3 MultiMedia eXtensions GCC interface library for IA32.
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.
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.
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.
20 1997-98 by H. Dietz and R. Fisher
23 97-98* R.Fisher Early versions
24 980501 R.Fisher Original Release
25 980611* H.Dietz Rewrite, correctly implementing inlines, and
26 R.Fisher including direct register accesses.
27 980616 R.Fisher Release of 980611 as 980616.
28 980714 R.Fisher Minor corrections to Makefile, etc.
29 980715 R.Fisher mmx_ok() now prevents optimizer from using
31 mmx_ok() now checks if cpuid instruction is
32 available before trying to use it.
33 980726* R.Fisher mm_support() searches for AMD 3DNow, Cyrix
34 Extended MMX, and standard MMX. It returns a
35 value which is positive if any of these are
36 supported, and can be masked with constants to
37 see which. mmx_ok() is now a call to this
38 980726* R.Fisher Added i2r support for shift functions
39 980919 R.Fisher Fixed AMD extended feature recognition bug.
40 980921 R.Fisher Added definition/check for _MMX_H.
41 Added "float s[2]" to mmx_t for use with
42 3DNow and EMMX. So same mmx_t can be used.
43 981013 R.Fisher Fixed cpuid function 1 bug (looked at wrong reg)
44 Fixed psllq_i2r error in mmxtest.c
46 * Unreleased (internal or interim) versions
49 It appears that the latest gas has the pand problem fixed, therefore
50 I'll undefine BROKEN_PAND by default.
51 String compares may be quicker than the multiple test/jumps in vendor
52 test sequence in mmx_ok(), but I'm not concerned with that right now.
55 Jussi Laako for pointing out the errors ultimately found to be
56 connected to the failure to notify the optimizer of clobbered values.
57 Roger Hardiman for reminding us that CPUID isn't everywhere, and that
58 someone may actually try to use this on a machine without CPUID.
59 Also for suggesting code for checking this.
60 Robert Dale for pointing out the AMD recognition bug.
61 Jimmy Mayfield and Carl Witty for pointing out the Intel recognition
63 Carl Witty for pointing out the psllq_i2r test bug.
69 /* Warning: at this writing, the version of GAS packaged
70 with most Linux distributions does not handle the
71 parallel AND operation mnemonic correctly. If the
72 symbol BROKEN_PAND is defined, a slower alternative
73 coding will be used. If execution of mmxtest results
74 in an illegal instruction fault, define this symbol.
79 /* The type of an value that fits in an MMX register
80 (note that long long constant values MUST be suffixed
81 by LL and unsigned long long values by ULL, lest
82 they be truncated by the compiler)
85 long long q; /* Quadword (64-bit) value */
86 unsigned long long uq; /* Unsigned Quadword */
87 int d[2]; /* 2 Doubleword (32-bit) values */
88 unsigned int ud[2]; /* 2 Unsigned Doubleword */
89 short w[4]; /* 4 Word (16-bit) values */
90 unsigned short uw[4]; /* 4 Unsigned Word */
91 char b[8]; /* 8 Byte (8-bit) values */
92 unsigned char ub[8]; /* 8 Unsigned Byte */
93 float s[2]; /* Single-precision (32-bit) value */
94 } __attribute__ ((aligned (8))) mmx_t;
96 /* Helper functions for the instruction macros that follow...
97 (note that memory-to-register, m2r, instructions are nearly
98 as efficient as register-to-register, r2r, instructions;
99 however, memory-to-memory instructions are really simulated
100 as a convenience, and are only 1/3 as efficient)
103 /* These macros are a lot simpler without the tracing...
