comparison src/video/mmx.h @ 689:5bb080d35049

Date: Tue, 19 Aug 2003 17:57:00 +0200 From: Stephane Marchesin Subject: Re: [SDL] [patch] MMX alpha blit patches with MMX detection I think everything is correct now. I've done as much testing as I could, but some real-world testing wouldn't hurt, I think. The patch is here : http://icps.u-strasbg.fr/~marchesin/sdl_mmxblit.patch If you do byte-by-byte comparison of the output between C and MMX functions, you'll notice that the results for 555 and 565 RGB alpha blits aren't exactly the same. This is because MMX functions for 555 and 565 RGB have an higher accuracy. If you want the exact same behaviour that's possible by masking the three lower alpha bits in the MMX functions. Just ask ! I removed one MMX function because after I fixed it to match its C equivalent, it revealed to be slower than the C version on a PIII (although a bit faster on an Athlon XP). I've also added MMX and PIII replacements for SDL_memcpy. Those provide some speed up in testvidinfo -benchmark (at least for me, under linux & X11).
author Sam Lantinga <slouken@libsdl.org>
date Fri, 22 Aug 2003 05:51:19 +0000
parents
children 82b85b731fe3
comparison
equal deleted inserted replaced
688:c0522010bb6d 689:5bb080d35049
1 /* mmx.h
2
3 MultiMedia eXtensions GCC interface library for IA32.
4
5 To use this library, simply include this header file
6 and compile with GCC. You MUST have inlining enabled
7 in order for mmx_ok() to work; this can be done by
8 simply using -O on the GCC command line.
9
10 Compiling with -DMMX_TRACE will cause detailed trace
11 output to be sent to stderr for each mmx operation.
12 This adds lots of code, and obviously slows execution to
13 a crawl, but can be very useful for debugging.
14
15 THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY
16 EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT
17 LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY
18 AND FITNESS FOR ANY PARTICULAR PURPOSE.
19
20 1997-99 by H. Dietz and R. Fisher
21
22 Notes:
23 It appears that the latest gas has the pand problem fixed, therefore
24 I'll undefine BROKEN_PAND by default.
25 */
26
27 #ifndef _MMX_H
28 #define _MMX_H
29
30
31 /* Warning: at this writing, the version of GAS packaged
32 with most Linux distributions does not handle the
33 parallel AND operation mnemonic correctly. If the
34 symbol BROKEN_PAND is defined, a slower alternative
35 coding will be used. If execution of mmxtest results
36 in an illegal instruction fault, define this symbol.
37 */
38 #undef BROKEN_PAND
39
40
41 /* The type of an value that fits in an MMX register
42 (note that long long constant values MUST be suffixed
43 by LL and unsigned long long values by ULL, lest
44 they be truncated by the compiler)
45 */
46 typedef union {
47 long long q; /* Quadword (64-bit) value */
48 unsigned long long uq; /* Unsigned Quadword */
49 int d[2]; /* 2 Doubleword (32-bit) values */
50 unsigned int ud[2]; /* 2 Unsigned Doubleword */
51 short w[4]; /* 4 Word (16-bit) values */
52 unsigned short uw[4]; /* 4 Unsigned Word */
53 char b[8]; /* 8 Byte (8-bit) values */
54 unsigned char ub[8]; /* 8 Unsigned Byte */
55 float s[2]; /* Single-precision (32-bit) value */
56 } __attribute__ ((aligned (8))) mmx_t; /* On an 8-byte (64-bit) boundary */
57
58
59 #if 0
60 /* Function to test if multimedia instructions are supported...
