Save (and load) compiled bytecode? #535
Comments
|
Ah, I was wondering when this would come up. :) I definitely think it's reasonable to want to separate out compilation from execution, especially for limited memory devices. There are a couple of challenges with it:
I'm going to leave this open because I do think there are valid use cases for it, and it would be cool to have. We could speed up start-up time of the VM if Wren's own core libraries were compiled ahead of time to bytecode and loaded directly from that.
I wonder how much of that is used by debug information: function names and line number buffers. We could potentially save memory by being able to turn those off (at the expense of losing stack traces at runtime). Also, the current representation of line info is very much not optimal. |
Yeah for some use cases this is more or less essential. Wren is such a good fit for what I am trying to do, but the performance and memory overhead of needing to compile at run time essentially quashes my ability to use it. Which is a shame.
Because sometimes there is no other choice ;) I would encourage you to take the approach that you use a version numbering scheme for the VM where minor numbers are improvements that don't break compatibility and major numbers do. The binary format for compiled binaries would carry this so the VM could check at load. Java does something like this. I can tell you that if I decide to go down the road of serializing the state of the VM to disk that I will have to write the version number to that image so that it would reject loading of an image with a mismatched version. Even that, with all the limitations it implies would be a fast improvement that shipping source and compiling it every single time the applet in question needs to be run. |
|
I just realized the situation of using Wren in a low memory environment is even worse then I first thought. On the page wren.io/embedding, you write:
So if I have a 8K source file I read into memory, this will consume 16K since you are making a copy of it. This is exactly the behavior I want in a desktop or server environment; but is not what you want to hear when you are running in a memory restricted environment. |
|
Yeah, that API would be easy to change (or to add another one). There's nothing fundamental about why Wren copies the source string. It just needs to ensure the passed in string has the right lifetime. But we could support an API where you pass ownership with the string and then it doesn't need to copy. |
|
I would personally vote for an additional "unsafe" api. Outside of resource constrained environments, the existing API is what you want |
|
Argument 1 is a false excuse. if you take a look at lua, it make zero promise on the VM opcodes to stay consistant. So technically only the scripts are valid precious source/data. Argument 2 and 3 are only true because of Meta.eval. Having the REPL builtin is nice, but maybe that should be optional. It would help to split the VM and minimal runtime from the quite noisy compilation infrastructure. For wrenInterpret I think we should go wrenInterpret(WrenVM*, const char*, unsigned flags) in the general API, this would brings opportunities to also pass other parameters. Along with the string life control flag, I would see some extra options like compiler output control, optional disassemble or debugger enabling. |
Fair enough.
The meta module is already optional. What I meant here was more things like:
The REPL does affect things because when compiling REPL expressions, we the compiler takes advantage of being able to tell what top-level variables are already in scope. I consider that an important use case for Wren in general. Of course, we don't need to support that for offline compiling. We could say that if you're going to save and load bytecode, you have to do an entire module at a time.
Yeah, I expect that API to get a little more sophisticated over time. |
I have already discussed this for my quick and dirty approach (still digging in to look at feasibility), but I think the idea of saving and loading the VM's configuration, state, and memory image would be a good approach. And is very similar to how most implementations of smalltalk work. So imagine we could issue the command:
So that wren will load and compile all the source and imported modules into memory — perhaps without the debugging information you mentioned previously — and then save that image out to a file. This would mean the user could write and test as much as possible on the desktop, then make an image that would be loaded in the final environment. This also means that the idea of having compile bytecode of mismatching versions would be less of an issue. It would be a good and useful step along the way, and would be immediately useful for people who don't want to distribute their core wren source for whatever reason. |
|
Le dim. 29 avr. 2018 à 20:41, Jock Murphy <notifications@github.com> a
écrit :
The meta module is already optional. What I meant here was more things
like:...
I have already discussed this for my quick and dirty approach (still
digging in to look at feasibility), but I think the idea of saving and
loading the VM's configuration, state, and memory image would be a good
approach. And is very similar to how most implementations of smalltalk work.
So imagine we could issue the command:
./wren --freeze [imageFileName] wrenFile {wrenFile...}
So that wren will load and compile all the source and imported modules
into memory — perhaps without the debugging information you mentioned
previously — and then save that image out to a file.
Does that mean imageFileName would contains all the required code frozen
after some running?
If so I don't particularly agree with that. Smalltalk does something like
that and it has quite a number of issues to consider and bring quite some
complexity.
