Question: Support for differentiating between volatile and non-volatile memory operations

[nayakajay]

Hi,

After lowering a CUDA program to mlir format using cgeist, I wanted to run some analysis on memory operations. I want to differentiate memory operations which are 'volatile.'

__global__ void test(volatile int* a, int* b) {
     a[some_idx] = data();
     b[another_idx] = data();
}

For example, I want to differentiate stores to array a and b as volatile has special semantics in CUDA. In the LLVM IR, it is possible (isVolatile method). Can something similar be achieved here?

[nayakajay]

Any comments on this?