186797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* 286797937017f52bff088d02edf64fb931177a7eaJun Nakajima * MMX/3DNow!/SSE/SSE2/SSE3/SSSE3/SSE4/PNI support 386797937017f52bff088d02edf64fb931177a7eaJun Nakajima * 486797937017f52bff088d02edf64fb931177a7eaJun Nakajima * Copyright (c) 2005 Fabrice Bellard 586797937017f52bff088d02edf64fb931177a7eaJun Nakajima * 686797937017f52bff088d02edf64fb931177a7eaJun Nakajima * This library is free software; you can redistribute it and/or 786797937017f52bff088d02edf64fb931177a7eaJun Nakajima * modify it under the terms of the GNU Lesser General Public 886797937017f52bff088d02edf64fb931177a7eaJun Nakajima * License as published by the Free Software Foundation; either 986797937017f52bff088d02edf64fb931177a7eaJun Nakajima * version 2 of the License, or (at your option) any later version. 1086797937017f52bff088d02edf64fb931177a7eaJun Nakajima * 1186797937017f52bff088d02edf64fb931177a7eaJun Nakajima * This library is distributed in the hope that it will be useful, 1286797937017f52bff088d02edf64fb931177a7eaJun Nakajima * but WITHOUT ANY WARRANTY; without even the implied warranty of 1386797937017f52bff088d02edf64fb931177a7eaJun Nakajima * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 1486797937017f52bff088d02edf64fb931177a7eaJun Nakajima * Lesser General Public License for more details. 1586797937017f52bff088d02edf64fb931177a7eaJun Nakajima * 1686797937017f52bff088d02edf64fb931177a7eaJun Nakajima * You should have received a copy of the GNU Lesser General Public 1786797937017f52bff088d02edf64fb931177a7eaJun Nakajima * License along with this library; if not, write to the Free Software 1886797937017f52bff088d02edf64fb931177a7eaJun Nakajima * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301 USA 1986797937017f52bff088d02edf64fb931177a7eaJun Nakajima */ 2086797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 0 2186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define Reg MMXReg 2286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SUFFIX _mmx 2386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#else 2486797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define Reg XMMReg 2586797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SUFFIX _xmm 2686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 2786797937017f52bff088d02edf64fb931177a7eaJun Nakajima 2886797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define dh_alias_Reg ptr 2986797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define dh_alias_XMMReg ptr 3086797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define dh_alias_MMXReg ptr 3186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define dh_ctype_Reg Reg * 3286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define dh_ctype_XMMReg XMMReg * 3386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define dh_ctype_MMXReg MMXReg * 34f1d9bf153726533acf659efd796aa484dfd0b412David 'Digit' Turner#define dh_is_signed_Reg dh_is_signed_ptr 35f1d9bf153726533acf659efd796aa484dfd0b412David 'Digit' Turner#define dh_is_signed_XMMReg dh_is_signed_ptr 36f1d9bf153726533acf659efd796aa484dfd0b412David 'Digit' Turner#define dh_is_signed_MMXReg dh_is_signed_ptr 3786797937017f52bff088d02edf64fb931177a7eaJun Nakajima 3886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psrlw, SUFFIX), void, Reg, Reg) 3986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psraw, SUFFIX), void, Reg, Reg) 4086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psllw, SUFFIX), void, Reg, Reg) 4186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psrld, SUFFIX), void, Reg, Reg) 4286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psrad, SUFFIX), void, Reg, Reg) 4386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pslld, SUFFIX), void, Reg, Reg) 4486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psrlq, SUFFIX), void, Reg, Reg) 4586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psllq, SUFFIX), void, Reg, Reg) 4686797937017f52bff088d02edf64fb931177a7eaJun Nakajima 4786797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 1 4886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psrldq, SUFFIX), void, Reg, Reg) 4986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pslldq, SUFFIX), void, Reg, Reg) 5086797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 5186797937017f52bff088d02edf64fb931177a7eaJun Nakajima 5286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SSE_HELPER_B(name, F)\ 5386797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg) 5486797937017f52bff088d02edf64fb931177a7eaJun Nakajima 5586797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SSE_HELPER_W(name, F)\ 5686797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg) 5786797937017f52bff088d02edf64fb931177a7eaJun Nakajima 5886797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SSE_HELPER_L(name, F)\ 5986797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg) 6086797937017f52bff088d02edf64fb931177a7eaJun Nakajima 6186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SSE_HELPER_Q(name, F)\ 6286797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg) 6386797937017f52bff088d02edf64fb931177a7eaJun Nakajima 6486797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(paddb, FADD) 6586797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(paddw, FADD) 6686797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_L(paddl, FADD) 6786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_Q(paddq, FADD) 6886797937017f52bff088d02edf64fb931177a7eaJun Nakajima 6986797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(psubb, FSUB) 