libbuild2-autoconf/0.4.0-a.0.20250910094552.e8631eeb8905

GNU Autoconf emulation build system module for build2.

Specifically, this module provides an in-based rule for processing config.h.in files. Besides the Autoconf special line flavor (#undef), it also supports the CMake (#cmakedefine, #cmakedefine01) and Meson (#mesondefine) variants. Note that the CMake ${VAR} style substitutions are not supported, only the @VAR@ style (see Modifying upstream source code during build for one way to deal with ${VAR}).

Similar to Autoconf, this module provides builtin support for a number of common HAVE_* configuration checks with projects being able to add custom ones. However, the values of these checks are not discovered by dynamic probing, such as trying to compile a test program to determine if the feature is present. Instead, they are set to static expected values based on the platform/compiler macro checks (see note at the beginning of Project Configuration for rationale).

See libbuild2/autoconf/checks/ for the list of available builtin checks. Submit requests for new checks as issues. Submit implementations of new checks (or any other improvements) as PRs or patches.

Using in your projects

This module is part of the standard pre-installed build2 modules and no extra integration steps are required other than the using directive in your buildfile. For example, for Autoconf config.h.in:

using autoconf

h{config}: in{config}

Or for CMake config.h.cmake:

using autoconf

h{config}: in{config.h.cmake}

The default flavor is autoconf but if the input file has the .cmake or .meson extension, then the cmake or meson flavors are selected automatically. If, however, the standard config.h.in file is re-used for CMake/Meson, then the flavor must be specified explicitly with the autoconf.flavor variable, for example:

using autoconf

h{config}: in{config}
{
  autoconf.flavor = meson
}

Besides the configuration checks, custom substitutions can be specified as buildfile variables or key-value pairs in the same way as with the in module. For example, as buildfile variables:

/* config.h.in */

#define PACKAGE_NAME @PACKAGE_NAME@
#define PACKAGE_VERSION @PACKAGE_VERSION@

#undef HAVE_STRLCPY
#undef HAVE_STRLCAT

h{config}: in{config}
{
  PACKAGE_NAME = $project
  PACKAGE_VERSION = $version
}

Or as key-value pairs in the autoconf.substitutions map (which is an alias for the in.substitutions variable; see the in module for details):

/* config.h.in */

#undef _GNU_SOURCE
#undef _POSIX_SOURCE

gnu_source = ($c.stdlib == 'glibc')
posix_source = ($c.target.class != 'windows' && !$gnu_source)

h{config}: in{config}
{
  autoconf.substitutions  = _GNU_SOURCE@$gnu_source
  autoconf.substitutions += _POSIX_SOURCE@$posix_source
}

In particular, the autoconf.substitutions mechanism is the only way to have substitutions that cannot be specified as buildfile variables because they start with an underscore (and thus are reserved, as in the above example) or refer to one of the predefined variables.

The custom substitutions can also be used to override the checks, for example:

h{config}: in{config}
{
  HAVE_STRLCPY = true
}

While this module provides widely used aliases for some checks, it doesn't attempt to cover every project's idiosyncrasies. Instead, it provides a mechanism for creating project-specific aliases for builtin checks. Specifically, the desired aliases can be specified as key-value pairs in the autoconf.aliases map with the key being the new name and the value -- old/existing. For example:

/* config.h.in */

#undef HAVE_AF_UNIX_H
#undef MY_SSIZE_T

h{config}: in{config}
{
  autoconf.aliases  = HAVE_AF_UNIX_H@HAVE_AFUNIX_H
  autoconf.aliases += MY_SSIZE_T@ssize_t
}

The checks can be prefixed in order to avoid clashes with similarly named macros in other headers. This is an especially good idea if the resulting header is public. To enable this, we specify the prefix with the autoconf.prefix variable and then use the prefixed versions of the checks in the config.h.in file. For example:

/* config.h.in */

#undef LIBFOO_HAVE_STRLCPY
#undef LIBFOO_HAVE_STRLCAT

h{config}: in{config}
{
  autoconf.prefix = LIBFOO_
}

Note that autoconf.prefix only affects the lookup of the checks in the catalogs (project and builtin). Custom substitutions and overrides of checks must include the prefix. Similarly, both names in autoconf.aliases must be specified with the prefix (unless unprefixable; see below). For example:

h{config}: in{config}
{
  autoconf.prefix = LIBFOO_

  LIBFOO_HAVE_STRLCPY = true

  autoconf.aliases = LIBFOO_SSIZE_T@ssize_t
}

Note also that some check names are unprefixable, usually because they are standard macro names (for example, BYTE_ORDER) that on some platforms come from system headers (for example, <sys/endian.h> on FreeBSD). Such checks have ! after their names on the first line of their implementation files (for example, // BYTE_ORDER!).

