%20 = "tf.Sum"(%loss, %6) {keep_dims = false} : (tensor<2xf32>, tensor<1xi32>) -> tensor<f32>
    %21 = "tf.Mul"(%20, %5) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %22 = "tf.Sum"(%21, %4) {keep_dims = false} : (tensor<f32>, tensor<0xi32>) -> tensor<f32>
    %23 = "tf.CrossReplicaSum"(%22, %3) : (tensor<f32>, tensor<1x2xi32>) -> tensor<f32>
    %24 = "tf.Softmax"(%18) : (tensor<2x1001xf32>) -> tensor<2x1001xf32>

//
// System calls and other sys.stuff for riscv64, Linux
//

#include "textflag.h"
#include "go_asm.h"

#define AT_FDCWD -100
#define CLOCK_REALTIME 0
#define CLOCK_MONOTONIC 1

#define SYS_brk			214
#define SYS_clock_gettime	113
#define SYS_clone		220
#define SYS_close		57
#define SYS_connect		203
#define SYS_exit		93
#define SYS_exit_group		94
#define SYS_faccessat		48
#define SYS_futex		98

    ByteArrayDataOutput out = ByteStreams.newDataOutput();
    out.writeChars("r\u00C9sum\u00C9");
    // need to remove byte order mark before comparing
    byte[] expected = Arrays.copyOfRange(utf16ExpectedWithBom, 2, 14);
    assertThat(out.toByteArray()).isEqualTo(expected);
  }

  @AndroidIncompatible // https://issuetracker.google.com/issues/37074504
  public void testUtf16Expected() {

#define SYS_gettid		178
#define SYS_futex		98
#define SYS_sched_getaffinity	123
#define SYS_exit_group		94
#define SYS_tgkill		131
#define SYS_openat		56
#define SYS_clock_gettime	113
#define SYS_brk			214
#define SYS_pipe2		59
#define SYS_timer_create	107
#define SYS_timer_settime	110
#define SYS_timer_delete	111

// func exit(code int32)
TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
	MOVW	code+0(FP), R4

  // resource or a constant.
  // The reason we optimized this code is because functions can have a lot of
  // captured arguments. For example, the backward pass of ResNet50 takes in all
  // 214 variables and a similar number of activations.
  SinglePassSearch constants_search(constant_arg_indices);
  SinglePassSearch resources_search(resource_arg_indices);
  for (size_t i = 0; i < fbody->arg_types.size(); ++i) {

    assertMean(-2, -3, -1);
    assertMean(0, -1, 1);
    assertMean(1, -1, 3);
    assertMean((1 << 30) - 1, -1, Integer.MAX_VALUE);

    // Even-sized ranges should prefer the lower mean
    assertMean(2, 1, 4);
    assertMean(-3, -4, -1);
    assertMean(0, -1, 2);
    assertMean(0, Integer.MIN_VALUE + 2, Integer.MAX_VALUE);
    assertMean(0, 0, 1);
    assertMean(-1, -1, 0);

  func.return %0 : tensor<2x5x4xf32>
  // CHECK-LABEL: einsum_matrixdotprod
  // CHECK-DAG: %[[cst:.*]] = arith.constant dense<[2, 5, 21]> : tensor<3xi64>
  // CHECK-DAG: %[[cst_1:.*]] = arith.constant dense<[21, 4]> : tensor<2xi64>
  // CHECK: %[[v0:.*]] = "tf.Reshape"(%arg0, %[[cst]]) : (tensor<2x5x7x3xf32>, tensor<3xi64>) -> tensor<2x5x21xf32>

#define SYS_futex		98
#define SYS_sched_getaffinity	123
#define SYS_exit_group		94
#define SYS_clock_gettime	113
#define SYS_faccessat		48
#define SYS_socket		198
#define SYS_connect		203
#define SYS_brk			214
#define SYS_timer_create	107
#define SYS_timer_settime	110
#define SYS_timer_delete	111

TEXT runtime·exit(SB),NOSPLIT|NOFRAME,$0-4
	MOVW	code+0(FP), R0
	MOVD	$SYS_exit_group, R8
	SVC
	RET

    assertMean(-2, -3, -1);
    assertMean(0, -1, 1);
    assertMean(1, -1, 3);
    assertMean((1L << 62) - 1, -1, Long.MAX_VALUE);

    // Even-sized ranges should prefer the lower mean
    assertMean(2, 1, 4);
    assertMean(-3, -4, -1);
    assertMean(0, -1, 2);
    assertMean(0, Long.MIN_VALUE + 2, Long.MAX_VALUE);
    assertMean(0, 0, 1);
    assertMean(-1, -1, 0);
    assertMean(-1, Long.MIN_VALUE, Long.MAX_VALUE);

    assertMean(-2, -3, -1);
    assertMean(0, -1, 1);
    assertMean(1, -1, 3);
    assertMean((1L << 62) - 1, -1, Long.MAX_VALUE);

    // Even-sized ranges should prefer the lower mean
    assertMean(2, 1, 4);
    assertMean(-3, -4, -1);
    assertMean(0, -1, 2);
    assertMean(0, Long.MIN_VALUE + 2, Long.MAX_VALUE);
    assertMean(0, 0, 1);
    assertMean(-1, -1, 0);
    assertMean(-1, Long.MIN_VALUE, Long.MAX_VALUE);