In practice, branch target predictors are also used for conditional branches too, and even unconditional, direct jumps. Yes, the target can (sort of by definition) be computed from the instruction and the pc, but the by using the branch target buffer, you avoid pipeline bubbles in the correctly predicted case by only having a data dependency on the pc, and predicting the target while the instruction is still being fetched. That way on the next cycle you're already fetching from the predicted target while still decoding the branch instruction itself in the next stage(s).

I wonder are there any compilers that recognizes switch-in-loop pattern of interprets and optimizes it to be branch-predictor friendly using multiple indirect jumps?

You really explain things in a way that's very easy to understand. I think I finally get it now. I understood the conditional branch part, but I think I was having trouble with the indirect branch predictor. So basically, when a branch instruction is executed, the CPU learns the actual target address. In the case of a switch statement, because the key is a single value like s, the history records get mixed up, whereas with a computed goto, since there are multiple keys, they don't get mixed up. Thank you for the clear explanation. It's a bit embarrassing that I didn't know this, but thanks to people like you, I feel like I'm learning more.