Google Git
Sign in
quiche / quiche / 7cd97cd649fc269da99f922aca30258c99f28fa6 / . / quic / core / quic_interval_set.h
blob: af64e29a5a6961e7a546d19fed0e8eba27d78ff1 [file] [log] [blame]
// Copyright (c) 2019 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#ifndef QUICHE_QUIC_CORE_QUIC_INTERVAL_SET_H_
#define QUICHE_QUIC_CORE_QUIC_INTERVAL_SET_H_
// QuicIntervalSet<T> is a data structure used to represent a sorted set of
// non-empty, non-adjacent, and mutually disjoint intervals. Mutations to an
// interval set preserve these properties, altering the set as needed. For
// example, adding [2, 3) to a set containing only [1, 2) would result in the
// set containing the single interval [1, 3).
//
// Supported operations include testing whether an Interval is contained in the
// QuicIntervalSet, comparing two QuicIntervalSets, and performing
// QuicIntervalSet union, intersection, and difference.
//
// QuicIntervalSet maintains the minimum number of entries needed to represent
// the set of underlying intervals. When the QuicIntervalSet is modified (e.g.
// due to an Add operation), other interval entries may be coalesced, removed,
// or otherwise modified in order to maintain this invariant. The intervals are
// maintained in sorted order, by ascending min() value.
//
// The reader is cautioned to beware of the terminology used here: this library
// uses the terms "min" and "max" rather than "begin" and "end" as is
// conventional for the STL. The terminology [min, max) refers to the half-open
// interval which (if the interval is not empty) contains min but does not
// contain max. An interval is considered empty if min >= max.
//
// T is required to be default- and copy-constructible, to have an assignment
// operator, a difference operator (operator-()), and the full complement of
// comparison operators (<, <=, ==, !=, >=, >). These requirements are inherited
// from value_type.
//
// QuicIntervalSet has constant-time move operations.
//
//
// Examples:
// QuicIntervalSet<int> intervals;
// intervals.Add(Interval<int>(10, 20));
// intervals.Add(Interval<int>(30, 40));
// // intervals contains [10,20) and [30,40).
// intervals.Add(Interval<int>(15, 35));
// // intervals has been coalesced. It now contains the single range [10,40).
// EXPECT_EQ(1, intervals.Size());
// EXPECT_TRUE(intervals.Contains(Interval<int>(10, 40)));
//
// intervals.Difference(Interval<int>(10, 20));
// // intervals should now contain the single range [20, 40).
// EXPECT_EQ(1, intervals.Size());
// EXPECT_TRUE(intervals.Contains(Interval<int>(20, 40)));
#include <stddef.h>
#include <algorithm>
#include <initializer_list>
#include <set>
#include <utility>
#include <vector>
#include <iterator>
#include <string>
#include "absl/types/variant.h"
#include "net/third_party/quiche/src/quic/core/quic_interval.h"
#include "net/third_party/quiche/src/quic/platform/api/quic_containers.h"
#include "net/third_party/quiche/src/quic/platform/api/quic_flags.h"
#include "net/third_party/quiche/src/quic/platform/api/quic_logging.h"
namespace quic {
namespace oldquic {
template <typename T>
class QUIC_NO_EXPORT QuicIntervalSet {
public:
using value_type = QuicInterval<T>;
private:
struct QUIC_NO_EXPORT IntervalLess {
bool operator()(const value_type& a, const value_type& b) const;
};
// TODO(wub): Switch to absl::btree_set when it is available in Chromium.
using Set = std::set<value_type, IntervalLess>;
public:
using const_iterator = typename Set::const_iterator;
using const_reverse_iterator = typename Set::const_reverse_iterator;
// Instantiates an empty QuicIntervalSet.
QuicIntervalSet() {}
// Instantiates an QuicIntervalSet containing exactly one initial half-open
// interval [min, max), unless the given interval is empty, in which case the
// QuicIntervalSet will be empty.
explicit QuicIntervalSet(const value_type& interval) { Add(interval); }
// Instantiates an QuicIntervalSet containing the half-open interval [min,
// max).
QuicIntervalSet(const T& min, const T& max) { Add(min, max); }
QuicIntervalSet(std::initializer_list<value_type> il) { assign(il); }
// Clears this QuicIntervalSet.
void Clear() { intervals_.clear(); }
// Returns the number of disjoint intervals contained in this QuicIntervalSet.
size_t Size() const { return intervals_.size(); }
// Returns the smallest interval that contains all intervals in this
// QuicIntervalSet, or the empty interval if the set is empty.
value_type SpanningInterval() const;
// Adds "interval" to this QuicIntervalSet. Adding the empty interval has no
// effect.
void Add(const value_type& interval);
// Adds the interval [min, max) to this QuicIntervalSet. Adding the empty
// interval has no effect.
void Add(const T& min, const T& max) { Add(value_type(min, max)); }
// Same semantics as Add(const value_type&), but optimized for the case where
// rbegin()->min() <= |interval|.min() <= rbegin()->max().
void AddOptimizedForAppend(const value_type& interval) {
if (Empty()) {
Add(interval);
return;
}
const_reverse_iterator last_interval = intervals_.rbegin();
// If interval.min() is outside of [last_interval->min, last_interval->max],
// we can not simply extend last_interval->max.
if (interval.min() < last_interval->min() ||
interval.min() > last_interval->max()) {
Add(interval);
return;
}
if (interval.max() <= last_interval->max()) {
// interval is fully contained by last_interval.
return;
}
// Extend last_interval.max to interval.max, in place.
//
// Set does not allow in-place updates due to the potential of violating its
// ordering requirements. But we know setting the max of the last interval
// is safe w.r.t set ordering and other invariants of QuicIntervalSet, so we
// force an in-place update for performance.
const_cast<value_type*>(&(*last_interval))->SetMax(interval.max());
}
// Same semantics as Add(const T&, const T&), but optimized for the case where
// rbegin()->max() == |min|.
void AddOptimizedForAppend(const T& min, const T& max) {
AddOptimizedForAppend(value_type(min, max));
}
// TODO(wub): Similar to AddOptimizedForAppend, we can also have a
// AddOptimizedForPrepend if there is a use case.
// Remove the first interval.
// REQUIRES: !Empty()
void PopFront() {
DCHECK(!Empty());
intervals_.erase(intervals_.begin());
}
// Trim all values that is smaller than |value|. Which means
// a) If all values in an interval is smaller than |value|, the entire
// interval is removed.
// b) If some but not all values in an interval is smaller than |value|, the
// min of that interval is raised to |value|.
// Returns true if some intervals are trimmed.
bool TrimLessThan(const T& value) {
// Number of intervals that are fully or partially trimmed.
size_t num_intervals_trimmed = 0;
while (!intervals_.empty()) {
const_iterator first_interval = intervals_.begin();
if (first_interval->min() >= value) {
break;
}
++num_intervals_trimmed;
if (first_interval->max() <= value) {
// a) Trim the entire interval.
intervals_.erase(first_interval);
continue;
}
// b) Trim a prefix of the interval.
//
// Set does not allow in-place updates due to the potential of violating
// its ordering requirements. But increasing the min of the first interval
// will not break the ordering, hence the const_cast.
const_cast<value_type*>(&(*first_interval))->SetMin(value);
break;
}
return num_intervals_trimmed != 0;
}
// Returns true if this QuicIntervalSet is empty.
bool Empty() const { return intervals_.empty(); }
// Returns true if any interval in this QuicIntervalSet contains the indicated
// value.
bool Contains(const T& value) const;
// Returns true if there is some interval in this QuicIntervalSet that wholly
// contains the given interval. An interval O "wholly contains" a non-empty
// interval I if O.Contains(p) is true for every p in I. This is the same
// definition used by value_type::Contains(). This method returns false on
// the empty interval, due to a (perhaps unintuitive) convention inherited
// from value_type.
// Example:
// Assume an QuicIntervalSet containing the entries { [10,20), [30,40) }.
// Contains(Interval(15, 16)) returns true, because [10,20) contains
// [15,16). However, Contains(Interval(15, 35)) returns false.
bool Contains(const value_type& interval) const;
// Returns true if for each interval in "other", there is some (possibly
// different) interval in this QuicIntervalSet which wholly contains it. See
// Contains(const value_type& interval) for the meaning of "wholly contains".
// Perhaps unintuitively, this method returns false if "other" is the empty
// set. The algorithmic complexity of this method is O(other.Size() *
// log(this->Size())). The method could be rewritten to run in O(other.Size()
// + this->Size()), and this alternative could be implemented as a free
// function using the public API.
bool Contains(const QuicIntervalSet<T>& other) const;
// Returns true if there is some interval in this QuicIntervalSet that wholly
// contains the interval [min, max). See Contains(const value_type&).
bool Contains(const T& min, const T& max) const {
return Contains(value_type(min, max));
}
// Returns true if for some interval in "other", there is some interval in
// this QuicIntervalSet that intersects with it. See value_type::Intersects()
// for the definition of interval intersection.
bool Intersects(const QuicIntervalSet& other) const;
// Returns an iterator to the value_type in the QuicIntervalSet that contains
// the given value. In other words, returns an iterator to the unique interval
// [min, max) in the QuicIntervalSet that has the property min <= value < max.
