Thesis: removing todo

[ifi-stolz-refaktor.git] / thesis / master-thesis-erlenkr.tex

\documentclass[USenglish,11pt]{ifimaster}
\usepackage{import}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc,url}
\usepackage{lmodern} % using Latin Modern to be able to use bold typewriter font
%\usepackage{mathpazo}
\urlstyle{sf}
\usepackage{listings}
\usepackage{booktabs}
\usepackage{tabularx}
\usepackage{tikz}
\usepackage{tikz-qtree}
\usetikzlibrary{shapes,snakes,trees,arrows,shadows,positioning,calc}
\usepackage{babel,textcomp,csquotes,ifimasterforside}

\usepackage{varioref}
\usepackage[hidelinks]{hyperref}
\usepackage{cleveref}
\usepackage[xindy]{glossaries}

\usepackage[style=alphabetic,backend=biber,doi=false,isbn=false]{biblatex}
\usepackage{amsthm}
\usepackage{mathtools}
\usepackage{graphicx}
% use 'disable' before printing:
\usepackage[]{todonotes}
\usepackage{xspace}
\usepackage{he-she}
\usepackage{verbatim}
\usepackage{minted}
\usepackage{multicol}
\usemintedstyle{bw}

\def\mintedframesep{11pt}

\usepackage{perpage} %the perpage package
\MakePerPage{footnote} %the perpage package command

\theoremstyle{definition}
\newtheorem*{wordDef}{Definition}
\newtheorem*{theorem}{Theorem}

\graphicspath{ {./figures/} }

\newcommand{\citing}[1]{~\cite{#1}}
% TODO: fix
\newcommand{\myref}[1]{\vref{#1}}
\newcommand{\Myref}[1]{\Vref{#1}}
%\newcommand{\myref}[1]{\ref{#1}}
%\newcommand{\Myref}[1]{\ref{#1}}
\newcommand{\mysimpleref}[1]{\cref{#1}}
\newcommand{\Mysimpleref}[1]{\Cref{#1}}

%\newcommand{\glossref}[1]{\textsuperscript{(\glsrefentry{#1})}}
%\newcommand{\gloss}[1]{\gls{#1}\glossref{#1}}
%\newcommand{\glosspl}[1]{\glspl{#1}\glossref{#1}}
\newcommand{\gloss}[1]{\gls{#1}}
\newcommand{\glosspl}[1]{\glspl{#1}}

\newcommand{\definition}[1]{\begin{wordDef}#1\end{wordDef}}
\newcommand{\see}[1]{(see \myref{#1})}
\newcommand{\explanation}[3]{\noindent\textbf{\textit{#1}}\\*\emph{When:} 
#2\\*\emph{How:} #3\\*[-7px]}

%\newcommand{\type}[1]{\lstinline{#1}}
\newcommand{\code}[1]{\texttt{\textbf{#1}}}
\newcommand{\type}[1]{\code{#1}}
\newcommand{\typeref}[1]{\footnote{\type{#1}}}
\newcommand{\typewithref}[2]{\type{#2}\typeref{#1.#2}}
\newcommand{\method}[1]{\type{#1}}
\newcommand{\methodref}[2]{\footnote{\type{#1}\method{\##2()}}}
\newcommand{\methodwithref}[2]{\method{#2}\footnote{\type{#1}\method{\##2()}}}
\newcommand{\var}[1]{\type{#1}}

\newcommand{\name}[1]{#1}
\newcommand{\tit}[1]{\emph{#1}}
\newcommand{\refa}[1]{\emph{#1}}
\newcommand{\pattern}[1]{\emph{#1}}
\newcommand{\metr}[1]{\emph{#1}}
\newcommand{\ExtractMethod}{\refa{Extract Method}\xspace}
\newcommand{\MoveMethod}{\refa{Move Method}\xspace}
\newcommand{\ExtractAndMoveMethod}{\refa{Extract and Move Method}\xspace}

\newcommand{\m}[1]{$#1$}

\newcommand\todoin[2][]{\todo[inline, caption={#2}, #1]{
\begin{minipage}{\textwidth-4pt}#2\end{minipage}}}

\title{Automated Composition of Refactorings}
\subtitle{Implementing and evaluating a search-based Extract and Move Method 
refactoring}
\author{Erlend Kristiansen}

\makeglossaries
\newglossaryentry{profiling}
{
  name=profiling,
  description={is to run a computer program through a profiler/with a profiler 
  attached}
}
\newglossaryentry{profiler}
{
  name=profiler,
  description={A profiler is a program for analyzing performance within an 
  application. It is used to analyze memory consumption, processing time and 
frequency of procedure calls and such}
}
\newglossaryentry{xUnit}
{
  name={xUnit framework},
  description={An xUnit framework is a framework for writing unit tests for a 
    computer program. It follows the patterns known from the JUnit framework for 
    Java\citing{fowlerXunit}
  },
  plural={xUnit frameworks}
}
\newglossaryentry{softwareObfuscation}
{
  name={software obfuscation},
  description={makes source code harder to read and analyze, while preserving 
  its semantics}
}
\newglossaryentry{extractClass}
{
  name=\refa{Extract Class},
  description={The \refa{Extract Class} refactoring works by creating a class, 
for then to move members from another class to that class and access them from 
the old class via a reference to the new class}
}
\newglossaryentry{designPattern}
{
  name={design pattern},
  description={A design pattern is a named abstraction that is meant to solve a 
  general design problem.  It describes the key aspects of a common problem and 
identifies its participators and how they collaborate},
  plural={design patterns}
}
\newglossaryentry{enclosingClass}
{
  name={enclosing class},
  description={An enclosing class is the class that surrounds any specific piece 
  of code that is written in the inner scope of this class},
}
\newglossaryentry{mementoPattern}
{
  name={memento pattern},
  description={The memento pattern is a software design pattern that is used to 
  capture an object's internal state so that it can be restored to this state 
  later\citing{designPatterns}},
}
%\newglossaryentry{extractMethod}
%{
%  name=\refa{Extract Method},
%  description={The \refa{Extract Method} refactoring is used to extract a 
%fragment of code from its context and into a new method. A call to the new 
%method is inlined where the fragment was before. It is used to break code into 
%logical units, with names that explain their purpose}
%}
%\newglossaryentry{moveMethod}
%{
%  name=\refa{Move Method},
%  description={The \refa{Move Method} refactoring is used to move a method from   
%  one class to another. This is useful if the method is using more features of 
%  another class than of the class which it is currently defined. Then all calls 
%  to this method must be updated, or the method must be copied, with the old 
%method delegating to the new method}
%}

\bibliography{bibliography/master-thesis-erlenkr-bibliography}
\DefineBibliographyStrings{english}{%
  bibliography = {References},
}
\newbibmacro{string+doi}[1]{%
  \iffieldundef{doi}{#1}{\href{http://dx.doi.org/\thefield{doi}}{#1}}}
\DeclareFieldFormat{title}{\usebibmacro{string+doi}{\mkbibemph{#1}}}
\DeclareFieldFormat[article]{title}{\usebibmacro{string+doi}{\mkbibquote{#1}}}

% UML comment in TikZ:
% ref: https://tex.stackexchange.com/questions/103688/folded-paper-shape-tikz
\makeatletter
\pgfdeclareshape{umlcomment}{
  \inheritsavedanchors[from=rectangle] % this is nearly a rectangle
  \inheritanchorborder[from=rectangle]
  \inheritanchor[from=rectangle]{center}
  \inheritanchor[from=rectangle]{north}
  \inheritanchor[from=rectangle]{south}
  \inheritanchor[from=rectangle]{west}
  \inheritanchor[from=rectangle]{east}
  % ... and possibly more
  \backgroundpath{% this is new
  % store lower right in xa/ya and upper right in xb/yb
  \southwest \pgf@xa=\pgf@x \pgf@ya=\pgf@y
  \northeast \pgf@xb=\pgf@x \pgf@yb=\pgf@y
  % compute corner of ‘‘flipped page’’
  \pgf@xc=\pgf@xb \advance\pgf@xc by-10pt % this should be a parameter
  \pgf@yc=\pgf@yb \advance\pgf@yc by-10pt
  % construct main path
  \pgfpathmoveto{\pgfpoint{\pgf@xa}{\pgf@ya}}
  \pgfpathlineto{\pgfpoint{\pgf@xa}{\pgf@yb}}
  \pgfpathlineto{\pgfpoint{\pgf@xc}{\pgf@yb}}
  \pgfpathlineto{\pgfpoint{\pgf@xb}{\pgf@yc}}
  \pgfpathlineto{\pgfpoint{\pgf@xb}{\pgf@ya}}
  \pgfpathclose
  % add little corner
  \pgfpathmoveto{\pgfpoint{\pgf@xc}{\pgf@yb}}
  \pgfpathlineto{\pgfpoint{\pgf@xc}{\pgf@yc}}
  \pgfpathlineto{\pgfpoint{\pgf@xb}{\pgf@yc}}
  \pgfpathlineto{\pgfpoint{\pgf@xc}{\pgf@yc}}
  }
}
\makeatother

\tikzstyle{comment}=[%
  draw,
  drop shadow,
  fill=white,
  align=center,
  shape=document,
  minimum width=20mm,
  minimum height=10mm,
  shape=umlcomment,
  inner sep=2ex,
  font=\ttfamily,
]

%\interfootnotelinepenalty=10000

% Space between table rows
\renewcommand{\arraystretch}{1.3}
% Multicolumns
\newcommand{\spancols}[2]{\multicolumn{#1}{@{}l@{}}{#2}}
% Column types
\newcolumntype{L}[1]{>{\hsize=#1\hsize\raggedright\arraybackslash}X}%
\newcolumntype{R}[1]{>{\hsize=#1\hsize\raggedleft\arraybackslash}X}%

\setcounter{tocdepth}{1}

\begin{document}
%\pagenumbering{arabic}
\mainmatter
\ififorside
%\frontmatter{}

%\setcounter{page}{3}

\chapter*{Abstract}
Complex source code impacts the cost of software maintenance in a negative way.  
In an object-oriented context, one class may depend on a high number of other 
classes, thus contributing to the complexity of a program and making changing
code prone to errors. Refactoring is a means to fight such complexity.

This thesis investigates whether automated refactoring can be used to lower the 
coupling between classes. A search-based composite refactoring combining the 
primitive refactorings \ExtractMethod and \MoveMethod is designed as a possible 
solution to this problem. Case studies are conducted to evaluate the effect of 
executing the search-based refactoring in a large code base.

\begin{comment}

The results from the case studies indicate that the refactoring has the 
potential of lowering coupling, but that the analysis performed before 
refactoring is not complete enough. This results in a large number of errors 
caused by the refactoring. The refactoring has not been properly evaluated for 
behavior changes. A complete pre-refactoring analysis for the search-based 
refactoring is still an open problem.

The efficiency and unreliability of the search-based refactoring makes it 
unsuitable for on-demand large-scale refactoring tasks. With some modifications, 
and extensive improvements, the work done for this thesis could be used to 
develop a tool for performing long-running analysis, which can be used to 
suggest code improvements.

\end{comment}

\todoin{\textbf{Remove all todos (including list) before delivery/printing!!!  
Can be done by removing ``draft'' from documentclass.}}
\todoin{Fix cross-references}

\tableofcontents{}
\listoffigures{}
\listoftables{}
\listoflistings{}

\chapter*{Acknowledgements}
\todoin{Write}

%\mainmatter
%\setcounter{page}{13}

\chapter{Introduction}

\section{Motivation and structure}

For large software projects, complex program source code is an issue.  It 
impacts the cost of maintenance in a negative 
way\citing{bankerMaintenanceCost1993}, and often stalls the implementation of 
new functionality and other program changes.  The code may be difficult to 
understand, the changes may introduce new bugs that are hard to find and its 
complexity can simply keep people from doing code changes, in fear of breaking 
some dependent piece of code.  All these problems are related, and often lead to 
a vicious circle that slowly degrades the overall quality of a project.

More specifically, and in an object-oriented context, a class may depend on a 
number of other classes. Sometimes these intimate relationships are appropriate, 
and sometimes they are not. Inappropriate \emph{coupling} between classes can 
make it difficult to know whether or not a change that is aimed at fixing a 
specific problem also alters the behavior of another part of a program.

One of the tools that are used to fight complexity and coupling in program 
source code is \emph{refactoring}. The intention for this master's thesis is 
therefore to create an automated composite refactoring that reduces coupling 
between classes. The refactoring shall be able to operate automatically in all 
phases of a refactoring, from performing analysis to executing changes. It is 
also a requirement that it should be able to process large quantities of source 
code in a reasonable amount of time.

The current chapter proceeds in \mysimpleref{sec:refactoring} by describing what 
refactoring is. Then the project is presented in \mysimpleref{sec:project}, 
before the chapter is concluded with a brief discussion of related work in 
\mysimpleref{sec:relatedWork}.

\Mysimpleref{ch:extractAndMoveMethod} shows the workings of our refactoring 
together with an example illustrating how it works for a specific case.  
\Mysimpleref{ch:eclipse} presents some of the APIs and frameworks that are 
relevant for source code analysis and change, and that are available when using 
the Eclipse Platform with the Java development tools installed.  
\Mysimpleref{ch:implementation} contains some implementation details
of what was presented in \mysimpleref{ch:extractAndMoveMethod}.  
\Mysimpleref{ch:caseStudies} presents a couple of case studies used to evaluate 
several aspects of our refactoring. The whole thesis is then winded up in 
\mysimpleref{ch:conclusions} with conclusions and future work.


\section{What is refactoring?}\label{sec:refactoring}

This question is best answered by first defining the concept of a 
\emph{refactoring}, what it is to \emph{refactor}, and then discuss what aspects 
of programming make people want to refactor their code.

\subsection{Defining refactoring}
Martin Fowler, in his classic book on refactoring\citing{refactoring}, defines a 
refactoring like this:

\begin{quote}
  \emph{Refactoring} (noun): a change made to the internal 
  structure\footnote{The structure observable by the programmer.} of software to 
  make it easier to understand and cheaper to modify without changing its 
  observable behavior.~\cite[p.~53]{refactoring}
\end{quote}

\noindent This definition assigns additional meaning to the word 
\emph{refactoring}, beyond the composition of the prefix \emph{re-}, usually 
meaning something like ``again'' or ``anew'', and the word \emph{factoring}, 
which can mean to isolate the \emph{factors} of something. Here a \emph{factor} 
would be close to the mathematical definition of something that divides a 
quantity, without leaving a remainder. Fowler is mixing the \emph{motivation} 
behind refactoring into his definition. Instead it could be more refined, formed 
to only consider the \emph{mechanical} and \emph{behavioral} aspects of 
refactoring. That is to factor the program again, putting it together in a 
different way than before, while preserving the behavior of the program. An 
alternative definition could then be: 

\definition{A \emph{refactoring} is a transformation
done to a program without altering its external behavior.}

From this we can conclude that a refactoring primarily changes how the 
\emph{code} of a program is perceived by the \emph{programmer}, and not the 
\emph{behavior} experienced by any user of the program. Although the logical 
meaning is preserved, such changes could potentially alter the program's 
behavior when it comes to performance gain or -penalties. So any logic depending 
on the performance of a program could make the program behave differently after 
a refactoring.

In the extreme case one could argue that \gloss{softwareObfuscation} is 
refactoring. It is often used to protect proprietary software. It restrains 
uninvited viewers, so they have a hard time analyzing code that they are not 
supposed to know how works. This could be a problem when using a language that 
is possible to decompile, such as Java. 

Obfuscation could be done composing many, more or less randomly chosen, 
refactorings. Then the question arises whether it can be called a 
\emph{composite refactoring} or not \see{compositeRefactorings}?  The answer is 
not obvious.  First, there is no way to describe the mechanics of software 
obfuscation, because there are infinitely many ways to do that. Second, 
obfuscation can be thought of as \emph{one operation}: Either the code is 
obfuscated, or it is not. Third, it makes no sense to call software obfuscation 
\emph{a refactoring}, since it holds different meaning to different people.

This last point is important, since one of the motivations behind defining 
different refactorings, is to establish a \emph{vocabulary} for software 
professionals to use when reasoning about and discussing programs, similar to 
the motivation behind \glosspl{designPattern}\citing{designPatterns}.  
\begin{comment}
So for describing \emph{software obfuscation}, it might be more appropriate to 
define what you do when performing it rather than precisely defining its 
mechanics in terms of other refactorings.
\end{comment}

\subsection{The etymology of 'refactoring'}
It is a little difficult to pinpoint the exact origin of the word 
``refactoring'', as it seems to have evolved as part of a colloquial 
terminology, more than a scientific term. There is no authoritative source for a 
formal definition of it. 

According to Martin Fowler\citing{etymology-refactoring}, there may also be more 
than one origin of the word. The most well-known source, when it comes to the 
origin of \emph{refactoring}, is the 
Smalltalk\footnote{\label{footNote}Programming language} community and their 
infamous \name{Refactoring 
Browser}\footnote{\url{http://st-www.cs.illinois.edu/users/brant/Refactory/RefactoringBrowser.html}} 
described in the article \tit{A Refactoring Tool for 
Smalltalk}\citing{refactoringBrowser1997}, published in 1997.  
Allegedly\citing{etymology-refactoring}, the metaphor of factoring programs was 
also present in the Forth\textsuperscript{\ref{footNote}} community, and the 
word ``refactoring'' is mentioned in a book by Leo Brodie, called \tit{Thinking 
Forth}\citing{brodie2004}, first published in 1984\footnote{\tit{Thinking Forth} 
was first published in 1984 by the \name{Forth Interest Group}.  Then it was 
reprinted in 1994 with minor typographical corrections, before it was 
transcribed into an electronic edition typeset in \LaTeX\ and published under a 
Creative Commons license in 
2004. The edition cited here is the 2004 edition, but the content should 
essentially be as in 1984.}. The exact word is only printed one 
place~\cite[p.~232]{brodie2004}, but the term \emph{factoring} is prominent in 
the book, which also contains a whole chapter dedicated to (re)factoring, and 
how to keep the (Forth) code clean and maintainable.

\begin{quote}
  \ldots good factoring technique is perhaps the most important skill for a 
  Forth programmer.~\cite[p.~172]{brodie2004}
\end{quote}

\noindent Brodie also express what \emph{factoring} means to him:

\begin{quote}
  Factoring means organizing code into useful fragments. To make a fragment 
  useful, you often must separate reusable parts from non-reusable parts. The  
  reusable parts become new definitions. The non-reusable parts become arguments 
  or parameters to the definitions.~\cite[p.~172]{brodie2004}
\end{quote}

Fowler claims that the usage of the word \emph{refactoring} did not pass between 
the \name{Forth} and \name{Smalltalk} communities, but that it emerged 
independently in each of the communities.

\subsection{Reasons for refactoring}
There are many reasons why people want to refactor their programs. They can for 
instance do it to remove duplication, break up long methods or to introduce 
design patterns into their software systems. The shared trait for all these is 
that peoples' intentions are to make their programs \emph{better}, in some 
sense.  But what aspects of their programs are becoming improved?

As just mentioned, people often refactor to get rid of duplication. They are 
moving identical or similar code into methods, and are pushing methods up or 
down in their class hierarchies. They are making template methods for 
overlapping algorithms/functionality, and so on. It is all about gathering what 
belongs together and putting it all in one place. The resulting code is then 
easier to maintain. When removing the implicit coupling\footnote{When 
  duplicating code, the duplicate pieces of code might not be coupled, apart 
from representing the same functionality. So if this functionality is going to 
change, it might need to change in more than one place, thus creating an 
implicit coupling between multiple pieces of code.} between code snippets, the 
location of a bug is limited to only one place, and new functionality need only 
to be added to this one place, instead of a number of places people might not 
even remember.

A problem you often encounter when programming, is that a program contains a lot 
of long and hard-to-grasp methods. It can then help to break the methods into 
smaller ones, using the \ExtractMethod refactoring\citing{refactoring}.  Then 
you may discover something about a program that you were not aware of before; 
revealing bugs you did not know about or could not find due to the complex 
structure of your program. Making the methods smaller and giving good names to 
the new ones clarifies the algorithms and enhances the \emph{understandability} 
of the program \see{magic_number_seven}. This makes refactoring an excellent 
method for exploring unknown program code, or code that you had forgotten that 
you wrote.

Most primitive refactorings are simple, and usually involves moving code 
around\citing{kerievsky2005}. The motivation behind them may first be revealed 
when they are combined into larger --- higher level --- refactorings, called 
\emph{composite refactorings} \see{compositeRefactorings}. Often the goal of 
such a series of refactorings is a design pattern. Thus the design can 
\emph{evolve} throughout the lifetime of a program, as opposed to designing 
up-front.  It is all about being structured and taking small steps to improve a 
program's design.

Many software design pattern are aimed at lowering the coupling between 
different classes and different layers of logic. One of the most famous is 
perhaps the \pattern{Model-View-Controller}\citing{designPatterns} pattern. It 
is aimed at lowering the coupling between the user interface, the business logic 
and the data representation of a program. This also has the added benefit that 
the business logic could much easier be the target of automated tests, thus 
increasing the productivity in the software development process.

Another effect of refactoring is that with the increased separation of concerns 
coming out of many refactorings, the \emph{performance} can be improved. When 
profiling programs, the problematic parts are narrowed down to smaller parts of 
the code, which are easier to tune, and optimization can be performed only where 
needed and in a more effective way\citing{refactoring}.

Last, but not least, and this should probably be the best reason to refactor, is 
to refactor to \emph{facilitate a program change}. If one has managed to keep 
one's code clean and tidy, and the code is not bloated with design patterns that 
are not ever going to be needed, then some refactoring might be needed to 
introduce a design pattern that is appropriate for the change that is going to 
happen.

Refactoring program code --- with a goal in mind --- can give the code itself 
more value. That is in the form of robustness to bugs, understandability and 
maintainability. Having robust code is an obvious advantage, but 
understandability and maintainability are both very important aspects of 
software development. By incorporating refactoring in the development process, 
bugs are found faster, new functionality is added more easily and code is easier 
to understand by the next person exposed to it, which might as well be the 
person who wrote it. The consequence of this, is that refactoring can increase 
the average productivity of the development process, and thus also add to the 
monetary value of a business in the long run. The perspective on productivity 
and money should also be able to open the eyes of the many nearsighted managers 
that seldom see beyond the next milestone.

\subsection{The magical number seven}\label{magic_number_seven}
The article \tit{The magical number seven, plus or minus two: some limits on our 
capacity for processing information}\citing{miller1956} by George A.  Miller, 
was published in the journal \name{Psychological Review} in 1956.  It presents 
evidence that support that the capacity of the number of objects a human being 
can hold in its working memory is roughly seven, plus or minus two objects. This 
number varies a bit depending on the nature and complexity of the objects, but 
is according to Miller ``\ldots never changing so much as to be 
unrecognizable.''

Miller's article culminates in the section called \emph{Recoding}, a term he 
borrows from communication theory. The central result in this section is that by 
recoding information, the capacity of the amount of information that a human can 
process at a time is increased. By \emph{recoding}, Miller means to group 
objects together in chunks, and give each chunk a new name that it can be 
remembered by. 

\begin{quote}
  \ldots recoding is an extremely powerful weapon for increasing the amount of 
  information that we can deal with.~\cite[p.~95]{miller1956}
\end{quote}

By organizing objects into patterns of ever growing depth, one can memorize and 
process a much larger amount of data than if it were to be represented as its 
basic pieces. This grouping and renaming is analogous to how many refactorings 
work, by grouping pieces of code and give them a new name.  Examples are the 
fundamental \ExtractMethod and \refa{Extract Class} 
refactorings\citing{refactoring}.

An example from the article addresses the problem of memorizing a sequence of 
binary digits. The example presented here is a slightly modified version of the 
one presented in the original article\citing{miller1956}, but it preserves the 
essence of it. Let us say we have the following sequence of 
16 binary digits: ``1010001001110011''. Most of us will have a hard time 
memorizing this sequence by only reading it once or twice. Imagine if we instead 
translate it to this sequence: ``A273''. If you have a background from computer 
science, it will be obvious that the latter sequence is the first sequence 
recoded to be represented by digits in base 16. Most people should be able to 
memorize this last sequence by only looking at it once.

Another result from the Miller article is that when the amount of information a 
human must interpret increases, it is crucial that the translation from one code 
to another must be almost automatic for the subject to be able to remember the 
translation, before \heshe is presented with new information to recode.  Thus 
learning and understanding how to best organize certain kinds of data is 
essential to efficiently handle that kind of data in the future. This is much 
like when humans learn to read. First they must learn how to recognize letters.  
Then they can learn distinct words, and later read sequences of words that form 
whole sentences. Eventually, most of them will be able to read whole books and 
briefly retell the important parts of its content. This suggests that the use of 
design patterns is a good idea when reasoning about computer programs. With 
extensive use of design patterns when creating complex program structures, one 
does not always have to read whole classes of code to comprehend how they 
function, it may be sufficient to only see the name of a class to almost fully 
understand its responsibilities.

\begin{quote}
  Our language is tremendously useful for repackaging material into a few chunks 
  rich in information.~\cite[p.~95]{miller1956}
\end{quote}

Without further evidence, these results at least indicate that refactoring 
source code into smaller units with higher cohesion and, when needed, 
introducing appropriate design patterns, should aid in the cause of creating 
computer programs that are easier to maintain and have code that is easier (and 
better) understood.

\subsection{Notable contributions to the refactoring literature}

\begin{description}
  \item[1992] William F. Opdyke submits his doctoral dissertation called 
    \tit{Refactoring Object-Oriented Frameworks}\citing{opdyke1992}. This work 
    defines a set of refactorings that are behavior-preserving given that their 
    preconditions are met. The dissertation is focused on the automation of 
    refactorings.
  \item[1999] Martin Fowler et al.: \tit{Refactoring: Improving the Design of 
    Existing Code}\citing{refactoring}. This is maybe the most influential text 
    on refactoring. It bares similarities with Opdykes thesis\citing{opdyke1992} 
    in the way that it provides a catalog of refactorings. But Fowler's book is 
    more about the craft of refactoring, as he focuses on establishing a 
    vocabulary for refactoring, together with the mechanics of different 
    refactorings and when to perform them. His methodology is also founded on 
    the principles of test-driven development.
  \item[2005] Joshua Kerievsky: \tit{Refactoring to 
    Patterns}\citing{kerievsky2005}. This book is heavily influenced by Fowler's 
    \tit{Refactoring}\citing{refactoring} and the ``Gang of Four'' \tit{Design 
    Patterns}\citing{designPatterns}. It is building on the refactoring 
    catalogue from Fowler's book, but is trying to bridge the gap between 
    \emph{refactoring} and \emph{design patterns} by providing a series of 
    higher-level composite refactorings, that makes code evolve toward or away 
    from certain design patterns. The book is trying to build up the reader's 
    intuition around \emph{why} one would want to use a particular design 
    pattern, and not just \emph{how}. The book is encouraging evolutionary 
    design \see{relationToDesignPatterns}.
\end{description}

\subsection{Tool support (for Java)}\label{toolSupport}
This section will briefly compare the refactoring support of the three IDEs 
\name{Eclipse}\footnote{\url{http://www.eclipse.org/}}, \name{IntelliJ 
IDEA}\footnote{The IDE under comparison is the \name{Community Edition}, 
\url{http://www.jetbrains.com/idea/}} and 
\name{NetBeans}\footnote{\url{https://netbeans.org/}}. These are the most 
popular Java IDEs\citing{javaReport2011}.

All three IDEs provide support for the most useful refactorings, like the 
different extract, move and rename refactorings. In fact, Java-targeted IDEs are 
known for their good refactoring support, so this did not appear as a big 
surprise.

The IDEs seem to have excellent support for the \ExtractMethod refactoring, so 
at least they have all passed the first ``refactoring 
rubicon''\citing{fowlerRubicon2001,secondRubicon2012}.

