[Buildroot] Toolchain issues on multiple architectures

[Matthew Weber <matthew.weber@rockwellcollins.com>]

On Thu, Sep 10, 2020 at 10:43 AM Matthew Weber
<matthew.weber@rockwellcollins.com> wrote:
>
> On Thu, Sep 10, 2020 at 9:43 AM Matthew Weber
> <matthew.weber@rockwellcollins.com> wrote:
> >
> > On Thu, Sep 10, 2020 at 5:06 AM Thomas Petazzoni
> > <thomas.petazzoni@bootlin.com> wrote:
> > >
> > > Hello,
> > >
> > > As part of the https://toolchains.bootlin.com project, I've tried to
> > > rebuild 175 toolchains on top of Buildroot 2020.08, and encountered a
> > > number of failures, which are summarized below.
> > >
> > > Damien, Alistair: there is an issue with the RISC-V 32-bit toolchain.
> > > Could you have a look, and tell us if you have some suggestions?
> > >
> > > ARC maintainers: all ARC toolchains are failing to build due to the use
> > > of a recent gdb where gdbserver was moved to the top-level. However, it
> > > doesn't seem to be that easy to build just gdbserver in those recent
> > > gcc versions. I have posted a question about this on the gdb@ mailing
> > > list: https://sourceware.org/pipermail/gdb/2020-September/048888.html.
> > >
> > > Matt, there are some PowerPC64 build issues as well. Have you ever seen
> > > them, do you have any hints?
> > >
> [snip]
> > > PowerPC64 (big and little endian) Power 8
> > > =========================================
> > >
> > > Both the glibc stable and bleeding edge are failing to build libstdc++
> > > in host-gcc-final. Interesingly, the musl configurations work fine.
> > >
> > > - https://gitlab.com/bootlin/toolchains-builder/-/jobs/729359625
> > > - https://gitlab.com/bootlin/toolchains-builder/-/jobs/729359626
> > > - https://gitlab.com/bootlin/toolchains-builder/-/jobs/729359621
> > > - https://gitlab.com/bootlin/toolchains-builder/-/jobs/729359622
> >
> > Checking out 2020.08 and using the following config from the CI job.
> > BR2_powerpc64le=y
> > BR2_powerpc_power8=y
> > BR2_KERNEL_HEADERS_4_9=y
> > BR2_TOOLCHAIN_BUILDROOT_CXX=y
> > BR2_TOOLCHAIN_BUILDROOT_FORTRAN=y
> > BR2_PACKAGE_HOST_GDB=y
> > BR2_INIT_NONE=y
> > BR2_SYSTEM_BIN_SH_NONE=y
> > # BR2_PACKAGE_BUSYBOX is not set
> > BR2_PACKAGE_GDB=y
> > # BR2_TARGET_ROOTFS_TAR is not set
> >
> > I see this error:
> > ../../../libgfortran/libgfortran.h:61:12: fatal error:
> > quadmath_weak.h: No such file or directory
> >
> > I did notice that if I did a basic build below using current defaults,
> > it doesn't fail.
> > BR2_powerpc64=y
> > BR2_powerpc_power8=y
> > BR2_TOOLCHAIN_BUILDROOT_CXX=y
> > BR2_INIT_NONE=y
> > # BR2_TARGET_ROOTFS_TAR is not set
>
> Disregard the successful build above, I forgot to enable fortran.
>
> It looks like the powerpc64 is detecting __float128 and is trying to
> use the quadmath library which we disable.  I'll check to see if the
> change should be in the libfortran configure stage or what is being
> passed to gcc final.

My original understanding was that just the IBM families with ISA 3.0+
could do float128 [1].  Turns out any Power ISA v.2.06+ (64bit powerpc
NXP & IBM) should be able to have quadmath enabled and the stdlib may
emulate as needed [2].

I'm still testing a series of builds (e5500, e6500, power6/7/8) but I
believe the fix is to just add power64(big/le) to the condition to
enable quadmath (BR2_TOOLCHAIN_HAS_LIBQUADMATH).

config BR2_TOOLCHAIN_HAS_LIBQUADMATH
        bool
        default y if BR2_i386 || BR2_x86_64 || BR2_powerpc64 || BR2_powerpc64le

Are you able to make this change in your toolchain builder branch of
buildroot and validate the IBM and e5500 toolchains create images that
can login?  I can send a patch to the mailing list as well


Regards,
Matt

[1] https://en.wikipedia.org/wiki/Power_ISA
[2] https://gcc.gnu.org/onlinedocs/gcc/RS_002f6000-and-PowerPC-Options.html