Porting to New Hardware or Operating Systems

This document explains how to port libstdc++ (the GNU C++ library) to a new target.

In order to make the GNU C++ library (libstdc++) work with a new target, you must edit some configuration files and provide some new header files. Unless this is done, libstdc++ will use generic settings which may not be correct for your target; even if they are correct, they will likely be inefficient.

Before you get started, make sure that you have a working C library on your target. The C library need not precisely comply with any particular standard, but should generally conform to the requirements imposed by the ANSI/ISO standard.

In addition, you should try to verify that the C++ compiler generally works. It is difficult to test the C++ compiler without a working library, but you should at least try some minimal test cases.

(Note that what we think of as a "target," the library refers to as a "host." The comment at the top of configure.ac explains why.)

If you are porting to a new operating system (as opposed to a new chip using an existing operating system), you will need to create a new directory in the config/os hierarchy. For example, the IRIX configuration files are all in config/os/irix. There is no set way to organize the OS configuration directory. For example, config/os/solaris/solaris-2.6 and config/os/solaris/solaris-2.7 are used as configuration directories for these two versions of Solaris. On the other hand, both Solaris 2.7 and Solaris 2.8 use the config/os/solaris/solaris-2.7 directory. The important information is that there needs to be a directory under config/os to store the files for your operating system.

You might have to change the configure.host file to ensure that your new directory is activated. Look for the switch statement that sets os_include_dir, and add a pattern to handle your operating system if the default will not suffice. The switch statement switches on only the OS portion of the standard target triplet; e.g., the solaris2.8 in sparc-sun-solaris2.8. If the new directory is named after the OS portion of the triplet (the default), then nothing needs to be changed.

The first file to create in this directory, should be called os_defines.h. This file contains basic macro definitions that are required to allow the C++ library to work with your C library.

Several libstdc++ source files unconditionally define the macro _POSIX_SOURCE. On many systems, defining this macro causes large portions of the C library header files to be eliminated at preprocessing time. Therefore, you may have to #undef this macro, or define other macros (like _LARGEFILE_SOURCE or __EXTENSIONS__). You won't know what macros to define or undefine at this point; you'll have to try compiling the library and seeing what goes wrong. If you see errors about calling functions that have not been declared, look in your C library headers to see if the functions are declared there, and then figure out what macros you need to define. You will need to add them to the CPLUSPLUS_CPP_SPEC macro in the GCC configuration file for your target. It will not work to simply define these macros in os_defines.h.

At this time, there are a few libstdc++-specific macros which may be defined:

_GLIBCXX_USE_C99_CHECK may be defined to 1 to check C99 function declarations (which are not covered by specialization below) found in system headers against versions found in the library headers derived from the standard.

_GLIBCXX_USE_C99_DYNAMIC may be defined to an expression that yields 0 if and only if the system headers are exposing proper support for C99 functions (which are not covered by specialization below). If defined, it must be 0 while bootstrapping the compiler/rebuilding the library.

_GLIBCXX_USE_C99_LONG_LONG_CHECK may be defined to 1 to check the set of C99 long long function declarations found in system headers against versions found in the library headers derived from the standard.

_GLIBCXX_USE_C99_LONG_LONG_DYNAMIC may be defined to an expression that yields 0 if and only if the system headers are exposing proper support for the set of C99 long long functions. If defined, it must be 0 while bootstrapping the compiler/rebuilding the library.

_GLIBCXX_USE_C99_FP_MACROS_DYNAMIC may be defined to an expression that yields 0 if and only if the system headers are exposing proper support for the related set of macros. If defined, it must be 0 while bootstrapping the compiler/rebuilding the library.

_GLIBCXX_USE_C99_FLOAT_TRANSCENDENTALS_CHECK may be defined to 1 to check the related set of function declarations found in system headers against versions found in the library headers derived from the standard.

_GLIBCXX_USE_C99_FLOAT_TRANSCENDENTALS_DYNAMIC may be defined to an expression that yields 0 if and only if the system headers are exposing proper support for the related set of functions. If defined, it must be 0 while bootstrapping the compiler/rebuilding the library.

Finally, you should bracket the entire file in an include-guard, like this:


#ifndef _GLIBCXX_OS_DEFINES
#define _GLIBCXX_OS_DEFINES
...
#endif

We recommend copying an existing os_defines.h to use as a starting point.

If you are porting to a new chip (as opposed to a new operating system running on an existing chip), you will need to create a new directory in the config/cpu hierarchy. Much like the Operating system setup, there are no strict rules on how to organize the CPU configuration directory, but careful naming choices will allow the configury to find your setup files without explicit help.

