The LC040 support, probably, even if it's deemed nearly impossible to do. The problem is not the LC040 per se; the problem is that there are quite a bunch of bad LC040 processors out there. These work incorrectly in that they do not correctly keep their registers when an FPU interrupt goes off; as a result, the kernel returns to an incorrect CPU state after handling the FPU, so stuff obviously doesn't work. Not being a kernel hacker myself, I've been told that it may be impossible to work around this processor bug.
Perhaps impossible to make FPU traps work, but not at all to circumvent the problem altogether. NetBSD now has support for m68k softfloat-built systems. It circumvents the CPU bug by not using the FPU; all floating point code generated by the toolchain is software based instead of using floating point instructions. Nothing to trap, and therefore no worries about broken trapping mechanisms.
Someone was working on bringing our changes back into gcc and into Debian, but I've forgotten where and who.
In the meanwhile, NetBSD support on LC040 systems is good. We have both 1.6.2 and 2.0 snapshots, plus many precompiled softfloat binary packages.
John Klos