...one of the most highly
regarded and expertly designed C++ library projects in the
world.
— Herb Sutter and Andrei
Alexandrescu, C++
Coding Standards
This section can be consulted selectively for specific topics of interest.
Old values require to copy the expression passed to BOOST_CONTRACT_OLDOF
thus the type of that expression needs to be copyable. More precisely, dereferencing
an old value pointer of type boost::contract::old_ptr
<T>
requires boost::contract::is_old_value_copyable
<T>::value
to be true
,
otherwise this library will generate a compile-time error.
In some cases it might be acceptable, or even desirable, to cause a compile-time
error when a program uses old value types that are not copyable (because
it is not possible to fully check the correctness of the program as stated
by the contract assertions that use these old values). In these cases, programmers
can declare old values using boost::contract::old_ptr
as seen so far.
However, in some other cases it might be desirable to simply skip assertions
that use old values when the respective old value types are not copyable,
without causing compile-time errors. Programmers can do this using boost::contract::old_ptr_if_copyable
instead of boost::contract::old_ptr
and checking if the old value pointer is not null before dereferencing it
in postconditions. For example, consider the following function template
that could in general be instantiated for types T
that are not copy constructible (that is for which boost::contract::is_old_value_copyable
<T>::value
is false
,
see old_if_copyable.cpp
):
[68]
template<typename T> // T might or might not be copyable. void offset(T& x, int count) { // No compiler error if T has no copy constructor... boost::contract::old_ptr_if_copyable<T> old_x = BOOST_CONTRACT_OLDOF(x); boost::contract::check c = boost::contract::function() .postcondition([&] { // ...but old value null if T has no copy constructor. if(old_x) BOOST_CONTRACT_ASSERT(x == *old_x + count); }) ; x += count; }
The old value pointer old_x
is programmed using boost::contract::old_ptr_if_copyable
.
When T
is copyable, boost::contract::old_ptr_if_copyable
<T>
behaves like boost::contract::old_ptr
<T>
.
When T
is not copyable instead,
boost::contract::old_ptr_if_copyable
<T>
will simply not copy x
at
run-time and leave old_x
initialized to a null pointer. Therefore, boost::contract::old_ptr_if_copyable
objects like old_x
must be
checked to be not null as in if(old_x) ...
before
they are dereferenced in postconditions and exception guarantees (to avoid
run-time errors of dereferencing null pointers).
If the above example used boost::contract::old_ptr
instead then this library would have generated a compile-time error when
offset
is instantiated with
types T
that are not copy
constructible (but only if old_x
is actually dereferenced somewhere, for example in the contract assertions
using *old_x
...
or old_x->...
). [69]
When C++11 auto
declarations
are used with BOOST_CONTRACT_OLDOF
,
this library always defaults to using the boost::contract::old_ptr
type (because its type requirements are more stringent than boost::contract::old_ptr_if_copyable
).
For example, the following statements are equivalent:
auto old_x = BOOST_CONTRACT_OLDOF(x); // C++11 auto declarations always use `old_ptr` (never `old_ptr_if_copyable`). boost::contract::old_ptr<decltype(x)> old_x = BOOST_CONTRACT_OLDOF(x);
If programmers want to relax the copyable type requirement, they must do
so explicitly by using the boost::contract::old_ptr_if_copyable
type instead of using auto
declarations.
This library uses boost::contract::is_old_value_copyable
to determine if an old value type is copyable or not, and then boost::contract::old_value_copy
to actually copy the old value.
By default, boost::contract::is_old_value_copyable
<T>
is equivalent to boost::is_copy_constructible<T>
and
boost::contract::old_value_copy
<T>
is implemented using T
's
copy constructor. However, these type traits can be specialized by programmers
for example to avoid making old value copies of types even when they have
a copy constructor (maybe because these copy constructors are too expensive),
or to make old value copies for types that do not have a copy constructor,
or for any other specific need programmers might have for the types in question.
For example, the following specialization of boost::contract::is_old_value_copyable
intentionally avoids making old value copies for all expressions of type
w
even if that type has a
copy constructor (see old_if_copyable.cpp
):
// Copyable type but... class w { public: w(w const&) { /* Some very expensive copy operation here... */ } /* ... */
// ...never copy old values for type `w` (because its copy is too expensive). namespace boost { namespace contract { template<> struct is_old_value_copyable<w> : boost::false_type {}; } }
On the flip side, the following specializations of boost::contract::is_old_value_copyable
and boost::contract::old_value_copy
make old value copies of expressions of type p
even if that type does not actually have a copy constructor (see old_if_copyable.cpp
):
// Non-copyable type but... class p : private boost::noncopyable { int* num_; friend struct boost::contract::old_value_copy<p>; /* ... */
// ...still copy old values for type `p` (using a deep copy). namespace boost { namespace contract { template<> struct old_value_copy<p> { explicit old_value_copy(p const& old) { *old_.num_ = *old.num_; // Deep copy pointed value. } p const& old() const { return old_; } private: p old_; }; template<> struct is_old_value_copyable<p> : boost::true_type {}; } }
In general, boost::is_copy_constructible
and therefore boost::contract::is_old_value_copyable
require C++11 decltype
and SFINAE
to automatically detect if a given type is copyable or not. On non-C++11
compilers, it is possible to inherit the old value type from boost::noncopyable
, or use BOOST_MOVABLE_BUT_NOT_COPYABLE
,
or specialize boost::is_copy_constructible
(see boost::is_copy_constructible
documentation
for more information), or just specialize boost::contract::is_old_value_copyable
.
For example, for a non-copyable type n
the following code will work also on non-C++11 compilers (see old_if_copyable.cpp
):
class n { // Do not want to use boost::noncopyable but... int num_; private: n(n const&); // ...unimplemented private copy constructor (so non-copyable). /* ... */
// Specialize `boost::is_copy_constructible` (no need for this on C++11). namespace boost { namespace contract { template<> struct is_old_value_copyable<n> : boost::false_type {}; } }
In general, assertions can introduce a new set of requirements on the types used by the program. Some of these type requirements might be necessary only to check the assertions and they would not be required by the program otherwise.
In some cases it might be acceptable, or even desirable, to cause a compile-time error when a program uses types that do not provide all the operations needed to check contract assertions (because it is not possible to fully check the correctness of the program as specified by its contracts). In these cases, programmers can specify contract assertions as we have seen so far, compilation will fail if user types do not provide all operations necessary to check the contracts.
However, in some other cases it might be desirable to not augment the type
requirements of a program just because of contract assertions and to simply
skip assertions when user types do not provide all the operations necessary
to check them, without causing compile-time errors. Programmers can do this
using if constexpr
on C++17 compilers, and boost::contract::condition_if
(or boost::contract::condition_if_c
)
on non-C++17 compilers.
For example, let's consider the following vector<T>
class template equivalent to std::vector<T>
.
C++ std::vector<T>
does
not require that its value type parameter T
has an equality operator ==
(it only requires T
to be
copy constructible, see std::vector
documentation). However, the contracts for the vector<T>::push_back(value)
public function include a postcondition back() == value
that introduces the new requirement that T
must also have an equality operator ==
.
Programmers can specify this postcondition as usual with BOOST_CONTRACT_ASSERT(back()
== value)
an let the program fail to compile when
users instantiate vector<T>
with a type T
that does not
provide an equality operator ==
.
