Posts Calling Super at Runtime in Swift
Post
Cancel

Calling Super at Runtime in Swift

While working on InterposeKit, I had a rather specific need: Create an implementation that simply calls super, but at runtime instead of at compile time. Doesn’t sound so hard, does it? Well, here we go again.

How Does Super Work?

Let’s say you have an empty UIViewController subclass and override viewDidLoad like this:

1
2
3
override func viewDidLoad() {
     super.viewDidLoad()
}

Simple enough! What the compiler creates for you is something along these lines:

1
2
3
4
5
6
7
 - (void)viewDidLoad {
     struct objc_super _super = {
         .receiver = self,
         .super_class = object_getClass(obj);
     };
     objc_msgSendSuper2(&_super, _cmd);
 }

In compiled code, there’s a lookup table so that object_getClass doesn’t need to be called, but you see the principle. A struct is created, and objc_msgSendSuper2 is called with it. The method is automatically dynamic, since UIKit is written in Objective-C, so the Swift compiler knows it needs to use dynamic dispatch.

objc_msgSendSuper vs. objc_msgSendSuper2

There are two versions of the super call, and the difference is minor but important: objc_msgSendSuper starts looking for an implementation of the selector at the current class, which would cause an endless loop in the above code, while objc_msgSendSuper2 looks for the superclass.

With objc_msgSendSuper it is possible to skip multiple overridden implementations and directly target what you want to call. objc_msgSendSuper2 is better tailored for a dynamic lookup, as it always fetches the superclass at runtime, so this even calls the correct super implementation if the class hierarchy is changed at runtime.

I’ve seen the compiler only emit objc_msgSendSuper2, but both need to be there forever, as they are both ABI.

Being the Compiler

Calling super seems fairly straightforward! Fill the struct, call the method, done — right? There are however a few problems with that.

For one, while objc_msgSendSuper is in the objc/message.h header, the -2 version is not included in the public headers of the runtime. It is not private API, since Clang creates these calls. The runtime is also open source, so we can copy the header and call it:

1
2
3
4
5
6
7
8
9
// https://opensource.apple.com/source/objc4/objc4-493.9/runtime/objc-abi.h
OBJC_EXPORT id objc_msgSendSuper2(struct objc_super *super, SEL op, ...);
class_addMethod(clazz, selector, imp_implementationWithBlock(^(__unsafe_unretained id self, va_list argp) {
        struct objc_super super = {
            .receiver = self,
            .super_class = class_getSuperclass(clazz)
        };
        return ((id(*)(struct objc_super *, SEL, va_list))objc_msgSendSuper2)(&super, selector, argp);
    }), types);

This works, and we’ve been shipping code like this for a while. For InterposeKit, I wanted to write the same in Swift.

Update: Marcel correctly pointed out that I am mixing up va_arg and va_list, which are absolutely not the same.

Compare the two definitions of printf:

1
2
int printf(const char * restrict format, …);
int vprintf(const char * restrict format, va_list ap);

It is possible to convert the first version into the second (via va_start), but not vice versa. So this code as-is works by pure luck.

Super in Swift

An additional goal was to be “pure” Swift, not because I’m a purist, but because SwiftPM doesn’t yet support mixed language projects.1 We can’t write a C header to import objc_msgSendSuper2, but we sure can look it up at runtime:

1
2
let handle = dlopen(nil, RTLD_LAZY);
let sendSuper2 = dlsym(handle, "objc_msgSendSuper2");

This works, and it’s a common trick to avoid C headers — it’s slightly slower, but as long as you cache the result, it should be hardly measurable:

1
2
3
4
5
6
7
8
9
10
let block: @convention(block) (AnyObject, va_list) -> AnyObject = { obj, vaList in
    let raw = Unmanaged<AnyObject>.passUnretained(obj)
    // https://bugs.swift.org/browse/SR-12945
    let superStruct = objc_super.self(receiver: raw, super_class: subclass)
    return withUnsafePointer(to: superStruct) { superStructPointer -> AnyObject in
        return unsafeBitCast(sendSuper2, to: (@convention(c) (UnsafePointer<objc_super>, Selector, va_list) -> AnyObject).self)(superStructPointer, self.selector, vaList)
    }
    // Equivalent in C:
    // return ((id(*)(struct objc_super *, SEL, va_list))objc_msgSendSuper2)(&super, selector, argp);
}

But this doesn’t compile. The Swift compiler crashes in various flavors, which I reported with examples. Since InterposeKit is small, they should be useful to reproduce:

I found a cursed workaround, and indeed the super call logic works. However, someone on the internet quickly told me that I’m wrong.