106 #define mmx_i2r(op, imm, reg) \
107 __asm__ __volatile__ (#op " $" #imm ", %%" #reg \
111 #define mmx_m2r(op, mem, reg) \
112 __asm__ __volatile__ (#op " %0, %%" #reg \
116 #define mmx_r2m(op, reg, mem) \
117 __asm__ __volatile__ (#op " %%" #reg ", %0" \
121 #define mmx_a2r(op, mem, reg) \
122 __asm__ __volatile__ (#op " %0, %%" #reg \
126 #define mmx_r2a(op, reg, mem) \
127 __asm__ __volatile__ (#op " %%" #reg ", %0" \
131 #define mmx_r2r(op, regs, regd) \
132 __asm__ __volatile__ (#op " %" #regs ", %" #regd)
134 #define mmx_m2m(op, mems, memd) \
135 __asm__ __volatile__ ("movq %0, %%mm0\n\t" \
136 #op " %1, %%mm0\n\t" \
141 /* 1x64 MOVE Quadword
142 (this is both a load and a store...
143 in fact, it is the only way to store)
145 #define movq_m2r(var, reg) mmx_m2r(movq, var, reg)
146 #define movq_r2m(reg, var) mmx_r2m(movq, reg, var)
147 #define movq_r2r(regs, regd) mmx_r2r(movq, regs, regd)
148 #define movq(vars, vard) \
149 __asm__ __volatile__ ("movq %1, %%mm0\n\t" \
153 #define movntq_r2m(reg, var) mmx_r2m(movntq, reg, var)
156 /* 1x32 MOVE Doubleword
157 (like movq, this is both load and store...
158 but is most useful for moving things between
159 mmx registers and ordinary registers)
161 #define movd_m2r(var, reg) mmx_a2r(movd, var, reg)
162 #define movd_r2m(reg, var) mmx_r2a(movd, reg, var)
163 #define movd_r2r(regs, regd) mmx_r2r(movd, regs, regd)
164 #define movd(vars, vard) \
165 __asm__ __volatile__ ("movd %1, %%mm0\n\t" \
171 /* 2x32, 4x16, and 8x8 Parallel ADDs
173 #define paddd_m2r(var, reg) mmx_m2r(paddd, var, reg)
174 #define paddd_r2r(regs, regd) mmx_r2r(paddd, regs, regd)
175 #define paddd(vars, vard) mmx_m2m(paddd, vars, vard)
177 #define paddw_m2r(var, reg) mmx_m2r(paddw, var, reg)
178 #define paddw_r2r(regs, regd) mmx_r2r(paddw, regs, regd)
179 #define paddw(vars, vard) mmx_m2m(paddw, vars, vard)
181 #define paddb_m2r(var, reg) mmx_m2r(paddb, var, reg)
182 #define paddb_r2r(regs, regd) mmx_r2r(paddb, regs, regd)
183 #define paddb(vars, vard) mmx_m2m(paddb, vars, vard)
186 /* 4x16 and 8x8 Parallel ADDs using Saturation arithmetic
188 #define paddsw_m2r(var, reg) mmx_m2r(paddsw, var, reg)
189 #define paddsw_r2r(regs, regd) mmx_r2r(paddsw, regs, regd)
190 #define paddsw(vars, vard) mmx_m2m(paddsw, vars, vard)
192 #define paddsb_m2r(var, reg) mmx_m2r(paddsb, var, reg)
193 #define paddsb_r2r(regs, regd) mmx_r2r(paddsb, regs, regd)
194 #define paddsb(vars, vard) mmx_m2m(paddsb, vars, vard)
197 /* 4x16 and 8x8 Parallel ADDs using Unsigned Saturation arithmetic
