x/tools/go/analysis/passes/nilness: detect wrapping of nil errors

[ainar-g]

As of 4874f863, the nilness check doesn't seem to trigger on fmt.Errorf calls with an error that is proven nil. Example:

func f() error {
	exists, err := g()
	if err != nil {
		return errors.Wrap(err, "error1")
	}

	if !exists {
		return fmt.Errorf("nothing: %w", err)
	}

	return nil
}

Same with the %v verb. I feel like most people would not want to actually produce a nothing: <nil> or a nothing: %!w(<nil>) message here.

Inspired by dominikh/go-tools#529.

[andybons]

@mpvl @jba @neild

[jba]

Currently the nilness check only knows about the Go language, not the standard library. We could extend it to check static calls to known functions, maybe along the lines of https://golang.org/cl/184939. Then we could borrow format-parsing code from the printf check to perform the %w check.

I don't know if any of this is a good idea or not. @matloob, @ianthehat, @alandonovan, your thoughts?

[alandonovan]

The nilness check could be made to return a result value to other analyses through which they could query whether a function call argument is nil. The Printf checker (and others) could then use it to refine their heuristics.

However, that would make the printf check depend on nilness, and the standard vet contains only the watered-down nilfunc check, not full nilness, because the latter depends on golang.org/x/tools/go/ssa, which is more expensive to compute. We could evaluate whether the benefit of better checks exceeds the cost of computing full SSA.