Borrow Checking, RC, GC, and Eleven Other Memory Safety Approaches
verdagon.dev
verdagon.dev
https://em-tg.github.io/csborrow/
These kinda-sorta fall under borrow checking or regions, just without any annotations. Then again, Ada/Spark's strategy also technically falls under Tofte-Talpin regions:
https://www.cs.cornell.edu/people/fluet/research/substruct-r...
Also, since it’s an error, I guess it must be different from gcc and clang (and, likely other C compilers) -Wreturn-local-addr warnings (https://www.emmtrix.com/wiki/Clang:Flag/-Wreturn-local-addr), which can have false positives.
What’s the difference? Is the language stricter, disallowing some constructs that are valid, but hard to prove or is the error not always triggered for buggy code?
That constraint is just a Tofte-Talpin region restriction, eg. lexically scoped regions.
https://github.com/prasannavl/WinApi
The performance of this interop layer is so close to native it's difficult to argue for doing things the much more painful way.
F# has other cool features like `inline` bindings and [<IlineIfLambda>] function attributes but I found it more difficult to work with for systems programming when attempting to stay within safe constructs.
But if you don't need byrefs, ref structs and spans, then regular pointer-based or even array-based code is quite pleasant to write, for example https://benchmarksgame-team.pages.debian.net/benchmarksgame/...
With that said - F# is an incredible language, but for systems programming you are likely to get better results in C#. I hope eventually someone contributes subsequent spec and compiler work to change this (and if you think you have time and could try it - please do! it's a nice small community).
With an acute accent, that should be roughly /ˌkɜːrˈseɪd/ “curse-ay-d”. (Think “café” or “sashayed”.)
The stylised pronunciation being evoked is roughly /ˈkɜːrˌsɛd/, “curse-ed”, and would be written with a grave accent: “cursèd”.
It is really good as food for thought though.
It reminds me of the early days of the web, when text was king and content was king. I particularly like the sidenotes in the margins approach.
(Hope the author sees this comment :) Hats off)
There's some great tooling for that via https://edwardtufte.github.io/tufte-css/ and https://tufte-latex.github.io/tufte-latex/.
> Wait a minute, this list goes to 17, yet the intro only mentions 14! I actually did that because a couple might overlap and a couple of them are half-approaches, and that last one is just here for fun. Besides, as I learn more approaches and add them to the list, the title will get more and more out of date anyway.
1. laissez-faire / manual memory management (c, c++, etc)
In this approach, the programmer decides everything.
2. dictatorship / garbage collection (java, go, etc)
In this approach, the runtime decides everything.
3. deterministic / lifetime memory management (rust, c with arenas, etc)
In this approach, the problem determines everything.
https://www.absint.com/astree/index.htm
For those who think static analyzers cannot do that, notice the word “sound”. This is a different type of static analyzer than the more common ones that do not catch everything.
Sadly, there is no open source option that works across a broad range of software. NASA’s IKOS is open source for example, but it does not support multithreading and some other things that I do not recall offhand, which makes it unable to catch all memory safety bugs in software using the features that it does not support. For now, people who want to use sound static analyzers need to use closed source tools or restrict themselves to a subset of what C/C++ can do so they can use IKOS:
* out-of-bounds array indexing,
* erroneous pointer manipulation and dereferencing (NULL, uninitialized and dangling pointers),
* read access to uninitialized variables,
NIST gave them a glowing review:
https://nvlpubs.nist.gov/nistpubs/ir/2020/NIST.IR.8304.pdf
It is possible to make sound static analyzers that can prove code to be free from memory safety bugs, but it is difficult and you need to implement checks that either complain or mathematically prove the absence of each class to do it.
These tools have been used for years in the aviation and nuclear industries, but almost nobody outside of those industries knows anything about them. If others had broader awareness of the existence of these tools, we could get open source equivalents and memory safety issues would be a thing of the past in C/C++ code.
Finally, your 1980s remark reveals enormous ignorance of what the field of formal methods has produced. Astree was not available in the 1980s. It took over 40 years of work to make and only became available about 20 years ago. C++ support took much longer for it to add, with it only adding it around 6 years ago if my recollection of the public documentation is correct. Some other things in this space are the Polyspace Code Prover and Frama-C.