7086797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(psubw, FSUB) 7186797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_L(psubl, FSUB) 7286797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_Q(psubq, FSUB) 7386797937017f52bff088d02edf64fb931177a7eaJun Nakajima 7486797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(paddusb, FADDUB) 7586797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(paddsb, FADDSB) 7686797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(psubusb, FSUBUB) 7786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(psubsb, FSUBSB) 7886797937017f52bff088d02edf64fb931177a7eaJun Nakajima 7986797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(paddusw, FADDUW) 8086797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(paddsw, FADDSW) 8186797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(psubusw, FSUBUW) 8286797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(psubsw, FSUBSW) 8386797937017f52bff088d02edf64fb931177a7eaJun Nakajima 8486797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(pminub, FMINUB) 8586797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(pmaxub, FMAXUB) 8686797937017f52bff088d02edf64fb931177a7eaJun Nakajima 8786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pminsw, FMINSW) 8886797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pmaxsw, FMAXSW) 8986797937017f52bff088d02edf64fb931177a7eaJun Nakajima 9086797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_Q(pand, FAND) 9186797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_Q(pandn, FANDN) 9286797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_Q(por, FOR) 9386797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_Q(pxor, FXOR) 9486797937017f52bff088d02edf64fb931177a7eaJun Nakajima 9586797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(pcmpgtb, FCMPGTB) 9686797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pcmpgtw, FCMPGTW) 9786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_L(pcmpgtl, FCMPGTL) 9886797937017f52bff088d02edf64fb931177a7eaJun Nakajima 9986797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(pcmpeqb, FCMPEQ) 10086797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pcmpeqw, FCMPEQ) 10186797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_L(pcmpeql, FCMPEQ) 10286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 10386797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pmullw, FMULLW) 10486797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 0 10586797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pmulhrw, FMULHRW) 10686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 10786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pmulhuw, FMULHUW) 10886797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pmulhw, FMULHW) 10986797937017f52bff088d02edf64fb931177a7eaJun Nakajima 11086797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_B(pavgb, FAVG) 11186797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_W(pavgw, FAVG) 11286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 11386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmuludq, SUFFIX), void, Reg, Reg) 11486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmaddwd, SUFFIX), void, Reg, Reg) 11586797937017f52bff088d02edf64fb931177a7eaJun Nakajima 11686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psadbw, SUFFIX), void, Reg, Reg) 11786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(maskmov, SUFFIX), void, Reg, Reg, tl) 11886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(movl_mm_T0, SUFFIX), void, Reg, i32) 11986797937017f52bff088d02edf64fb931177a7eaJun Nakajima#ifdef TARGET_X86_64 12086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(movq_mm_T0, SUFFIX), void, Reg, i64) 12186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 12286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 12386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 0 12486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pshufw, SUFFIX), void, Reg, Reg, int) 12586797937017f52bff088d02edf64fb931177a7eaJun Nakajima#else 12686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(shufps, void, Reg, Reg, int) 12786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(shufpd, void, Reg, Reg, int) 12886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pshufd, SUFFIX), void, Reg, Reg, int) 12986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pshuflw, SUFFIX), void, Reg, Reg, int) 13086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pshufhw, SUFFIX), void, Reg, Reg, int) 13186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 13286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 13386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 1 13486797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* FPU ops */ 13586797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* XXX: not accurate */ 13686797937017f52bff088d02edf64fb931177a7eaJun Nakajima 13786797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SSE_HELPER_S(name, F)\ 13886797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(name ## ps , void, Reg, Reg) \ 13986797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(name ## ss , void, Reg, Reg) \ 14086797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(name ## pd , void, Reg, Reg) \ 14186797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(name ## sd , void, Reg, Reg) 14286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 14386797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(add, FPU_ADD) 14486797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(sub, FPU_SUB) 