199 #define paddusw_m2r(var, reg) mmx_m2r(paddusw, var, reg)
200 #define paddusw_r2r(regs, regd) mmx_r2r(paddusw, regs, regd)
201 #define paddusw(vars, vard) mmx_m2m(paddusw, vars, vard)
203 #define paddusb_m2r(var, reg) mmx_m2r(paddusb, var, reg)
204 #define paddusb_r2r(regs, regd) mmx_r2r(paddusb, regs, regd)
205 #define paddusb(vars, vard) mmx_m2m(paddusb, vars, vard)
208 /* 2x32, 4x16, and 8x8 Parallel SUBs
210 #define psubd_m2r(var, reg) mmx_m2r(psubd, var, reg)
211 #define psubd_r2r(regs, regd) mmx_r2r(psubd, regs, regd)
212 #define psubd(vars, vard) mmx_m2m(psubd, vars, vard)
214 #define psubw_m2r(var, reg) mmx_m2r(psubw, var, reg)
215 #define psubw_r2r(regs, regd) mmx_r2r(psubw, regs, regd)
216 #define psubw(vars, vard) mmx_m2m(psubw, vars, vard)
218 #define psubb_m2r(var, reg) mmx_m2r(psubb, var, reg)
219 #define psubb_r2r(regs, regd) mmx_r2r(psubb, regs, regd)
220 #define psubb(vars, vard) mmx_m2m(psubb, vars, vard)
223 /* 4x16 and 8x8 Parallel SUBs using Saturation arithmetic
225 #define psubsw_m2r(var, reg) mmx_m2r(psubsw, var, reg)
226 #define psubsw_r2r(regs, regd) mmx_r2r(psubsw, regs, regd)
227 #define psubsw(vars, vard) mmx_m2m(psubsw, vars, vard)
229 #define psubsb_m2r(var, reg) mmx_m2r(psubsb, var, reg)
230 #define psubsb_r2r(regs, regd) mmx_r2r(psubsb, regs, regd)
231 #define psubsb(vars, vard) mmx_m2m(psubsb, vars, vard)
234 /* 4x16 and 8x8 Parallel SUBs using Unsigned Saturation arithmetic
236 #define psubusw_m2r(var, reg) mmx_m2r(psubusw, var, reg)
237 #define psubusw_r2r(regs, regd) mmx_r2r(psubusw, regs, regd)
238 #define psubusw(vars, vard) mmx_m2m(psubusw, vars, vard)
240 #define psubusb_m2r(var, reg) mmx_m2r(psubusb, var, reg)
241 #define psubusb_r2r(regs, regd) mmx_r2r(psubusb, regs, regd)
242 #define psubusb(vars, vard) mmx_m2m(psubusb, vars, vard)
245 /* 4x16 Parallel MULs giving Low 4x16 portions of results
247 #define pmullw_m2r(var, reg) mmx_m2r(pmullw, var, reg)
248 #define pmullw_r2r(regs, regd) mmx_r2r(pmullw, regs, regd)
249 #define pmullw(vars, vard) mmx_m2m(pmullw, vars, vard)
252 /* 4x16 Parallel MULs giving High 4x16 portions of results
254 #define pmulhw_m2r(var, reg) mmx_m2r(pmulhw, var, reg)
255 #define pmulhw_r2r(regs, regd) mmx_r2r(pmulhw, regs, regd)
256 #define pmulhw(vars, vard) mmx_m2m(pmulhw, vars, vard)
259 /* 4x16->2x32 Parallel Mul-ADD
260 (muls like pmullw, then adds adjacent 16-bit fields
261 in the multiply result to make the final 2x32 result)
263 #define pmaddwd_m2r(var, reg) mmx_m2r(pmaddwd, var, reg)
264 #define pmaddwd_r2r(regs, regd) mmx_r2r(pmaddwd, regs, regd)
