I think the hardest part is function calls. Implementing the standard calling convention is really fiddly and annoying when you have to save and restore registers around function calls if you want to do it efficiently (keeping track of which registers are dirty, doing parallel loads of registers into the right argument registers, spilling when appropriate, etc. It is fiddly.
The alternative is going to an IR and doing full register allocation but then you need to implement Lengauer-Tarjan to get into SSA, all the same parallel load stuff for phi functions/block arguments, out-of-SSA, reconstruct in optimisation passes all the information you discard by going to a linear IR, etc.
The alternative is going to an IR and doing full register allocation but then you need to implement Lengauer-Tarjan to get into SSA, all the same parallel load stuff for phi functions/block arguments, out-of-SSA, reconstruct in optimisation passes all the information you discard by going to a linear IR, etc.