to compilation errors in some edge cases
2. In case `unique_ptr` specified custom deleter with custom `deleter::pointer`, there could be compilation errors due to invalid cast from `deleter::pointer` to `void*`. Add extra overload of `model_set` to prevent this issue
Reviewed By: jberdine
Differential Revision: D5147071
fbshipit-source-id: 2586701
Summary:
Bufferoverrun-specific model for std::vector
Requires `--bufferoverrun` command line flag
Reviewed By: akotulski
Differential Revision: D4962136
fbshipit-source-id: f6b5f15
Summary:
`libstdc++` includes `bits/unique_ptr.h` via `include "unique_ptr.h"` in `bits/locale_conv.h` which infer can't redirect to model headers. This leads to compilation issues.
To work around that, redirect include of `locale_conv.h` to include `unique_ptr` via `include <...>` before normal `locale_conv.h`.
Reviewed By: mbouaziz
Differential Revision: D4962314
fbshipit-source-id: d4b9830
Summary: `__attribute__((annotate("")))` is better way of passing information to the frontend
Reviewed By: jberdine
Differential Revision: D4818805
fbshipit-source-id: 6e8add2
Summary:
Add support for Makefiles to the copyright linter. Makefiles are a bit
different than shell because they should start with the copyright notice
straight away (whereas shell starts with the #! stuff).
Reviewed By: mbouaziz
Differential Revision: D4786620
fbshipit-source-id: 504dc23
Summary:
It can be useful when debugging infer or the Makefiles themselves to see what
`make` is doing. Instead of editing Makefiles to remove `@` now you can `make
VERBOSE=1`.
This is just `git ls-files | grep -e Makefile -e '.*\.make' | xargs sed -e 's/^\t@/\t$(QUIET)/' -i`, and adding the definition of `QUIET` to Makefile.config.
Reviewed By: jeremydubreil
Differential Revision: D4779115
fbshipit-source-id: e6e4642
Summary:
This is helpful to make sure tests are up to date wrt the models.
Also made the Java deps depend on the models.jar instead of the model sources
as that's what the tests will be using. In particular, updating the sources of
the models will not update the results of a test unless someone rebuilds
models.jar, so rerunning the tests when the models haven't been rebuilt is
useless.
Reviewed By: akotulski
Differential Revision: D4635129
fbshipit-source-id: 75b4ab6
Summary:
Before: `make clean` followed by running `infer -- make`. If infer fails, it is
rerun automatically (by the `silent_on_success` Makefile function) to show the
output to the user, but by then there is nothing to build and `make` does
nothing.
Now: run directly `infer -- make clean all`. If infer fails, the command is
rerun and rebuilds all the source files, so there is a higher chance that the
same error will be displayed to the user than the one that originally caused
the command to fail.
Reviewed By: dulmarod
Differential Revision: D4578477
fbshipit-source-id: 774f45c
Summary:
Some classes may have deleted new operator for them. To fix it, run global `new` operator instead
```
struct X {
void* operator new(size_t) = delete;
};
X *p = new X; // compilation error
X *p = ::new X; // no compilation error
```
This change is following same strategy standard headers follow.
Reviewed By: jvillard
Differential Revision: D4500977
fbshipit-source-id: 20babfa
Summary:
Infer used to report null dereference when field was accessed later:
```
vector<int> v;
int& a = v[0]; // should be EMPTY_VECTOR_ACCESS here, but it wasn't reported
int b = a; // was NULL_DEREFERENCE here
```
To avoid this problem, model all accesses to vector as dereference of its internal `beginPtr` field.
Reviewed By: jberdine
Differential Revision: D4481942
fbshipit-source-id: 2142894
Summary:
Fixes issue with template argument deduction with enable_shared_from_this as argument
```
#include<memory>
template<class T>
void makeWeak(const std::shared_ptr<T>& x) {}
struct X : public std::enable_shared_from_this<X>{
};
void test() {
X x
makeWeak(x.shared_from_this()); // compilation failed here - it was unable to deduce template parameter of makeWeak
}
```
Reviewed By: jvillard
Differential Revision: D4414788
fbshipit-source-id: 4d19c53
Summary:
Turns out that swapping stdout and stderr using a temporary fd 3 was screwing
up with make's jobserver, who also uses fd 3!
Also, infer is partly to blame as it also calls `make`. Unsetting `MAKEFLAGS`
in infer tells `make` that the way infer calls `make` is independent from
parent `make` invocations.
Also, simplify the rules for direct tests and build system tests.
Reviewed By: jberdine
Differential Revision: D4328979
fbshipit-source-id: 96818e8
Summary: Our C++ model magic didn't work when instantiating smart pointers with volatile types. Fix it
Reviewed By: jvillard
Differential Revision: D4313271
fbshipit-source-id: 55ffb98
Summary: clang has very complicated logic what to translate based on `project_root` and filename. Add tests for different situations in regard of symbolic links in path/project_root
Reviewed By: jvillard
Differential Revision: D4168551
fbshipit-source-id: 586b364
Summary:
`install` will not do anything if the file didn't change, which should give
`make` more opportunities to not do work.
Reviewed By: jeremydubreil
Differential Revision: D4161918
fbshipit-source-id: 9b9061a
Summary:
Config.analyze_models, set by the INFER_ANALYZE_MODELS environment
variable, is redundant with Config.models_mode.