Astree uses the theory of abstract interpretation that most of those in the formal methods community publishing papers do not touch as far as I know. It is therefore a different kind of tool for verifying properties of code than the techniques put forth in most formal methods papers.
NIST evaluated Astree and it got a perfect score on their test suite. It was actually a better than perfect score, as it found flaws in the test suite that NIST had not known existed. I have not heard of runtime errors in the Astree verified C/C++ code running on Boeing airliners or ESA space probes, unlike the prominent Toyota issues where they did not use any such tools and their slogan “always moving forward” became literal. So far, those who can afford it seem to be avoiding runtime issues in C/C++ code, which would be impossible if it did not work as advertised.
I acknowledge that correctness by external observation is not airtight logic. However, substantially similar reasoning applies to the security of RSA, since the RSA hardness assumption is an assumption and not a verifiable proof. A key difference is that with Astree, a claim is made that the software can make proofs internally to justify its claims, but we can scrutinize neither the proofs nor its source code. I choose to believe claims about Astree until they are proven otherwise given that Astree appears to be preventing runtime errors in a great number of mission critical applications and the independent NIST review found no flaws in its ability to statically detect sources of runtime errors.
You can think of it as the rather classic "Vec of struct + numeric IDs" that is used a lot e.g. in Rust to represent complex graph-like structures.
This combined with bound checking is absolutely memory safe. It has a bunch of correctness issue that can arise due to index confusion but those are not safety issues. When combined with some kind of generational counters those correctness issue also go away but are only caught at runtime not at compile time (and they incur a runtime cost).
Rust's memory safety is about avoiding liveness issues (that become type confusions since all memory allocators will reuse memory for different types), nothing more, nothing less.
there are not memory safety issues. But they definitely can lead to security issues with some sort of confused deputy attack.
For example a capability based system that relied on just this form of memory safety would be pointless.
Of course this can be mitigated by adding version counters to objects and object selectors.
If a language is merely type-safe, then it might be OK to silently replace a value with a different one of the same type, sure, fine. Who cares if the program transmits the wrong message to the wrong recipient as long as it's definitely some message and some recipient?
But a value-safe language, I suggest, is one that doesn't pull this kind of switcheroo.
If you have an array of objects of the same type and you just pick the wrong one, then the data still has to be a valid bit pattern. Yes it might still be a security bug, but it's much less likely because you aren't completely subverting the language.
Surely you don't think all bugs are the same because they are all bugs?
To begin with, the whole point of classes is to maintain invariants. Guaranteeing that a location in memory matches the valid bit patterns of its members is far from sufficient.
Obviously. I never said otherwise. What's your point?
I am interested in Vale and it feels very promising, though because my interested in bootstrapping I don't like that it is written in Scala. I know, that is shallow, but that's a thing that limits my enthusiasm.
If you are like me and don't like jumping around between notes and text and you prefer to read the notes anyway, here is a little snippet you can run in Web Inspector's Console:
document.querySelectorAll(".slice-contents a[data-noteid]").forEach(e => {document.querySelectorAll('.slice-notes [data-noteid="' + e.attributes["data-noteid"].nodeValue + '"] p').forEach(p => {p.style.fontSize = 'smaller'; e.parentNode.insertBefore(p, e)}); e.remove() })
Borrow Checking, RC, GC, and the Eleven (!) Other Memory Safety Approaches - https://news.ycombinator.com/item?id=41974185 - Oct 2024 (1 comment)
Borrow Checking, RC, GC, and the Eleven () Other Memory Safety Approaches - https://news.ycombinator.com/item?id=40146615 - April 2024 (68 comments)
- Type safety > Memory management and type safety: https://en.wikipedia.org/wiki/Type_safety#Memory_management_...
- Memory safety > Classification of memory safety errors: https://en.wikipedia.org/wiki/Memory_safety#Classification_o...
- Template:Memory management https://en.wikipedia.org/wiki/Template:Memory_management
- Category:Memory_management https://en.wikipedia.org/wiki/Category:Memory_management
The only real problem with this approach is code reuse, because library writers will insist on opaque structs and malloc rather than letting the caller allocate.
https://www.absint.com/astree/index.htm
None make any mention of stack exhaustion that I can find in a cursory search.
https://en.wikipedia.org/wiki/Stack_overflow
> In software, a stack overflow occurs if the call stack pointer exceeds the stack bound.