// If there is no such interval, this method returns end().
const_iterator Find(const T& value) const;
// Returns an iterator to the value_type in the QuicIntervalSet that wholly
// contains the given interval. In other words, returns an iterator to the
// unique interval outer in the QuicIntervalSet that has the property that
// outer.Contains(interval). If there is no such interval, or if interval is
// empty, returns end().
const_iterator Find(const value_type& interval) const;
// Returns an iterator to the value_type in the QuicIntervalSet that wholly
// contains [min, max). In other words, returns an iterator to the unique
// interval outer in the QuicIntervalSet that has the property that
// outer.Contains(Interval<T>(min, max)). If there is no such interval, or if
// interval is empty, returns end().
const_iterator Find(const T& min, const T& max) const {
return Find(value_type(min, max));
}
// Returns an iterator pointing to the first value_type which contains or
// goes after the given value.
//
// Example:
// [10, 20) [30, 40)
// ^ LowerBound(10)
// ^ LowerBound(15)
// ^ LowerBound(20)
// ^ LowerBound(25)
const_iterator LowerBound(const T& value) const;
// Returns an iterator pointing to the first value_type which goes after
// the given value.
//
// Example:
// [10, 20) [30, 40)
// ^ UpperBound(10)
// ^ UpperBound(15)
// ^ UpperBound(20)
// ^ UpperBound(25)
const_iterator UpperBound(const T& value) const;
// Returns true if every value within the passed interval is not Contained
// within the QuicIntervalSet.
// Note that empty intervals are always considered disjoint from the
// QuicIntervalSet (even though the QuicIntervalSet doesn't `Contain` them).
bool IsDisjoint(const value_type& interval) const;
// Merges all the values contained in "other" into this QuicIntervalSet.
void Union(const QuicIntervalSet& other);
// Modifies this QuicIntervalSet so that it contains only those values that
// are currently present both in *this and in the QuicIntervalSet "other".
void Intersection(const QuicIntervalSet& other);
// Mutates this QuicIntervalSet so that it contains only those values that are
// currently in *this but not in "interval".
void Difference(const value_type& interval);
// Mutates this QuicIntervalSet so that it contains only those values that are
// currently in *this but not in the interval [min, max).
void Difference(const T& min, const T& max);
// Mutates this QuicIntervalSet so that it contains only those values that are
// currently in *this but not in the QuicIntervalSet "other".
void Difference(const QuicIntervalSet& other);
// Mutates this QuicIntervalSet so that it contains only those values that are
// in [min, max) but not currently in *this.
void Complement(const T& min, const T& max);
// QuicIntervalSet's begin() iterator. The invariants of QuicIntervalSet
// guarantee that for each entry e in the set, e.min() < e.max() (because the
// entries are non-empty) and for each entry f that appears later in the set,
// e.max() < f.min() (because the entries are ordered, pairwise-disjoint, and
// non-adjacent). Modifications to this QuicIntervalSet invalidate these
// iterators.
const_iterator begin() const { return intervals_.begin(); }
// QuicIntervalSet's end() iterator.
const_iterator end() const { return intervals_.end(); }
// QuicIntervalSet's rbegin() and rend() iterators. Iterator invalidation
// semantics are the same as those for begin() / end().
const_reverse_iterator rbegin() const { return intervals_.rbegin(); }
const_reverse_iterator rend() const { return intervals_.rend(); }
template <typename Iter>
void assign(Iter first, Iter last) {
Clear();
for (; first != last; ++first)
Add(*first);
}
void assign(std::initializer_list<value_type> il) {
assign(il.begin(), il.end());
}
// Returns a human-readable representation of this set. This will typically be
// (though is not guaranteed to be) of the form
// "[a1, b1) [a2, b2) ... [an, bn)"
// where the intervals are in the same order as given by traversal from
// begin() to end(). This representation is intended for human consumption;
// computer programs should not rely on the output being in exactly this form.
std::string ToString() const;
QuicIntervalSet& operator=(std::initializer_list<value_type> il) {
assign(il.begin(), il.end());
return *this;
}
friend bool operator==(const QuicIntervalSet& a, const QuicIntervalSet& b) {
return a.Size() == b.Size() &&
std::equal(a.begin(), a.end(), b.begin(), NonemptyIntervalEq());
}
friend bool operator!=(const QuicIntervalSet& a, const QuicIntervalSet& b) {
return !(a == b);
}
private:
// Simple member-wise equality, since all intervals are non-empty.
struct QUIC_NO_EXPORT NonemptyIntervalEq {
bool operator()(const value_type& a, const value_type& b) const {
return a.min() == b.min() && a.max() == b.max();
}
};
// Removes overlapping ranges and coalesces adjacent intervals as needed.
void Compact(const typename Set::iterator& begin,
const typename Set::iterator& end);
// Returns true if this set is valid (i.e. all intervals in it are non-empty,
// non-adjacent, and mutually disjoint). Currently this is used as an
// integrity check by the Intersection() and Difference() methods, but is only
// invoked for debug builds (via DCHECK).
bool Valid() const;
// Finds the first interval that potentially intersects 'other'.
const_iterator FindIntersectionCandidate(const QuicIntervalSet& other) const;
// Finds the first interval that potentially intersects 'interval'.
const_iterator FindIntersectionCandidate(const value_type& interval) const;
// Helper for Intersection() and Difference(): Finds the next pair of
// intervals from 'x' and 'y' that intersect. 'mine' is an iterator
// over x->intervals_. 'theirs' is an iterator over y.intervals_. 'mine'
// and 'theirs' are advanced until an intersecting pair is found.
// Non-intersecting intervals (aka "holes") from x->intervals_ can be
// optionally erased by "on_hole".
template <typename X, typename Func>
static bool FindNextIntersectingPairImpl(X* x,
const QuicIntervalSet& y,
const_iterator* mine,
const_iterator* theirs,
Func on_hole);
// The variant of the above method that doesn't mutate this QuicIntervalSet.
bool FindNextIntersectingPair(const QuicIntervalSet& other,
const_iterator* mine,
const_iterator* theirs) const {
return FindNextIntersectingPairImpl(
this, other, mine, theirs,
[](const QuicIntervalSet*, const_iterator, const_iterator) {});
}
// The variant of the above method that mutates this QuicIntervalSet by
// erasing holes.
bool FindNextIntersectingPairAndEraseHoles(const QuicIntervalSet& other,
const_iterator* mine,
const_iterator* theirs) {
return FindNextIntersectingPairImpl(
this, other, mine, theirs,
[](QuicIntervalSet* x, const_iterator from, const_iterator to) {
x->intervals_.erase(from, to);
});
}
// The representation for the intervals. The intervals in this set are
// non-empty, pairwise-disjoint, non-adjacent and ordered in ascending order
// by min().
Set intervals_;
};
template <typename T>
auto operator<<(std::ostream& out, const QuicIntervalSet<T>& seq)
-> decltype(out << *seq.begin()) {
out << "{";
for (const auto& interval : seq) {
out << " " << interval;
}
out << " }";
return out;
}
//==============================================================================
// Implementation details: Clients can stop reading here.
template <typename T>
typename QuicIntervalSet<T>::value_type QuicIntervalSet<T>::SpanningInterval()
const {
value_type result;
if (!intervals_.empty()) {
result.SetMin(intervals_.begin()->min());
result.SetMax(intervals_.rbegin()->max());
}
return result;
}
template <typename T>
void QuicIntervalSet<T>::Add(const value_type& interval) {
if (interval.Empty())
return;
std::pair<typename Set::iterator, bool> ins = intervals_.insert(interval);
if (!ins.second) {
// This interval already exists.
return;
}
// Determine the minimal range that will have to be compacted. We know that
// the QuicIntervalSet was valid before the addition of the interval, so only
// need to start with the interval itself (although Compact takes an open
// range so begin needs to be the interval to the left). We don't know how
// many ranges this interval may cover, so we need to find the appropriate
// interval to end with on the right.
typename Set::iterator begin = ins.first;
if (begin != intervals_.begin())
--begin;
const value_type target_end(interval.max(), interval.max());
const typename Set::iterator end = intervals_.upper_bound(target_end);
Compact(begin, end);
}
template <typename T>
bool QuicIntervalSet<T>::Contains(const T& value) const {
value_type tmp(value, value);
// Find the first interval with min() > value, then move back one step
const_iterator it = intervals_.upper_bound(tmp);
if (it == intervals_.begin())
return false;
--it;
return it->Contains(value);
}
template <typename T>
bool QuicIntervalSet<T>::Contains(const value_type& interval) const {
// Find the first interval with min() > value, then move back one step.
const_iterator it = intervals_.upper_bound(interval);
if (it == intervals_.begin())
return false;
--it;
return it->Contains(interval);
}
template <typename T>
bool QuicIntervalSet<T>::Contains(const QuicIntervalSet<T>& other) const {
if (!SpanningInterval().Contains(other.SpanningInterval())) {
return false;
}
for (const_iterator i = other.begin(); i != other.end(); ++i) {
// If we don't contain the interval, can return false now.
if (!Contains(*i)) {
return false;
}
}
return true;
}
// This method finds the interval that Contains() "value", if such an interval
// exists in the QuicIntervalSet. The way this is done is to locate the
// "candidate interval", the only interval that could *possibly* contain value,
// and test it using Contains(). The candidate interval is the interval with the
// largest min() having min() <= value.