14586797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(mul, FPU_MUL) 14686797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(div, FPU_DIV) 14786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(min, FPU_MIN) 14886797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(max, FPU_MAX) 14986797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_S(sqrt, FPU_SQRT) 15086797937017f52bff088d02edf64fb931177a7eaJun Nakajima 15186797937017f52bff088d02edf64fb931177a7eaJun Nakajima 15286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtps2pd, void, Reg, Reg) 15386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtpd2ps, void, Reg, Reg) 15486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtss2sd, void, Reg, Reg) 15586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtsd2ss, void, Reg, Reg) 15686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtdq2ps, void, Reg, Reg) 15786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtdq2pd, void, Reg, Reg) 15886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtpi2ps, void, XMMReg, MMXReg) 15986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtpi2pd, void, XMMReg, MMXReg) 16086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtsi2ss, void, XMMReg, i32) 16186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtsi2sd, void, XMMReg, i32) 16286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 16386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#ifdef TARGET_X86_64 16486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtsq2ss, void, XMMReg, i64) 16586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtsq2sd, void, XMMReg, i64) 16686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 16786797937017f52bff088d02edf64fb931177a7eaJun Nakajima 16886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtps2dq, void, XMMReg, XMMReg) 16986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtpd2dq, void, XMMReg, XMMReg) 17086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtps2pi, void, MMXReg, XMMReg) 17186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvtpd2pi, void, MMXReg, XMMReg) 17286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvtss2si, s32, XMMReg) 17386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvtsd2si, s32, XMMReg) 17486797937017f52bff088d02edf64fb931177a7eaJun Nakajima#ifdef TARGET_X86_64 17586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvtss2sq, s64, XMMReg) 17686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvtsd2sq, s64, XMMReg) 17786797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 17886797937017f52bff088d02edf64fb931177a7eaJun Nakajima 17986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvttps2dq, void, XMMReg, XMMReg) 18086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvttpd2dq, void, XMMReg, XMMReg) 18186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvttps2pi, void, MMXReg, XMMReg) 18286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(cvttpd2pi, void, MMXReg, XMMReg) 18386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvttss2si, s32, XMMReg) 18486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvttsd2si, s32, XMMReg) 18586797937017f52bff088d02edf64fb931177a7eaJun Nakajima#ifdef TARGET_X86_64 18686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvttss2sq, s64, XMMReg) 18786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(cvttsd2sq, s64, XMMReg) 18886797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 18986797937017f52bff088d02edf64fb931177a7eaJun Nakajima 19086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(rsqrtps, void, XMMReg, XMMReg) 19186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(rsqrtss, void, XMMReg, XMMReg) 19286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(rcpps, void, XMMReg, XMMReg) 19386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(rcpss, void, XMMReg, XMMReg) 19486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(haddps, void, XMMReg, XMMReg) 19586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(haddpd, void, XMMReg, XMMReg) 19686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(hsubps, void, XMMReg, XMMReg) 19786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(hsubpd, void, XMMReg, XMMReg) 19886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(addsubps, void, XMMReg, XMMReg) 19986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(addsubpd, void, XMMReg, XMMReg) 20086797937017f52bff088d02edf64fb931177a7eaJun Nakajima 20186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define SSE_HELPER_CMP(name, F)\ 20286797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2( name ## ps , void, Reg, Reg) \ 20386797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2( name ## ss , void, Reg, Reg) \ 20486797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2( name ## pd , void, Reg, Reg) \ 20586797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2( name ## sd , void, Reg, Reg) 20686797937017f52bff088d02edf64fb931177a7eaJun Nakajima 20786797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmpeq, FPU_CMPEQ) 20886797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmplt, FPU_CMPLT) 20986797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmple, FPU_CMPLE) 21086797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmpunord, FPU_CMPUNORD) 21186797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmpneq, FPU_CMPNEQ) 21286797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmpnlt, FPU_CMPNLT) 21386797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmpnle, FPU_CMPNLE) 21486797937017f52bff088d02edf64fb931177a7eaJun NakajimaSSE_HELPER_CMP(cmpord, FPU_CMPORD) 