mirror of
https://github.com/herumi/xbyak
synced 2024-11-20 16:06:14 -07:00
310 lines
7.5 KiB
Text
310 lines
7.5 KiB
Text
//
|
|
vcvtnebf162ibs(xm1, xm2);
|
|
vcvtnebf162ibs(xm1, ptr[rax+128]);
|
|
vcvtnebf162ibs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvtnebf162ibs(ym1, ym2);
|
|
vcvtnebf162ibs(ym1, ptr[rax+128]);
|
|
vcvtnebf162ibs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvtnebf162ibs(zm1, zm2);
|
|
vcvtnebf162ibs(zm1, ptr[rax+128]);
|
|
vcvtnebf162ibs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvtnebf162iubs(xm1, xm2);
|
|
vcvtnebf162iubs(xm1, ptr[rax+128]);
|
|
vcvtnebf162iubs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvtnebf162iubs(ym1, ym2);
|
|
vcvtnebf162iubs(ym1, ptr[rax+128]);
|
|
vcvtnebf162iubs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvtnebf162iubs(zm1, zm2);
|
|
vcvtnebf162iubs(zm1, ptr[rax+128]);
|
|
vcvtnebf162iubs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttnebf162ibs(xm1, xm2);
|
|
vcvttnebf162ibs(xm1, ptr[rax+128]);
|
|
vcvttnebf162ibs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttnebf162ibs(ym1, ym2);
|
|
vcvttnebf162ibs(ym1, ptr[rax+128]);
|
|
vcvttnebf162ibs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttnebf162ibs(zm1, zm2);
|
|
vcvttnebf162ibs(zm1, ptr[rax+128]);
|
|
vcvttnebf162ibs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttnebf162iubs(xm1, xm2);
|
|
vcvttnebf162iubs(xm1, ptr[rax+128]);
|
|
vcvttnebf162iubs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttnebf162iubs(ym1, ym2);
|
|
vcvttnebf162iubs(ym1, ptr[rax+128]);
|
|
vcvttnebf162iubs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttnebf162iubs(zm1, zm2);
|
|
vcvttnebf162iubs(zm1, ptr[rax+128]);
|
|
vcvttnebf162iubs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttpd2qqs(xm1, xm2);
|
|
vcvttpd2qqs(xm1, ptr[rax+128]);
|
|
vcvttpd2qqs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttpd2qqs(ym1, ym2);
|
|
vcvttpd2qqs(ym1, ym2|T_sae);
|
|
vcvttpd2qqs(ym1, ptr[rax+128]);
|
|
vcvttpd2qqs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttpd2qqs(zm1, zm2);
|
|
vcvttpd2qqs(zm1, zm2|T_sae);
|
|
vcvttpd2qqs(zm1, ptr[rax+128]);
|
|
vcvttpd2qqs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttpd2uqqs(xm1, xm2);
|
|
vcvttpd2uqqs(xm1, ptr[rax+128]);
|
|
vcvttpd2uqqs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttpd2uqqs(ym1, ym2);
|
|
vcvttpd2uqqs(ym1, ym2|T_sae);
|
|
vcvttpd2uqqs(ym1, ptr[rax+128]);
|
|