61 */
62 inline extern int
63 mm_support(void)
64 {
65 /* Returns 1 if MMX instructions are supported,
66 3 if Cyrix MMX and Extended MMX instructions are supported
67 5 if AMD MMX and 3DNow! instructions are supported
68 0 if hardware does not support any of these
69 */
70 register int rval = 0;
71
72 __asm__ __volatile__ (
73 /* See if CPUID instruction is supported ... */
74 /* ... Get copies of EFLAGS into eax and ecx */
75 "pushf\n\t"
76 "popl %%eax\n\t"
77 "movl %%eax, %%ecx\n\t"
78
79 /* ... Toggle the ID bit in one copy and store */
80 /* to the EFLAGS reg */
81 "xorl $0x200000, %%eax\n\t"
82 "push %%eax\n\t"
83 "popf\n\t"
84
85 /* ... Get the (hopefully modified) EFLAGS */
86 "pushf\n\t"
87 "popl %%eax\n\t"
88
89 /* ... Compare and test result */
90 "xorl %%eax, %%ecx\n\t"
91 "testl $0x200000, %%ecx\n\t"
92 "jz NotSupported1\n\t" /* CPUID not supported */
93
94
95 /* Get standard CPUID information, and
96 go to a specific vendor section */
97 "movl $0, %%eax\n\t"
98 "cpuid\n\t"
99
100 /* Check for Intel */
101 "cmpl $0x756e6547, %%ebx\n\t"
102 "jne TryAMD\n\t"
103 "cmpl $0x49656e69, %%edx\n\t"
104 "jne TryAMD\n\t"
105 "cmpl $0x6c65746e, %%ecx\n"
106 "jne TryAMD\n\t"
107 "jmp Intel\n\t"
108
109 /* Check for AMD */
110 "\nTryAMD:\n\t"
111 "cmpl $0x68747541, %%ebx\n\t"
112 "jne TryCyrix\n\t"
113 "cmpl $0x69746e65, %%edx\n\t"
114 "jne TryCyrix\n\t"
115 "cmpl $0x444d4163, %%ecx\n"
116 "jne TryCyrix\n\t"
117 "jmp AMD\n\t"
118
119 /* Check for Cyrix */
120 "\nTryCyrix:\n\t"
121 "cmpl $0x69727943, %%ebx\n\t"
122 "jne NotSupported2\n\t"
123 "cmpl $0x736e4978, %%edx\n\t"
124 "jne NotSupported3\n\t"
125 "cmpl $0x64616574, %%ecx\n\t"
126 "jne NotSupported4\n\t"
127 /* Drop through to Cyrix... */
128
129
130 /* Cyrix Section */
131 /* See if extended CPUID level 80000001 is supported */
132 /* The value of CPUID/80000001 for the 6x86MX is undefined
133 according to the Cyrix CPU Detection Guide (Preliminary
134 Rev. 1.01 table 1), so we'll check the value of eax for
135 CPUID/0 to see if standard CPUID level 2 is supported.
136 According to the table, the only CPU which supports level
137 2 is also the only one which supports extended CPUID levels.
138 */
139 "cmpl $0x2, %%eax\n\t"
140 "jne MMXtest\n\t" /* Use standard CPUID instead */
141
142 /* Extended CPUID supported (in theory), so get extended
143 features */
144 "movl $0x80000001, %%eax\n\t"
145 "cpuid\n\t"
146 "testl $0x00800000, %%eax\n\t" /* Test for MMX */
147 "jz NotSupported5\n\t" /* MMX not supported */
148 "testl $0x01000000, %%eax\n\t" /* Test for Ext'd MMX */
149 "jnz EMMXSupported\n\t"
150 "movl $1, %0:\n\n\t" /* MMX Supported */
151 "jmp Return\n\n"
152 "EMMXSupported:\n\t"
153 "movl $3, %0:\n\n\t" /* EMMX and MMX Supported */
154 "jmp Return\n\t"
155
156
157 /* AMD Section */
158 "AMD:\n\t"
159
160 /* See if extended CPUID is supported */
161 "movl $0x80000000, %%eax\n\t"
162 "cpuid\n\t"
163 "cmpl $0x80000000, %%eax\n\t"
164 "jl MMXtest\n\t" /* Use standard CPUID instead */
165
166 /* Extended CPUID supported, so get extended features */
167 "movl $0x80000001, %%eax\n\t"
168 "cpuid\n\t"
169 "testl $0x00800000, %%edx\n\t" /* Test for MMX */
170 "jz NotSupported6\n\t" /* MMX not supported */
171 "testl $0x80000000, %%edx\n\t" /* Test for 3DNow! */
172 "jnz ThreeDNowSupported\n\t"
173 "movl $1, %0:\n\n\t" /* MMX Supported */
174 "jmp Return\n\n"
175 "ThreeDNowSupported:\n\t"
176 "movl $5, %0:\n\n\t" /* 3DNow! and MMX Supported */
177 "jmp Return\n\t"
178
179
180 /* Intel Section */
181 "Intel:\n\t"
182
183 /* Check for MMX */
184 "MMXtest:\n\t"
185 "movl $1, %%eax\n\t"
186 "cpuid\n\t"
187 "testl $0x00800000, %%edx\n\t" /* Test for MMX */
188 "jz NotSupported7\n\t" /* MMX Not supported */
189 "movl $1, %0:\n\n\t" /* MMX Supported */
190 "jmp Return\n\t"
191
192 /* Nothing supported */
193 "\nNotSupported1:\n\t"
194 "#movl $101, %0:\n\n\t"
195 "\nNotSupported2:\n\t"
196 "#movl $102, %0:\n\n\t"
197 "\nNotSupported3:\n\t"
198 "#movl $103, %0:\n\n\t"
199 "\nNotSupported4:\n\t"
200 "#movl $104, %0:\n\n\t"
201 "\nNotSupported5:\n\t"
202 "#movl $105, %0:\n\n\t"
203 "\nNotSupported6:\n\t"
204 "#movl $106, %0:\n\n\t"
205 "\nNotSupported7:\n\t"
206 "#movl $107, %0:\n\n\t"
207 "movl $0, %0:\n\n\t"
208
209 "Return:\n\t"
210 : "=a" (rval)
211 : /* no input */
212 : "eax", "ebx", "ecx", "edx"
213 );
214
215 /* Return */
216 return(rval);
217 }
218
219 /* Function to test if mmx instructions are supported...
