Feature #15504: Freeze all Range objects

• Status: Open
• Priority: Normal
• Assignee: matz (Yukihiro Matsumoto)
• Target version: -

[ko1 (Koichi Sasada)]

Abstract

Range is currently non-frozen. How about freezing all Range objects?

Background

We froze some types of objects: Numerics (r47523) and Symbols [Feature #8906]. I believe that making objects immutable solves some kinds of programming difficulties.

Range is mutable at least when written as Range literal. So we can write the following weird program:

2.times{
  r = (1..3)
  p r.instance_variable_get(:@foo)
  #=> 1st time: nil
  #=> 2nd time: :bar
  r.instance_variable_set(:@foo, :bar)
}

In range.c, there is a comment (thanks znz-san):

static void
range_modify(VALUE range)
{
    rb_check_frozen(range);
    /* Ranges are immutable, so that they should be initialized only once. */
    if (RANGE_EXCL(range) != Qnil) {
    rb_name_err_raise("`initialize' called twice", range, ID2SYM(idInitialize));
    }
}

Patch

Index: range.c
===================================================================
--- range.c (リビジョン 66699)
+++ range.c (作業コピー)
@@ -45,6 +45,8 @@
     RANGE_SET_EXCL(range, exclude_end);
     RANGE_SET_BEG(range, beg);
     RANGE_SET_END(range, end);
+
+    rb_obj_freeze(range);
 }

 VALUE

Discussion

There are several usages of mutable Range in the tests.

• (1) Taint-flag
• (2) Add singleton methods.
• (3) Subclass with mutable states

Maybe (2) and (3) are crucial.

Thanks,
Koichi

[marcandre (Marc-Andre Lafortune)]

I think that (2) and (3) are indeed capital points. Freezing range litterals (only) might be a better idea? with an approach like frozen string literals?

Not that even frozen ranges aren't completely immutable:

r = ('a'..'z').freeze
r.end.upcase!
r # => "a".."Z" 

[Eregon (Benoit Daloze)]

I think it would make sense to freeze Range literals.

Adding methods to Range might be reasonable, but singleton methods, I would think much less so.

[mame (Yusuke Endoh)]

I guess Eregon (Benoit Daloze) came from #15950. Will ary[-3..] be as efficient as ary[-3, 3] by freezing and deduping a literal (-3..)? Looks good if we can confirm it by an experiment.

Some thoughts:

• Even if a range is frozen, ("a".."z") should not be deduped because of the reason marcandre (Marc-Andre Lafortune) said.
• I'm for freezing all Ranges, not only Range literals. I hate the idea of freezing only literals because casually mixing frozen and unfrozen objects leads to hard-to-debug bugs that depend upon data flow.
• It would be the most elegant if the combination of MJIT and escape analysis solves this kind of performance problems.

[ko1 (Koichi Sasada)]

I got an issue on Ractor.

def test
  (1..2)
end

r1 = test

Ractor.new do
  r2 = test
end.take

• compiler cached Range (1..2) because they begin and end are frozen literals. The test method returns a cached same range object.
• it means test returns not-immutable but same object. It violate the Ractor's memory model.

To solve it, there are two options.

(1) avoid cache at compile time.
(2) freeze Range objects which will be cached by th compiler.

For performance reason, I want to choose (2).

After that, could you please discuss all Range objects should be frozen or not.

Thanks,
Koichi

[sawa (Tsuyoshi Sawada)]

• Description updated (diff)
• Subject changed from Freeze all Range object to Freeze all Range objects

[Eregon (Benoit Daloze)]

Right, every object cached at parse time must be deeply immutable, I would think that was an oversight.