mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-12-05 22:12:21 +00:00
b7ef5a699c
* Get rid of non-existent vectorclass * Move FastMemcpy to contribs * Restore comments * Disable FastMemcpy on non-Linux * Fix cmake file * Don't build FastMemcpy for ARM64 * Replace FastMemcpy submodule with its contents * Fix cmake file * Move widechar_width to contrib/ * Move sumbur to contrib/ * Move consistent-hashing to contrib/ * Fix UBSan tests
21 lines
1.3 KiB
Markdown
21 lines
1.3 KiB
Markdown
Internal implementation of `memcpy` function.
|
|
|
|
It has the following advantages over `libc`-supplied implementation:
|
|
- it is linked statically, so the function is called directly, not through a `PLT` (procedure lookup table of shared library);
|
|
- it is linked statically, so the function can have position-dependent code;
|
|
- your binaries will not depend on `glibc`'s memcpy, that forces dependency on specific symbol version like `memcpy@@GLIBC_2.14` and consequently on specific version of `glibc` library;
|
|
- you can include `memcpy.h` directly and the function has the chance to be inlined, which is beneficial for small but unknown at compile time sizes of memory regions;
|
|
- this version of `memcpy` pretend to be faster (in our benchmarks, the difference is within few percents).
|
|
|
|
Currently it uses the implementation from **Linwei** (skywind3000@163.com).
|
|
Look at https://www.zhihu.com/question/35172305 for discussion.
|
|
|
|
Drawbacks:
|
|
- only use SSE 2, doesn't use wider (AVX, AVX 512) vector registers when available;
|
|
- no CPU dispatching; doesn't take into account actual cache size.
|
|
|
|
Also worth to look at:
|
|
- simple implementation from Facebook: https://github.com/facebook/folly/blob/master/folly/memcpy.S
|
|
- implementation from Agner Fog: http://www.agner.org/optimize/
|
|
- glibc source code.
|