220 */
221 inline extern int
222 mmx_ok(void)
223 {
224 /* Returns 1 if MMX instructions are supported, 0 otherwise */
225 return ( mm_support() & 0x1 );
226 }
227 #endif
228
229 /* Helper functions for the instruction macros that follow...
230 (note that memory-to-register, m2r, instructions are nearly
231 as efficient as register-to-register, r2r, instructions;
232 however, memory-to-memory instructions are really simulated
233 as a convenience, and are only 1/3 as efficient)
234 */
235 #ifdef MMX_TRACE
236
237 /* Include the stuff for printing a trace to stderr...
238 */
239
240 #include <stdio.h>
241
242 #define mmx_i2r(op, imm, reg) \
243 { \
244 mmx_t mmx_trace; \
245 mmx_trace.uq = (imm); \
246 printf(#op "_i2r(" #imm "=0x%08x%08x, ", \
247 mmx_trace.d[1], mmx_trace.d[0]); \
248 __asm__ __volatile__ ("movq %%" #reg ", %0" \
249 : "=X" (mmx_trace) \
250 : /* nothing */ ); \
251 printf(#reg "=0x%08x%08x) => ", \
252 mmx_trace.d[1], mmx_trace.d[0]); \
253 __asm__ __volatile__ (#op " %0, %%" #reg \
254 : /* nothing */ \
255 : "X" (imm)); \
256 __asm__ __volatile__ ("movq %%" #reg ", %0" \
257 : "=X" (mmx_trace) \
258 : /* nothing */ ); \
259 printf(#reg "=0x%08x%08x\n", \
260 mmx_trace.d[1], mmx_trace.d[0]); \
261 }
262
263 #define mmx_m2r(op, mem, reg) \
264 { \
265 mmx_t mmx_trace; \
266 mmx_trace = (mem); \
267 printf(#op "_m2r(" #mem "=0x%08x%08x, ", \
268 mmx_trace.d[1], mmx_trace.d[0]); \
269 __asm__ __volatile__ ("movq %%" #reg ", %0" \
270 : "=X" (mmx_trace) \
271 : /* nothing */ ); \
272 printf(#reg "=0x%08x%08x) => ", \
273 mmx_trace.d[1], mmx_trace.d[0]); \
274 __asm__ __volatile__ (#op " %0, %%" #reg \
275 : /* nothing */ \
276 : "X" (mem)); \
277 __asm__ __volatile__ ("movq %%" #reg ", %0" \
278 : "=X" (mmx_trace) \
279 : /* nothing */ ); \
280 printf(#reg "=0x%08x%08x\n", \
281 mmx_trace.d[1], mmx_trace.d[0]); \
282 }
283
284 #define mmx_r2m(op, reg, mem) \
285 { \
286 mmx_t mmx_trace; \
287 __asm__ __volatile__ ("movq %%" #reg ", %0" \
288 : "=X" (mmx_trace) \
289 : /* nothing */ ); \
290 printf(#op "_r2m(" #reg "=0x%08x%08x, ", \
291 mmx_trace.d[1], mmx_trace.d[0]); \
292 mmx_trace = (mem); \
293 printf(#mem "=0x%08x%08x) => ", \
294 mmx_trace.d[1], mmx_trace.d[0]); \
295 __asm__ __volatile__ (#op " %%" #reg ", %0" \
296 : "=X" (mem) \
297 : /* nothing */ ); \
298 mmx_trace = (mem); \
299 printf(#mem "=0x%08x%08x\n", \
300 mmx_trace.d[1], mmx_trace.d[0]); \
301 }
302
303 #define mmx_r2r(op, regs, regd) \
304 { \
305 mmx_t mmx_trace; \
306 __asm__ __volatile__ ("movq %%" #regs ", %0" \
307 : "=X" (mmx_trace) \
308 : /* nothing */ ); \
309 printf(#op "_r2r(" #regs "=0x%08x%08x, ", \
310 mmx_trace.d[1], mmx_trace.d[0]); \
311 __asm__ __volatile__ ("movq %%" #regd ", %0" \
312 : "=X" (mmx_trace) \
313 : /* nothing */ ); \
314 printf(#regd "=0x%08x%08x) => ", \
315 mmx_trace.d[1], mmx_trace.d[0]); \
316 __asm__ __volatile__ (#op " %" #regs ", %" #regd); \
317 __asm__ __volatile__ ("movq %%" #regd ", %0" \
318 : "=X" (mmx_trace) \
319 : /* nothing */ ); \
320 printf(#regd "=0x%08x%08x\n", \
321 mmx_trace.d[1], mmx_trace.d[0]); \
322 }
323
324 #define mmx_m2m(op, mems, memd) \
325 { \
326 mmx_t mmx_trace; \
327 mmx_trace = (mems); \
328 printf(#op "_m2m(" #mems "=0x%08x%08x, ", \
329 mmx_trace.d[1], mmx_trace.d[0]); \
330 mmx_trace = (memd); \
331 printf(#memd "=0x%08x%08x) => ", \
332 mmx_trace.d[1], mmx_trace.d[0]); \
333 __asm__ __volatile__ ("movq %0, %%mm0\n\t" \
334 #op " %1, %%mm0\n\t" \
335 "movq %%mm0, %0" \
336 : "=X" (memd) \
337 : "X" (mems)); \
338 mmx_trace = (memd); \
339 printf(#memd "=0x%08x%08x\n", \
340 mmx_trace.d[1], mmx_trace.d[0]); \
341 }
342
343 #else
344
345 /* These macros are a lot simpler without the tracing...
346 */
347
348 #define mmx_i2r(op, imm, reg) \
349 __asm__ __volatile__ (#op " %0, %%" #reg \
350 : /* nothing */ \
351 : "X" (imm) )
352
353 #define mmx_m2r(op, mem, reg) \
354 __asm__ __volatile__ (#op " %0, %%" #reg \
355 : /* nothing */ \
356 : "X" (mem))
357
358 #define mmx_r2m(op, reg, mem) \
359 __asm__ __volatile__ (#op " %%" #reg ", %0" \
360 : "=X" (mem) \
361 : /* nothing */ )
362
363 #define mmx_r2r(op, regs, regd) \
364 __asm__ __volatile__ (#op " %" #regs ", %" #regd)
365
366 #define mmx_m2m(op, mems, memd) \
367 __asm__ __volatile__ ("movq %0, %%mm0\n\t" \
368 #op " %1, %%mm0\n\t" \
369 "movq %%mm0, %0" \
370 : "=X" (memd) \
371 : "X" (mems))
372
373 #endif
374
375
376 /* 1x64 MOVe Quadword
377 (this is both a load and a store...
378 in fact, it is the only way to store)
379 */
380 #define movq_m2r(var, reg) mmx_m2r(movq, var, reg)
381 #define movq_r2m(reg, var) mmx_r2m(movq, reg, var)
382 #define movq_r2r(regs, regd) mmx_r2r(movq, regs, regd)
383 #define movq(vars, vard) \
384 __asm__ __volatile__ ("movq %1, %%mm0\n\t" \
385 "movq %%mm0, %0" \
386 : "=X" (vard) \
387 : "X" (vars))
388
389
390 /* 1x32 MOVe Doubleword
391 (like movq, this is both load and store...