Otherwise, programmers can guard this postcondition using C++17 if constexpr
to evaluate the asserted condition only for types T
that have an equality operator ==
and skip it otherwise. [70] For example:
template<typename T> class vector { public: void push_back(T const& value) { boost::contract::check c = boot::contract::public_function(this) .postcondition([&] { // Guard with `if constexpr` for T without `==`. if constexpr(boost::has_equal_to<T>::value) BOOST_CONTRACT_ASSERT(back() == value); }) ; vect_.push_back(value); } /* ... */
More in general, if constexpr
can be used to guard a mix of both old value copies and contract assertions
that introduce specific extra type requirements. For example, the following
swap
function can be called
on any type T
that is movable
but its old value copies and postcondition assertions are evaluated only
for types T
that are also
copyable and have an equality operator ==
(see if_constexpr.cpp
):
template<typename T> void swap(T& x, T& y) { constexpr bool b = boost::contract::is_old_value_copyable<T>::value && boost::has_equal_to<T>::value; boost::contract::old_ptr<T> old_x, old_y; if constexpr(b) { // Contract requires copyable T... old_x = BOOST_CONTRACT_OLDOF(x); old_y = BOOST_CONTRACT_OLDOF(y); } boost::contract::check c = boost::contract::function() .postcondition([&] { if constexpr(b) { // ... and T with `==`... BOOST_CONTRACT_ASSERT(x == *old_y); BOOST_CONTRACT_ASSERT(y == *old_x); } }) ; T t = std::move(x); // ...but body only requires movable T. x = std::move(y); y = std::move(t); }
if constexpr
(no C++17)
On non-C++17 compilers where if constexpr
is not available, it is possible
to use boost::contract::condition_if
to skip assertions based on type requirements (even if this code is less
readable and more verbose than using if
constexpr
). For example (see condition_if.cpp
):
template<typename T> class vector { public: void push_back(T const& value) { boost::contract::check c = boost::contract::public_function(this) .postcondition([&] { // Instead of `ASSERT(back() == value)` for T without `==`. BOOST_CONTRACT_ASSERT( boost::contract::condition_if<boost::has_equal_to<T> >( boost::bind(std::equal_to<T>(), boost::cref(back()), boost::cref(value) ) ) ); }) ; vect_.push_back(value); } /* ... */
More in general, boost::contract::condition_if
is used as follow:
boost::contract::condition_if<Pred>( cond )
Where Pred
is a nullary boolean
meta-function and cond
is
a nullary boolean functor. If Pred::value
is statically evaluated to be true
at compile-time then cond()
is called at run-time and its boolean result
is returned by the enclosing boost::contract::condition_if
call. Otherwise, if Pred::value
is statically evaluated to be false
at compile-time then boost::contract::condition_if
trivially returns true
. Therefore,
if cond
is a functor template
instantiation (not just a functor) then its call that contains the assertion
operations with the extra type requirements (e.g., the equality operator
==
) will not be actually instantiated
and compiled unless the compiler determines at compile-time that Pred::value
is true
(when used this way, functor templates like std::equal_to
and C++14 generic lambdas can be used to program cond
,
but C++11 lambdas cannot).
The boost::contract::condition_if
function template is a special case of the more general boost::contract::call_if
,
specifically boost::contract::condition_if
<Pred>(cond)
is logically equivalent to boost::contract::call_if
<Pred>(cond).else_([] { return
true; })
. [71] The boost::contract::call_if
function template also accepts a number of optional else-if
statements and one optional else statement:
boost::contract::call_if<Pred1>( t1 ).template else_if<Pred2>( // Optional. t2 ) ... // Optionally, other `else_if` statements. .else_( // Optional for `void` functors, otherwise required. e )
Where Pred1
, Pred2
, ... are nullary boolean meta-functions
and t1
, t2
,
..., e
are nullary functors.
The return types of the functor calls t1()
, t2()
, ..., e()
must either be all the same (including
possibly all void
) or be of
types implicitly convertible into one another. At run-time boost::contract::call_if
will
call the functor t1()
,
or t2()
,
..., or e()
depending on which meta-function Pred1::value
,
Pred2::value
, ... is statically evaluated to be
true
or false
at compile-time, and it will return the value returned by the functor being
called. If t1
, t2
, ..., e
are functor template instantiations (not just functors) then their code will
only be compiled if the compiler determines they need to be actually called
at run-time (so only if the related Pred1::value
,
Pred2::value
, ... are respectively evaluated to
be true
or false
at compile-time). All the else_if<...>(...)
statements are optional.
The else_(...)
statement is optional if the functor calls return void
,
otherwise it is required.
In general, boost::contract::call_if
can be used to program contract assertions that compile and check different
functor templates depending on related predicates being statically evaluated
to be true
or false
at compile-time (but in most cases
boost::contract::condition_if
should be sufficient and less verbose to use). The boost::contract::condition_if_c
,
boost::contract::call_if_c
,
and .else_if_c
function templates work similarly to their counterparts without the ..._c
postfix
seen so far, but they take their predicate template parameters as static
boolean values instead of nullary boolean meta-functions.
On compilers that support C++17 if
constexpr
there should be no need
to use boost::contract::condition_if
or boost::contract::call_if
because if constexpr
can be used instead (making the code more readable and easier to program).
[72]
This library allows to specify a different set of class invariants to check
for volatile public functions. These volatile class invariants
are programmed in a public const volatile
function, named invariant
,
taking no argument, and returning void
(see BOOST_CONTRACT_INVARIANT_FUNC
to name the invariant function differently from invariant
and Access Specifiers
to not have to declare it public
).
Classes that do no have invariants for their volatile public functions, simply
do not declare the void invariant() const volatile
function.
In general, const volatile
invariants work the same as const
invariants (see Class
Invariants) with the only difference that volatile
and const volatile
functions check const volatile
invariants while non-const
(i.e.,
neither const
nor volatile
) and const
functions check const
invariants.
A given class can specify any combination of static
,
const volatile
,
and const
invariant functions
(see Class Invariants):
[73]
static
invariants at entry and exit (even if an exception is thrown), plus
const volatile
and const
invariants in
that order at exit but only if no exception is thrown.
static
invariants at entry and exit (even if an exception is thrown), plus
const volatile
and const
invariants in
that order at entry (and at exit but only if an exception is thrown,
even is destructors should in general never throw in C++).
const
and const
public functions check static
and const
invariants at entry and at exit (even if an exception is thrown).
volatile
and const volatile
public functions check static
and const volatile
invariants at entry and at exit (even if an exception is thrown).
These rules ensure that volatile class invariants are correctly checked (see
Constructor
Calls, Destructor
Calls, and Public
Function Calls). For example (see volatile.cpp
):
class u { public: static void static_invariant(); // Static invariants. void invariant() const volatile; // Volatile invariants. void invariant() const; // Const invariants. u() { // Check static, volatile, and const invariants. boost::contract::check c= boost::contract::constructor(this); } ~u() { // Check static, volatile, and const invariants. boost::contract::check c = boost::contract::destructor(this); } void nc() { // Check static and const invariants. boost::contract::check c = boost::contract::public_function(this); } void c() const { // Check static and const invariants. boost::contract::check c = boost::contract::public_function(this); } void v() volatile { // Check static and volatile invariants. boost::contract::check c = boost::contract::public_function(this); } void cv() const volatile { // Check static and volatile invariants. boost::contract::check c = boost::contract::public_function(this); } static void s() { // Check static invariants only. boost::contract::check c = boost::contract::public_function<u>(); } };
This library does not automatically check const
volatile
invariants for non-volatile
functions. However, if the contract
specifications require it, programmers can explicitly call the const volatile
invariant function from the const
invariant function (preferably in that order to be consistent with the order
const volatile
and const
invariants are checked
for constructors and destructors). That way all public functions, volatile
or not, will check const volatile
invariants (while only const
and non-const
public functions
will check only const
invariants,
correctly so because the volatile
qualifier shall not be stripped away): [74]
class u { public: void invariant() const volatile { ... } // Volatile invariants. void invariant() const { auto const volatile* cv = this; cv->invariant(); // Call `const volatile` invariant function above. ... // Other non-volatile invariants. } ... };
(As usual, private and protected functions do not check any invariant, not
even when they are volatile
or const volatile
,
see Private
and Protected Functions).
