Ability to remove retain/autorelease ARC calls in Objective-C Binaries

[jonpalmisc]

There are numerous functions that produce a lot of noise in Objective-C code, examples include:

• _objc_retainAutoreleasedReturnValue
• _objc_retainAutorelease

These are not often the interesting bits of the function, and it might be helpful if these could be optionally hidden to produce more readable decompilation.

[0cyn]

PR Vector35/workflow_objc#52 addresses this, but is blocking on #5563

[0cyn]

Workarounds exist for this:

This plugin: https://github.com/bdash/bn-objc-extras/blob/dev/src/remove_memory_management.rs implements a workflow to remove these functions .

This very legacy PR to workflow_objc also implements this: https://github.com/Vector35/workflow_objc/pull/52/files However it is likely outdated and non-functional in current versions of the product.

[bdash]

My ARC removal pass works quite well. I'm happy to look at upstreaming it, though I'll need to figure out if it's more painful to rewrite it in C++ or to try to integrate Rust code with the Mach-O / shared cache plug-ins. If it stays in Rust it will need to be reworked to move off my custom instruction matching to directly using the low_level_il types.

There's one remaining rough edge where the resulting HLIL is confusing:

// TODO: Detect calls to `objc_release` that are immediately after a load of a struct field.
// It might be preferable to leave those in place since otherwise the load is left behind.

This shows up as a mystery reference to a struct field:

100069a14    id -[ScanResult init](struct ScanResult* self, SEL sel)

100069a14    {
100069a14        struct ScanResult* result = [super init];
100069a14        
100069a48        if (result) {
100069a54            struct XPScanResult* x0_1 = [XPScanResult new];
100069a58            result->xpScanResult
100069a5c            result->xpScanResult = x0_1;
100069a48        }
100069a48        
100069a78        return result;
100069a14    }

If you disable the pass you can see that the code is releasing any existing object before assigning to the field:

100069a14    id -[ScanResult init](struct ScanResult* self, SEL sel)

100069a14    {
100069a14        struct ScanResult* result = [super init];
100069a14        
100069a48        if (result) {
100069a54            struct XPScanResult* x0_1 = [XPScanResult new];
100069a58            struct XPScanResult* xpScanResult = result->xpScanResult;
100069a5c            result->xpScanResult = x0_1;
100069a64            [xpScanResult release];
100069a48        }
100069a48        
100069a78        return result;
100069a14    }

This can likely be improved after the fact.

[bdash]

#7440, first appearing in 5.2.8343-dev, added initial support for this. It is disabled by default since it rewrites IL in a way that changes program semantics. It can be enabled per-function via Obj-C: Remove reference counting calls in the Function Settings context menu or on a wider scale by setting core.function.objectiveC.removeMemoryManagement.

There are two limitations at present:

1. Only arm64/arm64e are supported.
2. The variants of objc_retain / objc_release that use a custom calling convention are not removed.

Supporting armv7/x86_64 requires generalizing the IL matching code. Supporting objc_retain_x8 and such requires generalizing the IL rewriting of it.

[bdash]

Closing as fixed since the most common cases are handled in Binary Ninja 5.2. #7660 tracks the cases that are not yet handled (support for x86_64 and runtime functions with custom calling conventions).