Maybe you are referring to something like the smalltalk image builder. So
something like '--build-image' would be a better name. I'm note
particularly against it, but I'm not a big fan of having all bulked in the
image, in particular having the expected built-in modules to be
*duplicated* in the image.
For me the simple approach to have a compiled image for one module is more
the way to go. It's simple and work as shared code does so it is well-known
for users.
This would mean the user could write and test as much as possible on the
… desktop, then make an image that would be loaded in the final environment.
This also means that the idea of having compile bytecode of mismatching
versions would be less of an issue. It would be a good and useful step
along the way, and would be immediately useful for people who don't want to
distribute their core wren source for whatever reason.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#535 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AY7blZ38KxM4XJScIRsIwCPWfHBf_Xn4ks5ttglXgaJpZM4TrI-U>
.
|
|
@mhermier I did call out the similarity to smalltalk and said it should be a step along the way, not the endpoint As for what the command line option should be called, that is semantics and I trust @munificent to pick something appropriate |
|
Looked at some WebAssembly videos, and found some interesting ideas. We are near to be able to produce binary modules. We need 3 changes:
|
|
Is there any news or plan on this issue? I was wondering if i could use wren on a microcontroller too and would be interested :) |
|
Nothing on the compilation process has changed at this time so the above discussion is where it's at for now. |
|
This is the main reason I haven’t used wren (and do use micropython) for
embedded projects.
This is extra sad because in the past munificent has bragged how good wren
would be on microcontrollers
On Thu, Jun 18, 2020 at 4:12 PM .ruby ***@***.***> wrote:
Nothing on the compilation process has changed at this time so the above
discussion is where it's at for now.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
<#535 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAE5T7ZC3BDI3HBW6WXHZQ3RXKNO5ANCNFSM4E5MR6KA>
.
--
Jock Murphy
http://www.jockmurphy.com
Photographer * Writer * Storyteller
|
|
I don't see how storing bytecode would be good for microcontroller?
Specifically what kind of micro controller are you targeting, because I can
see how it can help little with start-up time, and probably with memory
usage, but considering the small footprint wren has, I don't see how it
should be a requirement?
|
|
@mhermier You would compile the bytecode on a computer and then load it on the microcontroller. The bytecode is smaller and faster to load and also (and sometimes the most importantly) you aren't distributing your source code Also as Wren is implemented now you have to load the entire source into memory and then compile it, so you need enough RAM for both bytecode and source in memory at startup. This very well might not be possible in some embedded environments Additionally you don't have the compile time overhead at startup. Remember that this isn't a feature request in a vacuum: MicroPython already does all of this and runs in as little as 8K of SRAM; eLua is similar; there are multiple JavaScript interpreters that run in less than 64K (and compile and load from bytecode); Squirrel has been made to run on some microcontrollers (see the Electric Imp); Tessel cross compiles JavaScript to Lua, then compiles the lua to bytecode, and executes the bytecode on the microcontroller; etc. Nor in the desktop environment is it uncommon to be able to load and distribute just the compiled bytecode with scripting languages, because again sometime you don't want to distribute the source For example I do a lot of projects with microcontrollers that have sub 1K of SRAM, but also ones with 32K-256K of SRAM. |
|
Do you have some numbers on trying to use wren on these kind of systems as
this?
|
|
@mhermier Sadly no, because of the whole need to distribute source combined with the fact that the source had to be loaded into memory to compile it, I deemed wren unfit for purpose in 2018 after I raised this issue and it was clear @munificent wasn't interested in implementing this feature in the short term. At the time I could prove that the need to load all of source into memory would break the project I was working on. I switched to micropython with a couple of projects using DukTape (a Javascript interpreter). Both of which have other features that make it useful for running in memory constrained environments |
|
That's sad, I really would like to have some numbers to have some rough
idea of the situation.
That aside, the situation with current implementation the situation a
little bit complicated. Even stripping all non vital stuff, there is still
the bytecode patching, to specialise methods to a class that needs to
happen a consume quite some memory, but that's only a guess.