//
// Determining the candidate interval takes a couple of steps. First, since the
// underlying std::set stores intervals, not values, we need to create a "probe
// interval" suitable for use as a search key. The probe interval used is
// [value, value). Now we can restate the problem as finding the largest
// interval in the QuicIntervalSet that is <= the probe interval.
//
// This restatement only works if the set's comparator behaves in a certain way.
// In particular it needs to order first by ascending min(), and then by
// descending max(). The comparator used by this library is defined in exactly
// this way. To see why descending max() is required, consider the following
// example. Assume an QuicIntervalSet containing these intervals:
//
// [0, 5) [10, 20) [50, 60)
//
// Consider searching for the value 15. The probe interval [15, 15) is created,
// and [10, 20) is identified as the largest interval in the set <= the probe
// interval. This is the correct interval needed for the Contains() test, which
// will then return true.
//
// Now consider searching for the value 30. The probe interval [30, 30) is
// created, and again [10, 20] is identified as the largest interval <= the
// probe interval. This is again the correct interval needed for the Contains()
// test, which in this case returns false.
//
// Finally, consider searching for the value 10. The probe interval [10, 10) is
// created. Here the ordering relationship between [10, 10) and [10, 20) becomes
// vitally important. If [10, 10) were to come before [10, 20), then [0, 5)
// would be the largest interval <= the probe, leading to the wrong choice of
// interval for the Contains() test. Therefore [10, 10) needs to come after
// [10, 20). The simplest way to make this work in the general case is to order
// by ascending min() but descending max(). In this ordering, the empty interval
// is larger than any non-empty interval with the same min(). The comparator
// used by this library is careful to induce this ordering.
//
// Another detail involves the choice of which std::set method to use to try to
// find the candidate interval. The most appropriate entry point is
// set::upper_bound(), which finds the smallest interval which is > the probe
// interval. The semantics of upper_bound() are slightly different from what we
// want (namely, to find the largest interval which is <= the probe interval)
// but they are close enough; the interval found by upper_bound() will always be
// one step past the interval we are looking for (if it exists) or at begin()
// (if it does not). Getting to the proper interval is a simple matter of
// decrementing the iterator.
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::Find(
const T& value) const {
value_type tmp(value, value);
const_iterator it = intervals_.upper_bound(tmp);
if (it == intervals_.begin())
return intervals_.end();
--it;
if (it->Contains(value))
return it;
else
return intervals_.end();
}
// This method finds the interval that Contains() the interval "probe", if such
// an interval exists in the QuicIntervalSet. The way this is done is to locate
// the "candidate interval", the only interval that could *possibly* contain
// "probe", and test it using Contains(). The candidate interval is the largest
// interval that is <= the probe interval.
//
// The search for the candidate interval only works if the comparator used
// behaves in a certain way. In particular it needs to order first by ascending
// min(), and then by descending max(). The comparator used by this library is
// defined in exactly this way. To see why descending max() is required,
// consider the following example. Assume an QuicIntervalSet containing these
// intervals:
//
// [0, 5) [10, 20) [50, 60)
//
// Consider searching for the probe [15, 17). [10, 20) is the largest interval
// in the set which is <= the probe interval. This is the correct interval
// needed for the Contains() test, which will then return true, because [10, 20)
// contains [15, 17).
//
// Now consider searching for the probe [30, 32). Again [10, 20] is the largest
// interval <= the probe interval. This is again the correct interval needed for
// the Contains() test, which in this case returns false, because [10, 20) does
// not contain [30, 32).
//
// Finally, consider searching for the probe [10, 12). Here the ordering
// relationship between [10, 12) and [10, 20) becomes vitally important. If
// [10, 12) were to come before [10, 20), then [0, 5) would be the largest
// interval <= the probe, leading to the wrong choice of interval for the
// Contains() test. Therefore [10, 12) needs to come after [10, 20). The
// simplest way to make this work in the general case is to order by ascending
// min() but descending max(). In this ordering, given two intervals with the
// same min(), the wider one goes before the narrower one. The comparator used
// by this library is careful to induce this ordering.
//
// Another detail involves the choice of which std::set method to use to try to
// find the candidate interval. The most appropriate entry point is
// set::upper_bound(), which finds the smallest interval which is > the probe
// interval. The semantics of upper_bound() are slightly different from what we
// want (namely, to find the largest interval which is <= the probe interval)
// but they are close enough; the interval found by upper_bound() will always be
// one step past the interval we are looking for (if it exists) or at begin()
// (if it does not). Getting to the proper interval is a simple matter of
// decrementing the iterator.
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::Find(
const value_type& probe) const {
const_iterator it = intervals_.upper_bound(probe);
if (it == intervals_.begin())
return intervals_.end();
--it;
if (it->Contains(probe))
return it;
else
return intervals_.end();
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::LowerBound(
const T& value) const {
const_iterator it = intervals_.lower_bound(value_type(value, value));
if (it == intervals_.begin()) {
return it;
}
// The previous intervals_.lower_bound() checking is essentially based on
// interval.min(), so we need to check whether the `value` is contained in
// the previous interval.
--it;
if (it->Contains(value)) {
return it;
} else {
return ++it;
}
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::UpperBound(
const T& value) const {
return intervals_.upper_bound(value_type(value, value));
}
template <typename T>
bool QuicIntervalSet<T>::IsDisjoint(const value_type& interval) const {
if (interval.Empty())
return true;
value_type tmp(interval.min(), interval.min());
// Find the first interval with min() > interval.min()
const_iterator it = intervals_.upper_bound(tmp);
if (it != intervals_.end() && interval.max() > it->min())
return false;
if (it == intervals_.begin())
return true;
--it;
return it->max() <= interval.min();
}
template <typename T>
void QuicIntervalSet<T>::Union(const QuicIntervalSet& other) {
intervals_.insert(other.begin(), other.end());
Compact(intervals_.begin(), intervals_.end());
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator
QuicIntervalSet<T>::FindIntersectionCandidate(
const QuicIntervalSet& other) const {
return FindIntersectionCandidate(*other.intervals_.begin());
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator
QuicIntervalSet<T>::FindIntersectionCandidate(
const value_type& interval) const {
// Use upper_bound to efficiently find the first interval in intervals_
// where min() is greater than interval.min(). If the result
// isn't the beginning of intervals_ then move backwards one interval since
// the interval before it is the first candidate where max() may be
// greater than interval.min().
// In other words, no interval before that can possibly intersect with any
// of other.intervals_.
const_iterator mine = intervals_.upper_bound(interval);
if (mine != intervals_.begin()) {
--mine;
}
return mine;
}
template <typename T>
template <typename X, typename Func>
bool QuicIntervalSet<T>::FindNextIntersectingPairImpl(X* x,
const QuicIntervalSet& y,
const_iterator* mine,
const_iterator* theirs,
Func on_hole) {
CHECK(x != nullptr);
if ((*mine == x->intervals_.end()) || (*theirs == y.intervals_.end())) {
return false;
}
while (!(**mine).Intersects(**theirs)) {
const_iterator erase_first = *mine;
// Skip over intervals in 'mine' that don't reach 'theirs'.
while (*mine != x->intervals_.end() && (**mine).max() <= (**theirs).min()) {
++(*mine);
}
on_hole(x, erase_first, *mine);
// We're done if the end of intervals_ is reached.
if (*mine == x->intervals_.end()) {
return false;
}
// Skip over intervals 'theirs' that don't reach 'mine'.
while (*theirs != y.intervals_.end() &&
(**theirs).max() <= (**mine).min()) {
++(*theirs);
}
// If the end of other.intervals_ is reached, we're done.
if (*theirs == y.intervals_.end()) {
on_hole(x, *mine, x->intervals_.end());
return false;
}
}
return true;
}
template <typename T>
void QuicIntervalSet<T>::Intersection(const QuicIntervalSet& other) {
if (!SpanningInterval().Intersects(other.SpanningInterval())) {
intervals_.clear();
return;
}
const_iterator mine = FindIntersectionCandidate(other);
// Remove any intervals that cannot possibly intersect with other.intervals_.
intervals_.erase(intervals_.begin(), mine);
const_iterator theirs = other.FindIntersectionCandidate(*this);
while (FindNextIntersectingPairAndEraseHoles(other, &mine, &theirs)) {
// OK, *mine and *theirs intersect. Now, we find the largest
// span of intervals in other (starting at theirs) - say [a..b]
// - that intersect *mine, and we replace *mine with (*mine
// intersect x) for all x in [a..b] Note that subsequent
// intervals in this can't intersect any intervals in [a..b) --
// they may only intersect b or subsequent intervals in other.
