/**************************************************************************\ MODULE: vector SUMMARY: Template class for dynamic-sized vectors. The declaration Vec<T> v; creates a zero-length vector. To grow this vector to length n, execute v.SetLength(n) This causes space to be allocated for (at least) n elements, and also causes the default constructor for T to be called to initialize these elements. The current length of a vector is available as v.length(). Let n = v.length(). Calling v.SetLength(m) with m <= n sets the current length of v to m (but does not call any destructors or free any space). Calling v.SetLength(m) with m > n will allocate space and initialize as necessary, but will leave the values of the already allocated elements unchanged (although their addresses may change). If T has a user-defined default constructor, that is invoked. Otherwise, the new memory locations are "default initialized". In particular, this means that POD types may be uninitialized. v.MaxLength() is the largest value of n for which v.SetLength(n) was invoked, and is equal to the number of entries that have been initialized. v.SetMaxLength(n) will allocate space for and initialize up to n elements, without changing v.length(). When v's destructor is called, all constructed elements will be destructed, and all space will be relinquished. v.allocated() is the number of elements which have been allocated, which may be more than the number elements initialized. Note that if n <= v.allocated(), then v.SetLength(n) is guaranteed not to cause any memory allocation or movement of objects. Space is managed using malloc, realloc, and free. When a vector is grown, a bit more space may be allocated than was requested for efficiency reasons. Note that when a vector is grown, the objects in the vector may have moved, possibly creating dangling references to vector elements. One has to be especially careful of this when using vectors passed as reference parameters that may alias one another. ---------------------- IMPLEMENTATION DETAILS ---------------------- A Vec<T> object is just a pointer to the first element of the array. There is a control block immediately before the first element that keeps track of several parameters: len -- the logical length of the array (returned by length()) init -- the number of elements constructed (returned ny MaxLength()) alloc -- the number of elements for which space has been allocated (returned by allocated()) fixed -- flag that indicates that the length is fixed (returned by fixed()) Note that 0 <= len <= init <- alloc --------------- SPECIALIZATIONS --------------- NTL currently defines an optimized specialization Vec<GF2>. There are a few quirks to be aware of. See the documentation file for vec_GF2 for more details. --------------------- RANGE-BASED FOR LOOPS --------------------- NTL's vectors provide sufficient functionality to enable range-based for-loops (in C++11). The safest ways to use this functionality are to write: for (auto&& item : vec) { ... } // for read-only or read/write access or for (T item : vec) { ... } // for access via a copy This is especially true if vec may be of type Vec<GF2>. Again, see the documentation file for vec_GF2 for more details. -------------- OTHER FEATURES -------------- The i-th vector element (counting from 0) is accessed as v[i]. If the macro NTL_RANGE_CHECK is defined, code is emitted to test if 0 <= i < v.length(). This check is not performed by default. For old-time FORTRAN programmers, the i-th vector element (counting from 1) is accessed as v(i). -------------------- REALLOCATION DETAILS -------------------- When a vector is resized and not enough space has been allocated, then a reallocation process takes place. There are several strategies which can be used, which we list here from least to most efficient. COPY AND DESTROY STRATEGY: This is the same strategy used by the STL prior to C++11. New memory is allocated, and each element of the vector is copied (using a copy constructor) to the new memory, and then all the elements in the old memory is destroyed (using a destructor). While much safer than the Realloc Strategy, it can be a bit slower. MOVE STRATEGY: This is a strategy that is also used by the STL, assuming C++11. New memory is allocated, and each element of the vector is moved (using a move constructor) to the new memory, and then all the elements in the old memory is destroyed (using a destructor). This strategy is only viable if the underlying type has a "nothrow" move constructor. Typically, it will be significantly faster than the Copy and Destroy Strategy. REALLOC STRATEGY: This strategy simply calls the realloc function, which may result in objects being copied bit-wise from one location in memory to another. For most types of objects, even those with non-trivial constructors and destructors, this is perfectly safe, as most objects really don't care what address they are stored at. As an example of where this doesn't work, STL string implementations using a "short string" optimization may consist of an object containing a pointer to a sub-object, which would be invalidated if the object were bit-wise moved from one location to another. This strategy is the fastest of them all. CHOOSING A STRATEGY: In the implementation of Vec<T>, NTL will opt for the