__lfpd             69 libfftss/r4_bgl_n.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd             70 libfftss/r4_bgl_n.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd             71 libfftss/r4_bgl_n.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd             77 libfftss/r4_bgl_n.c       p2 = __lfpd(i2 + 2 * j);
__lfpd             78 libfftss/r4_bgl_n.c       p1 = __lfpd(i1 + 2 * j);
__lfpd             79 libfftss/r4_bgl_n.c       p3 = __lfpd(i3 + 2 * j);
__lfpd             80 libfftss/r4_bgl_n.c       p0 = __lfpd(i0 + 2 * j);
__lfpd            125 libfftss/r4_bgl_n.c     w2 = __lfpd(wn + 4 * i * bsize);
__lfpd            126 libfftss/r4_bgl_n.c     w1 = __lfpd(wn + 2 * i * bsize);
__lfpd            127 libfftss/r4_bgl_n.c     w3 = __lfpd(wn2 + 2 * i * bsize);
__lfpd            133 libfftss/r4_bgl_n.c       p2 = __lfpd(i2 + 2 * j);
__lfpd            134 libfftss/r4_bgl_n.c       p1 = __lfpd(i1 + 2 * j);
__lfpd            135 libfftss/r4_bgl_n.c       p3 = __lfpd(i3 + 2 * j);
__lfpd            136 libfftss/r4_bgl_n.c       p0 = __lfpd(i0 + 2 * j);
__lfpd             58 libfftss/r4_bgl_o.c     p0 = __lfpd(i0 + 2 * j);
__lfpd             59 libfftss/r4_bgl_o.c     p2 = __lfpd(i2 + 2 * j);
__lfpd             60 libfftss/r4_bgl_o.c     p1 = __lfpd(i1 + 2 * j);
__lfpd             61 libfftss/r4_bgl_o.c     p3 = __lfpd(i3 + 2 * j);
__lfpd             82 libfftss/r4_bgl_o.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd             83 libfftss/r4_bgl_o.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd             84 libfftss/r4_bgl_o.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd             90 libfftss/r4_bgl_o.c       p2 = __lfpd(i2 + 2 * j);
__lfpd             91 libfftss/r4_bgl_o.c       p1 = __lfpd(i1 + 2 * j);
__lfpd             92 libfftss/r4_bgl_o.c       p3 = __lfpd(i3 + 2 * j);
__lfpd             93 libfftss/r4_bgl_o.c       p0 = __lfpd(i0 + 2 * j);
__lfpd            140 libfftss/r4_bgl_o.c     p0 = __lfpd(i0 + 2 * j);
__lfpd            141 libfftss/r4_bgl_o.c     p2 = __lfpd(i2 + 2 * j);
__lfpd            142 libfftss/r4_bgl_o.c     p1 = __lfpd(i1 + 2 * j);
__lfpd            143 libfftss/r4_bgl_o.c     p3 = __lfpd(i3 + 2 * j);
__lfpd            164 libfftss/r4_bgl_o.c     w2 = __lfpd(wn + 4 * i * bsize);
__lfpd            165 libfftss/r4_bgl_o.c     w1 = __lfpd(wn + 2 * i * bsize);
__lfpd            166 libfftss/r4_bgl_o.c     w3 = __lfpd(wn2 + 2 * i * bsize);
__lfpd            172 libfftss/r4_bgl_o.c       p2 = __lfpd(i2 + 2 * j);
__lfpd            173 libfftss/r4_bgl_o.c       p1 = __lfpd(i1 + 2 * j);
__lfpd            174 libfftss/r4_bgl_o.c       p3 = __lfpd(i3 + 2 * j);
__lfpd            175 libfftss/r4_bgl_o.c       p0 = __lfpd(i0 + 2 * j);
__lfpd             78 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2);
__lfpd             79 libfftss/r4_bgl_pl_n.c   f17 = __lfpd(w2);
__lfpd             80 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3);
__lfpd             81 libfftss/r4_bgl_pl_n.c   f18 = __lfpd(w3);
__lfpd             82 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1);        
__lfpd             83 libfftss/r4_bgl_pl_n.c   f16 = __lfpd(w1);              f6 = __fxcxnpma(f2,f2,__creal(f17));
__lfpd             85 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0);              f7 = __fxcxnpma(f3,f3,__creal(f18));
__lfpd             86 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 2);          f5 = __fxcxnpma(f1,f1,__creal(f16));
__lfpd             88 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 2);
__lfpd             90 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 2);
__lfpd             97 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 2);          f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd             98 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 4);          f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            100 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 4);          f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            102 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 4);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            109 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 4);          f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            110 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 6);          f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            111 libfftss/r4_bgl_pl_n.c   f20 = __lfpd(w2);              f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            112 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 6);          f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            113 libfftss/r4_bgl_pl_n.c   f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            114 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 6);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            115 libfftss/r4_bgl_pl_n.c   f19 = __lfpd(w1);              f8  = __fxcxnsma(f0,f6,__cimag(f17));
__lfpd            128 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i - 2);  f7 = __fxcxnpma(f3,f3,__creal(f21));
__lfpd            136 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i);      f5  = __fxcxnpma(f1,f1,__creal(f19));
__lfpd            137 libfftss/r4_bgl_pl_n.c     f17 = __lfpd(w2);              f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            138 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i);      f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            139 libfftss/r4_bgl_pl_n.c     f18 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            140 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i);      f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            141 libfftss/r4_bgl_pl_n.c     f16 = __lfpd(w1);              f8  = __fxcxnsma(f0,f6,__cimag(f20));
__lfpd            147 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i);      f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            148 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 2);  f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            150 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 2);  f13 = __fxcxnpma(f8,f9,__cimag(f19));
__lfpd            152 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 2);  f15 = __fxcsmadd(f10,f11,__cimag(f19));
__lfpd            159 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 2);  f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            160 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 4);  f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            162 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 4);  f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            164 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 4);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            171 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 4);  f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            172 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 6);  f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            173 libfftss/r4_bgl_pl_n.c     f20 = __lfpd(w2);              f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            174 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 6);  f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            175 libfftss/r4_bgl_pl_n.c     f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            176 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 6);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            177 libfftss/r4_bgl_pl_n.c     f19 = __lfpd(w1);              f8  = __fxcxnsma(f0,f6,__cimag(f17));
__lfpd            192 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 - 2);           f7 = __fxcxnpma(f3,f3,__creal(f21));
__lfpd            250 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2);
__lfpd            251 libfftss/r4_bgl_pl_n.c   f17 = __lfpd(w2);
__lfpd            252 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3);
__lfpd            253 libfftss/r4_bgl_pl_n.c   f18 = __lfpd(w3);
__lfpd            254 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1);        
__lfpd            255 libfftss/r4_bgl_pl_n.c   f16 = __lfpd(w1);              f6 = __fxcxnsma(f2,f2,__creal(f17));
__lfpd            257 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0);              f7 = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            258 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 2);          f5 = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            260 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 2);
__lfpd            262 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 2);
__lfpd            269 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 2);          f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            270 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 4);          f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            272 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 4);          f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            274 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 4);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            281 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 4);          f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            282 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 6);          f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            283 libfftss/r4_bgl_pl_n.c   f20 = __lfpd(w2);              