StringScanner.new should use StringValue

[rubychan]

According to the original C implementation, StringScanner.new does not copy the string, but only coerce it:

static VALUE
strscan_initialize(int argc, VALUE *argv, VALUE self)
{
    struct strscanner *p;
    VALUE str, need_dup;

    Data_Get_Struct(self, struct strscanner, p);
    rb_scan_args(argc, argv, "11", &str, &need_dup);
    StringValue(str);
    p->str = str;

    return self;
}

So, the method in lib/strscan.rb should be:

def initialize(string, dup=false)
  @string = StringValue(string)
  reset_state
end

In addition, #string= and #concat should use StringValue to coerce the string. #concat may not need it since String#<< already uses StringValue, but it seems cleaner to apply it there, too.

The change speeds up this benchmark quite a bit:

def Bench.run
  a = ''
  1_000_000.times do
    StringScanner.new a
    a << '.'
  end
end

Runs for minutes without the fix.

[kronos]

Your variant is wrong because it fails specs where string is a subclass of String class. But this one works well:

def initialize(string, dup=false)
  @string = string.to_str
  reset_state
end

After doing this, the bench with rubinius runs faster than with 1.9.2/1.8.7

[rubychan]

Your variant is wrong

Isn't MRI doing the same?

[rubychan]

In MRI 1.9.2 and 1.8.7:

StringScanner.new(Class.new(String).new('test')).string.class == String
=> false 

[kronos]

Really, I don't know much about MRI. I just took your variant and ran specs)

[rubychan]

It seems there's a spec missing for StringScanner#string. Something like this:

describe "StringScanner#string" do
  ...
  it "returns an instance of the subclass when passed a String subclass" do
    cls = Class.new(String)
    sub = cls.new("abc")

    s = StringScanner.new(sub)

    string = s.string
    string.should be_an_instance_of(cls)
    string.should_not be_an_instance_of(String)
  end
end

[rubychan]

How about this: http://gist.github.com/836007

[evanphx]

Use coersion if possible in StringScanner. Closed by 7cbb344