[Mlir-commits] [mlir] [mlir][Vector] Improve support for vector.extract(broadcast) (PR #116234)

[Kunwar Grover]

================
@@ -710,29 +710,61 @@ func.func @fold_extract_transpose(
 
 // -----
 
-// CHECK-LABEL: fold_extract_broadcast
+// CHECK-LABEL: fold_extract_broadcast_same_input_output_scalar
 //  CHECK-SAME:   %[[A:.*]]: f32
 //       CHECK:   return %[[A]] : f32
-func.func @fold_extract_broadcast(%a : f32) -> f32 {
+func.func @fold_extract_broadcast_same_input_output_scalar(%a : f32, 
+  %idx0 : index, idx1 : index) -> f32 {
   %b = vector.broadcast %a : f32 to vector<1x2x4xf32>
-  %r = vector.extract %b[0, 1, 2] : f32 from vector<1x2x4xf32>
+  // The indices don't matter for this folder, so we use mixed indices.
+  %r = vector.extract %b[%idx0, %idx1, 2] : f32 from vector<1x2x4xf32>
   return %r : f32
 }
 
 // -----
 
-// CHECK-LABEL: fold_extract_broadcast_0dvec
+// CHECK-LABEL: fold_extract_broadcast_same_input_output_vec
+//  CHECK-SAME:   %[[A:.*]]: vector<4xf32>
+//       CHECK:   return %[[A]] : vector<4xf32>
+func.func @fold_extract_broadcast_same_input_output_vec(%a : vector<4xf32>, 
+  %idx0 : index) -> vector<4xf32> {
+  %b = vector.broadcast %a : vector<4xf32> to vector<1x2x4xf32>
+  // The indices don't matter for this folder, so we use mixed indices.
----------------
Groverkss wrote:

I think in my mind, I was trying not to not use dynamic indices for the innermost dimension because I didn't want to encourage using that (as it's harder to lower), but really, for the folder it doesn't matter. I'll switch them to dynamic indices.

https://github.com/llvm/llvm-project/pull/116234