Also your searching astree site reveals:
> StackAnalyzer automatically determines the worst-case stack usage of the tasks in your application. It lets you find any stack overflows, or formally prove the absence thereof.
https://en.wikipedia.org/wiki/Stack_buffer_overflow
What you call stack overflow appears to be what I call stack exhaustion. The two different cases are very different things when you look at what happens in memory. In computers, the stack grows down, so exhausting the stack occurs in a downward direction. When you overflow an object on the stack, this typically occurs in an upward direction, and continues until older stack frames. Downward is also possible for that case, but it is rare and when it happens it can also be the other issue at the same time.
Hearing stack overflow used to describe the other kind of issue is what prompted my reply. I had not known that others use these terms differently. In all cases, we are describing something going past a boundary, with the only difference being what, so the ambiguous usage makes sense. The ambiguous usage appears to be acknowledged by Wikipedia:
https://en.wikipedia.org/wiki/Stack_overflow_(disambiguation...
I will continue to use stack exhaustion for this, as it is more accurate.
There is also a third case, which is jumping the stack guard page. I am not sure if you consider this to be a “stack overflow” too. It is a third class of bug. Wikipedia appears to lump it together with stack exhaustion under stack overflow.
I never expected bug taxonomy to be controversial, yet here we are.
Something that's missing is full-on formal verification where you write unrestricted C code and then mathematically prove it doesn't have any bugs. Nobody does this because proving a C program is correct is harder than mining a bitcoin block by hand, but it's useful to anchor one end of the safety/freedom spectrum. Many other approaches (such as borrow checking) can also be viewed as variants of this where you restrict the allowed program constructs to ones that are easier to prove things about.
This isn't to say reference counting is without problems (there are plenty of them, inability to collect cyclical references being the most well known), but I don't normally think of it as a slow technique, particularly on modern CPUs.
[1] an atomic inc on x86 is typically ~30 clock cycles, doesn't really pipeline well and will stall at the very least other load operations.
Is that actually true? I'm pretty sure you need the borrow checker even for single threaded Rust to prevent use after frees.
Here's a good post that talks about why shared mutability even in single threaded contexts is dangerous: https://manishearth.github.io/blog/2015/05/17/the-problem-wi...
But in general, if threads can't communicate any sort of pointer or reference that allows direct, unmediated access to the same value that some other thread will see, there's no need for a "borrow checker" for thread safety.
(Note that "but what if I have a thing that is just a token for a value that people can potentially read and write from anywhere?" is not an exception to this, because in this context, such access would not be unmediated. This access would still require messages to and from the "holding" process. This sort of safety won't stop you from basically deliberately re-embedding your own new sorts of races and unsafe accesses in at the higher level of your own code, it just won't be a data race in the same way that multiple threads reading and writing through the same pointer is a data race at the lower level. The main solution to this problem is "Doctor, it hurts when I do this." -> "Don't do that.")
My understanding is that most (all?) GC languages are memory safe, but do not ensure statically verifiable thread safety at all. Like Java, Go, C#, Python, etc.
Also, no point in calling it “tracing garbage collection”. Its just “garbage collection”. If you’re garbage collecting, you’re tracing.
You're against more explicit naming just for the sake of it? In the literature reference counting is also referred to as a type of garbage collection, and doesn't involve tracing. If you talking about a specific context you can probably drop the "tracing", but in a general article like this it would just be very confusing?
This way, someone can google "tracing garbage collection", and will find the relevant wikipedia article: https://en.wikipedia.org/wiki/Tracing_garbage_collection
For example, nobody says that Objective-C is garbage collected just because it has ARC. Nobody says that C++ is garbage collected even though shared_ptr is widespread. And systems that do tracing GC just call it GC (see for example https://www.oracle.com/webfolder/technetwork/tutorials/obe/j...)
To think clearly about the tradeoff between GC and RC it’s important to acknowledge the semantic differences:
- GC definitely collects dead cycles.