Realloc Strategy if one of the two conditions hold: (i) T has a trivial copy constructor and a trivial destructor, or (iii) T is a class that has been explicit declared "relocatable". (See below for details on how to declare that a class is relocatable.) Otherwise, NTL will use the Move Strategy if T has a nothrow move costructor. Otherwise, NTL will use the Copy Strategy if T has a copy constructor that is accessible (i.e., not private and not deleted). Otherwise, no reallocation implementation is provided, and a runtime error will be raised if a reallocation is required (this provides maximum flexiblilty and also maximum backward compatibility with previous versions of NTL, even though sometimes a compile-time diagnostic would be preferable). DECLARING A CLASS RELOCATABLE: One can declare T to be relocatable by adding the following line of code immediately after the definition of class T: NTL_DECLARE_RELOCATABLE((T*)) Notice the double parentheses and the "*" --- these are unfortunately necessary. This macro expands as constexpr bool DeclareRelocatableType(T*) { return true; } Inside the class Vec<T>, this function is invoked with ((T*) 0) as the argument. By declaring relocatability via a function definition in this way, the Vec<T> class will always find (via "Koenig lookup", or ADL) the declaration for T in the namespace in which T is defined (assuming you put the relocatability declaration there as well). You can also declare a template class to be relocatable, as in: template<class X, class Y> NTL_DECLARE_RELOCATABLE((T<X,Y>*)) There is also a macro NTL_DECLARE_RELOCATABLE_WHEN, which leaves off the function body, if you want something other than { return true; } NTL also defines template functions: template<class T> constexpr bool DeclareRelocatableType(T*) { return ... ; } where "..." is an expression that returns true if T is a simple type that is safe to relocate, as determined by C++11 type traits. Such simple types include scalar types and simple classes with trivial copy constructor and destructor. NTL already declares as relocatable most if the fundamental types fot which it makes sense to do so. UNSAFE VECTOR MODE: The above describes the default behavior of NTL as of version 11.0. Prior to version 11.0, the default behavior was to only use the Realloc Strategy, which is potentially unsafe. You can revert to this old behavior by configuring NTL_SAFE_VECTORS=off. ------------------------- COMPARISON TO STL VECTORS ------------------------- When the length of an NTL vector is reduced, no objects are destroyed. In contrast, when the length of an STL vector is reduced, objects are destroyed (effectively, maintaining the invariant len == init). Also, STL vectors never use anything like the Realloc Strategy. ---------------------- HISTORICAL PERSPECTIVE ---------------------- When I first started working on NTL around 1990, C++ was a much simpler language. Among other things, there were no templates and no STL, so the Vec class started out (and remained, for many years) a set of macros (indeed, this approach was advocated in the first edition of Stroustrup's book). My thinking at that time was very much "C oriented". The idea of resizing a vector without using realloc never occured to me, and all of the classes I wanted to put in vectors were relocatable. Why would you ever bother making copies of vector elements and destroy the originals when you could just realloc? It wasn't until many years later that I even realized this was somehow questionable practice. Indeed, resizing via malloc is technically undefined behavior, but it's been working for me for over 25 years without problems. Furthermore, because of the poor performance of malloc in those days (which is much better now), I designed the Vec<ZZ_p> class (and a few others) to allocate the underlying ZZ's in blocks. This not only reduces the number of calls to malloc, but it also gives better locality of reference. Experiments in the last couple of years show that this is still benefit to doing this. With all of these design decisions baked into NTL's Vec class, transitioning to STL vectors would be problematic, and is unlikely to ever happen. But as things have evolved, NTL Vec's offers many of the same convenience and safety features as STL vector's. \**************************************************************************/ // EXCEPTIONS: all functions below do not throw any exceptions, // except as noted template<class T> class Vec { public: Vec(); // initially length 0 Vec(const Vec& a); // copy constructor; uses the assignment operator of T // for copying into locations that have already been initialized, // and uses the copy constructor for T for initializing new locations. // EXCEPTIONS: may throw Vec& operator=(const Vec& a); // assignment; uses the assignment operator of T // for copying into locations that have already been initialized, // and uses the copy constructor for T for initializing new locations. // EXCEPTIONS: weak ES (but if it throws, neither length nor MaxLength // will change, although some previously initialized elements // may have been assigned new values). Vec(Vec&& a); // move constructor (C++11 only) // declared noexcept unless NTL_EXCEPTIONS flag is set // will revert to copy constructor if a is fixed Vec& operator=(Vec&& a); // move