[Mlir-commits] [mlir] [mlir][vector] Add support for scalable vectors to VectorLinearize (PR #86786)

[Andrzej Warzyński]

================
@@ -97,3 +97,47 @@ func.func @test_tensor_no_linearize(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf3
 
     return %0, %arg0 : tensor<2x2xf32>, tensor<2x2xf32>
 }
+
+// -----
+
+// ALL-LABEL:   func.func @test_1_scalable_dim(
+// ALL-SAME:    %[[ARG_0:.*]]: vector<2x[4]xf32>) -> vector<2x[4]xf32> {
+func.func @test_1_scalable_dim(%arg0: vector<2x[4]xf32>) -> vector<2x[4]xf32> {
+  // DEFAULT:  %[[SC:.*]] = vector.shape_cast %[[ARG_0]] : vector<2x[4]xf32> to vector<[8]xf32>
+  // DEFAULT:  %[[CST:.*]] = arith.constant dense<3.000000e+00> : vector<[8]xf32>
+  // BW-128:  %[[CST:.*]] = arith.constant dense<3.000000e+00> : vector<2x[4]xf32>
----------------
banach-space wrote:

Not sure I follow :) This is intentionally `vector<2x[4]xf32>` so that `-test-vector-linearize` works and `-test-vector-linearize=target-vector-bitwidth=128` does not.

https://github.com/llvm/llvm-project/pull/86786