1 /*
2 
3    BLIS
4    An object-based framework for developing high-performance BLAS-like
5    libraries.
6 
7    Copyright (C) 2018, The University of Texas at Austin
8    Copyright (C) 2019, Advanced Micro Devices, Inc.
9 
10    Redistribution and use in source and binary forms, with or without
11    modification, are permitted provided that the following conditions are
12    met:
13     - Redistributions of source code must retain the above copyright
14       notice, this list of conditions and the following disclaimer.
15     - Redistributions in binary form must reproduce the above copyright
16       notice, this list of conditions and the following disclaimer in the
17       documentation and/or other materials provided with the distribution.
18     - Neither the name(s) of the copyright holder(s) nor the names of its
19       contributors may be used to endorse or promote products derived
20       from this software without specific prior written permission.
21 
22    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23    "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
25    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
26    HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27    SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28    LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29    DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30    THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32    OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 
34 */
35 
36 #ifndef BLIS_X86_ASM_MACROS_H
37 #define BLIS_X86_ASM_MACROS_H
38 
39 //
40 // Assembly macros to make inline x86 with AT&T syntax somewhat less painful
41 //
42 // "Private" macros end with _
43 //
44 
45 // Default syntax is Intel
46 #if !defined(BLIS_ASM_SYNTAX_ATT) && !defined(BLIS_ASM_SYNTAX_INTEL)
47 #define BLIS_ASM_SYNTAX_INTEL
48 #endif
49 
50 #define STRINGIFY_(...) #__VA_ARGS__
51 #define GET_MACRO_(_1_,_2_,_3_,_4_,NAME,...) NAME
52 
53 #if (defined(_WIN32) && !defined(__clang__) && !defined(__MINGW32__)) || defined(__MIC__)
54 
55 // Intel-style assembly blocks
56 
57 #define BEGIN_ASM __asm {
58 #define END_ASM(...) }
59 
60 #ifdef BLIS_ASM_SYNTAX_INTEL
61 
62 #define INSTR_4_(name,_0,_1,_2,_3) name _0,_1,_2,_3
63 #define INSTR_3_(name,_0,_1,_2) name _0,_1,_2
64 #define INSTR_2_(name,_0,_1) name _0,_1
65 #define INSTR_1_(name,_0) name _0
66 #define INSTR_0_(name) name
67 
68 #else
69 
70 #define INSTR_4_(name,_0,_1,_2,_3) name _3,_2,_1,_0
71 #define INSTR_3_(name,_0,_1,_2) name _2,_1,_0
72 #define INSTR_2_(name,_0,_1) name _1,_0
73 #define INSTR_1_(name,_0) name _0
74 #define INSTR_0_(name) name
75 
76 #endif
77 
78 #define LABEL(label) label:
79 #define REGISTER_(r) r
80 #define IMM(x) x
81 #define VAR(x) x
82 #define MASK_(x) {x}
83 #define JMP_(insn, target) insn target
84 
85 #define MEM_4_(reg,off,scale,disp) [reg + off*scale + disp]
86 #define MEM_3_(reg,off,scale) [reg + off*scale]
87 #define MEM_2_(reg,disp) [reg + disp]
88 #define MEM_1_(reg) [reg]
89 
90 #define ALIGN4 align 4
91 #define ALIGN8 align 8
92 #define ALIGN16 align 16
93 #define ALIGN32 align 32
94 
95 #else
96 
97 // GCC extended assembly with AT&T syntax
98 
99 #define COMMENT_BEGIN "#"
100 #define COMMENT_END
101 
102 #define BEGIN_ASM() __asm__ volatile (
103 #define END_ASM(...) __VA_ARGS__ );
104 
105 
106 #ifdef BLIS_ASM_SYNTAX_ATT
107 
108 #define INSTR_4_(name,_0,_1,_2,_3) STRINGIFY_(name) " " STRINGIFY_(_0,_1,_2,_3) "\n\t"
109 #define INSTR_3_(name,_0,_1,_2) STRINGIFY_(name) " " STRINGIFY_(_0,_1,_2) "\n\t"
110 #define INSTR_2_(name,_0,_1) STRINGIFY_(name) " " STRINGIFY_(_0,_1) "\n\t"
111 #define INSTR_1_(name,_0) STRINGIFY_(name) " " STRINGIFY_(_0) "\n\t"
112 #define INSTR_0_(name) STRINGIFY_(name) "\n\t"
113 
114 #else
115 
116 #define INSTR_4_(name,_0,_1,_2,_3) STRINGIFY_(name) " " STRINGIFY_(_3,_2,_1,_0) "\n\t"
117 #define INSTR_3_(name,_0,_1,_2) STRINGIFY_(name) " " STRINGIFY_(_2,_1,_0) "\n\t"
118 #define INSTR_2_(name,_0,_1) STRINGIFY_(name) " " STRINGIFY_(_1,_0) "\n\t"
119 #define INSTR_1_(name,_0) STRINGIFY_(name) " " STRINGIFY_(_0) "\n\t"
120 #define INSTR_0_(name) STRINGIFY_(name) "\n\t"
121 
122 #endif
123 
124 #if BLIS_OS_OSX
125 
126 #define LABEL_(label) "L" STRINGIFY_(label) "%="
127 
128 #else
129 
130 #define LABEL_(label) ".L" STRINGIFY_(label) "%="
131 
132 #endif
133 
134 #define REGISTER_(r) %%r
135 #define IMM(x) $##x
136 #define VAR(x) %[x]
137 #define MASK_(x) %{x%}
138 #define LABEL(target) LABEL_(target) ":\n\t"
139 #define JMP_(insn, target) STRINGIFY_(insn) " " LABEL_(target) "\n\t"
140 
141 #define MEM_4_(reg,off,scale,disp) disp(reg,off,scale)
142 #define MEM_3_(reg,off,scale) (reg,off,scale)
143 #define MEM_2_(reg,disp) disp(reg)
144 #define MEM_1_(reg) (reg)
145 
146 #define ALIGN4 ".p2align 2 \n\t"
147 #define ALIGN8 ".p2align 3 \n\t"
148 #define ALIGN16 ".p2align 4 \n\t"
149 #define ALIGN32 ".p2align 5 \n\t"
150 
151 #endif
152 
153 #define begin_asm() BEGIN_ASM()
154 #define end_asm(...) END_ASM(__VA_ARGS__)
155 
156 #define label(...) LABEL(__VA_ARGS__)
157 #define imm(...) IMM(__VA_ARGS__)
158 #define var(...) VAR(__VA_ARGS__)
159 #define align16 ALIGN16
160 #define align32 ALIGN32
161 
162 // General-purpose registers
163 
164 #define AL REGISTER_(al)
165 #define AH REGISTER_(ah)
166 #define BL REGISTER_(bl)
167 #define BH REGISTER_(bh)
168 #define CL REGISTER_(cl)
169 #define CH REGISTER_(ch)
170 #define DL REGISTER_(dl)
171 #define DH REGISTER_(dh)
172 #define R8B REGISTER_(r8b)
173 #define R9B REGISTER_(r9b)
174 #define R10B REGISTER_(r10b)
175 #define R11B REGISTER_(r11b)
176 #define R12B REGISTER_(r12b)
177 #define R13B REGISTER_(r13b)
178 #define R14B REGISTER_(r14b)
179 #define R15B REGISTER_(r15b)
180 
181 #define al AL
182 #define ah AH
183 #define bl BL
184 #define bh BH
185 #define cl CL
186 #define ch CH
187 #define dl DL
188 #define dh DH
189 #define r8b R8B
190 #define r9b R9B
191 #define r10b R10B
192 #define r11b R11B
193 #define r12b R12B
194 #define r13b R13B
195 #define r14b R14B
196 #define r15b R15B
197 
198 #define AX REGISTER_(ax)
199 #define BX REGISTER_(bx)
200 #define CX REGISTER_(cx)
201 #define DX REGISTER_(dx)
202 #define SI REGISTER_(si)
203 #define DI REGISTER_(di)
204 #define BP REGISTER_(bp)
205 #define SP REGISTER_(sp)
206 #define R8W REGISTER_(r8w)
207 #define R9W REGISTER_(r9w)
208 #define R10W REGISTER_(r10w)
209 #define R11W REGISTER_(r11w)
210 #define R12W REGISTER_(r12w)