Regarding the \MoveMethod refactoring, the \name{Eclipse} and \name{IntelliJ} 
IDEs do the job in very similar manners. In most situations they both do a 
satisfying job by producing the expected outcome. But they do nothing to check 
that the result does not break the semantics of the program 
\see{sec:correctness}.
The \name{NetBeans} IDE implements this refactoring in a somewhat 
unsophisticated way. For starters, the refactoring's default destination for the 
move, is the same class as the method already resides in, although it refuses to 
perform the refactoring if chosen.  But the worst part is, that if moving the 
method \method{f} of the class \type{C} to the class \type{X}, it will break the 
code.  The result is shown in \myref{lst:moveMethod_NetBeans}.

\begin{listing}
\begin{multicols}{2}
\begin{minted}[samepage]{java}
public class C {
    private X x;
    ...
    public void f() {
        x.m();
        x.n();
    }
}
\end{minted}

\columnbreak

\begin{minted}[samepage]{java}
public class X {
    ...
    public void f(C c) {
        c.x.m();
        c.x.n();
    }
}
\end{minted}
\end{multicols}
\caption{Moving method \method{f} from \type{C} to \type{X}.}
\label{lst:moveMethod_NetBeans}
\end{listing}

\name{NetBeans} will try to create code that call the methods \method{m} and \method{n} 
of \type{X} by accessing them through \var{c.x}, where \var{c} is a parameter of 
type \type{C} that is added the method \method{f} when it is moved. (This is 
seldom the desired outcome of this refactoring, but ironically, this ``feature'' 
keeps \name{NetBeans} from breaking the code in the example from 
\myref{sec:correctness}.) If \var{c.x} for some reason is inaccessible to 
\type{X}, as in this case, the refactoring breaks the code, and it will not 
compile. \name{NetBeans} presents a preview of the refactoring outcome, but the 
preview does not catch it if the IDE is about break the program. 

The IDEs under investigation seem to have fairly good support for primitive 
refactorings, but what about more complex ones, such as 
\gloss{extractClass}\citing{refactoring}? \name{IntelliJ} handles this in a 
fairly good manner, although, in the case of private methods, it leaves unused 
methods behind. These are methods that delegate to a field with the type of the 
new class, but are not used anywhere. \name{Eclipse} has added its own quirk to 
the \refa{Extract Class} refactoring, and only allows for \emph{fields} to be 
moved to a new class, \emph{not methods}. This makes it effectively only 
extracting a data structure, and calling it \refa{Extract Class} is a little 
misleading.  One would often be better off with textual extract and paste than 
using the \refa{Extract Class} refactoring in \name{Eclipse}. When it comes to 
\name{NetBeans}, it does not even show an attempt on providing this refactoring.  

\subsection{The relation to design patterns}\label{relationToDesignPatterns}

Refactoring and design patterns have at least one thing in common, they are both 
promoted by advocates of \emph{clean code}\citing{cleanCode} as fundamental 
tools on the road to more maintainable and extendable source code.

\begin{quote}
  Design patterns help you determine how to reorganize a design, and they can 
  reduce the amount of refactoring you need to do 
  later.~\cite[p.~353]{designPatterns}
\end{quote}

Although sometimes associated with 
over-engineering\citing{kerievsky2005,refactoring}, design patterns are in 
general assumed to be good for maintainability of source code.  That may be 
because many of them are designed to support the \emph{open/closed principle} of 
object-oriented programming. The principle was first formulated by Bertrand 
Meyer, the creator of the Eiffel programming language, like this: ``Modules 
should be both open and closed.''\citing{meyer1988} It has been popularized, 
with this as a common version: 

\begin{quote}
  Software entities (classes, modules, functions, etc.) should be open for 
  extension, but closed for modification.
\end{quote} 

Maintainability is often thought of as the ability to be able to introduce new 
functionality without having to change too much of the old code. When 
refactoring, the motivation is often to facilitate adding new functionality. It 
is about factoring the old code in a way that makes the new functionality being 
able to benefit from the functionality already residing in a software system, 
without having to copy old code into new. Then, next time someone shall add new 
functionality, it is less likely that the old code has to change. Assuming that 
a design pattern is the best way to get rid of duplication and assist in 
implementing new functionality, it is reasonable to conclude that a design 
pattern often is the target of a series of refactorings. Having a repertoire of 
design patterns can also help in knowing when and how to refactor a program to 
make it reflect certain desired characteristics.

\begin{quote}
  There is a natural relation between patterns and refactorings. Patterns are 
  where you want to be; refactorings are ways to get there from somewhere 
  else.~\cite[p.~107]{refactoring}
\end{quote}

This quote is wise in many contexts, but it is not always appropriate to say 
``Patterns are where you want to be\ldots''. \emph{Sometimes}, patterns are 
where you want to be, but only because it will benefit your design. It is not 
true that one should always try to incorporate as many design patterns as 
possible into a program. It is not like they have intrinsic value. They only add 
value to a system when they support its design. Otherwise, the use of design 
patterns may only lead to a program that is more complex than necessary.

\begin{quote}
  The overuse of patterns tends to result from being patterns happy. We are 
  \emph{patterns happy} when we become so enamored of patterns that we simply 
  must use them in our code.~\cite[p.~24]{kerievsky2005}
\end{quote}

This can easily happen when relying largely on up-front design. Then it is 
natural, in the very beginning, to try to build in all the flexibility that one 
believes will be necessary throughout the lifetime of a software system.  
According to Joshua Kerievsky ``That sounds reasonable --- if you happen to be 
psychic.''~\cite[p.~1]{kerievsky2005} He is advocating what he believes is a 
better approach: To let software continually evolve. To start with a simple 
design that meets today's needs, and tackle future needs by refactoring to 
satisfy them. He believes that this is a more economic approach than investing 
time and money into a design that inevitably is going to change. By relying on 
continuously refactoring a system, its design can be made simpler without 
sacrificing flexibility. To be able to fully rely on this approach, it is of 
utter importance to have a reliable suit of tests to lean on \see{testing}. This 
makes the design process more natural and less characterized by difficult 
decisions that has to be made before proceeding in the process, and that is 
going to define a project for all of its unforeseeable future.

\subsection{The impact on software quality}

\subsubsection{What is software quality?}
The term \emph{software quality} has many meanings. It all depends on the 
context we put it in. If we look at it with the eyes of a software developer, it 
usually means that the software is easily maintainable and testable, or in other 
words, that it is \emph{well designed}. This often correlates with the 
management scale, where \emph{keeping the schedule} and \emph{customer 
satisfaction} is at the center. From the customers point of view, in addition to 
good usability, \emph{performance} and \emph{lack of bugs} is always 
appreciated, measurements that are also shared by the software developer. (In 
addition, such things as good documentation could be measured, but this is out 
of the scope of this document.)

\subsubsection{The impact on performance}
\begin{quote}
  Refactoring certainly will make software go more slowly\footnote{With today's 
  compiler optimization techniques and performance tuning of e.g. the Java 
virtual machine, the penalties of object creation and method calls are 
debatable.}, but it also makes the software more amenable to performance 
tuning.~\cite[p.~69]{refactoring}
\end{quote}

\noindent There is a common belief that refactoring compromises performance, due 
to increased degree of indirection and that polymorphism is slower than 
conditionals.

In a survey, Demeyer\citing{demeyer2002} disproves this view in the case of 
polymorphism. He did an experiment on, what he calls, ``Transform Self Type 
Checks'' where you introduce a new polymorphic method and a new class hierarchy 
to get rid of a class' type checking of a ``type attribute``. He uses this kind 
of transformation to represent other ways of replacing conditionals with 
polymorphism as well. The experiment is performed on the C++ programming 
language and with three different compilers and platforms. Demeyer concludes 
that, with compiler optimization turned on, polymorphism beats middle to large 
sized if-statements and does as well as case-statements.  (In accordance with 
his hypothesis, due to similarities between the way C++ handles polymorphism and 
case-statements.)

\begin{quote}
  The interesting thing about performance is that if you analyze most programs, 
  you find that they waste most of their time in a small fraction of the 
  code.~\cite[p.~70]{refactoring}
\end{quote}

\noindent So, although an increased amount of method calls could potentially 
slow down programs, one should avoid premature optimization and sacrificing good 
design, leaving the performance tuning until after \gloss{profiling} the 
software and having isolated the actual problem areas.

\subsection{Composite refactorings}\label{compositeRefactorings}
Generally, when thinking about refactoring, at the mechanical level, there are 
essentially two kinds of refactorings. There are the \emph{primitive} 
refactorings, and the \emph{composite} refactorings. 

\definition{A \emph{primitive refactoring} is a refactoring that cannot be 
expressed in terms of other refactorings.}

\noindent Examples are the \refa{Pull Up Field} and \refa{Pull Up 
Method} refactorings\citing{refactoring}, that move members up in their class 
hierarchies.

\definition{A \emph{composite refactoring} is a refactoring that can be 
expressed in terms of two or more other refactorings.}

\noindent An example of a composite refactoring is the \refa{Extract 
Superclass} refactoring\citing{refactoring}. In its simplest form, it is composed 
of the previously described primitive refactorings, in addition to the 
\refa{Pull Up Constructor Body} refactoring\citing{refactoring}. It works 
by creating an abstract superclass that the target class(es) inherits from, then 
by applying \refa{Pull Up Field}, \refa{Pull Up Method} and 
\refa{Pull Up Constructor Body} on the members that are to be members of 
the new superclass. If there are multiple classes in play, their interfaces may 
need to be united with the help of some rename refactorings, before extracting 
the superclass. For an overview of the \refa{Extract Superclass} 
refactoring, see \myref{fig:extractSuperclass}.

\begin{figure}[h]
  \centering
  \includegraphics[angle=270,width=\linewidth]{extractSuperclassItalic.pdf}
  \caption{The Extract Superclass refactoring, with united interfaces. (Taken 
    from \url{http://refactoring.com/catalog/extractSuperclass.html}.)}
  \label{fig:extractSuperclass}
\end{figure}

\subsection{Manual vs. automated refactorings}
Refactoring is something every programmer does, even if \heshe does not known 
the term \emph{refactoring}. Every refinement of source code that does not alter 
the program's behavior is a refactoring. For small refactorings, such as 
\ExtractMethod, executing it manually is a manageable task, but is still prone 
to errors. Getting it right the first time is not easy, considering the method 
signature and all the other aspects of the refactoring that has to be in place.  

Consider the renaming of classes, methods and fields. For complex programs these 
refactorings are almost impossible to get right.  Attacking them with textual 
search and replace, or even regular expressions, will fall short on these tasks.  
Then it is crucial to have proper tool support that can perform them 
automatically. Tools that can parse source code and thus have semantic knowledge 
about which occurrences of which names belong to what construct in the program.  
For even trying to perform one of these complex tasks manually, one would have 
to be very confident on the existing test suite \see{testing}.

\subsection{Correctness of refactorings}\label{sec:correctness}
For automated refactorings to be truly useful, they must show a high degree of 
behavior preservation.  This last sentence might seem obvious, but there are 
examples of refactorings in existing tools that break programs. In an ideal 
world, every automated refactoring would be ``complete'', in the sense that it 
would never break a program. In an ideal world, every program would also be free 
from bugs. In modern IDEs the implemented automated refactorings are working for 
\emph{most} cases, which is enough for making them useful.

I will now present an example of a \emph{corner case} where a program breaks 
when a refactoring is applied. The example shows an \ExtractMethod refactoring 
followed by a \MoveMethod refactoring that breaks a program in both the 
\name{Eclipse} and \name{IntelliJ} IDEs\footnote{The \name{NetBeans} IDE handles this 
  particular situation without altering the program's behavior, mainly because 
  its \refa{Move Method} refactoring implementation is a bit flawed in other ways 
  \see{toolSupport}.}.  The target and the destination for the composed 
  refactoring are shown in \myref{lst:correctnessExtractAndMove}.  Note that the 
  method \method{m(C c)} of class \type{X} assigns to the field \var{x} of the 
  argument \var{c} that has type \type{C}.

\begin{listing}[h]
\begin{multicols}{2}
\begin{minted}[linenos,frame=topline,label={Refactoring 
  target},framesep=\mintedframesep]{java}
public class C {
  public X x = new X();

  public void f() {
    x.m(this);
    // Not the same x.
    x.n();
  }
}
\end{minted}

\columnbreak

\begin{minted}[frame=topline,label={Method 
  destination},framesep=\mintedframesep]{java}
public class X {
  public void m(C c) {
    c.x = new X();
    // If m is called from
    // c, then c.x no longer
    // equals 'this'.
  }
  public void n() {}
}
\end{minted}
\end{multicols}
\caption{The target and the destination for the composition of the Extract 
Method and \refa{Move Method} refactorings.}
\label{lst:correctnessExtractAndMove}
\end{listing}


The refactoring sequence works by extracting line 6 through 8 from the original 
class \type{C} into a method \method{f} with the statements from those lines as 
its method body (but with the comment left out, since it will no longer hold any 
meaning). The method is then moved to the class \type{X}.  The result is shown 
in \myref{lst:correctnessExtractAndMoveResult}.

Before the refactoring, the methods \method{m} and \method{n} of class \type{X} 
are called on different object instances (see line 6 and 8 of the original class 
\type{C} in \cref{lst:correctnessExtractAndMove}). After the refactoring, they 
are called on the same object, and the statement on line 
3 of class \type{X} (in \cref{lst:correctnessExtractAndMoveResult}) no longer 
  has the desired effect in our example. The method \method{f} of class \type{C} 
  is now calling the method \method{f} of class \type{X} (see line 5 of class 
  \type{C} in \cref{lst:correctnessExtractAndMoveResult}), and the program now 
  behaves different than before.

\begin{listing}[h]
\begin{multicols}{2}
\begin{minted}[linenos]{java}
public class C {
    public X x = new X();

    public void f() {
        x.f(this);
    }
}
\end{minted}

\columnbreak

\begin{minted}[linenos]{java}
public class X {
    public void m(C c) {
        c.x = new X();
    }
    public void n() {}
    // Extracted and 
    // moved method.
    public void f(C c) {
        m(c);
        n();
    }
}
\end{minted}
\end{multicols}
\caption{The result of the composed refactoring.}
\label{lst:correctnessExtractAndMoveResult}
\end{listing}

The bug introduced in the previous example is of such a nature\footnote{Caused 
by aliasing.} that it is very difficult to spot if the refactored code is not 
covered by tests.  It does not generate compilation errors, and will thus only 
result in a runtime error or corrupted data, which might be hard to detect.

\subsection{Refactoring and the importance of testing}\label{testing}
\begin{quote}
  If you want to refactor, the essential precondition is having solid 
  tests.\citing{refactoring}
\end{quote}

When refactoring, there are roughly three classes of errors that can be made.  
The first class of errors is the one that makes the code unable to compile.  
These \emph{compile-time} errors are of the nicer kind. They flash up at the 
moment they are made (at least when using an IDE), and are usually easy to fix.  
The second class is the \emph{runtime} errors. Although these errors take a bit 
longer to surface, they usually manifest after some time in an illegal argument 
exception, null pointer exception or similar during the program execution.  
These kinds of errors are a bit harder to handle, but at least they will show, 
eventually. Then there are the \emph{behavior-changing} errors. These errors are 
of the worst kind. They do not show up during compilation and they do not turn 
on a blinking red light during runtime either. The program can seem to work 
perfectly fine with them in play, but the business logic can be damaged in ways 
that will only show up over time.

For discovering runtime errors and behavior changes when refactoring, it is 
essential to have good test coverage. Testing in this context means writing 
automated tests. Manual testing may have its uses, but when refactoring, it is 
automated unit testing that dominate. For discovering behavior changes it is 
especially important to have tests that cover potential problems, since these 
kinds of errors do not reveal themselves.

Unit testing is not a way to \emph{prove} that a program is correct, but it is a 
way to make you confident that it \emph{probably} works as desired.  In the 
context of test-driven development (commonly known as TDD), the tests are even a 
way to define how the program is \emph{supposed} to work.  It is then, by 
definition, working if the tests are passing.  

If the test coverage for a code base is perfect, then it should, theoretically, 
be risk-free to perform refactorings on it. This is why automated tests and 
refactoring is such a great match.

\subsubsection{Testing the code from correctness section}
The worst thing that can happen when refactoring is to introduce changes to the 
behavior of a program, as in the example on \myref{sec:correctness}. This 
example may be trivial, but the essence is clear. The only problem with the 
example is that it is not clear how to create automated tests for it, without 
changing it in intrusive ways.

Unit tests, as they are known from the different \glosspl{xUnit} around, are 
only suitable to test the \emph{result} of isolated operations. They can not 
easily (if at all) observe the \emph{history} of a program.

This problem is still open.

\begin{comment}

Assuming a sequential (non-concurrent) program:

\begin{minted}{java}
tracematch (C c, X x) {
  sym m before:
    call(* X.m(C)) && args(c) && cflow(within(C));
  sym n before:
    call(* X.n()) && target(x) && cflow(within(C));
  sym setCx after:
    set(C.x) && target(c) && !cflow(m);

  m n

  { assert x == c.x; }
}
\end{minted}

%\begin{minted}{java}
%tracematch (X x1, X x2) {
%  sym m before:
%    call(* X.m(C)) && target(x1);
%  sym n before:
%    call(* X.n()) && target(x2);
%  sym setX after:
%    set(C.x) && !cflow(m) && !cflow(n);
%
%  m n
%
%  { assert x1 != x2; }
%}
%\end{minted}
\end{comment}


\section{The project}\label{sec:project}
In this section we look at the work that will be done for this project, its 
building blocks, pose some research questions and present some of the 
methodologies used.

\subsection{Project description}
The aim of this master's project will be to explore the relationship between the 
\ExtractMethod and the \MoveMethod refactorings. This will be done by composing 
the two into a composite refactoring. The refactoring will be called the 
\ExtractAndMoveMethod refactoring. 

The two primitive refactorings \ExtractMethod and \MoveMethod must already be 
implemented in a tool, so the \ExtractAndMoveMethod refactoring can be built on 
top of these.

The composition of the \ExtractMethod and \MoveMethod refactorings springs 
naturally out of the need to move procedures closer to the data they manipulate.  
This composed refactoring is not well described in the literature, but it is 
implemented in at least one tool called 
\href{https://help.devexpress.com/\#CodeRush/CustomDocument3519}{\name{CodeRush}}, 
which is an extension for \href{http://www.visualstudio.com/}{\name{MS Visual 
Studio}}. In CodeRush it is called \refa{Extract Method to 
Type}\footnote{\url{https://help.devexpress.com/\#CodeRush/CustomDocument6710}}, 
but I choose to call it \ExtractAndMoveMethod, since I feel this better 
communicates which primitive refactorings it is composed of. 

The project will consist of implementing the \ExtractAndMoveMethod refactoring, 
as well as executing it over a larger code base, as a case study. To be able to 
execute the refactoring automatically, I have to make it analyze code to 
determine the best selections to extract into new methods.

\subsection{The premises}
Before we can start manipulating source code, and write a tool for doing so, we 
need to decide on a programming language for the code we are going to 
manipulate. Also, since we do not want to start from scratch by implementing 
primitive refactorings ourselves, we need to choose an existing tool that 
provides the needed refactorings. In addition to be able to perform changes, we 
need a framework for analyzing source code for the language we select.

\subsubsection{Choosing the target language}
Choosing which programming language the code that will be manipulated shall be 
written in, is not a very difficult task. We choose to limit the possible 
languages to the object-oriented programming languages, since most of the 
terminology and literature regarding refactoring comes from the world of 
object-oriented programming. In addition, the language must have existing tool 
support for refactoring.

The \href{https://www.java.com/}{\name{Java} programming language} is the 
dominating language when it comes to example code in the literature of 
refactoring, and is thus a natural choice. Java is perhaps the most influential 
programming language in the world today, with its \name{Java Virtual Machine} 
that runs on all of the most popular computer architectures and also supports 
dozens of other programming languages\footnote{They compile to Java bytecode.}, 
with \name{Scala}, \name{Clojure} and \name{Groovy} as the most prominent ones.  
Java is currently the language that every other programming language is compared 
against. It is also the primary programming language for the author of this 
thesis.

\subsubsection{Choosing the tools}
When choosing a tool for manipulating Java, there are certain criteria that 
have to be met. First of all, the tool should have some existing refactoring 
support that this thesis can build upon. Secondly, it should provide some kind 
of framework for parsing and analyzing Java source code. Third, it should itself 
be open source. This is both because of the need to be able to browse the code 
for the existing refactorings that is contained in the tool, and also because 
open source projects hold value in them selves. Another important aspect to 
consider, is that open source projects of a certain size usually has large 
communities of people connected to them, which are committed to answering 
questions regarding the use and misuse of the products, that to a large degree 
is made by the community itself.

There is a certain class of tools that meet these criteria, namely the class of 
\emph{IDEs}\footnote{\emph{Integrated Development Environment}}. These are 
programs that are meant to support the whole production cycle of a computer 
program, and the most popular IDEs that support Java generally have quite good 
refactoring support.

The main contenders for this thesis are the \name{Eclipse IDE}, with the 
\name{Java development tools} (JDT), the \name{IntelliJ IDEA Community Edition} 
and the \name{NetBeans IDE} \see{toolSupport}. \name{Eclipse} and 
\name{NetBeans} are both free, open source and community driven, while the 
\name{IntelliJ IDEA} has an open-sourced community edition that is free of 
charge, but also offers an \name{Ultimate Edition} with an extended set of 
features, at additional cost.  All three IDEs supports adding plugins to extend 
their functionality and tools that can be used to parse and analyze Java source 
code. But one of the IDEs stand out as a favorite, and that is the \name{Eclipse 
IDE}. This is the most popular\citing{javaReport2011} among the three and seems 
to be de facto standard IDE for Java development regardless of platform.


\subsection{The primitive refactorings}
The refactorings presented here are the primitive refactorings used in this 
project. They are the abstract building blocks used by the \ExtractAndMoveMethod 
refactoring. 

\paragraph{The Extract Method refactoring.}
The \refa{Extract Method} refactoring\citing{refactoring} is used to extract 
a fragment of code from its context and into a new method. A call to the new 
method is inlined where the fragment was before. It is used to break code into 
logical units, with names that explain their purpose.

An example of an \ExtractMethod refactoring is shown in 
\myref{lst:extractMethodRefactoring}. It shows a method containing calls to the 
methods \method{foo} and \method{bar} of a type \type{X}. These statements are 
then extracted into the new method \method{fooBar}.

\begin{listing}[h]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\rightColX{32*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{2*\lineheight+5.8pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Before
  \draw[overlaybox] (\indent,\mintedtop+\lineheight*3) rectangle 
  +(19*\charwidth,\lineheight);

  % After
  \draw[overlaybox] (\rightColX+\indent,\mintedtop+\lineheight*3) rectangle 
  +(16*\charwidth,\lineheight);

  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*5) rectangle 
  +(21*\charwidth,3*\lineheight);
\end{tikzpicture}
  \begin{multicols}{2}
  \begin{minted}[samepage,frame=topline,label={Before},framesep=\mintedframesep]{java}
class C {
  void method() {
    X x = new X();
    x.foo(); x.bar();
  }
}
  \end{minted}

  \columnbreak

  \begin{minted}[samepage,frame=topline,label={After},framesep=\mintedframesep]{java}
class C {
  void method() {
    X x = new X();
    fooBar(x);
  }
  void fooBar(X x) {
    x.foo(); x.bar();
  }
}
  \end{minted}
  \end{multicols}
  \caption{An example of an \ExtractMethod refactoring.}
  \label{lst:extractMethodRefactoring}
\end{listing}

\paragraph{The Move Method refactoring.}
The \refa{Move Method} refactoring\citing{refactoring} is used to move a method 
from one class to another. This can be appropriate if the method is using more 
features of another class than of the class in which it is currently defined.  

\Myref{lst:moveMethodRefactoring} shows an example of this refactoring. Here, a 
method \method{fooBar} is moved from a class \type{C} to a class \type{X}.

\begin{listing}[h]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\rightColX{32*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{2*\lineheight+5.8pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Before
  \draw[overlaybox] (0,\mintedtop+\lineheight*5) rectangle 
  +(21*\charwidth,3*\lineheight);

  % After
  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*10) rectangle 
  +(22*\charwidth,3*\lineheight);
\end{tikzpicture}
\begin{multicols}{2}
  \begin{minted}[samepage,frame=topline,label={Before},framesep=\mintedframesep]{java}
class C {
  void method() {
    X x = new X();
    fooBar(x);
  }
  void fooBar(X x) {
    x.foo(); x.bar();
  }
}

class X {
  void foo(){/*...*/}
  void bar(){/*...*/}
}
  \end{minted}

  \columnbreak

  \begin{minted}[samepage,frame=topline,label={After},framesep=\mintedframesep]{java}
class C {
  void method() {
    X x = new X();
    x.fooBar();
  }
}

class X {
  void foo(){/*...*/}
  void bar(){/*...*/}
  void fooBar() {
    foo(); bar();
  }
}
  \end{minted}
\end{multicols}
\caption{An example of a \MoveMethod refactoring.}
\label{lst:moveMethodRefactoring}
\end{listing}

\subsection{The Extract and Move Method refactoring}
The \ExtractAndMoveMethod refactoring is a composite refactoring composed of the 
primitive refactorings \ExtractMethod and \MoveMethod. The effect of this 
refactoring on source code is the same as when extracting a method and moving it 
to another class. Conceptually, this is done without an intermediate step. In 
practice, as we shall see later, an intermediate step may be necessary.

An example of this composite refactoring is shown in 
\myref{lst:extractAndMoveMethodRefactoring}. The example joins the examples from 
\cref{lst:extractMethodRefactoring} and \cref{lst:moveMethodRefactoring}. This 
means that the selection consisting of the consecutive calls to the methods 
\method{foo} and \method{bar}, is extracted into a new method \method{fooBar} 
located in the class \type{X}.

\begin{listing}[h]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\rightColX{32*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{2*\lineheight+5.8pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Before
  \draw[overlaybox] (\indent,\mintedtop+\lineheight*3) rectangle 
  +(19*\charwidth,\lineheight);

  % After
  \draw[overlaybox] (\rightColX+\indent,\mintedtop+\lineheight*3) rectangle 
  +(16*\charwidth,\lineheight);

  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*10) rectangle 
  +(22*\charwidth,3*\lineheight);
\end{tikzpicture}
\begin{multicols}{2}
  \begin{minted}[samepage,frame=topline,label={Before},framesep=\mintedframesep]{java}
class C {
  void method() {
    X x = new X();
    x.foo(); x.bar();
  }
}

class X {
  void foo(){/*...*/}
  void bar(){/*...*/}
}
  \end{minted}

  \columnbreak

  \begin{minted}[samepage,frame=topline,label={After},framesep=\mintedframesep]{java}
class C {
  void method() {
    X x = new X();
    x.fooBar();
  }
}

class X {
  void foo(){/*...*/}
  void bar(){/*...*/}
  void fooBar() {
    foo(); bar();
  }
}
  \end{minted}
\end{multicols}
\caption{An example of the \ExtractAndMoveMethod refactoring.}
\label{lst:extractAndMoveMethodRefactoring}
\end{listing}

\subsection{The Coupling Between Object Classes metric}\label{sec:CBO}
The best known metric for measuring coupling between classes in object-oriented 
software is called \metr{Coupling Between Object Classes}, usually abbreviated 
as CBO. The metric is defined in the article \tit{A Metrics Suite for Object 
Oriented Design}\citing{metricsSuite1994} by Chidamber and Kemerer, published in
1994.

\definition{\emph{CBO} for a class is a count of the number of other classes to 
which it is coupled.}

An object is coupled to another object if one of them acts on the other by using 
methods or instance variables of the other object. This relation goes both ways, 
so both outgoing and incoming uses are counted. Each coupling relationship is 
only considered once when measuring CBO for a class.

\paragraph{How can the Extract and Move Method refactoring improve CBO?}
\Myref{lst:CBOExample} shows how CBO changes for a class when it is refactored 
with the \ExtractAndMoveMethod refactoring. In the example we consider only the 
CBO value of class \type{C}.

\begin{listing}[h]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\rightColX{32*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{2*\lineheight+5.8pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Before
  \draw[overlaybox] (\indent,\mintedtop+\lineheight*4) rectangle 
  +(15*\charwidth,2*\lineheight);

  % After
  \draw[overlaybox] (\rightColX+\indent,\mintedtop+\lineheight*4) rectangle 
  +(15*\charwidth,\lineheight);

  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*15) rectangle 
  +(19*\charwidth,4*\lineheight);
\end{tikzpicture}
\begin{multicols}{2}
\begin{minted}[linenos,samepage,frame=topline,label={Before},framesep=\mintedframesep]{java}
class C {
  A a; B b;
  X x;
  void method() {
    x.y.foo();
    x.y.bar();
  }
  /* Uses of A and B.
     No uses of other 
     classes. */
}

class X {
  Y y;
  /* No uses of C.
     Uses of Y. */
}

class Y {
  void foo(){
    /* No uses of C. */
  }
  void bar(){
    /* No uses of C. */
  }
}
\end{minted}

\columnbreak

\begin{minted}[linenos,samepage,frame=topline,label={After},framesep=\mintedframesep]{java}
class C {
  A a; B b;
  X x;
  void method() {
    x.fooBar();
  }
  /* Uses of A and B.
     No uses of other 
     classes. */
}

class X {
  Y y;
  /* No uses of C.
     Uses of Y. */
  void fooBar() {
    y.foo();
    y.bar();
  }
}

class Y {
  void foo(){
    /* No uses of C. */
  }
  void bar(){
    /* No uses of C. */
  }
}
  \end{minted}
\end{multicols}
\caption{An example of improving CBO. Class \type{C} has a CBO value of 4 
before refactoring it, and 3 after.}
\label{lst:CBOExample}
\end{listing}

Before refactoring the class \type{C} with the \ExtractAndMoveMethod 
refactoring, it has a CBO value of 4. The class uses members of the classes 
\type{A} and \type{B}, which accounts for 2 of the coupling relationships of 
class \type{C}. In addition to this, it uses its variable \var{x} with type 
\type{X} and also the methods \method{foo} and \method{bar} declared in class 
\type{Y}, giving it a total CBO value of 4.

