Optimize `is_permutation` for reversed sequences (#2043)

Co-authored-by: Stephan T. Lavavej <stl@nuwen.net>
Co-authored-by: Miya Natsuhara <46756417+mnatsuhara@users.noreply.github.com>

stl/inc/algorithm

@@ -985,8 +985,6 @@ namespace ranges {
                     --_Final1;
                     --_Final2;
                     if (!_STD invoke(_Pred, _STD invoke(_Proj1, *_Final1), _STD invoke(_Proj2, *_Final2))) { // mismatch
-                        ++_Final1;
-                        ++_Final2;
                         break;
                     }
 
@@ -996,6 +994,32 @@ namespace ranges {
                 }
                 // If we get here, _Count > 1, initial elements do not match, and final elements do not match.
 
+                // We've trimmed matching prefixes and matching suffixes.
+                // Now we need to compare each range's prefix to the other range's suffix.
+
+                const auto _ProjectedPred = [&]<class _Ty1, class _Ty2>(_Ty1&& _Left, _Ty2&& _Right) {
+                    return _STD invoke(_Pred, _STD invoke(_Proj1, _STD forward<_Ty1>(_Left)),
+                        _STD invoke(_Proj2, _STD forward<_Ty2>(_Right)));
+                };
+
+                const _TrimResult _Res = _Trim_completely(_First1, _Final1, _First2, _Final2, _ProjectedPred);
+
+                if (_Res != _TrimResult::_HaveWorkAfterTrimming) {
+                    return _Res == _TrimResult::_ReturnTrue;
+                }
+
+                ++_Final1;
+                ++_Final2;
+                // If we get here, initial elements do not match, final elements do not match, and ranges have length
+                // at least 2 and at most _Count.
+
+                // We've trimmed matching prefixes, matching suffixes,
+                // and each range's prefix matching the other range's suffix. That is, given:
+                // Range 1: [A, ..., B]
+                // Range 2: [X, ..., Y]
+                // we know that A != X, A != Y, B != X, and B != Y.
+                // (A == B and X == Y are possible but irrelevant.)
+
                 return _Match_counts(_STD move(_First1), _STD move(_Final1), _STD move(_First2), _STD move(_Final2),
                     _Pred, _Proj1, _Proj2);
             } else {
@@ -1053,13 +1077,37 @@ namespace ranges {
                     --_Final2; // since ranges have equal lengths, _Final2 cannot equal _First2
 
                     if (!_STD invoke(_Pred, _STD invoke(_Proj1, *_Final1), _STD invoke(_Proj2, *_Final2))) { // mismatch
-                        ++_Final1;
-                        ++_Final2;
                         break;
                     }
                 }
                 // If we get here, initial elements do not match, final elements do not match, and ranges have length
                 // at least 2.
+
+                // We've trimmed matching prefixes and matching suffixes.
+                // Now we need to compare each range's prefix to the other range's suffix.
+
+                const auto _ProjectedPred = [&]<class _Ty1, class _Ty2>(_Ty1&& _Left, _Ty2&& _Right) {
+                    return _STD invoke(_Pred, _STD invoke(_Proj1, _STD forward<_Ty1>(_Left)),
+                        _STD invoke(_Proj2, _STD forward<_Ty2>(_Right)));
+                };
+
+                const _TrimResult _Res = _Trim_completely(_First1, _Final1, _First2, _Final2, _ProjectedPred);
+
+                if (_Res != _TrimResult::_HaveWorkAfterTrimming) {
+                    return _Res == _TrimResult::_ReturnTrue;
+                }
+
+                ++_Final1;
+                ++_Final2;
+                // If we get here, initial elements do not match, final elements do not match, and ranges have length
+                // at least 2.
+
+                // We've trimmed matching prefixes, matching suffixes,
+                // and each range's prefix matching the other range's suffix. That is, given:
+                // Range 1: [A, ..., B]
+                // Range 2: [X, ..., Y]
+                // we know that A != X, A != Y, B != X, and B != Y.
+                // (A == B and X == Y are possible but irrelevant.)
             }
 
             return _Match_counts(

stl/inc/xutility

@@ -5348,9 +5348,106 @@ _NODISCARD constexpr _Iter_diff_t<_InIt> _Count_pr(_InIt _First, const _InIt _La
     return _Count;
 }
 
+enum class _TrimResult : unsigned char { _KeepTrimming, _HaveWorkAfterTrimming, _ReturnFalse, _ReturnTrue };
+
+template <class _BidIt1, class _BidIt2, class _Pr>
+_NODISCARD _CONSTEXPR20 _TrimResult _Trim_equal(
+    _BidIt1& _First1, _BidIt1& _Back1, _BidIt2& _First2, _BidIt2& _Back2, _Pr _Pred) {
+    // advances the iterators, trimming matching prefixes then matching suffixes
+    // from [_First1, _Back1] and [_First2, _Back2]
+    _STL_INTERNAL_CHECK(_First1 != _Back1);
+    _STL_INTERNAL_CHECK(_STD distance(_First1, _Back1) == _STD distance(_First2, _Back2));
+    if (_Pred(*_First1, *_First2)) {
+        do {
+            ++_First1;
+            ++_First2;
+            if (_First1 == _Back1) {
+                // only one element is left
+                return _Pred(*_First1, *_First2) ? _TrimResult::_ReturnTrue : _TrimResult::_ReturnFalse;
+            }
+        } while (_Pred(*_First1, *_First2));
+    } else {
+        if (!_Pred(*_Back1, *_Back2)) {
+            // nothing to trim
+            return _TrimResult::_HaveWorkAfterTrimming;
+        }
+        --_Back1;
+        --_Back2;
+    }
+
+    for (;;) {
+        if (_First1 == _Back1) {
+            // only one element is left, it can't match because it wasn't trimmed by the first loop
+            return _TrimResult::_ReturnFalse;
+        }
+
+        if (!_Pred(*_Back1, *_Back2)) {
+            return _TrimResult::_KeepTrimming;
+        }
+        --_Back1;
+        --_Back2;
+    }
+}
+
+template <class _BidIt1, class _BidIt2, class _Pr>
+_NODISCARD _CONSTEXPR20 _TrimResult _Trim_reversed(
+    _BidIt1& _First1, _BidIt1& _Back1, _BidIt2& _First2, _BidIt2& _Back2, _Pr _Pred) {
+    // advances the iterators, trimming each range's prefix that matches the other range's suffix
+    // from [_First1, _Back1] and [_First2, _Back2]
+    _STL_INTERNAL_CHECK(_First1 != _Back1);
+    _STL_INTERNAL_CHECK(_STD distance(_First1, _Back1) == _STD distance(_First2, _Back2));
+    if (_Pred(*_First1, *_Back2)) {
+        do {
+            ++_First1;
+            --_Back2;
+            if (_First1 == _Back1) {
+                // only one element is left
+                return _Pred(*_First1, *_First2) ? _TrimResult::_ReturnTrue : _TrimResult::_ReturnFalse;
+            }
+        } while (_Pred(*_First1, *_Back2));
+    } else {
+        if (!_Pred(*_Back1, *_First2)) {
+            // nothing to trim
+            return _TrimResult::_HaveWorkAfterTrimming;
+        }
+        --_Back1;
+        ++_First2;
+    }
+
+    for (;;) {
+        if (_First1 == _Back1) {
+            // only one element is left, it can't match because it wasn't trimmed by the first loop
+            return _TrimResult::_ReturnFalse;
+        }
+
+        if (!_Pred(*_Back1, *_First2)) {
+            return _TrimResult::_KeepTrimming;
+        }
+        --_Back1;
+        ++_First2;
+    }
+}
+
+template <class _BidIt1, class _BidIt2, class _Pr>
+_NODISCARD _CONSTEXPR20 _TrimResult _Trim_completely(
+    _BidIt1& _First1, _BidIt1& _Back1, _BidIt2& _First2, _BidIt2& _Back2, _Pr _Pred) {
+    // alternates between calling _Trim_reversed and _Trim_equal until no more trimming is possible
+    _TrimResult _Res = _TrimResult::_KeepTrimming;
+
+    for (bool _Check_reversed = true; _Res == _TrimResult::_KeepTrimming; _Check_reversed = !_Check_reversed) {
+        if (_Check_reversed) {
+            _Res = _Trim_reversed(_First1, _Back1, _First2, _Back2, _Pred);
+        } else {
+            _Res = _Trim_equal(_First1, _Back1, _First2, _Back2, _Pred);
+        }
+    }
+
+    return _Res;
+}
+
 template <class _FwdIt1, class _FwdIt2, class _Pr>
 _NODISCARD _CONSTEXPR20 bool _Check_match_counts(
-    const _FwdIt1 _First1, _FwdIt1 _Last1, const _FwdIt2 _First2, _FwdIt2 _Last2, _Pr _Pred) {
+    _FwdIt1 _First1, _FwdIt1 _Last1, _FwdIt2 _First2, _FwdIt2 _Last2, _Pr _Pred) {
     // test if [_First1, _Last1) == permuted [_First2, _Last2), after matching prefix removal
     _STL_INTERNAL_CHECK(!_Pred(*_First1, *_First2));
     _STL_INTERNAL_CHECK(_STD distance(_First1, _Last1) == _STD distance(_First2, _Last2));
@@ -5359,6 +5456,17 @@ _NODISCARD _CONSTEXPR20 bool _Check_match_counts(
             --_Last1;
             --_Last2;
         } while (_Pred(*_Last1, *_Last2));
+
+        if (_First1 == _Last1) {
+            return false;
+        }
+
+        const _TrimResult _Res = _Trim_completely(_First1, _Last1, _First2, _Last2, _Pred);
+
+        if (_Res != _TrimResult::_HaveWorkAfterTrimming) {
+            return _Res == _TrimResult::_ReturnTrue;
+        }
+
         ++_Last1;
         ++_Last2;
     }

tests/std/tests/Dev11_0000000_dual_range_algorithms/test.cpp

@@ -331,6 +331,32 @@ int main() {
         assert(!is_permutation(a.begin(), a.end(), b.begin(), b.end(), equal_to<int>()));
     }
 
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 8, 7, 6, 5, 4, 3, 2, 1, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 8, 7, 3, 4, 5, 6, 2, 1, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 1, 7, 3, 5, 4, 6, 2, 8, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 1, 7, 5, 6, 3, 4, 2, 8, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 10, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 1, 7, 3, 5, 11, 4, 6, 2, 8, 0};
+        assert(!is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+
     { // Test that _ITERATOR_DEBUG_ARRAY_OVERLOADS is not needed anymore
         int arr[8] = {};
         assert(mismatch(arr, arr, arr, arr) == make_pair(begin(arr), begin(arr)));

tests/std/tests/P0202R3_constexpr_algorithm_and_exchange/test.cpp

@@ -500,6 +500,10 @@ constexpr void test_permutations() {
         {40, 30, 20, 10},
     };
 
+    for (size_t i = 0; i < size(expected); ++i) {
+        assert(is_permutation(begin(buff), end(buff), begin(expected[i]), end(expected[i])));
+    }
+
     size_t cursor = 0;
     do {
         assert(equal(begin(buff), end(buff), begin(expected[cursor]), end(expected[cursor])));
@@ -518,6 +522,32 @@ constexpr void test_permutations() {
 
     assert(cursor == 0);
     assert(equal(begin(buff), end(buff), begin(expected[23]), end(expected[23])));
+
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 8, 7, 6, 5, 4, 3, 2, 1, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 8, 7, 3, 4, 5, 6, 2, 1, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 1, 7, 3, 5, 4, 6, 2, 8, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 1, 7, 5, 6, 3, 4, 2, 8, 0};
+        assert(is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
+    {
+        int arr1[] = {0, 1, 2, 3, 4, 10, 5, 6, 7, 8, 9};
+        int arr2[] = {9, 1, 7, 3, 5, 11, 4, 6, 2, 8, 0};
+        assert(!is_permutation(begin(arr1), end(arr1), begin(arr2), end(arr2)));
+    }
 }
 
 constexpr bool test() {

tests/std/tests/P0896R4_ranges_alg_is_permutation/test.cpp

@@ -88,6 +88,52 @@ struct instantiator {
             assert(!ranges::is_permutation(ranges::begin(r1), ranges::end(r1), ranges::begin(r2), ranges::end(r2),
                 ranges::equal_to{}, get_first, identity{}));
         }
+        {
+            int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+            int arr2[] = {9, 8, 7, 6, 5, 4, 3, 2, 1, 0};
+            const Fwd2 r1{arr1};
+            const Fwd2 r2{arr2};
+            assert(ranges::is_permutation(r1, r2));
+        }
+        {
+            int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+            int arr2[] = {9, 8, 7, 3, 4, 5, 6, 2, 1, 0};
+            const Fwd2 r1{arr1};
+            const Fwd2 r2{arr2};
+            assert(ranges::is_permutation(r1, r2));
+        }
+        {
+            int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+            int arr2[] = {9, 1, 7, 3, 5, 4, 6, 2, 8, 0};
+            const Fwd2 r1{arr1};
+            const Fwd2 r2{arr2};
+            assert(ranges::is_permutation(r1, r2));
+        }
+        {
+            int arr1[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+            int arr2[] = {9, 1, 7, 5, 6, 3, 4, 2, 8, 0};
+            const Fwd2 r1{arr1};
+            const Fwd2 r2{arr2};
+            assert(ranges::is_permutation(r1, r2));
+        }
+        {
+            int arr1[] = {0, 1, 2, 3, 4, 10, 5, 6, 7, 8, 9};
+            int arr2[] = {9, 1, 7, 3, 5, 11, 4, 6, 2, 8, 0};
+            const Fwd2 r1{arr1};
+            const Fwd2 r2{arr2};
+            assert(!ranges::is_permutation(r1, r2));
+        }
+        {
+            int arr1[] = {-1, 0, 1, 2, 3, 4, 5, 6, 7, 8};
+            int arr2[] = {10, 9, 8, 7, 6, 5, 4, 3, 2, 1};
+            const Fwd2 r1{arr1};
+            const Fwd2 r2{arr2};
+            assert(!ranges::is_permutation(r1, r2));
+            assert(!ranges::is_permutation(r1, r2, {}, {}, [](int n) { return n - 1; }));
+            assert(ranges::is_permutation(r1, r2, {}, {}, [](int n) { return n - 2; }));
+            assert(ranges::is_permutation(
+                r1, r2, {}, [](int n) { return n + 1; }, [](int n) { return n - 1; }));
+        }
     }
 };