163         q31_t R, S, T, U;  in arm_radix4_butterfly_q15()  local
206     T = read_q15x2 (pSi0);  in arm_radix4_butterfly_q15()
207     T = __SHADD16(T, 0); /* this is just a SIMD arithmetic shift right by 1 */  in arm_radix4_butterfly_q15()
208 …T = __SHADD16(T, 0); /* it turns out doing this twice is 2 cycles, the alternative takes 3 cycles …  in arm_radix4_butterfly_q15()
220     R = __QADD16(T, S);  in arm_radix4_butterfly_q15()
223     S = __QSUB16(T, S);  in arm_radix4_butterfly_q15()
227     T = read_q15x2 (pSi1);  in arm_radix4_butterfly_q15()
228     T = __SHADD16(T, 0);  in arm_radix4_butterfly_q15()
229     T = __SHADD16(T, 0);  in arm_radix4_butterfly_q15()
237     T = __QADD16(T, U);  in arm_radix4_butterfly_q15()
242     write_q15x2_ia (&pSi0, __SHADD16(R, T));  in arm_radix4_butterfly_q15()
245     R = __QSUB16(R, T);  in arm_radix4_butterfly_q15()
264     T = read_q15x2 (pSi1);  in arm_radix4_butterfly_q15()
265     T = __SHADD16(T, 0);  in arm_radix4_butterfly_q15()
266     T = __SHADD16(T, 0);  in arm_radix4_butterfly_q15()
279     T = __QSUB16(T, U);  in arm_radix4_butterfly_q15()
283     R = __QASX(S, T);  in arm_radix4_butterfly_q15()
285     S = __QSAX(S, T);  in arm_radix4_butterfly_q15()
288     R = __QSAX(S, T);  in arm_radix4_butterfly_q15()
290     S = __QASX(S, T);  in arm_radix4_butterfly_q15()
373         T = read_q15x2 (pSi0);  in arm_radix4_butterfly_q15()
379         R = __QADD16(T, S);  in arm_radix4_butterfly_q15()
382         S = __QSUB16(T, S);  in arm_radix4_butterfly_q15()
386         T = read_q15x2 (pSi1);  in arm_radix4_butterfly_q15()
392         T = __QADD16(T, U);  in arm_radix4_butterfly_q15()
398         out1 = __SHADD16(R, T);  in arm_radix4_butterfly_q15()
404         R = __SHSUB16(R, T);  in arm_radix4_butterfly_q15()
422         T = read_q15x2 (pSi1);  in arm_radix4_butterfly_q15()
436         T = __QSUB16(T, U);  in arm_radix4_butterfly_q15()
440         R = __SHASX(S, T);  in arm_radix4_butterfly_q15()
443         S = __SHSAX(S, T);  in arm_radix4_butterfly_q15()
451         R = __SHSAX(S, T);  in arm_radix4_butterfly_q15()
454         S = __SHASX(S, T);  in arm_radix4_butterfly_q15()
520     T = __QADD16(xbyb, xdyd);  in arm_radix4_butterfly_q15()
528     write_q15x2_ia (&ptr1, __SHADD16(R, T));  in arm_radix4_butterfly_q15()
531     T = __QADD16(xbyb, xdyd);  in arm_radix4_butterfly_q15()
535     write_q15x2_ia (&ptr1, __SHSUB16(R, T));  in arm_radix4_butterfly_q15()
1032         q31_t R, S, T, U;  in arm_radix4_butterfly_inverse_q15()  local
1075     T = read_q15x2 (pSi0);  in arm_radix4_butterfly_inverse_q15()
1076     T = __SHADD16(T, 0);  in arm_radix4_butterfly_inverse_q15()
1077     T = __SHADD16(T, 0);  in arm_radix4_butterfly_inverse_q15()
1085     R = __QADD16(T, S);  in arm_radix4_butterfly_inverse_q15()
1088     S = __QSUB16(T, S);  in arm_radix4_butterfly_inverse_q15()
1092     T = read_q15x2 (pSi1);  in arm_radix4_butterfly_inverse_q15()
1093     T = __SHADD16(T, 0);  in arm_radix4_butterfly_inverse_q15()
1094     T = __SHADD16(T, 0);  in arm_radix4_butterfly_inverse_q15()
1102     T = __QADD16(T, U);  in arm_radix4_butterfly_inverse_q15()
1107     write_q15x2_ia (&pSi0, __SHADD16(R, T));  in arm_radix4_butterfly_inverse_q15()
1110     R = __QSUB16(R, T);  in arm_radix4_butterfly_inverse_q15()
1129     T = read_q15x2 (pSi1);  in arm_radix4_butterfly_inverse_q15()
1130     T = __SHADD16(T, 0);  in arm_radix4_butterfly_inverse_q15()
1131     T = __SHADD16(T, 0);  in arm_radix4_butterfly_inverse_q15()
1144     T = __QSUB16(T, U);  in arm_radix4_butterfly_inverse_q15()
1148     R = __QSAX(S, T);  in arm_radix4_butterfly_inverse_q15()
1150     S = __QASX(S, T);  in arm_radix4_butterfly_inverse_q15()
1153     R = __QASX(S, T);  in arm_radix4_butterfly_inverse_q15()
1155     S = __QSAX(S, T);  in arm_radix4_butterfly_inverse_q15()
1238         T = read_q15x2 (pSi0);  in arm_radix4_butterfly_inverse_q15()
1244         R = __QADD16(T, S);  in arm_radix4_butterfly_inverse_q15()
1247         S = __QSUB16(T, S);  in arm_radix4_butterfly_inverse_q15()
1251         T = read_q15x2 (pSi1);  in arm_radix4_butterfly_inverse_q15()
1257         T = __QADD16(T, U);  in arm_radix4_butterfly_inverse_q15()
1263         out1 = __SHADD16(R, T);  in arm_radix4_butterfly_inverse_q15()
1269         R = __SHSUB16(R, T);  in arm_radix4_butterfly_inverse_q15()
1287         T = read_q15x2 (pSi1);  in arm_radix4_butterfly_inverse_q15()
1301         T = __QSUB16(T, U);  in arm_radix4_butterfly_inverse_q15()
1305         R = __SHSAX(S, T);  in arm_radix4_butterfly_inverse_q15()
1308         S = __SHASX(S, T);  in arm_radix4_butterfly_inverse_q15()
1315         R = __SHASX(S, T);  in arm_radix4_butterfly_inverse_q15()
1318         S = __SHSAX(S, T);  in arm_radix4_butterfly_inverse_q15()
1382     T = __QADD16(xbyb, xdyd);  in arm_radix4_butterfly_inverse_q15()
1390     write_q15x2_ia (&ptr1, __SHADD16(R, T));  in arm_radix4_butterfly_inverse_q15()
1393     T = __QADD16(xbyb, xdyd);  in arm_radix4_butterfly_inverse_q15()
1397     write_q15x2_ia (&ptr1, __SHSUB16(R, T));  in arm_radix4_butterfly_inverse_q15()