Revert "[mlir][linalg] Relax tensor.extract vectorization" (#102232)

Reverts #99299 because it breaks the lowering. To
repro: `mlir-opt -transform-interpreter ~/repro.mlir`

```mlir
#map = affine_map<(d0, d1) -> (d0)>
#map1 = affine_map<(d0, d1) -> (d1)>
#map2 = affine_map<(d0, d1) -> (d0, d1)>
#map3 = affine_map<(d0, d1) -> (d0 + d1)>
module {
  func.func @foo(%arg0: index, %arg1: tensor<2xf32>, %arg2: tensor<4xf32>, %arg3: tensor<1xf32>) -> tensor<4x1xf32> {
    %c0 = arith.constant 0 : index
    %cst = arith.constant 1.000000e+00 : f32
    %cst_0 = arith.constant 0.000000e+00 : f32
    %0 = tensor.empty() : tensor<4x1xf32>
    %1 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel"]} ins(%arg2, %arg3 : tensor<4xf32>, tensor<1xf32>) outs(%0 : tensor<4x1xf32>) {
    ^bb0(%in: f32, %in_1: f32, %out: f32):
      %2 = linalg.index 0 : index
      %3 = linalg.index 1 : index
      %4 = affine.apply #map3(%3, %arg0)
      %extracted = tensor.extract %arg1[%c0] : tensor<2xf32>
      %5 = arith.cmpi eq, %2, %c0 : index
      %6 = arith.cmpi ult, %2, %c0 : index
      %7 = arith.select %5, %cst, %in : f32
      %8 = arith.select %6, %cst_0, %7 : f32
      %9 = arith.cmpi eq, %4, %c0 : index
      %10 = arith.cmpi ult, %4, %c0 : index
      %11 = arith.select %9, %cst, %in_1 : f32
      %12 = arith.select %10, %cst_0, %11 : f32
      %13 = arith.mulf %8, %12 : f32
      %14 = arith.mulf %13, %extracted : f32
      %15 = arith.cmpi eq, %2, %4 : index
      %16 = arith.select %15, %cst, %cst_0 : f32
      %17 = arith.subf %16, %14 : f32
      linalg.yield %17 : f32
    } -> tensor<4x1xf32>
    return %1 : tensor<4x1xf32>
  }
}

module attributes {transform.with_named_sequence} {
  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
    %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
    transform.structured.vectorize %0 : !transform.any_op
    transform.yield
  }
}
```

Author: hanhanW

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

@@ -946,22 +946,27 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
   if (linalgOp.hasDynamicShape())
     return VectorMemoryAccessKind::Gather;
 
-  // True for vectors that are effectively 1D, e.g. `vector<1x4x1xi32>`, false
-  // otherwise.
-  bool isOutput1DVector = (llvm::count_if(targetShape, [](int64_t dimSize) {
-                             return dimSize > 1;
-                           }) == 1);
-
-  // 1. Assume that it's a gather load when reading non-1D vector.
-  if (!isOutput1DVector)
+  // 1. Assume that it's a gather load when reading _into_:
+  //    * an n-D "vector", like `tensor<1x2x4xi32` or `tensor<2x1x4xi32>`, or
+  //    * a 1-D "vector" with the trailing dim equal 1, e.g. `tensor<1x4x1xi32`.
+  // TODO: Relax these conditions.
+  // FIXME: This condition assumes non-dynamic sizes.
+  if ((llvm::count_if(targetShape,
+                      [](int64_t dimSize) { return dimSize > 1; }) != 1) ||
+      targetShape.back() == 1)
+    return VectorMemoryAccessKind::Gather;
+
+  // 2. Assume that it's a gather load when reading _from_ a tensor for which
+  // the trailing dimension is 1, e.g. `tensor<1x4x1xi32>`.
+  // TODO: Relax this condition.
+  if (inputShape.getShape().back() == 1)
     return VectorMemoryAccessKind::Gather;
 
   bool leadingIdxsLoopInvariant = true;
 
-  // 2. Analyze the leading indices of `extractOp`.
+  // 3. Analyze the leading indices of `extractOp`.
   // Look at the way each index is calculated and decide whether it is suitable
-  // for a contiguous load, i.e. whether it's loop invariant. If not, it's a
-  // gather load.
+  // for a contiguous load, i.e. whether it's loop invariant.
   auto indices = extractOp.getIndices();
   auto leadIndices = indices.drop_back(1);
 
@@ -977,13 +982,13 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
     return VectorMemoryAccessKind::Gather;
   }
 
-  // 3. Analyze the trailing index for `extractOp`.
+  // 4. Analyze the trailing index for `extractOp`.
   // At this point we know that the leading indices are loop invariant. This
   // means that is potentially a scalar or a contiguous load. We can decide
   // based on the trailing idx.
   auto extractOpTrailingIdx = indices.back();
 
-  // 3a. Scalar broadcast load
+  // 4a. Scalar broadcast load
   // If the trailing index is loop invariant then this is a scalar load.
   if (leadingIdxsLoopInvariant &&
       isLoopInvariantIdx(linalgOp, extractOpTrailingIdx)) {
@@ -992,7 +997,7 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
     return VectorMemoryAccessKind::ScalarBroadcast;
   }
 
