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, see <http://www.gnu.org/licenses/>.
18 */
19#if SHIFT == 0
20#define Reg MMXReg
21#define SUFFIX _mmx
22#else
23#define Reg XMMReg
24#define SUFFIX _xmm
25#endif
26
27#define dh_alias_Reg ptr
28#define dh_alias_XMMReg ptr
29#define dh_alias_MMXReg ptr
30#define dh_ctype_Reg Reg *
31#define dh_ctype_XMMReg XMMReg *
32#define dh_ctype_MMXReg MMXReg *
33#define dh_is_signed_Reg dh_is_signed_ptr
34#define dh_is_signed_XMMReg dh_is_signed_ptr
35#define dh_is_signed_MMXReg dh_is_signed_ptr
36
37DEF_HELPER_3(glue(psrlw, SUFFIX), void, env, Reg, Reg)
38DEF_HELPER_3(glue(psraw, SUFFIX), void, env, Reg, Reg)
39DEF_HELPER_3(glue(psllw, SUFFIX), void, env, Reg, Reg)
40DEF_HELPER_3(glue(psrld, SUFFIX), void, env, Reg, Reg)
41DEF_HELPER_3(glue(psrad, SUFFIX), void, env, Reg, Reg)
42DEF_HELPER_3(glue(pslld, SUFFIX), void, env, Reg, Reg)
43DEF_HELPER_3(glue(psrlq, SUFFIX), void, env, Reg, Reg)
44DEF_HELPER_3(glue(psllq, SUFFIX), void, env, Reg, Reg)
45
46#if SHIFT == 1
47DEF_HELPER_3(glue(psrldq, SUFFIX), void, env, Reg, Reg)
48DEF_HELPER_3(glue(pslldq, SUFFIX), void, env, Reg, Reg)
49#endif
50
51#define SSE_HELPER_B(name, F)\
52    DEF_HELPER_3(glue(name, SUFFIX), void, env, Reg, Reg)
53
54#define SSE_HELPER_W(name, F)\
55    DEF_HELPER_3(glue(name, SUFFIX), void, env, Reg, Reg)
56
57#define SSE_HELPER_L(name, F)\
58    DEF_HELPER_3(glue(name, SUFFIX), void, env, Reg, Reg)
59
60#define SSE_HELPER_Q(name, F)\
61    DEF_HELPER_3(glue(name, SUFFIX), void, env, Reg, Reg)
62
63SSE_HELPER_B(paddb, FADD)
64SSE_HELPER_W(paddw, FADD)
65SSE_HELPER_L(paddl, FADD)
66SSE_HELPER_Q(paddq, FADD)
67
68SSE_HELPER_B(psubb, FSUB)
69SSE_HELPER_W(psubw, FSUB)
70SSE_HELPER_L(psubl, FSUB)
71SSE_HELPER_Q(psubq, FSUB)
72
73SSE_HELPER_B(paddusb, FADDUB)
74SSE_HELPER_B(paddsb, FADDSB)
75SSE_HELPER_B(psubusb, FSUBUB)
76SSE_HELPER_B(psubsb, FSUBSB)
77
78SSE_HELPER_W(paddusw, FADDUW)
79SSE_HELPER_W(paddsw, FADDSW)
80SSE_HELPER_W(psubusw, FSUBUW)
81SSE_HELPER_W(psubsw, FSUBSW)
82
83SSE_HELPER_B(pminub, FMINUB)
84SSE_HELPER_B(pmaxub, FMAXUB)
85
86SSE_HELPER_W(pminsw, FMINSW)
87SSE_HELPER_W(pmaxsw, FMAXSW)
88
89SSE_HELPER_Q(pand, FAND)
90SSE_HELPER_Q(pandn, FANDN)
91SSE_HELPER_Q(por, FOR)
92SSE_HELPER_Q(pxor, FXOR)
93
94SSE_HELPER_B(pcmpgtb, FCMPGTB)
95SSE_HELPER_W(pcmpgtw, FCMPGTW)
96SSE_HELPER_L(pcmpgtl, FCMPGTL)
97
98SSE_HELPER_B(pcmpeqb, FCMPEQ)
99SSE_HELPER_W(pcmpeqw, FCMPEQ)
100SSE_HELPER_L(pcmpeql, FCMPEQ)
101
102SSE_HELPER_W(pmullw, FMULLW)
103#if SHIFT == 0
104SSE_HELPER_W(pmulhrw, FMULHRW)
105#endif
106SSE_HELPER_W(pmulhuw, FMULHUW)
107SSE_HELPER_W(pmulhw, FMULHW)
108
109SSE_HELPER_B(pavgb, FAVG)
110SSE_HELPER_W(pavgw, FAVG)
111