The after-part of the example code in \mysimpleref{lst:CBOExample} shows the 
result of extracting the lines 
5 and 6 of class \type{C} into a new method \method{fooBar}, with a subsequent 
  move of it to class \type{X}.

With respect to the CBO metric, the refactoring action accomplishes something 
important: It eliminates the uses of class \type{Y} from class \type{C}. This 
means that the class \type{C} is no longer coupled to \type{Y}, only the classes 
\type{A}, \type{B} and \type{X}. The CBO value of class \type{C} is therefore 3 
after the refactoring, while no other class have received any increase in CBO.

The example shown here is an ideal situation. Coupling is reduced for one class 
without any increase of coupling for another class. There is also another 
important point: It is the fact that to reduce the CBO value for a class, we 
need to remove \emph{all} its uses of another class. This is done for the class 
\type{C} in \myref{lst:CBOExample}, where all uses of class \type{Y} is removed 
by the \ExtractAndMoveMethod refactoring.


\subsection{Research questions}\label{sec:researchQuestions}
The main question that I seek an answer to in this thesis is:

\begin{quote}
  Is it possible to automate the analysis and execution of the 
  \ExtractAndMoveMethod refactoring, and do so for all of the code of a larger 
  project?
\end{quote}

\noindent The secondary questions will then be:

\paragraph{Can we do this efficiently?} Can we automate the analysis and 
execution of the refactoring so it can be run in a reasonable amount of time?  

\paragraph{Can we perform changes safely?} Can we take actions to prevent the 
refactoring from breaking code? By ``breaking code'' we mean to either do 
changes that do not compile, or make changes that alter the behavior of a 
program.

\paragraph{Can we improve the quality of source code?} Assuming that the 
refactoring is safe: Is it feasible to assure that code we refactor actually 
gets better in terms of coupling?

\paragraph{How can the automation of the refactoring be helpful?} Assuming the 
refactoring does in fact improve the quality of source code and is safe to use: 
What is the usefulness of the refactoring in a software development setting?  In 
what parts of the development process can the refactoring play a role?

\subsection{Methodology}
This section will present some of the methodologies used during the work of this 
thesis.

\subsubsection{Evolutionary design}
In the programming work for this project, I have tried using a design strategy 
called evolutionary design, also known as continuous or incremental 
design\citing{wiki_continuous_2014}. It is a software design strategy advocated 
by the Extreme Programming community. The essence of the strategy is that you 
should let the design of your program evolve naturally as your requirements 
change.  This is seen in contrast with up-front design, where design decisions 
are made early in the process. 

The motivation behind evolutionary design is to keep the design of software as 
simple as possible. This means not introducing unneeded functionality into a 
program. You should defer introducing flexibility into your software, until it 
is needed to be able to add functionality in a clean way.

Holding up design decisions, implies that the time will eventually come when 
decisions have to be made. The flexibility of the design then relies on the 
programmer's abilities to perform the necessary refactoring, and \his confidence 
in those abilities. From my experience working on this project, I can say that 
this confidence is greatly enhanced by having automated tests to rely on 
\see{tdd}.

The choice of going for evolutionary design developed naturally. As Fowler 
points out in his article \tit{Is Design Dead?}, evolutionary design much 
resembles the ``code and fix'' development strategy\citing{fowler_design_2004}.
A strategy which most of us have practiced in school. This was also the case 
when I first started this work. I had to learn the inner workings of Eclipse and 
its refactoring-related plugins. That meant a lot of fumbling around with code I 
did not know, in a trial and error fashion. Eventually I started writing tests 
for my code, and my design began to evolve.

\subsubsection{Test-driven development}\label{tdd}
As mentioned before, the project started out as a classic code and fix 
development process. My focus was aimed at getting something to work, rather 
than doing so according to best practice. This resulted in a project that got 
out of its starting blocks, but it was not accompanied by any tests. Hence it 
was soon difficult to make any code changes with the confidence that the program 
was still correct afterwards (assuming it was so before changing it). I always 
knew that I had to introduce some tests at one point, but this experience 
accelerated the process of leading me onto the path of testing.

I then wrote tests for the core functionality of the plugin, and thus gained 
more confidence in the correctness of my code. I could now perform quite drastic 
changes without ``wetting my pants``. After this, nearly all of the semantic 
changes done to the business logic of the project, or the addition of new 
functionality, were made in a test-driven manner. This means that before 
performing any changes, I would define the desired functionality through a set 
of tests. I would then run the tests to check that they were run and that they 
did not pass.  Then I would do any code changes necessary to make the tests 
pass.  The definition of how the program is supposed to operate is then captured 
by the tests.  However, this does not prove the correctness of the analysis 
leading to the test definitions.

\subsubsection{Case studies}
The case study methodology is used to show how the \ExtractAndMoveMethod 
refactoring performs on real code, not just toy examples. The case studies are 
used to analyze our project so we can conclude on its completeness and 
usefulness.

\subsubsection{Dogfooding}
Dogfooding is a methodology where you use your own tools to do your job, also 
referred to as ``eating your own dog food''\citing{harrisonDogfooding2006}. It 
is used in this project to see if we can refactor our own refactoring code and 
still use it to refactor other code.

\section{Related work}\label{sec:relatedWork}
Here, some work is presented that relate to automated composition of 
refactorings.

\subsection{Search-based refactoring}
\tit{Search-Based Refactoring: an
empirical study}\citing{okeeffeSearchBased2008} is a paper by Mark O'Keeffe and 
Mel Ó Cinnéide published in 2008. The authors present an empirical study of 
different algorithmic approaches to search-based refactoring.

The common approach for all these algorithms is to generate a set of changes to 
a program for then to use a ``fitness function'' to evaluate if they improve its
design or not.  The fitness function consists of a weighted sum of different 
object-oriented metrics.

Among other things, the authors conclude that even with small input programs, 
their solution representation is memory-intensive, at least for some algorithms.  
The programs they refactor on have in average 4,000 lines of code, spread over 
57 classes. I.e. considerably smaller than one of the programs that will be 
   subject to refactoring in this project.

\subsection{``Making Program Refactoring 
Safer''}\label{sec:saferRefactoringTests}
This is the name of an article\citing{soaresSafer2010} about providing a way to 
improve safety during refactoring. Soares et al. approaches the problem of 
preserving behavior during refactoring by analyzing a transformation and then 
generate a test suite for it, using static analysis. These tests are then run 
for both the before- and after-code, and is compared to assure that they are 
consistent.

\subsection{The compositional paradigm of refactoring}
This paradigm builds upon the observation of Vakilian et 
al.\citing{vakilian2012}, that of the many automated refactorings existing in 
modern IDEs, the simplest ones are dominating the usage statistics. The report 
mainly focuses on \name{Eclipse} as the tool under investigation.

The paradigm is described almost as the opposite of automated composition of 
refactorings \see{compositeRefactorings}. It works by providing the programmer 
with easily accessible primitive refactorings. These refactorings shall be 
accessed via keyboard shortcuts or quick-assist menus\footnote{Think 
quick-assist with Ctrl+1 in \name{Eclipse}} and be promptly executed, opposed to in the 
currently dominating wizard-based refactoring paradigm. They are meant to 
stimulate composing smaller refactorings into more complex changes, rather than 
doing a large upfront configuration of a wizard-based refactoring, before 
previewing and executing it. The compositional paradigm of refactoring is 
supposed to give control back to the programmer, by supporting \himher with an 
option of performing small rapid changes instead of large changes with a lesser 
degree of control. The report authors hope this will lead to fewer unsuccessful 
refactorings. It also could lower the bar for understanding the steps of a 
larger composite refactoring and thus also helps in figuring out what goes wrong 
if one should choose to opt in on a wizard-based refactoring.

Vakilian and his associates have performed a survey of the effectiveness of the 
compositional paradigm versus the wizard-based. They claim to have found 
evidence of the \emph{compositional paradigm} outperforming the 
\emph{wizard-based}. It does so by reducing automation, which seems 
counterintuitive. Therefore they ask the question ``What is an appropriate level 
of automation?'', and thus challenging what they feel is a rush toward more 
automation in the software engineering community.

\chapter{The search-based Extract and Move Method 
refactoring}\label{ch:extractAndMoveMethod}
In this chapter I will delve into the workings of the search-based 
\ExtractAndMoveMethod refactoring. We will see the choices it must make along 
the way and why it chooses a text selection as a candidate for refactoring or 
not.

After defining some concepts, I will introduce an example that will be used 
throughout the chapter to illustrate how the refactoring works in some simple 
situations.

\section{The inputs to the refactoring}
For executing an \ExtractAndMoveMethod refactoring, there are two simple 
requirements. The first thing the refactoring needs is a text selection, telling 
it what to extract. Its second requirement is a target for the subsequent move 
operation. 

When the refactoring performs changes to source code, the extracted method must 
be called in place of the selection that now makes up the method's body. Also, 
the method call has to be performed via a variable, since the method is not 
static \see{par:ignoringStatic}. Therefore, the move target must be a local 
variable or a field in the scope of the text selection.  The actual new location 
for the extracted method will be the class representing the type of the move 
target variable.

\section{Defining a text selection}
A text selection, in our context, is very similar to what you think of when 
selecting a bit of text in your editor or other text processing tool with your 
mouse or keyboard. It is an abstract construct that is meant to capture which 
specific portion of text we are about to deal with.

To be able to clearly reason about a text selection done to a portion of text in 
a computer file, which consists of pure text, we put up the following 
definition:

\definition{A \emph{text selection} in a text file is defined by two 
non-negative integers, in addition to a reference to the file itself. The first 
integer is an offset into the file, while the second reference is the length of 
the text selection.}

This means that the selected text consist of a number of characters equal to the 
length of the selection, where the first character is found at the specified 
offset.

\section{Where we look for text selections}

\subsection{Text selections are found in methods}
The text selections we are interested in are those that surround program 
statements. Therefore, the place we look for selections that can form candidates 
for an execution of the \ExtractAndMoveMethod refactoring, is within the body of 
a single method.

\paragraph{On ignoring static methods.}\label{par:ignoringStatic}
In this project we are not analyzing static methods for candidates to the 
\ExtractAndMoveMethod refactoring. The reason for this is that in the cases 
where we want to perform the refactoring for a selection within a static method, 
the first step is to extract the selection into a new method. Hence this method
also becomes static, since it must be possible to call it from a static context.  
It would then be difficult to move the method to another class, make it 
non-static and calling it through a variable. To avoid these obstacles, we 
simply ignore static methods.

\begin{listing}[htb]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{2*\lineheight+5.8pt}

\begin{tikzpicture}[overlay, yscale=-1, xshift=3.8pt+\charwidth*31]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Level 1
  \draw[overlaybox] (\indent,\mintedtop+\lineheight*4) rectangle 
  +(23*\charwidth,17*\lineheight);

  % Level 2
  \draw[overlaybox] (2*\indent,\mintedtop+5*\lineheight) rectangle 
  +(15*\charwidth,3*\lineheight);
  \draw[overlaybox] (2*\indent,\mintedtop+15*\lineheight) rectangle 
  +(15*\charwidth,3*\lineheight);
  \draw[overlaybox] (2*\indent,\mintedtop+19*\lineheight) rectangle 
  +(15*\charwidth,\lineheight);
\end{tikzpicture}
  \begin{multicols}{2}
  \begin{minted}[linenos,frame=topline,label=Clean,framesep=\mintedframesep]{java}
class C {
  A a; B b; boolean bool;

  void method(int val) {
    if (bool) {
      a.foo();
      a = new A();
      a.bar();
    }

    a.foo();
    a.bar();

    switch (val) {
    case 1:
      b.a.foo();
      b.a.bar();
      break;
    default:
      a.foo();
    }
  }
}
\end{minted}

\columnbreak

\begin{minted}[frame=topline,label={With statement 
  sequences},framesep=\mintedframesep]{java}
class C {
  A a; B b; boolean bool;

  void method(int val) {
    if (bool) {
      a.foo();
      a = new A();
      a.bar();
    }

    a.foo();
    a.bar();

    switch (val) {
    case 1:
      b.a.foo();
      b.a.bar();
      break;
    default:
      a.foo();
    }
  }
}
\end{minted}

  \end{multicols}
\caption{Classes \type{A} and \type{B} are both public.  The methods 
\method{foo} and \method{bar} are public members of class \type{A}.}
\label{lst:grandExample}
\end{listing}

\subsection{The possible text selections of a method body}
The number of possible text selections that can be made from the text in a 
method body, are equal to all the sub-sequences of characters within it. For our 
purposes, analyzing program source code, we must define what it means for a text 
selection to be valid.

\definition{A \emph{valid text selection} is a text selection that contains all 
of one or more consecutive program statements.}

For a sequence of statements, the text selections that can be made from it, are 
equal to all its sub-sequences. \Myref{lst:textSelectionsExample} show an 
example of all the text selections that can be made from the code in 
\myref{lst:grandExample}, lines 16-18. For convenience and the clarity of this 
example, the text selections are represented as tuples with the start and end 
line of all selections: $\{(16), (17), (18), (16,17), (16,18), (17,18)\}$.

\begin{listing}[htb]
\def\charwidth{5.7pt}
\def\indent{4*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{\lineheight-1pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]

  % First statement
  \draw[overlaybox] (2*\charwidth,\mintedtop) rectangle 
  +(16*\charwidth,\lineheight);

  % Second statement
  \draw[overlaybox] (2*\charwidth,\mintedtop+\lineheight) rectangle 
  +(16*\charwidth,\lineheight);

  % Third statement
  \draw[overlaybox] (2*\charwidth,\mintedtop+2*\lineheight) rectangle 
  +(16*\charwidth,\lineheight);

  \draw[overlaybox] (\indent-3*\charwidth,\mintedtop) rectangle 
  +(18*\charwidth,2*\lineheight);

  \draw[overlaybox] (3*\charwidth,\mintedtop+\lineheight) rectangle 
  +(14*\charwidth,2*\lineheight);

  % All
  \draw[overlaybox] (\indent,\mintedtop) rectangle 
  +(12*\charwidth,3*\lineheight);
\end{tikzpicture}
% indent should be 5 spaces
\begin{minted}[linenos,firstnumber=16]{java}
     b.a.foo();
     b.a.bar();
     break;
\end{minted}
\caption{Example of how the text selections generator would generate text 
  selections based on a lists of statements. Each highlighted rectangle 
represents a text selection.}
\label{lst:textSelectionsExample}
\end{listing}

Each nesting level of a method body can have many such sequences of statements.  
The outermost nesting level has one such sequence, and each branch contains
its own sequence of statements. \Myref{lst:grandExample} has a version of some 
code where all such sequences of statements are highlighted for a method body.

To complete our example of possible text selections, I will now list all 
possible text selections for the method in \myref{lst:grandExample}, by nesting 
level. There are 23 of them in total.

\begin{description}
  \item[Level 1 (10 selections)] \hfill \\
  $\{(5,9), (11), (12), (14,21), (5,11), (5,12), (5,21), (11,12),
  (11,21), \\(12,21)\}$

  \item[Level 2 (13 selections)] \hfill \\
  $\{(6), (7), (8), (6,7), (6,8), (7,8), (16), (17), (18), (16,17), (16,18), \\
  (17,18), (20)\}$
\end{description}

\subsubsection{The complexity}\label{sec:complexity} 
The complexity of how many text selections that need to be analyzed for a body 
of in total $n$ statements, is bounded by $O(n^2)$. A body of statements is here 
all the statements in all nesting levels of a sequence of statements. A method 
body (or a block) is a body of statements. To prove that the complexity is 
bounded by $O(n^2)$, I present a couple of theorems and prove them.

\begin{theorem}
The number of text selections that need to be analyzed for each list of 
statements of length $n$, is exactly

\begin{equation*}
  \sum_{i=1}^{n} i = \frac{n(n+1)}{2}
  \label{eq:complexityStatementList}
\end{equation*}
\label{thm:numberOfTextSelection}
\end{theorem}

\begin{proof}
  For $n=1$ this is trivial: $\frac{1(1+1)}{2} = \frac{2}{2} = 1$. One statement 
  equals one selection.

  For $n=2$, you get one text selection for the first statement, one selection 
  for the second statement, and one selection for the two of them combined.  
  This equals three selections. $\frac{2(2+1)}{2} = \frac{6}{2} = 3$.

  For $n=3$, you get 3 selections for the two first statements, as in the case 
  where $n=2$. In addition you get one selection for the third statement itself, 
  and two more statements for the combinations of it with the two previous 
  statements. This equals six selections. $\frac{3(3+1)}{2} = \frac{12}{2} = 6$.

  Assume that for $n=k$ there exists $\frac{k(k+1)}{2}$ text selections. Then we 
  want to add selections for another statement, following the previous $k$ 
  statements. So, for $n=k+1$, we get one additional selection for the statement 
  itself. Then we get one selection for each pair of the new selection and the 
  previous $k$ statements. So the total number of selections will be the number 
  of already generated selections, plus $k$ for every pair, plus one for the 
  statement itself: $\frac{k(k+1)}{2} + k + 
  1 = \frac{k(k+1)+2k+2}{2} = \frac{k(k+1)+2(k+1)}{2} = \frac{(k+1)(k+2)}{2} = 
    \frac{(k+1)((k+1)+1)}{2} = \sum_{i=1}^{k+1} i$
\end{proof}

%\definition{A \emph{body of statements} is a sequence of statements where every 
%statement may have sub-statements.}

\begin{theorem}
  The number of text selections for a body of statements is maximized if all the 
  statements are at the same level.
  \label{thm:textSelectionsMaximized}
\end{theorem}

\begin{proof}
 Assume we have a body of, in total, $k$ statements. Then, the sum of the 
 lengths of all the lists of statements in the body, is also $k$. Let 
 $\{l,\ldots,m,(k-l-\ldots-m)\}$ be the lengths of the lists of statements in 
 the body, with $l+\ldots+m<k \Rightarrow \forall i \in \{l,\ldots,m\} : i < k$.

 Then, the number of text selections that are generated for the $k$ statements 
 is 

 {
 \small
 \begin{align*}
   \frac{l(l+1)}{2} + \ldots + \frac{m(m+1)}{2} + 
   \frac{(k-l-\ldots-m)((k-l-\ldots-m)+ 1)}{2} = \\
   \frac{l^2+l}{2} + \ldots + \frac{m^2+m}{2} + \frac{k^2 - 2kl - \ldots - 2km + 
   l^2 + \ldots + m^2 + k - l - \ldots - m}{2} = \\
   \frac{2l^2 - 2kl + \ldots + 2m^2 - 2km + k^2 + k}{2}
 \end{align*}
 }

 \noindent It then remains to show that this inequality holds:

 \begin{align*}
   \frac{2l^2 - 2kl + \ldots + 2m^2 - 2km + k^2 + k}{2} < \frac{k(k+1)}{2} = 
   \frac{k^2 + k}{2}
 \end{align*}

 \noindent By multiplication by $2$ on both sides, and by removing the equal 
 parts, we get

 \begin{align*}
   2l^2 - 2kl + \ldots + 2m^2 - 2km < 0
 \end{align*}

 Since $\forall i \in \{l,\ldots,m\} : i < k$, we have that $\forall i \in 
 \{l,\ldots,m\} : 2ki > 2i^2$, so all the pairs of parts on the form $2i^2-2ki$ 
 are negative. In sum, the inequality holds.

\end{proof}

Therefore, the complexity for the number of selections that need to be analyzed 
for a body of $n$ statements is $O\bigl(\frac{n(n+1)}{2}\bigr) = O(n^2)$.

\section{Disqualifying a selection}
Certain text selections would lead to broken code if used as input to the 
\ExtractAndMoveMethod refactoring. To avoid this, we have to check all text 
selections for such conditions before they are further analyzed. This section
is therefore going to present some properties that make a selection unsuitable 
for our refactoring. When analyzing all these properties, it is assumed that the 
source code does not contain any compilation errors.

\subsection{A call to a protected or package-private method}
If a text selection contains a call to a protected or package-private method, it 
would not be safe to move it to another class. The reason for this, is that we 
cannot know if the called method is being overridden by some subclass of the 
\gloss{enclosingClass}, or not.

Imagine that the protected method \method{foo} is declared in class \m{A}, 
and overridden in class \m{B}. The method \method{foo} is called from within a 
selection done to a method in \m{A}. We want to extract and move this selection 
to another class. The method \method{foo} is not public, so the \MoveMethod 
refactoring must make it public, making the extracted method able to call it 
from the extracted method's new location. The problem is, that the now public
method \method{foo} is overridden in a subclass, where it has a protected 
status.  This makes the compiler complain that the subclass \m{B} is trying to 
reduce the visibility of a method declared in its superclass \m{A}. This is not 
allowed in Java, and for good reasons. It would make it possible to make a 
subclass that could not be a substitute for its superclass.

The problem this check helps to avoid, is a little subtle. The problem does not 
arise in the class where the change is done, but in a class derived from it.  
This shows that classes acting as superclasses are especially fragile to 
introducing errors in the context of automated refactoring.  
\begin{comment}
This is also shown in bug\ldots \todoin{File Eclipse bug report}
\end{comment}

\subsection{A double class instance creation}
The following is a problem caused solely by the underlying \MoveMethod 
refactoring.  The problem occurs if two classes are instantiated such that the 
first constructor invocation is an argument to a second, and that the first 
constructor invocation takes an argument that is built up using a field. As an 
example, say that \var{name} is a field of the enclosing class, and we have the 
expression \code{new A(new B(name))}. If this expression is located in a 
selection that is moved to another class, \var{name} will be left untouched, 
instead of being prefixed with a variable of the same type as it is declared in.  
If \var{name} is the destination for the move, it is not replaced by 
\code{this}, or removed if it is a prefix to a member access 
(\code{name.member}), but it is still left by itself.

Situations like this would lead to code that will not compile. Therefore, we 
have to avoid them by not allowing selections to contain such double class 
instance creations that also contain references to fields.
\begin{comment}
\todoin{File Eclipse bug report}
\end{comment}

\subsection{Instantiation of non-static inner class}
When a non-static inner class is instantiated, this must happen in the scope of 
its declaring class. This is because it must have access to the members of the 
declaring class. If the inner class is public, it is possible to instantiate it 
through an instance of its declaring class, but this is not handled by the 
underlying \MoveMethod refactoring.

Performing a move on a method that instantiates a non-static inner class, will 
break the code if the instantiation is not handled properly. For this reason, 
selections that contain instantiations of non-static inner classes are deemed 
unsuitable for the \ExtractAndMoveMethod refactoring.

\subsection{References to enclosing instances of the enclosing class}
To ``reference an enclosing instance of the enclosing class'' is to reference 
another instance than the one for the immediately enclosing class. Imagine there 
is a (non-static) class \m{C} that is declared in the inner scope of another 
class. That class can again be nested inside a third class, and so on. Hence, 
the nested class \m{C} can have access to many enclosing instances of its 
innermost enclosing class. A selection in a method declared in class \m{C} is 
disqualified if it contains a statement that contains a reference to one or more 
instances of these enclosing classes of \m{C}.

The problem with this, is that these references may not be valid if they are 
moved to another class. Theoretically, some situations could easily be solved by 
passing, to the moved method, a reference to the instance where the problematic 
referenced member is declared. This should work in the case where this member is 
publicly accessible. This is not done in the underlying \MoveMethod refactoring, 
so it cannot be allowed in the \ExtractAndMoveMethod refactoring either.

\subsection{Inconsistent return statements}
To verify that a text selection is consistent with respect to return statements, 
we must check that if a selection contains a return statement, then every 
possible execution path within the selection ends in either a return or a throw 
statement. This property is important regarding the \ExtractMethod refactoring.  
If it holds, it means that a method could be extracted from the selection, and a 
call to it could be substituted for the selection. If the method has a non-void 
return type, then a call to it would also be a valid return point for the 
calling method. If its return value is of the void type, then the \ExtractMethod 
refactoring will append an empty return statement to the back of the method 
call. Therefore, the analysis does not discriminate on either kind of return 
statements, with or without a return value.

A \emph{throw} statement is accepted anywhere a return statement is required.  
This is because a throw statement causes an immediate exit from the current 
block, together with all outer blocks in its control flow that does not catch 
the thrown exception.

We separate between explicit and implicit return statements. An \emph{explicit} 
return statement is formed by using the \code{return} keyword, while an 
\emph{implicit} return statement is a statement that is not formed using 
\code{return}, but must be the last statement of a method that can have any side 
effects. This can happen in methods with a void return type. An example is a 
statement that is inside one or more blocks. The last statement of a method 
could for instance be a synchronized statement, but the last statement that is 
executed in the method, and that can have any side effects, may be located 
inside the body of the synchronized statement.

We can start the check for this property by looking at the last statement of a 
selection to see if it is a return statement (explicit or implicit) or a throw 
statement.  If this is the case, then the property holds, assuming the selected 
code do not contain any compilation errors. All execution paths within the 
selection should end in either this, or another, return or throw statement.

If the last statement of the selection is not a \emph{return} or \emph{throw}, 
the execution of it must eventually end in one of these types of statements for 
the selection to be legal. This means that all branches of the last statement of 
every branch must end in a return or throw.  Given this recursive definition, 
there are only five types of statements that are guaranteed to end in a return 
or throw if their child branches do. All other statements would have to be 
considered illegal. The first three: Block-statements, labeled statements and 
do-statements are all kinds of fall-through statements that always get their 
body executed. Do-statements would not make much sense if written such that they 
always end after the first round of execution of their body, but that is not our 
concern. The remaining two statements that can end in a return or throw are 
if-statements and try-statements.

For an if-statement, the rule is that if its then-part does not contain any 
return or throw statements, this is considered illegal. If the then-part does 
contain a return or throw, the else-part is checked. If its else-part is 
non-existent, or it does not contain any return or throw statements, the 
statement is considered illegal. If an if-statement is not considered illegal, 
the bodies of its two parts must be checked. 

Try-statements are handled much the same way as if-statements. The body of a 
try-statement must contain a return or throw. The same applies to its catch 
clauses and finally body.  \todoin{finally body?}

\subsection{Ambiguous return values}
The problem with ambiguous return values arises when a selection is chosen to be 
extracted into a new method, but if refactored it needs to return more than one 
value from that method.

This problem occurs in two situations. The first situation arises when there is 
more than one local variable that is both assigned to within a selection and 
also referenced after the selection. The other situation occurs when there is 
only one such assignment, but the selection also contain return statements.

Therefore we must examine the selection for assignments to local variables that 
are referenced after the text selection. Then we must verify that not more than 
one such reference is done, or zero if any return statements are found.

\subsection{Illegal statements}
An illegal statement may be a statement that is of a type that is never allowed, 
or it may be a statement of a type that is only allowed if certain conditions 
are true.

Any use of the \var{super} keyword is prohibited, since its meaning is altered 
when moving a method to another class.

For a \emph{break} statement, there are two situations to consider: A break 
statement with or without a label. If the break statement has a label, it is 
checked that whole of the labeled statement is inside the selection. If the 
break statement does not have a label attached to it, it is checked that its 
innermost enclosing loop or switch statement also is inside the selection.

The situation for a \emph{continue} statement is the same as for a break 
statement, except that it is not allowed inside switch statements.

Regarding \emph{assignments}, two types of assignments are allowed: Assignments 
to non-final variables and assignments to array access. All other assignments 
are regarded illegal.

\paragraph{Incompleteness.}\label{par:incompleteness} The list of illegal 
statements is not complete, and a lot of situations that can lead to compilation 
errors or behavior changes are not considered. It is not feasible to consider 
all such situations within the limits of this master's project, and maybe not 
outside of them either. The feasibility of this problem could be explored 
further by others.

