Please read this document carefully before installing the GNU Compiler Collection on your machine.
We strongly recommend to upgrade to binutils 2.10 (or newer).
The following error:
Error: macro requires $at register while noat in effect
indicates that you should upgrade to a newer version of the assembler, 2.9 or later. If you can not upgrade the assembler, the compiler option "-Wa,-m21164a" may work around this problem.
If you install a shared libstdc++ and, when you link a non-trivial C++
program (for example, gcc/testsuite/g++.other/delete3.C
),
the linker reports a couple of errors about multiply-defined symbols
(for example, nothrow
, __throw
and
terminate(void)
), you've probably got a linker bug, for
which there's no known fix. The officially recommended work-around is
to remove the shared libstdc++.
An alternative solution is to arrange that all symbols from
libgcc
get copied to the shared libstdc++
;
see detailed solution below. (Surprising as it may seem, this does
indeed fix the problem!) Beware that this may bring you
binary-compatibility problems in the future, if you don't use the same
work-around next time you build libstdc++
: if programs
start to depend on libstdc++
to provide symbols that used
to be only in libgcc
, you must arrange that
libstdc++
keeps providing them, otherwise the programs
will have to be relinked.
The magic spell is to add -Wl,-all,-lgcc,-none
to the
definition of macro SHDEPS
in
libstdc++/config/dec-osf.ml
before
alpha*-dec-osf*/libstdc++/Makefile
is created (a patch that does just that is
available). If the Makefile already exists, run
./config.status
within directory
alpha*-dec-osf*/libstdc++
(and
alpha*-dec-osf*/ieee/libstdc++
, if it also exists).
Remove any existing libstdc++.so*
from such directories,
and run make all-target-libstdc++
in the top-level
directory, then make install-target-libstdc++
.
If you have already removed the build tree, you may just remove
libstdc++.so.2.10.0
from the install tree and re-create
it with the command gcc -shared -o libstdc++.so.2.10.0
-Wl,-all,-lstdc++,-lgcc,-none -lm
. If the ieee
sub-directory exists, repeat this command in it, with the additional
flag -mieee
.
Please have a look at our binaries page.
We highly recommend using gas/binutils-2.8 or newer on all hppa platforms; you may encounter a variety of problems when using the HP assembler.
Specifically, -g
does not work on HP-UX (since that system
uses a peculiar debugging format which GCC does not know about), unless you
use GAS and GDB and configure GCC with the --with-gnu-as
option.
If you wish to use pa-risc 2.0 architecture support, you must use either the HP assembler or a recent snapshot of gas.
More specific information to hppa*-hp-hpux* targets follows.
The HP assembler has major problems on this platform. We've tried to work around the worst of the problems. However, those workarounds may be causing linker crashes in some circumstances; the workarounds also probably prevent shared libraries from working. Use the GNU assembler to avoid these problems.
The configuration scripts for GCC will also trigger a bug in the hpux9 shell. To avoid this problem set CONFIG_SHELL to /bin/ksh and SHELL to /bin/ksh in your environment.
For hpux10.20, we highly recommend you pick up the latest sed
patch PHCO_19798
from HP. HP has two sites which provide patches free of charge:
The HP assembler on these systems is much better than the hpux9 assembler,
but still has some problems. Most notably the assembler inserts timestamps
into each object file it creates, causing the 3-stage comparison test to fail
during a `make bootstrap
'. You should be able to continue by
saying `make all
' after getting the failure from `make
bootstrap
'.
GCC 2.95.2 does not support HP-UX 11, and it cannot generate 64-bit object files. Current (as of late 2000) snapshots and GCC 3.0 do support HP-UX 11.
If you use glibc 2.2 (or 2.1.9x), GCC 2.95.2 won't install out-of-the-box. You'll get compile errors while building libstdc++. The patch glibc-2.2.patch, that is to be applied in the GCC source tree, fixes the compatibility problems.
You will need binutils-2.9.1.0.15 or newer for exception handling to work.
If you receive Signal 11 errors when building on GNU/Linux, then it is possible you have a hardware problem. Further information on this can be found on www.bitwizard.nl.
If you are building languages other than C, you must follow the instructions
about invoking `make bootstrap
' because the native OpenServer
compiler will build a cc1plus
that will not correctly parse many
valid C++ programs including those in libgcc.a
. You must do a `make bootstrap
' if you are building with the native compiler.
