A collection of clang plugins for safer C programming. The plugins detect problems at compile time.
- enums_conversion: Finds implicit casts to/from enums and integral types
warn_unused_result: Warns about functions declared without thewarn_unused_resultattribute- include_cleaner: Finds unused #includes
- large_assignment: Finds large copies in assignments and initializations (size is configurable)
- private: Prevents access to fields of structs that are defined in private.h files
See below for more information about each plugin.
Each plugin has a test directory with both positive and negative test cases, which also serve as good examples.
Building the plugin requires llvm/clang development libraries to be installed, it expects an llvm-config-64 to be available.
On debian or ubuntu, try installing: llvm-3.7-dev clang-3.7 libclang-3.7-dev. You may need to use a PPA.
Tested also on Centos 7.1, using the packages: llvm-static, llvm-libs, llvm-devel, clang-devel. You may need to get them from copr.
Once you have the prerequisites, build & test:
make
make tests
All the plugins are linked into a single .so file, for convenience, but are controlled individually via clang command line arguments.
Example usage (once clang_plugins.so is built):
clang -Xclang -load -Xclang ./clang_plugins.so \
-Xclang -add-plugin -Xclang include_cleaner \
-Xclang -add-plugin -Xclang enums_conversion \
-Xclang -add-plugin -Xclang large_assignment -Xclang -plugin-arg-large_assignment -Xclang 700 \
-c my_file.c -o my_file.o
See CFLAG_CLANG_PLUGINS in plugins.mk for the full command line argument for using all plugins.
(Check enums_conversion/README.md for more information about the enums_conversion plugin.)
In C, enum types are implicitly convertible to other integral types. The following code compiles without warnings, even
with -Wall -Wextra (tested on GCC 6.3 & clang 3.7):
enum Result { OK, NO_BUNNIES };
enum Result get_bunnies(void) {
return 3;
}
The enums_conversion plugin prevents such silly mistakes:
clang -Xclang -load -Xclang ./clang_plugins.so -Xclang -add-plugin -Xclang enums_conversion -c /tmp/test.c
/tmp/test.c:7:12: error: enum conversion to or from enum Result
return 3;
^
1 error generated.
Among other cases detected by the plugin:
- Passing, assigning or returning a value of the wrong type when an enum is expected
- Using enums as bool in conditional expressions
- Using values of wrong type in enum flag bitwise or
|expressions - etc.
There are cases (or styles of C coding) that use enums as actual numbers (such as array indices). In those cases you can use an explicit cast to make the fact that the enum needs to be converted, well, explicit.
Note: The enums_conversion plugin uses AST matching to find violations - it works by looking for specific problematic patterns. A more complete approach would be to traverse the entire AST and perform a complete type-checking pass (or to augment the existing clang type checker, but as far as I can tell, a clang plugin can't do that). However, we found the AST matching approach to be quite effective, and it was easy to implement.
Both clang and gcc support the attribute warn_unused_result on functions, but provide no way to treat all functions as if they have the attribute, nor is there a way to warn about functions that don't have the attribute.
The warn_unused_result plugin will emit a warning functions that lack a warn_unused_result attribute. The plugin only considers functions declared or defined in the current compilation unit.
The plugin accepts an optional arg, --static-only which causes it to warn only about static functions (for when changing external APIs is too much work).
Example:
int foo(void);
Compiler output:
/tmp/test.c:1:5: warning: missing attribute warn_unused_result
int foo(void);
The include_cleaner plugin helps you get rid of unnecessary includes, for example:
foo.h:
struct Foo { };
bar.c:
#include "foo.h"
// never uses anything from foo.h
In this case, the plugin will give an error:
clang -Xclang -load -Xclang ./clang_plugins.so -Xclang -add-plugin -Xclang include_cleaner -c /tmp/bar.c
/tmp/bar.c:1:1: warning: include cleaner: unused #include of '/tmp/foo.h'
#include "foo.h"
^
1 warning generated.
The plugin expects your code to follow a certain pattern:
foo.cis always allowed to #includefoo.horfoo_api.h- External files are always allowed to be #included (currently just anything in
/usr) - Any other file that's #included must be used, by referencing a type, function or global defined directly in that file.
Note that this prohibits "mega-h-files", which are .h files that just include other .h files in order to expose them.
Our intention with this plugin is to enforce #including only directly what you need. It makes dependencies easier to understand. Also, the compiler doesn't need to work on all those unused files - in one case we saved >30% compilation time in a large project by just cleaning up our #includes.
Files from /usr are always allowed because they don't follow any obvious pattern of who is responsible for
what. <errno.h> doesn't actually define errno - it does a bunch of nested includes which eventually bring in the
global variable.
Consider the following code:
struct Context {
char scratch_buffer[80*1024];
int field;
};
void init_field(struct Context *ctx, int x) {
*ctx = (struct Context) { /* 80 kb copy hiding here */
.field = x,
};
}
The large_assignment plugin helps to find assignments which look innocent but may incur a big copy. The size is configurable via a clang plugin command line argument.
With the plugin, the above example yields:
clang -Wall -Wextra -g -Xclang -load -Xclang ./clang_plugins.so \
-Xclang -add-plugin -Xclang large_assignment \
-Xclang -plugin-arg-large_assignment -Xclang 1024 \
-c /tmp/large.c -o /tmp/large.o
/tmp/large.c:7:16: warning: large assignment of 81924 bytes is more than allowed 1024 bytes
*ctx = (struct Context) {
^
1 warning generated.
To change the maximum allowed, use the command line arguments, e.g.
-Xclang -plugin-arg-large_assignment -Xclang 1024
The private plugin enforces a coding convention, where struct fields may not be accessed directly unless they are
exposed in a public .h file.
The plugin assumes that private .h files are suffixed with _private.h
The rule is: module.c is allowed to access privates from module_private.h, but not from any other private file
(e.g. stuff from other_private.h would not be allowed).
Example project:
foo.h:
#include "foo_private.h"
typedef struct _Foo Foo;
/* Public API, defined in foo.c, supposedly (implementation doesn't matter for the example) */
void foo_init(Foo *);
void foo_set_value(Foo *, int);
foo_private.h:
struct _Foo {
int field; /* supposed to be private, accessible via API function */
};
usage.c:
#include "foo.h"
void bar(Foo *foo) {
foo->field = 3; /* BAD! should use API foo_set_value() */
}
The plugin complains about this API circumvention:
clang -Wall -Wextra -g -Xclang -load -Xclang ./clang_plugins.so -Xclang -add-plugin -Xclang private -c /tmp/usage.c -o /tmp/usage.o
/tmp/usage.c:5:14: warning: access to private member of _Foo
foo->field = 3; /* BAD! should use API foo_set_value() */
^
/tmp/foo_private.h:1:12: note: declaration of _Foo
struct _Foo {
^
1 warning generated.
If a struct's internals are supposed to be exposed, you just need to define it in the non-private .h file.
You sometimes want to allow accessing a private field. Common examples are tests, or when a module is deliberately split
into several .c files all sharing the same _private.h. For those cases there is FRIEND_OF, a macro in private_api.h:
/* FRIEND_OF takes a single *type name* as an argument */
FRIEND_OF(Foo); /* Put this a file level. Now we are allowed in this .c file to access internals of Foo. */
The idea is not to abuse FRIEND_OF too much.