\section{Disqualifying selections from the 
example}\label{sec:disqualifyingExample}
Among the selections we found for the code in \myref{lst:grandExample}, not many 
of them must be disqualified on the basis of containing something illegal. The 
only statement causing trouble is the break statement in line 18. None of the 
selections on nesting level 2 can contain this break statement, since the 
innermost switch statement is not inside any of these selections.

This means that the text selections $(18)$, $(16,18)$ and $(17,18)$ can be 
excluded from further consideration, and we are left with the following 
selections.

\begin{description}
  \item[Level 1 (10 selections)] \hfill \\
  $\{(5,9), (11), (12), (14,21), (5,11), (5,12), (5,21), (11,12),
  (11,21), \\(12,21)\}$

  \item[Level 2 (10 selections)] \hfill \\
  $\{(6), (7), (8), (6,7), (6,8), (7,8), (16), (17), (16,17), (20)\}$
\end{description}

\section{Finding a move target}
In the analysis needed to perform the \ExtractAndMoveMethod refactoring 
automatically, the selection we choose is found among all the selections that 
have a possible move target. Therefore, the best possible move target must be 
found for all the candidate selections, so that we are able to sort out the 
selection that is best suited for the refactoring.

To find the best move target for a specific text selection, we first need to 
find all the possible targets. Since the target must be a local variable or a 
field, we are basically looking for names within the selection; names that 
represents references to variables.

The names we are looking for, we call prefixes. This is because we are not 
interested in names that occur in the middle of a dot-separated sequence of 
names. We are only interested in names constituting prefixes of other names, and 
possibly themselves. The reason for this, is that two lexically equal names need 
not be referencing the same variable, if they themselves are not referenced via 
the same prefix. Consider the two method calls \code{a.x.foo()} and 
\code{b.x.foo()}.  Here, the two references to \code{x}, in the middle of the 
qualified names both preceding \code{foo()}, are not referencing the same 
variable.  Even though the variables may share the type, and the method 
\method{foo} thus is the same for both, we would not know through which of the 
variables \var{a} or \var{b} we should call the extracted method.

The possible move targets are then the prefixes that are not among a subset of 
the prefixes that are not valid move targets \see{sec:unfixes}. Also, prefixes 
that are just simple names, and have only one occurrence, are left out. This is 
because they are not going to have any positive effect on coupling between 
classes, and are only going to increase the complexity of the code.

For finding the best move target among these safe prefixes, a simple heuristic 
is used. It is as simple as choosing the prefix that is most frequently 
referenced within the selection. 

\section{Unfixes}\label{sec:unfixes}
We will call the prefixes that are not valid as move targets for unfixes.

A name that is assigned to within a selection can be an unfix. The reason for 
this is that the result would be an assignment to the \type{this} keyword, which 
is not valid in Java \see{eclipse_bug_420726}.

Prefixes that originate from variable declarations within the same selection are 
also considered unfixes. The reason for this is that when a method is moved, it 
needs to be called through a variable. If this variable is also declared within 
the method that is to be moved, this obviously cannot be done.

Also considered as unfixes are variable references that are of types that are 
not suitable for moving methods to. This can either be because it is not 
physically possible to move a method to the desired class or that it will cause 
compilation errors by doing so.

If the type binding for a name is not resolved it is considered an unfix. The 
same applies to types that are only found in compiled code, so they have no 
underlying source that is accessible to us. (E.g. the \type{java.lang.String} 
class.)

Interface types are not suitable as targets. This is simply because interfaces 
in Java cannot contain methods with bodies. (This thesis does not deal with 
features of Java versions later than Java 7. Java 8 has interfaces with default 
implementations of methods.)

Neither are local types allowed. This accounts for both local and anonymous 
classes. Anonymous classes are effectively the same as interface types with 
respect to unfixes. Local classes could in theory be used as targets, but this 
is not possible due to limitations of the way the \refa{Extract and Move Method} 
refactoring has to be implemented. The problem is that the refactoring is done 
in two steps, so the intermediate state between the two refactorings would not 
be legal Java code. In the intermediate step for the case where a local class is 
the move target, the extracted method would need to take the local class as a 
parameter. This new method would need to live in the scope of the declaring 
class of the originating method. The local class would then not be in the scope 
of the extracted method, thus bringing the source code into an illegal state.  
This scenario is shown in \myref{lst:extractMethodLocalClass}. One could imagine 
that the method was extracted and moved in one operation, without an 
intermediate state.  Then it would make sense to include variables with types of 
local classes in the set of legal targets, since the local classes would then be 
in the scopes of the method calls. If this makes any difference for software 
metrics that measure coupling would be a different discussion.

\begin{listing}[htb]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\rightColX{32*\charwidth-1pt}
\def\lineheight{\baselineskip}
\def\mintedtop{2*\lineheight+5.8pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Before
  \draw[overlaybox] (0,\mintedtop+\lineheight*8) rectangle 
  +(27*\charwidth,2*\lineheight);

  % After
  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*8) rectangle 
  +(27*\charwidth,\lineheight);

  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*12) rectangle 
  +(30*\charwidth,4*\lineheight);

\end{tikzpicture}
\begin{multicols}{2}
\begin{minted}[frame=topline,label=Before,framesep=\mintedframesep]{java}
void declaresLocalClass() {
  class LocalClass {
    void foo() {}
    void bar() {}
  }

  LocalClass inst =
    new LocalClass();
  inst.foo();
  inst.bar();
}
\end{minted}

\columnbreak

\begin{minted}[frame=topline,label={After Extract 
  Method},framesep=\mintedframesep]{java}
void declaresLocalClass() {
  class LocalClass {
    void foo() {}
    void bar() {}
  }

  LocalClass inst =
    new LocalClass();
  fooBar(inst);
}

// Illegal intermediate step
void fooBar(LocalClass inst) {
  inst.foo();
  inst.bar();
}
\end{minted}
\end{multicols}
\caption{The \refa{Extract and Move Method} refactoring bringing the code into 
an illegal state with an intermediate step.}
\label{lst:extractMethodLocalClass}
\end{listing}

The last class of names that are considered unfixes are names used in null 
tests. These are tests that read like this: if \code{<name>} equals \var{null} 
then do something. If allowing variables used in those kinds of expressions as 
targets for moving methods, we would end up with code containing boolean 
expressions like \code{this == null}, which would always evaluate to 
\code{false}, since \var{this} would never be \var{null}. The existence of a 
null test indicates that a variable is expected to sometimes hold the value 
\var{null}.  By using a variable used in a null test as a move target, we could 
potentially end up with a
null pointer exception if the method is called on a variable with a null value.

\section{Finding the example selections that have possible targets}
We now pick up the thread from \myref{sec:disqualifyingExample} where we have a 
set of text selections that need to be analyzed to find out if some of them are 
suitable targets for the \ExtractAndMoveMethod refactoring.

We start by analyzing the text selections for nesting level 2, because these 
results can be used to reason about the selections for nesting level 1. First we 
have all the single-statement selections.

\begin{description}
  \item[Selections $(6)$, $(8)$ and $(20)$.] \hfill \\
    All these selections have a prefix that contains a possible target, namely 
    the variable \var{a}. The problem is that the prefixes are only one segment 
    long, and their frequency counts are only 1 as well. None of these 
    selections are therefore considered as suitable candidates for the 
    refactoring.

  \item[Selection $(7)$.] \hfill \\
    This selection contains the unfix \var{a}, and no other possible targets.  
    The reason for \var{a} being an unfix is that it is assigned to within the 
    selection. Selection $(7)$ is therefore unsuited as a refactoring candidate.

  \item[Selections $(16)$ and $(17)$.] \hfill \\
    These selections both have a possible target. The target for both selections 
    is the variable \var{b}. Both the prefixes have frequency 1. We denote this 
    with the new tuples $((16), \texttt{b.a}, f(1))$ and $((17), \texttt{b.a}, 
    f(1))$. They contain the selection, the prefix with the target and the 
    frequency for this prefix.

\end{description}

Then we have all the text selections from level 2 that are composed of multiple 
statements:

\begin{description}
  \item[Selections $(6,7)$, $(6,8)$ and $(7,8)$.] \hfill \\
    All these selections are disqualified for the reason that they contain the 
    unfix \var{a}, due to the assignment, and no other possible move targets.

  \item[Selection $(16,17)$.] \hfill \\
    This selection is the last selection that is not analyzed on nesting level 
    2. It contains only one possible move target, and that is the variable   
       \var{b}. It also contains only one prefix \var{b.a}, with frequency count 
    2. Therefore we have a new candidate $((16,17), \texttt{b.a}, f(2))$.

\end{description}

Moving on to the text selections for nesting level 1, starting with the 
single-statement selections:

\begin{description}
  \item[Selection $(5,9)$.] \hfill \\
    This selection contains two variable references that must be analyzed to see 
    if they are possible move candidates. The first one is the variable 
    \var{bool}. This variable is of type \type{boolean}, which is a primary type 
    and therefore not possible to make any changes to. The second variable is 
    \var{a}. The variable \var{a} is an unfix in $(5,9)$, for the same reason as 
    in the selections $(6,7)$, $(7,8)$ and $(6,8)$. So selection $(5,9)$ 
    contains no possible move targets.

  \item[Selections $(11)$ and $(12)$.] \hfill \\
    These selections are disqualified for the same reasons as selections $(6)$ 
    and $(8)$. Their prefixes are one segment long and are referenced only one 
    time.

  \item[Selection $(14,21)$] \hfill \\
    This is the switch statement from \myref{lst:grandExample}. It contains the 
    relevant variable references \var{val}, \var{a} and \var{b}. The variable 
    \var{val} is a primary type, just as \var{bool}. The variable \var{a} is 
    only found in one statement, and in a prefix with only one segment, so it is 
    not considered to be a possible move target. The only variable left is 
    \var{b}.  Just as in the selection $(16,17)$, \var{b} is part of the prefix 
    \code{b.a}, which has 2 appearances. We have found a new candidate 
    $((14,21), \texttt{b.a}, f(2))$.
    
\end{description}

It remains to see if we get any new candidates by analyzing the multi-statement 
text selections for nesting level 1:

\begin{description}
  \item[Selections $(5,11)$ and $(5,12)$.] \hfill \\
    These selections are disqualified for the same reason as $(5,9)$. The only 
    possible move target \var{a} is an unfix.

  \item[Selection $(5,21)$.] \hfill \\
    This is whole of the method body in \myref{lst:grandExample}. The variables 
    \var{a}, \var{bool} and \var{val} are either an unfix or primary types. The 
    variable \var{b} is the only possible move target, and we have, again, the 
    prefix \code{b.a} as the center of attention. The refactoring candidate is 
    $((5,21), \texttt{b.a}, f(2))$.

  \item[Selection $(11,12)$.] \hfill \\
    This small selection contains the prefix \code{a} with frequency 2, and no 
    unfixes. The candidate is $((11,12), \texttt{a}, f(2))$.

  \item[Selection $(11,21)$] \hfill \\
      This selection contains the selection $(11,12)$ in addition to the switch 
      statement. The selection has two possible move targets. The first one is 
      \var{b}, in a prefix with frequency 2. The second is \var{a}, although it 
      is in a simple prefix, it is referenced 3 times, and is therefore chosen
      as the target for the selection. The new candidate is $((11,21), 
      \texttt{a}, f(3))$.

    \item[Selection $(12,21)$.] \hfill \\
      This selection is similar to the previous $(11,21)$, only that \var{a} now 
      has frequency count 2. This means that the prefixes \code{a} and 
      \code{b.a} both have the count 2. Of the two, \code{b.a} is preferred, 
      since it has more segments than \code{a}. Thus the candidate for this 
      selection is $((12,21), \texttt{b.a}, f(2))$.

\end{description}

For the method in \myref{lst:grandExample} we therefore have the following 8 
candidates for the \ExtractAndMoveMethod refactoring: $\{((16), \texttt{b.a}, 
f(1)), ((17), \texttt{b.a}, f(1)), ((16,17), \texttt{b.a}, f(2)), ((14,21), 
\texttt{b.a}, f(2)), \\ ((5,21), \texttt{b.a}, f(2)), ((11,12), \texttt{a}, 
f(2)), ((11,21), \texttt{a}, f(3)), ((12,21), \texttt{b.a}, f(2))\}$.

It now only remains to specify an order for these candidates, so we can choose 
the most suitable one to refactor.


\section{Choosing the selection}\label{sec:choosingSelection}
When choosing a selection between the text selections that have possible move 
targets, the selections need to be ordered. The criteria below are presented in 
the order they are prioritized. If not one selection is favored over the other 
for a concrete criterion, the selections are evaluated by the next criterion.

\begin{enumerate}
  \item The first criterion that is evaluated is whether a selection contains 
    any unfixes or not. If selection \m{A} contains no unfixes, while selection 
    \m{B} does, selection \m{A} is favored over selection \m{B}. This is 
    because, if we can, we want to avoid moving such as assignments and variable 
    declarations. This is done under the assumption that, if possible, avoiding 
    selections containing unfixes will make the code moved a little cleaner.

  \item The second criterion that is evaluated is whether a selection contains 
    multiple possible targets or not. If selection \m{A} has only one possible 
    target, and selection \m{B} has multiple, selection \m{A} is favored. If 
    both selections have multiple possible targets, they are considered equal 
    with respect to this criterion. The rationale for this heuristic is that we 
    would prefer not to introduce new couplings between classes when performing 
    the \ExtractAndMoveMethod refactoring. 

  \item When evaluating this criterion, this is with the knowledge that
    selection \m{A} and \m{B} both have one possible target, or multiple 
    possible targets. Then, if the move target candidate of selection \m{A} has 
    a higher reference count than the target candidate of selection \m{B}, 
    selection \m{A} is favored.  The reason for this is that we would like to 
    move the selection that gets rid of the most references to another class. 

  \item The last criterion is that if the frequencies of the targets chosen for 
    both selections are equal, the selection with the target that is part of the 
    prefix with highest number of segments is favored. This is done to favor 
    indirection.

\end{enumerate}

If none of the above mentioned criteria favor one selection over another, the 
selections are considered to be equally good candidates for the 
\ExtractAndMoveMethod refactoring.

\section{Performing changes}
When a text selection and a move target is found for the \ExtractAndMoveMethod 
refactoring, the actual changes are executed by two existing primitive 
refactorings.  First the \ExtractMethod refactoring is used to extract the 
selection into a new method. Then the \MoveMethod refactoring is used to move 
that new method to the class determined by the move target.

If, at any point, an exception is thrown or the preconditions for one of the 
primitive refactorings are not satisfied, the composite refactoring is aborted, 
and the source code is left in its current state. This has the implication that 
the \ExtractAndMoveMethod refactoring could end up being partially executed.  
This happens if the \ExtractMethod refactoring is executed, but the \MoveMethod 
refactoring is being canceled. A partial execution is not considered a problem, 
since the code should still compile.

\section{Concluding the example}
For choosing one of the remaining selections, we need to order our candidates 
after the criteria in the previous section. Below we have the candidates ordered 
in descending order, with the ``best'' candidate first:

\begin{multicols}{2}
\begin{enumerate}
  \item $((16,17), \texttt{b.a}, f(2))$
  \item $((11,12), \texttt{a}, f(2))$
  \item $((16), \texttt{b.a}, f(1))$
  \item $((17), \texttt{b.a}, f(1))$

  % With unfixes:
  % Many possible targets
  \item $((11,21), \texttt{a}, f(3))$
  \item $((5,21), \texttt{b.a}, f(2))$
  \item $((12,21), \texttt{b.a}, f(2))$
  \item $((14,21), \texttt{b.a}, f(2))$

\end{enumerate}
\end{multicols}

\begin{comment}
5       if (bool) {
6           a.foo();
7           a = new A();
8           a.bar();
9       }
10
11      a.foo();
12      a.bar();
13
14      switch (val) {
15      case 1:
16          b.a.foo();
17          b.a.bar();
18          break;
19      default:
20          a.foo();
21      }
\end{comment}

The candidates ordered 5-8 all have unfixes in them, therefore they are ordered 
last. None of the candidates ordered 1-4 have multiple possible targets. The 
only interesting discussion is now why $(16,17)$ was favored over $(11,12)$.  
This is because the prefix \code{b.a} enclosing the move target of selection 
$(16,17)$ has one more segment than the prefix \code{a} of $(11,12)$.

The selection is now extracted into a new method \method{gen\_123} and then 
moved to class \type{B}, since that is the type of the variable \var{b} that is 
the target from the chosen refactoring candidate. The name of the method has a 
randomly generated suffix. In the refactoring implementation, the extracted 
methods follow the pattern \code{generated\_<long>}, where \code{<long>} is a 
pseudo-random long value. This is shortened here to make the example readable. 
The result after the refactoring is shown in \myref{lst:grandExampleResult}.

\begin{listing}[htb]
\def\charwidth{5.8pt}
\def\indent{2*\charwidth}
\def\rightColX{32*\charwidth-1pt}
\def\lineheight{\baselineskip}
\def\mintedtop{1*\lineheight+1.8pt}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]
  % Before
  \draw[overlaybox] (2*\indent,\mintedtop+\lineheight*15) rectangle 
  +(18*\charwidth,\lineheight);

  % After
  \draw[overlaybox] (\rightColX,\mintedtop+\lineheight*3) rectangle 
  +(29*\charwidth,4*\lineheight);

\end{tikzpicture}
  \begin{multicols}{2}
    \begin{minted}[linenos]{java}
class C {
  A a; B b; boolean bool;

  void method(int val) {
    if (bool) {
      a.foo();
      a = new A();
      a.bar();
    }

    a.foo();
    a.bar();

    switch (val) {
    case 1:
      b.gen_123(this);
      break;
    default:
      a.foo();
    }
  }
}
\end{minted}

\columnbreak

  \begin{minted}[]{java}
public class B {
  A a;

  public void gen_123(C c) {
    a.foo();
    a.bar();
  }
}
\end{minted}

  \end{multicols}
  \caption{The result after refactoring the class \type{C} of 
  \myref{lst:grandExample} with the \ExtractAndMoveMethod refactoring with 
  $((16,17), \texttt{b.a}, f(2))$ as input.}
\label{lst:grandExampleResult}
\end{listing}

\paragraph{Implementation details.} Implementation details for the various parts 
of this chapter are found in \myref{ch:implementation}.


\chapter{The Eclipse Platform with the Java development tools}\label{ch:eclipse}
The Eclipse Platform is an extensible platform. It can be used to build IDEs for 
many programming languages. For it to be a fully functional Java IDE, it must be 
equipped with the Java development tools plugin, abbreviated as JDT.

This chapter will present how to analyze and change Java source code by 
utilizing the APIs supplied by Eclipse and the JDT.

\section{Analyzing source code in Eclipse}
In this section we will see how to access Java source code in the Eclipse 
workspace. Then it is shown how this code is being represented when it is parsed 
and how to search this representation for the properties we are after.

\subsection{The Java model}\label{javaModel}
The Java model of \name{Eclipse} is its internal representation of a Java project. It 
is light-weight, and has only limited possibilities for manipulating source 
code. It is typically used as a basis for the Package Explorer in \name{Eclipse}.

The elements of the Java model are only handles to the underlying elements. This 
means that the underlying element of a handle does not need to actually exist.  
Hence the user of a handle must always check that it exist by calling the 
\method{exists} method of the handle.

The handles with descriptions are listed in \myref{tab:javaModel}, while the 
hierarchy of the Java Model is shown in \myref{fig:javaModel}.

\begin{table}[htb]
  \caption{The elements of the Java Model\citing{vogelEclipseJDT2012}.}
  \label{tab:javaModel}
  \centering
  % sum must equal number of columns (3)
  \begin{tabularx}{\textwidth}{@{} L{0.7}  L{1.1}  L{1.2} @{}}
    \toprule
    \textbf{Project Element} & \textbf{Java Model element} & 
    \textbf{Description} \\
    \midrule
    Java project & \type{IJavaProject} & The Java project which contains all other objects. \\
    \midrule
    Source folder /\linebreak[2] binary folder /\linebreak[3] external library & 
    \type{IPackageFragmentRoot} & Hold source or binary files, can be a folder 
    or a library (zip / jar file). \\
    \midrule
    Each package & \type{IPackageFragment} & Each package is below the 
    \type{IPackageFragmentRoot}, sub-packages are not leaves of the package, 
    they are listed directed under \type{IPackageFragmentRoot}. \\
    \midrule
    Java Source file & \type{ICompilationUnit} & The Source file is always below 
    the package node. \\
    \midrule
    Types / Fields /\linebreak[3] Methods & \type{IType} / \type{IField} 
    /\linebreak[3] \type{IMethod} & Types, fields and methods. \\
    \bottomrule
  \end{tabularx}
\end{table}


\begin{figure}[h]
  \centering
  \begin{tikzpicture}[%
  grow via three points={one child at (0,-0.7) and
  two children at (0,-0.7) and (0,-1.4)},
  edge from parent path={(\tikzparentnode.south west)+(0.5,0) |- 
  (\tikzchildnode.west)}]
  \tikzstyle{every node}=[draw=black,thick,anchor=west]
  \tikzstyle{selected}=[draw=red,fill=red!30]
  \tikzstyle{optional}=[dashed,fill=gray!50]
  \node {\type{IJavaProject}}
    child { node {\type{IPackageFragmentRoot}}
      child { node {\type{IPackageFragment}}
        child { node {\type{ICompilationUnit}}
          child { node {\type{IType}}
            child { node {\type{\{ IType \}*}}
              child { node {\type{\ldots}}}
            }
            child [missing] {}
            child { node {\type{\{ IField \}*}}}
            child { node {\type{IMethod}}
              child { node {\type{\{ IType \}*}}
                child { node {\type{\ldots}}}
              }
            }
            child [missing] {}
            child [missing] {}
            child { node {\type{\{ IMethod \}*}}}
          }
          child [missing] {}
          child [missing] {}
          child [missing] {}
          child [missing] {}
          child [missing] {}
          child [missing] {}
          child [missing] {}
          child { node {\type{\{ IType \}*}}}
        }
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child [missing] {}
        child { node {\type{\{ ICompilationUnit \}*}}}
      }
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child [missing] {}
      child { node {\type{\{ IPackageFragment \}*}}}
    }
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child [missing] {}
    child { node {\type{\{ IPackageFragmentRoot \}*}}}
    ;
  \end{tikzpicture}
  \caption{The Java model of \name{Eclipse}. ``\type{\{ SomeElement \}*}'' means 
  ``\type{SomeElement} zero or more times``. For recursive structures, 
  ``\type{\ldots}'' is used.}
  \label{fig:javaModel}
\end{figure}

\subsection{The abstract syntax tree}
\name{Eclipse} is following the common paradigm of using an abstract syntax tree for 
source code analysis and manipulation.

When parsing program source code into something that can be used as a foundation 
for analysis, the start of the process follows the same steps as in a compiler.  
This is all natural, because the way a compiler analyzes code is no different 
from how source manipulation programs would do it, except for some properties of 
code that is analyzed in the parser, and that they may be differing in what 
kinds of properties they analyze.  Thus the process of translation source code 
into a structure that is suitable for analyzing, can be seen as a kind of 
interrupted compilation process \see{fig:interruptedCompilationProcess}.

\begin{figure}[h]
  \centering
  \tikzset{
    base/.style={anchor=north, align=center, rectangle, minimum height=1.4cm},
    basewithshadow/.style={base, drop shadow, fill=white},
    outlined/.style={basewithshadow, draw, rounded corners, minimum 
    width=0.4cm},
    primary/.style={outlined, font=\bfseries},
    dashedbox/.style={outlined, dashed},
    arrowpath/.style={black, align=center, font=\small},
    processarrow/.style={arrowpath, ->, >=angle 90, shorten >=1pt},
  }
  \begin{tikzpicture}[node distance=1.3cm and 3cm, scale=1, every 
    node/.style={transform shape}]
    \node[base](AuxNode1){\small source code};
    \node[primary, right=of AuxNode1, xshift=-2.5cm](Scanner){Scanner};
    \node[primary, right=of Scanner, xshift=0.5cm](Parser){Parser};
    \node[dashedbox, below=of Parser](SemanticAnalyzer){Semantic\\Analyzer};
    \node[dashedbox, left=of SemanticAnalyzer](SourceCodeOptimizer){Source 
    Code\\Optimizer};
    \node[dashedbox, below=of SourceCodeOptimizer
    ](CodeGenerator){Code\\Generator};
    \node[dashedbox, right=of CodeGenerator](TargetCodeOptimizer){Target 
    Code\\Optimizer};
    \node[base, right=of TargetCodeOptimizer](AuxNode2){};

    \draw[processarrow](AuxNode1) -- (Scanner);

    \path[arrowpath] (Scanner) -- node [sloped](tokens){tokens}(Parser);
    \draw[processarrow](Scanner) -- (tokens) -- (Parser);

    \path[arrowpath] (Parser) -- node (syntax){syntax 
    tree}(SemanticAnalyzer);
    \draw[processarrow](Parser) -- (syntax) -- (SemanticAnalyzer);

    \path[arrowpath] (SemanticAnalyzer) -- node 
    [sloped](annotated){annotated\\tree}(SourceCodeOptimizer);
    \draw[processarrow, dashed](SemanticAnalyzer) -- (annotated) -- 
    (SourceCodeOptimizer);

    \path[arrowpath] (SourceCodeOptimizer) -- node 
    (intermediate){intermediate code}(CodeGenerator);
    \draw[processarrow, dashed](SourceCodeOptimizer) -- (intermediate) --
    (CodeGenerator);

    \path[arrowpath] (CodeGenerator) -- node [sloped](target1){target 
    code}(TargetCodeOptimizer);
    \draw[processarrow, dashed](CodeGenerator) -- (target1) --
    (TargetCodeOptimizer);

    \path[arrowpath](TargetCodeOptimizer) -- node [sloped](target2){target 
    code}(AuxNode2);
    \draw[processarrow, dashed](TargetCodeOptimizer) -- (target2) (AuxNode2);
  \end{tikzpicture}
  \caption{Interrupted compilation process. {\footnotesize (Full compilation 
    process borrowed from \emph{Compiler construction: principles and practice} 
    by Kenneth C.  Louden\citing{louden1997}.)}}
  \label{fig:interruptedCompilationProcess}
\end{figure}

The process starts with a \emph{scanner}, or lexer. The job of the scanner is to 
read the source code and divide it into tokens for the parser. Therefore, it is 
also sometimes called a tokenizer. A token is a logical unit, defined in the 
language specification, consisting of one or more consecutive characters.  In 
the Java language the tokens can for instance be the \var{this} keyword, a curly 
bracket \var{\{} or a \var{nameToken}. It is recognized by the scanner on the 
basis of something equivalent of a regular expression. This part of the process 
is often implemented with the use of a finite automata. In fact, it is common to 
specify the tokens in regular expressions, which in turn are translated into a 
finite automata lexer. This process can be automated.

The program component used to translate a stream of tokens into something 
meaningful, is called a parser. A parser is fed tokens from the scanner and 
performs an analysis of the structure of a program. It verifies that the syntax 
is correct according to the grammar rules of a language, that are usually 
specified in a context-free grammar, and often in a variant of the 
\name{Backus--Naur Form}. The result coming from the parser is in the form of an 
\emph{Abstract Syntax Tree}, AST for short. It is called \emph{abstract}, 
because the structure does not contain all of the tokens produced by the 
scanner. It only contains logical constructs, and because it forms a tree, all 
kinds of parentheses and brackets are implicit in the structure. It is this AST 
that is used when performing the semantic analysis of the code.

As an example we can think of the expression \code{(5 + 7) * 2}. The root of 
this tree would in \name{Eclipse} be an \type{InfixExpression} with the operator
\var{TIMES}, and a left operand, which is also an \type{InfixExpression} with 
the operator \var{PLUS}. The left operand \type{InfixExpression}, has in turn a 
left operand of type \type{NumberLiteral} with the value \var{``5''} and a right 
operand \type{NumberLiteral} with the value \var{``7''}.  The root will have a 
right operand of type \type{NumberLiteral} and value \var{``2''}. The AST for 
this expression is illustrated in \myref{fig:astInfixExpression}.