Greg worked on the Objective-C runtime for many years, so if he tells you something is only working by “blind luck,” it’s not something you should ship. He also has a really interesting blog called Hamster Emporium, if you’re into understanding low-level things.

Casting Objective-C Message Sends

The thing I got wrong is something many folks struggled with: The function looks like it takes a va_list, but it really doesn’t. This problem caused enough issues that Apple changed2 the method signature of both objc_msgSend and objc_msgSendSuper with the release of Xcode 11:

1
2
3
4
5
6
7
#if !OBJC_OLD_DISPATCH_PROTOTYPES
OBJC_EXPORT void objc_msgSend(void /* id self, SEL op, ... */ )
OBJC_EXPORT void objc_msgSendSuper(void /* struct objc_super *super, SEL op, ... */ )
#else
OBJC_EXPORT id _Nullable objc_msgSend(id _Nullable self, SEL _Nonnull op, ...)
OBJC_EXPORT id _Nullable objc_msgSendSuper(struct objc_super * _Nonnull super, SEL _Nonnull op, ...)
#endif

Previously, it was declared as a function that took id, SEL, and variadic arguments, returning id — now it takes and returns void. Why the change?

The short version is that there is no guarantee that the ABI for variadic function matches the ABI for a function with a mixed number of arguments. In the ARM64 ABI, variadic arguments are passed on the stack. However, Apple changed the way message sending works in ARM64 to not use the variadic ABI anymore, instead using the regular function-calling ABI.

Without casting, even a trivial use of objc_msgSend will result in a crash. There is an interesting article about this by Mike Ash entitled objc_msgSend’s New Prototype. Mike’s blog is brilliant, and I’m extremely happy that he still writes new posts from time to time, despite now working at Apple.

Accepting Assembly

The root problem is that objc_msgSend cannot be implemented in C, not at any speed.3 We cannot build dynamic parameter lists. Usually the compiler takes care of casting objc_msgSendSuper for us, but this isn’t something it can do when we try to do this at runtime. The only way to call this correctly without getting lucky is if we write the call in assembly.

First of all, assembly is hard, but it’s a useful skill that will make you better at debugging, so I’ve approached this entire thing as a “fun“ challenge. The most important part to know is what each register does. To keep things simple, we focus on ARM64 in this article, and we’re using the AT&T4 syntax.

Me trying to make sense of this via drawing

Caller-saved registers (“clobbered”) are registers you can freely work with and use as temporary variables. It’s normal that a call writes temporary values into these registers. Some of them (specifically x0-x7) are used to transport parameters when calling other functions.

Callee-saved registers (“call-preserved”) are registers that are expected to stay the same after your function returns. The best idea is to simply not touch them.5

There are some other special registers, such as the fp frame pointer (usually an offset of the stack pointer), the lp link register (holds the address to return to when a function completes), and the sp stack pointer (holds the address of the stack buffer).

There are also floating-point registers (q0q7), but we don’t need them for our task here.

Assembly and Swift

Unlike Rust, Swift doesn’t yet have a way to add inline assembly. Both have the approach to be a system language, and there are hacks to get inline assembly working, but I haven’t seen an evolution proposal so far.

Adding inline assembly is a niche feature, but it has valid use cases; even Chris Lattner is hoping that a future version of Swift will include it. For now, we can use C and the Swift/Obj-C interop to write assembly.

Perfectly Forwarding Arguments

Back to calling super: The goal is to perfectly forward all arguments from the caller to objc_msgSendSuper2, while also changing the first argument from self to struct objc_super, and potentially also filling this struct. Sounds easy enough!

My first inspiration was SGVSuperMessagingProxy. This uses an extremely clever trick of creating a proxy at runtime with exactly one ivar, which is the prefilled super struct. So the trampoline boils down to this:

1
2
3
4
5
6
__attribute__((__naked__)) \
void trampolineFunction(void) { \
asm volatile ("add " #selfLocation ", " #selfLocation ", #" #offset "\n\t" \
"b " #msgSendSuperFunction "\n\t" \
: : : "x0", "x1"); \
}

Class Variable Layout

What this code here so cleverly does is that it simply adds eight bytes to the location of self. The layout of classes looks like this:

Class Memory Layout: [[ISA] [IVARs]]

Remember, in ARM64, the caller arguments are in x0 to x7. x0 here is the pointer to self, the class object, which is where the isa pointer is located. isa means “is a.” Every Objective-C object (including every class) has an isa pointer as first variable.6 If we increment by 64-bits = 8 bytes, we get to the next storage location, which is where the object variables (ivars) are stored.

