[CakeML] Adding records to CakeML

Fri Apr 22 21:10:26 UTC 2016

On 22 April 2016 at 22:37, Francis Southern <francis.southern at gmail.com> wrote:
> Hi both,
>
> Thank you for your replies.
>
> Regarding my knowledge in PL theory and type systems, it's a difficult
> question to answer, because I don't want to under- or overstate my
> knowledge.  I first encountered type theory a couple of years ago when I was
> at university studying mathematical logic, where I attended a reading group
> discussing the simply typed lambda calculus, the Curry-Howard correspondence
> and related topics (using the lecture notes by Sørensen and Urzyczyn, if you
> know them).  Around the same time I began learning OCaml, which got me
> interested in more applied aspects of the subject as well (in fact, I first
> came across CakeML via Owens' paper on the semantics of OCaml light).
> Recently I've been reading Pierce's `Types and Programming Languages' to
> build a more solid foundation for my knowledge, and I'm comfortable with
> exposition at this level.  In short, I know I'm coming at this as somewhat
> of an amateur, but I'm not a complete outsider.  For example, I am able to
> easily follow the slides at
> <http://www.cl.cam.ac.uk/~mom22/fp/lec9-types.pdf> (although I found some of
> the other lectures were a bit beyond my current knowledge in parts).
>
> As for mechanised proof, I don't have much practical experience at all.  I
> did some tutorials with Coq a couple of years ago, and have done the same
> with HOL4 in the last week in preparation for working on CakeML.  I know
> this is the main skill I need to develop in order to contribute to CakeML,
> and it's one of the primary motivators of my interest in the project too!
> I'm more than happy to be told to come back after reading/doing X, Y or Z,
> if you think I need more background here.
>
> I am also interested in implementing and verifying compiler optimisations,
> because, to be honest, I'm excited by the thought of CakeML being a viable
> choice for industrial software development, but I'm even less knowledgeable
> about compiler optimisations than type systems and I don't want to keep
> biting while I'm still chewing.

There are everything from very simple optimisations to very complex
ones that one can implement. As such one can build up from simple ones
towards complicated ones.

I'll give an example of a reasonably simple one: wordLang is a
language where instruction selection and register allocation occurs.
Once the instruction allocator has run, the code can contain
instruction sequences such as the following:

  x := 600;
  a := a+5;
  t := mem[a];
  if t = 3 then ... else ...

A pipelined processor can stall for a cycle at the "if" in case the
memory load happens to be slow (e.g. doesn't hit an address in L1
cache or similar). One can improve the code by moving the unrelated
assignment to x down so that it falls between "t := mem[a];" and "if t
= 3 then". With this change to the code the processor has more time to
finish the load instruction. The optimised code would be:

  a := a+5;
  t := mem[a];
  x := 600;
  if t = 3 then ... else ...

Note that this optimisation shouldn't be exaggerated, e.g. one
shouldn't move more than two instructions between "t := mem[a];" and
"if t = 3 then". Having a too long artificial distance can cause
variables to be live for longer than necessary.

Implementing this optimisation should be a standard exercise in
functional programming within HOL. The proof should not be too
difficult either, but can potentially get messy for beginners.

