Feature #15663: Documenting autoload semantics

• Status: Open
• Assignee: -
• Target version: -

[Eregon (Benoit Daloze)]

The semantics of autoload are extremely complicated.

As far as I can see, they are unfortunately not documented.

ruby/spec tries to test many aspects of it, test/ruby/test_autoload.rb has a few tests, and e.g. zeitwerk tests some other parts.
One could of course read the MRI source code, but I find it very hard to follow around autoload.

For the context, I'm trying to implement autoload as correct as possible in TruffleRuby and finding it very difficult given the inconsistencies (see below) and lack of documentation.

There is nowhere a document on how it should behave, and given the complexity of it I am not even sure MRI behaves as expected.
Could we create this document?
For instance, there is such a document for refinements.

Here is an example how confusing autoload can be, and I would love to hear the rationale or have some written semantics on why it is that way.

main.rb:

require "pp"

$: << __dir__

Object.autoload(:Foo, "foo")

CHECK = -> state {
  checks = -> {
    {
      defined: defined?(Foo),
      const_defined: Object.const_defined?(:Foo),
      autoload?: Object.autoload?(:Foo),
      in_constants: Object.constants.include?(:Foo),
    }
  }

  pp when: state, **checks.call, other_thread: Thread.new { checks.call }.value
}

CHECK.call(:before_require)

if ARGV.first == "require"
  require "foo"
else
  Foo # trigger the autoload
end

CHECK.call(:after)

p Foo

foo.rb:

CHECK.call(:during_before_defining)

module Foo
end

CHECK.call(:during_after_defining)

Here are the results for MRI 2.6.1:

$ ruby main.rb        
{:when=>:before_require,
 :defined=>"constant",
 :const_defined=>true,
 :autoload?=>"foo",
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>"foo",
   :in_constants=>true}}
{:when=>:during_before_defining,
 :defined=>nil,
 :const_defined=>false,
 :autoload?=>nil,
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>"foo",
   :in_constants=>true}}
{:when=>:during_after_defining,
 :defined=>"constant",
 :const_defined=>true,
 :autoload?=>nil,
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>"foo",
   :in_constants=>true}}
{:when=>:after,
 :defined=>"constant",
 :const_defined=>true,
 :autoload?=>nil,
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>nil,
   :in_constants=>true}}
Foo

Looking at during_before_defining, the constant looks not defined during the autoload for the Thread loading it, but looks defined and as an autoload for other threads.

Now we can discover other subtle semantics, by using require on the autoload file instead of accessing the constant:

$ ruby main.rb require 
{:when=>:before_require,
 :defined=>"constant",
 :const_defined=>true,
 :autoload?=>"foo",
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>"foo",
   :in_constants=>true}}
{:when=>:during_before_defining,
 :defined=>nil,
 :const_defined=>false,
 :autoload?=>nil,
 :in_constants=>true,
 :other_thread=>
  {:defined=>nil, :const_defined=>false, :autoload?=>nil, :in_constants=>true}}
{:when=>:during_after_defining,
 :defined=>"constant",
 :const_defined=>true,
 :autoload?=>nil,
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>nil,
   :in_constants=>true}}
{:when=>:after,
 :defined=>"constant",
 :const_defined=>true,
 :autoload?=>nil,
 :in_constants=>true,
 :other_thread=>
  {:defined=>"constant",
   :const_defined=>true,
   :autoload?=>nil,
   :in_constants=>true}}
Foo

Looking at during_before_defining, now the other threads seem to see the constant not defined, although it is still in Object.constants.
But of course, the constant cannot be removed, as otherwise that would not be thread-safe and other threads would raise NameError when accessing the constant.
In fact, we can see other threads actually wait for the constant, by changing to Thread.new { Foo; checks.call }, and then we get a deadlock:

Traceback (most recent call last):
	2: from main.rb:20:in `<main>'
	1: from main.rb:17:in `block in <main>'
main.rb:17:in `value': No live threads left. Deadlock? (fatal)
3 threads, 3 sleeps current:0x00007f0124004cb0 main thread:0x000055929cc2c470
* #<Thread:0x000055929cc5b348 sleep_forever>
   rb_thread_t:0x000055929cc2c470 native:0x00007f013381d700 int:0
   main.rb:17:in `value'
   main.rb:17:in `block in <main>'
   main.rb:20:in `<main>'
* #<Thread:[e-mail]:17 sleep_forever>
   rb_thread_t:0x000055929ce026d0 native:0x00007f0129007700 int:0
    depended by: tb_thread_id:0x000055929cc2c470
   main.rb:17:in `value'
   main.rb:17:in `block in <main>'
   foo.rb:1:in `<top (required)>'
   /home/eregon/.rubies/ruby-2.6.1/lib/ruby/2.6.0/rubygems/core_ext/kernel_require.rb:54:in `require'
   /home/eregon/.rubies/ruby-2.6.1/lib/ruby/2.6.0/rubygems/core_ext/kernel_require.rb:54:in `require'
   main.rb:17:in `block (2 levels) in <main>'
* #<Thread:[e-mail]:17 sleep_forever>
   rb_thread_t:0x00007f0124004cb0 native:0x00007f0128e05700 int:0
    depended by: tb_thread_id:0x000055929ce026d0
   main.rb:17:in `block (2 levels) in <main>'

This is quite weird. Is the second behavior a bug?
Why should other threads suddenly see the constant as "not defined" while it is loading via require in the main thread?
It's also inconsistent with the first case.

