[lld] [lld] Merge GOT entries for symbols (PR #131630)

[Peter Smith via llvm-commits]

smithp35 wrote:

May I suggest considering an alternative approach of further limiting ICF.

I share some of @MaskRay 's concerns about complexity. In particularly I'm nervous about use cases like signed GOT and CHERI where separate GOT entries may be required even for aliases.

I do agree that unless the ABI prohibits it, which would *require* the unwinder to be changed, the linker should be able to handle GOT generating relocations from ADRP, (ADD/LDR) where each is in a separate section.

The simplest thing that I can think of is to prevent folding of sections where there is an AArch64 GOT generating relocation type that is used in pairs. For example `R_AARCH64_ADR_GOT_PAGE`, `R_AARCH64_LD64_GOT_LO12_NC`. There are more described in https://github.com/ARM-software/abi-aa/blob/main/aaelf64/aaelf64.rst (candidates are those with GDAT, GTPREL), the TLSDESC sequence should never be split up (ABI rules).

This would inhibit ICF on code like:
```
extern const int x;
extern const int y;

// GOT generating relocations used for x and y
int fn()
{
    return x + y;
}
```
That seems like it could be worth benchmarking to see if that is significant limitation.

A more complex heuristic would be to record in `ICF<ELFT>::constantEq` all GOT generating ADRP relocations from the section. If there's a GOT generating ADD/LDR to S in the section, but we can't find a GOT generating ADRP relocation to S as well then prevent the section containing S from being folded. That would only prevent folding if assembler, or the outliner had split apart the ADRP and ADD/LDR.


https://github.com/llvm/llvm-project/pull/131630