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

[Benjamin Maxwell]

================
@@ -100,13 +142,29 @@ func.func @non_overlapping_branches(%cond: i1) {
 
 // -----
 
-// Incorrect result! Everything assigned to tile 0 (which means values that are still live are overwritten).
-//
-// CHECK-BAD-LABEL: @constant_loop_init_with_multiple_users
-// CHECK-BAD: arm_sme.zero {tile_id = 0 : i32} : vector<[4]x[4]xf32>
-// CHECK-BAD: arm_sme.move_vector_to_tile_slice {{.*}} {tile_id = 0 : i32} : vector<[4]xf32> into vector<[4]x[4]xf32>
-// CHECK-BAD: arm_sme.move_vector_to_tile_slice {{.*}} {tile_id = 0 : i32} : vector<[4]xf32> into vector<[4]x[4]xf32>
+// Here %vecA and %vecB are not merged into the same live range (as they are unknown values).
+// This means that %vecA and %vecB are both allocated to different tiles (which is not legal).
----------------
MacDue wrote:

It's simply invalid code that would not produce correct results. Think about it, this is an `scf.if` that results in a value, no matter which branch you take, the result (used after the `scf.if`) has to be in the same register (or how would code after the `scf.if` know where the value is?). 

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