Apparent bounds checking failure

[Programmerino]

It's possible to read past the end of arrays without using unsafe, sometimes leading to segfaults. I tested it on the C backend on both 761c9de (current master) and 0.24.3 (latest release). My compiler is gcc version 13.1.1 and I am on x86-64 Linux.

I have two examples, the first one is simple but the returned value always seems to be x and I can't achieve a segfault with it:

entry main (x: i64) =
    let is = flatten [[x], [x]]
    in
    is[999999999]

Output:

> futhark c bad.fut
> echo "0" | ./bad --entry-point main
0i64

Here is the second one:

def helper inner outer =
    let is = flatten (map(\_ -> inner |> map id) outer)
    in
    is[999999999]

entry main (_: i64) = helper [(1i32, 0i32), (0, 0)] [(0, 0)]

Output:

> futhark c bad.fut
> echo "0" | ./bad --entry-point main
Segmentation fault (core dumped)

[athas]

The first one is benign, because the compiler realises that the output (if any) must be zero. We still normally try to preserve failing checks, though.

The second one is a real problem and is likely due to a sloppy rewrite rule that throws away safety dependencies.