Contrary to the Java Model, an abstract syntax tree is a heavy-weight 
representation of source code. It contains information about properties like 
type bindings for variables and variable bindings for names. 


\begin{figure}[h]
  \centering
  \begin{tikzpicture}[scale=0.8]
  \tikzset{level distance=40pt}
  \tikzset{sibling distance=5pt}
  \tikzstyle{thescale}=[scale=0.8]
  \tikzset{every tree node/.style={align=center}}
  \tikzset{edge from parent/.append style={thick}}
  \tikzstyle{inode}=[rectangle,rounded corners,draw,fill=lightgray,drop 
  shadow,align=center]
  \tikzset{every internal node/.style={inode}}
  \tikzset{every leaf node/.style={draw=none,fill=none}}

  \Tree [.\type{InfixExpression} [.\type{InfixExpression}
    [.\type{NumberLiteral} \var{``5''} ]  [.\type{Operator} \var{PLUS} ] 
    [.\type{NumberLiteral} \var{``7''} ] ]
  [.\type{Operator} \var{TIMES} ]
    [.\type{NumberLiteral} \var{``2''} ] 
  ]
  \end{tikzpicture}
  \caption{The abstract syntax tree for the expression \code{(5 + 7) * 2}.}
  \label{fig:astInfixExpression}
\end{figure}

\subsubsection{The AST in Eclipse}\label{astEclipse}
In \name{Eclipse}, every node in the AST is a child of the abstract superclass 
\typewithref{org.eclipse.jdt.core.dom}{ASTNode}. Every \type{ASTNode}, among a 
lot of other things, provides information about its position and length in the 
source code, as well as a reference to its parent and to the root of the tree.

The root of the AST is always of type \type{CompilationUnit}. It is not the same 
as an instance of an \type{ICompilationUnit}, which is the compilation unit 
handle of the Java model. The children of a \type{CompilationUnit} is an 
optional \type{PackageDeclaration}, zero or more nodes of type 
\type{ImportDecaration} and all its top-level type declarations that has node 
types \type{AbstractTypeDeclaration}.

An \type{AbstractType\-Declaration} can be one of the types 
\type{AnnotationType\-Declaration}, \type{Enum\-Declaration} or 
\type{Type\-Declaration}. The children of an \type{AbstractType\-Declaration} 
must be a subtype of a \type{BodyDeclaration}. These subtypes are: 
\type{AnnotationTypeMember\-Declaration}, \type{EnumConstant\-Declaration}, 
\type{Field\-Declaration}, \type{Initializer} and \type{Method\-Declaration}.

Of the body declarations, the \type{Method\-Declaration} is the most interesting 
one. Its children include lists of modifiers, type parameters, parameters and 
exceptions. It has a return type node and a body node. The body, if present, is 
of type \type{Block}. A \type{Block} is itself a \type{Statement}, and its 
children is a list of \type{Statement} nodes.

There are too many types of the abstract type \type{Statement} to list up, but 
there exists a subtype of \type{Statement} for every statement type of Java, as 
one would expect. This also applies to the abstract type \type{Expression}.  
However, the expression \type{Name} is a little special, since it is both used 
as an operand in compound expressions, as well as for names in type declarations 
and such.

There is an overview of some of the structure of an \name{Eclipse} AST in 
\myref{fig:astEclipse}.

\begin{figure}[h]
  \centering
  \begin{tikzpicture}[scale=0.8]
  \tikzset{level distance=50pt}
  \tikzset{sibling distance=5pt}
  \tikzstyle{thescale}=[scale=0.8]
  \tikzset{every tree node/.style={align=center}}
  \tikzset{edge from parent/.append style={thick}}
  \tikzstyle{inode}=[rectangle,rounded corners,draw,fill=lightgray,drop 
  shadow,align=center]
  \tikzset{every internal node/.style={inode}}
  \tikzset{every leaf node/.style={draw=none,fill=none}}

  \Tree [.\type{CompilationUnit} [.\type{[ PackageDeclaration ]} [.\type{Name} ] 
  [.\type{\{ Annotation \}*} ] ]
  [.\type{\{ ImportDeclaration \}*} [.\type{Name} ] ]
    [.\type{\{ AbstractTypeDeclaration \}+} [.\node(site){\type{\{ 
    BodyDeclaration \}*}}; ] [.\type{SimpleName} ] ]
  ]
  \begin{scope}[shift={(0.5,-6)}]
    \node[inode,thescale](root){\type{MethodDeclaration}};
    \node[inode,thescale](modifiers) at (4.5,-5){\type{\{ IExtendedModifier \}*} 
    \\ {\footnotesize (Of type \type{Modifier} or \type{Annotation})}};
    \node[inode,thescale](typeParameters) at (-6,-3.5){\type{\{ TypeParameter 
    \}*}};
    \node[inode,thescale](parameters) at (-5,-5){\type{\{ 
    SingleVariableDeclaration \}*} \\ {\footnotesize (Parameters)}};
    \node[inode,thescale](exceptions) at (5,-3){\type{\{ Name \}*} \\ 
    {\footnotesize (Exceptions)}};
    \node[inode,thescale](return) at (-6.5,-2){\type{Type} \\ {\footnotesize 
    (Return type)}};
    \begin{scope}[shift={(0,-5)}]
      \Tree [.\node(body){\type{[ Block ]} \\ {\footnotesize (Body)}};
      [.\type{\{ Statement \}*} [.\type{\{ Expression \}*} ]
        [.\type{\{ Statement \}*} [.\type{\ldots} ]]
      ]
      ]
    \end{scope}
  \end{scope}
  \draw[->,>=triangle 90,shorten >=1pt](root.east)..controls +(east:2) and 
  +(south:1)..(site.south);

  \draw (root.south) -- (modifiers);
  \draw (root.south) -- (typeParameters);
  \draw (root.south) -- ($ (parameters.north) + (2,0) $);
  \draw (root.south) -- (exceptions);
  \draw (root.south) -- (return);
  \draw (root.south) -- (body);

  \end{tikzpicture}
  \caption{The format of the abstract syntax tree in \name{Eclipse}.}
  \label{fig:astEclipse}
\end{figure}

\subsection{The ASTVisitor}\label{astVisitor}
So far, the only thing that has been addressed is how the data that is going to 
be the basis for our analysis is structured. Another aspect of it is how we are 
going to traverse the AST to gather the information we need, so we can conclude 
about the properties we are analyzing. It is of course possible to start at the 
top of the tree, and manually search through its nodes for the ones we are 
looking for, but that is a bit inconvenient. To be able to efficiently utilize 
such an approach, we would need to make our own framework for traversing the 
tree and visiting only the types of nodes we are after. Luckily, this 
functionality is already provided in \name{Eclipse}, by its 
\typewithref{org.eclipse.jdt.core.dom}{ASTVisitor}.

The \name{Eclipse} AST, together with its \type{ASTVisitor}, follows the 
\pattern{Visitor} pattern\citing{designPatterns}. The intent of this design 
pattern is to facilitate extending the functionality of classes without touching 
the classes themselves.

Let us say that there is a class hierarchy of elements. These elements all have 
a method \method{accept(Visitor visitor)}. In its simplest form, the 
\method{accept} method just calls the \method{visit} method of the visitor with 
itself as an argument, like this: \code{visitor.visit(this)}.  For the visitors 
to be able to extend the functionality of all the classes in the elements 
hierarchy, each \type{Visitor} must have one visit method for each concrete 
class in the hierarchy. Say the hierarchy consists of the concrete classes 
\type{ConcreteElementA} and \type{ConcreteElementB}. Then each visitor must have 
the (possibly empty) methods \method{visit(ConcreteElementA element)} and 
\method{visit(ConcreteElementB element)}. This scenario is depicted in 
\myref{fig:visitorPattern}.

\begin{figure}[h]
  \centering
  \tikzstyle{abstract}=[rectangle, draw=black, fill=white, drop shadow, text 
  centered, anchor=north, text=black, text width=6cm, every one node 
part/.style={align=center, font=\bfseries\itshape}]
  \tikzstyle{concrete}=[rectangle, draw=black, fill=white, drop shadow, text 
  centered, anchor=north, text=black, text width=6cm]
  \tikzstyle{inheritarrow}=[->, >=open triangle 90, thick]
  \tikzstyle{commentarrow}=[->, >=angle 90, dashed]
  \tikzstyle{line}=[-, thick]
  \tikzset{every one node part/.style={align=center, font=\bfseries}}
  \tikzset{every second node part/.style={align=center, font=\ttfamily}}
        
  \begin{tikzpicture}[node distance=1cm, scale=0.8, every node/.style={transform 
    shape}]
    \node (Element) [abstract, rectangle split, rectangle split parts=2]
        {
          \nodepart{one}{Element}
          \nodepart{second}{+accept(visitor: Visitor)}
        };
    \node (AuxNode01) [text width=0, minimum height=2cm, below=of Element] {};
    \node (ConcreteElementA) [concrete, rectangle split, rectangle split 
    parts=2, left=of AuxNode01]
        {
          \nodepart{one}{ConcreteElementA}
          \nodepart{second}{+accept(visitor: Visitor)}
        };
    \node (ConcreteElementB) [concrete, rectangle split, rectangle split 
    parts=2, right=of AuxNode01]
        {
          \nodepart{one}{ConcreteElementB}
          \nodepart{second}{+accept(visitor: Visitor)}
        };

    \node[comment, below=of ConcreteElementA] (CommentA) {visitor.visit(this)};

    \node[comment, below=of ConcreteElementB] (CommentB) {visitor.visit(this)};

    \node (AuxNodeX) [text width=0, minimum height=1cm, below=of AuxNode01] {};

    \node (Visitor) [abstract, rectangle split, rectangle split parts=2, 
    below=of AuxNodeX]
        {
          \nodepart{one}{Visitor}
          \nodepart{second}{+visit(ConcreteElementA)\\+visit(ConcreteElementB)}
        };
    \node (AuxNode02) [text width=0, minimum height=2cm, below=of Visitor] {};
    \node (ConcreteVisitor1) [concrete, rectangle split, rectangle split 
    parts=2, left=of AuxNode02]
        {
          \nodepart{one}{ConcreteVisitor1}
          \nodepart{second}{+visit(ConcreteElementA)\\+visit(ConcreteElementB)}
        };
    \node (ConcreteVisitor2) [concrete, rectangle split, rectangle split 
    parts=2, right=of AuxNode02]
        {
          \nodepart{one}{ConcreteVisitor2}
          \nodepart{second}{+visit(ConcreteElementA)\\+visit(ConcreteElementB)}
        };

    
    \draw[inheritarrow] (ConcreteElementA.north) -- ++(0,0.7) -| 
    (Element.south);
    \draw[line] (ConcreteElementA.north) -- ++(0,0.7) -| 
    (ConcreteElementB.north);

    \draw[inheritarrow] (ConcreteVisitor1.north) -- ++(0,0.7) -| 
    (Visitor.south);
    \draw[line] (ConcreteVisitor1.north) -- ++(0,0.7) -| 
    (ConcreteVisitor2.north);

    \draw[commentarrow] (CommentA.north) -- (ConcreteElementA.south);
    \draw[commentarrow] (CommentB.north) -- (ConcreteElementB.south);

    
  \end{tikzpicture}
  \caption{The Visitor Pattern.}
  \label{fig:visitorPattern}
\end{figure}

The use of the visitor pattern can be appropriate when the hierarchy of elements 
is mostly stable, but the family of operations over its elements is constantly 
growing. This is clearly the case for the \name{Eclipse} AST, since the 
hierarchy for the type \type{ASTNode} is very stable, but the functionality of 
its elements is extended every time someone need to operate on the AST. Another 
aspect of the \name{Eclipse} implementation is that it is a public API, and the 
visitor pattern is an easy way to provide access to the nodes in the tree.

The version of the visitor pattern implemented for the AST nodes in \name{Eclipse} also 
provides an elegant way to traverse the tree. It does so by following the 
convention that every node in the tree first let the visitor visit itself, 
before it also makes all its children accept the visitor. The children are only 
visited if the visit method of their parent returns \var{true}. This pattern 
then makes for a prefix traversal of the AST. If postfix traversal is desired, 
the visitors also have \method{endVisit} methods for each node type, which is 
called after the \method{visit} method for a node. In addition to these visit 
methods, there are also the methods \method{preVisit(ASTNode)}, 
\method{postVisit(ASTNode)} and \method{preVisit2(ASTNode)}. The 
\method{preVisit} method is called before the type-specific \method{visit} 
method. The \method{postVisit} method is called after the type-specific 
\method{endVisit}. The type specific \method{visit} is only called if 
\method{preVisit2} returns \var{true}. Overriding the \method{preVisit2} is also 
altering the behavior of \method{preVisit}, since the default implementation is 
responsible for calling it.

An example of a trivial \type{ASTVisitor} is shown in 
\myref{lst:astVisitorExample}.

\begin{listing}
\begin{minted}{java}
public class CollectNamesVisitor extends ASTVisitor {
    Collection<Name> names = new LinkedList<Name>();

    @Override
    public boolean visit(QualifiedName node) {
      names.add(node);
      return false;
    }

    @Override
    public boolean visit(SimpleName node) {
        names.add(node);
        return true;
    }
} 
\end{minted}
\caption{An \type{ASTVisitor} that visits all the names in a subtree and adds 
them to a collection, except those names that are children of any 
\type{QualifiedName}.}
\label{lst:astVisitorExample}
\end{listing}

\section{The refactoring API of Eclipse}\label{ch:jdt_refactorings}
This section will present the design behind the refactoring support in 
\name{Eclipse}, and the JDT in specific. After which it will follow a section 
about shortcomings of the refactoring API in terms of composition of 
refactorings.

\subsection{Design}
The refactoring world of \name{Eclipse} can in general be separated into two parts: The 
language independent part and the part written for a specific programming 
language -- the language that is the target of the supported refactorings.  

\subsubsection{The Language Toolkit.}
The Language Toolkit\footnote{The content of this section is a mixture of 
  written material from 
  \url{https://www.eclipse.org/articles/Article-LTK/ltk.html} and 
  \url{http://www.eclipse.org/articles/article.php?file=Article-Unleashing-the-Power-of-Refactoring/index.html}, 
the LTK source code and my own memory.}, or LTK for short, is the framework that 
is used to implement refactorings in \name{Eclipse}.  It is language independent and 
provides the abstractions of a refactoring and the change it generates, in the 
form of the classes \typewithref{org.eclipse.ltk.core.refactoring}{Refactoring} 
and \typewithref{org.eclipse.ltk.core.refactoring}{Change}.

There are also parts of the LTK that is concerned with user interaction, but 
they will not be discussed here, since they are of little value to us and our 
use of the framework. We are primarily interested in the parts that can be 
automated.

\paragraph{The Refactoring class.}
The abstract class \type{Refactoring} is the core of the LTK framework. Every 
refactoring that is going to be supported by the LTK has to end up creating an 
instance of one of its subclasses. The main responsibilities of subclasses of 
\type{Refactoring} are to implement template methods for condition checking 
(\methodwithref{org.eclipse.ltk.core.refactoring.Refactoring}{checkInitialConditions} 
and 
\methodwithref{org.eclipse.ltk.core.refactoring.Refactoring}{checkFinalConditions}), 
in addition to the 
\methodwithref{org.eclipse.ltk.core.refactoring.Refactoring}{createChange} 
method that creates and returns an instance of the \type{Change} class.

If the refactoring shall support that others participate in it when it is 
executed, the refactoring has to be a processor-based 
refactoring\typeref{org.eclipse.ltk.core.refactoring.participants.ProcessorBasedRefactoring}.  
It then delegates to its given 
\typewithref{org.eclipse.ltk.core.refactoring.participants}{RefactoringProcessor} 
for condition checking and change creation. Participating in a refactoring can 
be useful in cases where the changes done to programming source code affect
other related resources in the workspace. This can be names or paths in 
configuration files, or maybe one would like to perform additional logging of 
changes done in the workspace.

\paragraph{The Change class.}
This class is the base class for objects that is responsible for performing the 
actual workspace transformations in a refactoring. The main responsibilities for 
its subclasses are to implement the 
\methodwithref{org.eclipse.ltk.core.refactoring.Change}{perform} and 
\methodwithref{org.eclipse.ltk.core.refactoring.Change}{isValid} methods. The 
\method{isValid} method verifies that the change object is valid and thus can be 
executed by calling its \method{perform} method. The \method{perform} method 
performs the desired change and returns an undo change that can be executed to 
reverse the effect of the transformation done by its originating change object. 

\paragraph{Executing a refactoring}\label{executing_refactoring}
The life cycle of a refactoring generally follows two steps after creation: 
condition checking and change creation. By letting the refactoring object be 
handled by a 
\typewithref{org.eclipse.ltk.core.refactoring}{CheckConditionsOperation} that
in turn is handled by a 
\typewithref{org.eclipse.ltk.core.refactoring}{CreateChangeOperation}, it is 
assured that the change creation process is managed in a proper manner.

The actual execution of a change object has to follow a detailed life cycle.  
This life cycle is honored if the \type{CreateChangeOperation} is handled by a 
\typewithref{org.eclipse.ltk.core.refactoring}{PerformChangeOperation}. If also 
an undo manager\typeref{org.eclipse.ltk.core.refactoring.IUndoManager} is set 
for the \type{PerformChangeOperation}, the undo change is added into the undo 
history.

\subsubsection{The language specific refactorings}\label{sec:jdtRefactorings}
The language specific refactorings supplied by the JDT that are relevant for 
this project are presented below. It is the JDT-implementations of the two 
primitive refactorings \ExtractMethod and \MoveMethod. In the JDT, the 
implementations of these refactorings are found in the classes 
\typewithref{org.eclipse.jdt.internal.corext.refactoring.code}{ExtractMethodRefactoring} 
and 
\typewithref{org.eclipse.jdt.internal.corext.refactoring.structure}{MoveInstanceMethodProcessor}, 
where the last class is designed to be used together with the processor-based 
\typewithref{org.eclipse.ltk.core.refactoring.participants}{MoveRefactoring}.

\paragraph{The ExtractMethodRefactoring class.}
This class is quite simple in its use. The only parameters it requires for 
construction is a compilation 
unit\typeref{org.eclipse.jdt.core.ICompilationUnit}, the offset into the source 
code where the extraction shall start, and the length of the source to be 
extracted. Then you have to set the method name for the new method together with 
its visibility and some not so interesting parameters.

\paragraph{The MoveInstanceMethodProcessor class.}
For the \refa{Move Method}, the processor requires a little more advanced input than  
the class for the \refa{Extract Method}. For construction it requires a method 
handle\typeref{org.eclipse.jdt.core.IMethod} for the method that is to be moved.  
Then the target for the move has to be supplied as the variable binding from a 
chosen variable declaration. In addition to this, some parameters have to be set 
regarding setters/getters, as well as delegation.

To make the processor a working refactoring, a \type{MoveRefactoring} must be 
created with it as a parameter.


\subsection{Shortcomings}
This section is introduced naturally with a conclusion: The JDT refactoring 
implementations do not facilitate composition of refactorings. This section will 
try to explain why, and also try to identify other shortcomings of both the 
usability and readability of the JDT refactoring source code.

\subsubsection{Absence of generics in Eclipse source code}
This section is not only concerning the JDT refactoring API, but also large 
quantities of the \name{Eclipse} source code. The code shows a striking absence of the 
Java language feature of generics. It is hard to read a class' interface when 
methods return objects or takes parameters of raw types such as \type{List} or 
\type{Map}. This sometimes results in having to read a lot of source code to 
understand what is going on, instead of relying on the available interfaces. In 
addition, it results in a lot of ugly code, making the use of typecasting more 
of a rule than an exception.

\subsubsection{Composite refactorings will not appear as atomic actions}
When composing primitive refactorings from the JDT, it is not possible to make 
them appear as being executed as one change, but only as multiple small changes.

\paragraph{Missing Flexibility from JDT Refactorings.}
The JDT refactorings are not made with composition of refactorings in mind. When 
a JDT refactoring is executed, it assumes that all conditions for it to be 
applied successfully can be found by reading source files that have been 
persisted to disk. They can only operate on the actual source material, and not 
(in-memory) copies thereof. This constitutes a major disadvantage when trying to 
compose refactorings, since if an exception occurs in the middle of a sequence 
of refactorings, it can leave the project in a state where the composite 
refactoring was only partially executed. It makes it hard to discard the changes 
done without monitoring and consulting the undo manager, an approach that is not 
bullet proof.

\paragraph{Broken Undo History.}
When designing a composed refactoring that is to be performed as a sequence of 
refactorings, you would like it to appear as a single change to the workspace.  
This implies that you would also like to be able to undo all the changes done by 
the refactoring in a single step. This is not the way it appears when a sequence 
of JDT refactorings is executed. It leaves the undo history filled up with 
individual undo actions corresponding to every single JDT refactoring in the 
sequence. This problem is not trivial to handle in \name{Eclipse} 
\see{sec:hackingUndoHistory}.


\chapter{Source code organization and implementation 
details}\label{ch:implementation}

\section{Source code organization}
All the parts of this master's project are under version control with 
\name{Git}\footnote{\url{http://git-scm.com/}}.

The software written is organized as some \name{Eclipse} plugins. Writing a plugin is 
the natural way to utilize the API of \name{Eclipse}. This also makes it possible to 
provide a user interface to manually run operations on selections in program 
source code or whole projects/packages.

When writing a plugin in \name{Eclipse}, one has access to resources such as the 
current workspace, the open editor and the current selection.

The thesis work is contained in the following Eclipse projects:

\begin{description}
  \item[no.uio.ifi.refaktor] \hfill \\ This is the main Eclipse plugin 
    project, and contains all of the business logic for the plugin.

  \item[no.uio.ifi.refaktor.tests] \hfill \\
    This project contains the tests for the main plugin.

  \item[no.uio.ifi.refaktor.examples] \hfill \\
    Contains example code used in testing. It also contains code for managing 
    this example code, such as creating an Eclipse project from it before a test 
    run.

  \item[no.uio.ifi.refaktor.benchmark] \hfill \\
    This project contains code for running search based versions of the 
    composite refactoring over selected Eclipse projects.

  \item[no.uio.ifi.refaktor.releng] \hfill \\
    Contains the rmap, queries and target definitions needed by Buckminster on 
    the Jenkins continuous integration server.

\end{description}

\subsection{The no.uio.ifi.refaktor project}

\subsubsection{no.uio.ifi.refaktor.analyze}
This package, and its sub-packages, contains code that is used for analyzing 
Java source code. The most important sub-packages are presented below.

\begin{description}
  \item[no.uio.ifi.refaktor.analyze.analyzers] \hfill \\
    This package contains source code analyzers. These are usually responsible 
    for analyzing text selections or running specialized analyzers for different 
    kinds of entities.  Their structures are often hierarchical. This means that 
    you have an analyzer for text selections, that in turn is utilized by an 
    analyzer that analyzes all the selections of a method. Then there are 
    analyzers for analyzing all the methods of a type, all the types of a 
    compilation unit, all the compilation units of a package, and, at last, all 
    of the packages in a project.

  \item[no.uio.ifi.refaktor.analyze.checkers] \hfill \\
    A package containing checkers.  The checkers are classes used to validate 
    that a selection can be further analyzed and chosen as a candidate for a 
    refactoring. Invalidating properties can be such as usage of inner classes 
    or the need for multiple return values.  

  \item[no.uio.ifi.refaktor.analyze.collectors] \hfill \\
    This package contains the property collectors. Collectors are used to gather 
    properties from a text selection.  This is mostly properties regarding 
    referenced names and their occurrences. It is these properties that make up 
    the basis for finding the best candidates for a refactoring.
\end{description}

\subsubsection{no.uio.ifi.refaktor.change}
This package, and its sub-packages, contains functionality for manipulate source 
code.

\begin{description}
  \item[no.uio.ifi.refaktor.change.changers] \hfill \\
    This package contains source code changers. They are used to glue together 
    the analysis of source code and the actual execution of the changes.

  \item[no.uio.ifi.refaktor.change.executors] \hfill \\
    The executors that are responsible for making concrete changes are found in 
    this package. They are mostly used to create and execute one or more Eclipse 
    refactorings.

  \item[no.uio.ifi.refaktor.change.processors] \hfill \\
    Contains a refactoring processor for the \MoveMethod refactoring. The code 
    is stolen and modified to fix a bug. The related bug is described in
    \myref{eclipse_bug_429416}.

\end{description}

\subsubsection{no.uio.ifi.refaktor.handlers}
This package contains handlers for the commands defined in the plugin manifest. 

\subsubsection{no.uio.ifi.refaktor.prefix}
This package contains the \type{Prefix} type that is the data representation of 
the prefixes found by the \type{PrefixesCollector}. It also contains the prefix 
set for storing and working with prefixes.

\subsubsection{no.uio.ifi.refaktor.statistics}
The package contains statistics functionality. Its heart is the statistics 
aspect that is responsible for gathering statistics during the execution of the 
\ExtractAndMoveMethod refactoring.

\begin{description}
  \item[no.uio.ifi.refaktor.statistics.reports] \hfill \\
    This package contains a simple framework for generating reports from the 
    statistics data generated by the aspect. Currently, the only available 
    report type is a simple text report.

\end{description}


\subsubsection{no.uio.ifi.refaktor.textselection}
This package contains the two custom text selections that are used extensively 
throughout the project. One of them is just a subclass of the other, to support 
the use of the memento pattern to optimize the memory usage during benchmarking.

\subsubsection{no.uio.ifi.refaktor.debugging}
The package contains a debug utility class. I addition to this, the package 
\code{no.uio.ifi.refaktor.utils.aspects} contains a couple of aspects used for 
debugging purposes. 

\subsubsection{no.uio.ifi.refaktor.utils}
Utility package that contains all the functionality that has to do with parsing 
of source code. It also has utility classes for looking up handles to methods 
and types et cetera.

\begin{description}
  \item[no.uio.ifi.refaktor.utils.caching] \hfill \\
    This package contains the caching manager for compilation units, along with 
    classes for different caching strategies.

  \item[no.uio.ifi.refaktor.utils.nullobjects] \hfill \\
    Contains classes for creating different null objects. Most of the classes 
    are used to represent null objects of different handle types. These null 
    objects are returned from various utility classes instead of returning a 
    \var{null} value when other values are not available.

\end{description}


\section{Implementing source code analysis}
This section gathers implementation details for the most important parts of the 
source code analysis for the \ExtractAndMoveMethod refactoring.

\subsection{Representing prefixes}
This section shows the classes responsible for representing and working with 
prefixes.

\subsubsection{The Prefix class}
This class exists mainly for holding data about a prefix, such as the expression 
that the prefix represents and the occurrence count of the prefix within a 
selection. In addition to this, it has some functionality such as calculating 
its sub-prefixes and intersecting it with another prefix. The definition of the 
intersection between two prefixes is a prefix representing the longest common 
expression between the two.

\subsubsection{The PrefixSet class}
A prefix set holds elements of type \type{Prefix}. It is implemented with the 
help of a \typewithref{java.util}{HashMap} and contains some typical set 
operations, but it does not implement the \typewithref{java.util}{Set} 
interface, since the prefix set does not need all of the functionality a 
\type{Set} requires to be implemented. In addition It needs some other 
functionality not found in the \type{Set} interface. So due to the relatively 
limited use of prefix sets, and that it almost always needs to be referenced as 
such, and not a \type{Set<Prefix>}, it remains as an ad hoc solution to a 
concrete problem.

There are two ways adding prefixes to a \type{PrefixSet}. The first is through 
its \method{add} method. This works like one would expect from a set. It adds 
the prefix to the set if it does not already contain the prefix. The other way 
is to \emph{register} the prefix with the set. When registering a prefix, if the 
set does not contain the prefix, it is just added. If the set contains the 
prefix, its count gets incremented. This is how the occurrence count is handled.

The prefix set also computes the set of prefixes that is not enclosing any 
prefixes of another set. This is kind of a set difference operation only for 
enclosing prefixes.

\subsection{Property collectors}\label{propertyCollectors}
The prefixes and unfixes are found by property 
collectors\typeref{no.uio.ifi.refaktor.extractors.collectors.PropertyCollector}.  
A property collector is of the \type{ASTVisitor} type, and thus visits nodes of 
type \type{ASTNode} of the abstract syntax tree \see{astVisitor}.

\paragraph{The PrefixesCollector.}
The \typewithref{no.uio.ifi.refaktor.extractors.collectors}{PrefixesCollector} 
finds prefixes that makes up the basis for calculating move targets for the 
\refa{Extract and Move Method} refactoring. It visits expression 
statements\typeref{org.eclipse.jdt.core.dom.ExpressionStatement} and creates 
prefixes from its expressions in the case of method invocations. The prefixes 
found are registered with a prefix set, together with all its sub-prefixes.

