| 1 |
0) Summary |
| 2 |
|
| 3 |
New code in aptitude should be written, wherever practical, with |
| 4 |
fully abstract interfaces in .h files and hidden implementations in |
| 5 |
.cc files. |
| 6 |
|
| 7 |
1) Details |
| 8 |
|
| 9 |
A standard abstract interface in aptitude looks like this: |
| 10 |
|
| 11 |
class foo |
| 12 |
{ |
| 13 |
public: |
| 14 |
foo(); |
| 15 |
virtual ~foo(); |
| 16 |
|
| 17 |
virtual void method1(args...) = 0; |
| 18 |
virtual void method2(args...) = 0; |
| 19 |
// More methods ... |
| 20 |
|
| 21 |
// Static constructor? (see below) |
| 22 |
}; |
| 23 |
|
| 24 |
Note that both foo() and ~foo() are implemented in the .cc file. |
| 25 |
foo() should be public so that any module can implement the foo |
| 26 |
interface; this allows alternate implementations to be substituted |
| 27 |
during testing. |
| 28 |
|
| 29 |
Interfaces should contain only virtual methods, to aid multiple |
| 30 |
inheritance, and they should generally derive only from other pure |
| 31 |
virtual classes. |
| 32 |
|
| 33 |
It might be tempting to derive from sigc::trackable, so signals can |
| 34 |
be connected directly to interface methods. However, it is generally |
| 35 |
better to either have the object connect itself (so the implementation |
| 36 |
inherits trackable directly), or to connect the signal to a concrete |
| 37 |
object that holds a strong reference to the interface. If you do |
| 38 |
derive from trackable, please use virtual inheritance to avoid |
| 39 |
unnecesary extra copies of "trackable" if an object implements several |
| 40 |
interfaces. |
| 41 |
|
| 42 |
1.1) Constructors |
| 43 |
|
| 44 |
An abstract interface obviously can't be constructed. |
| 45 |
Implementation classes should be hidden in .cc files and constructed |
| 46 |
via either static methods on the interface, or free functions. Which |
| 47 |
one is used depends on the relationship of the interface to the |
| 48 |
implementation, as described below. |
| 49 |
|
| 50 |
1.1.1) Single-implementation interfaces |
| 51 |
|
| 52 |
Some interfaces in aptitude are only meant to have a single |
| 53 |
implementation (except possibly for an implementation used just for |
| 54 |
testing). These interfaces will expose their own creation via a |
| 55 |
::create() static member. If there are several overloads of the |
| 56 |
object's constructor, you should use different names for each static |
| 57 |
creation routine: create_empty(), create_full(), etc. |
| 58 |
|
| 59 |
The create() routine will construct an instance of a class that is |
| 60 |
only declared in the corresponding .cc file. |
| 61 |
|
| 62 |
For instance: |
| 63 |
|
| 64 |
// In foo.h: |
| 65 |
|
| 66 |
class foo |
| 67 |
{ |
| 68 |
// ... as above ... |
| 69 |
|
| 70 |
static boost::shared_ptr<foo> create(); |
| 71 |
}; |
| 72 |
|
| 73 |
// In foo.cc: |
| 74 |
|
| 75 |
class foo_impl : public foo |
| 76 |
{ |
| 77 |
public: |
| 78 |
// Declare implementations of private methods in foo: |
| 79 |
void method1(args...); |
| 80 |
// ... |
| 81 |
}; |
| 82 |
|
| 83 |
void foo_impl::method1(args...) { /* ... */ } |
| 84 |
|
| 85 |
boost::shared_ptr<foo> create() |
| 86 |
{ |
| 87 |
return make_shared<foo>(); |
| 88 |
} |
| 89 |
|
| 90 |
1.1.2) Multi-implementation interfaces |
| 91 |
|
| 92 |
Interfaces which are meant to have several implementations are |
| 93 |
defined in their own .h files, but without any constructor methods. |
| 94 |
Instead, they are constructed via free functions. For instance: |
| 95 |
|
| 96 |
// In foo.h: |
| 97 |
|
| 98 |
class foo |
| 99 |
{ |
| 100 |
// ... as above ... |
| 101 |
}; |
| 102 |
|
| 103 |
// In bar.h: |
| 104 |
|
| 105 |
boost::shared_ptr<foo> create_bar(); |
| 106 |
|
| 107 |
// In bar.cc: |
| 108 |
|
| 109 |
class bar : public foo |
| 110 |
{ |
| 111 |
// Declare implementations of methods in foo. |
| 112 |
}; |
| 113 |
|
| 114 |
// Define implementations of methods in foo. |
| 115 |
|
| 116 |
boost::shared_ptr<foo> create_bar() |
| 117 |
{ |
| 118 |
return make_shared<bar>(); |
| 119 |
} |
| 120 |
|
| 121 |
1.2) Signals and slots |
| 122 |
|
| 123 |
Instead of directly exposing signal members, signals should be |
| 124 |
exposed via appropriately-named routines. This provides better |
| 125 |
encapsulation (directly exposing sigc signals allows external code to |
| 126 |
invoke them) and makes it easier to test code that uses the |
| 127 |
interfaces. |
| 128 |
|
| 129 |
class foo |
| 130 |
{ |
| 131 |
/** \brief Register a callback to be invoked when the hidden |
| 132 |
* signal is invoked. |
| 133 |
*/ |
| 134 |
virtual sigc::connection connect_hidden(const sigc::slot<void> &slot) = 0; |
| 135 |
}; |
| 136 |
|
| 137 |
Code in the src/qt tree can use the corresponding Qt idioms instead. |
| 138 |
|
| 139 |
1.3) Mocks |
| 140 |
|
| 141 |
By convention, mock implementations of interfaces should be placed |
| 142 |
in a subdirectory mocks/ under the directory containing the .h of the |