265 #define pmaddwd(vars, vard) mmx_m2m(pmaddwd, vars, vard)
271 #define pand_m2r(var, reg) \
273 mmx_m2r(pandn, (mmx_t) -1LL, reg); \
274 mmx_m2r(pandn, var, reg); \
276 #define pand_r2r(regs, regd) \
278 mmx_m2r(pandn, (mmx_t) -1LL, regd); \
279 mmx_r2r(pandn, regs, regd) \
281 #define pand(vars, vard) \
283 movq_m2r(vard, mm0); \
284 mmx_m2r(pandn, (mmx_t) -1LL, mm0); \
285 mmx_m2r(pandn, vars, mm0); \
286 movq_r2m(mm0, vard); \
289 #define pand_m2r(var, reg) mmx_m2r(pand, var, reg)
290 #define pand_r2r(regs, regd) mmx_r2r(pand, regs, regd)
291 #define pand(vars, vard) mmx_m2m(pand, vars, vard)
295 /* 1x64 bitwise AND with Not the destination
297 #define pandn_m2r(var, reg) mmx_m2r(pandn, var, reg)
298 #define pandn_r2r(regs, regd) mmx_r2r(pandn, regs, regd)
299 #define pandn(vars, vard) mmx_m2m(pandn, vars, vard)
304 #define por_m2r(var, reg) mmx_m2r(por, var, reg)
305 #define por_r2r(regs, regd) mmx_r2r(por, regs, regd)
306 #define por(vars, vard) mmx_m2m(por, vars, vard)
309 /* 1x64 bitwise eXclusive OR
311 #define pxor_m2r(var, reg) mmx_m2r(pxor, var, reg)
312 #define pxor_r2r(regs, regd) mmx_r2r(pxor, regs, regd)
313 #define pxor(vars, vard) mmx_m2m(pxor, vars, vard)
316 /* 2x32, 4x16, and 8x8 Parallel CoMPare for EQuality
317 (resulting fields are either 0 or -1)
319 #define pcmpeqd_m2r(var, reg) mmx_m2r(pcmpeqd, var, reg)
320 #define pcmpeqd_r2r(regs, regd) mmx_r2r(pcmpeqd, regs, regd)
321 #define pcmpeqd(vars, vard) mmx_m2m(pcmpeqd, vars, vard)
323 #define pcmpeqw_m2r(var, reg) mmx_m2r(pcmpeqw, var, reg)
324 #define pcmpeqw_r2r(regs, regd) mmx_r2r(pcmpeqw, regs, regd)
325 #define pcmpeqw(vars, vard) mmx_m2m(pcmpeqw, vars, vard)
327 #define pcmpeqb_m2r(var, reg) mmx_m2r(pcmpeqb, var, reg)
328 #define pcmpeqb_r2r(regs, regd) mmx_r2r(pcmpeqb, regs, regd)
329 #define pcmpeqb(vars, vard) mmx_m2m(pcmpeqb, vars, vard)
332 /* 2x32, 4x16, and 8x8 Parallel CoMPare for Greater Than
333 (resulting fields are either 0 or -1)
335 #define pcmpgtd_m2r(var, reg) mmx_m2r(pcmpgtd, var, reg)
336 #define pcmpgtd_r2r(regs, regd) mmx_r2r(pcmpgtd, regs, regd)
337 #define pcmpgtd(vars, vard) mmx_m2m(pcmpgtd, vars, vard)
339 #define pcmpgtw_m2r(var, reg) mmx_m2r(pcmpgtw, var, reg)
340 #define pcmpgtw_r2r(regs, regd) mmx_r2r(pcmpgtw, regs, regd)
341 #define pcmpgtw(vars, vard) mmx_m2m(pcmpgtw, vars, vard)
343 #define pcmpgtb_m2r(var, reg) mmx_m2r(pcmpgtb, var, reg)
344 #define pcmpgtb_r2r(regs, regd) mmx_r2r(pcmpgtb, regs, regd)
345 #define pcmpgtb(vars, vard) mmx_m2m(pcmpgtb, vars, vard)