vcvttpd2uqqs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttpd2uqqs(zm1, zm2);
|
|
vcvttpd2uqqs(zm1, zm2|T_sae);
|
|
vcvttpd2uqqs(zm1, ptr[rax+128]);
|
|
vcvttpd2uqqs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvtph2ibs(xm1, xm2);
|
|
vcvtph2ibs(xm1, ptr[rax+128]);
|
|
vcvtph2ibs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvtph2ibs(ym1, ym2);
|
|
vcvtph2ibs(ym1, ym2|T_rd_sae);
|
|
vcvtph2ibs(ym1, ptr[rax+128]);
|
|
vcvtph2ibs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvtph2ibs(zm1, zm2);
|
|
vcvtph2ibs(zm1, zm2|T_ru_sae);
|
|
vcvtph2ibs(zm1, ptr[rax+128]);
|
|
vcvtph2ibs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvtph2iubs(xm1, xm2);
|
|
vcvtph2iubs(xm1, ptr[rax+128]);
|
|
vcvtph2iubs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvtph2iubs(ym1, ym2);
|
|
vcvtph2iubs(ym1, ym2|T_rd_sae);
|
|
vcvtph2iubs(ym1, ptr[rax+128]);
|
|
vcvtph2iubs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvtph2iubs(zm1, zm2);
|
|
vcvtph2iubs(zm1, zm2|T_ru_sae);
|
|
vcvtph2iubs(zm1, ptr[rax+128]);
|
|
vcvtph2iubs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttph2ibs(xm1, xm2);
|
|
vcvttph2ibs(xm1, ptr[rax+128]);
|
|
vcvttph2ibs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttph2ibs(ym1, ym2);
|
|
vcvttph2ibs(ym1, ym2|T_rd_sae);
|
|
vcvttph2ibs(ym1, ptr[rax+128]);
|
|
vcvttph2ibs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttph2ibs(zm1, zm2);
|
|
vcvttph2ibs(zm1, zm2|T_ru_sae);
|
|
vcvttph2ibs(zm1, ptr[rax+128]);
|
|
vcvttph2ibs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttph2iubs(xm1, xm2);
|
|
vcvttph2iubs(xm1, ptr[rax+128]);
|
|
vcvttph2iubs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttph2iubs(ym1, ym2);
|
|
vcvttph2iubs(ym1, ym2|T_rd_sae);
|
|
vcvttph2iubs(ym1, ptr[rax+128]);
|
|
vcvttph2iubs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttph2iubs(zm1, zm2);
|
|
vcvttph2iubs(zm1, zm2|T_ru_sae);
|
|
vcvttph2iubs(zm1, ptr[rax+128]);
|
|
vcvttph2iubs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttps2dqs(xm1, xm2);
|
|
vcvttps2dqs(xm1, ptr[rax+128]);
|
|
vcvttps2dqs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttps2dqs(ym1, ym2);
|
|
vcvttps2dqs(ym1, ym2|T_sae);
|
|
vcvttps2dqs(ym1, ptr[rax+128]);
|
|
vcvttps2dqs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttps2dqs(zm1, zm2);
|
|
vcvttps2dqs(zm1, zm2|T_sae);
|
|
vcvttps2dqs(zm1, ptr[rax+128]);
|
|
vcvttps2dqs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvtps2ibs(xm1, xm2);
|
|
