2337    __m128i row0, row1, row2, row3, row4, row5, row6, row7;  in stbi__idct_simd()  local
2394          dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \  in stbi__idct_simd()
2415          dct_bfly32o(row2,row5, x2,x5,bias,shift); \  in stbi__idct_simd()
2435    row2 = _mm_load_si128((const __m128i *) (data + 2*8));  in stbi__idct_simd()
2449       dct_interleave16(row2, row6);  in stbi__idct_simd()
2453       dct_interleave16(row0, row2);  in stbi__idct_simd()
2460       dct_interleave16(row2, row3);  in stbi__idct_simd()
2471       __m128i p1 = _mm_packus_epi16(row2, row3);  in stbi__idct_simd()
2517    int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;  in stbi__idct_simd()  local
2566       int16x8_t sum26 = vaddq_s16(row2, row6); \  in stbi__idct_simd()
2569       dct_long_mac(t3e, p1e, row2, rot0_2); \  in stbi__idct_simd()
2599       dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \  in stbi__idct_simd()
2606    row2 = vld1q_s16(data + 2*8);  in stbi__idct_simd()
2629       dct_trn16(row2, row3);  in stbi__idct_simd()
2634       dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4  in stbi__idct_simd()
2642       dct_trn64(row2, row6);  in stbi__idct_simd()
2660       uint8x8_t p2 = vqrshrun_n_s16(row2, 1);  in stbi__idct_simd()