[Mlir-commits] [mlir] [mlir][vector] Preserve dynamic mask bounds in shape cast folding (PR #229699)

[Andrzej Warzyński]

================
@@ -3523,6 +3523,23 @@ func.func @fold_shape_cast_with_mask_trailing_unit(%arg0: tensor<1x?xf32>) -> ve
 
 // -----
 
+// CHECK-LABEL:   func.func @fold_shape_cast_with_mask_trailing_unit_dynamic(
+// CHECK-SAME:     %[[N:.*]]: index) -> vector<8xi1> {
+func.func @fold_shape_cast_with_mask_trailing_unit_dynamic(%n: index) -> vector<8xi1> {
----------------
banach-space wrote:

I know _why_ you added "dynamic" in the test name, but to me "@fold_shape_cast_with_mask_trailing_unit" is _more_ "dynamic" - it contains a "dynamic shape":
```mlir
  %dim = tensor.dim %arg0, %c1 : tensor<1x?xf32>
```

In general, both tests are dynamic and I would like for you to use some other (more obvious) differentiating factor between the cases. I have the following suggestion:
* `@fold_shape_cast_with_mask_trailing_unit` -> `@fold_shape_cast_with_mask_trailing_unit_provable_true`
* `@fold_shape_cast_with_mask_trailing_unit_dynamic` -> `@no_fold_shape_cast_with_mask_trailing_unit_dynamic_unknown`

https://github.com/llvm/llvm-project/pull/229699