The Triton dialect load and store instructions (tt.load, tt.store) can take attributes. For example the load instruction may have the volatile and/or the cache attributes:

   tt.load %6 {isVolatile = true} : tensor<256x!tt.ptr<f32>, #blocked0>
   tt.load %6 cacheModifier = ca : tensor<256x!tt.ptr<f32>, #blocked0>

Similarly for the store instruction. During code generation to LLVM IR, these attributes are currently ignored so we need to handle them.