proposal-gsoc-2014.html

<h3>What is your e-mail address and IRC nick?</h3>

<p>bartwiegmans@gmail.com, brrt on freenode</p>


<h3>What is your web page, blog, or microblog?</h3>

<p><a href="https://plus.google.com/u/0/+BartWiegmans/posts">I have a google+ page</a> - although I don't actually post a lot of things there.</p>


<h3>What is your academic background?</h3>

<p>I'm studying for a bachelor's degree in biology at the university of Groningen in the Netherlands.</p>


<h3>Who is a possible mentor for the project you are proposing?</h3>

<p>Jonathan Worthington (jnthn)</p>


<h3>What is the ultimate goal of your proposal?</h3>p

<p>To create an optimising JIT compiler for the MoarVM virtual machine.</p>


<h3>What components/modules will the proposed work modify or create?</h3>

 <p>I will make substantial modifications and additions to the interpreter
   core of MoarVM.</p>p


<h3>What benefits does your proposed work have for Perl and its community?</h3>

 <p> It will allow us to run NQP and perl6 code much faster. And faster code
   is more useful code, because it then becomes feasible to use perl6 on
   larger projects.</p>


 <h3>Why are you the right person to work on this project?</h3>

  <p>Nobody else was crazy enough. Also, starting this year I've been
   programming for 10 years, starting with C and working my way through a
   variety of programming languages (including perl). I actually have some
   experience in working on interpreters and virtual machines: I created a
   scheme interpreter in high school and worked on the parrot project in
   2012. I'm also motivated, I like the perl6 community, and I'd love to
   learn more about compilers and virtual machines.</p>


<h3>How do you plan to achieve completion of your project?</h3>

   <p>Although a JIT compilter from scratch is quite an ambitious project,
   certainly for a single student in a single summer, I think it is also
   doable. The first reason is that there are many open-source
   implementations, such as LuaJIT, Dalvik VM, and v8, from which ideas can
   be stolen. The second reason is that there are also open source
   libraries available (such as dynasm) that take over many difficult and
   architecture-dependent parts of JIT compilation. Also, I plan to work
   together with my suggested mentor to create certain parts that require
   in-depth knowledge, such as forming a control flow graph out of MoarVM
   bytecode.</p>

   <p>I will work in public using github to publish all code and merge back to
   the MoarVM project repository using pull requests when appropriate.</p>


<h3>Please provide a schedule with dates and important milestones/deliverables (preferably in two week increments).</h3>

   <p>During the first two weeks, I'd like to take the time to research the
   structure of the MoarVM opcode set in further detail, as well as the
   functioning of the dynasm library. If dynasm turns out to be unsuitable,
   I hope to find out quickly.</p>

   <p>In the second two weeks (from June 2 to June 15) I will work on creating
   a bytecode-to-graph converter in MoarVM. I will probably also make a
   pretty-printer for this bytecode graph, perhaps as a part of the gdb
   support script. If possible, I'd also like to add (type) tracing
   operations and tables to the VM. This information will then be usable
   for optimisation routines.</p>

   <p>In the third two weeks I want work on compiling the code graph to a
   intermediate format that is suitable for just-in-time compilation (the
   'MoarVM low-level bytecode'. The v8 javascript interpreter does a very
   similar thing, I believe. If it is practical, I'd like that intermediate
   format to be to the 'regular' MoarVM bytecode restricted to a set of
   'core' operations. (MoarVM already defines a 'core' set of operations
   informally. However, the full set of operations is rather large, which
   is why I don't think it is suitable for 'direct' JIT compilation).</p>

   <p>During the fourth two weeks (June 30 to July 13) I plan to implement the
   'MoarVM low-level bytecode' in the form of JIT-able fragments using
   dynasm. When this is completed the next step is to JIT-compile these
   fragments using the dynasm library and implement an 'atomic code update'
   routine that replaces the old interpreted code with new JIT-compiled
   code. I'm least certain about the time this phase will take compared to
   others.</p>

   <p>In the next two weeks (July 14 to July 28) I think it is time to create
   explicit tests for the JIT system. Testing is obviously a continuous
   part of software development but taking two weeks 'off' for it will help
   find the weak spots and bugs that have been introduced while working. It
   will also give time to document the design and implementation and create
   an API for optimisation routines.</p>

   <p>In the last two 'real' weeks of the summer I will work on creating a
   'deoptimisation routine'. The function of a deoptimisation routines is
   to transfer the state from a running compiled routine to the state of
   the 'source' interpreted routine so that execution can continue
   there. It can be called when an assumption that was used to create
   optimised bytecode is no longer valid and is therefore required for the
   implementation of speculative type feedback optimisation.</p>


<h3>What will be showable two months into the project [1]?</h3>

 <p>The three important things I'd like to have by then are:</p>

   <ul>
   <li>An infrastructure for type and usage tracing in the interpreter.</li>
   <li>A (simplistic) JIT compiler.</li>
   <li>An API for bytecode optimisation routines</li>
   </ul>
  <p>Also, I'd like to have merged the work into the public MoarVM repository
   master branch by then.</p>


<h3>What are your past experiences with the open source world as a user and as a contributor?</h3>

 <p>I use many open source products (this proposal faithfully written in
   emacs :-)) and have in some occasions have had the chance to report bugs and
   submit fixes. (Notably in the django world were everything is always
   broken, but that is a digression).</p>

<h4>If available, please include links to any other code you wrote for Perl or other open source projects.</h4>

   <p>I've written mod_parrot, which embeds parrot into the apache httpd
   server. See https://github.com/bdw/mod_parrot. This hasn't been updated
   for quite some time, though, because I couldn't keep up with parrot code
   churn. I've also published some interesting solutions to problems I've
   encountered in https://github.com/bdw/lab.</p>


<h3>What other relevant projects have you worked on previously and what knowledge you gained from working on them?</h3>

   <p>I've created a scheme interpreter in C. And even though scheme is
   relatively simple it taught me many things about parsing, evaluation and
   garbage collection, which I think are still relevant skills.. More
   recently I've worked with the parrot virtual machine and learned a lot
   about the structure of (bytecode) virtual machines with more flexible
   object representations.</p>


<h3>What other time commitments, such as school work, another job, planned vacation, etc., will you have between May 19 and August 18?</h3>

 <p>The last course I'll be taking this year will end somewhere arround the end
   of may (not an exact date because no schedule is set). I have no other
   obligations or vacations planned.</p>