Re: building libobjc2 fPIC issue and ccmake . effectiveness

[Riccardo Mottola]

Hi,


On 5/31/20 1:05 PM, David Chisnall wrote:
> On 30 May 2020, at 22:51, Riccardo Mottola <riccardo.mottola@libero.it> wrote:
> > I have libstdc++.so and libstdc++.a available, and libsupc++.a
> >
> > so I use ccmake. and set
> >
> > CXX_RUNTIME_LIB to /usr/lib/libstdc++.so
> >
> > then I run (c)onfigure and (g)enerate - at that point libobjc2 fully built and 
> > installed.
> Great!  I am a bit confused as to why it’s finding the .a, because that should 
> be looking for only libraries that end with ${CMAKE_SHARED_LIBRARY_SUFFIX}, 
> which should be .so on ELF platforms.  We might need an extra check here.


Yes, it is strange indeed. Is there a configure log I can see or check? 
Like a config.log or so from autotools.


> This looks as if it is not finding the stringWithFormat:arguments: method.  
> This is in -base additions.  Are you linking with BFD ld?  This looks like 
> stack trace that I’d expect if Additions is not being loaded, which was the bug 
> we saw with BFD LD.  Can you try using a different linker for -base?

I always used the netbsd base linker. It does work with gcc and gcc 
runtime and gnustep-base.

GNU ld (NetBSD Binutils nb1) 2.31.1

I reconfigured base with:

./configure LD=/usr/pkg/gnu/bin/ld.gold

Is that enough?

./configure CFLAGS=-fuse-ld=gold

does not work with clang.


However, I still get the same crash. I wonder however if the old linker 
is really being used or not!

./configure LDFLAGS=-fuse-ld=gold

also does not work....

I have clang 9.0.1


> > Then I thought... let's try "make test" in libobjc2
> >
> > 98% tests passed, 4 tests failed out of 186
> >
> > Total Test time (real) =  68.32 sec
> >
> > The following tests FAILED:
> >           25 - BlockImpTest (SEGFAULT)
> >           26 - BlockImpTest_optimised (SEGFAULT)
> >           27 - BlockImpTest_legacy (SEGFAULT)
> >           28 - BlockImpTest_legacy_optimised (SEGFAULT)
> >
> > are these expected? I suppose... not, but I wonder if they can cause the issue 
> > I am seeing.
> These are four variations of the same test (-O0 / -O2, ABIv1 / ABIv2).  I would 
> expect the objc_msgSend tests to fail as well if it’s a problem with assembly 
> paths.  Is this OpenBSD?  As I recall, OpenBSD has some restrictions on JIT’d 
> code (which is what these paths are doing, effectively).  Can you look at the 
> mmap / mprotect calls in block_to_imp in these tests (these tests are all in 
> the Tests/ subdirectory of your build directory and the versions without the 
> _optimised suffix should work well in a debugger) and see if either of them are 
> failing?
>
> David

this is the "right track". I am on NetBSD, which has paxctl for memory 
access.

I did:

paxctl +m Test/BlockImpTest*


And rerun:

100% tests passed, 0 tests failed out of 186


so this looks great now! I hope that once finding the way to fix 
linking, everything will work on NetBSD too with clang. First time I try 
it (I used it 2 years ago, but in gnu compatibility mode)


Riccardo