Use of the `-march-pentiumpro
' flag can result in
unrecognized opcodes when using the native assembler. While
it's rather rare to see these emitted by GCC yet, errors of the basic form:
/usr/tmp/ccaNlqBc.s:22:unknown instruction: fcomip /usr/tmp/ccaNlqBc.s:50:unknown instruction: fucomip
are symptoms of this problem. You may work around this by not building affected files with that flag or by using the GNU assembler. Users of GNU assembler should see the note below for hazards on doing so.
If you choose to configure with --enable-shared
you should also
specify --with-gnu-as --disable-multilib
even if you are not
using the GNU assembler. In doing so you will give up the ability
to generate COFF executables as described below. This combination
of flags is necessary to suppress negative interactions with multilibing.
The native SCO assembler that is provided with the OS at no charge
is normally required. If, however, you must be able to use the GNU
assembler (perhaps you're compiling code with asms that require GAS syntax)
you may configure this package using the flags --with-gnu-as
.
You must use a recent version of GNU binutils; versions past 2.9.1 seems to work well.
If you select this option, you will be unable to reliably build COFF
images. In general, the --with-gnu-as
option isn't as well tested as the native assembler.
Unlike various prereleases of GCC that used -belf
and
defaulted to COFF, you must now use the -melf
and
-mcoff
flags to toggle between the two object file formats.
ELF is now the default.
Look in gcc/config/i386/sco5.h
(search for "messy") for additional OpenServer-specific flags.
Systems based on OpenServer before 5.0.4 (`uname -X
' will
tell you what you're running) require TLS597 from ftp.sco.com/TLS for
C++ constructors and destructors to work right.
The system linker in (at least) 5.0.4 and 5.0.5 will sometimes do the wrong thing for a construct that GCC will emit for PIC code. This can be seen as execution testsuite failures when using -fPIC on 921215-1.c, 931002-1.c, nestfunc-1.c, and gcov-1.c. For 5.0.5, an updated linker that will cure this problem is available. You must install both ftp://ftp.sco.com/Supplements/rs505a/ and OSS499A.
The dynamic linker in OpenServer 5.0.5 (earlier versions may show
the same problem) aborts on certain g77-compiled programs. It's particularly
likely to be triggered by building Fortran code with the -fPIC
flag.
Although it's conceivable that the error could be triggered by other
code, only G77-compiled code has been observed to cause this abort.
If you are getting core dumps immediately upon execution of your
g77 program - and especially if it's compiled with -fPIC - try applying
`sco_osr5_g77.patch'
to your libf2c and
rebuilding
GCC. Affected faults, when analyzed in a debugger, will show a stack
backtrace with a fault occurring in rtld()
and the program
running as /usr/lib/ld.so.1
. This problem has been reported to SCO engineering
and will hopefully be addressed in later releases.
GCC 2.95.2, when configured to use the GNU assembler, would invoke
it with the -s
switch, that GNU as up to 2.9.5.0.12 does
not support. If you'd rather not use a newer GNU as nor the native
assembler, you'll need the patch `x86-sol2-gas.patch'
.
This target emulates the SCO Universal Development Kit and requires that
package be installed. (If it is installed, you will have a /udk/usr/ccs/bin/cc
file present.) It's very much like the i?86-*-unixware7*
target
but is meant to be used when hosting on a system where UDK isn't the
default compiler such as OpenServer 5 or Unixware 2. This target will
generate binaries that will run on OpenServer, Unixware 2, or Unixware 7,
with the same warnings and caveats as the SCO UDK.
You can stage1 with either your native compiler or with UDK. If you don't do a full bootstrap when initially building with your native compiler you will have an utterly unusable pile of bits as your reward.
This target is a little tricky to build because we have to distinguish it from the native tools (so it gets headers, startups, and libraries from the right place) while making the tools not think we're actually building a cross compiler. The easiest way to do this is with a configure command like this:
CC=/udk/usr/ccs/bin/cc /your/path/to/gcc/configure --host=i686-pc-udk --target=i686-pc-udk --program-prefix=udk-
You should substitute 'i686' in the above command with the appropriate processor for your host.
You should follow this with a `make bootstrap
' then
`make install
'. You can then access the UDK-targeted GCC
tools by adding udk-
before the commonly known name. For example, to invoke the C compiler, you would use `udk-gcc
'. They will coexist peacefully with any native-target GCC tools you may have installed.
AIX Make frequently has problems with GCC makefiles. GNU Make 3.76 or newer is recommended to build on this platform.