21586797937017f52bff088d02edf64fb931177a7eaJun Nakajima 21686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(ucomiss, void, Reg, Reg) 21786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(comiss, void, Reg, Reg) 21886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(ucomisd, void, Reg, Reg) 21986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(comisd, void, Reg, Reg) 22086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(movmskps, i32, Reg) 22186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(movmskpd, i32, Reg) 22286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 22386797937017f52bff088d02edf64fb931177a7eaJun Nakajima 22486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_1(glue(pmovmskb, SUFFIX), i32, Reg) 22586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(packsswb, SUFFIX), void, Reg, Reg) 22686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(packuswb, SUFFIX), void, Reg, Reg) 22786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(packssdw, SUFFIX), void, Reg, Reg) 22886797937017f52bff088d02edf64fb931177a7eaJun Nakajima#define UNPCK_OP(base_name, base) \ 22986797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(punpck ## base_name ## bw, SUFFIX) , void, Reg, Reg) \ 23086797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(punpck ## base_name ## wd, SUFFIX) , void, Reg, Reg) \ 23186797937017f52bff088d02edf64fb931177a7eaJun Nakajima DEF_HELPER_2(glue(punpck ## base_name ## dq, SUFFIX) , void, Reg, Reg) 23286797937017f52bff088d02edf64fb931177a7eaJun Nakajima 23386797937017f52bff088d02edf64fb931177a7eaJun NakajimaUNPCK_OP(l, 0) 23486797937017f52bff088d02edf64fb931177a7eaJun NakajimaUNPCK_OP(h, 1) 23586797937017f52bff088d02edf64fb931177a7eaJun Nakajima 23686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 1 23786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(punpcklqdq, SUFFIX), void, Reg, Reg) 23886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(punpckhqdq, SUFFIX), void, Reg, Reg) 23986797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 24086797937017f52bff088d02edf64fb931177a7eaJun Nakajima 24186797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* 3DNow! float ops */ 24286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 0 24386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pi2fd, void, MMXReg, MMXReg) 24486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pi2fw, void, MMXReg, MMXReg) 24586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pf2id, void, MMXReg, MMXReg) 24686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pf2iw, void, MMXReg, MMXReg) 24786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfacc, void, MMXReg, MMXReg) 24886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfadd, void, MMXReg, MMXReg) 24986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfcmpeq, void, MMXReg, MMXReg) 25086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfcmpge, void, MMXReg, MMXReg) 25186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfcmpgt, void, MMXReg, MMXReg) 25286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfmax, void, MMXReg, MMXReg) 25386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfmin, void, MMXReg, MMXReg) 25486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfmul, void, MMXReg, MMXReg) 25586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfnacc, void, MMXReg, MMXReg) 25686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfpnacc, void, MMXReg, MMXReg) 25786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfrcp, void, MMXReg, MMXReg) 25886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfrsqrt, void, MMXReg, MMXReg) 25986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfsub, void, MMXReg, MMXReg) 26086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pfsubr, void, MMXReg, MMXReg) 26186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(pswapd, void, MMXReg, MMXReg) 26286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 26386797937017f52bff088d02edf64fb931177a7eaJun Nakajima 26486797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* SSSE3 op helpers */ 26586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phaddw, SUFFIX), void, Reg, Reg) 26686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phaddd, SUFFIX), void, Reg, Reg) 26786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phaddsw, SUFFIX), void, Reg, Reg) 26886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phsubw, SUFFIX), void, Reg, Reg) 26986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phsubd, SUFFIX), void, Reg, Reg) 27086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phsubsw, SUFFIX), void, Reg, Reg) 27186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pabsb, SUFFIX), void, Reg, Reg) 27286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pabsw, SUFFIX), void, Reg, Reg) 27386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pabsd, SUFFIX), void, Reg, Reg) 27486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmaddubsw, SUFFIX), void, Reg, Reg) 27586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmulhrsw, SUFFIX), void, Reg, Reg) 27686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pshufb, SUFFIX), void, Reg, Reg) 27786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psignb, SUFFIX), void, Reg, Reg) 27886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psignw, SUFFIX), void, Reg, Reg) 27986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(psignd, SUFFIX), void, Reg, Reg) 28086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(palignr, SUFFIX), void, Reg, Reg, s32) 28186797937017f52bff088d02edf64fb931177a7eaJun Nakajima 