Thank you for all the replies. Luite Stegeman - I was thinking that the LLVM IR code is optimized already or you can run LLVM IR optimization passes to get rid of such things. I think compiling with ghc -fllvm generates LLVM bitcode and then you can simply run emscripten on it to get Javascript - and it should work as expected. We are creating an online graphics (HTML5 canvas, webgl) processing tool - we want to use Haskell as our processing language and we want allow users to interactively process graphics in a functional, lazy manner. We want GHCI to be run inside browser, to allow for smooth work. So what about compiling GHC to JS - please look at: http://repl.it/languages There are a lot of interpreters (like python or ruby) compiled to javascript using emscripten. Would it be possible to compile the GHCI (written in haskell?) with GHC -fllvm to IR bitcode and then to javascript to get the online GHCI interpreter? (not such "online" tools like "Try Haskell", which under the hood compute everything on server) 2013/6/28 Tikhon Jelvis

My understanding is that Try Haskell actually runs the submitted code on a server with mueval rather than compiling it to JavaScript and running it in the client. This is different from some of the other "try" websites (like try.ocamlpro.com), so it's easy to get confused.

On Fri, Jun 28, 2013 at 11:24 AM, Henk-Jan van Tuyl wrote:

...

On Fri, 28 Jun 2013 15:25:59 +0200, B B wrote:

Hi!

...

Does anybody tried, or is there anywhere a project, of online ghc or ghci (compiled to JavaScript with Emscripten)?

There is Try Haskell![0], source code can be found on GitHub[1]

Regards, Henk-Jan van Tuyl

[0] http://tryhaskell.org/ [1] https://github.com/chrisdone/**tryhaskellhttps://github.com/chrisdone/tryhaskell

-- Folding@home What if you could share your unused computer power to help find a cure? In just 5 minutes you can join the world's biggest networked computer and get us closer sooner. Watch the video. http://folding.stanford.edu/

http://Van.Tuyl.eu/ http://members.chello.nl/**hjgtuyl/tourdemonad.htmlhttp://members.chello.nl/hjgtuyl/tourdemonad.html Haskell programming --

______________________________**_________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/**mailman/listinfo/haskell-cafehttp://www.haskell.org/mailman/listinfo/haskell-cafe

Emscripten is meant to translate ANY LLVM IR code to javascript and it should work (as I belive). I've tried to compile 'hello wrold' Haskell program to JS using Emscripten but I faced a problem, that in generated LLVM IR code there is no C-like main function (https://github.com/kripken/emscripten/issues/500) so there has to be a runtime library that has to be linked and will run the code - is this "RTS" or something else? If its RTS I understeand I should compile it to javascript and then provide it as a library to emscripten? Do you have a working RTS js version? I want to simply try emscripten and see how it performs on such GHC generater LLVM IRs. 2013/7/2 Luite Stegeman

On Tue, Jul 2, 2013 at 4:38 PM, B B wrote:

...

Thank you for all the replies. Luite Stegeman - I was thinking that the LLVM IR code is optimized already or you can run LLVM IR optimization passes to get rid of such things. I think compiling with ghc -fllvm generates LLVM bitcode and then you can simply run emscripten on it to get Javascript - and it should work as expected.

Most of the optimizations that GHC does are on Core, so you also get those if you use Core or STG as the source. You do miss out on the later optimization passes (GHC optimizes Cmm with Hoopl, and LLVM optimizes again). We implement this in GHCJS ourselves, optimizing the generated JavaScript.

Since GHC itself is written in Haskell, it would be a good goal to compile simple Haskell programs first and make sure that the RTS is working (you'll need to compile the Cmm and C RTS files to LLVM). You can try to get the GHC Testsuite [1] running. The GHCJS testsuite [2] might also be useful, it contains much of the GHC Testsuite and a runner program that checks the JavaScript results against native Haskell.

Like i said in the previous post, the generated code is a bit weird: stacks are allocated dynamically, functions never return, there are only tail calls. Emscripten might have a hard time with this.

