// `expression`.  It must be visible wherever `expression` is emitted.
    string variable_decl;
  };

  // Each cpp_shim corresponds to one embedded protocol buffer.
  std::vector<CPPShim> cpp_shims;

  // The contents of the object (".o") file the protocol buffers are embbed in.
  // This needs to be linked in to any program that wants to execute any of the
  // expressions in `cpp_shims`.
  string object_file_data;
};

TFCI_MACOS_CROSS_COMPILE_ENABLE=1
TFCI_MACOS_CROSS_COMPILE_SDK_DEST="tensorflow/tools/toolchains/cross_compile/cc/MacOSX.sdk"
TFCI_MACOS_CROSS_COMPILE_SDK_SOURCE="/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk"

// can be used to canonicalize IR, so there is always one `ParallelExecute`.
tf_device::ParallelExecuteOp BuildParallelExecuteOp(
    tf_device::ClusterFuncOp cluster_func, OpBuilder* builder);

// Unwrap `parallel_execute`'s contents if it only has one child.
LogicalResult RemoveSingletonParallelExecuteOp(
    tf_device::ParallelExecuteOp parallel_execute, OpBuilder* builder);

}  // namespace TF
}  // namespace mlir

# limitations under the License.
# ==============================================================================

name: ARM CI

on:
  push:
    branches:
      - master
      - r2.**
permissions:
  contents: read

jobs:
  build:
    # Don't do this in forks, and if labeled, only for 'kokoro:force-run'

std::optional<typename DeviceCompilationCache<ExecutableType>::Value>
DeviceCompilationCache<ExecutableType>::Lookup(const Key& key) const {
  // The outer lock protects the existence of the cache entry. It does not
  // protect the contents of the cache entry.
  Entry* entry;
  {
    mutex_lock lock(compile_cache_mu_);
    // Find cache entry.
    auto it = cache_.find(key);
    if (it == cache_.cend()) {
      return std::nullopt;
    }

import tensorflow.compat.v1 as tf
from tensorflow.compiler.mlir.tensorflow.tests.tf_saved_model import common_v1

# Verify that the tf.versions attribute exists. It is difficult to enforce
# contents, since the version numbers change over time. The conversion logic
# itself is verified in the common graphdef converter, so here just assert
# it is being invoked.
# CHECK: module
# CHECK-SAME: tf.versions

  // errors, the fallback path should be used whenever possible.
  kEnabledAfterGraphAnalysis,
};

// Analyzes the user requested policy as well as the contents of the graph and
// returns true when the MLIR Bridge should be run.
//
// If the user explicitly requests the bridge be enabled or disabled, this
// function will respect the request. If the user does not explicitly request

      describes model parameters. `flags_proto` describes conversion
      parameters (see relevant .protos for more information). Returns a string
      representing the contents of the converted model. When extended_return
      flag is set to true returns a dictionary that contains string representation
      of the converted model and some statistics like arithmetic ops count.

  const std::string src_file_path =
      absl::StrCat(testing::TempDir(), "/tmp_file");

  {
    std::ofstream ofs(src_file_path);
    ofs << "test_string";
  }

  // Test that the contents match.
  const absl::StatusOr<std::string> read_status =
      ReadFileToString(src_file_path);
  ASSERT_THAT(read_status, IsOk());
  EXPECT_THAT(*read_status, Eq("test_string"));
}

  tensorflow::tstring* output_value =
      static_cast<tensorflow::tstring*>(TF_TensorData(result));
  std::string file_contents(*output_value);
  EXPECT_NE(file_contents.find("TEST ASSET FILE CONTENTS"), std::string::npos);

  TF_DeleteTensor(result);
  TFE_DeleteTensorHandle(read_file_fn_outputs[0]);
  TFE_DeleteOp(read_file_op);
  TF_DeleteSavedModel(saved_model);
  TF_DeleteStatus(status);