unpack-bootstrap-tools.sh: use patchelf to undo nuke-refs and nothing else

[ghost]

Note: this PR includes two commits; the first commit is #169746 which needs to merge with or before this one in order to not break CI. Only the second commit is specific to this PR.

The current unpack-bootstrap-tools.sh wields patchelf selectively, guided by unexplained criteria. Specifically, it applies patchelf --set-interpreter to:

  $out/bin/* $out/libexec/gcc/*/*/*

(except */liblto*) and patchelf --set-interpreter --set-rpath to:

  $out/lib/librt-*.so $out/lib/libpcre*

The provenance of these lists is unclear. Some libraries which never had an rpath to begin with (like librt.so) are given an rpath by this script, while other libraries (like libbz2.so) are left with dangling /nix/store/eeee...'s.

This makes it very difficult to troubleshoot the behavior of the bootstrap-tools unpacker when porting to new platforms, where it is likely the case that patchelf is being debugged concurrently with nixpkgs.

This commit changes unpack-bootstrap-tools.sh to use patchelf only to replace /nix/store/eeee... references created by nuke-refs, and to replace all such references except where an exception is clearly described and explained. Currently there are two is one such exception:

1. libgcc_s.so is exempted from --set-rpath because its rpath leaks into the final stdenv, so we cannot patchelf its rpath without creating a fordbidden stdenv-final->bootstrap-tools requisite.
   2. libpthread.so and libc.so are exempted from --set-interpreter because the ELF interpreter of these files does not matter (they are libraries, not executable programs, and unlike ld.so they are not meant to play both roles), and because an unresolved bug in patchelf causes corruption if it is used to set the interpreter. Since setting the ELF interpreter of a library is seldom done (and indeed does not even need to be done here by nixpkgs), it is not likely that the patchelf bug will get fixed anytime soon, nor is it easy to argue for allocating resources to hunting it down. Edit: no longer required (see third patch in series)

Making sure to patchelf away all of the nuke-refs'd paths will make the bootstrapping process more robust and less fragile in general. In particular, it will prevent situations like #169746 where the only thing preventing a bug from manifesting was the fact that the arbitrary choice of which libraries to patchelf had been made in a particular way. Situations like that make things more fragile, because now the particular choice which used to be arbitrary has become mandatory, and neither that requirement nor the reason for it is documented anywhere.

Things done

• Built on platform(s)
  • x86_64-linux
  • mips64el-linux
  • aarch64-linux
  • powerpc64le-linux
• • nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD"
• Tested basic functionality of all binary files (usually in ./result/bin/)
• Fits CONTRIBUTING.md.

[SuperSandro2000]

@ofborg eval

[ghost]

This can wait until after staging closes tomorrow and reopens on 2022-May-15. But I would like to try to move forward with it at that point; it's been pending for a pretty long time now.

I've "minimized" bunch of my comments that either have been resolved or should be dealt with in a separate PR (if at all) in order to try to make the thread more readable to reviewers.

Sorry about bungling the base-branch change and all the noise/mass-review-requests caused by doing so.

[Mindavi]

In general this makes sense to me, but I don't feel that qualified to do a proper bootstrap review. I'm not at a pc now, but I'd like to do a build too to build some confidence in this. Planning to do that later when I find some time

[ghost]

I have significantly cleaned this up and rewritten the commit message to clarify the motivation (duplicated as the top comment in this PR).

Note that there are two commits in this PR; the first commit is #169746 which needs to be included here in order to pass CI. Only the second commit is specific to this PR.

[ghost]

Latest push just rebases on top of latest #169746; no other changes.