As with all public operations of a class, also public move operations should maintain class invariants (see [Stroustrup13], p. 520). Specifically, at a minimum C++ requires the following:
Therefore, both the move constructor and the move assignment operator need
to maintain the class invariants of the moved-from object so their contracts
can be programmed using boost::contract::constructor
and boost::contract::public_function
as usual. For example (see move.cpp
):
class circular_buffer : private boost::contract::constructor_precondition<circular_buffer> { public: void invariant() const { if(!moved()) { // Do not check (some) invariants for moved-from objects. BOOST_CONTRACT_ASSERT(index() < size()); } // More invariants here that hold also for moved-from objects (e.g., // all assertions necessary to successfully destroy moved-from objects). } // Move constructor. circular_buffer(circular_buffer&& other) : boost::contract::constructor_precondition<circular_buffer>([&] { BOOST_CONTRACT_ASSERT(!other.moved()); }) { boost::contract::check c = boost::contract::constructor(this) .postcondition([&] { BOOST_CONTRACT_ASSERT(!moved()); BOOST_CONTRACT_ASSERT(other.moved()); }) ; move(std::forward<circular_buffer>(other)); } // Move assignment. circular_buffer& operator=(circular_buffer&& other) { // Moved-from can be (move) assigned (so no pre `!moved()` here). boost::contract::check c = boost::contract::public_function(this) .precondition([&] { BOOST_CONTRACT_ASSERT(!other.moved()); }) .postcondition([&] { BOOST_CONTRACT_ASSERT(!moved()); BOOST_CONTRACT_ASSERT(other.moved()); }) ; return move(std::forward<circular_buffer>(other)); } ~circular_buffer() { // Moved-from can always be destroyed (in fact no preconditions). boost::contract::check c = boost::contract::destructor(this); } bool moved() const { boost::contract::check c = boost::contract::public_function(this); return moved_; } private: bool moved_; /* ... */
This example assumes that it is possible to call the public function moved()
on the moved-from object. [75]
Note | |
---|---|
The default move constructor and move assignment operator generated by
C++ will not automatically check contracts. Therefore, unless the move
operations are not public or they have no preconditions, no postconditions,
and their class has no invariants, programmers should manually define them
using |
As always, programmers can decide to not program contracts for a given type. Specifically, they might decide to not program contracts for a class that needs to be moved in order to avoid the run-time overhead of checking contract assertions and to maximize performance (see Benefits and Costs).
A C++ union
cannot have virtual
functions, base classes, and cannot be used as a base class thus subcontracting
(boost::contract::virtual_
,
BOOST_CONTRACT_OVERRIDE
,
etc.) do not apply to unions. Also a union
cannot inherit from boost::contract::constructor_precondition
(because it cannot have base classes), instead boost::contract::constructor_precondition
is used to declare a local object that checks constructor preconditions (at
the very beginning of the constructor before old value copies and other contracts,
see declaration of pre
in
the example below). A part from that, this library is used as usual to program
contracts for unions. For example (see union.cpp
):
union positive { public: static void static_invariant() { // Static class invariants (as usual). BOOST_CONTRACT_ASSERT(instances() >= 0); } void invariant() const { // Class invariants (as usual). BOOST_CONTRACT_ASSERT(i_ > 0); BOOST_CONTRACT_ASSERT(d_ > 0); } // Contracts for constructor, as usual but... explicit positive(int x) : d_(0) { // ...unions cannot have bases so constructor preconditions here. boost::contract::constructor_precondition<positive> pre([&] { BOOST_CONTRACT_ASSERT(x > 0); }); boost::contract::old_ptr<int> old_instances = BOOST_CONTRACT_OLDOF(instances()); boost::contract::check c = boost::contract::constructor(this) .postcondition([&] { { int y; get(y); BOOST_CONTRACT_ASSERT(y == x); } BOOST_CONTRACT_ASSERT(instances() == *old_instances + 1); }) ; i_ = x; ++instances_; } // Contracts for destructor (as usual). ~positive() { boost::contract::old_ptr<int> old_instances = BOOST_CONTRACT_OLDOF(instances()); boost::contract::check c = boost::contract::destructor(this) .postcondition([&] { BOOST_CONTRACT_ASSERT(instances() == *old_instances - 1); }) ; --instances_; } // Contracts for public function (as usual, but no virtual or override). void get(int& x) const { boost::contract::check c = boost::contract::public_function(this) .postcondition([&] { BOOST_CONTRACT_ASSERT(x > 0); }) ; x = i_; } // Contracts for static public function (as usual). static int instances() { boost::contract::check c = boost::contract::public_function<positive>(); return instances_; } private: int i_; double d_; /* ... */
This library provides three predefined assertion levels that can be used to selectively disable assertions depending on their computational complexity: [76]
BOOST_CONTRACT_ASSERT
is used to assert conditions that are not computationally expensive,
at least compared to the cost of executing the function body. These assertions
are the ones we have seen so far, they are always checked at run-time
and they cannot be disabled.
BOOST_CONTRACT_ASSERT_AUDIT
is used to assert conditions that are computationally expensive compared
to the cost of executing the function body. These assertions are not
checked at run-time unless programmers explicitly define BOOST_CONTRACT_AUDITS
(undefined by default), but the asserted conditions are always compiled
and therefore validated syntactically (even when they are not actually
evaluated and checked at run-time).
BOOST_CONTRACT_ASSERT_AXIOM
is used to assert conditions that are computationally prohibitive, at
least compared to the cost of executing the function body. These assertions
are never evaluated or checked at run-time, but the asserted conditions
are always compiled and therefore validated syntactically (so these assertions
can serve as formal comments to the code).
In addition, BOOST_CONTRACT_CHECK_AUDIT
and BOOST_CONTRACT_CHECK_AXIOM
are similar to BOOST_CONTRACT_ASSERT_AUDIT
and BOOST_CONTRACT_ASSERT_AXIOM
but they are used to program audit and axiom levels for implementation checks
instead of assertions (see Implementation
Checks).
For example, BOOST_CONTRACT_ASSERT_AUDIT
can be used to program computationally expensive assertions (see assertion_level.cpp
):
template<typename RandomIter, typename T> RandomIter random_binary_search(RandomIter first, RandomIter last, T const& value) { RandomIter result; boost::contract::check c = boost::contract::function() .precondition([&] { BOOST_CONTRACT_ASSERT(first <= last); // Default, not expensive. // Expensive O(n) assertion (use AXIOM if prohibitive instead). BOOST_CONTRACT_ASSERT_AUDIT(std::is_sorted(first, last)); }) .postcondition([&] { if(result != last) BOOST_CONTRACT_ASSERT(*result == value); }) ; /* ... */
Similarly, BOOST_CONTRACT_AUDITS
can be used to disable expensive old value copies and related assertions
that use them (see assertion_level.cpp
):
template<typename T> class vector {
public: void swap(vector& other) { boost::contract::old_ptr<vector> old_me, old_other; #ifdef BOOST_CONTRACT_AUDITS old_me = BOOST_CONTRACT_OLDOF(*this); old_other = BOOST_CONTRACT_OLDOF(other); #endif // Else, skip old value copies... boost::contract::check c = boost::contract::public_function(this) .postcondition([&] { // ...and also skip related assertions. BOOST_CONTRACT_ASSERT_AUDIT(*this == *old_other); BOOST_CONTRACT_ASSERT_AUDIT(other == *old_me); }) ; vect_.swap(other.vect_); }
/* ... */ private: std::vector<T> vect_; };
The condition passed to BOOST_CONTRACT_ASSERT_AXIOM
is compiled but not actually evaluated at run-time so this macro can be used
to program computationally prohibitive assertions but also assertions that
cannot actually be programmed in C++ using functions that are declared but
left undefined. For example, (see assertion_level.cpp
):
// If valid iterator range (cannot implement in C++ but OK to use in AXIOM). template<typename Iter> bool valid(Iter first, Iter last); // Only declared, not actually defined.
template<typename T> class vector {
public: iterator insert(iterator where, T const& value) { iterator result; boost::contract::old_ptr<unsigned> old_capacity = BOOST_CONTRACT_OLDOF(capacity()); boost::contract::check c = boost::contract::public_function(this) .postcondition([&] { BOOST_CONTRACT_ASSERT(capacity() >= *old_capacity); if(capacity() > *old_capacity) { BOOST_CONTRACT_ASSERT_AXIOM(!valid(begin(), end())); } else { BOOST_CONTRACT_ASSERT_AXIOM(!valid(where, end())); } }) ; return result = vect_.insert(where, value); }
/* ... */ private: std::vector<T> vect_; };
In addition to these assertion levels that are predefined by this library,
programmers are free to define their own. For example, the following macro
could be used to program and selectively disable assertions that have exponential
computational complexity O(e^n)
:
#ifdef EXPONENTIALLY_COMPLEX_ASSERTIONS
// Following will compile and also evaluate `cond`.