- RC knows exactly when objects die, which allows for sensible destructor semantics and things like CoW.
- it’s possible to use RC as an optimization in a GC, but then you end up with GC semantics and you still have tracing (hence: if it’s got tracing, it’s a garbage collector).
It’s a recent fad to say that RC is a kind of GC, but I don’t think it ever took off outside academia. Folks who write GCs call them GCs. Folks who do shared_ptr or ARC say that they don’t use GC.
And its good if this fad dies because saying that RC is a kind of GC causes folks to overlook the massive semantic elephant in the room: if you use a GC then you can’t swap it for RC because you’d leak memory (RC would fail to delete cycles), and if you use RC and swap it for a GC then you’d leak resources (your destructors would no longer get called when you expect them to).
On the other hand, it is possible to change the guts of an RC impl without anyone noticing. And it’s possible to change the guts of a GC while preserving compatibility. So these are really two different worlds.
Not always, but often enough that an introductory article that presents an overview of different memory managment techniques should maybe use the longer name to avoid confusion.
And I kinda agree with you, using the name "garbage collection" for RC doesn't really make sense, there is no metaphorical garbage truck driving around to collect your unused memory. :)
What's your opinion on the term "RC with cycle detection" that some use for things like Python's GC?
> And it’s possible to change the guts of a GC while preserving compatibility.
Moving to a conservative GC might also introduce memory leaks, if you're unlucky. But yes, "tracing" gc and rc obeviously behave very differently, and have very different performance considerations.
Referring to garbage collection as tracing garbage collection creates more confusion and should be avoided.
It confuses folks into thinking that there is some garbage collection that isn’t tracing. There’s no such thing.
> What's your opinion on the term "RC with cycle detection" that some use for things like Python's GC?
Depends on how you detect cycles. Python uses a garbage collector. Therefore I would say that python has a GC and is a GC’d language.
> Moving to a conservative GC might also introduce memory leaks, if you're unlucky.
Folks who adopt conservatism in production do so only if they have a story for avoiding those leaks. (That’s what we did in JavaScriptCore.)
> But yes, "tracing" gc and rc obeviously behave very differently, and have very different performance considerations.
Just call it “GC” and everyone will know what you mean. No need to be a contrarian and put “tracing” in front.
And it’s not perf considerations if it’s the difference between your program running at all and crashing. Failing to collect all cycles as RC does would cause a program written in a GC’d language to simply crash if it ran for more than just a short while. Failing to invoke destructors the way RC’d programs expect, which would happen if you tried to switch to GC, will cause observably different behavior in addition to possible performance issues.
It is standard to consider reference counting as garbage collection.
Bacon, D.F., Cheng, P. and Rajan, V.T. 2004. A unified theory of garbage collection. ACM SIGPLAN Notices. 39, 10 (Oct. 2004), 50–68. DOI: https://doi.org/10.1145/1035292.1028982
Abstract:
Tracing and reference counting are uniformly viewed as being fundamentally different approaches to garbage collection that possess very distinct performance properties. [...] Using this framework, we show that all high-performance collectors (for example, deferred reference counting and generational collection) are in fact hybrids of tracing and reference counting.
Yeah, that's what I meant to include by "behave very differently". I don't think we disagree on anything technical here. The problem is if you are currently googling for garbage collection you will mostly get garbage results. Here's duckduckgo to avoid my search bubble: https://duckduckgo.com/?q=garbage+collection+programming
The first result is the wikipedia article: https://en.wikipedia.org/wiki/Garbage_collection_%28computer... It's pretty bad, under "Strategies" it lists three: Tracing, Reference Counting and Escape Analysis. I'm sure these three are similar things.
The second result is this blog post, also listing rereference counting as gc: https://www.freecodecamp.org/news/a-guide-to-garbage-collect...
And the third result looks okay. Searching for "tracing garbage collection" has better results. The text in question already uses "gc" most of the time, and has a footnote saying:
> By "garbage collection", we're referring to tracing garbage collection.
