cmd/compile: missed prove inferences with uint conversions

[josharian]

[moved from private email with @brtzsnr]

Consider this code, adapted from test/prove.go:

package p

func f(a []int, i int) int {
    if i >= 0 && i < len(a) {
        return a[i]
    }
    if i >= 10 && i < len(a) {
        return a[i]
    }
    return 0
}

func g(a []int, i int) int {
    if uint(i) < uint(len(a)) {
        return a[i]
    }
    if i >= 10 && i < len(a) {
        return a[i]
    }
    return 0
}

Functions f and g are equivalent. The only difference between them is the expression i >= 0 && i < len(a) in f vs uint(i) < uint(len(a)). And these are equivalent, because we know that len(a) is non-negative. But compiling with go tool compile -d=ssa/prove/debug=3 x.go, we get:

x.go:5: Proved non-negative bounds IsInBounds
x.go:8: Proved non-negative bounds IsInBounds
x.go:15: Proved IsInBounds
x.go:18: Disproved IsInBounds

Going from "x.go:5: Proved non-negative bounds IsInBounds" to "x.go:15: Proved IsInBounds" is a bit unfortunate, but ok. Going from "x.go:8: Proved non-negative bounds IsInBounds" to "x.go:18: Disproved IsInBounds" seems incorrect.

[brtzsnr]

The problem in this case is that the second return a[i] is deadcode. The second condition can be reached only if i >= len(a) so the index is out of bounds which is why prove prints "Disproved".

There is indeed a missed optimization and the second condition is not fully removed. If you find a correctness issue, let me know. Otherwise, I'll address this once the go1.8 opens and I have some spare time.

edit: The difference between the output is because it looks at different facts to establish the truth value. For f it uses i >= 10 && i < len(a) and for g it uses uint(i) < uint(len(a)).

tl;dr: It's not a bug. It's dead code.

[josharian]

Thanks, Alexandru. I have a CL that I hope to mail soon (that you know about) that will make this kind of show up a lot more. Good to know that I don't have to hold that CL for this. :)

[gopherbot]

CL https://golang.org/cl/27652 mentions this issue.