#define ASSERT_EXP(cond) BOOST_CONTRACT_ASSERT(cond)
#else
// Following will compile but never actually evaluate `cond`.
#define ASSERT_EXP(cond) BOOST_CONTRACT_ASSERT(true || (cond))
#endif
...
ASSERT_EXP(some-exponentially-complex-boolean-condition
);
Checking contracts adds run-time overhead and can slow down program execution
(see Benefits
and Costs). Therefore, programmers can define any combination of the
following macros (-D
option in Clang and GCC, /D
option in MSVC, etc.) to instruct this library to not check specific groups
of contract conditions at run-time:
BOOST_CONTRACT_NO_PRECONDITIONS
to not check preconditions.
BOOST_CONTRACT_NO_POSTCONDITIONS
to not check postconditions.
BOOST_CONTRACT_NO_EXCEPTS
to not check exception guarantees.
BOOST_CONTRACT_NO_ENTRY_INVARIANTS
to not check class invariants at call entry.
BOOST_CONTRACT_NO_EXIT_INVARIANTS
to not check class invariants at call exit.
BOOST_CONTRACT_NO_INVARIANTS
to not check class invariants at both call entry and exit. (This is provided
for convenience, it is equivalent to defining both BOOST_CONTRACT_NO_ENTRY_INVARIANTS
and BOOST_CONTRACT_NO_EXIT_INVARIANTS
.)
BOOST_CONTRACT_NO_CHECKS
to not evaluate implementation checks.
Note | |
---|---|
Old values can be used by both postconditions and exception guarantees
so it is necessary to define both |
By default, none of these macros are defined so this library checks all contracts. When these macros are defined by the user, the implementation code of this library is internally optimized to minimize as much as possible any run-time and compile-time overhead associated with checking and compiling contracts (see Disable Contract Compilation for techniques to completely remove any run-time and compile-time overheads associated with contract code).
For example, programmers could decide to check all contracts during early
development builds, but later check only preconditions and maybe entry invariants
for release builds by defining BOOST_CONTRACT_NO_POSTCONDITIONS
,
BOOST_CONTRACT_NO_EXCEPTS
,
BOOST_CONTRACT_NO_EXIT_INVARIANTS
,
and BOOST_CONTRACT_NO_CHECKS
.
This library provides macros that can be used to completely disable compile-time
and run-time overhead introduced by contracts but at the cost of manually
programming #ifndef BOOST_CONTRACT_NO_...
statements around contract code:
BOOST_CONTRACT_NO_CONSTRUCTORS
when contract checking is disabled for constructors.
BOOST_CONTRACT_NO_DESTRUCTORS
when contract checking is disabled for destructors.
BOOST_CONTRACT_NO_PUBLIC_FUNCTIONS
when contract checking is disabled for public functions.
BOOST_CONTRACT_NO_FUNCTIONS
when contract checking is disabled for (non-public and non-member) functions.
BOOST_CONTRACT_NO_OLDS
when old value copies are disabled.
BOOST_CONTRACT_NO_ALL
when all contracts above and also implementation checks (see BOOST_CONTRACT_NO_CHECKS
)
are disabled.
These macros are not configuration macros and they should not be defined
directly by programmers (otherwise this library will generate compile-time
errors). Instead, these macros are automatically defined by this library
when programmers define BOOST_CONTRACT_NO_PRECONDITIONS
,
BOOST_CONTRACT_NO_POSTCONDITIONS
,
BOOST_CONTRACT_NO_EXCEPTS
,
BOOST_CONTRACT_NO_INVARIANTS
(or BOOST_CONTRACT_NO_ENTRY_INVARIANTS
and BOOST_CONTRACT_NO_EXIT_INVARIANTS
),
and BOOST_CONTRACT_NO_CHECKS
(see Disable
Contract Checking).
Alternatively, this library provides a macro-based interface defined in
boost/contract_macro.hpp
that can also be used to completely disable compile-time and run-time overheads
introduced by contracts but without the burden of manually writing the #ifndef BOOST_CONTRACT_NO_...
statements. For example, the following
code shows how to use both the boost/contract_macro.hpp
macro interface and the #ifndef
BOOST_CONTRACT_NO_...
statements to completely disable compile-time and run-time overheads for
non-member function contracts (see ifdef_macro.cpp
and ifdef.cpp
):
Macro Interface |
|
---|---|
// Use macro interface to completely disable contract code compilation. #include <boost/contract_macro.hpp> int inc(int& x) { int result; BOOST_CONTRACT_OLD_PTR(int)(old_x, x); BOOST_CONTRACT_FUNCTION() BOOST_CONTRACT_PRECONDITION([&] { BOOST_CONTRACT_ASSERT(x < std::numeric_limits<int>::max()); }) BOOST_CONTRACT_POSTCONDITION([&] { BOOST_CONTRACT_ASSERT(x == *old_x + 1); BOOST_CONTRACT_ASSERT(result == *old_x); }) ; return result = x++; }
|
// Use #ifdef to completely disable contract code compilation. #include <boost/contract/core/config.hpp> #ifndef BOOST_CONTRACT_NO_ALL #include <boost/contract.hpp> #endif int inc(int& x) { int result; #ifndef BOOST_CONTRACT_NO_OLDS boost::contract::old_ptr<int> old_x = BOOST_CONTRACT_OLDOF(x); #endif #ifndef BOOST_CONTRACT_NO_FUNCTIONS boost::contract::check c = boost::contract::function() #ifndef BOOST_CONTRACT_NO_PRECONDITIONS .precondition([&] { BOOST_CONTRACT_ASSERT(x < std::numeric_limits<int>::max()); }) #endif #ifndef BOOST_CONTRACT_NO_POSTCONDITIONS .postcondition([&] { BOOST_CONTRACT_ASSERT(x == *old_x + 1); BOOST_CONTRACT_ASSERT(result == *old_x); }) #endif ; #endif return result = x++; }
|
The same can be done to disable contract code complication for private and
protected functions. The BOOST_CONTRACT_OLD_PTR_IF_COPYABLE
macro is provided to handle non-copyable old value types (similar to boost::contract::old_ptr_if_copyable
).
For constructors, destructors, and public functions the boost/contract_macro.hpp
macro interface and the #ifndef
BOOST_CONTRACT_NO_...
