[Mlir-commits] [mlir] [mlir][vector] Improve shape_cast lowering (PR #140800)

[James Newling]

================
@@ -140,6 +140,59 @@ func.func @shape_cast_1d0d(%arg0 : vector<1xf32>) -> vector<f32> {
   return %s : vector<f32>
 }
 
+
+// The shapes have 2 inner dimension sizes in common, so the extract result is rank-2.
+// CHECK-LABEL:  func.func @squeeze_out_prefix_unit_dim(
+// CHECK-SAME:               %[[ARG0:.*]]: vector<1x2x3xf32>) -> vector<2x3xf32> {
+// CHECK:          %[[EXTRACTED:.*]] = vector.extract %[[ARG0]][0] :
+// CHECK-SAME:               vector<2x3xf32> from vector<1x2x3xf32>
----------------
newling wrote:

I did a significant test file refactor to make the pre-existing and new tests consistent. The pre-existing test logic is unchanged. I'm happy to postpone refactoring the old tests to make that clearer

https://github.com/llvm/llvm-project/pull/140800