value_type i(*mine);
intervals_.erase(mine);
mine = intervals_.end();
value_type intersection;
while (theirs != other.intervals_.end() &&
i.Intersects(*theirs, &intersection)) {
std::pair<typename Set::iterator, bool> ins =
intervals_.insert(intersection);
DCHECK(ins.second);
mine = ins.first;
++theirs;
}
DCHECK(mine != intervals_.end());
--theirs;
++mine;
}
DCHECK(Valid());
}
template <typename T>
bool QuicIntervalSet<T>::Intersects(const QuicIntervalSet& other) const {
if (!SpanningInterval().Intersects(other.SpanningInterval())) {
return false;
}
const_iterator mine = FindIntersectionCandidate(other);
if (mine == intervals_.end()) {
return false;
}
const_iterator theirs = other.FindIntersectionCandidate(*mine);
return FindNextIntersectingPair(other, &mine, &theirs);
}
template <typename T>
void QuicIntervalSet<T>::Difference(const value_type& interval) {
if (!SpanningInterval().Intersects(interval)) {
return;
}
Difference(QuicIntervalSet<T>(interval));
}
template <typename T>
void QuicIntervalSet<T>::Difference(const T& min, const T& max) {
Difference(value_type(min, max));
}
template <typename T>
void QuicIntervalSet<T>::Difference(const QuicIntervalSet& other) {
if (!SpanningInterval().Intersects(other.SpanningInterval())) {
return;
}
const_iterator mine = FindIntersectionCandidate(other);
// If no interval in mine reaches the first interval of theirs then we're
// done.
if (mine == intervals_.end()) {
return;
}
const_iterator theirs = other.FindIntersectionCandidate(*this);
while (FindNextIntersectingPair(other, &mine, &theirs)) {
// At this point *mine and *theirs overlap. Remove mine from
// intervals_ and replace it with the possibly two intervals that are
// the difference between mine and theirs.
value_type i(*mine);
intervals_.erase(mine++);
value_type lo;
value_type hi;
i.Difference(*theirs, &lo, &hi);
if (!lo.Empty()) {
// We have a low end. This can't intersect anything else.
std::pair<typename Set::iterator, bool> ins = intervals_.insert(lo);
DCHECK(ins.second);
}
if (!hi.Empty()) {
std::pair<typename Set::iterator, bool> ins = intervals_.insert(hi);
DCHECK(ins.second);
mine = ins.first;
}
}
DCHECK(Valid());
}
template <typename T>
void QuicIntervalSet<T>::Complement(const T& min, const T& max) {
QuicIntervalSet<T> span(min, max);
span.Difference(*this);
intervals_.swap(span.intervals_);
}
template <typename T>
std::string QuicIntervalSet<T>::ToString() const {
std::ostringstream os;
os << *this;
return os.str();
}
// This method compacts the QuicIntervalSet, merging pairs of overlapping
// intervals into a single interval. In the steady state, the QuicIntervalSet
// does not contain any such pairs. However, the way the Union() and Add()
// methods work is to temporarily put the QuicIntervalSet into such a state and
// then to call Compact() to "fix it up" so that it is no longer in that state.
//
// Compact() needs the interval set to allow two intervals [a,b) and [a,c)
// (having the same min() but different max()) to briefly coexist in the set at
// the same time, and be adjacent to each other, so that they can be efficiently
// located and merged into a single interval. This state would be impossible
// with a comparator which only looked at min(), as such a comparator would
// consider such pairs equal. Fortunately, the comparator used by
// QuicIntervalSet does exactly what is needed, ordering first by ascending
// min(), then by descending max().
template <typename T>
void QuicIntervalSet<T>::Compact(const typename Set::iterator& begin,
const typename Set::iterator& end) {
if (begin == end)
return;
typename Set::iterator next = begin;
typename Set::iterator prev = begin;
typename Set::iterator it = begin;
++it;
++next;
while (it != end) {
++next;
if (prev->max() >= it->min()) {
// Overlapping / coalesced range; merge the two intervals.
T min = prev->min();
T max = std::max(prev->max(), it->max());
value_type i(min, max);
intervals_.erase(prev);
intervals_.erase(it);
std::pair<typename Set::iterator, bool> ins = intervals_.insert(i);
DCHECK(ins.second);
prev = ins.first;
} else {
prev = it;
}
it = next;
}
}
template <typename T>
bool QuicIntervalSet<T>::Valid() const {
const_iterator prev = end();
for (const_iterator it = begin(); it != end(); ++it) {
// invalid or empty interval.
if (it->min() >= it->max())
return false;
// Not sorted, not disjoint, or adjacent.
if (prev != end() && prev->max() >= it->min())
return false;
prev = it;
}
return true;
}
// This comparator orders intervals first by ascending min() and then by
// descending max(). Readers who are satisified with that explanation can stop
// reading here. The remainder of this comment is for the benefit of future
// maintainers of this library.
//
// The reason for this ordering is that this comparator has to serve two
// masters. First, it has to maintain the intervals in its internal set in the
// order that clients expect to see them. Clients see these intervals via the
// iterators provided by begin()/end() or as a result of invoking Get(). For
// this reason, the comparator orders intervals by ascending min().
//
// If client iteration were the only consideration, then ordering by ascending
// min() would be good enough. This is because the intervals in the
// QuicIntervalSet are non-empty, non-adjacent, and mutually disjoint; such
// intervals happen to always have disjoint min() values, so such a comparator
// would never even have to look at max() in order to work correctly for this
// class.
//
// However, in addition to ordering by ascending min(), this comparator also has
// a second responsibility: satisfying the special needs of this library's
// peculiar internal implementation. These needs require the comparator to order
// first by ascending min() and then by descending max(). The best way to
// understand why this is so is to check out the comments associated with the
// Find() and Compact() methods.
template <typename T>
bool QuicIntervalSet<T>::IntervalLess::operator()(const value_type& a,
const value_type& b) const {
return a.min() < b.min() || (a.min() == b.min() && a.max() > b.max());
}
} // namespace oldquic
namespace newquic {
template <typename T>
class QUIC_NO_EXPORT QuicIntervalSet {
public:
typedef QuicInterval<T> value_type;
private:
struct QUIC_NO_EXPORT IntervalLess {
using is_transparent = void;
bool operator()(const value_type& a, const value_type& b) const;
// These transparent overloads are used when we do all of our searches (via
// Set::lower_bound() and Set::upper_bound()), which avoids the need to
// construct an interval when we are looking for a point and also avoids
// needing to worry about comparing overlapping intervals in the overload
// that takes two value_types (the one just above this comment).
bool operator()(const value_type& a, const T& point) const;
bool operator()(const value_type& a, T&& point) const;
bool operator()(const T& point, const value_type& a) const;
bool operator()(T&& point, const value_type& a) const;
};
typedef QuicOrderedSet<value_type,
IntervalLess,
QuicInlinedVector<value_type, 10>>
Set;
public:
typedef typename Set::const_iterator const_iterator;
typedef typename Set::const_reverse_iterator const_reverse_iterator;
// Instantiates an empty QuicIntervalSet.
QuicIntervalSet() = default;
// Instantiates a QuicIntervalSet containing exactly one initial half-open
// interval [min, max), unless the given interval is empty, in which case the
// QuicIntervalSet will be empty.
explicit QuicIntervalSet(const value_type& interval) { Add(interval); }
// Instantiates a QuicIntervalSet containing the half-open interval [min,
// max).
QuicIntervalSet(const T& min, const T& max) { Add(min, max); }
QuicIntervalSet(std::initializer_list<value_type> il) { assign(il); }
// Clears this QuicIntervalSet.
void Clear() { intervals_.clear(); }
// Returns the number of disjoint intervals contained in this QuicIntervalSet.
size_t Size() const { return intervals_.size(); }
// Returns the smallest interval that contains all intervals in this
// QuicIntervalSet, or the empty interval if the set is empty.
value_type SpanningInterval() const;
// Adds "interval" to this QuicIntervalSet. Adding the empty interval has no
// effect.
void Add(const value_type& interval);
// Adds the interval [min, max) to this QuicIntervalSet. Adding the empty
// interval has no effect.
void Add(const T& min, const T& max) { Add(value_type(min, max)); }
// Same semantics as Add(const value_type&), but optimized for the case where
// rbegin()->min() <= |interval|.min() <= rbegin()->max().
void AddOptimizedForAppend(const value_type& interval) {
if (Empty()) {
Add(interval);
return;
}
const_reverse_iterator last_interval = intervals_.rbegin();
// If interval.min() is outside of [last_interval->min, last_interval->max],
// we can not simply extend last_interval->max.
if (interval.min() < last_interval->min() ||
interval.min() > last_interval->max()) {
Add(interval);
return;
}
if (interval.max() <= last_interval->max()) {
// interval is fully contained by last_interval.
return;
}
// Extend last_interval.max to interval.max, in place.
