Hacker News Reader

Comment by eridius | Hacker News Reader

Doesn't at least objc_msgSend have to be assembly?

You can do it in plain C, but it will be quite slow.

It isn't only written in Assembly, it is written in clever optimised Assembly fine tuned across several compiler releases, to minimise as much as possible the impact of dynamic dispatch.

I don't remember why, but years ago I remember coming to the firm conclusion that it was impossible to write it in anything but assembly.

I don't see why the naive approach to a possible implementation cannot be done in C, even if using extensions to ANSI C.

Don't tell that to the GCC Objective C runtime.

With the NeXT/Apple runtime,

    [object selector: arguments...]

is syntactic sugar for

    objc_msgSend(object, @selector(selector:), arguments)

With the GCC runtime, that's syntactic sugar for:

    IMP tmp = objc_msg_lookup(object, @selector(selector:));
    tmp(object, @selector(selector:), arguments...)

No assembly necessary. I believe there are also compiler flags to assume the function won't be replaced behind your back so objc_msg_lookup can be hoisted out of a loop, etc.

objc_msg_lookup is very emphatically not objc_msgSend.

Lookup operates differently from the nominally tail-calling msgSend in that it cannot support transparent forwarding (where the receiver is switched to another object before dispatch.) Since objc_msgSend is given control over the final call, the destination can be swapped if necessary.

It's not possible[1] to implement objc_msgSend in plain (portable) C, as it needs to pass on an unknowable number of arguments without damaging the ones passed in registers, the stack frame, or the optional hidden return value pointer. Any methods attempting to do so using, say, naked functions and the method's type encoding will be more fragile than a perfect argument forwarding stub written in assembly.

[1]: __builtin_apply_args, __builtin_apply and __builtin_return (https://gcc.gnu.org/onlinedocs/gcc-5.1.0/gcc/Constructing-Ca...) notwithstanding.

GCC's C is emphatically not "plain" (ANSI/ISO) C, but that does not mean it isn't "portable" C, given that GCC itself targets many platforms, and other compilers (clang, at least) do attempt to support GCC's dialect.

I couldn't find documentation that supports Clang offering the __builtin_apply intrinsics, but their availability would make a C objc_msgSend somewhat more palatable.

Ah. I don't mean to be too argumentative. You've helped me to understand why objc_msgSend() could not be done in standard C, and I thank you for that :)

However, in the case of GCC's ObjC implementation, I don't see why it couldn't be implemented in GCC's dialect, thus obviating any need for assembly.

Pardon me while I reflect on (and lament) the fact that we're still stuck using such crude and ancient tools for systems programming...

FWIW, it appears that Mulle-objc gets around this by actually changing the calling convention for Obj-C methods, so the IMP takes a single parameter (beyond self and _cmd) which is a struct that contains all of the method parameters. This way the compiler can just synthesize the struct for you, and then objc_msgSend() itself can be written in C because it just needs to pass the pointer to the struct along to the IMP rather than messing about with forwarding arbitrary arguments.

Interesting. Thanks!

I don't see anything w.r.t. objc_msgSend() that obviously prevents it from being written in C, as opposed to objc_msg_lookup(). I mean, worst case scenario, could it not just be:

    id objc_msgSend(id object, id selector, ...) {
        IMP tmp = objc_msg_lookup(object, selector);
        return tmp(object, selector, ...);
    }

(forgive me for any mistyping here, as I'm not an Obj-C programmer, and for the varargs shorthand, to keep things simple)? Especially given that GCC supports some extensions to C that could reduce the overhead a bit (tail-recursion elimination, for example), I don't see why such a function couldn't be implementable in (GCC) C.

Would you be willing to enlighten me?

It's the ... that's the hard part; I've explained a little more here: https://news.ycombinator.com/item?id=13046391.

Conceptually, it is:

    IMP tmp = objc_msg_lookup(object, selector);
    goto *tmp;

but C can't handle that (and even with gcc computed gotos you need to preserve any argument registers and not screw up the stack).

objc_msgSend has a variable number of arguments and no way to determine how many there are.

No. Why? What's magic about assembly code that means you couldn't do the same thing in C?

While in this case I can't think of any specific reasons beyond performance to need to implement objc_msgSend in assembly, there are absolutely instances where assembly is the only option.

Example: when performing a context switch in a kernel, assembly is required since C abstracts away architecture-specific details necessary to swap out the register set, swap the page directory pointer, etc.

My ObjC-fu is pretty old --- the last time I looked at it seriously was when poc was still a thing --- but my understanding is:

Traditionally objc_msgSend is in assembly because it doesn't know how many and what types the message parameters are. So it can't portably preserve them while it does the computations to find out where to send the message to.

This project's alternative obj_msgSend always passes parameters in a structure. That helps, because now you know that all message calls have three parameters; the destination object, the selector, and the parameter block.

However... they don't always pass the return value in the message block; instead it coerces any integer types to void* and returns them in the usual way, only using an extra field in the parameter block for those values which aren't coerceable to void.

This does make me kinda unhappy. I don't think* it'll be a problem, but... it's definitely pushing the limits of what C allows; round-tripping arbitrary integers through pointers is one of those things that's not technically allowed but all real-world compilers support, so, um. I'd much rather they just pass everything through the parameter block.

But yeah, it's a decent enough idea, which simplifies things considerably at not a lot of performance hit. I'm kind of surprised that ObjC doesn't already do that.

My main disappointment is that this is a clang variant and not a standalone preprocessor. Since poc died there isn't a working ObjC preprocessor, which means if you want to use ObjC you're limited to gcc or clang. An ObjC->C preprocessor would solve that.

I think any real world systems C project will have undefined behaviour. I'm sure HotSpot, V8, .net etc are all totally undefined.

the objc_msgSend function doesn't know what it's return type is or how many arguments it has. It needs to lookup the actual function then jump to it without modifying the stack or any of the parameter registers. It simply doesn't follow the calling conventions and C can't handle it.

Right, true. I guess I'm used to working on languages where you can stick everything into a uint64, but that isn't the case with something like returned structs. if you controlled the calling convention you could do it in C.

It appears that controlling the calling convention is exactly what Mulle-objc is doing. It's layering a new convention on top of C calling conventions, where objc_msgSend() is given a pointer to a struct containing the arguments. This way the compiler can synthesize the struct for you, and objc_msgSend() itself can be written in vanilla C.

Reply on news.ycombinator.com