[lld] [lld][AArch64][ELF][PAC] Support `.relr.auth.dyn` section (PR #87635)

[Roland McGrath via llvm-commits]

frobtech wrote:

This is not at all the policy normally observed in LLVM.  Breakages that are not found by existing tests cannot be allowed to persist just because the test didn't exist before the breakage was landed.  It is quite normal for downstream reports of an actual regression in a real world use to be reason to revert.  It's not a sign that your change wasn't broken.  It's a sign that there was already insufficient testing, and your change did not add sufficient testing, to catch the real regression.

It is invalid output ever to produce an `ET_REL` file with any `SHT_REL` or `SHT_RELA` section whose `sh_info` and `sh_link` fields are not valid section indices for the symbol table and target sections.  This is why LLD refuses this as input.

There is a lack of existing testing coverage for `ld -r` output being correct in these ways, and a good test to add would be that `ld -r` output is accepted as `ld` input.  A general scheme of doing that test scenario for various kinds of links would ensure that bugs like the one introduced here are not easily overlooked.

https://github.com/llvm/llvm-project/pull/87635