This is what the assembly looks like without calling boilerplate:

1
2
add x0, x0, 8
b objc_msgSendSuper2

This is beautiful, since it’s very simple, and it doesn’t touch any of our calling registers — with the exception of the one that needs to be changed. This doesn’t work in my case though — the goal was to create a super call in an existing class hierarchy, not via creating a new proxy where we have exact control of the memory layout.

Trampolines Explained

In other architectures, we would just generate the assembly on the fly, changing the offset as needed. Having memory pages that are both writable (PROT_WRITE) and executable (PROT_EXEC) requires a dynamic-codesigning entitlement from Apple, which is something only very few system processes, such as JavaScriptCore, get — certainly not a third-party app. And while there are ways around this, jailbreaking or attaching a debugger aren’t realistic if we want to ship this.

Another solution is that of trampolines. The basic principle is that you have two pages next to each other with a fixed offset and a large number of entry points for your implementation:

1
2
3
4
5
6
7
8
9
10
11
12
13
               ┌───────────────────┐
            ┌──┤Trampoline Entry 1 │ 0x1000 Read, Execute
            │  ├───────────────────┤
            │  │Trampoline Entry 2 │ 0x2000 Read, Execute
 base + 0x3000 ├───────────────────┤
  (fix offset) │Trampoline Entry 3 │ 0x3000 Read, Execute
            │  ├───────────────────┤
            └─▶│Trampoline Data #1 │ 0x4000 Read, Write
               ├───────────────────┤
               │Trampoline Data #2 │ 0x5000 Read, Write
               ├───────────────────┤
               │Trampoline Data #3 │ 0x6000 Read, Write
               └───────────────────┘               

With a fixed offset, we can reach the corresponding data from the entry, and we can read variables as needed. This is how imp_implementationWithBlock works, and luckily it’s also open source — but there be dragons. Landon Fuller reimplemented this back when it was introduced in iOS 4.3, and he explains the principles really well.

Tail Calling

There’s a lot of logic required to correctly manage tables, and things need locking to make the code thread-safe. I decided that this is gonna be the backup plan and tried a more direct approach to just fetch everything at runtime.

The principle: We save the registers that we might spill, fill the struct at runtime, restore the registers, and then perform the tail call. This sounds simple now that I write it up, but it caused serious headaches at first.

Specifically, I tried to use the stack to generate the struct, which breaks stack-based parameter passing. I tried calling malloc in asm, but since that requires calling free, I couldn’t do the tail call optimization anymore. And I encountered oh so many crashes because I didn’t really understand what it means to align the stack pointer on 16 bytes.

Let’s start by saving registers:

1
2
3
4
5
6
// push {x0-x8, lr} (call params are: x0-x7)
"stp x8, lr, [sp, #-16]!\n" // lr = link register
"stp x6, x7, [sp, #-16]!\n"
"stp x4, x5, [sp, #-16]!\n"
"stp x2, x3, [sp, #-16]!\n" // push x3, then x2
"stp x0, x1, [sp, #-16]!\n" // push x1, then x0

stp saves a pair of registers (2*8=16 bytes) on the stack, and it also automatically decrements the stack pointer. The stack on most architectures grows downward from max to 0, so via decrementing, we reserve memory:

1
2
// fetch filled struct objc_super, call with self + _cmd
"bl _ITKReturnThreadSuper \n"

bl means “branch with link,” and it calls a function — in this case, a C function. The same call arguments exist here, so the first parameter will be self, and the second will be _cmd. bl and b are similar; however, bl stores the address of the next instruction into the lr register, and therefore the called function can jump back via ret.

C Helpers

1
2
3
4
5
6
7
8
9
// One thread local per thread should be enough.
_Thread_local struct objc_super _threadSuperStorage;

struct objc_super *ITKReturnThreadSuper(__unsafe_unretained id obj) {
    struct objc_super *_super = &_threadSuperStorage;
    _super->receiver = obj;
    _super->super_class = object_getClass(obj);
    return _super;
}

The C helper ITKReturnThreadSuper is fairly trivial; it fills the objc_super struct and returns it. In early versions, I simply called malloc() and used a dispatch_async to later free it — a pretty horrible first hack, but it worked. This version uses a thread-local storage. _Thread_local was added in C11 and only works on global variables, but for our use case, this should work just fine — even if we use objc_super multiple times in a call stack.

It’s important to not go wild here: We did not save floating-point registers — only the bare minimum — so don’t call random functions in here. This is more dangerous than you think! Even a simple memcpy could override floating-point parameters.

Also see that object_getClass, and not the class method, is used here. While the latter can be overridden so a class “lies” about its type, this always returns the correct type. This is important since Apple uses this trick for key-value observing, which creates a subclass at runtime but also overrides class to hide this fact.