I would have thought require autoload_path would basically do the same as triggering the autoload of the constant (such as Foo). But the results above show they differ.

There are many more complex cases for autoload, such as this spec, or how is thread-safety is achieved when methods are defined incrementally in Ruby but the module is defined immediately.

Who is knowledgeable about autoload and could answer these questions?
Could we start a document specifying the semantics?

[shevegen (Robert A. Heiler)]

May explain why matz wants to remove autoload in the long run. :)

[akr (Akira Tanaka)]

Eregon (Benoit Daloze) wrote:

I would have thought require autoload_path would basically do the same as triggering the autoload of the constant (such as Foo). But the results above show they differ.

Agreed.

I think no one seriously considered about requiring a library used for autoload.

Who is knowledgeable about autoload and could answer these questions?
Could we start a document specifying the semantics?

I think we should start to make autoload semantics simpler by introducing "global autoload lock"
as I described in https://bugs.ruby-lang.org/issues/15598 .
This is needed because ruby doesn't (cannot) know dependencies of autoloaded libraries before loading.
It makes autoload related procedure single threaded which is much simpler than multi threads.

[akr (Akira Tanaka)]

akr (Akira Tanaka) wrote:

Could we start a document specifying the semantics?

I think we should start to make autoload semantics simpler by introducing "global autoload lock"
as I described in https://bugs.ruby-lang.org/issues/15598 .

Of course, there is no problem to describe the single thread semantics.

[Eregon (Benoit Daloze)]

akr (Akira Tanaka) wrote:

Eregon (Benoit Daloze) wrote:

I would have thought require autoload_path would basically do the same as triggering the autoload of the constant (such as Foo). But the results above show they differ.

Agreed.

It seems tricky implementation-wise, constant resolution (e.g. #const_get) has the constant data structure, and #require has the expanded file path + related lock but none of them has both.

I was thinking maybe the require lock per path should be used for everything, but then since #autoload calls require dynamically, how to keep track which thread is loading the constant and so should observe the autoload constant as "not defined" while loading it, without deadlocks?
Other threads might try to load the constant too, or require the autoload path, and only one thread should be the loading thread for that constant.

I'd be tempted for constant resolution to basically do nothing more than call require, but then we need Kernel#require when it starts loading the file to also mark the autoload constant as being loaded by the current thread (such that the constant looks as "not defined").
How to pass that information (e.g., the constant data structure) from constant resolution down to Kernel#require?
The require feature could be changed by user-defined require. And a user-defined require might very well require other files for its own logic (e.g., rubygems files).

I think no one seriously considered about requiring a library used for autoload.

lib/net/http.rb has autoload :OpenSSL, 'openssl' and therefore just require "net/http"; require "openssl" produces such a case.

I think we should start to make autoload semantics simpler by introducing "global autoload lock"
as I described in https://bugs.ruby-lang.org/issues/15598 .
This is needed because ruby doesn't (cannot) know dependencies of autoloaded libraries before loading.
It makes autoload related procedure single threaded which is much simpler than multi threads.

That would simplify things, but then I think it would also need to be a global require lock, and I think that can be problematic for compatibility:
e.g., what if a required file starts a server and so the require never ends, and later on another thread wants to require some code?

[akr (Akira Tanaka)]

Eregon (Benoit Daloze) wrote:

I was thinking maybe the require lock per path should be used for everything, but then since #autoload calls require dynamically, how to keep track which thread is loading the constant and so should observe the autoload constant as "not defined" while loading it, without deadlocks?
Other threads might try to load the constant too, or require the autoload path, and only one thread should be the loading thread for that constant.

I think "lock per path" can cause deadlock with "mutual require" similar to
https://bugs.ruby-lang.org/issues/15598 .

That would simplify things, but then I think it would also need to be a global require lock, and I think that can be problematic for compatibility:
e.g., what if a required file starts a server and so the require never ends, and later on another thread wants to require some code?

Why "it would also need to be a global require lock"?

[fxn (Xavier Noria)]

Let me share some thoughts that won't help much, but would like to contribute anyway :).

To me it is a surprise that constants for which there is an autoload are treated as existing by the constants API. My basic observation is that you don't know if the constant will actually be there until you execute the require. Since the require could fail, from non-existing files, to syntax errors, to files not actually defining the constants. The constant may never materialize.

For me the semantics would be easier if autoloads were treated separately. For example, if const_defined? or defined? returned false for autoloads, constants would not include them, etc. You have actually existing constants, and autoloads, you have autoload? for autoloads if you need to introspect them. You would need remove_autoload perhaps... you see the mental model: two separate collections.

Of course, nothing of this is backwards compatible, so of no practical value for this thread surely.

