Hacker News Reader

Comment by DevelopingElk | Hacker News Reader

Good? Bad? Doesn't matter as long as you had fun.

Have you tested this GCs performance? Sometimes a baby GC can be fast enough.

Normally the reason to do a conservative GC is because it makes the VM’s implementation (and potentially FFI) simpler, not because it’s faster. Also for fun of course. (There are reasons a fully-conservative design might be faster, like not having to havr a ton of scanning functions if you have a ton of disparate types, but in general my perception is that production GCs are normally moving to at least some degree, which a fully conservative one can’t be.)

Yeah. I implemented the conservative collector because it simplified development of the language's primitive functions. Allowed me to put lisp values on the C stack without worrying about them being garbage collected.

The current version of the collector also compacts the heap and moves values around. All conservatively discovered values get pinned.

One recent interesting partly-conservative, moving design I remember reading about is Whippet[1] by Andy Wingo (Guile maintainer, SpiderMonkey contributor). He’s been partly forced into it, though, by the fact that Guile’s existing external API (unlike, say, Lua’s) exposes the fact that it uses a non-moving collector (by allowing the user to keep hold of raw pointers to Scheme objects), so I’m not sure if this should serve as an inspiration or a cautionary tale.

[1] http://wingolog.org/archives/2023/02/07/whippet-towards-a-ne... and other related posts

I really like his blog! I emailed him when I published my delimited continuations article because it addresses the overlapping native/lisp stacks problem he wrote about. Sadly I don't think he's seen it.

> One way is to inform the garbage collector of the locations of all roots [...] implicitly, in the form of a side table generated by the compiler associating code locations with root locations.

I wonder if there's a way to get GCC to do this.

Getting GCC to do things for you is fraught. Probably still possible; just... fraught.

I believe Clang was intended to be able to do this[1] and I remember seeing that stuff even back when it was a particularly spunky research project. The facility doesn’t seem to have really gone anywhere even if it technically works; I wonder why.

In general, the problem with stack maps of any kind—even the very minimal ones you need for stack unwinding—is that they’re liable to be large, slow to process, or both. Now that I’m thinking about it, I wonder if you could map the stack at runtime using those same unwinding tables. A C++ compiler does have to know where it put things so it can call their destructors, and perhaps you could make your GC root a word-sized but otherwise opaque uncopyable word-sized thingy so the compiler can’t put it in more than one place at once. (I can’t be the first to have the idea, even with how stupid it sounds and with how utterly miserable Itanium ABI unwinding is.)

[1] https://llvm.org/docs/StackMaps.html

Probably bad, right? I did have fun though.

I haven't measured the performance. I would like to. I'm especially interested in comparing it with the current version of the collector. It is now capable of heap compaction which will be the subject of a future article. I'm also interested in knowing how big a problem the conservative part of the collector is. The C stack depth has been minimized significantly since I got rid of the recursive evaluator. I don't think it's a significant problem but I could be wrong.

I need to implement the compiler's profiling functions in lone. Documentation on the matter seems scarce. Not even sure if those instrumentation APIs are stable. I got away with not implementing the ASAN and stack protector since they added flags that make the compiler emit trapping instructions instead of function calls. Profiling is a lot more complex.

Reply on news.ycombinator.com