[AArch64][ELF][lld] Support dynamic R_AARCH64_AUTH_* relocations

[kovdan01]

Summary

This patch adds support for dynamic R_AARCH64_AUTH_* relocations
(including RELR compressed AUTH relocations) as described here:
https://github.com/ARM-software/abi-aa/blob/main/pauthabielf64/pauthabielf64.rst#auth-variant-dynamic-relocations

Co-authored-by: Peter Collingbourne <[e-mail]>

[peter.smith]

Just a few small comments from me. Overall approach looks good to me. I expect MaskRay will want to take a look too.

lld/ELF/Relocations.cpp

1442

With -z rel as used by Fuchsia it can write out dynamic REL relocations but I think think this particular case is always from relocatable object files, which from clang/llvm-mc at least are always RELA.

Suggest changing the comment to:
// Assume relocations from relocatable objects are RELA.

1450

I think LLD would normally make this a flat if, else if, else. I've suggested some simplifications and comment changes below.

else if (part.relrAuthDyn && isInt<32>(sym.getVA(addend)))) {
  // Implicit addend is below 32-bits so we can use the compressed
  // relative relocation section. The R_AARCH64_AUTH_RELATIVE
  // has a smaller addend fielf as bits [63:32] encode the signing-schema.
  sec->addReloc({expr, type, offset, addend, &sym});
  part.relrAuthDyn->relocsVec[parallel::getThreadIndex()].push_back(
            {sec, offset});
} else {
   part.relaDyn->addReloc({R_AARCH64_AUTH_RELATIVE, sec, offset,
                            DynamicReloc::AddendOnlyWithTargetVA, sym,
                            addend, R_ABS});
}

1455

All AArch64 instructions require 4-byte alignment so sec->addralign >=2 and offset %2 == 0 will always be true otherwise code will get a fault at runtime.

For llvm it is unfortunate that a .section directive won't implicitly add the alignment so the user has to remember to add an align directive manually. This may lead to some assembler files having sections with sec->addralign being 1, but as long as these don't follow data sections (all AArch64 instructions are 4-bytes in size) then they effectively be 4-byte aligned.

[kovdan01]

• Rebase to current mainline and apply changes from D156505
• Fix review comments

lld/ELF/Relocations.cpp

1442

It makes sense, changed the comment.

1450

This is definitely better, changed, thanks.

1455

Got it, thanks

[kovdan01]

@MaskRay would be glad to see your comment on proposed changes

[MaskRay]

lld/ELF/Driver.cpp

1602

nopack-relative-auth-relocs is untested.

lld/ELF/Writer.cpp

402

lld/test/ELF/aarch64-ptrauth.s

4

omit -unknown-linux as this applies to generic ELF.

158

delete trailing blank line

llvm/include/llvm/BinaryFormat/DynamicTags.def

137

This suggests that the spec DT_AARCH64_AUTH_RELR* should be fixed first:) https://github.com/ARM-software/abi-aa/pull/214 has changed the values now.

[kovdan01]

Fix comments by @MaskRay

[kovdan01]

TODO: update the revision when https://reviews.llvm.org/D156505 is updated

lld/ELF/Driver.cpp

1602

It seems that it is, see lld/test/ELF/aarch64-ptrauth. Am I mistaken?

lld/ELF/Writer.cpp

402

Fixed

lld/test/ELF/aarch64-ptrauth.s

4

Fixed

158

Fixed

llvm/include/llvm/BinaryFormat/DynamicTags.def

137

Fixed, thanks!

[kovdan01]

Fix tests

[kovdan01]

Rebase to current mainline

[kovdan01]

@MaskRay A kind reminder regarding the revision. Would be glad to see your comments.

[kovdan01]

@MaskRay Would be glad to see your comments on the revision. It seems that the issues mentioned previously are resolved now.

[MaskRay]

llvm/tools/llvm-readobj/ELFDumper.cpp

2072

These llvm-readobj changes should be moved to D158574

2087

immediately invoked function expression

2103

if else if ... is better than immediately invoked function expression here

6233

processor-specific constants need to check isAArch64.