1/*
2 *  MMX/3DNow!/SSE/SSE2/SSE3/SSSE3/SSE4/PNI support
3 *
4 *  Copyright (c) 2005 Fabrice Bellard
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA  02110-1301 USA
19 */
20#if SHIFT == 0
21#define Reg MMXReg
22#define SUFFIX _mmx
23#else
24#define Reg XMMReg
25#define SUFFIX _xmm
26#endif
27
28#define dh_alias_Reg ptr
29#define dh_alias_XMMReg ptr
30#define dh_alias_MMXReg ptr
31#define dh_ctype_Reg Reg *
32#define dh_ctype_XMMReg XMMReg *
33#define dh_ctype_MMXReg MMXReg *
34#define dh_is_signed_Reg dh_is_signed_ptr
35#define dh_is_signed_XMMReg dh_is_signed_ptr
36#define dh_is_signed_MMXReg dh_is_signed_ptr
37
38DEF_HELPER_2(glue(psrlw, SUFFIX), void, Reg, Reg)
39DEF_HELPER_2(glue(psraw, SUFFIX), void, Reg, Reg)
40DEF_HELPER_2(glue(psllw, SUFFIX), void, Reg, Reg)
41DEF_HELPER_2(glue(psrld, SUFFIX), void, Reg, Reg)
42DEF_HELPER_2(glue(psrad, SUFFIX), void, Reg, Reg)
43DEF_HELPER_2(glue(pslld, SUFFIX), void, Reg, Reg)
44DEF_HELPER_2(glue(psrlq, SUFFIX), void, Reg, Reg)
45DEF_HELPER_2(glue(psllq, SUFFIX), void, Reg, Reg)
46
47#if SHIFT == 1
48DEF_HELPER_2(glue(psrldq, SUFFIX), void, Reg, Reg)
49DEF_HELPER_2(glue(pslldq, SUFFIX), void, Reg, Reg)
50#endif
51
52#define SSE_HELPER_B(name, F)\
53    DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg)
54
55#define SSE_HELPER_W(name, F)\
56    DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg)
57
58#define SSE_HELPER_L(name, F)\
59    DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg)
60
61#define SSE_HELPER_Q(name, F)\
62    DEF_HELPER_2(glue(name, SUFFIX), void, Reg, Reg)
63
64SSE_HELPER_B(paddb, FADD)
65SSE_HELPER_W(paddw, FADD)
66SSE_HELPER_L(paddl, FADD)
67SSE_HELPER_Q(paddq, FADD)
68
69SSE_HELPER_B(psubb, FSUB)
70SSE_HELPER_W(psubw, FSUB)
71SSE_HELPER_L(psubl, FSUB)
72SSE_HELPER_Q(psubq, FSUB)
73
74SSE_HELPER_B(paddusb, FADDUB)
75SSE_HELPER_B(paddsb, FADDSB)
76SSE_HELPER_B(psubusb, FSUBUB)
77SSE_HELPER_B(psubsb, FSUBSB)
78
79SSE_HELPER_W(paddusw, FADDUW)
80SSE_HELPER_W(paddsw, FADDSW)
81SSE_HELPER_W(psubusw, FSUBUW)
82SSE_HELPER_W(psubsw, FSUBSW)
83
84SSE_HELPER_B(pminub, FMINUB)
85SSE_HELPER_B(pmaxub, FMAXUB)
86
87SSE_HELPER_W(pminsw, FMINSW)
88SSE_HELPER_W(pmaxsw, FMAXSW)
89
90SSE_HELPER_Q(pand, FAND)
91SSE_HELPER_Q(pandn, FANDN)
92SSE_HELPER_Q(por, FOR)
93SSE_HELPER_Q(pxor, FXOR)
94
95SSE_HELPER_B(pcmpgtb, FCMPGTB)
96SSE_HELPER_W(pcmpgtw, FCMPGTW)
97SSE_HELPER_L(pcmpgtl, FCMPGTL)
98
99SSE_HELPER_B(pcmpeqb, FCMPEQ)