I think that's as clear as it gets, without going on rant about the names of things. You are clearly an expert in garbage collectors, but most people in the target audience of that article are not. The article compares the differences between rc and gc. If someone then goes and reads the wikipedia articles about either of those they will be very confused because wikipedia will tell them rc is gc. A "fad" like this can't be undone, once a usage of a word becomes this popular you can't undo it.
Okay, sorry, this was too long, and we agree to like 99% anyway. Have a nice day! :)
But concurrent GC is the basis for making deterministic GC, since it gives you the option of scheduling GC work whenever you like rather than pausing the world.
Some concurrent GCs are also deterministic while others aren’t. I’ve written both kinds.
Intuitively it feels like making it concurrent should do the opposite of making GC deterministic! I’d love to read something showing that intuition is wrong
If you want to see my latest concurrent GC, see
https://github.com/pizlonator/llvm-project-deluge/blob/delug...
https://github.com/pizlonator/llvm-project-deluge/blob/delug...
fn modify(val: &mut u8) {
// ...
}
This is automatically thread-safe, without any locks. It's guaranteed that there can't be any side effects that could affect this memory, no matter what code you call. You don't need any defensive coding copying the memory just in case. It optimizes well, because it's certain that it won't overlap with any other region.
C++ doesn't have that kind of strong no-alias guarantee. Even memory behind const pointers can be mutated by something else at distance. The closest equivalent is C's restrict pointers, but they're more coarse-grained, and aren't checked by the compiler.
There is no equivalent in Python, Javascript, Typescript, Java, or Kotlin.
If only C++ had a thing called "references", which conveniently also used the "&" symbol and are a way to do exactly this without pointers.
Yes. E.g. every language where you can pass pointers can do this. Reference are a construct around pointers.
>Of the ones I listed, I use all of, and am continuously frustrated by this limitation.
Why? This is only a limitation for a few basic types. All the language you mention allow you to modify objects passed to a function.
If it ever were any issue though you can always return the modified object and overwrite the original.
a = modify(a)
works even if you can only pass a value.
In practice this should never be a limitation, certainly I have never encountered it as one.
Hint: Python and Javascript will use the type passed to determine mutability. Kotlin will let you pass a mutableStateFlow etc that can be used for, e.g., top-level data structure, but not an arbitrary variable. (e.g. a data structure field, or free variable).
The most popular languages, outside of C and C++, cannot pass pointers. And, raw pointers are not ideal; they lose type safety, memory safety etc.
There are always canonical patterns for a given language to write a function like I did. I find them to be universally more complicated and less explicit/versatile than what I posted; they require code to permeate beyond the function signature.
class A:
a = 0
v.a = 2
a1.a = 1
print(a1.a)
modify(a1)
print(a1.a)
I have never seem a codebase where this was an issue. The most important concept in programming is defining your data structures. Professional programs almost never operate on primitive data structures, but always on larger structures, like classes or structs.
This is why this is only a problem for you. I don't want to be insulting, but you need to brush up on your data structures. The need to pass a modifiable data structure almost never arises, the idiom in your OP is extremely uncommon and would likely not pass any code review.
Think about why nobody but you is inconvenienced by this. I am sorry, this is a you problem, because you are violating common idioms of the language.
There is basically never a reason not to pass your data structure to a function, which is why what you are doing is an incredibly niche functionality.
Also functions should be particular to data structure. That is why types exists.
It is not a me problem, which is why Rust has fixed this. You are arguing that functions which mutate parameters generally are not useful, which isn't true. You have implied that I have a lack of knowledge of data structures, my code wouldn't pass code review. Reflect: You are proposing it is preferable to have a simple function accept a whole data structure when it may only act on one or more of its fields, and may be used in other places. That is a semantic misdirection, and provincialization.
The point is that with mutable references, you can modify only the part of the struct you need to, using a general function. You can use this same function to modify a different part of the structure, or a different structure, or a free variable.
Your inference that preferring mutable references to wrapper structs implies insufficient knowledge of data structures is puzzling and incorrect.
Safety requires unions to be safe, so unions have to become tagged enums. To have tagged enums usable, you have to have pattern matching, otherwise you'd get something awkward like C++ std::variant.
Borrow checking works only on borrowed values (as the name suggests), so you will need something else for long-lived/non-lexical storage. To avoid GC or automatic refcounting, you'll want moves with exclusive ownership.