348 /* 1x64, 2x32, and 4x16 Parallel Shift Left Logical
350 #define psllq_i2r(imm, reg) mmx_i2r(psllq, imm, reg)
351 #define psllq_m2r(var, reg) mmx_m2r(psllq, var, reg)
352 #define psllq_r2r(regs, regd) mmx_r2r(psllq, regs, regd)
353 #define psllq(vars, vard) mmx_m2m(psllq, vars, vard)
355 #define pslld_i2r(imm, reg) mmx_i2r(pslld, imm, reg)
356 #define pslld_m2r(var, reg) mmx_m2r(pslld, var, reg)
357 #define pslld_r2r(regs, regd) mmx_r2r(pslld, regs, regd)
358 #define pslld(vars, vard) mmx_m2m(pslld, vars, vard)
360 #define psllw_i2r(imm, reg) mmx_i2r(psllw, imm, reg)
361 #define psllw_m2r(var, reg) mmx_m2r(psllw, var, reg)
362 #define psllw_r2r(regs, regd) mmx_r2r(psllw, regs, regd)
363 #define psllw(vars, vard) mmx_m2m(psllw, vars, vard)
366 /* 1x64, 2x32, and 4x16 Parallel Shift Right Logical
368 #define psrlq_i2r(imm, reg) mmx_i2r(psrlq, imm, reg)
369 #define psrlq_m2r(var, reg) mmx_m2r(psrlq, var, reg)
370 #define psrlq_r2r(regs, regd) mmx_r2r(psrlq, regs, regd)
371 #define psrlq(vars, vard) mmx_m2m(psrlq, vars, vard)
373 #define psrld_i2r(imm, reg) mmx_i2r(psrld, imm, reg)
374 #define psrld_m2r(var, reg) mmx_m2r(psrld, var, reg)
375 #define psrld_r2r(regs, regd) mmx_r2r(psrld, regs, regd)
376 #define psrld(vars, vard) mmx_m2m(psrld, vars, vard)
378 #define psrlw_i2r(imm, reg) mmx_i2r(psrlw, imm, reg)
379 #define psrlw_m2r(var, reg) mmx_m2r(psrlw, var, reg)
380 #define psrlw_r2r(regs, regd) mmx_r2r(psrlw, regs, regd)
381 #define psrlw(vars, vard) mmx_m2m(psrlw, vars, vard)
384 /* 2x32 and 4x16 Parallel Shift Right Arithmetic
386 #define psrad_i2r(imm, reg) mmx_i2r(psrad, imm, reg)
387 #define psrad_m2r(var, reg) mmx_m2r(psrad, var, reg)
388 #define psrad_r2r(regs, regd) mmx_r2r(psrad, regs, regd)
389 #define psrad(vars, vard) mmx_m2m(psrad, vars, vard)
391 #define psraw_i2r(imm, reg) mmx_i2r(psraw, imm, reg)
392 #define psraw_m2r(var, reg) mmx_m2r(psraw, var, reg)
393 #define psraw_r2r(regs, regd) mmx_r2r(psraw, regs, regd)
394 #define psraw(vars, vard) mmx_m2m(psraw, vars, vard)
397 /* 2x32->4x16 and 4x16->8x8 PACK and Signed Saturate
398 (packs source and dest fields into dest in that order)
400 #define packssdw_m2r(var, reg) mmx_m2r(packssdw, var, reg)
401 #define packssdw_r2r(regs, regd) mmx_r2r(packssdw, regs, regd)
402 #define packssdw(vars, vard) mmx_m2m(packssdw, vars, vard)
404 #define packsswb_m2r(var, reg) mmx_m2r(packsswb, var, reg)
405 #define packsswb_r2r(regs, regd) mmx_r2r(packsswb, regs, regd)
406 #define packsswb(vars, vard) mmx_m2m(packsswb, vars, vard)
409 /* 4x16->8x8 PACK and Unsigned Saturate
