[Mlir-commits] [mlir] [mlir][vector]add foldConstantOp fold function and apply it to extractOp and insertOp. (PR #124399)

[lonely eagle]

================
@@ -4115,3 +4115,39 @@ func.func @step_scalable() -> vector<[4]xindex> {
   %0 = vector.step : vector<[4]xindex>
   return %0 : vector<[4]xindex>
 }
+
+// -----
+
+// CHECK-LABEL: @extract_arith_constnt
+func.func @extract_arith_constnt() -> i32 {
+  %v = arith.constant dense<0> : vector<32x1xi32>
+  %c_0 = arith.constant 0 : index
+  %elem = vector.extract %v[%c_0, %c_0] : i32 from vector<32x1xi32>
+  return %elem : i32
+}
+
+// CHECK: %[[VAL_0:.*]] = arith.constant dense<0> : vector<32x1xi32>
+// CHECK: %[[VAL_1:.*]] = builtin.unrealized_conversion_cast %[[VAL_0]] : vector<32x1xi32> to !llvm.array<32 x vector<1xi32>>
+// CHECK: %[[VAL_2:.*]] = llvm.extractvalue %[[VAL_1]][0] : !llvm.array<32 x vector<1xi32>>
+// CHECK: %[[VAL_3:.*]] = llvm.mlir.constant(0 : i64) : i64
+// CHECK: %{{.*}} = llvm.extractelement %[[VAL_2]]{{\[}}%[[VAL_3]] : i64] : vector<1xi32>
+
+// -----
+
+// CHECK-LABEL: @insert_arith_constnt()
+
+func.func @insert_arith_constnt() -> vector<32x1xi32> {
+  %v = arith.constant dense<0> : vector<32x1xi32>
+  %c_0 = arith.constant 0 : index
+  %c_1 = arith.constant 1 : i32
+  %v_1 = vector.insert %c_1, %v[%c_0, %c_0] : i32 into vector<32x1xi32>
+  return %v_1 : vector<32x1xi32>
+}
+    
+// CHECK: %[[VAL_0:.*]] = arith.constant dense<0> : vector<32x1xi32>
+// CHECK: %[[VAL_1:.*]] = builtin.unrealized_conversion_cast %[[VAL_0]] : vector<32x1xi32> to !llvm.array<32 x vector<1xi32>>
+// CHECK: %[[VAL_2:.*]] = arith.constant 1 : i32
+// CHECK: %[[VAL_3:.*]] = llvm.extractvalue %[[VAL_1]][0] : !llvm.array<32 x vector<1xi32>>
+// CHECK: %[[VAL_4:.*]] = llvm.mlir.constant(0 : i64) : i64
+// CHECK: %[[VAL_5:.*]] = llvm.insertelement %[[VAL_2]], %[[VAL_3]]{{\[}}%[[VAL_4]] : i64] : vector<1xi32>
+// CHECK: %{{.*}} = llvm.insertvalue %[[VAL_5]], %[[VAL_1]][0] : !llvm.array<32 x vector<1xi32>>
----------------
linuxlonelyeagle wrote:

https://github.com/llvm/llvm-project/blob/988978f964fb84cb99c83e6cd260dcc395afb6c2/mlir/test/Dialect/Vector/canonicalize.mlir#L469

As you can see from the test here, `canonicalize` will just eliminate `insertOp` and `extractOp`, so it seems pointless to add the test here. But I think it makes sense to include the test in `vector-to-llvm.mlir`, which itself is where I found the problem in running this pass, in addition to the other tests where this fold's functionality is demonstrated. 
It should determine if it can be folded after each run of the pattern.

https://github.com/llvm/llvm-project/pull/124399