lld/mac category behavior doesn't match ld64

[keith]

Bugzilla Link	52398
Version	unspecified
OS	All
CC	@gkmhub,@int3,@smeenai

Extended Description

With this Objective-C code:

#import <Foundation/Foundation.h>

@interface NSArray (first)
- (void)firstMethod;
@end

@implementation NSArray (first)
- (void)firstMethod {}
@end

@interface NSArray (second)
- (void)secondMethod;
@end

@implementation NSArray (second)
- (void)secondMethod {}
@end

int main() {
    return 0;
}

The produced binary has significantly different symbols than ld64:

--- ld64
+++ lld
@@ -1,7 +1,10 @@
 -[NSArray(first) firstMethod]
 -[NSArray(second) secondMethod]
 _OBJC_CLASS_$_NSArray
+__OBJC_$_CATEGORY_INSTANCE_METHODS_NSArray_$_first
+__OBJC_$_CATEGORY_INSTANCE_METHODS_NSArray_$_second
 __OBJC_$_CATEGORY_NSArray_$_first
-__OBJC_$_INSTANCE_METHODS_NSArray(first|second)
+__OBJC_$_CATEGORY_NSArray_$_second
 __mh_execute_header
 _main
+dyld_stub_binder

It's unclear to me if differences like this will be a major issue, but I assume we're hoping to match ld64 more closely than this?

[smeenai]

Linking with -no_objc_category_merging makes ld64 mimic LLD's behavior. I'm not well-versed enough with the Objective-C runtime to understand the implications of merging or not merging, but I've asked a coworker to comment.

[keith]

With this example and ld64 from Xcode 13 using that flag doesn't appear to make a difference in its output. I wonder if it matters if they're from different inputs to the link or not in that case.

[smeenai]

Ah, possibly. I was testing as follows:

$ cat cat.m

your source code

$ clang cat.m -framework Foundation -Wl,-no_objc_category_merging
$ nm -U a.out
0000000100003f50 t -[NSArray(first) firstMethod]
0000000100003f60 t -[NSArray(second) secondMethod]
0000000100008000 s _OBJC$CATEGORY_INSTANCE_METHODS_NSArray$_first
0000000100008060 s _OBJC$CATEGORY_INSTANCE_METHODS_NSArray$_second
0000000100008020 s _OBJC$CATEGORY_NSArray$_first
0000000100008080 s _OBJC$CATEGORY_NSArray$_second
0000000100000000 T __mh_execute_header
0000000100003f70 T _main

This is with Xcode 13, ld64-711.

To be clear, I imagine we'll want to implement category merging and other Obj-C optimizations in LLD as well, for parity with ld64. I was hoping for someone more knowledgeable about the Objective-C runtime to be able to give some sense of the importance of these optimizations though, so we can decide how to prioritize them.

[int3]

Possible dupe (or subset) of #48630

[keith]

Yep definitely, if you don't think we need this more specific one we can close

[int3]

The extra detail here is good, let's keep this open. (It's a pity GitHub doesn't support blocking dependencies between issues...)

[BertalanD]

This was implemented in #82928 🎉

I suppose this can be closed once it's enabled by default (cc @alx32)