112DEF_HELPER_3(glue(pmuludq, SUFFIX), void, env, Reg, Reg)
113DEF_HELPER_3(glue(pmaddwd, SUFFIX), void, env, Reg, Reg)
114
115DEF_HELPER_3(glue(psadbw, SUFFIX), void, env, Reg, Reg)
116DEF_HELPER_4(glue(maskmov, SUFFIX), void, env, Reg, Reg, tl)
117DEF_HELPER_2(glue(movl_mm_T0, SUFFIX), void, Reg, i32)
118#ifdef TARGET_X86_64
119DEF_HELPER_2(glue(movq_mm_T0, SUFFIX), void, Reg, i64)
120#endif
121
122#if SHIFT == 0
123DEF_HELPER_3(glue(pshufw, SUFFIX), void, Reg, Reg, int)
124#else
125DEF_HELPER_3(shufps, void, Reg, Reg, int)
126DEF_HELPER_3(shufpd, void, Reg, Reg, int)
127DEF_HELPER_3(glue(pshufd, SUFFIX), void, Reg, Reg, int)
128DEF_HELPER_3(glue(pshuflw, SUFFIX), void, Reg, Reg, int)
129DEF_HELPER_3(glue(pshufhw, SUFFIX), void, Reg, Reg, int)
130#endif
131
132#if SHIFT == 1
133/* FPU ops */
134/* XXX: not accurate */
135
136#define SSE_HELPER_S(name, F)                            \
137    DEF_HELPER_3(name ## ps, void, env, Reg, Reg)        \
138    DEF_HELPER_3(name ## ss, void, env, Reg, Reg)        \
139    DEF_HELPER_3(name ## pd, void, env, Reg, Reg)        \
140    DEF_HELPER_3(name ## sd, void, env, Reg, Reg)
141
142SSE_HELPER_S(add, FPU_ADD)
143SSE_HELPER_S(sub, FPU_SUB)
144SSE_HELPER_S(mul, FPU_MUL)
145SSE_HELPER_S(div, FPU_DIV)
146SSE_HELPER_S(min, FPU_MIN)
147SSE_HELPER_S(max, FPU_MAX)
148SSE_HELPER_S(sqrt, FPU_SQRT)
149
150
151DEF_HELPER_3(cvtps2pd, void, env, Reg, Reg)
152DEF_HELPER_3(cvtpd2ps, void, env, Reg, Reg)
153DEF_HELPER_3(cvtss2sd, void, env, Reg, Reg)
154DEF_HELPER_3(cvtsd2ss, void, env, Reg, Reg)
155DEF_HELPER_3(cvtdq2ps, void, env, Reg, Reg)
156DEF_HELPER_3(cvtdq2pd, void, env, Reg, Reg)
157DEF_HELPER_3(cvtpi2ps, void, env, XMMReg, MMXReg)
158DEF_HELPER_3(cvtpi2pd, void, env, XMMReg, MMXReg)
159DEF_HELPER_3(cvtsi2ss, void, env, XMMReg, i32)
160DEF_HELPER_3(cvtsi2sd, void, env, XMMReg, i32)
161
162#ifdef TARGET_X86_64
163DEF_HELPER_3(cvtsq2ss, void, env, XMMReg, i64)
164DEF_HELPER_3(cvtsq2sd, void, env, XMMReg, i64)
165#endif
166
167DEF_HELPER_3(cvtps2dq, void, env, XMMReg, XMMReg)
168DEF_HELPER_3(cvtpd2dq, void, env, XMMReg, XMMReg)
169DEF_HELPER_3(cvtps2pi, void, env, MMXReg, XMMReg)
170DEF_HELPER_3(cvtpd2pi, void, env, MMXReg, XMMReg)
171DEF_HELPER_2(cvtss2si, s32, env, XMMReg)
172DEF_HELPER_2(cvtsd2si, s32, env, XMMReg)
173#ifdef TARGET_X86_64
174DEF_HELPER_2(cvtss2sq, s64, env, XMMReg)
175DEF_HELPER_2(cvtsd2sq, s64, env, XMMReg)
176#endif
177
178DEF_HELPER_3(cvttps2dq, void, env, XMMReg, XMMReg)
179DEF_HELPER_3(cvttpd2dq, void, env, XMMReg, XMMReg)
180DEF_HELPER_3(cvttps2pi, void, env, MMXReg, XMMReg)
181DEF_HELPER_3(cvttpd2pi, void, env, MMXReg, XMMReg)
182DEF_HELPER_2(cvttss2si, s32, env, XMMReg)
183DEF_HELPER_2(cvttsd2si, s32, env, XMMReg)
184#ifdef TARGET_X86_64
