Cpp miscompilation when using mutable arguments

[truelossless]

Hello,

Thanks for the great work on CubeCL!
I've noticed that using mutable arguments on a #[cube] kernel causes a cpp miscompilation while using CUDA as the backend.

Example code:

This cube function...

#[cube]
pub fn foo(mut idx: u64) -> u64 {
    idx += 1;
    idx
}

...compiles to...

const uint64 l_2 = output_0[idxGlobal];
const uint64 l_2 = l_2 + uint64(1);
output_0[idxGlobal] = l_2;

...which causes

    default_program(23): error: "l_2" has already been declared in the current scope
      const uint64 l_2 = l_2 + uint64(1);
                   ^
    default_program(23): warning #549-D: variable "l_2" is used before its value is set
      const uint64 l_2 = l_2 + uint64(1);
                         ^
    Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"
    default_program(22): warning #177-D: variable "l_2" was declared but never referenced
      const uint64 l_2 = output_0[idxGlobal];
                   ^
    1 error detected in the compilation of "default_program".

Copying the argument in the function instead of mutating it works fine:

#[cube]
pub fn foo(idx: u64) -> u64 {
    let mut idx2 = idx;
    idx2 += 1;
    idx2
}

[maxtremblay]

Can you validate if you are running the latest version of cubecl? I fixed a bunch of these lately.

[truelossless]

Yes, it still happens in #a43015e2.
I haven't tested it toroughly, but I think variable shadowing can miscompile too. Nevermind, it was on my side.

[nathanielsimard]

I don't think we parse mutability correctly in CubeCL yet. It's because there is comptime mutability (which we handle correctly), referring to the manipulation of symbols at compile time, and "real" mutability, which involves actually changing a register's value at runtime. For now, you can use let mut variable = variable.runtime() to declare the latter.