.reason("If you don't declare the normalization, outputs can't be re-used between machines or locations on the same machine, therefore caching efficiency drops significantly")
            .solution("Declare the normalization strategy by annotating the property with either @PathSensitive, @Classpath or @CompileClasspath")
    }

 *       then as many exception-testing methods as there are exceptions the method can throw.
 *       Sometimes there are multiple tests per JSR166 method when the different "normal" behaviors
 *       differ significantly. And sometimes testcases cover multiple methods when they cannot be
 *       tested in isolation.
 *   <li>The documentation style for testcases is to provide as javadoc a simple sentence or two

 * which is optimal, and <i>not</i> the size of the internal hashtable, which could be much larger
 * than {@code size()}. Furthermore, this structure places significantly reduced load on the garbage
 * collector by only using a constant number of internal objects.
 *
 * <p>If there are no removals, then iteration order for the {@link #entrySet}, {@link #keySet}, and

 * The <<worker_api.adoc#worker_api,Worker API>> for enabling units of work to run in parallel.
 * A new API for <<#rel4.9:lazy_task_creation,creating and configuring tasks lazily>> that can significantly improve your build's configuration time.

Other notable changes to be aware of that may break your build include:

	// of queues and deal a hand of the size specified here.  The
	// request is put into one of the shortest queues in that hand.
	// `handSize` must be no larger than `queues`, and should be
	// significantly smaller (so that a few heavy flows do not
	// saturate most of the queues).  See the user-facing
	// documentation for more extensive guidance on setting this
	// field.  This field has a default value of 8.
	// +optional

	//
	// These limits may increase over time, but are currently at
	// most 48 bits except on s390x. On all architectures, Linux
	// starts placing mmap'd regions at addresses that are
	// significantly below 48 bits, so even if it's possible to
	// exceed Go's 48 bit limit, it's extremely unlikely in
	// practice.
	//
	// On 32-bit platforms, we accept the full 32-bit address
	// space because doing so is cheap.

	// that could be returned to the OS, but is being retained by
	// the runtime so it can grow the heap without requesting more
	// memory from the OS. If this difference is significantly
	// larger than the heap size, it indicates there was a recent
	// transient spike in live heap size.
	HeapIdle uint64

	// HeapInuse is bytes in in-use spans.
	//

 * which is optimal, and <i>not</i> the size of the internal hashtable, which could be much larger
 * than {@code size()}. Furthermore, this structure places significantly reduced load on the garbage
 * collector by only using a constant number of internal objects.
 *
 * <p>If there are no removals, then iteration order for the {@link #entrySet}, {@link #keySet}, and

        ' binary GPU code, or as "sm_xy" to only include the binary '
        'code.\nPlease note that each additional compute capability '
        'significantly increases your build time and binary size, and that '
        'TensorFlow only supports compute capabilities >= 3.5 [Default is: '
        '%s]: ' % default_cuda_compute_capabilities)

		end   int
	}
	spans := make([]span, 0, 32)

	// Find the field start and end indices.
	// Doing this in a separate pass (rather than slicing the string s
	// and collecting the result substrings right away) is significantly
	// more efficient, possibly due to cache effects.
	start := -1 // valid span start if >= 0
	for end, rune := range s {
		if f(rune) {
			if start >= 0 {
				spans = append(spans, span{start, end})