<arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.api=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.file=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.main=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.model=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.parser=ALL-UNNAMED</arg>

              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.api=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.file=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.main=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.model=ALL-UNNAMED</arg>
              <arg>-J--add-exports=jdk.compiler/com.sun.tools.javac.parser=ALL-UNNAMED</arg>

use explicit synchronization.
</p>

<h2 id="badcompiler">Incorrect compilation</h2>

<p>
The Go memory model restricts compiler optimizations as much as it does Go programs.
Some compiler optimizations that would be valid in single-threaded programs are not valid in all Go programs.
In particular, a compiler must not introduce writes that do not exist in the original program,
it must not allow a single read to observe multiple values,

    byte[] murmur3Hash = Hashing.murmur3_32().hashLong(42L).asBytes();
    byte[] combined = new byte[md5Hash.length + murmur3Hash.length];
    ByteBuffer buffer = ByteBuffer.wrap(combined);
    buffer.put(md5Hash);
    buffer.put(murmur3Hash);
    HashCode expected = HashCode.fromBytes(combined);

    assertEquals(
        expected, Hashing.concatenating(Hashing.md5(), Hashing.murmur3_32()).hashLong(42L));

## Compiler {#compiler}

## Assembler {#assembler}

            "@com_google_absl//absl/types:variant",
        ],
    }) + select({
        "//tensorflow:with_xla_support": [
            "//tensorflow/compiler/jit",
            "//tensorflow/compiler/jit:xla_device",
            "//tensorflow/compiler/tf2xla:xla_compiler",
        ],
        "//conditions:default": [],
    }) + [
        "//tensorflow/c:tf_status_helper",
        "//tensorflow/core:gpu_runtime",

  public static <E extends @Nullable Object> FluentIterable<E> from(FluentIterable<E> iterable) {
    return checkNotNull(iterable);
  }

  /**
   * Returns a fluent iterable that combines two iterables. The returned iterable has an iterator
   * that traverses the elements in {@code a}, followed by the elements in {@code b}. The source
   * iterators are not polled until necessary.
   *

INSTANTIATE_TEST_SUITE_P(
    UnifiedCAPI, CppGradients,
    ::testing::Combine(::testing::Values("graphdef", "mlir"),
                       /*tfrt*/ ::testing::Values(false),
                       /*executing_eagerly*/ ::testing::Values(true, false)));
#else
INSTANTIATE_TEST_SUITE_P(
    UnifiedCAPI, CppGradients,
    ::testing::Combine(::testing::Values("graphdef", "mlir"),

 * under GWT but with a slightly different signature.
 *
 * <p>We can't use {@code PartiallyGwtIncompatible} because then the GWT compiler wouldn't recognize
 * it as a {@code GwtIncompatible} annotation. And for {@code Futures.catching}, we need the GWT
 * compiler to autostrip the normal server method in order to expose the special, inherited GWT
 * version.
 */
@GwtCompatible
@ElementTypesAreNonnullByDefault

INSTANTIATE_TEST_SUITE_P(
    UnifiedCAPI, GradientCheckerTest,
    ::testing::Combine(::testing::Values("graphdef"),
                       /*tfrt*/ ::testing::Values(false),
                       /*use_function*/ ::testing::Values(true, false)));
#else
INSTANTIATE_TEST_SUITE_P(
    UnifiedCAPI, GradientCheckerTest,
    ::testing::Combine(::testing::Values("graphdef"),
                       /*tfrt*/ ::testing::Values(false),