We recommend that for a target triplet <CPU>-<vendor>-<OS>, you name your configuration directory config/cpu/<CPU>. If you do this, the configury will find the directory by itself. Otherwise you will need to edit the configure.host file and, in the switch statement that sets cpu_include_dir, add a pattern to handle your chip.

Note that some chip families share a single configuration directory, for example, alpha, alphaev5, and alphaev6 all use the config/cpu/alpha directory, and there is an entry in the configure.host switch statement to handle this.

The cpu_include_dir sets default locations for the files controlling Thread safety and Numeric limits, if the defaults are not appropriate for your chip.

The library requires that you provide three header files to implement character classification, analogous to that provided by the C libraries <ctype.h> header. You can model these on the files provided in config/os/generic. However, these files will almost certainly need some modification.

The first file to write is ctype_base.h. This file provides some very basic information about character classification. The libstdc++ library assumes that your C library implements <ctype.h> by using a table (indexed by character code) containing integers, where each of these integers is a bit-mask indicating whether the character is upper-case, lower-case, alphabetic, etc. The ctype_base.h file gives the type of the integer, and the values of the various bit masks. You will have to peer at your own <ctype.h> to figure out how to define the values required by this file.

The ctype_base.h header file does not need include guards. It should contain a single struct definition called ctype_base. This struct should contain two type declarations, and one enumeration declaration, like this example, taken from the IRIX configuration:

  struct ctype_base
     {
       typedef unsigned int 	mask;
       typedef int* 		__to_type;

       enum
       {
	 space = _ISspace,
	 print = _ISprint,
	 cntrl = _IScntrl,
	 upper = _ISupper,
	 lower = _ISlower,
	 alpha = _ISalpha,
	 digit = _ISdigit,
	 punct = _ISpunct,
	 xdigit = _ISxdigit,
	 alnum = _ISalnum,
	 graph = _ISgraph
       };
     };

The mask type is the type of the elements in the table. If your C library uses a table to map lower-case numbers to upper-case numbers, and vice versa, you should define __to_type to be the type of the elements in that table. If you don't mind taking a minor performance penalty, or if your library doesn't implement toupper and tolower in this way, you can pick any pointer-to-integer type, but you must still define the type.

The enumeration should give definitions for all the values in the above example, using the values from your native <ctype.h>. They can be given symbolically (as above), or numerically, if you prefer. You do not have to include <ctype.h> in this header; it will always be included before ctype_base.h is included.

The next file to write is ctype_noninline.h, which also does not require include guards. This file defines a few member functions that will be included in include/bits/locale_facets.h. The first function that must be written is the ctype<char>::ctype constructor. Here is the IRIX example:

ctype<char>::ctype(const mask* __table = 0, bool __del = false,
	   size_t __refs = 0)
       : _Ctype_nois<char>(__refs), _M_del(__table != 0 && __del),
	 _M_toupper(NULL),
	 _M_tolower(NULL),
	 _M_ctable(NULL),
	 _M_table(!__table
		  ? (const mask*) (__libc_attr._ctype_tbl->_class + 1)
		  : __table)
       { }

There are two parts of this that you might choose to alter. The first, and most important, is the line involving __libc_attr. That is IRIX system-dependent code that gets the base of the table mapping character codes to attributes. You need to substitute code that obtains the address of this table on your system. If you want to use your operating system's tables to map upper-case letters to lower-case, and vice versa, you should initialize _M_toupper and _M_tolower with those tables, in similar fashion.

Now, you have to write two functions to convert from upper-case to lower-case, and vice versa. Here are the IRIX versions:

     char
     ctype<char>::do_toupper(char __c) const
     { return _toupper(__c); }

     char
     ctype<char>::do_tolower(char __c) const
     { return _tolower(__c); }

Your C library provides equivalents to IRIX's _toupper and _tolower. If you initialized _M_toupper and _M_tolower above, then you could use those tables instead.

Finally, you have to provide two utility functions that convert strings of characters. The versions provided here will always work - but you could use specialized routines for greater performance if you have machinery to do that on your system:

     const char*
     ctype<char>::do_toupper(char* __low, const char* __high) const
     {
       while (__low < __high)
	 {
	   *__low = do_toupper(*__low);
	   ++__low;
	 }
       return __high;
     }

     const char*
     ctype<char>::do_tolower(char* __low, const char* __high) const
     {
       while (__low < __high)
	 {
	   *__low = do_tolower(*__low);
	   ++__low;
	 }
       return __high;
     }

You must also provide the ctype_inline.h file, which contains a few more functions. On most systems, you can just copy config/os/generic/ctype_inline.h and use it on your system.

In detail, the functions provided test characters for particular properties; they are analogous to the functions like isalpha and islower provided by the C library.

