hot code swapping
andrewrk opened this issue · 18 comments
- Have the compiler run continuously, watching the file system for source
changes and automatically perform multithreaded compilation to build projects
quickly. - Hot code swapping. When integrated with the previous feature, you could
press "save" in your editor and see the change immediately in your running
software.
Of course this only works for some kinds of changes - obviously if you make changes to your initialization code, you won't see a difference. But it would probably work at the function level, so you could swap out any individual function for a new one.
This feature (automatic rebuild in background & hot code swapping) is available for C++ in Projucer IDE, built upon LLVM infrastructure.
Old demo 1, old demo 2, few implementation details.
Be aware that it took several years to implement and the implementor said that "it was the most complicated thing I ever did".
@PavelVozenilek The Projucer link has died, sadly.
Information from the dead link is probably here:
https://forum.juce.com/t/demo-screencast-of-the-projucer/9218
and here:
http://web.archive.org/web/20130808104856/http://rawmaterialsoftware.com/viewtopic.php?f=2&t=9794
I always wonder how should hot-swap deal with data (data structure layout and long-lived in-memory state), it seems rather difficult compared to hot-swap stateless code (like a pure function).
In some languages which support closure, it's even more complicated.
Are there any articles on this subject?
here https://molecular-matters.com/products_livepp.html is a commercial product which realizes hot-reload for c/c++ projects
I wonder if looking at the way Common Lisp deals with these things would be useful at all. Sure, CL has a fatter runtime than Zig at the binary level, but:
- If there is a watcher like @andrewrk hints we could possibly have it keep track of the old structure of the code it is watching.
- CL contains a full-blown object system and is able to manage updating all live instances of a CLOS class. Updating structs would be comparatively simpler, size changes might be the hardest part to deal with?
@isaachier Thanks, that article is very helpful!
I can see it passes a DLL-specific 'game_state' around, but what if the layout of 'game_state' got changed between reloads? I can imagine there will be lots of crashes/segfaults/misbehavior.
Now i've integrated C live coding successfully with libtcc, which hot-swap pure code flawlessly.
About the state transition, I can think of a solution which serialize the state out (to a layout-agnostic format), allocate new state, then serialize in. It should work, but involves a lot of manual implementation of serializing code and is not very generic/automatic.
nim-lang/RFCs#511
Nim seems to have this implemented as a reference.
Regarding code swapping in C, see https://github.com/fungos/cr
This would be a killer feature for me in game dev. Commercial game engines (Unity/Unreal) tend to implement their own hotswapping layers for engine code or rely on scripting languages to get iteration times down.
To have this supported at the language level would definitely reinforce the "so productive you have to use it" narrative for zig in games.
@cshenton
One of the possible thing you can already do in C and C++ and probably in zig as well is have most of your core logic into a dll / so file and having the program check for it and reload it on change in the debug builds.
It isn't "native" hot code reloading supported into the language, but it allow for a similar workflow
as an example : https://youtu.be/LNRfbZlppLo?t=60
altough it isn't the same as if it was builtin and it would be great if it was, you might find it useful !
Hoping this wasn't a necrobump.
Lisps have a great interactive dev experience. You can eval the last expression, the current line, the selected region, the current 'cell' (a region delimited by special comments), the current file, etc. They also autocomplete based on the living program in the REPL, and you can inspect the variables.
It would be awesome to have this workflow supported. Julia does this (more or less; structs cannot yet be redefined) with its VSCode plugin, so it's not like emacs is needed to support this.
I always wonder how should hot-swap deal with data (data structure layout and long-lived in-memory state), it seems rather difficult compared to hot-swap stateless code (like a pure function). In some languages which support closure, it's even more complicated. Are there any articles on this subject?
You could have a look at how it is done in erlang. However, that's a different world..
Another thing, which I haven't seen in the comments so far would be security.
Would an attacker be able to patch production code at runtime?
I did some live coding on this feature recently, and we got to a proof-of-concept status:
It's in the hcs branch. It depends on further progress on the x86_64 backend to be generally useful. It's also for Linux only currently. Each OS will need its own hot code swapping strategy.
There has also been https://www.jakubkonka.com/2022/03/16/hcs-zig.html (hcs-macos branch) by @kubkon
I need HRC(Hot Reload Code) , If Zig has HRC It will be perfect language .
Use case: industrial control applications which require "online changes" (ability to change code between control cycles with effectivley zero downtime)
Would want to trigger the hot swap manually and be able to tell if a hot swap is possible or not before attempting.