| 1 | [[PageOutline]] |
| 2 | |
| 3 | = Portability Hints: Borland C++ 5.5.1 = #intro |
| 4 | |
| 5 | It is a general aim for boost libraries to be |
| 6 | [wiki:Guidelines/Requirements#Portability portable]. The |
| 7 | primary means for achieving this goal is to adhere to ISO |
| 8 | Standard C++. However, ISO C++ is a broad and complex standard |
| 9 | and most compilers are not fully conformant to ISO C++ yet. In |
| 10 | order to achieve portability in the light of this restriction, |
| 11 | it seems advisable to get acquainted with those language |
| 12 | features that some compilers do not fully implement yet. |
| 13 | |
| 14 | This page gives portability hints on some language features |
| 15 | of the Borland C++ version 5.5.1 compiler. Furthermore, the |
| 16 | appendix presents additional problems with Borland C++ version |
| 17 | 5.5. Borland C++ 5.5.1 is a freely available command-line |
| 18 | compiler for Win32 available at http://www.borland.com/. |
| 19 | |
| 20 | Each entry in the following list describes a particular |
| 21 | issue, complete with sample source code to demonstrate the |
| 22 | effect. Most sample code herein has been verified to compile |
| 23 | with gcc 2.95.2 and Comeau C++ 4.2.44. |
| 24 | |
| 25 | == Preprocessor symbol == #preprocessor |
| 26 | |
| 27 | The preprocessor symbol `__BORLANDC__` is defined |
| 28 | for all Borland C++ compilers. Its value is the version number |
| 29 | of the compiler interpreted as a hexadecimal number. The |
| 30 | following table lists some known values. |
| 31 | |
| 32 | ||Compiler ||`__BORLANDC__` value|| |
| 33 | ||Borland C++ Builder 4 ||0x0540 || |
| 34 | ||Borland C++ Builder 5 ||0x0550 || |
| 35 | ||Borland C++ 5.5 ||0x0550 || |
| 36 | ||Borland C++ 5.5.1 ||0x0551 || |
| 37 | ||Borland C++ Builder 6 ||0x0560 || |
| 38 | |
| 39 | == Core Language == #core |
| 40 | |
| 41 | === [using-directive] Mixing `using`-declarations and `using`-directives === #using-directive |
| 42 | |
| 43 | Mixing `using`-directives (which refer to whole |
| 44 | namespaces) and namespace-level `using`-declarations |
| 45 | (which refer to individual identifiers within foreign |
| 46 | namespaces) causes ambiguities where there are none. The |
| 47 | following code fragment illustrates this: |
| 48 | {{{ |
| 49 | namespace N { |
| 50 | int x(); |
| 51 | } |
| 52 | |
| 53 | using N::x; |
| 54 | using namespace N; |
| 55 | |
| 56 | int main() |
| 57 | { |
| 58 | &x; // Ambiguous overload |
| 59 | } |
| 60 | }}} |
| 61 | |
| 62 | === [using template] `using`-declarations for class templates === #using-template |
| 63 | |
| 64 | Identifiers for class templates can be used as arguments to |
| 65 | `using`-declarations as any other identifier. |
| 66 | However, the following code fails to compile with Borland |
| 67 | C++: |
| 68 | {{{ |
| 69 | template<class T> |
| 70 | class X { }; |
| 71 | |
| 72 | namespace N |
| 73 | { |
| 74 | // "cannot use template 'X<T>' without specifying specialization parameters" |
| 75 | using ::X; |
| 76 | }; |
| 77 | }}} |
| 78 | |
| 79 | === [template const arg] Deduction of constant arguments to function templates === #template-const-arg |
| 80 | |
| 81 | Template function type deduction should omit top-level |
| 82 | constness. However, this code fragment instantiates "f<const int>(int)": |
| 83 | {{{ |
| 84 | template<class T> |
| 85 | void f(T x) |
| 86 | { |
| 87 | x = 1; // works |
| 88 | (void) &x; |
| 89 | T y = 17; |
| 90 | y = 20; // "Cannot modify a const object in function f<const int>(int)" |
| 91 | (void) &y; |
| 92 | } |
| 93 | |
| 94 | int main() |
| 95 | { |
| 96 | const int i = 17; |
| 97 | f(i); |
| 98 | } |
| 99 | }}} |
| 100 | |
| 101 | === [function address] Resolving addresses of overloaded functions === #function-address |
| 102 | |
| 103 | Addresses of overloaded functions are not in all contexts |
