Based on the following samples:
- x86_64-unknown-linux-gnu (-m64/-m32/-mx32)
- powerpc64-unknown-linux-gnu (-m64/-m32)
- mipsel-unknown-linux-gnu (-mabi={32,n32,64})
- sh4-unknown-linux-gnu (-m4/-m4a)
- x86_64-unknown-linux-uclibc (-m64/-m32)
- mips64el-unknown-linux-uclibc (-mabi={32,n32,64})
New samples:
- sparc64-multilib-linux-gnu (-m64/-m32)
- sh4-multilib-linux-uclibc (-m4/-m4a/-m3)
- x86_64-multilib-linux-musl (-m64/-m32)
Notably missing is arm-unknown-linux-gnu (aprofile): GLIBC does not
compile in one of the variants in its multilib set (-march=armv4t
-mthumb) due to missing atomic ops implementation.
Signed-off-by: Alexey Neyman <stilor@att.net>
It turns out that core GCC on binfmt architectures (m68k, for example)
cannot produce the final executable (looks for ld.real in the wrong
place). Need to wait for the final gcc to become available.
Signed-off-by: Alexey Neyman <stilor@att.net>
'ld' does not search for dependency libraries in multi_os_directory, so
if there's both multi_os_directory and multi_root, and there is only one
configuration in each multi_root, forgo the multi_os_directory suffix.
Needed for sh4-multilib-linux-uclibc.
Signed-off-by: Alexey Neyman <stilor@att.net>
Create a separate 'libc_backend_once', install headers into a
subdirectory (different sets of headers are installed for 32- and 64-bit
architectures), and create a symlink for the dynamic linker location
expected by GCC.
Signed-off-by: Alexey Neyman <stilor@att.net>
Rather than echo-ing the new value, set the value into the variable with
the name passed as an argument (similar to CT_SanitizeVarDir). This
allows to use CT_DoLog in these functions.
Signed-off-by: Alexey Neyman <stilor@att.net>
This step was only used in uClibc. However, with upcoming multilib, the
config management will have to be done for each variant differently,
anyway.
uClibc was the only user of libc_check_config step, as well as
CT_CONFIG_DIR directory. Retire these.
Two other clean-ups in uClibc.sh:
- KERNEL_HEADERS check seems to be bogus, this config option is not
present even in 0.9.30 - which is not supported already.
- SHARED_LIB_LOADER_PREFIX was renamed to MULTILIB_DIR in 0.9.31,
according to ChangeLog - and MULTILIB_DIR is passed from command line
instead.
Signed-off-by: Alexey Neyman <stilor@att.net>
In preparation for multilib support, use the same "backend" model that
is already employed by glibc and musl.
Also, the verbosity setting descriptions were swapped. V=2 is actually
less verbose than V=1: V=1 prints full commands, while V=2 prints 'CC
<file> <defines>'.
Signed-off-by: Alexey Neyman <stilor@att.net>
- Dump CT_LIBC_EXTRA_CC_ARGS: instead, treat CT_LIBC_EXTRA_CFLAGS as
arguments to CC (or they are not applied to .S, for example).
Combine them with multi_flags and CT_TARGET_CFLAGS in proper order.
- Analyze thus combined flags to determine --with-fp/--without-fp.
Don't need to check CT_ARCH_FLOAT - it is reflected in
CT_TARGET_CFLAGS anyway. Check more soft/hard float options defined
on different architectures.
- Drop checking for endianness flags: they are not reflected in
configure arguments in any way, and they're already present in CFLAGS
(either via multi_flags or via CT_TARGET_CFLAGS). Besides,
CT_ARCH_ENDIAN_OPT was actually called CT_ARCH_ENDIAN_CFLAG, so this
was a no-op anyway.
Signed-off-by: Alexey Neyman <stilor@att.net>
Install startfiles for libc variants into the most specific combination
(suffixed sysroot, if applicable + suffixed multi-os dir, if
applicable). Install headers once in every suffixed sysroot (although it
seems that GCC picks up headers from top-level sysroot, GCC manual
claims that sysroot suffix affects headers search path).
In uClibc, this requires a better sanitization of the directory: it
creates symlinks from {sysroot}/usr/lib/{multi_os_dir} to
{sysroot}/lib/{multi_os_dir} and to do so, it counts the number of path
components in the libdir. This breaks if one of such components is `..'
- symlinks contain an extra `../..' then. Since such sanitization had to
be implemented anyway, use it in other places to print more sensible
directory names.
Also, fix the description of configure --host/--target per musl's
configure help message (and its actual code).
