[Mlir-commits] [mlir] [MLIR] extend `getCompressedMaskOp` support in `VectorEmulateNarrowType` (PR #116122)

[Andrzej Warzyński]

================
@@ -74,6 +74,55 @@ func.func @vector_cst_maskedload_i2(%passthru: vector<5xi2>) -> vector<3x5xi2> {
 
 // -----
 
+// This tests the correctness of generating compressed mask with `vector.create_mask` and a dynamic input.
+// Specifically, the program masked loads a vector<5xi2> from `vector<3x5xi2>[1, 0]`, with an unknown mask generator `m`.
+// After emulation transformation, it masked loads 2 bytes from linearized index `vector<4xi8>[1]`, with a new compressed mask
+// given by `ceildiv(m + 1, 4)`.
+func.func @check_unaligned_create_mask_dynamic_i2(%m : index, %passthru: vector<5xi2>) -> vector<5xi2> {
+    %0 = memref.alloc() : memref<3x5xi2>
+    %c0 = arith.constant 0 : index
+    %c1 = arith.constant 1 : index
+    %mask = vector.create_mask %m : vector<5xi1>
+    %1 = vector.maskedload %0[%c1, %c0], %mask, %passthru :
+      memref<3x5xi2>, vector<5xi1>, vector<5xi2> into vector<5xi2>
+    return %1 : vector<5xi2>
+}
+
+// CHECK-DAG: #[[MAP:.+]] = affine_map<()[s0] -> ((s0 + 1) ceildiv 4)>
+// CHECK: func @check_unaligned_create_mask_dynamic_i2(
+// CHECK-SAME:     %[[MASK:.+]]: index, %[[PASSTHRU:.+]]: vector<5xi2>)
----------------
banach-space wrote:

[nit] `MASK` -> `NUM_ELEMS_TO_READ`? Technically speaking, this value is used to create the mask (rather than being a mask).

https://github.com/llvm/llvm-project/pull/116122