Skip to content

mackyle/blocksruntime

Repository files navigation

Blocks Runtime

This project provides a convenient way to install the BlocksRuntime library from the compiler-rt project (see http://compiler-rt.llvm.org/).

Several systems (Linux, FreeBSD, MacPorts, etc.) provide the clang compiler either preinstalled or as an available package which has Blocks support (provided the -fblocks compiler option is used).

Unfortunately, those installer packages do not provide the Blocks runtime library.

On the other hand, the compiler-rt library can be downloaded and does contain the Blocks runtime support if properly enabled in the cmake build file.

By default, however, the compiler-rt library also builds a compiler runtime support library which is undesirable.

This project contains only the BlocksRuntime files (in the BlocksRuntime subdirectory) along with tests (in the BlocksRuntime/tests subdirectory) and the original CREDITS.TXT, LICENSE.TXT and README.txt from the top-level compiler-rt project (which have been placed in the BlocksRuntime subdirectory). Note that in 2014-02 the compiler-rt project moved the BlocksRuntime sources from the BlocksRuntime directory to the lib/BlocksRuntime directory and moved the tests from the BlocksRuntime/tests directory to the test/BlocksRuntime directory. The files themselves, however, remain unchanged and are still the same as they were in 2010-08.

This runtime can also be used with the gcc-apple-4.2 compiler built using the MacPorts.org apple-gcc42 package on Mac OS X.

License

The compiler-rt project (and hence the BlocksRuntime since it's a part of that project) has a very liberal dual-license of either the UIUC or MIT license. The MIT license is fully GPL compatible (and pretty much compatible with just about everything), so there should be no problems linking the libBlocksRuntime.a library with your executable. (Note that on the FSF's site http://www.gnu.org/licenses/license-list.html, you find the MIT license under the 'X11 License' section.) See the LICENSE.TXT file in the BlocksRuntime subdirectory for all the details.

Building

Since there are only two files to build, a makefile didn't seem warranted. A special config.h file has been created to make the build work. Build the libBlocksRuntime.a library by running:

    ./buildlib

The gcc compiler will be used by default, but you can do CC=clang ./buildlib for example to use the clang compiler instead. Note that neither make nor cmake are needed (but ar and ranlib will be used but they can also be changed with the AR and RANLIB environment variables similarly to the way the compiler can be changed).

IMPORTANT Mac OS X Note: If you are building this library on Mac OS X (presumably to use with a clang or gcc-apple-4.2 built with MacPorts or otherwise obtained), you probably want a fat library with multiple architectures in it. You can do that with the CFLAGS variable like so:

    CFLAGS='-O2 -arch x86_64 -arch ppc64 -arch i386 -arch ppc' ./buildlib

The buildlib-osx script will attempt to make an intelligent guess about building an OS X library and then run buildlib. If you're using Mac OS X you can do this to build a FAT OS X library:

    ./buildlib-osx

The buildlib (and buildlib-osx) script takes a single optional -shared argument. If given it will attempt to also build a shared library instead of just a static one in case you have a policy situation where use of a static library has been forbidden.

Testing

Skip this step at your peril! It's really quite painless. You see it's possible that the underlying blocks ABI between the blocks runtime files provided in this project and the blocks ABI used by your version of the clang compiler to implement the -fblocks option have diverged in an incompatible way. If this has happened, at least some of the tests should fail in spectacular ways (i.e. bus faults). For that reason skipping this step is not recommended.

You must have the clang compiler with -fblocks support installed for this step (if you don't have a clang compiler with -fblocks support available, why bother installing the Blocks runtime in the first place?)
Run the tests like so:

    ./checktests

By default checktests expects the clang compiler to be available in the PATH and named clang. If you are using gcc-apple-4.2 or your clang is named something different (such as clang-mp-2.9 or clang-mp-3.0) run the tests like this instead (replacing clang-mp-3.0 with your compiler's name):

    CC=clang-mp-3.0 ./checktests

Problems are indicated with a line that starts not ok. You will see a few of these. The ones that are marked with # TODO are expected to fail for the reason shown. The copy-block-literal-rdar6439600.c expected failure is a real failure. No it's not a bug in the Blocks runtime library, it's actually a bug in the compiler. You may want to examine the copy-block-literal-rdar6439600.c source file to make sure you fully grok the failure so you can avoid getting burned by it in your code. There may be a fix in the clang project by now (but as of the clang 3.2 release it still seems to fail), however it may be a while until it rolls downhill to your clang package.