The subject pop-up from time to time, and without someone that does some
proper experiments, it makes hard to discuss the pros and cons of the
current implementation.
|
|
@mhermier I hear you, but that person isn't me. I haven't looked at wren in years now. I only saw the updates to this issue because I forgot to turn notifications off. I appreciate your data driven approach but to my mind there are only two facts you need to know to make this a compelling feature:
As for the second, who cares if it benefits embedded or not? Most of the projects I work on (embedded or not) for clients that use a scripting language cannot distribute with source. Wren is cool, but not more compelling than Python, or Javascript, or... if it doesn't meet my needs |
|
I sense i might be of some help here, maybe in the coming months i'll catch up with the inner workings and see what i can contribute :D |
|
For curiosity, what happens does micro-python trigger an exception instead
of displayibg a stack-trace?
|
|
@mhermier Well micropython (note the lack of a dash) is written in C so there are no exceptions at the interpreter level per se. However the you will get an error code back and can generate stack trace using the C API |
|
What I really meant was: what can the interpreter display instead of the
stack trace if the code was stripped?
|
|
Yes though it isn't as useful since positions aren't known, just as in regular python |
|
Remember that bytecode compilation and saving is something common in many "full sized" languages as well. Nor does compilation have to happen on a file by file basis. DukTape — for example — lets you load all of your source and then save the current compiled state of everything to be loaded later. You suggested this as an option back in '18 IIRC This model might suit Wren better. There are multiple ways of solving the problem |
|
There are indeed various strategies. But until someone really stress/test
the implementation everything needs to be done.
It took me a long time to organize me, but I think I finally find a way to
organize the wren code, and experiment in making a fusion of reply and wren
analyzer, to make an experimental external compiler. Hope it gives some
results somehow.
|
|
Hi to all, Has there been any progress on sending only the byte code by externally compiling the program? It would also be interesting, for this application, that the compiler could have an optional compilation, since resources are scarce in a microcontroller. |
|
Afaik nothing changed. Small issue your repository is not publicly
available.
|
|
It saddens me to hear that. |
Look I understand and empathize with that statement, but it is also deeply frustrating to hear especially when @munificent when into various spaces to say that he thought Wren would be ideal for embedded devices... apparently without much understanding of the need, or why one wouldn't want to distribute source or have users endure runtime compilation in a microcontroller based device And while everything always needs to be done, the need to always distribute source means that Wren is disqualified for so many potential uses, and still is more than 4 years after I first brought this up. There is this chicken and egg problem with FLOSS projects, if Wren had been able to load precompiled bytecode (or a saved VM state) then I could have used it and been much more likely to test it, contribute to it and make it better. I don't know what decisions what went into the triage of what to address first, but it feels like a case of letting perfect be the enemy of good enough |
|
@jockm Please describe your use case. Embedded devices can rage from a system from 1kb of ram to Megabytes of ram and various clock speeds and other consideration. That said, updated wrenalyzer so it translate a (not finished) dialect near to official wren that suits my needs so it transpile to wren (for now). I don't plan to provide full compilation for now and even if reached it will not be public for some time. Compiling is far from an impossible task, but unless you bring solid arguments for it to happen, it will not magically happens because you think it is better. It should be tested to see if the benefits are marginals or not. Where are the numbers of the cost of a runtime compilation ? Can we have a better description of what we/you are trying to improve ? |
Just wanted to note I've seen people say this when Wren is ideal for embedding. Embedded != embedding contextually, where Wren is really good is embedding in other applications and games for scripting. This can apply to portable and embedded stuff but as mentioned it can vary wildly due to hardware. Here's Wren running on an Arduino M4 Grand Central board and the code for it that went into getting it to fit. Here's a thread on the sizes of things in the VM and how they brought it down to 50% and down to 39k etc. I mention this specifically because it shows that Wren is not designed for embedded, it makes trade offs to favour speed over memory use, simplicity over complexity and hackability and relies on 64 bit doubles etc. This fits a wide range of uses (including some embedded), and is primarily designed to be embedded in other applications, not designed for embedded hardware per se 💯 So while bytecode would be neat, it's just one part of the puzzle. It would almost surely make embedded stuff more possible, but since that's not really where Wren shines, there'll be a bunch more stuff come up immediately after and it's not going to be a straight path a lot of the time. |
Please refer to the original issue But again, even without space/performance issues: the need to distribute source is a showstopper for most non FLOSS embedded devices, no? |
|
Not really no. You don't have to distribute the plain source, you can obfuscate/encrypt whatever the contents first (I do this, but typically they're also packed into binary pack files anyway). The issues of the client having the code and it being in client memory will never go away. This means they can just dump the code from memory trivially, whether it's source or bytecode. Writing a bytecode -> wren converter is also trivial. It's a trade off on how much time you're willing to spend on something like that. For me it's quite low. |
|
Do you have some better description of the hardware ? an emulator or something else ? 64Kb is really constrained, I'm already surprised it fit there. I would analyze the binary to see where memory is really used, and start to made trades of from there. |
|
MicroPytjon, eLua, and at least 4 JavaScript VMs, amongst other scripting
languages can all run in most constrained environments than the one I
described
And again I first encountered Wren because the original author came onto
Hancker News (amongst others) into discussions of scripting languages for
microcontrollers with specs like I described and said that he thought Wren
would be better suited than something like MicroPython
And again: the need to distribute source
On Wed, Sep 28, 2022 at 3:00 PM Michel Hermier ***@***.***> wrote:
Do you have some better description of the hardware ? an emulator or
something else ? 64Kb is really constrained, I'm already surprised it fit
there. I would analyze the binary to see where memory is really used, and
start to made trades of from there.
