* weak references.
 *
 * <p><strong>Performance</strong>
 *
 * <p>The extra bookkeeping done by cycle detecting locks comes at some cost to performance.
 * Benchmarks (as of December 2011) show that:
 *
 * <ul>
 *   <li>for an unnested {@code lock()} and {@code unlock()}, a cycle detecting lock takes 38ns as
 *       opposed to the 24ns taken by a plain lock.

## Known Bugs

*   Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or
    `CUDA_ILLEGAL_ADDRESS` failures.

    Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA
    9 and CUDA 9.1 sometimes does not properly compute the carry bit when
    decomposing 64-bit address calculations with large offsets (e.g. `load [x +