Here's what's to be done for maintaining LLVM.

Apart from the target-specific configuration machinery, there shouldn't be any major differences within LLVM between the GNU/Hurd and GNU/Linux ports, for example. Especially all the compiler magic is all the same.

General information

Rebuild of the Debian archive with clang

From id:"":,



Last reviewed up to the Git mirror's sources: llvm 6f7becfe23f38e8c28e9422d883263974058aeed (2013-03-24), clang 495cfa46300979642acde8d93a1f21c9291dac98 (2013-03-23), compiler-rt a06fe9183fbffb78798a444da9bc3040fdd444aa (2013-03-23), test-suite 5a05572d268568fb15b48f374f9fc9f882adecd2 (2013-03-23).


  • [low] Some more PATH_MAX/MAXPATHLEN things.

  • gs: usage

    • lib/Target/X86/, test/CodeGen/X86/, test/MC/X86/.

    • tools/clang/

      tools/clang/docs/LanguageExtensions.rst:          movl    %gs:(%eax), %eax
      tools/clang/test/Sema/asm.c:  asm volatile ("movb %%gs:%P2,%b0" : "=q"(b) : "0"(0), "i"(5L));
    • compiler-rt and test-suite not yet examined.

  • [low] Linuxisms

    • In some CMake files (CMakeLists.txt, for example).

    • test/, unittests, tools/clang/test/, tools/clang/unittests/ not yet examined.

      • In clang's test/Driver/ a lot of testing can be generalized from Linux to GNU environment, adding GNU/Hurd testing, too.
    • LLVM commit 98fbe27ac8f0766ea94b89b8c03418131b72bea4 Support for HiPE-compatible code emission

      Only relevant for CallingConv::HiPE.

      +  assert(ST->isTargetLinux() &&
      +         "HiPE prologue is only supported on Linux operating systems.");
    • compiler-rt and test-suite not yet examined.

  • /proc usage

    • compiler-rt and test-suite not yet examined.
  • libc++ (not currently using)

    Some Hurd-porting work is said to have been done as Google Summer of Code 2012 Debian project,

  • sanitizers

    A lot of Linux-specific things.

  • IRC, OFTC, #debian-hurd, 2013-09-05:

    <gg0> how can this fix it on {kf,hurd}-i386?
    <pinotree> what makes you think it does?
    <pinotree> it fixes #714890, which has nothing to do with hurd or
    <gg0> i simple wouldn't add a patch that fixes it on one i386 arch
      only, being aware there are others
    <pinotree> meet sylvestre
  • IRC, freenode, #hurd, 2013-09-05:

    <pinotree> tschwinge: iirc you were working on llvm/clang, weren't you?
    <tschwinge> pinotree: That's right.  I have patches to
      follow-up/rework.  Stalled at the moment, as you probably already
      guessed...  %-)
    <pinotree> tschwinge: <Sylvestre> by the way, pinotree if you have time
      for hurd stuff, I would be glad to have your help to port
      llvm-toolchain-3.3 to hurd. I am having some issues with threading
    <pinotree> he's the debian packager of llvm
    <tschwinge> That paste is for LLDB -- which I'd not assume to be in a
      shape usable for Hurd.
    <tschwinge> (I didn't look at it at all.)
    <pinotree> tschwinge: if you look at the latest llvm-toolchain-3.3
      debian source, there's a lldb-hurd.diff patch, which starts some
      include header dance


Here's a log of a LLVM build run; this is from our Git repository's sources, llvm dc218fb6ae3241f4b66e9bf2c9d6352efecc0a14 (2013-03-24), clang 744290b5ecd48bddb4a6cf96d68cdc4a57e24e36 (2013-03-24), compiler-rt a06fe9183fbffb78798a444da9bc3040fdd444aa (2013-03-23), test-suite 1821ab0ef1c73430705356fdfde3769460092382 (2013-03-24), run on kepler.SCHWINGE and coulomb.SCHWINGE.

$ export LC_ALL=C
$ (cd ../Horace_Silver/ && ln -sfn ../../../clang/tschwinge/Hank_Mobley tools/clang)
$ (cd ../Horace_Silver/ && ln -sfn ../../../compiler-rt/tschwinge/Doug_Watkins projects/compiler-rt)
$ (cd ../Horace_Silver/ && ln -sfn ../../../test-suite/tschwinge/Art_Blakey projects/test-suite)
$ ../Horace_Silver/configure --prefix="$PWD".install --enable-optimized SHELL=/bin/dash CC=gcc-4.7 CXX=g++-4.7 2>&1 | tee log_build
$ make VERBOSE=1 2>&1 | tee log_build_

Different hosts may default to different shells and compiler versions; thus harmonized.

