Ruby

A MinGW build of ruby 2.5.0dev (2017-04-18 trunk 58397) [x64-mingw32] had a silent SEGV when running TestLambdaParameters#test_call_simple in test/ruby/test_lambda.rb.

The silent SEGV occurs on the following two commented lines:

  def test_call_simple
    assert_equal(1, lambda{|a| a}.call(1))
    assert_equal([1,2], lambda{|a, b| [a,b]}.call(1,2))
    assert_raise(ArgumentError) { lambda{|a|}.call(1,2) }
    assert_raise(ArgumentError) { lambda{|a|}.call() }
    assert_raise(ArgumentError) { lambda{}.call(1) }
#   assert_raise(ArgumentError) { lambda{|a, b|}.call(1,2,3) }

    assert_equal(1, ->(a){ a }.call(1))
    assert_equal([1,2], ->(a,b){ [a,b] }.call(1,2))
    assert_raise(ArgumentError) { ->(a){ }.call(1,2) }
    assert_raise(ArgumentError) { ->(a){ }.call() }
    assert_raise(ArgumentError) { ->(){ }.call(1) }
#   assert_raise(ArgumentError) { ->(a,b){ }.call(1,2,3) }
  end

This ocurred when running make test-all or in a test framework run directly from ruby.

But, the following runs fine, via either irb or ruby -e:

20.times { begin ; lambda{|a, b|}.call(1,2,3) ; rescue => e ; puts e.message ; end }

When tested with ruby 2.5.0dev (2017-04-17 trunk 58383) [x64-mingw32], it ran fine regardless of the test type.

This would seem to imply something is odd with the test framework, and not the actual test.

I've been building for while (trying to establish stable results), and many builds have gone fine using make -j3. My main desire was to get stable/repeatable test-all results.

The last two days, I had several builds that failed near the end with some form of rb_sysopen issue, either Permission denied or Invalid argument. When I switched to make -j1, I believe almost of all of the builds completed. I should have kept better logs, but I was getting a bit frustrated...

But now, the current svn 58397 build was done with make -j3. Not sure what changed...

Recently, the autoclose issue came up in discussion. Since I started building, I've need to add sock.autoclose = false to the starred line in TestSocket#test_closed_read in ruby/test/socket/test_socket.rb. Without it, another silent SEGV both inside of make test-all and using the ruby framework.

  def test_closed_read
    require 'timeout'
    require 'socket'
    bug4390 = '[ruby-core:35203]'
    server = TCPServer.new("localhost", 0)
    serv_thread = Thread.new {server.accept}
    begin sleep(0.1) end until serv_thread.stop?
*   sock = TCPSocket.new("localhost", server.addr[1]) ; sock.autoclose = false
    client_thread = Thread.new do
      sock.readline
    end
    begin sleep(0.1) end until client_thread.stop?
    Timeout.timeout(1) do
      sock.close
      sock = nil
      assert_raise(IOError, bug4390) {client_thread.join}
    end
  ensure
    serv_thread.value.close
    server.close
  end

Finally, I still require the patch for test/ruby/test_enum.rb, as mentioned in Bug 13298. This is odd because simply adding the to_s stops the SEGV...

Sorry for mentioning all of these items, but I can't help but believe they may be connected, or at least point to some underlying issue with Windows/MinGW builds...