| 104 | properly resolved (std:13.4 [over.over]); here is a small |
| 105 | example: |
| 106 | {{{ |
| 107 | template<class Arg> |
| 108 | void f( void(*g)(Arg) ); |
| 109 | |
| 110 | void h(int); |
| 111 | void h(double); |
| 112 | |
| 113 | template<class T> |
| 114 | void h2(T); |
| 115 | |
| 116 | int main() |
| 117 | { |
| 118 | void (*p)(int) = h; // this works (std:13.4-1.1) |
| 119 | void (*p2)(unsigned char) = h2; // this works as well (std:13.4-1.1) |
| 120 | f<int>(h2); // this also works (std:13.4-1.3) |
| 121 | |
| 122 | // "Cannot generate template specialization from h(int)", |
| 123 | // "Could not find a match for f<Arg>(void (*)(int))" |
| 124 | f<double>(h); // should work (std:13.4-1.3) |
| 125 | |
| 126 | f( (void(*)(double))h); // C-style cast works (std:13.4-1.6 with 5.4) |
| 127 | |
| 128 | // "Overloaded 'h' ambiguous in this context" |
| 129 | f(static_cast<void(*)(double)>(h)); // should work (std:13.4-1.6 with 5.2.9) |
| 130 | } |
| 131 | }}} |
| 132 | |
| 133 | '''Workaround:''' Always use C-style casts when |
| 134 | determining addresses of (potentially) overloaded |
| 135 | functions. |
| 136 | |
| 137 | === [string conversion] Converting `const char *` to `std::string` === #string-conversion |
| 138 | |
| 139 | Implicitly converting `const char *` parameters |
| 140 | to `std::string` arguments fails if template |
| 141 | functions are explicitly instantiated (it works in the usual |
| 142 | cases, though): |
| 143 | {{{ |
| 144 | #include <string> |
| 145 | |
| 146 | template<class T> |
| 147 | void f(const std::string & s) |
| 148 | {} |
| 149 | |
| 150 | int main() |
| 151 | { |
| 152 | f<double>("hello"); // "Could not find a match for f<T>(char *)" |
| 153 | } |
| 154 | |
| 155 | }}} |
| 156 | |
| 157 | '''Workaround:''' Avoid explicit template |
| 158 | function instantiations (they have significant problems with |
| 159 | Microsoft Visual C++) and pass default-constructed unused dummy |
| 160 | arguments with the appropriate type. Alternatively, if you wish |
| 161 | to keep to the explicit instantiation, you could use an |
| 162 | explicit conversion to `std::string` or declare the |
| 163 | template function as taking a `const char *` |
| 164 | parameter. |
| 165 | |
| 166 | === [template value defaults] Dependent default arguments for template value parameters === #template-value-defaults |
| 167 | |
| 168 | Template value parameters which default to an expression |
| 169 | dependent on previous template parameters don't work: |
| 170 | {{{ |
| 171 | template<class T> |
| 172 | struct A |
| 173 | { |
| 174 | static const bool value = true; |
| 175 | }; |
| 176 | |
| 177 | // "Templates must be classes or functions", "Declaration syntax error" |
| 178 | template<class T, bool v = A<T>::value> |
| 179 | struct B {}; |
| 180 | |
| 181 | int main() |
| 182 | { |
| 183 | B<int> x; |
| 184 | } |
| 185 | }}} |
| 186 | |
| 187 | '''Workaround:''' If the relevant non-type |
| 188 | template parameter is an implementation detail, use inheritance |
| 189 | and a fully qualified identifier (for example, |
| 190 | ::N::A<T>::value). |
| 191 | |
| 192 | === [function partial ordering] Partial ordering of function templates === #function-partial-ordering |
| 193 | |
| 194 | Partial ordering of function templates, as described in |
| 195 | std:14.5.5.2 [temp.func.order], does not work: |
| 196 | {{{ |
| 197 | #include <iostream> |
| 198 | |
| 199 | template<class T> struct A {}; |
| 200 | |
| 201 | template<class T1> |
| 202 | void f(const A<T1> &) |
| 203 | { |
| 204 | std::cout << "f(const A<T1>&)\n"; |
| 205 | } |
| 206 | |
| 207 | template<class T> |
| 208 | void f(T) |
| 209 | { |
| 210 | std::cout << "f(T)\n"; |
| 211 | } |
| 212 | |
| 213 | int main() |
| 214 | { |
| 215 | A<double> a; |
| 216 | f(a); // output: f(T) (wrong) |
| 217 | f(1); // output: f(T) (correct) |