Signed-off-by: Alexey Neyman <stilor@att.net>
On some arches (e.g. MIPS) the options like -mabi do not work if
specified more than once (see the comment in 100-gcc.sh). Therefore,
we need to determine which of the options produced by <arch>.sh can
be passed to multilib builds and which must be removed (i.e., which
options vary among the multilibs).
This presents a chicken-and-egg problem. GCC developers, in their
infinite wisdom, do not allow arbitrary multilib specification to be
supplied to GCC's configure. Instead, the target (and sometimes some
extra options) determine the set of multilibs - which may include
different CPUs, different ABIs, different endianness, different FPUs,
different floating-point ABIs, ... That is, we don't know which parts
vary until we build GCC and ask it.
So, the solution implemented here is:
- For multilib builds, start with empty CT_ARCH_TARGET_CFLAGS/LDFLAGS.
- For multilib builds, require core pass 1. Pass 1 does not build any
target binaries, so at that point, our target options have not been
used yet.
- Provide an API to modify the environment variables for the steps that
follow the current one.
- As a part of multilib-related housekeeping, determine the variable
part of multilibs and filter out these options; pass the rest into
CT_TARGET_CFLAGS/LDFLAGS.
This still does not handle extra dependencies between GCC options (like
-ma implying -mcpu=X -mtune=Y, etc.) but I feel that would complicate
matters too much. Let's leave this until there's a compelling case for
it.
Also, query GCC's sysroot suffix for targets that use it (SuperH,
for example) - the default multilib may not work if the command line
specifies the default option explicitly (%sysroot_suffix_spec is not
aware of multilib defaults).
Signed-off-by: Alexey Neyman <stilor@att.net>
Rather then building the manuals and locales for each multilib target, only
build the manuals on the last multilib target.
If you are not building a multilib toolchain, then the first libc build will
be the last.
Signed-off-by: Bryan Hundven <bryanhundven@gmail.com>
For 4 different folders:
${CT_PREFIX_DIR}
${CT_SYSROOT_DIR}
${CT_SYSROOT_DIR}/usr
${CT_PREFIX_DIR}/${CT_TARGET}
.. symlinks from 'lib32' and 'lib64' to 'lib' were created.
This was untidy and incorrect for multilib (the bitness of
the libraries in 'lib32' and 'lib64' will not be the same)
We can not know which folders this toolchain configuration
will require at this time so let them be created on-demand
instead.
Changed by Alexey Neyman: original change removed too much; we
still need to create the default directories because the os
directories are based off them (e.g. `lib/../lib64').
Signed-off-by: Ray Donnelly <mingw.android@gmail.com>
Signed-off-by: Alexey Neyman <stilor@att.net>
GCC makes the distinction between:
multilib (-print-multi-lib) and
multilib-os (--print-multi-os-directory)
as the GCC library and GCC sysroot library paths, respecitively.
Use this to build libc into the correct locations, the same
applies to the dummy libc.so
Changed by Alexey Neyman: restore missing CT_EndStep.
Signed-off-by: Ray Donnelly <mingw.android@gmail.com>
Signed-off-by: Alexey Neyman <stilor@att.net>
This commit makes the README.md more informative and basically jacks a
bunch of content from crosstool-ng.org (plus some extras).
I also added the waffle.io badges
Signed-off-by: Bryan Hundven <bryanhundven@gmail.com>
This change adds a powerpc64le-unknown-linux-gnu sample that can be used
to quickly create a little-endian tool-chain for powerpc64
architecture. This sample is based on the earlier work done by "Yann
E. MORIN" to add support for powerpc64 tool chain and implementing the
power64-unknown-linux-gnu sample. The existing sample however generates
a big-endian tool chain by default.
Signed-off-by: Vaibhav Jain <vaibhav@linux.vnet.ibm.com>
The previous patch added the function 'CT_DoMultilibTarget()' to
scripts/build/arch/*.sh.
This patch calls the common function to (currently) get just the target
tuple for the current multilib target.
This patch was originally by: Cody P Schafer
Changed by Alexey Neyman: first, try `gcc -print-multiarch`. If it is
supported, use whatever it reports. Otherwise, fall back to our
guesswork. Move "i486" quirk into glibc.sh, as it is specific to glibc
(e.g. uclibc will need i386, which is what GCC reports).
Signed-off-by: Bryan Hundven <bryanhundven@gmail.com>
Signed-off-by: Ray Donnelly <mingw.android@gmail.com>
Signed-off-by: Alexey Neyman <stilor@att.net>
i[34567]86-*-gnux32 is not a valid tuple.
