path: root/open_issues/gcc.mdwn

[[!meta copyright="Copyright © 2007, 2008, 2009, 2010, 2011, 2012 Free Software
Foundation, Inc."]]

[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]

[[!tag open_issue_gcc]]

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

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

[[!toc levels=2]]


# [[General information|/gcc]]


# [[Sources|source_repositories/gcc]]


# Configuration

<!--

git checkout reviewed
git log --reverse --pretty=fuller --stat=$COLUMNS,$COLUMNS -p -C --cc ..upstream/trunk
-i
/^commit |^---$|hurd|linux|nptl|glibc

-->

Last reviewed up to the [[Git mirror's dfed30bca14de84e0446cc02f5a27407dbfdc3e1
(2012-06-11) sources|source_repositories/gcc]].

<http://gcc.gnu.org/install/configure.html> has documentation for the
`configure` switches.

  * Configure fragments that have `*linux*` cases might/should often contain
    those for us (and GNU/k*BSD) as well.

      * `configure.ac`

      * `libgomp/configure.tgt`

      * `libstdc++-v3/configure.host`

        `abi_baseline_pair` etc. setting.

      * `libstdc++-v3/config/os/gnu-linux/*`

        Is used for all GNU systems, as per `libstdc++-v3/configure.host`.
        Should rename to `gnu-user` to reflect this?

      * `gcc/acinclude.m4`:`gcc_GAS_FLAGS`: always pass `--32` to assembler for
        x86 Linux.  (Why?)

  * `hurd/usr`

    `NATIVE_SYSTEM_HEADER_DIR`, `638454a19c1c08f01c10517bc72a114250fc4f33`,
    [[!message-id "mcrzkhcbftp.fsf@coign.corp.google.com"]].

    Debian.

      * Eventually: get rid of this special-casing.  [[!message-id
        "gckk1s$e0b$1@ger.gmane.org"]].

  * [[`libmudflap`|libmudflap]].

  * Might [`-fsplit-stack`](http://nickclifton.livejournal.com/6889.html) be
    worthwhile w.r.t. our [[multithreaded|multithreading]] libraries?

      * Also see `libgcc/config/i386/morestack.S`: comments w.r.t
        `TARGET_THREAD_SPLIT_STACK_OFFSET`; likely needs porting.

    As per `libgcc/config/i386/t-stack-i386`, the former file is only used for
    `-fsplit-stack` support -- which is currently enabled for us in
    `libgcc/config.host`, but not usable via GCC proper.

      * `gcc/config/gnu-user.h` defines `*SPLIT_STACK*` macros -- which aren't
        valid for us (yet), I think.

  * `--enable-languages=[...]`

      * GNAT is not yet ported / bootstrapped?

      * The Google Go's libgo (introduced in
        e440a3286bc89368b8d3a8fd6accd47191790bf2 (2010-12-03)) needs
        OS configuration / support.

  * `--enable-frame-pointer`

    `gcc/configure.ac`: `enable_frame_pointer=no`

  * `--with-dwarf2`?

  * `--enable-werror`

  * `--enable-checking`

  * `--enable-linker-build-id`

  * `--enable-gnu-unique-object`

  * `--enable-lto`, `--enable-gold`

    [[binutils_gold]]

  * `--enable-indirect-function`

    [[IFUNC]]

  * <http://gcc.gnu.org/ml/gcc/2007-11/msg00289.html>,
    <http://gcc.gnu.org/ml/gcc-patches/2010-12/msg00672.html>

      * `gcc/config/t-linux` should be named `gcc/config/t-gnu-user` or
        similar.  Likewise for `gcc/config/i386/t-linux`.

  * Debian's GCC package has Hurd-specific patches.  Some have been forwarded
    upstream (and have been ignored).  [[Thomas_Schwinge|tschwinge]] is working
    on getting them integrated.

  * [\[meta-bug\] bootstrap bugs for
    \*-gnu\*](http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21824)

  * [build system: gcc\_cv\_libc\_provides\_ssp and
    NATIVE\_SYSTEM\_HEADER\_DIR](http://gcc.gnu.org/ml/gcc/2008-10/msg00130.html)

  * [-fstack-protector shouldn't use TLS in freestanding
    mode](http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29838)

  * Check before/after Joseph changes.  (Should be fine.)

  * 34618b3190c110b8926cc2b1db4b4eac95451995

    What's this used for?  (Check ML.)  Ask to include i686-pc-gnu (once it is
    buildable out of the box)?  See also
    73905b5de0d9a086f22ded7638bb1c0ae1b91326.

