Rust - unsafely share between threads a mutable data without mutex - multithreading

Question

How can old school multi-threading (no wrapping mutex) can be achieve in Rust? And why is it undefined behavior?
I have to build a highly concurrent physic simulation. I am supposed to do it in C, but I chose Rust (I really needed higher level features).
By using Rust, I should opt for a safe communication between threads, however, I must use a mutable buffer shared between the threads. (actually, I have to implement different techniques and benchmark them)
First approach
Use Arc<Data> to share non-mutable state.
Use transmute to promote & to &mut when needed.
It was straightforward but the compiler would prevent this from compiling even with unsafe block. It is because the compiler can apply optimizations knowing this data is supposedly non-mutable (maybe cached and never updated, not an expert about that).
This kind of optimizations can be stopped by Cell wrapper and others.
Second approach
Use Arc<UnsafeCell<Data>>.
Then data.get() to access data.
This does not compile either. The reason is that UnsafeCell is not Send. The solution is to use SyncUnsafeCell but it is unstable for the moment (1.66), and the program will be compile and put to production on a machine with only the stable version.
Third approach
Use Arc<Mutex<Data>>.
At the beginning of each threads:
Lock the mutex.
Keep a *mut by coercing a &mut.
Release the mutex.
Use the *mut when needed
I haven't tried this one yet, but even if it compiles, is it safe (not talking about data race) as it would be with SyncUnsafeCell ?
PS: The values concurrently mutated are just f32, there are absolutely no memory allocation or any complex operations happening concurrently. Worst case scenario, I have scrambled some f32.

Answer 1

Disclaimer: There are probably many ways to solve this, this is just one of them, based on the idea of #Caesar.
Two main points of this post:
You can use AtomicU32 to share f32 between threads without any performance penalty (given an architecture where u32 is already atomic)
You can use std::thread::scope to avoid the overhead of Arc.
use std::{
fmt::Debug,
ops::Range,
sync::atomic::{AtomicU32, Ordering},
};
struct AtomicF32(AtomicU32);
impl AtomicF32 {
pub fn new(val: f32) -> Self {
Self(AtomicU32::new(val.to_bits()))
}
pub fn load(&self, order: Ordering) -> f32 {
f32::from_bits(self.0.load(order))
}
pub fn store(&self, val: f32, order: Ordering) {
self.0.store(val.to_bits(), order)
}
}
impl Debug for AtomicF32 {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
self.load(Ordering::Relaxed).fmt(f)
}
}
fn perform_action(data: &Vec<AtomicF32>, range: Range<usize>) {
for value_raw in &data[range] {
let mut value = value_raw.load(Ordering::Relaxed);
value *= 2.5;
value_raw.store(value, Ordering::Relaxed);
}
}
fn main() {
let data = (1..=10)
.map(|v| AtomicF32::new(v as f32))
.collect::<Vec<_>>();
println!("Before: {:?}", data);
std::thread::scope(|s| {
s.spawn(|| perform_action(&data, 0..5));
s.spawn(|| perform_action(&data, 5..10));
});
println!("After: {:?}", data);
}
Before: [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0]
After: [2.5, 5.0, 7.5, 10.0, 12.5, 15.0, 17.5, 20.0, 22.5, 25.0]
To demonstrate how leightweight this is, here is what this compiles to:
use std::{
sync::atomic::{AtomicU32, Ordering},
};
pub struct AtomicF32(AtomicU32);
impl AtomicF32 {
fn load(&self, order: Ordering) -> f32 {
f32::from_bits(self.0.load(order))
}
fn store(&self, val: f32, order: Ordering) {
self.0.store(val.to_bits(), order)
}
}
pub fn perform_action(value_raw: &AtomicF32) {
let mut value = value_raw.load(Ordering::Relaxed);
value *= 2.5;
value_raw.store(value, Ordering::Relaxed);
}
.LCPI0_0:
.long 0x40200000
example::perform_action:
movss xmm0, dword ptr [rdi]
mulss xmm0, dword ptr [rip + .LCPI0_0]
movss dword ptr [rdi], xmm0
ret
Note that while this contains zero undefined behaviour, it still is the programmer's responsibility to avoid read-modify-write race conditions.