Skip to content

TypeCombinator/uit

Repository files navigation

UIT

Introduction

UIT is a Unified Intrusive Linked List library. This library brings together 4 different linked lists based on a unified abstraction, the core trick is mock_head. In addition, some balanced trees have been implemented, what makes it different is that they're simplified with mock_sentinel, which is similar to the mock_head.

types number of head node pointers number of data node pointers order comments
uit::islist single single LIFO Singly Linked List.
uit::isdlist single double LIFO It's equivalent to the hlist in linux, and mainly used to implement hash tables.
uit::idslist double single LIFO or FIFO Data nodes have a smaller footprint.
uit::idlist double double LIFO or FIFO Doubly Linked List.

The trick named mock_head brings the following pros and cons.

Pros

  • Simple and easy-to-understand code
  • Fewer branches and better performance, e.g. there's no branch in the uit::idslist::push_back method
  • Secondary pointers are no longer required, even the hlist (uit::isdlist) is no exception
  • Truly zero overhead
  • No macros

Cons

  • There are some undefined behaviors in the implementation of the mock_head. The main branch uses type composition and container_of to implement the mock_head, which is completely UB, some test items cannot pass under high optimization level. The v2 branch uses type inheritance and static_cast to implement the mock_head, unfortunately, this implementation still has some UBs, such as the mock_head implementation of idslist, where the head and mock_head have no inheritance relationship with each other, and violate strict aliasing.
  • Among these, isdlist, idslist, and idlist are self-referential types, with isdlist and idlist being move-only.

Notice

This project is used for experimental exploration, I will not bear any consequences caused by the practical application of this project.

It must be said that the v2 branch is a better implementation under the current C++ standard definition, as it can currently (not guaranteed in the future) work on Clang and MSVC with high-level optimizations (such as O3), but for GCC, you must add an extra option-fno-strict-aliasing.

The main branch is mainly used to help understand, due to container_of being UB and having many unresolved conflicts with the C++ standard, which does not conform to the future development direction of C++, I am pessimistic about the possibility of container_of becoming a standard definition, despite the significant demands in many scenarios.

I have a private repository named uit_ng that still uses mock_head, it doesn't employ type inheritance, yet maintains zero-overhead abstraction. Despite this, it works correctly across Big Three (GCC, Clang and MSVC) at the highest optimization levels, and passes all test cases without requiring additional compiler flags or special attributes in the source code. So far, I haven’t been able to create a test case that it fails. If enough people express interest, I would consider making this repository public.

mock_head

For simplicity, I will use a singly linked list as an example in C language. Firstly, we define a singly linked list node type struct node, and implement the push_front, the code is as follows:

#include <stddef.h>

struct node {
    int sn;
    struct node *next;
};

static struct node *head = NULL;

void push_front(struct node *n) {
    n->next = head;
    head = n;
}

Traditionally, when we want to remove a node from a singly linked list, we need the help of a secondary pointer.

struct node *remove_with_secondary_pointer(struct node *n) {
    struct node **prev = &head;
    for (struct node *cur = *prev; cur != NULL;) {
        if (cur == n) {
            *prev = cur->next;
            return cur;
        }
        prev = &cur->next;
        cur = *prev;
    }
    return NULL;
}

Why do we need a secondary pointer? The core reason is that the type of the head node and data node are different, and the secondary pointer unifies them. If we define the head as static struct node head = {0, NULL};, the secondary pointer will no longer be needed, but there is a problem of memory waste, the member int sn; is redundant in the example, we only need the member struct node *next;.

The solution is to use head node to simulate data node through the container_of, the code is as follows:

#define container_of(_field_, _type_, _member_)                                                    \
    ((_type_ *) ((size_t) (_field_) -offsetof(_type_, _member_)))

#define MOCK_HEAD(_field_, _type_, _member_) container_of(_field_, _type_, _member_)

struct node *remove_with_mock_head(struct node *n) {
    struct node *prev = MOCK_HEAD(&head, struct node, next);
    for (struct node *cur = prev->next; cur != NULL;) {
        if (cur == n) {
            prev->next = cur->next;
            return cur;
        }
        prev = cur;
        cur = prev->next;
    }
    return NULL;
}

As you can see, mock_head is just a alias of the contianer_of, it unifies head node and data node, just like secondary pointer. It should be noted that when accessing head node members, all members except for struct node *next; are undefined in the example.

Of course, this is just an example. The implementation of container_of is slightly different in C++, please see the header in this library.

slist

For an empty slist: head.right = nullptr;.

slist

sdlist

For an empty sdlist: head.right = nullptr;.

sdlist

dslist

dslist_empty

dslist

dlist

dlist_empty

dlist

mock_sentinel

mock_sentinel

Example

For the main branch

#include <iostream>
#include <uit/islist.hpp>

class apple {
   public:
    apple(uint64_t weight, int sn)
        : weight(weight)
        , sn(sn) {
    }

    uint64_t weight;
    uit::isnode<apple> node;
    int sn;
};

int main(int argc, char *argv[]) {
    uit::islist<&apple::node> list{};

    apple a0{500, 0};
    apple a1{501, 1};
    apple a2{502, 2};
    apple a3{503, 3};

    list.push_front(&a3);
    list.push_front(&a2);
    list.push_front(&a1);
    list.push_front(&a0);

    for (const auto &i: list) {
        std::cout << "sn: " << i.sn << ", weight: " << i.weight << std::endl;
    }
    return 0;
}

For the v2 branch

#include <iostream>
#include <uit/islist.hpp>

class apple : public uit::isnode<apple, "0"> {
   public:
    apple(uint64_t weight, int sn)
        : weight(weight)
        , sn(sn) {
    }

    uint64_t weight;
    int sn;
};

int main(int argc, char *argv[]) {
    uit::islist<apple, "0"> list{};

    apple a0{500, 0};
    apple a1{501, 1};
    apple a2{502, 2};
    apple a3{503, 3};

    list.push_front(&a3);
    list.push_front(&a2);
    list.push_front(&a1);
    list.push_front(&a0);

    for (const auto &i: list) {
        std::cout << "sn: " << i.sn << ", weight: " << i.weight << std::endl;
    }
    return 0;
}

More

Please refer to the examples and tests folder.

TODO

  • Need more tests.

Acknowledgements

ykiko helped me answer the key questions I encountered in this project, and thank you all for your attention to this project.

About

Unified Intrusive Linked List library for C++20

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published