Project

General

Profile

Feature #15393

Add compilation flags to freeze Array and Hash literals

Added by tenderlovemaking (Aaron Patterson) 5 days ago. Updated about 9 hours ago.

Status:
Open
Priority:
Normal
Assignee:
-
Target version:
-
[ruby-core:90375]

Description

Hi,

I would like to add VM compilation options to freeze array and hash literals. For example:

frozen = RubyVM::InstructionSequence.compile(<<-eocode, __FILE__, nil, 0, frozen_string_literal: true, frozen_hash_and_array_literal: true)
  { 'a' => ['b', { 'c' => 'd' }] }
eocode
puts frozen.disasm

Output is:

$ ./ruby thing.rb
== disasm: #<ISeq:<compiled>@thing.rb:0 (0,0)-(0,34)> (catch: FALSE)
0000 putobject                    {"a"=>["b", {"c"=>"d"}]}
0002 leave

Anything nested in the hash that can't be "frozen" will cause it to not be frozen.

For example:

not_frozen = RubyVM::InstructionSequence.compile(<<-eocode, __FILE__, nil, 0, frozen_string_literal: true, frozen_hash_and_array_literal: true)
  { 'a' => some_method }
eocode
puts not_frozen.disasm

Output:

$ ./ruby thing.rb
== disasm: #<ISeq:<compiled>@thing.rb:0 (0,0)-(0,24)> (catch: FALSE)
0000 putobject                    "a"
0002 putself
0003 opt_send_without_block       <callinfo!mid:some_method, argc:0, FCALL|VCALL|ARGS_SIMPLE>, <callcache>
0006 newhash                      2
0008 leave

Eventually I would like to freeze array and hash literals in source code itself, but I think this is a good first step.

The reason I want this feature is I think we can reduce some object allocations, and once Guilds are implemented, easily create immutable data.

I've attached a patch that implements the above.

(Also I think maybe "frozen_literals" would be a better name, but I don't want to imply that numbers or booleans are frozen too)

Thanks!

History

#1 [ruby-core:90379] Updated by shevegen (Robert A. Heiler) 5 days ago

About the name - some suggestions if only to show which variants may be
better than others. :)

frozen: true # could say "ruby, freeze as much as you possibly can!"

In other words, that setting/flag could mean that strings, hashes and arrays
are frozen (if the above is approved).

Or:

frozen_string_array_hash_literals: true
frozen_string_hash_array_literals: true
frozen_three_literals: true

I am not saying any of these names are good names. Just putting them down. :)

frozen_literals: true

I think one advantage for frozen_literals is that it is short. People could
use that in .rb files too.

Another approach could of course also be to only allow strings to be frozen
(which will become the default in ruby 3.x if I remember correctly); and not
allow a change of hash and arrays in a .rb file, but to allow it for the
suggestion above, e. g. in RubyVM.

One could also split the above up into hash and array as separate calls,
so rather than:

RubyVM::InstructionSequence.compile(<<-eocode, __FILE__, nil, 0, frozen_string_literal: true, frozen_hash_and_array_literal: true)

to have:

RubyVM::InstructionSequence.compile(<<-eocode, __FILE__, nil, 0, frozen_string_literal: true, frozen_hash_literal: true, frozen_array_literal: true)

I mention the last primarily because we already have frozen string literals; so for
consistency it may be useful to have it split up into hash and arrays too. If this
is too cumbersome to use three separate calls, then one could always use keys that
unify these, such as in the example given by Aaron:

frozen_hash_and_array_literal: true

could also include Strings into it:

frozen_string_and_hash_and_array_literal: true

It's a bit cumbersome though. Perhaps a simple variant may then be:

frozen: true
frozen_literal: true
frozen_literals: true

This could then mean to "freeze all that can be frozen".

Anyway; I think the first step is to ask matz about the functionality
itself, and then see which API may be best if this is approved.

Personally I think "frozen_literals" would be ok, even if the name
may not be 100% "perfect". Ruby is not necessarily perfect in all
names e. g. I remember a few folks thinking that the word constant
implies "can not be changed", which may be the accurate meaning,
but from within ruby itself, I think it is perfectly fine to let
people change constants if they want to (e. g. the philosophy of
ruby to be practical and useful).

#2 [ruby-core:90380] Updated by Eregon (Benoit Daloze) 5 days ago

One alternative idea which would achieve a similar goal is having #deep_freeze on core types, and recognizing { 'a' => ['b', { 'c' => 'd' }] }.deep_freeze.
This would be more general, since it would also work for freezing non-constant/non-literal data structures.

I would think many [] and {} are meant to be mutated, so it seems frozen Array/Hash literals them would rather be the exception than the norm (which is less clear for Strings, so I think there the magic comment makes sense).

#3 [ruby-core:90387] Updated by shyouhei (Shyouhei Urabe) 3 days ago

Off topic:

Eregon (Benoit Daloze) wrote:

One alternative idea which would achieve a similar goal is having #deep_freeze on core types, and recognizing { 'a' => ['b', { 'c' => 'd' }] }.deep_freeze.
This would be more general, since it would also work for freezing non-constant/non-literal data structures.