Signed-off-by: Bryan Hundven <bryanhundven@gmail.com>
Signed-off-by: Ray Donnelly <ray.donnelly@gmail.com>
Signed-off-by: Alexey Neyman <stilor@att.net>
This code was abstracted out of Cody P Schafer's multilib patch.
It doesn't seem right having architecture dependent code in a
specific libc implementation script. So this patch breaks it out into
scripts/build/arch/<arch>.sh in a function:
multilib_target_to_build="$(CT_DoArchMultilibTarget 'multi_flags'
'target-in')"
Note that this function gets called on each multilib variant with
different sets of compiler flags supplied in 'multi_flags'. The caller
will first filter the flags so that there is no conflicting flags (e.g.,
no '-m32 -m64') supplied.
Changed by Alexey Neyman:
- make option analysis check specific option rather than match global
options string as a whole. Moreover, old code did not handle multiple
options in the same multilib, e.g. '-m64 -mlittle'.
- fixed substitutions in powerpc.sh (*le variants did not match the
pattern in the shell parameter expansion)
- make s390.sh actually apply the flags it gathered from the options.
- straighten the spaghetti in x86.sh by setting two flags, arch & abi.
Also, do not depend on "gnu" being the last part - we can have
'*-uclibcx32', for example.
Signed-off-by: Bryan Hundven <bryanhundven@gmail.com>
Signed-off-by: Ray Donnelly <ray.donnelly@gmail.com>
Signed-off-by: Alexey Neyman <stilor@att.net>
Written by Bryan Hundven.
Modified by Alexey Neyman to actually add the option to gcc.in.
Signed-off-by: Bryan Hundven <bryanhundven@gmail.com>
Signed-off-by: Ray Donnelly <mingw.android@gmail.com>
Signed-off-by: Alexey Neyman <stilor@att.net>
By default, it is 'auto' - which means, it is enabled if there are
multilibs directories detected in the installation location for libgcc.
Thus, it is not detected for pass-1 GCC: the installation location is
empty at this point.
Signed-off-by: Alexey Neyman <stilor@att.net>
If a multilib configuration contains an endianness option, the
${endian_extra} is set to, for example, 'mb' (note, no dash!). It is
then added to CFLAGS, resulting in bogus flags like 'mb -mb'. But it is
not even needed, as ${extra_flags} already contains the very same
option!
Found by experimenting with multilibs with different endianness on SH,
which still didn't work, but that's another story...
Signed-off-by: Alexey Neyman <stilor@att.net>
Now that libc backend installs the libraries into the directory reported
by gcc as 'multi-os-directory', sh4 libraries are installed into a '!m4'
subdirectory. This directory then confuses GNU ld, which assumes the
exclamation mark to be a word separator and attempts to link to
'/usr/lib' (a directory). However, if multilib is enabled, the default
libraries are installed into the [expected] '/usr/lib/./'. This looks
like an artifact of SuperH's unique way of specifying the multilibs to
be built in GCC (which may list exclusions, starting with '!').
Signed-off-by: Alexey Neyman <stilor@att.net>
By default, sparc64-*-linux is configured with -mcpu=v9. However,
according to https://sourceware.org/ml/libc-alpha/2005-12/msg00027.html:
"There is no Linux sparc64 port that runs on non-UltraSPARC-I+ ISA
CPUs."
v9 is such a "non-UltraSPARC-I+ ISA CPU", so it makes no sense to
default to v9 when targetting Linux.
Change the default to ultrasparc, even though it can suboptimally
schedule instructions for newer SPARC CPUs. See the pending patch:
https://patchwork.ozlabs.org/patch/409424/
Signed-off-by: Alexey Neyman <stilor@att.net>
If EXPERIMENTAL is not set, the only choice for version is the set of
released versions - currently, 1.1.14. But this only option is disabled
because it is also marked EXPERIMENTAL; this leaves no available choices
in the configuration.
Marking MUSL as experimental: it seems to have header issues which
prevent, for example, gdbserver from building. musl copied chunks of
ptrace.h code from the kernel into its own headers, which now clash with
Linux kernel headers. Manifests at least on SH4 target.
Also, musl breaks in powerpc builds: GCC balks at it with "unsupported
DEFAULT_LIBC" message. Also, 64-bit powerpc and mips are not supported.
So, until someone figures out the dependencies for musl in config/, mark
it experimental.
Signed-off-by: Alexey Neyman <stilor@att.net>
