Zig's toolchain work is continually impressive. While I still don't plan to write software in it, given that I believe memory safety is table stakes, all of this stuff is very, very good. Before the incremental work, it was the toolchain and cross-compiler work. The toolchain stuff has continually been fantastic. I'm very curious to see what they come up with next!
> Semantic analysis is the most difficult part of the compiler to handle incrementally. Perhaps unsurprisingly then, this is where language design starts to matter a lot: while I am pretty confident that most modern languages could support incremental compilation similar to how we do, certain design decisions can make that much more difficult. Zig has had its design tweaked over the years (sometimes controversially) specifically so that it is easier to support fast incremental compilation.
This is something I wish that we had done with Rust. It is impossible to do all of the things at once, though, and we already had a tremendous amount of things to do. This is also part of the "when do you ship 1.0" tradeoff; for our goals with the language, 2015 was the right moment to launch, but if had a few more years to bake things, maybe we could have made compile times way faster. Software engineering is hard.
I'm not sure what that means. Java lets you do many things programs may want to do in a memory-safe way but not everything. Rust lets you do fewer things than Java in a memory-safe way, but more things than Zig. Zig lets you do fewer things in a memory-safe way than Rust, but more than C. So among these four languages we already have four levels of memory safety, none of them is 100%, all of them give up something in exchange for what they offer, and different programmers have different preferences for the compromise they prefer, and even that preference is context-dependent. Which of those less-than-100% memory safety compromises is the table stakes? And given that all of these compromises require something that could be quite substantial, depending on the circumstance, in exchange, and consequently programmers with the highest level of knowledge and expertise choose every one of those four in different situations, to me it seems pretty obvious that none of these is "table stakes".
I'll say the same thing I said to you as I said to Andrew, last time he and I talked about this: the way that everyone talks about memory safety (with maybe two exceptions, one okay (go) and one I dislike (fil-c)) is that "memory safe language" is about there being a clear delineation between what is memory safe and what is not, and that the unsafe aspect is a superset. Rust and Java both are memory safe, except where explicitly demarcated as not (unsafe in Rust, JNI or sun.misc.unsafe or whatever in Java). Zig and C have no such separation. When I (and others) talk about wanting memory safety, this is the important aspect of the design. This is what enables the "I know statically that a large part of the code is safe, and I also know where to check if something goes wrong" aspect of things.
Would you mind sharing some thoughts about fil-c? AFAICT its claims mostly check out so besides implementation details (GC?) it seems directionally good.
My main problem with Fil-C is that it combines the worst parts of C with the worst parts of a GC language.
In a language like C (or Zig), you need to manually manage memory. This makes programming a lot more complex, and it's really easy to accidentally mess up. Especially in large projects which have a lot of separate modules.
The biggest advantage of using a garbage collector is that you don't have to think about freeing memory. The GC automatically frees objects when they're no longer referenced. This makes programming much much easier. The downside of using a garbage collector is that it hurts performance at runtime. GC languages are slower and use more RAM.
Fil-C is the worst of all worlds here. Like C, it forces you to manually manage your own memory. But you still pay the performance cost of having a runtime garbage collector. And that cost is (apparently) really high. The only performance numbers I've seen showed ~2x worse CPU performance and ~4x worse memory performance. There's no way Fil-C can compete with C, Zig and Rust for performance.
So with Fil-C, you have a language that's much slower than C, and much more difficult to program in than C#, Java, Go or Typescript.
Fil-C still has some wonderful uses. Fil-C could be a fabulous debugging tool for C programs. It could be a wonderful teaching tool if it had nice visualisations on top of the GC's view of the world. And it could be a great way to run legacy C code.
But it's not a "rust killer". Fil-C programs run too slowly to be able to compete head to head with rust. And Fil-C doesn't offer the language benefits of a GC that you get in C#, Go, and friends. It seems like a really bad deal.
This is not necessarily true. It depends on a language, e.g. Go is slow, Nim[0] is extremely fast with conventional GC and slightly faster with ARC/ORC[1].
GC programs can be faster than manually managed ones in some cases. It's just manual memory management gives you more control of where and when free is called. And a good type system is a privelege that gives Nim more control with destructors.
Another scarecrow of safe languages is GC pauses, which is also not a thing in Nim, see table in [2].
I'll come back and say some more in a bit, but in short: I like the idea of fil-c, but I also have some issues. I think it's overall a good and interesting project, but with some caveats.
2x slower and 4x less memory efficient were the numbers I heard a year or so ago. Reaching within 20% of full native performance while using a garbage collector sounds too good to be true.
Okay so: in general, as a rule of thumb: anything that makes stuff have more memory safety is good. And experiments towards that end are also good.
