Advanced Search

Log In | Not a Member?

Contact ADC

< Previous PageNext Page > Hide TOC

Compiling for Multiple CPU Architectures

Because the Macintosh platform includes more than one processor family, it is often important to compile software for multiple processor architectures. For example, libraries should generally be compiled as universal binaries even if you are exclusively targeting an Intel-based Macintosh computer, as your library may be used by a PowerPC binary running under Rosetta. For executables, if you plan to distribute compiled versions, you should generally create universal binaries for convenience.

When compiling programs for architectures other than your default host architecture, such as compiling for a ppc64 or Intel-based Macintosh target on a PowerPC-based build host, there are a few common problems that you may run into. Most of these problems result from one of the following mistakes:

• Assuming that the build host is architecturally similar to the target architecture and will thus be capable of executing intermediate build products

• Trying to determine target-processor-specific information at configuration time (by compiling and executing small code snippets) rather than at compile time (using macro tests) or execution time (for example, by using conditional byte swap functions)

Whenever cross-compiling occurs, extra care must be taken to ensure that the target architecture is detected correctly. This is particularly an issue when generating a binary containing object code for more than one architecture.

In many cases, binaries containing object code for more than one architecture can be generated simply by running the normal configuration script, then overriding the architecture flags at compile time.

For example, you might run

./configure

followed by

make CFLAGS="-isysroot /Developer/SDKs/MacOSX10.4u.sdk -arch ppc \

    -arch i386" LDFLAGS="-syslibroot /Developer/SDKs/MacOSX10.4u.sdk \

    -arch ppc -arch i386 -arch ppc -arch i386"

to generate a universal binary (for Intel-based and PowerPC-based Macintosh computers). To generate a 4-way universal binary that includes 64-bit versions, you would add -arch ppc64 and -arch x86_64 to the CFLAGS and LDFLAGS.

However, applications that make configuration-time decisions about the size of data structures will generally fail to build correctly in such an environment (since those sizes may need to be different depending on whether the compiler is executing a ppc pass, a ppc64 pass, or an i386 pass). When this happens, the tool must be configured and compiled for each architecture as separate executables, then glued together manually using lipo.

In rare cases, software not written with cross-compilation in mind will make configure-time decisions by executing code on the build host. In these cases, you will have to manually alter either the configuration scripts or the resulting headers to be appropriate for the actual target architecture (rather than the build architecture). In some cases, this can be solved by telling the configure script that you are cross-compiling using the --host, --build, and --target flags. However, this may simply result in defaults for the target platform being inserted, which doesn’t really solve the problem.

The best fix is to replace configure-time detection of endianness, data type sizes, and so on with compile-time or run-time detection. For example, instead of testing the architecture for endianness to obtain consistent byte order in a file, you should do one of the following:

• Use C preprocessor macros like __BIG_ENDIAN__ and __LITTLE_ENDIAN__ to test endianness at compile time.

• Use functions like htonl, htons, ntohl, and ntohs to guarantee a big-endian representation on any architecture.

• Extract individual bytes by bitwise masking and shifting (for example, lowbyte=word & 0xff; nextbyte = (word >> 8) & 0xff; and so on).

Similarly, instead of performing elaborate tests to determine whether to use int or long for a 4-byte piece of data, you should simply use a standard sized type such as uint32_t.

There are a few other caveats when working with universal binaries:

• The library archive utility, ar, cannot work with libraries containing code for more than one architecture (or single-architecture libraries generated with lipo) after ranlib has added a table of contents to them. Thus, if you need to add additional object files to a library, you must keep a separate copy without a TOC.

• The -M switch to gcc (to output dependency information) is not supported when multiple architectures are specified on the command line. Depending on your makefile, this may require substantial changes to your makefile rules. For autoconf-based configure scripts, the flag --disable-dependency-tracking should solve this problem.

  For projects using automake, it may be necessary to run automake with the -i flag to disable dependency checks or put no-dependencies in the AUTOMAKE_OPTIONS variable in each Makefile.am file.

• If you run into problems building a universal binary for an open source tool, the first thing you should do is to get the latest version of the source code. This does two things:

  • Ensures that the version of autoconf and automake used to generate the configuration scripts is reasonably current, reducing the likelihood of build failures, execution failures, backwards or forwards compatibility problems, and other idiosyncratic or downright broken behavior.

