objc object file generated for gnustep runtime for ELF target is too big

Dear community,

I'm using gnustep runtime -fobjc-runtime=gnustep with gnustep-libobjc2 (https://github.com/gnustep/libobjc2) and
    
Cocotron/Chameleon.
    
For following source file
    
#import <Foundation/Foundation.h>
      #import <UIKit/UIKit.h>
    
int main(void)
        {
         NSString *str = [NSString stringWithCString:"TEST"];
         NSLog(@"test object %@", str);
         return 0;
        }
      
Compilation in xcode's clang
xcrun -sdk iphoneos clang -arch armv7s -mios-version-min=6.1 -c test_foundation.m -o test_foundation_apple.o
gives following result:
1. object file has only 3156 bytes size
2. and in disassembler only main function
3. the final executable will be 49992 bytes size in this case

But compilation in llvm's clang downloaded from http://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-apple-darwin.tar.xz
    
(
clang version 3.8.0 (tags/RELEASE_380/final)
      Target: x86_64-apple-darwin15.0.0
      Thread model: posix
      )
gives another results:
1. object file has 159044 bytes size
2. in disassembler a lot of unnecessary .objc_property_list, .objc_method_list, .objc_protocol_list and .objc_protocol in .bss and .data section
3. the final binary size is 101837 bytes size
    
If comment in above sample #import <UIKit/UIKit.h, the size of object file and binary file in xcode compilation will be the same,
but size binary and object file in case of clang+llvm-3.8.0-x86_64-apple-darwin compilation will be reduced sufficiently:
binary 28569 and object 36952 accordingly.
                
Looks like all protocols from UIKit.h is placed into binary/object even it's not used there.
I got 3.9.0 from public git (branch release_39 of clang git repository) and found that all protocols interpret as definition (isThisDeclarationADefinition),
and in CGObjCGNU.cpp it's placed into object file whenever it's used or not. With the same open source compiler and -fobjc-runtime=ios protocols are not putting into
object file even in compilation for ELF (I used -target arm-linux-gnueabi).
I used default binutil linker, not llvm's ld.ldd, due to I have problem with relocation on ARM.
I didn't use LTO, due to my linker doesn't have plugin support (I didnt' yet recompile it).

Could somebody point me where to solve that problem, is it AST generation stage, or linkage stage or maybe it's not a problem and just LTO could be solution here?

p.s. It's not only one issue found by me in gnustep runtime generation, and I'll report these issues in next emails.
    
         Best regards,
        Alexey Perevalov

I checked with lto, unfortunately, it didn’t get rid of multiple .objc_protocol inclusions.

Which version?

От: mehdi.amini@apple.com <mehdi.amini@apple.com> от имени Mehdi Amini <mehdi.amini@apple.com>
Отправлено: 27 сентября 2016 г. 8:46
Кому: Perevalov Alexey
Копия: cfe-users@lists.llvm.org; llvm-dev@lists.llvm.org
Тема: Re: [llvm-dev] objc object file generated for gnustep runtime for ELF target is too big

Dear community,

I'm using gnustep runtime -fobjc-runtime=gnustep with gnustep-libobjc2 (https://github.com/gnustep/libobjc2) and

Cocotron/Chameleon.

For following source file

#import <Foundation/Foundation.h>
#import <UIKit/UIKit.h>

int main(void)
{
NSString *str = [NSString stringWithCString:"TEST"];
NSLog(@"test object %@", str);
return 0;
}

Compilation in xcode's clang

Which version?

I think, it doesn't matter, due to with xcode's clang all fine

but if you intersting,

xcrun -sdk iphoneos clang --version
Apple LLVM version 7.0.0 (clang-700.1.76)
Target: x86_64-apple-darwin15.0.0
Thread model: posix

also all fine with open source llvm clang if I choose ios as -fobjc-runtime, I got troubles only with gnustep runtime in clang+llvm-3.8.0 and 3.9.x.


Mehdi

xcrun -sdk iphoneos clang -arch armv7s -mios-version-min=6.1 -c test_foundation.m -o test_foundation_apple.o
gives following result:
1. object file has only 3156 bytes size
2. and in disassembler only main function
3. the final executable will be 49992 bytes size in this case

But compilation in llvm's clang downloaded from http://llvm.org/releases/3.8.0/clang+llvm-3.8.0-x86_64-apple-darwin.tar.xz

(
clang version 3.8.0 (tags/RELEASE_380/final)
Target: x86_64-apple-darwin15.0.0
Thread model: posix
)
gives another results:
1. object file has 159044 bytes size
2. in disassembler a lot of unnecessary .objc_property_list, .objc_method_list, .objc_protocol_list and .objc_protocol in .bss and .data section
3. the final binary size is 101837 bytes size

If comment in above sample #import <UIKit/UIKit.h, the size of object file and binary file in xcode compilation will be the same,
but size binary and object file in case of clang+llvm-3.8.0-x86_64-apple-darwin compilation will be reduced sufficiently:
binary 28569 and object 36952 accordingly.