211 #define R13W REGISTER_(r13w)
212 #define R14W REGISTER_(r14w)
213 #define R15W REGISTER_(r15w)
214 
215 #define ax AX
216 #define bx BX
217 #define cx CX
218 #define dx DX
219 #define si SI
220 #define di DI
221 #define bp BP
222 #define sp SP
223 #define r8w R8W
224 #define r9w R9W
225 #define r10w R10W
226 #define r11w R11W
227 #define r12w R12W
228 #define r13w R13W
229 #define r14w R14W
230 #define r15w R15W
231 
232 #define EAX REGISTER_(eax)
233 #define EBX REGISTER_(ebx)
234 #define ECX REGISTER_(ecx)
235 #define EDX REGISTER_(edx)
236 #define ESP REGISTER_(esp)
237 #define EBP REGISTER_(ebp)
238 #define EDI REGISTER_(edi)
239 #define ESI REGISTER_(esi)
240 #define R8D REGISTER_(r8d)
241 #define R9D REGISTER_(r9d)
242 #define R10D REGISTER_(r10d)
243 #define R11D REGISTER_(r11d)
244 #define R12D REGISTER_(r12d)
245 #define R13D REGISTER_(r13d)
246 #define R14D REGISTER_(r14d)
247 #define R15D REGISTER_(r15d)
248 
249 #define eax EAX
250 #define ebx EBX
251 #define ecx ECX
252 #define edx EDX
253 #define esp ESP
254 #define ebp EBP
255 #define edi EDI
256 #define esi ESI
257 #define r8d R8D
258 #define r9d R9D
259 #define r10d R10D
260 #define r11d R11D
261 #define r12d R12D
262 #define r13d R13D
263 #define r14d R14D
264 #define r15d R15D
265 
266 #define RAX REGISTER_(rax)
267 #define RBX REGISTER_(rbx)
268 #define RCX REGISTER_(rcx)
269 #define RDX REGISTER_(rdx)
270 #define RSP REGISTER_(rsp)
271 #define RBP REGISTER_(rbp)
272 #define RDI REGISTER_(rdi)
273 #define RSI REGISTER_(rsi)
274 #define R8 REGISTER_(r8)
275 #define R9 REGISTER_(r9)
276 #define R10 REGISTER_(r10)
277 #define R11 REGISTER_(r11)
278 #define R12 REGISTER_(r12)
279 #define R13 REGISTER_(r13)
280 #define R14 REGISTER_(r14)
281 #define R15 REGISTER_(r15)
282 
283 #define rax RAX
284 #define rbx RBX
285 #define rcx RCX
286 #define rdx RDX
287 #define rsp RSP
288 #define rbp RBP
289 #define rdi RDI
290 #define rsi RSI
291 #define r8 R8
292 #define r9 R9
293 #define r10 R10
294 #define r11 R11
295 #define r12 R12
296 #define r13 R13
297 #define r14 R14
298 #define r15 R15
299 
300 // Vector registers
301 
302 #define XMM(x) REGISTER_(Xmm##x)
303 #define YMM(x) REGISTER_(Ymm##x)
304 #define ZMM(x) REGISTER_(Zmm##x)
305 #define K(x) REGISTER_(k##x)
306 #define MASK_K(n) MASK_(K(n))
307 #define MASK_KZ(n) MASK_(K(n))MASK_(z)
308 
309 #define xmm(x) XMM(x)
310 #define ymm(x) YMM(x)
311 #define zmm(x) ZMM(x)
312 #define k(x) K(x)
313 #define mask_k(x) MASK_K(x)
314 #define mask_kz(x) MASK_KZ(x)
315 
316 #define XMM0 XMM(0)
317 #define XMM1 XMM(1)
318 #define XMM2 XMM(2)
319 #define XMM3 XMM(3)
320 #define XMM4 XMM(4)
321 #define XMM5 XMM(5)
322 #define XMM6 XMM(6)
323 #define XMM7 XMM(7)
324 #define XMM8 XMM(8)
325 #define XMM9 XMM(9)
326 #define XMM10 XMM(10)
327 #define XMM11 XMM(11)
328 #define XMM12 XMM(12)
329 #define XMM13 XMM(13)
330 #define XMM14 XMM(14)
331 #define XMM15 XMM(15)
332 #define XMM16 XMM(16)
333 #define XMM17 XMM(17)
334 #define XMM18 XMM(18)
335 #define XMM19 XMM(19)
336 #define XMM20 XMM(20)
337 #define XMM21 XMM(21)
338 #define XMM22 XMM(22)
339 #define XMM23 XMM(23)
340 #define XMM24 XMM(24)
341 #define XMM25 XMM(25)
342 #define XMM26 XMM(26)
343 #define XMM27 XMM(27)
344 #define XMM28 XMM(28)
345 #define XMM29 XMM(29)
346 #define XMM30 XMM(30)
347 #define XMM31 XMM(31)
348 
349 #define YMM0 YMM(0)
350 #define YMM1 YMM(1)
351 #define YMM2 YMM(2)
352 #define YMM3 YMM(3)
353 #define YMM4 YMM(4)
354 #define YMM5 YMM(5)
355 #define YMM6 YMM(6)
356 #define YMM7 YMM(7)
357 #define YMM8 YMM(8)
358 #define YMM9 YMM(9)
359 #define YMM10 YMM(10)
360 #define YMM11 YMM(11)
361 #define YMM12 YMM(12)
362 #define YMM13 YMM(13)
363 #define YMM14 YMM(14)
364 #define YMM15 YMM(15)
365 #define YMM16 YMM(16)
366 #define YMM17 YMM(17)
367 #define YMM18 YMM(18)
368 #define YMM19 YMM(19)
369 #define YMM20 YMM(20)
370 #define YMM21 YMM(21)
371 #define YMM22 YMM(22)
372 #define YMM23 YMM(23)
373 #define YMM24 YMM(24)
374 #define YMM25 YMM(25)
375 #define YMM26 YMM(26)
376 #define YMM27 YMM(27)
377 #define YMM28 YMM(28)
378 #define YMM29 YMM(29)
379 #define YMM30 YMM(30)
380 #define YMM31 YMM(31)
381 
382 #define ZMM0 ZMM(0)
383 #define ZMM1 ZMM(1)
384 #define ZMM2 ZMM(2)
385 #define ZMM3 ZMM(3)
386 #define ZMM4 ZMM(4)
387 #define ZMM5 ZMM(5)
388 #define ZMM6 ZMM(6)
389 #define ZMM7 ZMM(7)
390 #define ZMM8 ZMM(8)
391 #define ZMM9 ZMM(9)
392 #define ZMM10 ZMM(10)
393 #define ZMM11 ZMM(11)
394 #define ZMM12 ZMM(12)
395 #define ZMM13 ZMM(13)
396 #define ZMM14 ZMM(14)
397 #define ZMM15 ZMM(15)
398 #define ZMM16 ZMM(16)
399 #define ZMM17 ZMM(17)
400 #define ZMM18 ZMM(18)
401 #define ZMM19 ZMM(19)
402 #define ZMM20 ZMM(20)
403 #define ZMM21 ZMM(21)
404 #define ZMM22 ZMM(22)
405 #define ZMM23 ZMM(23)
406 #define ZMM24 ZMM(24)
407 #define ZMM25 ZMM(25)
408 #define ZMM26 ZMM(26)
409 #define ZMM27 ZMM(27)
410 #define ZMM28 ZMM(28)
411 #define ZMM29 ZMM(29)
412 #define ZMM30 ZMM(30)
413 #define ZMM31 ZMM(31)
414 
415 #define xmm0 xmm(0)
416 #define xmm1 xmm(1)
417 #define xmm2 xmm(2)
418 #define xmm3 xmm(3)
419 #define xmm4 xmm(4)
420 #define xmm5 xmm(5)
421 #define xmm6 xmm(6)
422 #define xmm7 xmm(7)
423 #define xmm8 xmm(8)
424 #define xmm9 xmm(9)
425 #define xmm10 xmm(10)
426 #define xmm11 xmm(11)
427 #define xmm12 xmm(12)
428 #define xmm13 xmm(13)
429 #define xmm14 xmm(14)
430 #define xmm15 xmm(15)
431 #define xmm16 xmm(16)
432 #define xmm17 xmm(17)
433 #define xmm18 xmm(18)
434 #define xmm19 xmm(19)
435 #define xmm20 xmm(20)
436 #define xmm21 xmm(21)
437 #define xmm22 xmm(22)
438 #define xmm23 xmm(23)
439 #define xmm24 xmm(24)
440 #define xmm25 xmm(25)
441 #define xmm26 xmm(26)
442 #define xmm27 xmm(27)
443 #define xmm28 xmm(28)
444 #define xmm29 xmm(29)
445 #define xmm30 xmm(30)
446 #define xmm31 xmm(31)
447 
448 #define ymm0 ymm(0)
449 #define ymm1 ymm(1)
450 #define ymm2 ymm(2)
451 #define ymm3 ymm(3)
452 #define ymm4 ymm(4)
453 #define ymm5 ymm(5)
454 #define ymm6 ymm(6)
455 #define ymm7 ymm(7)
456 #define ymm8 ymm(8)
457 #define ymm9 ymm(9)
458 #define ymm10 ymm(10)
459 #define ymm11 ymm(11)
460 #define ymm12 ymm(12)
461 #define ymm13 ymm(13)
462 #define ymm14 ymm(14)
463 #define ymm15 ymm(15)
464 #define ymm16 ymm(16)
465 #define ymm17 ymm(17)
466 #define ymm18 ymm(18)
467 #define ymm19 ymm(19)
468 #define ymm20 ymm(20)
469 #define ymm21 ymm(21)
470 #define ymm22 ymm(22)
471 #define ymm23 ymm(23)
472 #define ymm24 ymm(24)
473 #define ymm25 ymm(25)
474 #define ymm26 ymm(26)
