101   __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);  in compress_pre()  local
112   t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10  8  in compress_pre()
115   t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11  9  in compress_pre()
122   m3 = t3;  in compress_pre()
129   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));  in compress_pre()
134   t2 = _mm_unpacklo_epi64(m3, m1);  in compress_pre()
138   t3 = _mm_unpackhi_epi32(m1, m3);  in compress_pre()
146   m3 = t3;  in compress_pre()
152   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));  in compress_pre()
157   t2 = _mm_unpacklo_epi64(m3, m1);  in compress_pre()
161   t3 = _mm_unpackhi_epi32(m1, m3);  in compress_pre()
169   m3 = t3;  in compress_pre()
175   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));  in compress_pre()
180   t2 = _mm_unpacklo_epi64(m3, m1);  in compress_pre()
184   t3 = _mm_unpackhi_epi32(m1, m3);  in compress_pre()
192   m3 = t3;  in compress_pre()
198   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));  in compress_pre()
203   t2 = _mm_unpacklo_epi64(m3, m1);  in compress_pre()
207   t3 = _mm_unpackhi_epi32(m1, m3);  in compress_pre()
215   m3 = t3;  in compress_pre()
221   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));  in compress_pre()
226   t2 = _mm_unpacklo_epi64(m3, m1);  in compress_pre()
230   t3 = _mm_unpackhi_epi32(m1, m3);  in compress_pre()
238   m3 = t3;  in compress_pre()
244   t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));  in compress_pre()
249   t2 = _mm_unpacklo_epi64(m3, m1);  in compress_pre()
253   t3 = _mm_unpackhi_epi32(m1, m3);  in compress_pre()