The following two code samples achieve the same thing. They search in a string of 8 bit chars and return an index of the first match.
The first way uses a PCMPxSTRx instruction.
int Index = _mm_cmpistri(a, b, _SIDD_CMP_EQUAL_EACH);
The second way uses a PCMPEQ instruction.
__m128i VectorMask = _mm_cmpeq_epi8(a, b);
int Mask = _mm_movemask_epi8(VectorMask);
unsigned long Index {};
_BitScanForward(&Index, Mask);
Which way is better to use? Supposing that I know that the two vectors contain only 'valid characters' (no zeros) and that I am not interested in any manipulations of the result mask.
The first way looks more compact in the high level code, but I feel that it will produce a bigger number of micro-ops when decoded by the CPU, based on its description in the Intel intrinsics guide. So perhaps the second way is more efficient.