[1] https://github.com/ghc/testsuite [2] https://github.com/ghcjs/ghcjs/tree/master/test

Thank you for your response :) Could you please answer one additional question - why you, while creating GHCJS didn't base on emscripten? Why haven't you patched it and created custom solution? Is GHCJS "production ready"? Also - Can I use GHCJS to compile big projects (like GHC or GHCI) to Javascript? When you mentioned the file sizes I thought it would be good to further introduce what we are trying to do. We want to be able to send a haskell code to a server to compile it and return the resulted JS to user as a compiled library. We want users to be able to connect these compiled functions together to get some interactive results. for example: user is creating 2 functions (pseudocode): a(x) and b(x). They are send to server and compiled to JS. Then in an online tool user is connecting (visually with lines) a data (lets say a list of ints) [] -> a -> b. While connecting it we want this tool to "interpret" such connections and visualise the data on each step - so we want to have some kind of "runtime" or "interpreter" on client side. 2013/7/3 Luite Stegeman

On Wed, Jul 3, 2013 at 11:06 AM, B B wrote:

...

Emscripten is meant to translate ANY LLVM IR code to javascript and it should work (as I belive).

It cannot compile ANY LLVM code: It's heavily geared towards porting C and C++ code to JavaScript, and still there are some limitatons, see:

https://github.com/kripken/emscripten/wiki/CodeGuidelinesAndLimitations

Also the FAQ mentions how event loops should be converted to something that emscripten can deal with:

https://github.com/kripken/emscripten/wiki/FAQ

Since the GHC runtime never returns, you'd probably have to do something similar.

I've tried to compile 'hello wrold' Haskell program to JS using Emscripten

...

but I faced a problem, that in generated LLVM IR code there is no C-like main function (https://github.com/kripken/emscripten/issues/500) so there has to be a runtime library that has to be linked and will run the code - is this "RTS" or something else? If its RTS I understeand I should compile it to javascript and then provide it as a library to emscripten? Do you have a working RTS js version?

I don't have a working version. The files you need are probably mostly in the rts directory of the GHC repository: https://github.com/ghc/ghc/tree/master/rts , but you might need a bit more, like libffi and gmp (unless you use integer-simple).

...

I want to simply try emscripten and see how it performs on such GHC generater LLVM IRs.

I don't think it's that easy... I'd expect a few weeks of work minimum to get very simple examples like "Hello, world" working. And a few months for something big like GHC.

Also keep in mind that the code will probably be pretty big, GHC (and GHCi, which is the same file) is ~65MB on my linux system, so getting well over 100MB of JavaScript wouldn't be terribly surprising.

luite

...

I think GHCJS should be able to compile all Haskell code in GHC, but we haven't tested this yet. The tricky bit is probably getting foreign code work, and creating a working installation that includes all other things, like libraries and a package database. Usually, GHCi loads object files for the libraries when running Haskell code. Obviously you can't run machine code with JavaScript, so you'd have to find a way around it. GHCJS includes an IO layer, which can be used to set up a virtual filesystem [2], but the API is far from finished.

Nice to hear that - we are concidering using GHCJS heavly in our project. We dont have many people and this is free-time "driven" project for now, but we would love to cooperate with you - help with GHCJS development and work together to make it suitable for the project we are working on.

Ah it's much easier if the code can be compiled on the server. This is more or less what Daniil Frumin is doing for his Google Summer of Code project [3]. We are using GHC on the server for non-interactive things, and GHCJS to compile interactive code that is run on the client. Our main goal is to get interactive graphics with the diagrams [4] library working.

Currently he is working on building a good sandbox for the compiler on the

server (With SELinux, rlimits and cgroups) so that we can compile code, and run Template Haskell safely. I'm working on improving the GHCJS linker, so that we can support incremental code loading more easily (so users can download just the code for the new function they wrote, instead of the whole program every time)

That's very interesting. We want to use GHCJS EXACTLY for this - we want to compile the code on the server and run it in the users browesr BUT we dont want to load the whole program everytime a function is changed - only this particular function. Additional (which is NOT the same) - while writing a function - we want to interactively execute it's "body" on the client side - but this requires something like interpreter on the user client (please see below for explanation). But we are thinking right now about it and its architecture, so we will be in touch in this topic with you :) I have a very crude example of this type here:

http://hdiff.luite.com/reduce/

This is not exactly the kind of thing I'm talking about. Think about something like matlab simulink or labview - functions are representaed by "nodes" (rectangles) connected with other functions by lines. Lines are "sending data" between these nodes. So user is connecting these blocks - each block is like a function application on data - so adding a block is like adding a new line in Haskell .hs file. We want user to be able to interactively add such nodes (apply further functions on already processed and cached data - every "node" caches processed data) and process the data further - If I didn't explained this simple enought, I will try harder. Such thing needs something like client side interpreter I think ...

