Revert "[mlir][Vector] Support 0-d vectors natively in TransferOpReduceRank (llvm#112907)"

This reverts commit 1004865.

Failing CI as discussed here:
* iree-org/iree#19135

Author: banach-space

mlir/lib/Dialect/Vector/Transforms/LowerVectorTransfer.cpp

@@ -358,10 +358,31 @@ struct TransferOpReduceRank
           op, "map is not a minor identity with broadcasting");
     }
 
+    // TODO: support zero-dimension vectors natively.  See:
+    // https://llvm.discourse.group/t/should-we-have-0-d-vectors/3097.
+    // In the meantime, lower these to a scalar load when they pop up.
+    if (reducedShapeRank == 0) {
+      Value newRead;
+      if (isa<TensorType>(op.getShapedType())) {
+        newRead = rewriter.create<tensor::ExtractOp>(
+            op.getLoc(), op.getSource(), op.getIndices());
+      } else {
+        newRead = rewriter.create<memref::LoadOp>(
+            op.getLoc(), originalVecType.getElementType(), op.getSource(),
+            op.getIndices());
+      }
+      return rewriter
+          .create<vector::BroadcastOp>(op.getLoc(), originalVecType, newRead)
+          .getVector();
+    }
+
     SmallVector<int64_t> newShape(
         originalVecType.getShape().take_back(reducedShapeRank));
     SmallVector<bool> newScalableDims(
         originalVecType.getScalableDims().take_back(reducedShapeRank));
+    // Vector rank cannot be zero. Handled by TransferReadToVectorLoadLowering.
+    if (newShape.empty())
+      return rewriter.notifyMatchFailure(op, "rank-reduced vector is 0-d");
 
     VectorType newReadType = VectorType::get(
         newShape, originalVecType.getElementType(), newScalableDims);

mlir/test/Conversion/VectorToSCF/vector-to-scf.mlir

@@ -503,8 +503,8 @@ func.func @transfer_read_within_async_execute(%A : memref<2x2xf32>) -> !async.to
 
 // CHECK-LABEL: transfer_read_with_tensor
 func.func @transfer_read_with_tensor(%arg: tensor<f32>) -> vector<1xf32> {
-    // CHECK:      %[[EXTRACTED:.*]] = vector.transfer_read %{{.*}}[], %{{.*}} : tensor<f32>, vector<f32>
-    // CHECK-NEXT: %[[RESULT:.*]] = vector.broadcast %[[EXTRACTED]] : vector<f32> to vector<1xf32>
+    // CHECK:      %[[EXTRACTED:.*]] = tensor.extract %{{.*}}[] : tensor<f32>
+    // CHECK-NEXT: %[[RESULT:.*]] = vector.broadcast %[[EXTRACTED]] : f32 to vector<1xf32>
     // CHECK-NEXT: return %[[RESULT]] : vector<1xf32>
     %f0 = arith.constant 0.0 : f32
     %0 = vector.transfer_read %arg[], %f0 {permutation_map = affine_map<()->(0)>} :

mlir/test/Dialect/Linalg/vectorize-tensor-extract.mlir

@@ -138,7 +138,9 @@ func.func @vectorize_nd_tensor_extract_transfer_read_basic(
 // CHECK:   %[[READ:.*]] = vector.transfer_read %[[ARG0]][%[[IDX1]], %[[IDX2]], %[[IDX3]]], %[[CST]] {in_bounds = [true, true, true]} : tensor<3x3x3xf32>, vector<1x1x3xf32>
 // CHECK:   vector.transfer_write %[[READ]], %[[ARG1]][%[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<1x1x3xf32>, tensor<1x1x3xf32>
 
-// Same as example above, but reading into a column tensor.
+// Same as example above, but reading into a column tensor. Note that after the
+// vectorizatoin, the `TransferOpReduceRank` will replace
+// `vector.transfer_read` with `tensor.extract -> scalar`.
 
 // TODO: Currently this fails to vectorise when the indices are non-constant.
 
@@ -162,10 +164,9 @@ func.func @vectorize_nd_tensor_extract_transfer_read_basic_column(
 // CHECK-LABEL:   func.func @vectorize_nd_tensor_extract_transfer_read_basic_column(
 // CHECK-SAME:      %[[INPUT:.*]]: tensor<3x3x3xf32>,
 // CHECK-SAME:      %[[OUTPUT:.*]]: tensor<3x1x1xf32>)
-// CHECK-DAG:        %[[C0:.*]] = arith.constant 0 : index
-// CHECK-DAG:        %[[CST_0:.*]] = arith.constant 0.000000e+00 : f32
-// CHECK:           %[[READ:.*]] = vector.transfer_read %[[INPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]], %[[CST_0]] : tensor<3x3x3xf32>, vector<f32>
-// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[READ]] : vector<f32> to vector<3x1x1xf32>
+// CHECK:           %[[C0:.*]] = arith.constant 0 : index
+// CHECK:           %[[EXTRACT:.*]] = tensor.extract %[[INPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]] : tensor<3x3x3xf32>
+// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[EXTRACT]] : f32 to vector<3x1x1xf32>
 // CHECK:           %[[RES:.*]] = vector.transfer_write %[[BCAST]], %[[OUTPUT]]{{\[}}%[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<3x1x1xf32>, tensor<3x1x1xf32>
 // CHECK:           return %[[RES]] : tensor<3x1x1xf32>
 
@@ -747,8 +748,8 @@ func.func @vectorize_0d_tensor_extract(%arg0: tensor<f32>, %arg2: tensor<1x1x3xf
 
 // CHECK-LABEL:   func.func @vectorize_0d_tensor_extract(
 // CHECK-SAME:     %[[ARG_0:.*]]: tensor<f32>
-// CHECK:           %[[EXTRACT:.*]] = vector.transfer_read %[[ARG_0]][], %{{.+}} : tensor<f32>
-// CHECK:           vector.broadcast %[[EXTRACT]] : vector<f32> to vector<1x1x3xf32>
+// CHECK:           %[[EXTRACT:.*]] = tensor.extract %[[ARG_0]][] : tensor<f32>
+// CHECK:           vector.broadcast %[[EXTRACT]] : f32 to vector<1x1x3xf32>
 
 module attributes {transform.with_named_sequence} {
   transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {

mlir/test/Dialect/Vector/vector-transfer-to-vector-load-store.mlir

@@ -26,8 +26,8 @@ func.func @vector_transfer_ops_0d_memref(%mem: memref<f32>, %vec: vector<1x1x1xf
 func.func @vector_transfer_ops_0d_tensor(%src: tensor<f32>) -> vector<1xf32> {
     %f0 = arith.constant 0.0 : f32
 
-//  CHECK:   %[[S:.*]] = vector.transfer_read %[[SRC]][]
-//  CHECK:   %[[V:.*]] = vector.broadcast %[[S]] : vector<f32> to vector<1xf32>
+//  CHECK-NEXT:   %[[S:.*]] = tensor.extract %[[SRC]][] : tensor<f32>
+//  CHECK-NEXT:   %[[V:.*]] = vector.broadcast %[[S]] : f32 to vector<1xf32>
     %res = vector.transfer_read %src[], %f0 {in_bounds = [true], permutation_map = affine_map<()->(0)>} :
       tensor<f32>, vector<1xf32>