[CSL-1002] Add lld linker for macOS

[chshersh]

This commit is a part of CSL-1002 issues: speeding up linking to build project faster. gold linker is already added. Though there're some problems with adding lld. Waiting for @neongreen and @nh2 to comment.

[mention-bot]

@chshersh, thanks for your PR! By analyzing the history of the files in this pull request, we identified @neongreen, @flyingleafe and @volhovm to be potential reviewers.

[rockbmb]

Are we using lld in macOS because it doesn't have gold?

[neongreen]

And also because it's actually faster than gold.

[dcci]

I don't think lld is in a state where can be used on Mach-O, see my comment.
Also, it's not necessarily faster than gold. In many tests we run it goes from parity to 2x faster (in particular scenarios, e.g. when linking with debuginfo or with some options (e.g. -gdb-index). It's a safe claim to say that's generally at least as fast as gold.

[gromakovsky]

It's approved, but as we know it doesn't work. Now everybody can merge it and I am afraid that somebody will do it, so I am rejecting it just in case.

[deepfire]

What is the status of this, @gromakovsky? We need to know if this can be used to speed up Hydra builds.

[gromakovsky]

@deepfire we are already using gold for linux builds (in master). We are not using lld for MacOS because it doesn't work (that's the reason this PR is still open). Niklas can help us with it if it's a priority.

[neongreen]

Looks like we won't need it after switching to GHC 8.2:

 * GHC now tries to use the gold and lld linkers by default. These
   linkers are significantly faster than the BFD linker implementation
   that most Linux distributions use by default. If gold or lld are not
   available GHC will use the system's default linker. GHC can be forced
   to use the default linker by passing --disable-ld-override to
   configure.

[domenkozar]

Afaik this one depends on NixOS/nixpkgs#24744

[dcci]

LLVM/lld dev here. My $0.02 here.
I'm afraid there's a little of misunderstanding probably derived from the original scope of this project.
Originally, lld was an unified linker for Mach-O/COFF/ELF. This experiment was thrown away and now there are three separate backends (i.e. one for each target).

The MachO backend hasn't been pretty much touched since the split. All the timings for haskell, as far as I can tell have been taken on ELF. I'll be very surprised if the Mach-O backend links anything larger than an hello world (I'm pretty sure it doesn't self-host, because I tried myself). As such, I think having lld has a viable alternative on Nix (MacOS) would require a substantial amount of work and it's definitely not ready for production.

[dcci]

I don't think this is a good move :)

[dcci]

I don't think lld is in a state where can be used on Mach-O, see my comment.
Also, it's not necessarily faster than gold. In many tests we run it goes from parity to 2x faster (in particular scenarios, e.g. when linking with debuginfo or with some options (e.g. -gdb-index). It's a safe claim to say that's generally at least as fast as gold.

[dcci]

Oops, sorry, I thought this was for NixOS. Ignore my comment (although my reasoning is still valid :)