Exclusive ownership lets you find all the places where a value is used for the last time, and you will want to prevent double-free and uninitialized memory, which is a job for RAII and destructors.
Static analysis of manual malloc/realloc and casting to/from void* is difficult, slow, and in many cases provably impossible, so you'll want to have safely implemented standard collections, and for these you'll want generics.
Not all bounds checks can be eliminated, so you'll want to have iterators to implement typical patterns without redundant bounds checks. Iterators need closures to be ergonomic.
…and so on.
Every time you plug a safety hole, it needs a language feature to control it, and then it needs another language feature to make this control fast and ergonomic.
If you start with "C but safe", and keep pulling that thread, nearly all of Rust will come out.
So there are obviously different ways of addressing this.
Almost every programming language is memory-safe, but Smalltalk and most high-level dynamically typed languages require a fat runtime, and have a much higher overhead than C or Rust.
For languages that can be credible C and C++ competitors, the design space is much smaller.
And Objective-C is a credible C competitor, partly because it is a true superset of C, partly because you can get it to any performance level you want (frequently faster than equivalent practical C code [2]) and it was even used in the OS kernel in NeXTStep.
Now obviously it's not done, as it is a true superset and thus inherits all of C's non-safety, and if you were to just use the id-subset that is memory safe, you wouldn't be fully competitive.
However, it does show a fairly clear path forward: restrict the C part of Objective-C so that it remains safe, let all the tricky parts that would otherwise cause non-safety be handled by the id-subset.
That is the approach I am taking with the procedural part of Objective-S[3]: let the procedural part be like Smalltalk, with type-declarations allowing you to optimize that away to something like Pascal or Oberon. Use reference counting to keep references safe, but potentially leaky in the face of cycles. Optional lifetime annotations such as weak can be used to eliminate those leaks and to eliminate reference counting operations. Just like optional type declarations can reduce boxing and dynamic dispatch.
[1] https://blog.metaobject.com/2014/05/the-spidy-subset-or-avoi...
[2] https://www.amazon.com/gp/product/0321842847/ref=as_li_tl?ie...
Memory safety is just one aspect where the compiler can help making sure a program is correct. The more the compiler helps with static analysis, the less we need to rely on creating tests for edge cases.
I feel as though not enough attention is given to how std is designed. For example: [u8], str, Path, and OsStr may be confusing at first, but when you understand why they are there any other approach feels icky. std guides you down a path of caring about things that really should matter (at least if you're only unwrapping provably safe values).
Have you considered what happens if not-utf8 data winds up in an environment variable that you are writing to stdout? What if it contains malicious VT commands?
Unless you're talking about terminal bugs in parsing invalid UTF-8 - and parsing invalid UTF-8 is easier than rendering valid UTF-8 - VT commands are UTF-8 compatible. You just need to embed an ASCII escape character.
If an algorithm works for invalid Unicode, it should probably be an algorithm on bytes, not strings.
https://en.wikipedia.org/wiki/Maya_peoples
and secondly, the reason why the pre-Colombian cultural texts and script are not in use today, even by the people who speak the 28 Mayan languages currently in use, is because of genocide by Columbus and those that followed. The Catholic church destroyed every piece of Mayan script they could get their hands on.
The article reads like the author is not aware of these basic facts of American geography and history.
This is wrong, Interaction nets (and combinators) can model any kind of computational systems, including ones that use mutation. In fact, ICs are not really about types at all, although they do come from a generalization of Girard's proofs nets, which came from work in linear logic.
The interesting thing about ICs is that they are beta-optimal (any encoding of a computation will be done in the minimum number of steps required -- there is no useless work being done), and maximum-parallel with only local synchonization (all reduction steps are local, and all work that can be parallelized will be parallelized).
Additionally ICs have the property that any encoding of a different computational system in ICs will preserve the asymptotic behavior of all programs written for the encoded computational system. In fact, ICs are the only computational system with this property.
Interaction nets absolutely require garbage collection in the general sense. However, interaction combinators are linear and all garbage collection is explicit (but still exists). HVMs innovation is that by restricting the class of programs encoded in the ICs you can get very cheap lambda duplication and eschew the need for complex garbage collection while also reducing the overhead of implementing ICs on regular CPUs (no croissants or brackets, see Asperti[1] for what that means).