Passing --enable-debug-symbols, the GNU/Hurd build on coulomb.SCHWINGE terminates with a SIGBUS when linking Release+Debug+Asserts/bin/clang (which is bigger than 500 MiB for the corresponding GNU/Linux build). Using --enable-debug-symbols --enable-shared, the GNU/Hurd build hang after make[1]: Leaving directory [...]/tools/llvm-shlib, after (successfully) linking Release+Debug+Asserts/lib/ (which is bigger than 250 MiB for the corresponding GNU/Linux build). Also there is a separate --enable-debug-runtime.

This takes up around 3.2 GiB, and needs roughly 1.5 h on kepler.SCHWINGE and 5.25 h on coulomb.SCHWINGE.

Configuring without --enable-optimized even crashes mighty darnassus, probably because of too-big files when linking. Configuring with --enable-optimized --enable-expensive-checks --disable-threads --enable-debug-symbols --enable-debug-runtime is fine.


$ toolchain/logs/process llvm build

    -checking for mmap of files... yes
    +checking for mmap of files... no
     checking if /dev/zero is needed for mmap... no
    +configure: WARNING: mmap() of files required but not found

Due to mmap write-only.

    +In file included from [...]/lib/Support/Process.cpp:85:0:
    +[...]/lib/Support/Unix/ In function 'unsigned int getColumns(int)':
    +[...]/lib/Support/Unix/ warning: enumeral and non-enumeral type in conditional expression [enabled by default]

       225    // Try to determine the width of the terminal.
       226    struct winsize ws;
       227    if (ioctl(FileID, TIOCGWINSZ, &ws) == 0)




$ make VERBOSE=1 install 2>&1 | tee log_install

This takes up around 400 MiB, and needs roughly 1 min on kepler.SCHWINGE and 12 min on coulomb.SCHWINGE.


$ toolchain/logs/process llvm install



$ make -k VERBOSE=1 LIT_ARGS='-v --threads=1' check-all 2>&1 | tee log_test_check-all
$ make -k -C projects/test-suite/ 2>&1 | tee log_test_test-suite

LIT_ARGS=-v is default for VERBOSE=1, but we want only one worker thread, for stable order and usable test output log.

This needs roughly 10 min (check-all) + 150 min (test-suite) = 160 min on kepler.SCHWINGE and 45 min (check-all) + 165 min (test-suite) = 210 min on coulomb.SCHWINGE.


$ toolchain/logs/process llvm test
  • LLVM :: CodeGen/X86/mult-alt-generic-i686.ll

    This one, as well as a really large set of test from the test-suite fail on coulomb.SCHWINGE no matter whether a GNU/Hurd or GNU/Linux system is booted -- so all these are specific to the Athlon XP processor, hopefully.

  • Clang :: Index/crash-recovery-modules.m

    Also fails on GNU/Linux. Tested --enable-optimized --enable-expensive-checks --disable-threads --enable-debug-symbols --enable-debug-runtime. LLVM bug #11974.

  • Clang :: Misc/dev-fd-fs.c

    $ cat < [...]/test/Misc/dev-fd-fs.c | Release+Debug+Asserts+Checks/bin/clang -x c /dev/fd/0 -E
    clang: error: no such file or directory: '/dev/fd/0'
    clang: error: no input files

    Compare to:

    $ cat < [...]/test/Misc/dev-fd-fs.c | gcc -x c /dev/fd/0 -E
    gcc: error: /dev/fd/0: (ipc/mig) bad request message ID
    gcc: warning: '-x c' after last input file has no effect
    gcc: fatal error: no input files
    compilation terminated.

    These work:

    $ Release+Debug+Asserts+Checks/bin/clang -x c /dev/fd/0 -E < [...]/test/Misc/dev-fd-fs.c
    int x;
    $ gcc -x c /dev/fd/0 -E < [...]/test/Misc/dev-fd-fs.c
    int x;
    #include <stdio.h>
    #include <unistd.h>
    int main(int argc, char *argv[])
      while (argc > 0)
          int err;
          char *f = argv[argc -1];
      err = access(f, F_OK);
      if (err &lt; 0)
        printf("%s: %m\n", f);
    return 0; } $ ./a.out /dev/fd/0 < /dev/null $ cat < /dev/null | ./a.out /dev/fd/0 /dev/fd/0: (ipc/mig) bad request message ID

    file_check_access fails with MIG_BAD_ID, meaning this RPC is not implemented.

    The difference is that the former directly refers to the /dev/null instance, whereas the latter goes through an intermediate pflocal instance.


