lhmouse / mcf Goto Github PK

1. Use _mm_setzero_si128() instead of _mm_set1_epi8(0)( when 0 cannot be recognised as a constant ), which means don't forward strlen to memchr and friends.
   _mm_set1_epi8(c) means movd followed by pshufd or even worse. _mm_setzero_si128() simply means pxor.
   Yes GCC 7.1 does generate two functions for _mm_set1_epi8(0) as a special case, which was a surprise. But ( I suppose ) most compilers won't.

2. Use _mm_loadu_si128 for aligning pointers
   You may load unaligned 16-n bytes completely using _mm_loadu_si128 instead of writing a loop. This improves performance greatly when the buffer is unaligned and small. It helps more after migrating to avx.

3. Unroll loops manually.
   It's worth doing to make CPU executes them in parallel. Note twice is enough because in this case, 2 masks generated by _mm_movemask_epi8 ( 2 bytes once ) can be stored into one register (-m32) and then being scanned by bsf or tzcnt ( aka rep bsf ).

Test cases (use online compiler for fair):

1. https://ideone.com/1vwHi3
   Aligned large buffer: about 10% improvement.

2. https://ideone.com/5UExGr
   Unaligned small buffer: huge improvement.

3. https://ideone.com/vc8QIF
   Unaligned small buffer ( only 1 byte to scan ): almost the same, but still better.

NtDelayExecution 的第二个参数为nullptr 在Win10 上会返回无效的内存访问错误

Github is telling that this is C code.
I think a .gitignore will quickly fix this.

Hello lhmouse :)

Could you also include mingw32-make into your GCC-MCF distribution, please?

Thanks,
Arlen.