475 #define ymm27 ymm(27)
476 #define ymm28 ymm(28)
477 #define ymm29 ymm(29)
478 #define ymm30 ymm(30)
479 #define ymm31 ymm(31)
480 
481 #define zmm0 zmm(0)
482 #define zmm1 zmm(1)
483 #define zmm2 zmm(2)
484 #define zmm3 zmm(3)
485 #define zmm4 zmm(4)
486 #define zmm5 zmm(5)
487 #define zmm6 zmm(6)
488 #define zmm7 zmm(7)
489 #define zmm8 zmm(8)
490 #define zmm9 zmm(9)
491 #define zmm10 zmm(10)
492 #define zmm11 zmm(11)
493 #define zmm12 zmm(12)
494 #define zmm13 zmm(13)
495 #define zmm14 zmm(14)
496 #define zmm15 zmm(15)
497 #define zmm16 zmm(16)
498 #define zmm17 zmm(17)
499 #define zmm18 zmm(18)
500 #define zmm19 zmm(19)
501 #define zmm20 zmm(20)
502 #define zmm21 zmm(21)
503 #define zmm22 zmm(22)
504 #define zmm23 zmm(23)
505 #define zmm24 zmm(24)
506 #define zmm25 zmm(25)
507 #define zmm26 zmm(26)
508 #define zmm27 zmm(27)
509 #define zmm28 zmm(28)
510 #define zmm29 zmm(29)
511 #define zmm30 zmm(30)
512 #define zmm31 zmm(31)
513 
514 // Memory access
515 
516 // MEM(rax)            ->     (%rax)        or [rax]
517 // MEM(rax,0x80)       -> 0x80(%rax)        or [rax + 0x80]
518 // MEM(rax,rsi,4)      ->     (%rax,%rsi,4) or [rax + rsi*4]
519 // MEM(rax,rsi,4,0x80) -> 0x80(%rax,%rsi,4) or [rax + rsi*4 + 0x80]
520 
521 #define MEM(...) GET_MACRO_(__VA_ARGS__,MEM_4_,MEM_3_,MEM_2_,MEM_1_)(__VA_ARGS__)
522 #define MEM_1TO8(...) MEM(__VA_ARGS__) MASK_(1to8)
523 #define MEM_1TO16(...) MEM(__VA_ARGS__) MASK_(1to16)
524 #define MEM_BCAST(...) MEM(__VA_ARGS__) MASK_(b)
525 
526 #define mem(...) MEM(__VA_ARGS__)
527 #define mem_1to8(...) MEM_1TO8(__VA_ARGS__)
528 #define mem_1to16(...) MEM_1TO16(__VA_ARGS__)
529 #define mem_bcast(...) MEM_BCAST(__VA_ARGS__)
530 
531 #define VAR_1TO8(...) VAR(__VA_ARGS__) MASK_(1to8)
532 #define VAR_1TO16(...) VAR(__VA_ARGS__) MASK_(1to16)
533 #define VAR_BCAST(...) VAR(__VA_ARGS__) MASK_(b)
534 
535 #define var_1to8(...) VAR_1TO8(__VA_ARGS__)
536 #define var_1to16(...) VAR_1TO16(__VA_ARGS__)
537 #define var_bcast(...) VAR_BCAST(__VA_ARGS__)
538 
539 // Instructions
540 
541 #define INSTR_(name,...) GET_MACRO_(__VA_ARGS__,INSTR_4_,INSTR_3_,INSTR_2_, \
542                                     INSTR_1_,INSTR_0_)(name,__VA_ARGS__)
543 
544 // Jumps
545 
546 #define JC(_0) JMP_(jc, _0)
547 #define JB(_0) JC(_0)
548 #define JNAE(_0) JC(_0)
549 #define JNC(_0) JMP_(jnc, _0)
550 #define JNB(_0) JNC(_0)
551 #define JAE(_0) JNC(_0)
552 
553 #define jc(_0) JC(_0)
554 #define jb(_0) JB(_0)
555 #define jnae(_0) JNAE(_0)
556 #define jnc(_0) JNC(_0)
557 #define jnb(_0) JNB(_0)
558 #define jae(_0) JAE(_0)
559 
560 #define JO(_0) JMP_(jo, _0)
561 #define JNO(_0) JMP_(jno, _0)
562 
563 #define jo(_0) JO(_0)
564 #define jno(_0) JNO(_0)
565 
566 #define JP(_0) JMP_(jp, _0)
567 #define JPE(_0) JP(_0)
568 #define JNP(_0) JMP_(jnp, _0)
569 #define JPO(_0) JNP(_0)
570 
571 #define jp(_0) JP(_0)
572 #define jpe(_0) JPE(_0)
573 #define jnp(_0) JNP(_0)
574 #define jpo(_0) JPO(_0)
575 
576 #define JS(_0) JMP_(js, _0)
577 #define JNS(_0) JMP_(jns, _0)
578 
579 #define js(_0) JS(_0)
580 #define jns(_0) JNS(_0)
581 
582 #define JA(_0) JMP_(ja, _0)
583 #define JNBE(_0) JA(_0)
584 #define JNA(_0) JMP_(jna, _0)
585 #define JBE(_0) JNA(_0)
586 
587 #define ja(_0) JA(_0)
588 #define jnbe(_0) JNBE(_0)
589 #define jna(_0) JNA(_0)
590 #define jbe(_0) JBE(_0)
591 
592 #define JL(_0) JMP_(jl, _0)
593 #define JNGE(_0) JL(_0)
594 #define JNL(_0) JMP_(jnl, _0)
595 #define JGE(_0) JNL(_0)
596 
597 #define jl(_0) JL(_0)
598 #define jnge(_0) JNGE(_0)
599 #define jnl(_0) JNL(_0)
600 #define jge(_0) JGE(_0)
601 
602 #define JG(_0) JMP_(jg, _0)
603 #define JNLE(_0) JG(_0)
604 #define JNG(_0) JMP_(jng, _0)
605 #define JLE(_0) JNG(_0)
606 
607 #define jg(_0) JG(_0)
608 #define jnle(_0) JNLE(_0)
609 #define jng(_0) JNG(_0)
610 #define jle(_0) JLE(_0)
611 
612 #define JE(_0) JMP_(je, _0)
613 #define JZ(_0) JE(_0)
614 #define JNE(_0) JMP_(jne, _0)
615 #define JNZ(_0) JNE(_0)
616 
617 #define je(_0) JE(_0)
618 #define jz(_0) JZ(_0)
619 #define jne(_0) JNE(_0)
620 #define jnz(_0) JNZ(_0)
621 
622 #define JMP(_0) JMP_(jmp, _0)
623 
624 #define jmp(_0) JMP(_0)
625 
626 #define SETE(_0) INSTR_(sete, _0)
627 #define SETZ(_0) SETE(_0)
628 
629 #define sete(_0) SETE(_0)
630 #define setz(_0) SETZ(_0)
631 
632 // Comparisons
633 
634 #define CMP(_0, _1) INSTR_(cmp, _0, _1)
635 #define TEST(_0, _1) INSTR_(test, _0, _1)
636 
637 #define cmp(_0, _1) CMP(_0, _1)
638 #define test(_0, _1) TEST(_0, _1)
639 
640 // Integer math
641 
642 #define AND(_0, _1) INSTR_(and, _0, _1)
643 #define OR(_0, _1) INSTR_(or, _0, _1)
644 #define XOR(_0, _1) INSTR_(xor, _0, _1)
645 #define ADD(_0, _1) INSTR_(add, _0, _1)
646 #define SUB(_0, _1) INSTR_(sub, _0, _1)
647 #define IMUL(_0, _1) INSTR_(imul, _0, _1)
648 #define SAL(...) INSTR_(sal, __VA_ARGS__)
649 #define SAR(...) INSTR_(sar, __VA_ARGS__)
650 #define SHLX(_0, _1, _2) INSTR_(shlx, _0, _1, _2)
651 #define SHRX(_0, _1, _2) INSTR_(shrx, _0, _1, _2)
652 #define RORX(_0, _1, _2) INSTR_(rorx, _0, _1, _2)
653 #define DEC(_0) INSTR_(dec, _0)
654 #define INC(_0) INSTR_(inc, _0)
655 
656 #define and(_0, _1) AND(_0, _1)
657 #define or(_0, _1) OR(_0, _1)
658 #define xor(_0, _1) XOR(_0, _1)
659 #define add(_0, _1) ADD(_0, _1)
660 #define sub(_0, _1) SUB(_0, _1)
661 #define imul(_0, _1) IMUL(_0, _1)
662 #define sal(...) SAL(__VA_ARGS__)
663 #define sar(...) SAR(__VA_ARGS__)
664 #define shlx(_0, _1, _2) SHLX(_0, _1, _2)
665 #define shrx(_0, _1, _2) SHRX(_0, _1, _2)
666 #define rorx(_0, _1, _2) RORX(_0, _1, _2)
667 #define dec(_0) DEC(_0)
668 #define inc(_0) INC(_0)
669 
670 // Memory access
671 
672 #define LEA(_0, _1) INSTR_(lea, _0, _1)
673 #define MOV(_0, _1) INSTR_(mov, _0, _1)
674 #define MOVD(_0, _1) INSTR_(movd, _0, _1)
675 #define MOVL(_0, _1) INSTR_(movl, _0, _1)
676 #define MOVQ(_0, _1) INSTR_(movq, _0, _1)
677 #define CMOVA(_0, _1) INSTR_(cmova, _0, _1)
678 #define CMOVAE(_0, _1) INSTR_(cmovae, _0, _1)
679 #define CMOVB(_0, _1) INSTR_(cmovb, _0, _1)
680 #define CMOVBE(_0, _1) INSTR_(cmovbe, _0, _1)
681 #define CMOVC(_0, _1) INSTR_(cmovc, _0, _1)
682 #define CMOVP(_0, _1) INSTR_(cmovp, _0, _1)
683 #define CMOVO(_0, _1) INSTR_(cmovo, _0, _1)
684 #define CMOVS(_0, _1) INSTR_(cmovs, _0, _1)
685 #define CMOVE(_0, _1) INSTR_(cmove, _0, _1)
686 #define CMOVZ(_0, _1) INSTR_(cmovz, _0, _1)
687 #define CMOVG(_0, _1) INSTR_(cmovg, _0, _1)
688 #define CMOVGE(_0, _1) INSTR_(cmovge, _0, _1)
689 #define CMOVL(_0, _1) INSTR_(cmovl, _0, _1)
690 #define CMOVLE(_0, _1) INSTR_(cmovle, _0, _1)
