cmd/compile/internal/ssa/block.go:Block has a branch prediction field (called Likely).
You can use that in cmd/compile/internal/ppc64/ssa.go to populate a branch Prog with the right data.
For amd64 we use Prog.From=={Const,0} for unlikely taken and {Const,1} for likely taken. Use that encoding for PPC as well if it makes sense.
See cmd/compile/internal/amd64/ssa.go for an example of how to do that.

I will not be submitting a change for this in Go 1.8. The performance testing I've done for this had mixed results and will require more investigation to determine if it is worth it.

I've done some investigation on this, setting the likely bit based on the end of block information. However that does not always provide an improvement, it varies quite a bit and in many cases gets worse. I have not investigated further to try and determine if the information provided is not dependable for ppc64x or if there is some other reason.

If I were to set this at all, I'd prefer to only set it in the obvious cases, such as when there is an if check for a panic condition and then the panic path marked as unlikely. Not sure what the best way is to determine that.

Thanks for the research. I've done the analysis in the past (other languages, other compilers) to do the "this will certainly fail, so it's certainly not a critical path" likeliness check, and I think we could do that here, too.

I have an item on my to-do list to write up a list of little useful projects, and that's one of them.
(Tentative list: improve spill location in register allocation; improve escape analysis w/ interprocedural use info for interface-typed call args; spec out loop unrolling within SSA form for both optimization and the reschedule check; this finer-grained "likely" information. I think Keith has a similar list).

I'm going to close this. It looks like the cases I was concerned with are now marked with the correct likely flag.