What I do not like, primarily comes down to how the project is talked about and marketed. First, because it promotes an "us vs them" mindset, instead of a "we're all trying to improve memory safety" mindset, and second, because in doing so, it also overstates its case.
These things are sort of intertwined. Let's talk about the overstatement first. Fil-c has its own definition of memory safety that is slightly different than others. For example, I saw this recently:
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
struct User {
char name[8];
int is_root;
};
int main(int argc, char*argv[]) {
struct User* user = malloc(sizeof(struct User));
strcpy(user->name, argv[1]);
if (user->is_root) {
printf("I am root!\n");
} else {
printf("I am not root :(\n");
}
return 0;
}
This, when invoked with "012345678" passed in, will print "I am root!". In my understanding, this is deliberately allowed.
But beyond corners like this, fil-c's author will go on about "Rust has unsafe as a hatch, fil-c does not" while if you control-f for "zunsafe_" on https://fil-c.org/stdfil you get ... escape hatches.
The author regularly erases the difference between "traps at runtime" and "is prevented at compile time", which are legitimate tradeoffs where one or the other may be better depending on what you're doing. But they're presented as either equivalent, or one is superior, and I find this muddles the discourse. The performance issues also tie into this, "add a GC" is absolutely a valid way to handle these sorts of issues, but it is not the same thing as what Rust does. And that's okay! But presenting it as purely superior means that it's just hard to talk about.
Speaking of muddling the discourse, the author regularly trolls on X, providing tons of bad faith arguments and generally trying to rile up a "fil-c vs Rust" war that I think reduces our ability to talk about these differences in a calm, engineering focused context.
Finally, due to its design, fil-c is effectively Linux only. That's great for Linux, but many people also use other systems, and so it is not a meaningful option for them.
Anyway, after saying all that: I still think that it is a good project, and that it should exist and continue to be worked on. I just wish that the heat was turned down, and people could talk about the various approaches and their tradeoffs without turning it into a culture war.
I generally agree with all of this, but I'll add a few additional remarks. Because it's come up a bunch lately, I decided to do a bunch of code review/audit of the Fil-C codebase, and I'd say while it's got a lot of good bones, there's a long way to go to being a foundation I'd be ready to build on. I've reported a few UAF's upstream, and I've got a few PRs I'll add on, but if it only took me a day or two to find some of these big holes, I'm sure there's more lurking under the surface. I'd consider it to at this point be more of an engineering demo that this approach is feasible and tractable, but not a production ready language that I'd want to ship code in.
On the muddling the discourse, I'm not on twitter and don't engage there, so I don't have an opinion on that, but I did come across https://news.ycombinator.com/item?id=49044561 recently, and I just don't see how the author can make such bold claims while examples like the one Steve provided above are still in the language. Corrupting memory in Fil-C is still easy, type confusion is still easy, intra-object overflows are still easy. Fil-C prevents a range of classes of bugs from being exploitable, but it doesn't stop the bugs from happening.
> it promotes an "us vs them" mindset, instead of a "we're all trying to improve memory safety" mindset
But you did the same thing when, on the spectrum that ranges from C to ATS, with Zig, Rust, and Java somewhere in the middle (though all closer to C than to ATS), you declared the exact compromise that Rust makes "table stakes"! [1]
Zig improves on C's memory safety when it comes to spatial safety, possibly the more impactful kind, so it, too, could be part of the "we're all trying to improve memory safety", yet you exclude it.
You're trying to draw some hard line that passes exactly between Rust and Zig on the C to ATS spectrum, and I'm trying to say that that line isn't there (your attempt at a definition of delineating safe and unsafe code also applies to C). Obviously, C, Zig, Rust, Java, and ATS all make very different tradeoffs, all of which may be more or less attractive to different people and in different circumstances, but there is no sharp line, at least not one that is meaningful enough to be "table stakes". Your personal inclinations place a premium on the things Rust offers and Zig doesn't while mine are the opposite, but I make no claim to universality.
I'm happy to accept that not everyone shares my aesthetics and can understand why some people prefer Rust, but those claims to or hints at universality annoy me (as they did when they were made by Haskellers, and I actually find Haskell's aesthetics quite pleasing), as they are simply unsupported. I've spent a lot of time studying formal methods and software correctness in general (https://pron.github.io) and if there's one thing we know in that field is that things are never that simple (and, bringing this back to this posts topic, even something like incremental compilation can contribute to program correctness).
