Helped on arm64 macOS (3 -> 1.5 hour).

14-x86_64 was slower (1 -> 1.75 hour); however, may just be an unlucky run as 13-x86_64 is the same and LLVM will limit the max jobs to number of logical cores. So 64GB / 5GB ~ 12 jobs should be ignored.

May try a couple more things

For comparison, Fedora approximates 6.114 GB/process though that seems to be applied to total processes like our MAKEFLAGS - https://src.fedoraproject.org/rpms/flang/blob/rawhide/f/flang.spec#_72.

EDIT: Though I think GCC + ld.bfd usually has worse memory usage than Clang + Apple ld / lld

Would be nice to read the available RAM somehow instead of just hard-coding this.

On macOS, you can get this with

sysctl -n hw.memsize

On Linux, you can try reading /proc/meminfo, though this might not work in GitHub Actions/Docker.

It should be scaled on available RAM as it sets a limit to total processes based on available ÷ input clamped to [1, logical_cores].

The "best" input values will likely change per release (though not enough to impact CI runners). I may try to get a more accurate approximation by doing local builds but would still need to hard-code it.

Something like this (generated by ChatGPT, but tested locally and in a codespace):

def total_ram_bytes
  if OS.mac?
    Utils.safe_popen_read("sysctl", "-n", "hw.memsize").to_i
  elsif OS.linux?
    if File.readable?("/proc/meminfo")
      mem_kib = File.read("/proc/meminfo")[/MemTotal:\s+(\d+)\s+kB/, 1]
      mem_kib.to_i * 1024
    else
      pages = Utils.safe_popen_read("getconf", "_PHYS_PAGES").to_i
      page_size = Utils.safe_popen_read("getconf", "PAGE_SIZE").to_i
      pages * page_size
    end
  else
    odie "Unsupported OS for RAM detection"
  end
end

We can probably drop the getconf fallback on Linux, and maybe add it if finding the MemTotal entry of /proc/meminfo doesn't work.