392 but is most useful for moving things between
393 mmx registers and ordinary registers)
394 */
395 #define movd_m2r(var, reg) mmx_m2r(movd, var, reg)
396 #define movd_r2m(reg, var) mmx_r2m(movd, reg, var)
397 #define movd_r2r(regs, regd) mmx_r2r(movd, regs, regd)
398 #define movd(vars, vard) \
399 __asm__ __volatile__ ("movd %1, %%mm0\n\t" \
400 "movd %%mm0, %0" \
401 : "=X" (vard) \
402 : "X" (vars))
403
404
405 /* 2x32, 4x16, and 8x8 Parallel ADDs
406 */
407 #define paddd_m2r(var, reg) mmx_m2r(paddd, var, reg)
408 #define paddd_r2r(regs, regd) mmx_r2r(paddd, regs, regd)
409 #define paddd(vars, vard) mmx_m2m(paddd, vars, vard)
410
411 #define paddw_m2r(var, reg) mmx_m2r(paddw, var, reg)
412 #define paddw_r2r(regs, regd) mmx_r2r(paddw, regs, regd)
413 #define paddw(vars, vard) mmx_m2m(paddw, vars, vard)
414
415 #define paddb_m2r(var, reg) mmx_m2r(paddb, var, reg)
416 #define paddb_r2r(regs, regd) mmx_r2r(paddb, regs, regd)
417 #define paddb(vars, vard) mmx_m2m(paddb, vars, vard)
418
419
420 /* 4x16 and 8x8 Parallel ADDs using Saturation arithmetic
421 */
422 #define paddsw_m2r(var, reg) mmx_m2r(paddsw, var, reg)
423 #define paddsw_r2r(regs, regd) mmx_r2r(paddsw, regs, regd)
424 #define paddsw(vars, vard) mmx_m2m(paddsw, vars, vard)
425
426 #define paddsb_m2r(var, reg) mmx_m2r(paddsb, var, reg)
427 #define paddsb_r2r(regs, regd) mmx_r2r(paddsb, regs, regd)
428 #define paddsb(vars, vard) mmx_m2m(paddsb, vars, vard)
429
430
431 /* 4x16 and 8x8 Parallel ADDs using Unsigned Saturation arithmetic
432 */
433 #define paddusw_m2r(var, reg) mmx_m2r(paddusw, var, reg)
434 #define paddusw_r2r(regs, regd) mmx_r2r(paddusw, regs, regd)
435 #define paddusw(vars, vard) mmx_m2m(paddusw, vars, vard)
436
437 #define paddusb_m2r(var, reg) mmx_m2r(paddusb, var, reg)
438 #define paddusb_r2r(regs, regd) mmx_r2r(paddusb, regs, regd)
439 #define paddusb(vars, vard) mmx_m2m(paddusb, vars, vard)
440
441
442 /* 2x32, 4x16, and 8x8 Parallel SUBs
443 */
444 #define psubd_m2r(var, reg) mmx_m2r(psubd, var, reg)
445 #define psubd_r2r(regs, regd) mmx_r2r(psubd, regs, regd)
446 #define psubd(vars, vard) mmx_m2m(psubd, vars, vard)
447
448 #define psubw_m2r(var, reg) mmx_m2r(psubw, var, reg)
449 #define psubw_r2r(regs, regd) mmx_r2r(psubw, regs, regd)
450 #define psubw(vars, vard) mmx_m2m(psubw, vars, vard)
451
452 #define psubb_m2r(var, reg) mmx_m2r(psubb, var, reg)
453 #define psubb_r2r(regs, regd) mmx_r2r(psubb, regs, regd)
454 #define psubb(vars, vard) mmx_m2m(psubb, vars, vard)
455
456
457 /* 4x16 and 8x8 Parallel SUBs using Saturation arithmetic
458 */
459 #define psubsw_m2r(var, reg) mmx_m2r(psubsw, var, reg)
460 #define psubsw_r2r(regs, regd) mmx_r2r(psubsw, regs, regd)
461 #define psubsw(vars, vard) mmx_m2m(psubsw, vars, vard)
462
463 #define psubsb_m2r(var, reg) mmx_m2r(psubsb, var, reg)
464 #define psubsb_r2r(regs, regd) mmx_r2r(psubsb, regs, regd)
465 #define psubsb(vars, vard) mmx_m2m(psubsb, vars, vard)
466
467
468 /* 4x16 and 8x8 Parallel SUBs using Unsigned Saturation arithmetic