185DEF_HELPER_2(cvttss2sq, s64, env, XMMReg)
186DEF_HELPER_2(cvttsd2sq, s64, env, XMMReg)
187#endif
188
189DEF_HELPER_3(rsqrtps, void, env, XMMReg, XMMReg)
190DEF_HELPER_3(rsqrtss, void, env, XMMReg, XMMReg)
191DEF_HELPER_3(rcpps, void, env, XMMReg, XMMReg)
192DEF_HELPER_3(rcpss, void, env, XMMReg, XMMReg)
193DEF_HELPER_3(extrq_r, void, env, XMMReg, XMMReg)
194DEF_HELPER_4(extrq_i, void, env, XMMReg, int, int)
195DEF_HELPER_3(insertq_r, void, env, XMMReg, XMMReg)
196DEF_HELPER_4(insertq_i, void, env, XMMReg, int, int)
197DEF_HELPER_3(haddps, void, env, XMMReg, XMMReg)
198DEF_HELPER_3(haddpd, void, env, XMMReg, XMMReg)
199DEF_HELPER_3(hsubps, void, env, XMMReg, XMMReg)
200DEF_HELPER_3(hsubpd, void, env, XMMReg, XMMReg)
201DEF_HELPER_3(addsubps, void, env, XMMReg, XMMReg)
202DEF_HELPER_3(addsubpd, void, env, XMMReg, XMMReg)
203
204#define SSE_HELPER_CMP(name, F)                           \
205    DEF_HELPER_3(name ## ps, void, env, Reg, Reg)         \
206    DEF_HELPER_3(name ## ss, void, env, Reg, Reg)         \
207    DEF_HELPER_3(name ## pd, void, env, Reg, Reg)         \
208    DEF_HELPER_3(name ## sd, void, env, Reg, Reg)
209
210SSE_HELPER_CMP(cmpeq, FPU_CMPEQ)
211SSE_HELPER_CMP(cmplt, FPU_CMPLT)
212SSE_HELPER_CMP(cmple, FPU_CMPLE)
213SSE_HELPER_CMP(cmpunord, FPU_CMPUNORD)
214SSE_HELPER_CMP(cmpneq, FPU_CMPNEQ)
215SSE_HELPER_CMP(cmpnlt, FPU_CMPNLT)
216SSE_HELPER_CMP(cmpnle, FPU_CMPNLE)
217SSE_HELPER_CMP(cmpord, FPU_CMPORD)
218
219DEF_HELPER_3(ucomiss, void, env, Reg, Reg)
220DEF_HELPER_3(comiss, void, env, Reg, Reg)
221DEF_HELPER_3(ucomisd, void, env, Reg, Reg)
222DEF_HELPER_3(comisd, void, env, Reg, Reg)
223DEF_HELPER_2(movmskps, i32, env, Reg)
224DEF_HELPER_2(movmskpd, i32, env, Reg)
225#endif
226
227DEF_HELPER_2(glue(pmovmskb, SUFFIX), i32, env, Reg)
228DEF_HELPER_3(glue(packsswb, SUFFIX), void, env, Reg, Reg)
229DEF_HELPER_3(glue(packuswb, SUFFIX), void, env, Reg, Reg)
230DEF_HELPER_3(glue(packssdw, SUFFIX), void, env, Reg, Reg)
231#define UNPCK_OP(base_name, base)                                       \
232    DEF_HELPER_3(glue(punpck ## base_name ## bw, SUFFIX), void, env, Reg, Reg) \
233    DEF_HELPER_3(glue(punpck ## base_name ## wd, SUFFIX), void, env, Reg, Reg) \
234    DEF_HELPER_3(glue(punpck ## base_name ## dq, SUFFIX), void, env, Reg, Reg)
235
236UNPCK_OP(l, 0)
237UNPCK_OP(h, 1)
238
239#if SHIFT == 1
240DEF_HELPER_3(glue(punpcklqdq, SUFFIX), void, env, Reg, Reg)
241DEF_HELPER_3(glue(punpckhqdq, SUFFIX), void, env, Reg, Reg)
242#endif
243
244/* 3DNow! float ops */
245#if SHIFT == 0
246DEF_HELPER_3(pi2fd, void, env, MMXReg, MMXReg)
247DEF_HELPER_3(pi2fw, void, env, MMXReg, MMXReg)
248DEF_HELPER_3(pf2id, void, env, MMXReg, MMXReg)
