gcc: remove aligned_alloc hack on Darwin

[eliasnaur]

In a similar vein as PR #226048 this hack doesn't seem necessary anymore and introduces build differences for cross-compilation depending on build platform.

Description of changes

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandbox = true set in nix.conf? (See Nix manual)
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 23.05 Release Notes (or backporting 22.11 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
• Fits CONTRIBUTING.md.

[wegank]

Really? aligned_alloc isn't available on macOS 10.12. Have you tried to build stdenv on x86_64-darwin?

[eliasnaur]

Really? aligned_alloc isn't available on macOS 10.12. Have you tried to build stdenv on x86_64-darwin?

I built

nix build github:eliasnaur/nixpkgs/master#stdenv

successfully. https://github.com/eliasnaur/nixpkgs includes this PR: eliasnaur@b420b22.

[wegank]

Oh, sorry, I meant gcc:

$ nix build github:eliasnaur/nixpkgs/master#gcc --system x86_64-darwin

[eliasnaur]

Yes, this:

$ nix build github:eliasnaur/nixpkgs/master#gcc

completed successfully on an x86_64-darwin macbook running macOS 13.2.1.

[wegank]

Nice.

[trofi]

If I understand the comment the idea is to unconditionally disable the feature to make sure the final result does not depend on macos host. Do all macos versions now implement aligned_malloc? Do we use hermetic headers now? (AFAIU we don't). Or will we start requiring aligned_malloc for all macos users who pull from hydra?

[eliasnaur]

If I understand the comment the idea is to unconditionally disable the feature to make sure the final result does not depend on macos host. Do all macos versions now implement aligned_malloc? Do we use hermetic headers now? (AFAIU we don't). Or will we start requiring aligned_malloc for all macos users who pull from hydra?

Good points. I prepared #226290 which is a compromise: non-cross gcc should continue to run on macOS 10.12, but cross-gcc would produce the same output as Linux gcc (but require macOS 10.15 to run, something newer to build).

I must admit I'm surprised this is even an issue: why doesn't a cross gcc use the target standard library/headers for checking aligned_alloc? Is that a (fundamental) gcc problem or a nixpkgs problem? If the latter, I'd love to help fixing this properly.

[trofi]

If I understand the comment the idea is to unconditionally disable the feature to make sure the final result does not depend on macos host. Do all macos versions now implement aligned_malloc? Do we use hermetic headers now? (AFAIU we don't). Or will we start requiring aligned_malloc for all macos users who pull from hydra?

Good points. I prepared #226290 which is a compromise: non-cross gcc should continue to run on macOS 10.12, but cross-gcc would produce the same output as Linux gcc (but require macOS 10.15 to run, something newer to build).

I must admit I'm surprised this is even an issue: why doesn't a cross gcc use the target standard library/headers for checking aligned_alloc? Is that a (fundamental) gcc problem or a nixpkgs problem? If the latter, I'd love to help fixing this properly.

gcc runs multiple ./configure calls internally with various --build=/--host=/--target= options. Cor compiler itself it uses build's headers and/or host's headers. For target libraries it usually uses target headers.

It has more to do with limited export ac_cv_func_aligned_alloc=no interface that affects all of the ./configure calls at the same time.

[eliasnaur]

Superseded by #226290.