statements can be used as follow (see ifdef_macro.cpp
and ifdef.cpp
):
Macro Interface |
|
---|---|
class integers #define BASES public pushable<int> : // Left in code (almost no overhead). private boost::contract::constructor_precondition<integers>, BASES { // Followings left in code (almost no overhead). friend class boost::contract::access; typedef BOOST_CONTRACT_BASE_TYPES(BASES) base_types; #undef BASES BOOST_CONTRACT_INVARIANT({ BOOST_CONTRACT_ASSERT(size() <= capacity()); }) public: integers(int from, int to) : BOOST_CONTRACT_CONSTRUCTOR_PRECONDITION(integers)([&] { BOOST_CONTRACT_ASSERT(from <= to); }), vect_(to - from + 1) { BOOST_CONTRACT_CONSTRUCTOR(this) BOOST_CONTRACT_POSTCONDITION([&] { BOOST_CONTRACT_ASSERT(int(size()) == (to - from + 1)); }) ; for(int x = from; x <= to; ++x) vect_.at(x - from) = x; } virtual ~integers() { BOOST_CONTRACT_DESTRUCTOR(this); // Check invariants. } virtual void push_back( int const& x, boost::contract::virtual_* v = 0 // Left in code (almost no overhead). ) /* override */ { BOOST_CONTRACT_OLD_PTR(unsigned)(old_size); BOOST_CONTRACT_PUBLIC_FUNCTION_OVERRIDE(override_push_back)( v, &integers::push_back, this, x) BOOST_CONTRACT_PRECONDITION([&] { BOOST_CONTRACT_ASSERT(size() < max_size()); }) BOOST_CONTRACT_OLD([&] { old_size = BOOST_CONTRACT_OLDOF(v, size()); }) BOOST_CONTRACT_POSTCONDITION([&] { BOOST_CONTRACT_ASSERT(size() == *old_size + 1); }) BOOST_CONTRACT_EXCEPT([&] { BOOST_CONTRACT_ASSERT(size() == *old_size); }) ; vect_.push_back(x); } private: BOOST_CONTRACT_OVERRIDE(push_back) // Left in code (almost no overhead). /* ... */
|
class integers #define BASES public pushable<int> : #ifndef BOOST_CONTRACT_NO_PRECONDITIONS private boost::contract::constructor_precondition<integers>, BASES #else BASES #endif { #ifndef BOOST_CONTRACT_NO_ALL friend class boost::contract::access; #endif #ifndef BOOST_CONTRACT_NO_PUBLIC_FUNCTIONS typedef BOOST_CONTRACT_BASE_TYPES(BASES) base_types; #endif #undef BASES #ifndef BOOST_CONTRACT_NO_INVARIANTS void invariant() const { BOOST_CONTRACT_ASSERT(size() <= capacity()); } #endif public: integers(int from, int to) : #ifndef BOOST_CONTRACT_NO_PRECONDITIONS boost::contract::constructor_precondition<integers>([&] { BOOST_CONTRACT_ASSERT(from <= to); }), #endif vect_(to - from + 1) { #ifndef BOOST_CONTRACT_NO_CONSTRUCTORS boost::contract::check c = boost::contract::constructor(this) #ifndef BOOST_CONTRACT_NO_POSTCONDITIONS .postcondition([&] { BOOST_CONTRACT_ASSERT(int(size()) == (to - from + 1)); }) #endif ; #endif for(int x = from; x <= to; ++x) vect_.at(x - from) = x; } virtual ~integers() { #ifndef BOOST_CONTRACT_NO_DESTRUCTORS // Check invariants. boost::contract::check c = boost::contract::destructor(this); #endif } virtual void push_back( int const& x #ifndef BOOST_CONTRACT_NO_PUBLIC_FUNCTIONS , boost::contract::virtual_* v = 0 #endif ) /* override */ { #ifndef BOOST_CONTRACT_NO_OLDS boost::contract::old_ptr<unsigned> old_size; #endif #ifndef BOOST_CONTRACT_NO_PUBLIC_FUNCTIONS boost::contract::check c = boost::contract::public_function< override_push_back>(v, &integers::push_back, this, x) #ifndef BOOST_CONTRACT_NO_PRECONDITIONS .precondition([&] { BOOST_CONTRACT_ASSERT(size() < max_size()); }) #endif #ifndef BOOST_CONTRACT_NO_OLDS .old([&] { old_size = BOOST_CONTRACT_OLDOF(v, size()); }) #endif #ifndef BOOST_CONTRACT_NO_POSTCONDITIONS .postcondition([&] { BOOST_CONTRACT_ASSERT(size() == *old_size + 1); }) #endif #ifndef BOOST_CONTRACT_NO_EXCEPTS .except([&] { BOOST_CONTRACT_ASSERT(size() == *old_size); }) #endif ; #endif vect_.push_back(x); } private: #ifndef BOOST_CONTRACT_NO_PUBLIC_FUNCTIONS BOOST_CONTRACT_OVERRIDE(push_back) #endif /* ... */
|
Static and volatile class invariants can be programmed using BOOST_CONTRACT_STATIC_INVARIANT
and BOOST_CONTRACT_INVARIANT_VOLATILE
respectively (these macros expand code equivalent to the static
void BOOST_CONTRACT_STATIC_INVARIANT_FUNC()
and void
BOOST_CONTRACT_INVARIANT_FUNC()
const volatile
functions).
The boost/contract_macro.hpp
macro interface is usually preferred because more concise and easier to use
than programming #ifndef BOOST_CONTRACT_NO_...
statements by hand. However, C++ macros expand on a single line of code and
that can make compiler errors less useful when using this macro interface
plus all contract assertions within a given set of preconditions, postconditions,
exception guarantees, and class invariants will list the same line number
in error messages when assertions fail at run-time (but error messages still
list the assertion code and that should still allow programmers to identify
the specific assertion that failed). Finally, the macro interface leaves
a bit of contract decorations in the code but that should add no measurable
compile-time or run-time overhead (specifically, extra boost::contract::virtual_
*
parameters, calls to boost::contract::constructor_precondition
default constructor which does nothing, BOOST_CONTRACT_BASE_TYPES
typedef
s, and boost::contract::access
friendships are left in user code even when contracts are disabled unless
#ifndef BOOST_CONTRACT_NO_...
statements are used).
Disabling contract as shown in Disable
Contract Checking leaves the overhead of compiling contract code plus
some small run-time overhead due to the initialization of old value pointers
(even if those will be all null and no old value will be actually copied),
the calls to the contract functions used to initialize boost::contract::check
and boost::contract::constructor_precondition
(even if those calls will be internally optimized by this library to essentially
do nothing), etc. For truly performance critical code for which even such
small run-time overhead might not be acceptable, the macro interface (or
the #ifndef BOOST_CONTRACT_NO_...
statements) can be used to completely
disable compile-time and run-time overheads of contracts. However, for such
performance critical code even the overhead of checking simple preconditions
might be too much so it might be best to not program contracts at all.
Usually, if the overhead of checking preconditions and other assertions is
already considered acceptable for an application then the compile-time overhead
of contracts should not represent an issue and it should be sufficient to
disable contract checking at run-time as indicated in Disable
Contract Checking (without a real need to use the boost/contract_macro.hpp
macro interface or the #ifndef
BOOST_CONTRACT_NO_...
statements in most cases).
Contracts are part of the program specifications and not of its implementation (see Specifications vs. Implementation). However, this library uses function definitions to program contracts so contract code appears together with the function implementation code. This is not ideal (even if contracts programmed using this library will always appear at the very beginning of the function definition so programmers will easily be able to distinguish contract code from the rest of the function implementation code so this might not be real limitation in practise).
In some cases, it might be desirable to completely separate the contract code from the function implementation code. For example, this could be necessary for software that ships only header files and compiled object files to its users. If contracts are programmed in function definitions that are compiled in the object files, users will not be able to see the contract code to understand semantics and usage of the functions (again, this might not be a real problem in practice for example if contracts are already somehow extracted from the source code by some tool and presented as part of the documentation of the shipped software).
In any case, when it is truly important to separate contracts from function
implementation code, function implementations can be programmed in extra
body functions (here named ..._body
, but any other naming scheme could
be used) that are compiled in object files. Function definitions that remain
in header files instead will contain just contract code followed by calls
to the extra body functions. This technique allows to keep the contract code
in header files while separating the implementation code to source and object
files. However, this adds the overhead of manually programming an extra function
declaration for each body function (plus the limitation that constructor
member initialization lists must be programmed in header files because that
is where constructors need to be defined to list constructor contract code).