(Now, you may argue that you're only talking about "memory safety" and not correctness in general, but what gives memory safety value is that violations are causes of many dangerous vulnerabilities; but once, say, Java eliminates all of them, 100% of bugs/vulnerability - which are still numerous - will be caused by other problems, all potentially avoidable with ATS, so why isn't ATS table stakes? Of course, the answer is cost, but all the languages on the spectrum differ in their costs.)
[1]: I assume that you meant Rust's compromise, because you implied that Zig doesn't pass that bar but Rust does.
I’ve been watching this debate online and in my opinion both sides are guilty. Fil is intentionally trying to be funny or at least “interesting “ when he makes his points and I, for one, enjoy his humor, which includes having a go at Rust and other languages. It seems Rust people just can’t take a little criticism, even when it comes from a clearly trolling language! Yes it’s true Rust has an escape hatch, and we’ve seen serious memory safety bugs due to unsafe Rust in the wild. Fil-C does not have one, what you post seems to be internal or even temporary stuff given the author clearly has a goal of not providing one? I would say you and others need to just relax and not treat all and every Rust criticism as an offense to you.
So that could be a clear definition, but for it to be "table stakes" it needs to have some universal value and it doesn't (in fact, that very same definition could also classify even C as "memory safe"): https://news.ycombinator.com/item?id=49087458
I'm not exactly sure what you mean by "universal value", but I would say to be "table stakes" (i.e. not optional), it has to have overwhelming value. I think outside some fairly niche areas (e.g. programs that don't process untrusted data at all), it very very clearly has overwhelming value.
Now you might argue that the other features of Zig, like `defer`, are so good that they reduce the chance of memory errors and therefore memory safety has less value for Zig. But that seems highly dubious to me, especially for use-after-free. I guess we'll find out when Zig has more widespread use.
Yes, it is absolutely possible to build a language that uses Zig's compilation model and do borrow checking. Zig is not going to add a borrow checker though, so as a user, it's sort of a moot point.
no. let me be clearer; it is possible to intercept zigs ir from the compiler NOW (well, 15.2 proven) and have a third party package do borrow checking from the data that flow through, without changing zig (think "how miri works without changing rust"). this is not currently directly possible without changing the compiler (~ 50 loc), however the core team has indicated that exporting ir, the only change needed, will be a supported feature once the language stabilizes.
> I'm not sure what that means. Java lets you do many things in a memory-safe way but not everything. Rust lets you do fewer things than Java in a memory-safe way, but more things than Zig.
I don't think I agree with this framing. The question to me isn't "what can you do while being memory safe", it's "can you accidentally do something memory unsafe without noticing?" Rust and Java are the same here; you need to explicitly opt into using the language's mechanism for relaxing restrictions (Rust's `unsafe` blocks, Java's `Unsafe` class APIs), whereas from what I understand, neither Zig or C offers anything strict in that way.
That framing may seem intellectually satisfying, but it's not useful in practice. Consider the extreme edge case of C: We can clearly mechanically delineate between the empty program and a non-empty one, we call the empty program safe and any program that isn't empty unsafe (i.e. C is memory-safe if you want to do nothing and not if you want to do anything). And so, we also have this property that in C you can't do anything unsafe without noticing.
Now, that's ridiculous, but something not too different happens to me with Rust. I reach for a low-level language when I want to do low-level things in a more convenient way than in Java, but the very things that would make me reach for a low-level language in the first place are unsafe in Rust. So in ~100% of the programs I want to write in a low-level language, Rust and Zig offer the same level of memory safety (but I need to pay a higher price for Rust). That Rust reminds me that what I want to do is unsafe doesn't help me.
Of course, other people may want to reach for a low-level language in other situations and their perspective could be different, but if I pay the price and get little in return I can't see how that would be "table stakes". Table stakes imply some universality that is obviously not here.
Do you find most of your rust code is unsafe? Because I also find I do some unsafe stuff, because of the algorithms I work on I often end up with some unchecked array accesses and a couple of raw pointers into those arrays I pass around. But 98% of the code is safe and I find this makes it easier to reason about.
> That framing may seem intellectually satisfying, but it's not useful in practice.
Honestly, that's exactly how I feel in reverse. The framing you gave is more intellectually interesting, but it doesn't help explain the actual real-world outcomes where in practice, Rust and Java both don't have much problem with unsafety, whereas C does, and at least from what I've heard, Zig does as well.
> I reach for a low-level language when I want to do low-level things in a more convenient way than in Java, but the very things that would make me reach for a low-level language in the first place are unsafe in Rust. So in ~100% of the programs I want to write in a low-level language, Rust and Zig offer the same level of memory safety (but I need to pay a higher price for Rust). That Rust reminds me that what I want to do is unsafe doesn't help me.
I mean, sure, if you want to do things that are fundamentally not possible to validate because you think you're smart enough not to screw up, that's going to make Rust a tough sell. My issue with it is that history has shown that the best C and C++ programmers in the world still write code where memory safety rears its head, so I'm distrustful of the claim that being smart and diligent is enough to prevent the sort of bugs that we're still dealing with after half a century of us learning how not to write C. You need to have an excess of either talent or hubris to consider that a reasonably safe path, and given that the amount of talent needed is a lot higher than the amount of hubris, it seems way more likely that it's the latter.
The alternative is just learning how to write code that doesn't require expressing things in a way that can't be validated. While there are some things that fundamentally are not possible to, I'm dubious that it's anywhere close to as high as you seem to expect if your experience is that you literally can't reduce the amount of unsafe code you need in Rust below "literally my entire program is unsafe".
> Rust and Java both don't have much problem with unsafety, whereas C does, and at least from what I've heard, Zig does as well.
I'm not interested in the definition so much as I am in calling it "table stakes", and so the fact that these languages satisfy their promises is uninteresting in isolation. What matters is the value of their promises. The majority of Rust programs I see, I wouldn't have written in a low-level language, so the fact that it offers memory safety for the things I don't need it to do does nothing for me.
Now, clearly, Rust's originators didn't consider what Java offers (or at least what it offered 20 years ago when Rust was first conceived) to be table stakes or they wouldn't have wanted Rust. Java exacted some price in exchange for its memory safety that was unacceptable to Rust's originators and trumped its memory safety. But the same thing happens with Rust vs Zig. Rust exacts a heavy price for its memory safety, that - just as in Rust's case vs Java - is sometimes unacceptable. So I can't see how any of these could be "table stakes".
> I mean, sure, if you want to do things that are fundamentally not possible to validate because you think you're smart enough not to screw up, that's going to make Rust a tough sell.
What Rust can validate and what can fundamentally be validated are two very, very different things. Compared to what ATS can validate, what Rust can validate is almost indistinguishable from C. In Rust you have to do lots and lots of things that require you to be "smart enough not to screw up" that you could prove in ATS, and still no one (including Rust programmers) would say that what ATS offers is "table stakes" because, obviously, it comes at a high price that the people who choose Rust don't want to pay.
So clearly different languages offer different capabilities and charge a price for them. Sometimes the price is worth it and sometimes it isn't.
> so I'm distrustful of the claim that being smart and diligent is enough to prevent the sort of bugs that we're still dealing with after half a century of us learning how not to write C
But Java or Rust programs still suffer from a lot of bugs that ATS could eliminate, if you're willing to pay the price, and you're clearly unwilling. ATS programmers could say about Rust programmers what you say about C++ programmers. Clearly there's no universal table stakes here.
safe Rust is actually more memory safe than Java, since it guards against data races (in Java data races are not UB, but they are still one of the worst kinds of bugs because it leads to logically impossible program states)
Also note that Java has unsafe, but doesn't have the culture of plainly stating safety invariants like Rust. The unsafe features of Java are less widely used, but when they are you rarely know if a Java library has unsafe internals for performance, and if they do, it may be hard to audit
Java and Rust have actually very similar memory safety profiles. Rust let's you within the language escape the memory safety requirements whereas Java does not but both are considered memory safe languages. Rust also enforces thread safety as well which Java does not, but the slower JVM memory model doesn't let race conditions become memory safety issues whereas Rust is lower-level like Zig/C/C++ and thus thread safety could be a memory safety issue.
Zig has an identical memory safety profile to C. It has facilities to make it easier to stay memory safe, but those facilities are basically equivalent to what you have in C++ and that's equivalent memory safety profile as C.
> So among these four languages we already have four levels of memory safety, none of them is 100%
No, you've pretended like there's four when really it's Java / Rust which are safe by default and Zig/C/C++ which are unsafe by default.
One effective metric to evaluate is memory safety per LoC. Rust is ~0.2 vulnerabilities per MLoC. Java is effectively 0. C and C++ both seem to be about 1,000 vulnerabilities per MLoC. Zig is too new and hasn't had any analysis done on it, but generously it's likely at least 10-100.
So the table stakes could be defined as 1 memory safety vulnerability per MLoC.
> Java and Rust have actually very similar memory safety profiles
They really don't. Look at how many basic data structures (in the standard library or outside it) require unsafe features in Java vs Rust.
> Zig has an identical memory safety profile to C
It really doesn't. Zig gives you the same spatial memory safety as Rust and very much not like C (and violations of spatial memory safety are a bigger cause of vulnerabilities than violations of temporal memory safety).
> Look at how many basic data structures (in the standard library or outside it) require unsafe features in Java vs Rust.