\paragraph{The UnfixesCollector.}\label{unfixes}
The \typewithref{no.uio.ifi.refaktor.extractors.collectors}{UnfixesCollector} 
finds unfixes within a selection. Its semantics is described in 
\myref{sec:unfixes}.

\subsection{Checkers}\label{checkers}
The checkers are a range of classes that checks that text selections comply
with certain criteria. All checkers operates under the assumption that the code 
they check is free from compilation errors. If a 
\typewithref{no.uio.ifi.refaktor.analyze.analyzers}{Checker} fails, it throws a 
\type{CheckerException}. The checkers are managed by the 
\type{LegalStatementsChecker}, which does not, in fact, implement the 
\type{Checker} interface. It does, however, run all the checkers registered with 
it, and reports that all statements are considered legal if no 
\type{CheckerException} is thrown. Many of the checkers either extends the 
\type{PropertyCollector} or utilizes one or more property collectors to verify 
some criteria. The checkers registered with the \type{LegalStatementsChecker} 
are described next. They are run in the order presented below.

\subsubsection{The CallToProtectedOrPackagePrivateMethodChecker}
This checker is used to check that at selection does not contain a call to a 
method that is protected or package-private. Such a method either has the access 
modifier \code{protected} or it has no access modifier.

The workings of the \type{CallToProtectedOrPackagePrivateMethod\-Checker} is
very simple. It looks for calls to methods that are either protected or 
package-private within the selection, and throws an 
\type{IllegalExpressionFoundException} if one is found.

\subsubsection{The DoubleClassInstanceCreationChecker}
The \type{DoubleClassInstanceCreationChecker} checks that there are no double 
class instance creations where the inner constructor call takes an argument that 
is built up using field references.

The checker visits all nodes of type \type{ClassInstanceCreation} within a 
selection. For all of these nodes, if its parent also is a class instance 
creation, it accepts a visitor that throws a 
\type{IllegalExpressionFoundException} if it encounters a name that is a field 
reference.

\subsubsection{The InstantiationOfNonStaticInnerClassChecker}
The \type{InstantiationOfNonStaticInnerClassChecker} checks that selections
do not contain instantiations of non-static inner classes. The 
\type{MoveInstanceMethodProcessor} in \name{Eclipse} does not handle such 
instantiations gracefully when moving a method. This problem is also related to 
bug\ldots \todoin{File Eclipse bug report}

\subsubsection{The EnclosingInstanceReferenceChecker}
The purpose of this checker is to verify that the names in a text selection are 
not referencing any enclosing instances. In theory, the underlying problem could 
be solved in some situations, but our dependency on the 
\type{MoveInstanceMethodProcessor} prevents this.

The 
\typewithref{no.uio.ifi.refaktor.analyze.analyzers}{EnclosingInstanceReferenceChecker} 
is a modified version of the 
\typewithref{org.eclipse.jdt.internal.corext.refactoring.structure.MoveInstanceMethod\-Processor}{EnclosingInstanceReferenceFinder} 
from the \type{MoveInstanceMethodProcessor}. Wherever the 
\type{EnclosingInstanceReferenceFinder} would create a fatal error status, the
checker will throw a \type{CheckerException}.

The checker works by first finding all of the enclosing types of a selection.  
Thereafter, it visits all the simple names of the selection to check that they 
are not references to variables or methods declared in any of the enclosing 
types. In addition, the checker visits \var{this}-expressions to verify that no 
such expressions are qualified with any name.

\subsubsection{The ReturnStatementsChecker}\label{returnStatementsChecker}
The checker for return statements is meant to verify that a text selection is 
consistent regarding return statements.

If the selection is free from return statements, then the checker validates.  So 
this is the first thing the checker investigates.

If the checker proceeds any further, it is because the selection contains one 
or more return statements. The next test is therefore to check if the last 
statement of the selection ends in either a return or a throw statement. The 
responsibility for checking that the last statement of the selection eventually 
ends in a return or throw statement, is put on the 
\type{LastStatementOfSelectionEndsInReturnOrThrowChecker}. For every node 
visited, if the node is a statement, it does a test to see if the statement is a 
return, a throw or if it is an implicit return statement. If this is the case, 
no further checking is done. This checking is done in the \code{preVisit2} 
method \see{astVisitor}. If the node is not of a type that is being handled by 
its type-specific visit method, the checker performs a simple test. If the node 
being visited is not the last statement of its parent that is also enclosed by 
the selection, an \type{IllegalStatementFoundException} is thrown. This ensures 
that all statements are taken care of, one way or the other. It also ensures 
that the checker is conservative in the way it checks for legality of the 
selection.

To examine if a statement is an implicit return statement, the checker first 
finds the last statement declared in its enclosing method. If this statement is 
the same as the one under investigation, it is considered an implicit return 
statement. If the statements are not the same, the checker does a search to see 
if the statement examined is also the last statement of the method that can be 
reached. This includes the last statement of a block statement, a labeled 
statement, a synchronized statement or a try statement, that in turn is the last 
statement enclosed by one of the statement types listed. This search goes 
through all the parents of a statement until a statement is found that is not 
one of the mentioned acceptable parent statements. If the search ends in a 
method declaration, then the statement is considered to be the last reachable 
statement of the method, and thus it is an implicit return statement.

There are two kinds of statements that are handled explicitly: If-statements and 
try-statements. Block, labeled and do-statements are handled by fall-through to 
the other two.

If-statements are handled explicitly by overriding their type-specific visit 
method. If the then-part does not contain any return or throw statements an 
\type{IllegalStatementFoundException} is thrown. If it does contain a return or 
throw, its else-part is checked. If the else-part is non-existent, or it does 
not contain any return or throw statements an exception is thrown. If no 
exception is thrown while visiting the if-statement, its children are visited.

A try-statement is checked very similar to an if-statement. Its body must 
contain a return or throw. The same applies to its catch clauses and finally 
body. Failure to validate produces an \type{IllegalStatementFoundException}.

If the checker does not complain at any point, the selection is considered valid 
with respect to return statements.

\subsubsection{The AmbiguousReturnValueChecker}
This checker verifies that there are no ambiguous return values in a selection.

First, the checker needs to collect some data. Those data are the binding keys 
for all simple names that are assigned to within the selection, including 
variable declarations, but excluding fields. The checker also finds out whether 
a return statement is found in the selection or not. No further checks of return 
statements are needed, since, at this point, the selection is already checked 
for illegal return statements \see{returnStatementsChecker}.

After the binding keys of the assignees are collected, the checker searches the 
part of the enclosing method that is after the selection for references whose 
binding keys are among the collected keys. If more than one unique referral is 
found, or only one referral is found, but the selection also contains a return 
statement, we have a situation with an ambiguous return value, and an exception 
is thrown.

%\todoin{Explain why we do not need to consider variables assigned inside 
%local/anonymous classes. (The referenced variables need to be final and so 
%on\ldots)}

\subsubsection{The IllegalStatementsChecker}
This checker is designed to check for illegal statements.

Notice that labels in break and continue statements need some special treatment. 
Since a label does not have any binding information, we have to search upwards 
in the AST to find the \type{LabeledStatement} that corresponds to the label 
from the break or continue statement, and check that it is contained in the 
selection. If the break or continue statement does not have a label attached to 
it, it is checked that its innermost enclosing loop or switch statement (break 
statements only) also is contained in the selection.



\subsection{Source code analyzers}
The analyzers presented in this section are used to analyze source code. The 
\typewithref{no.uio.ifi.refaktor.analyze.analyzers}{ExtractAndMoveMethodAnalyzer} 
can be used to analyze a selection, while the 
\typewithref{no.uio.ifi.refaktor.analyze.analyzers}{SearchBasedExtractAndMoveMethodAnalyzer} 
analyzes all the text selection for a method.

\subsubsection{The 
  \type{ExtractAndMoveMethodAnalyzer}}\label{extractAndMoveMethodAnalyzer}
This analyzer can perform analysis and precondition checking for an 
\ExtractAndMoveMethod refactoring. First it checks whether a text selection is a 
valid selection or not, with respect to statement boundaries and that it 
actually contains any selections.  Then it checks the legality of both 
extracting the selection and also moving it to another class.  This checking of 
is performed by a range of checkers \see{checkers}.  If the selection is 
approved as legal, it is analyzed to find the presumably best target to move the 
extracted method to.

For finding the best suitable target the analyzer is using a 
\typewithref{no.uio.ifi.refaktor.analyze.collectors}{PrefixesCollector} that 
collects all the possible candidate targets for the refactoring. All the 
non-candidates are found by an 
\typewithref{no.uio.ifi.refaktor.analyze.collectors}{UnfixesCollector} that 
collects all the targets that will give some kind of error if used.  (For 
details about the property collectors, see \myref{propertyCollectors}.) All 
prefixes (and unfixes) are represented by a 
\typewithref{no.uio.ifi.refaktor.extractors}{Prefix}, and they are collected 
into sets of prefixes. The safe prefixes are found by subtracting from the set 
of candidate prefixes the prefixes that is enclosing any of the unfixes.  A 
prefix is enclosing an unfix if the unfix is in the set of its sub-prefixes.  As 
an example, \code{``a.b''} is enclosing \code{``a''}, as is \code{``a''}. The 
safe prefixes is unified in a \type{PrefixSet}. If a prefix has only one 
occurrence, and is a simple expression, it is considered unsuitable as a move 
target. This occurs in statements such as \code{``a.foo()''}. For such 
statements it bares no meaning to extract and move them. It only generates an 
extra method and the calling of it. 

The most suitable target for the refactoring is found by finding the prefix with 
the most occurrences. If two prefixes have the same occurrence count, but they 
differ in the number of segments, the one with most segments is chosen.


\subsubsection{The \type{SearchBasedExtractAndMoveMethodAnalyzer}}
This analyzer can be used for analyzing all the possible text selections of a 
method and then to choose the best result out of the analysis results that are, 
by the analyzer, considered to be the potential candidates for the 
\ExtractAndMoveMethod refactoring.

Before the analyzer is able to work with the text selections of a method, it 
needs to generate them. To do this, it parses the method to obtain a 
\type{MethodDeclaration} for it \see{astEclipse}. Then there is a statement 
lists creator that creates statements lists of the different groups of 
statements in the body of the method declaration. A text selections generator 
generates text selections of all the statement lists for the analyzer to work 
with.

\paragraph{The statement lists creator}
is responsible for generating lists of statements for all the possible nesting 
levels of statements in the method. The statement lists creator is implemented 
as an AST visitor \see{astVisitor}. It generates lists of statements by visiting 
all the blocks in the method declaration and stores their statements in a 
collection of statement lists. In addition, it visits all of the other 
statements that can have a statement as a child, such as the different control 
structures and the labeled statement.

The switch statement is the only kind of statement that is not straight forward 
to obtain the child statements from. It stores all of its children in a flat 
list. Its switch case statements are included in this list. This means that 
there are potential statement lists between all of these case statements. The 
list of statements from a switch statement is therefore traversed, and the 
statements between the case statements are grouped as separate lists.

The highlighted part of \myref{lst:grandExample} shows an example of how the 
statement lists creator would separate a method body into lists of statements.

\paragraph{The text selections generator} generates text selections for each 
list of statements from the statement lists creator. The generator generates a 
text selection for every sub-sequence of statements in a list. For a sequence of 
statements, the first statement and the last statement span out a text 
selection. 

In practice, the text selections are calculated by only one traversal of the 
statement list. There is a set of generated text selections. For each statement, 
there is created a temporary set of selections, in addition to a text selection 
based on the offset and length of the statement. This text selection is added to 
the temporary set. Then the new selection is added with every selection from the 
set of generated text selections. These new selections are added to the 
temporary set. Then the temporary set of selections is added to the set of 
generated text selections. The result of adding two text selections is a new 
text selection spanned out by the two addends. 

\begin{comment}
\begin{listing}[h]
\def\charwidth{5.7pt}
\def\indent{4*\charwidth}
\def\lineheight{\baselineskip}
\def\mintedtop{\lineheight}

\begin{tikzpicture}[overlay, yscale=-1]
  \tikzstyle{overlaybox}=[fill=lightgray,opacity=0.2]

  \draw[overlaybox] (2*\charwidth,\mintedtop) rectangle 
  +(18*\charwidth,\lineheight);

  \draw[overlaybox] (2*\charwidth,\mintedtop+\lineheight) rectangle 
  +(18*\charwidth,\lineheight);

  \draw[overlaybox] (2*\charwidth,\mintedtop+3*\lineheight) rectangle 
  +(18*\charwidth,\lineheight);

  \draw[overlaybox] (\indent-3*\charwidth,\mintedtop) rectangle 
  +(20*\charwidth,2*\lineheight);

  \draw[overlaybox] (3*\charwidth,\mintedtop+\lineheight) rectangle 
  +(16*\charwidth,3*\lineheight);

  \draw[overlaybox] (\indent,\mintedtop) rectangle 
  +(14*\charwidth,4*\lineheight);
\end{tikzpicture}
\begin{minted}{java}
    statement one;
    statement two;
    ...
    statement k;
\end{minted}
\caption{Example of how the text selections generator would generate text 
  selections based on a lists of statements. Each highlighted rectangle 
represents a text selection.}
\label{lst:textSelectionsExample}
\end{listing}
\todoin{fix \myref{lst:textSelectionsExample}? Text only? All 
sub-sequences\ldots}
\end{comment}

\paragraph{Finding the candidate} for the refactoring is done by analyzing all 
the generated text selections with an \type{ExtractAndMoveMethodAnalyzer} 
\see{extractAndMoveMethodAnalyzer}. If the analyzer generates a useful result, 
an \type{ExtractAndMoveMethod\-Candidate} is created from it, which is kept in a 
list of potential candidates. If no candidates are found, the 
\type{NoTargetFoundException} is thrown.

Since only one of the candidates can be chosen, the analyzer must sort out which 
candidate to choose. The sorting is done by the static \method{sort} method of 
\type{Collections}. The comparison in this sorting is done by an 
\type{ExtractAndMoveMethodCandidateComparator}. The implementation used is the 
\type{FavorNoUnfixesCandidateComparator}. Its sort criteria are the same as in 
\myref{sec:choosingSelection}.


\section{Composite refactoring implementations}
This section will present how composite refactorings are implemented within the 
bounds of the Eclipse platform and the JDT.

\subsection{A simple ad hoc model}
As pointed out in \myref{ch:jdt_refactorings}, the \name{Eclipse} JDT refactoring model 
is not very well suited for making composite refactorings. Therefore a simple 
model using changer objects (of type \type{RefaktorChanger}) is used as an 
abstraction layer on top of the existing \name{Eclipse} refactorings, instead of 
extending the \typewithref{org.eclipse.ltk.core.refactoring}{Refactoring} class.  

The use of an additional abstraction layer is a deliberate choice. It is due to 
the problem of creating a composite 
\typewithref{org.eclipse.ltk.core.refactoring}{Change} that can handle text 
changes that interfere with each other. Thus, a \type{RefaktorChanger} may, or 
may not, take advantage of one or more existing refactorings, but it is always 
intended to make a change to the workspace.

\paragraph{The typical \type{RefaktorChanger}.}
The typical refaktor changer class has two responsibilities: Checking 
preconditions and executing changes. This is not too different from the 
responsibilities of an LTK refactoring, with the distinction that a refaktor 
changer also executes the change, while an LTK refactoring is only responsible 
for creating the object that can later be used to do that job.

Checking of preconditions is typically done by an 
\typewithref{no.uio.ifi.refaktor.analyze.analyzers}{Analyzer}. If the 
preconditions validate, the upcoming changes are executed by an 
\typewithref{no.uio.ifi.refaktor.change.executors}{Executor}.

\subsection{A simple Extract and Move Method refactoring}
This section describes the implementation of a simple refactoring, that for a 
given text selection will analyze it and perform the \ExtractAndMoveMethod 
refactoring if a suitable move target is found within the selection.


\paragraph{The ExtractAndMoveMethodChanger.}
This 
changer\footnote{\type{no.uio.ifi.refaktor.changers.ExtractAndMoveMethodChanger}}
is a subclass of the class 
\typewithref{no.uio.ifi.refaktor.changers}{RefaktorChanger}. It is responsible 
for analyzing and finding the best target for, and also executing, an 
\ExtractAndMoveMethod refactoring.  This particular changer is the one of my 
changers that is closest to being a true LTK refactoring. It can be reworked to 
be one if the problems with overlapping changes are resolved.

The changer requires a text selection and the name of the new method, or else a 
method name will be generated. The selection has to be of the type
\typewithref{no.uio.ifi.refaktor.utils}{CompilationUnitTextSelection}. This 
class is a custom extension to 
\typewithref{org.eclipse.jface.text}{TextSelection}, that in addition to the 
basic offset, length and similar methods, also carry an instance of the 
underlying compilation unit handle for the selection.

The analysis and precondition checking for this changer is done by an 
\type{ExtractAndMoveMethodAnalyzer} \see{extractAndMoveMethodAnalyzer}, and the 
execution is done by an \type{ExtractAndMoveMethodExecutor}.

\paragraph{The 
  \type{ExtractAndMoveMethodExecutor}.}\label{extractAndMoveMethodExecutor}
If the analysis finds a possible target for the composite refactoring, it is 
executed by an 
\typewithref{no.uio.ifi.refaktor.change.executors}{ExtractAndMoveMethodExecutor}.  
It is composed of the two executors known as 
\typewithref{no.uio.ifi.refaktor.change.executors}{ExtractMethodRefactoringExecutor} 
and 
\typewithref{no.uio.ifi.refaktor.change.executors}{MoveMethod\-RefactoringExecutor}.  
The \type{ExtractAndMoveMethodExecutor} is responsible for gluing the two 
together by feeding the \type{MoveMethod\-RefactoringExecutor} with the 
resources needed after executing the extract method refactoring.

\paragraph{The \type{ExtractMethodRefactoringExecutor}.}
This executor is responsible for creating and executing an instance of the 
\type{ExtractMethodRefactoring} class. It is also responsible for collecting 
some post execution resources that can be used to find the method handle for the 
extracted method, as well as information about its parameters, including the 
variable they originated from.

\paragraph{The \type{MoveMethodRefactoringExecutor}.}
This executor is responsible for creating and executing an instance of the 
\type{MoveRefactoring}. The move refactoring is a processor-based refactoring, 
and for the \refa{Move Method} refactoring it is the \type{MoveInstanceMethodProcessor} 
that is used.

The handle for the method to be moved is found on the basis of the information 
gathered after the execution of the \refa{Extract Method} refactoring. The only 
information the \type{ExtractMethodRefactoring} is sharing after its execution, 
regarding finding the method handle, is the textual representation of the new 
method signature. Therefore it must be parsed, the strings for types of the 
parameters must be found and translated to a form that can be used to look up 
the method handle from its type handle. They have to be on the unresolved form. 
The name for the type is found from the original selection, since an extracted 
method must end up in the same type as the originating method.

When analyzing a selection prior to performing the \refa{Extract Method} refactoring, a 
target is chosen. It has to be a variable binding, so it is either a field or a 
local variable/parameter. If the target is a field, it can be used with the 
\type{MoveInstanceMethodProcessor} as it is, since the extracted method still is 
in its scope. But if the target is local to the originating method, the target 
that is to be used for the processor must be among its parameters. Thus the 
target must be found among the extracted method's parameters. This is done by 
finding the parameter information object that corresponds to the parameter that 
was declared on basis of the original target's variable when the method was 
extracted. (The extracted method must take one such parameter for each local 
variable that is declared outside the selection that is extracted.) To match the 
original target with the correct parameter information object, the key for the 
information object is compared to the key from the original target's binding.  
The source code must then be parsed to find the method declaration for the 
extracted method. The new target must be found by searching through the 
parameters of the declaration and choose the one that has the same type as the 
old binding from the parameter information object, as well as the same name that 
is provided by the parameter information object.

\subsection{An on-demand search-based Extract and Move Method refactoring} 
\label{searchBasedExtractAndMoveMethodChanger}
The 
\typewithref{no.uio.ifi.refaktor.change.changers}{SearchBasedExtractAndMoveMethodChanger} 
is a changer whose purpose is to automatically analyze a method, and execute the 
\ExtractAndMoveMethod refactoring on it if it is a suitable candidate for the 
refactoring.

First, the \type{SearchBasedExtractAndMoveMethodAnalyzer} is used to analyze the 
method. If the method is found to be a candidate, the result from the analysis 
is fed to the \type{ExtractAndMoveMethodExecutor}, whose job is to execute the 
refactoring \see{extractAndMoveMethodExecutor}.


\subsection{Hacking the refactoring undo history}\label{sec:hackingUndoHistory}

As an attempt to make multiple subsequent changes to the workspace appear as a 
single action (i.e. make the undo changes appear as such), I tried to alter 
the undo changes\typeref{org.eclipse.ltk.core.refactoring.Change} in the history 
of the refactorings.  

My first impulse was to remove the, in this case, last two undo changes from the 
undo manager\typeref{org.eclipse.ltk.core.refactoring.IUndoManager} for the 
\name{Eclipse} refactorings, and then add them to a composite 
change\typeref{org.eclipse.ltk.core.refactoring.CompositeChange} that could be 
added back to the manager. The interface of the undo manager does not offer a 
way to remove/pop the last added undo change, so a possible solution could be to 
decorate\citing{designPatterns} the undo manager, to intercept and collect the 
undo changes before delegating to the \method{addUndo} 
method\methodref{org.eclipse.ltk.core.refactoring.IUndoManager}{addUndo} of the 
manager. Instead of giving it the intended undo change, a null change could be 
given to prevent it from making any changes if run. Then one could let the 
collected undo changes form a composite change to be added to the manager.

There is a technical challenge with this approach, and it relates to the undo 
manager, and the concrete implementation 
\typewithref{org.eclipse.ltk.internal.core.refactoring}{UndoManager2}.  This 
implementation is designed in a way that it is not possible to just add an undo 
change, you have to do it in the context of an active 
operation\typeref{org.eclipse.core.commands.operations.TriggeredOperations}.  
One could imagine that it might be possible to trick the undo manager into 
believing that you are doing a real change, by executing a refactoring that is 
returning a kind of null change that is returning our composite change of undo 
refactorings when it is performed. But this is not the way to go.

Apart from the technical problems with this solution, there is a functional 
problem: If it all had worked out as planned, this would leave the undo history 
in a dirty state, with multiple empty undo operations corresponding to each of 
the sequentially executed refactoring operations, followed by a composite undo 
change corresponding to an empty change of the workspace for rounding of our 
composite refactoring. The solution to this particular problem could be to 
intercept the registration of the intermediate changes in the undo manager, and 
only register the last empty change.

Unfortunately, not everything works as desired with this solution. The grouping 
of the undo changes into the composite change does not make the undo operation 
appear as an atomic operation. The undo operation is still split up into 
separate undo actions, corresponding to the changes done by their originating
refactorings. And in addition, the undo actions have to be performed separate in 
all the editors involved. This makes it no solution at all, but a step toward 
something worse.

There might be a solution to this problem, but it remains to be found. The 
design of the refactoring undo management is partly to be blamed for this, as
it is too complex to be easily manipulated.

\section{Benchmarking}\label{sec:benchmarking}
This part of the master's project is located in the \name{Eclipse} project 
\code{no.uio.ifi.refaktor.benchmark}. The purpose of it is to run the equivalent 
of the \type{SearchBasedExtractAndMoveMethodChanger} 
\see{searchBasedExtractAndMoveMethodChanger} over a larger software project, 
both to test its robustness but also its effect on different software metrics.

\subsection{The benchmark setup}
The benchmark itself is set up as a \name{JUnit} test case. This is a convenient 
setup, and utilizes the \name{JUnit Plugin Test Launcher}. This provides us with 
a fully functional \name{Eclipse} workbench. Most importantly, this gives us 
access to the Java Model of \name{Eclipse} \see{javaModel}.

\subsubsection{The ProjectImporter}
The Java project that is going to be used as the data for the benchmark, must be 
imported into the JUnit workspace. This is done by the 
\typewithref{no.uio.ifi.refaktor.benchmark}{ProjectImporter}. The importer 
requires the absolute path to the project description file. This file is named 
\code{.project} and is located at the root of the project directory.

The project description is loaded to find the name of the project to be 
imported. The project that shall be the destination for the import is created in 
the workspace, on the base of the name from the description. Then an import 
operation is created, based on both the source and destination information. The 
import operation is run to perform the import.

I have found no simple API call to accomplish what the importer does, which 
tells me that it may not be too many people performing this particular action.  
The solution to the problem was found on \name{Stack 
Overflow}\footnote{\url{https://stackoverflow.com/questions/12401297}}. It 
contains enough dirty details to be considered inconvenient to use, if not 
wrapping it in a class like my \type{ProjectImporter}. One would probably have 
to delve into the source code for the import wizard to find out how the import 
operation works, if no one had already done it.

\subsection{Statistics}
Statistics for the analysis and changes is captured by the 
\typewithref{no.uio.ifi.refaktor.aspects}{StatisticsAspect}. This an 
\emph{aspect} written in \name{AspectJ}.

\subsubsection{AspectJ}
\href{http://eclipse.org/aspectj/}{AspectJ} is an extension to the Java 
language, and facilitates combining aspect-oriented programming with the 
object-oriented programming in Java.

Aspect-oriented programming is a programming paradigm that is meant to isolate 
so-called \emph{cross-cutting concerns} into their own modules. These 
cross-cutting concerns are functionalities that span over multiple classes, but 
may not belong naturally in any of them. It can be functionality that does not 
concern the business logic of an application, and thus may be a burden when 
entangled with parts of the source code it does not really belong. Examples 
include logging, debugging, optimization and security.

Aspects are interacting with other modules by defining advices. The concept of 
an \emph{advice} is known from both aspect-oriented and functional programming. 
It is a function that modifies another function when the latter is run. An 
advice in AspectJ is somewhat similar to a method in Java. It is meant to alter 
the behavior of other methods, and contains a body that is executed when it is 
applied.

An advice can be applied at a defined \emph{pointcut}. A pointcut picks out one 
or more \emph{join points}. A join point is a well-defined point in the 
execution of a program. It can occur when calling a method defined for a 
particular class, when calling all methods with the same name, 
accessing/assigning to a particular field of a given class and so on. An advice 
can be declared to run both before, after returning from a pointcut, when there 
is thrown an exception in the pointcut or after the pointcut either returns or 
throws an exception.  In addition to picking out join points, a pointcut can 
also bind variables from its context, so they can be accessed in the body of an 
advice. An example of a pointcut and an advice is found in 
\myref{lst:aspectjExample}.

\begin{listing}[h]
\begin{minted}{aspectj}
pointcut methodAnalyze(
  SearchBasedExtractAndMoveMethodAnalyzer analyzer) :
    call(* SearchBasedExtractAndMoveMethodAnalyzer.analyze()) 
      && target(analyzer);

after(SearchBasedExtractAndMoveMethodAnalyzer analyzer) : 
    methodAnalyze(analyzer) {
  statistics.methodCount++;
  debugPrintMethodAnalysisProgress(analyzer.method);
}
\end{minted}
\caption{An example of a pointcut named \method{methodAnalyze}, 
and an advice defined to be applied after it has occurred.}
\label{lst:aspectjExample}
\end{listing}

\subsubsection{The Statistics class}
The statistics aspect stores statistical information in an object of type 
\type{Statistics}. As of now, the aspect needs to be initialized at the point in 
time where it is desired that it starts its data gathering. At any point in time 
the statistics aspect can be queried for a snapshot of the current statistics.

The \type{Statistics} class also includes functionality for generating a report 
of its gathered statistics. The report can be given either as a string or it can 
be written to a file.

\subsubsection{Advices}
The statistics aspect contains advices for gathering statistical data from 
different parts of the benchmarking process. It captures statistics from both 
the analysis part and the execution part of the composite \ExtractAndMoveMethod 
refactoring.

For the analysis part, there are advices to count the number of text selections 
analyzed and the number of methods, types, compilation units and packages 
analyzed. There are also advices that counts for how many of the methods there 
are found a selection that is a candidate for the refactoring, and for how many 
methods there are not.

There exist advices for counting both the successful and unsuccessful executions 
of all the refactorings. Both for the \ExtractMethod and \MoveMethod 
refactorings in isolation, as well as for the combination of them.

\subsection{Optimizations}
When looking for possible optimizations for the benchmarking process, I used the 
\name{VisualVM}\footnote{\url{http://visualvm.java.net/}} \gloss{profiler} for 
the Java Virtual Machine to both profile the application and also to make memory 
dumps of its heap.

