Bump gcc to gcc10 and binutils to 2.34

[luc65r]

Motivation for this change

I wanted to have a try at updating binutils (I took the work of @lovesegfault #86954), and as all packages needed to be recompiled, I tried to make gcc10 default.

Things done

• Fix ld: multiple definition of (made gcc10 use -fcommon by default)
• Fix packages to work with -fnocommon
  • gdbm
• Fix python's _findLib_ld
• Fix libomxil-bellagio build
• Fix game-music-emu build (updated it)
• Fix rEFInd build (TODO: build it with gcc10)
• Fix dhcp build
• Fix pythonPackages.cffi (TODO: make tests pass)
• Fix pythonPackages.numpy (ImportError: liblapack.so.3: ELF load command address/offset not properly aligned)
• Fix systemd

Make clean fixes for

• rEFInd
• pythonPackages.cffi
• blas and lapack
• libopus
• qt512
• ncmpcpp
• libreoffice
• range-v3
• zimg

---

• Tested using sandboxing (nix.useSandbox on NixOS, or option sandbox in nix.conf on non-NixOS linux)
• Built on platform(s)
  • NixOS
  • macOS
  • other Linux distributions
• Tested via one or more NixOS test(s) if existing and applicable for the change (look inside nixos/tests)
• Tested compilation of all pkgs that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review wip"
• Tested execution of all binary files (usually in ./result/bin/)
• Determined the impact on package closure size (by running nix path-info -S before and after)
• Ensured that relevant documentation is up to date
• Fits CONTRIBUTING.md.

[FRidh]

We were at binutils 2.34 one or twice before but it got reverted.

[luc65r]

@FRidh We need to make sure that most of the things work before merging.
Is it fine to do multiple things in one PR? I don't want to compile my system 2 times, it takes already long enough.

[FRidh]

Please look at the past work that has been done on 2.34. This involves more work than you probably imagine right now.

[luc65r]

Yes, I saw that, I have a lot of free time (not a lot of skills), and I hope people will help.

[lovesegfault]

While I appreciate the effort to try and bump binutils, I have a few comments:

1. It's not great that you chose to rewrite the previous patches and submit them without any authorship of the work that was done before to try and get this to land.
2. Not sure the right approach is to go about reverting upstream commits. I recommend you go ask their mailing list so that, if not now in the future, we can avoid needing to maintain downstream patches.
3. If you need someone to test things that requires lots of rebuilds ping me on IRC.

[luc65r]

1. It's not great that you chose to rewrite the previous patches and submit them without any authorship of the work that was done before to try and get this to land.

Sorry, I was stupid.

2. Not sure the right approach is to go about reverting upstream commits. I recommend you go ask their mailing list so that, if not now in the future, we can avoid needing to maintain downstream patches.

You are right, should I try to patch python instead?

Also I'm a bit stuck with numpy (who would have thought...).

[luc65r]

That's not really a fix.

It isn't, but I do not have the technical abilities to make a proper fix, I don't understand the install phases of lapack and blas.
It enables me to try fixing other packages which depends on numpy.

[luc65r]

I'm really stuck on systemd...

This involves more work than you probably imagine right now.

I never thought systemd could brake so badly xD

It fails to start a lot a services like journald, logind or udevd with error code 127.
It seems related to binutils, as it has the same behavior compiled with gcc9 and gcc10.

Here is a strace systemctl restart systemd-journald on a minimal VM. I don't know why it is failing.

[fabianhjr]

Thanks @FRidh, @luc65r, et al. for working on this. :3

Will take a look at next-staging tomorrow and help stabilize it.