Errors involving "alloca" when building GCC generally are due to an incorrect definition of CC in the Makefile or mixing files compiled with the native C compiler and GCC. During the stage1 phase of the build, the native AIX compiler must be invoked as "cc" (not "xlc"). Once configure has been informed of "xlc", one needs to use "make distclean" to remove the configure cache files and ensure that $CC environment variable does not provide a definition that will confuse configure. If this error occurs during stage2 or later, then the problem most likely is the version of Make (see above).
Some versions of the AIX binder (linker) can fail with a relocation overflow severe error when the -bbigtoc option is used to link GCC-produced object files into an executable that overflows the TOC. A fix for APAR IX75823 (OVERFLOW DURING LINK WHEN USING GCC AND -BBIGTOC) is available from IBM Customer Support and from its service.boulder.ibm.com website as PTF U455193.
Binutils does not support AIX 4.3 (at least through release 2.9). GNU as and GNU ld will not work properly and one should not configure GCC to use those GNU utilities. Use the native AIX tools which do interoperate with GCC.
AIX 4.3 utilizes a new "large format" archive to support both 32-bit and 64-bit object modules. The routines provided in AIX 4.3.0 and AIX 4.3.1 to parse archive libraries did not handle the new format correctly. These routines are used by GCC and result in error messages during linking such as "not a COFF file". The version of the routines shipped with AIX 4.3.1 should work for a 32-bit environment. The -g option of the archive command may be used to create archives of 32-bit objects using the original "small format". A correct version of the routines is shipped with AIX 4.3.2.
The initial assembler shipped with AIX 4.3.0 generates incorrect object files. A fix for APAR IX74254 (64BIT DISASSEMBLED OUTPUT FROM COMPILER FAILS TO ASSEMBLE/BIND) is available from IBM Customer Support and from its service.boulder.ibm.com website as PTF U453956. This fix is incorporated in AIX 4.3.1 and above.
The AIX 4.3.2.1 linker (bos.rte.bind_cmds Level 4.3.2.1) will dump core with a segmentation fault when invoked by any version of GCC. A fix for APAR IX87327 is available from IBM Customer Support and from its service.boulder.ibm.com website as PTF U461879. This fix is incorporated in AIX 4.3.3 and above.
You absolutely must use GNU sed and GNU make on this platform.
On NEXTSTEP 3.x where x < 3 the build of GCC will abort during stage1 with an error message like this:
_eh /usr/tmp/ccbbsZ0U.s:987:Unknown pseudo-op: .section /usr/tmp/ccbbsZ0U.s:987:Rest of line ignored. 1st junk character valued 95 (_).
The reason for this is the fact that NeXT's assembler for these versions of the operating system does not support the .section pseudo op that's needed for full C++ exception functionality.
As NeXT's assembler is a derived work from GNU as, a free replacement that does can be obtained at ftp://ftp.next.peak.org:/next-ftp/next/apps/devtools/as.3.3.NIHS.s.tar.gz.
If you try to build the integrated C++ & C++ runtime libraries on this system you will run into trouble with include files. The way to get around this is to use the following sequence. Note you must have write permission to the directory prefix you specified in the configuration process of GCC for this sequence to work.
cd bld-gcc make all-texinfo all-bison all-byacc all-binutils all-gas all-ld cd gcc make bootstrap make install-headers-tar cd .. make bootstrap3
It is reported that you may need the GNU assembler on this platform.
You must use GAS on these platforms, as the native assembler can not handle the code for exception handling support. Either of these messages indicates that you are using the MIPS assembler when instead you should be using GAS:
as0: Error: ./libgcc2.c, line 1:Badly delimited numeric literal .4byte $LECIE1-$LSCIE1 as0: Error: ./libgcc2.c, line 1:malformed statement
or:
as0: Error: /src/bld-gcc/gcc/libgcc2.c, line 1:undefined symbol in expression .word $LECIE1-$LSCIE1
These systems don't have ranlib, which various components in GCC need; you should be able to avoid this problem by installing GNU binutils, which includes a functional ranlib for this system.
You may get the following warning on irix4 platforms, it can be safely ignored.
warning: foo.o does not have gp tables for all its sections.
When building GCC, the build process loops rebuilding cc1 over and
over again. This happens on mips-sgi-irix5.2, and possibly other platforms.
It has been reported that this is a known bug in the make shipped with
IRIX 5.2. We recommend you use GNU make instead of the vendor supplied
make program; however, you may have success with "smake" on IRIX 5.2 if
you do not have GNU make available.
See http://reality.sgi.com/ariel/freeware for more information about using GCC on IRIX platforms.
You must not use GAS on irix6 platforms; doing so will only cause problems.
