I guess that as soon as the Go compiler introduces more expensive optimization passes to the code the more it makes sense to ensure that these passes are run on

I guess that as soon as the Go compiler introduces more expensive optimization passes to the code the more it makes sense to ensure that these passes are run only once (i.e. when the source is compiled) and not many times (i.e. when the whole binary is "linked", which in the Go toolchain up to 1.2.x includes a translation from pseudo-instructions to actual instructions). TBH I don't know how much of the optimizations are currently doable before emitting pseudo-instructions, but I guess that moving the actual code generation in liblink makes sense nevertheless.