[Eregon (Benoit Daloze)]

• Related to Misc #21035: Clarify or redefine Module#autoload? and Module#const_defined? added

[Eregon (Benoit Daloze)]

@fxn Regarding the comment above, I think always treating autoload constants as if they already existed seems one way to be more consistent.
From that POV any autoload which when triggered does not define the constant should be an error, specifically a LoadError or so, which the program shouldn't catch and so should be considered a "fatal" error of that program.
Similar to e.g. a constant is defined but some bad code does remove_const of it later, those kind of broken code edge cases should be considered bugs of the program, and in fact anyone writing Ruby code relies on gems not doing such nasty things.

[fxn (Xavier Noria)]

Hi @Eregon (Benoit Daloze)!

From that POV any autoload which when triggered does not define the constant should be an error, specifically a LoadError

Not quite! Autoload is just a trigger for Kernel#require, there is no expectation on its side-effects. Once triggered, constant lookup continues and it is constant lookup what drives possible errors. For example, consider

# x.rb
M::X = 1

# m.rb
module M
  module N
    autoload :X, 'x'

    p X # prints 1!
  end
end

That works, because autoload does not care about what happened.

It triggered Kernel#require, which loaded the file and defined M::X, control back. N::X does not exist, up the nesting M::X exist, so the constant is resolved.

(Zeitwerk does require the constant to be defined in the receiver, but that is a contract enforced by the library to match file name conventions and have more strict rules.)

[fxn (Xavier Noria)]

A classic way to exploit that behavior I've seen out there (in the past):

module MyGem
  autoload :Date, 'date'
end

That meant, make Date available in case client code uses my gem in a way that hits an internal reference to the top-level Date from the standard library. But do it with autoload so that it is not eager loaded if not needed.

[fxn (Xavier Noria)]

BTW, I am describing how it works, not implying I like it :).

[Eregon (Benoit Daloze)]

Right, I recall those, I feel these should really give a NameError or LoadError when the autoload is triggered, because it's basically an incorrect usage of autoload.
It can result in a pretty weird state for M::N::X / MyGem::Date, what are those now, are they still autoload-registered constants (but then it might trigger the autoload again)?, are they removed (seems a bit weird if removed but then the access to the constant is still OK)? are they "undefined"? are they actually set to the same value as M::X/::Date (even though the Ruby code doesn't do that)?
IIRC it changed from undefined to removed in some Ruby version.

The fact the constant scope (IOW Module.nesting) is used when resolving an autoload is IMO wrong, because then resolving an autoload in one or another place can give a different result.
IOW, I'd like when an autoload is triggered to simplify the outcome to just two outcomes, when the require for the autoload is done:

• the constant which used to be the autoload constant has been replaced by a value, all good, autoloading succeeded, return that
• otherwise error like NameError or LoadError because the autoload was incorrectly registered or the loaded file didn't define the correct constant. Similar to require 'file_which_does_not_exist'.
  Right now we also have:
• the file defined a different constant, and that might still work due to the caller constant scope, but it depends on which constant access triggered the autoload first, which is obviously brittle and makes things really complicated.

[Eregon (Benoit Daloze)]

fxn (Xavier Noria) wrote in #note-9:

Autoload is just a trigger for Kernel#require, there is no expectation on its side-effects.

There certainly is for any Rubyist though (the whole point of an autoload is it lazily defines some constant on first access), but yeah the current implementation doesn't match that so strictly unfortunately.

[fxn (Xavier Noria)]

I see what you mean, but let me rephrase it a bit. The autoload does not depend on the nesting (depending on your definition of "autoload"). It is triggered by a missing constant reference when the constant is being looked up in the receiver, and the execution of the required file is not influenced by the nesting in the caller (you of course know this, only describing). Success/failure of the require call does not depend on the nesting.

Constant lookup does depend on the nesting, but constant lookup and autoload are independent of each other (except that one may trigger the other one).

I agree with you in those rules, I would also prefer this to be simpler and more strict. Either you define the constant in the receiver or error.

I'd go further for the sake of the exercise (but impossible to change). An autoload is a promise, but it may fail. In that sense, if this was day 0 and I was the author of the language, autoloads would not appear in the constants API. Only "actual" constants would.

[Eregon (Benoit Daloze)]

Right, that's a good summary. It's natural for the constant lookup to use the constant scope as usual, and that part is fine.
The part that's messy is when doing the require for an autoload, and it doesn't define the constant it "promised" to.
In such a case I think an exception would be warranted, with of course a migration path of first warning such cases.
We actually already have these warnings:

$ cat a.rb
A = 42
$ ruby -w -I. -e 'autoload :A, "a"; p A'
42
$ ruby -w -I. -e 'autoload :B, "a"; p B'
-e:1: warning: Expected a to define B but it didn't
-e:1:in `<main>': uninitialized constant B (NameError)

autoload :B, "a"; p B
                    ^
$ ruby -w -I. -e 'module M; autoload :A, "a"; p A; end'
-e:1: warning: Expected a to define M::A but it didn't
42

So I guess I should make a ticket about turning those warnings into errors now :)

At least it would be one complex aspect of autoload removed.