691 #define CMOVNA(_0, _1) INSTR_(cmovna, _0, _1)
692 #define CMOVNAE(_0, _1) INSTR_(cmovnae, _0, _1)
693 #define CMOVNB(_0, _1) INSTR_(cmovnb, _0, _1)
694 #define CMOVNBE(_0, _1) INSTR_(cmovnbe, _0, _1)
695 #define CMOVNC(_0, _1) INSTR_(cmovnc, _0, _1)
696 #define CMOVNP(_0, _1) INSTR_(cmovnp, _0, _1)
697 #define CMOVNO(_0, _1) INSTR_(cmovno, _0, _1)
698 #define CMOVNS(_0, _1) INSTR_(cmovns, _0, _1)
699 #define CMOVNE(_0, _1) INSTR_(cmovne, _0, _1)
700 #define CMOVNZ(_0, _1) INSTR_(cmovnz, _0, _1)
701 #define CMOVNG(_0, _1) INSTR_(cmovng, _0, _1)
702 #define CMOVNGE(_0, _1) INSTR_(cmovnge, _0, _1)
703 #define CMOVNL(_0, _1) INSTR_(cmovnl, _0, _1)
704 #define CMOVNLE(_0, _1) INSTR_(cmovnle, _0, _1)
705 
706 #define lea(_0, _1) LEA(_0, _1)
707 #define mov(_0, _1) MOV(_0, _1)
708 #define movd(_0, _1) MOVD(_0, _1)
709 #define movl(_0, _1) MOVL(_0, _1)
710 #define movq(_0, _1) MOVQ(_0, _1)
711 #define cmova(_0, _1) CMOVA(_0, _1)
712 #define cmovae(_0, _1) CMOVAE(_0, _1)
713 #define cmovb(_0, _1) CMOVB(_0, _1)
714 #define cmovbe(_0, _1) CMOVBE(_0, _1)
715 #define cmovc(_0, _1) CMOVC(_0, _1)
716 #define cmovp(_0, _1) CMOVP(_0, _1)
717 #define cmovo(_0, _1) CMOVO(_0, _1)
718 #define cmovs(_0, _1) CMOVS(_0, _1)
719 #define cmove(_0, _1) CMOVE(_0, _1)
720 #define cmovz(_0, _1) CMOVZ(_0, _1)
721 #define cmovg(_0, _1) CMOVG(_0, _1)
722 #define cmovge(_0, _1) CMOVGE(_0, _1)
723 #define cmovl(_0, _1) CMOVL(_0, _1)
724 #define cmovle(_0, _1) CMOVLE(_0, _1)
725 #define cmovna(_0, _1) CMOVNA(_0, _1)
726 #define cmovnae(_0, _1) CMOVNAE(_0, _1)
727 #define cmovnb(_0, _1) CMOVNB(_0, _1)
728 #define cmovnbe(_0, _1) CMOVNBE(_0, _1)
729 #define cmovnc(_0, _1) CMOVNC(_0, _1)
730 #define cmovnp(_0, _1) CMOVNP(_0, _1)
731 #define cmovno(_0, _1) CMOVNO(_0, _1)
732 #define cmovns(_0, _1) CMOVNS(_0, _1)
733 #define cmovne(_0, _1) CMOVNE(_0, _1)
734 #define cmovnz(_0, _1) CMOVNZ(_0, _1)
735 #define cmovng(_0, _1) CMOVNG(_0, _1)
736 #define cmovnge(_0, _1) CMOVNGE(_0, _1)
737 #define cmovnl(_0, _1) CMOVNL(_0, _1)
738 #define cmovnle(_0, _1) CMOVNLE(_0, _1)
739 
740 // Vector moves
741 
742 #define MOVSS(_0, _1) INSTR_(movss, _0, _1)
743 #define MOVSD(_0, _1) INSTR_(movsd, _0, _1)
744 #define MOVAPS(_0, _1) INSTR_(movaps, _0, _1)
745 #define MOVAPD(_0, _1) INSTR_(movaps, _0, _1) //use movaps because it is shorter
746 #define MOVDDUP(_0, _1) INSTR_(movddup, _0, _1)
747 #define MOVLPS(_0, _1) INSTR_(movlps, _0, _1)
748 #define MOVHPS(_0, _1) INSTR_(movhps, _0, _1)
749 #define MOVLPD(_0, _1) INSTR_(movlpd, _0, _1)
750 #define MOVHPD(_0, _1) INSTR_(movhpd, _0, _1)
751 
752 #define movss(_0, _1) MOVSS(_0, _1)
753 #define movsd(_0, _1) MOVSD(_0, _1)
754 #define movaps(_0, _1) MOVAPS(_0, _1)
755 #define movapd(_0, _1) MOVAPD(_0, _1)
756 #define movddup(_0, _1) MOVDDUP(_0, _1)
757 #define movlps(_0, _1) MOVLPS(_0, _1)
758 #define movhps(_0, _1) MOVHPS(_0, _1)
759 #define movlpd(_0, _1) MOVLPD(_0, _1)
760 #define movhpd(_0, _1) MOVHPD(_0, _1)
761 
762 #define VMOVDDUP(_0, _1) INSTR_(vmovddup, _0, _1)
763 #define VMOVSLDUP(_0, _1) INSTR_(vmovsldup, _0, _1)
764 #define VMOVSHDUP(_0, _1) INSTR_(vmovshdup, _0, _1)
765 #define VMOVD(_0, _1) INSTR_(vmovd, _0, _1)
766 #define VMOVQ(_0, _1) INSTR_(vmovq, _0, _1)
767 #define VMOVSS(_0, _1) INSTR_(vmovss, _0, _1)
768 #define VMOVSD(_0, _1) INSTR_(vmovsd, _0, _1)
769 #define VMOVAPS(_0, _1) INSTR_(vmovaps, _0, _1)
770 #define VMOVUPS(_0, _1) INSTR_(vmovups, _0, _1)
771 #define VMOVAPD(_0, _1) INSTR_(vmovapd, _0, _1)
772 #define VMOVUPD(_0, _1) INSTR_(vmovupd, _0, _1)
773 #define VMOVLPS(...) INSTR_(vmovlps, __VA_ARGS__)
774 #define VMOVHPS(...) INSTR_(vmovhps, __VA_ARGS__)
775 #define VMOVLPD(...) INSTR_(vmovlpd, __VA_ARGS__)
776 #define VMOVHPD(...) INSTR_(vmovhpd, __VA_ARGS__)
777 #define VMOVDQA(_0, _1) INSTR_(vmovdqa, _0, _1)
778 #define VMOVDQA32(_0, _1) INSTR_(vmovdqa32, _0, _1)
779 #define VMOVDQA64(_0, _1) INSTR_(vmovdqa64, _0, _1)
780 #define VBROADCASTSS(_0, _1) INSTR_(vbroadcastss, _0, _1)
781 #define VBROADCASTSD(_0, _1) INSTR_(vbroadcastsd, _0, _1)
782 #define VPBROADCASTD(_0, _1) INSTR_(vpbroadcastd, _0, _1)
783 #define VPBROADCASTQ(_0, _1) INSTR_(vpbroadcastq, _0, _1)
784 #define VBROADCASTF128(_0, _1) INSTR_(vbroadcastf128, _0, _1)
785 #define VBROADCASTF64X4(_0, _1) INSTR_(vbroadcastf64x4, _0, _1)
786 #define VGATHERDPS(...) INSTR_(vgatherdps, __VA_ARGS__)
787 #define VSCATTERDPS(_0, _1) INSTR_(vscatterdps, _0, _1)
788 #define VGATHERDPD(...) INSTR_(vgatherdpd, __VA_ARGS__)
789 #define VSCATTERDPD(_0, _1) INSTR_(vscatterdpd, _0, _1)
790 #define VGATHERQPS(...) INSTR_(vgatherqps, __VA_ARGS__)
791 #define VSCATTERQPS(_0, _1) INSTR_(vscatterqps, _0, _1)
792 #define VGATHERQPD(...) INSTR_(vgatherqpd, __VA_ARGS__)
793 #define VSCATTERQPD(_0, _1) INSTR_(vscatterqpd, _0, _1)
794 
795 #define vmovddup(_0, _1) VMOVDDUP(_0, _1)
796 #define vmovsldup(_0, _1) VMOVSLDUP(_0, _1)
797 #define vmovshdup(_0, _1) VMOVSHDUP(_0, _1)
798 #define vmovd(_0, _1) VMOVD(_0, _1)
799 #define vmovq(_0, _1) VMOVQ(_0, _1)
800 #define vmovss(_0, _1) VMOVSS(_0, _1)
801 #define vmovsd(_0, _1) VMOVSD(_0, _1)
802 #define vmovaps(_0, _1) VMOVAPS(_0, _1)
803 #define vmovups(_0, _1) VMOVUPS(_0, _1)
804 #define vmovapd(_0, _1) VMOVAPD(_0, _1)
805 #define vmovupd(_0, _1) VMOVUPD(_0, _1)
806 #define vmovlps(...) VMOVLPS(__VA_ARGS__)
807 #define vmovhps(...) VMOVHPS(__VA_ARGS__)
808 #define vmovlpd(...) VMOVLPD(__VA_ARGS__)
809 #define vmovhpd(...) VMOVHPD(__VA_ARGS__)
810 #define vmovdqa(_0, _1) VMOVDQA(_0, _1)
811 #define vmovdqa32(_0, _1) VMOVDQA32(_0, _1)
812 #define vmovdqa64(_0, _1) VMOVDQA64(_0, _1)
813 #define vbroadcastss(_0, _1) VBROADCASTSS(_0, _1)
814 #define vbroadcastsd(_0, _1) VBROADCASTSD(_0, _1)
815 #define vpbroadcastd(_0, _1) VPBROADCASTD(_0, _1)
816 #define vpbroadcastq(_0, _1) VPBROADCASTQ(_0, _1)
817 #define vbroadcastf128(_0, _1) VBROADCASTF128(_0, _1)
818 #define vbroadcastf64x4(_0, _1) VBROADCASTF64X4(_0, _1)
819 #define vgatherdps(...) VGATHERDPS(__VA_ARGS__)
820 #define vscatterdps(_0, _1) VSCATTERDPS(_0, _1)
821 #define vgatherdpd(...) VGATHERDPD(__VA_ARGS__)
822 #define vscatterdpd(_0, _1) VSCATTERDPD(_0, _1)
823 #define vgatherqps(...) VGATHERQPS(__VA_ARGS__)
824 #define vscatterqps(_0, _1) VSCATTERQPS(_0, _1)
825 #define vgatherqpd(...) VGATHERQPD(__VA_ARGS__)
826 #define vscatterqpd(_0, _1) VSCATTERQPD(_0, _1)
827 
828 // Vector comparisons
829 
830 #define VPCMPEQB(_0, _1, _2) INSTR_(vpcmpeqb, _0, _1, _2)
