VR0
"vpxor %" VR0(r) ", %" VR2(r)", %" VR2(r) "\n" \
"vmovdqa %" VR0(r) ", %" VR4(r) "\n" \
"vmovdqa %" VR0(r) ", %" VR2(r) "\n" \
"vmovdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"vmovdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"vmovdqa %%" VR0(r) ", 0x00(%[DST])\n" \
"vmovdqa %%" VR0(r) ", 0x00(%[DST])\n" \
"vpcmpgtb %" VR0(r)", %ymm15, %ymm12\n" \
"vpaddb %" VR0(r)", %" VR0(r)", %" VR0(r) "\n" \
"vpxor %ymm12, %" VR0(r)", %" VR0(r) "\n" \
"vpsraw $0x4, %%" VR0(r) ", %%"_as "\n" \
"vpand %%" _0f ", %%" VR0(r) ", %%" VR0(r) "\n" \
"vpshufb %%" VR0(r) ", %%" _ltmod ", %%" _ta "\n" \
"vpshufb %%" VR0(r) ", %%" _ltmul ", %%" VR0(r) "\n"\
"vpxor %%" _ta ", %%" VR0(r) ", %%" VR0(r) "\n" \
"vpxor %%" _as ", %%" VR0(r) ", %%" VR0(r) "\n" \
"vpxor 0x00(%[SRC]), %%" VR0(r)", %%" VR0(r) "\n" \
"vpxor 0x00(%[SRC]), %%" VR0(r)", %%" VR0(r) "\n" \
"vpxor %" VR0(r) ", %" VR4(r)", %" VR4(r) "\n" \
"pxor %" VR0(r) ", %" VR1(r)); \
"movdqa %" VR0(r) ", %" VR4(r) "\n" \
"movdqa %" VR0(r) ", %" VR2(r) "\n" \
"movdqa %" VR0(r) ", %" VR1(r)); \
"movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"movdqa %%" VR0(r)", 0x00(%[DST])\n" \
"movdqa %%" VR0(r)", 0x00(%[DST])\n" \
"movdqa %%" VR0(r)", 0x00(%[DST])\n" \
_MUL2_x2(VR0(r), VR1(r)); \
_MUL2_x2(VR0(r), VR1(r)); \
_MUL2_x1(VR0(r)); \
"pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
"pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
__asm("pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
"pxor %" VR0(r) ", %" VR4(r) "\n" \
"pxor %" VR0(r) ", %" VR2(r) "\n" \
"pxor %" VR0(r) ", %" VR2(r) "\n" \
"movdqa %" VR0(r) ", %" VR4(r) "\n" \
"movdqa %" VR0(r) ", %" VR2(r) "\n" \
"movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"movdqa 0x00(%[SRC]), %%" VR0(r) "\n" \
"movdqa %%" VR0(r)", 0x00(%[DST])\n" \
"movdqa %%" VR0(r)", 0x00(%[DST])\n" \
"pcmpgtb %" VR0(r)", %xmm14\n" \
"paddb %" VR0(r)", %" VR0(r) "\n" \
"pxor %xmm14, %" VR0(r) "\n" \
"movdqa %%" VR0(r) ", %%" _a_save "\n" \
"psraw $0x4, %%" VR0(r) "\n" \
"pand %%" _0f ", %%" VR0(r) "\n" \
"pshufb %%" VR0(r) ",%%" _lt_mod_a "\n" \
"pshufb %%" VR0(r) ",%%" _lt_clmul_a "\n" \
"movdqa %%" _lt_clmul_a ",%%" VR0(r) "\n" \
"pxor %%" _lt_mod_a ",%%" VR0(r) "\n" \
"pxor %%" _lt_clmul_a ",%%" VR0(r) "\n" \
"pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
"pxor 0x00(%[SRC]), %%" VR0(r) "\n" \
"pxor %" VR0(r) ", %" VR4(r) "\n" \