—
Reply to this email directly, view it on GitHub
<#535 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAE5T7ZV6APN5JVEYWJ64XLWAS5WXANCNFSM4E5MR6KA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
--
Jock Murphy
http://www.jockmurphy.com
Photographer * Writer * Storyteller
|
That is still distributing source, and I may have multiple reasons including requirements from my clients that prevent that. Wanting to distribute only bytecode is a perfectly reasonable request, as well as not wanting to go through startup compilation every time |
|
It's not unreasonable (I don't remember suggesting that :)). I want it to exist, it's on the list of things that would be nice to have. It's a lot of change to get there, the way the vm + compiler works doesn't allow it as is. |
There are multiple microcontrollers that fit the description i gave. At the time I first filed this issue I was using an STM32L432 microcontroller. 256K of FLASH, 64K of SRAM, operating at 80Mhz. The device I was making for the client also had a microsd that held data files, micropython bytecode, etc |
I'm half assed working on serializing/suspending the running image to disk and being able to restart it later again. Think about Smaltalk images. But it is a slow work in progress as I suffer a bid from post Covid "mind fog". What I did so far - on Mac OS only, no make/Cmake support yet, is translating the https://github.com/wren-lang/wren-cli project with C++ compiler settings. As such it is treating the star-dot-c files as C++. Caused a few hundred errors regarding assigning void-pointers (and a few simple pointer type mistakes) to a lhs-pointers of a well defined type. So I added casts. Wanted to script that with an awk or perl, but was to foggy - see above. So I did it manually. AFAIKT all works fine. Did not run the build in tests yet. My reasoning to switch to C++ is, to use a kind of smart pointer (implemented as template in C++), which is addressing the VM relatively from the VM memory base address. Not sure how I personally would treat source code, I actually find the idea to keep all sources (and probably even make them modifyable and recompilable kind of neat) - would a special garbage collection, or a special memory region for the compiler do the trick? I also upgraded the Wren-CLI to load an arbitrary number of files on command line - instead of only one. Executing all of them (in the same VM obviously), and staying in REPL mode afterwards. |
I'm pretty sure it can. It is just that your customer or his customer does not want, too. If part of the logic on the embedded device can be done in a very high level language it helps everyone. There is nothing really more easy than reverse engineer some compiled code on a micro controller. Most of the time you even know what C compiler one would use, and the "macros" that one is using to compile the code. Converting a binary into a readable C program on a small device is at max a few days work. Giving the source out usually does not hide any "trade secret". However mixing a scripting engine in - and only delievering byte code, might increase the effort needed for reversal. What I want to say: not giving access to the source on a small device, does not really protect a trade secret. A contract however will. |
|
Hi, |
|
Why not? If pure speed isn't a concern, then why not work in a language
that is more pleasant and less prone to memory issues and other landmines?
In my case I often use micropython with a native core and it allows for
much more rapid experimentation, I can often tweak a feature on the fly if
a client is curious what the effect would be. In the specific case I
wanted to use Wren, I wanted the client to be able to have their own
plugins without needing to compile the source and load it via SWD. They
could just write a file put it on a microsd and see what happens. I
definitely didn't want to write a dynamic loader for object code for the
microcontroller I was using to allow then to do it in C
But why do you think C is magically better? Why do you assume speed is of
paramount importance?
And why isn't "because I have thought about it and want to" a good enough
answer?
…On Thu, Aug 17, 2023 at 8:22 AM Jordi Espada ***@***.***> wrote:
Hi,
First my apologies if I post a question that it hasn't nothing to do with
this thread. I wonder why any interpreted language as wren, lua,
micropython, etc is really a requirement to be on a constrained hardware as
described in this thread (i.e 64K ROM and 20KB RAM). Even though any
interpreted languange it fits in the ROM and executes (like micropython) it
will not manage the memory and speed execution as better as it does in C.
Why don't do it in C directly ?