831 #define VPCMPEQW(_0, _1, _2) INSTR_(vpcmpeqw, _0, _1, _2)
832 #define VPCMPEQD(_0, _1, _2) INSTR_(vpcmpeqd, _0, _1, _2)
833 
834 #define vpcmpeqb(_0, _1, _2) VPCMPEQB(_0, _1, _2)
835 #define vpcmpeqw(_0, _1, _2) VPCMPEQW(_0, _1, _2)
836 #define vpcmpeqd(_0, _1, _2) VPCMPEQD(_0, _1, _2)
837 
838 // Vector integer math
839 
840 #define VPADDB(_0, _1, _2) INSTR_(vpaddb, _0, _1, _2)
841 #define VPADDW(_0, _1, _2) INSTR_(vpaddw, _0, _1, _2)
842 #define VPADDD(_0, _1, _2) INSTR_(vpaddd, _0, _1, _2)
843 #define VPADDQ(_0, _1, _2) INSTR_(vpaddq, _0, _1, _2)
844 
845 #define vpaddb(_0, _1, _2) VPADDB(_0, _1, _2)
846 #define vpaddw(_0, _1, _2) VPADDW(_0, _1, _2)
847 #define vpaddd(_0, _1, _2) VPADDD(_0, _1, _2)
848 #define vpaddq(_0, _1, _2) VPADDQ(_0, _1, _2)
849 
850 // Vector math
851 
852 #define ADDPS(_0, _1) INSTR_(addps, _0, _1)
853 #define ADDPD(_0, _1) INSTR_(addpd, _0, _1)
854 #define SUBPS(_0, _1) INSTR_(subps, _0, _1)
855 #define SUBPD(_0, _1) INSTR_(subpd, _0, _1)
856 #define MULPS(_0, _1) INSTR_(mulps, _0, _1)
857 #define MULPD(_0, _1) INSTR_(mulpd, _0, _1)
858 #define DIVPS(_0, _1) INSTR_(divps, _0, _1)
859 #define DIVPD(_0, _1) INSTR_(divpd, _0, _1)
860 #define XORPS(_0, _1) INSTR_(xorps, _0, _1)
861 #define XORPD(_0, _1) INSTR_(xorpd, _0, _1)
862 
863 #define UCOMISS(_0, _1) INSTR_(ucomiss, _0, _1)
864 #define UCOMISD(_0, _1) INSTR_(ucomisd, _0, _1)
865 #define COMISS(_0, _1) INSTR_(comiss, _0, _1)
866 #define COMISD(_0, _1) INSTR_(comisd, _0, _1)
867 
868 #define addps(_0, _1) ADDPS(_0, _1)
869 #define addpd(_0, _1) ADDPD(_0, _1)
870 #define subps(_0, _1) SUBPS(_0, _1)
871 #define subpd(_0, _1) SUBPD(_0, _1)
872 #define mulps(_0, _1) MULPS(_0, _1)
873 #define mulpd(_0, _1) MULPD(_0, _1)
874 #define divps(_0, _1) DIVPS(_0, _1)
875 #define divpd(_0, _1) DIVPD(_0, _1)
876 #define xorps(_0, _1) XORPS(_0, _1)
877 #define xorpd(_0, _1) XORPD(_0, _1)
878 
879 #define ucomiss(_0, _1) UCOMISS(_0, _1)
880 #define ucomisd(_0, _1) UCOMISD(_0, _1)
881 #define cmoiss(_0, _1) COMISS(_0, _1)
882 #define comisd(_0, _1) COMISD(_0, _1)
883 
884 #define VADDSUBPS(_0, _1, _2) INSTR_(vaddsubps, _0, _1, _2)
885 #define VADDSUBPD(_0, _1, _2) INSTR_(vaddsubpd, _0, _1, _2)
886 #define VHADDPD(_0, _1, _2) INSTR_(vhaddpd, _0, _1, _2)
887 #define VHADDPS(_0, _1, _2) INSTR_(vhaddps, _0, _1, _2)
888 #define VADDPS(_0, _1, _2) INSTR_(vaddps, _0, _1, _2)
889 #define VADDPD(_0, _1, _2) INSTR_(vaddpd, _0, _1, _2)
890 #define VSUBPS(_0, _1, _2) INSTR_(vsubps, _0, _1, _2)
891 #define VSUBPD(_0, _1, _2) INSTR_(vsubpd, _0, _1, _2)
892 #define VMULSS(_0, _1, _2) INSTR_(vmulss, _0, _1, _2)
893 #define VMULSD(_0, _1, _2) INSTR_(vmulsd, _0, _1, _2)
894 #define VMULPS(_0, _1, _2) INSTR_(vmulps, _0, _1, _2)
895 #define VMULPD(_0, _1, _2) INSTR_(vmulpd, _0, _1, _2)
896 #define VDIVSS(_0, _1, _2) INSTR_(vdivss, _0, _1, _2)
897 #define VDIVSD(_0, _1, _2) INSTR_(vdivsd, _0, _1, _2)
898 #define VDIVPS(_0, _1, _2) INSTR_(vdivps, _0, _1, _2)
899 #define VDIVPD(_0, _1, _2) INSTR_(vdivpd, _0, _1, _2)
900 #define VPMULLD(_0, _1, _2) INSTR_(vpmulld, _0, _1, _2)
901 #define VPMULLQ(_0, _1, _2) INSTR_(vpmullq, _0, _1, _2)
902 #define VPADDD(_0, _1, _2) INSTR_(vpaddd, _0, _1, _2)
903 #define VPSLLD(_0, _1, _2) INSTR_(vpslld, _0, _1, _2)
904 #define VXORPS(_0, _1, _2) INSTR_(vxorps, _0, _1, _2)
905 #define VXORPD(_0, _1, _2) INSTR_(vxorpd, _0, _1, _2)
906 #define VPXORD(_0, _1, _2) INSTR_(vpxord, _0, _1, _2)
907 
908 #define VUCOMISS(_0, _1) INSTR_(vucomiss, _0, _1)
909 #define VUCOMISD(_0, _1) INSTR_(vucomisd, _0, _1)
910 #define VCOMISS(_0, _1) INSTR_(vcomiss, _0, _1)
911 #define VCOMISD(_0, _1) INSTR_(vcomisd, _0, _1)
912 
913 #define VFMADD132SS(_0, _1, _2) INSTR_(vfmadd132ss, _0, _1, _2)
914 #define VFMADD213SS(_0, _1, _2) INSTR_(vfmadd213ss, _0, _1, _2)
915 #define VFMADD231SS(_0, _1, _2) INSTR_(vfmadd231ss, _0, _1, _2)
916 #define VFMADD132SD(_0, _1, _2) INSTR_(vfmadd132sd, _0, _1, _2)
917 #define VFMADD213SD(_0, _1, _2) INSTR_(vfmadd213sd, _0, _1, _2)
918 #define VFMADD231SD(_0, _1, _2) INSTR_(vfmadd231sd, _0, _1, _2)
919 #define VFMADD132PS(_0, _1, _2) INSTR_(vfmadd132ps, _0, _1, _2)
920 #define VFMADD213PS(_0, _1, _2) INSTR_(vfmadd213ps, _0, _1, _2)
921 #define VFMADD231PS(_0, _1, _2) INSTR_(vfmadd231ps, _0, _1, _2)
922 #define VFMADD132PD(_0, _1, _2) INSTR_(vfmadd132pd, _0, _1, _2)
923 #define VFMADD213PD(_0, _1, _2) INSTR_(vfmadd213pd, _0, _1, _2)
924 #define VFMADD231PD(_0, _1, _2) INSTR_(vfmadd231pd, _0, _1, _2)
925 #define VFMSUB132SS(_0, _1, _2) INSTR_(vfmsub132ss, _0, _1, _2)
926 #define VFMSUB213SS(_0, _1, _2) INSTR_(vfmsub213ss, _0, _1, _2)
927 #define VFMSUB231SS(_0, _1, _2) INSTR_(vfmsub231ss, _0, _1, _2)
928 #define VFMSUB132SD(_0, _1, _2) INSTR_(vfmsub132sd, _0, _1, _2)
929 #define VFMSUB213SD(_0, _1, _2) INSTR_(vfmsub213sd, _0, _1, _2)
930 #define VFMSUB231SD(_0, _1, _2) INSTR_(vfmsub231sd, _0, _1, _2)
931 #define VFMSUB132PS(_0, _1, _2) INSTR_(vfmsub132ps, _0, _1, _2)
932 #define VFMSUB213PS(_0, _1, _2) INSTR_(vfmsub213ps, _0, _1, _2)
933 #define VFMSUB231PS(_0, _1, _2) INSTR_(vfmsub231ps, _0, _1, _2)
934 #define VFMSUB132PD(_0, _1, _2) INSTR_(vfmsub132pd, _0, _1, _2)
935 #define VFMSUB213PD(_0, _1, _2) INSTR_(vfmsub213pd, _0, _1, _2)
936 #define VFMSUB231PD(_0, _1, _2) INSTR_(vfmsub231pd, _0, _1, _2)
937 #define VFNMADD132SS(_0, _1, _2) INSTR_(vfnmadd132ss, _0, _1, _2)
938 #define VFNMADD213SS(_0, _1, _2) INSTR_(vfnmadd213ss, _0, _1, _2)
939 #define VFNMADD231SS(_0, _1, _2) INSTR_(vfnmadd231ss, _0, _1, _2)
940 #define VFNMADD132SD(_0, _1, _2) INSTR_(vfnmadd132sd, _0, _1, _2)
941 #define VFNMADD213SD(_0, _1, _2) INSTR_(vfnmadd213sd, _0, _1, _2)
942 #define VFNMADD231SD(_0, _1, _2) INSTR_(vfnmadd231sd, _0, _1, _2)
943 #define VFNMADD132PS(_0, _1, _2) INSTR_(vfnmadd132ps, _0, _1, _2)
944 #define VFNMADD213PS(_0, _1, _2) INSTR_(vfnmadd213ps, _0, _1, _2)
945 #define VFNMADD231PS(_0, _1, _2) INSTR_(vfnmadd231ps, _0, _1, _2)
946 #define VFNMADD132PD(_0, _1, _2) INSTR_(vfnmadd132pd, _0, _1, _2)
947 #define VFNMADD213PD(_0, _1, _2) INSTR_(vfnmadd213pd, _0, _1, _2)
948 #define VFNMADD231PD(_0, _1, _2) INSTR_(vfnmadd231pd, _0, _1, _2)
949 #define VFNMSUB132SS(_0, _1, _2) INSTR_(vfnmsub132ss, _0, _1, _2)
950 #define VFNMSUB213SS(_0, _1, _2) INSTR_(vfnmsub213ss, _0, _1, _2)
951 #define VFNMSUB231SS(_0, _1, _2) INSTR_(vfnmsub231ss, _0, _1, _2)
952 #define VFNMSUB132SD(_0, _1, _2) INSTR_(vfnmsub132sd, _0, _1, _2)
953 #define VFNMSUB213SD(_0, _1, _2) INSTR_(vfnmsub213sd, _0, _1, _2)
