[mlir][bufferization] Fix insertion point issue in EliminateEmptyTensors

[springerm]

Summary

The replacement op insertion point was off by one.

[guraypp]

good catch

mlir/lib/Dialect/Bufferization/Transforms/EmptyTensorElimination.cpp

neededValuesDominateInsertionPoint(const DominanceInfo &domInfo,
                                    const SmallVector<Value> &neededValues) {
   for (Value val : neededValues) {
     if (auto bbArg = dyn_cast<BlockArgument>(val)) {
       Block *owner = bbArg.getOwner();
       if (!owner->findAncestorOpInBlock(*insertionPoint))
         return false;
     } else {
       auto opResult = cast<OpResult>(val);
-      if (!domInfo.dominates(opResult.getOwner(), insertionPoint))
+      if (!domInfo.properlyDominates(opResult.getOwner(), insertionPoint))
         return false;
     }
   }
   return true;
 }
 
 /// Return true if the given `insertionPoint` dominates all uses of
 /// `emptyTensorOp`.

mlir/test/Dialect/Bufferization/Transforms/one-shot-bufferize-empty-tensor-elimination.mlir

^bb0(%in: i32, %out: i8):
       %tr = arith.trunci %in : i32 to i8
       linalg.yield %tr : i8
     } -> tensor<?x8xi8>
     %inserted_slice_3 = tensor.insert_slice %32 into %arg5[%p, %arg4] [%sz, 8] [1, 1] : tensor<?x8xi8> into tensor<?x16xi8>
     scf.yield %inserted_slice_3 : tensor<?x16xi8>
   }
   func.return %28 : tensor<?x16xi8>
 }
+
+// -----
+
+// CHECK-LABEL: func.func @regression_multiple_insertion_points(
+//   CHECK-NOT:   memref.alloc
+func.func @regression_multiple_insertion_points(%t1: tensor<?x?xf32>) -> tensor<?x?xf32> {
+  %empty = tensor.empty() : tensor<2x5xf32>
+  %f0 = arith.constant 5.5 : f32
+  %0 = "test.foo"() : () -> (index)
+  %1 = "test.bar"() : () -> (index)
+  %filled = linalg.fill ins(%f0 : f32) outs(%empty : tensor<2x5xf32>) -> tensor<2x5xf32>
+  %2 = tensor.insert_slice %filled into %t1 [%0, %1] [2, 5] [1, 1] : tensor<2x5xf32> into tensor<?x?xf32>
+  return %2 : tensor<?x?xf32>
+}