This is a fundamental misunderstanding of how "unsafe" code relates to a platform's trusted computing base. Rust could move all of those unsafe data structures out of the standard library and into the compiler itself, thereby reducing the amount of occurrences of the string "unsafe" in the source, code, but this would do nothing to reduce the size of the trusted computing base that Rust presents. In fact, it would decrease our confidence in that code, because Rust libraries have a robust ecosystem of tools for validating their correctness, unlike whatever bespoke IR the Rust compiler itself is emitting. Java's own data structures are implemented with the support of an extensive runtime written in C++, which forms their own trusted computing base that every user of Java relies upon, and demands just as much careful auditing as any data structure in the Rust standard library.
> Rust is ~0.2 vulnerabilities per MLoC. Java is effectively 0. C and C++ both seem to be about 1,000 vulnerabilities per MLoC. Zig is too new and hasn't had any analysis done on it, but generously it's likely at least 10-100.
You have just described six orders of magnitude in your attempt to rebut pron pointing out the four languages have four levels of memory safety.
"two things are within an order of magnitude, and two other things are within an order of magnitude, and those two groups are three orders of magnitude apart" does sound like two groups to me.
> those two groups are three orders of magnitude apart
They aren't necessarily, though. Supposing that Zig were "10 issues per MLoC" (with just as much handwaving as the original poster), it would be equidistant from Rust and C. Java may also be more than one order of magnitude away from Rust; we say ~0 but is it 0.01, 0.001, 0.0001...? And why is "1 issue per MLoC" the acceptable metric that delineates what constitutes table-stakes memory-safe language? Because it's a nice, round-sounding number? I think 0 is a nicer, rounder number than 1, so let's call only Java table stakes and condemn all other languages to the garbage bin, tradeoffs be damned. Or would you say your arbitrary delineation point is worth more than mine?
10 years ago, I commented on the Rust issue for "Incremental recompilation", where it was suggested that Rust could at least adopt Haskell GHC's model of incrementality, which is currently file-level:
I recommend anybody who's interested in incremental recompilation to read what GHC does, because the effort to achieve that is relatively low.
Of course there's always desire for more:
GHC currently needs to parse+typecheck+codegen a file before it can process other files that import it. Codegen is slow. Thus, there's currently demand split compilation into "stages", so that the next file can be typechecked after its imports have been just typechecked (not codegenned).
I would also enjoy if recompilation avoidance were to happen at the function level, not the file level.
Macro systems are a key language feature that can destroy incremental recompilation. In theory, Haskell is well set up for that, as its macro system (TemplateHaskell) is fully AST based and _theoretically_ could distinguish "fully pure" macros from side-effectful macros (such as splicing the current git commit in as a string literal). But the recompilation avoidance system does not currently exploit such differences.
> Haskell GHC's model of incrementality, which is currently file-level
From what I see in Haskell files are the unit of compilation, and that's what allows incremental compilation to be file based (because it's really unit-of-compilation based)
I can see you can have circular dependencies between files with the `{-# SOURCE #-}` pragma, but I don't see documentation about how that affects incremental compilation.
A couple of issues I see with doing this in Rust are:
- in Rust the unit of compilaion is a crate, which can contain many fils/modules with circular imports, which is much more coarser than what can be done in Haskell.
- in Rust downstream crates can depend on function bodies upstream for running compile time functions; as such the crate/module interface is not enough to gate recompilation, but at the same time including all function bodies will also not give the wanted benefits. This is solvable but likely requires more work than what was done in Haskell.
In general you cannot take a language approach and blanket applying it to another one without considering their different quirks, which is likely why your proposal didn't get much attention. Or am I missing something that would make it easier to apply Haskell approach here?
Just to be clear about it, Rust today does do some amount of incremental compilation, and there is more work being done to continue to make it moreso. It's just very difficult to re-architect such a large and heavily used codebase. People are putting in heroic amounts of effort to improve things.
The article gestures at (and the author has made a comment in this thread about) how this is the case for Zig. You are right that there is tension here, and so that's exactly what you do: accept less performance for the gains in incremental, and then don't do incremental for final builds. It's a fine way to go about it, assuming that the lack of performance doesn't make the program unusuable. (Some people add some basic optimizations to their Rust debug builds, for example, because no optimizations is too painful to actually use.)
Rust's safety checks have basically nothing to do with its slow compile times. This is something that sounds intuitive but is just completely incorrect.
In particular, Rust made several good design decisions around this stuff that keeps those checks fast, like keeping checks local rather than being global.
That's interesting. Coming from C++ and Zig, the massive time "wasters" are metaprogramming features, i.e. Templates and comptime. Are Rust's macros the compile-time culprits?
