11 KiB
Why Include What You Use Is Difficult
This section is informational, for folks who are wondering why include-what-you-use requires so much code and yet still has so many errors.
Include-what-you-use has the most problems with templates and macros. If your code doesn't use either, IWYU will probably do great. And, you're probably not actually programming in C++...
Use Versus Forward Declare
Include-what-you-use has to be able to tell when a symbol is being used in a way that you can forward-declare it. Otherwise, if you wrote
vector<MyClass*> foo;
IWYU would tell you to #include "myclass.h", when perhaps the whole reason you're using a pointer here is to avoid the need for that #include.
In the above case, it's pretty easy for IWYU to tell that we can safely forward-declare MyClass. But now consider
vector<MyClass> foo; // requires full definition of MyClass
scoped_ptr<MyClass> foo; // forward-declaring MyClass is often ok
To distinguish these, clang has to instantiate the vector and scoped_ptr template classes, including analyzing all member variables and the bodies of the constructor and destructor (and recursively for superclasses).
But that's not enough: when instantiating the templates, we need to keep track of which symbols come from template arguments and which don't. For instance, suppose you call MyFunc<MyClass>(), where MyFunc looks like this:
template<typename T> void MyFunc() {
T* t;
MyClass myclass;
...
}
In this case, the caller of MyFunc is not using the full type of MyClass, because the template parameter is only used as a pointer. On the other hand, the file that defines MyFunc is using the full type information for MyClass. The end result is that the caller can forward-declare MyClass, but the file defining MyFunc has to #include "myclass.h".
Handling Template Arguments
Even figuring out what types are 'used' with a template can be difficult. Consider the following two declarations:
vector<MyClass> v;
hash_set<MyClass> h;
These both have default template arguments, so are parsed like
vector<MyClass, alloc<MyClass> > v;
hash_set<MyClass, hash<MyClass>, equal_to<MyClass>, alloc<MyClass> > h;
What symbols should we say are used? If we say alloc<MyClass> is used when you declare a vector, then every file that #includes <vector> will also need to #include <memory>.
So it's tempting to just ignore default template arguments. But that's not right either. What if hash<MyClass> is defined in some local myhash.h file (as hash<string> often is)? Then we want to make sure IWYU says to #include "myhash.h" when you create the hash_set (otherwise the code won't compile). That requires paying attention to the default template argument. Figuring out how to handle default template arguments can get very complex.
Even normal template arguments can be confusing. Consider this templated function:
template<typename A, typename B, typename C> void MyFunc(A (*fn)(B,C))
{ ... }
and you call MyFunc(FunctionReturningAFunctionPointer()). What types are being used where, in this case?
Who is Responsible for Dependent Template Types?
If you say vector<MyClass> v;, it's clear that you, and not vector.h are responsible for the use of MyClass, even though all the functions that use MyClass are defined in vector.h. (OK, technically, these functions are not "defined" in a particular location, they're instantiated from template methods written in vector.h, but for us it works out the same.)
When you say hash_map<MyClass, int> h;, you are likewise responsible for MyClass (and int), but are you responsible for pair<MyClass, int>? That is the type that hash_map uses to store your entries internally, and it depends on one of your template arguments, but even so it shouldn't be your responsibility -- it's an implementation detail of hash_map. Of course, if you say hash_map<pair<int, int>, int>, then you are responsible for the use of pair. Distinguishing these two cases from each other, and from the vector case, can be difficult.
Now suppose there's a template function like this:
template<typename T> void MyFunc(T t) {
strcat(t, 'a');
strchr(t, 'a');
cerr << t;
}
If you call MyFunc(some_char_star), which of these symbols are you responsible for, and which is the author of MyFunc responsible for: strcat, strchr, operator<<(ostream&, T)?
strcat is a normal function, and the author of MyFunc is responsible for its use. This is an easy case.
In C++, strchr is a templatized function (different impls for char* and const char*). Which version is called depends on the template argument. So, naively, we'd conclude that the caller is responsible for the use of strchr. However, that's ridiculous; we don't want caller of MyFunc to have to #include <string.h> just to call MyFunc. We have special code that (usually) handles this kind of case.
operator<< is also a templated function, but it's one that may be defined in lots of different files. It would be ridiculous in its own way if MyFunc was responsible for including every file that defines operator<<(ostream&, T) for all T. So, unlike the two cases above, the caller is the one responsible for the use of operator<<, and will have to #include the file that defines it. It's counter-intuitive, perhaps, but the alternatives are all worse.
As you can imagine, distinguishing all these cases is extremely difficult. To get it exactly right would require re-implementing C++'s (byzantine) lookup rules, which we have not yet tackled.
