Empty regexp replaceall and surrogate pairs results in corrupted utf16.

[Ulf Zibis]

I tend to agree Dawid.
Especially the comparison with Python behaviour is demonstrative.

Is there any spec weather the Java Regex API has a general contract with 16-bit chars or Unicode 
codepoints?

Thinking about the search pattern e.g. "[AB\uD840\uDC00C]"; what does it actually search for, the 
isolated occurence of each char, or the occurence of the codepoint "\uD840\uDC00" ?

Last, but not least, we should think about, which would be the common use case, an which would be 
more easy to work around.
(I think, the current view on isolated chars is more easy to work around:
StringBuilder sb = new StringBuilder(s1.length + 1).append('X');
for (char c : s1.getChars())
     sb.append(c).append('X');
String s2 = sb.toString();
)

Additionally I like to discuss: "any possible zero-width position of the target String"
If String length is l, maybe it's arguable, that position l is no valid position in the String.

 From the use case point of view, I think "P e t e r" as result of "Peter".replaceAll("", " ") is 
the most useful.

-Ulf


Am 08.06.2012 13:14, schrieb Dawid Weiss:
> I guess a lot depends on the point of view. From historical point of
> view (where a char[] and a String are basically unsigned values) that
> pattern should simply process every value (index) and work like you
> say. But from a practical point of view I think it is a bug -- it
> corrupts the string, transforming legal unicode into invalid values.
>
> I checked with Python (3) and the behavior there is the expected one
> (it work at the unicode codepoint level rather than surrogate level).
>
> Where is the behavior of "" that you mention defined? I admit I
> couldn't find any reference to this in the documentation:
>
>> Using an empty String "" as a regex for the replaceAll() takes the
>> advantage of the special meaning of "", in which it is interpreted as
>> it can match any possible zero-width position of the target String
> I'm not saying you're wrong (and that pattern is definitely not common
> so it's probably academic discussion) but I'd like some concrete
> reference as to how an empty pattern should behave. To me consistency
> with the rest of the Pattern specification would be that it operates
> at "zero width position between unicode characters" not between any
> char[] value, even an incorrect one or in the middle of a surrogate.
>
> Dawid
>
> On Fri, Jun 8, 2012 at 12:46 AM, Xueming Shen<xueming.shen at oracle.com>  wrote:
>> Personally I don't think it is a bug. A j.l.String represents a sequence of
>> UTF-16 chars. While
>> a pair of surrogates represents a supplementary character, a single
>> surrogate itself is still
>> a "legal" independent entity inside a String object and length of a String
>> is still defined as
>> the total number of char unit and an index value between a high surrogate
>> and a low
>> surrogate is still a legal index value that can be used to access the char
>> at that particular
>> position. Using an empty String "" as a regex for the replaceAll() takes the
>> advantage of the
>> special meaning of "", in which it is interpreted as it can match any
>> possible zero-width
>> position of the target String, it does  not imply anything regarding
>> "character"  or
>> "characters" around it, so I would not interpret it as a zero-with character
>> boundary,
>> therefor a "position" in between a pair surrogates is still a good "found"
>> for replacing.
>>
>> -Sherman
>>
>>
>> On 6/7/2012 1:07 PM, Dawid Weiss wrote:
>>> Hi, I'm a committer to the Apache Lucene project. We have randomized
>>> tests and one seed hit the following (simplified) scenario:
>>>
>>>     String s1 = "AB\uD840\uDC00C";
>>>     String s2 = s1.replaceAll("", "X");
>>>
>>> the input contains an extended unicode character (any surrogate pair
>>> will do). The pattern is an empty string (in fact, it was randomized
>>> as "]|" but it's the same problem so I omit the details). The problem
>>> is that after applying this pattern, replaceAll inserts X in between
>>> the surrogate pair characters and this results in invalid UTF-16:
>>>
>>> AB��C
>>> XAXBX?X?XCX
>>>
>>> I believe this is a bug in the regexp implementation (sorry, don't
>>> have a patch for it) but I'd like to confirm it's not something known.
>>> Pointers appreciated.
>>>
>>> Dawid