The first function is implemented like this on IRIX:

     bool
     ctype<char>::
     is(mask __m, char __c) const throw()
     { return (_M_table)[(unsigned char)(__c)] & __m; }

The _M_table is the table passed in above, in the constructor. This is the table that contains the bitmasks for each character. The implementation here should work on all systems.

The next function is:

     const char*
     ctype<char>::
     is(const char* __low, const char* __high, mask* __vec) const throw()
     {
       while (__low < __high)
	 *__vec++ = (_M_table)[(unsigned char)(*__low++)];
       return __high;
     }

This function is similar; it copies the masks for all the characters from __low up until __high into the vector given by __vec.

The last two functions again are entirely generic:

     const char*
     ctype<char>::
     scan_is(mask __m, const char* __low, const char* __high) const throw()
     {
       while (__low < __high && !this->is(__m, *__low))
	 ++__low;
       return __low;
     }

     const char*
     ctype<char>::
     scan_not(mask __m, const char* __low, const char* __high) const throw()
     {
       while (__low < __high && this->is(__m, *__low))
	 ++__low;
       return __low;
     }

The C++ library string functionality requires a couple of atomic operations to provide thread-safety. If you don't take any special action, the library will use stub versions of these functions that are not thread-safe. They will work fine, unless your applications are multi-threaded.

If you want to provide custom, safe, versions of these functions, there are two distinct approaches. One is to provide a version for your CPU, using assembly language constructs. The other is to use the thread-safety primitives in your operating system. In either case, you make a file called atomicity.h, and the variable ATOMICITYH must point to this file.

If you are using the assembly-language approach, put this code in config/cpu/<chip>/atomicity.h, where chip is the name of your processor (see CPU). No additional changes are necessary to locate the file in this case; ATOMICITYH will be set by default.

If you are using the operating system thread-safety primitives approach, you can also put this code in the same CPU directory, in which case no more work is needed to locate the file. For examples of this approach, see the atomicity.h file for IRIX or IA64.

Alternatively, if the primitives are more closely related to the OS than they are to the CPU, you can put the atomicity.h file in the Operating system directory instead. In this case, you must edit configure.host, and in the switch statement that handles operating systems, override the ATOMICITYH variable to point to the appropriate os_include_dir. For examples of this approach, see the atomicity.h file for AIX.

With those bits out of the way, you have to actually write atomicity.h itself. This file should be wrapped in an include guard named _GLIBCXX_ATOMICITY_H. It should define one type, and two functions.

The type is _Atomic_word. Here is the version used on IRIX:

typedef long _Atomic_word;

This type must be a signed integral type supporting atomic operations. If you're using the OS approach, use the same type used by your system's primitives. Otherwise, use the type for which your CPU provides atomic primitives.

Then, you must provide two functions. The bodies of these functions must be equivalent to those provided here, but using atomic operations:

     static inline _Atomic_word
     __attribute__ ((__unused__))
     __exchange_and_add (_Atomic_word* __mem, int __val)
     {
       _Atomic_word __result = *__mem;
       *__mem += __val;
       return __result;
     }

     static inline void
     __attribute__ ((__unused__))
     __atomic_add (_Atomic_word* __mem, int __val)
     {
       *__mem += __val;
     }

The C++ library requires information about the fundamental data types, such as the minimum and maximum representable values of each type. You can define each of these values individually, but it is usually easiest just to indicate how many bits are used in each of the data types and let the library do the rest. For information about the macros to define, see the top of include/bits/std_limits.h.

If you need to define any macros, you can do so in os_defines.h. However, if all operating systems for your CPU are likely to use the same values, you can provide a CPU-specific file instead so that you do not have to provide the same definitions for each operating system. To take that approach, create a new file called cpu_limits.h in your CPU configuration directory (see CPU).

The C++ library is compiled, archived and linked with libtool. Explaining the full workings of libtool is beyond the scope of this document, but there are a few, particular bits that are necessary for porting.

Some parts of the libstdc++ library are compiled with the libtool --tags CXX option (the C++ definitions for libtool). Therefore, ltcf-cxx.sh in the top-level directory needs to have the correct logic to compile and archive objects equivalent to the C version of libtool, ltcf-c.sh. Some libtool targets have definitions for C but not for C++, or C++ definitions which have not been kept up to date.

The C++ run-time library contains initialization code that needs to be run as the library is loaded. Often, that requires linking in special object files when the C++ library is built as a shared library, or taking other system-specific actions.

The libstdc++ library is linked with the C version of libtool, even though it is a C++ library. Therefore, the C version of libtool needs to ensure that the run-time library initializers are run. The usual way to do this is to build the library using gcc -shared.

If you need to change how the library is linked, look at ltcf-c.sh in the top-level directory. Find the switch statement that sets archive_cmds. Here, adjust the setting for your operating system.