f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            284 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 6);          f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            285 libfftss/r4_bgl_pl_n.c   f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            286 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 6);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            287 libfftss/r4_bgl_pl_n.c   f19 = __lfpd(w1);              f8  = __fxcxnpma(f0,f6,__cimag(f17));
__lfpd            301 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i - 2);  f7  = __fxcxnsma(f3,f3,__creal(f21));
__lfpd            309 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i);      f5  = __fxcxnsma(f1,f1,__creal(f19));
__lfpd            310 libfftss/r4_bgl_pl_n.c     f17 = __lfpd(w2);              f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            311 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i);      f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            312 libfftss/r4_bgl_pl_n.c     f18 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            313 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i);      f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            314 libfftss/r4_bgl_pl_n.c     f16 = __lfpd(w1);              f8  = __fxcxnpma(f0,f6,__cimag(f20));
__lfpd            320 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i);      f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            321 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 2);  f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            323 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 2);  f13 = __fxcxnsma(f8,f9,__cimag(f19));
__lfpd            325 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 2);  f15 = __fxcsmadd(f10,f11,__cimag(f19));
__lfpd            332 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 2);  f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            333 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 4);  f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            335 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 4);  f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            337 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 4);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            344 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 4);  f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            345 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 6);  f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            346 libfftss/r4_bgl_pl_n.c     f20 = __lfpd(w2);              f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            347 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 6);  f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            348 libfftss/r4_bgl_pl_n.c     f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            349 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 6);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            350 libfftss/r4_bgl_pl_n.c     f19 = __lfpd(w1);              f8  = __fxcxnpma(f0,f6,__cimag(f17));
__lfpd            365 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 - 2);           f7 = __fxcxnsma(f3,f3,__creal(f21));
__lfpd            418 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2);
__lfpd            419 libfftss/r4_bgl_pl_n.c   f17 = __lfpd(w2);
__lfpd            420 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3);
__lfpd            421 libfftss/r4_bgl_pl_n.c   f18 = __lfpd(w3);
__lfpd            422 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1);        
__lfpd            423 libfftss/r4_bgl_pl_n.c   f16 = __lfpd(w1);              f6 = __fxcxnpma(f2,f2,__creal(f17));
__lfpd            425 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0);              f7 = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            426 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 2);          f5 = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            428 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 2);
__lfpd            430 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 2);
__lfpd            437 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 2);          f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            438 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 4);          f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            440 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 4);          f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            442 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 4);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            449 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 