//
// Set does not allow in-place updates due to the potential of violating its
// ordering requirements. But we know setting the max of the last interval
// is safe w.r.t set ordering and other invariants of QuicIntervalSet, so we
// force an in-place update for performance.
const_cast<value_type*>(&(*last_interval))->SetMax(interval.max());
}
// Same semantics as Add(const T&, const T&), but optimized for the case where
// rbegin()->max() == |min|.
void AddOptimizedForAppend(const T& min, const T& max) {
AddOptimizedForAppend(value_type(min, max));
}
// TODO(wub): Similar to AddOptimizedForAppend, we can also have a
// AddOptimizedForPrepend if there is a use case.
// Remove the first interval.
// REQUIRES: !Empty()
void PopFront() {
DCHECK(!Empty());
intervals_.erase(intervals_.begin());
}
// Trim all values that are smaller than |value|. Which means
// a) If all values in an interval is smaller than |value|, the entire
// interval is removed.
// b) If some but not all values in an interval is smaller than |value|, the
// min of that interval is raised to |value|.
// Returns true if some intervals are trimmed.
bool TrimLessThan(const T& value) {
// Number of intervals that are fully or partially trimmed.
size_t num_intervals_trimmed = 0;
while (!intervals_.empty()) {
const_iterator first_interval = intervals_.begin();
if (first_interval->min() >= value) {
break;
}
++num_intervals_trimmed;
if (first_interval->max() <= value) {
// a) Trim the entire interval.
intervals_.erase(first_interval);
continue;
}
// b) Trim a prefix of the interval.
//
// Set does not allow in-place updates due to the potential of violating
// its ordering requirements. But increasing the min of the first interval
// will not break the ordering, hence the const_cast.
const_cast<value_type*>(&(*first_interval))->SetMin(value);
break;
}
return num_intervals_trimmed != 0;
}
// Returns true if this QuicIntervalSet is empty.
bool Empty() const { return intervals_.empty(); }
// Returns true if any interval in this QuicIntervalSet contains the indicated
// value.
bool Contains(const T& value) const;
// Returns true if there is some interval in this QuicIntervalSet that wholly
// contains the given interval. An interval O "wholly contains" a non-empty
// interval I if O.Contains(p) is true for every p in I. This is the same
// definition used by value_type::Contains(). This method returns false on
// the empty interval, due to a (perhaps unintuitive) convention inherited
// from value_type.
// Example:
// Assume an QuicIntervalSet containing the entries { [10,20), [30,40) }.
// Contains(Interval(15, 16)) returns true, because [10,20) contains
// [15,16). However, Contains(Interval(15, 35)) returns false.
bool Contains(const value_type& interval) const;
// Returns true if for each interval in "other", there is some (possibly
// different) interval in this QuicIntervalSet which wholly contains it. See
// Contains(const value_type& interval) for the meaning of "wholly contains".
// Perhaps unintuitively, this method returns false if "other" is the empty
// set. The algorithmic complexity of this method is O(other.Size() *
// log(this->Size())). The method could be rewritten to run in O(other.Size()
// + this->Size()), and this alternative could be implemented as a free
// function using the public API.
bool Contains(const QuicIntervalSet<T>& other) const;
// Returns true if there is some interval in this QuicIntervalSet that wholly
// contains the interval [min, max). See Contains(const value_type&).
bool Contains(const T& min, const T& max) const {
return Contains(value_type(min, max));
}
// Returns true if for some interval in "other", there is some interval in
// this QuicIntervalSet that intersects with it. See value_type::Intersects()
// for the definition of interval intersection. Runs in time O(n+m) where n
// is the number of intervals in this and m is the number of intervals in
// other.
bool Intersects(const QuicIntervalSet& other) const;
// Returns an iterator to the value_type in the QuicIntervalSet that contains
// the given value. In other words, returns an iterator to the unique interval
// [min, max) in the QuicIntervalSet that has the property min <= value < max.
// If there is no such interval, this method returns end().
const_iterator Find(const T& value) const;
// Returns an iterator to the value_type in the QuicIntervalSet that wholly
// contains the given interval. In other words, returns an iterator to the
// unique interval outer in the QuicIntervalSet that has the property that
// outer.Contains(interval). If there is no such interval, or if interval is
// empty, returns end().
const_iterator Find(const value_type& interval) const;
// Returns an iterator to the value_type in the QuicIntervalSet that wholly
// contains [min, max). In other words, returns an iterator to the unique
// interval outer in the QuicIntervalSet that has the property that
// outer.Contains(Interval<T>(min, max)). If there is no such interval, or if
// interval is empty, returns end().
const_iterator Find(const T& min, const T& max) const {
return Find(value_type(min, max));
}
// Returns an iterator pointing to the first value_type which contains or
// goes after the given value.
//
// Example:
// [10, 20) [30, 40)
// ^ LowerBound(10)
// ^ LowerBound(15)
// ^ LowerBound(20)
// ^ LowerBound(25)
const_iterator LowerBound(const T& value) const;
// Returns an iterator pointing to the first value_type which goes after
// the given value.
//
// Example:
// [10, 20) [30, 40)
// ^ UpperBound(10)
// ^ UpperBound(15)
// ^ UpperBound(20)
// ^ UpperBound(25)
const_iterator UpperBound(const T& value) const;
// Returns true if every value within the passed interval is not Contained
// within the QuicIntervalSet.
// Note that empty intervals are always considered disjoint from the
// QuicIntervalSet (even though the QuicIntervalSet doesn't `Contain` them).
bool IsDisjoint(const value_type& interval) const;
// Merges all the values contained in "other" into this QuicIntervalSet.
//
// Performance: Let n == Size() and m = other.Size(). Union() runs in O(m)
// Set operations, so that if Set is a tree, it runs in time O(m log(n+m)) and
// if Set is a flat_set it runs in time O(m(n+m)). In principle, for the
// flat_set, we should be able to make this run in time O(n+m).
//
// TODO(bradleybear): Make Union() run in time O(n+m) for flat_set. This may
// require an additional template parameter to indicate that the Set is a
// linear-time data structure instead of a log-time data structure.
void Union(const QuicIntervalSet& other);
// Modifies this QuicIntervalSet so that it contains only those values that
// are currently present both in *this and in the QuicIntervalSet "other".
void Intersection(const QuicIntervalSet& other);
// Mutates this QuicIntervalSet so that it contains only those values that are
// currently in *this but not in "interval".
void Difference(const value_type& interval);
// Mutates this QuicIntervalSet so that it contains only those values that are
// currently in *this but not in the interval [min, max).
void Difference(const T& min, const T& max);
// Mutates this QuicIntervalSet so that it contains only those values that are
// currently in *this but not in the QuicIntervalSet "other". Runs in time
// O(n+m) where n is this->Size(), m is other.Size(), regardless of whether
// the Set is a flat_set or a std::set.
void Difference(const QuicIntervalSet& other);
// Mutates this QuicIntervalSet so that it contains only those values that are
// in [min, max) but not currently in *this.
void Complement(const T& min, const T& max);
// QuicIntervalSet's begin() iterator. The invariants of QuicIntervalSet
// guarantee that for each entry e in the set, e.min() < e.max() (because the
// entries are non-empty) and for each entry f that appears later in the set,
// e.max() < f.min() (because the entries are ordered, pairwise-disjoint, and
// non-adjacent). Modifications to this QuicIntervalSet invalidate these
// iterators.
const_iterator begin() const { return intervals_.begin(); }
// QuicIntervalSet's end() iterator.
const_iterator end() const { return intervals_.end(); }
// QuicIntervalSet's rbegin() and rend() iterators. Iterator invalidation
// semantics are the same as those for begin() / end().
const_reverse_iterator rbegin() const { return intervals_.rbegin(); }
const_reverse_iterator rend() const { return intervals_.rend(); }
template <typename Iter>
void assign(Iter first, Iter last) {
Clear();
for (; first != last; ++first)
Add(*first);
}
void assign(std::initializer_list<value_type> il) {
assign(il.begin(), il.end());
}
// Returns a human-readable representation of this set. This will typically be
// (though is not guaranteed to be) of the form
// "[a1, b1) [a2, b2) ... [an, bn)"
// where the intervals are in the same order as given by traversal from
// begin() to end(). This representation is intended for human consumption;
// computer programs should not rely on the output being in exactly this form.
std::string ToString() const;
QuicIntervalSet& operator=(std::initializer_list<value_type> il) {
assign(il.begin(), il.end());
return *this;
}
friend bool operator==(const QuicIntervalSet& a, const QuicIntervalSet& b) {
return a.Size() == b.Size() &&
std::equal(a.begin(), a.end(), b.begin(), NonemptyIntervalEq());
}
friend bool operator!=(const QuicIntervalSet& a, const QuicIntervalSet& b) {
return !(a == b);
}
private:
// Simple member-wise equality, since all intervals are non-empty.
struct QUIC_NO_EXPORT NonemptyIntervalEq {
bool operator()(const value_type& a, const value_type& b) const {
return a.min() == b.min() && a.max() == b.max();
}
};
// Returns true if this set is valid (i.e. all intervals in it are non-empty,
// non-adjacent, and mutually disjoint). Currently this is used as an
// integrity check by the Intersection() and Difference() methods, but is only
// invoked for debug builds (via DCHECK).
bool Valid() const;
// Finds the first interval that potentially intersects 'other'.
const_iterator FindIntersectionCandidate(const QuicIntervalSet& other) const;
// Finds the first interval that potentially intersects 'interval'. More
// precisely, return an interator it pointing at the last interval J such that
// interval <= J. If all the intervals are > J then return begin().
const_iterator FindIntersectionCandidate(const value_type& interval) const;
// Helper for Intersection() and Difference(): Finds the next pair of
// intervals from 'x' and 'y' that intersect. 'mine' is an iterator
// over x->intervals_. 'theirs' is an iterator over y.intervals_. 'mine'
// and 'theirs' are advanced until an intersecting pair is found.
