Vectorize alignment algorithm for x86-64

Ideally you would want to use in this case __m128i _mm_blendv_epi8(__m128i a, __m128i b, __m128i mask)
where the mask could be created with anything in the _mm_cmp**_epi8 range.

But blendv is a SSE4.1 instruction. Leading to compile headaches. However, this can be done using SSE2 instructions only:

<include "emmintrin.h">

static inline __m128i vector_blend_128(__m128i a, __m128i b, __m128i mask) {
     return _mm_or_si128(
         _mm_and_si128(mask, a);
         _mm_andnot_si128(mask, b);
     );
}

So this might open up opportunities for vectorization, using only #ifdef __SSE2__ compile guards.

EDIT: This would work for other than epi8 data types as well of course.

Start by locating the alignment algorithm and any existing x86-64 or SIMD implementation in the repository. Check how CPU feature guards are handled, then evaluate whether an SSE2-only blend using the proposed intrinsics can be integrated without SSE4.1 requirements; done means the alignment path is vectorized and remains buildable on supported x86-64 targets.