[77]
For example, the following header file only contains function declarations,
contract code, and constructor member initializations, but it does not contain
the code implementing the function bodies (see separate_body.hpp
):
class iarray : private boost::contract::constructor_precondition<iarray> { public: void invariant() const { BOOST_CONTRACT_ASSERT(size() <= capacity()); } explicit iarray(unsigned max, unsigned count = 0) : boost::contract::constructor_precondition<iarray>([&] { BOOST_CONTRACT_ASSERT(count <= max); }), // Still, member initializations must be here. values_(new int[max]), capacity_(max) { boost::contract::check c = boost::contract::constructor(this) .postcondition([&] { BOOST_CONTRACT_ASSERT(capacity() == max); BOOST_CONTRACT_ASSERT(size() == count); }) ; constructor_body(max, count); // Separate constructor body impl. } virtual ~iarray() { boost::contract::check c = boost::contract::destructor(this); // Inv. destructor_body(); // Separate destructor body implementation. } virtual void push_back(int value, boost::contract::virtual_* v = 0) { boost::contract::old_ptr<unsigned> old_size = BOOST_CONTRACT_OLDOF(v, size()); boost::contract::check c = boost::contract::public_function(v, this) .precondition([&] { BOOST_CONTRACT_ASSERT(size() < capacity()); }) .postcondition([&] { BOOST_CONTRACT_ASSERT(size() == *old_size + 1); }) ; push_back_body(value); // Separate member function body implementation. } private: // Contracts in class declaration (above), but body implementations are not. void constructor_body(unsigned max, unsigned count); void destructor_body(); void push_back_body(int value); /* ... */
Instead, the function bodies are implemented in a separate source file (see
separate_body.cpp
):
void iarray::constructor_body(unsigned max, unsigned count) { for(unsigned i = 0; i < count; ++i) values_[i] = int(); size_ = count; } void iarray::destructor_body() { delete[] values_; } void iarray::push_back_body(int value) { values_[size_++] = value; } /* ... */
The same technique can be used for non-member, private, and protected functions, etc.
Note | |
---|---|
When contracts are programmed only in
On the flip side, if contracts are programmed only in header files (e.g.,
using extra |
This section shows how to use this library without C++11 lambda functions. This has some advantages:
..._precondition
, ..._old
, and ..._postcondition
functions in the example
below) can be programmed to fully enforce constant-correctness and other
contract requirements at compile-time (see Constant-Correctness).
[79]
However, not using C++11 lambda functions comes at the significant cost of
having to manually program the extra contract functions and related boiler-plate
code. For example, the header file (see no_lambdas.hpp
):
class iarray : private boost::contract::constructor_precondition<iarray> { public: static void static_invariant() { BOOST_CONTRACT_ASSERT(instances() >= 0); } void invariant() const { BOOST_CONTRACT_ASSERT(size() <= capacity()); } explicit iarray(unsigned max, unsigned count = 0); static void constructor_precondition(unsigned const max, unsigned const count) { BOOST_CONTRACT_ASSERT(count <= max); } static void constructor_old(boost::contract::old_ptr<int>& old_instances) { old_instances = BOOST_CONTRACT_OLDOF(instances()); } void constructor_postcondition(unsigned const max, unsigned const count, boost::contract::old_ptr<int> const old_instances) const { BOOST_CONTRACT_ASSERT(capacity() == max); BOOST_CONTRACT_ASSERT(size() == count); BOOST_CONTRACT_ASSERT(instances() == *old_instances + 1); } virtual ~iarray(); void destructor_old(boost::contract::old_ptr<int>& old_instances) const { old_instances = BOOST_CONTRACT_OLDOF(instances()); } static void destructor_postcondition(boost::contract::old_ptr<int> const old_instances) { BOOST_CONTRACT_ASSERT(instances() == *old_instances - 1); } virtual void push_back(int value, boost::contract::virtual_* v = 0); void push_back_precondition() const { BOOST_CONTRACT_ASSERT(size() < capacity()); } void push_back_old(boost::contract::virtual_* v, boost::contract::old_ptr<unsigned>& old_size) const { old_size = BOOST_CONTRACT_OLDOF(v, size()); } void push_back_postcondition( boost::contract::old_ptr<unsigned> const old_size) const { BOOST_CONTRACT_ASSERT(size() == *old_size + 1); } unsigned capacity() const; unsigned size() const; static int instances(); private: int* values_; unsigned capacity_; unsigned size_; static int instances_; };
And, the source file (see no_lambdas.cpp
):
iarray::iarray(unsigned max, unsigned count) : boost::contract::constructor_precondition<iarray>(boost::bind( &iarray::constructor_precondition, max, count)), values_(new int[max]), // Member initializations can be here. capacity_(max) { boost::contract::old_ptr<int> old_instances; boost::contract::check c = boost::contract::constructor(this) .old(boost::bind(&iarray::constructor_old, boost::ref(old_instances))) .postcondition(boost::bind( &iarray::constructor_postcondition, this, boost::cref(max), boost::cref(count), boost::cref(old_instances) )) ; for(unsigned i = 0; i < count; ++i) values_[i] = int(); size_ = count; ++instances_; } iarray::~iarray() { boost::contract::old_ptr<int> old_instances; boost::contract::check c = boost::contract::destructor(this) .old(boost::bind(&iarray::destructor_old, this, boost::ref(old_instances))) .postcondition(boost::bind(&iarray::destructor_postcondition, boost::cref(old_instances))) ; delete[] values_; --instances_; } void iarray::push_back(int value, boost::contract::virtual_* v) { boost::contract::old_ptr<unsigned> old_size; boost::contract::check c = boost::contract::public_function(v, this) .precondition(boost::bind(&iarray::push_back_precondition, this)) .old(boost::bind(&iarray::push_back_old, this, boost::cref(v), boost::ref(old_size))) .postcondition(boost::bind(&iarray::push_back_postcondition, this, boost::cref(old_size))) ; values_[size_++] = value; } unsigned iarray::capacity() const { // Check invariants. boost::contract::check c = boost::contract::public_function(this); return capacity_; } unsigned iarray::size() const { // Check invariants. boost::contract::check c = boost::contract::public_function(this); return size_; } int iarray::instances() { // Check static invariants. boost::contract::check c = boost::contract::public_function<iarray>(); return instances_; } int iarray::instances_ = 0;
If programmers also want to fully enforce all contract programming constant-correctness requirements at compile-time, they should follow these rules when programming the contract functions (see Constant-Correctness):
..._precondition
functions in the example
above) can take their arguments either by const
value or by const&
,
and when they are member functions they should be either static
or const
functions.
..._postcondition
functions in the example
above) should take their arguments by const&
, and when they are member functions
they should be either static
or const
functions.
const&
, and when they are member functions
they should be either static
or const
functions.
..._old
functions in the example above)
should take their arguments by const&
a part from old value pointers that
should be taken by &
(so only old value pointers can be modified), and when they are member
functions they should be either static
or const
functions.
static
(because
there is no valid object this
if the constructor body does not run successfully, see Constructor
Calls).
static
(because there is no valid object this
after the destructor body runs successfully, but exception guarantee
functions do not have to be static
since the object this
is
still valid because the destructor body did not run successfully, see
Destructor
Calls).
Note that the extra contract functions also allow to keep the contract code
in the header file while all function bodies are implemented in a separate
source file (including the constructor member initialization list, that could
not be done with the techniques shown in Separate
Body Implementation). [80] Also note that the contract functions can always be declared
private
if programmers need
to exactly control the public members of the class (this was not done in
this example only for brevity).
The authors think this library is most useful when used together with C++11 lambda functions (because of the large amount of boiler-plate code required when C++11 lambdas are not used as also shown by the example above).
It is possible to specify contracts without using most of the macros provided
by this library and programming the related code manually instead (the only
macros that cannot be programmed manually are BOOST_CONTRACT_OVERRIDE
,
BOOST_CONTRACT_OVERRIDES
,
and BOOST_CONTRACT_NAMED_OVERRIDE
).
Note | |
---|---|
Some of this library macros are variadic macros, others are not (see below). Variadic macros were officially added to the language in C++11 but most compilers have been supporting them as an extension for a long time, plus all compilers that support C++11 lambda functions should also support C++11 variadic macros (and this library might rarely be used without the convenience of C++11 lambda functions, see No Lambda Functions). [81] Therefore, the rest of this section can be considered mainly a curiosity because programmers should seldom, if ever, need to use this library without using its macros. |
As shown in Public
Function Overrides and Named
Overrides, this library provides the BOOST_CONTRACT_OVERRIDE
and BOOST_CONTRACT_NAMED_OVERRIDE
macros to program contracts for overriding public functions (see BOOST_CONTRACT_MAX_ARGS
for compilers
that do not support variadic templates). [82] These macro cannot be programmed manually but they are not variadic
macros (so programmers should be able to use them on any C++ compiler with
a sound support for SFINAE). [83] The BOOST_CONTRACT_OVERRIDES
macro is a variadic macro instead but programmes can manually repeat the
non-variadic macro BOOST_CONTRACT_OVERRIDE
for each overriding public function name on compilers that do not support
variadic macros.