469 */
470 #define psubusw_m2r(var, reg) mmx_m2r(psubusw, var, reg)
471 #define psubusw_r2r(regs, regd) mmx_r2r(psubusw, regs, regd)
472 #define psubusw(vars, vard) mmx_m2m(psubusw, vars, vard)
473
474 #define psubusb_m2r(var, reg) mmx_m2r(psubusb, var, reg)
475 #define psubusb_r2r(regs, regd) mmx_r2r(psubusb, regs, regd)
476 #define psubusb(vars, vard) mmx_m2m(psubusb, vars, vard)
477
478
479 /* 4x16 Parallel MULs giving Low 4x16 portions of results
480 */
481 #define pmullw_m2r(var, reg) mmx_m2r(pmullw, var, reg)
482 #define pmullw_r2r(regs, regd) mmx_r2r(pmullw, regs, regd)
483 #define pmullw(vars, vard) mmx_m2m(pmullw, vars, vard)
484
485
486 /* 4x16 Parallel MULs giving High 4x16 portions of results
487 */
488 #define pmulhw_m2r(var, reg) mmx_m2r(pmulhw, var, reg)
489 #define pmulhw_r2r(regs, regd) mmx_r2r(pmulhw, regs, regd)
490 #define pmulhw(vars, vard) mmx_m2m(pmulhw, vars, vard)
491
492
493 /* 4x16->2x32 Parallel Mul-ADD
494 (muls like pmullw, then adds adjacent 16-bit fields
495 in the multiply result to make the final 2x32 result)
496 */
497 #define pmaddwd_m2r(var, reg) mmx_m2r(pmaddwd, var, reg)
498 #define pmaddwd_r2r(regs, regd) mmx_r2r(pmaddwd, regs, regd)
499 #define pmaddwd(vars, vard) mmx_m2m(pmaddwd, vars, vard)
500
501
502 /* 1x64 bitwise AND
503 */
504 #ifdef BROKEN_PAND
505 #define pand_m2r(var, reg) \
506 { \
507 mmx_m2r(pandn, (mmx_t) -1LL, reg); \
508 mmx_m2r(pandn, var, reg); \
509 }
510 #define pand_r2r(regs, regd) \
511 { \
512 mmx_m2r(pandn, (mmx_t) -1LL, regd); \
513 mmx_r2r(pandn, regs, regd) \
514 }
515 #define pand(vars, vard) \
516 { \
517 movq_m2r(vard, mm0); \
518 mmx_m2r(pandn, (mmx_t) -1LL, mm0); \
519 mmx_m2r(pandn, vars, mm0); \
520 movq_r2m(mm0, vard); \
521 }
522 #else
523 #define pand_m2r(var, reg) mmx_m2r(pand, var, reg)
524 #define pand_r2r(regs, regd) mmx_r2r(pand, regs, regd)
525 #define pand(vars, vard) mmx_m2m(pand, vars, vard)
526 #endif
527
528
529 /* 1x64 bitwise AND with Not the destination
530 */
531 #define pandn_m2r(var, reg) mmx_m2r(pandn, var, reg)
532 #define pandn_r2r(regs, regd) mmx_r2r(pandn, regs, regd)
533 #define pandn(vars, vard) mmx_m2m(pandn, vars, vard)
534
535
536 /* 1x64 bitwise OR
537 */
538 #define por_m2r(var, reg) mmx_m2r(por, var, reg)
539 #define por_r2r(regs, regd) mmx_r2r(por, regs, regd)
540 #define por(vars, vard) mmx_m2m(por, vars, vard)
541
542
543 /* 1x64 bitwise eXclusive OR
544 */
545 #define pxor_m2r(var, reg) mmx_m2r(pxor, var, reg)
546 #define pxor_r2r(regs, regd) mmx_r2r(pxor, regs, regd)
547 #define pxor(vars, vard) mmx_m2m(pxor, vars, vard)
548
549
550 /* 2x32, 4x16, and 8x8 Parallel CoMPare for EQuality
551 (resulting fields are either 0 or -1)
552 */
553 #define pcmpeqd_m2r(var, reg) mmx_m2r(pcmpeqd, var, reg)
554 #define pcmpeqd_r2r(regs, regd) mmx_r2r(pcmpeqd, regs, regd)
555 #define pcmpeqd(vars, vard) mmx_m2m(pcmpeqd, vars, vard)
556
557 #define pcmpeqw_m2r(var, reg) mmx_m2r(pcmpeqw, var, reg)