410 (packs source and dest fields into dest in that order)
412 #define packuswb_m2r(var, reg) mmx_m2r(packuswb, var, reg)
413 #define packuswb_r2r(regs, regd) mmx_r2r(packuswb, regs, regd)
414 #define packuswb(vars, vard) mmx_m2m(packuswb, vars, vard)
417 /* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK Low
418 (interleaves low half of dest with low half of source
419 as padding in each result field)
421 #define punpckldq_m2r(var, reg) mmx_m2r(punpckldq, var, reg)
422 #define punpckldq_r2r(regs, regd) mmx_r2r(punpckldq, regs, regd)
423 #define punpckldq(vars, vard) mmx_m2m(punpckldq, vars, vard)
425 #define punpcklwd_m2r(var, reg) mmx_m2r(punpcklwd, var, reg)
426 #define punpcklwd_r2r(regs, regd) mmx_r2r(punpcklwd, regs, regd)
427 #define punpcklwd(vars, vard) mmx_m2m(punpcklwd, vars, vard)
429 #define punpcklbw_m2r(var, reg) mmx_m2r(punpcklbw, var, reg)
430 #define punpcklbw_r2r(regs, regd) mmx_r2r(punpcklbw, regs, regd)
431 #define punpcklbw(vars, vard) mmx_m2m(punpcklbw, vars, vard)
434 /* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK High
435 (interleaves high half of dest with high half of source
436 as padding in each result field)
438 #define punpckhdq_m2r(var, reg) mmx_m2r(punpckhdq, var, reg)
439 #define punpckhdq_r2r(regs, regd) mmx_r2r(punpckhdq, regs, regd)
440 #define punpckhdq(vars, vard) mmx_m2m(punpckhdq, vars, vard)
442 #define punpckhwd_m2r(var, reg) mmx_m2r(punpckhwd, var, reg)
443 #define punpckhwd_r2r(regs, regd) mmx_r2r(punpckhwd, regs, regd)
444 #define punpckhwd(vars, vard) mmx_m2m(punpckhwd, vars, vard)
446 #define punpckhbw_m2r(var, reg) mmx_m2r(punpckhbw, var, reg)
447 #define punpckhbw_r2r(regs, regd) mmx_r2r(punpckhbw, regs, regd)
448 #define punpckhbw(vars, vard) mmx_m2m(punpckhbw, vars, vard)
450 #define MOVE_8DWORDS_MMX(src,dst) \
452 "movq (%1), %%mm0 \n" \
453 "movq 0x8(%1), %%mm1 \n" \
454 "movq 0x10(%1), %%mm2 \n" \
455 "movq 0x18(%1), %%mm3 \n" \
456 "movq %%mm0, (%0) \n" \
457 "movq %%mm1, 0x8(%0) \n" \
458 "movq %%mm2, 0x10(%0) \n" \
459 "movq %%mm3, 0x18(%0) \n" \
461 : "q" (dst), "r" (src) \
464 #define MOVE_10DWORDS_MMX(src,dst) \
466 "movq (%1), %%mm0 \n" \
467 "movq 0x8(%1), %%mm1 \n" \
468 "movq 0x10(%1), %%mm2 \n" \
469 "movq 0x18(%1), %%mm3 \n" \
470 "movq 0x20(%1), %%mm4 \n" \
471 "movq %%mm0, (%0) \n" \
472 "movq %%mm1, 0x8(%0) \n" \
473 "movq %%mm2, 0x10(%0) \n" \
474 "movq %%mm3, 0x18(%0) \n" \
475 "movq %%mm4, 0x20(%0) \n" \
477 : "q" (dst), "r" (src) \
480 #define MOVE_16DWORDS_MMX(src,dst) \
482 "movq (%1), %%mm0 \n" \