  * Various testsuite bits should include `*-*-gnu*`, too.

  * [low] [[toolchain/cross-gnu]] toolchain bootstrap vs. `fenv.h` in libgcc's
    libbid:

        [...]/xgcc [...] -DIN_LIBGCC2 -fbuilding-libgcc [...] -Dinhibit_libc [...] -o bid_decimal_globals.o [...] -c [...]/libgcc/config/libbid/bid_decimal_globals.c
        [...]/libgcc/config/libbid/bid_decimal_globals.c:47:18: fatal error: fenv.h: No such file or directory
        compilation terminated.
        make[1]: *** [bid_decimal_globals.o] Error 1
        make[1]: Leaving directory `/media/boole-data/thomas/tmp/gnu-0/src/gcc.obj/i686-pc-gnu/libgcc'
        make: *** [all-target-libgcc] Error 2

    See threads at [[!message-id
    "AANLkTinY1Cd4_qO_9euYJN8zev4hdr7_ANpjNG+yGRMn@mail.gmail.com"]],
    [[!message-id "20110328225532.GE5293@synopsys.com"]], [[!message-id
    "4D52D522.1040804@gmail.com"]].  Can simply configure the first GCC with
    `--disable-decimal-float`.

    Alternatively, can we use `#ifndef inhibit_libc` for this (these?) file(s)?
    See `generic-nonstrack.c`, for example.  The latter (and also
    `generic-morestack-thread.c`) also has a nice explanation of `inhibit_libc`
    which could be centralized at one place, for example definition of
    `inhibit_libc`.

  * [low] [[toolchain/cross-gnu]]

        The directory that should contain system headers does not exist:
          /media/boole-data/thomas/tmp/gnu-0/sys_root/usr/include
        make[2]: *** [stmp-fixinc] Error 1
        make[2]: Leaving directory `/media/boole-data/thomas/tmp/gnu-0/src/gcc.obj/gcc'
        make[1]: *** [all-gcc] Error 2
        make[1]: Leaving directory `/media/boole-data/thomas/tmp/gnu-0/src/gcc.obj'

    `mkdir` the directory for now, but what is really going on?  GCC has *use
    `/usr/include` patch*, but glibc still installs into `/include/`?

  * `__GLIBC__`

    IRC, freenode, #hurd, 2012-01-05:

        <civodul> on GNU/kFreeBSD, it's GCC that defines __GLIBC__, funny
        <youpi> ??
        <youpi> not from features.h ?
        <civodul> in gcc/config/kfreebsd-gnu.h
        <civodul> :-)
        <pinotree> correct, it's enabled in gcc's config
        <pinotree> i discovered that after banging my head on the wall trying
          to find out why some stuff wasn't compiling even after kfreebsd
          porting patches adding preprocessors checks for __GLIBC__

    IRC, freenode, #hurd, 2012-05-25:

        <gnu_srs> Hi, looks like __GLIBC__ is not defined by default for GNU?
        <gnu_srs> touch foo.h; cpp -dM foo.h|grep LIBC: empty
        <braunr> gnu_srs: well, this only tells your the compiler defaults
        <tschwinge> gnu_srs: See the email I just sent.

    [[!message-id "87396od3ej.fsf@schwinge.name"]]

        <braunr> __GLIBC__ would probably be introduced by a glibc header
        <gnu_srs> tschwinge: I saw your email. I wonder if features.h is
          included in the kFreeBSD build of webkit.
        <gnu_srs> It is defined in their build, but not in the Hurd build.
        <pinotree> gcc on kfreebsd unconditionally defines __GLIBC__
        <pinotree> (a bit stupid choice imho, but hardly something that could
          be changed now...)
        <braunr> :/
        <braunr> personally i don't consider this only "a bit" stupid, as
          kfreebsd is one of the various efforts pushing towards portability
        <braunr> and using such hacks actually hinders portability ...
        <pinotree> yeah don't tell me, i can remember at least half dozen of
          occasions when a code wouldn't have been compiling at all on other
          glibc platforms otherwise
        <pinotree> sure, i have nothing against kfreebsd's efforts, but making
          gcc define something which is proper of the libc used is stupid
        <braunr> it is
        <pinotree> i spotted changes like:
        <pinotree> -#ifdef __linux
        <pinotree> +#if defined(__linux__) || defined(__GLIBC__)
        <pinotree> and wondered why they wouldn't work at all for us... and
          then realized there were no #include in that file before that
          preprocessor check
        <tschwinge> This is even in upstream GCC gcc/config/kfreebsd-gnu.h:
        <tschwinge> #define GNU_USER_TARGET_OS_CPP_BUILTINS()               \
        <tschwinge>   do                                            \
        <tschwinge>     {                                           \
        <tschwinge>         builtin_define ("__FreeBSD_kernel__");  \
        <tschwinge>         builtin_define ("__GLIBC__");           \
        <tschwinge>         builtin_define_std ("unix");            \
        <tschwinge>         builtin_assert ("system=unix");         \
        <tschwinge>         builtin_assert ("system=posix");        \
        <tschwinge>     }                                           \
        <tschwinge>   while (0)
        <tschwinge> I might raise this upstream at some point.
        <pinotree> tschwinge: i could guess the change was proposed by the
          kfreebsd people, so asking them before at d-bsd@d.o would be a start
        <tschwinge> pinotree: Ack.
        <pinotree> especially that they would need to fix stuff afterwards
        <pinotree> imho we could propose them the change, and if they agree put
          that as local patch to debian's gcc4.6/.7 after wheezy, so there is
          plenty of time for them to fix stuff
        <pinotree> what should be done first is, however, find out why that
          define has been added to gcc