—
Reply to this email directly, view it on GitHub
<#535 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAE5T74LFKCZUIBVKUEL4DTXVYZMBANCNFSM4E5MR6KA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
--
Jock Murphy
http://www.jockmurphy.com
Photographer * Writer * Storyteller
|
I very much agree with you that interpret language highs productivity , so I would use that too.
I don't say the C is magically better but if the execution of the main program is a for-while-loop that runs forever in interpreted language it would consumes more CPU cycles and more memory rather if it was on native code ... so in a long term of using it will consume more power and resources and so the lifetime of the component. But if you are telling me that the main loop is done in C and other parts done on interpreted language to load or test some parts of the program this is perfect :-) I posted this question from my ignorance. I've never tried execute an interpreted language on a limited hardware. |
|
Your wording really came off as presupposing that C was better, at least to me. But also you seem to be defining better as fast. Very few embedded programs are compute bound. Most programs spend most of their time waiting for input, either from another part of the system or from a user. If it is from a user then the time between inputs is glacial to the CPU. The microcontroller I used in my example was a ARM Cortex M4F, which means it ran at around 1.25/Mhz. So even if a user hammered on a button 100 times a second, the vast majority of the time would still be spent idling waiting on an interrupt. The next most intensive thing it did was write to a display, but that was just setting up a DMA transfer and sending it out, so it effectively happened in the background. So the compute bound tasks were building the display, and a bunch of decision logic. While it would be faster in C, the user and overall system would never see the difference between compiled and interpreted code. Finally on the topic of performance. I just want to share some information to give some context on speed. The Xerox Alto, the machine the some of the first GUIs were written for, had 64Kilowords of RAM (so 128K) and ran at about 1/500th the speed of that M4F (if I am doing my math right). Speed is relative. Would I want to compute pi to 10,000 digits in micropython? No. But then I wouldn't want to use a microcontroller for that task as well. |
|
Out of curiosity, do you have any number of energy consumption, between
equivalent projects on these platform & interpreters.
|
|
@mhermier Not of the top of my head, and it gets a bit complicated. Since we are talking about in order, little to no speculative execution, no burst, non superscalar, etc cores; most of the power consumption comes from: the io subsystems, how much current is being pulled by the gpio pins, and how long the CPU is allowed to sleep waiting for an interrupt. Computational load doesn't really matter for current consumption as much as the duration between sleeps. Sleeping wouldn't be a function of the interpreter, but a native call run by the program when it wants to sleep. Though I have seen some interpreters get clever and set a timer and then sleep for things like So in theory an interpreted program would run longer between calls to put the system to sleep, but in the grand scheme of things you aren't talking about that much difference. This also assumes the embedded system is trying to maximize for power consumption, this is far from the case. To give you a real world example, the microcontroller in my original example was pulling about 10ma most of the time when active, and that included the display. Something like a microSD can pull as much as 500ma, when active, but most I tested are less than 5ma on read. But something like the Nordic nRF53 series (a very popular bluetooth microcontroller) pulls 0.5ma when the UART is on, and much less than that when it is off at 64Mhz. |
|
Still interesting. I wonder if we can have an optimized version of the
environment that would react on interruptions and do the loop from C...
|
|
Very possible, but it is going to be platform/vendor dependent. So you want
to build an infrastructure for those kinds of calls and then have device
specific code that can be compiled in
On Thu, Aug 17, 2023 at 10:55 AM Michel Hermier ***@***.***> wrote:
Still interesting. I wonder if we can have an optimized version of the
environment that would react on interruptions and do the loop from C...
—
Reply to this email directly, view it on GitHub
<#535 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAE5T76IEAK7IHNO3ODE36TXVZLIRANCNFSM4E5MR6KA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
--
Jock Murphy
http://www.jockmurphy.com
Photographer * Writer * Storyteller
|
|
👍 I'm really keen to have it as well - I'd really like to ship binary code to the production. |
I am looking to use Wren on a microcontroller. I have done a quick and dirty test to prove that it runs and that reasonable enough scripts can run within the limited memory, however since I want to load scripts from an SD card I am left with an issue:
I am giving the VM 50K (out of 64), but that means I don't have a lot of space left over for buffers and strings and such.
So what I would really like is a way to stream compiled bytecode into the system.
I may try an tackle this myself. It occurs to me that the "cheap and easy" way to tackle this might be to load and compile the source, and then save the VM as a binary that can be loaded. This isn't as ideal as having a
wrenccompiler (alaluac), but it would be a useful first step.Are there any impediments to this working, such as address fixups when saving/loading or the like?
The text was updated successfully, but these errors were encountered: