Porting and Maintenance

The configure process begins the act of building libstdc++, and is started via:

configure

The configure file is a script generated (via autoconf) from the file configure.ac.

After the configure process is complete,

make all

in the build directory starts the build process. The all target comes from the Makefile file, which is generated via configure from the Makefile.in file, which is in turn generated (via automake) from the file Makefile.am.

Figure B.1. Configure and Build File Dependencies

Regenerate all generated files by using the command autoreconf at the top level of the libstdc++ source directory.

Until that glorious day when we can use AC_TRY_LINK with a cross-compiler, we have to hardcode the results of what the tests would have shown if they could be run. So we have an inflexible mess like crossconfig.m4.

Wouldn't it be nice if we could store that information in files like configure.host, which can be modified without needing to regenerate anything, and can even be tweaked without really knowing how the configury all works? Perhaps break the pieces of crossconfig.m4 out and place them in their appropriate config/{cpu,os} directory.

Alas, writing macros like "AC_DEFINE(HAVE_A_NICE_DAY)" can only be done inside files which are passed through autoconf. Files which are pure shell script can be source'd at configure time. Files which contain autoconf macros must be processed with autoconf. We could still try breaking the pieces out into "config/*/cross.m4" bits, for instance, but then we would need arguments to aclocal/autoconf to properly find them all when generating configure. I would discourage that.

Most comments should use {octothorpes, shibboleths, hash marks, pound signs, whatever} rather than "dnl". Nearly all comments in configure.ac should. Comments inside macros written in ancillary .m4 files should. About the only comments which should not use #, but use dnl instead, are comments outside our own macros in the ancillary files. The difference is that # comments show up in configure (which is most helpful for debugging), while dnl'd lines just vanish. Since the macros in ancillary files generate code which appears in odd places, their "outside" comments tend to not be useful while reading configure.

Do not use any $target* variables, such as $target_alias. The single exception is in configure.ac, for automake+dejagnu's sake.

The nice thing about acinclude.m4/aclocal.m4 is that macros aren't actually performed/called/expanded/whatever here, just loaded. So we can arrange the contents however we like. As of this writing, acinclude.m4 is arranged as follows:

    GLIBCXX_CHECK_HOST
    GLIBCXX_TOPREL_CONFIGURE
    GLIBCXX_CONFIGURE
  

All the major variable "discovery" is done here. CXX, multilibs, etc.

    fragments included from elsewhere
  

Right now, "fragments" == "the math/linkage bits".

    GLIBCXX_CHECK_COMPILER_FEATURES
    GLIBCXX_CHECK_LINKER_FEATURES
    GLIBCXX_CHECK_WCHAR_T_SUPPORT

Next come extra compiler/linker feature tests. Wide character support was placed here because I couldn't think of another place for it. It will probably get broken apart like the math tests, because we're still disabling wchars on systems which could actually support them.

    GLIBCXX_CHECK_SETRLIMIT_ancilliary
    GLIBCXX_CHECK_SETRLIMIT
    GLIBCXX_CHECK_S_ISREG_OR_S_IFREG
    GLIBCXX_CHECK_POLL
    GLIBCXX_CHECK_WRITEV

    GLIBCXX_CONFIGURE_TESTSUITE

Feature tests which only get used in one place. Here, things used only in the testsuite, plus a couple bits used in the guts of I/O.

    GLIBCXX_EXPORT_INCLUDES
    GLIBCXX_EXPORT_FLAGS
    GLIBCXX_EXPORT_INSTALL_INFO

Installation variables, multilibs, working with the rest of the compiler. Many of the critical variables used in the makefiles are set here.

    GLIBGCC_ENABLE
    GLIBCXX_ENABLE_C99
    GLIBCXX_ENABLE_CHEADERS
    GLIBCXX_ENABLE_CLOCALE
    GLIBCXX_ENABLE_CONCEPT_CHECKS
    GLIBCXX_ENABLE_CSTDIO
    GLIBCXX_ENABLE_CXX_FLAGS
    GLIBCXX_ENABLE_C_MBCHAR
    GLIBCXX_ENABLE_DEBUG
    GLIBCXX_ENABLE_DEBUG_FLAGS
    GLIBCXX_ENABLE_LONG_LONG
    GLIBCXX_ENABLE_PCH
    GLIBCXX_ENABLE_SYMVERS
    GLIBCXX_ENABLE_THREADS

All the features which can be controlled with enable/disable configure options. Note how they're alphabetized now? Keep them like that. :-)

    AC_LC_MESSAGES
    libtool bits

Things which we don't seem to use directly, but just has to be present otherwise stuff magically goes wonky.