  * [low] Does `-mcpu=native` etc. work?  (For example,
    2ae1f0cc764e998bfc684d662aba0497e8723e52.)

  * transactional memory, 4c0315d05fa0f707875686abc4f91f7a979a7c7b

      * `config/mmap.m4`

      * In `libitm/config/`, is the generic stuff (`tls.h`, etc.) enough for
        us?

      * f29a2041f32773464e226a83f41762c2e9cf658e
        (e53a96c2136f7cdff4699475fea41afeed9dece3)

    Testresults same as for GNU/Linux.

  * [high] 3efc00f6f17778172d3fa7ac737fa1473b3b4d5a, `Check __GLIBC__ when
    using __SIGRTMIN`.  GCC PR52390.  Fixed by
    8d2259c83f94c082ad8a00b5d00bb639ce24efce.

  * 15ac1e637ad0cb92bf7629205c617ea847a4b810 `Build 64-bit libffi multilib for
    i?86-linux`.

  * `libstdc++`: uses `_GLIBCXX_HAVE_TLS`, but where is this defined?  Supposed
    to come from `config/tls.m4:GCC_CHECK_TLS`?

  * `libgcc/gthr-posix.h:__gthread_active_p` -- is this suitable for us?  This
    is used in libgcc for ObjC wrapper stuff and similar in libstdc++.
    C.f. [[!message-id "x57jobtqx89w.fsf@frobland.mtv.corp.google.com"]],
    [[!message-id "x57jd359fkx3.fsf@frobland.mtv.corp.google.com"]] as well as
    [[!debbug 629866]]/[[!message-id
    "20110609002620.GA16719@const.famille.thibault.fr"]].


# Build

Here's a log of a GCC build run; this is from our [[Git repository's
2e2db3f92b534460c68c2f9ae64455884424beb6 (2012-06-15; 2012-06-06)
sources|source_repositories/gcc]], run on kepler.SCHWINGE and coulomb.SCHWINGE.

    $ export LC_ALL=C
    $ (cd ../master/ && contrib/gcc_update --touch)
    $ ../master/configure --prefix="$PWD".install SHELL=/bin/dash CC=gcc-4.6 CXX=g++-4.6 --enable-build-with-cxx --enable-languages=all,ada 2>&1 | tee log_build
    [...]
    $ make 2>&1 | tee log_build_
    [...]

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

This takes up around 3.1 GiB, and needs roughly 3.0 h on kepler.SCHWINGE and
12.75 h on coulomb.SCHWINGE.

<!--

    $ (make && touch .go-install) 2>&1 | tee log_build_ && test -f .go-install && (make install && touch .go-check) 2>&1 | tee log_install && test -f .go-check && make -k RUNTESTFLAGS=-v check 2>&1 | tee log_check

-->


## Analysis

    $ toolchain/logs/process gcc build

  * [[`checking if gcc static flag -static
    works... no`|glibc_madvise_vs_static_linking]]

    Addressed in Debian glibc.

  * DFP

    Addressed in *hurd/decimal_floating_point* branch.