An implementation of a check may depend on another check. As a result, substitutions should not be conditional at the preprocessor level (unless all the checks are part of the same condition). Nor should the results of checks be adjusted until after the last check. For example:

#ifndef _WIN32
#  cmakedefine HAVE_EXPLICIT_BZERO // Conditional substitution.
#endif

#cmakedefine HAVE_EXPLICIT_MEMSET  // Shares implementation with BZERO.

#cmakedefine BYTE_ORDER

#if BYTE_ORDER == LITTLE_ENDIAN
#  undef BYTE_ORDER               // Adjusting the result.
#endif

#cmakedefine WORDS_BIGENDIAN      // Based on BYTE_ORDER.

Below is the correct way to achieve the above semantics:

#cmakedefine HAVE_EXPLICIT_BZERO
#cmakedefine HAVE_EXPLICIT_MEMSET

#cmakedefine BYTE_ORDER
#cmakedefine WORDS_BIGENDIAN

#ifdef _WIN32
#  undef HAVE_EXPLICIT_BZERO
#endif

#if BYTE_ORDER == LITTLE_ENDIAN
#  undef BYTE_ORDER
#endif

Communicating check results to buildfiles

The results of the checks are represented as preprocessor macros and are normally used in source files. However, sometimes we may need these results to be available in buildfiles. For example, we may need them to decide whether to use a certain compiler option or to exclude certain source files from the build.

The results of the checks can be communicated to buildfiles by combining two build2 mechanisms: the c.predefs rule, which allows us to extract macro values from header files, and the update directive, which allow us to update a target while loading a buildfile. Specifically, we can extract macro values from the autoconf-generated header into a buildfile fragment (or a JSON file) and then include this fragment (or load this JSON file) into our buildfile.

As an example, let's make the value of the BYTE_ORDER macro available in our buildfile in order to decide which source file to include into the build. First we prepare byte-order.h.in:

/* byte-order.h.in */

#undef BYTE_ORDER

#if BYTE_ORDER == LITTLE_ENDIAN
#  define BYTE_ORDER_LITTLE_ENDIAN true
#elif BYTE_ORDER == BIG_ENDIAN
#  define BYTE_ORDER_LITTLE_ENDIAN false
#else
#  error unexpected byte order
#endif

Then we load the c.predefs submodule after c in our root.build:

# root.build
#
using c
using c.predefs

Finally, at the beginning of our buildfile we add the following fragment:

# Detect target endianness.
#
using autoconf

h{byte-order}: in{byte-order}

[rule_hint=c.predefs] buildfile{byte-order}: h{byte-order}
{
  c.predefs.poptions = false
  c.predefs.macros = BYTE_ORDER_LITTLE_ENDIAN@little_endian
}

./: buildfile{byte-order} # Make sure it gets cleaned.

if ($build.meta_operation == 'perform')
{
  update buildfile{byte-order}
  source $path(buildfile{byte-order})
}
else
  little_endian = false

After this fragment we can use the little_endian variable to make decisions in our buildfile:

/: exe{hello}: cxx{hello}

exe{hello}: cxx{hello-big}:    include = (!$little_endian)
exe{hello}: cxx{hello-little}: include = $little_endian

Adding new checks

There are two check catalogs: builtin, which is part of the autoconf module, and project-specific. Ideally, common checks which can be used by multiple projects should be added to the builtin rather than project catalog (see "Which checks are considered common?" below for details). In particular, this makes sure that fixes and improvements only need to be applied in one place rather than in all the projects that use the check.

Practically, however, there are several valid reasons why we may want to add a project-specific check:

1. Test a common check on CI before proposing it for the builtin catalog.

2. Release a project with a common check without waiting on the autoconf release. Once the check is available as builtin, it can be dropped from the project catalog.

3. Override a broken builtin check with a fixed version. Similar to the previous case, the check can be dropped once the fix is available in the builtin version.

4. Have a project-private check that is unlikely to be useful to any other project.

Before adding a new check, verify the same checks does not already exist in the builtin catalog, potentially with a different name. Note that adding a project check because the same builtin check has a different name is not a valid reason: this case should be handled with autoconf.aliases discussed above.

Checks from the project catalog take precedence over the builtin checks. However, to help detect cases where a project check can be dropped (items 1-3 in the above list), the autoconf module issues a warning when a check with the same name exists in both catalogs. It's recommended that private checks (item 4) use names that can never clash with builtin checks since such checks could be used as bases by other builtin checks (see below for details on base checks). This can be achieved, for example, by embedding the project name in the check name (use autoconf.aliases to retain the original name in output).

To add a new configuration check <NAME> create the <NAME>.h header file (preserving the case) which will contain the check's implementation (use existing checks for inspiration).

