Mentor for a JVM backend for GHC

Wed May 11 11:10:02 UTC 2016

Hi Rahul,

Thanks for the feedback, and good to hear you are thinking about using it
as an assembler!
First thing I'll have to do is to extract the assembler itself from "Kuna",
so it can be reused by yourself. I'll be happy to give support and
implement new features as needed, I know some important operations are
still missing but have a clear idea how to ad dthem.

Then a bit of history, where this project comes from, and why it had no
commit since December...

Long story short, I got trapped working on a FP library in Scala, and had
left this project aside in the meantime (had strong interest from folks
asking for me to look into this Scala stuff, and at the time not much
interest in Kuna... which the opposite of my personal preference... so I
was extremely happy to read your initial message here!).

Now back to our common interest, the reason I "chose" Core, it's because it
seems fundamentally impossible to avoid having an interpreter in order to
get stack safe execution of STG on the JVM. That is a fundamental
limitation, that is the core thing I want to experiment with... Creating a
JVM assembler and a minimal Lambda core calculus was kind of just an excuse
to get to this core issue (I'm thinking about control flow analysis and
complex program transformation/optimization in order to get there).

That being said, the path I would like to take is quite complex, full of
unknowns, I would love to share the ideas I have in mind, but I fear it
might be out of scope of your actual goal: getting a workable
implementation ASAP.

So, I suppose you are fine writing a small (hopefully performant) STG-like
Java interpreter used during runtime (that's basically what I'm trying to
avoid with my experiment), which I think is a reasonable path to follow in
order to get a GHC JVM implementation soon. I would be very interested to
help you on this, first by giving support on the JVM assembler (that will
be extracted from Kuna), then also by proposing myself as a mentor. I'm not
sure exactly how things goes there, how the process is handled... but I'll
learn about it, I did notify Edward Kmett of my interest.

My gut feeling is that it should be fine to use STG with the approach you
follow, I also feel some technics could be shared between GHC JVM and the
Kuna experiment (where the later could be seen as the a sandbox/test-bed
for the former).

Anyway as you said: "Again, all these issues can be looked at once a
performance baseline has been established.".

Totally agree, that's why I don't want to share too much of my experimental
ideas atm, let's focus on a realistic initial implementation (for which I
have lot of ideas to share as well). Again, I'm ready to assist you as much
as possible, implementation missing features in the assembler, I already
read so much of this specification..... ;-)

In case you are using IRC/IM/Gitter let me knows, so we could discuss more
directly.

Looking forward hearing from you,

Aloïs

On 10 May 2016 at 04:43, <rahulmutt at gmail.com> wrote:

> Hi Alois,
>
> I just checked out Kuna, and it looks like a great project. For others the
> link to the repo is https://github.com/aloiscochard/kuna. I think I'll go
> with it since not having to implement StackMapTables will save me a lot of
> time.
>
> I'm interested in your approach, can you explain more, especially the
> stack-safe bytecode part? And I noticed the last commit was in December. Is
> there any particular reason you stopped the project?
>
> I chose STG over Core because Core has an embedded language of coercions
> which complicates the code generation (or maybe they can simply be
> ignored?), and the terms are not in administrative normal form which
> requires more effort to manage. But a thought did cross my mind that
> certain Core optimizations might actually need to be turned off ‎because
> the resulting STG might be in a form that might not get translated to the
> most efficient JVM bytecode. Again, all these issues can be looked at once
> a performance baseline has been established.
>
> Thanks,
> Rahul Muttineni
>
> Sent from my BlackBerry 10 smartphone.
> *From: *Alois Cochard
> *Sent: *Monday 9 May 2016 10:21 PM
> *To: *Edward Kmett
> *Cc: *ghc-devs at haskell.org
> *Subject: *Re: Mentor for a JVM backend for GHC
>
> Totally agree, Cmm is too late.
>
> My aim in Kuna was to start the transformation from Core, targeting
> stack-safe JVM bytecode without using Graal or the like.
>
> Yes, I am quite an optimistic person ;-) but I believe it's the path to
> take. I'm not interested in Frege like approach for various reasons,
> performance being one of them.
> On 7 May 2016 6:19 pm, "Edward Kmett" <ekmett at gmail.com> wrote:
>
>> By the time it has made it down to Cmm there are a lot of assumptions
>> about layout in memory -- everything is assumed to be a flat object
>> made out of 32-bit or 64-bit slots. These assumptions aren't really
>> suitable for the JVM.
>>
>> -Edward
>>
>> On Sat, May 7, 2016 at 11:32 AM, Thomas Jakway <tjakway at nyu.edu> wrote:
>> > This is a strange coincidence.  I'm definitely no expert GHC hacker but
>> I
>> > started (highly preliminary) work on a JVM backend for GHC a few weeks
>> ago.
>> > It's here:
>> https://github.com/tjakway/ghcjvm/tree/jvm/compiler/jvmGen/Jvm
>> > (The memory runtime is here: https://github.com/tjakway/lljvm)
>> >
>> > I'm very new to this so pardon my ignorance, but I don't understand
>> what the
>> > benefit is of intercepting STG code and translating that to bytecode vs.
>> > translating Cmm to bytecode (or Jasmin assembly, as I'd prefer)?  It
>> seems
>> > like Cmm is designed for backends and the obvious choice.  Or have I got
>> > this really mixed up?
>> >
>> > I hope this isn't out of line considering my overall lack of experience
>> but
>> > I think I can give some advice:
>> >
>> > read the JVM 7 spec cover-to-cover.
>> > I highly suggest outputting Jasmin assembly instead of raw bytecode.
>> The
>> > classfile format is complicated and you will have to essentially rewrite
>> > Jasmin in Haskell if you don't want to reuse it.  Jasmin is also the de
>> > facto standard assembler and much more thoroughly tested than any
>> homegrown
>> > solution we might make.
>> > read the LLVM code generator.  This project is more like the LLVM
>> backend
>> > than the native code generator.
>> > Don't go for speed.  The approach that I've begun is to emulate a C
>> stack
>> > and memory system the RTS can run on top of
>> > (
>> https://github.com/tjakway/lljvm/blob/master/src/main/java/lljvm/runtime/Memory.java
>> ).
>> > This will make getting something working much faster and also solves the
>> > problem of how to deal with memcpy/memset/memmove on the JVM.  This
>> will of
>> > course be very slow (I think) and is not a permanent solution.  Can't do
>> > everything at once.  Any other approach will probably require rewriting
>> the
>> > entire RTS from the beginning.
>> > I don't think Frege is especially useful to this project, though I'd
>> love to
>> > be proven wrong.  Frege's compilation model is completely different from
>> > GHC's: they compile Haskell to Java and then send that to javac.
>> Porting
>> > GHC to the JVM is really more like writing a Cmm to JVM compiler.
>> >
>> >
>> > I've heard of the LambdaVM project but couldn't find the actual code
>> > anywhere.  The site where it was hosted appears to be offline.  I'd
>> > certainly like to look at it if anyone knows where to find it.
>> >
>> > Information on Jasmin:
>> > http://web.mit.edu/javadev/packages/jasmin/doc/
>> > http://web.mit.edu/javadev/packages/jasmin/doc/instructions.html
>> > http://web.mit.edu/javadev/packages/jasmin/doc/about.html
>> >
>> > Once you've tried manually dealing with constant pools you'll appreciate
>> > Jonathan Meyer's work!
>> >
>> > I forked davidar's extended version of Jasmin.  The differences versus
>> the
>> > original Jasmin are detailed here.  Some nice additions:
>> >
>> > supports invokedynamic
>> > supports .annotation, .inner, .attribute, .deprecated directives
>> > better handling of the ldc_w instruction
>> > multi-line fields
>> > .debug directives
>> > signatures for local variables
>> > .bytecode directive to specify bytecode version
>> > (most importantly, I think): support for the StackMap attribute.  If we
>> > eventually want to use new JVM instructions like invokedynamic, we need
>> > stack map frames or the JVM will reject our bytecode.  JVM 7 has
>> options to
>> > bypass this (but it's a hack), but they're deprecated and I believe not
>> > optional going forward.  Alternatively we can stick with older bytecode
>> > versions indefinitely and not use the new features.
>> >
>> > (Just to be clear, I forked it in case it was deleted.  I didn't write
>> those
>> > features, the credit belongs to him).
>> >
>> > I think the biggest risk is taking too much on at once.  Any one of
>> these
>> > subtasks, writing a bytecode assembler, porting the RTS, etc. could
>> consume
>> > the whole summer if you're not careful.
>> >
>> > I'd love to help out with this project!
>> >
>> > Sincerely,
>> > Thomas Jakway
>> >
>> > -------
>> >