assignment (C++11 only) // declared noexcept unless NTL_EXCEPTIONS flag is set // will revert to copy assignment if *this or a is fixed // NOTE: If neither source nor destination are fixed, these operations will // be fast pointer moves and no exceptions will be thrown. If either are // fixed (say, the row in a Mat<T>), these operations revert to copy // operations, which may either throw an exception (if NTL_EXCEPTIONS is on) // or abort the program (if NTL_EXCEPTIONS is off). // NOTE: In a vector copy assignment x=a, all iterators, references, // and pointers related to x may be invalidated by the assignment. // In a vector move assignment, the same applies to a as well. // BACKWARD COMPATIBILITY: after introducing of "move constructors" in NTL // 10.4, there have been some reports of incompatibilties. Some legacy NTL // clients have been known to assume that x=a does not invalidate pointers // into x, provided x.length() >= a.length() before the assignment. This // assumption is not valid, however, for a move assignment. Because of this // problem, as of NTL 11.0, move assignments have been disabled by default // (this is controlled by the configuration flag NTL_DISABLE_MOVE_ASSIGN). // This affects vectors and matrices, as well as many other types (such as // polynomials) that are implemented in terms of vectors. Move assignment // is not disabled. ~Vec(); // destructor: calls T's destructor for all initialized // elements in the vector, and then frees the vector itself void SetLength(long n); // set current length to n, growing vector if necessary // new objects are initialized using the default contructor for T // EXCEPTIONS: strong ES (but the vector may have been // reallocated) void SetLength(long n, const T& a); // set current length to n, growing vector if necessary // new objects are initialized using the copy contructor for T // EXCEPTIONS: strong ES (but the vector may have been // reallocated) template<class F> void SetLengthAndApply(long n, F f); // set current length to n, growing vector if necessary // any new objects are initialized using defauly constructor // for T, and after that, f is applied to each new object x // as f(x). // EXCEPTIONS: strong ES (but the vector may have been // reallocated) long length() const; // current length T& operator[](long i); const T& operator[](long i) const; // indexing operation, starting from 0. // The first version is applied to non-const Vec<T>, // and returns a non-const reference to a T, while the second version // is applied to a const Vec<T> and returns a const reference to a T. // EXCEPTIONS: may throw if range checking turned on, strong ES T& operator()(long i); const T& operator()(long i) const; // indexing operation, starting from 1 // The first version is applied to non-const Vec<T>, // and returns a non-const reference to a T, while the second version // is applied to a const Vec<T> and returns a const reference to a T. // EXCEPTIONS: may throw if range checking turned on, strong ES T* elts(); const T* elts() const; // returns address of first vector element (or 0 if no space has been // allocated for this vector). If a vector potentially has length 0, it is // safer to write v.elts() instead of &v[0]: the latter is not well defined // by the C++ standard (although this is likely an academic concern). // // The first version is applied to non-const Vec<T>, and returns a non-const // pointer to a T, while the second version is applied to a const Vec<T> and // returns a const reference to a T. void swap(Vec& y); // swap with y (fast: just swaps pointers) // EXCEPTIONS: throws if vectors are fixed and lengths do not match, strong ES void move(Vec& y); // move y into *this, killing y (fast: just moves pointers) // EXCEPTIONS: strong ES, raises an error if // &y != this and either y or *this are fixed void append(const T& a); // append a to end of vector; uses the assignment operator of T // for copying into locations that have already been initialized, // and uses the copy constructor for T for initializing new locations. // EXCEPTIONS: strong ES if initializing a new element (and in any // case, if an exception throws, length and MaxLength remain // unchanged). void append(const Vec& w); // append w to end of vector; uses the assignment operator of T // for copying into locations that have already been initialized, // and uses the copy constructor for T for initializing new locations. // EXCEPTIONS: strong ES if initializing new elements (and in any // case, if an exception throws, length and MaxLength remain // unchanged). // Alternative access interface const T& get(long i) const; // v.get(i) returns v[i] void put(long i, const T& a); // v.put(i, a) equivalent to v[i] = q // Some STL compatibility typedef T value_type; typedef value_type& reference; typedef const value_type& const_reference; typedef value_type *iterator; typedef const value_type *const_iterator; T* data(); const T* data() const; // v.data() same as v.elts() T* begin(); const T* begin() const; // v.begin() same as v.elts() T* end(); const T* end() const; // pointer to (one past) last element (or NULL) T& at(long i); const T& at(long i) const; // indexing with range checking // the remaining member functions are a bit esoteric (skip on first // reading) Vec(INIT_SIZE_TYPE, long n); // Vec(INIT_SIZE, n) initializes