Feature #21000: A way to avoid loading constant required by a type check

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

[Dan0042 (Daniel DeLorme)]

There is this pattern I encounter sometimes:

if defined?(NameSpace::ClassName) and obj.is_a?(NameSpace::ClassName)

Searching in gems, the pattern is fairly common: https://pastebin.com/VGfjRWNu

I would like a way to avoid the repetition of NameSpace::ClassName above. I can think of a number of ways to approach the issue, each with different tradeoffs...

Pattern match ignores uninitialized constant

Pattern match like obj in XYZ could return false if XYZ is not defined. The danger here is that a typo could go undetected and just silently ignore the error even when the constants is expected to be defined.

Pattern match has special syntax to ignore uninitialized constant

Pattern match such as obj in XYZ? (or some other syntax) could return false if XYZ is not defined. The downside is that we're adding yet more new syntax. But it could be obj in defined?(XYZ) and then it doesn't really feel like new syntax.

Do not autoload constants required by pattern match

If we have autoload :XYZ, "xyz" then obj in XYZ could skip the autoload and return false. There is a possibility that XYZ might be defined as a regexp or other matcher that return true, but in general autoload is only used for classes/modules. And if the class/module is not yet loaded, obviously an object of that type cannot exist so we can avoid loading it. But this would only work for autoloaded constants, so can't be used to check a library that might not be loaded, ex: obj in ActiveRecord::Base

defined?(mod) returns mod if it's a class/module

If XYZ is a module, defined?(XYZ) could return XYZ instead of returning "constant". So it can be used in expressions like

case obj
when nil
when defined?(XYZ)

if obj and defined?(XYZ) === obj

if defined?(Gem::Specification)&.respond_to?(:each)

Very versatile, with the downside that it's a small backward incompatibiliy.

[nobu (Nobuyoshi Nakada)]

Dan0042 (Daniel DeLorme) wrote:

defined?(mod) returns mod if it's a class/module

If XYZ is a module, defined?(XYZ) could return XYZ instead of returning "constant". So it can be used in expressions like

case obj
when nil
when defined?(XYZ)

if obj and defined?(XYZ) === obj

if defined?(Gem::Specification)&.respond_to?(:each)

Very versatile, with the downside that it's a small backward incompatibiliy.

I like this, but we will need to consider the incompatibility and inconsistency with other constants.

[Eregon (Benoit Daloze)]

defined? so far always returned nil or frozen strings, so I think it's not so great to use for this.

Also defined? is particularly error-prone for constants & autoloading, for instance:

$DEBUG = true # to show swallowed exceptions

autoload :Foo, "foo"

p defined?(Foo) # "constant"

p defined?(Foo::Bar)
# Exception 'LoadError' at <internal:/home/eregon/.rubies/ruby-3.4.1/lib/ruby/3.4.0/rubygems/core_ext/kernel_require.rb>:136 - cannot load such file -- foo
# Exception 'LoadError' at <internal:/home/eregon/.rubies/ruby-3.4.1/lib/ruby/3.4.0/rubygems/core_ext/kernel_require.rb>:144 - cannot load such file -- foo
nil

So defined?(Foo::Bar) will actually try to load Foo.

[Dan0042 (Daniel DeLorme)]

Eregon (Benoit Daloze) wrote in #note-2:

So defined?(Foo::Bar) will actually try to load Foo.

It would be a good idea to fix that right?

[Eregon (Benoit Daloze)]

I'm unsure, I suppose it could be a compatibility concern.
If you want to know if Foo::Bar exists and Foo is an autoload the only way to know is to load Foo and see if it has a constant Bar.
So while surprising these semantics do make sense since defined? doesn't have a "it's an autoload (not yet loaded), I don't know" return value, it's either nil (doesn't exist) or "constant" (exists).

That said, p defined?(Foo) # nil is weird though.

[Dan0042 (Daniel DeLorme)]

While these semantics make a certain sense, the opposite also makes just as much sense. If Foo is not yet autoloaded, that means neither Foo nor Foo::Bar are yet defined, and it makes sense to return falsy for both. At the very least defined?(Foo) and defined?(Foo::Bar) should be consistent.

[nobu (Nobuyoshi Nakada)]

• Description updated (diff)

[mame (Yusuke Endoh)]

When you create a ticket, please state clearly the problem you wish to solve.

Here is my understanding. You see two problems with the following idiom

if defined?(Foo) and obj.is_a?(Foo)

• This idiom fires autoload. Before the actual load, there is no way that obj could be an instance of Foo, so we can tell "false" without firing autoload.
• It is redundant to write the constant name twice.

Therefore, you want a way to write a decision similar to this idiom without firing autoload. Preferably without writing the constant name twice.

Is my understanding right?

If so, how would your last proposal that defined?(XYZ) return XYZ solve this problem?
Currently, for the constant XYZ for which autoload is set, defined?(XYZ) returns "constant" even before the actual load.
If it were to return XYZ itself, I don't think it would solve the autoload problem.
Or does your proposal involve changing defined?(XYZ) to return nil if it is before the actual load? I don't think that is a small incompatibility.

[Dan0042 (Daniel DeLorme)]

mame (Yusuke Endoh) wrote in #note-7:

When you create a ticket, please state clearly the problem you wish to solve.