954 #define VFNMSUB231SD(_0, _1, _2) INSTR_(vfnmsub231sd, _0, _1, _2)
955 #define VFNMSUB132PS(_0, _1, _2) INSTR_(vfnmsub132ps, _0, _1, _2)
956 #define VFNMSUB213PS(_0, _1, _2) INSTR_(vfnmsub213ps, _0, _1, _2)
957 #define VFNMSUB231PS(_0, _1, _2) INSTR_(vfnmsub231ps, _0, _1, _2)
958 #define VFNMSUB132PD(_0, _1, _2) INSTR_(vfnmsub132pd, _0, _1, _2)
959 #define VFNMSUB213PD(_0, _1, _2) INSTR_(vfnmsub213pd, _0, _1, _2)
960 #define VFNMSUB231PD(_0, _1, _2) INSTR_(vfnmsub231pd, _0, _1, _2)
961 #define VFMADDSUB132SS(_0, _1, _2) INSTR_(vfmaddsub132ss, _0, _1, _2)
962 #define VFMADDSUB213SS(_0, _1, _2) INSTR_(vfmaddsub213ss, _0, _1, _2)
963 #define VFMADDSUB231SS(_0, _1, _2) INSTR_(vfmaddsub231ss, _0, _1, _2)
964 #define VFMADDSUB132SD(_0, _1, _2) INSTR_(vfmaddsub132sd, _0, _1, _2)
965 #define VFMADDSUB213SD(_0, _1, _2) INSTR_(vfmaddsub213sd, _0, _1, _2)
966 #define VFMADDSUB231SD(_0, _1, _2) INSTR_(vfmaddsub231sd, _0, _1, _2)
967 #define VFMADDSUB132PS(_0, _1, _2) INSTR_(vfmaddsub132ps, _0, _1, _2)
968 #define VFMADDSUB213PS(_0, _1, _2) INSTR_(vfmaddsub213ps, _0, _1, _2)
969 #define VFMADDSUB231PS(_0, _1, _2) INSTR_(vfmaddsub231ps, _0, _1, _2)
970 #define VFMADDSUB132PD(_0, _1, _2) INSTR_(vfmaddsub132pd, _0, _1, _2)
971 #define VFMADDSUB213PD(_0, _1, _2) INSTR_(vfmaddsub213pd, _0, _1, _2)
972 #define VFMADDSUB231PD(_0, _1, _2) INSTR_(vfmaddsub231pd, _0, _1, _2)
973 #define VFMSUBADD132SS(_0, _1, _2) INSTR_(vfmsubadd132ss, _0, _1, _2)
974 #define VFMSUBADD213SS(_0, _1, _2) INSTR_(vfmsubadd213ss, _0, _1, _2)
975 #define VFMSUBADD231SS(_0, _1, _2) INSTR_(vfmsubadd231ss, _0, _1, _2)
976 #define VFMSUBADD132SD(_0, _1, _2) INSTR_(vfmsubadd132sd, _0, _1, _2)
977 #define VFMSUBADD213SD(_0, _1, _2) INSTR_(vfmsubadd213sd, _0, _1, _2)
978 #define VFMSUBADD231SD(_0, _1, _2) INSTR_(vfmsubadd231sd, _0, _1, _2)
979 #define VFMSUBADD132PS(_0, _1, _2) INSTR_(vfmsubadd132ps, _0, _1, _2)
980 #define VFMSUBADD213PS(_0, _1, _2) INSTR_(vfmsubadd213ps, _0, _1, _2)
981 #define VFMSUBADD231PS(_0, _1, _2) INSTR_(vfmsubadd231ps, _0, _1, _2)
982 #define VFMSUBADD132PD(_0, _1, _2) INSTR_(vfmsubadd132pd, _0, _1, _2)
983 #define VFMSUBADD213PD(_0, _1, _2) INSTR_(vfmsubadd213pd, _0, _1, _2)
984 #define VFMSUBADD231PD(_0, _1, _2) INSTR_(vfmsubadd231pd, _0, _1, _2)
985 #define VFMADDSS(_0, _1, _2, _3) INSTR_(vfmaddss, _0, _1, _2, _3)
986 #define VFMADDSD(_0, _1, _2, _3) INSTR_(vfmaddsd, _0, _1, _2, _3)
987 #define VFMADDPS(_0, _1, _2, _3) INSTR_(vfmaddps, _0, _1, _2, _3)
988 #define VFMADDPD(_0, _1, _2, _3) INSTR_(vfmaddpd, _0, _1, _2, _3)
989 #define VFMSUBSS(_0, _1, _2, _3) INSTR_(vfmsubss, _0, _1, _2, _3)
990 #define VFMSUBSD(_0, _1, _2, _3) INSTR_(vfmsubsd, _0, _1, _2, _3)
991 #define VFMSUBPS(_0, _1, _2, _3) INSTR_(vfmsubps, _0, _1, _2, _3)
992 #define VFMSUBPD(_0, _1, _2, _3) INSTR_(vfmsubpd, _0, _1, _2, _3)
993 #define VFNMADDSS(_0, _1, _2, _3) INSTR_(vfnmaddss, _0, _1, _2, _3)
994 #define VFNMADDSD(_0, _1, _2, _3) INSTR_(vfnmaddsd, _0, _1, _2, _3)
995 #define VFNMADDPS(_0, _1, _2, _3) INSTR_(vfnmaddps, _0, _1, _2, _3)
996 #define VFNMADDPD(_0, _1, _2, _3) INSTR_(vfnmaddpd, _0, _1, _2, _3)
997 #define VFNMSUBSS(_0, _1, _2, _3) INSTR_(vfnmsubss, _0, _1, _2, _3)
998 #define VFNMSUBSD(_0, _1, _2, _3) INSTR_(vfnmsubsd, _0, _1, _2, _3)
999 #define VFNMSUBPS(_0, _1, _2, _3) INSTR_(vfnmsubps, _0, _1, _2, _3)
1000 #define VFNMSUBPD(_0, _1, _2, _3) INSTR_(vfnmsubpd, _0, _1, _2, _3)
1001 #define VFMADDSUBSS(_0, _1, _2, _3) INSTR_(vfmaddsubss, _0, _1, _2, _3)
1002 #define VFMADDSUBSD(_0, _1, _2, _3) INSTR_(vfmaddsubsd, _0, _1, _2, _3)
1003 #define VFMADDSUBPS(_0, _1, _2, _3) INSTR_(vfmaddsubps, _0, _1, _2, _3)
1004 #define VFMADDSUBPD(_0, _1, _2, _3) INSTR_(vfmaddsubpd, _0, _1, _2, _3)
1005 #define VFMSUBADDSS(_0, _1, _2, _3) INSTR_(vfmsubaddss, _0, _1, _2, _3)
1006 #define VFMSUBADDSD(_0, _1, _2, _3) INSTR_(vfmsubaddsd, _0, _1, _2, _3)
1007 #define VFMSUBADDPS(_0, _1, _2, _3) INSTR_(vfmsubaddps, _0, _1, _2, _3)
1008 #define VFMSUBADDPD(_0, _1, _2, _3) INSTR_(vfmsubaddpd, _0, _1, _2, _3)
1009 #define V4FMADDSS(_0, _1, _2) INSTR_(v4fmaddss, _0, _1, _2)
1010 #define V4FMADDPS(_0, _1, _2) INSTR_(v4fmaddps, _0, _1, _2)
1011 #define V4FNMADDSS(_0, _1, _2) INSTR_(v4fnmaddss, _0, _1, _2)
1012 #define V4FNMADDPS(_0, _1, _2) INSTR_(v4fnmaddps, _0, _1, _2)
1013 
1014 #define vaddsubps(_0, _1, _2) VADDSUBPS(_0, _1, _2)
1015 #define vaddsubpd(_0, _1, _2) VADDSUBPD(_0, _1, _2)
1016 #define vhaddpd(_0, _1, _2) VHADDPD(_0, _1, _2)
1017 #define vhaddps(_0, _1, _2) VHADDPS(_0, _1, _2)
1018 #define vaddps(_0, _1, _2) VADDPS(_0, _1, _2)
1019 #define vaddpd(_0, _1, _2) VADDPD(_0, _1, _2)
1020 #define vsubps(_0, _1, _2) VSUBPS(_0, _1, _2)
1021 #define vsubpd(_0, _1, _2) VSUBPD(_0, _1, _2)
1022 #define vmulss(_0, _1, _2) VMULSS(_0, _1, _2)
1023 #define vmulps(_0, _1, _2) VMULPS(_0, _1, _2)
1024 #define vmulsd(_0, _1, _2) VMULSD(_0, _1, _2)
1025 #define vmulpd(_0, _1, _2) VMULPD(_0, _1, _2)
1026 #define vdivss(_0, _1, _2) VDIVSS(_0, _1, _2)
1027 #define vdivps(_0, _1, _2) VDIVPS(_0, _1, _2)
1028 #define vdivsd(_0, _1, _2) VDIVSD(_0, _1, _2)
1029 #define vdivpd(_0, _1, _2) VDIVPD(_0, _1, _2)
1030 #define vpmulld(_0, _1, _2) VPMULLD(_0, _1, _2)
1031 #define vpmullq(_0, _1, _2) VPMULLQ(_0, _1, _2)
1032 #define vpaddd(_0, _1, _2) VPADDD(_0, _1, _2)
1033 #define vpslld(_0, _1, _2) VPSLLD(_0, _1, _2)
1034 #define vxorps(_0, _1, _2) VXORPS(_0, _1, _2)
1035 #define vxorpd(_0, _1, _2) VXORPD(_0, _1, _2)
1036 #define vpxord(_0, _1, _2) VPXORD(_0, _1, _2)
1037 
1038 #define vucomiss(_0, _1) VUCOMISS(_0, _1)
1039 #define vucomisd(_0, _1) VUCOMISD(_0, _1)
1040 #define vcomiss(_0, _1) VCOMISS(_0, _1)
1041 #define vcomisd(_0, _1) VCOMISD(_0, _1)
1042 
1043 #define vfmadd132ss(_0, _1, _2) VFMADD132SS(_0, _1, _2)
1044 #define vfmadd213ss(_0, _1, _2) VFMADD213SS(_0, _1, _2)
1045 #define vfmadd231ss(_0, _1, _2) VFMADD231SS(_0, _1, _2)
1046 #define vfmadd132sd(_0, _1, _2) VFMADD132SD(_0, _1, _2)
1047 #define vfmadd213sd(_0, _1, _2) VFMADD213SD(_0, _1, _2)
1048 #define vfmadd231sd(_0, _1, _2) VFMADD231SD(_0, _1, _2)
1049 #define vfmadd132ps(_0, _1, _2) VFMADD132PS(_0, _1, _2)
1050 #define vfmadd213ps(_0, _1, _2) VFMADD213PS(_0, _1, _2)
1051 #define vfmadd231ps(_0, _1, _2) VFMADD231PS(_0, _1, _2)
1052 #define vfmadd132pd(_0, _1, _2) VFMADD132PD(_0, _1, _2)
1053 #define vfmadd213pd(_0, _1, _2) VFMADD213PD(_0, _1, _2)
1054 #define vfmadd231pd(_0, _1, _2) VFMADD231PD(_0, _1, _2)
1055 #define vfmadd132ss(_0, _1, _2) VFMADD132SS(_0, _1, _2)
