58 } m256d;  typedef
60 typedef m256d v4sf;
86 FORCE_INLINE m256d mm256_setzero_pd(void)  in mm256_setzero_pd()
88     m256d ret;  in mm256_setzero_pd()
93 FORCE_INLINE m256d mm256_mul_pd(m256d a, m256d b)  in mm256_mul_pd()
95     m256d ret;  in mm256_mul_pd()
101 FORCE_INLINE m256d mm256_add_pd(m256d a, m256d b)  in mm256_add_pd()
103     m256d ret;  in mm256_add_pd()
109 FORCE_INLINE m256d mm256_sub_pd(m256d a, m256d b)  in mm256_sub_pd()
111     m256d ret;  in mm256_sub_pd()
117 FORCE_INLINE m256d mm256_set1_pd(double a)  in mm256_set1_pd()
119     m256d ret;  in mm256_set1_pd()
124 FORCE_INLINE m256d mm256_load_pd (double const * mem_addr)  in mm256_load_pd()
126     m256d res;  in mm256_load_pd()
131 FORCE_INLINE m256d mm256_loadu_pd (double const * mem_addr)  in mm256_loadu_pd()
133     m256d res;  in mm256_loadu_pd()
152 FORCE_INLINE __m128d mm256_castpd256_pd128(m256d a)  in mm256_castpd256_pd128()
157 FORCE_INLINE __m128d mm256_extractf128_pd (m256d a, const int imm8)  in mm256_extractf128_pd()
162 FORCE_INLINE m256d mm256_insertf128_pd_1(m256d a, __m128d b)  in mm256_insertf128_pd_1()
164     m256d res;  in mm256_insertf128_pd_1()
169 FORCE_INLINE m256d mm256_castpd128_pd256(__m128d a)  in mm256_castpd128_pd256()
171     m256d res;  in mm256_castpd128_pd256()
176 FORCE_INLINE m256d mm256_shuffle_pd_00(m256d a, m256d b)  in mm256_shuffle_pd_00()
178     m256d res;  in mm256_shuffle_pd_00()
184 FORCE_INLINE m256d mm256_shuffle_pd_11(m256d a, m256d b)  in mm256_shuffle_pd_11()
186     m256d res;  in mm256_shuffle_pd_11()
192 FORCE_INLINE m256d mm256_permute2f128_pd_0x20(m256d a, m256d b) {  in mm256_permute2f128_pd_0x20()
193     m256d res;  in mm256_permute2f128_pd_0x20()
200 FORCE_INLINE m256d mm256_permute2f128_pd_0x31(m256d a, m256d b)  in mm256_permute2f128_pd_0x31()
202     m256d res;  in mm256_permute2f128_pd_0x31()
208 FORCE_INLINE m256d mm256_reverse(m256d x)  in mm256_reverse()
210     m256d res;  in mm256_reverse()
225 	m256d tmp__ = mm256_insertf128_pd_1(								\
243 	m256d tmp__ = mm256_insertf128_pd_1(								\
253         m256d tmp3, tmp2, tmp1, tmp0;                     			\