That's only part of the issue. First, declarative macros are of course pure. Most procedural macros are pure too, although there are exceptions (an infamous case is sqlx). And while rustc does expand them every time (although a config for declaring "my macro is pure" was considered multiple times and the teams generally view it favorably), rust-analyzer just declares "we don't support impure macros" and caches them anyway.
Macros can be, but in part because they can produce new items (top level declarations, to sort of make the same handwave as the article does) and so that means you have to do macro expansion and stuff before you can even start to check some things, and similar issues. See the link I posted above for some details on a related issue.
There's also stuff around name resolution.
Proc macros are just an inherently very slow way to do what they do.
Because Rust commits to the traditional compilation model and pipleine (which I think is overall a good thing, or at least, a good thing to support), it does a lot of work that will eventually be thrown away. Consider this example: I have a library with a function foo that returns a simple 42. I have a binary which calls foo from that library and prints the result. Now imagine the library is a hundred thousand lines of unrelated code to what the binary needs, but is useful for other people. Because compilation works in the "produce libraries, produce binary, link them all together" style model, you have to compile the entire library with all of that code, when all you need is really one function. That intermediate work is useful, and I'm picking an example that's deliberately extreme, of course.
There's a bunch of stuff like this, and I do not have time to really say more than that right now. But yeah, monomorphized generics also produce a lot of compile time pressure too, in various ways.
Anyway I just also want to reiterate a few things: first of all, all of these decisions were made for good reasons, and there are pros to what Rust does and why. It's just that compile times suffer because of it. What I wish was that we had taken compile times into more consideration when deciding what to do and why in a more serious way. The same decisions might have been made, but at least it would have been known, rather than the situation now, where there's just a tremendous amount of work to try to optimize what exists, rather than having the freedom to maybe tweak some things to make that job way easier.
The traditional model of compilation is the biggest issue holding back fast incremental compilation IMHO. Swift suffers from this as well.
Even a simple change to one file results in re-parsing the whole library because definitions come from anywhere and we have to obey the 1970s single file compilation model. The result is a driver spawns 8 threads and each one wastes time re-parsing every file in the library looking for definitions. AFAIK Rust doesn't really track dependencies at the file or function level either so it doesn't really know what changed.
To me compilers should be content-addressed databases. Each declaration and its associated content generate hashes that roll up to its containing type or namespace, then to the file, then to the library as a whole, along with hashes of the dependencies. Changing the type signature of a single function should result in the compiler being able to cheaply determine whether that has any visibility and if so to what other files in the same library or if it affects the public interface.
A file that hasn't changed and whos inputs hasn't changed should re-use the IR from the prior compilation. Even for an individual type that should be the case so changing the internals of a function in a struct only regnerates that one function and nothing else. The compiler knows deterministically that change can't have affected anything else.
That has major benefits for code completion and editing as prior compilations can feed into generating errors or suggested corrections.
Then you can take things a step further and JIT a changed function, injecting the new machine code on the fly so long as the shapes of the types don't change. Very useful for debugging.
Compilers are mostly held back because the people who write compilers are stuck on certain ideas about how compilers should be written.
A really dumb 1 AM question. If there is a lot of work thrown away because stuff is compiled even if not needed, would making every function generic and delaying compilation until instantiation help here?
Note that it's not a serious suggestion, but I wonder what effect it would have on build times.
> Because Rust commits to the traditional compilation model and pipleine (which I think is overall a good thing, or at least, a good thing to support), it does a lot of work that will eventually be thrown away. Consider this example: I have a library with a function foo that returns a simple 42. I have a binary which calls foo from that library and prints the result. Now imagine the library is a hundred thousand lines of unrelated code to what the binary needs, but is useful for other people. Because compilation works in the "produce libraries, produce binary, link them all together" style model, you have to compile the entire library with all of that code, when all you need is really one function. That intermediate work is useful, and I'm picking an example that's deliberately extreme, of course.
I feel extremely validated reading this! I've had a pet theory for a while that compile times would be drastically reduced if every single item in a crate were implicitly under a cargo feature flag and then they only got enabled if they were imported (and used in non-dead code). I know that making something like that work isn't anywhere close to as simple as I'm describing it, and there are probably a million edge cases, but I've long felt that the ergonomics of Cargo features basically making it too annoying to expose everything conditionally (and then transitively expose all of the features from all of the direct dependencies as well so that things depending on your library could also only conditionally enable them) is secretly the reason that people think Rust compile times are slow, and seeing someone who has way more direct knowledge than me of how all of it works under the hood give a similar take makes me more confident that I might have been on to something all along.
Yet that doesn't prevent C++ to have REPL and hot reloading tools, use binary libraries instead of compiling the world from scratch, all of which allow for a much better experience.
(tongue in cheek) It seems recent history has shown that zig can get you to working software faster, then you can port it to rust once you are acquired or find market fit?
Maybe at some point in the future zig could add a rust compilation target ( like with `-ofmt=c` )...
I know this is mostly a joke but I have seriously wondered if “no hidden behavior” made it easier to port from Zig. Like, regardless of how easy or successful the port was in general, I think Zig’s explicitness may have worked in its favor.
This post is really interesting. As a member of the rust-analyzer team, I cannot avoid comparing it to the situation in Rust land. Rust famously has not less (or even more) sophisticated system for incremental compilation, yet its compilation is way slower. I attribute that to two main things:
- Language design. Zig was designed for fast and incremental compilation, Rust is just not. For instance, the post states that Zig has four properties (layout, type, value, body) that the compiler has to track for changes. Rust has much more, to the point that tracking them statically is just impossible, so the compiler uses a query system that tracks them dynamically, which adds overhead.
- Compiler implementation. Rust is much more complicated to compile than Zig, and rustc is both older and bigger (10x-20x LOC) than the Zig compiler, making changing it way harder.
There is something that I don't fully understand about this design: why are they insisting on building a giant binary for debug builds that contains all of the code? From my perspective, a simpler approach is to generate many smaller shared libraries (perhaps at the file level) and link them in to the final binary. With this approach, the program binary would have a tiny text section and a (potentially long) list of shared libraries to load. But even with thousands of shared libraries to load, the resulting program binary would not be all that long and there would be no need for binary patching.
I understand that for a release mode a single giant binary may be desirable, but I am struggling to understand this design for debug builds. Moreover, while reading this article, I found myself wondering what happens if the main binary becomes corrupted. Maybe the user cancels compilation with ctrl+c while it is patching the binary. Even if they have a story for avoiding and/or detecting corruption, it is simpler to not patch in the first place and always generate a new main binary. Again, this is reasonable because the new binary is mainly just a list of shared libraries to link which will not take up much space and can be written to disk quickly. Moreover, this process can be done recursively, e.g. at the subdirectory level, so that during incremental linking a few quite small shared libraries may be produced rather than patching in the new code and writing cascading relocations.
A quick test (C, clang) gives me that a binary depending on 1000 shared libraries, each containing a single function returning an integer, with a main function summing up the results of all those functions, takes ~270ms to run from the dynamic linker overhead. So you'd definitely want a good chunk below thousands.
(a process with 100 shared libraries takes 6ms to run, which is a lot better (0.9ms for 1 library, for reference), but, especially with in-place patching skipping work on unchanged values, static linking still has a good shot at beating dynamic linking, especially if you run the binary multiple times)
Incremental compilation generally already depends on its stored intermediate data not getting corrupted, and the final binary need not be any differently handled in that aspect.
If you choose to build static or dynamic libraries, I expect you’d still get to do that, and I expect as long as those don’t change and only the main binary needs patching it’d work the same. (Though I’m waiting for the next tagged release to really try it out myself, so that’s not like a guarantee or anything.)
I'm becoming a big fan of Zig every since learning about `zig cc` as a way of dipping my toes in it. I was already impressed by the build caching so I'm keen to play with this
It works through the fact that I specified "runtime" function ;)
A bit after that quote I have a note about `inline` functions in Zig, where I mention that they perform semantic inlining, which means dependencies triggered by the function actually get associated with the call site. Well, `comptime` function calls work just the same way---in fact, to the compiler, `comptime` calls are almost exactly identical to `inline` calls. So when we encounter a comptime function call, we start analyzing the ZIR for that function's body, but we don't switch our analysis unit, so comptime stuff doesn't really complicate the dependency graph at all (aside from the fact that it means you can depend on any number of source code hashes, instead of everything depending on exactly one).
With all that being said, there actually is a (completely unrelated) way in Zig to can depend on the body of a runtime function (hence why the quote includes "at least in the simplified view I'm presenting here"). It's to do with "inferred error sets" (IESes for short). If a function's return type is written `!T`, that means it can return an error, but we're asking the compiler to figure out exactly which errors are possible. So if at some point we need to know that set of errors (e.g. because the user has done some reflection to try and access the list of errors), that's where we get a dependency on a runtime function body, because we need to analyze the function body to learn about all the places it might return an error.
not quite! if your project is mixed C/Zig, then editing Zig would work but editing C would not. the Zig compiler caches Zig AIR but no information from C. plus, the C compilation can only really be done with LLVM and writing a self-hosted backend would be PITA for C
Not quite true, there is already a capable C compiler written in Zig (Aro/arocc), and a plan to transition to it for C compilation: https://codeberg.org/ziglang/translate-c
Using this native (written in Zig) C compiler to translate C source into Zig source as a part of the build, would presumably lend itself trivially to all the incremental logic in TFA, as updating C would update the generated Zig, and the incremental logic would detect differences just like it detects differences made by a human in an editor. Maybe there are aspects of the generated Zig that would complicate that somewhat, but I don't know -- just a warning about my ignorance.