// Non-intersecting intervals (aka "holes") from x->intervals_ can be
// optionally erased by "on_hole". "on_hole" must return an iterator to the
// first element in 'x' after the hole, or x->intervals_.end() if no elements
// exist after the hole.
template <typename X, typename Func>
static bool FindNextIntersectingPairImpl(X* x,
const QuicIntervalSet& y,
const_iterator* mine,
const_iterator* theirs,
Func on_hole);
// The variant of the above method that doesn't mutate this QuicIntervalSet.
bool FindNextIntersectingPair(const QuicIntervalSet& other,
const_iterator* mine,
const_iterator* theirs) const {
return FindNextIntersectingPairImpl(
this, other, mine, theirs,
[](const QuicIntervalSet*, const_iterator, const_iterator end) {
return end;
});
}
// The variant of the above method that mutates this QuicIntervalSet by
// erasing holes.
bool FindNextIntersectingPairAndEraseHoles(const QuicIntervalSet& other,
const_iterator* mine,
const_iterator* theirs) {
return FindNextIntersectingPairImpl(
this, other, mine, theirs,
[](QuicIntervalSet* x, const_iterator from, const_iterator to) {
return x->intervals_.erase(from, to);
});
}
// The representation for the intervals. The intervals in this set are
// non-empty, pairwise-disjoint, non-adjacent and ordered in ascending order
// by min().
Set intervals_;
};
template <typename T>
auto operator<<(std::ostream& out, const QuicIntervalSet<T>& seq)
-> decltype(out << *seq.begin()) {
out << "{";
for (const auto& interval : seq) {
out << " " << interval;
}
out << " }";
return out;
}
//==============================================================================
// Implementation details: Clients can stop reading here.
template <typename T>
typename QuicIntervalSet<T>::value_type QuicIntervalSet<T>::SpanningInterval()
const {
value_type result;
if (!intervals_.empty()) {
result.SetMin(intervals_.begin()->min());
result.SetMax(intervals_.rbegin()->max());
}
return result;
}
template <typename T>
void QuicIntervalSet<T>::Add(const value_type& interval) {
if (interval.Empty())
return;
const_iterator it = intervals_.lower_bound(interval.min());
value_type the_union = interval;
if (it != intervals_.begin()) {
--it;
if (it->Separated(the_union)) {
++it;
}
}
// Don't erase the elements one at a time, since that will produce quadratic
// work on a flat_set, and apparently an extra log-factor of work for a
// tree-based set. Instead identify the first and last intervals that need to
// be erased, and call erase only once.
const_iterator start = it;
while (it != intervals_.end() && !it->Separated(the_union)) {
the_union.SpanningUnion(*it);
++it;
}
intervals_.erase(start, it);
intervals_.insert(the_union);
}
template <typename T>
bool QuicIntervalSet<T>::Contains(const T& value) const {
// Find the first interval with min() > value, then move back one step
const_iterator it = intervals_.upper_bound(value);
if (it == intervals_.begin())
return false;
--it;
return it->Contains(value);
}
template <typename T>
bool QuicIntervalSet<T>::Contains(const value_type& interval) const {
// Find the first interval with min() > value, then move back one step.
const_iterator it = intervals_.upper_bound(interval.min());
if (it == intervals_.begin())
return false;
--it;
return it->Contains(interval);
}
template <typename T>
bool QuicIntervalSet<T>::Contains(const QuicIntervalSet<T>& other) const {
if (!SpanningInterval().Contains(other.SpanningInterval())) {
return false;
}
for (const_iterator i = other.begin(); i != other.end(); ++i) {
// If we don't contain the interval, can return false now.
if (!Contains(*i)) {
return false;
}
}
return true;
}
// This method finds the interval that Contains() "value", if such an interval
// exists in the QuicIntervalSet. The way this is done is to locate the
// "candidate interval", the only interval that could *possibly* contain value,
// and test it using Contains(). The candidate interval is the interval with the
// largest min() having min() <= value.
//
// Another detail involves the choice of which Set method to use to try to find
// the candidate interval. The most appropriate entry point is
// Set::upper_bound(), which finds the least interval with a min > the
// value. The semantics of upper_bound() are slightly different from what we
// want (namely, to find the greatest interval which is <= the probe interval)
// but they are close enough; the interval found by upper_bound() will always be
// one step past the interval we are looking for (if it exists) or at begin()
// (if it does not). Getting to the proper interval is a simple matter of
// decrementing the iterator.
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::Find(
const T& value) const {
const_iterator it = intervals_.upper_bound(value);
if (it == intervals_.begin())
return intervals_.end();
--it;
if (it->Contains(value))
return it;
else
return intervals_.end();
}
// This method finds the interval that Contains() the interval "probe", if such
// an interval exists in the QuicIntervalSet. The way this is done is to locate
// the "candidate interval", the only interval that could *possibly* contain
// "probe", and test it using Contains(). We use the same algorithm as for
// Find(value), except that instead of checking that the value is contained, we
// check that the probe is contained.
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::Find(
const value_type& probe) const {
const_iterator it = intervals_.upper_bound(probe.min());
if (it == intervals_.begin())
return intervals_.end();
--it;
if (it->Contains(probe))
return it;
else
return intervals_.end();
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::LowerBound(
const T& value) const {
const_iterator it = intervals_.lower_bound(value);
if (it == intervals_.begin()) {
return it;
}
// The previous intervals_.lower_bound() checking is essentially based on
// interval.min(), so we need to check whether the `value` is contained in
// the previous interval.
--it;
if (it->Contains(value)) {
return it;
} else {
return ++it;
}
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator QuicIntervalSet<T>::UpperBound(
const T& value) const {
return intervals_.upper_bound(value);
}
template <typename T>
bool QuicIntervalSet<T>::IsDisjoint(const value_type& interval) const {
if (interval.Empty())
return true;
// Find the first interval with min() > interval.min()
const_iterator it = intervals_.upper_bound(interval.min());
if (it != intervals_.end() && interval.max() > it->min())
return false;
if (it == intervals_.begin())
return true;
--it;
return it->max() <= interval.min();
}
template <typename T>
void QuicIntervalSet<T>::Union(const QuicIntervalSet& other) {
for (const value_type& interval : other.intervals_) {
Add(interval);
}
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator
QuicIntervalSet<T>::FindIntersectionCandidate(
const QuicIntervalSet& other) const {
return FindIntersectionCandidate(*other.intervals_.begin());
}
template <typename T>
typename QuicIntervalSet<T>::const_iterator
QuicIntervalSet<T>::FindIntersectionCandidate(
const value_type& interval) const {
// Use upper_bound to efficiently find the first interval in intervals_
// where min() is greater than interval.min(). If the result
// isn't the beginning of intervals_ then move backwards one interval since
// the interval before it is the first candidate where max() may be
// greater than interval.min().
// In other words, no interval before that can possibly intersect with any
// of other.intervals_.
const_iterator mine = intervals_.upper_bound(interval.min());
if (mine != intervals_.begin()) {
--mine;
}
return mine;
}
template <typename T>
template <typename X, typename Func>
bool QuicIntervalSet<T>::FindNextIntersectingPairImpl(X* x,
const QuicIntervalSet& y,
const_iterator* mine,
const_iterator* theirs,
Func on_hole) {
CHECK(x != nullptr);
if ((*mine == x->intervals_.end()) || (*theirs == y.intervals_.end())) {
return false;
}
while (!(**mine).Intersects(**theirs)) {
const_iterator erase_first = *mine;
// Skip over intervals in 'mine' that don't reach 'theirs'.
while (*mine != x->intervals_.end() && (**mine).max() <= (**theirs).min()) {
++(*mine);
}
*mine = on_hole(x, erase_first, *mine);
// We're done if the end of intervals_ is reached.
if (*mine == x->intervals_.end()) {
return false;
}
// Skip over intervals 'theirs' that don't reach 'mine'.
while (*theirs != y.intervals_.end() &&
(**theirs).max() <= (**mine).min()) {
++(*theirs);
}
// If the end of other.intervals_ is reached, we're done.
if (*theirs == y.intervals_.end()) {
on_hole(x, *mine, x->intervals_.end());
return false;
}
}
return true;
}
template <typename T>
void QuicIntervalSet<T>::Intersection(const QuicIntervalSet& other) {
if (!SpanningInterval().Intersects(other.SpanningInterval())) {
intervals_.clear();
return;
}
const_iterator mine = FindIntersectionCandidate(other);
// Remove any intervals that cannot possibly intersect with other.intervals_.
mine = intervals_.erase(intervals_.begin(), mine);
const_iterator theirs = other.FindIntersectionCandidate(*this);
while (FindNextIntersectingPairAndEraseHoles(other, &mine, &theirs)) {
// OK, *mine and *theirs intersect. Now, we find the largest
// span of intervals in other (starting at theirs) - say [a..b]
// - that intersect *mine, and we replace *mine with (*mine
// intersect x) for all x in [a..b] Note that subsequent
// intervals in this can't intersect any intervals in [a..b) --
// they may only intersect b or subsequent intervals in other.
value_type i(*mine);
intervals_.erase(mine);
mine = intervals_.end();
value_type intersection;
while (theirs != other.intervals_.end() &&
i.Intersects(*theirs, &intersection)) {
std::pair<const_iterator, bool> ins = intervals_.insert(intersection);
DCHECK(ins.second);
mine = ins.first;
++theirs;
}
DCHECK(mine != intervals_.end());
--theirs;
++mine;
}
DCHECK(Valid());
}
template <typename T>
bool QuicIntervalSet<T>::Intersects(const QuicIntervalSet& other) const {
// Don't bother to handle nonoverlapping spanning intervals as a special case.
// This code runs in time O(n+m), as guaranteed, even for that case .
// Handling the nonoverlapping spanning intervals as a special case doesn't
// improve the asymptotics but does make the code more complex.
auto mine = intervals_.begin();
auto theirs = other.intervals_.begin();
while (mine != intervals_.end() && theirs != other.intervals_.end()) {
if (mine->Intersects(*theirs))
return true;
else if (*mine < *theirs)
++mine;
else
++theirs;
}
return false;
}
template <typename T>
void QuicIntervalSet<T>::Difference(const value_type& interval) {
if (!SpanningInterval().Intersects(interval)) {
return;
}
Difference(QuicIntervalSet<T>(interval));
}
template <typename T>
void QuicIntervalSet<T>::Difference(const T& min, const T& max) {
Difference(value_type(min, max));
}
template <typename T>
void QuicIntervalSet<T>::Difference(const QuicIntervalSet& other) {
// In order to avoid quadratic-time when using a flat set, we don't try to
// update intervals_ in place. Instead we build up a new result_, always
// inserting at the end which is O(1) time per insertion. Since the number of
// elements in the result is O(Size() + other.Size()), the cost for all the
// insertions is also O(Size() + other.Size()).
//
// We look at all the elements of intervals_, so that's O(Size()).
//
// We also look at all the elements of other.intervals_, for O(other.Size()).
if (Empty())
return;
Set result;
const_iterator mine = intervals_.begin();
value_type myinterval = *mine;
const_iterator theirs = other.intervals_.begin();
while (mine != intervals_.end()) {
// Loop invariants:
// myinterval is nonempty.
// mine points at a range that is a suffix of myinterval.
DCHECK(!myinterval.Empty());
DCHECK(myinterval.max() == mine->max());
// There are 3 cases.
// myinterval is completely before theirs (treat theirs==end() as if it is
// infinity).
// --> consume myinterval into result.
// myinterval is completely after theirs
// --> theirs can no longer affect us, so ++theirs.
// myinterval touches theirs with a prefix of myinterval not touching
// *theirs.
// --> consume the prefix of myinterval into the result.
// myinterval touches theirs, with the first element of myinterval in
// *theirs.
// -> reduce myinterval
if (theirs == other.intervals_.end() || myinterval.max() <= theirs->min()) {
// Keep all of my_interval.
result.insert(result.end(), myinterval);
myinterval.Clear();
} else if (theirs->max() <= myinterval.min()) {
++theirs;
} else if (myinterval.min() < theirs->min()) {
// Keep a nonempty prefix of my interval.
result.insert(result.end(), value_type(myinterval.min(), theirs->min()));
myinterval.SetMin(theirs->max());
} else {
// myinterval starts at or after *theirs, chop down myinterval.
myinterval.SetMin(theirs->max());
}
// if myinterval became empty, find the next interval
if (myinterval.Empty()) {
++mine;
if (mine != intervals_.end()) {
myinterval = *mine;
}
}
}
std::swap(result, intervals_);
DCHECK(Valid());
}
template <typename T>
void QuicIntervalSet<T>::Complement(const T& min, const T& max) {
QuicIntervalSet<T> span(min, max);
span.Difference(*this);
intervals_.swap(span.intervals_);
}
template <typename T>
std::string QuicIntervalSet<T>::ToString() const {
std::ostringstream os;
os << *this;
return os.str();
}
template <typename T>
bool QuicIntervalSet<T>::Valid() const {
const_iterator prev = end();
for (const_iterator it = begin(); it != end(); ++it) {
// invalid or empty interval.
if (it->min() >= it->max())
return false;
// Not sorted, not disjoint, or adjacent.
if (prev != end() && prev->max() >= it->min())
return false;
prev = it;
}
return true;
}
// This comparator orders intervals first by ascending min(). The Set never
// contains overlapping intervals, so that suffices.
template <typename T>
bool QuicIntervalSet<T>::IntervalLess::operator()(const value_type& a,
const value_type& b) const {
// This overload is probably used only by Set::insert().
return a.min() < b.min();
}
// It appears that the Set::lower_bound(T) method uses only two overloads of the
// comparison operator that take a T as the second argument.. In contrast
// Set::upper_bound(T) uses the two overloads that take T as the first argument.
template <typename T>
bool QuicIntervalSet<T>::IntervalLess::operator()(const value_type& a,
const T& point) const {
// Compare an interval to a point.
return a.min() < point;
}
template <typename T>
bool QuicIntervalSet<T>::IntervalLess::operator()(const value_type& a,
T&& point) const {
// Compare an interval to a point
return a.min() < point;
}
// It appears that the Set::upper_bound(T) method uses only the next two
// overloads of the comparison operator.
template <typename T>
bool QuicIntervalSet<T>::IntervalLess::operator()(const T& point,
const value_type& a) const {
// Compare an interval to a point.
return point < a.min();
}
template <typename T>
bool QuicIntervalSet<T>::IntervalLess::operator()(T&& point,
const value_type& a) const {
// Compare an interval to a point.
return point < a.min();
}
} // namespace newquic
// TODO(bradleybear): Get rid of this when we get rid of
// quic_startup_faster_interval_set_flag.
class QUIC_NO_EXPORT QuicIntervalSetParameterSetter {
private:
// friend the tests so that we can SetUseFasterIntervalSet in the tests.
template <typename T>
friend class QuicIntervalSet;
friend void Quic_Test_Set_Fast(bool fast);
static bool* UseFasterIntervalSetAddress() {
static bool result = GetQuicRestartFlag(quic_startup_faster_interval_set);
return &result;
}
static void SetUseFasterIntervalSet(bool use) {
*UseFasterIntervalSetAddress() = use;
}
static bool UseFasterIntervalSet() { return *UseFasterIntervalSetAddress(); }
};
template <typename T>
class QUIC_NO_EXPORT QuicIntervalSet {
template <class old_iterator, class new_iterator>
class InternalIterator;
using OldSet = oldquic::QuicIntervalSet<T>;
using NewSet = newquic::QuicIntervalSet<T>;
using QISet = absl::variant<OldSet, NewSet>;
public:
typedef QuicInterval<T> value_type;
using const_iterator = InternalIterator<typename OldSet::const_iterator,
typename NewSet::const_iterator>;
using const_reverse_iterator =
InternalIterator<typename OldSet::const_reverse_iterator,
typename NewSet::const_reverse_iterator>;
QuicIntervalSet()
: qiset_(QuicIntervalSetParameterSetter::UseFasterIntervalSet()
? QISet(NewSet())
: QISet(OldSet())) {}
explicit QuicIntervalSet(const value_type& interval)
: qiset_(QuicIntervalSetParameterSetter::UseFasterIntervalSet()
? QISet(NewSet(interval))
: QISet(OldSet(interval))) {}
QuicIntervalSet(const T& min, const T& max)
: qiset_(QuicIntervalSetParameterSetter::UseFasterIntervalSet()
? QISet(NewSet(min, max))
: QISet(OldSet(min, max))) {}
QuicIntervalSet(std::initializer_list<value_type> il)
: qiset_(QuicIntervalSetParameterSetter::UseFasterIntervalSet()
? QISet(NewSet(il))
: QISet(OldSet(il))) {}
void Clear() {
return absl::visit([](auto& s) { s.Clear(); }, qiset_);
}
size_t Size() const {
return absl::visit([](const auto& s) { return s.Size(); }, qiset_);
}
value_type SpanningInterval() const {
return absl::visit([](auto& s) { return s.SpanningInterval(); }, qiset_);
}
void Add(const value_type& interval) {
return absl::visit([&](auto& s) { return s.Add(interval); }, qiset_);
}
void Add(const T& min, const T& max) {
return absl::visit([&](auto& s) { return s.Add(min, max); }, qiset_);
}
void AddOptimizedForAppend(const value_type& interval) {
return absl::visit(
[&](auto& s) { return s.AddOptimizedForAppend(interval); }, qiset_);
}
void AddOptimizedForAppend(const T& min, const T& max) {
return absl::visit(
[&](auto& s) { return s.AddOptimizedForAppend(min, max); }, qiset_);
}
void PopFront() {
return absl::visit([](auto& s) { return s.PopFront(); }, qiset_);
}
bool Empty() const {
return absl::visit([](auto& s) { return s.Empty(); }, qiset_);
}
bool Contains(const T& value) const {
return absl::visit([&](auto& s) { return s.Contains(value); }, qiset_);
}
bool Contains(const value_type& interval) const {
return absl::visit([&](auto& s) { return s.Contains(interval); }, qiset_);
}
bool Contains(const T& min, const T& max) const {
return absl::visit([&](auto& s) { return s.Contains(min, max); }, qiset_);
}
private:
template <class A, class B, class C>
struct overloader : A, B, C {
overloader(A a, B b, C c) : A(a), B(b), C(c) {}
using A::operator();
using B::operator();
using C::operator();
};
template <class A, class B, class C>
static auto make_visitor(A a, B b, C c) {
return overloader<A, B, C>(a, b, c);
}
public:
bool Contains(const QuicIntervalSet<T>& other) const {
return absl::visit(
make_visitor(
// Sadly C++11's templated lambda system isn't quite powerful enough
// to specify that the two auto&& arguments are the same, so we have
// to specify two functions.