If you are using CC=gcc-apple-4.2, you will probably get two additional expect failure compiler bugs in the cast.c and josh.C tests. These extra failures are not failures in the blocks runtime itself, just gcc not accepting some source files that clang accepts. You can still use the libBlocksRuntime.a library just fine.

Note that if you have an earlier version of clang (anything before version 2.8 see clang -v) then clang (C support) is either incomplete or missing and the few C tests (.C extension) will be automatically skipped (if clang is missing) or possibly fail in odd ways (if clang is present but older than version 2.8).

Note that the ./checktests output is TAP (Test Anything Protocol) compliant and can therefore be run with Perl's prove utility like so:

    prove -v checktests

Optionally first setting CC like so:

    CC=gcc-apple-4.2 prove -v checktests

Omit the -v option for more succinct output.

ARM Hard Float Bug

When running on a system that uses the ARM hard float ABI (e.g. RaspberryPi), the clang compiler has a bug. When passing float arguments to a vararg function they must also be passed on the stack, not just in hardware floating point registers. The clang compiler does this correctly for normal vararg functions, but fails to do this for block vararg functions.

If you really need this, a workaround is to call a normal vararg function that takes a block and ... arguments. It can then package up the ... arguments into a va_list and then call the block it was passed as an argument passing the block the va_list. This works fine and avoids the clang bug even though it's fugly.

The checktests script marks this test (variadic.c) as expect fail when running the tests on an ARM hard float ABI system if it's able to detect that the ARM hard float ABI is in use.

clang -fblocks failure

If clang is not using the integrated assembler (option -integrated-as) then it will incorrectly pass options such as -fblocks down to the assembler which will probably not like it. One example of an error caused by this bug is:

    gcc: error: unrecognized command line option '-fblocks'

In this case clang is not using the integrated assembler (which is not supported on all platforms) and passes the -fblocks option down to the gcc assembler which does not like that option at all.

The following references talk about this:

The ugly workaround for this problem is to compile the sources using both the -S and -fblocks options to produce a .s file which can then be compiled into whatever is desired without needing to use the -fblocks option.

If checktests detects this situation it will emit a line similar to this:

    WARNING: -S required for -fblocks with clang

If this is the case, then rules to compile .c into .s and then compile .s into .o (or whatever) will be needed instead of the usual compile .c into .o (or whatever).

Note that this workaround is required to use -fblocks with the version of clang included with cygwin.

Installing

Assuming that you have built the library (with ./buildlib) and are satisfied it works (./checktests) then it can be installed with:

    sudo ./installlib

The default installation prefix is /usr/local, but can be changed to /myprefix like so:

    sudo env prefix=/myprefix ./installlib

The include file (Block.h) is installed into $prefix/include and the library (libBlocksRuntime.a) into $prefix/lib by default. (Those can also be changed by setting and exporting includedir and/or libdir in the same way prefix can be changed.)

If you want to see what will be installed without actually installing use:

    ./installlib --dry-run

Note that DESTDIR is supported by the installlib script if that's needed. Just set DESTDIR before running installlib the same way prefix can be set.

Note that if the shared library exists it will also be installed. Add a single optional -shared or -static option to install only one or the other.

Sample Code

After you have installed the Blocks runtime header and library, you can check to make sure everything's working by building the sample.c file. The instructions are at the top of the file (use head sample.c to see them) or just do this (replace clang with the name of the compiler you're using):

    clang -o sample -fblocks sample.c -lBlocksRuntime && ./sample

If the above line outputs Hello world 2 then your Blocks runtime support is correctly installed and fully usable. Have fun!

Note that if you have the problem described above in the section named "clang -fblocks failure", then you'll need to do this instead:

    clang -S -o sample.s -fblocks sample.c && \
    clang -o sample sample.s -lBlocksRuntime && ./sample

Note that it's possible to use the Blocks runtime without installing it into the system directories. You simply need to add an appropriate -I option to find the Block.h header when you compile your source(s). And a -L option to find the libBlocksRuntime.a library when you link your executable. Since libBlocksRuntime.a is a static library no special system support will be needed to run the resulting executable.

Glibc Problem

The unistd.h header from older versions of glibc has an incompatibility with the -fblocks option. See http://mackyle.github.io/blocksruntime/#glibc for a workaround.

This problem was corrected with commit 84ae135d3282dc362bed0a5c9a575319ef336884 (http://repo.or.cz/w/glibc.git/commit/84ae135d) on 2013-11-21 and first appears in glibc-2.19 released on 2014-02-07. Since ldd is part of glibc you can check to see what version of glibc you have with:

    ldd --version

Documentation

You can find information on the Blocks language extension at these URLs