Ubuntu GCC 15 defaults to -Wa,-gsframe, causing "relocation refers to a discarded section: " linker error

[mstorsjo]

This can be reproduced by building LLVM itself, with GCC 15. Not sure if this specifically is caused by GCC 15 or something else. It can be reproduced on Ubuntu 25.10 at least.

It's easily reproducible in almost any environment, with the following docker file:

FROM ubuntu:25.10

RUN apt-get update -qq && \
    DEBIAN_FRONTEND="noninteractive" apt-get install -qqy --no-install-recommends \
    git build-essential lld cmake ninja-build ca-certificates python3 && \
    apt-get clean -y && \
    rm -rf /var/lib/apt/lists/*

WORKDIR /build

RUN git clone -b release/21.x --depth 1 https://github.com/llvm/llvm-project

RUN cd llvm-project/llvm && \
    mkdir build && \
    cd build && \
    cmake .. \
        -G Ninja \
        -DCMAKE_BUILD_TYPE=Release \
        -DLLVM_TARGETS_TO_BUILD=X86 \
        -DLLVM_USE_LINKER=lld && \
    ninja llc

Currently, this ends up with:

ld.lld: error: relocation refers to a discarded section: .text._ZN4llvm7remarks14EndOfFileErrorD2Ev
>>> defined in lib/libLLVMRemarks.a(YAMLRemarkParser.cpp.o)
>>> referenced by YAMLRemarkParser.cpp
>>>               YAMLRemarkParser.cpp.o:(.sframe+0x1c) in archive lib/libLLVMRemarks.a

[... multiple similar errors ...]

Compiling the source with Clang rather than with the distro-provided GCC 15, or by not passing -DLLVM_USE_LINKER=lld, avoids the issue.

[MaskRay]

This is a .sframe issue. I believe Ubuntu prematurely enabled -Wa,-gsframe by default for GCC despite severe linker garbage collection and COMDAT issues. sourceware.org/bugzilla/show_bug.cgi?id=33370 ("gas: monolithic .sframe violates COMDAT group rule")

The ELF provides mechanisms to support the desired linker garbage collection behavior, and it's essential to follow its rules. The current .sframe design requires significant code within linkers and every consumer, as seen with .eh_frame.

---

I can build llvm 21 with stock gcc 15. Perhaps the Ubuntu gcc defaults to -Wa,-gsframe.

cat > a.cc <<'eof'
[[gnu::noinline]] inline int inl() { return 0; }
auto *fa = inl;
eof
cat > b.cc <<'eof'
[[gnu::noinline]] inline int inl() { return 0; }
auto *fb = inl;
eof
~/opt/gcc-15/bin/g++ -Wa,--gsframe -c a.cc b.cc

% ld.lld a.o b.o
ld.lld: error: relocation refers to a discarded section: .text._Z3inlv
>>> defined in b.o
>>> referenced by b.cc
>>>               b.o:(.sframe+0x1c)
% gold a.o b.o
b.o(.sframe+0x1c): error: relocation refers to local symbol ".text._Z3inlv" [2], which is defined in a discarded section
  section group signature: "inl()"
  prevailing definition is from a.o

The generated sframe is invalid https://maskray.me/blog/2021-07-25-comdat-and-section-group#group-signature-localization

If a non-local symbol (e.g. __llvm_retpoline_r11) is defined relative to a section in a group, and referenced from outside the section group, it cannot be localized. The generic ABI says:

A symbol table entry with STB_LOCAL binding that is defined relative to one of a group's sections, and that is contained in a symbol table section that is not part of the group, must be discarded if the group members are discarded. References to this symbol table entry from outside the group are not allowed.

[mstorsjo]

I can build llvm 21 with stock gcc 15. Perhaps the Ubuntu gcc defaults to -Wa,-gsframe.

Yes, this does seem to be the root cause. Thanks!

cat > a.cc <<'eof'
[[gnu::noinline]] inline int inl() { return 0; }
auto *fa = inl;
eof
cat > b.cc <<'eof'
[[gnu::noinline]] inline int inl() { return 0; }
auto *fb = inl;
eof
~/opt/gcc-15/bin/g++ -Wa,--gsframe -c a.cc b.cc

% ld.lld a.o b.o
ld.lld: error: relocation refers to a discarded section: .text._Z3inlv
>>> defined in b.o
>>> referenced by b.cc
>>>               b.o:(.sframe+0x1c)

Indeed, this reproduces on current Ubuntu 25.10, by compiling the objects with a plain g++ -Wa,--gsframe -c a.cc b.cc. (For reference, I also tested with mold, and that doesn't error out on this case.)

The generated sframe is invalid https://maskray.me/blog/2021-07-25-comdat-and-section-group#group-signature-localization

If a non-local symbol (e.g. __llvm_retpoline_r11) is defined relative to a section in a group, and referenced from outside the section group, it cannot be localized. The generic ABI says:

A symbol table entry with STB_LOCAL binding that is defined relative to one of a group's sections, and that is contained in a symbol table section that is not part of the group, must be discarded if the group members are discarded. References to this symbol table entry from outside the group are not allowed.

Ok, I see. As you're more familiar with the concepts involved here, whom should the bug report go to; gcc or binutils? Can you file the relevant upstream bugs there? (And once such a bug is placed, I guess we should tell Ubuntu that things maybe aren't ready for enabling -Wa,--gsframe quite yet?)

[thesamesam]

Can you file the relevant upstream bugs there?

https://sourceware.org/PR33370

[thesamesam]

Note that Ubuntu reverted this change. I believe it was only enabled for a prerelease.

It was enabled by https://launchpad.net/ubuntu/+source/gcc-15/15.2.0-1ubuntu1 and disabled in https://launchpad.net/ubuntu/+source/gcc-15/15.2.0-4ubuntu2.