How does triton reason about whether to issue vectorized loads/stores or not ?

[Sh0g0-1758]

Hey! I am working on porting some kernels from cuda to triton and I noticed that triton is not emitting vectorized loads/stores. I understand that my starting base and mask are not a multiple of 4 and hence triton can't reason about the outliers but I think for some specific blocks, a mask is not really needed. I am hoping to write some compiler pass that can give a hint to the compiler about this. Can someone please help me figure out how to do that ?