1056 #define vfmsub213ss(_0, _1, _2) VFMSUB213SS(_0, _1, _2)
1057 #define vfmsub231ss(_0, _1, _2) VFMSUB231SS(_0, _1, _2)
1058 #define vfmsub132sd(_0, _1, _2) VFMSUB132SD(_0, _1, _2)
1059 #define vfmsub213sd(_0, _1, _2) VFMSUB213SD(_0, _1, _2)
1060 #define vfmsub231sd(_0, _1, _2) VFMSUB231SD(_0, _1, _2)
1061 #define vfmsub132ps(_0, _1, _2) VFMSUB132PS(_0, _1, _2)
1062 #define vfmsub213ps(_0, _1, _2) VFMSUB213PS(_0, _1, _2)
1063 #define vfmsub231ps(_0, _1, _2) VFMSUB231PS(_0, _1, _2)
1064 #define vfmsub132pd(_0, _1, _2) VFMSUB132PD(_0, _1, _2)
1065 #define vfmsub213pd(_0, _1, _2) VFMSUB213PD(_0, _1, _2)
1066 #define vfmsub231pd(_0, _1, _2) VFMSUB231PD(_0, _1, _2)
1067 #define vfnmadd132ss(_0, _1, _2) VFNMADD132SS(_0, _1, _2)
1068 #define vfnmadd213ss(_0, _1, _2) VFNMADD213SS(_0, _1, _2)
1069 #define vfnmadd231ss(_0, _1, _2) VFNMADD231SS(_0, _1, _2)
1070 #define vfnmadd132sd(_0, _1, _2) VFNMADD132SD(_0, _1, _2)
1071 #define vfnmadd213sd(_0, _1, _2) VFNMADD213SD(_0, _1, _2)
1072 #define vfnmadd231sd(_0, _1, _2) VFNMADD231SD(_0, _1, _2)
1073 #define vfnmadd132ps(_0, _1, _2) VFNMADD132PS(_0, _1, _2)
1074 #define vfnmadd213ps(_0, _1, _2) VFNMADD213PS(_0, _1, _2)
1075 #define vfnmadd231ps(_0, _1, _2) VFNMADD231PS(_0, _1, _2)
1076 #define vfnmadd132pd(_0, _1, _2) VFNMADD132PD(_0, _1, _2)
1077 #define vfnmadd213pd(_0, _1, _2) VFNMADD213PD(_0, _1, _2)
1078 #define vfnmadd231pd(_0, _1, _2) VFNMADD231PD(_0, _1, _2)
1079 #define vfnmadd132ss(_0, _1, _2) VFNMADD132SS(_0, _1, _2)
1080 #define vfnmsub213ss(_0, _1, _2) VFNMSUB213SS(_0, _1, _2)
1081 #define vfnmsub231ss(_0, _1, _2) VFNMSUB231SS(_0, _1, _2)
1082 #define vfnmsub132sd(_0, _1, _2) VFNMSUB132SD(_0, _1, _2)
1083 #define vfnmsub213sd(_0, _1, _2) VFNMSUB213SD(_0, _1, _2)
1084 #define vfnmsub231sd(_0, _1, _2) VFNMSUB231SD(_0, _1, _2)
1085 #define vfnmsub132ps(_0, _1, _2) VFNMSUB132PS(_0, _1, _2)
1086 #define vfnmsub213ps(_0, _1, _2) VFNMSUB213PS(_0, _1, _2)
1087 #define vfnmsub231ps(_0, _1, _2) VFNMSUB231PS(_0, _1, _2)
1088 #define vfnmsub132pd(_0, _1, _2) VFNMSUB132PD(_0, _1, _2)
1089 #define vfnmsub213pd(_0, _1, _2) VFNMSUB213PD(_0, _1, _2)
1090 #define vfnmsub231pd(_0, _1, _2) VFNMSUB231PD(_0, _1, _2)
1091 #define vfmaddsub132ss(_0, _1, _2) VFMADDSUB132SS(_0, _1, _2)
1092 #define vfmaddsub213ss(_0, _1, _2) VFMADDSUB213SS(_0, _1, _2)
1093 #define vfmaddsub231ss(_0, _1, _2) VFMADDSUB231SS(_0, _1, _2)
1094 #define vfmaddsub132sd(_0, _1, _2) VFMADDSUB132SD(_0, _1, _2)
1095 #define vfmaddsub213sd(_0, _1, _2) VFMADDSUB213SD(_0, _1, _2)
1096 #define vfmaddsub231sd(_0, _1, _2) VFMADDSUB231SD(_0, _1, _2)
1097 #define vfmaddsub132ps(_0, _1, _2) VFMADDSUB132PS(_0, _1, _2)
1098 #define vfmaddsub213ps(_0, _1, _2) VFMADDSUB213PS(_0, _1, _2)
1099 #define vfmaddsub231ps(_0, _1, _2) VFMADDSUB231PS(_0, _1, _2)
1100 #define vfmaddsub132pd(_0, _1, _2) VFMADDSUB132PD(_0, _1, _2)
1101 #define vfmaddsub213pd(_0, _1, _2) VFMADDSUB213PD(_0, _1, _2)
1102 #define vfmaddsub231pd(_0, _1, _2) VFMADDSUB231PD(_0, _1, _2)
1103 #define vfmsubadd132ss(_0, _1, _2) VFMSUBADD132SS(_0, _1, _2)
1104 #define vfmsubadd213ss(_0, _1, _2) VFMSUBADD213SS(_0, _1, _2)
1105 #define vfmsubadd231ss(_0, _1, _2) VFMSUBADD231SS(_0, _1, _2)
1106 #define vfmsubadd132sd(_0, _1, _2) VFMSUBADD132SD(_0, _1, _2)
1107 #define vfmsubadd213sd(_0, _1, _2) VFMSUBADD213SD(_0, _1, _2)
1108 #define vfmsubadd231sd(_0, _1, _2) VFMSUBADD231SD(_0, _1, _2)
1109 #define vfmsubadd132ps(_0, _1, _2) VFMSUBADD132PS(_0, _1, _2)
1110 #define vfmsubadd213ps(_0, _1, _2) VFMSUBADD213PS(_0, _1, _2)
1111 #define vfmsubadd231ps(_0, _1, _2) VFMSUBADD231PS(_0, _1, _2)
1112 #define vfmsubadd132pd(_0, _1, _2) VFMSUBADD132PD(_0, _1, _2)
1113 #define vfmsubadd213pd(_0, _1, _2) VFMSUBADD213PD(_0, _1, _2)
1114 #define vfmsubadd231pd(_0, _1, _2) VFMSUBADD231PD(_0, _1, _2)
1115 #define vfmaddss(_0, _1, _2, _3) VFMADDSS(_0, _1, _2, _3)
1116 #define vfmaddsd(_0, _1, _2, _3) VFMADDSD(_0, _1, _2, _3)
1117 #define vfmaddps(_0, _1, _2, _3) VFMADDPS(_0, _1, _2, _3)
1118 #define vfmaddpd(_0, _1, _2, _3) VFMADDPD(_0, _1, _2, _3)
1119 #define vfmsubss(_0, _1, _2, _3) VFMSUBSS(_0, _1, _2, _3)
1120 #define vfmsubsd(_0, _1, _2, _3) VFMSUBSD(_0, _1, _2, _3)
1121 #define vfmsubps(_0, _1, _2, _3) VFMSUBPS(_0, _1, _2, _3)
1122 #define vfmsubpd(_0, _1, _2, _3) VFMSUBPD(_0, _1, _2, _3)
1123 #define vfnmaddss(_0, _1, _2, _3) VFNMADDSS(_0, _1, _2, _3)
1124 #define vfnmaddsd(_0, _1, _2, _3) VFNMADDSD(_0, _1, _2, _3)
1125 #define vfnmaddps(_0, _1, _2, _3) VFNMADDPS(_0, _1, _2, _3)
1126 #define vfnmaddpd(_0, _1, _2, _3) VFNMADDPD(_0, _1, _2, _3)
1127 #define vfnmsubss(_0, _1, _2, _3) VFNMSUBSS(_0, _1, _2, _3)
1128 #define vfnmsubsd(_0, _1, _2, _3) VFNMSUBSD(_0, _1, _2, _3)
1129 #define vfnmsubps(_0, _1, _2, _3) VFNMSUBPS(_0, _1, _2, _3)
1130 #define vfnmsubpd(_0, _1, _2, _3) VFNMSUBPD(_0, _1, _2, _3)
1131 #define vfmaddsubss(_0, _1, _2, _3) VFMADDSUBSS(_0, _1, _2, _3)
1132 #define vfmaddsubsd(_0, _1, _2, _3) VFMADDSUBSD(_0, _1, _2, _3)
1133 #define vfmaddsubps(_0, _1, _2, _3) VFMADDSUBPS(_0, _1, _2, _3)
1134 #define vfmaddsubpd(_0, _1, _2, _3) VFMADDSUBPD(_0, _1, _2, _3)
1135 #define vfmsubaddss(_0, _1, _2, _3) VFMSUBADDSS(_0, _1, _2, _3)
1136 #define vfmsubaddsd(_0, _1, _2, _3) VFMSUBADDSD(_0, _1, _2, _3)
1137 #define vfmsubaddps(_0, _1, _2, _3) VFMSUBADDPS(_0, _1, _2, _3)
1138 #define vfmsubaddpd(_0, _1, _2, _3) VFMSUBADDPD(_0, _1, _2, _3)
1139 #define v4fmaddss(_0, _1, _2) V4FMADDSS(_0, _1, _2)
1140 #define v4fmaddps(_0, _1, _2) V4FMADDPS(_0, _1, _2)
1141 #define v4fnmaddss(_0, _1, _2) V4FNMADDSS(_0, _1, _2)
1142 #define v4fnmaddps(_0, _1, _2) V4FNMADDPS(_0, _1, _2)
1143 
1144 // Conversions
1145 
1146 #define CVTSS2SD(_0, _1) INSTR_(cvtss2sd, _0, _1)
1147 #define CVTSD2SS(_0, _1) INSTR_(cvtsd2ss, _0, _1)
1148 #define CVTPS2PD(_0, _1) INSTR_(cvtps2pd, _0, _1)
1149 #define CVTPD2PS(_0, _1) INSTR_(cvtpd2ps, _0, _1)
1150 
1151 #define cvtss2sd(_0, _1) CVTSS2SD(_0, _1)
1152 #define cvtsd2ss(_0, _1) CVTSD2SS(_0, _1)
1153 #define cvtps2pd(_0, _1) CVTPS2PD(_0, _1)
1154 #define cvtpd2ps(_0, _1) CVTPD2PS(_0, _1)
1155 
1156 #define VCVTSS2SD(_0, _1) INSTR_(vcvtss2sd, _0, _1)
1157 #define VCVTSD2SS(_0, _1) INSTR_(vcvtsd2ss, _0, _1)
1158 #define VCVTPS2PD(_0, _1) INSTR_(vcvtps2pd, _0, _1)
1159 #define VCVTPD2PS(_0, _1) INSTR_(vcvtpd2ps, _0, _1)
1160 
1161 #define vcvtss2sd(_0, _1) VCVTSS2SD(_0, _1)
1162 #define vcvtsd2ss(_0, _1) VCVTSD2SS(_0, _1)
1163 #define vcvtps2pd(_0, _1) VCVTPS2PD(_0, _1)
1164 #define vcvtpd2ps(_0, _1) VCVTPD2PS(_0, _1)
1165 