    $ stat /dev/fd/0 < /dev/null
      File: `/dev/fd/0'
      Size: 0               Blocks: 0          IO Block: 1048576 character special file
    Device: 17h/23d Inode: 342820      Links: 1     Device type: 0,0
    Access: (0666/crw-rw-rw-)  Uid: (    0/    root)   Gid: (    0/    root)
    Access: 2012-11-27 16:03:19.000000000 +0100
    Modify: 2012-11-27 16:03:19.000000000 +0100
    Change: 2012-11-27 16:03:19.000000000 +0100
     Birth: -
    $ cat < /dev/null | stat /dev/fd/0
      File: `/dev/fd/0'
      Size: 0               Blocks: 0          IO Block: 65536  fifo
    Device: 9h/9d   Inode: 0           Links: 0
    Access: (0000/p---------)  Uid: (    0/    root)   Gid: (    0/    root)
    Access: 1970-01-01 01:00:00.000000000 +0100
    Modify: 1970-01-01 01:00:00.000000000 +0100
    Change: 1970-01-01 01:00:00.000000000 +0100
     Birth: -

    io_stat_request fills in these values.

  • Clang :: Tooling/clang-check-builtin-headers.cpp

    Fails: fatal error: 'stddef.h' file not found; succeeds when ran manually.

  • With --enable-optimized --enable-expensive-checks --disable-threads --enable-debug-symbols --enable-debug-runtime, there are a few new FAILs for both GNU/Linux and GNU/Hurd:

    • Clang :: Tooling/auto-detect-from-source-parent-of-cwd.cpp

    • Clang :: Tooling/auto-detect-from-source-parent.cpp

    • Clang :: Tooling/clang-check-autodetect-dir.cpp

    For all three, the clang-check invocation fails. LLVM bug #15194.

  • Several tests are not considered on GNU/Hurd.

    -PASS: Clang-Unit :: ASTMatchers/[...]/tschwinge/
    -PASS: Clang-Unit :: ASTMatchers/[...]/tschwinge/
    -PASS: Clang-Unit :: ASTMatchers/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: Support/[...]/tschwinge/
    -PASS: LLVM-Unit :: Support/[...]/tschwinge/
    GTEST_HAS_DEATH_TEST utils/unittest/googletest/include/gtest/internal/gtest-port.h
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    -PASS: LLVM-Unit :: ADT/[...]/tschwinge/
    EXPECT_DEBUG_DEATH utils/unittest/googletest/include/gtest/gtest-death-test.h
  • Differences in test-suite, that are not evidently floating-point issues, GNU/Linux vs. GNU/Hurd on coulomb.SCHWINGE:

     gcc-4.6 -I/media/erich/home/thomas/tmp/llvm/tschwinge/ -I/media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/SingleSource/UnitTests -I/home/thomas/tmp/llvm/tschwinge/ -I../../include -I/home/thomas/tmp/llvm/tschwinge/ -I/home/thomas/tmp/llvm/tschwinge/Horace_Silver/include -D_GNU_SOURCE -D__STDC_LIMIT_MACROS -DNDEBUG  -O3  /media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/SingleSource/UnitTests/2007-04-25-weak.c -lm -o Output/2007-04-25-weak.native  -lstdc++ 
    +/media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/SingleSource/UnitTests/2007-04-25-weak.c:3:1: warning: 'weak_import' attribute directive ignored [-Wattributes]
    +/tmp/ccWGwKvo.o: In function `main':
    +2007-04-25-weak.c:(.text.startup+0x7): undefined reference to `test_weak'
    +collect2: ld returned 1 exit status
    +make[2]: [Output/2007-04-25-weak.native] Error 1 (ignored)

    On GNU/Linux, the clamav tests are compiled with -DC_LINUX.

    +/media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/MultiSource/Applications/lambda-0.1.3/ error: use of undeclared identifier 'MAXPATHLEN'
    +                char buf[MAXPATHLEN+1];

    ..., with follow-up failures.

    projects/test-suite/MultiSource/Applications/obsequi is not at all considered for GNU/Hurd.

    +/media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/MultiSource/Benchmarks/Olden/voronoi/newvor.c:178:25: warning: implicit declaration of function 'memalign' is invalid in C99 [-Wimplicit-function-declaration]
    +    char* base = (char*)memalign(align_size, alloc_size);
    +                        ^
    +1 warning generated.
    +/media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/MultiSource/Benchmarks/Prolangs-C/archie-client/get_vdir.c:213:20: error: use of undeclared identifier 'MAXPATHLEN'
    +                    char        l_name[MAX_DIR_LINESIZE];
    +                                       ^
    +/media/erich/home/thomas/tmp/llvm/test-suite/tschwinge/Art_Blakey/MultiSource/Benchmarks/Prolangs-C/archie-client/pprot.h:39:37: note: expanded from macro 'MAX_DIR_LINESIZE'
    +#define        MAX_DIR_LINESIZE 160+MAXPATHLEN /* Max linesize in directory */
    +                                    ^

    ..., and several more.