  • Reduces the likelihood of building a version of an open source tool that contains known security holes or other serious bugs.

• Older versions of autoconf do not handle the case where --target, --host, and --build are not handled gracefully. Different versions also behave differently when you specify only one or two of these flags. Thus, you should always specify all three of these options if you are running an autoconf-generated configure script with intent to cross-compile.

• Some earlier versions of autoconf handle cross-compiling poorly. If your tool contains a configure script generated by an early autoconf, you may be able to significantly improve things by replacing some of the config.* files (and config.guess in particular) with updated copies from the version of autoconf that comes with Mac OS X.

  This will not always work, however, in which case it may be necessary to actually regenerate the configure script by running autoconf. To do this, simply change into the root directory of the project and run /usr/bin/autoconf. It will automatically detect and use the configure.in file and use it to generate a new configure script. If you get warnings, you should first try a web search for the error message, as someone else may have already run into the problem (possibly on a different tool) and found a solution.

  If you get errors about missing AC_ macros, you may need to download a copy of libraries on which your tool depends and copy their .m4 autoconf configuration files into /usr/share/autoconf. Alternately, you can add the macros to the file acinclude.m4 in your project’s main directory and autoconf should automatically pick up those macros.

  You may, in some cases, need to rerun automake and/or autoheader if your tool uses them. Be prepared to run into missing AM_ and AH_ macros if you do, however. Because of the added risk of missing macros, this should generally only be done if running autoconf by itself does not correct a build problem.

  Important: Be sure to make a backup copy of the original scripts, headers, and other generated files (or, ideally, the entire project directory) before running autoheader or automake.

• Different makefiles and configure scripts handle command-line overrides in different ways. The most consistent way to force these overrides is to specify them prior to the command. For example:

  make CFLAGS="-isysroot /Developer/SDKs/MacOSX10.4u.sdk -arch i386 -arch ppc"

  should generally result in the above being added to CFLAGS during compilation. However, this behavior is not completely consistent across makefiles from different projects.

For additional information about autoconf, automake, and autoheader, you can view the autoconf documentation at http://www.gnu.org/software/autoconf/manual/index.html.

For additional information on compiler flags for Intel-based Macintosh computers, modifying code to support little-endian CPUs, and other porting concerns, you should read Universal Binary Programming Guidelines, Second Edition, available from the ADC Reference Library.

Cross-Compiling a Self-Bootstrapping Tool

Probably the most difficult situation you may experience is that of a self-bootstrapping tool—a tool that uses a (possibly stripped-down) copy of itself to either compile the final version of itself or to construct support files or libraries. Some examples include TeX, Perl, and gcc.

Ideally, you should be able to build the executable as a universal binary in a single build pass. If that is possible, everything “just works”, since the universal binary can execute on the host. However, this is not always possible. If you have to cross-compile and glue the pieces together with lipo, this obviously will not work.

If the build system is written well, the tool will bootstrap itself by building a version compiled for the host, then use that to build the pieces for the target, and finally compile a version of the binary for the target. In that case, you should not have to do anything special for the build to succeed.

In some cases, however, it is not possible to simultaneously compile for multiple architectures and the build system wasn’t designed for cross-compiling. In those cases, the recommended solution is to pre-install a version of the tool for the host architecture, then modify the build scripts to rename the target’s intermediate copy of the tool and copy the host’s copy in place of that intermediate build product (for example, mv miniperl miniperl-target; cp /usr/bin/perl miniperl).

By doing this, later parts of the build script will execute the version of the tool built for the host architecture. Assuming there are no architecture dependencies in the dependent tools or support files, they should build correctly using the host’s copy of the tool. Once the dependent build is complete, you should swap back in the original target copy in the final build phase. The trick is in figuring out when to have each copy in place.

< Previous PageNext Page > Hide TOC

Did this document help you? Yes: Tell us what works for you. It’s good, but: Report typos, inaccuracies, and so forth. It wasn’t helpful: Tell us what would have helped.

	Get information on Apple products. Visit the Apple Store online or at retail locations. 1-800-MY-APPLE Copyright © 2007 Apple Inc. All rights reserved. | Terms of use | Privacy Notice