\subsubsection{Caching}
When \gloss{profiling} the benchmark process before making any optimizations, it 
early became apparent that the parsing of source code was a place to direct 
attention toward. This discovery was done when only \emph{analyzing} source 
code, before trying to do any \emph{manipulation} of it. Caching of the parsed 
ASTs seemed like the best way to save some time, as expected. With only a simple 
cache of the most recently used AST, the analysis time was speeded up by a 
factor of around 20. This number depends a little upon which type of system the 
analysis is run.

The caching is managed by a cache manager, that now, by default, utilizes the 
not so well known feature of Java called a \emph{soft reference}. Soft 
references are best explained in the context of weak references. A \emph{weak 
reference} is a reference to an object instance that is only guaranteed to 
persist as long as there is a \emph{strong reference} or a soft reference 
referring the same object. If no such reference is found, its referred object is 
garbage collected. A strong reference is basically the same as a regular Java 
reference. A soft reference has the same guarantees as a week reference when it 
comes to its relation to strong references, but it is not necessarily garbage 
collected if there are no strong references to it. A soft reference \emph{may} 
reside in memory as long as the JVM has enough free memory in the heap. A soft 
reference will therefore usually perform better than a weak reference when used 
for simple caching and similar tasks. The way to use a soft/weak reference is to 
as it for its referent. The return value then has to be tested to check that it 
is not \var{null}. For the basic usage of soft references, see 
\myref{lst:softReferenceExample}. For a more thorough explanation of weak 
references in general, see\citing{weakRef2006}.

\begin{listing}[h]
\begin{minted}{java}
// Strong reference
Object strongRef = new Object();

// Soft reference
SoftReference<Object> softRef =
    new SoftReference<Object>(new Object());

// Using the soft reference
Object obj = softRef.get();
if (obj != null) {
    // Use object here
}
\end{minted}
\caption{Showing the basic usage of soft references. Weak references is used the 
  same way. {\footnotesize (The references are part of the \code{java.lang.ref} 
package.)}}
\label{lst:softReferenceExample}
\end{listing}

The cache based on soft references has no limit for how many ASTs it caches. It 
is generally not advisable to keep references to ASTs for prolonged periods of
time, since they are expensive structures to hold on to. For regular plugin
development, \name{Eclipse} recommends not creating more than one AST at a time to 
limit memory consumption. Since the benchmarking has nothing to do with user 
experience, and throughput is everything, these advices are intentionally 
ignored. This means that during the benchmarking process, the target \name{Eclipse} 
application may very well work close to its memory limit for the heap space for 
long periods during the benchmark.

\subsubsection{Candidates stored as mementos}
When performing large scale analysis of source code for finding candidates to 
the \ExtractAndMoveMethod refactoring, memory is an issue. One of the inputs to 
the refactoring is a variable binding. This variable binding indirectly retains 
a whole AST. Since ASTs are large structures, this quickly leads to an 
\type{OutOfMemoryError} if trying to analyze a large project without optimizing 
how we store the candidates' data. This means that the JVM cannot allocate more 
memory for our benchmark, and it exits disgracefully.

A possible solution could be to just allow the JVM to allocate even more memory, 
but this is not a dependable solution. The allocated memory could easily 
supersede the physical memory of a machine, which would make the benchmark go 
really slow.

Thus, the candidates' data must be stored in another format. Therefore, we use 
the \gloss{mementoPattern} to store variable binding information. This is done 
in a way that makes it possible to retrieve a variable binding at a later point.  
The data that is stored to achieve this, is the key to the original variable 
binding. In addition to the key, we know which method and text selection the 
variable is referenced in, so that we can find it by parsing the source code and 
search for it when it is needed.

\subsection{Handling failures}
Failures during the benchmarking process are logged and then ignored. The 
failures are represented in the statistics gathered.


\section{Continuous integration}
The continuous integration server 
\name{Jenkins}\footnote{\url{http://jenkins-ci.org/}} has been set up for the 
project\footnote{A work mostly done by the supervisor.}. It is used as a way to 
run tests and perform code coverage analysis. 

To be able to build the \name{Eclipse} plugins and run tests for them with Jenkins, the 
component assembly project 
\name{Buckminster}\footnote{\url{http://www.eclipse.org/buckminster/}} is used, 
through its plugin for Jenkins. Buckminster provides for a way to specify the 
resources needed for building a project and where and how to find them.  
Buckminster also handles the setup of a target environment to run the tests in.  
All this is needed because the code to build depends on an \name{Eclipse} 
installation with various plugins.

\subsection{Problems with AspectJ}
The Buckminster build worked fine until introducing AspectJ into the project.  
When building projects using AspectJ, there are some additional steps that need
to be performed. First of all, the aspects themselves must be compiled. Then the 
aspects need to be woven with the classes they affect. This demands a process 
that does multiple passes over the source code.

When using AspectJ with \name{Eclipse}, the specialized compilation and the 
weaving can be handled by the \name{AspectJ Development 
Tools}\footnote{\url{https://www.eclipse.org/ajdt/}}. This works all fine, but 
it complicates things when trying to build a project depending on \name{Eclipse} 
plugins outside of \name{Eclipse}. There is supposed to be a way to specify a 
compiler adapter for javac, together with the file extensions for the file types 
it shall operate. The AspectJ compiler adapter is called 
\typewithref{org.aspectj.tools.ant.taskdefs}{Ajc11CompilerAdapter}, and it works 
with files that has the extensions \code{*.java} and \code{*.aj}. I tried to 
setup this in the build properties file for the project containing the aspects, 
but to no avail. The project containing the aspects does not seem to be built at 
all, and the projects that depend on it complain that they cannot find certain 
classes.

I then managed to write an \name{Ant}\footnote{\url{https://ant.apache.org/}} 
build file that utilizes the AspectJ compiler adapter, for the 
\code{no.uio.ifi.refaktor} plugin. The problem was then that it could no longer 
take advantage of the environment set up by Buckminster. The solution to this 
particular problem was of a ``hacky'' nature. It involves exporting the plugin 
dependencies for the project to an Ant build file, and copy the exported path 
into the existing build script. But then the Ant script needs to know where the 
local \name{Eclipse} installation is located. This is no problem when building 
on a local machine, but to utilize the setup done by Buckminster is a problem 
still unsolved. To get the classpath for the build setup correctly, and here 
comes the most ``hacky'' part of the solution, the Ant script has a target for 
copying the classpath elements into a directory relative to the project 
directory and checking it into Git. When no \code{ECLIPSE\_HOME} property is set 
while running Ant, the script uses the copied plugins instead of the ones 
provided by the \name{Eclipse} installation when building the project. This 
obviously creates some problems with maintaining the list of dependencies in the 
Ant file, as well as remembering to copy the plugins every time the list of 
dependencies changes.

The Ant script described above is run by Jenkins before the Buckminster setup 
and build. When setup like this, the Buckminster build succeeds for the projects 
not using AspectJ, and the tests are run as normal. This is all good, but it 
feels a little scary, since the reason for Buckminster not working with AspectJ 
is still unknown.

The problems with building with AspectJ on the Jenkins server lasted for a 
while, before they were solved. This is reflected in the ``Test Result Trend'' 
and ``Code Coverage Trend'' reported by Jenkins.



\chapter{Case studies}\label{ch:caseStudies}

In this chapter I will present a two case studies. This is done to give an 
impression of how the search-based \ExtractAndMoveMethod refactoring performs 
when giving it a larger project to take on. I will try to answer where it lacks, 
in terms of completeness, as well as showing its effect on refactored source 
code.

The first and primary case, is refactoring source code from the \name{Eclipse 
JDT UI} project. The project is chosen because it is a well-known open-source 
project, still in development, with a large code base that is written by many 
different people over several years. The code is installed in a large number of 
\name{Eclipse} applications worldwide, and many other projects build on the 
Eclipse platform. For a long time, it was even the official IDE for Android 
development. All this means that Eclipse must be seen as a good representative 
for professionally written Java source code. It is also the home for most of the 
JDT refactoring code.

For the second case, the \ExtractAndMoveMethod refactoring is fed the 
\code{no.uio.ifi.refaktor} project. This is done as a variation of the 
``dogfooding'' methodology.

\section{The tools}
For conducting these experiments, three software tools are used. Two of the 
tools both use Eclipse as their platform. The first is our own tool, described 
in \myref{sec:benchmarking}, written to be able to run the \ExtractAndMoveMethod 
refactoring as a batch process. It analyzes and refactors all the methods of a 
project in sequence. The second is JUnit, which is used for running the 
project's own unit tests on the target code both before and after it is 
refactored. The last tool that is used is a code quality management tool, called 
\name{SonarQube}. It can be used to perform different tasks for assuring code 
quality, but we are only going to take advantage of one of its main features, 
namely quality profiles.

A quality profile is used to define a set of coding rules that a project is 
supposed to comply with. Failure to following these rules will be recorded as 
so-called ``issues'', marked as having one of several degrees of severities, 
ranging from ``info'' to ``blocker'', where the latter one is the most severe.  
The measurements done for these case studies are therefore not presented as 
fine-grained software metrics results, but rather as the number of issues for 
each defined rule.  

In its analysis, \name{SonarQube} discriminates between functions and accessors.  
Accessors are methods that are recognized as setters or getters. 

In addition to the coding rules defined through quality profiles, 
\name{SonarQube} calculates the complexity of source code. The metric that is 
used is cyclomatic complexity, developed by Thomas J. McCabe in 
1976\citing{mccabeComplexity1976}. In this metric, functions have an initial 
complexity of 1, and whenever the control flow of a function splits, the 
complexity increases by
one\footnote{\url{http://docs.codehaus.org/display/SONAR/Metric+definitions}}. 
Accessors are not counted in the complexity analysis. 

Specifications for the computer used during the experiments are shown in 
\myref{tab:experimentComputerSpecs}.

\begin{table}[htb]
  \caption{Specifications for experiment computer.}
  \label{tab:experimentComputerSpecs}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{0.35}R{1.65}@{}}
    \toprule
    \spancols{2}{Hardware} \\
    \midrule
    Model & Lenovo ThinkPad Edge S430 \\
    Processor & Intel\textregistered{} Core\texttrademark{} 
    i5-3210M\linebreak[4] (2.5 GHz/3.1 GHz (turbo), 
    2 cores, 4 threads, 3 MB Cache) \\
    Memory & 8 GB DDR3 1600 MHz \\
    Storage & 500 GB HDD (7200 RPM) + 16 GB SSD Cache for Lenovo Hard Disk Drive 
    Performance Booster \\
    \midrule
    \spancols{2}{Operating system} \\
    \midrule
    Distribution & Ubuntu 12.10 \\
    Kernel & Linux 3.5.0-49-generic (x86\_64) \\
    
    \bottomrule
  \end{tabularx}
\end{table}


\section{The \name{SonarQube} quality profile}
The quality profile that is used with \name{SonarQube} in these case studies has got 
the name \name{IFI Refaktor Case Study} (version 6). The rules defined in the 
profile are chosen because they are the available rules found in \name{SonarQube} that 
measures complexity and coupling. Now follows a description of the rules in the 
quality profile. The values that are set for these rules are listed in 
\myref{tab:qualityProfile1}.

\begin{description}
  \item[Avoid too complex class] is a rule that measures cyclomatic complexity 
    for every statement in the body of a class, except for setters and getter.  
    The threshold value set is its default value of 200.

  \item[Classes should not be coupled to too many other classes ] is a rule that 
    measures how many other classes a class depends upon. It does not count the 
    dependencies of nested classes. It is meant to promote the Single 
    Responsibility Principle. The metric for the rule resembles the CBO metric 
    that is described in \myref{sec:CBO}, but is only considering outgoing 
    dependencies. The max value for the rule is chosen on the basis of an 
    empirical study by Raed Shatnawi, which concludes that the number 9 is the 
    most useful threshold for the CBO metric\citing{shatnawiQuantitative2010}.  
    This study is also performed on Eclipse source code, so this threshold value 
    should be particularly well suited for the Eclipse JDT UI case in this 
    chapter.

  \item[Control flow statements \ldots{} should not be nested too deeply] is 
    a rule that is meant to counter ``Spaghetti code''. It measures the nesting 
    level of \emph{if}, \emph{for}, \emph{while}, \emph{switch} and \emph{try} 
    statements.  The nesting levels start at 1. The max value set is its default 
    value of 3.

  \item[Methods should not be too complex] is a rule that measures cyclomatic 
    complexity the same way as the ``Avoid too complex class'' rule. The max 
    value used is 10, which ``seems like a reasonable, but not magical, upper 
    limit``\citing{mccabeComplexity1976}.

  \item[Methods should not have too many lines] is a rule that simply measures 
    the number of lines in methods. A threshold value of 20 is used for this 
    metric. This is based on my own subjective opinions, as the default value of 
    100 describes method bodies that do not even fit on most screens.

  \item[NPath Complexity] is a rule that measures the number of possible 
    execution paths through a function. The value used is the default value of 
    200, which seems like a recognized threshold for this metric.

  \item[Too many methods] is a rule that measures the number of methods in a 
    class. The threshold value used is the default value of 10. 

\end{description}


\begin{table}[htb]
  \caption{The \name{IFI Refaktor Case Study} quality profile (version 6).}
  \label{tab:qualityProfile1}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1.5}R{0.5}@{}}
    \toprule
    \textbf{Rule} & \textbf{Max value} \\
    \midrule
    Avoid too complex class & 200 \\
    Classes should not be coupled to too many other classes (Single 
    Responsibility Principle) & 9 \\
    Control flow statements \ldots{} should not be nested too deeply & 
    3 \\
    Methods should not be too complex & 10 \\
    Methods should not have too many lines & 20 \\
    NPath Complexity & 200 \\
    Too many methods & 10 \\
    
    \bottomrule
  \end{tabularx}
\end{table}

\section{The input}
A precondition for the source code that is going to be the target for a series 
of \ExtractAndMoveMethod refactorings, is that it is organized as an Eclipse 
project. It is also assumed that the code is free from compilation errors.

\section{The experiment}
For a given project, the first job that is done, is to refactor its source code. 
The refactoring batch job produces three things: The refactored project, 
statistics gathered during the execution of the series of refactorings, and an 
error log describing any errors happening during this execution. See 
\myref{sec:benchmarking} for more information about how the refactorings are 
performed. 

After the refactoring process is done, the before- and after-code is analyzed 
with \name{SonarQube}. The analysis results are then stored in a database and 
displayed through a \name{SonarQube} server with a web interface.

The before- and after-code is also tested with their own unit tests. This is 
done to discover any changes in the semantic behavior of the refactored code, 
within the limits of these tests.

\section{Case 1: The Eclipse JDT UI project}
This case is the ultimate test for our \ExtractAndMoveMethod refactoring. The 
target source code is massive. With its over 300,000 lines of code\footnote{For 
  all uses of ``lines of code'' we follow the definition from \name{SonarQube}.  
LOC = the number of physical lines containing a character which is neither 
whitespace or part of a comment.} and more than 25,000 methods, it is a 
formidable task to perform automated changes on it. There should be plenty of 
situations where things can go wrong, and, as we shall see later, they do. 

I will start by presenting some statistics from the refactoring execution, 
before I pick apart the \name{SonarQube} analysis and conclude by commenting on 
the results from the unit tests. The configuration for the experiment is 
specified in \myref{tab:configurationCase1}.

\begin{table}[htb]
  \caption{Configuration for Case 1.}
  \label{tab:configurationCase1}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{0.67}L{1.33}@{}}
    \toprule
    \spancols{2}{Benchmark data} \\
    \midrule
    Launch configuration & CaseStudy.launch \\
    Project & no.uio.ifi.refaktor.benchmark \\
    Repository & gitolite@git.uio.no:ifi-stolz-refaktor \\
    Commit & 43c16c04520746edd75f8dc2a1935781d3d9de6c \\
    \midrule
    \spancols{2}{Input data} \\
    \midrule
    Project & org.eclipse.jdt.ui \\
    Repository & git://git.eclipse.org/gitroot/jdt/eclipse.jdt.ui.git \\
    Commit & f218388fea6d4ec1da7ce22432726c244888bb6b \\
    Branch & R3\_8\_maintenance \\
    Tests suites & org.eclipse.jdt.ui.tests.AutomatedSuite, 
    org.eclipse.jdt.ui.tests.refactoring.all.\-AllAllRefactoringTests \\
    
    \bottomrule
  \end{tabularx}
\end{table}
\subsection{Statistics}
The statistics gathered during the refactoring execution is presented in 
\myref{tab:case1Statistics}.

\begin{table}[htb]
  \caption{Statistics after batch refactoring the Eclipse JDT UI project with 
  the \ExtractAndMoveMethod refactoring.}
  \label{tab:case1Statistics}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1.5}R{0.5}@{}}
    \toprule
    \spancols{2}{Time used} \\
    \midrule
    Total time & 98m38s \\
    Analysis time & 14m41s (15\%) \\
    Change time & 74m20s (75\%) \\
    Miscellaneous tasks & 9m37s (10\%) \\
    \midrule
    \spancols{2}{Numbers of each type of entity analyzed} \\
    \midrule
    Packages & 110 \\
    Compilation units & 2,097 \\
    Types & 3,152 \\
    Methods & 27,667 \\
    Text selections & 591,500 \\
    \midrule
    \spancols{2}{Numbers for \ExtractAndMoveMethod refactoring candidates} \\
    \midrule
    Methods chosen as candidates & 2,552 \\
    Methods NOT chosen as candidates & 25,115 \\
    Candidate selections (multiple per method) & 36,843 \\
    \midrule
    \spancols{2}{\ExtractAndMoveMethod refactorings executed} \\
    \midrule
    Fully executed & 2,469 \\
    Not fully executed & 83 \\
    Total attempts & 2,552 \\
    \midrule
    \spancols{2}{Primitive refactorings executed} \\
    \spancols{2}{\small \ExtractMethod refactorings} \\
    \midrule
    Performed & 2,483 \\
    Not performed & 69 \\
    Total attempts & 2,552 \\
    \midrule
    \spancols{2}{\small \MoveMethod refactorings} \\
    \midrule
    Performed & 2469 \\
    Not performed & 14 \\
    Total attempts & 2,483 \\

    \bottomrule
  \end{tabularx}
\end{table}

\subsubsection{Execution time}\label{sec:case1ExecutionTime}
I consider the total execution time of approximately 1.5 hours, on a mid-level 
laptop computer, as being acceptable. It clearly makes the batch process 
unsuitable for doing any on-demand analysis or changes, but it is good enough 
for running periodic jobs, like over-night analysis. In comparison, the 
SonarQube analysis for the same project consumes about the same amount of time. 

As the statistics show, 75\% of the total time goes into making the actual code 
changes.  The time consumers are here the primitive \ExtractMethod and 
\MoveMethod refactorings. Included in the change time is the parsing and 
precondition checking done by the refactorings, as well as textual changes done 
to files on disk. All this parsing and disk access is time-consuming, and 
constitutes a large part of the change time.

The pure analysis time, which is the time used on finding suitable refactoring 
candidates, only makes up for 15\% of the total time consumed. This includes 
analyzing almost 600,000 text selections, while the number of attempted 
executions of the \ExtractAndMoveMethod refactoring is only about 2,500. So the 
number of executed primitive refactorings is approximately 5,000. Assuming the 
time used on miscellaneous tasks are used mostly for parsing source code for the 
analysis, we can say that the time used for analyzing code is at most 25\% of 
the total time.  This means that for every primitive refactoring executed, we 
can analyze about 360 text selections. So, with an average of about 21 text 
selections per method, it is reasonable to say that we can analyze over 15 
methods in the time it takes to perform a primitive refactoring.

\subsubsection{Refactoring candidates}
Out of the 27,667 methods that were analyzed, 2,552 methods contained selections 
that were considered candidates for the \ExtractAndMoveMethod refactoring. This 
is roughly 9\% off the methods in the project. These 9\% of the methods had on 
average 14.4 text selections that were considered possible refactoring 
candidates.

\subsubsection{Executed refactorings}
2,469 out of 2,552 attempts on executing the \ExtractAndMoveMethod refactoring 
were successful, giving a success rate of 96.7\%. The failure rate of 3.3\% 
stems from situations where the analysis finds a candidate selection, but the 
change execution fails. This failure could be an exception that was thrown, and 
the refactoring aborts. It could also be the precondition checking for one of 
the primitive refactorings that gives us an error status, meaning that if the 
refactoring proceeds, the code will contain compilation errors afterwards, 
forcing the composite refactoring to abort.

Out of the 2,552 \ExtractMethod refactorings that were attempted executed, 69 of 
them failed. This gives a failure rate of 2.7\% for the primitive refactoring.  
In comparison, the \MoveMethod refactoring had a failure rate of 0.6 \% of the 
2,483 attempts on the refactoring.

The failure rates for the refactorings are not that bad, if we also take into 
account that the pre-refactoring analysis is incomplete 
\see{par:incompleteness}.

\subsection{\name{SonarQube} analysis}
Results from the \name{SonarQube} analysis are shown in 
\myref{tab:case1ResultsProfile1}.

\begin{table}[htb]
  \caption{Results for analyzing the Eclipse JDT UI project, before and after 
    the refactoring, with \name{SonarQube} and the \name{IFI Refaktor Case Study} 
  quality profile.  (Bold numbers are better.)}
  \label{tab:case1ResultsProfile1}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1.5}R{0.25}R{0.25}@{}}
    \toprule
    \textnormal{Number of issues for each rule} & Before & After \\
    \midrule
    Avoid too complex class & 81 & \textbf{79} \\
    Classes should not be coupled to too many other classes (Single 
    Responsibility Principle) & \textbf{1,098} & 1,199 \\
    Control flow statements \ldots{} should not be nested too deeply & 1,375 & 
    \textbf{1,285} \\
    Methods should not be too complex & 1,518 & \textbf{1,452} \\
    Methods should not have too many lines & 3,396 & \textbf{3,291} \\
    NPath Complexity & 348 & \textbf{329} \\
    Too many methods & \textbf{454} & 520 \\
    \midrule
    Total number of issues & 8,270 & \textbf{8,155} \\
    \midrule
    \midrule
    \spancols{3}{Complexity} \\
    \midrule
    Per function & 3.6 & \textbf{3.3} \\
    Per class & \textbf{29.5} & 30.4 \\
    Per file & \textbf{44.0} & 45.3 \\
    \midrule
    Total complexity & \textbf{84,765} & 87,257 \\
    \midrule
    \midrule
    \spancols{3}{Numbers of each type of entity analyzed} \\
    \midrule
    Files & 1,926 & 1,926 \\
    Classes & 2,875 & 2,875 \\
    Functions & 23,744 & 26,332 \\
    Accessors & 1,296 & 1,019 \\
    Statements & 162,768 & 165,145 \\
    Lines of code & 320,941 & 329,112 \\
    \midrule
    Technical debt (in days) & \textbf{1,003.4} & 1,032.7 \\
    \bottomrule
  \end{tabularx}
\end{table}

\subsubsection{Diversity in the number of entities analyzed}
The analysis performed by \name{SonarQube} is reporting fewer methods than found 
by the pre-refactoring analysis. \name{SonarQube} discriminates between 
functions (methods) and accessors, so the 1,296 accessors play a part in this 
calculation.  \name{SonarQube} also has the same definition as our plugin when 
it comes to how a class is defined. Therefore it seems like \name{SonarQube} 
misses 277 classes that our plugin handles. This can explain why the {SonarQube} 
report differs from our numbers by approximately 2,500 methods.

\subsubsection{Complexity}
On all complexity rules that works on the method level, the number of issues 
decreases with between 3.1\% and 6.5\% from before to after the refactoring. The 
average complexity of a method decreases from 3.6 to 3.3, which is an 
improvement of about 8.3\%. So, on the method level, the refactoring must be 
said to have a slightly positive impact. This is due to the extraction of a lot 
of methods, making the average method size smaller.

The improvement in complexity on the method level is somewhat traded for 
complexity on the class level. The complexity per class metric is worsened by 
3\% from before to after. The issues for the ``Too many methods'' rule also 
increases by 14.5\%. These numbers indicate that the refactoring makes quite a 
lot of the classes a little more complex overall. This is the expected outcome, 
since the \ExtractAndMoveMethod refactoring introduces almost 2,500 new methods 
into the project.

The only number that can save the refactoring's impact on complexity on the 
class level, is the ``Avoid too complex class'' rule. It improves with 2.5\%, 
thus indicating that the complexity is moderately better distributed between the 
classes after the refactoring than before.

\subsubsection{Coupling}
One of the hopes when starting this project, was to be able to make a 
refactoring that could lower the coupling between classes. Better complexity at 
the method level is a not very unexpected byproduct of dividing methods into 
smaller parts. Lowering the coupling on the other hand, is a far greater task.  
This is also reflected in the results for the only coupling rule defined in the 
\name{SonarQube} quality profile, namely the ``Classes should not be coupled to 
too many
other classes (Single Responsibility Principle)'' rule. 

The number of issues for the coupling rule is 1,098 before the refactoring, and 
1,199 afterwards. This is an increase in issues of 9.2\%. These numbers can be 
interpreted two ways. The first possibility is that our assumptions are wrong, 
and that increasing indirection does not decrease coupling between classes. The 
other possibility is that our analysis and choices of candidate text selections 
are not good enough. I vote for the second possibility. (Voting against the 
public opinion may also be a little bold.)

\subsubsection{An example of what makes the number of coupling issues 
grow}\label{sec:case1IssuesExample}
\Myref{lst:sonarJDTExampleBefore} shows a portion of the class 
\typewithref{org.eclipse.jdt.ui.actions}{ShowActionGroup} from the JDT UI 
project before it is refactored with the search-based \ExtractAndMoveMethod 
refactoring. Before the refactoring, the \type{ShowActionGroup} class has 12 
outgoing dependencies (reported by \name{SonarQube}).

\begin{listing}[htb]
\begin{minted}[linenos,samepage]{java}
public class ShowActionGroup extends ActionGroup {
  /* ... */
  private void initialize(IWorkbenchSite site,
                          boolean isJavaEditor) {
    fSite= site;
    ISelectionProvider provider= fSite.getSelectionProvider();
    ISelection selection= provider.getSelection();
    fShowInPackagesViewAction.update(selection);
    if (!isJavaEditor) {
      provider.addSelectionChangedListener(
                                   fShowInPackagesViewAction);
    }
  }
}
\end{minted}
\caption{Portion of the \type{ShowActionGroup} class before refactoring.}
\label{lst:sonarJDTExampleBefore}
\end{listing}

During the benchmark process, the search-based \ExtractAndMoveMethod refactoring 
extracts the lines 6 to 12 of the code in \myref{lst:sonarJDTExampleBefore}, and 
moves the new method to the move target, which is the field 
\var{fShowInPackagesViewAction} with type 
\typewithref{org.eclipse.jdt.ui.actions}{ShowInPackageViewAction}. The result is 
shown in \myref{lst:sonarJDTExampleAfter}.

\begin{listing}[htb]
\begin{minted}[linenos,samepage]{java}
public class ShowActionGroup extends ActionGroup {
  /* ... */
  private void initialize(IWorkbenchSite site,
                          boolean isJavaEditor) {
    fSite= site;
    fShowInPackagesViewAction.generated_8019497110545412081(
                                           this, isJavaEditor);
  }
}
\end{minted}

\begin{minted}[linenos,samepage]{java}
public class ShowInPackageViewAction
        extends SelectionDispatchAction {
  /* ... */
  public void generated_8019497110545412081(
      ShowActionGroup showactiongroup, boolean isJavaEditor) {
    ISelectionProvider provider=
                 showactiongroup.fSite.getSelectionProvider();
    ISelection selection= provider.getSelection();
    update(selection);
    if (!isJavaEditor) {
      provider.addSelectionChangedListener(this);
    }
  }
}
\end{minted}
\caption{Portions of the classes \type{ShowActionGroup} and 
\type{ShowInPackageViewAction} after refactoring.}
\label{lst:sonarJDTExampleAfter}
\end{listing}

After the refactoring, the \type{ShowActionGroup} has only 11 outgoing 
dependencies. It no longer depends on the \type{ISelection} type.  So our 
refactoring managed to get rid of one dependency, which is exactly what we 
wanted. The only problem is, that now the \type{ShowInPackageViewAction} class 
has got two new dependencies, in the \type{ISelectionProvider} and the 
\type{ISelection} types.  The bottom line is that we eliminated one dependency, 
but introduced two more, ending up with a program that has more dependencies now 
than when we started.