Therefore, you want a way to write a decision similar to this idiom without firing autoload. Preferably without writing the constant name twice.

Therefore, I want a way to write a decision similar to this idiom without writing the constant name twice. Preferably without firing autoload.

Apologies if this wasn't clear enough in the original description—I may not have formatted it effectively.

In the third idea I mentioned autoload as a possible solution in the context of pattern matching, but I hadn't really considered autoload in the context of defined?(XYZ) in the fourth idea.

To clarify, my fourth idea was:

defined?(NothingHere) #=> nil
N = nil
defined?(N) #=> "constant"
XYZ = Class.new
defined?(XYZ) #=> XYZ

I hadn’t accounted for how defined? should behave with an autoloaded constant. @Eregon (Benoit Daloze) mentioned p defined?(Foo) # nil but it actually returns "constant", which I wasn't aware of.

I think it would make sense for defined?(Foo) to return nil for a not-yet-loaded autoload constant. Alternatively, it could trigger autoload and return Foo if it resolves to a module. But as you said that is no longer a small incompatibility.

[Eregon (Benoit Daloze)]

Dan0042 (Daniel DeLorme) wrote in #note-8:

I hadn’t accounted for how defined? should behave with an autoloaded constant. @Eregon (Benoit Daloze) mentioned p defined?(Foo) # nil but it actually returns "constant", which I wasn't aware of.

Sorry about that, it is indeed "constant", I must have made a copy-paste mistake (I also updated the original post for clarity).
Here is the full code and output:

$ ruby -de 'autoload :Foo, "foo"; p defined?(Foo); p defined?(Foo::Bar)'
Exception `LoadError' at /home/eregon/.rubies/ruby-3.3.5/lib/ruby/3.3.0/rubygems.rb:1346 - cannot load such file -- rubygems/defaults/operating_system
Exception `LoadError' at /home/eregon/.rubies/ruby-3.3.5/lib/ruby/3.3.0/rubygems.rb:1363 - cannot load such file -- rubygems/defaults/ruby
"constant"
Exception `LoadError' at <internal:/home/eregon/.rubies/ruby-3.3.5/lib/ruby/3.3.0/rubygems/core_ext/kernel_require.rb>:136 - cannot load such file -- foo
Exception `LoadError' at <internal:/home/eregon/.rubies/ruby-3.3.5/lib/ruby/3.3.0/rubygems/core_ext/kernel_require.rb>:144 - cannot load such file -- foo
nil

So indeed defined?(Foo) returns "constant" without trying to load it, and defined?(Foo::Bar) triggers the autoload.

[Dan0042 (Daniel DeLorme)]

After some further thought, I believe the fourth idea could work if autoload constants simply returned "constant". Because if a constant isn’t yet autoloaded, no objects of that type can exist. To recap:

defined?(NothingHere) #=> nil

N = nil
defined?(N) #=> "constant"

XYZ = Class.new
defined?(XYZ) #=> XYZ

autoload :Foo, "foo"
defined?(Foo) #=> "constant"

With this approach, we’re back to only a small incompatibility. I describe it as "small" because defined? is almost always used as a boolean condition. I've only found 1 gem that uses the return value of defined? beyond checking its truthiness, and even then, it's for "method", not "constant"

oj-3.13.14/test/helper.rb
22:if defined?(GC.verify_compaction_references) == 'method'

oj-3.13.14/test/json_gem/test_helper.rb
18:  if defined?(GC.verify_compaction_references) == 'method'

[mame (Yusuke Endoh)]

If you just want to avoid repetition and don't care autoload, I found a good (and magical) way to do it.

class TrueClass
  def true?
  end
end

if defined?(obj.is_a?(NameSpace::ClassName).true?)
  # obj is a NameSpace::ClassName
else
  # obj is not a NameSpace::ClassName
end

# or, more magically

if defined?(-(obj.is_a?(NameSpace::ClassName)&&1))
  # obj is a NameSpace::ClassName
else
  # obj is not a NameSpace::ClassName
end

The behavior of defined? that actually evaluates the some part is sometimes convenient :-)

I don't think my idiom is nice, though. It is too magical. And, I think if obj and defined?(XYZ) === obj is too magical as well.

[Eregon (Benoit Daloze)]

IMO the pattern is not frequent enough to justify adding new syntax for this or breaking compatibility.
In https://pastebin.com/VGfjRWNu many cases are actually not classes, e.g. defined?(::JSON::JSON_LOADED) and ::JSON::JSON_LOADED, defined?(RUBY_ENGINE) and RUBY_ENGINE == 'jruby', etc.

Maybe a new method or kwarg to const_get which has some special behavior regarding autoload (to not trigger them)?

[Dan0042 (Daniel DeLorme)]

Eregon (Benoit Daloze) wrote in #note-12:

Maybe a new method or kwarg to const_get which has some special behavior regarding autoload (to not trigger them)?

That sounds good to me. What about Object.const_dig(:JSON, :JSON_LOADED) which returns nil if at any point the constant is not defined or not yet autoloaded.

It can even mirror the #dig behavior:

Object.const_dig(:JSON, :JSON_LOADED, :SOMETHING)
# TrueClass does not have #const_dig method (TypeError)