Then, if this is a project-specific check, place it into the build/autoconf/checks/ subdirectory of your project (or build2/autoconf/checks/ if using the alternative naming scheme). And if this is a builtin check -- into libbuild2/autoconf/checks/ of libbuild2-autoconf. For the latter, see also "Stylistic guidelines for builtin checks" below.

The format and semantics of this header file are exactly the same for both project and builtin catalogs. Its first line must be in the form:

// <NAME>[!] [: <BASE>...]

If the name is followed by the ! modifier, then it is unprefixable (see the previous section for details). The name can also be followed by : and a space-separated list of base checks. Such checks are automatically inserted before the rest of the lines in the resulting substitution. One notable check that you may want to use as a base is BUILD2_AUTOCONF_LIBC_VERSION (see comments for details).

Subsequent lines should be C-style comments or preprocessor directives that #define or #undef <NAME> depending on whether the feature is available (though there can be idiosyncrasies; see const.h, for example). The file may also contain C++-style comment lines, which (along with the first line) are excluded from the output. Note also that there should be no double-quotes or backslashes except for line continuations.

For example, to add a new check HAVE_BAR, we could create the HAVE_BAR.h header file with the following content:

// HAVE_BAR

// TODO: maybe add support for Cygwin?

#undef HAVE_BAR

/* No bar on Windows except with MinGW. */
#if !defined(_WIN32) || \
     defined(__MINGW32__)
#  define HAVE_BAR 1
#endif

Note also that the module implementation may need to replace <NAME> with its prefixed version (unless it is unprefixable) if the autoconf.prefix functionality is in use (see above). This is done by textually substituting every occurrence of <NAME> that is separated on both left and right hand sides (that is, both characters immediately before and after <NAME> are not [A-Za-z0-9_]).

Within a file duplicate checks are automatically suppressed. And if multiple files are involved, then the user is expected to employ the autoconf.prefix functionality to avoid clashes across files. However, this does not help unprefixable names and, as a result, such checks should be implemented in ways that deal with duplication (for example, include guards).

The duplicate suppression is incompatible with conditional (at the preprocessor level) checks, for example, assuming both HAVE_EXPLICIT_* checks are based on BUILD2_AUTOCONF_LIBC_VERSION:

#ifndef _WIN32
#  undef HAVE_EXPLICIT_BZERO
#endif

#undef HAVE_EXPLICIT_MEMSET

In this example, the autoconf module will omit the second copy of the BUILD2_AUTOCONF_LIBC_VERSION check as part of the HAVE_EXPLICIT_MEMSET substitution because it was already inserted as part of the HAVE_EXPLICIT_BZERO substitution. But the first copy will not be preprocessed on Windows.

While there is no bulletproof way to detect such situations (because the unconditional check could be BUILD2_AUTOCONF_LIBC_VERSION itself; perhaps we should only have private bases that are only accessed by the user via derived public checks), it is a good idea for checks that are based on other checks to verify that the base macros are in fact defined, for example:

// HAVE_EXPLICIT_BZERO : BUILD2_AUTOCONF_LIBC_VERSION

#ifndef BUILD2_AUTOCONF_LIBC_VERSION
#  error BUILD2_AUTOCONF_LIBC_VERSION appears to be conditionally included
#endif

...

Which checks are considered common?

As mentioned above, common checks should preferably be added to the builtin catalog rather than being part of multiple projects. But which checks are considered common? The key criterion is the likelihood of a check being used by multiple projects. However, when adding a check in a specific single project, it may not always be obvious whether it is likely to be useful to someone else. In fact, since you found it useful, chances are someone else will as well, in some potentially distant future. But we feel the bar for inclusion into the builtin catalog should be higher than that.

If both of the following points hold then it's a strong indication the check in question is too obscure and should be rather kept in the project:

1. No other project is using the same check (search the Internet for the check name and its plausible alternative spellings).

2. The feature a check tries to detect is only available on one platform and is not likely to ever become available anywhere else. For example, a check detects presence of an idiosyncratic, OS-specific API (quite common on Mac OS) that has no counterparts on other platforms.

Stylistic guidelines for builtin checks

When writing checks for the builtin catalog we require that you follow a number of stylistic guidelines described below in addition to the correctness rules described in the "Adding new checks" section above. This helps with keeping the checks maintainable. The following example of a check illustrate many of the points discussed next:

// HAVE_STRLCPY

#undef HAVE_STRLCPY

// TODO: available in glibc since 2.38.

// NOTE: keep consistent with HAVE_STRLCAT.

