<div dir="ltr"><div dir="ltr">Related to the discussion is this rather dated thread: <a href="http://lists.squeakfoundation.org/pipermail/squeak-dev/2010-March/147241.html" target="_blank">http://lists.squeakfoundation.org/pipermail/squeak-dev/2010-March/147241.html</a> I haven't yet seen much in the way of objective analysis since then to render it obsolete. (if it exists and I've missed it, please point it out!)</div><div dir="ltr"><div> </div><div>Conventional wisdom (prior to Cog & Spur VMs... that likely changes the equation at least a bit along with whatever (de-)evolution has occurred in the respective approaches over the years) used to be that Plugin > FFI > Alien from a raw performance standpoint. But Alien had callbacks... and when you need them, you need them. Also, IIRC Alien utilizes introspection and therefore wouldn't run without a sources file which is an issue for some deployment scenarios.</div></div><div> </div>As you mention the (at least theoretical) advantage of abstracting functionality by pushing it to the VM shouldn't be dismissed. I've experienced both the good and bad side of this with things like the sound and ssl plugins. When they work, it's great: you don't care what platform you're on: it just works. With an FFI solution, you're dealing with different libraries/locations and possibly entirely different APIs entirely on different platforms and it's usually up to you to sort it out. However, when there is a plugin problem (i.e. mismatched library versions etc), you have all of the issues of dealing with plugin and FFI code. So a lot of this potential advantage comes down to how well plugins are maintained. If they aren't, you may be better off with FFI since sometimes fixing a broken plugin can be more work than just rolling your own solution.<div> </div><div>Thanks,</div><div>Phil <div> <div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, May 27, 2020 at 1:49 PM Juan Vuletich via Cuis-dev <<a href="mailto:cuis-dev@lists.cuis.st">cuis-dev@lists.cuis.st</a>> wrote: </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Hi Philip, On 5/27/2020 2:17 AM, Philip Bernhart via Cuis-dev wrote: > Hi, > > I recently asked on the squeak vm-beginners list about the stability > of the external plugin interface. The general vibe from Eliot was: > don't! Use FFI instead! > > See: > <a href="http://forum.world.st/Stability-of-the-external-plugin-interface-td5117112.html" rel="noreferrer" target="_blank">http://forum.world.st/Stability-of-the-external-plugin-interface-td5117112.html</a> Ok. Specifically on this. I've used buth alternatives over the years. I developed or fixed bugs in JPEGReadWriterPlugin2, BitBltPlugin, ExtendedClipboardPlugin, SoundPlugin, (old experiment) VectorCanvasPlugin. I ported FFI to 64 bit Cuis (the 32 bit FFI was inherited from Squeak), and used it for OpenCL. I don't have a favorite, and I think there shouldn't be one. So, today I spent a couple of hours writing this: ================================ FFI or Vm Plugins? ------------------ There are two main ways to call external code in the Cuis / Squeak world. One is FFI (what is usually used in other dynamic languages) and VM Plugins (a mechanism that is specific to the Squeak VM). There might be use cases where one of them is preferable to the other. Let's take a look at what they do, to see their pros and cons. FFI --- FFI is a general mechanism to call external platform libraries. DLLs and the like. The mechanism is general, so for each different function we might call, we need a way to build the native platform call stack (function arguments). This is different for each platform, and different for each function. This function+platform specific specification lives in the Smalltalk image. FFI break with the idea that the VM defines the boundary between the Smalltalk world and the external world: - Any platform specific difference in functions called via FFI is handled in the Smalltalk image - The Smalltalk image is aware of each platform we might want to run on For a non trivial example, evaluate `Feature require: 'OpenCL'` and browse the OpenCLPrimitiveInterface hierarchy. VM Plugins ---------- VM Plugins are a Squeak VM specific way to call C code. The call mechanism is general, and handled by the VM. The C code is specific to the problem to solve, and must follow the rules for building plugins. In many cases it is not platform specific. This function specific code in plugin form lives in Slang or C code. VM plugins follow the idea that the VM defines the boundary between the Smalltalk world and the external world: - Smalltalk image is platform