1.  API changes require discussion, use cases, etc. Code comes later.
2.  Pull requests are great for small fixes for bugs, documentation, etc.
3.  Pull requests are not merged directly into the master branch.
4.  Code contributions require signing a Google CLA.

API changes
-----------

We make changes to Guava's public [APIs][], including adding new APIs, very

          - os: windows-latest
            java: 21
            root-pom: pom.xml
    runs-on: ${{ matrix.os }}
    env:
      ROOT_POM: ${{ matrix.root-pom }}
    steps:
      # Cancel any previous runs for the same branch that are still running.
      - name: 'Cancel previous runs'
        uses: styfle/cancel-workflow-action@85880fa0301c86cca9da44039ee3bb12d3bedbfa # 0.12.1
        with:
          access_token: ${{ github.token }}

  @CheckForNull
  public abstract C previous(C value);

  /**
   * Returns a signed value indicating how many nested invocations of {@link #next} (if positive) or
   * {@link #previous} (if negative) are needed to reach {@code end} starting from {@code start}.
   * For example, if {@code end = next(next(next(start)))}, then {@code distance(start, end) == 3}
   * and {@code distance(end, start) == -3}. As well, {@code distance(a, a)} is always zero.

    return x > 0 & (x & (x - 1)) == 0;
  }

  /**
   * Returns 1 if {@code x < y} as unsigned longs, and 0 otherwise. Assumes that x - y fits into a
   * signed long. The implementation is branch-free, and benchmarks suggest it is measurably faster
   * than the straightforward ternary expression.
   */
  @VisibleForTesting
  static int lessThanBranchFree(long x, long y) {
    // Returns the sign bit of x - y.

 *       SHORT is always discriminable from zero time, and always allows enough time for the small
 *       amounts of computation (creating a thread, calling a few methods, etc) needed to reach a
 *       timeout point. Similarly, a SMALL is always discriminable as larger than SHORT and smaller
 *       than MEDIUM. And so on. These constants are set to conservative values, but even so, if

    @Override
    public boolean apply(CharSequence t) {
      return pattern.matcher(t).find();
    }

    @Override
    public int hashCode() {
      // Pattern uses Object.hashCode, so we have to reach
      // inside to build a hashCode consistent with equals.

      return Objects.hashCode(pattern.pattern(), pattern.flags());
    }

    @Override
    public boolean equals(@CheckForNull Object obj) {

   * than CPUs, supposing that each thread were bound to a CPU,
   * there would exist a perfect hash function mapping threads to
   * slots that eliminates collisions. When we reach capacity, we
   * search for this mapping by randomly varying the hash codes of
   * colliding threads.  Because search is random, and collisions
   * only become known via CAS failures, convergence can be slow,

            .put(key1, 1)
            .put(key2, 2)
            .put(key1, 3)
            .put(key2, 4)
            .buildKeepingLast();
    assertMapEquals(map, key1, 3, key2, 4);
  }

  // The java7 branch has different code depending on whether the entry indexes fit in a byte,
  // short, or int. The small table in testBuildKeepingLast_allowsOverwrite will test the byte
  // case. This method tests the short case.

 *       SHORT is always discriminable from zero time, and always allows enough time for the small
 *       amounts of computation (creating a thread, calling a few methods, etc) needed to reach a
 *       timeout point. Similarly, a SMALL is always discriminable as larger than SHORT and smaller
 *       than MEDIUM. And so on. These constants are set to conservative values, but even so, if

  /** Unboxes {@code obj}. Assumes that obj is not {@code null} or a {@link SetFuture}. */
  @ParametricNullness
  private V getDoneValue(Object obj) throws ExecutionException {
    // While this seems like it might be too branch-y, simple benchmarking proves it to be
    // unmeasurable (comparing done AbstractFutures with immediateFuture)
    if (obj instanceof Cancellation) {