483 "movq 0x8(%1), %%mm1 \n" \
484 "movq 0x10(%1), %%mm2 \n" \
485 "movq 0x18(%1), %%mm3 \n" \
486 "movq 0x20(%1), %%mm4 \n" \
487 "movq 0x28(%1), %%mm5 \n" \
488 "movq 0x30(%1), %%mm6 \n" \
489 "movq 0x38(%1), %%mm7 \n" \
490 "movq %%mm0, (%0) \n" \
491 "movq %%mm1, 0x8(%0) \n" \
492 "movq %%mm2, 0x10(%0) \n" \
493 "movq %%mm3, 0x18(%0) \n" \
494 "movq %%mm4, 0x20(%0) \n" \
495 "movq %%mm5, 0x28(%0) \n" \
496 "movq %%mm6, 0x30(%0) \n" \
497 "movq %%mm7, 0x38(%0) \n" \
499 : "q" (dst), "r" (src) \
502 #define MOVE_16DWORDS_MMX2(src,dst) \
504 "movq (%1), %%mm0 \n" \
505 "movq 0x8(%1), %%mm1 \n" \
506 "movq 0x10(%1), %%mm2 \n" \
507 "movq 0x18(%1), %%mm3 \n" \
508 "movq 0x20(%1), %%mm4 \n" \
509 "movq 0x28(%1), %%mm5 \n" \
510 "movq 0x30(%1), %%mm6 \n" \
511 "movq 0x38(%1), %%mm7 \n" \
512 "movntq %%mm0, (%0) \n" \
513 "movntq %%mm1, 0x8(%0) \n" \
514 "movntq %%mm2, 0x10(%0) \n" \
515 "movntq %%mm3, 0x18(%0) \n" \
516 "movntq %%mm4, 0x20(%0) \n" \
517 "movntq %%mm5, 0x28(%0) \n" \
518 "movntq %%mm6, 0x30(%0) \n" \
519 "movntq %%mm7, 0x38(%0) \n" \
521 : "q" (dst), "r" (src) \
524 #define MOVE_32DWORDS_SSE2(src,dst) \
526 "movdqu (%1), %%xmm0 \n" \
527 "movdqu 0x10(%1), %%xmm1 \n" \
528 "movdqu 0x20(%1), %%xmm2 \n" \
529 "movdqu 0x30(%1), %%xmm3 \n" \
530 "movdqu 0x40(%1), %%xmm4 \n" \
531 "movdqu 0x50(%1), %%xmm5 \n" \
532 "movdqu 0x60(%1), %%xmm6 \n" \
533 "movdqu 0x70(%1), %%xmm7 \n" \
534 "movntdq %%xmm0, (%0) \n" \
535 "movntdq %%xmm1, 0x10(%0) \n" \
536 "movntdq %%xmm2, 0x20(%0) \n" \
537 "movntdq %%xmm3, 0x30(%0) \n" \
538 "movntdq %%xmm4, 0x40(%0) \n" \
539 "movntdq %%xmm5, 0x50(%0) \n" \
540 "movntdq %%xmm6, 0x60(%0) \n" \
541 "movntdq %%xmm7, 0x70(%0) \n" \
543 : "q" (dst), "r" (src) \
546 #define MOVE_32DWORDS_ALIGNED_SSE2(src,dst) \
548 "movdqa (%1), %%xmm0 \n" \
549 "movdqa 0x10(%1), %%xmm1 \n" \
550 "movdqa 0x20(%1), %%xmm2 \n" \
551 "movdqa 0x30(%1), %%xmm3 \n" \
552 "movdqa 0x40(%1), %%xmm4 \n" \
553 "movdqa 0x50(%1), %%xmm5 \n" \
554 "movdqa 0x60(%1), %%xmm6 \n" \
555 "movdqa 0x70(%1), %%xmm7 \n" \
556 "movntdq %%xmm0, (%0) \n" \
557 "movntdq %%xmm1, 0x10(%0) \n" \
558 "movntdq %%xmm2, 0x20(%0) \n" \
559 "movntdq %%xmm3, 0x30(%0) \n" \
560 "movntdq %%xmm4, 0x40(%0) \n" \
561 "movntdq %%xmm5, 0x50(%0) \n" \
562 "movntdq %%xmm6, 0x60(%0) \n" \
563 "movntdq %%xmm7, 0x70(%0) \n" \
565 : "q" (dst), "r" (src) \
569 (used to clean-up when going from mmx to float use
570 of the registers that are shared by both; note that
571 there is no float-to-mmx operation needed, because
