diff options
author | Abseil Team <absl-team@google.com> | 2020-03-03T19·22-0800 |
---|---|---|
committer | Andy Soffer <asoffer@google.com> | 2020-03-03T22·32-0500 |
commit | b19ba96766db08b1f32605cb4424a0e7ea0c7584 (patch) | |
tree | c4ba295b067b000b9d84410ec81e0095715641a5 /absl/strings | |
parent | 06f0e767d13d4d68071c4fc51e25724e0fc8bc74 (diff) |
Export of internal Abseil changes
-- a3e58c1870a9626039f4d178d2d599319bd9f8a8 by Matt Kulukundis <kfm@google.com>: Allow MakeCordFromExternal to take a zero arg releaser. PiperOrigin-RevId: 298650274 -- 01897c4a9bb99f3dc329a794019498ad345ddebd by Samuel Benzaquen <sbenza@google.com>: Reduce library bloat for absl::Flag by moving the definition of base virtual functions to a .cc file. This removes the duplicate symbols in user translation units and has the side effect of moving the vtable definition too (re key function) PiperOrigin-RevId: 298617920 -- 190f0d3782c63aed01046886d7fbc1be5bca2de9 by Derek Mauro <dmauro@google.com>: Import GitHub #596: Unbreak stacktrace code for UWP apps PiperOrigin-RevId: 298600834 -- cd5cf6f8c87b35b85a9584e94da2a99057345b73 by Gennadiy Rozental <rogeeff@google.com>: Use union of heap allocated pointer, one word atomic and two word atomic to represent flags value. Any type T, which is trivially copy-able and with with sizeof(T) <= 8, will be stored in atomic int64_t. Any type T, which is trivially copy-able and with with 8 < sizeof(T) <= 16, will be stored in atomic AlignedTwoWords. We also introducing value storage type to distinguish these cases. PiperOrigin-RevId: 298497200 -- f8fe7bd53bfed601f002f521e34ab4bc083fc28b by Matthew Brown <matthewbr@google.com>: Ensure a deep copy and proper equality on absl::Status::ErasePayload PiperOrigin-RevId: 298482742 -- a5c9ccddf4b04f444e3f7e27dbc14faf1fcb5373 by Gennadiy Rozental <rogeeff@google.com>: Change ChunkIterator implementation to use fixed capacity collection of CordRep*. We can now assume that depth never exceeds 91. That makes comparison operator exception safe. I've tested that with this CL we do not observe an overhead of chunk_end. Compiler optimized this iterator completely. PiperOrigin-RevId: 298458472 -- 327ea5e8910bc388b03389c730763f9823abfce5 by Abseil Team <absl-team@google.com>: Minor cleanups in b-tree code: - Rename some variables: fix issues of different param names between definition/declaration, move away from `x` as a default meaningless variable name. - Make init_leaf/init_internal be non-static methods (they already take the node as the first parameter). - In internal_emplace/try_shrink, update root/rightmost the same way as in insert_unique/insert_multi. - Replace a TODO with a comment. PiperOrigin-RevId: 298432836 -- 8020ce9ec8558ee712d9733ae3d660ac1d3ffe1a by Abseil Team <absl-team@google.com>: Guard against unnecessary copy in case the buffer is empty. This is important in cases were the user is explicitly tuning their chunks to match PiecewiseChunkSize(). PiperOrigin-RevId: 298366044 -- 89324441d1c0c697c90ba7d8fc63639805fcaa9d by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 298219363 GitOrigin-RevId: a3e58c1870a9626039f4d178d2d599319bd9f8a8 Change-Id: I28dffc684b6fd0292b94807b88ec6664d5d0e183
Diffstat (limited to 'absl/strings')
-rw-r--r-- | absl/strings/cord.cc | 242 | ||||
-rw-r--r-- | absl/strings/cord.h | 100 | ||||
-rw-r--r-- | absl/strings/cord_test.cc | 56 |
3 files changed, 265 insertions, 133 deletions
diff --git a/absl/strings/cord.cc b/absl/strings/cord.cc index 5cc68539941d..9b32b3cc46b6 100644 --- a/absl/strings/cord.cc +++ b/absl/strings/cord.cc @@ -30,7 +30,6 @@ #include "absl/base/internal/raw_logging.h" #include "absl/base/port.h" #include "absl/container/fixed_array.h" -#include "absl/container/inlined_vector.h" #include "absl/strings/escaping.h" #include "absl/strings/internal/cord_internal.h" #include "absl/strings/internal/resize_uninitialized.h" @@ -132,6 +131,14 @@ inline const CordRepExternal* CordRep::external() const { return static_cast<const CordRepExternal*>(this); } +using CordTreeConstPath = CordTreePath<const CordRep*, MaxCordDepth()>; + +// This type is used to store the list of pending nodes during re-balancing. +// Its maximum size is 2 * MaxCordDepth() because the tree has a maximum +// possible depth of MaxCordDepth() and every concat node along a tree path +// could theoretically be split during rebalancing. +using RebalancingStack = CordTreePath<CordRep*, 2 * MaxCordDepth()>; + } // namespace cord_internal static const size_t kFlatOverhead = offsetof(CordRep, data); @@ -180,8 +187,8 @@ static constexpr size_t TagToLength(uint8_t tag) { // Enforce that kMaxFlatSize maps to a well-known exact tag value. static_assert(TagToAllocatedSize(224) == kMaxFlatSize, "Bad tag logic"); -constexpr uint64_t Fibonacci(unsigned char n, uint64_t a = 0, uint64_t b = 1) { - return n == 0 ? a : Fibonacci(n - 1, b, a + b); +constexpr uint64_t Fibonacci(uint8_t n, uint64_t a = 0, uint64_t b = 1) { + return n == 0 ? a : n == 1 ? b : Fibonacci(n - 1, b, a + b); } static_assert(Fibonacci(63) == 6557470319842, @@ -189,89 +196,68 @@ static_assert(Fibonacci(63) == 6557470319842, // Minimum length required for a given depth tree -- a tree is considered // balanced if -// length(t) >= min_length[depth(t)] -// The root node depth is allowed to become twice as large to reduce rebalancing -// for larger strings (see IsRootBalanced). -static constexpr uint64_t min_length[] = { - Fibonacci(2), - Fibonacci(3), - Fibonacci(4), - Fibonacci(5), - Fibonacci(6), - Fibonacci(7), - Fibonacci(8), - Fibonacci(9), - Fibonacci(10), - Fibonacci(11), - Fibonacci(12), - Fibonacci(13), - Fibonacci(14), - Fibonacci(15), - Fibonacci(16), - Fibonacci(17), - Fibonacci(18), - Fibonacci(19), - Fibonacci(20), - Fibonacci(21), - Fibonacci(22), - Fibonacci(23), - Fibonacci(24), - Fibonacci(25), - Fibonacci(26), - Fibonacci(27), - Fibonacci(28), - Fibonacci(29), - Fibonacci(30), - Fibonacci(31), - Fibonacci(32), - Fibonacci(33), - Fibonacci(34), - Fibonacci(35), - Fibonacci(36), - Fibonacci(37), - Fibonacci(38), - Fibonacci(39), - Fibonacci(40), - Fibonacci(41), - Fibonacci(42), - Fibonacci(43), - Fibonacci(44), - Fibonacci(45), - Fibonacci(46), - Fibonacci(47), - 0xffffffffffffffffull, // Avoid overflow -}; - -static const int kMinLengthSize = ABSL_ARRAYSIZE(min_length); - -// The inlined size to use with absl::InlinedVector. -// -// Note: The InlinedVectors in this file (and in cord.h) do not need to use -// the same value for their inlined size. The fact that they do is historical. -// It may be desirable for each to use a different inlined size optimized for -// that InlinedVector's usage. -// -// TODO(jgm): Benchmark to see if there's a more optimal value than 47 for -// the inlined vector size (47 exists for backward compatibility). -static const int kInlinedVectorSize = 47; - -static inline bool IsRootBalanced(CordRep* node) { - if (node->tag != CONCAT) { - return true; - } else if (node->concat()->depth() <= 15) { - return true; - } else if (node->concat()->depth() > kMinLengthSize) { - return false; - } else { - // Allow depth to become twice as large as implied by fibonacci rule to - // reduce rebalancing for larger strings. - return (node->length >= min_length[node->concat()->depth() / 2]); - } +// length(t) >= kMinLength[depth(t)] +// The node depth is allowed to become larger to reduce rebalancing +// for larger strings (see ShouldRebalance). +constexpr uint64_t kMinLength[] = { + Fibonacci(2), Fibonacci(3), Fibonacci(4), Fibonacci(5), Fibonacci(6), + Fibonacci(7), Fibonacci(8), Fibonacci(9), Fibonacci(10), Fibonacci(11), + Fibonacci(12), Fibonacci(13), Fibonacci(14), Fibonacci(15), Fibonacci(16), + Fibonacci(17), Fibonacci(18), Fibonacci(19), Fibonacci(20), Fibonacci(21), + Fibonacci(22), Fibonacci(23), Fibonacci(24), Fibonacci(25), Fibonacci(26), + Fibonacci(27), Fibonacci(28), Fibonacci(29), Fibonacci(30), Fibonacci(31), + Fibonacci(32), Fibonacci(33), Fibonacci(34), Fibonacci(35), Fibonacci(36), + Fibonacci(37), Fibonacci(38), Fibonacci(39), Fibonacci(40), Fibonacci(41), + Fibonacci(42), Fibonacci(43), Fibonacci(44), Fibonacci(45), Fibonacci(46), + Fibonacci(47), Fibonacci(48), Fibonacci(49), Fibonacci(50), Fibonacci(51), + Fibonacci(52), Fibonacci(53), Fibonacci(54), Fibonacci(55), Fibonacci(56), + Fibonacci(57), Fibonacci(58), Fibonacci(59), Fibonacci(60), Fibonacci(61), + Fibonacci(62), Fibonacci(63), Fibonacci(64), Fibonacci(65), Fibonacci(66), + Fibonacci(67), Fibonacci(68), Fibonacci(69), Fibonacci(70), Fibonacci(71), + Fibonacci(72), Fibonacci(73), Fibonacci(74), Fibonacci(75), Fibonacci(76), + Fibonacci(77), Fibonacci(78), Fibonacci(79), Fibonacci(80), Fibonacci(81), + Fibonacci(82), Fibonacci(83), Fibonacci(84), Fibonacci(85), Fibonacci(86), + Fibonacci(87), Fibonacci(88), Fibonacci(89), Fibonacci(90), Fibonacci(91), + Fibonacci(92), Fibonacci(93)}; + +static_assert(sizeof(kMinLength) / sizeof(uint64_t) == + (cord_internal::MaxCordDepth() + 1), + "Not enough elements in kMinLength array to cover all the " + "supported Cord depth(s)"); + +inline bool ShouldRebalance(const CordRep* node) { + if (node->tag != CONCAT) return false; + + size_t node_depth = node->concat()->depth(); + + if (node_depth <= 15) return false; + + // Rebalancing Cords is expensive, so we reduce how often rebalancing occurs + // by allowing shallow Cords to have twice the depth that the Fibonacci rule + // would otherwise imply. Deep Cords need to follow the rule more closely, + // however to ensure algorithm correctness. We implement this with linear + // interpolation. Cords of depth 16 are treated as though they have a depth + // of 16 * 1/2, and Cords of depth MaxCordDepth() interpolate to + // MaxCordDepth() * 1. + return node->length < + kMinLength[(node_depth * (cord_internal::MaxCordDepth() - 16)) / + (2 * cord_internal::MaxCordDepth() - 16 - node_depth)]; +} + +// Unlike root balancing condition this one is part of the re-balancing +// algorithm and has to be always matching against right depth for +// algorithm to be correct. +inline bool IsNodeBalanced(const CordRep* node) { + if (node->tag != CONCAT) return true; + + size_t node_depth = node->concat()->depth(); + + return node->length >= kMinLength[node_depth]; } static CordRep* Rebalance(CordRep* node); -static void DumpNode(CordRep* rep, bool include_data, std::ostream* os); -static bool VerifyNode(CordRep* root, CordRep* start_node, +static void DumpNode(const CordRep* rep, bool include_data, std::ostream* os); +static bool VerifyNode(const CordRep* root, const CordRep* start_node, bool full_validation); static inline CordRep* VerifyTree(CordRep* node) { @@ -318,7 +304,8 @@ __attribute__((preserve_most)) static void UnrefInternal(CordRep* rep) { assert(rep != nullptr); - absl::InlinedVector<CordRep*, kInlinedVectorSize> pending; + cord_internal::RebalancingStack pending; + while (true) { if (rep->tag == CONCAT) { CordRepConcat* rep_concat = rep->concat(); @@ -400,6 +387,11 @@ static void SetConcatChildren(CordRepConcat* concat, CordRep* left, concat->length = left->length + right->length; concat->set_depth(1 + std::max(Depth(left), Depth(right))); + + ABSL_INTERNAL_CHECK(concat->depth() <= cord_internal::MaxCordDepth(), + "Cord depth exceeds max"); + ABSL_INTERNAL_CHECK(concat->length >= left->length, "Cord is too long"); + ABSL_INTERNAL_CHECK(concat->length >= right->length, "Cord is too long"); } // Create a concatenation of the specified nodes. @@ -425,7 +417,7 @@ static CordRep* RawConcat(CordRep* left, CordRep* right) { static CordRep* Concat(CordRep* left, CordRep* right) { CordRep* rep = RawConcat(left, right); - if (rep != nullptr && !IsRootBalanced(rep)) { + if (rep != nullptr && ShouldRebalance(rep)) { rep = Rebalance(rep); } return VerifyTree(rep); @@ -916,7 +908,7 @@ void Cord::Prepend(absl::string_view src) { static CordRep* RemovePrefixFrom(CordRep* node, size_t n) { if (n >= node->length) return nullptr; if (n == 0) return Ref(node); - absl::InlinedVector<CordRep*, kInlinedVectorSize> rhs_stack; + cord_internal::CordTreeMutablePath rhs_stack; while (node->tag == CONCAT) { assert(n <= node->length); @@ -957,7 +949,7 @@ static CordRep* RemovePrefixFrom(CordRep* node, size_t n) { static CordRep* RemoveSuffixFrom(CordRep* node, size_t n) { if (n >= node->length) return nullptr; if (n == 0) return Ref(node); - absl::InlinedVector<CordRep*, kInlinedVectorSize> lhs_stack; + absl::cord_internal::CordTreeMutablePath lhs_stack; bool inplace_ok = node->refcount.IsOne(); while (node->tag == CONCAT) { @@ -1028,6 +1020,7 @@ void Cord::RemoveSuffix(size_t n) { // Work item for NewSubRange(). struct SubRange { + SubRange() = default; SubRange(CordRep* a_node, size_t a_pos, size_t a_n) : node(a_node), pos(a_pos), n(a_n) {} CordRep* node; // nullptr means concat last 2 results. @@ -1036,8 +1029,11 @@ struct SubRange { }; static CordRep* NewSubRange(CordRep* node, size_t pos, size_t n) { - absl::InlinedVector<CordRep*, kInlinedVectorSize> results; - absl::InlinedVector<SubRange, kInlinedVectorSize> todo; + cord_internal::CordTreeMutablePath results; + // The algorithm below in worst case scenario adds up to 3 nodes to the `todo` + // list, but we also pop one out on every cycle. If original tree has depth d + // todo list can grew up to 2*d in size. + cord_internal::CordTreePath<SubRange, 2 * cord_internal::MaxCordDepth()> todo; todo.push_back(SubRange(node, pos, n)); do { const SubRange& sr = todo.back(); @@ -1074,7 +1070,7 @@ static CordRep* NewSubRange(CordRep* node, size_t pos, size_t n) { } } while (!todo.empty()); assert(results.size() == 1); - return results[0]; + return results.back(); } Cord Cord::Subcord(size_t pos, size_t new_size) const { @@ -1113,11 +1109,12 @@ Cord Cord::Subcord(size_t pos, size_t new_size) const { class CordForest { public: - explicit CordForest(size_t length) - : root_length_(length), trees_(kMinLengthSize, nullptr) {} + explicit CordForest(size_t length) : root_length_(length), trees_({}) {} void Build(CordRep* cord_root) { - std::vector<CordRep*> pending = {cord_root}; + // We are adding up to two nodes to the `pending` list, but we also popping + // one, so the size of `pending` will never exceed `MaxCordDepth()`. + cord_internal::CordTreeMutablePath pending(cord_root); while (!pending.empty()) { CordRep* node = pending.back(); @@ -1129,21 +1126,20 @@ class CordForest { } CordRepConcat* concat_node = node->concat(); - if (concat_node->depth() >= kMinLengthSize || - concat_node->length < min_length[concat_node->depth()]) { - pending.push_back(concat_node->right); - pending.push_back(concat_node->left); - - if (concat_node->refcount.IsOne()) { - concat_node->left = concat_freelist_; - concat_freelist_ = concat_node; - } else { - Ref(concat_node->right); - Ref(concat_node->left); - Unref(concat_node); - } - } else { + if (IsNodeBalanced(concat_node)) { AddNode(node); + continue; + } + pending.push_back(concat_node->right); + pending.push_back(concat_node->left); + + if (concat_node->refcount.IsOne()) { + concat_node->left = concat_freelist_; + concat_freelist_ = concat_node; + } else { + Ref(concat_node->right); + Ref(concat_node->left); + Unref(concat_node); } } } @@ -1175,7 +1171,7 @@ class CordForest { // Collect together everything with which we will merge node int i = 0; - for (; node->length > min_length[i + 1]; ++i) { + for (; node->length > kMinLength[i + 1]; ++i) { auto& tree_at_i = trees_[i]; if (tree_at_i == nullptr) continue; @@ -1186,7 +1182,7 @@ class CordForest { sum = AppendNode(node, sum); // Insert sum into appropriate place in the forest - for (; sum->length >= min_length[i]; ++i) { + for (; sum->length >= kMinLength[i]; ++i) { auto& tree_at_i = trees_[i]; if (tree_at_i == nullptr) continue; @@ -1194,7 +1190,7 @@ class CordForest { tree_at_i = nullptr; } - // min_length[0] == 1, which means sum->length >= min_length[0] + // kMinLength[0] == 1, which means sum->length >= kMinLength[0] assert(i > 0); trees_[i - 1] = sum; } @@ -1227,9 +1223,7 @@ class CordForest { } size_t root_length_; - - // use an inlined vector instead of a flat array to get bounds checking - absl::InlinedVector<CordRep*, kInlinedVectorSize> trees_; + std::array<cord_internal::CordRep*, cord_internal::MaxCordDepth()> trees_; // List of concat nodes we can re-use for Cord balancing. CordRepConcat* concat_freelist_ = nullptr; @@ -1841,18 +1835,18 @@ absl::string_view Cord::FlattenSlowPath() { } } -static void DumpNode(CordRep* rep, bool include_data, std::ostream* os) { +static void DumpNode(const CordRep* rep, bool include_data, std::ostream* os) { const int kIndentStep = 1; int indent = 0; - absl::InlinedVector<CordRep*, kInlinedVectorSize> stack; - absl::InlinedVector<int, kInlinedVectorSize> indents; + cord_internal::CordTreeConstPath stack; + cord_internal::CordTreePath<int, cord_internal::MaxCordDepth()> indents; for (;;) { *os << std::setw(3) << rep->refcount.Get(); *os << " " << std::setw(7) << rep->length; *os << " ["; - if (include_data) *os << static_cast<void*>(rep); + if (include_data) *os << static_cast<const void*>(rep); *os << "]"; - *os << " " << (IsRootBalanced(rep) ? 'b' : 'u'); + *os << " " << (IsNodeBalanced(rep) ? 'b' : 'u'); *os << " " << std::setw(indent) << ""; if (rep->tag == CONCAT) { *os << "CONCAT depth=" << Depth(rep) << "\n"; @@ -1873,7 +1867,7 @@ static void DumpNode(CordRep* rep, bool include_data, std::ostream* os) { } else { *os << "FLAT cap=" << TagToLength(rep->tag) << " ["; if (include_data) - *os << absl::CEscape(std::string(rep->data, rep->length)); + *os << absl::CEscape(absl::string_view(rep->data, rep->length)); *os << "]\n"; } if (stack.empty()) break; @@ -1886,19 +1880,19 @@ static void DumpNode(CordRep* rep, bool include_data, std::ostream* os) { ABSL_INTERNAL_CHECK(indents.empty(), ""); } -static std::string ReportError(CordRep* root, CordRep* node) { +static std::string ReportError(const CordRep* root, const CordRep* node) { std::ostringstream buf; buf << "Error at node " << node << " in:"; DumpNode(root, true, &buf); return buf.str(); } -static bool VerifyNode(CordRep* root, CordRep* start_node, +static bool VerifyNode(const CordRep* root, const CordRep* start_node, bool full_validation) { - absl::InlinedVector<CordRep*, 2> worklist; + cord_internal::CordTreeConstPath worklist; worklist.push_back(start_node); do { - CordRep* node = worklist.back(); + const CordRep* node = worklist.back(); worklist.pop_back(); ABSL_INTERNAL_CHECK(node != nullptr, ReportError(root, node)); @@ -1948,7 +1942,7 @@ static bool VerifyNode(CordRep* root, CordRep* start_node, // Iterate over the tree. cur_node is never a leaf node and leaf nodes will // never be appended to tree_stack. This reduces overhead from manipulating // tree_stack. - absl::InlinedVector<const CordRep*, kInlinedVectorSize> tree_stack; + cord_internal::CordTreeConstPath tree_stack; const CordRep* cur_node = rep; while (true) { const CordRep* next_node = nullptr; diff --git a/absl/strings/cord.h b/absl/strings/cord.h index 40566cbaa011..68a7e52feb7a 100644 --- a/absl/strings/cord.h +++ b/absl/strings/cord.h @@ -41,13 +41,13 @@ #include <iostream> #include <iterator> #include <string> +#include <type_traits> #include "absl/base/internal/endian.h" #include "absl/base/internal/invoke.h" #include "absl/base/internal/per_thread_tls.h" #include "absl/base/macros.h" #include "absl/base/port.h" -#include "absl/container/inlined_vector.h" #include "absl/functional/function_ref.h" #include "absl/meta/type_traits.h" #include "absl/strings/internal/cord_internal.h" @@ -66,6 +66,73 @@ template <typename H> H HashFragmentedCord(H, const Cord&); } +namespace cord_internal { + +// It's expensive to keep a tree perfectly balanced, so instead we keep trees +// approximately balanced. A tree node N of depth D(N) that contains a string +// of L(N) characters is considered balanced if L >= Fibonacci(D + 2). +// The "+ 2" is used to ensure that every leaf node contains at least one +// character. Here we presume that +// Fibonacci(0) = 0 +// Fibonacci(1) = 1 +// Fibonacci(2) = 1 +// Fibonacci(3) = 2 +// ... +// +// Fibonacci numbers are convenient because it means when two balanced trees of +// the same depth are made the children of a new node, the resulting tree is +// guaranteed to also be balanced: +// +// +// L(left) >= Fibonacci(D(left) + 2) +// L(right) >= Fibonacci(D(right) + 2) +// +// L(left) + L(right) >= Fibonacci(D(left) + 2) + Fibonacci(D(right) + 2) +// L(left) + L(right) == L(new_tree) +// +// L(new_tree) >= 2 * Fibonacci(D(child) + 2) +// D(child) == D(new_tree) - 1 +// +// L(new_tree) >= 2 * Fibonacci(D(new_tree) + 1) +// 2 * Fibonacci(N) >= Fibonacci(N + 1) +// +// L(new_tree) >= Fibonacci(D(new_tree) + 2) +// +// +// The 93rd Fibonacci number is the largest Fibonacci number that can be +// represented in 64 bits, so the size of a balanced Cord of depth 92 is too big +// for an unsigned 64 bit integer to hold. Therefore we can safely assume that +// the maximum depth of a Cord is 91. +constexpr size_t MaxCordDepth() { return 91; } + +// This class models fixed max size stack of CordRep pointers. +// The elements are being pushed back and popped from the back. +template <typename CordRepPtr, size_t N> +class CordTreePath { + public: + CordTreePath() {} + explicit CordTreePath(CordRepPtr root) { push_back(root); } + + bool empty() const { return size_ == 0; } + size_t size() const { return size_; } + void clear() { size_ = 0; } + + CordRepPtr back() { return data_[size_ - 1]; } + + void pop_back() { + --size_; + assert(size_ < N); + } + void push_back(CordRepPtr elem) { data_[size_++] = elem; } + + private: + CordRepPtr data_[N]; + size_t size_ = 0; +}; + +using CordTreeMutablePath = CordTreePath<CordRep*, MaxCordDepth()>; +} // namespace cord_internal + // A Cord is a sequence of characters. class Cord { private: @@ -114,7 +181,8 @@ class Cord { // finished with `data`. The data must remain live and unchanging until the // releaser is called. The requirements for the releaser are that it: // * is move constructible, - // * supports `void operator()(absl::string_view) const`, + // * supports `void operator()(absl::string_view) const` or + // `void operator()() const`, // * does not have alignment requirement greater than what is guaranteed by // ::operator new. This is dictated by alignof(std::max_align_t) before // C++17 and __STDCPP_DEFAULT_NEW_ALIGNMENT__ if compiling with C++17 or @@ -127,8 +195,8 @@ class Cord { // FillBlock(block); // return absl::MakeCordFromExternal( // block->ToStringView(), - // [pool, block](absl::string_view /*ignored*/) { - // pool->FreeBlock(block); + // [pool, block](absl::string_view v) { + // pool->FreeBlock(block, v); // }); // } // @@ -282,8 +350,7 @@ class Cord { absl::cord_internal::CordRep* current_leaf_ = nullptr; // The number of bytes left in the `Cord` over which we are iterating. size_t bytes_remaining_ = 0; - absl::InlinedVector<absl::cord_internal::CordRep*, 4> - stack_of_right_children_; + absl::cord_internal::CordTreeMutablePath stack_of_right_children_; }; // Returns an iterator to the first chunk of the `Cord`. @@ -667,6 +734,21 @@ ExternalRepReleaserPair NewExternalWithUninitializedReleaser( absl::string_view data, ExternalReleaserInvoker invoker, size_t releaser_size); +struct Rank1 {}; +struct Rank0 : Rank1 {}; + +template <typename Releaser, typename = ::absl::base_internal::InvokeT< + Releaser, absl::string_view>> +void InvokeReleaser(Rank0, Releaser&& releaser, absl::string_view data) { + ::absl::base_internal::Invoke(std::forward<Releaser>(releaser), data); +} + +template <typename Releaser, + typename = ::absl::base_internal::InvokeT<Releaser>> +void InvokeReleaser(Rank1, Releaser&& releaser, absl::string_view) { + ::absl::base_internal::Invoke(std::forward<Releaser>(releaser)); +} + // Creates a new `CordRep` that owns `data` and `releaser` and returns a pointer // to it, or `nullptr` if `data` was empty. template <typename Releaser> @@ -684,14 +766,14 @@ CordRep* NewExternalRep(absl::string_view data, Releaser&& releaser) { using ReleaserType = absl::decay_t<Releaser>; if (data.empty()) { // Never create empty external nodes. - ::absl::base_internal::Invoke( - ReleaserType(std::forward<Releaser>(releaser)), data); + InvokeReleaser(Rank0{}, ReleaserType(std::forward<Releaser>(releaser)), + data); return nullptr; } auto releaser_invoker = [](void* type_erased_releaser, absl::string_view d) { auto* my_releaser = static_cast<ReleaserType*>(type_erased_releaser); - ::absl::base_internal::Invoke(std::move(*my_releaser), d); + InvokeReleaser(Rank0{}, std::move(*my_releaser), d); my_releaser->~ReleaserType(); return sizeof(Releaser); }; diff --git a/absl/strings/cord_test.cc b/absl/strings/cord_test.cc index 434f3a247e08..68515cbf15b9 100644 --- a/absl/strings/cord_test.cc +++ b/absl/strings/cord_test.cc @@ -1032,6 +1032,19 @@ TEST(ConstructFromExternal, MoveOnlyReleaser) { EXPECT_TRUE(invoked); } +TEST(ConstructFromExternal, NoArgLambda) { + bool invoked = false; + (void)absl::MakeCordFromExternal("dummy", [&invoked]() { invoked = true; }); + EXPECT_TRUE(invoked); +} + +TEST(ConstructFromExternal, StringViewArgLambda) { + bool invoked = false; + (void)absl::MakeCordFromExternal( + "dummy", [&invoked](absl::string_view) { invoked = true; }); + EXPECT_TRUE(invoked); +} + TEST(ConstructFromExternal, NonTrivialReleaserDestructor) { struct Releaser { explicit Releaser(bool* destroyed) : destroyed(destroyed) {} @@ -1346,6 +1359,49 @@ TEST(CordChunkIterator, Operations) { VerifyChunkIterator(subcords, 128); } +TEST(CordChunkIterator, MaxLengthFullTree) { + absl::Cord cord; + size_t size = 1; + AddExternalMemory("x", &cord); + EXPECT_EQ(cord.size(), size); + + for (int i = 0; i < 63; ++i) { + cord.Prepend(absl::Cord(cord)); + size <<= 1; + + EXPECT_EQ(cord.size(), size); + + auto chunk_it = cord.chunk_begin(); + EXPECT_EQ(*chunk_it, "x"); + } + + EXPECT_DEATH_IF_SUPPORTED( + (cord.Prepend(absl::Cord(cord)), *cord.chunk_begin()), + "Cord is too long"); +} + +TEST(CordChunkIterator, MaxDepth) { + // By reusing nodes, it's possible in pathological cases to build a Cord that + // exceeds both the maximum permissible length and depth. In this case, the + // violation of the maximum depth is reported. + absl::Cord left_child; + AddExternalMemory("x", &left_child); + absl::Cord root = left_child; + + for (int i = 0; i < 91; ++i) { + size_t new_size = left_child.size() + root.size(); + root.Prepend(left_child); + EXPECT_EQ(root.size(), new_size); + + auto chunk_it = root.chunk_begin(); + EXPECT_EQ(*chunk_it, "x"); + + std::swap(left_child, root); + } + + EXPECT_DEATH_IF_SUPPORTED(root.Prepend(left_child), "Cord depth exceeds max"); +} + TEST(CordCharIterator, Traits) { static_assert(std::is_copy_constructible<absl::Cord::CharIterator>::value, ""); |