Consider:

func f(x uint64) uint64 {
  return uint64(uint32(x))
}

If, when inlined, x is known to be small, then f can be replaced by the identity function.

I was curious, so I sketched this quickly for a specific case by adding this case to func simplifyBlock in prove.go:

		case OpZeroExt32to64:
			if trunc := v.Args[0]; trunc.Op == OpTrunc64to32 {
				orig := trunc.Args[0]
				lim, ok := ft.limits[orig.ID]
				if !ok {
					continue
				}
				if lim.umax < (1<<32) || (lim.max < (1<<31) && ft.isNonNegative(orig)) {
					if b.Func.pass.debug > 0 {
						b.Func.Warnl(v.Pos, "Proved %v bounded", v.Op)
					}
					v.reset(OpCopy)
					v.AddArg(orig)
				}
			}

It triggers just a few times during make.bash.

This issue is to explore whether adding more cases makes this trigger enough times to be worth doing. There are lots more variants of signed/zero extension, and maybe other ways to generalize this optimization. (It also needs tests, both of correctness and that the optimization triggers.) Also, my limit tests in the code above could easily have fencepost problems. :)

cc @rasky @randall77