Re: runtime 2.0 issues.

[David Chisnall]

On 05/03/2019 12:45, Andreas Fink wrote:

> Making all for tool cvtenc...
>   Linking tool cvtenc ...
> ./obj/cvtenc.obj/cvtenc.m.o:(.data..objc_init[.objc_init]+0x18): 
> undefined reference to `__start___objc_classes'
> ./obj/cvtenc.obj/cvtenc.m.o:(.data..objc_init[.objc_init]+0x20): 
> undefined reference to `__stop___objc_classes'

As I have said before, this is a bug in clang 7.x  It is fixed in the 
8.x release branch and the fix is back-ported in the FreeBSD llvm70 
package.  If your favourite operating system's LLVM 7 package does not 
work, please ask them to incorporate this fix:

https://svnweb.freebsd.org/ports/head/devel/llvm70/files/clang/patch-tools_clang_lib_CodeGen_CGObjCGNU.cpp?revision=489195&view=markup

This bug doesn't affect real code, but it will be a problem for any 
programs that use Objective-C but don't contain any classes, which 
unfortunately includes most of the GNUstep base configure scripts.

I have tested clang 8 on FreeBSD and am able to build all of the GNUstep 
applications that I've tried (all of the ones in the FreeBSD ports 
collection) with it without issues.  This version already contains the 
bug fix.  It does cause a regression in one of the libobjc2 tests, which 
I'm working on tracking down (it's triggered by a change to how LLVM 
does ARC optimisation, but that's not actually the cause), but that 
particular pattern is unlikely to occur in real code.

David