[Mlir-commits] [mlir] [mlir][ArmSME] Use liveness information in the tile allocator (PR #90448)

[Andrzej Warzyński]

================
@@ -23,12 +29,17 @@ func.func @constant_with_multiple_users(%a: vector<[4]xf32>, %b: vector<[4]xf32>
 
 // -----
 
-// (No tile IDs -- the current tile allocator ignores this case)
+//  CHECK-LIVE-RANGE-LABEL: @value_with_multiple_users
+//        CHECK-LIVE-RANGE: ========== Coalesced Live Ranges:
+//        CHECK-LIVE-RANGE: ^bb0:
+//   CHECK-LIVE-RANGE-NEXT: |S arm_sme.move_vector_to_tile_slice
+//   CHECK-LIVE-RANGE-NEXT: || arm_sme.move_vector_to_tile_slice
+//   CHECK-LIVE-RANGE-NEXT: |E test.some_use
+//   CHECK-LIVE-RANGE-NEXT: E  test.some_use
 
-// CHECK-BAD-LABEL: @value_with_multiple_users
-// CHECK-BAD-NOT: tile_id
+// expected-note at below {{tile operand is: <block argument> of type 'vector<[4]x[4]xf32>'}}
 func.func @value_with_multiple_users(%tile: vector<[4]x[4]xf32>, %a: vector<[4]xf32>, %b: vector<[4]xf32>, %index: index) {
-  // A future allocator should error here (as `%tile` would need to be copied).
+  // expected-error at below {{op tile operand allocated to different SME virtial tile (move required)}}
----------------
banach-space wrote:

Why is move required for this case? I see only 3 tiles (and there are 4 SME virtual tiles for f32)

https://github.com/llvm/llvm-project/pull/90448