>> > Woops, after scrolling back through the emails it looks like someone
>> sent
>> > out the LambdaVM source.  I'll have to take a look at that.
>> >
>> >
>> >
>> > On 05/02/2016 11:26 AM, Rahul Muttineni wrote:
>> >
>> > Hi GHC Developers,
>> >
>> > I've started working on a JVM backend for GHC [1] and I'd love to work
>> on it
>> > as my Summer of Haskell project.
>> >
>> > Currently, the build system is setup using a mix of Shake (for the RTS
>> > build) and Stack (for the main compiler build) and I ensure that most
>> > commits build successfully. I have ported the core part of the
>> scheduler and
>> > ported over the fundamental types (Capability, StgTSO, Task, StgClosure,
>> > etc.) taking advantage of OOP in the implementation when I could.
>> >
>> > Additionally, I performed a non-trivial refactor of the hs-java package
>> > adding support for inner classes and fields which was very cumbersome
>> to do
>> > in the original package. On the frontend, I have tapped into the STG
>> code
>> > from the GHC 7.10.3 library and setup a CodeGen monad for generating JVM
>> > bytecode. The main task of generating the actual bytecode, porting the
>> more
>> > critical parts of the RTS, and adding support for the threaded RTS
>> remain.
>> >
>> > The strategy for compilation is as follows:
>> > - Intercept the STG code in the GHC pipeline
>> > - Convert from STG->JVM bytecode [2] in a similar manner as STG->Cmm
>> > preserving semantics as best as possible [3]
>> > - Port the GHC RTS (normal & threaded) to Java [4]
>> > - Put all the generated class files + RTS into a single jar to be run
>> > directly by the JVM.
>> >
>> > My objectives for the project during the summer are:
>> > - To implement the compilation strategy mentioned above
>> > - Implement the Java FFI for foreign imports. [5]
>> > - Implement the most important [6] PrimOps that GHC supports.
>> > - Port the base package replacing the C FFI imports with equivalent
>> Java FFI
>> > imports. [7]
>> >
>> > A little bit about myself: I spent a lot of time studying functional
>> > language implementation by reading SPJ's famous book and reading
>> research
>> > papers on related topics last summer as self-study.
>> >
>> > I took a break and resumed a couple months ago where I spent a lot of
>> time
>> > plowing through the STG->Cmm code generator as well as the RTS and going
>> > back and forth between them to get a clear understanding of how
>> everything
>> > works.
>> >
>> > Moreover, I compiled simple Haskell programs and observed the STG, Cmm,
>> and
>> > assembly output (by decompiling the final executable with objdump) to
>> > understand bits of the code generator where the source code wasn't that
>> > clear.
>> >
>> > I also spent a great deal of time studying the JVM internals, reading
>> the
>> > JVM spec, looking for any new features that could facilitate a high
>> > performance implementation [8].
>> >
>> > It would be great if someone with an understanding of nuances of the
>> RTS and
>> > code generator could mentor me for this project. It has been a blast so
>> far
>> > learning all the prerequisites and contemplating the design. I'd be very
>> > excited to take this on as a summer project.
>> >
>> > Also, given that I have hardly 5 days remaining, does anyone have
>> > suggestions on how I can structure the proposal without getting into too
>> > many details? There are still some parts of the design I haven't figured
>> > out, but I know I could find some solution when I get to it during the
>> > porting process.
>> >
>> > Thanks,
>> > Rahul Muttineni
>> >
>> > [1] http://github.com/rahulmutt/ghcvm
>> >
>> > [2] I intend to organically derive an IR at a later stage to allow for
>> some
>> > optimizations by looking at the final working implementation without an
>> IR
>> > and looking for patterns of repeated sequences of bytecode and assigning
>> > each sequence its own instruction in the IR.
>> >
>> > [3] Obviously, the lack of control of memory layouts (besides
>> allocating off
>> > the JVM heap using DirectByteBuffers) and lack of general tail calls
>> makes
>> > it tough to match the semantics of Cmm, but there are many solutions
>> around
>> > it, as can be found in the few papers on translating STG to Java/JVM
>> > bytecode.
>> >
>> > [4] This is the GHC RTS without GC and profiling since the JVM has great
>> > support for those already. Also, lots of care must be taken to ensure
>> that
>> > the lock semantics stays in tact during the port.
>> >
>> > [5] foreign exports will be dealt at a later stage, but I am taking
>> care of
>> > naming the closures nicely so that in the future you don't have to type
>> long
>> > names like the labels GHC compiles to call a Haskell function in Java.
>> >
>> > [6] Basically all the PrimOps that would be required to provide
>> plumbing for
>> > the Prelude functions that can compile beginner-level programs found in
>> > books such as Learn You a Haskell for Great Good.
>> >
>> > [7] I know that it's a lot more complicated than just replacing FFI
>> calls.
>> > I'd have to change around a lot of the code in base as well.
>> >
>> > [8] I found that the new "invokedynamic" instruction as well as the
>> > MethodHandle API (something like function pointers) that were
>> introduced in
>> > JDK 7 could fit the bill. But as of now, I want to get a baseline
>> > implementation that is compatible with Java 5 so I will not be utilizing
>> > these newer features.
>> >
>> >
>> >
>> > _______________________________________________
>> > ghc-devs mailing list
>> > ghc-devs at haskell.org
>> > http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>> >
>> >
>> >
>> > _______________________________________________
>> > ghc-devs mailing list
>> > ghc-devs at haskell.org
>> > http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>> >
>> _______________________________________________
>> ghc-devs mailing list
>> ghc-devs at haskell.org
>> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>>
>
>

-- 
*Λ\ois*
http://twitter.com/aloiscochard
http://github.com/aloiscochard
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20160511/b029a1f6/attachment-0001.html>