[llvm-branch-commits] [lld] ELF: Only rewrite non-preemptible IFUNCs to IPLT functions if a non-IRELATIVE relocation is needed. (PR #133531)

[Peter Smith via llvm-branch-commits]

smithp35 wrote:

I have some small reservations about using ifunc resolvers like this. Mostly in that we are using a mechanism invented for a different purpose, and relying on some specific linker behaviour to make this case work.

This is similar to comments made in the Discourse post https://discourse.llvm.org/t/rfc-structure-protection-a-family-of-uaf-mitigation-techniques/85555/9 but repeating them here as this is closest to the implementation.

As I understand it, this has a more limited and more specific use case than ifuncs. Traditional ifuncs which can be address taken or called, possibly in multiple ways, so it makes sense to use a symbol type STT_GNU_IFUNC rather than special relocation directives. The initializers for structure field protection are compiler generated, can not be legally called or address taken from user code, and only have one relocation type R_*_ABS64 (or 32 on a 32-bit platform). With an addition of a single relocation, something like R_*_ADDRINIT64 which would target a STT_FUNC resolver symbol. We can isolate the structure field initialization use case from an actual ifunc.

I guess it all comes down to whether structure field initialization needs, or benefits from being distinguished from an ifunc. Ifuncs seem to be quite easy to get wrong so being able to isolate this case has some attraction to me at least. It also handles the structure field that points to an ifunc relatively gracefully.

As you pointed out in your response, this does mean adding 2 relocations to every psABI that supports structure field protection rather than just one. Although I'd expect the alternative of having relocations that alternatively write "directly call" ifunc resolver or take address of function might require new relocations too?

https://github.com/llvm/llvm-project/pull/133531