vcvtps2ibs(xm1, ptr[rax+128]);
|
|
vcvtps2ibs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvtps2ibs(ym1, ym2);
|
|
vcvtps2ibs(ym1, ym2|T_rd_sae);
|
|
vcvtps2ibs(ym1, ptr[rax+128]);
|
|
vcvtps2ibs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvtps2ibs(zm1, zm2);
|
|
vcvtps2ibs(zm1, zm2|T_ru_sae);
|
|
vcvtps2ibs(zm1, ptr[rax+128]);
|
|
vcvtps2ibs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvtps2iubs(xm1, xm2);
|
|
vcvtps2iubs(xm1, ptr[rax+128]);
|
|
vcvtps2iubs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvtps2iubs(ym1, ym2);
|
|
vcvtps2iubs(ym1, ym2|T_rd_sae);
|
|
vcvtps2iubs(ym1, ptr[rax+128]);
|
|
vcvtps2iubs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvtps2iubs(zm1, zm2);
|
|
vcvtps2iubs(zm1, zm2|T_ru_sae);
|
|
vcvtps2iubs(zm1, ptr[rax+128]);
|
|
vcvtps2iubs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttps2ibs(xm1, xm2);
|
|
vcvttps2ibs(xm1, ptr[rax+128]);
|
|
vcvttps2ibs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttps2ibs(ym1, ym2);
|
|
vcvttps2ibs(ym1, ym2|T_rd_sae);
|
|
vcvttps2ibs(ym1, ptr[rax+128]);
|
|
vcvttps2ibs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttps2ibs(zm1, zm2);
|
|
vcvttps2ibs(zm1, zm2|T_ru_sae);
|
|
vcvttps2ibs(zm1, ptr[rax+128]);
|
|
vcvttps2ibs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttps2iubs(xm1, xm2);
|
|
vcvttps2iubs(xm1, ptr[rax+128]);
|
|
vcvttps2iubs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttps2iubs(ym1, ym2);
|
|
vcvttps2iubs(ym1, ym2|T_rd_sae);
|
|
vcvttps2iubs(ym1, ptr[rax+128]);
|
|
vcvttps2iubs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttps2iubs(zm1, zm2);
|
|
vcvttps2iubs(zm1, zm2|T_ru_sae);
|
|
vcvttps2iubs(zm1, ptr[rax+128]);
|
|
vcvttps2iubs(zm1, ptr_b[rax+128]);
|
|
//
|
|
vcvttps2udqs(xm1, xm2);
|
|
vcvttps2udqs(xm1, ptr[rax+128]);
|
|
vcvttps2udqs(xm1, ptr_b[rax+128]);
|
|
|
|
vcvttps2udqs(ym1, ym2);
|
|
vcvttps2udqs(ym1, ym2|T_sae);
|
|
vcvttps2udqs(ym1, ptr[rax+128]);
|
|
vcvttps2udqs(ym1, ptr_b[rax+128]);
|
|
|
|
vcvttps2udqs(zm1, zm2);
|
|
vcvttps2udqs(zm1, zm2|T_sae);
|
|
vcvttps2udqs(zm1, ptr[rax+128]);
|
|
vcvttps2udqs(zm1, ptr_b[rax+128]);
|
|
|
|
//
|
|
vcvttpd2dqs(xm1|k1|T_z, xm2);
|
|
vcvttpd2dqs(xm1|k1|T_z, xword [rax+128]);
|
|
vcvttpd2dqs(xm1|k1|T_z, xword_b[rax+128]);
|
|
|
|
vcvttpd2dqs(xm1|k1|T_z, ym2);
|
|
vcvttpd2dqs(xm1|k1|T_z, ym2|T_sae);
|
|
vcvttpd2dqs(xm1|k1|T_z, yword [rax+128]);
|
|
vcvttpd2dqs(xm1|k1|T_z, yword_b[rax+128]);
|
|
|
|
vcvttpd2dqs(ym1|k1|T_z, zm2);
