[Mlir-commits] [mlir] [mlir][ArmSME] Skip tile allocation for functions with no SME tile values (PR #227153)

[Benjamin Maxwell]

================

----------------
MacDue wrote:

Ah not in this case. The allocator is clever enough to sink the `get_tile` ops into the branches (as they have no side effects) which mean both can be assigned the same tile ID. To reproduce the error you need to add some ops that can't be sunk (e.g. a load_slice):

```mlir
func.func @overlapping_branches(%cond: i1, %src : memref<?x?xi8>, %mask : vector<[16]xi1>) {
  %tile = arm_sme.get_tile : vector<[16]x[16]xi8>
  %c0 = arith.constant 0 : index
  %c1 = arith.constant 1 : index
  %vecA = arm_sme.load_tile_slice %src[%c0], %mask, %tile, %c0 : memref<?x?xi8>, vector<[16]xi1>, vector<[16]x[16]xi8>
  %vecB = arm_sme.load_tile_slice %src[%c0], %mask, %tile, %c1 : memref<?x?xi8>, vector<[16]xi1>, vector<[16]x[16]xi8>
  %ret = scf.if %cond -> vector<[16]x[16]xi8> {
    scf.yield %vecA : vector<[16]x[16]xi8>
  } else {
    scf.yield %vecB : vector<[16]x[16]xi8>
  }
  "test.some_use"(%ret) : (vector<[16]x[16]xi8>) -> ()
  return
}
```

https://github.com/llvm/llvm-project/pull/227153