}
		worst := mmuCurve.Examples(test.window, 2)
		// Which exact windows are returned is unspecified
		// (and depends on the exact banding), so we just
		// check that we got the right number with the right
		// utilizations.
		if len(worst) != len(test.worst) {
			t.Errorf("for %s window, want worst %v, got %v", test.window, test.worst, worst)
		} else {
			for i := range worst {

         })
         .done(function(worst) {
            details.text('Lowest mutator utilization in ' + niceDuration(windowNS) + ' windows:');
            for (var i = 0; i < worst.length; i++) {
              details.append($('<br>'));
              var text = worst[i].MutatorUtil.toFixed(3) + ' at time ' + niceDuration(worst[i].Time);
              details.append($('<a/>').text(text).attr('href', worst[i].URL));
            }
         });

		acc.mmu = mu
	}
	acc.bound = acc.mmu

	if acc.nWorst == 0 {
		// If the minimum has reached zero, it can't go any
		// lower, so we can stop early.
		return mu == 0
	}

	// Consider adding this window to the n worst.
	if len(acc.wHeap) < acc.nWorst || mu < acc.wHeap[0].MutatorUtil {
		// This window is lower than the K'th worst window.
		//
		// Check if there's any overlapping window

 * offering expected O(n + k log k) performance (worst case O(n log k)) for n calls to {@link
 * #offer} and a call to {@link #topK}, with O(k) memory. In comparison, quickselect has the same
 * asymptotics but requires O(n) memory, and a {@code PriorityQueue} implementation takes O(n log
 * k). In benchmarks, this implementation performs at least as well as either implementation, and
 * degrades more gracefully for worst-case input.
 *

	Error   *apiservercel.Error
	// If true, the compiled expression contains a reference to the identifier "oldSelf".
	UsesOldSelf bool
	// Represents the worst-case cost of the compiled expression in terms of CEL's cost units, as used by cel.EstimateCost.
	MaxCost uint64
	// MaxCardinality represents the worse case number of times this validation rule could be invoked if contained under an
	// unbounded map or list in an OpenAPIv3 schema.
	MaxCardinality uint64

 * offering expected O(n + k log k) performance (worst case O(n log k)) for n calls to {@link
 * #offer} and a call to {@link #topK}, with O(k) memory. In comparison, quickselect has the same
 * asymptotics but requires O(n) memory, and a {@code PriorityQueue} implementation takes O(n log
 * k). In benchmarks, this implementation performs at least as well as either implementation, and
 * degrades more gracefully for worst-case input.
 *

	constValue.reset(OpInvalid)
	return lattice{bottom, nil}
}

func (t *worklist) visitValue(val *Value) {
	if !possibleConst(val) {
		// fast fail for always worst Values, i.e. there is no lowering happen
		// on them, their lattices must be initially worse Bottom.
		return
	}

	oldLt := t.getLatticeCell(val)
	defer func() {
		// re-visit all uses of value if its lattice is changed
		newLt := t.getLatticeCell(val)

   *     .collect(least(2, comparingInt(String::length)))
   * // returns {"foo", "quux"}
   * }</pre>
   *
   * <p>This {@code Collector} uses O(k) memory and takes expected time O(n) (worst-case O(n log
   * k)), as opposed to e.g. {@code Stream.sorted(comparator).limit(k)}, which currently takes O(n
   * log n) time and O(n) space.
   *
   * @throws IllegalArgumentException if {@code k < 0}

	// on all our VirtualAlloc calls, try freeing successively smaller pieces until
	// we manage to free something, and then repeat. This ends up being O(n log n)
	// in the worst case, but that's fast enough.
	for n > 0 {
		small := n
		for small >= 4096 && stdcall3(_VirtualFree, uintptr(v), small, _MEM_DECOMMIT) == 0 {
			small /= 2
			small &^= 4096 - 1
		}
		if small < 4096 {

  snprintf(buf, kBufSize, "Mean of %2.0f%% best:", best_ratio * 100);
  std::string label_best(buf);
  std::vector<std::pair<std::string, double>> groups = {
      {"Best:", sorted_us.front()},
      {"Worst:", sorted_us.back()},
      {"Median:", sorted_us[count_us / 2]},
      {"Mean:", sum_us / count_us},
      {std::move(label_trimmed), sum_us_trimmed / count_us_trimmed},
      {std::move(label_best), sum_us_best / count_us_best},