| 218 | } |
| 219 | }}} |
| 220 | |
| 221 | '''Workaround:''' Declare all such functions |
| 222 | uniformly as either taking a value or a reference |
| 223 | parameter. |
| 224 | |
| 225 | === [instantiate memfun ptr] Instantiation with member function pointer === #instantiate-memfun-ptr |
| 226 | |
| 227 | When directly instantiating a template with some member |
| 228 | function pointer, which is itself dependent on some template |
| 229 | parameter, the compiler cannot cope: |
| 230 | |
| 231 | {{{ |
| 232 | template<class U> class C { }; |
| 233 | template<class T> |
| 234 | class A |
| 235 | { |
| 236 | static const int v = C<void (T::*)()>::value; |
| 237 | }; |
| 238 | }}} |
| 239 | |
| 240 | '''Workaround:''' Use an intermediate `typedef`: |
| 241 | {{{ |
| 242 | template<class U> class C { }; |
| 243 | template<class T> |
| 244 | class A |
| 245 | { |
| 246 | typedef void (T::*my_type)(); |
| 247 | static const int v = C<my_type>::value; |
| 248 | }; |
| 249 | }}} |
| 250 | |
| 251 | (Extracted from e-mail exchange of David Abrahams, Fernando |
| 252 | Cacciola, and Peter Dimov; not actually tested.) |
| 253 | |
| 254 | == Library == #library |
| 255 | |
| 256 | === [cmath.abs] Function `double std::abs(double)` missing === #cmath-abs |
| 257 | |
| 258 | The function `double std::abs(double)` should be |
| 259 | defined (std:26.5-5 [lib.c.math]), but it is not: |
| 260 | {{{ |
| 261 | #include <cmath> |
| 262 | |
| 263 | int main() |
| 264 | { |
| 265 | double (*p)(double) = std::abs; // error |
| 266 | } |
| 267 | }}} |
| 268 | |
| 269 | Note that `int std::abs(int)` will be used |
| 270 | without warning if you write `std::abs(5.1)`. |
| 271 | |
| 272 | Similar remarks apply to seemingly all of the other standard |
| 273 | math functions, where Borland C++ fails to provide |
| 274 | `float` and `long double` overloads. |
| 275 | |
| 276 | '''Workaround:''' Use `std::fabs` instead if type genericity is not required. |
| 277 | |
| 278 | == Appendix: Additional issues with Borland C++ version 5.5 == #5.5-issues |
| 279 | |
| 280 | These issues are documented mainly for historic reasons. If |
| 281 | you are still using Borland C++ version 5.5, you are strongly |
| 282 | encouraged to obtain an upgrade to version 5.5.1, which fixes |
| 283 | the issues described in this section. |
| 284 | |
| 285 | === [inline friend] Inline friend functions in template classes === #inline-friend |
| 286 | |
| 287 | If a friend function of some class has not been declared |
| 288 | before the friend function declaration, the function is |
| 289 | declared at the namespace scope surrounding the class |
| 290 | definition. Together with class templates and inline |
| 291 | definitions of friend functions, the code in the following |
| 292 | fragment should declare (and define) a non-template function |
| 293 | "bool N::f(int,int)", which is a friend of class |
| 294 | N::A<int>. However, Borland C++ v5.5 expects the function |
| 295 | f to be declared beforehand: |
| 296 | {{{ |
| 297 | namespace N { |
| 298 | template<class T> |
| 299 | class A |
| 300 | { |
| 301 | // "f is not a member of 'N' in function main()" |
| 302 | friend bool f(T x, T y) { return x < y; } |
| 303 | }; |
| 304 | } |
| 305 | |
| 306 | int main() |
| 307 | { |
| 308 | N::A<int> a; |
| 309 | } |
| 310 | }}} |
| 311 | |
| 312 | This technique is extensively used in boost/operators.hpp. |
| 313 | Giving in to the wish of the compiler doesn't work in this |
| 314 | case, because then the "instantiate one template, get lots of |
| 315 | helper functions at namespace scope" approach doesn't work |
| 316 | anymore. Defining BOOST_NO_OPERATORS_IN_NAMESPACE (a define |
| 317 | BOOST_NO_INLINE_FRIENDS_IN_CLASS_TEMPLATES would match this |
| 318 | case better) works around this problem and leads to another |
| 319 | one, see [using-template]. |