/* Check for the strlcpy() function.
 *
 * Available in BSDs and Mac OS since the beginning. Not available
 * on Windows including MinGW or Linux/glibc. */
 */
#if defined(__FreeBSD__) || \
    defined(__OpenBSD__) || \
    defined(__NetBSD__)  || \
    defined(__APPLE__)
#  define HAVE_STRLCPY 1
#endif

1. Make sure the same check is not already present in the builtin catalog, potentially with a different name. To accomplish this, search for a component of a check name that is unlikely to vary (for example strlcpy) in the contents of the existing checks.

2. Use C-style comments except for information that should not be copied to config.h (for example, TODO notes).

3. Use proper sentences in comments. That is, start each sentence with a capital letter and end with a period.

4. Use canonical check name even if your project doesn't (see autoconf.aliases above).

   For headers, the canonical name is the header name with / replaced with _ and ending with _H. For example, for <sys/stat.h> the canonical check name is HAVE_SYS_STAT_H.

   For functions and macros, the canonical name is just the function/macro name. For example, for strlcpy() the canonical check name is HAVE_STRLCPY.

   For structs, the canonical name is the STRUCT_ prefix followed by the struct name. For example, for struct stat the canonical check name is HAVE_STRUCT_STAT.

   For struct data members, the canonical name is the STRUCT_ prefix followed by the struct name followed by the data member name. For example, for struct stat's st_mtim.tv_nsec data member the canonical check name is HAVE_STRUCT_STAT_ST_MTIM_TV_NSEC.

   For other entities (builtin types, instruction sets, etc) the canonical name is typically just the entity name. For example, HAVE_SSE4_2.

   Note that if the entity includes leading underscore(s), include it (them) in the name (since there could also be a version without). For example: __bswap_32() becomes HAVE___BSWAP_32.

5. Describe what the check is checking for. Examples:

   /* Check for the <sys/stat.h> header.
    *
    * ...
    */
   

   /* Check for the strlcpy() function.
    *
    * ...
    */
   

   /* Check for the stat struct.
    *
    * ...
    */
   

   /* Check for the st_mtim.tv_nsec data member in the stat struct.
    *
    *  ...
    */
   

   For more obscure entities you may want to expand of what it is about. For example:

   /* Check for the availability of the kCMVideoCodecType_VP9 constant.
    * VP9 is an open-source video codec developed by Google.
    *
    *  ...
    */
   

6. Describe in a comment on which platforms/versions the checked functionality is present. Also list explicitly if it's not present on one of the main platforms: Linux with glibc, Windows (including MinGW), and Mac OS. For example:

   /* Check for the addrinfo struct.
    *
    * Available since Linux/glibc 2.4, OpenBSD 2.9, FreeBSD 3.5, NetBSD 1.5,
    * and Mac OS (exact version is unclear but for a while now). Not available
    * on Windows including MinGW.
    */
   

   The reason for essentially describing what the check below does is two-fold: Firstly, some check implementations might be hard to decipher. For example, the OpenBSD version in the check is specified as a date, not a version. More importantly, if there is a bug in the check, the description allows one to distinguish between an incorrect implementation of a correct assumption and an incorrect assumption.

   On Windows, it is fairly common for the functionality to be not available in MSVC with vanilla PlatformSDK but available in MinGW. This should be noted. The standard Windows description lines are:

   Not available on Windows including MinGW.
   Not available on Windows except MinGW.
   Available on Windows including MinGW.
   Available on Windows except MinGW.
   

7. Avoid redundant macro checks.

   The most common offender is checking for _WIN32 or __MINGW32__:

   /* Available on Windows including MinGW. */
   #if defined(_WIN32) || defined(__MINGW32__)
   

   This is redundant since if __MINGW32__ is defined, _WIN32 is always defined as well. The correct version would be:

   /* Available on Windows including MinGW. */
   #if defined(_WIN32)
   

8. Do not use one check to implement another (except for specific "base" checks, like BUILD2_AUTOCONF_LIBC_VERSION, that are meant to be used as bases). For example, this is incorrect:

   // HAVE_STRLCAT: HAVE_STRLCPY
   
   #undef HAVE_STRLCAT
   
   /* The same as strlcat() so just define it in its terms. */
   #ifdef HAVE_STRLCPY
   #  define HAVE_STRLCAT 1
   #endif
   

   Instead, duplicate the same (for now) check in all places, potentially adding a note to keep them consistent.

9. If you are relying on another macro in your check, you should either include the header that defines it (potentially different for different platforms) or you should explicitly document in which situations (for example, compiler options) it is predefined by the compiler.

10. When submitting a pull request with a number of checks, squash them into a single commit with the subject reading "Add number of common checks" and the body listing all the added checks one per line. For example:

    Add number of common checks
    
    HAVE_ACCESS
    HAVE_FILENO
    ...