| 143 |
interface. The mock class has the same name as the interface, but |
| 144 |
exists in a "mocks" namespace. |
| 145 |
|
| 146 |
// In foo.h: |
| 147 |
|
| 148 |
namespace aptitude |
| 149 |
{ |
| 150 |
class foo |
| 151 |
{ |
| 152 |
// Declarations of interface methods on foo. |
| 153 |
}; |
| 154 |
} |
| 155 |
|
| 156 |
// In mocks/foo.h: |
| 157 |
|
| 158 |
namespace aptitude |
| 159 |
{ |
| 160 |
namespace mocks |
| 161 |
{ |
| 162 |
class foo : public aptitude::foo |
| 163 |
{ |
| 164 |
// Mock out the methods of foo. |
| 165 |
}; |
| 166 |
} |
| 167 |
} |
| 168 |
|
| 169 |
1.5) Global variables vs singleton classes vs static methods |
| 170 |
|
| 171 |
Some functionality is truly global, and could be exposed via global |
| 172 |
variables of class type, singleton classes with a hidden instance |
| 173 |
variable, or static methods on a class. |
| 174 |
|
| 175 |
Between these three approaches, singleton classes are preferred. |
| 176 |
Using a singleton class behind an interface means that the |
| 177 |
functionality can be swapped out, in test code if nowhere else. To |
| 178 |
support this, it should be possible to pass in the interface as a |
| 179 |
constructor parameter (although a convenience constructor that uses |
| 180 |
the most common implementation of that interface is a good idea). |
| 181 |
|
| 182 |
This is important because these modules represent access to system |
| 183 |
functionality that might be awkward to use in test code. Existing |
| 184 |
aptitude code violates this principle, with the result that it is |
| 185 |
difficult to verify modules that use, for instance, the global |
| 186 |
download cache or file queue. |
| 187 |
|
| 188 |
1.6) Dependencies: constructor parameters vs created. |
| 189 |
|
| 190 |
If an object accesses another object through an interface, it can |
| 191 |
either take it as a constructor parameter or create it itself. Where |
| 192 |
practical, it is preferred to pass it as a constructor parameter, to |
| 193 |
make it possible to swap in different implementations of the |
| 194 |
interface. However, it might be a good idea to provide a convenience |
| 195 |
constructor that instantiates the most common implementations of the |
| 196 |
interfaces that the class uses. |
| 197 |
|
| 198 |
2) Motivation |
| 199 |
|
| 200 |
Obviously, it is not necessary to use this many virtual interfaces. |
| 201 |
There are a number of reasons that this approach is preferred: |
| 202 |
|
| 203 |
2.1) Increasing testability |
| 204 |
|
| 205 |
Unit tests are good. They let you verify your code up-front, and |
| 206 |
they let you catch stealth changes to behavior early, before they |
| 207 |
turn into bug reports. It's also difficult to write good unit |
| 208 |
tests for complex pieces of code with many dependencies; you end |
| 209 |
up testing the dependencies instead of the code you care about. |
| 210 |
|
| 211 |
Breaking code into logically separated modules and dropping |
| 212 |
abstraction layers between them makes it possible to separate |
| 213 |
modules from the other code that they interact with. It also |
| 214 |
makes it easy to test corner cases, particularly error |
| 215 |
conditions, without having to generate synthetic system states |
| 216 |
(for instance, what happens if the disk is full?), and it |
| 217 |
provides a way of testing interactive code without requiring user |
| 218 |
interaction (real or simulated). |
| 219 |
|
| 220 |
2.2) Clear statements of internal interfaces |
| 221 |
|
| 222 |
In a header file full of class definitions, with instance |
| 223 |
variables, private methods, private member classes, and so on, it |
| 224 |
can be tough to spot the actual interface methods. The coding |
| 225 |
style described above places the public interface and (almost) |
| 226 |
nothing else into the header file of a class. |
| 227 |
|
| 228 |
I believe that this also leads to simpler interfaces, as it's |
| 229 |
harder to hide poorly designed interface methods when they aren't |
| 230 |
mixed with implementation concerns. |
| 231 |
|
| 232 |
2.3) Fewer inter-header dependencies |
| 233 |
|
| 234 |
Removing implementations from header files means that you don't |
| 235 |
need to include the headers that are used by the implementation |
| 236 |
from the header file. This can speed up compile times of client |
| 237 |
code, in some cases significantly (for instance, when the |
| 238 |
implementation requires some templates to be instantiated). |
| 239 |
Furthermore, it means that modules using the interface don't have |
| 240 |
to be recompiled just because the implementation changed, which |
| 241 |
can speed up compiles significantly (old aptitude code was very |
| 242 |
incestuous this way). |
| 243 |
|
| 244 |
In extreme cases, this can break circular dependencies between |
| 245 |
headers, although those tend to be signs that something e,se is |
| 246 |
wrong anyway. |