        +configure: WARNING: decimal float is not supported for this target, ignored

    ... and later on:

        -checking for decimal floating point... bid
        +checking for decimal floating point... configure: WARNING: decimal float is not supported for this target, ignored
        +dpd

    ... and later on:

        -checking whether decimal floating point is supported... yes
        +checking whether decimal floating point is supported... no
        +configure: WARNING: decimal float is not supported for this target, ignored

      * `libstdc++-v3/acinclude.m4`: ISO/IEC TR 24733

            -checking for ISO/IEC TR 24733 ... yes
            +checking for ISO/IEC TR 24733 ... no

      * `--enable-decimal-float`, `--enable-fixed-point`, `--with-long-double-128`

        `configure: WARNING: decimal float is not supported for this target,
        ignored`

          * `libgcc/configure.ac` [might
            need](http://gcc.gnu.org/ml/gcc-patches/2008-10/msg00315.html) to be
            aligned for us to the `*linux*` cases.  As well as at the end of
            `libgcc/config.host`.  Check.

                checking whether decimal floating point is supported... no
                checking whether fixed-point is supported... no

  * `host-linux.c` vs. `host-default.c`

  * *fixincludes* stuff

  * malloc?

        -cat ../../hurd/gcc/config/i386/pmm_malloc.h > mm_malloc.h
        +cat ../../hurd/gcc/config/i386/gmm_malloc.h > mm_malloc.h

    Comes from `gcc/config.gcc`: `i386/t-pmm_malloc` vs. `i386/t-gmm_malloc`
    for `i[34567]86-*-linux*` vs. `i[34567]86-*-*`.

  * *libgomp*

      * `libgomp/config/linux/`, `libgomp/config/linux/x86`

    `sed`ed away.

      * `-ftls-model=initial-exec -march=i486 -mtune=i686`

    `sed`ed away.

  * Missing `EOWNERDEAD`, `ENOTRECOVERABLE`.  What're they used for?

  * `RLIMIT_VMEM`.  Usage kosher?

  * `libtool: link: ar rc .libs/libstdc++.a [...]`

    Just different order of object files, or another problem?  TODO

  * `libobjc/encoding.c`: 

         libtool: compile:  [...]/hurd/master.build/./gcc/xgcc [...] [...]/hurd/master/libobjc/encoding.c -c [...]
        +[...]/hurd/master/libobjc/encoding.c:128:1: warning: '_darwin_rs6000_special_round_type_align' defined but not used [-Wunused-function]

  * `libobjc/thr.c`: `gcc/gthr-posix.h`

         libtool: compile:  [...]/hurd/master.build/./gcc/xgcc [...] [...]/hurd/master/libobjc/thr.c -c [...]
        +In file included from [...]/hurd/master/libobjc/../libgcc/gthr.h:142:0,
        +                 from [...]/hurd/master/libobjc/thr.c:45:
        +../libgcc/gthr-default.h: In function '__gthread_objc_thread_set_priority':
        +../libgcc/gthr-default.h:388:41: warning: unused parameter 'priority' [-Wunused-parameter]

  * `/proc/self/*`

        -checking for /proc/self/exe... yes
        -checking for /proc/self/maps... yes
        +checking for /proc/self/exe... no
        +checking for /proc/self/maps... no

  * GCJ: `java-signal.h`, `java-signal-aux.h`

        -config.status: linking ../../../hurd/libjava/include/i386-signal.h to include/java-signal.h
        -config.status: linking ../../../hurd/libjava/include/i386-signal.h to include/java-signal-aux.h
        +config.status: linking ../../../hurd/libjava/include/default-signal.h to include/java-signal.h
        +config.status: linking ../../../hurd/libjava/include/default-signal.h to include/java-signal-aux.h

  * GCJ: `jni_md.h`

        -checking jni_md.h support... yes
        +checking jni_md.h support... configure: WARNING: no

  * *default library search path*

        -checking for the default library search path... /lib /usr/lib /lib/i386-linux-gnu /usr/lib/i386-linux-gnu /lib/i486-linux-gnu /usr/lib/i486-linux-gnu /usr/local/lib /lib64 /usr/lib64 
        +checking for the default library search path... /lib /usr/lib

  * `./classpath/[...]/*.properties`

    Just different order of files, or another problem?

  * `libjava/gnu/gcj/util/natGCInfo.cc`

         libtool: compile:  [...]/hurd/master.build/./gcc/xgcc [...] -c ../../../master/libjava/gnu/gcj/util/natGCInfo.cc [...]
        +../../../master/libjava/gnu/gcj/util/natGCInfo.cc:440:1: warning: unused parameter 'name' [-Wunused-parameter]
        +../../../master/libjava/gnu/gcj/util/natGCInfo.cc:446:1: warning: unused parameter 'name' [-Wunused-parameter]
        +../../../master/libjava/gnu/gcj/util/natGCInfo.cc:452:1: warning: unused parameter 'name' [-Wunused-parameter]

  * `libgcj.la`

    Just different order of object files, or another problem?