independent - VM + Plugins are platform dependent As examples of their use, you might browse all methods containing string `module: 'FloatArrayPlugin'>` or module: 'JPEGReadWriter2Plugin'> Use cases --------- Both FFI and VM Plugins can be used in a variety of scenarios, with possible different needs: (1) Using platform specific functionality (for instance, the Windows API) (2) Using third party code distributed as dynamic libraries (for example, TensorFlow) (3) Using third party code distributed as source code or static linked libraries (for example, libJPEG, stdlib.h memcpy(), math.h float stuff) (4) Own code, written in Slang or C for performance reasons (BitBlt, FloatArray) (5) Own code, needing to deal with Smalltalk object internals, or VM services. Additionally, some calls need to be done with strict real time restrictions (like music / audio), or execution might be so quick and so frequent that any call overhead needs to be minimized (for example, BitBlt, LargeInteger). How do they compare? -------------------- a) Ease of development and prototyping. For the scenarios where FFI is well suited, (1) or (2), working with FFI is much faster and easier. It is possible to reuse API docs and general know how from outside the Smalltalk world. On the other hand, code that needs to deal with Smalltalk object internals and VM services (5) is easier to do as a plugins. For code written in C (and not regular Smalltalk) for performance reasons (4), if you already have the required setup, writing a plugin is easier and faster, as you don't need to deal with the platform stack. -- Bottom line: It depends. b) Ease of modifications, ease of updating users installations. Modifying a plugin usually requires the developer to compile a new dll. Modifying FFI calls can be done with just Smalltalk code. -- Bottom line: Clear win for FFI. c) Call Speed. Plugin calls are almost as fast as regular numbered primitives. FFI calls range from terribly slow to just plain slow, when compared to that. -- Bottom line: Plugin wins, if you really need low call overhead. d) Support for callbacks Recent implementations of FFI do support real callbacks. With VM plugins, the best we can do is to register a semaphore for the plugin to signal. This is usually safer, as the callback is done as a Smalltalk Process switch, making it easier to protect shared state. But there might be a large delay from the moment the semaphore is signaled to the moment where the "callback" is actually ran. -- Bottom line: FFI (callbacks enabled) clear win, if you need callbacks. e) Access to VM services and data structures to deal with Smalltalk objects VM plugins have a lot of services provided by available VM functions, and types provided by VM maker. Slang/C code needing to deal with the internals of Smalltalk objects is usually much easily done in plugins. -- Bottom line: Plugin wins, if you need this. What is each alternative good for? ---------------------------------- FFI is good for: - Tight integration with host platform (Host windowing system, native widget libraries) - 3rd party code that is meant to be linked dynamically, and everyone else calls via FFI - Application specific stuff (where dealing with platform specific details is not a problem) - Functionality that requires real callbacks - Stuff that is in development, APIs that are not yet stable, experimental code Plugins are good for: - Own code written for performance - 3rd party code that is meant to be linked statically - Kernel Smalltalk functionality - Functions that take very short time to run, and overhead becomes dominant time - Stable functionality that will be used by many people over a long time ================================ I hope I haven't left out too many details. Feel free to argue, correct, discuss, etc. Thanks, -- Juan Vuletich <a href="http://www.cuis-smalltalk.org" rel="noreferrer" target="_blank">www.cuis-smalltalk.org</a> <a href="https://github.com/Cuis-Smalltalk/Cuis-Smalltalk-Dev" rel="noreferrer" target="_blank">https://github.com/Cuis-Smalltalk/Cuis-Smalltalk-Dev</a> <a href="https://github.com/jvuletich" rel="noreferrer" target="_blank">https://github.com/jvuletich</a> <a href="https://www.linkedin.com/in/juan-vuletich-75611b3" rel="noreferrer" target="_blank">https://www.linkedin.com/in/juan-vuletich-75611b3</a> @JuanVuletich -- Cuis-dev mailing list <a href="mailto:Cuis-dev@lists.cuis.st" target="_blank">Cuis-dev@lists.cuis.st</a> <a href="https://lists.cuis.st/mailman/listinfo/cuis-dev" rel="noreferrer" target="_blank">https://lists.cuis.st/mailman/listinfo/cuis-dev</a> </blockquote></div></div></div></div>