No need to memset to 0 after mmap to /dev/zero

[Tianlong Wu <rockuw@…>]

In boost source /libs/coroutine/detail/standard_stack_allocator_posix.cpp:120, it has the following snippet:

    void * limit =
# if defined(macintosh) || defined(__APPLE__) || defined(__APPLE_CC__)
    ::mmap( 0, size_, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
# else
    ::mmap( 0, size_, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
# endif
    ::close( fd);
    if ( ! limit) throw std::bad_alloc();

    std::memset( limit, '\0', size_);

Since mmap() with the MAP_PRIVATE flag implies a copy-on-write mechanism which ensures that memory pointed by limit will be initialized to all-zero. The additional memset slows down coroutine creation and makes coroutine not that lightweight. In this case of memseting ~64KB memory takes ~40us on my machine. And creating a coroutine takes ~50us.

[olli]

I've added class standard_stack_allocator (using malloc/free) to the library. the class is the standard stack-allocator and on my machine (Core2 Q6700) creating a coroutine consumes ca. 200ns.

[Tianlong Wu <rockuw@…>]

Replying to olli:

I've added class standard_stack_allocator (using malloc/free) to the library. the class is the standard stack-allocator and on my machine (Core2 Q6700) creating a coroutine consumes ca. 200ns.

Good. When can we use it? Will it be checked into 1.55 or only trunk?

[olli]

will be in boost-1.56 thx, Oliver