[FRidh]

Several aarch64 builds fail with is referenced by DSO followed by bin/ld: final link failed: bad value

ar: `u' modifier ignored since `D' is the default (see `U')
  CXXLD    libprotoc.la
libtool: warning: '/nix/store/spps8qis3wmbr5051g64l1lzv82d7zsc-gcc-10.2.0-lib/lib/libstdc++.la' seems to be moved
ar: `u' modifier ignored since `D' is the default (see `U')
  CXXLD    protoc
/nix/store/d3a24znms2blwv2fvqcrmrhwqq08f425-binutils-2.34/bin/ld: .libs/protoc: hidden symbol `__aarch64_cas8_rel' in /nix/store/9zvbxwpzfp9k6xl9px67y0l2m8xy2d1w-gcc-10.2.0/lib/gcc/aarch64-unknown-linux-gnu/10.2.0/libgcc.a(cas_8_3.o) is referenced by DSO
/nix/store/d3a24znms2blwv2fvqcrmrhwqq08f425-binutils-2.34/bin/ld: final link failed: bad value
collect2: error: ld returned 1 exit status
make[2]: *** [Makefile:3892: protoc] Error 1
make[2]: Leaving directory '/build/source/src'
make[1]: *** [Makefile:1874: all-recursive] Error 1
make[1]: Leaving directory '/build/source'
make: *** [Makefile:1781: all] Error 2
builder for '/nix/store/cl5fmbdqxnkvmfw21s5wwz2lrqpc72lq-protobuf-3.14.0.drv' failed with exit code 2

[luc65r]

I didn't test on aarch64 and darwin at all, we might have to force gcc9 on them for now.

[vcunat]

Darwin uses clang as default.

[flokli]

I can reproduce the build error with nix-build -A protobuf on a aarch64-linux box.

I can't reproduce it when cross-compiling from x86_64.

[kira-bruneau]

Thanks everyone for working on this!

I noticed two small issues building my system with 4008fd3:

• python-prctl fails on test_mpx even with pythonPackages.python-prctl: disabled tests causing build failure #106875 cherry-picked:
  It looks like mpx support should have been removed in GCC 9.1, so it's strange we're now only seeing this issue with GCC 10: https://www.phoronix.com/scan.php?page=news_item&px=MPX-Removed-From-GCC9
• ccls fails to build since llvmPackages_8.llvm fails to build:
  ccls-failure.log

I noticed that stripping was re-enabled on wine, so I will also confirm if binutils 2.34 fixes the original issue.

EDIT: strip with binutils 2.34 now keeps the original DOS stub instead of resetting it, so Wine builtin DLL & Wine placeholder DLL markers are no longer corrupted!

[FRidh]

cc @NixOS/exotic-platform-maintainers for assistance in fixing up aarch64. See #107783.

[FRidh]

possibly related aarch64 fix https://bugzilla.redhat.com/show_bug.cgi?id=1830472#c1

On aarch64 -moutline-atomics has been turned on by default, and those symbols are solely in libgcc.a, not in libgcc_s.so.*.
So, libgcc_s.so linker script is needed instead of libgcc_s.so symlink like on some other arches, gcc-10.0.1-0.14.fc33 with a fix is building in koji.

found via https://www.spinics.net/lists/fedora-devel/msg270592.html

[vcunat]

Sounds it could be "the same problem". Their patch: https://src.fedoraproject.org/rpms/gcc/c/95507e8b68

[FRidh]

They remove libgcc_s.so and then replace it with

GROUP ( /%{_lib}/libgcc_s.so.1 libgcc.a )' > $FULLPATH/libgcc_s.so

If I am correct libgcc_s.so is actually a statically linked file despite the .so, and this line here merges the two libraries into a new one, which is I think also what they meant with the message I quoted.

We have libgcc_s.so in the glibc store path, not in the gcc path. It is pulled in by the wrapper in the file /nix/store/y78inja64xaaq6bx33nxn2h1lgw0jswk-gcc-wrapper-10.2.0/nix-support/libc-crt1-cflags which contains

-B/nix/store/rqrklqsvw4ydpcg5kdcvn506fhcbqxk2-glibc-2.32-10/lib/

This is defined in pkgs/build-support/cc-wrapper/default.nix

echo "-B${libc_lib}${libc.libdir or "/lib/"}" >> $out/nix-support/libc-crt1-cflags

Note also the comment

# The "-B${libc_lib}/lib/" flag is a quick hack

[luc65r]

The tests for pkgsi686Linux.libopus pass with xnorpx/opus@6e2209d.