Currently, the internal toolchain backend does a three stage gcc
build, with the following sequence of builds:
- build gcc-initial
- configure libc, install headers and start files
- build gcc-intermediate
- build libc
- build gcc-final
However, it turns out that this is not necessary, and only a two stage
gcc build is needed. At some point, it was believed that a three stage
gcc build was needed for NPTL based toolchains with old gcc versions,
but even a gcc 4.4 build with a NPTL toolchain works fine.
So, this commit switches the internal toolchain backend to use a two
stage gcc build: just gcc-initial and gcc-final. It does so by:
* Removing the custom dependency of all C libraries build step to
host-gcc-intermediate. Now the C library packages simply have to
depend on host-gcc-initial as a normal dependency (which they
already do), and that's it.
* Build and install both gcc *and* libgcc in
host-gcc-initial. Previously, only gcc was built and installed in
host-gcc-initial. libgcc was only done in host-gcc-intermediate,
but now we need libgcc to build the C library.
* Pass appropriate environment variables to get SSP (Stack Smashing
Protection) to work properly:
- Tell the compiler that the libc will provide the SSP support, by
passing gcc_cv_libc_provides_ssp=yes. In Buildroot, we have
chosen to use the SSP support from the C library instead of the
SSP support from the compiler (this is not changed by this patch
series, it was already the case).
- Tell glibc to *not* build its own programs with SSP support. The
issue is that if glibc detects that the compiler supports
-fstack-protector, then glibc uses it to build a few things with
SSP. However, at this point, the support is not complete (we
only have host-gcc-initial, and the C library is not completely
built). So, we pass libc_cv_ssp=no to tell the C library to not
use SSP support itself. Note that this is not a big loss: only a
few parts of the C library were built with -fstack-protector,
not the entire library.
* A special change is needed for ARC, because its libgcc depends on
the C library, which breaks building libgcc in
host-gcc-initial. This looks like a bug in the ARC compiler, as it
does not obey the inhibit_libc variable which tells the compiler
build process to *not* enable things that depend on the C
library. So for now, in host-gcc-initial, we simply disable the
build of libgmon.a for ARC. It's going to be built as part of
host-gcc-final, so the final compiler will have gmon support.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
gcc build scripts use wrong variable name to specify xtensa overlay
application command. As a result gcc is built with the default overlay,
which leads to obscure failures later in the build process.
xtensa toolchain needs an additional configuration for a specific core
variant we're building for. This configuration is called 'overlay' and
is an archive with files for binutils, gcc and gdb that replace
corresponding files in toolchain components.
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
Acked-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
While the idea of skipping the intermediate gcc step seems to work
fine in most situations, it causes problems with the SSP
support. Until we can figure out a proper solution for this problem,
we need to revert back to the previous solution of a three stages
build.
This reverts commit 2babed4a50.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <peter@korsgaard.com>
When NPTL support was introduced, gcc required a three stages build
process. Since gcc 4.7, this is no longer necessary, and it is
possible to get back to a two stages build process. This patch takes
advantage of this, by doing a two stages build process when possible.
We introduce a few hidden kconfig options:
* BR2_GCC_VERSION_NEEDS_THREE_STAGE_BUILD, which is set by the gcc
Config.in logic to indicate that the compiler might need a three
stages build. Currently, all versions prior to 4.7.x are selecting
this kconfig option.
* BR2_TOOLCHAIN_LIBC_NEEDS_THREE_STAGE_BUILD, which indicates whether
the C library might need a three stages build. This is the case for
eglibc, and uClibc when NPTL is enabled.
* BR2_TOOLCHAIN_NEEDS_THREE_STAGE_BUILD finally is enabled when both
of the previous options are enabled. It indicates that a three
stages build is actually needed.
In addition to those options, the uClibc/gcc build logic is changed to
use only a two stages build process when possible.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
Fixes a build issue with the avr32 toolchain:
http://jenkins.free-electrons.com/job/buildroot/config=atngw100_defconfig/104/
Invalid configuration `MAKEINFO=missing': machine `MAKEINFO=missing' not
recognized
Instead pass it in the environment of ./configure, similar to how it was
done originally in 62322acb2c (toolchain/gcc: disable makeinfo).
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
When refactoring the internal toolchain backend logic, the code was
changed to pass the custom configure options given through
BR2_EXTRA_GCC_CONFIG_OPTIONS only for the gcc final pass, with the
idea that we're only interested by user customization for the final
compiler.
However, the beaglebone_defconfig was passing --with-float=hard
--with-fpu=vfpv3-d16 as BR2_EXTRA_GCC_CONFIG_OPTIONS, and since the
refactoring, it was causing build failures of the beaglebone_defconfig
(with messages saying that Busybox is built to use VFP arguments, but
libc/libm are not). This is due to the fact that the gcc intermediate,
which is used to build the C library, wasn't built to generate hard
float, while the final compiler was generating hard float.
So, we get back to the original situation where the options in
BR2_EXTRA_GCC_CONFIG_OPTIONS are passed to all of the compiler
passes. Of course, the specific case of hard float will be fixed by
following patches in this area, but the idea still remains: the three
gcc should have the same options, if those options affected the ABI of
the generated code.
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
Several sub-directories of the gcc code base are in fact not needed
for the Buildroot build: libjava/, libgo/ and gcc/testsuite/ being the
biggest ones. Avoiding their extraction saves quite a bit of disk
space, and compensates a bit the fact that we now extract three times
the gcc source code.
This requires changing the 100-uclibc-conf.patch to no longer patch
files from the libjava/ directory, since this directory is no longer
extracted.
[Peter: add comment about why this is done]
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: Peter Korsgaard <jacmet@sunsite.dk>
