[Cuis-dev] FFI vs. Plugins - Re: Alien? ThreadedFFIPlugin?

[Juan Vuletich]

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:
>    http://forum.world.st/Stability-of-the-external-plugin-interface-td5117112.html

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
www.cuis-smalltalk.org
https://github.com/Cuis-Smalltalk/Cuis-Smalltalk-Dev
https://github.com/jvuletich
https://www.linkedin.com/in/juan-vuletich-75611b3
@JuanVuletich