Resizable structs in Zig

https://tristanpemble.com/resizable-structs-in-zig/

by rvrb • 12 months ago

152 73 12 months ago

rvrb7• 12 months ago

I am the author of this post, let me know if you have any questions or feedback :)

azemetre• 12 months ago

I sincerely mean this when I write this: please make more Girls on the Beach albums. I really love Splif Tape, reminds me of Julie Ruin during that timeframe as well (2015ish).

rvrb7• 12 months ago

oh, man! you made my day with this comment. those days are long behind me. we were young, dumb, and inebriated. if you're looking for the bands we were trying to sound like, there was a resurgence of this surf/girl group sound in the early 2010s. look for bands like Shannon and the Clams, Hunx and His Punx, Nobunny, Harlem, Ty Segall, Thee Oh Sees.. basically anyone on the now defunct Burger Records

atmikemikeb3• 12 months ago

I thought about dynamically sized types (DSTs) in zig recently. Was considering writing about it. I came to a different conclusion. Why not use zig's opaque?

It's pretty clean at this imo: Less metaprogramming but I think nicer to use in some cases.

  const Connection = opaque {
    pub const Header = struct {
      host_len: usize,
      // add more static fields or dynamic field lengths here
      //buff_len: usize,
    };

    pub fn init(a: std.mem.Allocator, args: struct { host: []const u8 }) *@This() {
      var this = a.allocWithOptions(u8, @sizeOf(Header) + host.len, @alignOf(Header), null);
      @memcpy(this[@sizeOf(Header)..], host);
      return this.ptr; 
    }

    pub fn host(self: *const @This()) []const u8 {
      const bytes: *u8 = @ptrCast(self);
      const header: *Header = @ptrCast(self);
      const data = bytes[@sizeOf(Header)..];
      const host = data[0..header.host_len];
      return host;
    }
  };

going off memory so I expect it to not actually compile, but I've definitely done something like this before.

rvrb7• 12 months ago

I would describe this approach as 'intrusive' - you're storing the lengths of the arrays behind the pointer, enforcing a certain layout of the memory being allocated.

Because the solution outlined in the article stores the lengths alongside the pointer, instead of behind it, there is room for it to work across an ABI (though it currently does not). It's more like a slice in this way.

You could in theory implement your opaque approach using this as a utility to avoid the headache of alignment calculations. For this reason, I think that makes the approach outlined in the article more suitable as a candidate for inclusion in the standard library.

atmikemikeb3• 12 months ago

Yeah I think mine is more about being able to provide a `host()` helper function instead of a `.get(.host)` meta function. It is somewhat boilerplate-y. I think it's really a matter of taste haha. Likely yours would be useful regardless if this is done a lot, since it abstracts some of it, if one wants that.

ethan_smith• 12 months ago

The opaque keyword in Zig doesn't support method definitions - it creates an incomplete type that must be implemented elsewhere, not a type that can have inline methods and fields like your example attempts.

pbaam• 12 months ago

https://ziglang.org/documentation/master/#opaque

opaque {} declares a new type with an unknown (but non-zero) size and alignment. It can contain declarations the same as structs, unions, and enums

So it can contain methods, just like structs. The only thing it cannot contain is fields (which the example above doesn't contain). An opaque type can only be used as a pointer, it's like a `typedef void* ...` in C but with the possibility to add declarations within a namespace

Edit: the documentation doesn't contain any example of an opaque type which happens to have declarations inside. But here is one: https://codeberg.org/andrewrk/daw/src/commit/38d3f0513bf9bfc...

mananaysiempre3• 12 months ago

Zig does not, and will not, have VLAs in the language spec. Instead, you can allocate a slice on the heap. If you want to have the data on the stack, use an array as a bounded backing store, and work with a slice into it[.]

Too bad, aligned byte-typed VLAs (and a license to retype them as a struct) are what you need to get stack allocation across ABI boundaries the way Swift does it. (A long long time ago, SOM, IBM’s answer to Microsoft’s COM, did this in C with alloca instead of VLAs, but that’s the same thing.) I guess I’ll have to use something else instead.

rvrb7• 12 months ago

I think there may be room to expand this implementation to support such a use case. Right now it enforces an `.auto` layout of the struct provided in order to ensure alignment, but its easy to imagine supporting an `extern struct` with a defined layout.

