Skip to content

Latest commit

 

History

History
50 lines (36 loc) · 2.16 KB

readme.md

File metadata and controls

50 lines (36 loc) · 2.16 KB
Raw

stackalloc

A quick prototype for vector-like containers based on address space reservation.

This has the following advantages:

  1. Allows interior pointers, since memory never moves (!).
  2. Super cheap buffer allocation (pointer bump).
  3. Super cheap push_back (pointer bump, no bounds check needed).

The quick benchmark in test.cpp shows the following speedup over std::vector (on a 8700K CPU on Windows):

  • Clang 10 release:

    [5 elems] stackalloc: 0.0032, stl naive: 0.1376, stl opt: 0.0302, ratio: 42.47x / 9.32x faster!
[50 elems] stackalloc: 0.0328, stl naive: 0.3030, stl opt: 0.0782, ratio: 9.25x / 2.39x faster!
[500 elems] stackalloc: 0.2966, stl naive: 1.0705, stl opt: 0.6607, ratio: 3.61x / 2.23x faster!

  • VS2019 release:

    [5 elems] stackalloc: 0.0057, stl naive: 0.1555, stl opt: 0.0275, ratio: 27.46x / 4.85x faster!
[50 elems] stackalloc: 0.0657, stl naive: 0.3761, stl opt: 0.0883, ratio: 5.72x / 1.34x faster!
[500 elems] stackalloc: 0.6485, stl naive: 1.2304, stl opt: 0.7924, ratio: 1.90x / 1.22x faster!

  • VS2019 debug:

    [5 elems] stackalloc: 0.0802, stl naive: 1.2429, stl opt: 0.5559, ratio: 15.49x / 6.93x faster!
[50 elems] stackalloc: 0.6124, stl naive: 4.4576, stl opt: 2.6846, ratio: 7.28x / 4.38x faster!
[500 elems] stackalloc: 5.5520, stl naive: 26.9562, stl opt: 23.5752, ratio: 4.86x / 4.25x faster!

Since on Windows the minimum address space reservation is 64KB, the code in this repo is such that these "stacks" that back vector memory are shared between vectors where possible (thru RAII), such that even small vectors are cheap.

Disadvantages:

  • You need to manage all your memory in vectors owned by local variables, rather than using dynamic memory directly. This is actually not much of a limitation anymore, since these vectors allow interior pointers, so they can function as general allocators with no limitations.
  • Since it relies on address space reservation, it does not work on platforms that do not support mmap (or VirtualAlloc), like some IOT / embedded platforms, and currently also WebAssembly (WebAssembly/memory64#4)