100SSE_HELPER_W(pcmpeqw, FCMPEQ)
101SSE_HELPER_L(pcmpeql, FCMPEQ)
102
103SSE_HELPER_W(pmullw, FMULLW)
104#if SHIFT == 0
105SSE_HELPER_W(pmulhrw, FMULHRW)
106#endif
107SSE_HELPER_W(pmulhuw, FMULHUW)
108SSE_HELPER_W(pmulhw, FMULHW)
109
110SSE_HELPER_B(pavgb, FAVG)
111SSE_HELPER_W(pavgw, FAVG)
112
113DEF_HELPER_2(glue(pmuludq, SUFFIX), void, Reg, Reg)
114DEF_HELPER_2(glue(pmaddwd, SUFFIX), void, Reg, Reg)
115
116DEF_HELPER_2(glue(psadbw, SUFFIX), void, Reg, Reg)
117DEF_HELPER_3(glue(maskmov, SUFFIX), void, Reg, Reg, tl)
118DEF_HELPER_2(glue(movl_mm_T0, SUFFIX), void, Reg, i32)
119#ifdef TARGET_X86_64
120DEF_HELPER_2(glue(movq_mm_T0, SUFFIX), void, Reg, i64)
121#endif
122
123#if SHIFT == 0
124DEF_HELPER_3(glue(pshufw, SUFFIX), void, Reg, Reg, int)
125#else
126DEF_HELPER_3(shufps, void, Reg, Reg, int)
127DEF_HELPER_3(shufpd, void, Reg, Reg, int)
128DEF_HELPER_3(glue(pshufd, SUFFIX), void, Reg, Reg, int)
129DEF_HELPER_3(glue(pshuflw, SUFFIX), void, Reg, Reg, int)
130DEF_HELPER_3(glue(pshufhw, SUFFIX), void, Reg, Reg, int)
131#endif
132
133#if SHIFT == 1
134/* FPU ops */
135/* XXX: not accurate */
136
137#define SSE_HELPER_S(name, F)\
138    DEF_HELPER_2(name ## ps , void, Reg, Reg)        \
139    DEF_HELPER_2(name ## ss , void, Reg, Reg)        \
140    DEF_HELPER_2(name ## pd , void, Reg, Reg)        \
141    DEF_HELPER_2(name ## sd , void, Reg, Reg)
142
143SSE_HELPER_S(add, FPU_ADD)
144SSE_HELPER_S(sub, FPU_SUB)
145SSE_HELPER_S(mul, FPU_MUL)
146SSE_HELPER_S(div, FPU_DIV)
147SSE_HELPER_S(min, FPU_MIN)
148SSE_HELPER_S(max, FPU_MAX)
149SSE_HELPER_S(sqrt, FPU_SQRT)
150
151
152DEF_HELPER_2(cvtps2pd, void, Reg, Reg)
153DEF_HELPER_2(cvtpd2ps, void, Reg, Reg)
154DEF_HELPER_2(cvtss2sd, void, Reg, Reg)
155DEF_HELPER_2(cvtsd2ss, void, Reg, Reg)
156DEF_HELPER_2(cvtdq2ps, void, Reg, Reg)
157DEF_HELPER_2(cvtdq2pd, void, Reg, Reg)
158DEF_HELPER_2(cvtpi2ps, void, XMMReg, MMXReg)
159DEF_HELPER_2(cvtpi2pd, void, XMMReg, MMXReg)
160DEF_HELPER_2(cvtsi2ss, void, XMMReg, i32)
161DEF_HELPER_2(cvtsi2sd, void, XMMReg, i32)
162
163#ifdef TARGET_X86_64
164DEF_HELPER_2(cvtsq2ss, void, XMMReg, i64)
165DEF_HELPER_2(cvtsq2sd, void, XMMReg, i64)
166#endif
167
168DEF_HELPER_2(cvtps2dq, void, XMMReg, XMMReg)
169DEF_HELPER_2(cvtpd2dq, void, XMMReg, XMMReg)
170DEF_HELPER_2(cvtps2pi, void, MMXReg, XMMReg)
171DEF_HELPER_2(cvtpd2pi, void, MMXReg, XMMReg)
172DEF_HELPER_1(cvtss2si, s32, XMMReg)
173DEF_HELPER_1(cvtsd2si, s32, XMMReg)
