On Fri, 12 Feb 2010, Leandro wrote:

> 
> Levente Uzonyi wrote:
>
>       You can also get the word based diff using TextDiffBuilder by subclassing it and overriding #split:. Actually ClassDiffBuilder does exactly
>       this, so using ClassDiffBuilder instead of TextDiffBuilder will give you the expected result.
> 
> 
> I think that the difference with the implementation Valeria and I did is that ours switches from lines to words automatically. It also defines "words"
> as Smalltalk words, meaning that it understands keyword limits, which gives better results. The differentiator starts comparing lines, and if the
> proportion of line differences isn't too high, then it reutilizes the work already done (with lines) and refines it to get differences in between
> keywords. The refining strategy was crucial for achieving good performance.

If you subclass ClassDiffBuilder (let's call the subclass KeywordDiffBuilder)
and implement KeywordDiffBuilder >> #split: just like DifferenceFinder >> 
#keywordsAndBlanksFrom: you get even better results.

DifferenceFinder has to use such tricks/hacks to have good performance, 
because it uses the simplest lcs algorithm which requires n*m time and 
space in all cases, while for practical cases TextDiffBuilder requires 
O(n+m) time and space.


Levente

> 
> Not sure if the full functionality I'm describing here is in the last Cuis version though.
> 
> /Leandro
> 
>