Looks like all protocols from UIKit.h is placed into binary/object even it's not used there.
I got 3.9.0 from public git (branch release_39 of clang git repository) and found that all protocols interpret as definition (isThisDeclarationADefinition),
and in CGObjCGNU.cpp it's placed into object file whenever it's used or not. With the same open source compiler and -fobjc-runtime=ios protocols are not putting into
object file even in compilation for ELF (I used -target arm-linux-gnueabi).
I used default binutil linker, not llvm's ld.ldd, due to I have problem with relocation on ARM.
I didn't use LTO, due to my linker doesn't have plugin support (I didnt' yet recompile it).

Could somebody point me where to solve that problem, is it AST generation stage, or linkage stage or maybe it's not a problem and just LTO could be solution here?

p.s. It's not only one issue found by me in gnustep runtime generation, and I'll report these issues in next emails.

     Best regards,
    Alexey Perevalov

_______________________________________________
LLVM Developers mailing list
llvm-dev@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

llvm-dev Info Page
lists.llvm.org
To see the collection of prior postings to the list, visit the llvm-dev Archives. Using llvm-dev: To post a message to all the list members, send ...

First, please don’t expect the sizes for different runtimes to be the same (or, necessarily, to be even very close). The Mac runtimes have a bit more support from the linker for various things. To ensure that property introspection is consistent without this support, the GNUstep runtime IR generation emits property definitions in all compilation units and then uniques them later.

We’re currently working on a new ABI for the GNUstep runtime that should address the size issues, but until then this looks pretty much what I’d expect.

David

От: mehdi.amini@apple.com <mehdi.amini@apple.com> от имени Mehdi Amini <mehdi.amini@apple.com>
Отправлено: 27 сентября 2016 г. 8:46
Кому: Perevalov Alexey
Копия: cfe-users@lists.llvm.org; llvm-dev@lists.llvm.org
Тема: Re: [llvm-dev] objc object file generated for gnustep runtime for ELF target is too big
    

Dear community,

I'm using gnustep runtime -fobjc-runtime=gnustep with gnustep-libobjc2 (https://github.com/gnustep/libobjc2) and

Cocotron/Chameleon.

For following source file

#import <Foundation/Foundation.h>
      #import <UIKit/UIKit.h>

int main(void)
        {
                NSString *str = [NSString stringWithCString:"TEST"];
                NSLog(@"test object %@", str);
                return 0;
        }

Compilation in xcode's clang

Which version?

I think, it doesn't matter, due to with xcode's clang all fine

It does: if there is a recent regression in the compiler, it may be in Xcode 8 as well. Can you update and try?

От: Dr D. Chisnall <dc552@hermes.cam.ac.uk> от имени David Chisnall <David.Chisnall@cl.cam.ac.uk>
Отправлено: 27 сентября 2016 г. 10:42
Кому: Perevalov Alexey
Копия: cfe-users@lists.llvm.org; llvm-dev@lists.llvm.org
Тема: Re: [llvm-dev] objc object file generated for gnustep runtime for ELF target is too big

Could somebody point me where to solve that problem, is it AST generation stage, or linkage stage or maybe it's not a problem and just LTO could be solution here?

First, please don’t expect the sizes for different runtimes to be the same (or, necessarily, to be even very close). The Mac runtimes have a bit more support from the linker for >various things. To ensure that property introspection is consistent without this support, the GNUstep runtime IR generation emits property definitions in all compilation units and >then uniques them later.

We’re currently working on a new ABI for the GNUstep runtime that should address the size issues, but until then this looks pretty much what I’d expect.

David

Thank you David,
I didn't expect the same size for executable file, but I got binary (real application) for gnustep runtime 5 times bigger than for ios runtime, and then I decided to figure out it.
(neither -flto nor -fdata-sections -ffunction-sections -Wl,--gc-sections helped).
5x bigber it's not due to only this issue :wink: I think.

For small sample (listed above in thread) I found that all protocol's declaration is interprets as definitions, and I see a lot of
.objc_protocol:
        .long 3
        .long .L.objc_protocol_name
        .long .objc_protocol_list
        .long .objc_method_list
        .long .objc_method_list.1
        .long .objc_method_list.2
        .long .objc_method_list.3
        .long .objc_property_list
        .long .objc_property_list.4
        .size .objc_protocol, 36

        .type .L__unnamed_37,%object @ @34
        .section .rodata.str1.1,"aMS",%progbits,1
.L__unnamed_37:
        .asciz "copyWithZone:"
        .size .L__unnamed_37, 14

        .type .L__unnamed_38,%object @ @35
...

in disassembler output and in final binary, but these protocol aren't used in sample. Neither front end nor linker got rid of it.
You wrote definitions will be unique, but it didn't strip out unnecessary .objc_protocol and depended fields.
(frankly saying after strip -s I didn't see .objc_protocol, but still find all selector and protocol names in strings output :wink:

And it's look like a bug.

This is a known issue and is unavoidable with the current ABI. I’m currently working on a replacement GNUstep Objective-C ABI that takes advantage of some features of linkers that weren’t available when the original GNU Objective-C ABI was created, but it’s spare-time (unfunded) work, so is taking a while. With the current prototype, there are a lot of space savings but it’s nowhere near ready for release.

David

P.S. This thread is being sent to two mailing lists, neither of which is the correct one. For future reference, cfe-dev is the correct list for development issues relating to the clang front end.