This also implies that the solution here is NOT a general super call, but will only work on the outermost object level. If you subclass again, this will no longer work correctly:

1
2
3
// first param is now struct objc_super (x0)
// protect returned new value when we restore the pairs
"mov x9, x0\n"

Once we return, in ARM64, the return value is in x0.7 We temporarily store this in the “scratch space” register set; x9-x15 are free to use. Another word for this is caller-saved or clobbered. Why do we do that when x0 is already exactly what we want? Because on ARM64, we can only operate on the stack in 16 bytes, so we always restore pairs of registers:

1
2
3
4
5
6
// pop {x0-x8, lr}
"ldp x0, x1, [sp], #16\n"
"ldp x2, x3, [sp], #16\n"
"ldp x4, x5, [sp], #16\n"
"ldp x6, x7, [sp], #16\n"
"ldp x8, lr, [sp], #16\n"

While there are ways around this, they are less elegant and require even more assembly. After restoring the registers, we’re now almost ready for the super call:

1
2
3
4
// get new return (adr of the objc_super class)
"mov x0, x9\n"
// tail call
"b _objc_msgSendSuper2 \n"

We copy x9 back to x0 and then call objc_msgSendSuper2 with b, not saving the link registry and thus performing a tail call.

Assembly Notes

That’s it. You can see the result for both architectures and also the objc_msgSendSuper2_stret variant for struct returns on GitHub.

Luckily, we currently only need x86_64 and arm64, and one day we might even be able to drop Intel altogether. Apple removed support for armv7 (32-bit arm) in iOS 11 and i386 with macOS Catalina, so I didn’t write variants, although it wouldn’t be so hard, as the principles are the same.

After being almost done with this, Joe Groff pointed out that there’s another (although less efficient) way to not need assembly for my specific case — but having a generic super logic has many other useful possibilities, and it was a great learning experience.

Now I’d love to hear from you. Is what I do here correct? Does this make sense? Is there a better way?

Bonus Content: Using Your New Assembly Superpowers

Now that you understand assembly, we can look at what the compiler really generates with a super call:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
// Stack preservation
sub	sp, sp, #48             ; =48
stp	x29, x30, [sp, #32]     ; 16-byte Folded Spill
add	x29, sp, #32            ; =32
// Fetch data
adrp	 x8, _OBJC_SELECTOR_REFERENCES_@PAGE
add	x8, x8, _OBJC_SELECTOR_REFERENCES_@PAGEOFF
adrp	 x9, _OBJC_CLASSLIST_SUP_REFS_$_@PAGE
add	x9, x9, _OBJC_CLASSLIST_SUP_REFS_$_@PAGEOFF

// build objc_super struct on the stack
stur 	x0, [x29, #-8]
str	x1, [sp, #16]
ldur x10, [x29, #-8]
str	x10, [sp]
ldr	x9, [x9]
str	x9, [sp, #8]
ldr	x1, [x8]
mov	x0, sp

// regular call
bl	 _objc_msgSendSuper2
ldp	x29, x30, [sp, #32]     ; 16-byte Folded Reload
add	sp, sp, #48             ; =48
ret

You can debug this via Debug > Debug Workflow > Always Show Disassembly. Step through commands via ni and is.

Xcode Assembly Debugger

Right before the call to _objc_msgSendSuper2, we can inspect the objc_super struct:

1
2
3
4
5
(lldb) po *(id *)$x0
<ViewController: 0x102607ca0>

(lldb) po *(id *)($x0+8)
ViewController

Further Resources

Thanks to everyone helping me with this post, especially @badlogicgames, @DavidJGoldman, and @mpweiher.

I used these resources to learn:

  1. SwiftPM supports multiple modules, so things can be split up. There are also plans to lift this limitation eventually. 

  2. OBJC_OLD_DISPATCH_PROTOTYPES has been an option for many years, but Apple only recently changed the default. 

  3. The old GNU runtime used objc_lookup(receiver-class, SEL)(receiver, SEL, …), a different approach altogether. 

  4. Yes, there are two concurring syntax branches. Intel syntax is popular in the DOS and Windows world, and AT&T syntax is for Unix. AT&T is source before destination, while Intel is destination before source. 

  5. There’s no need to deal with this register type for our code here, but compilers sure use these. There’s a good answer on Stack Overflow that explains this in more detail. 

  6. Swift uses the same concept, but it has a second variable in there, so the offset would be 16. SGVSuperMessagingProxy works with any function marked as dynamic, not just Objective-C. Pretty amazing to see how new things still map to old concepts! 

  7. This is true for our simple C function — large structs might use an indirect return (x8). 

This post is licensed under CC BY 4.0 by the author.