17 #define LOAD_MSG_0_1(b0, b1) \  argument
21 b1 = _mm_unpacklo_epi64(m2, m3); \
25 #define LOAD_MSG_0_2(b0, b1) \  argument
29 b1 = _mm_unpackhi_epi64(m2, m3); \
33 #define LOAD_MSG_0_3(b0, b1) \  argument
37 b1 = _mm_unpacklo_epi64(m6, m7); \
41 #define LOAD_MSG_0_4(b0, b1) \  argument
45 b1 = _mm_unpackhi_epi64(m6, m7); \
49 #define LOAD_MSG_1_1(b0, b1) \  argument
53 b1 = _mm_unpackhi_epi64(m4, m6); \
57 #define LOAD_MSG_1_2(b0, b1) \  argument
61 b1 = _mm_alignr_epi8(m3, m7, 8); \
65 #define LOAD_MSG_1_3(b0, b1) \  argument
69 b1 = _mm_unpackhi_epi64(m5, m2); \
73 #define LOAD_MSG_1_4(b0, b1) \  argument
77 b1 = _mm_unpackhi_epi64(m3, m1); \
81 #define LOAD_MSG_2_1(b0, b1) \  argument
85 b1 = _mm_unpackhi_epi64(m2, m7); \
89 #define LOAD_MSG_2_2(b0, b1) \  argument
93 b1 = _mm_blend_epi16(m1, m6, 0xF0); \
97 #define LOAD_MSG_2_3(b0, b1) \  argument
101 b1 = _mm_unpackhi_epi64(m3, m4); \
105 #define LOAD_MSG_2_4(b0, b1) \  argument
109 b1 = _mm_alignr_epi8(m2, m0, 8); \
113 #define LOAD_MSG_3_1(b0, b1) \  argument
117 b1 = _mm_unpackhi_epi64(m6, m5); \
121 #define LOAD_MSG_3_2(b0, b1) \  argument
125 b1 = _mm_unpacklo_epi64(m6, m7); \
129 #define LOAD_MSG_3_3(b0, b1) \  argument
133 b1 = _mm_blend_epi16(m2, m7, 0xF0); \
137 #define LOAD_MSG_3_4(b0, b1) \  argument
141 b1 = _mm_unpacklo_epi64(m0, m4); \
145 #define LOAD_MSG_4_1(b0, b1) \  argument
149 b1 = _mm_unpacklo_epi64(m1, m5); \
153 #define LOAD_MSG_4_2(b0, b1) \  argument
157 b1 = _mm_blend_epi16(m2, m7, 0xF0); \
161 #define LOAD_MSG_4_3(b0, b1) \  argument
165 b1 = _mm_blend_epi16(m3, m1, 0xF0); \
169 #define LOAD_MSG_4_4(b0, b1) \  argument
173 b1 = _mm_blend_epi16(m4, m6, 0xF0); \
177 #define LOAD_MSG_5_1(b0, b1) \  argument
181 b1 = _mm_unpacklo_epi64(m0, m4); \
185 #define LOAD_MSG_5_2(b0, b1) \  argument
189 b1 = _mm_unpackhi_epi64(m5, m1); \
193 #define LOAD_MSG_5_3(b0, b1) \  argument
197 b1 = _mm_unpackhi_epi64(m7, m0); \
201 #define LOAD_MSG_5_4(b0, b1) \  argument
205 b1 = _mm_blend_epi16(m7, m4, 0xF0); \
209 #define LOAD_MSG_6_1(b0, b1) \  argument
213 b1 = _mm_unpacklo_epi64(m7, m2); \
217 #define LOAD_MSG_6_2(b0, b1) \  argument
221 b1 = _mm_alignr_epi8(m5, m6, 8); \
225 #define LOAD_MSG_6_3(b0, b1) \  argument
229 b1 = _mm_shuffle_epi32(m4, _MM_SHUFFLE(1,0,3,2)); \
233 #define LOAD_MSG_6_4(b0, b1) \  argument
237 b1 = _mm_blend_epi16(m1, m5, 0xF0); \
241 #define LOAD_MSG_7_1(b0, b1) \  argument
245 b1 = _mm_blend_epi16(m6, m1, 0xF0); \
249 #define LOAD_MSG_7_2(b0, b1) \  argument
253 b1 = _mm_unpackhi_epi64(m0, m4); \
257 #define LOAD_MSG_7_3(b0, b1) \  argument
261 b1 = _mm_unpacklo_epi64(m4, m1); \
265 #define LOAD_MSG_7_4(b0, b1) \  argument
269 b1 = _mm_unpacklo_epi64(m3, m5); \
273 #define LOAD_MSG_8_1(b0, b1) \  argument
277 b1 = _mm_alignr_epi8(m0, m5, 8); \
281 #define LOAD_MSG_8_2(b0, b1) \  argument
285 b1 = _mm_alignr_epi8(m4, m1, 8); \
289 #define LOAD_MSG_8_3(b0, b1) \  argument
293 b1 = _mm_alignr_epi8(m5, m0, 8); \
297 #define LOAD_MSG_8_4(b0, b1) \  argument
301 b1 = m2; \
305 #define LOAD_MSG_9_1(b0, b1) \  argument
309 b1 = _mm_unpackhi_epi64(m3, m0); \
313 #define LOAD_MSG_9_2(b0, b1) \  argument
317 b1 = _mm_blend_epi16(m3, m2, 0xF0); \
321 #define LOAD_MSG_9_3(b0, b1) \  argument
325 b1 = _mm_unpackhi_epi64(m1, m6); \
329 #define LOAD_MSG_9_4(b0, b1) \  argument
333 b1 = _mm_unpacklo_epi64(m6, m0); \
337 #define LOAD_MSG_10_1(b0, b1) \  argument
341 b1 = _mm_unpacklo_epi64(m2, m3); \
345 #define LOAD_MSG_10_2(b0, b1) \  argument
349 b1 = _mm_unpackhi_epi64(m2, m3); \
353 #define LOAD_MSG_10_3(b0, b1) \  argument
357 b1 = _mm_unpacklo_epi64(m6, m7); \
361 #define LOAD_MSG_10_4(b0, b1) \  argument
365 b1 = _mm_unpackhi_epi64(m6, m7); \
369 #define LOAD_MSG_11_1(b0, b1) \  argument
373 b1 = _mm_unpackhi_epi64(m4, m6); \
377 #define LOAD_MSG_11_2(b0, b1) \  argument
381 b1 = _mm_alignr_epi8(m3, m7, 8); \
385 #define LOAD_MSG_11_3(b0, b1) \  argument
389 b1 = _mm_unpackhi_epi64(m5, m2); \
393 #define LOAD_MSG_11_4(b0, b1) \  argument
397 b1 = _mm_unpackhi_epi64(m3, m1); \