[lld] [MachO LLD] Respect -all_load with --start-lib --end-lib style archives (PR #93993)

[Fangrui Song via llvm-commits]

MaskRay wrote:

> > For Mach-O, it might make sense for the non-positional -all_load to override --start-lib.
> > But I expect more reasoning behind this change.
> 
> I think the flag being global makes a difference.
> 
> Since it is global, a user can supply the flag to a build system and expect it to work without inspecting the rest of the linker invocation. The same is not true of --whole-archive. The user already needs to see the entire invocation to position the —whole-archive / —no-whole-archive flags correctly, at which point they can just remove —start-lib / —end-lib as required. —We use —start-lib / —end-lib for distributed thin-lto. Users should be able to safely switch between thin-lto and distributed thin-lto without having to audit the custom flags they’ve provided (within reason).
> 
> To accomplish this without this change, the build system would need to inspect each linker invocation and remove —start-lib / —end-lib if -all_load is passed. That seems needlessly complex. It seems to me for a global flag, the behavior of a real archive, and —start-lib —end-lib should be the same.
> 
> > Btw, I'm curious how we handle the -force_load case (case OPT_force_load) for d-thinlto.
> 
> That is an attribute of the library being consumed (link_whole). Libraries with this attribute are passed to the linker without `--start-lib / --end-lib`.

Yes, I agree with the reasoning. It's worth capturing some of the thinking into the commit message (first comment).

https://github.com/llvm/llvm-project/pull/93993