[Mlir-commits] [mlir] [MLIR] `vector.constant_mask` to support unaligned cases (PR #116520)

[Andrzej Warzyński]

================
@@ -268,18 +310,17 @@ func.func @vector_maskedload_i4_constant_mask_unaligned(%passthru: vector<5xi2>)
 // CHECK: %[[MASK:.+]] = arith.constant dense<[false, true, true, true, false]> : vector<5xi1>
 
 // CHECK: %[[COMPRESSED_MASK:.+]] = arith.constant dense<true> : vector<2xi1>
+
+// Emulated masked load from alloc:
 // CHECK: %[[EMPTY:.+]] = arith.constant dense<0> : vector<8xi2>
 // CHECK: %[[PTH_PADDED:.+]] = vector.insert_strided_slice %[[PTH]], %[[EMPTY]]
 // CHECK-SAME: {offsets = [1], strides = [1]} : vector<5xi2> into vector<8xi2>
-
-// Emulated masked load from alloc:
 // CHECK: %[[PTH_PADDED_UPCAST:.+]] = vector.bitcast %[[PTH_PADDED]] : vector<8xi2> to vector<2xi8>
 // CHECK: %[[C1:.+]] = arith.constant 1 : index
 // CHECK: %[[MASKLOAD:.+]] = vector.maskedload %[[ALLOC]][%[[C1]]], %[[COMPRESSED_MASK]], %[[PTH_PADDED_UPCAST]]
 // CHECK: %[[MASKLOAD_DOWNCAST:.+]] = vector.bitcast %[[MASKLOAD]] : vector<2xi8> to vector<8xi2>
 
-// Select from emulated loaded vector and passthru vector:
-// TODO: fold this part if possible.
+// Select from emulated loaded vector and passthru vector: (TODO: fold this part if possible)
----------------
banach-space wrote:

I'd reduce noise and keep TODO on a separate line.

https://github.com/llvm/llvm-project/pull/116520