249DEF_HELPER_3(pf2iw, void, env, MMXReg, MMXReg)
250DEF_HELPER_3(pfacc, void, env, MMXReg, MMXReg)
251DEF_HELPER_3(pfadd, void, env, MMXReg, MMXReg)
252DEF_HELPER_3(pfcmpeq, void, env, MMXReg, MMXReg)
253DEF_HELPER_3(pfcmpge, void, env, MMXReg, MMXReg)
254DEF_HELPER_3(pfcmpgt, void, env, MMXReg, MMXReg)
255DEF_HELPER_3(pfmax, void, env, MMXReg, MMXReg)
256DEF_HELPER_3(pfmin, void, env, MMXReg, MMXReg)
257DEF_HELPER_3(pfmul, void, env, MMXReg, MMXReg)
258DEF_HELPER_3(pfnacc, void, env, MMXReg, MMXReg)
259DEF_HELPER_3(pfpnacc, void, env, MMXReg, MMXReg)
260DEF_HELPER_3(pfrcp, void, env, MMXReg, MMXReg)
261DEF_HELPER_3(pfrsqrt, void, env, MMXReg, MMXReg)
262DEF_HELPER_3(pfsub, void, env, MMXReg, MMXReg)
263DEF_HELPER_3(pfsubr, void, env, MMXReg, MMXReg)
264DEF_HELPER_3(pswapd, void, env, MMXReg, MMXReg)
265#endif
266
267/* SSSE3 op helpers */
268DEF_HELPER_3(glue(phaddw, SUFFIX), void, env, Reg, Reg)
269DEF_HELPER_3(glue(phaddd, SUFFIX), void, env, Reg, Reg)
270DEF_HELPER_3(glue(phaddsw, SUFFIX), void, env, Reg, Reg)
271DEF_HELPER_3(glue(phsubw, SUFFIX), void, env, Reg, Reg)
272DEF_HELPER_3(glue(phsubd, SUFFIX), void, env, Reg, Reg)
273DEF_HELPER_3(glue(phsubsw, SUFFIX), void, env, Reg, Reg)
274DEF_HELPER_3(glue(pabsb, SUFFIX), void, env, Reg, Reg)
275DEF_HELPER_3(glue(pabsw, SUFFIX), void, env, Reg, Reg)
276DEF_HELPER_3(glue(pabsd, SUFFIX), void, env, Reg, Reg)
277DEF_HELPER_3(glue(pmaddubsw, SUFFIX), void, env, Reg, Reg)
278DEF_HELPER_3(glue(pmulhrsw, SUFFIX), void, env, Reg, Reg)
279DEF_HELPER_3(glue(pshufb, SUFFIX), void, env, Reg, Reg)
280DEF_HELPER_3(glue(psignb, SUFFIX), void, env, Reg, Reg)
281DEF_HELPER_3(glue(psignw, SUFFIX), void, env, Reg, Reg)
282DEF_HELPER_3(glue(psignd, SUFFIX), void, env, Reg, Reg)
283DEF_HELPER_4(glue(palignr, SUFFIX), void, env, Reg, Reg, s32)
284
285/* SSE4.1 op helpers */
286#if SHIFT == 1
287DEF_HELPER_3(glue(pblendvb, SUFFIX), void, env, Reg, Reg)
288DEF_HELPER_3(glue(blendvps, SUFFIX), void, env, Reg, Reg)
289DEF_HELPER_3(glue(blendvpd, SUFFIX), void, env, Reg, Reg)
290DEF_HELPER_3(glue(ptest, SUFFIX), void, env, Reg, Reg)
291DEF_HELPER_3(glue(pmovsxbw, SUFFIX), void, env, Reg, Reg)
292DEF_HELPER_3(glue(pmovsxbd, SUFFIX), void, env, Reg, Reg)
293DEF_HELPER_3(glue(pmovsxbq, SUFFIX), void, env, Reg, Reg)
294DEF_HELPER_3(glue(pmovsxwd, SUFFIX), void, env, Reg, Reg)
295DEF_HELPER_3(glue(pmovsxwq, SUFFIX), void, env, Reg, Reg)
296DEF_HELPER_3(glue(pmovsxdq, SUFFIX), void, env, Reg, Reg)
297DEF_HELPER_3(glue(pmovzxbw, SUFFIX), void, env, Reg, Reg)
298DEF_HELPER_3(glue(pmovzxbd, SUFFIX), void, env, Reg, Reg)
299DEF_HELPER_3(glue(pmovzxbq, SUFFIX), void, env, Reg, Reg)
300DEF_HELPER_3(glue(pmovzxwd, SUFFIX), void, env, Reg, Reg)
301DEF_HELPER_3(glue(pmovzxwq, SUFFIX), void, env, Reg, Reg)
302DEF_HELPER_3(glue(pmovzxdq, SUFFIX), void, env, Reg, Reg)
303DEF_HELPER_3(glue(pmuldq, SUFFIX), void, env, Reg, Reg)
304DEF_HELPER_3(glue(pcmpeqq, SUFFIX), void, env, Reg, Reg)
305DEF_HELPER_3(glue(packusdw, SUFFIX), void, env, Reg, Reg)
306DEF_HELPER_3(glue(pminsb, SUFFIX), void, env, Reg, Reg)
307DEF_HELPER_3(glue(pminsd, SUFFIX), void, env, Reg, Reg)
308DEF_HELPER_3(glue(pminuw, SUFFIX), void, env, Reg, Reg)
309DEF_HELPER_3(glue(pminud, SUFFIX), void, env, Reg, Reg)
310DEF_HELPER_3(glue(pmaxsb, SUFFIX), void, env, Reg, Reg)
311DEF_HELPER_3(glue(pmaxsd, SUFFIX), void, env, Reg, Reg)
312DEF_HELPER_3(glue(pmaxuw, SUFFIX), void, env, Reg, Reg)
313DEF_HELPER_3(glue(pmaxud, SUFFIX), void, env, Reg, Reg)
314DEF_HELPER_3(glue(pmulld, SUFFIX), void, env, Reg, Reg)
315DEF_HELPER_3(glue(phminposuw, SUFFIX), void, env, Reg, Reg)
316DEF_HELPER_4(glue(roundps, SUFFIX), void, env, Reg, Reg, i32)
317DEF_HELPER_4(glue(roundpd, SUFFIX), void, env, Reg, Reg, i32)
318DEF_HELPER_4(glue(roundss, SUFFIX), void, env, Reg, Reg, i32)
319DEF_HELPER_4(glue(roundsd, SUFFIX), void, env, Reg, Reg, i32)
320DEF_HELPER_4(glue(blendps, SUFFIX), void, env, Reg, Reg, i32)
321DEF_HELPER_4(glue(blendpd, SUFFIX), void, env, Reg, Reg, i32)
322DEF_HELPER_4(glue(pblendw, SUFFIX), void, env, Reg, Reg, i32)
323DEF_HELPER_4(glue(dpps, SUFFIX), void, env, Reg, Reg, i32)
324DEF_HELPER_4(glue(dppd, SUFFIX), void, env, Reg, Reg, i32)
325DEF_HELPER_4(glue(mpsadbw, SUFFIX), void, env, Reg, Reg, i32)
326#endif
327
328/* SSE4.2 op helpers */
329#if SHIFT == 1
330DEF_HELPER_3(glue(pcmpgtq, SUFFIX), void, env, Reg, Reg)
331DEF_HELPER_4(glue(pcmpestri, SUFFIX), void, env, Reg, Reg, i32)
332DEF_HELPER_4(glue(pcmpestrm, SUFFIX), void, env, Reg, Reg, i32)
333DEF_HELPER_4(glue(pcmpistri, SUFFIX), void, env, Reg, Reg, i32)
334DEF_HELPER_4(glue(pcmpistrm, SUFFIX), void, env, Reg, Reg, i32)
335DEF_HELPER_3(crc32, tl, i32, tl, i32)
336DEF_HELPER_3(popcnt, tl, env, tl, i32)
337#endif
338
339/* AES-NI op helpers */
340#if SHIFT == 1
341DEF_HELPER_3(glue(aesdec, SUFFIX), void, env, Reg, Reg)
342DEF_HELPER_3(glue(aesdeclast, SUFFIX), void, env, Reg, Reg)
343DEF_HELPER_3(glue(aesenc, SUFFIX), void, env, Reg, Reg)
344DEF_HELPER_3(glue(aesenclast, SUFFIX), void, env, Reg, Reg)
345DEF_HELPER_3(glue(aesimc, SUFFIX), void, env, Reg, Reg)
346DEF_HELPER_4(glue(aeskeygenassist, SUFFIX), void, env, Reg, Reg, i32)
347DEF_HELPER_4(glue(pclmulqdq, SUFFIX), void, env, Reg, Reg, i32)
348#endif
349
350#undef SHIFT
351#undef Reg
352#undef SUFFIX
353
354#undef SSE_HELPER_B
355#undef SSE_HELPER_W
356#undef SSE_HELPER_L
357#undef SSE_HELPER_Q
358#undef SSE_HELPER_S
359#undef SSE_HELPER_CMP
360#undef UNPCK_OP
361