// CHECK: "mhlo.dot"(%[[UPDATED_A]], %[[UPDATED_B]])
  %0 = "tf.MatMul"(%a, %b) {transpose_a = true, transpose_b = true} : (tensor<7x5xf32>, tensor<11x7xf32>) -> tensor<5x11xf32>

  func.return %0 : tensor<5x11xf32>
}

// Verify that MatMul with ranked inputs are lowered to HLO.
// CHECK-LABEL: matmul_ranked

  // - rhs: [RHSBATCHDIMS..., RHSROWS, RHSCOLS]
  // - result: [broadcast(LHSBATCHDIMS, RHSBATCHDIMS)..., LHSROWS, RHSCOLS]
  // To perform the matmul, we need to first broadcast lhs and rhs to a common
  // set of leading dimensions before doing the actual matmul.
  // That's what the code below does.
  // In particular, we populate out_lhs and out_rhs to have dimension structure: