886   __m256d            a0, a1, a2, a3, a4, a5, w0, w1, w2, w3, s0, s1, s2, v0, v1, v2, v3;  in MatSolve_SeqBAIJ_9_NaturalOrdering()  local
914       a2 = _mm256_loadu_pd(&v[8]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
915       s2 = _mm256_fnmadd_pd(a2, w0, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
930       a2 = _mm256_loadu_pd(&v[26]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
931       s2 = _mm256_fnmadd_pd(a2, w2, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
946       a2 = _mm256_loadu_pd(&v[44]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
947       s2 = _mm256_fnmadd_pd(a2, w0, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
962       a2 = _mm256_loadu_pd(&v[62]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
963       s2 = _mm256_fnmadd_pd(a2, w2, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
978       a2 = _mm256_maskload_pd(v + 80, _mm256_set_epi64x(0LL, 0LL, 0LL, 1LL << 63));  in MatSolve_SeqBAIJ_9_NaturalOrdering()
979       s2 = _mm256_fnmadd_pd(a2, w0, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1005       a2 = _mm256_loadu_pd(&v[8]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1006       s2 = _mm256_fnmadd_pd(a2, w0, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1023       a2 = _mm256_loadu_pd(&v[26]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1024       s2 = _mm256_fnmadd_pd(a2, w2, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1041       a2 = _mm256_loadu_pd(&v[44]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1042       s2 = _mm256_fnmadd_pd(a2, w0, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1059       a2 = _mm256_loadu_pd(&v[62]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1060       s2 = _mm256_fnmadd_pd(a2, w2, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1077       a2 = _mm256_maskload_pd(v + 80, _mm256_set_epi64x(0LL, 0LL, 0LL, 1LL << 63));  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1078       s2 = _mm256_fnmadd_pd(a2, w0, s2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1096     a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[8]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1097     w2 = _mm256_fmadd_pd(a2, v0, w2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1114     a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[26]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1115     w2 = _mm256_fmadd_pd(a2, v2, w2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1132     a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[44]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1133     w2 = _mm256_fmadd_pd(a2, v0, w2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1150     a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[62]);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1151     w2 = _mm256_fmadd_pd(a2, v2, w2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1168 …  a2 = _mm256_maskload_pd(&(aa + bs2 * adiag[i])[80], _mm256_set_epi64x(0LL, 0LL, 0LL, 1LL << 63));  in MatSolve_SeqBAIJ_9_NaturalOrdering()
1169     w2 = _mm256_fmadd_pd(a2, v0, w2);  in MatSolve_SeqBAIJ_9_NaturalOrdering()