gcc: install info files serially

[raboof]

Description of changes

installing info files in parallel is dangerous, because install-info will write to a dir-file as a side-effect, and it has no protection against multiple install-info processes running in parallel and overwriting each others' changes.

Local fix until we can fix the Makefile.in generation upstream

Fixes #229470

Things done

Tested the concept locally with the following Makefile:

all: install-info

.NOTPARALLEL: install-info

install-info:: 1.to 2.to 3.to 4.to

%.to: %.from
	echo "converting $^"
	sleep 10
	echo "converted $^"

clean:
	rm *.to

• 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.

[trofi]

The workaround looks good. We will probably want to apply it to all gcc versions.

[raboof]

We will probably want to apply it to all gcc versions.

Are you sure we want this? I'd like to follow upstream as much as possible, and as for 12 we have a particular motivation to want it to be reproducible (e.g. we use this to build our installation media), it seems most conservative not to also apply it to the older versions?

[trofi]

We will probably want to apply it to all gcc versions.

Are you sure we want this? ... and as for 12 we have a particular motivation to want it to be reproducible (e.g. we use this to build our installation media), it seems most conservative not to also apply it to the older versions?

Out of curiosity who do you mean by "we"? I would normally expect all the nixpkgs packages should be determinitstic. The change is very safe and well-contained. If you are not comfortable of doing it I can also it to over gcc versions as a separate PR.

I'd like to follow upstream as much as possible

gcc is an example of very mangled packages in nixpkgs compared to what upstream suggests doing. I would say this patch does not make anything worse.

[raboof]

Out of curiosity who do you mean by "we"?

'NixOS', which is admittedly an ill-defined 'we' :)

I would normally expect all the nixpkgs packages should be determinitstic.

True, I was thinking "let's focus on the currently-widely-used things and not worry about older versions too much"

The change is very safe and well-contained. gcc is an example of very mangled packages in nixpkgs compared to what upstream suggests doing. I would say this patch does not make anything worse.

OK, I'll give it a go

[raboof]

OK, I'll give it a go

That applied surprisingly painlessly!

[trofi]

LGTM. All the $ nix build --no-link -f. gcc48 gcc49 gcc6 gcc7 gcc8 gcc9 gcc10 gcc11 gcc12 succeeded.

[ghost]

Hi, this is causing the entire gcc build to use only a single core.

According to the GNU Make manual, 5.4.1 Disabling Parallel Execution the patch added by this PR marks install-info and all of its prerequisites as needing to be built serially. I really don't think that's what we want to do.

[raboof]

Hi, this is causing the entire gcc build to use only a single core.

According to the GNU Make manual, 5.4.1 Disabling Parallel Execution the patch added by this PR marks install-info and all of its prerequisites as needing to be built serially. I really don't think that's what we want to do.

Ouch, that indeed was not the intent - I had read that documentation, but assumed it would only run the direct prerequisite builds in parallel, not everything transitively. I thought I had checked that but perhaps not sufficiently/correctly. Will look into this (unless someone beats me to it)

[trofi]

make-4.4 is expected to behave like that. The problem is that gcc in nixpkgs does not use gnumake from nixpkgs:

$ nix develop -f. gcc.cc
$ make --version
GNU Make 4.2.1

[trofi]

We discovered the same in https://gcc.gnu.org/PR109898#c4

[trofi]

enableParallelInstalling = false; is probably the best non-invasive alternative we have compared to serial build :(