cmd/compile/ssa: mark if { panic } blocks as unlikely

[josharian]

Given a block like

if condition {
  // basic block
  panic(...)
}

we should mark condition as unlikely. Unlikely to make a huge difference, but many things in the runtime have guards like this; we may as well help the branch predictor. Inspired by CL 11541.

[josharian]

This should probably happen at the SSA stage, perhaps as part of a general branch prediction pass.

We should also make sure that SSA preserves likelihood hints passed in by the frontend.

[dr2chase]

Generalization (from years ago working on Java) -- everything post-dominated by (a) panic(s) is marked as unlikely.

In years past, people have proposed training neural networks to infer branch frequencies from program characteristics -- I recall a presentation by Ben Zorn where he showed how well a by-the-book evaluation worked, doing 20 trials of train-on-19-test-on-remaining-1, with the bonus unethical slide at the end where he does 20 trials of train-on-20 -- which is what all the compiler writers would use, because benchmarks.

[tzneal]

Is there any good way to determine what function the Aux value of an OpStaticCall refers to? I'm working on a branch prediction pass that works the way David suggested (branches to blocks post-dominated by a panic are unlikely).

CC @randall77

[josharian]

(general advice) Poke around syntax.go. Request docs/clarification as desired. Off the top of my head, you probably want v.Aux.(*Node).Func.Nname.Sym.Name and ....Sym.Pkg.Path, but take those with a grain of salt.

[gopherbot]

CL https://golang.org/cl/19995 mentions this issue.