[Mlir-commits] [mlir] [mlir][scf] Relax requirements for loops fusion (PR #79187)

[Ivan Butygin]

Hardcode84 wrote:

I think first case with multiple writes to the same indices is fine, 
but I believe I have a counter-example for the second case with func arguments:
```
func.func @fuse_two(%A: memref, %B: memref) {
    %temp = memref.alloc
    scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
        %A_elem = memref.load %A[%i, %j]
        memref.store %A_elem, %temp[%i, %j] : memref<2x2xf32>
    }
    scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
        %elem = memref.load %temp[%i, %j]
        memref.store %elem, %B[%i, %j] : memref<2x2xf32>
    }
}
```
If %A and %B alias same array with an different offset, the fusion will not be legal.
And generally, I would prefer to be conservative about handling functions arguments.
What you probably want is some sort of `restrict` attribute on function args, so user can 
guarantee args do not alias. Upstream alias analysis doesn't support this, but this is the
reason I did aliasing hook configurable.

Regarding broken tests, they probably can be fixed by making some memrefs local allocations instead
of functions args.

https://github.com/llvm/llvm-project/pull/79187