All the GLIBCXX_ENABLE_FOO macros use a common helper, GLIBCXX_ENABLE. (You don't have to use it, but it's easy.) The helper does two things for us:

Doing these things correctly takes some extra autoconf/autom4te code, which made our macros nearly illegible. So all the ugliness is factored out into this one helper macro.

Many of the macros take an argument, passed from when they are expanded in configure.ac. The argument controls the default value of the enable/disable switch. Previously, the arguments themselves had defaults. Now they don't, because that's extra complexity with zero gain for us.

There are three "overloaded signatures". When reading the descriptions below, keep in mind that the brackets are autoconf's quotation characters, and that they will be stripped. Examples of just about everything occur in acinclude.m4, if you want to look.

    GLIBCXX_ENABLE (FEATURE, DEFAULT, HELP-ARG, HELP-STRING)
    GLIBCXX_ENABLE (FEATURE, DEFAULT, HELP-ARG, HELP-STRING, permit a|b|c)
    GLIBCXX_ENABLE (FEATURE, DEFAULT, HELP-ARG, HELP-STRING, SHELL-CODE-HANDLER)

With no other arguments, only the standard autoconf patterns are allowed: "--{enable,disable}-foo[={yes,no}]" The $enable_FEATURE variable is guaranteed to equal either "yes" or "no" after the macro. If the user tries to pass something else, an explanatory error message will be given, and configure will halt.

The second signature takes a fifth argument, "[permit a | b | c | ...]" This allows a or b or ... after the equals sign in the option, and $enable_FEATURE is guaranteed to equal one of them after the macro. Note that if you want to allow plain --enable/--disable with no "=whatever", you must include "yes" and "no" in the list of permitted values. Also note that whatever you passed as DEFAULT must be in the list. If the user tries to pass something not on the list, a semi-explanatory error message will be given, and configure will halt. Example: [permit generic|gnu|ieee_1003.1-2001|yes|no|auto]

The third signature takes a fifth argument. It is arbitrary shell code to execute if the user actually passes the enable/disable option. (If the user does not, the default is used. Duh.) No argument checking at all is done in this signature. See GLIBCXX_ENABLE_CXX_FLAGS for an example of handling, and an error message.

The libstdc++.so shared library must be carefully managed to maintain binary compatible with older versions of the library. This ensures a new version of the library is still usable by programs that were linked against an older version.

Dependent on the target supporting it, the library uses ELF symbol versioning for all exported symbols. The symbol versions are defined by a linker script that assigns a version to every symbol. The set of symbols in each version is fixed when a GCC release is made, and must not change after that.

When new symbols are added to the library they must be added to a new symbol version, which must be created the first time new symbols are added after a release. Adding a new symbol version involves the following steps:

Once the new symbol version has been added you can add the names of your new symbols in the new version node:

    GLIBCXX_3.4.23 {

      # basic_string<C, T, A>::_Alloc_hider::_Alloc_hider(C*, A&&)
      _ZNSt7__cxx1112basic_stringI[cw]St11char_traitsI[cw]ESaI[cw]EE12_Alloc_hiderC[12]EP[cw]OS3_;

    } GLIBCXX_3.4.22;

You can either use mangled names, or demangled names inside an extern "C++" block. You might find that the new symbol matches an existing pattern in an old symbol version (causing the check-abi test target to fail). If that happens then the existing pattern must be adjusted to be more specific so that it doesn't match the new symbol.

For an example of these steps, including adjusting old patterns to be less greedy, see https://gcc.gnu.org/ml/gcc-patches/2016-07/msg01926.html and the attached patch.

If it wasn't done for the last release, you might also need to regenerate the baseline_symbols.txt file that defines the set of expected symbols for old symbol versions. A new baseline file can be generated by running make new-abi-baseline in the libbuilddir/testsuite directory. Be sure to generate the baseline from a clean build using unmodified sources, or you will incorporate your local changes into the baseline file.