These systems don't have ranlib, which various components in GCC need; you should be able to avoid this problem by making a dummy script called ranlib which just exits with zero status and placing it in your path.
If you are using Irix cc as your bootstrap compiler, you must
ensure that the N32 ABI is in use. To test this, compile a simple C
file with cc
and then run file
on the
resulting object file. The output should look like:
test.o: ELF N32 MSB ...
If you see:
test.o: ELF 32-bit MSB
then your version of cc
uses the O32 ABI default. You
should set the environment variable CC
to 'cc -n32'
before configuring GCC.
GCC does not currently support generating O32 ABI binaries in the mips-sgi-irix6 configurations. It used to be possible to create a GCC with O32 ABI only support by configuring it for the mips-sgi-irix5 target. See the link below for details.
GCC does not correctly pass/return structures which are smaller than 16 bytes and which are not 8 bytes. The problem is very involved and difficult to fix. It affects a number of other targets also, but IRIX 6 is affected the most, because it is a 64 bit target, and 4 byte structures are common. The exact problem is that structures are being padded at the wrong end, e.g. a 4 byte structure is loaded into the lower 4 bytes of the register when it should be loaded into the upper 4 bytes of the register.
GCC is consistent with itself, but not consistent with the SGI C compiler (and the SGI supplied runtime libraries), so the only failures that can happen are when there are library functions that take/return such structures. There are very few such library functions. I can only recall seeing two of them: inet_ntoa, and semctl.
See http://reality.sgi.com/ariel/freeware for more information about using GCC on IRIX platforms.
You will need binutils-2.9.4.0.8 or newer for a working GCC. It is strongly recommended to recompile binutils if you initially built it with gcc-2.7.2.x.
Starting with Solaris, Sun does not ship a C compiler any more. To bootstrap and install GCC you first have to install a pre-built compiler, see our binaries page for details.
Sun as 4.X is broken in that it cannot cope with long symbol names. A typical error message might look similar to the following:
/usr/ccs/bin/as: "/var/tmp/ccMsw135.s", line 11041: error: can't compute
value of an expression involving an external symbol.
See the How to work around too long C++ symbol names? FAQ entry for further information.
Sun make in all known Solaris 1 (SunOS 4) and Solaris 2 releases has a broken VPATH mechanism, which means you must either:
binutils 2.9.1 has known bugs on this platform. We recommend to use the vendor tools (Sun as, Sun ld) until these have been fixed. Unfortunately, C++ shared libraries, including libstdc++, won't work properly if assembled with Sun as: the linker will complain about relocations in read-only sections, in the definition of virtual tables. Some possible work-arounds: use some development release of binutils, wait for the next stable binutils release or refrain from creating C++ shared libraries.
Sun patch 107058-01 (1999-01-13) for SPARC Solaris 7 triggers a bug in the dynamic linker. This problem (Sun bug 4210064) affects GCC 2.8 and later, including all EGCS releases. Sun formerly recommended 107058-01 for all Solaris 7 users, but around 1999-09-01 it started to recommend it only for people who use Sun's compilers.
Here are some workarounds to this problem:
/usr/ccs/bin/as
into
/usr/local/lib/gcc-lib/sparc-sun-solaris2.7/2.95.1/as
,
adjusting the latter name to fit your local conventions and software
version numbers.A bug in the SunOS4 linker will cause it to crash when linking -fPIC compiled objects (and will therefore not allow you to build shared libraries).
To fix this problem you can either use the most recent version of binutils or get the latest SunOS4 linker patch (patch ID 100170-10) from Sun's patch site.
It has been reported that you might need binutils-2.8.1.0.23 for this platform, too.
GCC version 2.95 is not able to compile code correctly for
sparc64
targets. Users of the Linux kernel, at least,
can use the sparc32
program to start up a new shell
invocation with an environment that causes configure
to
recognize (via uname -a
) the system as
sparc-*-*
instead.
GCC currently builds under the Cygwin
environment. However you need to delete (or rename) the texinfo
directory included with the GCC sources. You will also need to configure
with the option --with-included-gettext
.
GCC does not currently support OS/2. However, Andrew Zabolotny has been working on a generic OS/2 port with pgcc. The current code code can be found at http://www.goof.com/pcg/os2/.
An older copy of GCC 2.8.1 is included with the EMX tools available at ftp://ftp.leo.org/pub/comp/os/os2/leo/devtools/emx+gcc/.
C++ support is significantly better on ELF targets if you use the GNU linker; duplicate copies of inlines, vtables and template instantiations will be discarded automatically.