Object#deep_freeze can be written without any extensions right now, like this: https://github.com/shyouhei/optdown/blob/master/lib/optdown/deeply_frozen.rb
I heard this trick from akr (Akira Tanaka) .

#4 [ruby-core:90404] Updated by Eregon (Benoit Daloze) 3 days ago

Interesting code :)
But also inefficient of course, making extra copies (changing
identity) and allocations.
On Mon, Dec 10, 2018 at 1:55 AM shyouhei@ruby-lang.org wrote:

Issue #15393 has been updated by shyouhei (Shyouhei Urabe).

Off topic:

Eregon (Benoit Daloze) wrote:

One alternative idea which would achieve a similar goal is having #deep_freeze on core types, and recognizing { 'a' => ['b', { 'c' => 'd' }] }.deep_freeze.
This would be more general, since it would also work for freezing non-constant/non-literal data structures.

Object#deep_freeze can be written without any extensions right now, like this: https://github.com/shyouhei/optdown/blob/master/lib/optdown/deeply_frozen.rb
I heard this trick from akr (Akira Tanaka) .


Feature #15393: Add compilation flags to freeze Array and Hash literals
https://bugs.ruby-lang.org/issues/15393#change-75509

  • Author: tenderlovemaking (Aaron Patterson)
  • Status: Open
  • Priority: Normal
  • Assignee:

* Target version:

Hi,

I would like to add VM compilation options to freeze array and hash literals. For example:

frozen = RubyVM::InstructionSequence.compile(<<-eocode, __FILE__, nil, 0, frozen_string_literal: true, frozen_hash_and_array_literal: true)
  { 'a' => ['b', { 'c' => 'd' }] }
eocode
puts frozen.disasm

Output is:

$ ./ruby thing.rb
== disasm: #<ISeq:<compiled>@thing.rb:0 (0,0)-(0,34)> (catch: FALSE)
0000 putobject                    {"a"=>["b", {"c"=>"d"}]}
0002 leave

Anything nested in the hash that can't be "frozen" will cause it to not be frozen.

For example:

not_frozen = RubyVM::InstructionSequence.compile(<<-eocode, __FILE__, nil, 0, frozen_string_literal: true, frozen_hash_and_array_literal: true)
  { 'a' => some_method }
eocode
puts not_frozen.disasm

Output:

$ ./ruby thing.rb
== disasm: #<ISeq:<compiled>@thing.rb:0 (0,0)-(0,24)> (catch: FALSE)
0000 putobject                    "a"
0002 putself
0003 opt_send_without_block       <callinfo!mid:some_method, argc:0, FCALL|VCALL|ARGS_SIMPLE>, <callcache>
0006 newhash                      2
0008 leave

Eventually I would like to freeze array and hash literals in source code itself, but I think this is a good first step.

The reason I want this feature is I think we can reduce some object allocations, and once Guilds are implemented, easily create immutable data.

I've attached a patch that implements the above.

(Also I think maybe "frozen_literals" would be a better name, but I don't want to imply that numbers or booleans are frozen too)

Thanks!

---Files--------------------------------
0001-Add-compile-options-for-freezing-hash-and-array-lite.patch (6.14 KB)

--
https://bugs.ruby-lang.org/

#5 [ruby-core:90408] Updated by tenderlovemaking (Aaron Patterson) 2 days ago

Eregon (Benoit Daloze) wrote:

One alternative idea which would achieve a similar goal is having #deep_freeze on core types, and recognizing { 'a' => ['b', { 'c' => 'd' }] }.deep_freeze.
This would be more general, since it would also work for freezing non-constant/non-literal data structures.

I thought about doing this with ".freeze", introducing a special instruction the same way we do for the "string".freeze optimization, but dealing with deoptization in the case someone monkey patches the method seems like a pain.

The reason I went this direction first is that it side steps the deoptimization problem, and if we want to support a ".deep_freeze" method later we can. I imagine the implementation would be a branch where one of the branches is the same putobject instruction that I have in this patch.

Anyway, I think the advantages of this patch are:

  1. No new methods like "deep_freeze", so you can try this with existing code
  2. Any code you write that works with these flags enabled will also work with them disabled (vs deep_freeze that isn't on existing Rubys)
  3. This patch should be forward compatible when we figure out what "deep_freeze" solution we're going to have in the future

I would think many [] and {} are meant to be mutated, so it seems frozen Array/Hash literals them would rather be the exception than the norm (which is less clear for Strings, so I think there the magic comment makes sense).

I thought the same thing, but I'm not 100% convinced. Lots of code in Rails (as well as our application at work) is merging some kind of "default hash" like:

def method(some_hash)
  some_hash = { :defaults => 'thing' }.merge(some_hash)
end

I'll try to get some numbers on this though. :D

#6 [ruby-core:90457] Updated by shan (Shannon Skipper) about 9 hours ago

I had the same thought as shevy, that it'd be nice to have a:

# frozen_literals: true

It might also be worth considering a separate frozen_array_literal and frozen_hash_literal. I have files where I'd initially only want to enable one of the above.

Also available in: Atom PDF