thread_local int a; int main() { return a; } % g++ -fno-plt -fpic a.cc -Wa,-mrelax-relocations=no -fuse-ld=bfd /usr/bin/ld.bfd: /tmp/ccLggG8h.o: TLS transition from R_X86_64_TLSGD to R_X86_64_GOTTPOFF against `a' at 0x8 in section `.text' failed /usr/bin/ld.bfd: final link failed: Nonrepresentable section on output collect2: error: ld returned 1 exit status static thread_local int a; int main() { return a; } % g++ -fno-plt -fpic a.cc -Wa,-mrelax-relocations=no -fuse-ld=bfd /usr/bin/ld.bfd: /tmp/ccO9KvlL.o: TLS transition from R_X86_64_TLSGD to R_X86_64_TPOFF32 against `_ZL1a' at 0x8 in section `.text' failed /usr/bin/ld.bfd: final link failed: Nonrepresentable section on output collect2: error: ld returned 1 exit status Due to -Wa,-mrelax-relocations=no, R_X86_64_GOTPCREL instead of R_X86_64_GOTPCRELX is generated. a9: callq *(%rip) 00000000000000ad: R_X86_64_GOTPCREL __tls_get_addr-4 // bfd/elf-x86-64.c:elf_x86_64_check_tls_transition ... if (largepic) return r_type == R_X86_64_PLTOFF64; else if (indirect_call) return r_type == R_X86_64_GOTPCRELX; // It should allow R_X86_64_GOTPCREL else return (r_type == R_X86_64_PC32 || r_type == R_X86_64_PLT32);
Since the new linker is required to support no-PLT TLS sequence, R_X86_64_GOTPCRELX should be used here. R_X86_64_GOTPCREL won't work with the old linker anyway.
(In reply to H.J. Lu from comment #1) > Since the new linker is required to support no-PLT TLS sequence, > R_X86_64_GOTPCRELX should be used here. R_X86_64_GOTPCREL won't > work with the old linker anyway. I agree, but for the sake of completeness there is still value supporting GOTPCREL. https://sourceware.org/bugzilla/show_bug.cgi?id=20216 GOTPCREL was added to gold in commit 9b3777ae068e67e08b27914abbab11297f1f8d1b.
This was discovered by a Rust user. They use clang integrated assembler (not GNU as) with GNU ld. The integrated assembler isn't configured with -Wa,-mrelax-relocations=yes on by default. I worked around that in https://reviews.llvm.org/D64304 but it'd be nice if binutils 2.33 (or 2.34) will fix the issue.
(In reply to Fangrui Song from comment #3) > This was discovered by a Rust user. They use clang integrated assembler (not > GNU as) with GNU ld. The integrated assembler isn't configured with > -Wa,-mrelax-relocations=yes on by default. I worked around that in > https://reviews.llvm.org/D64304 but it'd be nice if binutils 2.33 (or 2.34) > will fix the issue. The bug should be fixed in clang.
I don't see how this bug is "invalid". I've already worked around the bug in clang, but I hope this hack can be deleted in a few years. > g++ -fno-plt -fpic a.cc -Wa,-mrelax-relocations=no -fuse-ld=bfd I mentioned the combination of -fno-plt and -Wa,-mrelax-relocations=no doesn't work. So we have 3 resolutions: 1. Ignore it. gcc -fno-plt -fpic -Wa,-mrelax-relocations=no has the linker error 2. Document in Chapter 11 Alternate Code Sequences For Security of x86-64 psABI, call *__tls_get_addr@GOTPCREL(%rip) must use R_X86_64_GOTPCRELX. This kinda justifies 1., though it is not obvious why R_X86_64_GOTPCREL can't be used. 3. One line change to bfd/elf-x86-64.c:elf_x86_64_check_tls_transition - return r_type == R_X86_64_GOTPCRELX; // It should allow R_X86_64_GOTPCREL + return r_type == R_X86_64_GOTPCREL || r_type == R_X86_64_GOTPCRELX; https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=9b3777ae068e67e08b27914abbab11297f1f8d1b With 3, the `gcc -fno-plt -fpic -Wa,-mrelax-relocations=no` issue can be fixed.
