%r_min_without_zero = "tf.Min"(%input_1d, %dim) { keep_dims = true }: (tensor<?xf32>, tensor<1xi64>) -> tensor<1xf32>
    %r_min = "tf.Minimum"(%zero, %r_min_without_zero) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    %r_max_f64 = "tf.Cast"(%r_max) : (tensor<1xf32>) -> tensor<1xf64>
    %r_min_f64 = "tf.Cast"(%r_min) : (tensor<1xf32>) -> tensor<1xf64>

    %i8_min = "tf.Const"() {value = dense<-128.0> : tensor<f32>} : () -> tensor<f32>

    for (int64_t i = 0; i < call.getNumOperands(); ++i) {
      auto arg_it = info.buffer_arg_to_size_arg.find(i);
      if (arg_it == info.buffer_arg_to_size_arg.end()) continue;
      auto it = buffer_to_size->find(call.getOperand(i));
      if (it == buffer_to_size->end()) {
        call.emitOpError("unknown tensor list.");
        return failure();
      }
      assert(arg_it->second == new_operands.size());

# LIMITATIONS
# 1. Exit code is probably not always correct.
# 2. There are no unittests.
# 3. Will not work if the total length of paths to addons is greater than
#    bash can handle. Probably it is not a problem: ARG_MAX=2097152 on GCE.

# cosmetic improvements to be done
# 1. Improve the log function; add timestamp, file name, etc.
# 2. Logging doesn't work from files that print things out.

      xla::PrintHloProfile(fn.hlo_profile_printer_data(), fn.profile_counters(),
                           /*clock_rate_ghz=*/1.0);
  VLOG(1) << "Original HLO profile string:\n" << hlo_profile_as_string;

  // Replace Arg_n with argn when the MLIR bridge is used.
#if defined(ENABLE_MLIR_BRIDGE_TEST)
  RE2::GlobalReplace(&hlo_profile_as_string, "(Arg_)([0-9].)", "arg\\2");
#endif

and offsets match.
On 32-bit systems, the low and high 32 bits of a 64-bit value are distinguished by adding
a <code>_lo</code> or <code>_hi</code> suffix to the name, as in <code>arg_lo+0(FP)</code> or <code>arg_hi+4(FP)</code>.
If a Go prototype does not name its result, the expected assembly name is <code>ret</code>.
</p>

<p>
The <code>SP</code> pseudo-register is a virtual stack pointer