Conceivably, an implementation of this `ResizableStruct` that uses an array buffer as backing rather than a heap allocation, and supports the defined layout of an extern struct, could be used to work across the ABI

throwawaymaths3• 12 months ago

you can certainly allocate on the stack (like alloca). you just have to overallocate a compile-time known size and have some sort of fallback mechanism or fail if the size requested exceeds the amount created.

moreover since the stack allocator is just an allocator, you can use it with any std (or user) datastructure that takes an allocator.

AndyKelley8• 12 months ago

Note that not having runtime-known stack allocations is a key piece of the puzzle in Zig's upcoming async I/O strategy because it allows the compiler to calculate upper bound stack usage for a given function call.

At a fundamental level, runtime-known stack allocation harms code reusability.

Edit: commenters identified 2 more puzzle pieces below, but there's still one that didn't get asked about yet :P

Conscat• 12 months ago

Does anything stop a user from doing this with inline assembly?

AshamedCaptain• 12 months ago

How does this work given... recursion?

Even on languages without VLAs one can implement a simulacra of them with recursion.

omnicognate2• 12 months ago

there's still one that didn't get asked about yet :P

C libraries?

mananaysiempre3• 12 months ago

Note that not having runtime-known stack allocations is a key piece of the puzzle in Zig's upcoming async I/O strategy because it allows the compiler to calculate upper bound stack usage for a given function call.

That’s a genuinely interesting point. I don’t think known sizes for locals are a hard requirement here, though threading this needle in a lower-level fashion than Swift would need some subtle language design.

Fundamentally, what you want to do is construct an (inevitably) runtime-sized type (the coroutine) out of (by problem statement) runtime-sized pieces (the activation frames, itself composed out of individual, possibly runtime-sized locals). It’s true that you can’t then allow the activations to perform arbitrary allocas. You can, however, allow them to do allocas whose sizes (and alignments) are known at the time the coroutine is constructed, with some bookkeeping burden morally equivalent to maintaining a frame pointer, which seems fair. (In Swift terms, you can construct a generic type if you know what type arguments are passed to it.) And that’s enough to have a local of type of unknown size pulled in from a dynamic library, for example.

Again, I’m not sure how a language could express this constraint on allocas without being Swift (and hiding the whole thing from the user completely) or C (and forcing the user to maintain the frames by hand), so thank you for drawing my attention to this question. But I’m not ready to give up on it just yet.

At a fundamental level, runtime-known stack allocation harms code reusability.

This is an assertion, not an argument, so it doesn’t really have any points I could respond to. I guess my view is this: there are programs that can be written with alloca and can’t be written without (unless you introduce a fully general allocator, which brings fragmentation problems, or a parallel stack, which is silly but was in fact used to implement alloca historically). One other example I can give in addition to locals of dynamically-linked types is a bytecode interpreter that allocates virtual frames on the host stack. So I guess that’s the other side of being opinionated—those whose opinions don’t match are turned away.

Frankly, I don’t even know why I’m defending alloca this hard. I’m not actually happy with the status quo of just yoloing a hopefully maybe sufficiently large stack. I guess the sticking point is that you seem to think alloca is obviously the wrong thing, when it’s not even close to obvious to me what the right thing is.

do_not_redeem4• 12 months ago

A comptime_int-bounded alloca would achieve those goals, plus would be more space-efficient on average than the current strategy of always pessimistically allocating for the worst case scenario.

  @alloca(T: type, count: usize, upper_bound_count: comptime_int)

with the added bonus that if `count` is small, you can avoid splitting the stack around a big chunk of unused bytes. Don't underestimate the important of memory locality on modern CPUs.

travisgriggs• 12 months ago

Note that not having runtime-known stack allocations is a key piece of the puzzle in Zig's upcoming async I/O strategy because it allows the compiler to calculate upper bound stack usage for a given function call.

Sigh. So I have to choose between something I think might be useful, for something that too many languages have already soiled themselves with. Hopes that Zig has a better solution, but not optimistic.

Our stack compels me to work in Swift, Kotlin, Elixir, and Python. I use the async feature of Swift and Kotlin when some library forces me to. I actually preferred working with GCD before Swift had to join the async crowd. Elixir of course just has this problem solved already.

I frequently ask others who work in these languages how often they themselves reach for the async abilities of their languages, and the best I ever get from the more adventurous type is “I did a play thing to experiment with what I could do with it”.

ManDeJan• 12 months ago

How does this work with interrupts, say in an embedded context, that execute on the current stack, and that may in some cases, be interrupted themselves. Do you add the maximum stack depth of all interrupt routines that could go off at the same time?

h4ck_th3_pl4n3t• 12 months ago

I am wondering if this is more of an unclearly defined memory ownership problem rather than a problem of what types you have to use to interact with C ABIs or FFI calls.