Template Template Types
Let's say you have a function
template<template<typename U> T> void MyFunc() {
T<string> t;
}
And you call MyFunc<hash_set>. Who is responsible for the 'use' of hash<string>, and thus needs to #include "myhash.h"? I think it has to be the caller, even if the caller never uses the string type in its file at all. This is rather counter-intuitive. Luckily, it's also rather rare.
Typedefs
Suppose you #include a file "foo.h" that has typedef hash_map<Foo, Bar> MyMap;. And you have this code:
for (MyMap::iterator it = ...)
Who, if anyone, is using the symbol hash_map<Foo, Bar>::iterator? If we say you, as the author of the for-loop, are the user, then you must #include <hash_map>, which undoubtedly goes against the goal of the typedef (you shouldn't even have to know you're using a hash_map). So we want to say the author of the typedef is responsible for the use. But how could the author of the typedef know that you were going to use MyMap::iterator? It can't predict that. That means it has to be responsible for every possible use of the typedef type. This can be complicated to figure out. It requires instantiating all methods of the underlying type, some of which might not even be legal C++ (if, say, the class uses SFINAE).
Worse, when the language auto-derives template types, it loses typedef information. Suppose you wrote this:
MyMap m;
find(m.begin(), m.end(), some_foo);
The compiler sees this as syntactic sugar for find<hash_map<Foo, Bar, hash<Foo>, equal_to<Foo>, alloc<Foo> >(m.begin(), m.end(), some_foo);
Not only is the template argument hash_map instead of MyMap, it includes all the default template arguments, with no indication they're default arguments. All the tricks we used above to intelligently ignore default template arguments are worthless here. We have to jump through lots of hoops so this code doesn't require you to #include not only <hash_map>, but <alloc> and <utility> as well.
Macros
It's no surprise macros cause a huge problem for include-what-you-use. Basically, all the problems of templates also apply to macros, but worse: with templates you can analyze the uninstantiated template, but with macros, you can't analyze the uninstantiated macro -- it likely doesn't even parse cleanly in isolation. As a result, we have very few tools to distinguish when the author of a macro is responsible for a symbol used in a macro, and when the caller of the macro is responsible.
Includes with Side Effects
While not a major problem, this indicates the myriad "gotchas" that exist around include-what-you-use: removing an #include and replacing it with a forward-declare may be dangerous even if no symbols are fully used from the #include. Consider the following code:
foo.h:
namespace ns { class Foo {}; }
using ns::Foo;
foo.cc:
#include "foo.h"
Foo* foo;`
If IWYU just blindly replaces the #include with a forward declare such as namespace ns { class Foo; }, the code will break because of the lost using declaration. Include-what-you-use has to watch out for this case.
Another case is a header file like this:
foo.h:
#define MODULE_NAME MyModule
#include "module_writer.h"
We might think we can remove an #include of foo.h and replace it by #include "module_writer.h", but that is likely to break the build if module_writer.h requires MODULE_NAME be defined. Since my file doesn't participate in this dependency at all, it won't even notice it. IWYU needs to keep track of dependencies between files it's not even trying to analyze!
Private Includes
Suppose you write vector<int> v;. You are using vector, and thus have to #include <vector>. Even this seemingly easy case is difficult, because vector isn't actually defined in <vector>; it's defined in <bits/stl_vector.h>. The C++ standard library has hundreds of private files that users are not supposed to #include directly. Third party libraries have hundreds more. There's no general way to distinguish private from public headers; we have to manually construct the proper mapping.
In the future, we hope to provide a way for users to annotate if a file is public or private, either a comment or a #pragma. For now, we hard-code it in the IWYU tool.
The mappings themselves can be ambiguous. For instance, NULL is provided by many files, including stddef.h, stdlib.h, and more. If you use NULL, what header file should IWYU suggest? We have rules to try to minimize the number of #includes you have to add; it can get rather involved.
Unparsed Code
Conditional #includes are a problem for IWYU when the condition is false:
#if defined(LOG_VERBOSE)
#include <verbose_logger.h>
#endif
...
void StartProcess() {
#if defined(LOG_VERBOSE)
LogVerbose("Starting process");
#endif
...
}
If you're running IWYU without that preprocessor definition set, it has no way of telling if verbose_logger.h is a necessary #include or not.
Placing New Includes and Forward-Declares
Figuring out where to insert new #includes and forward-declares is a complex problem of its own (one that is the responsibility of fix_includes.py). In general, we want to put new #includes with existing #includes. But the existing #includes may be broken up into sections, either because of conditional #includes (with #ifdefs), or macros (such as #define __GNU_SOURCE), or for other reasons. Some forward-declares may need to come early in the file, and some may prefer to come later (after we're in an appropriate namespace, for instance).
fix_includes.py tries its best to give pleasant-looking output, while being conservative about putting code in a place where it might not compile. It uses heuristics to do this, which are not yet perfect.