This is part of plans to remove the hard LLVM dependency. AFAIK, the LLVM dependency will still be a variant many will use for the convenience of Zig as a much better clang, but removing the hard dependency is part of enabling all these great features like incremental compilation.
Thanks for the correction, I knew I was about 2 decisions out of date when I had the impulse to write my post, but my research to fill in my blanks was a bit faulty. I added the 2nd paragraph and "plan to" ending of 1st as an edit because I thought I had learned enough good info when writing the rest, but I was wrong.
It only works with our self-hosted code generation backends (the main one being for x86_64), which right now don't have any optimisation passes. It's planned that they will in future, but that's a long-term goal. Also, any optimisations which propagate information between functions (the most obvious and important one is inlining) are more-or-less incompatible with incremental compilation (I touch on this in the post IIRC), so once this does work it'll probably still be limited to a subset of optimisations.
TL;DR: only debug builds for now, could extend to release builds once we have our own optimisation passes one day, but some optimisations will still be inapplicable.
I've always thought this was fascinating, but the only incremental compilation I knew was obscure programming languages and Rust. Oh yeah, I guess Roslyn?
Yeah I remember being in sophomore year of college, watching Anders Hejlsberg's video on "the new way to build compilers" or something and having my mind blown. But I only ever looked at the source code for Rust when it came to something actually implementing this, so that came to mind first.
Roslyn also has an extra constraint of integrating with live editing on the fly; I think you can get simpler and/or have different constraints if your requirement is only incremental compilation.
It disappoints me how unseriously the industry has taken compilation speed for so long. I'm glad to see Zig doing incredibly valuable, high-impact work here.
Well, I remember back in the day Java was winning a lot of love from C++ devs over its much-much faster compilation than C++. It was important, even 15+ years ago. People praised faster iteration time while coding.
The incremental linking part sounds pretty hackish to me and I wonder what the price is in increased code complexity and maintenance effort.
It also does not mention how it deals with the patching of debug information.
That's a lot to follow, just to output a plan-text message, especially after this line: "The primary goal of Zig is to be a better solution to the sorts of tasks that are currently solved with C. A primary concern in that respect is readability…"
I don’t think you can judge a programming language based on its “Hello World.” AppleSoft BASIC has this:
10 PRINT “Hello World”
Beautifully simple and readable. But it’s not a good language by modern standards.
In your example, I see a lot of complexity being surfaced: output streams, locals instead of globals, error handling. I don’t know Zig but all of those are things that are important to address, and I like that the example doesn’t sweep them under the rug in pursuit of a false readability.
Everything it's doing it clear and readable. It's just not as easy to write. It streams the bytes to stdout using the default IO interface, and it can fail.
Alternatively, here's a simpler version (prints to stderr).
the given complicated version's complexity actually just comes from the fact that it's a "more correct" way to write a hello world program, as it manually acquires the stdout File object and acknowledges that printing to stdout can fail. the complexity has nothing to do with stdout vs stderr, you could just use `.stderr()` instead of `.stdout()` (same "friction", it's even the same number of characters). `std.debug.print` is only meant for debugging/development as it gets stderr for you and discards the errors that could happen when writing to stderr.
While most hello worlds do not check that the message was printed (which I assume writeStreamingAll does for you), dismissing the rest of the differences as "the others aren't correct" isn't really accurate.
Explicitly passing IO in is a fine design choice, but it's not a correctness issue to say others are wrong to not do so.
> Semantic analysis is the most difficult part of the compiler to handle incrementally. Perhaps unsurprisingly then, this is where language design starts to matter a lot: while I am pretty confident that most modern languages could support incremental compilation similar to how we do, certain design decisions can make that much more difficult. Zig has had its design tweaked over the years (sometimes controversially) specifically so that it is easier to support fast incremental compilation.
This is something I wish that we had done with Rust. It is impossible to do all of the things at once, though, and we already had a tremendous amount of things to do. This is also part of the "when do you ship 1.0" tradeoff; for our goals with the language, 2015 was the right moment to launch, but if had a few more years to bake things, maybe we could have made compile times way faster. Software engineering is hard.