The psABI also says For the occurrence of name@GOTPCREL in the following assembler instructions: call *name@GOTPCREL(%rip) ... the R_X86_64_GOTPCRELX relocation, or the R_X86_64_REX_GOTPCRELX relocation if the REX prefix is present, should be generated. R_X86_64_GOTPCRELX should be generated for: call __tls_get_addr@GOTPCREL(%rip)
> For the occurrence of name@GOTPCREL in the following assembler instructions: > call *name@GOTPCREL(%rip) I agree, yet it also says: > (PIC) the base of GOT is within 2GB an indirect call to the GOT entry might be implemented like so: > foo (); call *(foo@GOT); R_X86_64_GOTPCREL :)
(In reply to Fangrui Song from comment #7) > > For the occurrence of name@GOTPCREL in the following assembler instructions: > > > call *name@GOTPCREL(%rip) > > I agree, yet it also says: > > > (PIC) the base of GOT is within 2GB an indirect call to the GOT entry might be implemented like so: > > foo (); call *(foo@GOT); R_X86_64_GOTPCREL > > :) There is a typo. Relocation should be R_X86_64_GOT32, not R_X86_64_GOTPCREL.
rustc runs into this problem as well as it currently somehow defaults to disable R_X86_64_*GOTPCRELX/R_386_GOT32X. Since the fix is so trivial, I sent https://sourceware.org/pipermail/binutils/2023-January/125506.html https://sourceware.org/pipermail/binutils/2023-January/125507.html
Since the new TLS sequence was added after R_X86_64_GOTPCRELX was required for call, R_X86_64_GOTPCREL is invalid in this TLS sequence.
The master branch has been updated by Fangrui Song <maskray@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=d58854b6dd88e05dbf2a5d1c32c5acb7bd6ea274 commit d58854b6dd88e05dbf2a5d1c32c5acb7bd6ea274 Author: Fangrui Song <maskray@google.com> Date: Fri Mar 10 11:53:32 2023 -0800 ld: Allow R_X86_64_GOTPCREL for call *__tls_get_addr@GOTPCREL(%rip) _Thread_local int a; int main() { return a; } % gcc -fno-plt -fpic a.c -fuse-ld=bfd -Wa,-mrelax-relocations=no /usr/bin/ld.bfd: /tmp/ccSSBgrg.o: TLS transition from R_X86_64_TLSGD to R_X86_64_GOTTPOFF against `a' at 0xd in section `.text' failed /usr/bin/ld.bfd: failed to set dynamic section sizes: bad value collect2: error: ld returned 1 exit status This commit fixes the issue. There is an argument that the -fno-plt TLS sequence was added after R_X86_64_GOTPCRELX was required for call, so R_X86_64_GOTPCREL was intended to be unsupported. Unfortunately this standpoint has caused interop difficulty: some projects specify -mrelax-relocations=no to build relocatable object files compatible with older linkers (e.g. https://github.com/IHaskell/IHaskell/issues/636) or do so by accident (e.g. https://github.com/rust-lang/rust/pull/106511 not addressed as of today). Many uses have not been cleaned up in practice, and compiling with -fno-plt will lead to the `TLS transition from R_X86_64_TLSGD ...` error which is hard to reason about. There is another argument which may be weaker but relevant to the necessity of -mrelax-relocations=no: HWAddressSanitizer x86-64 will likely need some assembler support to disable relaxation. Without the support and if the compiler needs to support many gas version, the simplest solution would be to use -Wa,-mrelax-relocations=no. PR ld/24784 * bfd/elf64-x86-64.c (elf_x86_64_check_tls_transition): Allow R_X86_64_GOTPCREL.
The master branch has been updated by Fangrui Song <maskray@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=2cef48423031d23d319692ceced5ad0110625f69 commit 2cef48423031d23d319692ceced5ad0110625f69 Author: Fangrui Song <maskray@google.com> Date: Fri Mar 10 11:55:09 2023 -0800 ld: Allow R_386_GOT32 for call *__tls_get_addr@GOT(%reg) Similar to d58854b6dd88e05dbf2a5d1c32c5acb7bd6ea274 for x86_64. _Thread_local int a; int main() { return a; } % gcc -m32 -fno-plt -fpic a.c -fuse-ld=bfd -Wa,-mrelax-relocations=no /usr/bin/ld.bfd: /tmp/ccR8Yexy.o: TLS transition from R_386_TLS_GD to R_386_TLS_IE_32 against `a' at 0x15 in section `.text' failed /usr/bin/ld.bfd: failed to set dynamic section sizes: bad value collect2: error: ld returned 1 exit status This commit fixes the issue. There is an argument that the -fno-plt TLS sequence was added after R_386_GOT32X was required for call *func@GOT(%ebx), so R_386_GOT32 was intended to be unsupported. Unfortunately this standpoint has caused interop difficulty: some projects specify -mrelax-relocations=no to build relocatable object files compatible with older linkers (e.g. https://github.com/IHaskell/IHaskell/issues/636) or do so by accident (e.g. https://github.com/rust-lang/rust/pull/106511 not addressed as of today). Many uses have not been cleaned up in practice, and compiling with -fno-plt will lead to the `TLS transition from R_386_TLS_GD ...` error which is hard to reason about. It seems easier to apply this simple change to prevent the footgun. PR ld/24784 * bfd/elf32-i386.c (elf_i386_check_tls_transition): Allow R_386_GOT32.
Fixed for 2.41