Having a linear language with the above restriction allows for a very efficient implementation with a very simple GC, while maximizing the benefits of ICs. In principle any language can be implemented on top of ICs, but to get most benefits you want a language with these properties. It's not that HVM starts with affine types and an efficient lazy clone operation, it's that a linear language allows extremely efficient lazy cloning (including lambda cloning) to be implemented on top of ICs, and the result of that is HVM.
> The HVM runtime implements this for Haskell.
This is very wrong. HVM has nothing to do with Haskell. HVM3 is written in C[2], HVM2 has three implementations, one in C[3], one in Rust[4], and a CUDA[5] one. HVM1 was just a prototype and was written in Rust[6].
HOC[7], the company behing HVM provides two languages that compile to HVM, Bend[8], and Kind[9]. Bend is a usual functional language, while Kind is a theorem prover based on self types.
Haskell is not involved in any of these things except that the HVM compiler (not runtime) is written in Haskell, but that is irrelevant, before Haskell it used to be written in TypeScript and then in Agda (Twitter discussion, sorry, no reference). It's an implementation detail, it's not something the user sees.
Please note that HVM adds some stuff on top of ICs that makes it not strictly beta-optimal, but nevertheless the stuff added is useful in practice and the practical downgrade from theoretical behaviour is minimal.
[1] Andrea Asperti, The Optimal Implementation of Functional Programming Languages, ISBN-13: 978-0060815424
[2] https://github.com/HigherOrderCO/HVM3/blob/main/src/HVML/Run...
[3] https://github.com/HigherOrderCO/HVM/blob/main/src/hvm.c
[4] https://github.com/HigherOrderCO/HVM/blob/main/src/hvm.rs
[5] https://github.com/HigherOrderCO/HVM/blob/main/src/hvm.cu
[6] https://github.com/HigherOrderCO/HVM1
A generalized RCU is just a tracing GC.
The problem is that it is very easy to write non-GC'd code in a GC'd language, but the other way around it is much much harder.
Therefore, I think the fundamental choice of Rust to not support a GC is wrong.
If GC is an option and you want all the nice parts of Rust, use OCaml
Having said that, there are still OCaml (as you noted), Haskell, .NET languages with Native AOT, JVM languages with GraalVM/OpenJ9, D, Nim, Swift, ....
And if one wants to get fancy with type systems, Idris, Dafny, FStar,..
So are you saying it would be possible to use a hypothetical "non-GC-enabled" OCaml compiler that complains if GC'd code is invoked/generated, and it would be a similar experience as using Rust?
In order to do away with Garbage collection you'd need to completely change the type system, and while you might get a result which ends up looking a lot like ocaml, it won't be ocaml any more than ocaml is sml.
This is what Rust started out as, by the way. In my opinion, it should have stayed that way. Post-Mozilla Rust has been nothing but a gigantic disappointment.
Do you think that projects that have a large GUI component should be written in Rust?
What if a project has both a "systems" and a GUI component to it?
I was talking with a colleague about that, he said “in C I know exactly where things are when” And I replied that under any OS with virtual memory, you have basically no clue where are things at any time, in the N levels of cache, and you cannot do accurate time predictions anyway… [1]
I’m convinced today GC is the way to go for almost all. And I was until 5 years ago or so, totally opposed to that view.
Many of their customers are factory processes and military deployments with weapons control, two scenarios where any kind of stall might produce deadly results.
Yet they surely appreciate the income they earn by using them.
It just runs at compile time. Bonus feature, it helpfully prevents a number of common bugs too.
If you have to do it yourself, then it does not "have" a GC.
Use after free is important, but in my experience not common and not too hard to track down when it happens (maybe I'm lucky? - we generally used a referenced counted GC for the cases where ownership is hard to track down in C++)
I'm more worried about other issues of memory safety that are not addressed: write into someone else's buffer - which is generally caused by write off the end of your buffer.
Yes. A garbage collector is only safe if it works correctly. What an irrelevant observation. Nothing can guarantee that something works correctly if it doesn't work correctly.