// CHECK-DAG:  %[[CONSTANT1:.*]] = arith.constant dense<0.000000e+00> : tensor<4xf32>
//              variance + epsilon
// CHECK:  %[[ADD1:.*]] = "tf.Add"(%[[ARG4:.*]], %[[CONSTANT]])
//              rsqrt(variance + epsilon)
// CHECK:  %[[RSQRT:.*]] = "tf.Rsqrt"(%[[ADD1]])
//              scale * rsqrt(variance + epsilon)
// CHECK:  %[[MUL1:.*]] = "tf.Mul"(%[[ARG1:.*]], %[[RSQRT]])

            %island1:3 = tf_executor.island {
              %add1 = "tf.Add"(%arg0, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
              %add2 = "tf.Add"(%add1, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
              %res = "tf.Print"(%add2) { message = "add result" } : (tensor<*xi32>) -> (tensor<*xi32>)
              tf_executor.yield %add1, %add2 : tensor<*xi32>, tensor<*xi32>
            }

        %island1:3 = tf_executor.island {
          %add1 = "tf.Add"(%arg0, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
          %add2 = "tf.Add"(%add1, %arg1) : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
          %res = "tf.Print"(%add2) { message = "add result" } : (tensor<*xi32>) -> (tensor<*xi32>)
          tf_executor.yield %add1, %add2 : tensor<*xi32>, tensor<*xi32>
        }

// Fusing:  %[[add:[0-9].*]] = tfl.add %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<1xf32>
// Fusing:  %[[add1:[0-9].*]] = tfl.add %arg0, %[[add]] {fused_activation_function = "RELU"} : tensor<1xf32>
// Fusing:  %[[relu:[0-9].*]] = "tfl.relu"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
// Fusing:  %[[add2:[0-9].*]] = tfl.add %[[relu]], %[[add1]] {fused_activation_function = "RELU6"} : tensor<1xf32>

Robert Griesemer <******@****.***> 1664399604 -0700

  }, {
    ^bb0(%barg0: tensor<*xf32>, %barg1: tensor<i32>):
      %add0 = "tf.Add"(%barg0, %barg0) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
      %add1 = "tf.Add"(%barg1, %barg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
      "tf.Yield"(%add0, %add1) : (tensor<*xf32>, tensor<i32>) -> ()
  }) { is_stateless = false } : (tensor<*xf32>, tensor<i32>) -> (tensor<*xf32>, tensor<i32>)

  func.return %0#0 : tensor<*xf32>
}

openAVX2Tail128LoopB:
	ADDQ     $16, itr2
	chachaQR_AVX2(AA1, BB1, CC1, DD1, TT0)
	VPALIGNR $4, BB1, BB1, BB1
	VPALIGNR $8, CC1, CC1, CC1
	VPALIGNR $12, DD1, DD1, DD1
	chachaQR_AVX2(AA1, BB1, CC1, DD1, TT0)
	VPALIGNR $12, BB1, BB1, BB1
	VPALIGNR $8, CC1, CC1, CC1
	VPALIGNR $4, DD1, DD1, DD1
	CMPQ     itr2, itr1
	JB       openAVX2Tail128LoopA
	CMPQ     itr2, $160
	JNE      openAVX2Tail128LoopB

  // CHECK-DAG: %[[ADD1:.*]] = chlo.broadcast_add %[[FEATURES]], %[[ONE]] {broadcast_dimensions = array<i64>}
  // CHECK-DAG: %[[MULGRAD:.*]] = mhlo.multiply %[[GRADIENTS]], %[[ADD1]] : (tensor<4x8xf32>, tensor<?x?xf32>) -> tensor<4x8xf32>
  // CHECK: %[[RESULT:.*]] = mhlo.select %[[PRED]], %[[GRADIENTS]], %[[MULGRAD]]

	t := &ret
	printnl() // ERROR "live at call to printnl: ret$"
	// Note: ret is live at the printnl because the compiler moves &ret
	// from before the printnl to after.
	useT40(t)
}

func ddd1(x, y *int) { // ERROR "live at entry to ddd1: x y$"
	ddd2(x, y) // ERROR "stack object .autotmp_[0-9]+ \[2\]\*int$"
	printnl()
	// Note: no .?autotmp live at printnl.  See issue 16996.
}

	t := &ret
	printnl() // ERROR "live at call to printnl: ret$"
	// Note: ret is live at the printnl because the compiler moves &ret
	// from before the printnl to after.
	useT40(t)
}

func ddd1(x, y *int) { // ERROR "live at entry to ddd1: x y$"
	ddd2(x, y) // ERROR "stack object .autotmp_[0-9]+ \[2\]\*int$"
	printnl()
	// Note: no .?autotmp live at printnl.  See issue 16996.
}