lld mishandling R_X86_64_PC32 relocations

Hi,

I've tracked down what I believe is a bug in lld's relocation processing for R_X86_64_PC32 REL relocations.

I'm producing the object file in a slightly unusual way: I'm using objcopy on a relocatable i386 ELF object file to convert it to x86_64 which transforms a R_386_PC32 into a R_X86_64_PC32.

Steps to reproduce:

1. Assemble the attached bug.asm using nasm and note the R_386_PC32 REL.

$ nasm -felf32 -o bug.o bug.asm
$ x86_64-elf-objdump -dr bug.o | grep -A1 -e '<_start>:'
00000000 <_start>:
   0: e8 fc ff ff ff call 1 <_start+0x1>
$ x86_64-elf-readelf -r bug.o

Relocation section '.rel.text._start' at offset 0x260 contains 1 entry:
Offset Info Type Sym.Value Sym. Name
00000001 00000302 R_386_PC32 00000000 .text.foo

2. Convert bug.o to a 64-bit ELF object file and note the R_X86_64_PC32 REL (not RELA!)

$ x86_64-elf-objcopy -I elf32-i386 -O elf64-x86-64 bug.o bug-64.o
$ x86_64-elf-objdump -M i386 -dr bug-64.o | grep -A1 -e '<_start>:'
0000000000000000 <_start>:
   0: e8 fc ff ff ff call 1 <_start+0x1>
$ x86_64-elf-readelf -r bug-64.o

Relocation section '.rel.text._start' at offset 0x128 contains 1 entry:
  Offset Info Type Sym. Value Sym. Name
000000000001 000300000002 R_X86_64_PC32 0000000000000000 .text.foo

3. Link with a just-built ld.lld and note that the relocation has been misapplied. It's now calling foo+4 rather than foo.

$ ./llvm-build/bin/ld.lld -melf_x86_64 -o bug-lld bug-64.o
$ x86_64-elf-objdump -M i386 -d bug-64-lld | grep -A1 -e '<_start>:'
0000000000201000 <_start>:
  201000: e8 0f 00 00 00 call 201014 <foo+0x4>

If you link with GNU ld instead, the relocation is applied correctly.

$ x86_64-elf-ld -melf_x86_64 -o bug-64-ld bug-64.o
$ x86_64-elf-objdump -M i386 -d bug-64-ld | grep -A1 -e '<_start>:'
0000000000400080 <_start>:
  400080: e8 0b 00 00 00 call 400090 <foo>

Linking the 32-bit object file works correctly with both GNU ld and ld.lld.

$ ./llvm-build/bin/ld.lld -melf_i386 -o bug-lld bug.o
$ x86_64-elf-objdump -d bug-lld | grep -A1 -e '<_start>:'
00011000 <_start>:
   11000: e8 0b 00 00 00 call 11010 <foo>
$ x86_64-elf-ld -melf_i386 -o bug-ld bug.o
$ x86_64-elf-objdump -d bug-ld | grep -A1 -e '<_start>:'
08048060 <_start>:
8048060: e8 0b 00 00 00 call 8048070 <foo>

I'm not at all familiar with the lld source, but this looks a lot like getImplicitAddend() needs to be implemented for the X86_64 class.

Alternatively (but much less useful for me) would be an error message that REL relocations are not supported on x86-64.

I've attached all of the files created above, in case anyone wants to examine them.

Thank you,

Steve

bug-64-ld (848 Bytes)

bug-64-lld (8.61 KB)

bug-64.o (816 Bytes)

bug-ld (604 Bytes)

bug-lld (8.42 KB)

bug.asm (229 Bytes)

bug.o (624 Bytes)

Hi Stephen,

I think the bug is in objcopy, as it is creating a file that does not comply with the x86-64 psABI, which requires all relocations to be RELA:
https://github.com/hjl-tools/x86-psABI/blob/hjl/master/object-files.tex#L429

I also tried passing your bug-64.o to ld.gold, and it rejects it with an internal error.

$ ld.gold -r -o bug-64-2.o bug-64.o
ld.gold: internal error in scan_relocatable_relocs, at …/…/gold/x86_64.cc:5118
$ ld.gold -o bug-64-2.o bug-64.o
ld.gold: error: bug-64.o: unsupported REL reloc section
ld.gold: internal error in relocate_section, at …/…/gold/x86_64.cc:5051

Probably what needs to happen is that we need to start rejecting files which use the wrong relocation type and a bug needs to be filed against objcopy.

Peter

Hi Peter,

I hadn't noticed that line about RELA in the ABI document, thanks! I agree that this is a bug in objcopy. I'll file a bug against objcopy.

It seems that lld already supports relocations that don't comply with the standard (namely R_X86_64_16 and R_X86_64_8 <https://github.com/hjl-tools/x86-psABI/blob/hjl/master/object-files.tex#L554>), but I guess that's different than supporting an entirely different format.

Thanks for looking into this.

Steve

Hi Peter,

I hadn’t noticed that line about RELA in the ABI document, thanks! I agree that this is a bug in objcopy. I’ll file a bug against objcopy.

It seems that lld already supports relocations that don’t comply with the standard (namely R_X86_64_16 and R_X86_64_8 <https://github.com/hjl-tools/x86-psABI/blob/hjl/master/object-files.tex#L554>), but I guess that’s different than supporting an entirely different format.

lld supports the extension because LLVM supports the extension too (https://llvm.org/docs/Extensions.html#x86). But yeah, as you said, that’s a small extension to the standard, and that’s different from supporting an entirely different format.