What can happen in many situations where the \ExtractAndMoveMethod refactoring 
is performed, is that the \MoveMethod refactoring ``drags'' with it references 
to classes that are unknown to the method destination. If the refactoring 
happens to be so lucky that it removes a dependency from one class, it might as 
well introduce a couple of new dependencies to another class, as shown in the 
previous example. In those situations where a destination class does not know 
about the originating class of a moved method, the \MoveMethod refactoring most 
certainly will introduce a dependency.  This is because there is a 
bug\footnote{\href{https://bugs.eclipse.org/bugs/show\_bug.cgi?id=228635}{Eclipse 
Bug 228635 - [move method] unnecessary reference to source}} in the refactoring, 
making it pass an instance of the originating class as a reference to the moved 
method, regardless of whether the reference is used in the method body or not.

There is also the possibility that the heuristics used to find candidate text 
selections are not good enough. There is work to be done with fine-tuning the 
heuristics and to complete the analysis part of this project.  

\subsubsection{Totals}
On the bright side, the total number of issues is lower after the refactoring 
than it was before. Before the refactoring, the total number of issues was
8,270, and after it is 8,155. This is an improvement of 1.4\%.

The down side is that \name{SonarQube} shows that the total cyclomatic 
complexity has increased by 2.9\%, and that the (more questionable) ``technical 
debt'' has increased from 1,003.4 to 1,032.7 days, also a deterioration of 
2.9\%.  Although these numbers are similar, no correlation has been found 
between them.

\subsection{Unit tests}
The tests that have been run for the \name{Eclipse JDT UI} project, are the
test suites specified as the main test suites on the JDT UI wiki page on how to 
contribute to the 
project\footnote{\url{https://wiki.eclipse.org/JDT\_UI/How\_to\_Contribute\#Unit\_Testing}}.  
The results from these tests are shown in \myref{tab:case1UnitTests}.

\begin{table}[htb]
  \caption{Results from the unit tests run for the Eclipse JDT UI project, 
  before and after the refactoring.}
  \label{tab:case1UnitTests}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1}R{0.5}R{0.5}@{}}
    \toprule
    \textnormal{AutomatedSuite} & Before & After \\
    \midrule
    Runs & 2007/2007 & 2007/2007 \\
    Errors & 4 & 565 \\
    Failures & 3 & 5 \\
    \midrule
    \spancols{2}{AllAllRefactoringTests} \\
    \midrule
    Runs & 3815/3816 & 3815/3816 \\
    Errors & 2 & 2257 \\
    Failures & 3 & 0 \\
    \bottomrule
  \end{tabularx}
\end{table}

\subsubsection{Before the refactoring}
Running the tests for the before-code of Eclipse JDT UI yielded 4 errors and 3 
failures for the \type{AutomatedSuite} test suite (2,007 test cases), and 2 
errors and 3 failures for the \type{AllAllRefactoringTests} test suite (3,816 
test cases).  

\subsubsection{After the refactoring}
For the after-code of the Eclipse JDT UI project, Eclipse reports that the 
project contains 322 compilation errors, and a lot of other errors that
follow from these. All of the errors are caused by the \ExtractAndMoveMethod 
refactoring. Had these errors originated from only one bug, it would not have 
been much of a problem, but this is not the case. By only looking at some random 
compilation problems in the refactored code, I came up with at least four 
different bugs \todo{write bug reports?} that caused those problems. I then 
stopped looking for more, since some of the bugs would take more time to fix 
than I could justify using on them at this point. 

One thing that can be said in my defense, is that all the compilation errors 
could have been avoided if the types of situations that cause them were properly 
handled by the primitive refactorings, which again are supplied by the Eclipse 
JDT UI project. All four bugs that I mentioned before are weaknesses of the 
\MoveMethod refactoring. If the primitive refactorings had detected the 
up-coming errors in their precondition checking phase, the refactorings would 
have been aborted, since this is how the \ExtractAndMoveMethod refactoring 
handles such situations. This shows that it is not safe to completely rely upon 
the primitive refactorings to save us if our own pre-refactoring analysis fails 
to detect that a compilation error will be introduced. A problem is that the 
source code analysis done by both the JDT refactorings and our own tool is
incomplete.

Of course, taking into account all possible situations that could lead to 
compilation errors is an immense task. A complete analysis of these situations 
is too big of a problem for this master's project to solve. Looking at it now, 
this comes as no surprise, since the task is obviously also too big for the 
creators of the primitive \MoveMethod refactoring. 

Considering all these problems, it is difficult to know how to interpret the 
unit test results from after refactoring the Eclipse JDT UI. The 
\type{AutomatedSuite} reported 565 errors and 5 failures, which means that 1437, 
or 71.6\%, of the tests still passed. Three of the failures were the same as 
reported before the refactoring took place, so two of them are new. For these 
two cases it is not immediately apparent what makes them behave differently. The 
program is so complex that to analyze it to find this out, we might need more 
powerful methods than just manually analyzing its source code.  This is somewhat 
characteristic for imperative programming: The programs are often hard to 
analyze and understand. 

For the \type{AllAllRefactoringTests} test suite, the three failures are gone, 
but the two errors have grown to 2,257 errors. I will not try to analyze those 
errors.

What I can say at this point, is that it is likely that the 
\ExtractAndMoveMethod refactoring has introduced some unintentional behavioral 
changes. Let us say that the refactoring introduces at least two 
behavior-altering changes for every 2,500 executions.  More than that is 
difficult to say about the behavior-preserving properties of the 
\ExtractAndMoveMethod refactoring, at this point. What is clear, is that it 
would benefit from a strategy for making it perform refactoring in a safer 
manner. 

\subsection{Conclusions}
After automatically analyzing and executing the \ExtractAndMoveMethod 
refactoring for all the methods in the Eclipse JDT UI project, the results do
not look that promising. For this case, the refactoring seems almost unusable as 
it is now. The error rate and measurements tell us this.

The refactoring makes the code a little less complex at the method level. But 
this is merely a side effect of extracting methods. When it comes to the overall 
complexity, it is increased, although it is slightly better spread among the 
classes.

The pre-refactoring analysis of the \ExtractAndMoveMethod refactoring, is 
currently not complete enough to make the refactoring useful. It introduces too 
many errors in the code, and the code may change its behavior. It also remains 
to prove that large-scale refactoring with it can decrease the overall coupling 
between classes, although there are individual examples 
\see{sec:case1IssuesExample}.

On the bright side, the performance of the refactoring process is not that bad.  
It shows that it is possible to make a tool the way we do, if we can make the 
tool do anything useful. As long as the analysis phase is not going to involve 
anything that uses too much disk access, a lot of analysis can be done in a 
reasonable amount of time.

The time used on performing the actual changes excludes a trial and error 
approach with the tools used in this master's project. In a trial and error 
approach, you could for instance be using the primitive refactorings used in 
this project to refactor code, and only then make decisions based on the effect, 
possibly shown by traditional software metrics. The problem with the approach 
taken in this project, compared to a trial and error approach, is that using 
heuristics beforehand is much more complicated. But on the other hand, a trial 
and error approach would still need to face the challenges of producing code 
that does compile without errors. If using refactorings that could produce 
in-memory changes, a trial and error approach could be made more efficient.

\section{Case 2: The \type{no.uio.ifi.refaktor} project}
In this case we will see a form of the ``dogfooding'' methodology used, when 
refactoring our own \type{no.uio.ifi.refaktor} project with the 
\ExtractAndMoveMethod refactoring.

In this case I will try to point out some differences from the first case, and 
how they impact the execution of the benchmark. The refaktor project is 39 times 
smaller than the Eclipse JDT UI project, measured in lines of code. This will 
make things a bit more transparent. It will therefore be interesting to see if 
this case can shed light on any aspect of our project that were lost in the 
larger case 1.

The configuration for the experiment is specified in 
\myref{tab:configurationCase2}.

\begin{table}[htb]
  \caption{Configuration for Case 2.}
  \label{tab:configurationCase2}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{0.67}L{1.33}@{}}
    \toprule
    \spancols{2}{Benchmark data} \\
    \midrule
    Launch configuration & CaseStudyDogfooding.launch \\
    Project & no.uio.ifi.refaktor.benchmark \\
    Repository & gitolite@git.uio.no:ifi-stolz-refaktor \\
    Commit & 43c16c04520746edd75f8dc2a1935781d3d9de6c \\
    \midrule
    \spancols{2}{Input data} \\
    \midrule
    Project & no.uio.ifi.refaktor \\
    Repository & gitolite@git.uio.no:ifi-stolz-refaktor \\
    Commit & 43c16c04520746edd75f8dc2a1935781d3d9de6c \\
    Branch & master \\
    Test configuration & no.uio.ifi.refaktor.tests/ExtractTest.launch \\
    \bottomrule
  \end{tabularx}
\end{table}

\subsection{Statistics}
The statistics gathered during the refactoring execution is presented in 
\myref{tab:case2Statistics}.

\begin{table}[htb]
  \caption{Statistics after batch refactoring the \type{no.uio.ifi.refaktor} 
project with the \ExtractAndMoveMethod refactoring.}
  \label{tab:case2Statistics}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1.5}R{0.5}@{}}
    \toprule
    \spancols{2}{Time used} \\
    \midrule
    Total time & 1m15s \\
    Analysis time & 0m18s (24\%) \\
    Change time & 0m47s (63\%) \\
    Miscellaneous tasks & 0m10s (14\%) \\
    \midrule
    \spancols{2}{Numbers of each type of entity analyzed} \\
    \midrule
    Packages & 33 \\
    Compilation units & 154 \\
    Types & 168 \\
    Methods & 1,070 \\
    Text selections & 8,609 \\
    \midrule
    \spancols{2}{Numbers for \ExtractAndMoveMethod refactoring candidates} \\
    \midrule
    Methods chosen as candidates & 58 \\
    Methods NOT chosen as candidates & 1,012 \\
    Candidate selections (multiple per method) & 227 \\
    \midrule
    \spancols{2}{\ExtractAndMoveMethod refactorings executed} \\
    \midrule
    Fully executed & 53 \\
    Not fully executed & 5 \\
    Total attempts & 58 \\
    \midrule
    \spancols{2}{Primitive refactorings executed} \\
    \spancols{2}{\small \ExtractMethod refactorings} \\
    \midrule
    Performed & 56 \\
    Not performed & 2 \\
    Total attempts & 58 \\
    \midrule
    \spancols{2}{\small \MoveMethod refactorings} \\
    \midrule
    Performed & 53 \\
    Not performed & 3 \\
    Total attempts & 56 \\

    \bottomrule
  \end{tabularx}
\end{table}

\subsubsection{Differences}
There are some differences between the two projects that make them a little 
difficult to compare by performance.

\paragraph{Different complexity.} 
Although the JDT UI project is 39 times greater than the refaktor project in 
terms of lines of code, it is only about 26 times its size measured in number of 
methods. This means that the average method size is smaller in the refaktor 
project than in the JDT project. This is also reflected in the \name{SonarQube} 
report, where the complexity per method for the JDT project is 
3.6, while the refaktor project has a complexity per method of 2.1.

\paragraph{Number of selections per method.}
The analysis for the JDT project processed 21 text selections per method in 
average. This number for the refaktor project is only 8 selections per method 
analyzed. This is a direct consequence of smaller methods.

\paragraph{Different candidates to methods ratio.} 
The differences in how the projects are factored are also reflected in the 
ratios for how many methods that are chosen as candidates compared to the total 
number of methods analyzed. For the JDT project, 9\% of the methods were 
considered to be candidates, while for the refaktor project, only 5\% of the 
methods were chosen.

\paragraph{The average number of possible candidate selection.} 
For the methods that are chosen as candidates, the average number of possible 
candidate selections for these methods differ quite much. For the JDT project, 
the number of possible candidate selections for these methods was 14.44 
selections per method, while the candidate methods in the refaktor project had 
only 3.91 candidate selections to choose from, in average.

\subsubsection{Execution time}
The differences in complexity, and the different candidate methods to total 
number of methods ratios, is shown in the distributions of the execution times.  
For the JDT project, 75\% of the total time was used on the actual changes, 
while for the refaktor project, this number was only 63\%.

For the JDT project, the benchmark used on average 0.21 seconds per method in 
the project, while for the refaktor project it used only 0.07 seconds per 
method. So the process used 3 times as much time per method for the JDT project 
than for the refaktor project.

While the JDT project is 39 times larger than the refaktor project measured in 
lines of code, the benchmark used about 79 times as long time on it than for the 
refaktor project. Relatively, this is about twice as long.

Since the details of these execution times are not that relevant to this 
master's project, only their magnitude, I will leave them here.

\subsubsection{Executed refactorings}
For the composite \ExtractAndMoveMethod refactoring performed in case 2, 53 
successful attempts out of 58 gives a success rate of 91.4\%. This is 5.3 
percentage points worse than for the first case.

\subsection{\name{SonarQube} analysis}
Results from the \name{SonarQube} analysis are shown in 
\myref{tab:case2ResultsProfile1}.

Not much is to be said about these results. The trends in complexity and 
coupling are the same. We end up a little worse after the refactoring process 
than before.

\begin{table}[htb]
  \caption{Results for analyzing the \var{no.uio.ifi.refaktor} project, before 
  and after the refactoring, with \name{SonarQube} and the \name{IFI Refaktor 
  Case Study} quality profile.  (Bold numbers are better.)}
  \label{tab:case2ResultsProfile1}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1.5}R{0.25}R{0.25}@{}}
    \toprule
    \textnormal{Number of issues for each rule} & Before & After \\
    \midrule
    Avoid too complex class & 1 & 1 \\
    Classes should not be coupled to too many other classes (Single 
    Responsibility Principle) & \textbf{29} & 34 \\
    Control flow statements \ldots{} should not be nested too deeply & 24 & 
    \textbf{21} \\
    Methods should not be too complex & 17 & \textbf{15} \\
    Methods should not have too many lines & 41 & \textbf{40} \\
    NPath Complexity & 3 & 3 \\
    Too many methods & \textbf{13} & 15 \\
    \midrule
    Total number of issues & \textbf{128} & 129 \\
    \midrule
    \midrule
    \spancols{3}{Complexity} \\
    \midrule
    Per function & 2.1 & 2.1 \\
    Per class & \textbf{12.5} & 12.9 \\
    Per file & \textbf{13.8} & 14.2 \\
    \midrule
    Total complexity & \textbf{2,089} & 2,148 \\
    \midrule
    \midrule
    \spancols{3}{Numbers of each type of entity analyzed} \\
    \midrule
    Files & 151 & 151 \\
    Classes & 167 & 167 \\
    Functions & 987 & 1,045 \\
    Accessors & 35 & 30 \\
    Statements & 3,355 & 3,416 \\
    Lines of code & 8,238 & 8,460 \\
    \midrule
    Technical debt (in days) & \textbf{19.0} & 20.7 \\
    \bottomrule
  \end{tabularx}
\end{table}

\subsection{Unit tests}
The tests used for this case are the same that has been developed throughout 
this master's project.

The code that was refactored for this case suffered from some of the problems 
discovered in the first case. This means that the after-code for this case also 
contained compilation errors, but they were not as many. The code contained only 
6 errors that made the code not compile.

All of the six errors originated from the same bug. The bug arises in a
situation where a class instance creation is moved between packages, and the 
class for the instance is package-private.  The \MoveMethod refactoring does not 
detect that there will be a visibility problem, and neither does it promote the 
package-private class to be public.

Since the errors in the refactored refaktor code were easy to fix manually, I 
corrected them and ran the unit tests as planned. The unit test results are 
shown in \myref{tab:case2UnitTests}. Before the refactoring, all tests passed.  
All tests also passed after the refactoring, with the six error corrections.  
Since the corrections done are not of a kind that could make the behavior of the 
program change, it is likely that the refactorings done to the 
\type{no.uio.ifi.refaktor} project did not change its behavior. This is also 
supported by the informal experiment presented next.

\begin{table}[htb]
  \caption{Results from the unit tests run for the \type{no.uio.ifi.refaktor} 
project, before and after the refactoring (with 6 corrections done to the 
refactored code).}
  \label{tab:case2UnitTests}
  \centering
  \begin{tabularx}{\textwidth}{@{}>{\bfseries}L{1}R{0.5}R{0.5}@{}}
    \toprule
    & Before & After \\
    \midrule
    Runs & 148/148 & 148/148 \\
    Errors & 0 & 0 \\
    Failures & 0 & 0 \\
    \bottomrule
  \end{tabularx}
\end{table}

\subsection{An additional experiment}
To complete the task of ``eating my own dog food'', I conducted an experiment 
where I used the refactored version of the \type{no.uio.ifi.refaktor} project, 
with the corrections, to again refaktor ``itself''.  

The experiment produced code containing the same six errors as after the 
previous experiment.  I also compared the after-code from the two experiments 
with a diff-tool. The only differences found were different method names. This 
is expected, since the method names are randomly generated by the 
\ExtractAndMoveMethod refactoring.

The outcome of this simple experiment makes me more confident that the 
\ExtractAndMoveMethod refactoring made only behavior-preserving changes to the 
\type{no.uio.ifi.refaktor} project, apart from the compilation errors.

\subsection{Conclusions}
The differences in complexity between the Eclipse JDT UI project and the 
\type{no.uio.ifi.refaktor} project, clearly influenced the differences in their 
execution times. This is mostly because fewer of the methods were chosen to be 
refactored for the refaktor project than for the JDT project. This makes it 
difficult to know if there are any severe performance penalties associated with 
refactoring on a large project compared to a small one.

The trends in the \name{SonarQube} analysis are the same for this case as for 
the previous one. This gives more confidence in the these results.

By refactoring our own code and using it again to refactor our code, we showed 
that it is possible to write an automated composite refactoring that works for 
many cases. That it probably did not alter the behavior of a smaller project 
shows us nothing more than that though, and might just be a coincidence. 

\section{Threats to validity}
\todoin{Tool not fine grained enough.}
\todoin{Only one large project.}
\todoin{Only performing each experiment once. (Performance.)}
\todoin{SonarQube analysis does only show number issues. The overall coupling 
may still be lowered.}
\todoin{Could need better metric, like ``the number of references``.}


\chapter{Conclusions and future work}\label{ch:conclusions}
This chapter will conclude this master's thesis. I will try to give justified 
answers to the research questions posed \see{sec:researchQuestions} and present 
some future work that could be done to take this project to the next level.

\section{Conclusions}
Some of the motivation for this thesis was to create a fully automated composite 
refactoring that could be used to make program source code better in terms of 
coupling between classes. Earlier, in \mysimpleref{sec:CBO}, it was shown that a 
composition of the \ExtractMethod and the \MoveMethod refactorings reduces the 
coupling between two classes in an ideal situation. The Eclipse JDT plugin 
implements both these refactorings, and also provides a framework for analyzing 
source code, so it was considered a suitable tool to build upon for our project.

The search-based \ExtractAndMoveMethod refactoring was created by utilizing the 
analysis and refactoring support of Eclipse, and a small framework was built
for executing large scale refactoring with it. The refactoring was set up to 
analyze and execute changes on the Eclipse JDT UI project. Statistics was 
gathered during this process and the resulting code was analyzed through 
SonarQube. The project's own unit tests were also performed to find out whether 
our refactoring introduces any behavior-altering changes in the code it 
refactors.

\paragraph{Answering the main research question.}
The first and greatest challenge was to find out if the \ExtractAndMoveMethod 
refactoring could be automated, in all tasks ranging from analysis to executing 
changes. It is now confirmed that this can be done, since it has been 
implemented as a part of the work done for this project. It has also been shown 
that the refactoring can be used to refactor large code bases, through the case 
study done on the Eclipse JDT UI project.

Asking whether the existing Eclipse refactorings are well suited for this task 
is another question. The refactorings provided by the JDT UI project are clearly 
not meant to be combined in any way. The preconditions for one refactoring are 
not always easily retrievable after the execution of another. Also, the 
refactorings are all assuming that the code they are going to refactor is 
textualized.  This means that the source code must be parsed between the 
executions of each refactoring. Another problem with this dependency on textual 
changes, is that you cannot make a composition of two refactorings appear as one 
change if the two refactorings' changes overlap. This will make the undo-history 
of the composite refactoring show two changes instead of one, and is not nice 
for usability if the refactoring would be used as an on-demand refactoring in an 
IDE.

Apart from the problems with implementing the actual refactoring, the analysis 
framework is quite nicely solved in Eclipse. The AST generated when parsing 
source code, supports using visitors to traverse it, and this works without 
problems.

\paragraph{Is the refactoring efficient enough?}
Since we have concluded that the search-based \ExtractAndMoveMethod refactoring 
is not suitable for on-demand large-scale refactoring, but may be put to better 
use as a kind of analysis tool, superb performance is not crucial. In 
\myref{sec:case1ExecutionTime} we conclude that the refactoring performs well 
enough for this purpose. If performed on demand for a single method, the 
performance of the \ExtractAndMoveMethod refactoring is not an issue.

\paragraph{What about breaking the source code?}
The case studies shows that our safety measures, which rely on the precondition 
checking of the existing primitive refactorings, are not good enough in 
practice.  If we were going to assure that code we refactor compiles, we would 
need to consider all possible situations where the refactoring could fail, and 
search for them in our analysis. It is an open question if this is even 
feasible. Our analysis is incomplete, and so are the analyses for the 
\ExtractMethod and \MoveMethod refactorings.

Our refactoring does not take any precautions to preserve behavior. A few 
running and failing unit test for the JDT UI project after the refactoring 
indicate that our refactoring causes some changes to the way a program behaves.

\paragraph{Is the quality of the code improved?}
For coupling, there is no evidence that the refactoring improves the quality of 
source code. Shall we believe the SonarQube analysis from the case studies, our 
refactoring makes classes more coupled after the refactoring than before, in the 
general case. Examples exist where the \ExtractAndMoveMethod refactoring 
improves coupling. The problem is that it introduces too many dependencies 
overall \see{sec:case1IssuesExample}. The essence is that our analysis and 
heuristics for finding the best candidates for the refactoring are not adequate.

\paragraph{Is the refactoring useful?}
In its present state, the refactoring cannot be said to be very useful. It 
generates too many compilation errors for it to fall into that category. On the 
other hand, if the problems with the search-based \ExtractAndMoveMethod 
refactoring were to be solved, it could be put to use in some situations.

If the refactoring was perfected, it could of course be used as a regular 
on-demand automated refactoring on a per method base (or per class, package or 
project).

As it is now, the refactoring is not well suited for performing unattended 
refactoring. But if we could find a way to filter out the changes that create 
compilation errors, we could use the refactoring to look for improvement points 
in a software project. This process could for instance be scheduled to run at 
regular intervals, possibly after a nightly build or the like.  Then the results 
could be made available, and an administrator could be set to review them and 
choose whether or not they should be performed.

\section{Future work}
An important part that is missing for making the search-based 
\ExtractAndMoveMethod refactoring more usable, is to complete the 
pre-refactoring analysis of the source code, to make sure that the refactoring 
does not introduce compilation errors when it is performed.

The first point of making the static analysis complete, brings up the next 
question: Is it feasible to complete such an analysis? And can this feasibility 
be proven, or disproved?

Another shortcoming of this project is that we have no strategy for assuring 
safety when refactoring, so a program may end up behaving differently after a 
refactoring than it behaved before. One approach toward safer refactorings is
mentioned in \myref{sec:saferRefactoringTests}, and includes generating tests 
for the refactored code. Another approach that can be considered for making 
refactorings safer is part of the original thesis proposal for this thesis, 
which diverged somewhat from the original proposal. The proposal is about 
detecting behavioral changes during refactoring, and the work done in this 
thesis can be used as a basis if one would like to engage in that proposal. The 
proposed solution for exposing behavioral changes, is to insert assertions into 
source code when refactoring it. For the example in 
\myref{lst:correctnessExtractAndMoveResult}, which is the result of a 
refactoring, it is suggested that we insert an assert statement between lines 9 
and 10. In the example, the assert statement would be 
\mint{java}|assert c.x == this;| which would discover the aliasing problems of 
the example.

The final important improvement that I would suggest making to this project, is 
to refine the heuristics that are used to find suitable refactoring candidates.  
This effort should in particular be directed toward making the heuristics choose 
candidates that do not introduce new dependencies to their destination classes.


\appendix


\chapter{Eclipse bugs submitted}
\newcommand{\submittedBugReport}[1]{The submitted bug report can be found on 
  \url{#1}.}

\section{Eclipse bug 420726: Code is broken when moving a method that is 
assigning to the parameter that is also the move 
destination}\label{eclipse_bug_420726}
This bug
was found when analyzing what kinds of names that were to be considered as 
\emph{unfixes} \see{sec:unfixes}.

\paragraph{The bug.}
The bug emerges when trying to move a method from one class to another, and when 
the target for the move (must be a variable, local or field) is both a parameter 
variable and also is assigned to within the method body. \name{Eclipse} allows this to 
happen, although it is the sure path to a compilation error. This is because we 
would then have an assignment to a \var{this} expression, which is not allowed 
in Java. 
\submittedBugReport{https://bugs.eclipse.org/bugs/show\_bug.cgi?id=420726}  

\paragraph{The solution.}
The solution to this problem is to add all simple names that are assigned to in 
a method body to the set of unfixes.

\section{Eclipse bug 429416: IAE when moving method from anonymous 
class}\label{eclipse_bug_429416}
I discovered
this bug during a batch change on the \type{org.eclipse.jdt.ui} project.

\paragraph{The bug.}
This bug surfaces when trying to use the \refa{Move Method} refactoring to move a 
method from an anonymous class to another class. This happens both for my 
simulation as well as in \name{Eclipse}, through the user interface. It only occurs 
when \name{Eclipse} analyzes the program and finds it necessary to pass an 
instance of the originating class as a parameter to the moved method. I.e. it 
wants to pass a \var{this} expression. The execution ends in an 
\typewithref{java.lang}{IllegalArgumentException} in 
\typewithref{org.eclipse.jdt.core.dom}{SimpleName} and its 
\method{setIdentifier(String)} method. The simple name is attempted created in 
the method
\methodwithref{org.eclipse.jdt.internal.corext.refactoring.structure.\\MoveInstanceMethodProcessor}{createInlinedMethodInvocation} 
so the \type{MoveInstanceMethodProcessor} was early a clear suspect.

The \method{createInlinedMethodInvocation} is the method that creates a method 
invocation where the previous invocation to the method that was moved was 
located. From its code it can be read that when a \var{this} expression is going 
to be passed in to the invocation, it shall be qualified with the name of the 
original method's declaring class, if the declaring class is either an anonymous 
class or a member class. The problem with this, is that an anonymous class does 
not have a name, hence the term \emph{anonymous} class! Therefore, when its 
name, an empty string, is passed into 
\methodwithref{org.eclipse.jdt.core.dom.AST}{newSimpleName} it all ends in an 
\type{IllegalArgumentException}.  
\submittedBugReport{https://bugs.eclipse.org/bugs/show\_bug.cgi?id=429416} 

\paragraph{How I solved the problem.}
Since the \type{MoveInstanceMethodProcessor} is instantiated in the 
\typewithref{no.uio.ifi.refaktor.change.executors}{MoveMethod\-RefactoringExecutor}, 
and only need to be a 
\typewithref{org.eclipse.ltk.core.refactoring.participants}{MoveProcessor}, I 
was able to copy the code for the original move processor and modify it so that 
it works better for me. It is now called 
\typewithref{no.uio.ifi.refaktor.change.processors}{ModifiedMoveInstanceMethodProcessor}.  
The only modification done (in addition to some imports and suppression of 
warnings), is in the \method{createInlinedMethodInvocation}. When the declaring 
class of the method to move is anonymous, the \var{this} expression in the 
parameter list is not qualified with the declaring class' (empty) name.

\section{Eclipse bug 429954: Extracting statement with reference to local type 
breaks code}\label{eclipse_bug_429954}
The bug was discovered when doing some changes to the way unfixes is computed.

\paragraph{The bug.}
The problem is that \name{Eclipse} is allowing selections that references variables of 
local types to be extracted. When this happens the code is broken, since the 
extracted method must take a parameter of a local type that is not in the 
methods scope. The problem is illustrated in 
\myref{lst:extractMethodLocalClass}, but there in another setting.  
\submittedBugReport{https://bugs.eclipse.org/bugs/show\_bug.cgi?id=429954}

\paragraph{Actions taken.}
There are no actions directly springing out of this bug, since the Extract 
Method refactoring cannot be meant to be this way. This is handled on the 
analysis stage of our \refa{Extract and Move Method} refactoring. So names 
representing variables of local types are considered unfixes \see{sec:unfixes}.



\backmatter{}
\printglossaries
\printbibliography
\listoftodos
\end{document}