1166 // Vector shuffles
1167 
1168 #define PSHUFD(_0, _1, _2) INSTR_(pshufd, _0, _1, _2)
1169 #define SHUFPS(_0, _1, _2) INSTR_(shufps, _0, _1, _2)
1170 #define SHUFPD(_0, _1, _2) INSTR_(shufpd, _0, _1, _2)
1171 #define UNPCKLPS(_0, _1) INSTR_(unpcklps, _0, _1)
1172 #define UNPCKHPS(_0, _1) INSTR_(unpckhps, _0, _1)
1173 #define UNPCKLPD(_0, _1) INSTR_(unpcklpd, _0, _1)
1174 #define UNPCKHPD(_0, _1) INSTR_(unpckhpd, _0, _1)
1175 
1176 #define pshufd(_0, _1, _2) PSHUFD(_0, _1, _2)
1177 #define shufps(_0, _1, _2) SHUFPS(_0, _1, _2)
1178 #define shufpd(_0, _1, _2) SHUFPD(_0, _1, _2)
1179 #define unpcklps(_0, _1) UNPCKLPS(_0, _1)
1180 #define unpckhps(_0, _1) UNPCKHPS(_0, _1)
1181 #define unpcklpd(_0, _1) UNPCKLPD(_0, _1)
1182 #define unpckhpd(_0, _1) UNPCKHPD(_0, _1)
1183 
1184 #define VSHUFPS(_0, _1, _2, _3) INSTR_(vshufps, _0, _1, _2, _3)
1185 #define VSHUFPD(_0, _1, _2, _3) INSTR_(vshufpd, _0, _1, _2, _3)
1186 #define VPERMILPS(_0, _1, _2) INSTR_(vpermilps, _0, _1, _2)
1187 #define VPERMILPD(_0, _1, _2) INSTR_(vpermilpd, _0, _1, _2)
1188 #define VPERM2F128(_0, _1, _2, _3) INSTR_(vperm2f128, _0, _1, _2, _3)
1189 #define VPERMPD(_0, _1, _2) INSTR_(vpermpd, _0, _1, _2)
1190 #define VUNPCKLPS(_0, _1, _2) INSTR_(vunpcklps, _0, _1, _2)
1191 #define VUNPCKHPS(_0, _1, _2) INSTR_(vunpckhps, _0, _1, _2)
1192 #define VUNPCKLPD(_0, _1, _2) INSTR_(vunpcklpd, _0, _1, _2)
1193 #define VUNPCKHPD(_0, _1, _2) INSTR_(vunpckhpd, _0, _1, _2)
1194 #define VSHUFF32X4(_0, _1, _2, _3) INSTR_(vshuff32x4, _0, _1, _2, _3)
1195 #define VSHUFF64X2(_0, _1, _2, _3) INSTR_(vshuff64x2, _0, _1, _2, _3)
1196 #define VINSERTF128(_0, _1, _2, _3) INSTR_(vinsertf128, _0, _1, _2, _3)
1197 #define VINSERTF32X4(_0, _1, _2, _3) INSTR_(vinsertf32x4, _0, _1, _2, _3)
1198 #define VINSERTF32X8(_0, _1, _2, _3) INSTR_(vinsertf32x8, _0, _1, _2, _3)
1199 #define VINSERTF64X2(_0, _1, _2, _3) INSTR_(vinsertf64x2, _0, _1, _2, _3)
1200 #define VINSERTF64X4(_0, _1, _2, _3) INSTR_(vinsertf64x4, _0, _1, _2, _3)
1201 #define VEXTRACTF128(_0, _1, _2) INSTR_(vextractf128, _0, _1, _2)
1202 #define VEXTRACTF32X4(_0, _1, _2) INSTR_(vextractf32x4, _0, _1, _2)
1203 #define VEXTRACTF32X8(_0, _1, _2) INSTR_(vextractf32x8, _0, _1, _2)
1204 #define VEXTRACTF64X2(_0, _1, _2) INSTR_(vextractf64x4, _0, _1, _2)
1205 #define VEXTRACTF64X4(_0, _1, _2) INSTR_(vextractf64x4, _0, _1, _2)
1206 #define VBLENDPS(_0, _1, _2, _3) INSTR_(vblendps, _0, _1, _2, _3)
1207 #define VBLENDPD(_0, _1, _2, _3) INSTR_(vblendpd, _0, _1, _2, _3)
1208 #define VBLENDMPS(_0, _1, _2) INSTR_(vblendmps, _0, _1, _2)
1209 #define VBLENDMPD(_0, _1, _2) INSTR_(vblendmpd, _0, _1, _2)
1210 
1211 #define vshufps(_0, _1, _2, _3) VSHUFPS(_0, _1, _2, _3)
1212 #define vshufpd(_0, _1, _2, _3) VSHUFPD(_0, _1, _2, _3)
1213 #define vpermilps(_0, _1, _2) VPERMILPS(_0, _1, _2)
1214 #define vpermilpd(_0, _1, _2) VPERMILPD(_0, _1, _2)
1215 #define vperm2f128(_0, _1, _2, _3) VPERM2F128(_0, _1, _2, _3)
1216 #define vpermpd(_0, _1, _2) VPERMPD(_0, _1, _2)
1217 #define vunpcklps(_0, _1, _2) VUNPCKLPS(_0, _1, _2)
1218 #define vunpckhps(_0, _1, _2) VUNPCKHPS(_0, _1, _2)
1219 #define vunpcklpd(_0, _1, _2) VUNPCKLPD(_0, _1, _2)
1220 #define vunpckhpd(_0, _1, _2) VUNPCKHPD(_0, _1, _2)
1221 #define vshuff32x4(_0, _1, _2, _3) VSHUFF32x4(_0, _1, _2, _3)
1222 #define vshuff64x2(_0, _1, _2, _3) VSHUFF64x2(_0, _1, _2, _3)
1223 #define vinsertf128(_0, _1, _2, _3) VINSERTF128(_0, _1, _2, _3)
1224 #define vinsertf32x4(_0, _1, _2, _3) VINSERTF32x4(_0, _1, _2, _3)
1225 #define vinsertf32x8(_0, _1, _2, _3) VINSERTF32x8(_0, _1, _2, _3)
1226 #define vinsertf64x2(_0, _1, _2, _3) VINSERTF64x2(_0, _1, _2, _3)
1227 #define vinsertf64x4(_0, _1, _2, _3) VINSERTF64x4(_0, _1, _2, _3)
1228 #define vextractf128(_0, _1, _2) VEXTRACTF128(_0, _1, _2)
1229 #define vextractf32x4(_0, _1, _2) VEXTRACTF32x4(_0, _1, _2)
1230 #define vextractf32x8(_0, _1, _2) VEXTRACTF32x8(_0, _1, _2)
1231 #define vextractf64x2(_0, _1, _2) VEXTRACTF64x2(_0, _1, _2)
1232 #define vextractf64x4(_0, _1, _2) VEXTRACTF64x4(_0, _1, _2)
1233 #define vblendps(_0, _1, _2, _3) VBLENDPS(_0, _1, _2, _3)
1234 #define vblendpd(_0, _1, _2, _3) VBLENDPD(_0, _1, _2, _3)
1235 #define vblendmps(_0, _1, _2) VBLENDMSD(_0, _1, _2)
1236 #define vblendmpd(_0, _1, _2) VBLENDMPD(_0, _1, _2)
1237 
1238 // Prefetches
1239 
1240 #define PREFETCH(_0, _1) INSTR_(prefetcht##_0, _1)
1241 #define PREFETCHW0(_0) INSTR_(prefetchw, _0)
1242 #define PREFETCHW1(_0) INSTR_(prefetchwt1, _0)
1243 #define VGATHERPFDPS(_0, _1) INSTR_(vgatherpf##_0##dps, _1)
1244 #define VSCATTERPFDPS(_0, _1) INSTR_(vscatterpf##_0##dps, _1)
1245 #define VGATHERPFDPD(_0, _1) INSTR_(vgatherpf##_0##dpd, _1)
1246 #define VSCATTERPFDPD(_0, _1) INSTR_(vscatterpf##_0##dpd, _1)
1247 #define VGATHERPFQPS(_0, _1) INSTR_(vgatherpf##_0##qps, _1)
1248 #define VSCATTERPFQPS(_0, _1) INSTR_(vscatterpf##_0##qps, _1)
1249 #define VGATHERPFQPD(_0, _1) INSTR_(vgatherpf##_0##qpd, _1)
1250 #define VSCATTERPFQPD(_0, _1) INSTR_(vscatterpf##_0##qpd, _1)
1251 
1252 #define prefetch(_0, _1) PREFETCH(_0, _1)
1253 #define prefetchw0(_0) PREFETCHW0(_0)
1254 #define prefetchw1(_0) PREFETCHW1(_0)
1255 #define vgatherpfdps(_0, _1) VGATHERPFDPS(_0, _1)
1256 #define vscatterpfdps(_0, _1) VSCATTERPFDPS(_0, _1)
1257 #define vgatherpfdpd(_0, _1) VGATHERPFDPD(_0, _1)
1258 #define vscatterpfdpd(_0, _1) VSCATTERPFDPD(_0, _1)
1259 #define vgatherpfqps(_0, _1) VGATHERPFQPS(_0, _1)
1260 #define vscatterpfqps(_0, _1) VSCATTERPFQPS(_0, _1)
1261 #define vgatherpfqpd(_0, _1) VGATHERPFQPD(_0, _1)
1262 #define vscatterpfqpd(_0, _1) VSCATTERPFQPD(_0, _1)
1263 
1264 // Mask operations
1265 
1266 #ifdef __MIC__
1267 
1268 #define KMOVW(_0, _1) INSTR_(kmov, _0, _1)
1269 #define JKNZD(_0, _1) INSTR_(jknzd, _0, _1)
1270 
1271 #else
1272 
1273 #define KMOVW(_0, _1) INSTR_(kmovw, _0, _1)
1274 #define JKNZD(_0, _1) INSTR_(kortestw, _0, _0) INSTR_(jnz, _1)
1275 
1276 #endif
1277 
1278 #define KXNORW(_0, _1, _2) INSTR_(kxnorw, _0, _1, _2)
1279 #define KSHIFTRW(_0, _1, _2) INSTR_(kshiftrw, _0, _1, _2)
1280 
1281 #define kmovw(_0, _1) KMOVW(_0, _1)
1282 #define jknzd(_0, _1) JKNZD(_0, _1)
1283 #define kxnorw(_0, _1, _2) KXNORW(_0, _1, _2)
1284 #define kshiftrw(_0, _1, _2) KSHIFTRW(_0, _1, _2)
1285 
1286 // Other
1287 
1288 #define RDTSC() INSTR_(rdtsc)
1289 #define VZEROALL() INSTR_(vzeroall)
1290 #define VZEROUPPER() INSTR_(vzeroupper)
1291 
1292 #define rdtsc() RDTSC()
1293 #define vzeroall() VZEROALL()
1294 #define vzeroupper() VZEROUPPER()
1295 
1296 #endif
1297