-  // 3b. Contiguous loads
+  // 4b. Contiguous loads
   // The trailing `extractOp` index should increment with every loop iteration.
   // This effectively means that it must be based on the trailing loop index.
   // This is what the following bool captures.
@@ -1006,7 +1011,7 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
     return VectorMemoryAccessKind::Contiguous;
   }
 
-  // 4. Fallback case - gather load.
+  // 5. Fallback case - gather load.
   LDBG("Found gather load: " << extractOp);
   return VectorMemoryAccessKind::Gather;
 }

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

@@ -595,59 +595,3 @@ module attributes {transform.with_named_sequence} {
      transform.yield
    }
 }
-
-
-// -----
-
-func.func @vectorize_scalar_broadcast_column_tensor(%in: tensor<1x1x4xi32>) -> tensor<1x1x4xi32> {
-  %c4 = arith.constant 4 : index
-  %c0 = arith.constant 0 : index
-  %cst = arith.constant dense<[[0], [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]]> : tensor<15x1xi32>
-
-  %out = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} outs(%in : tensor<1x1x4xi32>) {
-  ^bb0(%out: i32):
-    %8 = linalg.index 0 : index
-    %idx_0 = linalg.index 0 : index
-    %extracted = tensor.extract %cst[%idx_0, %c0] : tensor<15x1xi32>
-    linalg.yield %extracted : i32
-  } -> tensor<1x1x4xi32>
-
-  return %out:tensor<1x1x4xi32>
-}
-
-// CHECK: #[[$ATTR_1:.+]] = affine_map<(d0, d1) -> (0, 0, 0)>
-// CHECK-LABEL:   func.func @vectorize_scalar_broadcast_column_tensor(
-// CHECK-SAME:      %[[VAL_0:.*]]: tensor<1x1x4xi32>) -> tensor<1x1x4xi32> {
-// CHECK:           %[[VAL_1:.*]] = arith.constant 4 : index
-// CHECK:           %[[VAL_2:.*]] = arith.constant 0 : index
-// CHECK:           %[[VAL_3:.*]] = arith.constant dense<{{\[\[}}0], [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]]> : tensor<15x1xi32>
-// CHECK:           %[[VAL_4:.*]] = arith.constant 1 : index
-// CHECK:           %[[VAL_5:.*]] = arith.constant 1 : index
-// CHECK:           %[[VAL_6:.*]] = arith.constant 4 : index
-// CHECK:           %[[VAL_7:.*]] = arith.constant 0 : index
-// CHECK:           %[[VAL_8:.*]] = arith.constant 0 : i32
-// CHECK:           %[[VAL_9:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_7]], %[[VAL_7]], %[[VAL_7]]], %[[VAL_8]] : tensor<1x1x4xi32>, vector<1x1x4xi32>
-// CHECK:           %[[VAL_10:.*]] = vector.step : vector<1xindex>
-// CHECK:           %[[VAL_11:.*]] = vector.broadcast %[[VAL_10]] : vector<1xindex> to vector<4x1x1xindex>
-// CHECK:           %[[VAL_12:.*]] = vector.transpose %[[VAL_11]], [2, 1, 0] : vector<4x1x1xindex> to vector<1x1x4xindex>
-// CHECK:           %[[VAL_13:.*]] = vector.step : vector<1xindex>
-// CHECK:           %[[VAL_14:.*]] = vector.broadcast %[[VAL_13]] : vector<1xindex> to vector<4x1x1xindex>
-// CHECK:           %[[VAL_15:.*]] = vector.transpose %[[VAL_14]], [2, 1, 0] : vector<4x1x1xindex> to vector<1x1x4xindex>
-// CHECK:           %[[VAL_16:.*]] = arith.constant dense<true> : vector<1x1x4xi1>
-// CHECK:           %[[VAL_17:.*]] = arith.constant dense<0> : vector<1x1x4xi32>
-// CHECK:           %[[VAL_18:.*]] = arith.constant 0 : index
-// CHECK:           %[[VAL_19:.*]] = arith.constant 0 : i32
-// CHECK:           %[[VAL_20:.*]] = vector.shape_cast %[[VAL_15]] : vector<1x1x4xindex> to vector<4xindex>
-// CHECK:           %[[VAL_21:.*]] = vector.extractelement %[[VAL_20]]{{\[}}%[[VAL_19]] : i32] : vector<4xindex>
-// CHECK:           %[[VAL_22:.*]] = arith.constant 0 : i32
-// CHECK:           %[[VAL_23:.*]] = vector.transfer_read %[[VAL_3]]{{\[}}%[[VAL_21]], %[[VAL_2]]], %[[VAL_22]] {in_bounds = [true, true, true], permutation_map = #[[$ATTR_1]]} : tensor<15x1xi32>, vector<1x1x4xi32>
-// CHECK:           %[[VAL_24:.*]] = arith.constant 0 : index
-// CHECK:           %[[VAL_25:.*]] = vector.transfer_write %[[VAL_23]], %[[VAL_0]]{{\[}}%[[VAL_24]], %[[VAL_24]], %[[VAL_24]]] : vector<1x1x4xi32>, tensor<1x1x4xi32>
-
-module attributes {transform.with_named_sequence} {
-  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
-    %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
-    transform.structured.vectorize %0  vector_sizes [1, 1, 4]{ vectorize_nd_extract } : !transform.any_op
-    transform.yield
-  }
-}