As shown in Preconditions,
Postconditions,
Exception Guarantees,
Class Invariants,
etc. this library provides the BOOST_CONTRACT_ASSERT
macro to assert contract conditions. This is not a variadic macro and programmers
should be able to use it on all C++ compilers. In any case, the invocation
BOOST_CONTRACT_ASSERT(
cond
)
simply expands to code equivalent to the
following: [84]
if(!(cond
)) { throw boost::contract::assertion_failure(__FILE__, __LINE__, BOOST_PP_STRINGIZE(cond
)); }
In fact, this library considers any exception thrown from within preconditions,
postconditions, exception guarantees, and class invariants as a contract
failure and reports it calling the related contract failure handler (boost::contract::precondition_failure
,
etc.). If there is a need for it, programmers can always program contract
assertions that throw specific user-defined exceptions as follow (see Throw
on Failures):
if(!cond
) throwexception-object
;
However, using BOOST_CONTRACT_ASSERT
is convenient because it always allows this library to show an informative
message in case of assertion failure containing the assertion code, file
name, line number, etc.
As shown in Assertion
Levels, this library pre-defines BOOST_CONTRACT_ASSERT_AUDIT
and BOOST_CONTRACT_ASSERT_AXIOM
assertion levels. These macros are not variadic macros and programmers should
be able to use them on all C++ compilers. In any case, their implementations
are equivalent to the following:
#ifdef BOOST_CONTRACT_AUDITS #define BOOST_CONTRACT_ASSERT_AUDIT(cond
) \ BOOST_CONTRACT_ASSERT(cond
) #else #define BOOST_CONTRACT_ASSERT_AUDIT(cond
) \ BOOST_CONTRACT_ASSERT(true || (cond
)) #endif #define BOOST_CONTRACT_ASSERT_AXIOM(cond
) \ BOOST_CONTRACT_ASSERT(true || (cond
))
As shown in Base
Classes, this library provides the BOOST_CONTRACT_BASE_TYPES
variadic macro to declare the base_types
member type that will expand to the list of all public bases for a derived
class. Programmers can also declare base_types
without using BOOST_CONTRACT_BASE_TYPES
at the cost of writing a bit more code and increase maintenance efforts.
For example (see base_types_no_macro.cpp
):
#include <boost/mpl/vector.hpp> class chars : private boost::contract::constructor_precondition<chars>, public unique_chars, public virtual pushable<char>, virtual protected has_size, private has_empty { public: // Program `base_types` without macros (list only public bases). typedef boost::mpl::vector<unique_chars, pushable<char> > base_types; /* ... */
The base_types
member type
must be a boost::mpl::vector
which must list all and only public
base classes (because only public bases subcontract, see Function
Calls), and in the same order these public base classes appear in
the derived class inheritance list. If the BOOST_CONTRACT_BASE_TYPES
macro is not used, it is the responsibility of the programmers to maintain
the correct list of bases in the boost::mpl::vector
each time the derived class inheritance
list changes (this might significantly complicate maintenance).
In general, it is recommended to use the BOOST_CONTRACT_BASE_TYPES
macro whenever possible.
As shown in Old Values,
this library provides the BOOST_CONTRACT_OLDOF
variadic macro to assign old value copies. Programmers can also assign old
values without using BOOST_CONTRACT_OLDOF
at the cost of writing a bit more code manually. For example (see old_no_macro.cpp
):
template<typename T> class vector { public: virtual void push_back(T const& value, boost::contract::virtual_* v = 0) { // Program old value instead of using `OLD(size())` macro. boost::contract::old_ptr<unsigned> old_size = boost::contract::make_old(v, boost::contract::copy_old(v) ? size() : boost::contract::null_old()) ; boost::contract::check c = boost::contract::public_function(v, this) .postcondition([&] { BOOST_CONTRACT_ASSERT(size() == *old_size + 1); }) ; vect_.push_back(value); } /* ... */
The ternary operator boost::contract::copy_old(v)
? size() : boost::contract::null_old()
must be used here to avoid evaluating and
copying the old value expression size()
when boost::contract::copy_old
returns false
(because old values
are not being copied when postcondition and exception guarantee checking
is disabled at run-time, an overridden virtual function call is not checking
postconditions or exception guarantees yet, etc.). The enclosing boost::contract::make_old
copies the old value expression and creates an old value pointer. Otherwise,
boost::contract::null_old
indicates that a null old value pointer should be created.
The boost::contract::make_old
and boost::contract::copy_old
functions are used exactly as shown above but without the extra v
parameter when they are called from within
non-virtual functions (see Public
Function Overrides). The old value pointer returned by boost::contract::make_old
can be assigned to either boost::contract::old_ptr
or boost::contract::old_ptr_if_copyable
(see Old
Value Requirements).
In general, it is recommended to use the BOOST_CONTRACT_OLDOF
macro whenever possible.
Almost all macros defined in boost/contract_macro.hpp
are variadic macros. On compilers that do not support variadic macros, programmers
can manually disable contract code compilation using #ifndef
BOOST_CONTRACT_NO_...
statements as shown in Disable
Contract Compilation.
[68]
Rationale: [N1962]
and other proposals to add contracts to C++ do not provide a mechanism
to selectively disable copies only for old value types that are not copy
constructible. However, this library provides such a mechanism to allow
to program contracts for template code without necessarily adding extra
copy constructible type requirements that would not be present if it were
not for copying old values (so compiling the code with and without contracts
will not necessarily alter the type requirements of the program). Something
similar could be achieved combing C++17 if
constexpr
with [N1962]
or [P0380] so that old value expressions
within template code can be guarded by if
constexpr
statements checking if
the old value types are copyable or not. For example, assuming old values
are added to [P0380] (using some kind
of oldof(...)
syntax) and that C++17 if constexpr
can be used within [P0380]
contracts:
template<typename T> void offset(T& x, int count) [[ensures: if constexpr(std::is_copy_constructible<T>::value) x == oldof(x) + count]] ...
[69]
Technically, on C++17 it is possible to use boost::contract::old_ptr
together with if constexpr
instead of using boost::contract::old_ptr_if_copyable
,
for example:
template<typename T> void offset(T& x, int count) { boost::contract::old_ptr<T> old_x; if constexpr(boost::contract::is_old_value_copyable<T>::value) old_x = BOOST_CONTRACT_OLDOF(x); boost::contract::check c = boost::contract::function() .postcondition([&] { if constexpr(boost::contract::is_old_value_copyable<T>::value) BOOST_CONTRACT_ASSERT(x == *old_x + count); }) ; x += count; }
However, the authors find this code less readable and more verbose than
its equivalent that uses boost::contract::old_ptr_if_copyable
.
Guarding old value copies and related assertions with if
constexpr
is useful instead when
the guard condition checks type requirements more complex than just boost::contract::is_old_value_copyable
(as shown later in this documentation).
[70]
Rationale: [N1962]
and other proposals to add contracts to C++ do not provide a mechanism
to selectively disable assertions based on their type requirements. However,
this library provides such a mechanism to allow to program contracts for
template code without necessarily adding extra type requirements that would
not be present if it was not for the contracts (so compiling the code with
and without contracts will not alter the type requirements of the program).
