I suggest NOT using the built in intrinsics and implicit vectors. This only makes sense if you don't use the non GCC intrinsics (e.g. _mm_cmpeq_epi32) and only want to stick to GCC. You can do what you want like this
__m128i key2 = _mm_set1_epi64x(key);
__m128i v1 = _mm_loadu_si128((const __m128i *)&data[start + i + 0]);
__m128i v2 = _mm_loadu_si128((const __m128i *)&data[start + i + 2]);
__m128i cmp0 = _mm_cmpeq_epi64(key2, v1);
__m128i cmp1 = _mm_cmpeq_epi64(key2, v2);
__m128i low2 = _mm_shuffle_epi32(cmp0,0xD8);
__m128i high2 = _mm_shuffle_epi32(cmp1,0xD8);
__m128i pack = _mm_unpacklo_epi64(low2,high2);
__m128i pack01 = _mm_packs_epi32(pack, _mm_setzero_si128());
__m128i pack0123 = _mm_packs_epi16(pack01, _mm_setzero_si128());
int res = _mm_movemask_epi8(pack0123);
You can probably find a more efficient version that avoids the packing but then you would have to use a different function than __builtin_ctz
.
For 32-bit ints I suggest
__m128i key4 = _mm_set1_epi32(key);
__m128i v1 = _mm_loadu_si128((const __m128i *)&data[start + i + 0]);
__m128i v2 = _mm_loadu_si128((const __m128i *)&data[start + i + 4]);
__m128i v3 = _mm_loadu_si128((const __m128i *)&data[start + i + 8]);
__m128i v4 = _mm_loadu_si128((const __m128i *)&data[start + i + 12]);
__m128i cmp0 = _mm_cmpeq_epi32(key4, v1);
__m128i cmp1 = _mm_cmpeq_epi32(key4, v2);
__m128i cmp2 = _mm_cmpeq_epi32(key4, v3);
__m128i cmp3 = _mm_cmpeq_epi32(key4, v4);
__m128i pack01 = _mm_packs_epi32(cmp0, cmp1);
__m128i pack23 = _mm_packs_epi32(cmp2, cmp3);
__m128i pack0123 = _mm_packs_epi16(pack01, pack23);
int res = _mm_movemask_epi8(pack0123);