|
|
vcvttpd2dqs(ym1|k1|T_z, zm2|T_sae);
|
|
vcvttpd2dqs(ym1|k1|T_z, zword [rax+128]);
|
|
vcvttpd2dqs(ym1|k1|T_z, zword_b[rax+128]);
|
|
|
|
//
|
|
vcvttpd2udqs(xm1|k1|T_z, xm2);
|
|
vcvttpd2udqs(xm1|k1|T_z, xword [rax+128]);
|
|
vcvttpd2udqs(xm1|k1|T_z, xword_b[rax+128]);
|
|
|
|
vcvttpd2udqs(xm1|k1|T_z, ym2);
|
|
vcvttpd2udqs(xm1|k1|T_z, ym2|T_sae);
|
|
vcvttpd2udqs(xm1|k1|T_z, yword [rax+128]);
|
|
vcvttpd2udqs(xm1|k1|T_z, yword_b[rax+128]);
|
|
|
|
vcvttpd2udqs(ym1|k1|T_z, zm2);
|
|
vcvttpd2udqs(ym1|k1|T_z, zm2|T_sae);
|
|
vcvttpd2udqs(ym1|k1|T_z, zword [rax+128]);
|
|
vcvttpd2udqs(ym1|k1|T_z, zword_b[rax+128]);
|
|
//
|
|
vcvttps2qqs(xm1|k1|T_z, xm2);
|
|
vcvttps2qqs(xm1|k1|T_z, ptr [rax+128]);
|
|
vcvttps2qqs(xm1|k1|T_z, ptr_b[rax+128]);
|
|
|
|
vcvttps2qqs(ym1|k1|T_z, xm2);
|
|
vcvttps2qqs(ym1|k1|T_z, xm2|T_sae);
|
|
vcvttps2qqs(ym1|k1|T_z, ptr [rax+128]);
|
|
vcvttps2qqs(ym1|k1|T_z, ptr_b[rax+128]);
|
|
|
|
vcvttps2qqs(zm1, ym2);
|
|
vcvttps2qqs(zm1|k1|T_z, ym2);
|
|
vcvttps2qqs(zm1|k1|T_z|T_sae, ym2);
|
|
vcvttps2qqs(zm1|k1|T_z, ptr [rax+128]);
|
|
vcvttps2qqs(zm1|k1|T_z, ptr_b[rax+128]);
|
|
|
|
//
|
|
vcvttps2uqqs(xm1|k1|T_z, xm2);
|
|
vcvttps2uqqs(xm1|k1|T_z, ptr [rax+128]);
|
|
vcvttps2uqqs(xm1|k1|T_z, ptr_b[rax+128]);
|
|
|
|
vcvttps2uqqs(ym1|k1|T_z, xm2);
|
|
vcvttps2uqqs(ym1|k1|T_z, xm2|T_sae);
|
|
vcvttps2uqqs(ym1|k1|T_z, ptr [rax+128]);
|
|
vcvttps2uqqs(ym1|k1|T_z, ptr_b[rax+128]);
|
|
|
|
vcvttps2uqqs(zm1, ym2);
|
|
vcvttps2uqqs(zm1|k1|T_z, ym2);
|
|
vcvttps2uqqs(zm1|k1|T_z|T_sae, ym2);
|
|
vcvttps2uqqs(zm1|k1|T_z, ptr [rax+128]);
|
|
vcvttps2uqqs(zm1|k1|T_z, ptr_b[rax+128]);
|
|
|
|
//
|
|
vcvttsd2sis(eax, xm1);
|
|
vcvttsd2sis(eax, xm1|T_sae);
|
|
vcvttsd2sis(eax, ptr[rax+128]);
|
|
|
|
vcvttsd2sis(r30, xm1);
|
|
vcvttsd2sis(r30, xm1|T_sae);
|
|
vcvttsd2sis(r30, ptr[rax+128]);
|
|
//
|
|
vcvttsd2usis(eax, xm1);
|
|
vcvttsd2usis(eax, xm1|T_sae);
|
|
vcvttsd2usis(eax, ptr[rax+128]);
|
|
|
|
vcvttsd2usis(r30, xm1);
|
|
vcvttsd2usis(r30, xm1|T_sae);
|
|
vcvttsd2usis(r30, ptr[rax+128]);
|
|
//
|
|
vcvttss2sis(eax, xm1);
|
|
vcvttss2sis(eax, xm1|T_sae);
|
|
vcvttss2sis(eax, ptr[rax+128]);
|
|
|
|
vcvttss2sis(r30, xm1);
|
|
vcvttss2sis(r30, xm1|T_sae);
|
|
vcvttss2sis(r30, ptr[rax+128]);
|
|
//
|
|
vcvttss2usis(eax, xm1);
|
|
vcvttss2usis(eax, xm1|T_sae);
|
|
vcvttss2usis(eax, ptr[rax+128]);
|
|
|
|
vcvttss2usis(r30, xm1);
|
|
vcvttss2usis(r30, xm1|T_sae);
|
|
vcvttss2usis(r30, ptr[rax+128]);
|