Something similar could be achieved combing C++17 if
constexpr
with [N1962]
or [P0380] so that contract assertions
within template code could be guarded by if
constexpr
statements checking the
related type requirements ([N1962]
already allows of if
statements
in contracts under the name of select assertions,
[P0380] does not so probably if
statements should be added to [P0380]
as well). For example, assuming C++17 if
constexpr
can be used within [P0380] contracts:
template<typename T> class vector { public: void push_back(T const& value) [[ensures: if constexpr(boost::has_equal_to<T>::value) back() == value]] ... };
[71]
The internal implementation of boost::contract::condition_if
is optimized and it does not actually use boost::contract::call_if
.
A part from its use within contracts, boost::contract::call_if
can be used together with C++14 generic lambdas to emulate C++17 if constexpr
(boost::hana::if_
and probably other approaches can
also be used together with generic lambdas to emulate C++17 if constexpr
on C++14 compilers). For example, the following implementation of myadvance
will compile since C++14
and it is more concise, easier to read and maintain than the usual implementation
of std::advance
that uses tag dispatching (see
call_if_cxx14.cpp
):
template<typename Iter, typename Dist> void myadvance(Iter& i, Dist n) { Iter* p = &i; // So captures change actual pointed iterator value. boost::contract::call_if<is_random_access_iterator<Iter> >( std::bind([] (auto p, auto n) { // C++14 generic lambda. *p += n; }, p, n) ).template else_if<is_bidirectional_iterator<Iter> >( std::bind([] (auto p, auto n) { if(n >= 0) while(n--) ++*p; else while(n++) --*p; }, p, n) ).template else_if<is_input_iterator<Iter> >( std::bind([] (auto p, auto n) { while(n--) ++*p; }, p, n) ).else_( std::bind([] (auto false_) { static_assert(false_, "requires at least input iterator"); }, std::false_type()) // Use constexpr value. ); }
Of course, since C++17 the implementation that uses if
constexpr
is even more readable
and concise:
template<typename Iter, typename Dist> void myadvance(Iter& i, Dist n) { if constexpr(is_random_access_iterator<Iter>::value) { i += n; } else if constexpr(is_bidirectional_iterator<Iter>::value) { if(n >= 0) while(n--) ++i; else while(n++) --i; } else if constexpr(is_input_iterator<Iter>::value) { while(n--) ++i; } else { static_assert(false, "requires at least input iterator"); } }
[73]
Rationale: Constructors and destructors
check const volatile
and const
invariants in that
order because the qualifier that can be applied to more calls is checked
first (note that const volatile
calls can be made on any object while const
calls cannot be made on volatile
objects, in that sense the const
volatile
qualifier can be applied
to more calls than const
alone
can). This is consistent with static
class invariants that are checked even before const
volatile
invariants (the static
classifier can be applied to even
more calls than const volatile
,
in fact an object is not even needed to make static calls).
[74]
Rationale: Note that while all public
functions can be made to check const
volatile
invariants, it is never
possible to make volatile public functions check const
non-volatile invariants. That is because both const
and volatile
can always be
added but never stripped in C++ (a part from forcefully via const_cast
) but const
is always automatically added by this library in order to enforce contract
constant-correctness (see Constant-Correctness).
That said, it would be too stringent for this library to also automatically
add volatile
and require all
functions to check const volatile
(not just const
)
invariants because only volatile
members can be accessed from const
volatile
invariants so there could
be many const
(but not const volatile
)
members that are accessible from const
invariants but not from const volatile
invariants. To avoid this confusion,
this library has chosen to draw a clear dichotomy between const
and const
volatile
invariants so that only
volatile public functions check const
volatile
invariants and only non-volatile
public functions check const
(but not const volatile
)
invariants. This is a clear distinction and it should serve most cases.
If programmers need non-volatile public functions to also check const volatile
invariants, they can explicitly do so by calling the const
volatile
invariant function from
the const
invariant function
as shown in this documentation.
[75]
In this example, the moved()
function is simple enough that programmers
could decide to not even call boost::contract::public_function
from it for optimization reasons. However, calling boost::contract::public_function
from moved()
has no negative impact, a part from run-time overhead, because this library
automatically disables contract checking while checking other contracts
(so this call will not cause infinite recursion).
[76] The assertion levels predefined by this library are similar to the default, audit, and axiom levels from [P0380].
[77] When used as default parameter values, lambda functions allow to program code statements within function declarations. However, these lambadas cannot be effectively used to program contracts in function declarations instead of definitions. That is because the C++11 standard does not allow lambdas in function declarations to capture any variable (for the good reason that it is not at all obvious how to correctly define the semantics of such captures). For example, the following code is not valid C++ and it does not compile:
// Specifications (in declaration). int inc(int& x, // Error: Lambdas in default parameters cannot capture `this`, `x`, or any other variable. std::function<void ()> pre = [&] { BOOST_CONTRACT_ASSERT(x < std::numeric_limits<int>::max()); }, std::function<void (int const&, boost::contract::old_ptr<int> const&)> post = [&] (int const& result, boost::contract::old_ptr<int> const& old_x) { BOOST_CONTRACT_ASSERT(x == *old_x + 1); BOOST_CONTRACT_ASSERT(result == *old_x); } ); // Implementation (in definition). int inc(int& x, std::function<void ()> pre, std::function<void (int const&, boost::contract::old_ptr<int> const&)> post ) { int result; boost::contract::old_ptr<int> old_x = BOOST_CONTRACT_OLDOF(x); boost::contract::check c = boost::contract::function() .precondition(pre) .postcondition(std::bind(post, std::cref(result), std::cref(old_x))) ; return result = x++; // Function body. }
In any case, even if the above code compiled, it would require significant boiler-plate code to bind return and old values.
[78]
Alternatively, on compilers that do not support C++11 lambda functions,
Boost.LocalFunction
could be used to program the contract functors still within the function
definitions (for example, see no_lambda_local_func.cpp
).
In general, such a code is less verbose than the example shown in this
section that uses contract functions programmed outside of the original
function definitions (about 30% less lines of code) but the contract
code is hard to read. Other libraries could also be used to program
the contract functors without C++11 lambda functions (Boost.Lambda,
Boost.Fusion, etc.) but again all these techniques will result in contract
code either more verbose, or harder to read and maintain than the code
that uses C++11 lambda functions.
[79]
If C++ allowed lambda functions to capture variables by constant reference
(for example allowing a syntax like this [const&]
{ ...
}
and [const&
variable-name
] { ... }
,
see https://groups.google.com/a/isocpp.org/forum/#!topic/std-proposals/0UKQw9eo3N0)
also lambdas could be used to program contract functors that fully
enforce Constant-Correctness
at compile-time. Note that C++11 lambdas allow to capture variables
by value (using [=] {
... }
and [
variable-name
] { ... }
)
and these value captures are const
(unless the lambda is explicitly declared mutable
)
but they are not suitable to program postconditions and exception guarantees
using this library (because those require capturing by reference, see
Postconditions
and Exception
Guarantees), plus they introduce a copy of the captured value
that might be too expensive in general and therefore not suitable for
preconditions either.
[80]
In this example, bind
was
used to generate nullary functors from the contract functions. As always
with bind
, cref
and ref
must be used to bind arguments by const&
and &
respectively, plus it might be necessary to explicitly static_cast
the function pointer passed to bind
for overloaded functions.
[81] Compilation times of this library were measured to be comparable between compilers that support variadic macros and compilers that do not.
[82]
Rationale: The BOOST_CONTRACT_MAX_ARGS
macro is named after BOOST_FUNCTION_MAX_ARGS
.
[83]
Rationale: These macros expand to SFINAE-based
introspection template code that are too complex to be programmed manually
by users (that remains the case even if C++14 generic lambdas were to be
used here). On a related note, in theory using C++14 generic lambdas, the
BOOST_CONTRACT_OVERRIDE
macro could be re-implemented in a way that can be expanded at function
scope, instead of class scope (but there is not really a need to do that).
[84]
Rationale: There is no need for the code
expanded by BOOST_CONTRACT_ASSERT
to also use C++11 __func__
.
That is because __func__
will always expand to the name operator()
of the functor used to program the contract
assertions (e.g., the internal name the compiler assigns to lambda functions)
and it will not expand to the name of the actual function enclosing the
contract declaration.