174#ifdef TARGET_X86_64
175DEF_HELPER_1(cvtss2sq, s64, XMMReg)
176DEF_HELPER_1(cvtsd2sq, s64, XMMReg)
177#endif
178
179DEF_HELPER_2(cvttps2dq, void, XMMReg, XMMReg)
180DEF_HELPER_2(cvttpd2dq, void, XMMReg, XMMReg)
181DEF_HELPER_2(cvttps2pi, void, MMXReg, XMMReg)
182DEF_HELPER_2(cvttpd2pi, void, MMXReg, XMMReg)
183DEF_HELPER_1(cvttss2si, s32, XMMReg)
184DEF_HELPER_1(cvttsd2si, s32, XMMReg)
185#ifdef TARGET_X86_64
186DEF_HELPER_1(cvttss2sq, s64, XMMReg)
187DEF_HELPER_1(cvttsd2sq, s64, XMMReg)
188#endif
189
190DEF_HELPER_2(rsqrtps, void, XMMReg, XMMReg)
191DEF_HELPER_2(rsqrtss, void, XMMReg, XMMReg)
192DEF_HELPER_2(rcpps, void, XMMReg, XMMReg)
193DEF_HELPER_2(rcpss, void, XMMReg, XMMReg)
194DEF_HELPER_2(haddps, void, XMMReg, XMMReg)
195DEF_HELPER_2(haddpd, void, XMMReg, XMMReg)
196DEF_HELPER_2(hsubps, void, XMMReg, XMMReg)
197DEF_HELPER_2(hsubpd, void, XMMReg, XMMReg)
198DEF_HELPER_2(addsubps, void, XMMReg, XMMReg)
199DEF_HELPER_2(addsubpd, void, XMMReg, XMMReg)
200
201#define SSE_HELPER_CMP(name, F)\
202    DEF_HELPER_2( name ## ps , void, Reg, Reg)        \
203    DEF_HELPER_2( name ## ss , void, Reg, Reg)        \
204    DEF_HELPER_2( name ## pd , void, Reg, Reg)        \
205    DEF_HELPER_2( name ## sd , void, Reg, Reg)
206
207SSE_HELPER_CMP(cmpeq, FPU_CMPEQ)
208SSE_HELPER_CMP(cmplt, FPU_CMPLT)
209SSE_HELPER_CMP(cmple, FPU_CMPLE)
210SSE_HELPER_CMP(cmpunord, FPU_CMPUNORD)
211SSE_HELPER_CMP(cmpneq, FPU_CMPNEQ)
212SSE_HELPER_CMP(cmpnlt, FPU_CMPNLT)
213SSE_HELPER_CMP(cmpnle, FPU_CMPNLE)
214SSE_HELPER_CMP(cmpord, FPU_CMPORD)
215
216DEF_HELPER_2(ucomiss, void, Reg, Reg)
217DEF_HELPER_2(comiss, void, Reg, Reg)
218DEF_HELPER_2(ucomisd, void, Reg, Reg)
219DEF_HELPER_2(comisd, void, Reg, Reg)
220DEF_HELPER_1(movmskps, i32, Reg)
221DEF_HELPER_1(movmskpd, i32, Reg)
222#endif
223
224DEF_HELPER_1(glue(pmovmskb, SUFFIX), i32, Reg)
225DEF_HELPER_2(glue(packsswb, SUFFIX), void, Reg, Reg)
226DEF_HELPER_2(glue(packuswb, SUFFIX), void, Reg, Reg)
227DEF_HELPER_2(glue(packssdw, SUFFIX), void, Reg, Reg)
228#define UNPCK_OP(base_name, base)                               \
229    DEF_HELPER_2(glue(punpck ## base_name ## bw, SUFFIX) , void, Reg, Reg) \
230    DEF_HELPER_2(glue(punpck ## base_name ## wd, SUFFIX) , void, Reg, Reg) \
231    DEF_HELPER_2(glue(punpck ## base_name ## dq, SUFFIX) , void, Reg, Reg)
232
233UNPCK_OP(l, 0)
234UNPCK_OP(h, 1)
235
236#if SHIFT == 1
