[PATCH] D59594: [ELF] Change GOT*_FROM_END (relative to end(.got)) to GOTPLT* (start(.got.plt))

[Fangrui Song via Phabricator via llvm-commits]

MaskRay marked an inline comment as done.
MaskRay added inline comments.


================
Comment at: ELF/Relocations.cpp:1073-1075
+    // These x86 and x86_64 relocation types are relative to GOTPLT but their
+    // description uses GOT (the address of the global offset table), thus also
+    // ensure GOT is created.
----------------
sivachandra wrote:
> ruiu wrote:
> > Is this logically x86-64 only? I think it is fine to create a .got when you create .got.plt, but the comment should explain that.
> One need not emit `.got` even if emitting `.got.plt`. For example, for non-PIE but statically linked excutables, ld.bfd does not emit `.got` but still emits `.got.plt`.
> 
> That said, I am not sure if not emitting a `.got` (when emitting `.got.plt`) for other kinds of executables is a good/bad thing. IMO, emitting a `.got` even when not required will only add a few bytes at most.
`R_GOTPLT, R_GOTPLTONLY_PC, and R_GOTPLTREL` are x86-32 x86-64 only. It currently emits `.got` because we need to compute the distance between `.got` and `.got.plt`. The empty `.got` will serve as an anchor.

I read some history in the past and I remember that `*_FROM_END` was chosen to make the GOT base in the middle of `.got`

AMD64 psABI p.75

> The symbol _GLOBAL_OFFSET_TABLE_ may reside in the middle of the .got section,
allowing both negative and non-negative offsets into the array of addresses.

But in reality that rule doesn't have to be followed strictly.. @sivachandra mentioned https://fedoraproject.org/wiki/Changes/.got.plt_Isolation Now I'm thinking ld.bfd may in some configuration allow `.got` and `.got.plt` to be separated. If that is true, The start of `.got.plt` would still be the GOT base as `end(.got)` would break things...


Repository:
  rLLD LLVM Linker

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D59594/new/

https://reviews.llvm.org/D59594