558 #define pcmpeqw_r2r(regs, regd) mmx_r2r(pcmpeqw, regs, regd)
559 #define pcmpeqw(vars, vard) mmx_m2m(pcmpeqw, vars, vard)
560
561 #define pcmpeqb_m2r(var, reg) mmx_m2r(pcmpeqb, var, reg)
562 #define pcmpeqb_r2r(regs, regd) mmx_r2r(pcmpeqb, regs, regd)
563 #define pcmpeqb(vars, vard) mmx_m2m(pcmpeqb, vars, vard)
564
565
566 /* 2x32, 4x16, and 8x8 Parallel CoMPare for Greater Than
567 (resulting fields are either 0 or -1)
568 */
569 #define pcmpgtd_m2r(var, reg) mmx_m2r(pcmpgtd, var, reg)
570 #define pcmpgtd_r2r(regs, regd) mmx_r2r(pcmpgtd, regs, regd)
571 #define pcmpgtd(vars, vard) mmx_m2m(pcmpgtd, vars, vard)
572
573 #define pcmpgtw_m2r(var, reg) mmx_m2r(pcmpgtw, var, reg)
574 #define pcmpgtw_r2r(regs, regd) mmx_r2r(pcmpgtw, regs, regd)
575 #define pcmpgtw(vars, vard) mmx_m2m(pcmpgtw, vars, vard)
576
577 #define pcmpgtb_m2r(var, reg) mmx_m2r(pcmpgtb, var, reg)
578 #define pcmpgtb_r2r(regs, regd) mmx_r2r(pcmpgtb, regs, regd)
579 #define pcmpgtb(vars, vard) mmx_m2m(pcmpgtb, vars, vard)
580
581
582 /* 1x64, 2x32, and 4x16 Parallel Shift Left Logical
583 */
584 #define psllq_i2r(imm, reg) mmx_i2r(psllq, imm, reg)
585 #define psllq_m2r(var, reg) mmx_m2r(psllq, var, reg)
586 #define psllq_r2r(regs, regd) mmx_r2r(psllq, regs, regd)
587 #define psllq(vars, vard) mmx_m2m(psllq, vars, vard)
588
589 #define pslld_i2r(imm, reg) mmx_i2r(pslld, imm, reg)
590 #define pslld_m2r(var, reg) mmx_m2r(pslld, var, reg)
591 #define pslld_r2r(regs, regd) mmx_r2r(pslld, regs, regd)
592 #define pslld(vars, vard) mmx_m2m(pslld, vars, vard)
593
594 #define psllw_i2r(imm, reg) mmx_i2r(psllw, imm, reg)
595 #define psllw_m2r(var, reg) mmx_m2r(psllw, var, reg)
596 #define psllw_r2r(regs, regd) mmx_r2r(psllw, regs, regd)
597 #define psllw(vars, vard) mmx_m2m(psllw, vars, vard)
598
599
600 /* 1x64, 2x32, and 4x16 Parallel Shift Right Logical
601 */
602 #define psrlq_i2r(imm, reg) mmx_i2r(psrlq, imm, reg)
603 #define psrlq_m2r(var, reg) mmx_m2r(psrlq, var, reg)
604 #define psrlq_r2r(regs, regd) mmx_r2r(psrlq, regs, regd)
605 #define psrlq(vars, vard) mmx_m2m(psrlq, vars, vard)
606
607 #define psrld_i2r(imm, reg) mmx_i2r(psrld, imm, reg)
608 #define psrld_m2r(var, reg) mmx_m2r(psrld, var, reg)
609 #define psrld_r2r(regs, regd) mmx_r2r(psrld, regs, regd)
610 #define psrld(vars, vard) mmx_m2m(psrld, vars, vard)
611
612 #define psrlw_i2r(imm, reg) mmx_i2r(psrlw, imm, reg)
613 #define psrlw_m2r(var, reg) mmx_m2r(psrlw, var, reg)
614 #define psrlw_r2r(regs, regd) mmx_r2r(psrlw, regs, regd)
615 #define psrlw(vars, vard) mmx_m2m(psrlw, vars, vard)
616
617
618 /* 2x32 and 4x16 Parallel Shift Right Arithmetic
619 */
620 #define psrad_i2r(imm, reg) mmx_i2r(psrad, imm, reg)
621 #define psrad_m2r(var, reg) mmx_m2r(psrad, var, reg)
622 #define psrad_r2r(regs, regd) mmx_r2r(psrad, regs, regd)
623 #define psrad(vars, vard) mmx_m2m(psrad, vars, vard)
624
625 #define psraw_i2r(imm, reg) mmx_i2r(psraw, imm, reg)
626 #define psraw_m2r(var, reg) mmx_m2r(psraw, var, reg)