Kind regards,
Magnus

> On 22 April 2016 at 14:38, Scott Owens <S.A.Owens at kent.ac.uk> wrote:
>>
>> There are several possible ways to approach extending the CakeML type
>> system, and several different directions that extensions could go in. I’d be
>> happy to outline some possible projects, but would like to get a sense of
>> your expertise in PL theory and type systems and/or mechanised proof. There
>> are projects suitable for a variety of interests and expertise, but it might
>> be helpful to know before suggesting a concrete approach.
>>
>> Scott
>>
>> > On 2016/04/22, at 06:55, Magnus Myreen <magnus.myreen at gmail.com> wrote:
>> >
>> > Hi Francis,
>> >
>> > I'm very happy to hear that you are interested in contributing to
>> > CakeML. For records and type annotations, Yong Kiam, Scott and Ramana
>> > are the right people to consult. They should all be on this mailing
>> > list.
>> >
>> > I can imagine that records might be a difficult one to start with,
>> > while type annotations are probably a more gentle introduction.
>> >
>> > There are also some obvious optimisations missing in the compiler. Let
>> > me know if you are interested in implementing and verifying compiler
>> > optimisations. Some of the optimisations should be well contained
>> > within specific intermediate languages.
>> >
>> > Cheers,
>> > Magnus
>> >
>> >
>> >
>> >
>> >
>> > On 21 April 2016 at 22:24, Francis Southern <francis.southern at gmail.com>
>> > wrote:
>> >> Dear CakeML developers,
>> >>
>> >> I want to start by saying that I find your work on CakeML fascinating
>> >> and
>> >> I'm eager to find ways to begin contributing, so I'm looking for a
>> >> small
>> >> project to get my feet wet with (since I haven't got that much
>> >> experience
>> >> with formal development or programming language implementation).  Of
>> >> course
>> >> I've seen the list at <https://cakeml.org/projects.html> and I am
>> >> hopeful I
>> >> can start working on one of these projects (possibly related to pattern
>> >> matching) in the near future, but I'm trying to find something smaller
>> >> that
>> >> I can make progress on more quickly.  Since I'm working independently
>> >> at the
>> >> moment, I suppose I want to get feedback and that buzz of contribution
>> >> as
>> >> soon as possible!
>> >>
>> >> So, whilst poking around in the code looking for interesting leads, I
>> >> came
>> >> across a list of unimplemented features including records and type
>> >> annotations in the file documentation/reference.tex.  These seem to me
>> >> like
>> >> self-evidently desirable features which are hopefully reasonably simple
>> >> to
>> >> implement.  I don't mean to suggest they'll be easy, but they're the
>> >> easiest
>> >> meaningful contributions I've been able to identify!
>> >>
>> >> Adding records seemed the more interesting of the two, so I started
>> >> thinking
>> >> about it and discussed it a little in the IRC channel.  First off,
>> >> since I
>> >> understand CakeML has diverged from Standard ML in other areas, there
>> >> are
>> >> choices to be made about the semantics we want (basically, along the
>> >> lines
>> >> of SML vs OCaml).  Do we want them to be declared or anonymous?  Do we
>> >> want
>> >> unique field names?  Do we want to support field reordering?  Mutable
>> >> fields?  Subtyping?!  There are probably other possibilities I've
>> >> either
>> >> forgotten about or am ignorant of.  The answers to some of these
>> >> questions
>> >> also have implications for pattern matching and type inference, of
>> >> course.
>> >>
>> >> And then there's the implementation.  I think the main decision here is
>> >> whether we implement records on top of tuples or vice versa.  The
>> >> implementation will obviously require changes to several parts of the
>> >> codebase (the type system and inference, the parser, the compiler,
>> >> etc).
>> >> This is, of course, the purpose of the whole exercise, but still rather
>> >> intimidating to me, so any advice on how to approach this (which files
>> >> would
>> >> need to be worked on, which could be enlightening to study, etc) would
>> >> be
>> >> much appreciated.
>> >>
>> >> In fact, after thinking about all that, I'm wondering if it's actually
>> >> a
>> >> better idea to start with type annotations instead!  But anyway, I'm
>> >> looking
>> >> forward to hearing your opinions.  I hope that I don't come across as
>> >> too
>> >> clueless, and naïve only in the optimistic sense.
>> >>
>> >> Cheers,
>> >> Francis
>> >>
>> >> P.S.  Thanks to xrchz and YK from the #cakeml IRC channel for the
>> >> advice and
>> >> encouragement that led me to write this email.
>> >>
>> >> _______________________________________________
>> >> Users mailing list
>> >> Users at cakeml.org
>> >> https://lists.cakeml.org/listinfo/users
>> >>
>> >
>> > _______________________________________________
>> > Users mailing list
>> > Users at cakeml.org
>> > https://lists.cakeml.org/listinfo/users
>>
>