572 only the float tag word info is corruptible)
575 #define emms() __asm__ __volatile__ ("emms":::"memory")
576 #define sfence() __asm__ __volatile__ ("sfence":::"memory")
578 /* additions to detect mmx - */
579 /* Raster <raster@rasterman.com> */
581 #define CPUID_MMX (1 << 23) /* flags: mmx */
582 #define CPUID_SSE (1 << 25) /* flags: xmm */
583 #define CPUID_SSE2 (1 << 26) /* flags: ? */
587 #define have_cpuid(cpuid_ret) \
588 __asm__ __volatile__ ( \
593 " movq %%rax, %%rbx \n" \
594 " xorq $0x200000, %%rax \n" \
599 " cmpq %%rax, %%rbx \n" \
610 #define get_cpuid(cpuid_ret) \
611 __asm__ __volatile__ ( \
614 " movl $1, %%eax \n" \
616 " test $0x00800000, %%edx\n" \
618 " movl %%edx, %0 \n" \
626 #define have_cpuid(cpuid_ret) \
627 __asm__ __volatile__ ( \
632 " movl %%eax, %%ebx \n" \
633 " xorl $0x200000, %%eax \n" \
638 " cmpl %%eax, %%ebx \n" \
649 #define get_cpuid(cpuid_ret) \
650 __asm__ __volatile__ ( \
653 " movl $1, %%eax \n" \
655 " test $0x00800000, %%edx\n" \
657 " movl %%edx, %0 \n" \
667 #define prefetch(var) \
668 __asm__ __volatile__ ( \
669 "prefetchnta (%0) \n" \
673 #define prefetch0(var) \
674 __asm__ __volatile__ ( \
675 "prefetcht0 (%0) \n" \
679 #define prefetch1(var) \
680 __asm__ __volatile__ ( \
681 "prefetcht1 (%0) \n" \
685 #define prefetch2(var) \
686 __asm__ __volatile__ ( \
687 "prefetcht2 (%0) \n" \
691 #define pshufw(r1, r2, imm) \
692 __asm__ __volatile__ ( \
693 "pshufw $" #imm ", %" #r1 ", %" #r2 " \n" \
696 #define pshufhw(r1, r2, imm) \
697 __asm__ __volatile__ ( \
698 "pshufhw $" #imm ", %" #r1 ", %" #r2 " \n" \
701 #define pshuflw(r1, r2, imm) \
702 __asm__ __volatile__ ( \
703 "pshuflw $" #imm ", %" #r1 ", %" #r2 " \n" \
705 #define pshufd(r1, r2, imm) \
706 __asm__ __volatile__ ( \
707 "pshufd $" #imm ", %" #r1 ", %" #r2 " \n" \
710 /* 1x238 MOVE Doouble Quadword
711 (this is both a load and a store...
712 in fact, it is the only way to store)
714 #define movdqu_m2r(var, reg) mmx_m2r(movdqu, var, reg)
715 #define movdqu_r2m(reg, var) mmx_r2m(movdqu, reg, var)
716 #define movdqu_r2r(regs, regd) mmx_r2r(movdqu, regs, regd)
717 #define movdqu(vars, vard) \
718 __asm__ __volatile__ ("movdqu %1, %%xmm0\n\t" \
719 "movdqu %%xmm0, %0" \
722 #define movdqa_m2r(var, reg) mmx_m2r(movdqa, var, reg)
723 #define movdqa_r2m(reg, var) mmx_r2m(movdqa, reg, var)
724 #define movdqa_r2r(regs, regd) mmx_r2r(movdqa, regs, regd)
725 #define movdqa(vars, vard) \
726 __asm__ __volatile__ ("movdqa %1, %%xmm0\n\t" \
727 "movdqa %%xmm0, %0" \
730 #define movntdq_r2m(reg, var) mmx_r2m(movntdq, reg, var)