I mean you could also just abstract the allocation away and handle it after the function pointer to your bridge, right?

pjmlp2• 12 months ago

A long long time ago, SOM, IBM’s answer to Microsoft’s COM, did this in C with alloca instead of VLAs, but that’s the same thing.

Was kind of the other way around, given the whole OS/2 versus Windows history, and that COM started as the evolution of OLE and VBX technologies, Windows 9X and Windows NT weren't as COM heavy as OS/2 was with SOM.

There was no COM to worry about on Windows 3.x back in 1991.

https://www.edm2.com/index.php/SOM_%26_DSOM_-_An_Introductio...

Also SOM was so much better, bettwen C++, Smalltalk and C, with support for meta-classes and proper inheritance implementation across such disparate languages.

konstantinua002• 12 months ago

one thing I never understood about VLAs - discussion about them always hits a "can't put it on stack safely" and gets halted, forever

why not to make it heap-only type? it seems such a useful addition to type system, why ignore it due to one usecase?

Out_of_Characte2• 12 months ago

Because arrays simply do not deal with fragmentation. Yes, you could probaly get decent performance on a modern system that has memory overcommit strategy where you could allocate sparse adress ranges where you would probaly never run out of pointers unless you actually write to your variable array.

But its just kind of mediocre and you're better off actually dealing with the stack if you can actually deal with certain fixed sizes.

konstantinua002• 12 months ago

...what are you talking about?

array-like storage with dynamic size has existed since forever - it's vector. over or undercommitting is a solved problem

VLA is the way to bring that into type system, so that it can be it's own variable or struct member, with compiler auto-magic-ing size reading to access members after it

ori_b• 12 months ago

Those effectively exist. They're called slices.

uecker2• 12 months ago

You can also put them safely on the stack. The VLA is discussion is just irrational.

quotemstr• 12 months ago

Zig articles tend to get a little too excited about rediscovering longstanding techniques.

The author has described a metaprogramming utility for allocating a contiguous hunk of memory, carving this hunk into fields (in the article's example, a fixed-sized Client header, then some number of bytes for host, then some number of bytes for read_buffer, and then some for write_buffer). I'll acknowledge the syntax is convenient, but

1. we've done this since time immemorial in C. See https://learn.microsoft.com/en-us/windows/win32/api/evntcons...

2. you can implement this pattern ergonomically in C++, and even moreso once the C++26 reflection stuff comes online

3. the zig implementation is inefficient. It desugars to

  const Connection = struct {
    ptr: [*]u8,
    lens: struct {
      host: usize,
      read_buffer: usize,
      write_buffer: usize,
    }
  }

That first pointer is needless indirection and probably a cache miss. You should (unless you have specific performance data showing otherwise) store the sizes in the object header, not in an obese pointer to it. (It's bigger than even a fat pointer.)

rvrb7• 12 months ago

You have raised a good discussion point or two, but I am not inclined to engage with them due to the tone you've created with the rest of your comment.

Would it be productive to jump into a thread on a Ruby article and puff your feathers about how you've always been able to do this in Perl, and also in Python 4 you can do XYZ? I don't think so.

For whatever reason, inevitably in threads on systems languages, someone comes in and postures themselves defensively like this. You might want to reflect on that.

AndyKelley8• 12 months ago

I think you're misinterpreting OP's excitement here. The technique isn't novel but the ergonomics are.

you can implement this pattern ergonomically in C++,

lmao

rvrb7• 12 months ago

I think you're misinterpreting OP's excitement here. The technique isn't novel but the ergonomics are.

This, of course, but I also consciously made the decision to write for an audience less familiar with the language and concepts being discussed. That excitement can translate to engaging a less disgruntled and more curious reader than the OP.

pjmlp2• 12 months ago

While C++ isn't without warts, and WG21 does indeed a good job adding a few more, many time people aren't able to do ergonomic stuff in C++, because they insist in using it like C with Classes.

qalmakka• 12 months ago

The author is literally proposing to implement arrays of variant types

edem• 12 months ago

Should I try Zig? I wanted to learn something that's more low level than the JVM/Node and i have been contemplating rust and go. zig never occurred to me until now.

hiccuphippo• 12 months ago

I think it's simpler than rust, the basics are easier to learn and will get you a long way before you need the more advanced stuff. Go might be easier because it's garbage collected. If you want something closer to C, but not C itself, Zig is a good choice.

Resizable structs in Zig

Related Stories