vector with an intial length of n. // new objects are initialized using the default contructor for T // EXCEPTIONS: may throw Vec(INIT_SIZE_TYPE, long n, const T& a); // Vec(INIT_SIZE, n, a) initializes vector with an intial length of n. // new objects are initialized using the copy contructor for T // EXCEPTIONS: may throw void kill(); // release space and set to length 0 void SetMaxLength(long n); // allocates space and initializes up to n elements. Does not change // current length // EXCEPTIONS: may throw, strong ES void FixLength(long n); // sets length to n and prohibits all future length changes. // FixLength may only be invoked immediately after the default // construction or kill. // The kill operation is also subsequently prohibited, and swap is // allowed on fixed length vectors of the same length. // FixLength is provided mainly to implement Mat<T>, to enforce // the restriction that all rows have the same length. // EXCEPTIONS: may throw, strong ES void FixAtCurrentLength(); // fixes the length at the cuurent length and prohibits // all future length changes. // It is required that length() == MaxLength() when called. // EXCEPTIONS: if length() != MaxLength() and error is raised; // if length() == 0, a memory allocation error may be raised. // Strong ES. long fixed() const; // test if length has been fixed by FixLength() or FixAtCurrentLength() long MaxLength() const; // maximum length, i.e., number of allocated and initialized elements long allocated() const; // the number of objects for which space has been allocated, but not // necessarily initialized; this may be larger than MaxLength(). T& RawGet(long i); const T& RawGet(long i) const; // indexing with no range checking long position(const T& a) const; // returns position of a in the vector, or -1 if it is not there. // The search is conducted from position 0 to allocated()-1 the vector, // and an error is raised if the object is found at position MaxLength() // or higher (in which case a references an uninitialized object). // Note that if NTL_CLEAN_PTR flag is set, this routine takes // linear time, and otherwise, it takes constant time. // EXCEPTIONS: may throw (as indicated above) long position1(const T& a) const; // returns position of a in the vector, or -1 if it is not there. // The search is conducted from position 0 to length()-1 of the vector. // Note that if NTL_CLEAN_PTR flag is set, this routine takes // linear time, and otherwise, it takes constant time. }; /**************************************************************************\ Some utility routines \**************************************************************************/ template<class T> void swap(Vec<T>& x, Vec<T>& y); // swaps x & y; same as x.swap(y) // EXCEPTIONS: same as for swap member function template<class T> void append(Vec<T>& v, const T& a); // appends a to the end of v; same as v.append(a) // EXCEPTIONS: same as for append member function template<class T> void append(Vec<T>& v, const Vec<T>& w); // appends w to the end of v; same as v.append(w) // EXCEPTIONS: same as for append member function /**************************************************************************\ Input/Output The I/O format for a vector v with n elements is: [v[0] v[1] ... v[n-1]] Uses corresponding I/O operators for T \**************************************************************************/ template<class T> istream& operator>>(istream&, Vec<T>&); // EXCEPTIONS: may throw, weak ES template<class T> ostream& operator<<(ostream&, const Vec<T>&); // EXCEPTIONS: may throw, weak ES /**************************************************************************\ Equality Testing \**************************************************************************/ template<class T> long operator==(const Vec<T>& a, const Vec<T>& b); template<class T> long operator!=(const Vec<T>& a, const Vec<T>& b); /**************************************************************************\ Customized Constructors and Destructors Esoteric: skip on first reading...also these interfaces are subject to change When new elements in a vector need to be constructed, one of the following routines is called: void BlockConstruct(T* p, long n); // invokes T() to initialize p[i] for i = 0..n-1 void BlockConstructFromVec(T* p, long n, const T* q); // invokes T(q[i]) to initialize p[i] for i = 0..n-1; // q points to elements from a Vec<T> void BlockConstructFromObj(T* p, long n, const T& q); // invokes T(q) to initialize p[i] for i = 0..n-1 When a vector is destroyed, the following routine is called: void BlockDestroy(T* p, long n); // invokes ~T() on p[i] for i = 0..n-1 The default behavior of these routines may be modified by overloading these functions with a custom implementation. EXCEPTIONS: In order to provide exception safe code, the Construct routines should provide strong ES; in particular, if any constructor throws, all newly constructed objects should be destroyed. Moreover, the BlockDestroy routine should not throw at all. In NTL, these routines are overridden for the ZZ_p and GF2E classes, so that many vector entries will be packed into contiguous storage locations. This reduces the number of invocations of malloc, and increases locality of reference. \**************************************************************************/