[](const OldSet& a, const OldSet& b) { return a.Contains(b); },
[](const NewSet& a, const NewSet& b) { return a.Contains(b); },
// If the types mismatch then return false. This shouldn't happen
// in production since we capture the
// quic_startup_faster_interval_set flag very early. But tests can
// make it happen.
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return false;
}),
qiset_, other.qiset_);
}
bool Intersects(const QuicIntervalSet& other) const {
return absl::visit(
make_visitor(
[](const OldSet& a, const OldSet& b) { return a.Intersects(b); },
[](const NewSet& a, const NewSet& b) { return a.Intersects(b); },
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return false;
}),
qiset_, other.qiset_);
}
const_iterator Find(const T& value) const {
return absl::visit([&](auto& s) { return const_iterator(s.Find(value)); },
qiset_);
}
const_iterator Find(const T& min, const T& max) const {
return absl::visit(
[&](auto& s) { return const_iterator(s.Find(min, max)); }, qiset_);
}
const_iterator LowerBound(const T& value) const {
return absl::visit(
[&](auto& s) { return const_iterator(s.LowerBound(value)); }, qiset_);
}
const_iterator UpperBound(const T& value) const {
return absl::visit(
[&](auto& s) { return const_iterator(s.UpperBound(value)); }, qiset_);
}
bool IsDisjoint(const value_type& interval) const {
return absl::visit([&](auto& s) { return s.IsDisjoint(interval); }, qiset_);
}
void Union(const QuicIntervalSet& other) {
return absl::visit(
make_visitor([](OldSet& a, const OldSet& b) { return a.Union(b); },
[](NewSet& a, const NewSet& b) { return a.Union(b); },
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return void();
}),
qiset_, other.qiset_);
}
void Intersection(const QuicIntervalSet& other) {
return absl::visit(
make_visitor(
[](OldSet& a, const OldSet& b) { return a.Intersection(b); },
[](NewSet& a, const NewSet& b) { return a.Intersection(b); },
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return void();
}),
qiset_, other.qiset_);
}
void Difference(const value_type& interval) {
return absl::visit([&](auto& s) { return s.Difference(interval); }, qiset_);
}
void Difference(const T& min, const T& max) {
return absl::visit([&](auto& s) { return s.Difference(min, max); }, qiset_);
}
void Difference(const QuicIntervalSet& other) {
return absl::visit(
make_visitor([](OldSet& a, const OldSet& b) { return a.Difference(b); },
[](NewSet& a, const NewSet& b) { return a.Difference(b); },
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return void();
}),
qiset_, other.qiset_);
}
void Complement(const T& min, const T& max) {
return absl::visit([&](auto& s) { return s.Complement(min, max); }, qiset_);
}
bool TrimLessThan(const T& value) {
return absl::visit([&](auto& s) { return s.TrimLessThan(value); }, qiset_);
}
const_iterator begin() const {
return absl::visit([](auto& s) { return const_iterator(s.begin()); },
qiset_);
}
const_iterator end() const {
return absl::visit([](auto& s) { return const_iterator(s.end()); }, qiset_);
}
const_reverse_iterator rbegin() const {
return absl::visit(
[](auto& s) { return const_reverse_iterator(s.rbegin()); }, qiset_);
}
const_reverse_iterator rend() const {
return absl::visit([](auto& s) { return const_reverse_iterator(s.rend()); },
qiset_);
}
template <typename Iter>
void assign(Iter first, Iter last) {
return absl::visit([&](auto& s) { return s.assign(first, last); }, qiset_);
}
void assign(std::initializer_list<value_type> il) {
return absl::visit([&](auto& s) { return s.assign(il); }, qiset_);
}
std::string ToString() const {
return absl::visit([](auto& s) { return s.ToString(); }, qiset_);
}
friend bool operator==(const QuicIntervalSet& a, const QuicIntervalSet& b) {
return absl::visit(
make_visitor([](const OldSet& a, const OldSet& b) { return a == b; },
[](const NewSet& a, const NewSet& b) { return a == b; },
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return false;
}),
a.qiset_, b.qiset_);
}
friend bool operator!=(const QuicIntervalSet& a, const QuicIntervalSet& b) {
return absl::visit(
make_visitor([](const OldSet& a, const OldSet& b) { return a != b; },
[](const NewSet& a, const NewSet& b) { return a != b; },
[]([[maybe_unused]] auto&& a, [[maybe_unused]] auto&& b) {
return true;
}),
a.qiset_, b.qiset_);
}
private:
template <class old_iterator, class new_iterator>
// This same idea appeared in cl/338686797 as IteratorWrapper. Perhaps we
// should implement it once and for all.
class InternalIterator
: public std::iterator<std::forward_iterator_tag, // iterator category
QuicIntervalSet::value_type, // value_type
ptrdiff_t, // difference_type
const value_type*, // pointer
const value_type& // reference
> {
public:
const value_type* operator->() const {
return absl::visit([](auto& it) { return &*it; }, it_);
}
const value_type& operator*() const {
// Must return &*it (rather than just returning *it) since, for some
// iterators (especially const_iterators) return a value_type instead of a
// value_type&. The standard says that it must return a "reference,
// convertible to T" but doesn't actually say that the "reference" must be
// a reference type. The standard isn't very clear, but the simpler code
// doesn't work.
return *absl::visit([](auto& it) { return &*it; }, it_);
}
InternalIterator& operator++() {
absl::visit([](auto& it) { ++it; }, it_);
return *this;
}
InternalIterator& operator--() {
absl::visit([](auto& it) { --it; }, it_);
return *this;
}
private:
friend QuicIntervalSet;
// The following doesn't work for some compiler combinations:
// explicit InternalIterator(old_iterator it) : it_(std::move(it)) {}
// explicit InternalIterator(new_iterator it) : it_(std::move(it)) {}
// So we are using emplace instead.
explicit InternalIterator(old_iterator it) {
it_.template emplace<0>(std::move(it));
}
explicit InternalIterator(new_iterator it) {
it_.template emplace<1>(std::move(it));
}
friend bool operator==(const InternalIterator& a,
const InternalIterator& b) {
return absl::visit(
make_visitor([](const old_iterator& a,
const old_iterator& b) { return a == b; },
[](const new_iterator& a, const new_iterator& b) {
return a == b;
},
[]([[maybe_unused]] auto&& a,
[[maybe_unused]] auto&& b) { return false; }),
a.it_, b.it_);
}
friend bool operator!=(const InternalIterator& a,
const InternalIterator& b) {
return absl::visit(
make_visitor([](const old_iterator& a,
const old_iterator& b) { return a != b; },
[](const new_iterator& a, const new_iterator& b) {
return a != b;
},
[]([[maybe_unused]] auto&& a,
[[maybe_unused]] auto&& b) { return true; }),
a.it_, b.it_);
}
absl::variant<old_iterator, new_iterator> it_;
};
QISet qiset_;
};
template <typename T>
auto operator<<(std::ostream& out, const QuicIntervalSet<T>& seq)
-> decltype(out << *seq.begin()) {
out << "{";
for (const auto& interval : seq) {
out << " " << interval;
}
out << " }";
return out;
}
} // namespace quic
#endif // QUICHE_QUIC_CORE_QUIC_INTERVAL_SET_H_