When we have the SELinux sandbox working, I plan to work on this again, so that users can enter/compile their own code.

When do you plan roughtly to release it? We are strongly interested in testing and using it :)

On Jul 3, 2013, at 5:13 PM, B B wrote:

...

I think GHCJS should be able to compile all Haskell code in GHC, but we haven't tested this yet. The tricky bit is probably getting foreign code work, and creating a working installation that includes all other things, like libraries and a package database. Usually, GHCi loads object files for the libraries when running Haskell code. Obviously you can't run machine code with JavaScript, so you'd have to find a way around it. GHCJS includes an IO layer, which can be used to set up a virtual filesystem [2], but the API is far from finished.

Nice to hear that - we are concidering using GHCJS heavly in our project. We dont have many people and this is free-time "driven" project for now, but we would love to cooperate with you - help with GHCJS development and work together to make it suitable for the project we are working on.

...

Ah it's much easier if the code can be compiled on the server. This is more or less what Daniil Frumin is doing for his Google Summer of Code project [3]. We are using GHC on the server for non-interactive things, and GHCJS to compile interactive code that is run on the client. Our main goal is to get interactive graphics with the diagrams [4] library working. Currently he is working on building a good sandbox for the compiler on the server (With SELinux, rlimits and cgroups) so that we can compile code, and run Template Haskell safely. I'm working on improving the GHCJS linker, so that we can support incremental code loading more easily (so users can download just the code for the new function they wrote, instead of the whole program every time)

That's very interesting. We want to use GHCJS EXACTLY for this - we want to compile the code on the server and run it in the users browesr BUT we dont want to load the whole program everytime a function is changed - only this particular function. Additional (which is NOT the same) - while writing a function - we want to interactively execute it's "body" on the client side - but this requires something like interpreter on the user client (please see below for explanation). But we are thinking right now about it and its architecture, so we will be in touch in this topic with you :)

...

I have a very crude example of this type here:

http://hdiff.luite.com/reduce/

This is not exactly the kind of thing I'm talking about. Think about something like matlab simulink or labview - functions are representaed by "nodes" (rectangles) connected with other functions by lines. Lines are "sending data" between these nodes. So user is connecting these blocks - each block is like a function application on data - so adding a block is like adding a new line in Haskell .hs file. We want user to be able to interactively add such nodes (apply further functions on already processed and cached data - every "node" caches processed data) and process the data further - If I didn't explained this simple enought, I will try harder. Such thing needs something like client side interpreter I think ...

...

When we have the SELinux sandbox working, I plan to work on this again, so that users can enter/compile their own code.

When do you plan roughtly to release it? We are strongly interested in testing and using it :)

As Luite has pointed out we plan on releasing the first public beta within a month and we are hoping to ship full-featured easily extensible documented version by the end of the summer.

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

Similarly I would expect that generating any sort of sensible Javascript would require something fairly tightly tied to GHC; otherwise the output's going to have horrible performance because it's not going to understand the input and will fall back to the slowest but most general translation. (If it even has such a fallback, instead of simply failing on code that it doesn't recognize.)

Emscripten produces a limited subset of JavaScript (asm.js), and Firefox has an optimizer specifically for that. It could result in good performance, but the CPS transformed code is still a problem indeed, so it's unclear if GHC generated code would work at all. Asm.js is really limited, it's like a far less convenient LLVM without modern instructions in JavaScript syntax. You can't use objects, everything is stored in a global heap array. Interestingly, Firefox performs far worse than Chrome and Safari on the current GHCJS code, probably because their garbage collector is rather poor (non-generational), and Haskell code produces lots of short-lived immutable objects. luite