19 #define LOAD_MSG_0_1(b0, b1) \  argument
23 b1 = _mm_unpacklo_epi64(m2, m3); \
27 #define LOAD_MSG_0_2(b0, b1) \  argument
31 b1 = _mm_unpackhi_epi64(m2, m3); \
35 #define LOAD_MSG_0_3(b0, b1) \  argument
39 b1 = _mm_unpacklo_epi64(m6, m7); \
43 #define LOAD_MSG_0_4(b0, b1) \  argument
47 b1 = _mm_unpackhi_epi64(m6, m7); \
51 #define LOAD_MSG_1_1(b0, b1) \  argument
55 b1 = _mm_unpackhi_epi64(m4, m6); \
59 #define LOAD_MSG_1_2(b0, b1) \  argument
63 b1 = _mm_alignr_epi8(m3, m7, 8); \
67 #define LOAD_MSG_1_3(b0, b1) \  argument
71 b1 = _mm_unpackhi_epi64(m5, m2); \
75 #define LOAD_MSG_1_4(b0, b1) \  argument
79 b1 = _mm_unpackhi_epi64(m3, m1); \
83 #define LOAD_MSG_2_1(b0, b1) \  argument
87 b1 = _mm_unpackhi_epi64(m2, m7); \
91 #define LOAD_MSG_2_2(b0, b1) \  argument
95 b1 = _mm_blend_epi16(m1, m6, 0xF0); \
99 #define LOAD_MSG_2_3(b0, b1) \  argument
103 b1 = _mm_unpackhi_epi64(m3, m4); \
107 #define LOAD_MSG_2_4(b0, b1) \  argument
111 b1 = _mm_alignr_epi8(m2, m0, 8); \
115 #define LOAD_MSG_3_1(b0, b1) \  argument
119 b1 = _mm_unpackhi_epi64(m6, m5); \
123 #define LOAD_MSG_3_2(b0, b1) \  argument
127 b1 = _mm_unpacklo_epi64(m6, m7); \
131 #define LOAD_MSG_3_3(b0, b1) \  argument
135 b1 = _mm_blend_epi16(m2, m7, 0xF0); \
139 #define LOAD_MSG_3_4(b0, b1) \  argument
143 b1 = _mm_unpacklo_epi64(m0, m4); \
147 #define LOAD_MSG_4_1(b0, b1) \  argument
151 b1 = _mm_unpacklo_epi64(m1, m5); \
155 #define LOAD_MSG_4_2(b0, b1) \  argument
159 b1 = _mm_blend_epi16(m2, m7, 0xF0); \
163 #define LOAD_MSG_4_3(b0, b1) \  argument
167 b1 = _mm_blend_epi16(m3, m1, 0xF0); \
171 #define LOAD_MSG_4_4(b0, b1) \  argument
175 b1 = _mm_blend_epi16(m4, m6, 0xF0); \
179 #define LOAD_MSG_5_1(b0, b1) \  argument
183 b1 = _mm_unpacklo_epi64(m0, m4); \
187 #define LOAD_MSG_5_2(b0, b1) \  argument
191 b1 = _mm_unpackhi_epi64(m5, m1); \
195 #define LOAD_MSG_5_3(b0, b1) \  argument
199 b1 = _mm_unpackhi_epi64(m7, m0); \
203 #define LOAD_MSG_5_4(b0, b1) \  argument
207 b1 = _mm_blend_epi16(m7, m4, 0xF0); \
211 #define LOAD_MSG_6_1(b0, b1) \  argument
215 b1 = _mm_unpacklo_epi64(m7, m2); \
219 #define LOAD_MSG_6_2(b0, b1) \  argument
223 b1 = _mm_alignr_epi8(m5, m6, 8); \
227 #define LOAD_MSG_6_3(b0, b1) \  argument
231 b1 = _mm_shuffle_epi32(m4, _MM_SHUFFLE(1,0,3,2)); \
235 #define LOAD_MSG_6_4(b0, b1) \  argument
239 b1 = _mm_blend_epi16(m1, m5, 0xF0); \
243 #define LOAD_MSG_7_1(b0, b1) \  argument
247 b1 = _mm_blend_epi16(m6, m1, 0xF0); \
251 #define LOAD_MSG_7_2(b0, b1) \  argument
255 b1 = _mm_unpackhi_epi64(m0, m4); \
259 #define LOAD_MSG_7_3(b0, b1) \  argument
263 b1 = _mm_unpacklo_epi64(m4, m1); \
267 #define LOAD_MSG_7_4(b0, b1) \  argument
271 b1 = _mm_unpacklo_epi64(m3, m5); \
275 #define LOAD_MSG_8_1(b0, b1) \  argument
279 b1 = _mm_alignr_epi8(m0, m5, 8); \
283 #define LOAD_MSG_8_2(b0, b1) \  argument
287 b1 = _mm_alignr_epi8(m4, m1, 8); \
291 #define LOAD_MSG_8_3(b0, b1) \  argument
295 b1 = _mm_alignr_epi8(m5, m0, 8); \
299 #define LOAD_MSG_8_4(b0, b1) \  argument
303 b1 = m2; \
307 #define LOAD_MSG_9_1(b0, b1) \  argument
311 b1 = _mm_unpackhi_epi64(m3, m0); \
315 #define LOAD_MSG_9_2(b0, b1) \  argument
319 b1 = _mm_blend_epi16(m3, m2, 0xF0); \
323 #define LOAD_MSG_9_3(b0, b1) \  argument
327 b1 = _mm_unpackhi_epi64(m1, m6); \
331 #define LOAD_MSG_9_4(b0, b1) \  argument
335 b1 = _mm_unpacklo_epi64(m6, m0); \
339 #define LOAD_MSG_10_1(b0, b1) \  argument
343 b1 = _mm_unpacklo_epi64(m2, m3); \
347 #define LOAD_MSG_10_2(b0, b1) \  argument
351 b1 = _mm_unpackhi_epi64(m2, m3); \
355 #define LOAD_MSG_10_3(b0, b1) \  argument
359 b1 = _mm_unpacklo_epi64(m6, m7); \
363 #define LOAD_MSG_10_4(b0, b1) \  argument
367 b1 = _mm_unpackhi_epi64(m6, m7); \
371 #define LOAD_MSG_11_1(b0, b1) \  argument
375 b1 = _mm_unpackhi_epi64(m4, m6); \
379 #define LOAD_MSG_11_2(b0, b1) \  argument
383 b1 = _mm_alignr_epi8(m3, m7, 8); \
387 #define LOAD_MSG_11_3(b0, b1) \  argument
391 b1 = _mm_unpackhi_epi64(m5, m2); \
395 #define LOAD_MSG_11_4(b0, b1) \  argument
399 b1 = _mm_unpackhi_epi64(m3, m1); \