627 #define psraw_r2r(regs, regd) mmx_r2r(psraw, regs, regd)
628 #define psraw(vars, vard) mmx_m2m(psraw, vars, vard)
629
630
631 /* 2x32->4x16 and 4x16->8x8 PACK and Signed Saturate
632 (packs source and dest fields into dest in that order)
633 */
634 #define packssdw_m2r(var, reg) mmx_m2r(packssdw, var, reg)
635 #define packssdw_r2r(regs, regd) mmx_r2r(packssdw, regs, regd)
636 #define packssdw(vars, vard) mmx_m2m(packssdw, vars, vard)
637
638 #define packsswb_m2r(var, reg) mmx_m2r(packsswb, var, reg)
639 #define packsswb_r2r(regs, regd) mmx_r2r(packsswb, regs, regd)
640 #define packsswb(vars, vard) mmx_m2m(packsswb, vars, vard)
641
642
643 /* 4x16->8x8 PACK and Unsigned Saturate
644 (packs source and dest fields into dest in that order)
645 */
646 #define packuswb_m2r(var, reg) mmx_m2r(packuswb, var, reg)
647 #define packuswb_r2r(regs, regd) mmx_r2r(packuswb, regs, regd)
648 #define packuswb(vars, vard) mmx_m2m(packuswb, vars, vard)
649
650
651 /* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK Low
652 (interleaves low half of dest with low half of source
653 as padding in each result field)
654 */
655 #define punpckldq_m2r(var, reg) mmx_m2r(punpckldq, var, reg)
656 #define punpckldq_r2r(regs, regd) mmx_r2r(punpckldq, regs, regd)
657 #define punpckldq(vars, vard) mmx_m2m(punpckldq, vars, vard)
658
659 #define punpcklwd_m2r(var, reg) mmx_m2r(punpcklwd, var, reg)
660 #define punpcklwd_r2r(regs, regd) mmx_r2r(punpcklwd, regs, regd)
661 #define punpcklwd(vars, vard) mmx_m2m(punpcklwd, vars, vard)
662
663 #define punpcklbw_m2r(var, reg) mmx_m2r(punpcklbw, var, reg)
664 #define punpcklbw_r2r(regs, regd) mmx_r2r(punpcklbw, regs, regd)
665 #define punpcklbw(vars, vard) mmx_m2m(punpcklbw, vars, vard)
666
667
668 /* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK High
669 (interleaves high half of dest with high half of source
670 as padding in each result field)
671 */
672 #define punpckhdq_m2r(var, reg) mmx_m2r(punpckhdq, var, reg)
673 #define punpckhdq_r2r(regs, regd) mmx_r2r(punpckhdq, regs, regd)
674 #define punpckhdq(vars, vard) mmx_m2m(punpckhdq, vars, vard)
675
676 #define punpckhwd_m2r(var, reg) mmx_m2r(punpckhwd, var, reg)
677 #define punpckhwd_r2r(regs, regd) mmx_r2r(punpckhwd, regs, regd)
678 #define punpckhwd(vars, vard) mmx_m2m(punpckhwd, vars, vard)
679
680 #define punpckhbw_m2r(var, reg) mmx_m2r(punpckhbw, var, reg)
681 #define punpckhbw_r2r(regs, regd) mmx_r2r(punpckhbw, regs, regd)
682 #define punpckhbw(vars, vard) mmx_m2m(punpckhbw, vars, vard)
683
684
685 /* Empty MMx State
686 (used to clean-up when going from mmx to float use
687 of the registers that are shared by both; note that
688 there is no float-to-mmx operation needed, because
689 only the float tag word info is corruptible)
690 */
691 #ifdef MMX_TRACE
692
693 #define emms() \
694 { \
695 printf("emms()\n"); \
696 __asm__ __volatile__ ("emms"); \
697 }
698
699 #else
700
701 #define emms() __asm__ __volatile__ ("emms")
702
703 #endif
704
705 #endif
706