On 2013-03-31, at 15:26, Greg Troxel <[email protected]> wrote: > > I suppose the --sysroot argument should have forced the use of the > correct arguments. I ktrace'd the link step and discovered that the > i386 build reads /lib/libcrypt.so too but the immediately reads > version in /usr/src/obj.i386/destdir.i386/lib. Presumably because the > linker realize it's not the correct version. The amd64 build, on the > other hand, appears happy with the /lib version and skips the > /usr/src/obj.amd64/destdir.amd64/lib version. Removing > /lib/libcrypt.so "fixes" the problem. > > This seems like a bug in the build system. > > I am not 100% clear on how --sysroot is supposed to work. It seems > obvious to me that --sysroot should suppress looking in the host > fileystem. But it could also be that -L/lib should be > -L/usr/src/obj.amd64/destdir.amd64/lib instead - try that and see via > ktrace if it skips the host lookup.
With -L/usr/src/obj.amd64/destdir.amd64/lib, I still get the error and ktrace shows that /lib/libcrypt.so is being used. I have to change the -Wl,-rpath,/lib argument for the compiler to use the build version. With the rpath change, /lib/libcrypt.so is not even accessed (according to ktrace). Is no-one else seeing this? My system is pretty plain NetBSD-current and I'm attempting to compile an unmodified snapshot of the netbsd-6 branch. My /etc/mk.conf contains OBJMACHINE=yes MKX11=yes and I'm using pkgsrc bash as interactive shell (building as root). Regards, Sverre
