Only add extra resolver spec group for Ruby platform when needed

[deivid-rodriguez]

What was the end-user or developer problem that led to this PR?

Our platform resolution logic rely on considering platform specific gems and "RUBY" gems different resolution candidates, because they may have different dependencies.

This works, but duplicates the number of candidates considered for every version, if the version has platform specific gems. This may seem harmless but may completely change resolution logic, because the number of candidates for a given dependency is actually used in order to decide which dependency the resolver should consider next. See

rubygems/bundler/lib/bundler/resolver.rb

Lines 188 to 202 in 342c7f8

	def sort_dependencies(dependencies, activated, conflicts)
	dependencies.sort_by do |dependency|
	name = name_for(dependency)
	vertex = activated.vertex_named(name)
	[
	@base[name].any? ? 0 : 1,
	vertex.payload ? 0 : 1,
	vertex.root? ? 0 : 1,
	amount_constrained(dependency),
	conflicts[name] ? 0 : 1,
	vertex.payload ? 0 : search_for(dependency).count,
	self.class.platform_sort_key(dependency.__platform),
	]
	end
	end

and

rubygems/bundler/lib/bundler/resolver.rb

Lines 237 to 259 in 342c7f8

	# returns an integer \in (-\infty, 0]
	# a number closer to 0 means the dependency is less constraining
	#
	# dependencies w/ 0 or 1 possibilities (ignoring version requirements)
	# are given very negative values, so they _always_ sort first,
	# before dependencies that are unconstrained
	def amount_constrained(dependency)
	@amount_constrained ||= {}
	@amount_constrained[dependency.name] ||= if (base = @base[dependency.name]) && !base.empty?
	dependency.requirement.satisfied_by?(base.first.version) ? 0 : 1
	else
	all = index_for(dependency).search(dependency.name).size
	if all <= 1
	all - 1_000_000
	else
	search = search_for(dependency)
	search = prerelease_specified[dependency.name] ? search.count : search.count {|s| !s.version.prerelease? }
	search - all
	end
	end
	end

This has caused some cases that worked fine before we started considering platforms no longer resolve, for example, https://github.com/duckinator/rubygems-4539-testcase.

What is your fix for the problem, implemented in this PR?

My fix is to only consider extra resolver groups when the dependencies for the platform specific variants of a particular gem differ from the dependencies of the RUBY variant.

This PR includes some other refactoring that led to figuring out this solution, and was extracted from #5960 to make things more easily reviewable.

Fixes #4539.

Make sure the following tasks are checked

• Describe the problem / feature
• Write tests for features and bug fixes
• Write code to solve the problem
• Make sure you follow the current code style and write meaningful commit messages without tags

[indirect]

Wow, nice investigation work. 👍🏻

[indirect]

I like the custom matcher, it's much nicer. 👍🏻 Looks good to me generally, and I like the way this simplified a lot of manually generating comparison criteria also.

[indirect]

the second conditional is now true 100% of the time, right? because generic local platform will always exist, so empty will always be false, so the condition will always be true?

[deivid-rodriguez]

If the current generic local platform is "ruby" and the dependency is for jruby (like gem "activerecord-jdbcsqlite3-adapter", :platform => :java), then d.gem_platforms([generic_local_platform]) will be empty (and the dependency will be filtered out).

[deivid-rodriguez]

Although on the other hand, I think that case is already filtered out by d.should_include?. Let me run the specs removing that second condition and see what fails (if anything).

[deivid-rodriguez]

When removing the second condition, only one spec fails (although maybe more will fail on Windows/JRuby CI, since I only run specs locally)

rubygems/bundler/spec/runtime/platform_spec.rb

Lines 364 to 379 in dd23fa1

	it "allows specifying only-ruby-platform on windows with dependency platforms" do
	simulate_windows do
	install_gemfile <<-G
	source "#{file_uri_for(gem_repo1)}"
	gem "nokogiri", :platforms => [:mingw, :mswin, :x64_mingw, :jruby]
	gem "platform_specific"
	G
	bundle "config set force_ruby_platform true"
	bundle "install"
	expect(the_bundle).to include_gems "platform_specific 1.0 RUBY"
	expect(the_bundle).to not_include_gems "nokogiri"
	end
	end

However, I'm not sure if that may be due to another issue, or whether the expectation may be wrong.

If that's ok, I'll reticket this to not block this PR on that.

[indirect]

Yep, totally ok. No need to remove the condition, since my reading on it was wrong anyway!

[deivid-rodriguez]

Awesome, I'm merging this now then!