237DEF_HELPER_2(glue(punpcklqdq, SUFFIX), void, Reg, Reg)
238DEF_HELPER_2(glue(punpckhqdq, SUFFIX), void, Reg, Reg)
239#endif
240
241/* 3DNow! float ops */
242#if SHIFT == 0
243DEF_HELPER_2(pi2fd, void, MMXReg, MMXReg)
244DEF_HELPER_2(pi2fw, void, MMXReg, MMXReg)
245DEF_HELPER_2(pf2id, void, MMXReg, MMXReg)
246DEF_HELPER_2(pf2iw, void, MMXReg, MMXReg)
247DEF_HELPER_2(pfacc, void, MMXReg, MMXReg)
248DEF_HELPER_2(pfadd, void, MMXReg, MMXReg)
249DEF_HELPER_2(pfcmpeq, void, MMXReg, MMXReg)
250DEF_HELPER_2(pfcmpge, void, MMXReg, MMXReg)
251DEF_HELPER_2(pfcmpgt, void, MMXReg, MMXReg)
252DEF_HELPER_2(pfmax, void, MMXReg, MMXReg)
253DEF_HELPER_2(pfmin, void, MMXReg, MMXReg)
254DEF_HELPER_2(pfmul, void, MMXReg, MMXReg)
255DEF_HELPER_2(pfnacc, void, MMXReg, MMXReg)
256DEF_HELPER_2(pfpnacc, void, MMXReg, MMXReg)
257DEF_HELPER_2(pfrcp, void, MMXReg, MMXReg)
258DEF_HELPER_2(pfrsqrt, void, MMXReg, MMXReg)
259DEF_HELPER_2(pfsub, void, MMXReg, MMXReg)
260DEF_HELPER_2(pfsubr, void, MMXReg, MMXReg)
261DEF_HELPER_2(pswapd, void, MMXReg, MMXReg)
262#endif
263
264/* SSSE3 op helpers */
265DEF_HELPER_2(glue(phaddw, SUFFIX), void, Reg, Reg)
266DEF_HELPER_2(glue(phaddd, SUFFIX), void, Reg, Reg)
267DEF_HELPER_2(glue(phaddsw, SUFFIX), void, Reg, Reg)
268DEF_HELPER_2(glue(phsubw, SUFFIX), void, Reg, Reg)
269DEF_HELPER_2(glue(phsubd, SUFFIX), void, Reg, Reg)
270DEF_HELPER_2(glue(phsubsw, SUFFIX), void, Reg, Reg)
271DEF_HELPER_2(glue(pabsb, SUFFIX), void, Reg, Reg)
272DEF_HELPER_2(glue(pabsw, SUFFIX), void, Reg, Reg)
273DEF_HELPER_2(glue(pabsd, SUFFIX), void, Reg, Reg)
274DEF_HELPER_2(glue(pmaddubsw, SUFFIX), void, Reg, Reg)
275DEF_HELPER_2(glue(pmulhrsw, SUFFIX), void, Reg, Reg)
276DEF_HELPER_2(glue(pshufb, SUFFIX), void, Reg, Reg)
277DEF_HELPER_2(glue(psignb, SUFFIX), void, Reg, Reg)
278DEF_HELPER_2(glue(psignw, SUFFIX), void, Reg, Reg)
279DEF_HELPER_2(glue(psignd, SUFFIX), void, Reg, Reg)
280DEF_HELPER_3(glue(palignr, SUFFIX), void, Reg, Reg, s32)
281
282/* SSE4.1 op helpers */
283#if SHIFT == 1
284DEF_HELPER_2(glue(pblendvb, SUFFIX), void, Reg, Reg)
285DEF_HELPER_2(glue(blendvps, SUFFIX), void, Reg, Reg)
286DEF_HELPER_2(glue(blendvpd, SUFFIX), void, Reg, Reg)
287DEF_HELPER_2(glue(ptest, SUFFIX), void, Reg, Reg)
288DEF_HELPER_2(glue(pmovsxbw, SUFFIX), void, Reg, Reg)
289DEF_HELPER_2(glue(pmovsxbd, SUFFIX), void, Reg, Reg)
290DEF_HELPER_2(glue(pmovsxbq, SUFFIX), void, Reg, Reg)
291DEF_HELPER_2(glue(pmovsxwd, SUFFIX), void, Reg, Reg)
292DEF_HELPER_2(glue(pmovsxwq, SUFFIX), void, Reg, Reg)
293DEF_HELPER_2(glue(pmovsxdq, SUFFIX), void, Reg, Reg)
294DEF_HELPER_2(glue(pmovzxbw, SUFFIX), void, Reg, Reg)
295DEF_HELPER_2(glue(pmovzxbd, SUFFIX), void, Reg, Reg)
296DEF_HELPER_2(glue(pmovzxbq, SUFFIX), void, Reg, Reg)
297DEF_HELPER_2(glue(pmovzxwd, SUFFIX), void, Reg, Reg)
298DEF_HELPER_2(glue(pmovzxwq, SUFFIX), void, Reg, Reg)
299DEF_HELPER_2(glue(pmovzxdq, SUFFIX), void, Reg, Reg)
300DEF_HELPER_2(glue(pmuldq, SUFFIX), void, Reg, Reg)
301DEF_HELPER_2(glue(pcmpeqq, SUFFIX), void, Reg, Reg)
302DEF_HELPER_2(glue(packusdw, SUFFIX), void, Reg, Reg)
303DEF_HELPER_2(glue(pminsb, SUFFIX), void, Reg, Reg)
304DEF_HELPER_2(glue(pminsd, SUFFIX), void, Reg, Reg)
305DEF_HELPER_2(glue(pminuw, SUFFIX), void, Reg, Reg)
306DEF_HELPER_2(glue(pminud, SUFFIX), void, Reg, Reg)
307DEF_HELPER_2(glue(pmaxsb, SUFFIX), void, Reg, Reg)
308DEF_HELPER_2(glue(pmaxsd, SUFFIX), void, Reg, Reg)
309DEF_HELPER_2(glue(pmaxuw, SUFFIX), void, Reg, Reg)
310DEF_HELPER_2(glue(pmaxud, SUFFIX), void, Reg, Reg)
311DEF_HELPER_2(glue(pmulld, SUFFIX), void, Reg, Reg)
312DEF_HELPER_2(glue(phminposuw, SUFFIX), void, Reg, Reg)
313DEF_HELPER_3(glue(roundps, SUFFIX), void, Reg, Reg, i32)
314DEF_HELPER_3(glue(roundpd, SUFFIX), void, Reg, Reg, i32)
315DEF_HELPER_3(glue(roundss, SUFFIX), void, Reg, Reg, i32)
316DEF_HELPER_3(glue(roundsd, SUFFIX), void, Reg, Reg, i32)
317DEF_HELPER_3(glue(blendps, SUFFIX), void, Reg, Reg, i32)
318DEF_HELPER_3(glue(blendpd, SUFFIX), void, Reg, Reg, i32)
319DEF_HELPER_3(glue(pblendw, SUFFIX), void, Reg, Reg, i32)
320DEF_HELPER_3(glue(dpps, SUFFIX), void, Reg, Reg, i32)
321DEF_HELPER_3(glue(dppd, SUFFIX), void, Reg, Reg, i32)
322DEF_HELPER_3(glue(mpsadbw, SUFFIX), void, Reg, Reg, i32)
323#endif
324
325/* SSE4.2 op helpers */
326#if SHIFT == 1
327DEF_HELPER_2(glue(pcmpgtq, SUFFIX), void, Reg, Reg)
328DEF_HELPER_3(glue(pcmpestri, SUFFIX), void, Reg, Reg, i32)
329DEF_HELPER_3(glue(pcmpestrm, SUFFIX), void, Reg, Reg, i32)
330DEF_HELPER_3(glue(pcmpistri, SUFFIX), void, Reg, Reg, i32)
331DEF_HELPER_3(glue(pcmpistrm, SUFFIX), void, Reg, Reg, i32)
332DEF_HELPER_3(crc32, tl, i32, tl, i32)
333DEF_HELPER_2(popcnt, tl, tl, i32)
334#endif
335
336#undef SHIFT
337#undef Reg
338#undef SUFFIX
339
340#undef SSE_HELPER_B
341#undef SSE_HELPER_W
342#undef SSE_HELPER_L
343#undef SSE_HELPER_Q
344#undef SSE_HELPER_S
345#undef SSE_HELPER_CMP
346#undef UNPCK_OP
347