4);          f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            450 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 6);          f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            451 libfftss/r4_bgl_pl_n.c   f20 = __lfpd(w2);              f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            452 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 6);          f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            453 libfftss/r4_bgl_pl_n.c   f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            454 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 6);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            455 libfftss/r4_bgl_pl_n.c   f19 = __lfpd(w1);              f8  = __fxcxnsma(f0,f6,__cimag(f17));
__lfpd            467 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i - 2);  f7 = __fxcxnpma(f3,f3,__creal(f21));
__lfpd            475 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i);      f5  = __fxcxnpma(f1,f1,__creal(f19));
__lfpd            476 libfftss/r4_bgl_pl_n.c     f17 = __lfpd(w2);              f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            477 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i);      f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            478 libfftss/r4_bgl_pl_n.c     f18 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            479 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i);      f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            480 libfftss/r4_bgl_pl_n.c     f16 = __lfpd(w1);              f8  = __fxcxnsma(f0,f6,__cimag(f20));
__lfpd            486 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i);      f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            487 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 2);  f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            489 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 2);  f13 = __fxcxnpma(f8,f9,__cimag(f19));
__lfpd            491 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 2);  f15 = __fxcsmadd(f10,f11,__cimag(f19));
__lfpd            498 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 2);  f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            499 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 4);  f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            501 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 4);  f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            503 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 4);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            510 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 4);  f7  = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            511 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 6);  f5  = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            512 libfftss/r4_bgl_pl_n.c     f20 = __lfpd(w2);              f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            513 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 6);  f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            514 libfftss/r4_bgl_pl_n.c     f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            515 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 6);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            516 libfftss/r4_bgl_pl_n.c     f19 = __lfpd(w1);              f8  = __fxcxnsma(f0,f6,__cimag(f17));
__lfpd            531 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 - 2);           f7 = __fxcxnpma(f3,f3,__creal(f21));
__lfpd            585 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2);
__lfpd            586 libfftss/r4_bgl_pl_n.c   f17 = __lfpd(w2);
__lfpd            587 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3);
__lfpd            588 libfftss/r4_bgl_pl_n.c   f18 = __lfpd(w3);
__lfpd            589 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1);        
__lfpd            590 libfftss/r4_bgl_pl_n.c   f16 = __lfpd(w1);              f6 = __fxcxnsma(f2,f2,__creal(f17));
__lfpd            592 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0);              f7 = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            593 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 2);          f5 = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            595 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 2);
__lfpd            597 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 2);
__lfpd            604 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 2);          f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            605 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 4);          f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            