I am trying to make a prime sieve in Rust.
I need several mutable references to the same element of the array, not mutable references to different parts of the array.
For the way this works, I know data races are not a relevant problem, so multiple mutable references are acceptable, but the Rust compiler does not accept my unsafe code.
I am using crossbeam 0.8.0.
fn extend_helper(
primes: &Vec<usize>,
true_block: &mut Vec<bool>,
segment_min: usize,
segment_len: usize,
) {
crossbeam::scope(|scope| {
for prime in primes {
let prime = prime.clone();
let segment_min = &segment_min;
let segment_len = &segment_len;
let shared = unsafe { &mut *true_block };
scope.spawn(move |_| {
let tmp = smallest_multiple_of_n_geq_m(prime, *segment_min) - *segment_min;
for j in (tmp..*segment_len).step_by(prime) {
shared[j] = false;
}
});
}
})
.unwrap();
}
fn smallest_multiple_of_n_geq_m(n: usize, m: usize) -> usize {
m + ((n - (m % n)) % n)
}
error[E0499]: cannot borrow `*true_block` as mutable more than once at a time
--> src/lib.rs:12:35
|
7 | crossbeam::scope(|scope| {
| ----- has type `&Scope<'1>`
...
12 | let shared = unsafe { &mut *true_block };
| ^^^^^^^^^^^^^^^^ `*true_block` was mutably borrowed here in the previous iteration of the loop
13 |
14 | / scope.spawn(move |_| {
15 | | let tmp = smallest_multiple_of_n_geq_m(prime, *segment_min) - *segment_min;
16 | | for j in (tmp..*segment_len).step_by(prime) {
17 | | shared[j] = false;
18 | | }
19 | | });
| |______________- argument requires that `*true_block` is borrowed for `'1`
warning: unnecessary `unsafe` block
--> src/lib.rs:12:26
|
12 | let shared = unsafe { &mut *true_block };
| ^^^^^^ unnecessary `unsafe` block
|
= note: `#[warn(unused_unsafe)]` on by default
How I should write the unsafe in a way that Rust accepts?
Rust does not allow that in its model. You want
AtomicBool
with relaxed ordering.This is a somewhat common edge case in most concurrency models - if you have two non-atomic writes of the same value to one location, is that well-defined? In Rust (and C++) it is not, and you need to explicitly use atomic.
On any platform you care about, the atomic bool store with relaxed ordering will have no impact.
For a reason why this matters, consider:
At setting
x
to true, the compiler is free to assume (due to no shared mutable references) thatx
is still false. It may implement this assignment as, e.g.,inc x
in assembly, moving its value from 0 to 1.If two threads came through this and both reached
*x = true
, its value may become something other than 0 or 1, which could violate other assumed invariants.