    Is there a pattern that GNU/Hurd hands out the files alphabetically sorted
    where it wouldn't need to ([[!taglink open_issue_hurd]])?

  * `libjvm.la`, `.libs/libjvm.so`, `libgij.la`, `.libs/libgij.so.12.0.0`

    `-Wl,-Bsymbolic` vs. `-Wl,-Bsymbolic-functions`

  * `jar`

         make[2]: Entering directory `[...]/hurd/master.build/[ARCH]/libjava'
        -: make ; exec make "AR_FLAGS=rc" [...] "RANLIB=ranlib" "DESTDIR=" "JAR=[...]/hurd/master.build/[ARCH]/libjava/scripts/jar" DO=all multi-do
        +: make ; exec make "AR_FLAGS=rc" [...] "RANLIB=ranlib" "DESTDIR=" "JAR=jar" DO=all multi-do

    Probably because kepler.SCHWINGE has an OpenJDK `/usr/bin/jar`, and
    coulomb.SCHWINGE a GCJ one.

    There are other instances of this in the following.

  * *default library search path*

        -checking for the default library search path... /lib /usr/lib /lib/[MULTIARCH] /usr/lib/[MULTIARCH] /lib/i486-linux-gnu /usr/lib/i486-linux-gnu /usr/local/lib /lib64 /usr/lib64 
        +checking for the default library search path... /lib /usr/lib

    Should be aligned by Samuel's binutils patch.

  * `value-unwind.h`

        -DEFINES='' HEADERS='../../../master/libgcc/config/i386/value-unwind.h' \
        +DEFINES='' HEADERS='' \
                        ../../../master/libgcc/mkheader.sh > tmp-libgcc_tm.h

    Comes from `gcc/config.gcc`: for `i[34567]86-*-linux*`
    vs. `i[34567]86-*-*`, but apparently is important only for *x86_64* anyway.

  * `soft-fp` prototypes

         ../../../master/libgcc/soft-fp/eqtf2.c:34:9: warning: no previous prototype for '__eqtf2' [-Wmissing-prototypes]
        +../../../master/libgcc/soft-fp/eqtf2.c:50:1: warning: no previous prototype for '__netf2' [-Wmissing-prototypes]

         ../../../master/libgcc/soft-fp/getf2.c:34:9: warning: no previous prototype for '__getf2' [-Wmissing-prototypes]
        +../../../master/libgcc/soft-fp/getf2.c:50:1: warning: no previous prototype for '__gttf2' [-Wmissing-prototypes]

         ../../../master/libgcc/soft-fp/letf2.c:34:9: warning: no previous prototype for '__letf2' [-Wmissing-prototypes]
        +../../../master/libgcc/soft-fp/letf2.c:50:1: warning: no previous prototype for '__lttf2' [-Wmissing-prototypes]

  * `libatomic` on GNU/Linux compiles several more files than on GNU/Hurd.  Is
    that correct?  Probably futex support.


# Install

    $ make install 2>&1 | tee log_install
    [...]

This takes up around 850 MiB, and needs roughly 4 min on kepler.SCHWINGE and 45
min on coulomb.SCHWINGE.


## Analysis

    $ toolchain/logs/process gcc install

  * `libtool: finish`: `ldconfig` is not run for the Hurd.

    [[libtool]].

  * `libjvm.la`, `.libs/libjvm.so`, `libgij.la`, `.libs/libgij.so.12.0.0`

    `-Wl,-Bsymbolic` vs. `-Wl,-Bsymbolic-functions` (as above)

  * `jar`: as above.


# Testsuite

<http://gcc.gnu.org/install/test.html>

Testing on GNU/Hurd is blocked on
[[fork_mach_port_mod_refs_ekern_urefs_owerflow]].

TODO.  Can use parallel testing, see [[!message-id
"20110331070322.GI11563@sunsite.ms.mff.cuni.cz"]].

    $ make -k RUNTESTFLAGS=-v check 2>&1 | tee log_check
    [...]

This needs roughly 6.5 h on kepler.SCHWINGE and 50.25 h on coulomb.SCHWINGE.


## Analysis

    $ toolchain/logs/process gcc test

TODO.


# Specific Languages

  * [[GNAT]]

  * [[gccgo]]