VSLDOI $8, ZERO, T1, T2 // 0x0000..0808 VADDUBM T0, T2, HIPERM // 0x0001..1617 VADDUBM T1, HIPERM, LOPERM // 0x0809..1e1f SRD $4, LEN, LEN // this allows to use sign bit as carry LXVD2X (INP)(R0), VIN0 // load input LXVD2X (INP)(R8), VIN1 SUBCCC $8, LEN, LEN LXVD2X (INP)(R9), VIN2 LXVD2X (INP)(R10), VIN3 ADD $0x40, INP, INP #ifdef GOARCH_ppc64le VPERM IN0, IN0, LEMASK, IN0 VPERM IN1, IN1, LEMASK, IN1 VPERM IN2, IN2, LEMASK, IN2 VPERM IN3, IN3, LEMASK, IN3 #endif VXOR IN0, XL, XH VPMSUMD IN1, H3L, XL1 VPMSUMD...