93     f16x8_t vecTmp0, vecTmp1;  in _arm_radix4_butterfly_f16_mve()  local
151                 vecTmp0 = vecSum0 + vecSum1;  in _arm_radix4_butterfly_f16_mve()
152                 vst1q(inA, vecTmp0);  in _arm_radix4_butterfly_f16_mve()
158                 vecTmp0 = vecSum0 - vecSum1;  in _arm_radix4_butterfly_f16_mve()
164                 vecTmp1 = MVE_CMPLX_MULT_FLT_Conj_AxB(vecW, vecTmp0);  in _arm_radix4_butterfly_f16_mve()
171                 vecTmp0 = MVE_CMPLX_SUB_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_f16_mve()
177                 vecTmp1 = MVE_CMPLX_MULT_FLT_Conj_AxB(vecW, vecTmp0);  in _arm_radix4_butterfly_f16_mve()
184                 vecTmp0 = MVE_CMPLX_ADD_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_f16_mve()
190                 vecTmp1 = MVE_CMPLX_MULT_FLT_Conj_AxB(vecW, vecTmp0);  in _arm_radix4_butterfly_f16_mve()
233         vecTmp0 = vecSum0 + vecSum1;  in _arm_radix4_butterfly_f16_mve()
234         vstrwq_scatter_base_f32(vecScGathAddr, -64, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_f16_mve()
236         vecTmp0 = vecSum0 - vecSum1;  in _arm_radix4_butterfly_f16_mve()
237         vstrwq_scatter_base_f32(vecScGathAddr, -64 + 4, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_f16_mve()
239         vecTmp0 = MVE_CMPLX_SUB_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_f16_mve()
240         vstrwq_scatter_base_f32(vecScGathAddr, -64 + 8, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_f16_mve()
242         vecTmp0 = MVE_CMPLX_ADD_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_f16_mve()
243         vstrwq_scatter_base_f32(vecScGathAddr, -64 + 12, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_f16_mve()
299     f16x8_t vecTmp0, vecTmp1;  in _arm_radix4_butterfly_inverse_f16_mve()  local
359                 vecTmp0 = vecSum0 + vecSum1;  in _arm_radix4_butterfly_inverse_f16_mve()
360                 vst1q(inA, vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
365                 vecTmp0 = vecSum0 - vecSum1;  in _arm_radix4_butterfly_inverse_f16_mve()
371                 vecTmp1 = MVE_CMPLX_MULT_FLT_AxB(vecW, vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
378                 vecTmp0 = MVE_CMPLX_ADD_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_inverse_f16_mve()
384                 vecTmp1 = MVE_CMPLX_MULT_FLT_AxB(vecW, vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
391                 vecTmp0 = MVE_CMPLX_SUB_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_inverse_f16_mve()
397                 vecTmp1 = MVE_CMPLX_MULT_FLT_AxB(vecW, vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
441         vecTmp0 = vecSum0 + vecSum1;  in _arm_radix4_butterfly_inverse_f16_mve()
442         vecTmp0 = vecTmp0 * onebyfftLen;  in _arm_radix4_butterfly_inverse_f16_mve()
443         vstrwq_scatter_base_f32(vecScGathAddr, -64, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
445         vecTmp0 = vecSum0 - vecSum1;  in _arm_radix4_butterfly_inverse_f16_mve()
446         vecTmp0 = vecTmp0 * onebyfftLen;  in _arm_radix4_butterfly_inverse_f16_mve()
447         vstrwq_scatter_base_f32(vecScGathAddr, -64 + 4, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
449         vecTmp0 = MVE_CMPLX_ADD_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_inverse_f16_mve()
450         vecTmp0 = vecTmp0 * onebyfftLen;  in _arm_radix4_butterfly_inverse_f16_mve()
451         vstrwq_scatter_base_f32(vecScGathAddr, -64 + 8, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()
453         vecTmp0 = MVE_CMPLX_SUB_A_ixB(vecDiff0, vecDiff1);  in _arm_radix4_butterfly_inverse_f16_mve()
454         vecTmp0 = vecTmp0 * onebyfftLen;  in _arm_radix4_butterfly_inverse_f16_mve()
455         vstrwq_scatter_base_f32(vecScGathAddr, -64 + 12, (f32x4_t)vecTmp0);  in _arm_radix4_butterfly_inverse_f16_mve()