607 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 4);          f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            609 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 4);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            616 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 + 4);          f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            617 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi2 + 6);          f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            618 libfftss/r4_bgl_pl_n.c   f20 = __lfpd(w2);              f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            619 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi3 + 6);          f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            620 libfftss/r4_bgl_pl_n.c   f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            621 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi1 + 6);          f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            622 libfftss/r4_bgl_pl_n.c   f19 = __lfpd(w1);              f8  = __fxcxnpma(f0,f6,__cimag(f17));
__lfpd            634 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i - 2);  f7  = __fxcxnsma(f3,f3,__creal(f21));
__lfpd            642 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i);      f5  = __fxcxnsma(f1,f1,__creal(f19));
__lfpd            643 libfftss/r4_bgl_pl_n.c     f17 = __lfpd(w2);              f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            644 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i);      f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            645 libfftss/r4_bgl_pl_n.c     f18 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            646 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i);      f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            647 libfftss/r4_bgl_pl_n.c     f16 = __lfpd(w1);              f8  = __fxcxnpma(f0,f6,__cimag(f20));
__lfpd            653 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i);      f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            654 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 2);  f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            656 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 2);  f13 = __fxcxnsma(f8,f9,__cimag(f19));
__lfpd            658 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 2);  f15 = __fxcsmadd(f10,f11,__cimag(f19));
__lfpd            665 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 2);  f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            666 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 4);  f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            668 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 4);  f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            670 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 4);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            677 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi0 + 8 * i + 4);  f7  = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            678 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi2 + 8 * i + 6);  f5  = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            679 libfftss/r4_bgl_pl_n.c     f20 = __lfpd(w2);              f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            680 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi3 + 8 * i + 6);  f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            681 libfftss/r4_bgl_pl_n.c     f21 = __lfpd(w3);              f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            682 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi1 + 8 * i + 6);  f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            683 libfftss/r4_bgl_pl_n.c     f19 = __lfpd(w1);              f8  = __fxcxnpma(f0,f6,__cimag(f17));
__lfpd            698 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi0 - 2);           f7 = __fxcxnsma(f3,f3,__creal(f21));
__lfpd            754 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi + 4);
__lfpd            755 libfftss/r4_bgl_pl_n.c   f17 = __lfpd(w2); w2 += 4;
__lfpd            756 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi + 6);
__lfpd            757 libfftss/r4_bgl_pl_n.c   f18 = __lfpd(w3); w3 += 2;
__lfpd            758 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi + 2);        
__lfpd            759 libfftss/r4_bgl_pl_n.c   f16 = __lfpd(w1); w1 += 2;  f6 = __fxcxnpma(f2,f2,__creal(f17));
__lfpd            761 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi);            f7 = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            762 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi + 12);       f5 = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            763 libfftss/r4_bgl_pl_n.c   f20 = __lfpd(w2); w2 += 4;  
__lfpd            764 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi + 14);
__lfpd            765 libfftss/r4_bgl_pl_n.c   f21 = __lfpd(w3); w3 += 2;
__lfpd            766 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi + 10);
__lfpd            767 libfftss/r4_bgl_pl_n.c   f19 = __lfpd(w1); w1 += 2;  f8 = __fxcxnsma(f0,f6,__cimag(f17));
__lfpd            776 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi + 16 * i - 8); f7 = __fxcxnpma(f3,f3,__creal(f21));
__lfpd            777 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi + 16 * i + 4); f5 = __fxcxnpma(f1,f1,__creal(f19));
__lfpd            778 libfftss/r4_bgl_pl_n.c     f17 = __lfpd(w2 + 8 * i);     f12 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            779 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi + 16 * i + 6); f13 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            780 libfftss/r4_bgl_pl_n.c     f18 = __lfpd(w3 + 4 * i);     f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            781 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi + 16 * i + 2); f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            782 libfftss/r4_bgl_pl_n.c     f16 = __lfpd(w1 + 4 * i);     f8 = __fxcxnsma(f0,f6,__cimag(f20));
__lfpd            788 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi + 16 * i); f7 = __fxcxnpma(f3,f3,__creal(f18));
__lfpd            789 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi + 16 * i + 12); f5 = __fxcxnpma(f1,f1,__creal(f16));
__lfpd            790 libfftss/r4_bgl_pl_n.c     f20 = __lfpd(w2 + 8 * i + 4);  f12 = __fxcxnsma(f8,f9,__cimag(f19));
__lfpd            791 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi + 16 * i + 14); f13 = __fxcxnpma(f8,f9,__cimag(f19));
__lfpd            792 libfftss/r4_bgl_pl_n.c     f21 = __lfpd(w3 + 4 * i + 2);  f14 = __fxcsnmsub(f10,f11,__cimag(f19));
__lfpd            793 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi + 16 * i + 10); f15 = __fxcsmadd(f10,f11,__cimag(f19));
__lfpd            794 libfftss/r4_bgl_pl_n.c     f19 = __lfpd(w1 + 4 * i + 2);  f8 = __fxcxnsma(f0,f6,__cimag(f17));
__lfpd            806 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi - 8);          f7 = __fxcxnpma(f3,f3,__creal(f21));
__lfpd            861 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi + 4);
__lfpd            862 libfftss/r4_bgl_pl_n.c   f17 = __lfpd(w2); w2 += 4;
__lfpd            863 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi + 6);
__lfpd            864 libfftss/r4_bgl_pl_n.c   f18 = __lfpd(w3); w3 += 2;
__lfpd            865 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi + 2);        
__lfpd            866 libfftss/r4_bgl_pl_n.c   f16 = __lfpd(w1); w1 += 2;  f6 = __fxcxnsma(f2,f2,__creal(f17));
__lfpd            868 libfftss/r4_bgl_pl_n.c   f0 = __lfpd(vi);            f7 = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            869 libfftss/r4_bgl_pl_n.c   f2 = __lfpd(vi + 12);       f5 = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            870 libfftss/r4_bgl_pl_n.c   f20 = __lfpd(w2); w2 += 4;  
__lfpd            871 libfftss/r4_bgl_pl_n.c   f3 = __lfpd(vi + 14);
__lfpd            872 libfftss/r4_bgl_pl_n.c   f21 = __lfpd(w3); w3 += 2;
__lfpd            873 libfftss/r4_bgl_pl_n.c   f1 = __lfpd(vi + 10);
__lfpd            874 libfftss/r4_bgl_pl_n.c   f19 = __lfpd(w1); w1 += 2;  f8 = __fxcxnpma(f0,f6,__cimag(f17));
__lfpd            883 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi + 16 * i - 8); f7 = __fxcxnsma(f3,f3,__creal(f21));
__lfpd            884 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi + 16 * i + 4); f5 = __fxcxnsma(f1,f1,__creal(f19));
__lfpd            885 libfftss/r4_bgl_pl_n.c     f17 = __lfpd(w2 + 8 * i);     f12 = __fxcxnpma(f8,f9,__cimag(f16));
__lfpd            886 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi + 16 * i + 6); f13 = __fxcxnsma(f8,f9,__cimag(f16));
__lfpd            887 libfftss/r4_bgl_pl_n.c     f18 = __lfpd(w3 + 4 * i);     f14 = __fxcsnmsub(f10,f11,__cimag(f16));
__lfpd            888 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi + 16 * i + 2); f15 = __fxcsmadd(f10,f11,__cimag(f16));
__lfpd            889 libfftss/r4_bgl_pl_n.c     f16 = __lfpd(w1 + 4 * i);     f8 = __fxcxnpma(f0,f6,__cimag(f20));
__lfpd            895 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi + 16 * i); f7 = __fxcxnsma(f3,f3,__creal(f18));
__lfpd            896 libfftss/r4_bgl_pl_n.c     f2 = __lfpd(vi + 16 * i + 12); f5 = __fxcxnsma(f1,f1,__creal(f16));
__lfpd            897 libfftss/r4_bgl_pl_n.c     f20 = __lfpd(w2 + 8 * i + 4);  f12 = __fxcxnpma(f8,f9,__cimag(f19));
__lfpd            898 libfftss/r4_bgl_pl_n.c     f3 = __lfpd(vi + 16 * i + 14); f13 = __fxcxnsma(f8,f9,__cimag(f19));
__lfpd            899 libfftss/r4_bgl_pl_n.c     f21 = __lfpd(w3 + 4 * i + 2);  f14 = __fxcsnmsub(f10,f11,__cimag(f19));
__lfpd            900 libfftss/r4_bgl_pl_n.c     f1 = __lfpd(vi + 16 * i + 10); f15 = __fxcsmadd(f10,f11,__cimag(f19));
__lfpd            901 libfftss/r4_bgl_pl_n.c     f19 = __lfpd(w1 + 4 * i + 2);  f8 = __fxcxnpma(f0,f6,__cimag(f17));
__lfpd            912 libfftss/r4_bgl_pl_n.c     f0 = __lfpd(vi - 8);          f7 = __fxcxnsma(f3,f3,__creal(f21));
__lfpd             70 libfftss/r4_bgl_u1.c     p2 = __lfpd(i2 + 8 * i);
__lfpd             71 libfftss/r4_bgl_u1.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd             72 libfftss/r4_bgl_u1.c     p1 = __lfpd(i1 + 8 * i);
__lfpd             73 libfftss/r4_bgl_u1.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd             74 libfftss/r4_bgl_u1.c     p3 = __lfpd(i3 + 8 * i);
__lfpd             75 libfftss/r4_bgl_u1.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd             76 libfftss/r4_bgl_u1.c     p0 = __lfpd(i0 + 8 * i);
__lfpd            128 libfftss/r4_bgl_u1.c     p2 = __lfpd(i2 + 8 * i);
__lfpd            129 libfftss/r4_bgl_u1.c     w2 = __lfpd(wn + 4 * i * bsize);
__lfpd            130 libfftss/r4_bgl_u1.c     p1 = __lfpd(i1 + 8 * i);
__lfpd            131 libfftss/r4_bgl_u1.c     w1 = __lfpd(wn + 2 * i * bsize);
__lfpd            132 libfftss/r4_bgl_u1.c     p3 = __lfpd(i3 + 8 * i);
__lfpd            133 libfftss/r4_bgl_u1.c     w3 = __lfpd(wn2 + 2 * i * bsize);
__lfpd            134 libfftss/r4_bgl_u1.c     p0 = __lfpd(i0 + 8 * i);
__lfpd             61 libfftss/r8_bgl_n.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd             62 libfftss/r8_bgl_n.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd             63 libfftss/r8_bgl_n.c     w4 = __lfpd(wn + i * bsize * 8);
__lfpd             64 libfftss/r8_bgl_n.c     w5 = __lfpd(wn + i * bsize * 2 + 2 * blocks * bsize);
__lfpd             65 libfftss/r8_bgl_n.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd             66 libfftss/r8_bgl_n.c     w6 = __lfpd(wn2 + i * bsize * 4);
__lfpd             67 libfftss/r8_bgl_n.c     w7 = __lfpd(wn2 + i * bsize * 2 + blocks * bsize * 2);
__lfpd             75 libfftss/r8_bgl_n.c       p5 = __lfpd(i5 + 2 * j);
__lfpd             76 libfftss/r8_bgl_n.c       p7 = __lfpd(i7 + 2 * j);
__lfpd             77 libfftss/r8_bgl_n.c       p4 = __lfpd(i4 + 2 * j);
__lfpd             78 libfftss/r8_bgl_n.c       p2 = __lfpd(i2 + 2 * j);
__lfpd             79 libfftss/r8_bgl_n.c       p6 = __lfpd(i6 + 2 * j);
__lfpd             81 libfftss/r8_bgl_n.c       p1 = __lfpd(i1 + 2 * j);
__lfpd             83 libfftss/r8_bgl_n.c       p3 = __lfpd(i3 + 2 * j);
__lfpd             85 libfftss/r8_bgl_n.c       p0 = __lfpd(i0 + 2 * j);
__lfpd            169 libfftss/r8_bgl_n.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd            170 libfftss/r8_bgl_n.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd            171 libfftss/r8_bgl_n.c     w4 = __lfpd(wn + i * bsize * 8);
__lfpd            172 libfftss/r8_bgl_n.c     w5 = __lfpd(wn + i * bsize * 2 + 2 * blocks * bsize);
__lfpd            173 libfftss/r8_bgl_n.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd            174 libfftss/r8_bgl_n.c     w6 = __lfpd(wn2 + i * bsize * 4);
__lfpd            175 libfftss/r8_bgl_n.c     w7 = __lfpd(wn2 + i * bsize * 2 + blocks * bsize * 2);
__lfpd            184 libfftss/r8_bgl_n.c       p5 = __lfpd(i5 + 2 * j);
__lfpd            185 libfftss/r8_bgl_n.c       p7 = __lfpd(i7 + 2 * j);
__lfpd            186 libfftss/r8_bgl_n.c       p4 = __lfpd(i4 + 2 * j);
__lfpd            187 libfftss/r8_bgl_n.c       p2 = __lfpd(i2 + 2 * j);
__lfpd            188 libfftss/r8_bgl_n.c       p6 = __lfpd(i6 + 2 * j);
__lfpd            190 libfftss/r8_bgl_n.c       p1 = __lfpd(i1 + 2 * j);
__lfpd            192 libfftss/r8_bgl_n.c       p3 = __lfpd(i3 + 2 * j);
__lfpd            194 libfftss/r8_bgl_n.c       p0 = __lfpd(i0 + 2 * j);
__lfpd             67 libfftss/r8_bgl_o.c     p1 = __lfpd(i1 + 2 * j);
__lfpd             68 libfftss/r8_bgl_o.c     p5 = __lfpd(i5 + 2 * j);
__lfpd             69 libfftss/r8_bgl_o.c     p3 = __lfpd(i3 + 2 * j);
__lfpd             70 libfftss/r8_bgl_o.c     p7 = __lfpd(i7 + 2 * j);
__lfpd             72 libfftss/r8_bgl_o.c     p0 = __lfpd(i0 + 2 * j);
__lfpd             73 libfftss/r8_bgl_o.c     p4 = __lfpd(i4 + 2 * j);
__lfpd             75 libfftss/r8_bgl_o.c     p2 = __lfpd(i2 + 2 * j);
__lfpd             77 libfftss/r8_bgl_o.c     p6 = __lfpd(i6 + 2 * j);
__lfpd            128 libfftss/r8_bgl_o.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd            129 libfftss/r8_bgl_o.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd            130 libfftss/r8_bgl_o.c     w4 = __lfpd(wn + i * bsize * 8);
__lfpd            131 libfftss/r8_bgl_o.c     w5 = __lfpd(wn + i * bsize * 2 + 2 * blocks * bsize);
__lfpd            132 libfftss/r8_bgl_o.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd            133 libfftss/r8_bgl_o.c     w6 = __lfpd(wn2 + i * bsize * 4);
__lfpd            134 libfftss/r8_bgl_o.c     w7 = __lfpd(wn2 + i * bsize * 2 + blocks * bsize * 2);
__lfpd            142 libfftss/r8_bgl_o.c       p5 = __lfpd(i5 + 2 * j);
__lfpd            143 libfftss/r8_bgl_o.c       p7 = __lfpd(i7 + 2 * j);
__lfpd            144 libfftss/r8_bgl_o.c       p4 = __lfpd(i4 + 2 * j);
__lfpd            145 libfftss/r8_bgl_o.c       p2 = __lfpd(i2 + 2 * j);
__lfpd            146 libfftss/r8_bgl_o.c       p6 = __lfpd(i6 + 2 * j);
__lfpd            148 libfftss/r8_bgl_o.c       p1 = __lfpd(i1 + 2 * j);
__lfpd            150 libfftss/r8_bgl_o.c       p3 = __lfpd(i3 + 2 * j);
__lfpd            152 libfftss/r8_bgl_o.c       p0 = __lfpd(i0 + 2 * j);
__lfpd            242 libfftss/r8_bgl_o.c     p1 = __lfpd(i1 + 2 * j);
__lfpd            243 libfftss/r8_bgl_o.c     p5 = __lfpd(i5 + 2 * j);
__lfpd            244 libfftss/r8_bgl_o.c     p3 = __lfpd(i3 + 2 * j);
__lfpd            245 libfftss/r8_bgl_o.c     p7 = __lfpd(i7 + 2 * j);
__lfpd            247 libfftss/r8_bgl_o.c     p0 = __lfpd(i0 + 2 * j);
__lfpd            248 libfftss/r8_bgl_o.c     p4 = __lfpd(i4 + 2 * j);
__lfpd            250 libfftss/r8_bgl_o.c     p2 = __lfpd(i2 + 2 * j);
__lfpd            252 libfftss/r8_bgl_o.c     p6 = __lfpd(i6 + 2 * j);
__lfpd            304 libfftss/r8_bgl_o.c     w1 = __lfpd(wn + i * bsize * 2);
__lfpd            305 libfftss/r8_bgl_o.c     w2 = __lfpd(wn + i * bsize * 4);
__lfpd            306 libfftss/r8_bgl_o.c     w4 = __lfpd(wn + i * bsize * 8);
__lfpd            307 libfftss/r8_bgl_o.c     w5 = __lfpd(wn + i * bsize * 2 + 2 * blocks * bsize);
__lfpd            308 libfftss/r8_bgl_o.c     w3 = __lfpd(wn2 + i * bsize * 2);
__lfpd            309 libfftss/r8_bgl_o.c     w6 = __lfpd(wn2 + i * bsize * 4);
__lfpd            310 libfftss/r8_bgl_o.c     w7 = __lfpd(wn2 + i * bsize * 2 + blocks * bsize * 2);
__lfpd            319 libfftss/r8_bgl_o.c       p5 = __lfpd(i5 + 2 * j);
__lfpd            320 libfftss/r8_bgl_o.c       p7 = __lfpd(i7 + 2 * j);
__lfpd            321 libfftss/r8_bgl_o.c       p4 = __lfpd(i4 + 2 * j);
__lfpd            322 libfftss/r8_bgl_o.c       p2 = __lfpd(i2 + 2 * j);
__lfpd            323 libfftss/r8_bgl_o.c       p6 = __lfpd(i6 + 2 * j);
__lfpd            325 libfftss/r8_bgl_o.c       p1 = __lfpd(i1 + 2 * j);
__lfpd            327 libfftss/r8_bgl_o.c       p3 = __lfpd(i3 + 2 * j);
__lfpd            329 libfftss/r8_bgl_o.c       p0 = __lfpd(i0 + 2 * j);
__lfpd             63 libfftss/r8_bgl_u1.c     w4 = __lfpd(wn + i * 8);
__lfpd             65 libfftss/r8_bgl_u1.c     p5 = __lfpd(i5 + 16 * i);
__lfpd             66 libfftss/r8_bgl_u1.c     p7 = __lfpd(i7 + 16 * i);
__lfpd             67 libfftss/r8_bgl_u1.c     p4 = __lfpd(i4 + 16 * i);
__lfpd             68 libfftss/r8_bgl_u1.c     w2 = __lfpd(wn + i * 4);
__lfpd             70 libfftss/r8_bgl_u1.c     p2 = __lfpd(i2 + 16 * i);
__lfpd             71 libfftss/r8_bgl_u1.c     w6 = __lfpd(wn2 + 4 * i); 
__lfpd             72 libfftss/r8_bgl_u1.c     p6 = __lfpd(i6 + 16 * i);
__lfpd             73 libfftss/r8_bgl_u1.c     p1 = __lfpd(i1 + 16 * i);
__lfpd             75 libfftss/r8_bgl_u1.c     p3 = __lfpd(i3 + 16 * i);
__lfpd             77 libfftss/r8_bgl_u1.c     p0 = __lfpd(i0 + 16 * i);
__lfpd             85 libfftss/r8_bgl_u1.c     w1 = __lfpd(wn + i * 2);
__lfpd             87 libfftss/r8_bgl_u1.c     w3 = __lfpd(wn2 + 2 * i);
__lfpd             92 libfftss/r8_bgl_u1.c     w5 = __lfpd(wn + i * 2 + 2 * blocks);
__lfpd             94 libfftss/r8_bgl_u1.c     w7 = __lfpd(wn2 + 2 * i + 2 * blocks);
__lfpd            165 libfftss/r8_bgl_u1.c     w4 = __lfpd(wn + i * 8);
__lfpd            166 libfftss/r8_bgl_u1.c     p5 = __lfpd(i5 + 16 * i);
__lfpd            167 libfftss/r8_bgl_u1.c     p7 = __lfpd(i7 + 16 * i);
__lfpd            168 libfftss/r8_bgl_u1.c     p4 = __lfpd(i4 + 16 * i);
__lfpd            169 libfftss/r8_bgl_u1.c     w2 = __lfpd(wn + i * 4);
__lfpd            170 libfftss/r8_bgl_u1.c     p2 = __lfpd(i2 + 16 * i);
__lfpd            171 libfftss/r8_bgl_u1.c     w6 = __lfpd(wn2 + 4 * i);
__lfpd            172 libfftss/r8_bgl_u1.c     p6 = __lfpd(i6 + 16 * i);
__lfpd            174 libfftss/r8_bgl_u1.c     p1 = __lfpd(i1 + 16 * i);
__lfpd            176 libfftss/r8_bgl_u1.c     p3 = __lfpd(i3 + 16 * i);
__lfpd            178 libfftss/r8_bgl_u1.c     p0 = __lfpd(i0 + 16 * i);
__lfpd            185 libfftss/r8_bgl_u1.c     w1 = __lfpd(wn + i * 2);
__lfpd            187 libfftss/r8_bgl_u1.c     w3 = __lfpd(wn2 + 2 * i); 
__lfpd            189 libfftss/r8_bgl_u1.c     w5 = __lfpd(wn + i * 2 + 2 * blocks);
__lfpd            191 libfftss/r8_bgl_u1.c     w7 = __lfpd(wn2 + 2 * i + 2 * blocks);