Reviewed By: jvillard
Differential Revision: D4047338
fbshipit-source-id: 4522d65
Summary:
1. models no longer need access to private fields (shared_ptr needed that)
2. create macro for __attribute__((deprecated("__infer_replace_with_deref_first_arg"))) and use it in models
Reviewed By: jberdine
Differential Revision: D3791113
fbshipit-source-id: 532dd33
Summary:
Follow strategy that was done to `std::shared_ptr` model and translate
`std::unique_ptr<T>` as raw pointer `T*`.
As a bonus, model `operator[]` of array overload as dereference
Reviewed By: jvillard
Differential Revision: D3785031
fbshipit-source-id: 2c5b0a4
Summary:
This reverts commit 7e3f8ffc84980ea5bcb4dae63a1588c07460c946.
Workaround is no longer needed since new clang we use has fix for this:
https://llvm.org/bugs/show_bug.cgi?id=28519
Reviewed By: dulmarod
Differential Revision: D3791083
fbshipit-source-id: ec339f1
Summary:
1. Add capability to clang frontend to replace some function calls with another SIL code based on `__deprecated__` attribute.
2. Given this capability, use those attributes for shared_ptr getters to generate `Sil.Load` instruction instead of method call
3. Add test that mimics shared_ptr model, but it doesn't have that much scary C++ templated code
Reviewed By: jvillard, jberdine
Differential Revision: D3729176
fbshipit-source-id: 2a330d5
Summary:
Make std::shared_ptr<T> translated as T* inside infer. This will make reporting better
since smart pointers are really pointers not structs - this form is much easier for the analyzer to understand.
This requires changes to the model of shared_ptr as well.
Reviewed By: jvillard
Differential Revision: D3587255
fbshipit-source-id: b86fb36
Summary:
So far infer had very fragile mechanism to detect smart pointers. It was looking for "std" and "(shared|unique)_ptr" inside name string.
This is easy to trick (like mystd::shared_ptr) and not something we want.
Instead, inside models create models inside infer_std_model namespace. Then just "export" that model into std namespace
via `using shared_ptr = infer_std_model<T>;`
Reviewed By: jvillard
Differential Revision: D3703827
fbshipit-source-id: 9640fc2
Summary:
Use Itanium mangling for C++ functions/methods instead of raw type name. This is a step towards removing expensive `ti_raw` field from `type_info`.
For virtual methods, use mangled name of the method from base class in order for dynamic dispatch to work.
Reviewed By: dulmarod
Differential Revision: D3556118
fbshipit-source-id: e45edb5
Summary:
There is nice hook inside glog/logging.h for all CHECK_(GT|LT|GE|...) macros.
This simplifies AST significantly which makes infer way more happy:
cda16b3443/src/glog/logging.h.in (L722-L724)
Reviewed By: jberdine
Differential Revision: D3522110
fbshipit-source-id: 70c94cb
Summary:
Assume that std::vector::resize will always create nonempty vector. While this is clearly
wrong for resize(0), it removes many FPs for `resize(n)` calls, where value of `n` is unknown.
Without it, infer was thinking that `n` could be 0 and reported empty vector access.
Reviewed By: jvillard
Differential Revision: D3424355
fbshipit-source-id: cb476de
Summary:
Reuse data fields of model's superclass (which is actual implementation).
Not very pretty, but makes sizeof shared_ptr same as in actual libraries
Reviewed By: jvillard
Differential Revision: D3398628
fbshipit-source-id: bdb9418
Summary:
Hardcoding `variable@` in Makefiles is Bad™ because it prevents the users from
overwriting them easily with `make variable="my custom value"`. The right way
to do it is thus:
```
variable = variable@
# then use $(variable) everywhere
```
This diff puts all the `variable = variable@` lines in Makefile.config.in, and
changes every occurrence of a `variable@` to `$(variable)` everywhere else.
I mostly automated generating this diff. Here are the steps I did:
- find out which `variable@`s we use:
find . -name 'Makefile*' -exec grep -e '@[^@ []\+@' -o -h \{\} \+ | sort | uniq > config_variables
- write this `replace.sh` script to replace every `variable@` with `$(variable)`:
```
#!/bin/sh
config_vars_file=$1
shift
for line in $(cat $config_vars_file); do
var=$(echo $line | tr -d @)
sed -i -e "s/$line/\$($var)/g" $@ > /dev/null
done
```
- run the script as such:
find . -name 'Makefile.*in' \( -not -wholename './Makefile.config.in' \) -exec ./replace.sh config_variables \{\} \+
- put all the `VARIABLE = VARIABLE@` lines in Makefile.config.in
- move all `Makefile.in` to `Makefile`, since they don't need to be generated by `./configure` anymore:
```
for i in $(find . -name 'Makefile.*in' \( -not -wholename './Makefile.config.in' \)); do \
rm $(dirname $i)/$(basename $i .in) && git mv $i $(dirname $i)/$(basename $i .in) ; \
done
```
- delete all Makefile except Makefile.config from configure.ac
- manually inspect and remove remaining instances of `VAR = $(VAR)` in makefiles, looking at the output of `git grep '^\(\w\+\) = $(\1)'`
Reviewed By: jberdine
Differential Revision: D3358379
fbshipit-source-id: 5d37f02
Summary:
Turns out, analyzer was getting confused with complicated
model and it was reporting empty access in places it
shouldn't. Fixing backend is not trivial (tracing mode is the answer),
but the model can be simplified.
It introduces the problem that get() method doesn't return fresh value
every time, but we should be able to change backend later to deal with it.
Reviewed By: sblackshear
Differential Revision: D3328228
fbshipit-source-id: dddbaf8
Martino Luca
Jeremy Dubreil
Jia Chen
Mehdi Bouaziz
Andrzej Kotulski
Jules Villard
Josh Berdine