28286797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* SSE4.1 op helpers */ 28386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 1 28486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pblendvb, SUFFIX), void, Reg, Reg) 28586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(blendvps, SUFFIX), void, Reg, Reg) 28686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(blendvpd, SUFFIX), void, Reg, Reg) 28786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(ptest, SUFFIX), void, Reg, Reg) 28886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovsxbw, SUFFIX), void, Reg, Reg) 28986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovsxbd, SUFFIX), void, Reg, Reg) 29086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovsxbq, SUFFIX), void, Reg, Reg) 29186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovsxwd, SUFFIX), void, Reg, Reg) 29286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovsxwq, SUFFIX), void, Reg, Reg) 29386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovsxdq, SUFFIX), void, Reg, Reg) 29486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovzxbw, SUFFIX), void, Reg, Reg) 29586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovzxbd, SUFFIX), void, Reg, Reg) 29686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovzxbq, SUFFIX), void, Reg, Reg) 29786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovzxwd, SUFFIX), void, Reg, Reg) 29886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovzxwq, SUFFIX), void, Reg, Reg) 29986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmovzxdq, SUFFIX), void, Reg, Reg) 30086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmuldq, SUFFIX), void, Reg, Reg) 30186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pcmpeqq, SUFFIX), void, Reg, Reg) 30286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(packusdw, SUFFIX), void, Reg, Reg) 30386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pminsb, SUFFIX), void, Reg, Reg) 30486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pminsd, SUFFIX), void, Reg, Reg) 30586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pminuw, SUFFIX), void, Reg, Reg) 30686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pminud, SUFFIX), void, Reg, Reg) 30786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmaxsb, SUFFIX), void, Reg, Reg) 30886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmaxsd, SUFFIX), void, Reg, Reg) 30986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmaxuw, SUFFIX), void, Reg, Reg) 31086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmaxud, SUFFIX), void, Reg, Reg) 31186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pmulld, SUFFIX), void, Reg, Reg) 31286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(phminposuw, SUFFIX), void, Reg, Reg) 31386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(roundps, SUFFIX), void, Reg, Reg, i32) 31486797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(roundpd, SUFFIX), void, Reg, Reg, i32) 31586797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(roundss, SUFFIX), void, Reg, Reg, i32) 31686797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(roundsd, SUFFIX), void, Reg, Reg, i32) 31786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(blendps, SUFFIX), void, Reg, Reg, i32) 31886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(blendpd, SUFFIX), void, Reg, Reg, i32) 31986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pblendw, SUFFIX), void, Reg, Reg, i32) 32086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(dpps, SUFFIX), void, Reg, Reg, i32) 32186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(dppd, SUFFIX), void, Reg, Reg, i32) 32286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(mpsadbw, SUFFIX), void, Reg, Reg, i32) 32386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 32486797937017f52bff088d02edf64fb931177a7eaJun Nakajima 32586797937017f52bff088d02edf64fb931177a7eaJun Nakajima/* SSE4.2 op helpers */ 32686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#if SHIFT == 1 32786797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(glue(pcmpgtq, SUFFIX), void, Reg, Reg) 32886797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pcmpestri, SUFFIX), void, Reg, Reg, i32) 32986797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pcmpestrm, SUFFIX), void, Reg, Reg, i32) 33086797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pcmpistri, SUFFIX), void, Reg, Reg, i32) 33186797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(glue(pcmpistrm, SUFFIX), void, Reg, Reg, i32) 33286797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_3(crc32, tl, i32, tl, i32) 33386797937017f52bff088d02edf64fb931177a7eaJun NakajimaDEF_HELPER_2(popcnt, tl, tl, i32) 33486797937017f52bff088d02edf64fb931177a7eaJun Nakajima#endif 33586797937017f52bff088d02edf64fb931177a7eaJun Nakajima 33686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SHIFT 33786797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef Reg 33886797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SUFFIX 33986797937017f52bff088d02edf64fb931177a7eaJun Nakajima 34086797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SSE_HELPER_B 34186797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SSE_HELPER_W 34286797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SSE_HELPER_L 34386797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SSE_HELPER_Q 34486797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SSE_HELPER_S 34586797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef SSE_HELPER_CMP 34686797937017f52bff088d02edf64fb931177a7eaJun Nakajima#undef UNPCK_OP 347