I don't have much experience with writing a custom gcc backend, but my experience with LLVM is that its model ends up being a somewhat poor fit for smaller or weirder architectures. For example, LLVM has this annoying tendency to really aggressively promote the size of everything--if you have a 32-bit core with 64-bit memory indexing, LLVM tends to make anything that eventually becomes a memory index a 64-bit computation, even if it would be more efficient to keep everything as 32-bit.